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