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 case KMP_HW_UNKNOWN:
131 case KMP_HW_LAST:
132 return KMP_I18N_STR(Unknown);
133 }
134 KMP_ASSERT2(false, "Unhandled kmp_hw_t enumeration");
135 KMP_BUILTIN_UNREACHABLE;
136}
137
138const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural) {
139 switch (type) {
140 case KMP_HW_SOCKET:
141 return ((plural) ? "sockets" : "socket");
142 case KMP_HW_DIE:
143 return ((plural) ? "dice" : "die");
144 case KMP_HW_MODULE:
145 return ((plural) ? "modules" : "module");
146 case KMP_HW_TILE:
147 return ((plural) ? "tiles" : "tile");
148 case KMP_HW_NUMA:
149 return ((plural) ? "numa_domains" : "numa_domain");
150 case KMP_HW_L3:
151 return ((plural) ? "l3_caches" : "l3_cache");
152 case KMP_HW_L2:
153 return ((plural) ? "l2_caches" : "l2_cache");
154 case KMP_HW_L1:
155 return ((plural) ? "l1_caches" : "l1_cache");
156 case KMP_HW_LLC:
157 return ((plural) ? "ll_caches" : "ll_cache");
158 case KMP_HW_CORE:
159 return ((plural) ? "cores" : "core");
160 case KMP_HW_THREAD:
161 return ((plural) ? "threads" : "thread");
162 case KMP_HW_PROC_GROUP:
163 return ((plural) ? "proc_groups" : "proc_group");
164 case KMP_HW_UNKNOWN:
165 case KMP_HW_LAST:
166 return ((plural) ? "unknowns" : "unknown");
167 }
168 KMP_ASSERT2(false, "Unhandled kmp_hw_t enumeration");
169 KMP_BUILTIN_UNREACHABLE;
170}
171
172const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type) {
173 switch (type) {
174 case KMP_HW_CORE_TYPE_UNKNOWN:
175 case KMP_HW_MAX_NUM_CORE_TYPES:
176 return "unknown";
177#if KMP_ARCH_X86 || KMP_ARCH_X86_64
178 case KMP_HW_CORE_TYPE_ATOM:
179 return "Intel Atom(R) processor";
180 case KMP_HW_CORE_TYPE_CORE:
181 return "Intel(R) Core(TM) processor";
182#endif
183 }
184 KMP_ASSERT2(false, "Unhandled kmp_hw_core_type_t enumeration");
185 KMP_BUILTIN_UNREACHABLE;
186}
187
188#if KMP_AFFINITY_SUPPORTED
189// If affinity is supported, check the affinity
190// verbose and warning flags before printing warning
191#define KMP_AFF_WARNING(s, ...) \
192 if (s.flags.verbose || (s.flags.warnings && (s.type != affinity_none))) { \
193 KMP_WARNING(__VA_ARGS__); \
194 }
195#else
196#define KMP_AFF_WARNING(s, ...) KMP_WARNING(__VA_ARGS__)
197#endif
198
200// kmp_hw_thread_t methods
201int kmp_hw_thread_t::compare_ids(const void *a, const void *b) {
202 const kmp_hw_thread_t *ahwthread = (const kmp_hw_thread_t *)a;
203 const kmp_hw_thread_t *bhwthread = (const kmp_hw_thread_t *)b;
204 int depth = __kmp_topology->get_depth();
205 for (int level = 0; level < depth; ++level) {
206 // Reverse sort (higher efficiencies earlier in list) cores by core
207 // efficiency if available.
208 if (__kmp_is_hybrid_cpu() &&
209 __kmp_topology->get_type(level) == KMP_HW_CORE &&
210 ahwthread->attrs.is_core_eff_valid() &&
211 bhwthread->attrs.is_core_eff_valid()) {
212 if (ahwthread->attrs.get_core_eff() < bhwthread->attrs.get_core_eff())
213 return 1;
214 if (ahwthread->attrs.get_core_eff() > bhwthread->attrs.get_core_eff())
215 return -1;
216 }
217 if (ahwthread->ids[level] == bhwthread->ids[level])
218 continue;
219 // If the hardware id is unknown for this level, then place hardware thread
220 // further down in the sorted list as it should take last priority
221 if (ahwthread->ids[level] == UNKNOWN_ID)
222 return 1;
223 else if (bhwthread->ids[level] == UNKNOWN_ID)
224 return -1;
225 else if (ahwthread->ids[level] < bhwthread->ids[level])
226 return -1;
227 else if (ahwthread->ids[level] > bhwthread->ids[level])
228 return 1;
229 }
230 if (ahwthread->os_id < bhwthread->os_id)
231 return -1;
232 else if (ahwthread->os_id > bhwthread->os_id)
233 return 1;
234 return 0;
235}
236
237#if KMP_AFFINITY_SUPPORTED
238int kmp_hw_thread_t::compare_compact(const void *a, const void *b) {
239 int i;
240 const kmp_hw_thread_t *aa = (const kmp_hw_thread_t *)a;
241 const kmp_hw_thread_t *bb = (const kmp_hw_thread_t *)b;
242 int depth = __kmp_topology->get_depth();
243 int compact = __kmp_topology->compact;
244 KMP_DEBUG_ASSERT(compact >= 0);
245 KMP_DEBUG_ASSERT(compact <= depth);
246 for (i = 0; i < compact; i++) {
247 int j = depth - i - 1;
248 if (aa->sub_ids[j] < bb->sub_ids[j])
249 return -1;
250 if (aa->sub_ids[j] > bb->sub_ids[j])
251 return 1;
252 }
253 for (; i < depth; i++) {
254 int j = i - compact;
255 if (aa->sub_ids[j] < bb->sub_ids[j])
256 return -1;
257 if (aa->sub_ids[j] > bb->sub_ids[j])
258 return 1;
259 }
260 return 0;
261}
262#endif
263
264void kmp_hw_thread_t::print() const {
265 int depth = __kmp_topology->get_depth();
266 printf("%4d ", os_id);
267 for (int i = 0; i < depth; ++i) {
268 printf("%4d (%d) ", ids[i], sub_ids[i]);
269 }
270 if (attrs) {
271 if (attrs.is_core_type_valid())
272 printf(" (%s)", __kmp_hw_get_core_type_string(attrs.get_core_type()));
273 if (attrs.is_core_eff_valid())
274 printf(" (eff=%d)", attrs.get_core_eff());
275 }
276 if (leader)
277 printf(" (leader)");
278 printf("\n");
279}
280
282// kmp_topology_t methods
283
284// Add a layer to the topology based on the ids. Assume the topology
285// is perfectly nested (i.e., so no object has more than one parent)
286void kmp_topology_t::insert_layer(kmp_hw_t type, const int *ids) {
287 // Figure out where the layer should go by comparing the ids of the current
288 // layers with the new ids
289 int target_layer;
290 int previous_id = kmp_hw_thread_t::UNKNOWN_ID;
291 int previous_new_id = kmp_hw_thread_t::UNKNOWN_ID;
292
293 // Start from the highest layer and work down to find target layer
294 // If new layer is equal to another layer then put the new layer above
295 for (target_layer = 0; target_layer < depth; ++target_layer) {
296 bool layers_equal = true;
297 bool strictly_above_target_layer = false;
298 for (int i = 0; i < num_hw_threads; ++i) {
299 int id = hw_threads[i].ids[target_layer];
300 int new_id = ids[i];
301 if (id != previous_id && new_id == previous_new_id) {
302 // Found the layer we are strictly above
303 strictly_above_target_layer = true;
304 layers_equal = false;
305 break;
306 } else if (id == previous_id && new_id != previous_new_id) {
307 // Found a layer we are below. Move to next layer and check.
308 layers_equal = false;
309 break;
310 }
311 previous_id = id;
312 previous_new_id = new_id;
313 }
314 if (strictly_above_target_layer || layers_equal)
315 break;
316 }
317
318 // Found the layer we are above. Now move everything to accommodate the new
319 // layer. And put the new ids and type into the topology.
320 for (int i = depth - 1, j = depth; i >= target_layer; --i, --j)
321 types[j] = types[i];
322 types[target_layer] = type;
323 for (int k = 0; k < num_hw_threads; ++k) {
324 for (int i = depth - 1, j = depth; i >= target_layer; --i, --j)
325 hw_threads[k].ids[j] = hw_threads[k].ids[i];
326 hw_threads[k].ids[target_layer] = ids[k];
327 }
328 equivalent[type] = type;
329 depth++;
330}
331
332#if KMP_GROUP_AFFINITY
333// Insert the Windows Processor Group structure into the topology
334void kmp_topology_t::_insert_windows_proc_groups() {
335 // Do not insert the processor group structure for a single group
336 if (__kmp_num_proc_groups == 1)
337 return;
338 kmp_affin_mask_t *mask;
339 int *ids = (int *)__kmp_allocate(sizeof(int) * num_hw_threads);
340 KMP_CPU_ALLOC(mask);
341 for (int i = 0; i < num_hw_threads; ++i) {
342 KMP_CPU_ZERO(mask);
343 KMP_CPU_SET(hw_threads[i].os_id, mask);
344 ids[i] = __kmp_get_proc_group(mask);
345 }
346 KMP_CPU_FREE(mask);
347 insert_layer(KMP_HW_PROC_GROUP, ids);
348 __kmp_free(ids);
349
350 // sort topology after adding proc groups
351 __kmp_topology->sort_ids();
352}
353#endif
354
355// Remove layers that don't add information to the topology.
356// This is done by having the layer take on the id = UNKNOWN_ID (-1)
357void kmp_topology_t::_remove_radix1_layers() {
358 int preference[KMP_HW_LAST];
359 int top_index1, top_index2;
360 // Set up preference associative array
361 preference[KMP_HW_SOCKET] = 110;
362 preference[KMP_HW_PROC_GROUP] = 100;
363 preference[KMP_HW_CORE] = 95;
364 preference[KMP_HW_THREAD] = 90;
365 preference[KMP_HW_NUMA] = 85;
366 preference[KMP_HW_DIE] = 80;
367 preference[KMP_HW_TILE] = 75;
368 preference[KMP_HW_MODULE] = 73;
369 preference[KMP_HW_L3] = 70;
370 preference[KMP_HW_L2] = 65;
371 preference[KMP_HW_L1] = 60;
372 preference[KMP_HW_LLC] = 5;
373 top_index1 = 0;
374 top_index2 = 1;
375 while (top_index1 < depth - 1 && top_index2 < depth) {
376 kmp_hw_t type1 = types[top_index1];
377 kmp_hw_t type2 = types[top_index2];
378 KMP_ASSERT_VALID_HW_TYPE(type1);
379 KMP_ASSERT_VALID_HW_TYPE(type2);
380 // Do not allow the three main topology levels (sockets, cores, threads) to
381 // be compacted down
382 if ((type1 == KMP_HW_THREAD || type1 == KMP_HW_CORE ||
383 type1 == KMP_HW_SOCKET) &&
384 (type2 == KMP_HW_THREAD || type2 == KMP_HW_CORE ||
385 type2 == KMP_HW_SOCKET)) {
386 top_index1 = top_index2++;
387 continue;
388 }
389 bool radix1 = true;
390 bool all_same = true;
391 int id1 = hw_threads[0].ids[top_index1];
392 int id2 = hw_threads[0].ids[top_index2];
393 int pref1 = preference[type1];
394 int pref2 = preference[type2];
395 for (int hwidx = 1; hwidx < num_hw_threads; ++hwidx) {
396 if (hw_threads[hwidx].ids[top_index1] == id1 &&
397 hw_threads[hwidx].ids[top_index2] != id2) {
398 radix1 = false;
399 break;
400 }
401 if (hw_threads[hwidx].ids[top_index2] != id2)
402 all_same = false;
403 id1 = hw_threads[hwidx].ids[top_index1];
404 id2 = hw_threads[hwidx].ids[top_index2];
405 }
406 if (radix1) {
407 // Select the layer to remove based on preference
408 kmp_hw_t remove_type, keep_type;
409 int remove_layer, remove_layer_ids;
410 if (pref1 > pref2) {
411 remove_type = type2;
412 remove_layer = remove_layer_ids = top_index2;
413 keep_type = type1;
414 } else {
415 remove_type = type1;
416 remove_layer = remove_layer_ids = top_index1;
417 keep_type = type2;
418 }
419 // If all the indexes for the second (deeper) layer are the same.
420 // e.g., all are zero, then make sure to keep the first layer's ids
421 if (all_same)
422 remove_layer_ids = top_index2;
423 // Remove radix one type by setting the equivalence, removing the id from
424 // the hw threads and removing the layer from types and depth
425 set_equivalent_type(remove_type, keep_type);
426 for (int idx = 0; idx < num_hw_threads; ++idx) {
427 kmp_hw_thread_t &hw_thread = hw_threads[idx];
428 for (int d = remove_layer_ids; d < depth - 1; ++d)
429 hw_thread.ids[d] = hw_thread.ids[d + 1];
430 }
431 for (int idx = remove_layer; idx < depth - 1; ++idx)
432 types[idx] = types[idx + 1];
433 depth--;
434 } else {
435 top_index1 = top_index2++;
436 }
437 }
438 KMP_ASSERT(depth > 0);
439}
440
441void kmp_topology_t::_set_last_level_cache() {
442 if (get_equivalent_type(KMP_HW_L3) != KMP_HW_UNKNOWN)
443 set_equivalent_type(KMP_HW_LLC, KMP_HW_L3);
444 else if (get_equivalent_type(KMP_HW_L2) != KMP_HW_UNKNOWN)
445 set_equivalent_type(KMP_HW_LLC, KMP_HW_L2);
446#if KMP_MIC_SUPPORTED
447 else if (__kmp_mic_type == mic3) {
448 if (get_equivalent_type(KMP_HW_L2) != KMP_HW_UNKNOWN)
449 set_equivalent_type(KMP_HW_LLC, KMP_HW_L2);
450 else if (get_equivalent_type(KMP_HW_TILE) != KMP_HW_UNKNOWN)
451 set_equivalent_type(KMP_HW_LLC, KMP_HW_TILE);
452 // L2/Tile wasn't detected so just say L1
453 else
454 set_equivalent_type(KMP_HW_LLC, KMP_HW_L1);
455 }
456#endif
457 else if (get_equivalent_type(KMP_HW_L1) != KMP_HW_UNKNOWN)
458 set_equivalent_type(KMP_HW_LLC, KMP_HW_L1);
459 // Fallback is to set last level cache to socket or core
460 if (get_equivalent_type(KMP_HW_LLC) == KMP_HW_UNKNOWN) {
461 if (get_equivalent_type(KMP_HW_SOCKET) != KMP_HW_UNKNOWN)
462 set_equivalent_type(KMP_HW_LLC, KMP_HW_SOCKET);
463 else if (get_equivalent_type(KMP_HW_CORE) != KMP_HW_UNKNOWN)
464 set_equivalent_type(KMP_HW_LLC, KMP_HW_CORE);
465 }
466 KMP_ASSERT(get_equivalent_type(KMP_HW_LLC) != KMP_HW_UNKNOWN);
467}
468
469// Gather the count of each topology layer and the ratio
470void kmp_topology_t::_gather_enumeration_information() {
471 int previous_id[KMP_HW_LAST];
472 int max[KMP_HW_LAST];
473
474 for (int i = 0; i < depth; ++i) {
475 previous_id[i] = kmp_hw_thread_t::UNKNOWN_ID;
476 max[i] = 0;
477 count[i] = 0;
478 ratio[i] = 0;
479 }
480 int core_level = get_level(KMP_HW_CORE);
481 for (int i = 0; i < num_hw_threads; ++i) {
482 kmp_hw_thread_t &hw_thread = hw_threads[i];
483 for (int layer = 0; layer < depth; ++layer) {
484 int id = hw_thread.ids[layer];
485 if (id != previous_id[layer]) {
486 // Add an additional increment to each count
487 for (int l = layer; l < depth; ++l) {
488 if (hw_thread.ids[l] != kmp_hw_thread_t::UNKNOWN_ID)
489 count[l]++;
490 }
491 // Keep track of topology layer ratio statistics
492 if (hw_thread.ids[layer] != kmp_hw_thread_t::UNKNOWN_ID)
493 max[layer]++;
494 for (int l = layer + 1; l < depth; ++l) {
495 if (max[l] > ratio[l])
496 ratio[l] = max[l];
497 max[l] = 1;
498 }
499 // Figure out the number of different core types
500 // and efficiencies for hybrid CPUs
501 if (__kmp_is_hybrid_cpu() && core_level >= 0 && layer <= core_level) {
502 if (hw_thread.attrs.is_core_eff_valid() &&
503 hw_thread.attrs.core_eff >= num_core_efficiencies) {
504 // Because efficiencies can range from 0 to max efficiency - 1,
505 // the number of efficiencies is max efficiency + 1
506 num_core_efficiencies = hw_thread.attrs.core_eff + 1;
507 }
508 if (hw_thread.attrs.is_core_type_valid()) {
509 bool found = false;
510 for (int j = 0; j < num_core_types; ++j) {
511 if (hw_thread.attrs.get_core_type() == core_types[j]) {
512 found = true;
513 break;
514 }
515 }
516 if (!found) {
517 KMP_ASSERT(num_core_types < KMP_HW_MAX_NUM_CORE_TYPES);
518 core_types[num_core_types++] = hw_thread.attrs.get_core_type();
519 }
520 }
521 }
522 break;
523 }
524 }
525 for (int layer = 0; layer < depth; ++layer) {
526 previous_id[layer] = hw_thread.ids[layer];
527 }
528 }
529 for (int layer = 0; layer < depth; ++layer) {
530 if (max[layer] > ratio[layer])
531 ratio[layer] = max[layer];
532 }
533}
534
535int kmp_topology_t::_get_ncores_with_attr(const kmp_hw_attr_t &attr,
536 int above_level,
537 bool find_all) const {
538 int current, current_max;
539 int previous_id[KMP_HW_LAST];
540 for (int i = 0; i < depth; ++i)
541 previous_id[i] = kmp_hw_thread_t::UNKNOWN_ID;
542 int core_level = get_level(KMP_HW_CORE);
543 if (find_all)
544 above_level = -1;
545 KMP_ASSERT(above_level < core_level);
546 current_max = 0;
547 current = 0;
548 for (int i = 0; i < num_hw_threads; ++i) {
549 kmp_hw_thread_t &hw_thread = hw_threads[i];
550 if (!find_all && hw_thread.ids[above_level] != previous_id[above_level]) {
551 if (current > current_max)
552 current_max = current;
553 current = hw_thread.attrs.contains(attr);
554 } else {
555 for (int level = above_level + 1; level <= core_level; ++level) {
556 if (hw_thread.ids[level] != previous_id[level]) {
557 if (hw_thread.attrs.contains(attr))
558 current++;
559 break;
560 }
561 }
562 }
563 for (int level = 0; level < depth; ++level)
564 previous_id[level] = hw_thread.ids[level];
565 }
566 if (current > current_max)
567 current_max = current;
568 return current_max;
569}
570
571// Find out if the topology is uniform
572void kmp_topology_t::_discover_uniformity() {
573 int num = 1;
574 for (int level = 0; level < depth; ++level)
575 num *= ratio[level];
576 flags.uniform = (num == count[depth - 1]);
577}
578
579// Set all the sub_ids for each hardware thread
580void kmp_topology_t::_set_sub_ids() {
581 int previous_id[KMP_HW_LAST];
582 int sub_id[KMP_HW_LAST];
583
584 for (int i = 0; i < depth; ++i) {
585 previous_id[i] = -1;
586 sub_id[i] = -1;
587 }
588 for (int i = 0; i < num_hw_threads; ++i) {
589 kmp_hw_thread_t &hw_thread = hw_threads[i];
590 // Setup the sub_id
591 for (int j = 0; j < depth; ++j) {
592 if (hw_thread.ids[j] != previous_id[j]) {
593 sub_id[j]++;
594 for (int k = j + 1; k < depth; ++k) {
595 sub_id[k] = 0;
596 }
597 break;
598 }
599 }
600 // Set previous_id
601 for (int j = 0; j < depth; ++j) {
602 previous_id[j] = hw_thread.ids[j];
603 }
604 // Set the sub_ids field
605 for (int j = 0; j < depth; ++j) {
606 hw_thread.sub_ids[j] = sub_id[j];
607 }
608 }
609}
610
611void kmp_topology_t::_set_globals() {
612 // Set nCoresPerPkg, nPackages, __kmp_nThreadsPerCore, __kmp_ncores
613 int core_level, thread_level, package_level;
614 package_level = get_level(KMP_HW_SOCKET);
615#if KMP_GROUP_AFFINITY
616 if (package_level == -1)
617 package_level = get_level(KMP_HW_PROC_GROUP);
618#endif
619 core_level = get_level(KMP_HW_CORE);
620 thread_level = get_level(KMP_HW_THREAD);
621
622 KMP_ASSERT(core_level != -1);
623 KMP_ASSERT(thread_level != -1);
624
625 __kmp_nThreadsPerCore = calculate_ratio(thread_level, core_level);
626 if (package_level != -1) {
627 nCoresPerPkg = calculate_ratio(core_level, package_level);
628 nPackages = get_count(package_level);
629 } else {
630 // assume one socket
631 nCoresPerPkg = get_count(core_level);
632 nPackages = 1;
633 }
634#ifndef KMP_DFLT_NTH_CORES
635 __kmp_ncores = get_count(core_level);
636#endif
637}
638
639kmp_topology_t *kmp_topology_t::allocate(int nproc, int ndepth,
640 const kmp_hw_t *types) {
641 kmp_topology_t *retval;
642 // Allocate all data in one large allocation
643 size_t size = sizeof(kmp_topology_t) + sizeof(kmp_hw_thread_t) * nproc +
644 sizeof(int) * (size_t)KMP_HW_LAST * 3;
645 char *bytes = (char *)__kmp_allocate(size);
646 retval = (kmp_topology_t *)bytes;
647 if (nproc > 0) {
648 retval->hw_threads = (kmp_hw_thread_t *)(bytes + sizeof(kmp_topology_t));
649 } else {
650 retval->hw_threads = nullptr;
651 }
652 retval->num_hw_threads = nproc;
653 retval->depth = ndepth;
654 int *arr =
655 (int *)(bytes + sizeof(kmp_topology_t) + sizeof(kmp_hw_thread_t) * nproc);
656 retval->types = (kmp_hw_t *)arr;
657 retval->ratio = arr + (size_t)KMP_HW_LAST;
658 retval->count = arr + 2 * (size_t)KMP_HW_LAST;
659 retval->num_core_efficiencies = 0;
660 retval->num_core_types = 0;
661 retval->compact = 0;
662 for (int i = 0; i < KMP_HW_MAX_NUM_CORE_TYPES; ++i)
663 retval->core_types[i] = KMP_HW_CORE_TYPE_UNKNOWN;
664 KMP_FOREACH_HW_TYPE(type) { retval->equivalent[type] = KMP_HW_UNKNOWN; }
665 for (int i = 0; i < ndepth; ++i) {
666 retval->types[i] = types[i];
667 retval->equivalent[types[i]] = types[i];
668 }
669 return retval;
670}
671
672void kmp_topology_t::deallocate(kmp_topology_t *topology) {
673 if (topology)
674 __kmp_free(topology);
675}
676
677bool kmp_topology_t::check_ids() const {
678 // Assume ids have been sorted
679 if (num_hw_threads == 0)
680 return true;
681 for (int i = 1; i < num_hw_threads; ++i) {
682 kmp_hw_thread_t &current_thread = hw_threads[i];
683 kmp_hw_thread_t &previous_thread = hw_threads[i - 1];
684 bool unique = false;
685 for (int j = 0; j < depth; ++j) {
686 if (previous_thread.ids[j] != current_thread.ids[j]) {
687 unique = true;
688 break;
689 }
690 }
691 if (unique)
692 continue;
693 return false;
694 }
695 return true;
696}
697
698void kmp_topology_t::dump() const {
699 printf("***********************\n");
700 printf("*** __kmp_topology: ***\n");
701 printf("***********************\n");
702 printf("* depth: %d\n", depth);
703
704 printf("* types: ");
705 for (int i = 0; i < depth; ++i)
706 printf("%15s ", __kmp_hw_get_keyword(types[i]));
707 printf("\n");
708
709 printf("* ratio: ");
710 for (int i = 0; i < depth; ++i) {
711 printf("%15d ", ratio[i]);
712 }
713 printf("\n");
714
715 printf("* count: ");
716 for (int i = 0; i < depth; ++i) {
717 printf("%15d ", count[i]);
718 }
719 printf("\n");
720
721 printf("* num_core_eff: %d\n", num_core_efficiencies);
722 printf("* num_core_types: %d\n", num_core_types);
723 printf("* core_types: ");
724 for (int i = 0; i < num_core_types; ++i)
725 printf("%3d ", core_types[i]);
726 printf("\n");
727
728 printf("* equivalent map:\n");
729 KMP_FOREACH_HW_TYPE(i) {
730 const char *key = __kmp_hw_get_keyword(i);
731 const char *value = __kmp_hw_get_keyword(equivalent[i]);
732 printf("%-15s -> %-15s\n", key, value);
733 }
734
735 printf("* uniform: %s\n", (is_uniform() ? "Yes" : "No"));
736
737 printf("* num_hw_threads: %d\n", num_hw_threads);
738 printf("* hw_threads:\n");
739 for (int i = 0; i < num_hw_threads; ++i) {
740 hw_threads[i].print();
741 }
742 printf("***********************\n");
743}
744
745void kmp_topology_t::print(const char *env_var) const {
746 kmp_str_buf_t buf;
747 int print_types_depth;
748 __kmp_str_buf_init(&buf);
749 kmp_hw_t print_types[KMP_HW_LAST + 2];
750
751 // Num Available Threads
752 if (num_hw_threads) {
753 KMP_INFORM(AvailableOSProc, env_var, num_hw_threads);
754 } else {
755 KMP_INFORM(AvailableOSProc, env_var, __kmp_xproc);
756 }
757
758 // Uniform or not
759 if (is_uniform()) {
760 KMP_INFORM(Uniform, env_var);
761 } else {
762 KMP_INFORM(NonUniform, env_var);
763 }
764
765 // Equivalent types
766 KMP_FOREACH_HW_TYPE(type) {
767 kmp_hw_t eq_type = equivalent[type];
768 if (eq_type != KMP_HW_UNKNOWN && eq_type != type) {
769 KMP_INFORM(AffEqualTopologyTypes, env_var,
770 __kmp_hw_get_catalog_string(type),
771 __kmp_hw_get_catalog_string(eq_type));
772 }
773 }
774
775 // Quick topology
776 KMP_ASSERT(depth > 0 && depth <= (int)KMP_HW_LAST);
777 // Create a print types array that always guarantees printing
778 // the core and thread level
779 print_types_depth = 0;
780 for (int level = 0; level < depth; ++level)
781 print_types[print_types_depth++] = types[level];
782 if (equivalent[KMP_HW_CORE] != KMP_HW_CORE) {
783 // Force in the core level for quick topology
784 if (print_types[print_types_depth - 1] == KMP_HW_THREAD) {
785 // Force core before thread e.g., 1 socket X 2 threads/socket
786 // becomes 1 socket X 1 core/socket X 2 threads/socket
787 print_types[print_types_depth - 1] = KMP_HW_CORE;
788 print_types[print_types_depth++] = KMP_HW_THREAD;
789 } else {
790 print_types[print_types_depth++] = KMP_HW_CORE;
791 }
792 }
793 // Always put threads at very end of quick topology
794 if (equivalent[KMP_HW_THREAD] != KMP_HW_THREAD)
795 print_types[print_types_depth++] = KMP_HW_THREAD;
796
797 __kmp_str_buf_clear(&buf);
798 kmp_hw_t numerator_type;
799 kmp_hw_t denominator_type = KMP_HW_UNKNOWN;
800 int core_level = get_level(KMP_HW_CORE);
801 int ncores = get_count(core_level);
802
803 for (int plevel = 0, level = 0; plevel < print_types_depth; ++plevel) {
804 int c;
805 bool plural;
806 numerator_type = print_types[plevel];
807 KMP_ASSERT_VALID_HW_TYPE(numerator_type);
808 if (equivalent[numerator_type] != numerator_type)
809 c = 1;
810 else
811 c = get_ratio(level++);
812 plural = (c > 1);
813 if (plevel == 0) {
814 __kmp_str_buf_print(&buf, "%d %s", c,
815 __kmp_hw_get_catalog_string(numerator_type, plural));
816 } else {
817 __kmp_str_buf_print(&buf, " x %d %s/%s", c,
818 __kmp_hw_get_catalog_string(numerator_type, plural),
819 __kmp_hw_get_catalog_string(denominator_type));
820 }
821 denominator_type = numerator_type;
822 }
823 KMP_INFORM(TopologyGeneric, env_var, buf.str, ncores);
824
825 // Hybrid topology information
826 if (__kmp_is_hybrid_cpu()) {
827 for (int i = 0; i < num_core_types; ++i) {
828 kmp_hw_core_type_t core_type = core_types[i];
829 kmp_hw_attr_t attr;
830 attr.clear();
831 attr.set_core_type(core_type);
832 int ncores = get_ncores_with_attr(attr);
833 if (ncores > 0) {
834 KMP_INFORM(TopologyHybrid, env_var, ncores,
835 __kmp_hw_get_core_type_string(core_type));
836 KMP_ASSERT(num_core_efficiencies <= KMP_HW_MAX_NUM_CORE_EFFS)
837 for (int eff = 0; eff < num_core_efficiencies; ++eff) {
838 attr.set_core_eff(eff);
839 int ncores_with_eff = get_ncores_with_attr(attr);
840 if (ncores_with_eff > 0) {
841 KMP_INFORM(TopologyHybridCoreEff, env_var, ncores_with_eff, eff);
842 }
843 }
844 }
845 }
846 }
847
848 if (num_hw_threads <= 0) {
849 __kmp_str_buf_free(&buf);
850 return;
851 }
852
853 // Full OS proc to hardware thread map
854 KMP_INFORM(OSProcToPhysicalThreadMap, env_var);
855 for (int i = 0; i < num_hw_threads; i++) {
856 __kmp_str_buf_clear(&buf);
857 for (int level = 0; level < depth; ++level) {
858 if (hw_threads[i].ids[level] == kmp_hw_thread_t::UNKNOWN_ID)
859 continue;
860 kmp_hw_t type = types[level];
861 __kmp_str_buf_print(&buf, "%s ", __kmp_hw_get_catalog_string(type));
862 __kmp_str_buf_print(&buf, "%d ", hw_threads[i].ids[level]);
863 }
864 if (__kmp_is_hybrid_cpu())
865 __kmp_str_buf_print(
866 &buf, "(%s)",
867 __kmp_hw_get_core_type_string(hw_threads[i].attrs.get_core_type()));
868 KMP_INFORM(OSProcMapToPack, env_var, hw_threads[i].os_id, buf.str);
869 }
870
871 __kmp_str_buf_free(&buf);
872}
873
874#if KMP_AFFINITY_SUPPORTED
875void kmp_topology_t::set_granularity(kmp_affinity_t &affinity) const {
876 const char *env_var = __kmp_get_affinity_env_var(affinity);
877 // If requested hybrid CPU attributes for granularity (either OMP_PLACES or
878 // KMP_AFFINITY), but none exist, then reset granularity and have below method
879 // select a granularity and warn user.
880 if (!__kmp_is_hybrid_cpu()) {
881 if (affinity.core_attr_gran.valid) {
882 // OMP_PLACES with cores:<attribute> but non-hybrid arch, use cores
883 // instead
884 KMP_AFF_WARNING(
885 affinity, AffIgnoringNonHybrid, env_var,
886 __kmp_hw_get_catalog_string(KMP_HW_CORE, /*plural=*/true));
887 affinity.gran = KMP_HW_CORE;
888 affinity.gran_levels = -1;
889 affinity.core_attr_gran = KMP_AFFINITY_ATTRS_UNKNOWN;
890 affinity.flags.core_types_gran = affinity.flags.core_effs_gran = 0;
891 } else if (affinity.flags.core_types_gran ||
892 affinity.flags.core_effs_gran) {
893 // OMP_PLACES=core_types|core_effs but non-hybrid, use cores instead
894 if (affinity.flags.omp_places) {
895 KMP_AFF_WARNING(
896 affinity, AffIgnoringNonHybrid, env_var,
897 __kmp_hw_get_catalog_string(KMP_HW_CORE, /*plural=*/true));
898 } else {
899 // KMP_AFFINITY=granularity=core_type|core_eff,...
900 KMP_AFF_WARNING(affinity, AffGranularityBad, env_var,
901 "Intel(R) Hybrid Technology core attribute",
902 __kmp_hw_get_catalog_string(KMP_HW_CORE));
903 }
904 affinity.gran = KMP_HW_CORE;
905 affinity.gran_levels = -1;
906 affinity.core_attr_gran = KMP_AFFINITY_ATTRS_UNKNOWN;
907 affinity.flags.core_types_gran = affinity.flags.core_effs_gran = 0;
908 }
909 }
910 // Set the number of affinity granularity levels
911 if (affinity.gran_levels < 0) {
912 kmp_hw_t gran_type = get_equivalent_type(affinity.gran);
913 // Check if user's granularity request is valid
914 if (gran_type == KMP_HW_UNKNOWN) {
915 // First try core, then thread, then package
916 kmp_hw_t gran_types[3] = {KMP_HW_CORE, KMP_HW_THREAD, KMP_HW_SOCKET};
917 for (auto g : gran_types) {
918 if (get_equivalent_type(g) != KMP_HW_UNKNOWN) {
919 gran_type = g;
920 break;
921 }
922 }
923 KMP_ASSERT(gran_type != KMP_HW_UNKNOWN);
924 // Warn user what granularity setting will be used instead
925 KMP_AFF_WARNING(affinity, AffGranularityBad, env_var,
926 __kmp_hw_get_catalog_string(affinity.gran),
927 __kmp_hw_get_catalog_string(gran_type));
928 affinity.gran = gran_type;
929 }
930#if KMP_GROUP_AFFINITY
931 // If more than one processor group exists, and the level of
932 // granularity specified by the user is too coarse, then the
933 // granularity must be adjusted "down" to processor group affinity
934 // because threads can only exist within one processor group.
935 // For example, if a user sets granularity=socket and there are two
936 // processor groups that cover a socket, then the runtime must
937 // restrict the granularity down to the processor group level.
938 if (__kmp_num_proc_groups > 1) {
939 int gran_depth = get_level(gran_type);
940 int proc_group_depth = get_level(KMP_HW_PROC_GROUP);
941 if (gran_depth >= 0 && proc_group_depth >= 0 &&
942 gran_depth < proc_group_depth) {
943 KMP_AFF_WARNING(affinity, AffGranTooCoarseProcGroup, env_var,
944 __kmp_hw_get_catalog_string(affinity.gran));
945 affinity.gran = gran_type = KMP_HW_PROC_GROUP;
946 }
947 }
948#endif
949 affinity.gran_levels = 0;
950 for (int i = depth - 1; i >= 0 && get_type(i) != gran_type; --i)
951 affinity.gran_levels++;
952 }
953}
954#endif
955
956void kmp_topology_t::canonicalize() {
957#if KMP_GROUP_AFFINITY
958 _insert_windows_proc_groups();
959#endif
960 _remove_radix1_layers();
961 _gather_enumeration_information();
962 _discover_uniformity();
963 _set_sub_ids();
964 _set_globals();
965 _set_last_level_cache();
966
967#if KMP_MIC_SUPPORTED
968 // Manually Add L2 = Tile equivalence
969 if (__kmp_mic_type == mic3) {
970 if (get_level(KMP_HW_L2) != -1)
971 set_equivalent_type(KMP_HW_TILE, KMP_HW_L2);
972 else if (get_level(KMP_HW_TILE) != -1)
973 set_equivalent_type(KMP_HW_L2, KMP_HW_TILE);
974 }
975#endif
976
977 // Perform post canonicalization checking
978 KMP_ASSERT(depth > 0);
979 for (int level = 0; level < depth; ++level) {
980 // All counts, ratios, and types must be valid
981 KMP_ASSERT(count[level] > 0 && ratio[level] > 0);
982 KMP_ASSERT_VALID_HW_TYPE(types[level]);
983 // Detected types must point to themselves
984 KMP_ASSERT(equivalent[types[level]] == types[level]);
985 }
986}
987
988// Canonicalize an explicit packages X cores/pkg X threads/core topology
989void kmp_topology_t::canonicalize(int npackages, int ncores_per_pkg,
990 int nthreads_per_core, int ncores) {
991 int ndepth = 3;
992 depth = ndepth;
993 KMP_FOREACH_HW_TYPE(i) { equivalent[i] = KMP_HW_UNKNOWN; }
994 for (int level = 0; level < depth; ++level) {
995 count[level] = 0;
996 ratio[level] = 0;
997 }
998 count[0] = npackages;
999 count[1] = ncores;
1000 count[2] = __kmp_xproc;
1001 ratio[0] = npackages;
1002 ratio[1] = ncores_per_pkg;
1003 ratio[2] = nthreads_per_core;
1004 equivalent[KMP_HW_SOCKET] = KMP_HW_SOCKET;
1005 equivalent[KMP_HW_CORE] = KMP_HW_CORE;
1006 equivalent[KMP_HW_THREAD] = KMP_HW_THREAD;
1007 types[0] = KMP_HW_SOCKET;
1008 types[1] = KMP_HW_CORE;
1009 types[2] = KMP_HW_THREAD;
1010 //__kmp_avail_proc = __kmp_xproc;
1011 _discover_uniformity();
1012}
1013
1014#if KMP_AFFINITY_SUPPORTED
1015static kmp_str_buf_t *
1016__kmp_hw_get_catalog_core_string(const kmp_hw_attr_t &attr, kmp_str_buf_t *buf,
1017 bool plural) {
1018 __kmp_str_buf_init(buf);
1019 if (attr.is_core_type_valid())
1020 __kmp_str_buf_print(buf, "%s %s",
1021 __kmp_hw_get_core_type_string(attr.get_core_type()),
1022 __kmp_hw_get_catalog_string(KMP_HW_CORE, plural));
1023 else
1024 __kmp_str_buf_print(buf, "%s eff=%d",
1025 __kmp_hw_get_catalog_string(KMP_HW_CORE, plural),
1026 attr.get_core_eff());
1027 return buf;
1028}
1029
1030bool kmp_topology_t::restrict_to_mask(const kmp_affin_mask_t *mask) {
1031 // Apply the filter
1032 bool affected;
1033 int new_index = 0;
1034 for (int i = 0; i < num_hw_threads; ++i) {
1035 int os_id = hw_threads[i].os_id;
1036 if (KMP_CPU_ISSET(os_id, mask)) {
1037 if (i != new_index)
1038 hw_threads[new_index] = hw_threads[i];
1039 new_index++;
1040 } else {
1041 KMP_CPU_CLR(os_id, __kmp_affin_fullMask);
1042 __kmp_avail_proc--;
1043 }
1044 }
1045
1046 KMP_DEBUG_ASSERT(new_index <= num_hw_threads);
1047 affected = (num_hw_threads != new_index);
1048 num_hw_threads = new_index;
1049
1050 // Post hardware subset canonicalization
1051 if (affected) {
1052 _gather_enumeration_information();
1053 _discover_uniformity();
1054 _set_globals();
1055 _set_last_level_cache();
1056#if KMP_OS_WINDOWS
1057 // Copy filtered full mask if topology has single processor group
1058 if (__kmp_num_proc_groups <= 1)
1059#endif
1060 __kmp_affin_origMask->copy(__kmp_affin_fullMask);
1061 }
1062 return affected;
1063}
1064
1065// Apply the KMP_HW_SUBSET envirable to the topology
1066// Returns true if KMP_HW_SUBSET filtered any processors
1067// otherwise, returns false
1068bool kmp_topology_t::filter_hw_subset() {
1069 // If KMP_HW_SUBSET wasn't requested, then do nothing.
1070 if (!__kmp_hw_subset)
1071 return false;
1072
1073 // First, sort the KMP_HW_SUBSET items by the machine topology
1074 __kmp_hw_subset->sort();
1075
1076 __kmp_hw_subset->canonicalize(__kmp_topology);
1077
1078 // Check to see if KMP_HW_SUBSET is a valid subset of the detected topology
1079 bool using_core_types = false;
1080 bool using_core_effs = false;
1081 bool is_absolute = __kmp_hw_subset->is_absolute();
1082 int hw_subset_depth = __kmp_hw_subset->get_depth();
1083 kmp_hw_t specified[KMP_HW_LAST];
1084 int *topology_levels = (int *)KMP_ALLOCA(sizeof(int) * hw_subset_depth);
1085 KMP_ASSERT(hw_subset_depth > 0);
1086 KMP_FOREACH_HW_TYPE(i) { specified[i] = KMP_HW_UNKNOWN; }
1087 int core_level = get_level(KMP_HW_CORE);
1088 for (int i = 0; i < hw_subset_depth; ++i) {
1089 int max_count;
1090 const kmp_hw_subset_t::item_t &item = __kmp_hw_subset->at(i);
1091 int num = item.num[0];
1092 int offset = item.offset[0];
1093 kmp_hw_t type = item.type;
1094 kmp_hw_t equivalent_type = equivalent[type];
1095 int level = get_level(type);
1096 topology_levels[i] = level;
1097
1098 // Check to see if current layer is in detected machine topology
1099 if (equivalent_type != KMP_HW_UNKNOWN) {
1100 __kmp_hw_subset->at(i).type = equivalent_type;
1101 } else {
1102 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetNotExistGeneric,
1103 __kmp_hw_get_catalog_string(type));
1104 return false;
1105 }
1106
1107 // Check to see if current layer has already been
1108 // specified either directly or through an equivalent type
1109 if (specified[equivalent_type] != KMP_HW_UNKNOWN) {
1110 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetEqvLayers,
1111 __kmp_hw_get_catalog_string(type),
1112 __kmp_hw_get_catalog_string(specified[equivalent_type]));
1113 return false;
1114 }
1115 specified[equivalent_type] = type;
1116
1117 // Check to see if each layer's num & offset parameters are valid
1118 max_count = get_ratio(level);
1119 if (!is_absolute) {
1120 if (max_count < 0 ||
1121 (num != kmp_hw_subset_t::USE_ALL && num + offset > max_count)) {
1122 bool plural = (num > 1);
1123 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetManyGeneric,
1124 __kmp_hw_get_catalog_string(type, plural));
1125 return false;
1126 }
1127 }
1128
1129 // Check to see if core attributes are consistent
1130 if (core_level == level) {
1131 // Determine which core attributes are specified
1132 for (int j = 0; j < item.num_attrs; ++j) {
1133 if (item.attr[j].is_core_type_valid())
1134 using_core_types = true;
1135 if (item.attr[j].is_core_eff_valid())
1136 using_core_effs = true;
1137 }
1138
1139 // Check if using a single core attribute on non-hybrid arch.
1140 // Do not ignore all of KMP_HW_SUBSET, just ignore the attribute.
1141 //
1142 // Check if using multiple core attributes on non-hyrbid arch.
1143 // Ignore all of KMP_HW_SUBSET if this is the case.
1144 if ((using_core_effs || using_core_types) && !__kmp_is_hybrid_cpu()) {
1145 if (item.num_attrs == 1) {
1146 if (using_core_effs) {
1147 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetIgnoringAttr,
1148 "efficiency");
1149 } else {
1150 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetIgnoringAttr,
1151 "core_type");
1152 }
1153 using_core_effs = false;
1154 using_core_types = false;
1155 } else {
1156 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetAttrsNonHybrid);
1157 return false;
1158 }
1159 }
1160
1161 // Check if using both core types and core efficiencies together
1162 if (using_core_types && using_core_effs) {
1163 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetIncompat, "core_type",
1164 "efficiency");
1165 return false;
1166 }
1167
1168 // Check that core efficiency values are valid
1169 if (using_core_effs) {
1170 for (int j = 0; j < item.num_attrs; ++j) {
1171 if (item.attr[j].is_core_eff_valid()) {
1172 int core_eff = item.attr[j].get_core_eff();
1173 if (core_eff < 0 || core_eff >= num_core_efficiencies) {
1174 kmp_str_buf_t buf;
1175 __kmp_str_buf_init(&buf);
1176 __kmp_str_buf_print(&buf, "%d", item.attr[j].get_core_eff());
1177 __kmp_msg(kmp_ms_warning,
1178 KMP_MSG(AffHWSubsetAttrInvalid, "efficiency", buf.str),
1179 KMP_HNT(ValidValuesRange, 0, num_core_efficiencies - 1),
1180 __kmp_msg_null);
1181 __kmp_str_buf_free(&buf);
1182 return false;
1183 }
1184 }
1185 }
1186 }
1187
1188 // Check that the number of requested cores with attributes is valid
1189 if ((using_core_types || using_core_effs) && !is_absolute) {
1190 for (int j = 0; j < item.num_attrs; ++j) {
1191 int num = item.num[j];
1192 int offset = item.offset[j];
1193 int level_above = core_level - 1;
1194 if (level_above >= 0) {
1195 max_count = get_ncores_with_attr_per(item.attr[j], level_above);
1196 if (max_count <= 0 ||
1197 (num != kmp_hw_subset_t::USE_ALL && num + offset > max_count)) {
1198 kmp_str_buf_t buf;
1199 __kmp_hw_get_catalog_core_string(item.attr[j], &buf, num > 0);
1200 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetManyGeneric, buf.str);
1201 __kmp_str_buf_free(&buf);
1202 return false;
1203 }
1204 }
1205 }
1206 }
1207
1208 if ((using_core_types || using_core_effs) && item.num_attrs > 1) {
1209 for (int j = 0; j < item.num_attrs; ++j) {
1210 // Ambiguous use of specific core attribute + generic core
1211 // e.g., 4c & 3c:intel_core or 4c & 3c:eff1
1212 if (!item.attr[j]) {
1213 kmp_hw_attr_t other_attr;
1214 for (int k = 0; k < item.num_attrs; ++k) {
1215 if (item.attr[k] != item.attr[j]) {
1216 other_attr = item.attr[k];
1217 break;
1218 }
1219 }
1220 kmp_str_buf_t buf;
1221 __kmp_hw_get_catalog_core_string(other_attr, &buf, item.num[j] > 0);
1222 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetIncompat,
1223 __kmp_hw_get_catalog_string(KMP_HW_CORE), buf.str);
1224 __kmp_str_buf_free(&buf);
1225 return false;
1226 }
1227 // Allow specifying a specific core type or core eff exactly once
1228 for (int k = 0; k < j; ++k) {
1229 if (!item.attr[j] || !item.attr[k])
1230 continue;
1231 if (item.attr[k] == item.attr[j]) {
1232 kmp_str_buf_t buf;
1233 __kmp_hw_get_catalog_core_string(item.attr[j], &buf,
1234 item.num[j] > 0);
1235 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetAttrRepeat, buf.str);
1236 __kmp_str_buf_free(&buf);
1237 return false;
1238 }
1239 }
1240 }
1241 }
1242 }
1243 }
1244
1245 // For keeping track of sub_ids for an absolute KMP_HW_SUBSET
1246 // or core attributes (core type or efficiency)
1247 int prev_sub_ids[KMP_HW_LAST];
1248 int abs_sub_ids[KMP_HW_LAST];
1249 int core_eff_sub_ids[KMP_HW_MAX_NUM_CORE_EFFS];
1250 int core_type_sub_ids[KMP_HW_MAX_NUM_CORE_TYPES];
1251 for (size_t i = 0; i < KMP_HW_LAST; ++i) {
1252 abs_sub_ids[i] = -1;
1253 prev_sub_ids[i] = -1;
1254 }
1255 for (size_t i = 0; i < KMP_HW_MAX_NUM_CORE_EFFS; ++i)
1256 core_eff_sub_ids[i] = -1;
1257 for (size_t i = 0; i < KMP_HW_MAX_NUM_CORE_TYPES; ++i)
1258 core_type_sub_ids[i] = -1;
1259
1260 // Determine which hardware threads should be filtered.
1261
1262 // Helpful to determine if a topology layer is targeted by an absolute subset
1263 auto is_targeted = [&](int level) {
1264 if (is_absolute) {
1265 for (int i = 0; i < hw_subset_depth; ++i)
1266 if (topology_levels[i] == level)
1267 return true;
1268 return false;
1269 }
1270 // If not absolute KMP_HW_SUBSET, then every layer is seen as targeted
1271 return true;
1272 };
1273
1274 // Helpful to index into core type sub Ids array
1275 auto get_core_type_index = [](const kmp_hw_thread_t &t) {
1276 switch (t.attrs.get_core_type()) {
1277 case KMP_HW_CORE_TYPE_UNKNOWN:
1278 case KMP_HW_MAX_NUM_CORE_TYPES:
1279 return 0;
1280#if KMP_ARCH_X86 || KMP_ARCH_X86_64
1281 case KMP_HW_CORE_TYPE_ATOM:
1282 return 1;
1283 case KMP_HW_CORE_TYPE_CORE:
1284 return 2;
1285#endif
1286 }
1287 KMP_ASSERT2(false, "Unhandled kmp_hw_thread_t enumeration");
1288 KMP_BUILTIN_UNREACHABLE;
1289 };
1290
1291 // Helpful to index into core efficiencies sub Ids array
1292 auto get_core_eff_index = [](const kmp_hw_thread_t &t) {
1293 return t.attrs.get_core_eff();
1294 };
1295
1296 int num_filtered = 0;
1297 kmp_affin_mask_t *filtered_mask;
1298 KMP_CPU_ALLOC(filtered_mask);
1299 KMP_CPU_COPY(filtered_mask, __kmp_affin_fullMask);
1300 for (int i = 0; i < num_hw_threads; ++i) {
1301 kmp_hw_thread_t &hw_thread = hw_threads[i];
1302
1303 // Figure out the absolute sub ids and core eff/type sub ids
1304 if (is_absolute || using_core_effs || using_core_types) {
1305 for (int level = 0; level < get_depth(); ++level) {
1306 if (hw_thread.sub_ids[level] != prev_sub_ids[level]) {
1307 bool found_targeted = false;
1308 for (int j = level; j < get_depth(); ++j) {
1309 bool targeted = is_targeted(j);
1310 if (!found_targeted && targeted) {
1311 found_targeted = true;
1312 abs_sub_ids[j]++;
1313 if (j == core_level && using_core_effs)
1314 core_eff_sub_ids[get_core_eff_index(hw_thread)]++;
1315 if (j == core_level && using_core_types)
1316 core_type_sub_ids[get_core_type_index(hw_thread)]++;
1317 } else if (targeted) {
1318 abs_sub_ids[j] = 0;
1319 if (j == core_level && using_core_effs)
1320 core_eff_sub_ids[get_core_eff_index(hw_thread)] = 0;
1321 if (j == core_level && using_core_types)
1322 core_type_sub_ids[get_core_type_index(hw_thread)] = 0;
1323 }
1324 }
1325 break;
1326 }
1327 }
1328 for (int level = 0; level < get_depth(); ++level)
1329 prev_sub_ids[level] = hw_thread.sub_ids[level];
1330 }
1331
1332 // Check to see if this hardware thread should be filtered
1333 bool should_be_filtered = false;
1334 for (int hw_subset_index = 0; hw_subset_index < hw_subset_depth;
1335 ++hw_subset_index) {
1336 const auto &hw_subset_item = __kmp_hw_subset->at(hw_subset_index);
1337 int level = topology_levels[hw_subset_index];
1338 if (level == -1)
1339 continue;
1340 if ((using_core_effs || using_core_types) && level == core_level) {
1341 // Look for the core attribute in KMP_HW_SUBSET which corresponds
1342 // to this hardware thread's core attribute. Use this num,offset plus
1343 // the running sub_id for the particular core attribute of this hardware
1344 // thread to determine if the hardware thread should be filtered or not.
1345 int attr_idx;
1346 kmp_hw_core_type_t core_type = hw_thread.attrs.get_core_type();
1347 int core_eff = hw_thread.attrs.get_core_eff();
1348 for (attr_idx = 0; attr_idx < hw_subset_item.num_attrs; ++attr_idx) {
1349 if (using_core_types &&
1350 hw_subset_item.attr[attr_idx].get_core_type() == core_type)
1351 break;
1352 if (using_core_effs &&
1353 hw_subset_item.attr[attr_idx].get_core_eff() == core_eff)
1354 break;
1355 }
1356 // This core attribute isn't in the KMP_HW_SUBSET so always filter it.
1357 if (attr_idx == hw_subset_item.num_attrs) {
1358 should_be_filtered = true;
1359 break;
1360 }
1361 int sub_id;
1362 int num = hw_subset_item.num[attr_idx];
1363 int offset = hw_subset_item.offset[attr_idx];
1364 if (using_core_types)
1365 sub_id = core_type_sub_ids[get_core_type_index(hw_thread)];
1366 else
1367 sub_id = core_eff_sub_ids[get_core_eff_index(hw_thread)];
1368 if (sub_id < offset ||
1369 (num != kmp_hw_subset_t::USE_ALL && sub_id >= offset + num)) {
1370 should_be_filtered = true;
1371 break;
1372 }
1373 } else {
1374 int sub_id;
1375 int num = hw_subset_item.num[0];
1376 int offset = hw_subset_item.offset[0];
1377 if (is_absolute)
1378 sub_id = abs_sub_ids[level];
1379 else
1380 sub_id = hw_thread.sub_ids[level];
1381 if (hw_thread.ids[level] == kmp_hw_thread_t::UNKNOWN_ID ||
1382 sub_id < offset ||
1383 (num != kmp_hw_subset_t::USE_ALL && sub_id >= offset + num)) {
1384 should_be_filtered = true;
1385 break;
1386 }
1387 }
1388 }
1389 // Collect filtering information
1390 if (should_be_filtered) {
1391 KMP_CPU_CLR(hw_thread.os_id, filtered_mask);
1392 num_filtered++;
1393 }
1394 }
1395
1396 // One last check that we shouldn't allow filtering entire machine
1397 if (num_filtered == num_hw_threads) {
1398 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetAllFiltered);
1399 return false;
1400 }
1401
1402 // Apply the filter
1403 restrict_to_mask(filtered_mask);
1404 return true;
1405}
1406
1407bool kmp_topology_t::is_close(int hwt1, int hwt2,
1408 const kmp_affinity_t &stgs) const {
1409 int hw_level = stgs.gran_levels;
1410 if (hw_level >= depth)
1411 return true;
1412 bool retval = true;
1413 const kmp_hw_thread_t &t1 = hw_threads[hwt1];
1414 const kmp_hw_thread_t &t2 = hw_threads[hwt2];
1415 if (stgs.flags.core_types_gran)
1416 return t1.attrs.get_core_type() == t2.attrs.get_core_type();
1417 if (stgs.flags.core_effs_gran)
1418 return t1.attrs.get_core_eff() == t2.attrs.get_core_eff();
1419 for (int i = 0; i < (depth - hw_level); ++i) {
1420 if (t1.ids[i] != t2.ids[i])
1421 return false;
1422 }
1423 return retval;
1424}
1425
1427
1428bool KMPAffinity::picked_api = false;
1429
1430void *KMPAffinity::Mask::operator new(size_t n) { return __kmp_allocate(n); }
1431void *KMPAffinity::Mask::operator new[](size_t n) { return __kmp_allocate(n); }
1432void KMPAffinity::Mask::operator delete(void *p) { __kmp_free(p); }
1433void KMPAffinity::Mask::operator delete[](void *p) { __kmp_free(p); }
1434void *KMPAffinity::operator new(size_t n) { return __kmp_allocate(n); }
1435void KMPAffinity::operator delete(void *p) { __kmp_free(p); }
1436
1437void KMPAffinity::pick_api() {
1438 KMPAffinity *affinity_dispatch;
1439 if (picked_api)
1440 return;
1441#if KMP_USE_HWLOC
1442 // Only use Hwloc if affinity isn't explicitly disabled and
1443 // user requests Hwloc topology method
1444 if (__kmp_affinity_top_method == affinity_top_method_hwloc &&
1445 __kmp_affinity.type != affinity_disabled) {
1446 affinity_dispatch = new KMPHwlocAffinity();
1447 } else
1448#endif
1449 {
1450 affinity_dispatch = new KMPNativeAffinity();
1451 }
1452 __kmp_affinity_dispatch = affinity_dispatch;
1453 picked_api = true;
1454}
1455
1456void KMPAffinity::destroy_api() {
1457 if (__kmp_affinity_dispatch != NULL) {
1458 delete __kmp_affinity_dispatch;
1459 __kmp_affinity_dispatch = NULL;
1460 picked_api = false;
1461 }
1462}
1463
1464#define KMP_ADVANCE_SCAN(scan) \
1465 while (*scan != '\0') { \
1466 scan++; \
1467 }
1468
1469// Print the affinity mask to the character array in a pretty format.
1470// The format is a comma separated list of non-negative integers or integer
1471// ranges: e.g., 1,2,3-5,7,9-15
1472// The format can also be the string "{<empty>}" if no bits are set in mask
1473char *__kmp_affinity_print_mask(char *buf, int buf_len,
1474 kmp_affin_mask_t *mask) {
1475 int start = 0, finish = 0, previous = 0;
1476 bool first_range;
1477 KMP_ASSERT(buf);
1478 KMP_ASSERT(buf_len >= 40);
1479 KMP_ASSERT(mask);
1480 char *scan = buf;
1481 char *end = buf + buf_len - 1;
1482
1483 // Check for empty set.
1484 if (mask->begin() == mask->end()) {
1485 KMP_SNPRINTF(scan, end - scan + 1, "{<empty>}");
1486 KMP_ADVANCE_SCAN(scan);
1487 KMP_ASSERT(scan <= end);
1488 return buf;
1489 }
1490
1491 first_range = true;
1492 start = mask->begin();
1493 while (1) {
1494 // Find next range
1495 // [start, previous] is inclusive range of contiguous bits in mask
1496 for (finish = mask->next(start), previous = start;
1497 finish == previous + 1 && finish != mask->end();
1498 finish = mask->next(finish)) {
1499 previous = finish;
1500 }
1501
1502 // The first range does not need a comma printed before it, but the rest
1503 // of the ranges do need a comma beforehand
1504 if (!first_range) {
1505 KMP_SNPRINTF(scan, end - scan + 1, "%s", ",");
1506 KMP_ADVANCE_SCAN(scan);
1507 } else {
1508 first_range = false;
1509 }
1510 // Range with three or more contiguous bits in the affinity mask
1511 if (previous - start > 1) {
1512 KMP_SNPRINTF(scan, end - scan + 1, "%u-%u", start, previous);
1513 } else {
1514 // Range with one or two contiguous bits in the affinity mask
1515 KMP_SNPRINTF(scan, end - scan + 1, "%u", start);
1516 KMP_ADVANCE_SCAN(scan);
1517 if (previous - start > 0) {
1518 KMP_SNPRINTF(scan, end - scan + 1, ",%u", previous);
1519 }
1520 }
1521 KMP_ADVANCE_SCAN(scan);
1522 // Start over with new start point
1523 start = finish;
1524 if (start == mask->end())
1525 break;
1526 // Check for overflow
1527 if (end - scan < 2)
1528 break;
1529 }
1530
1531 // Check for overflow
1532 KMP_ASSERT(scan <= end);
1533 return buf;
1534}
1535#undef KMP_ADVANCE_SCAN
1536
1537// Print the affinity mask to the string buffer object in a pretty format
1538// The format is a comma separated list of non-negative integers or integer
1539// ranges: e.g., 1,2,3-5,7,9-15
1540// The format can also be the string "{<empty>}" if no bits are set in mask
1541kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
1542 kmp_affin_mask_t *mask) {
1543 int start = 0, finish = 0, previous = 0;
1544 bool first_range;
1545 KMP_ASSERT(buf);
1546 KMP_ASSERT(mask);
1547
1548 __kmp_str_buf_clear(buf);
1549
1550 // Check for empty set.
1551 if (mask->begin() == mask->end()) {
1552 __kmp_str_buf_print(buf, "%s", "{<empty>}");
1553 return buf;
1554 }
1555
1556 first_range = true;
1557 start = mask->begin();
1558 while (1) {
1559 // Find next range
1560 // [start, previous] is inclusive range of contiguous bits in mask
1561 for (finish = mask->next(start), previous = start;
1562 finish == previous + 1 && finish != mask->end();
1563 finish = mask->next(finish)) {
1564 previous = finish;
1565 }
1566
1567 // The first range does not need a comma printed before it, but the rest
1568 // of the ranges do need a comma beforehand
1569 if (!first_range) {
1570 __kmp_str_buf_print(buf, "%s", ",");
1571 } else {
1572 first_range = false;
1573 }
1574 // Range with three or more contiguous bits in the affinity mask
1575 if (previous - start > 1) {
1576 __kmp_str_buf_print(buf, "%u-%u", start, previous);
1577 } else {
1578 // Range with one or two contiguous bits in the affinity mask
1579 __kmp_str_buf_print(buf, "%u", start);
1580 if (previous - start > 0) {
1581 __kmp_str_buf_print(buf, ",%u", previous);
1582 }
1583 }
1584 // Start over with new start point
1585 start = finish;
1586 if (start == mask->end())
1587 break;
1588 }
1589 return buf;
1590}
1591
1592// Return (possibly empty) affinity mask representing the offline CPUs
1593// Caller must free the mask
1594kmp_affin_mask_t *__kmp_affinity_get_offline_cpus() {
1595 kmp_affin_mask_t *offline;
1596 KMP_CPU_ALLOC(offline);
1597 KMP_CPU_ZERO(offline);
1598#if KMP_OS_LINUX
1599 int n, begin_cpu, end_cpu;
1600 kmp_safe_raii_file_t offline_file;
1601 auto skip_ws = [](FILE *f) {
1602 int c;
1603 do {
1604 c = fgetc(f);
1605 } while (isspace(c));
1606 if (c != EOF)
1607 ungetc(c, f);
1608 };
1609 // File contains CSV of integer ranges representing the offline CPUs
1610 // e.g., 1,2,4-7,9,11-15
1611 int status = offline_file.try_open("/sys/devices/system/cpu/offline", "r");
1612 if (status != 0)
1613 return offline;
1614 while (!feof(offline_file)) {
1615 skip_ws(offline_file);
1616 n = fscanf(offline_file, "%d", &begin_cpu);
1617 if (n != 1)
1618 break;
1619 skip_ws(offline_file);
1620 int c = fgetc(offline_file);
1621 if (c == EOF || c == ',') {
1622 // Just single CPU
1623 end_cpu = begin_cpu;
1624 } else if (c == '-') {
1625 // Range of CPUs
1626 skip_ws(offline_file);
1627 n = fscanf(offline_file, "%d", &end_cpu);
1628 if (n != 1)
1629 break;
1630 skip_ws(offline_file);
1631 c = fgetc(offline_file); // skip ','
1632 } else {
1633 // Syntax problem
1634 break;
1635 }
1636 // Ensure a valid range of CPUs
1637 if (begin_cpu < 0 || begin_cpu >= __kmp_xproc || end_cpu < 0 ||
1638 end_cpu >= __kmp_xproc || begin_cpu > end_cpu) {
1639 continue;
1640 }
1641 // Insert [begin_cpu, end_cpu] into offline mask
1642 for (int cpu = begin_cpu; cpu <= end_cpu; ++cpu) {
1643 KMP_CPU_SET(cpu, offline);
1644 }
1645 }
1646#endif
1647 return offline;
1648}
1649
1650// Return the number of available procs
1651int __kmp_affinity_entire_machine_mask(kmp_affin_mask_t *mask) {
1652 int avail_proc = 0;
1653 KMP_CPU_ZERO(mask);
1654
1655#if KMP_GROUP_AFFINITY
1656
1657 if (__kmp_num_proc_groups > 1) {
1658 int group;
1659 KMP_DEBUG_ASSERT(__kmp_GetActiveProcessorCount != NULL);
1660 for (group = 0; group < __kmp_num_proc_groups; group++) {
1661 int i;
1662 int num = __kmp_GetActiveProcessorCount(group);
1663 for (i = 0; i < num; i++) {
1664 KMP_CPU_SET(i + group * (CHAR_BIT * sizeof(DWORD_PTR)), mask);
1665 avail_proc++;
1666 }
1667 }
1668 } else
1669
1670#endif /* KMP_GROUP_AFFINITY */
1671
1672 {
1673 int proc;
1674 kmp_affin_mask_t *offline_cpus = __kmp_affinity_get_offline_cpus();
1675 for (proc = 0; proc < __kmp_xproc; proc++) {
1676 // Skip offline CPUs
1677 if (KMP_CPU_ISSET(proc, offline_cpus))
1678 continue;
1679 KMP_CPU_SET(proc, mask);
1680 avail_proc++;
1681 }
1682 KMP_CPU_FREE(offline_cpus);
1683 }
1684
1685 return avail_proc;
1686}
1687
1688// All of the __kmp_affinity_create_*_map() routines should allocate the
1689// internal topology object and set the layer ids for it. Each routine
1690// returns a boolean on whether it was successful at doing so.
1691kmp_affin_mask_t *__kmp_affin_fullMask = NULL;
1692// Original mask is a subset of full mask in multiple processor groups topology
1693kmp_affin_mask_t *__kmp_affin_origMask = NULL;
1694
1695#if KMP_USE_HWLOC
1696static inline bool __kmp_hwloc_is_cache_type(hwloc_obj_t obj) {
1697#if HWLOC_API_VERSION >= 0x00020000
1698 return hwloc_obj_type_is_cache(obj->type);
1699#else
1700 return obj->type == HWLOC_OBJ_CACHE;
1701#endif
1702}
1703
1704// Returns KMP_HW_* type derived from HWLOC_* type
1705static inline kmp_hw_t __kmp_hwloc_type_2_topology_type(hwloc_obj_t obj) {
1706
1707 if (__kmp_hwloc_is_cache_type(obj)) {
1708 if (obj->attr->cache.type == HWLOC_OBJ_CACHE_INSTRUCTION)
1709 return KMP_HW_UNKNOWN;
1710 switch (obj->attr->cache.depth) {
1711 case 1:
1712 return KMP_HW_L1;
1713 case 2:
1714#if KMP_MIC_SUPPORTED
1715 if (__kmp_mic_type == mic3) {
1716 return KMP_HW_TILE;
1717 }
1718#endif
1719 return KMP_HW_L2;
1720 case 3:
1721 return KMP_HW_L3;
1722 }
1723 return KMP_HW_UNKNOWN;
1724 }
1725
1726 switch (obj->type) {
1727 case HWLOC_OBJ_PACKAGE:
1728 return KMP_HW_SOCKET;
1729 case HWLOC_OBJ_NUMANODE:
1730 return KMP_HW_NUMA;
1731 case HWLOC_OBJ_CORE:
1732 return KMP_HW_CORE;
1733 case HWLOC_OBJ_PU:
1734 return KMP_HW_THREAD;
1735 case HWLOC_OBJ_GROUP:
1736#if HWLOC_API_VERSION >= 0x00020000
1737 if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_DIE)
1738 return KMP_HW_DIE;
1739 else if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_TILE)
1740 return KMP_HW_TILE;
1741 else if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_MODULE)
1742 return KMP_HW_MODULE;
1743 else if (obj->attr->group.kind == HWLOC_GROUP_KIND_WINDOWS_PROCESSOR_GROUP)
1744 return KMP_HW_PROC_GROUP;
1745#endif
1746 return KMP_HW_UNKNOWN;
1747#if HWLOC_API_VERSION >= 0x00020100
1748 case HWLOC_OBJ_DIE:
1749 return KMP_HW_DIE;
1750#endif
1751 }
1752 return KMP_HW_UNKNOWN;
1753}
1754
1755// Returns the number of objects of type 'type' below 'obj' within the topology
1756// tree structure. e.g., if obj is a HWLOC_OBJ_PACKAGE object, and type is
1757// HWLOC_OBJ_PU, then this will return the number of PU's under the SOCKET
1758// object.
1759static int __kmp_hwloc_get_nobjs_under_obj(hwloc_obj_t obj,
1760 hwloc_obj_type_t type) {
1761 int retval = 0;
1762 hwloc_obj_t first;
1763 for (first = hwloc_get_obj_below_by_type(__kmp_hwloc_topology, obj->type,
1764 obj->logical_index, type, 0);
1765 first != NULL && hwloc_get_ancestor_obj_by_type(__kmp_hwloc_topology,
1766 obj->type, first) == obj;
1767 first = hwloc_get_next_obj_by_type(__kmp_hwloc_topology, first->type,
1768 first)) {
1769 ++retval;
1770 }
1771 return retval;
1772}
1773
1774// This gets the sub_id for a lower object under a higher object in the
1775// topology tree
1776static int __kmp_hwloc_get_sub_id(hwloc_topology_t t, hwloc_obj_t higher,
1777 hwloc_obj_t lower) {
1778 hwloc_obj_t obj;
1779 hwloc_obj_type_t ltype = lower->type;
1780 int lindex = lower->logical_index - 1;
1781 int sub_id = 0;
1782 // Get the previous lower object
1783 obj = hwloc_get_obj_by_type(t, ltype, lindex);
1784 while (obj && lindex >= 0 &&
1785 hwloc_bitmap_isincluded(obj->cpuset, higher->cpuset)) {
1786 if (obj->userdata) {
1787 sub_id = (int)(RCAST(kmp_intptr_t, obj->userdata));
1788 break;
1789 }
1790 sub_id++;
1791 lindex--;
1792 obj = hwloc_get_obj_by_type(t, ltype, lindex);
1793 }
1794 // store sub_id + 1 so that 0 is differed from NULL
1795 lower->userdata = RCAST(void *, sub_id + 1);
1796 return sub_id;
1797}
1798
1799static bool __kmp_affinity_create_hwloc_map(kmp_i18n_id_t *const msg_id) {
1800 kmp_hw_t type;
1801 int hw_thread_index, sub_id;
1802 int depth;
1803 hwloc_obj_t pu, obj, root, prev;
1804 kmp_hw_t types[KMP_HW_LAST];
1805 hwloc_obj_type_t hwloc_types[KMP_HW_LAST];
1806
1807 hwloc_topology_t tp = __kmp_hwloc_topology;
1808 *msg_id = kmp_i18n_null;
1809 if (__kmp_affinity.flags.verbose) {
1810 KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
1811 }
1812
1813 if (!KMP_AFFINITY_CAPABLE()) {
1814 // Hack to try and infer the machine topology using only the data
1815 // available from hwloc on the current thread, and __kmp_xproc.
1816 KMP_ASSERT(__kmp_affinity.type == affinity_none);
1817 // hwloc only guarantees existance of PU object, so check PACKAGE and CORE
1818 hwloc_obj_t o = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PACKAGE, 0);
1819 if (o != NULL)
1820 nCoresPerPkg = __kmp_hwloc_get_nobjs_under_obj(o, HWLOC_OBJ_CORE);
1821 else
1822 nCoresPerPkg = 1; // no PACKAGE found
1823 o = hwloc_get_obj_by_type(tp, HWLOC_OBJ_CORE, 0);
1824 if (o != NULL)
1825 __kmp_nThreadsPerCore = __kmp_hwloc_get_nobjs_under_obj(o, HWLOC_OBJ_PU);
1826 else
1827 __kmp_nThreadsPerCore = 1; // no CORE found
1828 if (__kmp_nThreadsPerCore == 0)
1829 __kmp_nThreadsPerCore = 1;
1830 __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
1831 if (nCoresPerPkg == 0)
1832 nCoresPerPkg = 1; // to prevent possible division by 0
1833 nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
1834 return true;
1835 }
1836
1837#if HWLOC_API_VERSION >= 0x00020400
1838 // Handle multiple types of cores if they exist on the system
1839 int nr_cpu_kinds = hwloc_cpukinds_get_nr(tp, 0);
1840
1841 typedef struct kmp_hwloc_cpukinds_info_t {
1842 int efficiency;
1843 kmp_hw_core_type_t core_type;
1844 hwloc_bitmap_t mask;
1845 } kmp_hwloc_cpukinds_info_t;
1846 kmp_hwloc_cpukinds_info_t *cpukinds = nullptr;
1847
1848 if (nr_cpu_kinds > 0) {
1849 unsigned nr_infos;
1850 struct hwloc_info_s *infos;
1851 cpukinds = (kmp_hwloc_cpukinds_info_t *)__kmp_allocate(
1852 sizeof(kmp_hwloc_cpukinds_info_t) * nr_cpu_kinds);
1853 for (unsigned idx = 0; idx < (unsigned)nr_cpu_kinds; ++idx) {
1854 cpukinds[idx].efficiency = -1;
1855 cpukinds[idx].core_type = KMP_HW_CORE_TYPE_UNKNOWN;
1856 cpukinds[idx].mask = hwloc_bitmap_alloc();
1857 if (hwloc_cpukinds_get_info(tp, idx, cpukinds[idx].mask,
1858 &cpukinds[idx].efficiency, &nr_infos, &infos,
1859 0) == 0) {
1860 for (unsigned i = 0; i < nr_infos; ++i) {
1861 if (__kmp_str_match("CoreType", 8, infos[i].name)) {
1862#if KMP_ARCH_X86 || KMP_ARCH_X86_64
1863 if (__kmp_str_match("IntelAtom", 9, infos[i].value)) {
1864 cpukinds[idx].core_type = KMP_HW_CORE_TYPE_ATOM;
1865 break;
1866 } else if (__kmp_str_match("IntelCore", 9, infos[i].value)) {
1867 cpukinds[idx].core_type = KMP_HW_CORE_TYPE_CORE;
1868 break;
1869 }
1870#endif
1871 }
1872 }
1873 }
1874 }
1875 }
1876#endif
1877
1878 root = hwloc_get_root_obj(tp);
1879
1880 // Figure out the depth and types in the topology
1881 depth = 0;
1882 obj = hwloc_get_pu_obj_by_os_index(tp, __kmp_affin_fullMask->begin());
1883 while (obj && obj != root) {
1884#if HWLOC_API_VERSION >= 0x00020000
1885 if (obj->memory_arity) {
1886 hwloc_obj_t memory;
1887 for (memory = obj->memory_first_child; memory;
1888 memory = hwloc_get_next_child(tp, obj, memory)) {
1889 if (memory->type == HWLOC_OBJ_NUMANODE)
1890 break;
1891 }
1892 if (memory && memory->type == HWLOC_OBJ_NUMANODE) {
1893 types[depth] = KMP_HW_NUMA;
1894 hwloc_types[depth] = memory->type;
1895 depth++;
1896 }
1897 }
1898#endif
1899 type = __kmp_hwloc_type_2_topology_type(obj);
1900 if (type != KMP_HW_UNKNOWN) {
1901 types[depth] = type;
1902 hwloc_types[depth] = obj->type;
1903 depth++;
1904 }
1905 obj = obj->parent;
1906 }
1907 KMP_ASSERT(depth > 0);
1908
1909 // Get the order for the types correct
1910 for (int i = 0, j = depth - 1; i < j; ++i, --j) {
1911 hwloc_obj_type_t hwloc_temp = hwloc_types[i];
1912 kmp_hw_t temp = types[i];
1913 types[i] = types[j];
1914 types[j] = temp;
1915 hwloc_types[i] = hwloc_types[j];
1916 hwloc_types[j] = hwloc_temp;
1917 }
1918
1919 // Allocate the data structure to be returned.
1920 __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
1921
1922 hw_thread_index = 0;
1923 pu = NULL;
1924 while ((pu = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, pu))) {
1925 int index = depth - 1;
1926 bool included = KMP_CPU_ISSET(pu->os_index, __kmp_affin_fullMask);
1927 kmp_hw_thread_t &hw_thread = __kmp_topology->at(hw_thread_index);
1928 if (included) {
1929 hw_thread.clear();
1930 hw_thread.ids[index] = pu->logical_index;
1931 hw_thread.os_id = pu->os_index;
1932 hw_thread.original_idx = hw_thread_index;
1933 // If multiple core types, then set that attribute for the hardware thread
1934#if HWLOC_API_VERSION >= 0x00020400
1935 if (cpukinds) {
1936 int cpukind_index = -1;
1937 for (int i = 0; i < nr_cpu_kinds; ++i) {
1938 if (hwloc_bitmap_isset(cpukinds[i].mask, hw_thread.os_id)) {
1939 cpukind_index = i;
1940 break;
1941 }
1942 }
1943 if (cpukind_index >= 0) {
1944 hw_thread.attrs.set_core_type(cpukinds[cpukind_index].core_type);
1945 hw_thread.attrs.set_core_eff(cpukinds[cpukind_index].efficiency);
1946 }
1947 }
1948#endif
1949 index--;
1950 }
1951 obj = pu;
1952 prev = obj;
1953 while (obj != root && obj != NULL) {
1954 obj = obj->parent;
1955#if HWLOC_API_VERSION >= 0x00020000
1956 // NUMA Nodes are handled differently since they are not within the
1957 // parent/child structure anymore. They are separate children
1958 // of obj (memory_first_child points to first memory child)
1959 if (obj->memory_arity) {
1960 hwloc_obj_t memory;
1961 for (memory = obj->memory_first_child; memory;
1962 memory = hwloc_get_next_child(tp, obj, memory)) {
1963 if (memory->type == HWLOC_OBJ_NUMANODE)
1964 break;
1965 }
1966 if (memory && memory->type == HWLOC_OBJ_NUMANODE) {
1967 sub_id = __kmp_hwloc_get_sub_id(tp, memory, prev);
1968 if (included) {
1969 hw_thread.ids[index] = memory->logical_index;
1970 hw_thread.ids[index + 1] = sub_id;
1971 index--;
1972 }
1973 }
1974 prev = obj;
1975 }
1976#endif
1977 type = __kmp_hwloc_type_2_topology_type(obj);
1978 if (type != KMP_HW_UNKNOWN) {
1979 sub_id = __kmp_hwloc_get_sub_id(tp, obj, prev);
1980 if (included) {
1981 hw_thread.ids[index] = obj->logical_index;
1982 hw_thread.ids[index + 1] = sub_id;
1983 index--;
1984 }
1985 prev = obj;
1986 }
1987 }
1988 if (included)
1989 hw_thread_index++;
1990 }
1991
1992#if HWLOC_API_VERSION >= 0x00020400
1993 // Free the core types information
1994 if (cpukinds) {
1995 for (int idx = 0; idx < nr_cpu_kinds; ++idx)
1996 hwloc_bitmap_free(cpukinds[idx].mask);
1997 __kmp_free(cpukinds);
1998 }
1999#endif
2000 __kmp_topology->sort_ids();
2001 return true;
2002}
2003#endif // KMP_USE_HWLOC
2004
2005// If we don't know how to retrieve the machine's processor topology, or
2006// encounter an error in doing so, this routine is called to form a "flat"
2007// mapping of os thread id's <-> processor id's.
2008static bool __kmp_affinity_create_flat_map(kmp_i18n_id_t *const msg_id) {
2009 *msg_id = kmp_i18n_null;
2010 int depth = 3;
2011 kmp_hw_t types[] = {KMP_HW_SOCKET, KMP_HW_CORE, KMP_HW_THREAD};
2012
2013 if (__kmp_affinity.flags.verbose) {
2014 KMP_INFORM(UsingFlatOS, "KMP_AFFINITY");
2015 }
2016
2017 // Even if __kmp_affinity.type == affinity_none, this routine might still
2018 // be called to set __kmp_ncores, as well as
2019 // __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
2020 if (!KMP_AFFINITY_CAPABLE()) {
2021 KMP_ASSERT(__kmp_affinity.type == affinity_none);
2022 __kmp_ncores = nPackages = __kmp_xproc;
2023 __kmp_nThreadsPerCore = nCoresPerPkg = 1;
2024 return true;
2025 }
2026
2027 // When affinity is off, this routine will still be called to set
2028 // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
2029 // Make sure all these vars are set correctly, and return now if affinity is
2030 // not enabled.
2031 __kmp_ncores = nPackages = __kmp_avail_proc;
2032 __kmp_nThreadsPerCore = nCoresPerPkg = 1;
2033
2034 // Construct the data structure to be returned.
2035 __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
2036 int avail_ct = 0;
2037 int i;
2038 KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
2039 // Skip this proc if it is not included in the machine model.
2040 if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
2041 continue;
2042 }
2043 kmp_hw_thread_t &hw_thread = __kmp_topology->at(avail_ct);
2044 hw_thread.clear();
2045 hw_thread.os_id = i;
2046 hw_thread.original_idx = avail_ct;
2047 hw_thread.ids[0] = i;
2048 hw_thread.ids[1] = 0;
2049 hw_thread.ids[2] = 0;
2050 avail_ct++;
2051 }
2052 if (__kmp_affinity.flags.verbose) {
2053 KMP_INFORM(OSProcToPackage, "KMP_AFFINITY");
2054 }
2055 return true;
2056}
2057
2058#if KMP_GROUP_AFFINITY
2059// If multiple Windows* OS processor groups exist, we can create a 2-level
2060// topology map with the groups at level 0 and the individual procs at level 1.
2061// This facilitates letting the threads float among all procs in a group,
2062// if granularity=group (the default when there are multiple groups).
2063static bool __kmp_affinity_create_proc_group_map(kmp_i18n_id_t *const msg_id) {
2064 *msg_id = kmp_i18n_null;
2065 int depth = 3;
2066 kmp_hw_t types[] = {KMP_HW_PROC_GROUP, KMP_HW_CORE, KMP_HW_THREAD};
2067 const static size_t BITS_PER_GROUP = CHAR_BIT * sizeof(DWORD_PTR);
2068
2069 if (__kmp_affinity.flags.verbose) {
2070 KMP_INFORM(AffWindowsProcGroupMap, "KMP_AFFINITY");
2071 }
2072
2073 // If we aren't affinity capable, then use flat topology
2074 if (!KMP_AFFINITY_CAPABLE()) {
2075 KMP_ASSERT(__kmp_affinity.type == affinity_none);
2076 nPackages = __kmp_num_proc_groups;
2077 __kmp_nThreadsPerCore = 1;
2078 __kmp_ncores = __kmp_xproc;
2079 nCoresPerPkg = nPackages / __kmp_ncores;
2080 return true;
2081 }
2082
2083 // Construct the data structure to be returned.
2084 __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
2085 int avail_ct = 0;
2086 int i;
2087 KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
2088 // Skip this proc if it is not included in the machine model.
2089 if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
2090 continue;
2091 }
2092 kmp_hw_thread_t &hw_thread = __kmp_topology->at(avail_ct);
2093 hw_thread.clear();
2094 hw_thread.os_id = i;
2095 hw_thread.original_idx = avail_ct;
2096 hw_thread.ids[0] = i / BITS_PER_GROUP;
2097 hw_thread.ids[1] = hw_thread.ids[2] = i % BITS_PER_GROUP;
2098 avail_ct++;
2099 }
2100 return true;
2101}
2102#endif /* KMP_GROUP_AFFINITY */
2103
2104#if KMP_ARCH_X86 || KMP_ARCH_X86_64
2105
2106template <kmp_uint32 LSB, kmp_uint32 MSB>
2107static inline unsigned __kmp_extract_bits(kmp_uint32 v) {
2108 const kmp_uint32 SHIFT_LEFT = sizeof(kmp_uint32) * 8 - 1 - MSB;
2109 const kmp_uint32 SHIFT_RIGHT = LSB;
2110 kmp_uint32 retval = v;
2111 retval <<= SHIFT_LEFT;
2112 retval >>= (SHIFT_LEFT + SHIFT_RIGHT);
2113 return retval;
2114}
2115
2116static int __kmp_cpuid_mask_width(int count) {
2117 int r = 0;
2118
2119 while ((1 << r) < count)
2120 ++r;
2121 return r;
2122}
2123
2124class apicThreadInfo {
2125public:
2126 unsigned osId; // param to __kmp_affinity_bind_thread
2127 unsigned apicId; // from cpuid after binding
2128 unsigned maxCoresPerPkg; // ""
2129 unsigned maxThreadsPerPkg; // ""
2130 unsigned pkgId; // inferred from above values
2131 unsigned coreId; // ""
2132 unsigned threadId; // ""
2133};
2134
2135static int __kmp_affinity_cmp_apicThreadInfo_phys_id(const void *a,
2136 const void *b) {
2137 const apicThreadInfo *aa = (const apicThreadInfo *)a;
2138 const apicThreadInfo *bb = (const apicThreadInfo *)b;
2139 if (aa->pkgId < bb->pkgId)
2140 return -1;
2141 if (aa->pkgId > bb->pkgId)
2142 return 1;
2143 if (aa->coreId < bb->coreId)
2144 return -1;
2145 if (aa->coreId > bb->coreId)
2146 return 1;
2147 if (aa->threadId < bb->threadId)
2148 return -1;
2149 if (aa->threadId > bb->threadId)
2150 return 1;
2151 return 0;
2152}
2153
2154class cpuid_cache_info_t {
2155public:
2156 struct info_t {
2157 unsigned level = 0;
2158 unsigned mask = 0;
2159 bool operator==(const info_t &rhs) const {
2160 return level == rhs.level && mask == rhs.mask;
2161 }
2162 bool operator!=(const info_t &rhs) const { return !operator==(rhs); }
2163 };
2164 cpuid_cache_info_t() : depth(0) {
2165 table[MAX_CACHE_LEVEL].level = 0;
2166 table[MAX_CACHE_LEVEL].mask = 0;
2167 }
2168 size_t get_depth() const { return depth; }
2169 info_t &operator[](size_t index) { return table[index]; }
2170 const info_t &operator[](size_t index) const { return table[index]; }
2171 bool operator==(const cpuid_cache_info_t &rhs) const {
2172 if (rhs.depth != depth)
2173 return false;
2174 for (size_t i = 0; i < depth; ++i)
2175 if (table[i] != rhs.table[i])
2176 return false;
2177 return true;
2178 }
2179 bool operator!=(const cpuid_cache_info_t &rhs) const {
2180 return !operator==(rhs);
2181 }
2182 // Get cache information assocaited with L1, L2, L3 cache, etc.
2183 // If level does not exist, then return the "NULL" level (level 0)
2184 const info_t &get_level(unsigned level) const {
2185 for (size_t i = 0; i < depth; ++i) {
2186 if (table[i].level == level)
2187 return table[i];
2188 }
2189 return table[MAX_CACHE_LEVEL];
2190 }
2191
2192 static kmp_hw_t get_topology_type(unsigned level) {
2193 KMP_DEBUG_ASSERT(level >= 1 && level <= MAX_CACHE_LEVEL);
2194 switch (level) {
2195 case 1:
2196 return KMP_HW_L1;
2197 case 2:
2198 return KMP_HW_L2;
2199 case 3:
2200 return KMP_HW_L3;
2201 }
2202 return KMP_HW_UNKNOWN;
2203 }
2204 void get_leaf4_levels() {
2205 unsigned level = 0;
2206 while (depth < MAX_CACHE_LEVEL) {
2207 unsigned cache_type, max_threads_sharing;
2208 unsigned cache_level, cache_mask_width;
2209 kmp_cpuid buf2;
2210 __kmp_x86_cpuid(4, level, &buf2);
2211 cache_type = __kmp_extract_bits<0, 4>(buf2.eax);
2212 if (!cache_type)
2213 break;
2214 // Skip instruction caches
2215 if (cache_type == 2) {
2216 level++;
2217 continue;
2218 }
2219 max_threads_sharing = __kmp_extract_bits<14, 25>(buf2.eax) + 1;
2220 cache_mask_width = __kmp_cpuid_mask_width(max_threads_sharing);
2221 cache_level = __kmp_extract_bits<5, 7>(buf2.eax);
2222 table[depth].level = cache_level;
2223 table[depth].mask = ((-1) << cache_mask_width);
2224 depth++;
2225 level++;
2226 }
2227 }
2228 static const int MAX_CACHE_LEVEL = 3;
2229
2230private:
2231 size_t depth;
2232 info_t table[MAX_CACHE_LEVEL + 1];
2233};
2234
2235// On IA-32 architecture and Intel(R) 64 architecture, we attempt to use
2236// an algorithm which cycles through the available os threads, setting
2237// the current thread's affinity mask to that thread, and then retrieves
2238// the Apic Id for each thread context using the cpuid instruction.
2239static bool __kmp_affinity_create_apicid_map(kmp_i18n_id_t *const msg_id) {
2240 kmp_cpuid buf;
2241 *msg_id = kmp_i18n_null;
2242
2243 if (__kmp_affinity.flags.verbose) {
2244 KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(DecodingLegacyAPIC));
2245 }
2246
2247 // Check if cpuid leaf 4 is supported.
2248 __kmp_x86_cpuid(0, 0, &buf);
2249 if (buf.eax < 4) {
2250 *msg_id = kmp_i18n_str_NoLeaf4Support;
2251 return false;
2252 }
2253
2254 // The algorithm used starts by setting the affinity to each available thread
2255 // and retrieving info from the cpuid instruction, so if we are not capable of
2256 // calling __kmp_get_system_affinity() and _kmp_get_system_affinity(), then we
2257 // need to do something else - use the defaults that we calculated from
2258 // issuing cpuid without binding to each proc.
2259 if (!KMP_AFFINITY_CAPABLE()) {
2260 // Hack to try and infer the machine topology using only the data
2261 // available from cpuid on the current thread, and __kmp_xproc.
2262 KMP_ASSERT(__kmp_affinity.type == affinity_none);
2263
2264 // Get an upper bound on the number of threads per package using cpuid(1).
2265 // On some OS/chps combinations where HT is supported by the chip but is
2266 // disabled, this value will be 2 on a single core chip. Usually, it will be
2267 // 2 if HT is enabled and 1 if HT is disabled.
2268 __kmp_x86_cpuid(1, 0, &buf);
2269 int maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
2270 if (maxThreadsPerPkg == 0) {
2271 maxThreadsPerPkg = 1;
2272 }
2273
2274 // The num cores per pkg comes from cpuid(4). 1 must be added to the encoded
2275 // value.
2276 //
2277 // The author of cpu_count.cpp treated this only an upper bound on the
2278 // number of cores, but I haven't seen any cases where it was greater than
2279 // the actual number of cores, so we will treat it as exact in this block of
2280 // code.
2281 //
2282 // First, we need to check if cpuid(4) is supported on this chip. To see if
2283 // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n or
2284 // greater.
2285 __kmp_x86_cpuid(0, 0, &buf);
2286 if (buf.eax >= 4) {
2287 __kmp_x86_cpuid(4, 0, &buf);
2288 nCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
2289 } else {
2290 nCoresPerPkg = 1;
2291 }
2292
2293 // There is no way to reliably tell if HT is enabled without issuing the
2294 // cpuid instruction from every thread, can correlating the cpuid info, so
2295 // if the machine is not affinity capable, we assume that HT is off. We have
2296 // seen quite a few machines where maxThreadsPerPkg is 2, yet the machine
2297 // does not support HT.
2298 //
2299 // - Older OSes are usually found on machines with older chips, which do not
2300 // support HT.
2301 // - The performance penalty for mistakenly identifying a machine as HT when
2302 // it isn't (which results in blocktime being incorrectly set to 0) is
2303 // greater than the penalty when for mistakenly identifying a machine as
2304 // being 1 thread/core when it is really HT enabled (which results in
2305 // blocktime being incorrectly set to a positive value).
2306 __kmp_ncores = __kmp_xproc;
2307 nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
2308 __kmp_nThreadsPerCore = 1;
2309 return true;
2310 }
2311
2312 // From here on, we can assume that it is safe to call
2313 // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
2314 // __kmp_affinity.type = affinity_none.
2315
2316 // Save the affinity mask for the current thread.
2317 kmp_affinity_raii_t previous_affinity;
2318
2319 // Run through each of the available contexts, binding the current thread
2320 // to it, and obtaining the pertinent information using the cpuid instr.
2321 //
2322 // The relevant information is:
2323 // - Apic Id: Bits 24:31 of ebx after issuing cpuid(1) - each thread context
2324 // has a uniqie Apic Id, which is of the form pkg# : core# : thread#.
2325 // - Max Threads Per Pkg: Bits 16:23 of ebx after issuing cpuid(1). The value
2326 // of this field determines the width of the core# + thread# fields in the
2327 // Apic Id. It is also an upper bound on the number of threads per
2328 // package, but it has been verified that situations happen were it is not
2329 // exact. In particular, on certain OS/chip combinations where Intel(R)
2330 // Hyper-Threading Technology is supported by the chip but has been
2331 // disabled, the value of this field will be 2 (for a single core chip).
2332 // On other OS/chip combinations supporting Intel(R) Hyper-Threading
2333 // Technology, the value of this field will be 1 when Intel(R)
2334 // Hyper-Threading Technology is disabled and 2 when it is enabled.
2335 // - Max Cores Per Pkg: Bits 26:31 of eax after issuing cpuid(4). The value
2336 // of this field (+1) determines the width of the core# field in the Apic
2337 // Id. The comments in "cpucount.cpp" say that this value is an upper
2338 // bound, but the IA-32 architecture manual says that it is exactly the
2339 // number of cores per package, and I haven't seen any case where it
2340 // wasn't.
2341 //
2342 // From this information, deduce the package Id, core Id, and thread Id,
2343 // and set the corresponding fields in the apicThreadInfo struct.
2344 unsigned i;
2345 apicThreadInfo *threadInfo = (apicThreadInfo *)__kmp_allocate(
2346 __kmp_avail_proc * sizeof(apicThreadInfo));
2347 unsigned nApics = 0;
2348 KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
2349 // Skip this proc if it is not included in the machine model.
2350 if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
2351 continue;
2352 }
2353 KMP_DEBUG_ASSERT((int)nApics < __kmp_avail_proc);
2354
2355 __kmp_affinity_dispatch->bind_thread(i);
2356 threadInfo[nApics].osId = i;
2357
2358 // The apic id and max threads per pkg come from cpuid(1).
2359 __kmp_x86_cpuid(1, 0, &buf);
2360 if (((buf.edx >> 9) & 1) == 0) {
2361 __kmp_free(threadInfo);
2362 *msg_id = kmp_i18n_str_ApicNotPresent;
2363 return false;
2364 }
2365 threadInfo[nApics].apicId = (buf.ebx >> 24) & 0xff;
2366 threadInfo[nApics].maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
2367 if (threadInfo[nApics].maxThreadsPerPkg == 0) {
2368 threadInfo[nApics].maxThreadsPerPkg = 1;
2369 }
2370
2371 // Max cores per pkg comes from cpuid(4). 1 must be added to the encoded
2372 // value.
2373 //
2374 // First, we need to check if cpuid(4) is supported on this chip. To see if
2375 // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n
2376 // or greater.
2377 __kmp_x86_cpuid(0, 0, &buf);
2378 if (buf.eax >= 4) {
2379 __kmp_x86_cpuid(4, 0, &buf);
2380 threadInfo[nApics].maxCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
2381 } else {
2382 threadInfo[nApics].maxCoresPerPkg = 1;
2383 }
2384
2385 // Infer the pkgId / coreId / threadId using only the info obtained locally.
2386 int widthCT = __kmp_cpuid_mask_width(threadInfo[nApics].maxThreadsPerPkg);
2387 threadInfo[nApics].pkgId = threadInfo[nApics].apicId >> widthCT;
2388
2389 int widthC = __kmp_cpuid_mask_width(threadInfo[nApics].maxCoresPerPkg);
2390 int widthT = widthCT - widthC;
2391 if (widthT < 0) {
2392 // I've never seen this one happen, but I suppose it could, if the cpuid
2393 // instruction on a chip was really screwed up. Make sure to restore the
2394 // affinity mask before the tail call.
2395 __kmp_free(threadInfo);
2396 *msg_id = kmp_i18n_str_InvalidCpuidInfo;
2397 return false;
2398 }
2399
2400 int maskC = (1 << widthC) - 1;
2401 threadInfo[nApics].coreId = (threadInfo[nApics].apicId >> widthT) & maskC;
2402
2403 int maskT = (1 << widthT) - 1;
2404 threadInfo[nApics].threadId = threadInfo[nApics].apicId & maskT;
2405
2406 nApics++;
2407 }
2408
2409 // We've collected all the info we need.
2410 // Restore the old affinity mask for this thread.
2411 previous_affinity.restore();
2412
2413 // Sort the threadInfo table by physical Id.
2414 qsort(threadInfo, nApics, sizeof(*threadInfo),
2415 __kmp_affinity_cmp_apicThreadInfo_phys_id);
2416
2417 // The table is now sorted by pkgId / coreId / threadId, but we really don't
2418 // know the radix of any of the fields. pkgId's may be sparsely assigned among
2419 // the chips on a system. Although coreId's are usually assigned
2420 // [0 .. coresPerPkg-1] and threadId's are usually assigned
2421 // [0..threadsPerCore-1], we don't want to make any such assumptions.
2422 //
2423 // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
2424 // total # packages) are at this point - we want to determine that now. We
2425 // only have an upper bound on the first two figures.
2426 //
2427 // We also perform a consistency check at this point: the values returned by
2428 // the cpuid instruction for any thread bound to a given package had better
2429 // return the same info for maxThreadsPerPkg and maxCoresPerPkg.
2430 nPackages = 1;
2431 nCoresPerPkg = 1;
2432 __kmp_nThreadsPerCore = 1;
2433 unsigned nCores = 1;
2434
2435 unsigned pkgCt = 1; // to determine radii
2436 unsigned lastPkgId = threadInfo[0].pkgId;
2437 unsigned coreCt = 1;
2438 unsigned lastCoreId = threadInfo[0].coreId;
2439 unsigned threadCt = 1;
2440 unsigned lastThreadId = threadInfo[0].threadId;
2441
2442 // intra-pkg consist checks
2443 unsigned prevMaxCoresPerPkg = threadInfo[0].maxCoresPerPkg;
2444 unsigned prevMaxThreadsPerPkg = threadInfo[0].maxThreadsPerPkg;
2445
2446 for (i = 1; i < nApics; i++) {
2447 if (threadInfo[i].pkgId != lastPkgId) {
2448 nCores++;
2449 pkgCt++;
2450 lastPkgId = threadInfo[i].pkgId;
2451 if ((int)coreCt > nCoresPerPkg)
2452 nCoresPerPkg = coreCt;
2453 coreCt = 1;
2454 lastCoreId = threadInfo[i].coreId;
2455 if ((int)threadCt > __kmp_nThreadsPerCore)
2456 __kmp_nThreadsPerCore = threadCt;
2457 threadCt = 1;
2458 lastThreadId = threadInfo[i].threadId;
2459
2460 // This is a different package, so go on to the next iteration without
2461 // doing any consistency checks. Reset the consistency check vars, though.
2462 prevMaxCoresPerPkg = threadInfo[i].maxCoresPerPkg;
2463 prevMaxThreadsPerPkg = threadInfo[i].maxThreadsPerPkg;
2464 continue;
2465 }
2466
2467 if (threadInfo[i].coreId != lastCoreId) {
2468 nCores++;
2469 coreCt++;
2470 lastCoreId = threadInfo[i].coreId;
2471 if ((int)threadCt > __kmp_nThreadsPerCore)
2472 __kmp_nThreadsPerCore = threadCt;
2473 threadCt = 1;
2474 lastThreadId = threadInfo[i].threadId;
2475 } else if (threadInfo[i].threadId != lastThreadId) {
2476 threadCt++;
2477 lastThreadId = threadInfo[i].threadId;
2478 } else {
2479 __kmp_free(threadInfo);
2480 *msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
2481 return false;
2482 }
2483
2484 // Check to make certain that the maxCoresPerPkg and maxThreadsPerPkg
2485 // fields agree between all the threads bounds to a given package.
2486 if ((prevMaxCoresPerPkg != threadInfo[i].maxCoresPerPkg) ||
2487 (prevMaxThreadsPerPkg != threadInfo[i].maxThreadsPerPkg)) {
2488 __kmp_free(threadInfo);
2489 *msg_id = kmp_i18n_str_InconsistentCpuidInfo;
2490 return false;
2491 }
2492 }
2493 // When affinity is off, this routine will still be called to set
2494 // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
2495 // Make sure all these vars are set correctly
2496 nPackages = pkgCt;
2497 if ((int)coreCt > nCoresPerPkg)
2498 nCoresPerPkg = coreCt;
2499 if ((int)threadCt > __kmp_nThreadsPerCore)
2500 __kmp_nThreadsPerCore = threadCt;
2501 __kmp_ncores = nCores;
2502 KMP_DEBUG_ASSERT(nApics == (unsigned)__kmp_avail_proc);
2503
2504 // Now that we've determined the number of packages, the number of cores per
2505 // package, and the number of threads per core, we can construct the data
2506 // structure that is to be returned.
2507 int idx = 0;
2508 int pkgLevel = 0;
2509 int coreLevel = 1;
2510 int threadLevel = 2;
2511 //(__kmp_nThreadsPerCore <= 1) ? -1 : ((coreLevel >= 0) ? 2 : 1);
2512 int depth = (pkgLevel >= 0) + (coreLevel >= 0) + (threadLevel >= 0);
2513 kmp_hw_t types[3];
2514 if (pkgLevel >= 0)
2515 types[idx++] = KMP_HW_SOCKET;
2516 if (coreLevel >= 0)
2517 types[idx++] = KMP_HW_CORE;
2518 if (threadLevel >= 0)
2519 types[idx++] = KMP_HW_THREAD;
2520
2521 KMP_ASSERT(depth > 0);
2522 __kmp_topology = kmp_topology_t::allocate(nApics, depth, types);
2523
2524 for (i = 0; i < nApics; ++i) {
2525 idx = 0;
2526 unsigned os = threadInfo[i].osId;
2527 kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
2528 hw_thread.clear();
2529
2530 if (pkgLevel >= 0) {
2531 hw_thread.ids[idx++] = threadInfo[i].pkgId;
2532 }
2533 if (coreLevel >= 0) {
2534 hw_thread.ids[idx++] = threadInfo[i].coreId;
2535 }
2536 if (threadLevel >= 0) {
2537 hw_thread.ids[idx++] = threadInfo[i].threadId;
2538 }
2539 hw_thread.os_id = os;
2540 hw_thread.original_idx = i;
2541 }
2542
2543 __kmp_free(threadInfo);
2544 __kmp_topology->sort_ids();
2545 if (!__kmp_topology->check_ids()) {
2546 kmp_topology_t::deallocate(__kmp_topology);
2547 __kmp_topology = nullptr;
2548 *msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
2549 return false;
2550 }
2551 return true;
2552}
2553
2554// Hybrid cpu detection using CPUID.1A
2555// Thread should be pinned to processor already
2556static void __kmp_get_hybrid_info(kmp_hw_core_type_t *type, int *efficiency,
2557 unsigned *native_model_id) {
2558 kmp_cpuid buf;
2559 __kmp_x86_cpuid(0x1a, 0, &buf);
2560 *type = (kmp_hw_core_type_t)__kmp_extract_bits<24, 31>(buf.eax);
2561 switch (*type) {
2562 case KMP_HW_CORE_TYPE_ATOM:
2563 *efficiency = 0;
2564 break;
2565 case KMP_HW_CORE_TYPE_CORE:
2566 *efficiency = 1;
2567 break;
2568 default:
2569 *efficiency = 0;
2570 }
2571 *native_model_id = __kmp_extract_bits<0, 23>(buf.eax);
2572}
2573
2574// Intel(R) microarchitecture code name Nehalem, Dunnington and later
2575// architectures support a newer interface for specifying the x2APIC Ids,
2576// based on CPUID.B or CPUID.1F
2577/*
2578 * CPUID.B or 1F, Input ECX (sub leaf # aka level number)
2579 Bits Bits Bits Bits
2580 31-16 15-8 7-4 4-0
2581---+-----------+--------------+-------------+-----------------+
2582EAX| reserved | reserved | reserved | Bits to Shift |
2583---+-----------|--------------+-------------+-----------------|
2584EBX| reserved | Num logical processors at level (16 bits) |
2585---+-----------|--------------+-------------------------------|
2586ECX| reserved | Level Type | Level Number (8 bits) |
2587---+-----------+--------------+-------------------------------|
2588EDX| X2APIC ID (32 bits) |
2589---+----------------------------------------------------------+
2590*/
2591
2592enum {
2593 INTEL_LEVEL_TYPE_INVALID = 0, // Package level
2594 INTEL_LEVEL_TYPE_SMT = 1,
2595 INTEL_LEVEL_TYPE_CORE = 2,
2596 INTEL_LEVEL_TYPE_MODULE = 3,
2597 INTEL_LEVEL_TYPE_TILE = 4,
2598 INTEL_LEVEL_TYPE_DIE = 5,
2599 INTEL_LEVEL_TYPE_LAST = 6,
2600};
2601KMP_BUILD_ASSERT(INTEL_LEVEL_TYPE_LAST < sizeof(unsigned) * CHAR_BIT);
2602#define KMP_LEAF_1F_KNOWN_LEVELS ((1u << INTEL_LEVEL_TYPE_LAST) - 1u)
2603
2604static kmp_hw_t __kmp_intel_type_2_topology_type(int intel_type) {
2605 switch (intel_type) {
2606 case INTEL_LEVEL_TYPE_INVALID:
2607 return KMP_HW_SOCKET;
2608 case INTEL_LEVEL_TYPE_SMT:
2609 return KMP_HW_THREAD;
2610 case INTEL_LEVEL_TYPE_CORE:
2611 return KMP_HW_CORE;
2612 case INTEL_LEVEL_TYPE_TILE:
2613 return KMP_HW_TILE;
2614 case INTEL_LEVEL_TYPE_MODULE:
2615 return KMP_HW_MODULE;
2616 case INTEL_LEVEL_TYPE_DIE:
2617 return KMP_HW_DIE;
2618 }
2619 return KMP_HW_UNKNOWN;
2620}
2621
2622static int __kmp_topology_type_2_intel_type(kmp_hw_t type) {
2623 switch (type) {
2624 case KMP_HW_SOCKET:
2625 return INTEL_LEVEL_TYPE_INVALID;
2626 case KMP_HW_THREAD:
2627 return INTEL_LEVEL_TYPE_SMT;
2628 case KMP_HW_CORE:
2629 return INTEL_LEVEL_TYPE_CORE;
2630 case KMP_HW_TILE:
2631 return INTEL_LEVEL_TYPE_TILE;
2632 case KMP_HW_MODULE:
2633 return INTEL_LEVEL_TYPE_MODULE;
2634 case KMP_HW_DIE:
2635 return INTEL_LEVEL_TYPE_DIE;
2636 default:
2637 return INTEL_LEVEL_TYPE_INVALID;
2638 }
2639}
2640
2641struct cpuid_level_info_t {
2642 unsigned level_type, mask, mask_width, nitems, cache_mask;
2643};
2644
2645class cpuid_topo_desc_t {
2646 unsigned desc = 0;
2647
2648public:
2649 void clear() { desc = 0; }
2650 bool contains(int intel_type) const {
2651 KMP_DEBUG_ASSERT(intel_type >= 0 && intel_type < INTEL_LEVEL_TYPE_LAST);
2652 if ((1u << intel_type) & desc)
2653 return true;
2654 return false;
2655 }
2656 bool contains_topology_type(kmp_hw_t type) const {
2657 KMP_DEBUG_ASSERT(type >= 0 && type < KMP_HW_LAST);
2658 int intel_type = __kmp_topology_type_2_intel_type(type);
2659 return contains(intel_type);
2660 }
2661 bool contains(cpuid_topo_desc_t rhs) const {
2662 return ((desc | rhs.desc) == desc);
2663 }
2664 void add(int intel_type) { desc |= (1u << intel_type); }
2665 void add(cpuid_topo_desc_t rhs) { desc |= rhs.desc; }
2666};
2667
2668struct cpuid_proc_info_t {
2669 // Topology info
2670 int os_id;
2671 unsigned apic_id;
2672 unsigned depth;
2673 // Hybrid info
2674 unsigned native_model_id;
2675 int efficiency;
2676 kmp_hw_core_type_t type;
2677 cpuid_topo_desc_t description;
2678
2679 cpuid_level_info_t levels[INTEL_LEVEL_TYPE_LAST];
2680};
2681
2682// This function takes the topology leaf, an info pointer to store the levels
2683// detected, and writable descriptors for the total topology.
2684// Returns whether total types, depth, or description were modified.
2685static bool __kmp_x2apicid_get_levels(int leaf, cpuid_proc_info_t *info,
2686 kmp_hw_t total_types[KMP_HW_LAST],
2687 int *total_depth,
2688 cpuid_topo_desc_t *total_description) {
2689 unsigned level, levels_index;
2690 unsigned level_type, mask_width, nitems;
2691 kmp_cpuid buf;
2692 cpuid_level_info_t(&levels)[INTEL_LEVEL_TYPE_LAST] = info->levels;
2693 bool retval = false;
2694
2695 // New algorithm has known topology layers act as highest unknown topology
2696 // layers when unknown topology layers exist.
2697 // e.g., Suppose layers were SMT <X> CORE <Y> <Z> PACKAGE, where <X> <Y> <Z>
2698 // are unknown topology layers, Then SMT will take the characteristics of
2699 // (SMT x <X>) and CORE will take the characteristics of (CORE x <Y> x <Z>).
2700 // This eliminates unknown portions of the topology while still keeping the
2701 // correct structure.
2702 level = levels_index = 0;
2703 do {
2704 __kmp_x86_cpuid(leaf, level, &buf);
2705 level_type = __kmp_extract_bits<8, 15>(buf.ecx);
2706 mask_width = __kmp_extract_bits<0, 4>(buf.eax);
2707 nitems = __kmp_extract_bits<0, 15>(buf.ebx);
2708 if (level_type != INTEL_LEVEL_TYPE_INVALID && nitems == 0) {
2709 info->depth = 0;
2710 return retval;
2711 }
2712
2713 if (KMP_LEAF_1F_KNOWN_LEVELS & (1u << level_type)) {
2714 // Add a new level to the topology
2715 KMP_ASSERT(levels_index < INTEL_LEVEL_TYPE_LAST);
2716 levels[levels_index].level_type = level_type;
2717 levels[levels_index].mask_width = mask_width;
2718 levels[levels_index].nitems = nitems;
2719 levels_index++;
2720 } else {
2721 // If it is an unknown level, then logically move the previous layer up
2722 if (levels_index > 0) {
2723 levels[levels_index - 1].mask_width = mask_width;
2724 levels[levels_index - 1].nitems = nitems;
2725 }
2726 }
2727 level++;
2728 } while (level_type != INTEL_LEVEL_TYPE_INVALID);
2729 KMP_ASSERT(levels_index <= INTEL_LEVEL_TYPE_LAST);
2730 info->description.clear();
2731 info->depth = levels_index;
2732
2733 // If types, depth, and total_description are uninitialized,
2734 // then initialize them now
2735 if (*total_depth == 0) {
2736 *total_depth = info->depth;
2737 total_description->clear();
2738 for (int i = *total_depth - 1, j = 0; i >= 0; --i, ++j) {
2739 total_types[j] =
2740 __kmp_intel_type_2_topology_type(info->levels[i].level_type);
2741 total_description->add(info->levels[i].level_type);
2742 }
2743 retval = true;
2744 }
2745
2746 // Ensure the INTEL_LEVEL_TYPE_INVALID (Socket) layer isn't first
2747 if (levels_index == 0 || levels[0].level_type == INTEL_LEVEL_TYPE_INVALID)
2748 return 0;
2749
2750 // Set the masks to & with apicid
2751 for (unsigned i = 0; i < levels_index; ++i) {
2752 if (levels[i].level_type != INTEL_LEVEL_TYPE_INVALID) {
2753 levels[i].mask = ~((-1) << levels[i].mask_width);
2754 levels[i].cache_mask = (-1) << levels[i].mask_width;
2755 for (unsigned j = 0; j < i; ++j)
2756 levels[i].mask ^= levels[j].mask;
2757 } else {
2758 KMP_DEBUG_ASSERT(i > 0);
2759 levels[i].mask = (-1) << levels[i - 1].mask_width;
2760 levels[i].cache_mask = 0;
2761 }
2762 info->description.add(info->levels[i].level_type);
2763 }
2764
2765 // If this processor has level type not on other processors, then make
2766 // sure to include it in total types, depth, and description.
2767 // One assumption here is that the first type, i.e. socket, is known.
2768 // Another assumption is that types array is always large enough to fit any
2769 // new layers since its length is KMP_HW_LAST.
2770 if (!total_description->contains(info->description)) {
2771 for (int i = info->depth - 1, j = 0; i >= 0; --i, ++j) {
2772 // If this level is known already, then skip it.
2773 if (total_description->contains(levels[i].level_type))
2774 continue;
2775 // Unknown level, insert before last known level
2776 kmp_hw_t curr_type =
2777 __kmp_intel_type_2_topology_type(levels[i].level_type);
2778 KMP_ASSERT(j != 0 && "Bad APIC Id information");
2779 // Move over all known levels to make room for new level
2780 for (int k = info->depth - 1; k >= j; --k) {
2781 KMP_DEBUG_ASSERT(k + 1 < KMP_HW_LAST);
2782 total_types[k + 1] = total_types[k];
2783 }
2784 // Insert new level
2785 total_types[j] = curr_type;
2786 (*total_depth)++;
2787 }
2788 total_description->add(info->description);
2789 retval = true;
2790 }
2791 return retval;
2792}
2793
2794static bool __kmp_affinity_create_x2apicid_map(kmp_i18n_id_t *const msg_id) {
2795
2796 kmp_hw_t types[INTEL_LEVEL_TYPE_LAST];
2797 kmp_cpuid buf;
2798 int topology_leaf, highest_leaf;
2799 int num_leaves;
2800 int depth = 0;
2801 cpuid_topo_desc_t total_description;
2802 static int leaves[] = {0, 0};
2803
2804 // If affinity is disabled, __kmp_avail_proc may be zero
2805 int ninfos = (__kmp_avail_proc > 0 ? __kmp_avail_proc : 1);
2806 cpuid_proc_info_t *proc_info = (cpuid_proc_info_t *)__kmp_allocate(
2807 (sizeof(cpuid_proc_info_t) + sizeof(cpuid_cache_info_t)) * ninfos);
2808 cpuid_cache_info_t *cache_info = (cpuid_cache_info_t *)(proc_info + ninfos);
2809
2810 kmp_i18n_id_t leaf_message_id;
2811
2812 *msg_id = kmp_i18n_null;
2813 if (__kmp_affinity.flags.verbose) {
2814 KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(Decodingx2APIC));
2815 }
2816
2817 // Get the highest cpuid leaf supported
2818 __kmp_x86_cpuid(0, 0, &buf);
2819 highest_leaf = buf.eax;
2820
2821 // If a specific topology method was requested, only allow that specific leaf
2822 // otherwise, try both leaves 31 and 11 in that order
2823 num_leaves = 0;
2824 if (__kmp_affinity_top_method == affinity_top_method_x2apicid) {
2825 num_leaves = 1;
2826 leaves[0] = 11;
2827 leaf_message_id = kmp_i18n_str_NoLeaf11Support;
2828 } else if (__kmp_affinity_top_method == affinity_top_method_x2apicid_1f) {
2829 num_leaves = 1;
2830 leaves[0] = 31;
2831 leaf_message_id = kmp_i18n_str_NoLeaf31Support;
2832 } else {
2833 num_leaves = 2;
2834 leaves[0] = 31;
2835 leaves[1] = 11;
2836 leaf_message_id = kmp_i18n_str_NoLeaf11Support;
2837 }
2838
2839 // Check to see if cpuid leaf 31 or 11 is supported.
2840 __kmp_nThreadsPerCore = nCoresPerPkg = nPackages = 1;
2841 topology_leaf = -1;
2842 for (int i = 0; i < num_leaves; ++i) {
2843 int leaf = leaves[i];
2844 if (highest_leaf < leaf)
2845 continue;
2846 __kmp_x86_cpuid(leaf, 0, &buf);
2847 if (buf.ebx == 0)
2848 continue;
2849 topology_leaf = leaf;
2850 __kmp_x2apicid_get_levels(leaf, &proc_info[0], types, &depth,
2851 &total_description);
2852 if (depth == 0)
2853 continue;
2854 break;
2855 }
2856 if (topology_leaf == -1 || depth == 0) {
2857 *msg_id = leaf_message_id;
2858 __kmp_free(proc_info);
2859 return false;
2860 }
2861 KMP_ASSERT(depth <= INTEL_LEVEL_TYPE_LAST);
2862
2863 // The algorithm used starts by setting the affinity to each available thread
2864 // and retrieving info from the cpuid instruction, so if we are not capable of
2865 // calling __kmp_get_system_affinity() and __kmp_get_system_affinity(), then
2866 // we need to do something else - use the defaults that we calculated from
2867 // issuing cpuid without binding to each proc.
2868 if (!KMP_AFFINITY_CAPABLE()) {
2869 // Hack to try and infer the machine topology using only the data
2870 // available from cpuid on the current thread, and __kmp_xproc.
2871 KMP_ASSERT(__kmp_affinity.type == affinity_none);
2872 for (int i = 0; i < depth; ++i) {
2873 if (proc_info[0].levels[i].level_type == INTEL_LEVEL_TYPE_SMT) {
2874 __kmp_nThreadsPerCore = proc_info[0].levels[i].nitems;
2875 } else if (proc_info[0].levels[i].level_type == INTEL_LEVEL_TYPE_CORE) {
2876 nCoresPerPkg = proc_info[0].levels[i].nitems;
2877 }
2878 }
2879 __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
2880 nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
2881 __kmp_free(proc_info);
2882 return true;
2883 }
2884
2885 // From here on, we can assume that it is safe to call
2886 // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
2887 // __kmp_affinity.type = affinity_none.
2888
2889 // Save the affinity mask for the current thread.
2890 kmp_affinity_raii_t previous_affinity;
2891
2892 // Run through each of the available contexts, binding the current thread
2893 // to it, and obtaining the pertinent information using the cpuid instr.
2894 unsigned int proc;
2895 int hw_thread_index = 0;
2896 bool uniform_caches = true;
2897
2898 KMP_CPU_SET_ITERATE(proc, __kmp_affin_fullMask) {
2899 // Skip this proc if it is not included in the machine model.
2900 if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
2901 continue;
2902 }
2903 KMP_DEBUG_ASSERT(hw_thread_index < __kmp_avail_proc);
2904
2905 // Gather topology information
2906 __kmp_affinity_dispatch->bind_thread(proc);
2907 __kmp_x86_cpuid(topology_leaf, 0, &buf);
2908 proc_info[hw_thread_index].os_id = proc;
2909 proc_info[hw_thread_index].apic_id = buf.edx;
2910 __kmp_x2apicid_get_levels(topology_leaf, &proc_info[hw_thread_index], types,
2911 &depth, &total_description);
2912 if (proc_info[hw_thread_index].depth == 0) {
2913 *msg_id = kmp_i18n_str_InvalidCpuidInfo;
2914 __kmp_free(proc_info);
2915 return false;
2916 }
2917 // Gather cache information and insert afterwards
2918 cache_info[hw_thread_index].get_leaf4_levels();
2919 if (uniform_caches && hw_thread_index > 0)
2920 if (cache_info[0] != cache_info[hw_thread_index])
2921 uniform_caches = false;
2922 // Hybrid information
2923 if (__kmp_is_hybrid_cpu() && highest_leaf >= 0x1a) {
2924 __kmp_get_hybrid_info(&proc_info[hw_thread_index].type,
2925 &proc_info[hw_thread_index].efficiency,
2926 &proc_info[hw_thread_index].native_model_id);
2927 }
2928 hw_thread_index++;
2929 }
2930 KMP_ASSERT(hw_thread_index > 0);
2931 previous_affinity.restore();
2932
2933 // Allocate the data structure to be returned.
2934 __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
2935
2936 // Create topology Ids and hybrid types in __kmp_topology
2937 for (int i = 0; i < __kmp_topology->get_num_hw_threads(); ++i) {
2938 kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
2939 hw_thread.clear();
2940 hw_thread.os_id = proc_info[i].os_id;
2941 hw_thread.original_idx = i;
2942 unsigned apic_id = proc_info[i].apic_id;
2943 // Put in topology information
2944 for (int j = 0, idx = depth - 1; j < depth; ++j, --idx) {
2945 if (!(proc_info[i].description.contains_topology_type(
2946 __kmp_topology->get_type(j)))) {
2947 hw_thread.ids[idx] = kmp_hw_thread_t::UNKNOWN_ID;
2948 } else {
2949 hw_thread.ids[idx] = apic_id & proc_info[i].levels[j].mask;
2950 if (j > 0) {
2951 hw_thread.ids[idx] >>= proc_info[i].levels[j - 1].mask_width;
2952 }
2953 }
2954 }
2955 hw_thread.attrs.set_core_type(proc_info[i].type);
2956 hw_thread.attrs.set_core_eff(proc_info[i].efficiency);
2957 }
2958
2959 __kmp_topology->sort_ids();
2960
2961 // Change Ids to logical Ids
2962 for (int j = 0; j < depth - 1; ++j) {
2963 int new_id = 0;
2964 int prev_id = __kmp_topology->at(0).ids[j];
2965 int curr_id = __kmp_topology->at(0).ids[j + 1];
2966 __kmp_topology->at(0).ids[j + 1] = new_id;
2967 for (int i = 1; i < __kmp_topology->get_num_hw_threads(); ++i) {
2968 kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
2969 if (hw_thread.ids[j] == prev_id && hw_thread.ids[j + 1] == curr_id) {
2970 hw_thread.ids[j + 1] = new_id;
2971 } else if (hw_thread.ids[j] == prev_id &&
2972 hw_thread.ids[j + 1] != curr_id) {
2973 curr_id = hw_thread.ids[j + 1];
2974 hw_thread.ids[j + 1] = ++new_id;
2975 } else {
2976 prev_id = hw_thread.ids[j];
2977 curr_id = hw_thread.ids[j + 1];
2978 hw_thread.ids[j + 1] = ++new_id;
2979 }
2980 }
2981 }
2982
2983 // First check for easy cache placement. This occurs when caches are
2984 // equivalent to a layer in the CPUID leaf 0xb or 0x1f topology.
2985 if (uniform_caches) {
2986 for (size_t i = 0; i < cache_info[0].get_depth(); ++i) {
2987 unsigned cache_mask = cache_info[0][i].mask;
2988 unsigned cache_level = cache_info[0][i].level;
2989 KMP_ASSERT(cache_level <= cpuid_cache_info_t::MAX_CACHE_LEVEL);
2990 kmp_hw_t cache_type = cpuid_cache_info_t::get_topology_type(cache_level);
2991 __kmp_topology->set_equivalent_type(cache_type, cache_type);
2992 for (int j = 0; j < depth; ++j) {
2993 unsigned hw_cache_mask = proc_info[0].levels[j].cache_mask;
2994 if (hw_cache_mask == cache_mask && j < depth - 1) {
2995 kmp_hw_t type = __kmp_intel_type_2_topology_type(
2996 proc_info[0].levels[j + 1].level_type);
2997 __kmp_topology->set_equivalent_type(cache_type, type);
2998 }
2999 }
3000 }
3001 } else {
3002 // If caches are non-uniform, then record which caches exist.
3003 for (int i = 0; i < __kmp_topology->get_num_hw_threads(); ++i) {
3004 for (size_t j = 0; j < cache_info[i].get_depth(); ++j) {
3005 unsigned cache_level = cache_info[i][j].level;
3006 kmp_hw_t cache_type =
3007 cpuid_cache_info_t::get_topology_type(cache_level);
3008 if (__kmp_topology->get_equivalent_type(cache_type) == KMP_HW_UNKNOWN)
3009 __kmp_topology->set_equivalent_type(cache_type, cache_type);
3010 }
3011 }
3012 }
3013
3014 // See if any cache level needs to be added manually through cache Ids
3015 bool unresolved_cache_levels = false;
3016 for (unsigned level = 1; level <= cpuid_cache_info_t::MAX_CACHE_LEVEL;
3017 ++level) {
3018 kmp_hw_t cache_type = cpuid_cache_info_t::get_topology_type(level);
3019 // This also filters out caches which may not be in the topology
3020 // since the equivalent type might be KMP_HW_UNKNOWN.
3021 if (__kmp_topology->get_equivalent_type(cache_type) == cache_type) {
3022 unresolved_cache_levels = true;
3023 break;
3024 }
3025 }
3026
3027 // Insert unresolved cache layers into machine topology using cache Ids
3028 if (unresolved_cache_levels) {
3029 int num_hw_threads = __kmp_topology->get_num_hw_threads();
3030 int *ids = (int *)__kmp_allocate(sizeof(int) * num_hw_threads);
3031 for (unsigned l = 1; l <= cpuid_cache_info_t::MAX_CACHE_LEVEL; ++l) {
3032 kmp_hw_t cache_type = cpuid_cache_info_t::get_topology_type(l);
3033 if (__kmp_topology->get_equivalent_type(cache_type) != cache_type)
3034 continue;
3035 for (int i = 0; i < num_hw_threads; ++i) {
3036 int original_idx = __kmp_topology->at(i).original_idx;
3037 ids[i] = kmp_hw_thread_t::UNKNOWN_ID;
3038 const cpuid_cache_info_t::info_t &info =
3039 cache_info[original_idx].get_level(l);
3040 // if cache level not in topology for this processor, then skip
3041 if (info.level == 0)
3042 continue;
3043 ids[i] = info.mask & proc_info[original_idx].apic_id;
3044 }
3045 __kmp_topology->insert_layer(cache_type, ids);
3046 }
3047 }
3048
3049 if (!__kmp_topology->check_ids()) {
3050 kmp_topology_t::deallocate(__kmp_topology);
3051 __kmp_topology = nullptr;
3052 *msg_id = kmp_i18n_str_x2ApicIDsNotUnique;
3053 __kmp_free(proc_info);
3054 return false;
3055 }
3056 __kmp_free(proc_info);
3057 return true;
3058}
3059#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3060
3061#define osIdIndex 0
3062#define threadIdIndex 1
3063#define coreIdIndex 2
3064#define pkgIdIndex 3
3065#define nodeIdIndex 4
3066
3067typedef unsigned *ProcCpuInfo;
3068static unsigned maxIndex = pkgIdIndex;
3069
3070static int __kmp_affinity_cmp_ProcCpuInfo_phys_id(const void *a,
3071 const void *b) {
3072 unsigned i;
3073 const unsigned *aa = *(unsigned *const *)a;
3074 const unsigned *bb = *(unsigned *const *)b;
3075 for (i = maxIndex;; i--) {
3076 if (aa[i] < bb[i])
3077 return -1;
3078 if (aa[i] > bb[i])
3079 return 1;
3080 if (i == osIdIndex)
3081 break;
3082 }
3083 return 0;
3084}
3085
3086#if KMP_USE_HIER_SCHED
3087// Set the array sizes for the hierarchy layers
3088static void __kmp_dispatch_set_hierarchy_values() {
3089 // Set the maximum number of L1's to number of cores
3090 // Set the maximum number of L2's to either number of cores / 2 for
3091 // Intel(R) Xeon Phi(TM) coprocessor formally codenamed Knights Landing
3092 // Or the number of cores for Intel(R) Xeon(R) processors
3093 // Set the maximum number of NUMA nodes and L3's to number of packages
3094 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1] =
3095 nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
3096 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L1 + 1] = __kmp_ncores;
3097#if KMP_ARCH_X86_64 && \
3098 (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_DRAGONFLY || \
3099 KMP_OS_WINDOWS) && \
3100 KMP_MIC_SUPPORTED
3101 if (__kmp_mic_type >= mic3)
3102 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores / 2;
3103 else
3104#endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
3105 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores;
3106 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L3 + 1] = nPackages;
3107 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_NUMA + 1] = nPackages;
3108 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_LOOP + 1] = 1;
3109 // Set the number of threads per unit
3110 // Number of hardware threads per L1/L2/L3/NUMA/LOOP
3111 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_THREAD + 1] = 1;
3112 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L1 + 1] =
3113 __kmp_nThreadsPerCore;
3114#if KMP_ARCH_X86_64 && \
3115 (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_DRAGONFLY || \
3116 KMP_OS_WINDOWS) && \
3117 KMP_MIC_SUPPORTED
3118 if (__kmp_mic_type >= mic3)
3119 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
3120 2 * __kmp_nThreadsPerCore;
3121 else
3122#endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
3123 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
3124 __kmp_nThreadsPerCore;
3125 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L3 + 1] =
3126 nCoresPerPkg * __kmp_nThreadsPerCore;
3127 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_NUMA + 1] =
3128 nCoresPerPkg * __kmp_nThreadsPerCore;
3129 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_LOOP + 1] =
3130 nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
3131}
3132
3133// Return the index into the hierarchy for this tid and layer type (L1, L2, etc)
3134// i.e., this thread's L1 or this thread's L2, etc.
3135int __kmp_dispatch_get_index(int tid, kmp_hier_layer_e type) {
3136 int index = type + 1;
3137 int num_hw_threads = __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1];
3138 KMP_DEBUG_ASSERT(type != kmp_hier_layer_e::LAYER_LAST);
3139 if (type == kmp_hier_layer_e::LAYER_THREAD)
3140 return tid;
3141 else if (type == kmp_hier_layer_e::LAYER_LOOP)
3142 return 0;
3143 KMP_DEBUG_ASSERT(__kmp_hier_max_units[index] != 0);
3144 if (tid >= num_hw_threads)
3145 tid = tid % num_hw_threads;
3146 return (tid / __kmp_hier_threads_per[index]) % __kmp_hier_max_units[index];
3147}
3148
3149// Return the number of t1's per t2
3150int __kmp_dispatch_get_t1_per_t2(kmp_hier_layer_e t1, kmp_hier_layer_e t2) {
3151 int i1 = t1 + 1;
3152 int i2 = t2 + 1;
3153 KMP_DEBUG_ASSERT(i1 <= i2);
3154 KMP_DEBUG_ASSERT(t1 != kmp_hier_layer_e::LAYER_LAST);
3155 KMP_DEBUG_ASSERT(t2 != kmp_hier_layer_e::LAYER_LAST);
3156 KMP_DEBUG_ASSERT(__kmp_hier_threads_per[i1] != 0);
3157 // (nthreads/t2) / (nthreads/t1) = t1 / t2
3158 return __kmp_hier_threads_per[i2] / __kmp_hier_threads_per[i1];
3159}
3160#endif // KMP_USE_HIER_SCHED
3161
3162static inline const char *__kmp_cpuinfo_get_filename() {
3163 const char *filename;
3164 if (__kmp_cpuinfo_file != nullptr)
3165 filename = __kmp_cpuinfo_file;
3166 else
3167 filename = "/proc/cpuinfo";
3168 return filename;
3169}
3170
3171static inline const char *__kmp_cpuinfo_get_envvar() {
3172 const char *envvar = nullptr;
3173 if (__kmp_cpuinfo_file != nullptr)
3174 envvar = "KMP_CPUINFO_FILE";
3175 return envvar;
3176}
3177
3178// Parse /proc/cpuinfo (or an alternate file in the same format) to obtain the
3179// affinity map. On AIX, the map is obtained through system SRAD (Scheduler
3180// Resource Allocation Domain).
3181static bool __kmp_affinity_create_cpuinfo_map(int *line,
3182 kmp_i18n_id_t *const msg_id) {
3183 *msg_id = kmp_i18n_null;
3184
3185#if KMP_OS_AIX
3186 unsigned num_records = __kmp_xproc;
3187#else
3188 const char *filename = __kmp_cpuinfo_get_filename();
3189 const char *envvar = __kmp_cpuinfo_get_envvar();
3190
3191 if (__kmp_affinity.flags.verbose) {
3192 KMP_INFORM(AffParseFilename, "KMP_AFFINITY", filename);
3193 }
3194
3195 kmp_safe_raii_file_t f(filename, "r", envvar);
3196
3197 // Scan of the file, and count the number of "processor" (osId) fields,
3198 // and find the highest value of <n> for a node_<n> field.
3199 char buf[256];
3200 unsigned num_records = 0;
3201 while (!feof(f)) {
3202 buf[sizeof(buf) - 1] = 1;
3203 if (!fgets(buf, sizeof(buf), f)) {
3204 // Read errors presumably because of EOF
3205 break;
3206 }
3207
3208 char s1[] = "processor";
3209 if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
3210 num_records++;
3211 continue;
3212 }
3213
3214 // FIXME - this will match "node_<n> <garbage>"
3215 unsigned level;
3216 if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
3217 // validate the input fisrt:
3218 if (level > (unsigned)__kmp_xproc) { // level is too big
3219 level = __kmp_xproc;
3220 }
3221 if (nodeIdIndex + level >= maxIndex) {
3222 maxIndex = nodeIdIndex + level;
3223 }
3224 continue;
3225 }
3226 }
3227
3228 // Check for empty file / no valid processor records, or too many. The number
3229 // of records can't exceed the number of valid bits in the affinity mask.
3230 if (num_records == 0) {
3231 *msg_id = kmp_i18n_str_NoProcRecords;
3232 return false;
3233 }
3234 if (num_records > (unsigned)__kmp_xproc) {
3235 *msg_id = kmp_i18n_str_TooManyProcRecords;
3236 return false;
3237 }
3238
3239 // Set the file pointer back to the beginning, so that we can scan the file
3240 // again, this time performing a full parse of the data. Allocate a vector of
3241 // ProcCpuInfo object, where we will place the data. Adding an extra element
3242 // at the end allows us to remove a lot of extra checks for termination
3243 // conditions.
3244 if (fseek(f, 0, SEEK_SET) != 0) {
3245 *msg_id = kmp_i18n_str_CantRewindCpuinfo;
3246 return false;
3247 }
3248#endif // KMP_OS_AIX
3249
3250 // Allocate the array of records to store the proc info in. The dummy
3251 // element at the end makes the logic in filling them out easier to code.
3252 unsigned **threadInfo =
3253 (unsigned **)__kmp_allocate((num_records + 1) * sizeof(unsigned *));
3254 unsigned i;
3255 for (i = 0; i <= num_records; i++) {
3256 threadInfo[i] =
3257 (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3258 }
3259
3260#define CLEANUP_THREAD_INFO \
3261 for (i = 0; i <= num_records; i++) { \
3262 __kmp_free(threadInfo[i]); \
3263 } \
3264 __kmp_free(threadInfo);
3265
3266 // A value of UINT_MAX means that we didn't find the field
3267 unsigned __index;
3268
3269#define INIT_PROC_INFO(p) \
3270 for (__index = 0; __index <= maxIndex; __index++) { \
3271 (p)[__index] = UINT_MAX; \
3272 }
3273
3274 for (i = 0; i <= num_records; i++) {
3275 INIT_PROC_INFO(threadInfo[i]);
3276 }
3277
3278#if KMP_OS_AIX
3279 int smt_threads;
3280 lpar_info_format1_t cpuinfo;
3281 unsigned num_avail = __kmp_xproc;
3282
3283 if (__kmp_affinity.flags.verbose)
3284 KMP_INFORM(AffParseFilename, "KMP_AFFINITY", "system info for topology");
3285
3286 // Get the number of SMT threads per core.
3287 smt_threads = syssmt(GET_NUMBER_SMT_SETS, 0, 0, NULL);
3288
3289 // Allocate a resource set containing available system resourses.
3290 rsethandle_t sys_rset = rs_alloc(RS_SYSTEM);
3291 if (sys_rset == NULL) {
3292 CLEANUP_THREAD_INFO;
3293 *msg_id = kmp_i18n_str_UnknownTopology;
3294 return false;
3295 }
3296 // Allocate a resource set for the SRAD info.
3297 rsethandle_t srad = rs_alloc(RS_EMPTY);
3298 if (srad == NULL) {
3299 rs_free(sys_rset);
3300 CLEANUP_THREAD_INFO;
3301 *msg_id = kmp_i18n_str_UnknownTopology;
3302 return false;
3303 }
3304
3305 // Get the SRAD system detail level.
3306 int sradsdl = rs_getinfo(NULL, R_SRADSDL, 0);
3307 if (sradsdl < 0) {
3308 rs_free(sys_rset);
3309 rs_free(srad);
3310 CLEANUP_THREAD_INFO;
3311 *msg_id = kmp_i18n_str_UnknownTopology;
3312 return false;
3313 }
3314 // Get the number of RADs at that SRAD SDL.
3315 int num_rads = rs_numrads(sys_rset, sradsdl, 0);
3316 if (num_rads < 0) {
3317 rs_free(sys_rset);
3318 rs_free(srad);
3319 CLEANUP_THREAD_INFO;
3320 *msg_id = kmp_i18n_str_UnknownTopology;
3321 return false;
3322 }
3323
3324 // Get the maximum number of procs that may be contained in a resource set.
3325 int max_procs = rs_getinfo(NULL, R_MAXPROCS, 0);
3326 if (max_procs < 0) {
3327 rs_free(sys_rset);
3328 rs_free(srad);
3329 CLEANUP_THREAD_INFO;
3330 *msg_id = kmp_i18n_str_UnknownTopology;
3331 return false;
3332 }
3333
3334 int cur_rad = 0;
3335 int num_set = 0;
3336 for (int srad_idx = 0; cur_rad < num_rads && srad_idx < VMI_MAXRADS;
3337 ++srad_idx) {
3338 // Check if the SRAD is available in the RSET.
3339 if (rs_getrad(sys_rset, srad, sradsdl, srad_idx, 0) < 0)
3340 continue;
3341
3342 for (int cpu = 0; cpu < max_procs; cpu++) {
3343 // Set the info for the cpu if it is in the SRAD.
3344 if (rs_op(RS_TESTRESOURCE, srad, NULL, R_PROCS, cpu)) {
3345 threadInfo[cpu][osIdIndex] = cpu;
3346 threadInfo[cpu][pkgIdIndex] = cur_rad;
3347 threadInfo[cpu][coreIdIndex] = cpu / smt_threads;
3348 ++num_set;
3349 if (num_set >= num_avail) {
3350 // Done if all available CPUs have been set.
3351 break;
3352 }
3353 }
3354 }
3355 ++cur_rad;
3356 }
3357 rs_free(sys_rset);
3358 rs_free(srad);
3359
3360 // The topology is already sorted.
3361
3362#else // !KMP_OS_AIX
3363 unsigned num_avail = 0;
3364 *line = 0;
3365#if KMP_ARCH_S390X
3366 bool reading_s390x_sys_info = true;
3367#endif
3368 while (!feof(f)) {
3369 // Create an inner scoping level, so that all the goto targets at the end of
3370 // the loop appear in an outer scoping level. This avoids warnings about
3371 // jumping past an initialization to a target in the same block.
3372 {
3373 buf[sizeof(buf) - 1] = 1;
3374 bool long_line = false;
3375 if (!fgets(buf, sizeof(buf), f)) {
3376 // Read errors presumably because of EOF
3377 // If there is valid data in threadInfo[num_avail], then fake
3378 // a blank line in ensure that the last address gets parsed.
3379 bool valid = false;
3380 for (i = 0; i <= maxIndex; i++) {
3381 if (threadInfo[num_avail][i] != UINT_MAX) {
3382 valid = true;
3383 }
3384 }
3385 if (!valid) {
3386 break;
3387 }
3388 buf[0] = 0;
3389 } else if (!buf[sizeof(buf) - 1]) {
3390 // The line is longer than the buffer. Set a flag and don't
3391 // emit an error if we were going to ignore the line, anyway.
3392 long_line = true;
3393
3394#define CHECK_LINE \
3395 if (long_line) { \
3396 CLEANUP_THREAD_INFO; \
3397 *msg_id = kmp_i18n_str_LongLineCpuinfo; \
3398 return false; \
3399 }
3400 }
3401 (*line)++;
3402
3403#if KMP_ARCH_LOONGARCH64
3404 // The parsing logic of /proc/cpuinfo in this function highly depends on
3405 // the blank lines between each processor info block. But on LoongArch a
3406 // blank line exists before the first processor info block (i.e. after the
3407 // "system type" line). This blank line was added because the "system
3408 // type" line is unrelated to any of the CPUs. We must skip this line so
3409 // that the original logic works on LoongArch.
3410 if (*buf == '\n' && *line == 2)
3411 continue;
3412#endif
3413#if KMP_ARCH_S390X
3414 // s390x /proc/cpuinfo starts with a variable number of lines containing
3415 // the overall system information. Skip them.
3416 if (reading_s390x_sys_info) {
3417 if (*buf == '\n')
3418 reading_s390x_sys_info = false;
3419 continue;
3420 }
3421#endif
3422
3423#if KMP_ARCH_S390X
3424 char s1[] = "cpu number";
3425#else
3426 char s1[] = "processor";
3427#endif
3428 if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
3429 CHECK_LINE;
3430 char *p = strchr(buf + sizeof(s1) - 1, ':');
3431 unsigned val;
3432 if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3433 goto no_val;
3434 if (threadInfo[num_avail][osIdIndex] != UINT_MAX)
3435#if KMP_ARCH_AARCH64
3436 // Handle the old AArch64 /proc/cpuinfo layout differently,
3437 // it contains all of the 'processor' entries listed in a
3438 // single 'Processor' section, therefore the normal looking
3439 // for duplicates in that section will always fail.
3440 num_avail++;
3441#else
3442 goto dup_field;
3443#endif
3444 threadInfo[num_avail][osIdIndex] = val;
3445#if KMP_OS_LINUX && !(KMP_ARCH_X86 || KMP_ARCH_X86_64)
3446 char path[256];
3447 KMP_SNPRINTF(
3448 path, sizeof(path),
3449 "/sys/devices/system/cpu/cpu%u/topology/physical_package_id",
3450 threadInfo[num_avail][osIdIndex]);
3451 __kmp_read_from_file(path, "%u", &threadInfo[num_avail][pkgIdIndex]);
3452
3453#if KMP_ARCH_S390X
3454 // Disambiguate physical_package_id.
3455 unsigned book_id;
3456 KMP_SNPRINTF(path, sizeof(path),
3457 "/sys/devices/system/cpu/cpu%u/topology/book_id",
3458 threadInfo[num_avail][osIdIndex]);
3459 __kmp_read_from_file(path, "%u", &book_id);
3460 threadInfo[num_avail][pkgIdIndex] |= (book_id << 8);
3461
3462 unsigned drawer_id;
3463 KMP_SNPRINTF(path, sizeof(path),
3464 "/sys/devices/system/cpu/cpu%u/topology/drawer_id",
3465 threadInfo[num_avail][osIdIndex]);
3466 __kmp_read_from_file(path, "%u", &drawer_id);
3467 threadInfo[num_avail][pkgIdIndex] |= (drawer_id << 16);
3468#endif
3469
3470 KMP_SNPRINTF(path, sizeof(path),
3471 "/sys/devices/system/cpu/cpu%u/topology/core_id",
3472 threadInfo[num_avail][osIdIndex]);
3473 __kmp_read_from_file(path, "%u", &threadInfo[num_avail][coreIdIndex]);
3474 continue;
3475#else
3476 }
3477 char s2[] = "physical id";
3478 if (strncmp(buf, s2, sizeof(s2) - 1) == 0) {
3479 CHECK_LINE;
3480 char *p = strchr(buf + sizeof(s2) - 1, ':');
3481 unsigned val;
3482 if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3483 goto no_val;
3484 if (threadInfo[num_avail][pkgIdIndex] != UINT_MAX)
3485 goto dup_field;
3486 threadInfo[num_avail][pkgIdIndex] = val;
3487 continue;
3488 }
3489 char s3[] = "core id";
3490 if (strncmp(buf, s3, sizeof(s3) - 1) == 0) {
3491 CHECK_LINE;
3492 char *p = strchr(buf + sizeof(s3) - 1, ':');
3493 unsigned val;
3494 if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3495 goto no_val;
3496 if (threadInfo[num_avail][coreIdIndex] != UINT_MAX)
3497 goto dup_field;
3498 threadInfo[num_avail][coreIdIndex] = val;
3499 continue;
3500#endif // KMP_OS_LINUX && USE_SYSFS_INFO
3501 }
3502 char s4[] = "thread id";
3503 if (strncmp(buf, s4, sizeof(s4) - 1) == 0) {
3504 CHECK_LINE;
3505 char *p = strchr(buf + sizeof(s4) - 1, ':');
3506 unsigned val;
3507 if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3508 goto no_val;
3509 if (threadInfo[num_avail][threadIdIndex] != UINT_MAX)
3510 goto dup_field;
3511 threadInfo[num_avail][threadIdIndex] = val;
3512 continue;
3513 }
3514 unsigned level;
3515 if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
3516 CHECK_LINE;
3517 char *p = strchr(buf + sizeof(s4) - 1, ':');
3518 unsigned val;
3519 if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3520 goto no_val;
3521 // validate the input before using level:
3522 if (level > (unsigned)__kmp_xproc) { // level is too big
3523 level = __kmp_xproc;
3524 }
3525 if (threadInfo[num_avail][nodeIdIndex + level] != UINT_MAX)
3526 goto dup_field;
3527 threadInfo[num_avail][nodeIdIndex + level] = val;
3528 continue;
3529 }
3530
3531 // We didn't recognize the leading token on the line. There are lots of
3532 // leading tokens that we don't recognize - if the line isn't empty, go on
3533 // to the next line.
3534 if ((*buf != 0) && (*buf != '\n')) {
3535 // If the line is longer than the buffer, read characters
3536 // until we find a newline.
3537 if (long_line) {
3538 int ch;
3539 while (((ch = fgetc(f)) != EOF) && (ch != '\n'))
3540 ;
3541 }
3542 continue;
3543 }
3544
3545 // A newline has signalled the end of the processor record.
3546 // Check that there aren't too many procs specified.
3547 if ((int)num_avail == __kmp_xproc) {
3548 CLEANUP_THREAD_INFO;
3549 *msg_id = kmp_i18n_str_TooManyEntries;
3550 return false;
3551 }
3552
3553 // Check for missing fields. The osId field must be there, and we
3554 // currently require that the physical id field is specified, also.
3555 if (threadInfo[num_avail][osIdIndex] == UINT_MAX) {
3556 CLEANUP_THREAD_INFO;
3557 *msg_id = kmp_i18n_str_MissingProcField;
3558 return false;
3559 }
3560 if (threadInfo[0][pkgIdIndex] == UINT_MAX) {
3561 CLEANUP_THREAD_INFO;
3562 *msg_id = kmp_i18n_str_MissingPhysicalIDField;
3563 return false;
3564 }
3565
3566 // Skip this proc if it is not included in the machine model.
3567 if (KMP_AFFINITY_CAPABLE() &&
3568 !KMP_CPU_ISSET(threadInfo[num_avail][osIdIndex],
3569 __kmp_affin_fullMask)) {
3570 INIT_PROC_INFO(threadInfo[num_avail]);
3571 continue;
3572 }
3573
3574 // We have a successful parse of this proc's info.
3575 // Increment the counter, and prepare for the next proc.
3576 num_avail++;
3577 KMP_ASSERT(num_avail <= num_records);
3578 INIT_PROC_INFO(threadInfo[num_avail]);
3579 }
3580 continue;
3581
3582 no_val:
3583 CLEANUP_THREAD_INFO;
3584 *msg_id = kmp_i18n_str_MissingValCpuinfo;
3585 return false;
3586
3587 dup_field:
3588 CLEANUP_THREAD_INFO;
3589 *msg_id = kmp_i18n_str_DuplicateFieldCpuinfo;
3590 return false;
3591 }
3592 *line = 0;
3593
3594#if KMP_MIC && REDUCE_TEAM_SIZE
3595 unsigned teamSize = 0;
3596#endif // KMP_MIC && REDUCE_TEAM_SIZE
3597
3598 // check for num_records == __kmp_xproc ???
3599
3600 // If it is configured to omit the package level when there is only a single
3601 // package, the logic at the end of this routine won't work if there is only a
3602 // single thread
3603 KMP_ASSERT(num_avail > 0);
3604 KMP_ASSERT(num_avail <= num_records);
3605
3606 // Sort the threadInfo table by physical Id.
3607 qsort(threadInfo, num_avail, sizeof(*threadInfo),
3608 __kmp_affinity_cmp_ProcCpuInfo_phys_id);
3609
3610#endif // KMP_OS_AIX
3611
3612 // The table is now sorted by pkgId / coreId / threadId, but we really don't
3613 // know the radix of any of the fields. pkgId's may be sparsely assigned among
3614 // the chips on a system. Although coreId's are usually assigned
3615 // [0 .. coresPerPkg-1] and threadId's are usually assigned
3616 // [0..threadsPerCore-1], we don't want to make any such assumptions.
3617 //
3618 // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
3619 // total # packages) are at this point - we want to determine that now. We
3620 // only have an upper bound on the first two figures.
3621 unsigned *counts =
3622 (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3623 unsigned *maxCt =
3624 (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3625 unsigned *totals =
3626 (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3627 unsigned *lastId =
3628 (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3629
3630 bool assign_thread_ids = false;
3631 unsigned threadIdCt;
3632 unsigned index;
3633
3634restart_radix_check:
3635 threadIdCt = 0;
3636
3637 // Initialize the counter arrays with data from threadInfo[0].
3638 if (assign_thread_ids) {
3639 if (threadInfo[0][threadIdIndex] == UINT_MAX) {
3640 threadInfo[0][threadIdIndex] = threadIdCt++;
3641 } else if (threadIdCt <= threadInfo[0][threadIdIndex]) {
3642 threadIdCt = threadInfo[0][threadIdIndex] + 1;
3643 }
3644 }
3645 for (index = 0; index <= maxIndex; index++) {
3646 counts[index] = 1;
3647 maxCt[index] = 1;
3648 totals[index] = 1;
3649 lastId[index] = threadInfo[0][index];
3650 ;
3651 }
3652
3653 // Run through the rest of the OS procs.
3654 for (i = 1; i < num_avail; i++) {
3655 // Find the most significant index whose id differs from the id for the
3656 // previous OS proc.
3657 for (index = maxIndex; index >= threadIdIndex; index--) {
3658 if (assign_thread_ids && (index == threadIdIndex)) {
3659 // Auto-assign the thread id field if it wasn't specified.
3660 if (threadInfo[i][threadIdIndex] == UINT_MAX) {
3661 threadInfo[i][threadIdIndex] = threadIdCt++;
3662 }
3663 // Apparently the thread id field was specified for some entries and not
3664 // others. Start the thread id counter off at the next higher thread id.
3665 else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
3666 threadIdCt = threadInfo[i][threadIdIndex] + 1;
3667 }
3668 }
3669 if (threadInfo[i][index] != lastId[index]) {
3670 // Run through all indices which are less significant, and reset the
3671 // counts to 1. At all levels up to and including index, we need to
3672 // increment the totals and record the last id.
3673 unsigned index2;
3674 for (index2 = threadIdIndex; index2 < index; index2++) {
3675 totals[index2]++;
3676 if (counts[index2] > maxCt[index2]) {
3677 maxCt[index2] = counts[index2];
3678 }
3679 counts[index2] = 1;
3680 lastId[index2] = threadInfo[i][index2];
3681 }
3682 counts[index]++;
3683 totals[index]++;
3684 lastId[index] = threadInfo[i][index];
3685
3686 if (assign_thread_ids && (index > threadIdIndex)) {
3687
3688#if KMP_MIC && REDUCE_TEAM_SIZE
3689 // The default team size is the total #threads in the machine
3690 // minus 1 thread for every core that has 3 or more threads.
3691 teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
3692#endif // KMP_MIC && REDUCE_TEAM_SIZE
3693
3694 // Restart the thread counter, as we are on a new core.
3695 threadIdCt = 0;
3696
3697 // Auto-assign the thread id field if it wasn't specified.
3698 if (threadInfo[i][threadIdIndex] == UINT_MAX) {
3699 threadInfo[i][threadIdIndex] = threadIdCt++;
3700 }
3701
3702 // Apparently the thread id field was specified for some entries and
3703 // not others. Start the thread id counter off at the next higher
3704 // thread id.
3705 else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
3706 threadIdCt = threadInfo[i][threadIdIndex] + 1;
3707 }
3708 }
3709 break;
3710 }
3711 }
3712 if (index < threadIdIndex) {
3713 // If thread ids were specified, it is an error if they are not unique.
3714 // Also, check that we waven't already restarted the loop (to be safe -
3715 // shouldn't need to).
3716 if ((threadInfo[i][threadIdIndex] != UINT_MAX) || assign_thread_ids) {
3717 __kmp_free(lastId);
3718 __kmp_free(totals);
3719 __kmp_free(maxCt);
3720 __kmp_free(counts);
3721 CLEANUP_THREAD_INFO;
3722 *msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
3723 return false;
3724 }
3725
3726 // If the thread ids were not specified and we see entries that
3727 // are duplicates, start the loop over and assign the thread ids manually.
3728 assign_thread_ids = true;
3729 goto restart_radix_check;
3730 }
3731 }
3732
3733#if KMP_MIC && REDUCE_TEAM_SIZE
3734 // The default team size is the total #threads in the machine
3735 // minus 1 thread for every core that has 3 or more threads.
3736 teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
3737#endif // KMP_MIC && REDUCE_TEAM_SIZE
3738
3739 for (index = threadIdIndex; index <= maxIndex; index++) {
3740 if (counts[index] > maxCt[index]) {
3741 maxCt[index] = counts[index];
3742 }
3743 }
3744
3745 __kmp_nThreadsPerCore = maxCt[threadIdIndex];
3746 nCoresPerPkg = maxCt[coreIdIndex];
3747 nPackages = totals[pkgIdIndex];
3748
3749 // When affinity is off, this routine will still be called to set
3750 // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
3751 // Make sure all these vars are set correctly, and return now if affinity is
3752 // not enabled.
3753 __kmp_ncores = totals[coreIdIndex];
3754 if (!KMP_AFFINITY_CAPABLE()) {
3755 KMP_ASSERT(__kmp_affinity.type == affinity_none);
3756 return true;
3757 }
3758
3759#if KMP_MIC && REDUCE_TEAM_SIZE
3760 // Set the default team size.
3761 if ((__kmp_dflt_team_nth == 0) && (teamSize > 0)) {
3762 __kmp_dflt_team_nth = teamSize;
3763 KA_TRACE(20, ("__kmp_affinity_create_cpuinfo_map: setting "
3764 "__kmp_dflt_team_nth = %d\n",
3765 __kmp_dflt_team_nth));
3766 }
3767#endif // KMP_MIC && REDUCE_TEAM_SIZE
3768
3769 KMP_DEBUG_ASSERT(num_avail == (unsigned)__kmp_avail_proc);
3770
3771 // Count the number of levels which have more nodes at that level than at the
3772 // parent's level (with there being an implicit root node of the top level).
3773 // This is equivalent to saying that there is at least one node at this level
3774 // which has a sibling. These levels are in the map, and the package level is
3775 // always in the map.
3776 bool *inMap = (bool *)__kmp_allocate((maxIndex + 1) * sizeof(bool));
3777 for (index = threadIdIndex; index < maxIndex; index++) {
3778 KMP_ASSERT(totals[index] >= totals[index + 1]);
3779 inMap[index] = (totals[index] > totals[index + 1]);
3780 }
3781 inMap[maxIndex] = (totals[maxIndex] > 1);
3782 inMap[pkgIdIndex] = true;
3783 inMap[coreIdIndex] = true;
3784 inMap[threadIdIndex] = true;
3785
3786 int depth = 0;
3787 int idx = 0;
3788 kmp_hw_t types[KMP_HW_LAST];
3789 int pkgLevel = -1;
3790 int coreLevel = -1;
3791 int threadLevel = -1;
3792 for (index = threadIdIndex; index <= maxIndex; index++) {
3793 if (inMap[index]) {
3794 depth++;
3795 }
3796 }
3797 if (inMap[pkgIdIndex]) {
3798 pkgLevel = idx;
3799 types[idx++] = KMP_HW_SOCKET;
3800 }
3801 if (inMap[coreIdIndex]) {
3802 coreLevel = idx;
3803 types[idx++] = KMP_HW_CORE;
3804 }
3805 if (inMap[threadIdIndex]) {
3806 threadLevel = idx;
3807 types[idx++] = KMP_HW_THREAD;
3808 }
3809 KMP_ASSERT(depth > 0);
3810
3811 // Construct the data structure that is to be returned.
3812 __kmp_topology = kmp_topology_t::allocate(num_avail, depth, types);
3813
3814 for (i = 0; i < num_avail; ++i) {
3815 unsigned os = threadInfo[i][osIdIndex];
3816 int src_index;
3817 kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
3818 hw_thread.clear();
3819 hw_thread.os_id = os;
3820 hw_thread.original_idx = i;
3821
3822 idx = 0;
3823 for (src_index = maxIndex; src_index >= threadIdIndex; src_index--) {
3824 if (!inMap[src_index]) {
3825 continue;
3826 }
3827 if (src_index == pkgIdIndex) {
3828 hw_thread.ids[pkgLevel] = threadInfo[i][src_index];
3829 } else if (src_index == coreIdIndex) {
3830 hw_thread.ids[coreLevel] = threadInfo[i][src_index];
3831 } else if (src_index == threadIdIndex) {
3832 hw_thread.ids[threadLevel] = threadInfo[i][src_index];
3833 }
3834 }
3835 }
3836
3837 __kmp_free(inMap);
3838 __kmp_free(lastId);
3839 __kmp_free(totals);
3840 __kmp_free(maxCt);
3841 __kmp_free(counts);
3842 CLEANUP_THREAD_INFO;
3843 __kmp_topology->sort_ids();
3844
3845 int tlevel = __kmp_topology->get_level(KMP_HW_THREAD);
3846 if (tlevel > 0) {
3847 // If the thread level does not have ids, then put them in.
3848 if (__kmp_topology->at(0).ids[tlevel] == kmp_hw_thread_t::UNKNOWN_ID) {
3849 __kmp_topology->at(0).ids[tlevel] = 0;
3850 }
3851 for (int i = 1; i < __kmp_topology->get_num_hw_threads(); ++i) {
3852 kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
3853 if (hw_thread.ids[tlevel] != kmp_hw_thread_t::UNKNOWN_ID)
3854 continue;
3855 kmp_hw_thread_t &prev_hw_thread = __kmp_topology->at(i - 1);
3856 // Check if socket, core, anything above thread level changed.
3857 // If the ids did change, then restart thread id at 0
3858 // Otherwise, set thread id to prev thread's id + 1
3859 for (int j = 0; j < tlevel; ++j) {
3860 if (hw_thread.ids[j] != prev_hw_thread.ids[j]) {
3861 hw_thread.ids[tlevel] = 0;
3862 break;
3863 }
3864 }
3865 if (hw_thread.ids[tlevel] == kmp_hw_thread_t::UNKNOWN_ID)
3866 hw_thread.ids[tlevel] = prev_hw_thread.ids[tlevel] + 1;
3867 }
3868 }
3869
3870 if (!__kmp_topology->check_ids()) {
3871 kmp_topology_t::deallocate(__kmp_topology);
3872 __kmp_topology = nullptr;
3873 *msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
3874 return false;
3875 }
3876 return true;
3877}
3878
3879// Create and return a table of affinity masks, indexed by OS thread ID.
3880// This routine handles OR'ing together all the affinity masks of threads
3881// that are sufficiently close, if granularity > fine.
3882template <typename FindNextFunctionType>
3883static void __kmp_create_os_id_masks(unsigned *numUnique,
3884 kmp_affinity_t &affinity,
3885 FindNextFunctionType find_next) {
3886 // First form a table of affinity masks in order of OS thread id.
3887 int maxOsId;
3888 int i;
3889 int numAddrs = __kmp_topology->get_num_hw_threads();
3890 int depth = __kmp_topology->get_depth();
3891 const char *env_var = __kmp_get_affinity_env_var(affinity);
3892 KMP_ASSERT(numAddrs);
3893 KMP_ASSERT(depth);
3894
3895 i = find_next(-1);
3896 // If could not find HW thread location that satisfies find_next conditions,
3897 // then return and fallback to increment find_next.
3898 if (i >= numAddrs)
3899 return;
3900
3901 maxOsId = 0;
3902 for (i = numAddrs - 1;; --i) {
3903 int osId = __kmp_topology->at(i).os_id;
3904 if (osId > maxOsId) {
3905 maxOsId = osId;
3906 }
3907 if (i == 0)
3908 break;
3909 }
3910 affinity.num_os_id_masks = maxOsId + 1;
3911 KMP_CPU_ALLOC_ARRAY(affinity.os_id_masks, affinity.num_os_id_masks);
3912 KMP_ASSERT(affinity.gran_levels >= 0);
3913 if (affinity.flags.verbose && (affinity.gran_levels > 0)) {
3914 KMP_INFORM(ThreadsMigrate, env_var, affinity.gran_levels);
3915 }
3916 if (affinity.gran_levels >= (int)depth) {
3917 KMP_AFF_WARNING(affinity, AffThreadsMayMigrate);
3918 }
3919
3920 // Run through the table, forming the masks for all threads on each core.
3921 // Threads on the same core will have identical kmp_hw_thread_t objects, not
3922 // considering the last level, which must be the thread id. All threads on a
3923 // core will appear consecutively.
3924 int unique = 0;
3925 int j = 0; // index of 1st thread on core
3926 int leader = 0;
3927 kmp_affin_mask_t *sum;
3928 KMP_CPU_ALLOC_ON_STACK(sum);
3929 KMP_CPU_ZERO(sum);
3930
3931 i = j = leader = find_next(-1);
3932 KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
3933 kmp_full_mask_modifier_t full_mask;
3934 for (i = find_next(i); i < numAddrs; i = find_next(i)) {
3935 // If this thread is sufficiently close to the leader (within the
3936 // granularity setting), then set the bit for this os thread in the
3937 // affinity mask for this group, and go on to the next thread.
3938 if (__kmp_topology->is_close(leader, i, affinity)) {
3939 KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
3940 continue;
3941 }
3942
3943 // For every thread in this group, copy the mask to the thread's entry in
3944 // the OS Id mask table. Mark the first address as a leader.
3945 for (; j < i; j = find_next(j)) {
3946 int osId = __kmp_topology->at(j).os_id;
3947 KMP_DEBUG_ASSERT(osId <= maxOsId);
3948 kmp_affin_mask_t *mask = KMP_CPU_INDEX(affinity.os_id_masks, osId);
3949 KMP_CPU_COPY(mask, sum);
3950 __kmp_topology->at(j).leader = (j == leader);
3951 }
3952 unique++;
3953
3954 // Start a new mask.
3955 leader = i;
3956 full_mask.include(sum);
3957 KMP_CPU_ZERO(sum);
3958 KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
3959 }
3960
3961 // For every thread in last group, copy the mask to the thread's
3962 // entry in the OS Id mask table.
3963 for (; j < i; j = find_next(j)) {
3964 int osId = __kmp_topology->at(j).os_id;
3965 KMP_DEBUG_ASSERT(osId <= maxOsId);
3966 kmp_affin_mask_t *mask = KMP_CPU_INDEX(affinity.os_id_masks, osId);
3967 KMP_CPU_COPY(mask, sum);
3968 __kmp_topology->at(j).leader = (j == leader);
3969 }
3970 full_mask.include(sum);
3971 unique++;
3972 KMP_CPU_FREE_FROM_STACK(sum);
3973
3974 // See if the OS Id mask table further restricts or changes the full mask
3975 if (full_mask.restrict_to_mask() && affinity.flags.verbose) {
3976 __kmp_topology->print(env_var);
3977 }
3978
3979 *numUnique = unique;
3980}
3981
3982// Stuff for the affinity proclist parsers. It's easier to declare these vars
3983// as file-static than to try and pass them through the calling sequence of
3984// the recursive-descent OMP_PLACES parser.
3985static kmp_affin_mask_t *newMasks;
3986static int numNewMasks;
3987static int nextNewMask;
3988
3989#define ADD_MASK(_mask) \
3990 { \
3991 if (nextNewMask >= numNewMasks) { \
3992 int i; \
3993 numNewMasks *= 2; \
3994 kmp_affin_mask_t *temp; \
3995 KMP_CPU_INTERNAL_ALLOC_ARRAY(temp, numNewMasks); \
3996 for (i = 0; i < numNewMasks / 2; i++) { \
3997 kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i); \
3998 kmp_affin_mask_t *dest = KMP_CPU_INDEX(temp, i); \
3999 KMP_CPU_COPY(dest, src); \
4000 } \
4001 KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks / 2); \
4002 newMasks = temp; \
4003 } \
4004 KMP_CPU_COPY(KMP_CPU_INDEX(newMasks, nextNewMask), (_mask)); \
4005 nextNewMask++; \
4006 }
4007
4008#define ADD_MASK_OSID(_osId, _osId2Mask, _maxOsId) \
4009 { \
4010 if (((_osId) > _maxOsId) || \
4011 (!KMP_CPU_ISSET((_osId), KMP_CPU_INDEX((_osId2Mask), (_osId))))) { \
4012 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, _osId); \
4013 } else { \
4014 ADD_MASK(KMP_CPU_INDEX(_osId2Mask, (_osId))); \
4015 } \
4016 }
4017
4018// Re-parse the proclist (for the explicit affinity type), and form the list
4019// of affinity newMasks indexed by gtid.
4020static void __kmp_affinity_process_proclist(kmp_affinity_t &affinity) {
4021 int i;
4022 kmp_affin_mask_t **out_masks = &affinity.masks;
4023 unsigned *out_numMasks = &affinity.num_masks;
4024 const char *proclist = affinity.proclist;
4025 kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
4026 int maxOsId = affinity.num_os_id_masks - 1;
4027 const char *scan = proclist;
4028 const char *next = proclist;
4029
4030 // We use malloc() for the temporary mask vector, so that we can use
4031 // realloc() to extend it.
4032 numNewMasks = 2;
4033 KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
4034 nextNewMask = 0;
4035 kmp_affin_mask_t *sumMask;
4036 KMP_CPU_ALLOC(sumMask);
4037 int setSize = 0;
4038
4039 for (;;) {
4040 int start, end, stride;
4041
4042 SKIP_WS(scan);
4043 next = scan;
4044 if (*next == '\0') {
4045 break;
4046 }
4047
4048 if (*next == '{') {
4049 int num;
4050 setSize = 0;
4051 next++; // skip '{'
4052 SKIP_WS(next);
4053 scan = next;
4054
4055 // Read the first integer in the set.
4056 KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad proclist");
4057 SKIP_DIGITS(next);
4058 num = __kmp_str_to_int(scan, *next);
4059 KMP_ASSERT2(num >= 0, "bad explicit proc list");
4060
4061 // Copy the mask for that osId to the sum (union) mask.
4062 if ((num > maxOsId) ||
4063 (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
4064 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, num);
4065 KMP_CPU_ZERO(sumMask);
4066 } else {
4067 KMP_CPU_COPY(sumMask, KMP_CPU_INDEX(osId2Mask, num));
4068 setSize = 1;
4069 }
4070
4071 for (;;) {
4072 // Check for end of set.
4073 SKIP_WS(next);
4074 if (*next == '}') {
4075 next++; // skip '}'
4076 break;
4077 }
4078
4079 // Skip optional comma.
4080 if (*next == ',') {
4081 next++;
4082 }
4083 SKIP_WS(next);
4084
4085 // Read the next integer in the set.
4086 scan = next;
4087 KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
4088
4089 SKIP_DIGITS(next);
4090 num = __kmp_str_to_int(scan, *next);
4091 KMP_ASSERT2(num >= 0, "bad explicit proc list");
4092
4093 // Add the mask for that osId to the sum mask.
4094 if ((num > maxOsId) ||
4095 (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
4096 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, num);
4097 } else {
4098 KMP_CPU_UNION(sumMask, KMP_CPU_INDEX(osId2Mask, num));
4099 setSize++;
4100 }
4101 }
4102 if (setSize > 0) {
4103 ADD_MASK(sumMask);
4104 }
4105
4106 SKIP_WS(next);
4107 if (*next == ',') {
4108 next++;
4109 }
4110 scan = next;
4111 continue;
4112 }
4113
4114 // Read the first integer.
4115 KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
4116 SKIP_DIGITS(next);
4117 start = __kmp_str_to_int(scan, *next);
4118 KMP_ASSERT2(start >= 0, "bad explicit proc list");
4119 SKIP_WS(next);
4120
4121 // If this isn't a range, then add a mask to the list and go on.
4122 if (*next != '-') {
4123 ADD_MASK_OSID(start, osId2Mask, maxOsId);
4124
4125 // Skip optional comma.
4126 if (*next == ',') {
4127 next++;
4128 }
4129 scan = next;
4130 continue;
4131 }
4132
4133 // This is a range. Skip over the '-' and read in the 2nd int.
4134 next++; // skip '-'
4135 SKIP_WS(next);
4136 scan = next;
4137 KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
4138 SKIP_DIGITS(next);
4139 end = __kmp_str_to_int(scan, *next);
4140 KMP_ASSERT2(end >= 0, "bad explicit proc list");
4141
4142 // Check for a stride parameter
4143 stride = 1;
4144 SKIP_WS(next);
4145 if (*next == ':') {
4146 // A stride is specified. Skip over the ':" and read the 3rd int.
4147 int sign = +1;
4148 next++; // skip ':'
4149 SKIP_WS(next);
4150 scan = next;
4151 if (*next == '-') {
4152 sign = -1;
4153 next++;
4154 SKIP_WS(next);
4155 scan = next;
4156 }
4157 KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
4158 SKIP_DIGITS(next);
4159 stride = __kmp_str_to_int(scan, *next);
4160 KMP_ASSERT2(stride >= 0, "bad explicit proc list");
4161 stride *= sign;
4162 }
4163
4164 // Do some range checks.
4165 KMP_ASSERT2(stride != 0, "bad explicit proc list");
4166 if (stride > 0) {
4167 KMP_ASSERT2(start <= end, "bad explicit proc list");
4168 } else {
4169 KMP_ASSERT2(start >= end, "bad explicit proc list");
4170 }
4171 KMP_ASSERT2((end - start) / stride <= 65536, "bad explicit proc list");
4172
4173 // Add the mask for each OS proc # to the list.
4174 if (stride > 0) {
4175 do {
4176 ADD_MASK_OSID(start, osId2Mask, maxOsId);
4177 start += stride;
4178 } while (start <= end);
4179 } else {
4180 do {
4181 ADD_MASK_OSID(start, osId2Mask, maxOsId);
4182 start += stride;
4183 } while (start >= end);
4184 }
4185
4186 // Skip optional comma.
4187 SKIP_WS(next);
4188 if (*next == ',') {
4189 next++;
4190 }
4191 scan = next;
4192 }
4193
4194 *out_numMasks = nextNewMask;
4195 if (nextNewMask == 0) {
4196 *out_masks = NULL;
4197 KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
4198 return;
4199 }
4200 KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
4201 for (i = 0; i < nextNewMask; i++) {
4202 kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
4203 kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
4204 KMP_CPU_COPY(dest, src);
4205 }
4206 KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
4207 KMP_CPU_FREE(sumMask);
4208}
4209
4210/*-----------------------------------------------------------------------------
4211Re-parse the OMP_PLACES proc id list, forming the newMasks for the different
4212places. Again, Here is the grammar:
4213
4214place_list := place
4215place_list := place , place_list
4216place := num
4217place := place : num
4218place := place : num : signed
4219place := { subplacelist }
4220place := ! place // (lowest priority)
4221subplace_list := subplace
4222subplace_list := subplace , subplace_list
4223subplace := num
4224subplace := num : num
4225subplace := num : num : signed
4226signed := num
4227signed := + signed
4228signed := - signed
4229-----------------------------------------------------------------------------*/
4230static void __kmp_process_subplace_list(const char **scan,
4231 kmp_affinity_t &affinity, int maxOsId,
4232 kmp_affin_mask_t *tempMask,
4233 int *setSize) {
4234 const char *next;
4235 kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
4236
4237 for (;;) {
4238 int start, count, stride, i;
4239
4240 // Read in the starting proc id
4241 SKIP_WS(*scan);
4242 KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
4243 next = *scan;
4244 SKIP_DIGITS(next);
4245 start = __kmp_str_to_int(*scan, *next);
4246 KMP_ASSERT(start >= 0);
4247 *scan = next;
4248
4249 // valid follow sets are ',' ':' and '}'
4250 SKIP_WS(*scan);
4251 if (**scan == '}' || **scan == ',') {
4252 if ((start > maxOsId) ||
4253 (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
4254 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, start);
4255 } else {
4256 KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
4257 (*setSize)++;
4258 }
4259 if (**scan == '}') {
4260 break;
4261 }
4262 (*scan)++; // skip ','
4263 continue;
4264 }
4265 KMP_ASSERT2(**scan == ':', "bad explicit places list");
4266 (*scan)++; // skip ':'
4267
4268 // Read count parameter
4269 SKIP_WS(*scan);
4270 KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
4271 next = *scan;
4272 SKIP_DIGITS(next);
4273 count = __kmp_str_to_int(*scan, *next);
4274 KMP_ASSERT(count >= 0);
4275 *scan = next;
4276
4277 // valid follow sets are ',' ':' and '}'
4278 SKIP_WS(*scan);
4279 if (**scan == '}' || **scan == ',') {
4280 for (i = 0; i < count; i++) {
4281 if ((start > maxOsId) ||
4282 (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
4283 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, start);
4284 break; // don't proliferate warnings for large count
4285 } else {
4286 KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
4287 start++;
4288 (*setSize)++;
4289 }
4290 }
4291 if (**scan == '}') {
4292 break;
4293 }
4294 (*scan)++; // skip ','
4295 continue;
4296 }
4297 KMP_ASSERT2(**scan == ':', "bad explicit places list");
4298 (*scan)++; // skip ':'
4299
4300 // Read stride parameter
4301 int sign = +1;
4302 for (;;) {
4303 SKIP_WS(*scan);
4304 if (**scan == '+') {
4305 (*scan)++; // skip '+'
4306 continue;
4307 }
4308 if (**scan == '-') {
4309 sign *= -1;
4310 (*scan)++; // skip '-'
4311 continue;
4312 }
4313 break;
4314 }
4315 SKIP_WS(*scan);
4316 KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
4317 next = *scan;
4318 SKIP_DIGITS(next);
4319 stride = __kmp_str_to_int(*scan, *next);
4320 KMP_ASSERT(stride >= 0);
4321 *scan = next;
4322 stride *= sign;
4323
4324 // valid follow sets are ',' and '}'
4325 SKIP_WS(*scan);
4326 if (**scan == '}' || **scan == ',') {
4327 for (i = 0; i < count; i++) {
4328 if ((start > maxOsId) ||
4329 (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
4330 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, start);
4331 break; // don't proliferate warnings for large count
4332 } else {
4333 KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
4334 start += stride;
4335 (*setSize)++;
4336 }
4337 }
4338 if (**scan == '}') {
4339 break;
4340 }
4341 (*scan)++; // skip ','
4342 continue;
4343 }
4344
4345 KMP_ASSERT2(0, "bad explicit places list");
4346 }
4347}
4348
4349static void __kmp_process_place(const char **scan, kmp_affinity_t &affinity,
4350 int maxOsId, kmp_affin_mask_t *tempMask,
4351 int *setSize) {
4352 const char *next;
4353 kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
4354
4355 // valid follow sets are '{' '!' and num
4356 SKIP_WS(*scan);
4357 if (**scan == '{') {
4358 (*scan)++; // skip '{'
4359 __kmp_process_subplace_list(scan, affinity, maxOsId, tempMask, setSize);
4360 KMP_ASSERT2(**scan == '}', "bad explicit places list");
4361 (*scan)++; // skip '}'
4362 } else if (**scan == '!') {
4363 (*scan)++; // skip '!'
4364 __kmp_process_place(scan, affinity, maxOsId, tempMask, setSize);
4365 KMP_CPU_COMPLEMENT(maxOsId, tempMask);
4366 } else if ((**scan >= '0') && (**scan <= '9')) {
4367 next = *scan;
4368 SKIP_DIGITS(next);
4369 int num = __kmp_str_to_int(*scan, *next);
4370 KMP_ASSERT(num >= 0);
4371 if ((num > maxOsId) ||
4372 (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
4373 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, num);
4374 } else {
4375 KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, num));
4376 (*setSize)++;
4377 }
4378 *scan = next; // skip num
4379 } else {
4380 KMP_ASSERT2(0, "bad explicit places list");
4381 }
4382}
4383
4384// static void
4385void __kmp_affinity_process_placelist(kmp_affinity_t &affinity) {
4386 int i, j, count, stride, sign;
4387 kmp_affin_mask_t **out_masks = &affinity.masks;
4388 unsigned *out_numMasks = &affinity.num_masks;
4389 const char *placelist = affinity.proclist;
4390 kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
4391 int maxOsId = affinity.num_os_id_masks - 1;
4392 const char *scan = placelist;
4393 const char *next = placelist;
4394
4395 numNewMasks = 2;
4396 KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
4397 nextNewMask = 0;
4398
4399 // tempMask is modified based on the previous or initial
4400 // place to form the current place
4401 // previousMask contains the previous place
4402 kmp_affin_mask_t *tempMask;
4403 kmp_affin_mask_t *previousMask;
4404 KMP_CPU_ALLOC(tempMask);
4405 KMP_CPU_ZERO(tempMask);
4406 KMP_CPU_ALLOC(previousMask);
4407 KMP_CPU_ZERO(previousMask);
4408 int setSize = 0;
4409
4410 for (;;) {
4411 __kmp_process_place(&scan, affinity, maxOsId, tempMask, &setSize);
4412
4413 // valid follow sets are ',' ':' and EOL
4414 SKIP_WS(scan);
4415 if (*scan == '\0' || *scan == ',') {
4416 if (setSize > 0) {
4417 ADD_MASK(tempMask);
4418 }
4419 KMP_CPU_ZERO(tempMask);
4420 setSize = 0;
4421 if (*scan == '\0') {
4422 break;
4423 }
4424 scan++; // skip ','
4425 continue;
4426 }
4427
4428 KMP_ASSERT2(*scan == ':', "bad explicit places list");
4429 scan++; // skip ':'
4430
4431 // Read count parameter
4432 SKIP_WS(scan);
4433 KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
4434 next = scan;
4435 SKIP_DIGITS(next);
4436 count = __kmp_str_to_int(scan, *next);
4437 KMP_ASSERT(count >= 0);
4438 scan = next;
4439
4440 // valid follow sets are ',' ':' and EOL
4441 SKIP_WS(scan);
4442 if (*scan == '\0' || *scan == ',') {
4443 stride = +1;
4444 } else {
4445 KMP_ASSERT2(*scan == ':', "bad explicit places list");
4446 scan++; // skip ':'
4447
4448 // Read stride parameter
4449 sign = +1;
4450 for (;;) {
4451 SKIP_WS(scan);
4452 if (*scan == '+') {
4453 scan++; // skip '+'
4454 continue;
4455 }
4456 if (*scan == '-') {
4457 sign *= -1;
4458 scan++; // skip '-'
4459 continue;
4460 }
4461 break;
4462 }
4463 SKIP_WS(scan);
4464 KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
4465 next = scan;
4466 SKIP_DIGITS(next);
4467 stride = __kmp_str_to_int(scan, *next);
4468 KMP_DEBUG_ASSERT(stride >= 0);
4469 scan = next;
4470 stride *= sign;
4471 }
4472
4473 // Add places determined by initial_place : count : stride
4474 for (i = 0; i < count; i++) {
4475 if (setSize == 0) {
4476 break;
4477 }
4478 // Add the current place, then build the next place (tempMask) from that
4479 KMP_CPU_COPY(previousMask, tempMask);
4480 ADD_MASK(previousMask);
4481 KMP_CPU_ZERO(tempMask);
4482 setSize = 0;
4483 KMP_CPU_SET_ITERATE(j, previousMask) {
4484 if (!KMP_CPU_ISSET(j, previousMask)) {
4485 continue;
4486 }
4487 if ((j + stride > maxOsId) || (j + stride < 0) ||
4488 (!KMP_CPU_ISSET(j, __kmp_affin_fullMask)) ||
4489 (!KMP_CPU_ISSET(j + stride,
4490 KMP_CPU_INDEX(osId2Mask, j + stride)))) {
4491 if (i < count - 1) {
4492 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, j + stride);
4493 }
4494 continue;
4495 }
4496 KMP_CPU_SET(j + stride, tempMask);
4497 setSize++;
4498 }
4499 }
4500 KMP_CPU_ZERO(tempMask);
4501 setSize = 0;
4502
4503 // valid follow sets are ',' and EOL
4504 SKIP_WS(scan);
4505 if (*scan == '\0') {
4506 break;
4507 }
4508 if (*scan == ',') {
4509 scan++; // skip ','
4510 continue;
4511 }
4512
4513 KMP_ASSERT2(0, "bad explicit places list");
4514 }
4515
4516 *out_numMasks = nextNewMask;
4517 if (nextNewMask == 0) {
4518 *out_masks = NULL;
4519 KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
4520 return;
4521 }
4522 KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
4523 KMP_CPU_FREE(tempMask);
4524 KMP_CPU_FREE(previousMask);
4525 for (i = 0; i < nextNewMask; i++) {
4526 kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
4527 kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
4528 KMP_CPU_COPY(dest, src);
4529 }
4530 KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
4531}
4532
4533#undef ADD_MASK
4534#undef ADD_MASK_OSID
4535
4536// This function figures out the deepest level at which there is at least one
4537// cluster/core with more than one processing unit bound to it.
4538static int __kmp_affinity_find_core_level(int nprocs, int bottom_level) {
4539 int core_level = 0;
4540
4541 for (int i = 0; i < nprocs; i++) {
4542 const kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
4543 for (int j = bottom_level; j > 0; j--) {
4544 if (hw_thread.ids[j] > 0) {
4545 if (core_level < (j - 1)) {
4546 core_level = j - 1;
4547 }
4548 }
4549 }
4550 }
4551 return core_level;
4552}
4553
4554// This function counts number of clusters/cores at given level.
4555static int __kmp_affinity_compute_ncores(int nprocs, int bottom_level,
4556 int core_level) {
4557 return __kmp_topology->get_count(core_level);
4558}
4559// This function finds to which cluster/core given processing unit is bound.
4560static int __kmp_affinity_find_core(int proc, int bottom_level,
4561 int core_level) {
4562 int core = 0;
4563 KMP_DEBUG_ASSERT(proc >= 0 && proc < __kmp_topology->get_num_hw_threads());
4564 for (int i = 0; i <= proc; ++i) {
4565 if (i + 1 <= proc) {
4566 for (int j = 0; j <= core_level; ++j) {
4567 if (__kmp_topology->at(i + 1).sub_ids[j] !=
4568 __kmp_topology->at(i).sub_ids[j]) {
4569 core++;
4570 break;
4571 }
4572 }
4573 }
4574 }
4575 return core;
4576}
4577
4578// This function finds maximal number of processing units bound to a
4579// cluster/core at given level.
4580static int __kmp_affinity_max_proc_per_core(int nprocs, int bottom_level,
4581 int core_level) {
4582 if (core_level >= bottom_level)
4583 return 1;
4584 int thread_level = __kmp_topology->get_level(KMP_HW_THREAD);
4585 return __kmp_topology->calculate_ratio(thread_level, core_level);
4586}
4587
4588static int *procarr = NULL;
4589static int __kmp_aff_depth = 0;
4590static int *__kmp_osid_to_hwthread_map = NULL;
4591
4592static void __kmp_affinity_get_mask_topology_info(const kmp_affin_mask_t *mask,
4593 kmp_affinity_ids_t &ids,
4594 kmp_affinity_attrs_t &attrs) {
4595 if (!KMP_AFFINITY_CAPABLE())
4596 return;
4597
4598 // Initiailze ids and attrs thread data
4599 for (int i = 0; i < KMP_HW_LAST; ++i)
4600 ids.ids[i] = kmp_hw_thread_t::UNKNOWN_ID;
4601 attrs = KMP_AFFINITY_ATTRS_UNKNOWN;
4602
4603 // Iterate through each os id within the mask and determine
4604 // the topology id and attribute information
4605 int cpu;
4606 int depth = __kmp_topology->get_depth();
4607 KMP_CPU_SET_ITERATE(cpu, mask) {
4608 int osid_idx = __kmp_osid_to_hwthread_map[cpu];
4609 ids.os_id = cpu;
4610 const kmp_hw_thread_t &hw_thread = __kmp_topology->at(osid_idx);
4611 for (int level = 0; level < depth; ++level) {
4612 kmp_hw_t type = __kmp_topology->get_type(level);
4613 int id = hw_thread.sub_ids[level];
4614 if (ids.ids[type] == kmp_hw_thread_t::UNKNOWN_ID || ids.ids[type] == id) {
4615 ids.ids[type] = id;
4616 } else {
4617 // This mask spans across multiple topology units, set it as such
4618 // and mark every level below as such as well.
4619 ids.ids[type] = kmp_hw_thread_t::MULTIPLE_ID;
4620 for (; level < depth; ++level) {
4621 kmp_hw_t type = __kmp_topology->get_type(level);
4622 ids.ids[type] = kmp_hw_thread_t::MULTIPLE_ID;
4623 }
4624 }
4625 }
4626 if (!attrs.valid) {
4627 attrs.core_type = hw_thread.attrs.get_core_type();
4628 attrs.core_eff = hw_thread.attrs.get_core_eff();
4629 attrs.valid = 1;
4630 } else {
4631 // This mask spans across multiple attributes, set it as such
4632 if (attrs.core_type != hw_thread.attrs.get_core_type())
4633 attrs.core_type = KMP_HW_CORE_TYPE_UNKNOWN;
4634 if (attrs.core_eff != hw_thread.attrs.get_core_eff())
4635 attrs.core_eff = kmp_hw_attr_t::UNKNOWN_CORE_EFF;
4636 }
4637 }
4638}
4639
4640static void __kmp_affinity_get_thread_topology_info(kmp_info_t *th) {
4641 if (!KMP_AFFINITY_CAPABLE())
4642 return;
4643 const kmp_affin_mask_t *mask = th->th.th_affin_mask;
4644 kmp_affinity_ids_t &ids = th->th.th_topology_ids;
4645 kmp_affinity_attrs_t &attrs = th->th.th_topology_attrs;
4646 __kmp_affinity_get_mask_topology_info(mask, ids, attrs);
4647}
4648
4649// Assign the topology information to each place in the place list
4650// A thread can then grab not only its affinity mask, but the topology
4651// information associated with that mask. e.g., Which socket is a thread on
4652static void __kmp_affinity_get_topology_info(kmp_affinity_t &affinity) {
4653 if (!KMP_AFFINITY_CAPABLE())
4654 return;
4655 if (affinity.type != affinity_none) {
4656 KMP_ASSERT(affinity.num_os_id_masks);
4657 KMP_ASSERT(affinity.os_id_masks);
4658 }
4659 KMP_ASSERT(affinity.num_masks);
4660 KMP_ASSERT(affinity.masks);
4661 KMP_ASSERT(__kmp_affin_fullMask);
4662
4663 int max_cpu = __kmp_affin_fullMask->get_max_cpu();
4664 int num_hw_threads = __kmp_topology->get_num_hw_threads();
4665
4666 // Allocate thread topology information
4667 if (!affinity.ids) {
4668 affinity.ids = (kmp_affinity_ids_t *)__kmp_allocate(
4669 sizeof(kmp_affinity_ids_t) * affinity.num_masks);
4670 }
4671 if (!affinity.attrs) {
4672 affinity.attrs = (kmp_affinity_attrs_t *)__kmp_allocate(
4673 sizeof(kmp_affinity_attrs_t) * affinity.num_masks);
4674 }
4675 if (!__kmp_osid_to_hwthread_map) {
4676 // Want the +1 because max_cpu should be valid index into map
4677 __kmp_osid_to_hwthread_map =
4678 (int *)__kmp_allocate(sizeof(int) * (max_cpu + 1));
4679 }
4680
4681 // Create the OS proc to hardware thread map
4682 for (int hw_thread = 0; hw_thread < num_hw_threads; ++hw_thread) {
4683 int os_id = __kmp_topology->at(hw_thread).os_id;
4684 if (KMP_CPU_ISSET(os_id, __kmp_affin_fullMask))
4685 __kmp_osid_to_hwthread_map[os_id] = hw_thread;
4686 }
4687
4688 for (unsigned i = 0; i < affinity.num_masks; ++i) {
4689 kmp_affinity_ids_t &ids = affinity.ids[i];
4690 kmp_affinity_attrs_t &attrs = affinity.attrs[i];
4691 kmp_affin_mask_t *mask = KMP_CPU_INDEX(affinity.masks, i);
4692 __kmp_affinity_get_mask_topology_info(mask, ids, attrs);
4693 }
4694}
4695
4696// Called when __kmp_topology is ready
4697static void __kmp_aux_affinity_initialize_other_data(kmp_affinity_t &affinity) {
4698 // Initialize other data structures which depend on the topology
4699 if (__kmp_topology && __kmp_topology->get_num_hw_threads()) {
4700 machine_hierarchy.init(__kmp_topology->get_num_hw_threads());
4701 __kmp_affinity_get_topology_info(affinity);
4702#if KMP_WEIGHTED_ITERATIONS_SUPPORTED
4703 __kmp_first_osid_with_ecore = __kmp_get_first_osid_with_ecore();
4704#endif
4705 }
4706}
4707
4708// Create a one element mask array (set of places) which only contains the
4709// initial process's affinity mask
4710static void __kmp_create_affinity_none_places(kmp_affinity_t &affinity) {
4711 KMP_ASSERT(__kmp_affin_fullMask != NULL);
4712 KMP_ASSERT(affinity.type == affinity_none);
4713 KMP_ASSERT(__kmp_avail_proc == __kmp_topology->get_num_hw_threads());
4714 affinity.num_masks = 1;
4715 KMP_CPU_ALLOC_ARRAY(affinity.masks, affinity.num_masks);
4716 kmp_affin_mask_t *dest = KMP_CPU_INDEX(affinity.masks, 0);
4717 KMP_CPU_COPY(dest, __kmp_affin_fullMask);
4718 __kmp_aux_affinity_initialize_other_data(affinity);
4719}
4720
4721static void __kmp_aux_affinity_initialize_masks(kmp_affinity_t &affinity) {
4722 // Create the "full" mask - this defines all of the processors that we
4723 // consider to be in the machine model. If respect is set, then it is the
4724 // initialization thread's affinity mask. Otherwise, it is all processors that
4725 // we know about on the machine.
4726 int verbose = affinity.flags.verbose;
4727 const char *env_var = affinity.env_var;
4728
4729 // Already initialized
4730 if (__kmp_affin_fullMask && __kmp_affin_origMask)
4731 return;
4732
4733 if (__kmp_affin_fullMask == NULL) {
4734 KMP_CPU_ALLOC(__kmp_affin_fullMask);
4735 }
4736 if (__kmp_affin_origMask == NULL) {
4737 KMP_CPU_ALLOC(__kmp_affin_origMask);
4738 }
4739 if (KMP_AFFINITY_CAPABLE()) {
4740 __kmp_get_system_affinity(__kmp_affin_fullMask, TRUE);
4741 // Make a copy before possible expanding to the entire machine mask
4742 __kmp_affin_origMask->copy(__kmp_affin_fullMask);
4743 if (affinity.flags.respect) {
4744 // Count the number of available processors.
4745 unsigned i;
4746 __kmp_avail_proc = 0;
4747 KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
4748 if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
4749 continue;
4750 }
4751 __kmp_avail_proc++;
4752 }
4753 if (__kmp_avail_proc > __kmp_xproc) {
4754 KMP_AFF_WARNING(affinity, ErrorInitializeAffinity);
4755 affinity.type = affinity_none;
4756 KMP_AFFINITY_DISABLE();
4757 return;
4758 }
4759
4760 if (verbose) {
4761 char buf[KMP_AFFIN_MASK_PRINT_LEN];
4762 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4763 __kmp_affin_fullMask);
4764 KMP_INFORM(InitOSProcSetRespect, env_var, buf);
4765 }
4766 } else {
4767 if (verbose) {
4768 char buf[KMP_AFFIN_MASK_PRINT_LEN];
4769 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4770 __kmp_affin_fullMask);
4771 KMP_INFORM(InitOSProcSetNotRespect, env_var, buf);
4772 }
4773 __kmp_avail_proc =
4774 __kmp_affinity_entire_machine_mask(__kmp_affin_fullMask);
4775#if KMP_OS_WINDOWS
4776 if (__kmp_num_proc_groups <= 1) {
4777 // Copy expanded full mask if topology has single processor group
4778 __kmp_affin_origMask->copy(__kmp_affin_fullMask);
4779 }
4780 // Set the process affinity mask since threads' affinity
4781 // masks must be subset of process mask in Windows* OS
4782 __kmp_affin_fullMask->set_process_affinity(true);
4783#endif
4784 }
4785 }
4786}
4787
4788static bool __kmp_aux_affinity_initialize_topology(kmp_affinity_t &affinity) {
4789 bool success = false;
4790 const char *env_var = affinity.env_var;
4791 kmp_i18n_id_t msg_id = kmp_i18n_null;
4792 int verbose = affinity.flags.verbose;
4793
4794 // For backward compatibility, setting KMP_CPUINFO_FILE =>
4795 // KMP_TOPOLOGY_METHOD=cpuinfo
4796 if ((__kmp_cpuinfo_file != NULL) &&
4797 (__kmp_affinity_top_method == affinity_top_method_all)) {
4798 __kmp_affinity_top_method = affinity_top_method_cpuinfo;
4799 }
4800
4801 if (__kmp_affinity_top_method == affinity_top_method_all) {
4802// In the default code path, errors are not fatal - we just try using
4803// another method. We only emit a warning message if affinity is on, or the
4804// verbose flag is set, an the nowarnings flag was not set.
4805#if KMP_USE_HWLOC
4806 if (!success &&
4807 __kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC) {
4808 if (!__kmp_hwloc_error) {
4809 success = __kmp_affinity_create_hwloc_map(&msg_id);
4810 if (!success && verbose) {
4811 KMP_INFORM(AffIgnoringHwloc, env_var);
4812 }
4813 } else if (verbose) {
4814 KMP_INFORM(AffIgnoringHwloc, env_var);
4815 }
4816 }
4817#endif
4818
4819#if KMP_ARCH_X86 || KMP_ARCH_X86_64
4820 if (!success) {
4821 success = __kmp_affinity_create_x2apicid_map(&msg_id);
4822 if (!success && verbose && msg_id != kmp_i18n_null) {
4823 KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4824 }
4825 }
4826 if (!success) {
4827 success = __kmp_affinity_create_apicid_map(&msg_id);
4828 if (!success && verbose && msg_id != kmp_i18n_null) {
4829 KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4830 }
4831 }
4832#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4833
4834#if KMP_OS_LINUX || KMP_OS_AIX
4835 if (!success) {
4836 int line = 0;
4837 success = __kmp_affinity_create_cpuinfo_map(&line, &msg_id);
4838 if (!success && verbose && msg_id != kmp_i18n_null) {
4839 KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4840 }
4841 }
4842#endif /* KMP_OS_LINUX */
4843
4844#if KMP_GROUP_AFFINITY
4845 if (!success && (__kmp_num_proc_groups > 1)) {
4846 success = __kmp_affinity_create_proc_group_map(&msg_id);
4847 if (!success && verbose && msg_id != kmp_i18n_null) {
4848 KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4849 }
4850 }
4851#endif /* KMP_GROUP_AFFINITY */
4852
4853 if (!success) {
4854 success = __kmp_affinity_create_flat_map(&msg_id);
4855 if (!success && verbose && msg_id != kmp_i18n_null) {
4856 KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4857 }
4858 KMP_ASSERT(success);
4859 }
4860 }
4861
4862// If the user has specified that a paricular topology discovery method is to be
4863// used, then we abort if that method fails. The exception is group affinity,
4864// which might have been implicitly set.
4865#if KMP_USE_HWLOC
4866 else if (__kmp_affinity_top_method == affinity_top_method_hwloc) {
4867 KMP_ASSERT(__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC);
4868 success = __kmp_affinity_create_hwloc_map(&msg_id);
4869 if (!success) {
4870 KMP_ASSERT(msg_id != kmp_i18n_null);
4871 KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4872 }
4873 }
4874#endif // KMP_USE_HWLOC
4875
4876#if KMP_ARCH_X86 || KMP_ARCH_X86_64
4877 else if (__kmp_affinity_top_method == affinity_top_method_x2apicid ||
4878 __kmp_affinity_top_method == affinity_top_method_x2apicid_1f) {
4879 success = __kmp_affinity_create_x2apicid_map(&msg_id);
4880 if (!success) {
4881 KMP_ASSERT(msg_id != kmp_i18n_null);
4882 KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4883 }
4884 } else if (__kmp_affinity_top_method == affinity_top_method_apicid) {
4885 success = __kmp_affinity_create_apicid_map(&msg_id);
4886 if (!success) {
4887 KMP_ASSERT(msg_id != kmp_i18n_null);
4888 KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4889 }
4890 }
4891#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4892
4893 else if (__kmp_affinity_top_method == affinity_top_method_cpuinfo) {
4894 int line = 0;
4895 success = __kmp_affinity_create_cpuinfo_map(&line, &msg_id);
4896 if (!success) {
4897 KMP_ASSERT(msg_id != kmp_i18n_null);
4898 const char *filename = __kmp_cpuinfo_get_filename();
4899 if (line > 0) {
4900 KMP_FATAL(FileLineMsgExiting, filename, line,
4901 __kmp_i18n_catgets(msg_id));
4902 } else {
4903 KMP_FATAL(FileMsgExiting, filename, __kmp_i18n_catgets(msg_id));
4904 }
4905 }
4906 }
4907
4908#if KMP_GROUP_AFFINITY
4909 else if (__kmp_affinity_top_method == affinity_top_method_group) {
4910 success = __kmp_affinity_create_proc_group_map(&msg_id);
4911 KMP_ASSERT(success);
4912 if (!success) {
4913 KMP_ASSERT(msg_id != kmp_i18n_null);
4914 KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4915 }
4916 }
4917#endif /* KMP_GROUP_AFFINITY */
4918
4919 else if (__kmp_affinity_top_method == affinity_top_method_flat) {
4920 success = __kmp_affinity_create_flat_map(&msg_id);
4921 // should not fail
4922 KMP_ASSERT(success);
4923 }
4924
4925 // Early exit if topology could not be created
4926 if (!__kmp_topology) {
4927 if (KMP_AFFINITY_CAPABLE()) {
4928 KMP_AFF_WARNING(affinity, ErrorInitializeAffinity);
4929 }
4930 if (nPackages > 0 && nCoresPerPkg > 0 && __kmp_nThreadsPerCore > 0 &&
4931 __kmp_ncores > 0) {
4932 __kmp_topology = kmp_topology_t::allocate(0, 0, NULL);
4933 __kmp_topology->canonicalize(nPackages, nCoresPerPkg,
4934 __kmp_nThreadsPerCore, __kmp_ncores);
4935 if (verbose) {
4936 __kmp_topology->print(env_var);
4937 }
4938 }
4939 return false;
4940 }
4941
4942 // Canonicalize, print (if requested), apply KMP_HW_SUBSET
4943 __kmp_topology->canonicalize();
4944 if (verbose)
4945 __kmp_topology->print(env_var);
4946 bool filtered = __kmp_topology->filter_hw_subset();
4947 if (filtered && verbose)
4948 __kmp_topology->print("KMP_HW_SUBSET");
4949 return success;
4950}
4951
4952static void __kmp_aux_affinity_initialize(kmp_affinity_t &affinity) {
4953 bool is_regular_affinity = (&affinity == &__kmp_affinity);
4954 bool is_hidden_helper_affinity = (&affinity == &__kmp_hh_affinity);
4955 const char *env_var = __kmp_get_affinity_env_var(affinity);
4956
4957 if (affinity.flags.initialized) {
4958 KMP_ASSERT(__kmp_affin_fullMask != NULL);
4959 return;
4960 }
4961
4962 if (is_regular_affinity && (!__kmp_affin_fullMask || !__kmp_affin_origMask))
4963 __kmp_aux_affinity_initialize_masks(affinity);
4964
4965 if (is_regular_affinity && !__kmp_topology) {
4966 bool success = __kmp_aux_affinity_initialize_topology(affinity);
4967 if (success) {
4968 KMP_ASSERT(__kmp_avail_proc == __kmp_topology->get_num_hw_threads());
4969 } else {
4970 affinity.type = affinity_none;
4971 KMP_AFFINITY_DISABLE();
4972 }
4973 }
4974
4975 // If KMP_AFFINITY=none, then only create the single "none" place
4976 // which is the process's initial affinity mask or the number of
4977 // hardware threads depending on respect,norespect
4978 if (affinity.type == affinity_none) {
4979 __kmp_create_affinity_none_places(affinity);
4980#if KMP_USE_HIER_SCHED
4981 __kmp_dispatch_set_hierarchy_values();
4982#endif
4983 affinity.flags.initialized = TRUE;
4984 return;
4985 }
4986
4987 __kmp_topology->set_granularity(affinity);
4988 int depth = __kmp_topology->get_depth();
4989
4990 // Create the table of masks, indexed by thread Id.
4991 unsigned numUnique = 0;
4992 int numAddrs = __kmp_topology->get_num_hw_threads();
4993 // If OMP_PLACES=cores:<attribute> specified, then attempt
4994 // to make OS Id mask table using those attributes
4995 if (affinity.core_attr_gran.valid) {
4996 __kmp_create_os_id_masks(&numUnique, affinity, [&](int idx) {
4997 KMP_ASSERT(idx >= -1);
4998 for (int i = idx + 1; i < numAddrs; ++i)
4999 if (__kmp_topology->at(i).attrs.contains(affinity.core_attr_gran))
5000 return i;
5001 return numAddrs;
5002 });
5003 if (!affinity.os_id_masks) {
5004 const char *core_attribute;
5005 if (affinity.core_attr_gran.core_eff != kmp_hw_attr_t::UNKNOWN_CORE_EFF)
5006 core_attribute = "core_efficiency";
5007 else
5008 core_attribute = "core_type";
5009 KMP_AFF_WARNING(affinity, AffIgnoringNotAvailable, env_var,
5010 core_attribute,
5011 __kmp_hw_get_catalog_string(KMP_HW_CORE, /*plural=*/true))
5012 }
5013 }
5014 // If core attributes did not work, or none were specified,
5015 // then make OS Id mask table using typical incremental way with
5016 // checking for validity of each id at granularity level specified.
5017 if (!affinity.os_id_masks) {
5018 int gran = affinity.gran_levels;
5019 int gran_level = depth - 1 - affinity.gran_levels;
5020 if (gran >= 0 && gran_level >= 0 && gran_level < depth) {
5021 __kmp_create_os_id_masks(
5022 &numUnique, affinity, [depth, numAddrs, &affinity](int idx) {
5023 KMP_ASSERT(idx >= -1);
5024 int gran = affinity.gran_levels;
5025 int gran_level = depth - 1 - affinity.gran_levels;
5026 for (int i = idx + 1; i < numAddrs; ++i)
5027 if ((gran >= depth) ||
5028 (gran < depth && __kmp_topology->at(i).ids[gran_level] !=
5029 kmp_hw_thread_t::UNKNOWN_ID))
5030 return i;
5031 return numAddrs;
5032 });
5033 }
5034 }
5035 // Final attempt to make OS Id mask table using typical incremental way.
5036 if (!affinity.os_id_masks) {
5037 __kmp_create_os_id_masks(&numUnique, affinity, [](int idx) {
5038 KMP_ASSERT(idx >= -1);
5039 return idx + 1;
5040 });
5041 }
5042
5043 switch (affinity.type) {
5044
5045 case affinity_explicit:
5046 KMP_DEBUG_ASSERT(affinity.proclist != NULL);
5047 if (is_hidden_helper_affinity ||
5048 __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) {
5049 __kmp_affinity_process_proclist(affinity);
5050 } else {
5051 __kmp_affinity_process_placelist(affinity);
5052 }
5053 if (affinity.num_masks == 0) {
5054 KMP_AFF_WARNING(affinity, AffNoValidProcID);
5055 affinity.type = affinity_none;
5056 __kmp_create_affinity_none_places(affinity);
5057 affinity.flags.initialized = TRUE;
5058 return;
5059 }
5060 break;
5061
5062 // The other affinity types rely on sorting the hardware threads according to
5063 // some permutation of the machine topology tree. Set affinity.compact
5064 // and affinity.offset appropriately, then jump to a common code
5065 // fragment to do the sort and create the array of affinity masks.
5066 case affinity_logical:
5067 affinity.compact = 0;
5068 if (affinity.offset) {
5069 affinity.offset =
5070 __kmp_nThreadsPerCore * affinity.offset % __kmp_avail_proc;
5071 }
5072 goto sortTopology;
5073
5074 case affinity_physical:
5075 if (__kmp_nThreadsPerCore > 1) {
5076 affinity.compact = 1;
5077 if (affinity.compact >= depth) {
5078 affinity.compact = 0;
5079 }
5080 } else {
5081 affinity.compact = 0;
5082 }
5083 if (affinity.offset) {
5084 affinity.offset =
5085 __kmp_nThreadsPerCore * affinity.offset % __kmp_avail_proc;
5086 }
5087 goto sortTopology;
5088
5089 case affinity_scatter:
5090 if (affinity.compact >= depth) {
5091 affinity.compact = 0;
5092 } else {
5093 affinity.compact = depth - 1 - affinity.compact;
5094 }
5095 goto sortTopology;
5096
5097 case affinity_compact:
5098 if (affinity.compact >= depth) {
5099 affinity.compact = depth - 1;
5100 }
5101 goto sortTopology;
5102
5103 case affinity_balanced:
5104 if (depth <= 1 || is_hidden_helper_affinity) {
5105 KMP_AFF_WARNING(affinity, AffBalancedNotAvail, env_var);
5106 affinity.type = affinity_none;
5107 __kmp_create_affinity_none_places(affinity);
5108 affinity.flags.initialized = TRUE;
5109 return;
5110 } else if (!__kmp_topology->is_uniform()) {
5111 // Save the depth for further usage
5112 __kmp_aff_depth = depth;
5113
5114 int core_level =
5115 __kmp_affinity_find_core_level(__kmp_avail_proc, depth - 1);
5116 int ncores = __kmp_affinity_compute_ncores(__kmp_avail_proc, depth - 1,
5117 core_level);
5118 int maxprocpercore = __kmp_affinity_max_proc_per_core(
5119 __kmp_avail_proc, depth - 1, core_level);
5120
5121 int nproc = ncores * maxprocpercore;
5122 if ((nproc < 2) || (nproc < __kmp_avail_proc)) {
5123 KMP_AFF_WARNING(affinity, AffBalancedNotAvail, env_var);
5124 affinity.type = affinity_none;
5125 __kmp_create_affinity_none_places(affinity);
5126 affinity.flags.initialized = TRUE;
5127 return;
5128 }
5129
5130 procarr = (int *)__kmp_allocate(sizeof(int) * nproc);
5131 for (int i = 0; i < nproc; i++) {
5132 procarr[i] = -1;
5133 }
5134
5135 int lastcore = -1;
5136 int inlastcore = 0;
5137 for (int i = 0; i < __kmp_avail_proc; i++) {
5138 int proc = __kmp_topology->at(i).os_id;
5139 int core = __kmp_affinity_find_core(i, depth - 1, core_level);
5140
5141 if (core == lastcore) {
5142 inlastcore++;
5143 } else {
5144 inlastcore = 0;
5145 }
5146 lastcore = core;
5147
5148 procarr[core * maxprocpercore + inlastcore] = proc;
5149 }
5150 }
5151 if (affinity.compact >= depth) {
5152 affinity.compact = depth - 1;
5153 }
5154
5155 sortTopology:
5156 // Allocate the gtid->affinity mask table.
5157 if (affinity.flags.dups) {
5158 affinity.num_masks = __kmp_avail_proc;
5159 } else {
5160 affinity.num_masks = numUnique;
5161 }
5162
5163 if ((__kmp_nested_proc_bind.bind_types[0] != proc_bind_intel) &&
5164 (__kmp_affinity_num_places > 0) &&
5165 ((unsigned)__kmp_affinity_num_places < affinity.num_masks) &&
5166 !is_hidden_helper_affinity) {
5167 affinity.num_masks = __kmp_affinity_num_places;
5168 }
5169
5170 KMP_CPU_ALLOC_ARRAY(affinity.masks, affinity.num_masks);
5171
5172 // Sort the topology table according to the current setting of
5173 // affinity.compact, then fill out affinity.masks.
5174 __kmp_topology->sort_compact(affinity);
5175 {
5176 int i;
5177 unsigned j;
5178 int num_hw_threads = __kmp_topology->get_num_hw_threads();
5179 kmp_full_mask_modifier_t full_mask;
5180 for (i = 0, j = 0; i < num_hw_threads; i++) {
5181 if ((!affinity.flags.dups) && (!__kmp_topology->at(i).leader)) {
5182 continue;
5183 }
5184 int osId = __kmp_topology->at(i).os_id;
5185
5186 kmp_affin_mask_t *src = KMP_CPU_INDEX(affinity.os_id_masks, osId);
5187 if (KMP_CPU_ISEMPTY(src))
5188 continue;
5189 kmp_affin_mask_t *dest = KMP_CPU_INDEX(affinity.masks, j);
5190 KMP_ASSERT(KMP_CPU_ISSET(osId, src));
5191 KMP_CPU_COPY(dest, src);
5192 full_mask.include(src);
5193 if (++j >= affinity.num_masks) {
5194 break;
5195 }
5196 }
5197 KMP_DEBUG_ASSERT(j == affinity.num_masks);
5198 // See if the places list further restricts or changes the full mask
5199 if (full_mask.restrict_to_mask() && affinity.flags.verbose) {
5200 __kmp_topology->print(env_var);
5201 }
5202 }
5203 // Sort the topology back using ids
5204 __kmp_topology->sort_ids();
5205 break;
5206
5207 default:
5208 KMP_ASSERT2(0, "Unexpected affinity setting");
5209 }
5210 __kmp_aux_affinity_initialize_other_data(affinity);
5211 affinity.flags.initialized = TRUE;
5212}
5213
5214void __kmp_affinity_initialize(kmp_affinity_t &affinity) {
5215 // Much of the code above was written assuming that if a machine was not
5216 // affinity capable, then affinity type == affinity_none.
5217 // We now explicitly represent this as affinity type == affinity_disabled.
5218 // There are too many checks for affinity type == affinity_none in this code.
5219 // Instead of trying to change them all, check if
5220 // affinity type == affinity_disabled, and if so, slam it with affinity_none,
5221 // call the real initialization routine, then restore affinity type to
5222 // affinity_disabled.
5223 int disabled = (affinity.type == affinity_disabled);
5224 if (!KMP_AFFINITY_CAPABLE())
5225 KMP_ASSERT(disabled);
5226 if (disabled)
5227 affinity.type = affinity_none;
5228 __kmp_aux_affinity_initialize(affinity);
5229 if (disabled)
5230 affinity.type = affinity_disabled;
5231}
5232
5233void __kmp_affinity_uninitialize(void) {
5234 for (kmp_affinity_t *affinity : __kmp_affinities) {
5235 if (affinity->masks != NULL)
5236 KMP_CPU_FREE_ARRAY(affinity->masks, affinity->num_masks);
5237 if (affinity->os_id_masks != NULL)
5238 KMP_CPU_FREE_ARRAY(affinity->os_id_masks, affinity->num_os_id_masks);
5239 if (affinity->proclist != NULL)
5240 __kmp_free(affinity->proclist);
5241 if (affinity->ids != NULL)
5242 __kmp_free(affinity->ids);
5243 if (affinity->attrs != NULL)
5244 __kmp_free(affinity->attrs);
5245 *affinity = KMP_AFFINITY_INIT(affinity->env_var);
5246 }
5247 if (__kmp_affin_origMask != NULL) {
5248 if (KMP_AFFINITY_CAPABLE()) {
5249#if KMP_OS_AIX
5250 // Uninitialize by unbinding the thread.
5251 bindprocessor(BINDTHREAD, thread_self(), PROCESSOR_CLASS_ANY);
5252#else
5253 __kmp_set_system_affinity(__kmp_affin_origMask, FALSE);
5254#endif
5255 }
5256 KMP_CPU_FREE(__kmp_affin_origMask);
5257 __kmp_affin_origMask = NULL;
5258 }
5259 __kmp_affinity_num_places = 0;
5260 if (procarr != NULL) {
5261 __kmp_free(procarr);
5262 procarr = NULL;
5263 }
5264 if (__kmp_osid_to_hwthread_map) {
5265 __kmp_free(__kmp_osid_to_hwthread_map);
5266 __kmp_osid_to_hwthread_map = NULL;
5267 }
5268#if KMP_USE_HWLOC
5269 if (__kmp_hwloc_topology != NULL) {
5270 hwloc_topology_destroy(__kmp_hwloc_topology);
5271 __kmp_hwloc_topology = NULL;
5272 }
5273#endif
5274 if (__kmp_hw_subset) {
5275 kmp_hw_subset_t::deallocate(__kmp_hw_subset);
5276 __kmp_hw_subset = nullptr;
5277 }
5278 if (__kmp_topology) {
5279 kmp_topology_t::deallocate(__kmp_topology);
5280 __kmp_topology = nullptr;
5281 }
5282 KMPAffinity::destroy_api();
5283}
5284
5285static void __kmp_select_mask_by_gtid(int gtid, const kmp_affinity_t *affinity,
5286 int *place, kmp_affin_mask_t **mask) {
5287 int mask_idx;
5288 bool is_hidden_helper = KMP_HIDDEN_HELPER_THREAD(gtid);
5289 if (is_hidden_helper)
5290 // The first gtid is the regular primary thread, the second gtid is the main
5291 // thread of hidden team which does not participate in task execution.
5292 mask_idx = gtid - 2;
5293 else
5294 mask_idx = __kmp_adjust_gtid_for_hidden_helpers(gtid);
5295 KMP_DEBUG_ASSERT(affinity->num_masks > 0);
5296 *place = (mask_idx + affinity->offset) % affinity->num_masks;
5297 *mask = KMP_CPU_INDEX(affinity->masks, *place);
5298}
5299
5300// This function initializes the per-thread data concerning affinity including
5301// the mask and topology information
5302void __kmp_affinity_set_init_mask(int gtid, int isa_root) {
5303
5304 kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
5305
5306 // Set the thread topology information to default of unknown
5307 for (int id = 0; id < KMP_HW_LAST; ++id)
5308 th->th.th_topology_ids.ids[id] = kmp_hw_thread_t::UNKNOWN_ID;
5309 th->th.th_topology_attrs = KMP_AFFINITY_ATTRS_UNKNOWN;
5310
5311 if (!KMP_AFFINITY_CAPABLE()) {
5312 return;
5313 }
5314
5315 if (th->th.th_affin_mask == NULL) {
5316 KMP_CPU_ALLOC(th->th.th_affin_mask);
5317 } else {
5318 KMP_CPU_ZERO(th->th.th_affin_mask);
5319 }
5320
5321 // Copy the thread mask to the kmp_info_t structure. If
5322 // __kmp_affinity.type == affinity_none, copy the "full" mask, i.e.
5323 // one that has all of the OS proc ids set, or if
5324 // __kmp_affinity.flags.respect is set, then the full mask is the
5325 // same as the mask of the initialization thread.
5326 kmp_affin_mask_t *mask;
5327 int i;
5328 const kmp_affinity_t *affinity;
5329 bool is_hidden_helper = KMP_HIDDEN_HELPER_THREAD(gtid);
5330
5331 if (is_hidden_helper)
5332 affinity = &__kmp_hh_affinity;
5333 else
5334 affinity = &__kmp_affinity;
5335
5336 if (KMP_AFFINITY_NON_PROC_BIND || is_hidden_helper) {
5337 if ((affinity->type == affinity_none) ||
5338 (affinity->type == affinity_balanced) ||
5339 KMP_HIDDEN_HELPER_MAIN_THREAD(gtid)) {
5340#if KMP_GROUP_AFFINITY
5341 if (__kmp_num_proc_groups > 1) {
5342 return;
5343 }
5344#endif
5345 KMP_ASSERT(__kmp_affin_fullMask != NULL);
5346 i = 0;
5347 mask = __kmp_affin_fullMask;
5348 } else {
5349 __kmp_select_mask_by_gtid(gtid, affinity, &i, &mask);
5350 }
5351 } else {
5352 if (!isa_root || __kmp_nested_proc_bind.bind_types[0] == proc_bind_false) {
5353#if KMP_GROUP_AFFINITY
5354 if (__kmp_num_proc_groups > 1) {
5355 return;
5356 }
5357#endif
5358 KMP_ASSERT(__kmp_affin_fullMask != NULL);
5359 i = KMP_PLACE_ALL;
5360 mask = __kmp_affin_fullMask;
5361 } else {
5362 __kmp_select_mask_by_gtid(gtid, affinity, &i, &mask);
5363 }
5364 }
5365
5366 th->th.th_current_place = i;
5367 if (isa_root && !is_hidden_helper) {
5368 th->th.th_new_place = i;
5369 th->th.th_first_place = 0;
5370 th->th.th_last_place = affinity->num_masks - 1;
5371 } else if (KMP_AFFINITY_NON_PROC_BIND) {
5372 // When using a Non-OMP_PROC_BIND affinity method,
5373 // set all threads' place-partition-var to the entire place list
5374 th->th.th_first_place = 0;
5375 th->th.th_last_place = affinity->num_masks - 1;
5376 }
5377 // Copy topology information associated with the place
5378 if (i >= 0) {
5379 th->th.th_topology_ids = __kmp_affinity.ids[i];
5380 th->th.th_topology_attrs = __kmp_affinity.attrs[i];
5381 }
5382
5383 if (i == KMP_PLACE_ALL) {
5384 KA_TRACE(100, ("__kmp_affinity_set_init_mask: setting T#%d to all places\n",
5385 gtid));
5386 } else {
5387 KA_TRACE(100, ("__kmp_affinity_set_init_mask: setting T#%d to place %d\n",
5388 gtid, i));
5389 }
5390
5391 KMP_CPU_COPY(th->th.th_affin_mask, mask);
5392}
5393
5394void __kmp_affinity_bind_init_mask(int gtid) {
5395 if (!KMP_AFFINITY_CAPABLE()) {
5396 return;
5397 }
5398 kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
5399 const kmp_affinity_t *affinity;
5400 const char *env_var;
5401 bool is_hidden_helper = KMP_HIDDEN_HELPER_THREAD(gtid);
5402
5403 if (is_hidden_helper)
5404 affinity = &__kmp_hh_affinity;
5405 else
5406 affinity = &__kmp_affinity;
5407 env_var = __kmp_get_affinity_env_var(*affinity, /*for_binding=*/true);
5408 /* to avoid duplicate printing (will be correctly printed on barrier) */
5409 if (affinity->flags.verbose && (affinity->type == affinity_none ||
5410 (th->th.th_current_place != KMP_PLACE_ALL &&
5411 affinity->type != affinity_balanced)) &&
5412 !KMP_HIDDEN_HELPER_MAIN_THREAD(gtid)) {
5413 char buf[KMP_AFFIN_MASK_PRINT_LEN];
5414 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5415 th->th.th_affin_mask);
5416 KMP_INFORM(BoundToOSProcSet, env_var, (kmp_int32)getpid(), __kmp_gettid(),
5417 gtid, buf);
5418 }
5419
5420#if KMP_OS_WINDOWS
5421 // On Windows* OS, the process affinity mask might have changed. If the user
5422 // didn't request affinity and this call fails, just continue silently.
5423 // See CQ171393.
5424 if (affinity->type == affinity_none) {
5425 __kmp_set_system_affinity(th->th.th_affin_mask, FALSE);
5426 } else
5427#endif
5428#ifndef KMP_OS_AIX
5429 // Do not set the full mask as the init mask on AIX.
5430 __kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
5431#endif
5432}
5433
5434void __kmp_affinity_bind_place(int gtid) {
5435 // Hidden helper threads should not be affected by OMP_PLACES/OMP_PROC_BIND
5436 if (!KMP_AFFINITY_CAPABLE() || KMP_HIDDEN_HELPER_THREAD(gtid)) {
5437 return;
5438 }
5439
5440 kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
5441
5442 KA_TRACE(100, ("__kmp_affinity_bind_place: binding T#%d to place %d (current "
5443 "place = %d)\n",
5444 gtid, th->th.th_new_place, th->th.th_current_place));
5445
5446 // Check that the new place is within this thread's partition.
5447 KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
5448 KMP_ASSERT(th->th.th_new_place >= 0);
5449 KMP_ASSERT((unsigned)th->th.th_new_place <= __kmp_affinity.num_masks);
5450 if (th->th.th_first_place <= th->th.th_last_place) {
5451 KMP_ASSERT((th->th.th_new_place >= th->th.th_first_place) &&
5452 (th->th.th_new_place <= th->th.th_last_place));
5453 } else {
5454 KMP_ASSERT((th->th.th_new_place <= th->th.th_first_place) ||
5455 (th->th.th_new_place >= th->th.th_last_place));
5456 }
5457
5458 // Copy the thread mask to the kmp_info_t structure,
5459 // and set this thread's affinity.
5460 kmp_affin_mask_t *mask =
5461 KMP_CPU_INDEX(__kmp_affinity.masks, th->th.th_new_place);
5462 KMP_CPU_COPY(th->th.th_affin_mask, mask);
5463 th->th.th_current_place = th->th.th_new_place;
5464
5465 if (__kmp_affinity.flags.verbose) {
5466 char buf[KMP_AFFIN_MASK_PRINT_LEN];
5467 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5468 th->th.th_affin_mask);
5469 KMP_INFORM(BoundToOSProcSet, "OMP_PROC_BIND", (kmp_int32)getpid(),
5470 __kmp_gettid(), gtid, buf);
5471 }
5472 __kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
5473}
5474
5475int __kmp_aux_set_affinity(void **mask) {
5476 int gtid;
5477 kmp_info_t *th;
5478 int retval;
5479
5480 if (!KMP_AFFINITY_CAPABLE()) {
5481 return -1;
5482 }
5483
5484 gtid = __kmp_entry_gtid();
5485 KA_TRACE(
5486 1000, (""); {
5487 char buf[KMP_AFFIN_MASK_PRINT_LEN];
5488 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5489 (kmp_affin_mask_t *)(*mask));
5490 __kmp_debug_printf(
5491 "kmp_set_affinity: setting affinity mask for thread %d = %s\n",
5492 gtid, buf);
5493 });
5494
5495 if (__kmp_env_consistency_check) {
5496 if ((mask == NULL) || (*mask == NULL)) {
5497 KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
5498 } else {
5499 unsigned proc;
5500 int num_procs = 0;
5501
5502 KMP_CPU_SET_ITERATE(proc, ((kmp_affin_mask_t *)(*mask))) {
5503 if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5504 KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
5505 }
5506 if (!KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask))) {
5507 continue;
5508 }
5509 num_procs++;
5510 }
5511 if (num_procs == 0) {
5512 KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
5513 }
5514
5515#if KMP_GROUP_AFFINITY
5516 if (__kmp_get_proc_group((kmp_affin_mask_t *)(*mask)) < 0) {
5517 KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
5518 }
5519#endif /* KMP_GROUP_AFFINITY */
5520 }
5521 }
5522
5523 th = __kmp_threads[gtid];
5524 KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
5525 retval = __kmp_set_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
5526 if (retval == 0) {
5527 KMP_CPU_COPY(th->th.th_affin_mask, (kmp_affin_mask_t *)(*mask));
5528 }
5529
5530 th->th.th_current_place = KMP_PLACE_UNDEFINED;
5531 th->th.th_new_place = KMP_PLACE_UNDEFINED;
5532 th->th.th_first_place = 0;
5533 th->th.th_last_place = __kmp_affinity.num_masks - 1;
5534
5535 // Turn off 4.0 affinity for the current tread at this parallel level.
5536 th->th.th_current_task->td_icvs.proc_bind = proc_bind_false;
5537
5538 return retval;
5539}
5540
5541int __kmp_aux_get_affinity(void **mask) {
5542 int gtid;
5543 int retval;
5544#if KMP_OS_WINDOWS || KMP_OS_AIX || KMP_DEBUG
5545 kmp_info_t *th;
5546#endif
5547 if (!KMP_AFFINITY_CAPABLE()) {
5548 return -1;
5549 }
5550
5551 gtid = __kmp_entry_gtid();
5552#if KMP_OS_WINDOWS || KMP_OS_AIX || KMP_DEBUG
5553 th = __kmp_threads[gtid];
5554#else
5555 (void)gtid; // unused variable
5556#endif
5557 KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
5558
5559 KA_TRACE(
5560 1000, (""); {
5561 char buf[KMP_AFFIN_MASK_PRINT_LEN];
5562 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5563 th->th.th_affin_mask);
5564 __kmp_printf(
5565 "kmp_get_affinity: stored affinity mask for thread %d = %s\n", gtid,
5566 buf);
5567 });
5568
5569 if (__kmp_env_consistency_check) {
5570 if ((mask == NULL) || (*mask == NULL)) {
5571 KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity");
5572 }
5573 }
5574
5575#if !KMP_OS_WINDOWS && !KMP_OS_AIX
5576
5577 retval = __kmp_get_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
5578 KA_TRACE(
5579 1000, (""); {
5580 char buf[KMP_AFFIN_MASK_PRINT_LEN];
5581 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5582 (kmp_affin_mask_t *)(*mask));
5583 __kmp_printf(
5584 "kmp_get_affinity: system affinity mask for thread %d = %s\n", gtid,
5585 buf);
5586 });
5587 return retval;
5588
5589#else
5590 (void)retval;
5591
5592 KMP_CPU_COPY((kmp_affin_mask_t *)(*mask), th->th.th_affin_mask);
5593 return 0;
5594
5595#endif /* !KMP_OS_WINDOWS && !KMP_OS_AIX */
5596}
5597
5598int __kmp_aux_get_affinity_max_proc() {
5599 if (!KMP_AFFINITY_CAPABLE()) {
5600 return 0;
5601 }
5602#if KMP_GROUP_AFFINITY
5603 if (__kmp_num_proc_groups > 1) {
5604 return (int)(__kmp_num_proc_groups * sizeof(DWORD_PTR) * CHAR_BIT);
5605 }
5606#endif
5607 return __kmp_xproc;
5608}
5609
5610int __kmp_aux_set_affinity_mask_proc(int proc, void **mask) {
5611 if (!KMP_AFFINITY_CAPABLE()) {
5612 return -1;
5613 }
5614
5615 KA_TRACE(
5616 1000, (""); {
5617 int gtid = __kmp_entry_gtid();
5618 char buf[KMP_AFFIN_MASK_PRINT_LEN];
5619 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5620 (kmp_affin_mask_t *)(*mask));
5621 __kmp_debug_printf("kmp_set_affinity_mask_proc: setting proc %d in "
5622 "affinity mask for thread %d = %s\n",
5623 proc, gtid, buf);
5624 });
5625
5626 if (__kmp_env_consistency_check) {
5627 if ((mask == NULL) || (*mask == NULL)) {
5628 KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity_mask_proc");
5629 }
5630 }
5631
5632 if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5633 return -1;
5634 }
5635 if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5636 return -2;
5637 }
5638
5639 KMP_CPU_SET(proc, (kmp_affin_mask_t *)(*mask));
5640 return 0;
5641}
5642
5643int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask) {
5644 if (!KMP_AFFINITY_CAPABLE()) {
5645 return -1;
5646 }
5647
5648 KA_TRACE(
5649 1000, (""); {
5650 int gtid = __kmp_entry_gtid();
5651 char buf[KMP_AFFIN_MASK_PRINT_LEN];
5652 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5653 (kmp_affin_mask_t *)(*mask));
5654 __kmp_debug_printf("kmp_unset_affinity_mask_proc: unsetting proc %d in "
5655 "affinity mask for thread %d = %s\n",
5656 proc, gtid, buf);
5657 });
5658
5659 if (__kmp_env_consistency_check) {
5660 if ((mask == NULL) || (*mask == NULL)) {
5661 KMP_FATAL(AffinityInvalidMask, "kmp_unset_affinity_mask_proc");
5662 }
5663 }
5664
5665 if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5666 return -1;
5667 }
5668 if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5669 return -2;
5670 }
5671
5672 KMP_CPU_CLR(proc, (kmp_affin_mask_t *)(*mask));
5673 return 0;
5674}
5675
5676int __kmp_aux_get_affinity_mask_proc(int proc, void **mask) {
5677 if (!KMP_AFFINITY_CAPABLE()) {
5678 return -1;
5679 }
5680
5681 KA_TRACE(
5682 1000, (""); {
5683 int gtid = __kmp_entry_gtid();
5684 char buf[KMP_AFFIN_MASK_PRINT_LEN];
5685 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5686 (kmp_affin_mask_t *)(*mask));
5687 __kmp_debug_printf("kmp_get_affinity_mask_proc: getting proc %d in "
5688 "affinity mask for thread %d = %s\n",
5689 proc, gtid, buf);
5690 });
5691
5692 if (__kmp_env_consistency_check) {
5693 if ((mask == NULL) || (*mask == NULL)) {
5694 KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity_mask_proc");
5695 }
5696 }
5697
5698 if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5699 return -1;
5700 }
5701 if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5702 return 0;
5703 }
5704
5705 return KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask));
5706}
5707
5708#if KMP_WEIGHTED_ITERATIONS_SUPPORTED
5709// Returns first os proc id with ATOM core
5710int __kmp_get_first_osid_with_ecore(void) {
5711 int low = 0;
5712 int high = __kmp_topology->get_num_hw_threads() - 1;
5713 int mid = 0;
5714 while (high - low > 1) {
5715 mid = (high + low) / 2;
5716 if (__kmp_topology->at(mid).attrs.get_core_type() ==
5717 KMP_HW_CORE_TYPE_CORE) {
5718 low = mid + 1;
5719 } else {
5720 high = mid;
5721 }
5722 }
5723 if (__kmp_topology->at(mid).attrs.get_core_type() == KMP_HW_CORE_TYPE_ATOM) {
5724 return mid;
5725 }
5726 return -1;
5727}
5728#endif
5729
5730// Dynamic affinity settings - Affinity balanced
5731void __kmp_balanced_affinity(kmp_info_t *th, int nthreads) {
5732 KMP_DEBUG_ASSERT(th);
5733 bool fine_gran = true;
5734 int tid = th->th.th_info.ds.ds_tid;
5735 const char *env_var = "KMP_AFFINITY";
5736
5737 // Do not perform balanced affinity for the hidden helper threads
5738 if (KMP_HIDDEN_HELPER_THREAD(__kmp_gtid_from_thread(th)))
5739 return;
5740
5741 switch (__kmp_affinity.gran) {
5742 case KMP_HW_THREAD:
5743 break;
5744 case KMP_HW_CORE:
5745 if (__kmp_nThreadsPerCore > 1) {
5746 fine_gran = false;
5747 }
5748 break;
5749 case KMP_HW_SOCKET:
5750 if (nCoresPerPkg > 1) {
5751 fine_gran = false;
5752 }
5753 break;
5754 default:
5755 fine_gran = false;
5756 }
5757
5758 if (__kmp_topology->is_uniform()) {
5759 int coreID;
5760 int threadID;
5761 // Number of hyper threads per core in HT machine
5762 int __kmp_nth_per_core = __kmp_avail_proc / __kmp_ncores;
5763 // Number of cores
5764 int ncores = __kmp_ncores;
5765 if ((nPackages > 1) && (__kmp_nth_per_core <= 1)) {
5766 __kmp_nth_per_core = __kmp_avail_proc / nPackages;
5767 ncores = nPackages;
5768 }
5769 // How many threads will be bound to each core
5770 int chunk = nthreads / ncores;
5771 // How many cores will have an additional thread bound to it - "big cores"
5772 int big_cores = nthreads % ncores;
5773 // Number of threads on the big cores
5774 int big_nth = (chunk + 1) * big_cores;
5775 if (tid < big_nth) {
5776 coreID = tid / (chunk + 1);
5777 threadID = (tid % (chunk + 1)) % __kmp_nth_per_core;
5778 } else { // tid >= big_nth
5779 coreID = (tid - big_cores) / chunk;
5780 threadID = ((tid - big_cores) % chunk) % __kmp_nth_per_core;
5781 }
5782 KMP_DEBUG_ASSERT2(KMP_AFFINITY_CAPABLE(),
5783 "Illegal set affinity operation when not capable");
5784
5785 kmp_affin_mask_t *mask = th->th.th_affin_mask;
5786 KMP_CPU_ZERO(mask);
5787
5788 if (fine_gran) {
5789 int osID =
5790 __kmp_topology->at(coreID * __kmp_nth_per_core + threadID).os_id;
5791 KMP_CPU_SET(osID, mask);
5792 } else {
5793 for (int i = 0; i < __kmp_nth_per_core; i++) {
5794 int osID;
5795 osID = __kmp_topology->at(coreID * __kmp_nth_per_core + i).os_id;
5796 KMP_CPU_SET(osID, mask);
5797 }
5798 }
5799 if (__kmp_affinity.flags.verbose) {
5800 char buf[KMP_AFFIN_MASK_PRINT_LEN];
5801 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
5802 KMP_INFORM(BoundToOSProcSet, env_var, (kmp_int32)getpid(), __kmp_gettid(),
5803 tid, buf);
5804 }
5805 __kmp_affinity_get_thread_topology_info(th);
5806 __kmp_set_system_affinity(mask, TRUE);
5807 } else { // Non-uniform topology
5808
5809 kmp_affin_mask_t *mask = th->th.th_affin_mask;
5810 KMP_CPU_ZERO(mask);
5811
5812 int core_level =
5813 __kmp_affinity_find_core_level(__kmp_avail_proc, __kmp_aff_depth - 1);
5814 int ncores = __kmp_affinity_compute_ncores(__kmp_avail_proc,
5815 __kmp_aff_depth - 1, core_level);
5816 int nth_per_core = __kmp_affinity_max_proc_per_core(
5817 __kmp_avail_proc, __kmp_aff_depth - 1, core_level);
5818
5819 // For performance gain consider the special case nthreads ==
5820 // __kmp_avail_proc
5821 if (nthreads == __kmp_avail_proc) {
5822 if (fine_gran) {
5823 int osID = __kmp_topology->at(tid).os_id;
5824 KMP_CPU_SET(osID, mask);
5825 } else {
5826 int core =
5827 __kmp_affinity_find_core(tid, __kmp_aff_depth - 1, core_level);
5828 for (int i = 0; i < __kmp_avail_proc; i++) {
5829 int osID = __kmp_topology->at(i).os_id;
5830 if (__kmp_affinity_find_core(i, __kmp_aff_depth - 1, core_level) ==
5831 core) {
5832 KMP_CPU_SET(osID, mask);
5833 }
5834 }
5835 }
5836 } else if (nthreads <= ncores) {
5837
5838 int core = 0;
5839 for (int i = 0; i < ncores; i++) {
5840 // Check if this core from procarr[] is in the mask
5841 int in_mask = 0;
5842 for (int j = 0; j < nth_per_core; j++) {
5843 if (procarr[i * nth_per_core + j] != -1) {
5844 in_mask = 1;
5845 break;
5846 }
5847 }
5848 if (in_mask) {
5849 if (tid == core) {
5850 for (int j = 0; j < nth_per_core; j++) {
5851 int osID = procarr[i * nth_per_core + j];
5852 if (osID != -1) {
5853 KMP_CPU_SET(osID, mask);
5854 // For fine granularity it is enough to set the first available
5855 // osID for this core
5856 if (fine_gran) {
5857 break;
5858 }
5859 }
5860 }
5861 break;
5862 } else {
5863 core++;
5864 }
5865 }
5866 }
5867 } else { // nthreads > ncores
5868 // Array to save the number of processors at each core
5869 int *nproc_at_core = (int *)KMP_ALLOCA(sizeof(int) * ncores);
5870 // Array to save the number of cores with "x" available processors;
5871 int *ncores_with_x_procs =
5872 (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
5873 // Array to save the number of cores with # procs from x to nth_per_core
5874 int *ncores_with_x_to_max_procs =
5875 (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
5876
5877 for (int i = 0; i <= nth_per_core; i++) {
5878 ncores_with_x_procs[i] = 0;
5879 ncores_with_x_to_max_procs[i] = 0;
5880 }
5881
5882 for (int i = 0; i < ncores; i++) {
5883 int cnt = 0;
5884 for (int j = 0; j < nth_per_core; j++) {
5885 if (procarr[i * nth_per_core + j] != -1) {
5886 cnt++;
5887 }
5888 }
5889 nproc_at_core[i] = cnt;
5890 ncores_with_x_procs[cnt]++;
5891 }
5892
5893 for (int i = 0; i <= nth_per_core; i++) {
5894 for (int j = i; j <= nth_per_core; j++) {
5895 ncores_with_x_to_max_procs[i] += ncores_with_x_procs[j];
5896 }
5897 }
5898
5899 // Max number of processors
5900 int nproc = nth_per_core * ncores;
5901 // An array to keep number of threads per each context
5902 int *newarr = (int *)__kmp_allocate(sizeof(int) * nproc);
5903 for (int i = 0; i < nproc; i++) {
5904 newarr[i] = 0;
5905 }
5906
5907 int nth = nthreads;
5908 int flag = 0;
5909 while (nth > 0) {
5910 for (int j = 1; j <= nth_per_core; j++) {
5911 int cnt = ncores_with_x_to_max_procs[j];
5912 for (int i = 0; i < ncores; i++) {
5913 // Skip the core with 0 processors
5914 if (nproc_at_core[i] == 0) {
5915 continue;
5916 }
5917 for (int k = 0; k < nth_per_core; k++) {
5918 if (procarr[i * nth_per_core + k] != -1) {
5919 if (newarr[i * nth_per_core + k] == 0) {
5920 newarr[i * nth_per_core + k] = 1;
5921 cnt--;
5922 nth--;
5923 break;
5924 } else {
5925 if (flag != 0) {
5926 newarr[i * nth_per_core + k]++;
5927 cnt--;
5928 nth--;
5929 break;
5930 }
5931 }
5932 }
5933 }
5934 if (cnt == 0 || nth == 0) {
5935 break;
5936 }
5937 }
5938 if (nth == 0) {
5939 break;
5940 }
5941 }
5942 flag = 1;
5943 }
5944 int sum = 0;
5945 for (int i = 0; i < nproc; i++) {
5946 sum += newarr[i];
5947 if (sum > tid) {
5948 if (fine_gran) {
5949 int osID = procarr[i];
5950 KMP_CPU_SET(osID, mask);
5951 } else {
5952 int coreID = i / nth_per_core;
5953 for (int ii = 0; ii < nth_per_core; ii++) {
5954 int osID = procarr[coreID * nth_per_core + ii];
5955 if (osID != -1) {
5956 KMP_CPU_SET(osID, mask);
5957 }
5958 }
5959 }
5960 break;
5961 }
5962 }
5963 __kmp_free(newarr);
5964 }
5965
5966 if (__kmp_affinity.flags.verbose) {
5967 char buf[KMP_AFFIN_MASK_PRINT_LEN];
5968 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
5969 KMP_INFORM(BoundToOSProcSet, env_var, (kmp_int32)getpid(), __kmp_gettid(),
5970 tid, buf);
5971 }
5972 __kmp_affinity_get_thread_topology_info(th);
5973 __kmp_set_system_affinity(mask, TRUE);
5974 }
5975}
5976
5977#if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_DRAGONFLY || \
5978 KMP_OS_AIX
5979// We don't need this entry for Windows because
5980// there is GetProcessAffinityMask() api
5981//
5982// The intended usage is indicated by these steps:
5983// 1) The user gets the current affinity mask
5984// 2) Then sets the affinity by calling this function
5985// 3) Error check the return value
5986// 4) Use non-OpenMP parallelization
5987// 5) Reset the affinity to what was stored in step 1)
5988#ifdef __cplusplus
5989extern "C"
5990#endif
5991 int
5992 kmp_set_thread_affinity_mask_initial()
5993// the function returns 0 on success,
5994// -1 if we cannot bind thread
5995// >0 (errno) if an error happened during binding
5996{
5997 int gtid = __kmp_get_gtid();
5998 if (gtid < 0) {
5999 // Do not touch non-omp threads
6000 KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
6001 "non-omp thread, returning\n"));
6002 return -1;
6003 }
6004 if (!KMP_AFFINITY_CAPABLE() || !__kmp_init_middle) {
6005 KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
6006 "affinity not initialized, returning\n"));
6007 return -1;
6008 }
6009 KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
6010 "set full mask for thread %d\n",
6011 gtid));
6012 KMP_DEBUG_ASSERT(__kmp_affin_fullMask != NULL);
6013#if KMP_OS_AIX
6014 return bindprocessor(BINDTHREAD, thread_self(), PROCESSOR_CLASS_ANY);
6015#else
6016 return __kmp_set_system_affinity(__kmp_affin_fullMask, FALSE);
6017#endif
6018}
6019#endif
6020
6021#endif // KMP_AFFINITY_SUPPORTED
int try_open(const char *filename, const char *mode)
Definition kmp.h:4759