1 /*
2 * kmp_affinity.cpp -- affinity management
3 */
4
5 //===----------------------------------------------------------------------===//
6 //
7 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
8 // See https://llvm.org/LICENSE.txt for license information.
9 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
10 //
11 //===----------------------------------------------------------------------===//
12
13 #include "kmp.h"
14 #include "kmp_affinity.h"
15 #include "kmp_i18n.h"
16 #include "kmp_io.h"
17 #include "kmp_str.h"
18 #include "kmp_wrapper_getpid.h"
19 #if KMP_USE_HIER_SCHED
20 #include "kmp_dispatch_hier.h"
21 #endif
22 #if KMP_USE_HWLOC
23 // Copied from hwloc
24 #define HWLOC_GROUP_KIND_INTEL_MODULE 102
25 #define HWLOC_GROUP_KIND_INTEL_TILE 103
26 #define HWLOC_GROUP_KIND_INTEL_DIE 104
27 #define HWLOC_GROUP_KIND_WINDOWS_PROCESSOR_GROUP 220
28 #endif
29 #include <ctype.h>
30
31 // The machine topology
32 kmp_topology_t *__kmp_topology = nullptr;
33 // KMP_HW_SUBSET environment variable
34 kmp_hw_subset_t *__kmp_hw_subset = nullptr;
35
36 // Store the real or imagined machine hierarchy here
37 static hierarchy_info machine_hierarchy;
38
__kmp_cleanup_hierarchy()39 void __kmp_cleanup_hierarchy() { machine_hierarchy.fini(); }
40
41 #if KMP_AFFINITY_SUPPORTED
42 // Helper class to see if place lists further restrict the fullMask
43 class kmp_full_mask_modifier_t {
44 kmp_affin_mask_t *mask;
45
46 public:
kmp_full_mask_modifier_t()47 kmp_full_mask_modifier_t() {
48 KMP_CPU_ALLOC(mask);
49 KMP_CPU_ZERO(mask);
50 }
~kmp_full_mask_modifier_t()51 ~kmp_full_mask_modifier_t() {
52 KMP_CPU_FREE(mask);
53 mask = nullptr;
54 }
include(const kmp_affin_mask_t * other)55 void include(const kmp_affin_mask_t *other) { KMP_CPU_UNION(mask, other); }
56 // If the new full mask is different from the current full mask,
57 // then switch them. Returns true if full mask was affected, false otherwise.
restrict_to_mask()58 bool restrict_to_mask() {
59 // See if the new mask further restricts or changes the full mask
60 if (KMP_CPU_EQUAL(__kmp_affin_fullMask, mask) || KMP_CPU_ISEMPTY(mask))
61 return false;
62 return __kmp_topology->restrict_to_mask(mask);
63 }
64 };
65
66 static inline const char *
__kmp_get_affinity_env_var(const kmp_affinity_t & affinity,bool for_binding=false)67 __kmp_get_affinity_env_var(const kmp_affinity_t &affinity,
68 bool for_binding = false) {
69 if (affinity.flags.omp_places) {
70 if (for_binding)
71 return "OMP_PROC_BIND";
72 return "OMP_PLACES";
73 }
74 return affinity.env_var;
75 }
76 #endif // KMP_AFFINITY_SUPPORTED
77
__kmp_get_hierarchy(kmp_uint32 nproc,kmp_bstate_t * thr_bar)78 void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar) {
79 kmp_uint32 depth;
80 // The test below is true if affinity is available, but set to "none". Need to
81 // init on first use of hierarchical barrier.
82 if (TCR_1(machine_hierarchy.uninitialized))
83 machine_hierarchy.init(nproc);
84
85 // Adjust the hierarchy in case num threads exceeds original
86 if (nproc > machine_hierarchy.base_num_threads)
87 machine_hierarchy.resize(nproc);
88
89 depth = machine_hierarchy.depth;
90 KMP_DEBUG_ASSERT(depth > 0);
91
92 thr_bar->depth = depth;
93 __kmp_type_convert(machine_hierarchy.numPerLevel[0] - 1,
94 &(thr_bar->base_leaf_kids));
95 thr_bar->skip_per_level = machine_hierarchy.skipPerLevel;
96 }
97
98 static int nCoresPerPkg, nPackages;
99 static int __kmp_nThreadsPerCore;
100 #ifndef KMP_DFLT_NTH_CORES
101 static int __kmp_ncores;
102 #endif
103
__kmp_hw_get_catalog_string(kmp_hw_t type,bool plural)104 const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural) {
105 switch (type) {
106 case KMP_HW_SOCKET:
107 return ((plural) ? KMP_I18N_STR(Sockets) : KMP_I18N_STR(Socket));
108 case KMP_HW_DIE:
109 return ((plural) ? KMP_I18N_STR(Dice) : KMP_I18N_STR(Die));
110 case KMP_HW_MODULE:
111 return ((plural) ? KMP_I18N_STR(Modules) : KMP_I18N_STR(Module));
112 case KMP_HW_TILE:
113 return ((plural) ? KMP_I18N_STR(Tiles) : KMP_I18N_STR(Tile));
114 case KMP_HW_NUMA:
115 return ((plural) ? KMP_I18N_STR(NumaDomains) : KMP_I18N_STR(NumaDomain));
116 case KMP_HW_L3:
117 return ((plural) ? KMP_I18N_STR(L3Caches) : KMP_I18N_STR(L3Cache));
118 case KMP_HW_L2:
119 return ((plural) ? KMP_I18N_STR(L2Caches) : KMP_I18N_STR(L2Cache));
120 case KMP_HW_L1:
121 return ((plural) ? KMP_I18N_STR(L1Caches) : KMP_I18N_STR(L1Cache));
122 case KMP_HW_LLC:
123 return ((plural) ? KMP_I18N_STR(LLCaches) : KMP_I18N_STR(LLCache));
124 case KMP_HW_CORE:
125 return ((plural) ? KMP_I18N_STR(Cores) : KMP_I18N_STR(Core));
126 case KMP_HW_THREAD:
127 return ((plural) ? KMP_I18N_STR(Threads) : KMP_I18N_STR(Thread));
128 case KMP_HW_PROC_GROUP:
129 return ((plural) ? KMP_I18N_STR(ProcGroups) : KMP_I18N_STR(ProcGroup));
130 case KMP_HW_UNKNOWN:
131 case KMP_HW_LAST:
132 return KMP_I18N_STR(Unknown);
133 }
134 KMP_ASSERT2(false, "Unhandled kmp_hw_t enumeration");
135 KMP_BUILTIN_UNREACHABLE;
136 }
137
__kmp_hw_get_keyword(kmp_hw_t type,bool plural)138 const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural) {
139 switch (type) {
140 case KMP_HW_SOCKET:
141 return ((plural) ? "sockets" : "socket");
142 case KMP_HW_DIE:
143 return ((plural) ? "dice" : "die");
144 case KMP_HW_MODULE:
145 return ((plural) ? "modules" : "module");
146 case KMP_HW_TILE:
147 return ((plural) ? "tiles" : "tile");
148 case KMP_HW_NUMA:
149 return ((plural) ? "numa_domains" : "numa_domain");
150 case KMP_HW_L3:
151 return ((plural) ? "l3_caches" : "l3_cache");
152 case KMP_HW_L2:
153 return ((plural) ? "l2_caches" : "l2_cache");
154 case KMP_HW_L1:
155 return ((plural) ? "l1_caches" : "l1_cache");
156 case KMP_HW_LLC:
157 return ((plural) ? "ll_caches" : "ll_cache");
158 case KMP_HW_CORE:
159 return ((plural) ? "cores" : "core");
160 case KMP_HW_THREAD:
161 return ((plural) ? "threads" : "thread");
162 case KMP_HW_PROC_GROUP:
163 return ((plural) ? "proc_groups" : "proc_group");
164 case KMP_HW_UNKNOWN:
165 case KMP_HW_LAST:
166 return ((plural) ? "unknowns" : "unknown");
167 }
168 KMP_ASSERT2(false, "Unhandled kmp_hw_t enumeration");
169 KMP_BUILTIN_UNREACHABLE;
170 }
171
__kmp_hw_get_core_type_string(kmp_hw_core_type_t type)172 const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type) {
173 switch (type) {
174 case KMP_HW_CORE_TYPE_UNKNOWN:
175 case KMP_HW_MAX_NUM_CORE_TYPES:
176 return "unknown";
177 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
178 case KMP_HW_CORE_TYPE_ATOM:
179 return "Intel Atom(R) processor";
180 case KMP_HW_CORE_TYPE_CORE:
181 return "Intel(R) Core(TM) processor";
182 #endif
183 }
184 KMP_ASSERT2(false, "Unhandled kmp_hw_core_type_t enumeration");
185 KMP_BUILTIN_UNREACHABLE;
186 }
187
188 #if KMP_AFFINITY_SUPPORTED
189 // If affinity is supported, check the affinity
190 // verbose and warning flags before printing warning
191 #define KMP_AFF_WARNING(s, ...) \
192 if (s.flags.verbose || (s.flags.warnings && (s.type != affinity_none))) { \
193 KMP_WARNING(__VA_ARGS__); \
194 }
195 #else
196 #define KMP_AFF_WARNING(s, ...) KMP_WARNING(__VA_ARGS__)
197 #endif
198
199 ////////////////////////////////////////////////////////////////////////////////
200 // kmp_hw_thread_t methods
compare_ids(const void * a,const void * b)201 int kmp_hw_thread_t::compare_ids(const void *a, const void *b) {
202 const kmp_hw_thread_t *ahwthread = (const kmp_hw_thread_t *)a;
203 const kmp_hw_thread_t *bhwthread = (const kmp_hw_thread_t *)b;
204 int depth = __kmp_topology->get_depth();
205 for (int level = 0; level < depth; ++level) {
206 if (ahwthread->ids[level] < bhwthread->ids[level])
207 return -1;
208 else if (ahwthread->ids[level] > bhwthread->ids[level])
209 return 1;
210 }
211 if (ahwthread->os_id < bhwthread->os_id)
212 return -1;
213 else if (ahwthread->os_id > bhwthread->os_id)
214 return 1;
215 return 0;
216 }
217
218 #if KMP_AFFINITY_SUPPORTED
compare_compact(const void * a,const void * b)219 int kmp_hw_thread_t::compare_compact(const void *a, const void *b) {
220 int i;
221 const kmp_hw_thread_t *aa = (const kmp_hw_thread_t *)a;
222 const kmp_hw_thread_t *bb = (const kmp_hw_thread_t *)b;
223 int depth = __kmp_topology->get_depth();
224 int compact = __kmp_topology->compact;
225 KMP_DEBUG_ASSERT(compact >= 0);
226 KMP_DEBUG_ASSERT(compact <= depth);
227 for (i = 0; i < compact; i++) {
228 int j = depth - i - 1;
229 if (aa->sub_ids[j] < bb->sub_ids[j])
230 return -1;
231 if (aa->sub_ids[j] > bb->sub_ids[j])
232 return 1;
233 }
234 for (; i < depth; i++) {
235 int j = i - compact;
236 if (aa->sub_ids[j] < bb->sub_ids[j])
237 return -1;
238 if (aa->sub_ids[j] > bb->sub_ids[j])
239 return 1;
240 }
241 return 0;
242 }
243 #endif
244
print() const245 void kmp_hw_thread_t::print() const {
246 int depth = __kmp_topology->get_depth();
247 printf("%4d ", os_id);
248 for (int i = 0; i < depth; ++i) {
249 printf("%4d ", ids[i]);
250 }
251 if (attrs) {
252 if (attrs.is_core_type_valid())
253 printf(" (%s)", __kmp_hw_get_core_type_string(attrs.get_core_type()));
254 if (attrs.is_core_eff_valid())
255 printf(" (eff=%d)", attrs.get_core_eff());
256 }
257 if (leader)
258 printf(" (leader)");
259 printf("\n");
260 }
261
262 ////////////////////////////////////////////////////////////////////////////////
263 // kmp_topology_t methods
264
265 // Add a layer to the topology based on the ids. Assume the topology
266 // is perfectly nested (i.e., so no object has more than one parent)
_insert_layer(kmp_hw_t type,const int * ids)267 void kmp_topology_t::_insert_layer(kmp_hw_t type, const int *ids) {
268 // Figure out where the layer should go by comparing the ids of the current
269 // layers with the new ids
270 int target_layer;
271 int previous_id = kmp_hw_thread_t::UNKNOWN_ID;
272 int previous_new_id = kmp_hw_thread_t::UNKNOWN_ID;
273
274 // Start from the highest layer and work down to find target layer
275 // If new layer is equal to another layer then put the new layer above
276 for (target_layer = 0; target_layer < depth; ++target_layer) {
277 bool layers_equal = true;
278 bool strictly_above_target_layer = false;
279 for (int i = 0; i < num_hw_threads; ++i) {
280 int id = hw_threads[i].ids[target_layer];
281 int new_id = ids[i];
282 if (id != previous_id && new_id == previous_new_id) {
283 // Found the layer we are strictly above
284 strictly_above_target_layer = true;
285 layers_equal = false;
286 break;
287 } else if (id == previous_id && new_id != previous_new_id) {
288 // Found a layer we are below. Move to next layer and check.
289 layers_equal = false;
290 break;
291 }
292 previous_id = id;
293 previous_new_id = new_id;
294 }
295 if (strictly_above_target_layer || layers_equal)
296 break;
297 }
298
299 // Found the layer we are above. Now move everything to accommodate the new
300 // layer. And put the new ids and type into the topology.
301 for (int i = depth - 1, j = depth; i >= target_layer; --i, --j)
302 types[j] = types[i];
303 types[target_layer] = type;
304 for (int k = 0; k < num_hw_threads; ++k) {
305 for (int i = depth - 1, j = depth; i >= target_layer; --i, --j)
306 hw_threads[k].ids[j] = hw_threads[k].ids[i];
307 hw_threads[k].ids[target_layer] = ids[k];
308 }
309 equivalent[type] = type;
310 depth++;
311 }
312
313 #if KMP_GROUP_AFFINITY
314 // Insert the Windows Processor Group structure into the topology
_insert_windows_proc_groups()315 void kmp_topology_t::_insert_windows_proc_groups() {
316 // Do not insert the processor group structure for a single group
317 if (__kmp_num_proc_groups == 1)
318 return;
319 kmp_affin_mask_t *mask;
320 int *ids = (int *)__kmp_allocate(sizeof(int) * num_hw_threads);
321 KMP_CPU_ALLOC(mask);
322 for (int i = 0; i < num_hw_threads; ++i) {
323 KMP_CPU_ZERO(mask);
324 KMP_CPU_SET(hw_threads[i].os_id, mask);
325 ids[i] = __kmp_get_proc_group(mask);
326 }
327 KMP_CPU_FREE(mask);
328 _insert_layer(KMP_HW_PROC_GROUP, ids);
329 __kmp_free(ids);
330 }
331 #endif
332
333 // Remove layers that don't add information to the topology.
334 // This is done by having the layer take on the id = UNKNOWN_ID (-1)
_remove_radix1_layers()335 void kmp_topology_t::_remove_radix1_layers() {
336 int preference[KMP_HW_LAST];
337 int top_index1, top_index2;
338 // Set up preference associative array
339 preference[KMP_HW_SOCKET] = 110;
340 preference[KMP_HW_PROC_GROUP] = 100;
341 preference[KMP_HW_CORE] = 95;
342 preference[KMP_HW_THREAD] = 90;
343 preference[KMP_HW_NUMA] = 85;
344 preference[KMP_HW_DIE] = 80;
345 preference[KMP_HW_TILE] = 75;
346 preference[KMP_HW_MODULE] = 73;
347 preference[KMP_HW_L3] = 70;
348 preference[KMP_HW_L2] = 65;
349 preference[KMP_HW_L1] = 60;
350 preference[KMP_HW_LLC] = 5;
351 top_index1 = 0;
352 top_index2 = 1;
353 while (top_index1 < depth - 1 && top_index2 < depth) {
354 kmp_hw_t type1 = types[top_index1];
355 kmp_hw_t type2 = types[top_index2];
356 KMP_ASSERT_VALID_HW_TYPE(type1);
357 KMP_ASSERT_VALID_HW_TYPE(type2);
358 // Do not allow the three main topology levels (sockets, cores, threads) to
359 // be compacted down
360 if ((type1 == KMP_HW_THREAD || type1 == KMP_HW_CORE ||
361 type1 == KMP_HW_SOCKET) &&
362 (type2 == KMP_HW_THREAD || type2 == KMP_HW_CORE ||
363 type2 == KMP_HW_SOCKET)) {
364 top_index1 = top_index2++;
365 continue;
366 }
367 bool radix1 = true;
368 bool all_same = true;
369 int id1 = hw_threads[0].ids[top_index1];
370 int id2 = hw_threads[0].ids[top_index2];
371 int pref1 = preference[type1];
372 int pref2 = preference[type2];
373 for (int hwidx = 1; hwidx < num_hw_threads; ++hwidx) {
374 if (hw_threads[hwidx].ids[top_index1] == id1 &&
375 hw_threads[hwidx].ids[top_index2] != id2) {
376 radix1 = false;
377 break;
378 }
379 if (hw_threads[hwidx].ids[top_index2] != id2)
380 all_same = false;
381 id1 = hw_threads[hwidx].ids[top_index1];
382 id2 = hw_threads[hwidx].ids[top_index2];
383 }
384 if (radix1) {
385 // Select the layer to remove based on preference
386 kmp_hw_t remove_type, keep_type;
387 int remove_layer, remove_layer_ids;
388 if (pref1 > pref2) {
389 remove_type = type2;
390 remove_layer = remove_layer_ids = top_index2;
391 keep_type = type1;
392 } else {
393 remove_type = type1;
394 remove_layer = remove_layer_ids = top_index1;
395 keep_type = type2;
396 }
397 // If all the indexes for the second (deeper) layer are the same.
398 // e.g., all are zero, then make sure to keep the first layer's ids
399 if (all_same)
400 remove_layer_ids = top_index2;
401 // Remove radix one type by setting the equivalence, removing the id from
402 // the hw threads and removing the layer from types and depth
403 set_equivalent_type(remove_type, keep_type);
404 for (int idx = 0; idx < num_hw_threads; ++idx) {
405 kmp_hw_thread_t &hw_thread = hw_threads[idx];
406 for (int d = remove_layer_ids; d < depth - 1; ++d)
407 hw_thread.ids[d] = hw_thread.ids[d + 1];
408 }
409 for (int idx = remove_layer; idx < depth - 1; ++idx)
410 types[idx] = types[idx + 1];
411 depth--;
412 } else {
413 top_index1 = top_index2++;
414 }
415 }
416 KMP_ASSERT(depth > 0);
417 }
418
_set_last_level_cache()419 void kmp_topology_t::_set_last_level_cache() {
420 if (get_equivalent_type(KMP_HW_L3) != KMP_HW_UNKNOWN)
421 set_equivalent_type(KMP_HW_LLC, KMP_HW_L3);
422 else if (get_equivalent_type(KMP_HW_L2) != KMP_HW_UNKNOWN)
423 set_equivalent_type(KMP_HW_LLC, KMP_HW_L2);
424 #if KMP_MIC_SUPPORTED
425 else if (__kmp_mic_type == mic3) {
426 if (get_equivalent_type(KMP_HW_L2) != KMP_HW_UNKNOWN)
427 set_equivalent_type(KMP_HW_LLC, KMP_HW_L2);
428 else if (get_equivalent_type(KMP_HW_TILE) != KMP_HW_UNKNOWN)
429 set_equivalent_type(KMP_HW_LLC, KMP_HW_TILE);
430 // L2/Tile wasn't detected so just say L1
431 else
432 set_equivalent_type(KMP_HW_LLC, KMP_HW_L1);
433 }
434 #endif
435 else if (get_equivalent_type(KMP_HW_L1) != KMP_HW_UNKNOWN)
436 set_equivalent_type(KMP_HW_LLC, KMP_HW_L1);
437 // Fallback is to set last level cache to socket or core
438 if (get_equivalent_type(KMP_HW_LLC) == KMP_HW_UNKNOWN) {
439 if (get_equivalent_type(KMP_HW_SOCKET) != KMP_HW_UNKNOWN)
440 set_equivalent_type(KMP_HW_LLC, KMP_HW_SOCKET);
441 else if (get_equivalent_type(KMP_HW_CORE) != KMP_HW_UNKNOWN)
442 set_equivalent_type(KMP_HW_LLC, KMP_HW_CORE);
443 }
444 KMP_ASSERT(get_equivalent_type(KMP_HW_LLC) != KMP_HW_UNKNOWN);
445 }
446
447 // Gather the count of each topology layer and the ratio
_gather_enumeration_information()448 void kmp_topology_t::_gather_enumeration_information() {
449 int previous_id[KMP_HW_LAST];
450 int max[KMP_HW_LAST];
451
452 for (int i = 0; i < depth; ++i) {
453 previous_id[i] = kmp_hw_thread_t::UNKNOWN_ID;
454 max[i] = 0;
455 count[i] = 0;
456 ratio[i] = 0;
457 }
458 int core_level = get_level(KMP_HW_CORE);
459 for (int i = 0; i < num_hw_threads; ++i) {
460 kmp_hw_thread_t &hw_thread = hw_threads[i];
461 for (int layer = 0; layer < depth; ++layer) {
462 int id = hw_thread.ids[layer];
463 if (id != previous_id[layer]) {
464 // Add an additional increment to each count
465 for (int l = layer; l < depth; ++l)
466 count[l]++;
467 // Keep track of topology layer ratio statistics
468 max[layer]++;
469 for (int l = layer + 1; l < depth; ++l) {
470 if (max[l] > ratio[l])
471 ratio[l] = max[l];
472 max[l] = 1;
473 }
474 // Figure out the number of different core types
475 // and efficiencies for hybrid CPUs
476 if (__kmp_is_hybrid_cpu() && core_level >= 0 && layer <= core_level) {
477 if (hw_thread.attrs.is_core_eff_valid() &&
478 hw_thread.attrs.core_eff >= num_core_efficiencies) {
479 // Because efficiencies can range from 0 to max efficiency - 1,
480 // the number of efficiencies is max efficiency + 1
481 num_core_efficiencies = hw_thread.attrs.core_eff + 1;
482 }
483 if (hw_thread.attrs.is_core_type_valid()) {
484 bool found = false;
485 for (int j = 0; j < num_core_types; ++j) {
486 if (hw_thread.attrs.get_core_type() == core_types[j]) {
487 found = true;
488 break;
489 }
490 }
491 if (!found) {
492 KMP_ASSERT(num_core_types < KMP_HW_MAX_NUM_CORE_TYPES);
493 core_types[num_core_types++] = hw_thread.attrs.get_core_type();
494 }
495 }
496 }
497 break;
498 }
499 }
500 for (int layer = 0; layer < depth; ++layer) {
501 previous_id[layer] = hw_thread.ids[layer];
502 }
503 }
504 for (int layer = 0; layer < depth; ++layer) {
505 if (max[layer] > ratio[layer])
506 ratio[layer] = max[layer];
507 }
508 }
509
_get_ncores_with_attr(const kmp_hw_attr_t & attr,int above_level,bool find_all) const510 int kmp_topology_t::_get_ncores_with_attr(const kmp_hw_attr_t &attr,
511 int above_level,
512 bool find_all) const {
513 int current, current_max;
514 int previous_id[KMP_HW_LAST];
515 for (int i = 0; i < depth; ++i)
516 previous_id[i] = kmp_hw_thread_t::UNKNOWN_ID;
517 int core_level = get_level(KMP_HW_CORE);
518 if (find_all)
519 above_level = -1;
520 KMP_ASSERT(above_level < core_level);
521 current_max = 0;
522 current = 0;
523 for (int i = 0; i < num_hw_threads; ++i) {
524 kmp_hw_thread_t &hw_thread = hw_threads[i];
525 if (!find_all && hw_thread.ids[above_level] != previous_id[above_level]) {
526 if (current > current_max)
527 current_max = current;
528 current = hw_thread.attrs.contains(attr);
529 } else {
530 for (int level = above_level + 1; level <= core_level; ++level) {
531 if (hw_thread.ids[level] != previous_id[level]) {
532 if (hw_thread.attrs.contains(attr))
533 current++;
534 break;
535 }
536 }
537 }
538 for (int level = 0; level < depth; ++level)
539 previous_id[level] = hw_thread.ids[level];
540 }
541 if (current > current_max)
542 current_max = current;
543 return current_max;
544 }
545
546 // Find out if the topology is uniform
_discover_uniformity()547 void kmp_topology_t::_discover_uniformity() {
548 int num = 1;
549 for (int level = 0; level < depth; ++level)
550 num *= ratio[level];
551 flags.uniform = (num == count[depth - 1]);
552 }
553
554 // Set all the sub_ids for each hardware thread
_set_sub_ids()555 void kmp_topology_t::_set_sub_ids() {
556 int previous_id[KMP_HW_LAST];
557 int sub_id[KMP_HW_LAST];
558
559 for (int i = 0; i < depth; ++i) {
560 previous_id[i] = -1;
561 sub_id[i] = -1;
562 }
563 for (int i = 0; i < num_hw_threads; ++i) {
564 kmp_hw_thread_t &hw_thread = hw_threads[i];
565 // Setup the sub_id
566 for (int j = 0; j < depth; ++j) {
567 if (hw_thread.ids[j] != previous_id[j]) {
568 sub_id[j]++;
569 for (int k = j + 1; k < depth; ++k) {
570 sub_id[k] = 0;
571 }
572 break;
573 }
574 }
575 // Set previous_id
576 for (int j = 0; j < depth; ++j) {
577 previous_id[j] = hw_thread.ids[j];
578 }
579 // Set the sub_ids field
580 for (int j = 0; j < depth; ++j) {
581 hw_thread.sub_ids[j] = sub_id[j];
582 }
583 }
584 }
585
_set_globals()586 void kmp_topology_t::_set_globals() {
587 // Set nCoresPerPkg, nPackages, __kmp_nThreadsPerCore, __kmp_ncores
588 int core_level, thread_level, package_level;
589 package_level = get_level(KMP_HW_SOCKET);
590 #if KMP_GROUP_AFFINITY
591 if (package_level == -1)
592 package_level = get_level(KMP_HW_PROC_GROUP);
593 #endif
594 core_level = get_level(KMP_HW_CORE);
595 thread_level = get_level(KMP_HW_THREAD);
596
597 KMP_ASSERT(core_level != -1);
598 KMP_ASSERT(thread_level != -1);
599
600 __kmp_nThreadsPerCore = calculate_ratio(thread_level, core_level);
601 if (package_level != -1) {
602 nCoresPerPkg = calculate_ratio(core_level, package_level);
603 nPackages = get_count(package_level);
604 } else {
605 // assume one socket
606 nCoresPerPkg = get_count(core_level);
607 nPackages = 1;
608 }
609 #ifndef KMP_DFLT_NTH_CORES
610 __kmp_ncores = get_count(core_level);
611 #endif
612 }
613
allocate(int nproc,int ndepth,const kmp_hw_t * types)614 kmp_topology_t *kmp_topology_t::allocate(int nproc, int ndepth,
615 const kmp_hw_t *types) {
616 kmp_topology_t *retval;
617 // Allocate all data in one large allocation
618 size_t size = sizeof(kmp_topology_t) + sizeof(kmp_hw_thread_t) * nproc +
619 sizeof(int) * (size_t)KMP_HW_LAST * 3;
620 char *bytes = (char *)__kmp_allocate(size);
621 retval = (kmp_topology_t *)bytes;
622 if (nproc > 0) {
623 retval->hw_threads = (kmp_hw_thread_t *)(bytes + sizeof(kmp_topology_t));
624 } else {
625 retval->hw_threads = nullptr;
626 }
627 retval->num_hw_threads = nproc;
628 retval->depth = ndepth;
629 int *arr =
630 (int *)(bytes + sizeof(kmp_topology_t) + sizeof(kmp_hw_thread_t) * nproc);
631 retval->types = (kmp_hw_t *)arr;
632 retval->ratio = arr + (size_t)KMP_HW_LAST;
633 retval->count = arr + 2 * (size_t)KMP_HW_LAST;
634 retval->num_core_efficiencies = 0;
635 retval->num_core_types = 0;
636 retval->compact = 0;
637 for (int i = 0; i < KMP_HW_MAX_NUM_CORE_TYPES; ++i)
638 retval->core_types[i] = KMP_HW_CORE_TYPE_UNKNOWN;
639 KMP_FOREACH_HW_TYPE(type) { retval->equivalent[type] = KMP_HW_UNKNOWN; }
640 for (int i = 0; i < ndepth; ++i) {
641 retval->types[i] = types[i];
642 retval->equivalent[types[i]] = types[i];
643 }
644 return retval;
645 }
646
deallocate(kmp_topology_t * topology)647 void kmp_topology_t::deallocate(kmp_topology_t *topology) {
648 if (topology)
649 __kmp_free(topology);
650 }
651
check_ids() const652 bool kmp_topology_t::check_ids() const {
653 // Assume ids have been sorted
654 if (num_hw_threads == 0)
655 return true;
656 for (int i = 1; i < num_hw_threads; ++i) {
657 kmp_hw_thread_t ¤t_thread = hw_threads[i];
658 kmp_hw_thread_t &previous_thread = hw_threads[i - 1];
659 bool unique = false;
660 for (int j = 0; j < depth; ++j) {
661 if (previous_thread.ids[j] != current_thread.ids[j]) {
662 unique = true;
663 break;
664 }
665 }
666 if (unique)
667 continue;
668 return false;
669 }
670 return true;
671 }
672
dump() const673 void kmp_topology_t::dump() const {
674 printf("***********************\n");
675 printf("*** __kmp_topology: ***\n");
676 printf("***********************\n");
677 printf("* depth: %d\n", depth);
678
679 printf("* types: ");
680 for (int i = 0; i < depth; ++i)
681 printf("%15s ", __kmp_hw_get_keyword(types[i]));
682 printf("\n");
683
684 printf("* ratio: ");
685 for (int i = 0; i < depth; ++i) {
686 printf("%15d ", ratio[i]);
687 }
688 printf("\n");
689
690 printf("* count: ");
691 for (int i = 0; i < depth; ++i) {
692 printf("%15d ", count[i]);
693 }
694 printf("\n");
695
696 printf("* num_core_eff: %d\n", num_core_efficiencies);
697 printf("* num_core_types: %d\n", num_core_types);
698 printf("* core_types: ");
699 for (int i = 0; i < num_core_types; ++i)
700 printf("%3d ", core_types[i]);
701 printf("\n");
702
703 printf("* equivalent map:\n");
704 KMP_FOREACH_HW_TYPE(i) {
705 const char *key = __kmp_hw_get_keyword(i);
706 const char *value = __kmp_hw_get_keyword(equivalent[i]);
707 printf("%-15s -> %-15s\n", key, value);
708 }
709
710 printf("* uniform: %s\n", (is_uniform() ? "Yes" : "No"));
711
712 printf("* num_hw_threads: %d\n", num_hw_threads);
713 printf("* hw_threads:\n");
714 for (int i = 0; i < num_hw_threads; ++i) {
715 hw_threads[i].print();
716 }
717 printf("***********************\n");
718 }
719
print(const char * env_var) const720 void kmp_topology_t::print(const char *env_var) const {
721 kmp_str_buf_t buf;
722 int print_types_depth;
723 __kmp_str_buf_init(&buf);
724 kmp_hw_t print_types[KMP_HW_LAST + 2];
725
726 // Num Available Threads
727 if (num_hw_threads) {
728 KMP_INFORM(AvailableOSProc, env_var, num_hw_threads);
729 } else {
730 KMP_INFORM(AvailableOSProc, env_var, __kmp_xproc);
731 }
732
733 // Uniform or not
734 if (is_uniform()) {
735 KMP_INFORM(Uniform, env_var);
736 } else {
737 KMP_INFORM(NonUniform, env_var);
738 }
739
740 // Equivalent types
741 KMP_FOREACH_HW_TYPE(type) {
742 kmp_hw_t eq_type = equivalent[type];
743 if (eq_type != KMP_HW_UNKNOWN && eq_type != type) {
744 KMP_INFORM(AffEqualTopologyTypes, env_var,
745 __kmp_hw_get_catalog_string(type),
746 __kmp_hw_get_catalog_string(eq_type));
747 }
748 }
749
750 // Quick topology
751 KMP_ASSERT(depth > 0 && depth <= (int)KMP_HW_LAST);
752 // Create a print types array that always guarantees printing
753 // the core and thread level
754 print_types_depth = 0;
755 for (int level = 0; level < depth; ++level)
756 print_types[print_types_depth++] = types[level];
757 if (equivalent[KMP_HW_CORE] != KMP_HW_CORE) {
758 // Force in the core level for quick topology
759 if (print_types[print_types_depth - 1] == KMP_HW_THREAD) {
760 // Force core before thread e.g., 1 socket X 2 threads/socket
761 // becomes 1 socket X 1 core/socket X 2 threads/socket
762 print_types[print_types_depth - 1] = KMP_HW_CORE;
763 print_types[print_types_depth++] = KMP_HW_THREAD;
764 } else {
765 print_types[print_types_depth++] = KMP_HW_CORE;
766 }
767 }
768 // Always put threads at very end of quick topology
769 if (equivalent[KMP_HW_THREAD] != KMP_HW_THREAD)
770 print_types[print_types_depth++] = KMP_HW_THREAD;
771
772 __kmp_str_buf_clear(&buf);
773 kmp_hw_t numerator_type;
774 kmp_hw_t denominator_type = KMP_HW_UNKNOWN;
775 int core_level = get_level(KMP_HW_CORE);
776 int ncores = get_count(core_level);
777
778 for (int plevel = 0, level = 0; plevel < print_types_depth; ++plevel) {
779 int c;
780 bool plural;
781 numerator_type = print_types[plevel];
782 KMP_ASSERT_VALID_HW_TYPE(numerator_type);
783 if (equivalent[numerator_type] != numerator_type)
784 c = 1;
785 else
786 c = get_ratio(level++);
787 plural = (c > 1);
788 if (plevel == 0) {
789 __kmp_str_buf_print(&buf, "%d %s", c,
790 __kmp_hw_get_catalog_string(numerator_type, plural));
791 } else {
792 __kmp_str_buf_print(&buf, " x %d %s/%s", c,
793 __kmp_hw_get_catalog_string(numerator_type, plural),
794 __kmp_hw_get_catalog_string(denominator_type));
795 }
796 denominator_type = numerator_type;
797 }
798 KMP_INFORM(TopologyGeneric, env_var, buf.str, ncores);
799
800 // Hybrid topology information
801 if (__kmp_is_hybrid_cpu()) {
802 for (int i = 0; i < num_core_types; ++i) {
803 kmp_hw_core_type_t core_type = core_types[i];
804 kmp_hw_attr_t attr;
805 attr.clear();
806 attr.set_core_type(core_type);
807 int ncores = get_ncores_with_attr(attr);
808 if (ncores > 0) {
809 KMP_INFORM(TopologyHybrid, env_var, ncores,
810 __kmp_hw_get_core_type_string(core_type));
811 KMP_ASSERT(num_core_efficiencies <= KMP_HW_MAX_NUM_CORE_EFFS)
812 for (int eff = 0; eff < num_core_efficiencies; ++eff) {
813 attr.set_core_eff(eff);
814 int ncores_with_eff = get_ncores_with_attr(attr);
815 if (ncores_with_eff > 0) {
816 KMP_INFORM(TopologyHybridCoreEff, env_var, ncores_with_eff, eff);
817 }
818 }
819 }
820 }
821 }
822
823 if (num_hw_threads <= 0) {
824 __kmp_str_buf_free(&buf);
825 return;
826 }
827
828 // Full OS proc to hardware thread map
829 KMP_INFORM(OSProcToPhysicalThreadMap, env_var);
830 for (int i = 0; i < num_hw_threads; i++) {
831 __kmp_str_buf_clear(&buf);
832 for (int level = 0; level < depth; ++level) {
833 kmp_hw_t type = types[level];
834 __kmp_str_buf_print(&buf, "%s ", __kmp_hw_get_catalog_string(type));
835 __kmp_str_buf_print(&buf, "%d ", hw_threads[i].ids[level]);
836 }
837 if (__kmp_is_hybrid_cpu())
838 __kmp_str_buf_print(
839 &buf, "(%s)",
840 __kmp_hw_get_core_type_string(hw_threads[i].attrs.get_core_type()));
841 KMP_INFORM(OSProcMapToPack, env_var, hw_threads[i].os_id, buf.str);
842 }
843
844 __kmp_str_buf_free(&buf);
845 }
846
847 #if KMP_AFFINITY_SUPPORTED
set_granularity(kmp_affinity_t & affinity) const848 void kmp_topology_t::set_granularity(kmp_affinity_t &affinity) const {
849 const char *env_var = __kmp_get_affinity_env_var(affinity);
850 // If requested hybrid CPU attributes for granularity (either OMP_PLACES or
851 // KMP_AFFINITY), but none exist, then reset granularity and have below method
852 // select a granularity and warn user.
853 if (!__kmp_is_hybrid_cpu()) {
854 if (affinity.core_attr_gran.valid) {
855 // OMP_PLACES with cores:<attribute> but non-hybrid arch, use cores
856 // instead
857 KMP_AFF_WARNING(
858 affinity, AffIgnoringNonHybrid, env_var,
859 __kmp_hw_get_catalog_string(KMP_HW_CORE, /*plural=*/true));
860 affinity.gran = KMP_HW_CORE;
861 affinity.gran_levels = -1;
862 affinity.core_attr_gran = KMP_AFFINITY_ATTRS_UNKNOWN;
863 affinity.flags.core_types_gran = affinity.flags.core_effs_gran = 0;
864 } else if (affinity.flags.core_types_gran ||
865 affinity.flags.core_effs_gran) {
866 // OMP_PLACES=core_types|core_effs but non-hybrid, use cores instead
867 if (affinity.flags.omp_places) {
868 KMP_AFF_WARNING(
869 affinity, AffIgnoringNonHybrid, env_var,
870 __kmp_hw_get_catalog_string(KMP_HW_CORE, /*plural=*/true));
871 } else {
872 // KMP_AFFINITY=granularity=core_type|core_eff,...
873 KMP_AFF_WARNING(affinity, AffGranularityBad, env_var,
874 "Intel(R) Hybrid Technology core attribute",
875 __kmp_hw_get_catalog_string(KMP_HW_CORE));
876 }
877 affinity.gran = KMP_HW_CORE;
878 affinity.gran_levels = -1;
879 affinity.core_attr_gran = KMP_AFFINITY_ATTRS_UNKNOWN;
880 affinity.flags.core_types_gran = affinity.flags.core_effs_gran = 0;
881 }
882 }
883 // Set the number of affinity granularity levels
884 if (affinity.gran_levels < 0) {
885 kmp_hw_t gran_type = get_equivalent_type(affinity.gran);
886 // Check if user's granularity request is valid
887 if (gran_type == KMP_HW_UNKNOWN) {
888 // First try core, then thread, then package
889 kmp_hw_t gran_types[3] = {KMP_HW_CORE, KMP_HW_THREAD, KMP_HW_SOCKET};
890 for (auto g : gran_types) {
891 if (get_equivalent_type(g) != KMP_HW_UNKNOWN) {
892 gran_type = g;
893 break;
894 }
895 }
896 KMP_ASSERT(gran_type != KMP_HW_UNKNOWN);
897 // Warn user what granularity setting will be used instead
898 KMP_AFF_WARNING(affinity, AffGranularityBad, env_var,
899 __kmp_hw_get_catalog_string(affinity.gran),
900 __kmp_hw_get_catalog_string(gran_type));
901 affinity.gran = gran_type;
902 }
903 #if KMP_GROUP_AFFINITY
904 // If more than one processor group exists, and the level of
905 // granularity specified by the user is too coarse, then the
906 // granularity must be adjusted "down" to processor group affinity
907 // because threads can only exist within one processor group.
908 // For example, if a user sets granularity=socket and there are two
909 // processor groups that cover a socket, then the runtime must
910 // restrict the granularity down to the processor group level.
911 if (__kmp_num_proc_groups > 1) {
912 int gran_depth = get_level(gran_type);
913 int proc_group_depth = get_level(KMP_HW_PROC_GROUP);
914 if (gran_depth >= 0 && proc_group_depth >= 0 &&
915 gran_depth < proc_group_depth) {
916 KMP_AFF_WARNING(affinity, AffGranTooCoarseProcGroup, env_var,
917 __kmp_hw_get_catalog_string(affinity.gran));
918 affinity.gran = gran_type = KMP_HW_PROC_GROUP;
919 }
920 }
921 #endif
922 affinity.gran_levels = 0;
923 for (int i = depth - 1; i >= 0 && get_type(i) != gran_type; --i)
924 affinity.gran_levels++;
925 }
926 }
927 #endif
928
canonicalize()929 void kmp_topology_t::canonicalize() {
930 #if KMP_GROUP_AFFINITY
931 _insert_windows_proc_groups();
932 #endif
933 _remove_radix1_layers();
934 _gather_enumeration_information();
935 _discover_uniformity();
936 _set_sub_ids();
937 _set_globals();
938 _set_last_level_cache();
939
940 #if KMP_MIC_SUPPORTED
941 // Manually Add L2 = Tile equivalence
942 if (__kmp_mic_type == mic3) {
943 if (get_level(KMP_HW_L2) != -1)
944 set_equivalent_type(KMP_HW_TILE, KMP_HW_L2);
945 else if (get_level(KMP_HW_TILE) != -1)
946 set_equivalent_type(KMP_HW_L2, KMP_HW_TILE);
947 }
948 #endif
949
950 // Perform post canonicalization checking
951 KMP_ASSERT(depth > 0);
952 for (int level = 0; level < depth; ++level) {
953 // All counts, ratios, and types must be valid
954 KMP_ASSERT(count[level] > 0 && ratio[level] > 0);
955 KMP_ASSERT_VALID_HW_TYPE(types[level]);
956 // Detected types must point to themselves
957 KMP_ASSERT(equivalent[types[level]] == types[level]);
958 }
959 }
960
961 // Canonicalize an explicit packages X cores/pkg X threads/core topology
canonicalize(int npackages,int ncores_per_pkg,int nthreads_per_core,int ncores)962 void kmp_topology_t::canonicalize(int npackages, int ncores_per_pkg,
963 int nthreads_per_core, int ncores) {
964 int ndepth = 3;
965 depth = ndepth;
966 KMP_FOREACH_HW_TYPE(i) { equivalent[i] = KMP_HW_UNKNOWN; }
967 for (int level = 0; level < depth; ++level) {
968 count[level] = 0;
969 ratio[level] = 0;
970 }
971 count[0] = npackages;
972 count[1] = ncores;
973 count[2] = __kmp_xproc;
974 ratio[0] = npackages;
975 ratio[1] = ncores_per_pkg;
976 ratio[2] = nthreads_per_core;
977 equivalent[KMP_HW_SOCKET] = KMP_HW_SOCKET;
978 equivalent[KMP_HW_CORE] = KMP_HW_CORE;
979 equivalent[KMP_HW_THREAD] = KMP_HW_THREAD;
980 types[0] = KMP_HW_SOCKET;
981 types[1] = KMP_HW_CORE;
982 types[2] = KMP_HW_THREAD;
983 //__kmp_avail_proc = __kmp_xproc;
984 _discover_uniformity();
985 }
986
987 // Represents running sub IDs for a single core attribute where
988 // attribute values have SIZE possibilities.
989 template <size_t SIZE, typename IndexFunc> struct kmp_sub_ids_t {
990 int last_level; // last level in topology to consider for sub_ids
991 int sub_id[SIZE]; // The sub ID for a given attribute value
992 int prev_sub_id[KMP_HW_LAST];
993 IndexFunc indexer;
994
995 public:
kmp_sub_ids_tkmp_sub_ids_t996 kmp_sub_ids_t(int last_level) : last_level(last_level) {
997 KMP_ASSERT(last_level < KMP_HW_LAST);
998 for (size_t i = 0; i < SIZE; ++i)
999 sub_id[i] = -1;
1000 for (size_t i = 0; i < KMP_HW_LAST; ++i)
1001 prev_sub_id[i] = -1;
1002 }
updatekmp_sub_ids_t1003 void update(const kmp_hw_thread_t &hw_thread) {
1004 int idx = indexer(hw_thread);
1005 KMP_ASSERT(idx < (int)SIZE);
1006 for (int level = 0; level <= last_level; ++level) {
1007 if (hw_thread.sub_ids[level] != prev_sub_id[level]) {
1008 if (level < last_level)
1009 sub_id[idx] = -1;
1010 sub_id[idx]++;
1011 break;
1012 }
1013 }
1014 for (int level = 0; level <= last_level; ++level)
1015 prev_sub_id[level] = hw_thread.sub_ids[level];
1016 }
get_sub_idkmp_sub_ids_t1017 int get_sub_id(const kmp_hw_thread_t &hw_thread) const {
1018 return sub_id[indexer(hw_thread)];
1019 }
1020 };
1021
1022 #if KMP_AFFINITY_SUPPORTED
1023 static kmp_str_buf_t *
__kmp_hw_get_catalog_core_string(const kmp_hw_attr_t & attr,kmp_str_buf_t * buf,bool plural)1024 __kmp_hw_get_catalog_core_string(const kmp_hw_attr_t &attr, kmp_str_buf_t *buf,
1025 bool plural) {
1026 __kmp_str_buf_init(buf);
1027 if (attr.is_core_type_valid())
1028 __kmp_str_buf_print(buf, "%s %s",
1029 __kmp_hw_get_core_type_string(attr.get_core_type()),
1030 __kmp_hw_get_catalog_string(KMP_HW_CORE, plural));
1031 else
1032 __kmp_str_buf_print(buf, "%s eff=%d",
1033 __kmp_hw_get_catalog_string(KMP_HW_CORE, plural),
1034 attr.get_core_eff());
1035 return buf;
1036 }
1037
restrict_to_mask(const kmp_affin_mask_t * mask)1038 bool kmp_topology_t::restrict_to_mask(const kmp_affin_mask_t *mask) {
1039 // Apply the filter
1040 bool affected;
1041 int new_index = 0;
1042 for (int i = 0; i < num_hw_threads; ++i) {
1043 int os_id = hw_threads[i].os_id;
1044 if (KMP_CPU_ISSET(os_id, mask)) {
1045 if (i != new_index)
1046 hw_threads[new_index] = hw_threads[i];
1047 new_index++;
1048 } else {
1049 KMP_CPU_CLR(os_id, __kmp_affin_fullMask);
1050 __kmp_avail_proc--;
1051 }
1052 }
1053
1054 KMP_DEBUG_ASSERT(new_index <= num_hw_threads);
1055 affected = (num_hw_threads != new_index);
1056 num_hw_threads = new_index;
1057
1058 // Post hardware subset canonicalization
1059 if (affected) {
1060 _gather_enumeration_information();
1061 _discover_uniformity();
1062 _set_globals();
1063 _set_last_level_cache();
1064 #if KMP_OS_WINDOWS
1065 // Copy filtered full mask if topology has single processor group
1066 if (__kmp_num_proc_groups <= 1)
1067 #endif
1068 __kmp_affin_origMask->copy(__kmp_affin_fullMask);
1069 }
1070 return affected;
1071 }
1072
1073 // Apply the KMP_HW_SUBSET envirable to the topology
1074 // Returns true if KMP_HW_SUBSET filtered any processors
1075 // otherwise, returns false
filter_hw_subset()1076 bool kmp_topology_t::filter_hw_subset() {
1077 // If KMP_HW_SUBSET wasn't requested, then do nothing.
1078 if (!__kmp_hw_subset)
1079 return false;
1080
1081 // First, sort the KMP_HW_SUBSET items by the machine topology
1082 __kmp_hw_subset->sort();
1083
1084 // Check to see if KMP_HW_SUBSET is a valid subset of the detected topology
1085 bool using_core_types = false;
1086 bool using_core_effs = false;
1087 int hw_subset_depth = __kmp_hw_subset->get_depth();
1088 kmp_hw_t specified[KMP_HW_LAST];
1089 int *topology_levels = (int *)KMP_ALLOCA(sizeof(int) * hw_subset_depth);
1090 KMP_ASSERT(hw_subset_depth > 0);
1091 KMP_FOREACH_HW_TYPE(i) { specified[i] = KMP_HW_UNKNOWN; }
1092 int core_level = get_level(KMP_HW_CORE);
1093 for (int i = 0; i < hw_subset_depth; ++i) {
1094 int max_count;
1095 const kmp_hw_subset_t::item_t &item = __kmp_hw_subset->at(i);
1096 int num = item.num[0];
1097 int offset = item.offset[0];
1098 kmp_hw_t type = item.type;
1099 kmp_hw_t equivalent_type = equivalent[type];
1100 int level = get_level(type);
1101 topology_levels[i] = level;
1102
1103 // Check to see if current layer is in detected machine topology
1104 if (equivalent_type != KMP_HW_UNKNOWN) {
1105 __kmp_hw_subset->at(i).type = equivalent_type;
1106 } else {
1107 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetNotExistGeneric,
1108 __kmp_hw_get_catalog_string(type));
1109 return false;
1110 }
1111
1112 // Check to see if current layer has already been
1113 // specified either directly or through an equivalent type
1114 if (specified[equivalent_type] != KMP_HW_UNKNOWN) {
1115 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetEqvLayers,
1116 __kmp_hw_get_catalog_string(type),
1117 __kmp_hw_get_catalog_string(specified[equivalent_type]));
1118 return false;
1119 }
1120 specified[equivalent_type] = type;
1121
1122 // Check to see if each layer's num & offset parameters are valid
1123 max_count = get_ratio(level);
1124 if (max_count < 0 ||
1125 (num != kmp_hw_subset_t::USE_ALL && num + offset > max_count)) {
1126 bool plural = (num > 1);
1127 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetManyGeneric,
1128 __kmp_hw_get_catalog_string(type, plural));
1129 return false;
1130 }
1131
1132 // Check to see if core attributes are consistent
1133 if (core_level == level) {
1134 // Determine which core attributes are specified
1135 for (int j = 0; j < item.num_attrs; ++j) {
1136 if (item.attr[j].is_core_type_valid())
1137 using_core_types = true;
1138 if (item.attr[j].is_core_eff_valid())
1139 using_core_effs = true;
1140 }
1141
1142 // Check if using a single core attribute on non-hybrid arch.
1143 // Do not ignore all of KMP_HW_SUBSET, just ignore the attribute.
1144 //
1145 // Check if using multiple core attributes on non-hyrbid arch.
1146 // Ignore all of KMP_HW_SUBSET if this is the case.
1147 if ((using_core_effs || using_core_types) && !__kmp_is_hybrid_cpu()) {
1148 if (item.num_attrs == 1) {
1149 if (using_core_effs) {
1150 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetIgnoringAttr,
1151 "efficiency");
1152 } else {
1153 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetIgnoringAttr,
1154 "core_type");
1155 }
1156 using_core_effs = false;
1157 using_core_types = false;
1158 } else {
1159 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetAttrsNonHybrid);
1160 return false;
1161 }
1162 }
1163
1164 // Check if using both core types and core efficiencies together
1165 if (using_core_types && using_core_effs) {
1166 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetIncompat, "core_type",
1167 "efficiency");
1168 return false;
1169 }
1170
1171 // Check that core efficiency values are valid
1172 if (using_core_effs) {
1173 for (int j = 0; j < item.num_attrs; ++j) {
1174 if (item.attr[j].is_core_eff_valid()) {
1175 int core_eff = item.attr[j].get_core_eff();
1176 if (core_eff < 0 || core_eff >= num_core_efficiencies) {
1177 kmp_str_buf_t buf;
1178 __kmp_str_buf_init(&buf);
1179 __kmp_str_buf_print(&buf, "%d", item.attr[j].get_core_eff());
1180 __kmp_msg(kmp_ms_warning,
1181 KMP_MSG(AffHWSubsetAttrInvalid, "efficiency", buf.str),
1182 KMP_HNT(ValidValuesRange, 0, num_core_efficiencies - 1),
1183 __kmp_msg_null);
1184 __kmp_str_buf_free(&buf);
1185 return false;
1186 }
1187 }
1188 }
1189 }
1190
1191 // Check that the number of requested cores with attributes is valid
1192 if (using_core_types || using_core_effs) {
1193 for (int j = 0; j < item.num_attrs; ++j) {
1194 int num = item.num[j];
1195 int offset = item.offset[j];
1196 int level_above = core_level - 1;
1197 if (level_above >= 0) {
1198 max_count = get_ncores_with_attr_per(item.attr[j], level_above);
1199 if (max_count <= 0 ||
1200 (num != kmp_hw_subset_t::USE_ALL && num + offset > max_count)) {
1201 kmp_str_buf_t buf;
1202 __kmp_hw_get_catalog_core_string(item.attr[j], &buf, num > 0);
1203 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetManyGeneric, buf.str);
1204 __kmp_str_buf_free(&buf);
1205 return false;
1206 }
1207 }
1208 }
1209 }
1210
1211 if ((using_core_types || using_core_effs) && item.num_attrs > 1) {
1212 for (int j = 0; j < item.num_attrs; ++j) {
1213 // Ambiguous use of specific core attribute + generic core
1214 // e.g., 4c & 3c:intel_core or 4c & 3c:eff1
1215 if (!item.attr[j]) {
1216 kmp_hw_attr_t other_attr;
1217 for (int k = 0; k < item.num_attrs; ++k) {
1218 if (item.attr[k] != item.attr[j]) {
1219 other_attr = item.attr[k];
1220 break;
1221 }
1222 }
1223 kmp_str_buf_t buf;
1224 __kmp_hw_get_catalog_core_string(other_attr, &buf, item.num[j] > 0);
1225 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetIncompat,
1226 __kmp_hw_get_catalog_string(KMP_HW_CORE), buf.str);
1227 __kmp_str_buf_free(&buf);
1228 return false;
1229 }
1230 // Allow specifying a specific core type or core eff exactly once
1231 for (int k = 0; k < j; ++k) {
1232 if (!item.attr[j] || !item.attr[k])
1233 continue;
1234 if (item.attr[k] == item.attr[j]) {
1235 kmp_str_buf_t buf;
1236 __kmp_hw_get_catalog_core_string(item.attr[j], &buf,
1237 item.num[j] > 0);
1238 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetAttrRepeat, buf.str);
1239 __kmp_str_buf_free(&buf);
1240 return false;
1241 }
1242 }
1243 }
1244 }
1245 }
1246 }
1247
1248 struct core_type_indexer {
1249 int operator()(const kmp_hw_thread_t &t) const {
1250 switch (t.attrs.get_core_type()) {
1251 case KMP_HW_CORE_TYPE_UNKNOWN:
1252 case KMP_HW_MAX_NUM_CORE_TYPES:
1253 return 0;
1254 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1255 case KMP_HW_CORE_TYPE_ATOM:
1256 return 1;
1257 case KMP_HW_CORE_TYPE_CORE:
1258 return 2;
1259 #endif
1260 }
1261 KMP_ASSERT2(false, "Unhandled kmp_hw_thread_t enumeration");
1262 KMP_BUILTIN_UNREACHABLE;
1263 }
1264 };
1265 struct core_eff_indexer {
1266 int operator()(const kmp_hw_thread_t &t) const {
1267 return t.attrs.get_core_eff();
1268 }
1269 };
1270
1271 kmp_sub_ids_t<KMP_HW_MAX_NUM_CORE_TYPES, core_type_indexer> core_type_sub_ids(
1272 core_level);
1273 kmp_sub_ids_t<KMP_HW_MAX_NUM_CORE_EFFS, core_eff_indexer> core_eff_sub_ids(
1274 core_level);
1275
1276 // Determine which hardware threads should be filtered.
1277 int num_filtered = 0;
1278 kmp_affin_mask_t *filtered_mask;
1279 KMP_CPU_ALLOC(filtered_mask);
1280 KMP_CPU_COPY(filtered_mask, __kmp_affin_fullMask);
1281 for (int i = 0; i < num_hw_threads; ++i) {
1282 kmp_hw_thread_t &hw_thread = hw_threads[i];
1283 // Update type_sub_id
1284 if (using_core_types)
1285 core_type_sub_ids.update(hw_thread);
1286 if (using_core_effs)
1287 core_eff_sub_ids.update(hw_thread);
1288
1289 // Check to see if this hardware thread should be filtered
1290 bool should_be_filtered = false;
1291 for (int hw_subset_index = 0; hw_subset_index < hw_subset_depth;
1292 ++hw_subset_index) {
1293 const auto &hw_subset_item = __kmp_hw_subset->at(hw_subset_index);
1294 int level = topology_levels[hw_subset_index];
1295 if (level == -1)
1296 continue;
1297 if ((using_core_effs || using_core_types) && level == core_level) {
1298 // Look for the core attribute in KMP_HW_SUBSET which corresponds
1299 // to this hardware thread's core attribute. Use this num,offset plus
1300 // the running sub_id for the particular core attribute of this hardware
1301 // thread to determine if the hardware thread should be filtered or not.
1302 int attr_idx;
1303 kmp_hw_core_type_t core_type = hw_thread.attrs.get_core_type();
1304 int core_eff = hw_thread.attrs.get_core_eff();
1305 for (attr_idx = 0; attr_idx < hw_subset_item.num_attrs; ++attr_idx) {
1306 if (using_core_types &&
1307 hw_subset_item.attr[attr_idx].get_core_type() == core_type)
1308 break;
1309 if (using_core_effs &&
1310 hw_subset_item.attr[attr_idx].get_core_eff() == core_eff)
1311 break;
1312 }
1313 // This core attribute isn't in the KMP_HW_SUBSET so always filter it.
1314 if (attr_idx == hw_subset_item.num_attrs) {
1315 should_be_filtered = true;
1316 break;
1317 }
1318 int sub_id;
1319 int num = hw_subset_item.num[attr_idx];
1320 int offset = hw_subset_item.offset[attr_idx];
1321 if (using_core_types)
1322 sub_id = core_type_sub_ids.get_sub_id(hw_thread);
1323 else
1324 sub_id = core_eff_sub_ids.get_sub_id(hw_thread);
1325 if (sub_id < offset ||
1326 (num != kmp_hw_subset_t::USE_ALL && sub_id >= offset + num)) {
1327 should_be_filtered = true;
1328 break;
1329 }
1330 } else {
1331 int num = hw_subset_item.num[0];
1332 int offset = hw_subset_item.offset[0];
1333 if (hw_thread.sub_ids[level] < offset ||
1334 (num != kmp_hw_subset_t::USE_ALL &&
1335 hw_thread.sub_ids[level] >= offset + num)) {
1336 should_be_filtered = true;
1337 break;
1338 }
1339 }
1340 }
1341 // Collect filtering information
1342 if (should_be_filtered) {
1343 KMP_CPU_CLR(hw_thread.os_id, filtered_mask);
1344 num_filtered++;
1345 }
1346 }
1347
1348 // One last check that we shouldn't allow filtering entire machine
1349 if (num_filtered == num_hw_threads) {
1350 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetAllFiltered);
1351 return false;
1352 }
1353
1354 // Apply the filter
1355 restrict_to_mask(filtered_mask);
1356 return true;
1357 }
1358
is_close(int hwt1,int hwt2,const kmp_affinity_t & stgs) const1359 bool kmp_topology_t::is_close(int hwt1, int hwt2,
1360 const kmp_affinity_t &stgs) const {
1361 int hw_level = stgs.gran_levels;
1362 if (hw_level >= depth)
1363 return true;
1364 bool retval = true;
1365 const kmp_hw_thread_t &t1 = hw_threads[hwt1];
1366 const kmp_hw_thread_t &t2 = hw_threads[hwt2];
1367 if (stgs.flags.core_types_gran)
1368 return t1.attrs.get_core_type() == t2.attrs.get_core_type();
1369 if (stgs.flags.core_effs_gran)
1370 return t1.attrs.get_core_eff() == t2.attrs.get_core_eff();
1371 for (int i = 0; i < (depth - hw_level); ++i) {
1372 if (t1.ids[i] != t2.ids[i])
1373 return false;
1374 }
1375 return retval;
1376 }
1377
1378 ////////////////////////////////////////////////////////////////////////////////
1379
1380 bool KMPAffinity::picked_api = false;
1381
operator new(size_t n)1382 void *KMPAffinity::Mask::operator new(size_t n) { return __kmp_allocate(n); }
operator new[](size_t n)1383 void *KMPAffinity::Mask::operator new[](size_t n) { return __kmp_allocate(n); }
operator delete(void * p)1384 void KMPAffinity::Mask::operator delete(void *p) { __kmp_free(p); }
operator delete[](void * p)1385 void KMPAffinity::Mask::operator delete[](void *p) { __kmp_free(p); }
operator new(size_t n)1386 void *KMPAffinity::operator new(size_t n) { return __kmp_allocate(n); }
operator delete(void * p)1387 void KMPAffinity::operator delete(void *p) { __kmp_free(p); }
1388
pick_api()1389 void KMPAffinity::pick_api() {
1390 KMPAffinity *affinity_dispatch;
1391 if (picked_api)
1392 return;
1393 #if KMP_USE_HWLOC
1394 // Only use Hwloc if affinity isn't explicitly disabled and
1395 // user requests Hwloc topology method
1396 if (__kmp_affinity_top_method == affinity_top_method_hwloc &&
1397 __kmp_affinity.type != affinity_disabled) {
1398 affinity_dispatch = new KMPHwlocAffinity();
1399 } else
1400 #endif
1401 {
1402 affinity_dispatch = new KMPNativeAffinity();
1403 }
1404 __kmp_affinity_dispatch = affinity_dispatch;
1405 picked_api = true;
1406 }
1407
destroy_api()1408 void KMPAffinity::destroy_api() {
1409 if (__kmp_affinity_dispatch != NULL) {
1410 delete __kmp_affinity_dispatch;
1411 __kmp_affinity_dispatch = NULL;
1412 picked_api = false;
1413 }
1414 }
1415
1416 #define KMP_ADVANCE_SCAN(scan) \
1417 while (*scan != '\0') { \
1418 scan++; \
1419 }
1420
1421 // Print the affinity mask to the character array in a pretty format.
1422 // The format is a comma separated list of non-negative integers or integer
1423 // ranges: e.g., 1,2,3-5,7,9-15
1424 // The format can also be the string "{<empty>}" if no bits are set in mask
__kmp_affinity_print_mask(char * buf,int buf_len,kmp_affin_mask_t * mask)1425 char *__kmp_affinity_print_mask(char *buf, int buf_len,
1426 kmp_affin_mask_t *mask) {
1427 int start = 0, finish = 0, previous = 0;
1428 bool first_range;
1429 KMP_ASSERT(buf);
1430 KMP_ASSERT(buf_len >= 40);
1431 KMP_ASSERT(mask);
1432 char *scan = buf;
1433 char *end = buf + buf_len - 1;
1434
1435 // Check for empty set.
1436 if (mask->begin() == mask->end()) {
1437 KMP_SNPRINTF(scan, end - scan + 1, "{<empty>}");
1438 KMP_ADVANCE_SCAN(scan);
1439 KMP_ASSERT(scan <= end);
1440 return buf;
1441 }
1442
1443 first_range = true;
1444 start = mask->begin();
1445 while (1) {
1446 // Find next range
1447 // [start, previous] is inclusive range of contiguous bits in mask
1448 for (finish = mask->next(start), previous = start;
1449 finish == previous + 1 && finish != mask->end();
1450 finish = mask->next(finish)) {
1451 previous = finish;
1452 }
1453
1454 // The first range does not need a comma printed before it, but the rest
1455 // of the ranges do need a comma beforehand
1456 if (!first_range) {
1457 KMP_SNPRINTF(scan, end - scan + 1, "%s", ",");
1458 KMP_ADVANCE_SCAN(scan);
1459 } else {
1460 first_range = false;
1461 }
1462 // Range with three or more contiguous bits in the affinity mask
1463 if (previous - start > 1) {
1464 KMP_SNPRINTF(scan, end - scan + 1, "%u-%u", start, previous);
1465 } else {
1466 // Range with one or two contiguous bits in the affinity mask
1467 KMP_SNPRINTF(scan, end - scan + 1, "%u", start);
1468 KMP_ADVANCE_SCAN(scan);
1469 if (previous - start > 0) {
1470 KMP_SNPRINTF(scan, end - scan + 1, ",%u", previous);
1471 }
1472 }
1473 KMP_ADVANCE_SCAN(scan);
1474 // Start over with new start point
1475 start = finish;
1476 if (start == mask->end())
1477 break;
1478 // Check for overflow
1479 if (end - scan < 2)
1480 break;
1481 }
1482
1483 // Check for overflow
1484 KMP_ASSERT(scan <= end);
1485 return buf;
1486 }
1487 #undef KMP_ADVANCE_SCAN
1488
1489 // Print the affinity mask to the string buffer object in a pretty format
1490 // The format is a comma separated list of non-negative integers or integer
1491 // ranges: e.g., 1,2,3-5,7,9-15
1492 // The format can also be the string "{<empty>}" if no bits are set in mask
__kmp_affinity_str_buf_mask(kmp_str_buf_t * buf,kmp_affin_mask_t * mask)1493 kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
1494 kmp_affin_mask_t *mask) {
1495 int start = 0, finish = 0, previous = 0;
1496 bool first_range;
1497 KMP_ASSERT(buf);
1498 KMP_ASSERT(mask);
1499
1500 __kmp_str_buf_clear(buf);
1501
1502 // Check for empty set.
1503 if (mask->begin() == mask->end()) {
1504 __kmp_str_buf_print(buf, "%s", "{<empty>}");
1505 return buf;
1506 }
1507
1508 first_range = true;
1509 start = mask->begin();
1510 while (1) {
1511 // Find next range
1512 // [start, previous] is inclusive range of contiguous bits in mask
1513 for (finish = mask->next(start), previous = start;
1514 finish == previous + 1 && finish != mask->end();
1515 finish = mask->next(finish)) {
1516 previous = finish;
1517 }
1518
1519 // The first range does not need a comma printed before it, but the rest
1520 // of the ranges do need a comma beforehand
1521 if (!first_range) {
1522 __kmp_str_buf_print(buf, "%s", ",");
1523 } else {
1524 first_range = false;
1525 }
1526 // Range with three or more contiguous bits in the affinity mask
1527 if (previous - start > 1) {
1528 __kmp_str_buf_print(buf, "%u-%u", start, previous);
1529 } else {
1530 // Range with one or two contiguous bits in the affinity mask
1531 __kmp_str_buf_print(buf, "%u", start);
1532 if (previous - start > 0) {
1533 __kmp_str_buf_print(buf, ",%u", previous);
1534 }
1535 }
1536 // Start over with new start point
1537 start = finish;
1538 if (start == mask->end())
1539 break;
1540 }
1541 return buf;
1542 }
1543
1544 // Return (possibly empty) affinity mask representing the offline CPUs
1545 // Caller must free the mask
__kmp_affinity_get_offline_cpus()1546 kmp_affin_mask_t *__kmp_affinity_get_offline_cpus() {
1547 kmp_affin_mask_t *offline;
1548 KMP_CPU_ALLOC(offline);
1549 KMP_CPU_ZERO(offline);
1550 #if KMP_OS_LINUX
1551 int n, begin_cpu, end_cpu;
1552 kmp_safe_raii_file_t offline_file;
1553 auto skip_ws = [](FILE *f) {
1554 int c;
1555 do {
1556 c = fgetc(f);
1557 } while (isspace(c));
1558 if (c != EOF)
1559 ungetc(c, f);
1560 };
1561 // File contains CSV of integer ranges representing the offline CPUs
1562 // e.g., 1,2,4-7,9,11-15
1563 int status = offline_file.try_open("/sys/devices/system/cpu/offline", "r");
1564 if (status != 0)
1565 return offline;
1566 while (!feof(offline_file)) {
1567 skip_ws(offline_file);
1568 n = fscanf(offline_file, "%d", &begin_cpu);
1569 if (n != 1)
1570 break;
1571 skip_ws(offline_file);
1572 int c = fgetc(offline_file);
1573 if (c == EOF || c == ',') {
1574 // Just single CPU
1575 end_cpu = begin_cpu;
1576 } else if (c == '-') {
1577 // Range of CPUs
1578 skip_ws(offline_file);
1579 n = fscanf(offline_file, "%d", &end_cpu);
1580 if (n != 1)
1581 break;
1582 skip_ws(offline_file);
1583 c = fgetc(offline_file); // skip ','
1584 } else {
1585 // Syntax problem
1586 break;
1587 }
1588 // Ensure a valid range of CPUs
1589 if (begin_cpu < 0 || begin_cpu >= __kmp_xproc || end_cpu < 0 ||
1590 end_cpu >= __kmp_xproc || begin_cpu > end_cpu) {
1591 continue;
1592 }
1593 // Insert [begin_cpu, end_cpu] into offline mask
1594 for (int cpu = begin_cpu; cpu <= end_cpu; ++cpu) {
1595 KMP_CPU_SET(cpu, offline);
1596 }
1597 }
1598 #endif
1599 return offline;
1600 }
1601
1602 // Return the number of available procs
__kmp_affinity_entire_machine_mask(kmp_affin_mask_t * mask)1603 int __kmp_affinity_entire_machine_mask(kmp_affin_mask_t *mask) {
1604 int avail_proc = 0;
1605 KMP_CPU_ZERO(mask);
1606
1607 #if KMP_GROUP_AFFINITY
1608
1609 if (__kmp_num_proc_groups > 1) {
1610 int group;
1611 KMP_DEBUG_ASSERT(__kmp_GetActiveProcessorCount != NULL);
1612 for (group = 0; group < __kmp_num_proc_groups; group++) {
1613 int i;
1614 int num = __kmp_GetActiveProcessorCount(group);
1615 for (i = 0; i < num; i++) {
1616 KMP_CPU_SET(i + group * (CHAR_BIT * sizeof(DWORD_PTR)), mask);
1617 avail_proc++;
1618 }
1619 }
1620 } else
1621
1622 #endif /* KMP_GROUP_AFFINITY */
1623
1624 {
1625 int proc;
1626 kmp_affin_mask_t *offline_cpus = __kmp_affinity_get_offline_cpus();
1627 for (proc = 0; proc < __kmp_xproc; proc++) {
1628 // Skip offline CPUs
1629 if (KMP_CPU_ISSET(proc, offline_cpus))
1630 continue;
1631 KMP_CPU_SET(proc, mask);
1632 avail_proc++;
1633 }
1634 KMP_CPU_FREE(offline_cpus);
1635 }
1636
1637 return avail_proc;
1638 }
1639
1640 // All of the __kmp_affinity_create_*_map() routines should allocate the
1641 // internal topology object and set the layer ids for it. Each routine
1642 // returns a boolean on whether it was successful at doing so.
1643 kmp_affin_mask_t *__kmp_affin_fullMask = NULL;
1644 // Original mask is a subset of full mask in multiple processor groups topology
1645 kmp_affin_mask_t *__kmp_affin_origMask = NULL;
1646
1647 #if KMP_USE_HWLOC
__kmp_hwloc_is_cache_type(hwloc_obj_t obj)1648 static inline bool __kmp_hwloc_is_cache_type(hwloc_obj_t obj) {
1649 #if HWLOC_API_VERSION >= 0x00020000
1650 return hwloc_obj_type_is_cache(obj->type);
1651 #else
1652 return obj->type == HWLOC_OBJ_CACHE;
1653 #endif
1654 }
1655
1656 // Returns KMP_HW_* type derived from HWLOC_* type
__kmp_hwloc_type_2_topology_type(hwloc_obj_t obj)1657 static inline kmp_hw_t __kmp_hwloc_type_2_topology_type(hwloc_obj_t obj) {
1658
1659 if (__kmp_hwloc_is_cache_type(obj)) {
1660 if (obj->attr->cache.type == HWLOC_OBJ_CACHE_INSTRUCTION)
1661 return KMP_HW_UNKNOWN;
1662 switch (obj->attr->cache.depth) {
1663 case 1:
1664 return KMP_HW_L1;
1665 case 2:
1666 #if KMP_MIC_SUPPORTED
1667 if (__kmp_mic_type == mic3) {
1668 return KMP_HW_TILE;
1669 }
1670 #endif
1671 return KMP_HW_L2;
1672 case 3:
1673 return KMP_HW_L3;
1674 }
1675 return KMP_HW_UNKNOWN;
1676 }
1677
1678 switch (obj->type) {
1679 case HWLOC_OBJ_PACKAGE:
1680 return KMP_HW_SOCKET;
1681 case HWLOC_OBJ_NUMANODE:
1682 return KMP_HW_NUMA;
1683 case HWLOC_OBJ_CORE:
1684 return KMP_HW_CORE;
1685 case HWLOC_OBJ_PU:
1686 return KMP_HW_THREAD;
1687 case HWLOC_OBJ_GROUP:
1688 #if HWLOC_API_VERSION >= 0x00020000
1689 if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_DIE)
1690 return KMP_HW_DIE;
1691 else if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_TILE)
1692 return KMP_HW_TILE;
1693 else if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_MODULE)
1694 return KMP_HW_MODULE;
1695 else if (obj->attr->group.kind == HWLOC_GROUP_KIND_WINDOWS_PROCESSOR_GROUP)
1696 return KMP_HW_PROC_GROUP;
1697 #endif
1698 return KMP_HW_UNKNOWN;
1699 #if HWLOC_API_VERSION >= 0x00020100
1700 case HWLOC_OBJ_DIE:
1701 return KMP_HW_DIE;
1702 #endif
1703 }
1704 return KMP_HW_UNKNOWN;
1705 }
1706
1707 // Returns the number of objects of type 'type' below 'obj' within the topology
1708 // tree structure. e.g., if obj is a HWLOC_OBJ_PACKAGE object, and type is
1709 // HWLOC_OBJ_PU, then this will return the number of PU's under the SOCKET
1710 // object.
__kmp_hwloc_get_nobjs_under_obj(hwloc_obj_t obj,hwloc_obj_type_t type)1711 static int __kmp_hwloc_get_nobjs_under_obj(hwloc_obj_t obj,
1712 hwloc_obj_type_t type) {
1713 int retval = 0;
1714 hwloc_obj_t first;
1715 for (first = hwloc_get_obj_below_by_type(__kmp_hwloc_topology, obj->type,
1716 obj->logical_index, type, 0);
1717 first != NULL && hwloc_get_ancestor_obj_by_type(__kmp_hwloc_topology,
1718 obj->type, first) == obj;
1719 first = hwloc_get_next_obj_by_type(__kmp_hwloc_topology, first->type,
1720 first)) {
1721 ++retval;
1722 }
1723 return retval;
1724 }
1725
1726 // This gets the sub_id for a lower object under a higher object in the
1727 // topology tree
__kmp_hwloc_get_sub_id(hwloc_topology_t t,hwloc_obj_t higher,hwloc_obj_t lower)1728 static int __kmp_hwloc_get_sub_id(hwloc_topology_t t, hwloc_obj_t higher,
1729 hwloc_obj_t lower) {
1730 hwloc_obj_t obj;
1731 hwloc_obj_type_t ltype = lower->type;
1732 int lindex = lower->logical_index - 1;
1733 int sub_id = 0;
1734 // Get the previous lower object
1735 obj = hwloc_get_obj_by_type(t, ltype, lindex);
1736 while (obj && lindex >= 0 &&
1737 hwloc_bitmap_isincluded(obj->cpuset, higher->cpuset)) {
1738 if (obj->userdata) {
1739 sub_id = (int)(RCAST(kmp_intptr_t, obj->userdata));
1740 break;
1741 }
1742 sub_id++;
1743 lindex--;
1744 obj = hwloc_get_obj_by_type(t, ltype, lindex);
1745 }
1746 // store sub_id + 1 so that 0 is differed from NULL
1747 lower->userdata = RCAST(void *, sub_id + 1);
1748 return sub_id;
1749 }
1750
__kmp_affinity_create_hwloc_map(kmp_i18n_id_t * const msg_id)1751 static bool __kmp_affinity_create_hwloc_map(kmp_i18n_id_t *const msg_id) {
1752 kmp_hw_t type;
1753 int hw_thread_index, sub_id;
1754 int depth;
1755 hwloc_obj_t pu, obj, root, prev;
1756 kmp_hw_t types[KMP_HW_LAST];
1757 hwloc_obj_type_t hwloc_types[KMP_HW_LAST];
1758
1759 hwloc_topology_t tp = __kmp_hwloc_topology;
1760 *msg_id = kmp_i18n_null;
1761 if (__kmp_affinity.flags.verbose) {
1762 KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
1763 }
1764
1765 if (!KMP_AFFINITY_CAPABLE()) {
1766 // Hack to try and infer the machine topology using only the data
1767 // available from hwloc on the current thread, and __kmp_xproc.
1768 KMP_ASSERT(__kmp_affinity.type == affinity_none);
1769 // hwloc only guarantees existance of PU object, so check PACKAGE and CORE
1770 hwloc_obj_t o = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PACKAGE, 0);
1771 if (o != NULL)
1772 nCoresPerPkg = __kmp_hwloc_get_nobjs_under_obj(o, HWLOC_OBJ_CORE);
1773 else
1774 nCoresPerPkg = 1; // no PACKAGE found
1775 o = hwloc_get_obj_by_type(tp, HWLOC_OBJ_CORE, 0);
1776 if (o != NULL)
1777 __kmp_nThreadsPerCore = __kmp_hwloc_get_nobjs_under_obj(o, HWLOC_OBJ_PU);
1778 else
1779 __kmp_nThreadsPerCore = 1; // no CORE found
1780 __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
1781 if (nCoresPerPkg == 0)
1782 nCoresPerPkg = 1; // to prevent possible division by 0
1783 nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
1784 return true;
1785 }
1786
1787 #if HWLOC_API_VERSION >= 0x00020400
1788 // Handle multiple types of cores if they exist on the system
1789 int nr_cpu_kinds = hwloc_cpukinds_get_nr(tp, 0);
1790
1791 typedef struct kmp_hwloc_cpukinds_info_t {
1792 int efficiency;
1793 kmp_hw_core_type_t core_type;
1794 hwloc_bitmap_t mask;
1795 } kmp_hwloc_cpukinds_info_t;
1796 kmp_hwloc_cpukinds_info_t *cpukinds = nullptr;
1797
1798 if (nr_cpu_kinds > 0) {
1799 unsigned nr_infos;
1800 struct hwloc_info_s *infos;
1801 cpukinds = (kmp_hwloc_cpukinds_info_t *)__kmp_allocate(
1802 sizeof(kmp_hwloc_cpukinds_info_t) * nr_cpu_kinds);
1803 for (unsigned idx = 0; idx < (unsigned)nr_cpu_kinds; ++idx) {
1804 cpukinds[idx].efficiency = -1;
1805 cpukinds[idx].core_type = KMP_HW_CORE_TYPE_UNKNOWN;
1806 cpukinds[idx].mask = hwloc_bitmap_alloc();
1807 if (hwloc_cpukinds_get_info(tp, idx, cpukinds[idx].mask,
1808 &cpukinds[idx].efficiency, &nr_infos, &infos,
1809 0) == 0) {
1810 for (unsigned i = 0; i < nr_infos; ++i) {
1811 if (__kmp_str_match("CoreType", 8, infos[i].name)) {
1812 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1813 if (__kmp_str_match("IntelAtom", 9, infos[i].value)) {
1814 cpukinds[idx].core_type = KMP_HW_CORE_TYPE_ATOM;
1815 break;
1816 } else if (__kmp_str_match("IntelCore", 9, infos[i].value)) {
1817 cpukinds[idx].core_type = KMP_HW_CORE_TYPE_CORE;
1818 break;
1819 }
1820 #endif
1821 }
1822 }
1823 }
1824 }
1825 }
1826 #endif
1827
1828 root = hwloc_get_root_obj(tp);
1829
1830 // Figure out the depth and types in the topology
1831 depth = 0;
1832 pu = hwloc_get_pu_obj_by_os_index(tp, __kmp_affin_fullMask->begin());
1833 KMP_ASSERT(pu);
1834 obj = pu;
1835 types[depth] = KMP_HW_THREAD;
1836 hwloc_types[depth] = obj->type;
1837 depth++;
1838 while (obj != root && obj != NULL) {
1839 obj = obj->parent;
1840 #if HWLOC_API_VERSION >= 0x00020000
1841 if (obj->memory_arity) {
1842 hwloc_obj_t memory;
1843 for (memory = obj->memory_first_child; memory;
1844 memory = hwloc_get_next_child(tp, obj, memory)) {
1845 if (memory->type == HWLOC_OBJ_NUMANODE)
1846 break;
1847 }
1848 if (memory && memory->type == HWLOC_OBJ_NUMANODE) {
1849 types[depth] = KMP_HW_NUMA;
1850 hwloc_types[depth] = memory->type;
1851 depth++;
1852 }
1853 }
1854 #endif
1855 type = __kmp_hwloc_type_2_topology_type(obj);
1856 if (type != KMP_HW_UNKNOWN) {
1857 types[depth] = type;
1858 hwloc_types[depth] = obj->type;
1859 depth++;
1860 }
1861 }
1862 KMP_ASSERT(depth > 0);
1863
1864 // Get the order for the types correct
1865 for (int i = 0, j = depth - 1; i < j; ++i, --j) {
1866 hwloc_obj_type_t hwloc_temp = hwloc_types[i];
1867 kmp_hw_t temp = types[i];
1868 types[i] = types[j];
1869 types[j] = temp;
1870 hwloc_types[i] = hwloc_types[j];
1871 hwloc_types[j] = hwloc_temp;
1872 }
1873
1874 // Allocate the data structure to be returned.
1875 __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
1876
1877 hw_thread_index = 0;
1878 pu = NULL;
1879 while ((pu = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, pu))) {
1880 int index = depth - 1;
1881 bool included = KMP_CPU_ISSET(pu->os_index, __kmp_affin_fullMask);
1882 kmp_hw_thread_t &hw_thread = __kmp_topology->at(hw_thread_index);
1883 if (included) {
1884 hw_thread.clear();
1885 hw_thread.ids[index] = pu->logical_index;
1886 hw_thread.os_id = pu->os_index;
1887 // If multiple core types, then set that attribute for the hardware thread
1888 #if HWLOC_API_VERSION >= 0x00020400
1889 if (cpukinds) {
1890 int cpukind_index = -1;
1891 for (int i = 0; i < nr_cpu_kinds; ++i) {
1892 if (hwloc_bitmap_isset(cpukinds[i].mask, hw_thread.os_id)) {
1893 cpukind_index = i;
1894 break;
1895 }
1896 }
1897 if (cpukind_index >= 0) {
1898 hw_thread.attrs.set_core_type(cpukinds[cpukind_index].core_type);
1899 hw_thread.attrs.set_core_eff(cpukinds[cpukind_index].efficiency);
1900 }
1901 }
1902 #endif
1903 index--;
1904 }
1905 obj = pu;
1906 prev = obj;
1907 while (obj != root && obj != NULL) {
1908 obj = obj->parent;
1909 #if HWLOC_API_VERSION >= 0x00020000
1910 // NUMA Nodes are handled differently since they are not within the
1911 // parent/child structure anymore. They are separate children
1912 // of obj (memory_first_child points to first memory child)
1913 if (obj->memory_arity) {
1914 hwloc_obj_t memory;
1915 for (memory = obj->memory_first_child; memory;
1916 memory = hwloc_get_next_child(tp, obj, memory)) {
1917 if (memory->type == HWLOC_OBJ_NUMANODE)
1918 break;
1919 }
1920 if (memory && memory->type == HWLOC_OBJ_NUMANODE) {
1921 sub_id = __kmp_hwloc_get_sub_id(tp, memory, prev);
1922 if (included) {
1923 hw_thread.ids[index] = memory->logical_index;
1924 hw_thread.ids[index + 1] = sub_id;
1925 index--;
1926 }
1927 prev = memory;
1928 }
1929 prev = obj;
1930 }
1931 #endif
1932 type = __kmp_hwloc_type_2_topology_type(obj);
1933 if (type != KMP_HW_UNKNOWN) {
1934 sub_id = __kmp_hwloc_get_sub_id(tp, obj, prev);
1935 if (included) {
1936 hw_thread.ids[index] = obj->logical_index;
1937 hw_thread.ids[index + 1] = sub_id;
1938 index--;
1939 }
1940 prev = obj;
1941 }
1942 }
1943 if (included)
1944 hw_thread_index++;
1945 }
1946
1947 #if HWLOC_API_VERSION >= 0x00020400
1948 // Free the core types information
1949 if (cpukinds) {
1950 for (int idx = 0; idx < nr_cpu_kinds; ++idx)
1951 hwloc_bitmap_free(cpukinds[idx].mask);
1952 __kmp_free(cpukinds);
1953 }
1954 #endif
1955 __kmp_topology->sort_ids();
1956 return true;
1957 }
1958 #endif // KMP_USE_HWLOC
1959
1960 // If we don't know how to retrieve the machine's processor topology, or
1961 // encounter an error in doing so, this routine is called to form a "flat"
1962 // mapping of os thread id's <-> processor id's.
__kmp_affinity_create_flat_map(kmp_i18n_id_t * const msg_id)1963 static bool __kmp_affinity_create_flat_map(kmp_i18n_id_t *const msg_id) {
1964 *msg_id = kmp_i18n_null;
1965 int depth = 3;
1966 kmp_hw_t types[] = {KMP_HW_SOCKET, KMP_HW_CORE, KMP_HW_THREAD};
1967
1968 if (__kmp_affinity.flags.verbose) {
1969 KMP_INFORM(UsingFlatOS, "KMP_AFFINITY");
1970 }
1971
1972 // Even if __kmp_affinity.type == affinity_none, this routine might still
1973 // be called to set __kmp_ncores, as well as
1974 // __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
1975 if (!KMP_AFFINITY_CAPABLE()) {
1976 KMP_ASSERT(__kmp_affinity.type == affinity_none);
1977 __kmp_ncores = nPackages = __kmp_xproc;
1978 __kmp_nThreadsPerCore = nCoresPerPkg = 1;
1979 return true;
1980 }
1981
1982 // When affinity is off, this routine will still be called to set
1983 // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
1984 // Make sure all these vars are set correctly, and return now if affinity is
1985 // not enabled.
1986 __kmp_ncores = nPackages = __kmp_avail_proc;
1987 __kmp_nThreadsPerCore = nCoresPerPkg = 1;
1988
1989 // Construct the data structure to be returned.
1990 __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
1991 int avail_ct = 0;
1992 int i;
1993 KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
1994 // Skip this proc if it is not included in the machine model.
1995 if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
1996 continue;
1997 }
1998 kmp_hw_thread_t &hw_thread = __kmp_topology->at(avail_ct);
1999 hw_thread.clear();
2000 hw_thread.os_id = i;
2001 hw_thread.ids[0] = i;
2002 hw_thread.ids[1] = 0;
2003 hw_thread.ids[2] = 0;
2004 avail_ct++;
2005 }
2006 if (__kmp_affinity.flags.verbose) {
2007 KMP_INFORM(OSProcToPackage, "KMP_AFFINITY");
2008 }
2009 return true;
2010 }
2011
2012 #if KMP_GROUP_AFFINITY
2013 // If multiple Windows* OS processor groups exist, we can create a 2-level
2014 // topology map with the groups at level 0 and the individual procs at level 1.
2015 // This facilitates letting the threads float among all procs in a group,
2016 // if granularity=group (the default when there are multiple groups).
__kmp_affinity_create_proc_group_map(kmp_i18n_id_t * const msg_id)2017 static bool __kmp_affinity_create_proc_group_map(kmp_i18n_id_t *const msg_id) {
2018 *msg_id = kmp_i18n_null;
2019 int depth = 3;
2020 kmp_hw_t types[] = {KMP_HW_PROC_GROUP, KMP_HW_CORE, KMP_HW_THREAD};
2021 const static size_t BITS_PER_GROUP = CHAR_BIT * sizeof(DWORD_PTR);
2022
2023 if (__kmp_affinity.flags.verbose) {
2024 KMP_INFORM(AffWindowsProcGroupMap, "KMP_AFFINITY");
2025 }
2026
2027 // If we aren't affinity capable, then use flat topology
2028 if (!KMP_AFFINITY_CAPABLE()) {
2029 KMP_ASSERT(__kmp_affinity.type == affinity_none);
2030 nPackages = __kmp_num_proc_groups;
2031 __kmp_nThreadsPerCore = 1;
2032 __kmp_ncores = __kmp_xproc;
2033 nCoresPerPkg = nPackages / __kmp_ncores;
2034 return true;
2035 }
2036
2037 // Construct the data structure to be returned.
2038 __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
2039 int avail_ct = 0;
2040 int i;
2041 KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
2042 // Skip this proc if it is not included in the machine model.
2043 if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
2044 continue;
2045 }
2046 kmp_hw_thread_t &hw_thread = __kmp_topology->at(avail_ct++);
2047 hw_thread.clear();
2048 hw_thread.os_id = i;
2049 hw_thread.ids[0] = i / BITS_PER_GROUP;
2050 hw_thread.ids[1] = hw_thread.ids[2] = i % BITS_PER_GROUP;
2051 }
2052 return true;
2053 }
2054 #endif /* KMP_GROUP_AFFINITY */
2055
2056 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2057
2058 template <kmp_uint32 LSB, kmp_uint32 MSB>
__kmp_extract_bits(kmp_uint32 v)2059 static inline unsigned __kmp_extract_bits(kmp_uint32 v) {
2060 const kmp_uint32 SHIFT_LEFT = sizeof(kmp_uint32) * 8 - 1 - MSB;
2061 const kmp_uint32 SHIFT_RIGHT = LSB;
2062 kmp_uint32 retval = v;
2063 retval <<= SHIFT_LEFT;
2064 retval >>= (SHIFT_LEFT + SHIFT_RIGHT);
2065 return retval;
2066 }
2067
__kmp_cpuid_mask_width(int count)2068 static int __kmp_cpuid_mask_width(int count) {
2069 int r = 0;
2070
2071 while ((1 << r) < count)
2072 ++r;
2073 return r;
2074 }
2075
2076 class apicThreadInfo {
2077 public:
2078 unsigned osId; // param to __kmp_affinity_bind_thread
2079 unsigned apicId; // from cpuid after binding
2080 unsigned maxCoresPerPkg; // ""
2081 unsigned maxThreadsPerPkg; // ""
2082 unsigned pkgId; // inferred from above values
2083 unsigned coreId; // ""
2084 unsigned threadId; // ""
2085 };
2086
__kmp_affinity_cmp_apicThreadInfo_phys_id(const void * a,const void * b)2087 static int __kmp_affinity_cmp_apicThreadInfo_phys_id(const void *a,
2088 const void *b) {
2089 const apicThreadInfo *aa = (const apicThreadInfo *)a;
2090 const apicThreadInfo *bb = (const apicThreadInfo *)b;
2091 if (aa->pkgId < bb->pkgId)
2092 return -1;
2093 if (aa->pkgId > bb->pkgId)
2094 return 1;
2095 if (aa->coreId < bb->coreId)
2096 return -1;
2097 if (aa->coreId > bb->coreId)
2098 return 1;
2099 if (aa->threadId < bb->threadId)
2100 return -1;
2101 if (aa->threadId > bb->threadId)
2102 return 1;
2103 return 0;
2104 }
2105
2106 class kmp_cache_info_t {
2107 public:
2108 struct info_t {
2109 unsigned level, mask;
2110 };
kmp_cache_info_t()2111 kmp_cache_info_t() : depth(0) { get_leaf4_levels(); }
get_depth() const2112 size_t get_depth() const { return depth; }
operator [](size_t index)2113 info_t &operator[](size_t index) { return table[index]; }
operator [](size_t index) const2114 const info_t &operator[](size_t index) const { return table[index]; }
2115
get_topology_type(unsigned level)2116 static kmp_hw_t get_topology_type(unsigned level) {
2117 KMP_DEBUG_ASSERT(level >= 1 && level <= MAX_CACHE_LEVEL);
2118 switch (level) {
2119 case 1:
2120 return KMP_HW_L1;
2121 case 2:
2122 return KMP_HW_L2;
2123 case 3:
2124 return KMP_HW_L3;
2125 }
2126 return KMP_HW_UNKNOWN;
2127 }
2128
2129 private:
2130 static const int MAX_CACHE_LEVEL = 3;
2131
2132 size_t depth;
2133 info_t table[MAX_CACHE_LEVEL];
2134
get_leaf4_levels()2135 void get_leaf4_levels() {
2136 unsigned level = 0;
2137 while (depth < MAX_CACHE_LEVEL) {
2138 unsigned cache_type, max_threads_sharing;
2139 unsigned cache_level, cache_mask_width;
2140 kmp_cpuid buf2;
2141 __kmp_x86_cpuid(4, level, &buf2);
2142 cache_type = __kmp_extract_bits<0, 4>(buf2.eax);
2143 if (!cache_type)
2144 break;
2145 // Skip instruction caches
2146 if (cache_type == 2) {
2147 level++;
2148 continue;
2149 }
2150 max_threads_sharing = __kmp_extract_bits<14, 25>(buf2.eax) + 1;
2151 cache_mask_width = __kmp_cpuid_mask_width(max_threads_sharing);
2152 cache_level = __kmp_extract_bits<5, 7>(buf2.eax);
2153 table[depth].level = cache_level;
2154 table[depth].mask = ((-1) << cache_mask_width);
2155 depth++;
2156 level++;
2157 }
2158 }
2159 };
2160
2161 // On IA-32 architecture and Intel(R) 64 architecture, we attempt to use
2162 // an algorithm which cycles through the available os threads, setting
2163 // the current thread's affinity mask to that thread, and then retrieves
2164 // the Apic Id for each thread context using the cpuid instruction.
__kmp_affinity_create_apicid_map(kmp_i18n_id_t * const msg_id)2165 static bool __kmp_affinity_create_apicid_map(kmp_i18n_id_t *const msg_id) {
2166 kmp_cpuid buf;
2167 *msg_id = kmp_i18n_null;
2168
2169 if (__kmp_affinity.flags.verbose) {
2170 KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(DecodingLegacyAPIC));
2171 }
2172
2173 // Check if cpuid leaf 4 is supported.
2174 __kmp_x86_cpuid(0, 0, &buf);
2175 if (buf.eax < 4) {
2176 *msg_id = kmp_i18n_str_NoLeaf4Support;
2177 return false;
2178 }
2179
2180 // The algorithm used starts by setting the affinity to each available thread
2181 // and retrieving info from the cpuid instruction, so if we are not capable of
2182 // calling __kmp_get_system_affinity() and _kmp_get_system_affinity(), then we
2183 // need to do something else - use the defaults that we calculated from
2184 // issuing cpuid without binding to each proc.
2185 if (!KMP_AFFINITY_CAPABLE()) {
2186 // Hack to try and infer the machine topology using only the data
2187 // available from cpuid on the current thread, and __kmp_xproc.
2188 KMP_ASSERT(__kmp_affinity.type == affinity_none);
2189
2190 // Get an upper bound on the number of threads per package using cpuid(1).
2191 // On some OS/chps combinations where HT is supported by the chip but is
2192 // disabled, this value will be 2 on a single core chip. Usually, it will be
2193 // 2 if HT is enabled and 1 if HT is disabled.
2194 __kmp_x86_cpuid(1, 0, &buf);
2195 int maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
2196 if (maxThreadsPerPkg == 0) {
2197 maxThreadsPerPkg = 1;
2198 }
2199
2200 // The num cores per pkg comes from cpuid(4). 1 must be added to the encoded
2201 // value.
2202 //
2203 // The author of cpu_count.cpp treated this only an upper bound on the
2204 // number of cores, but I haven't seen any cases where it was greater than
2205 // the actual number of cores, so we will treat it as exact in this block of
2206 // code.
2207 //
2208 // First, we need to check if cpuid(4) is supported on this chip. To see if
2209 // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n or
2210 // greater.
2211 __kmp_x86_cpuid(0, 0, &buf);
2212 if (buf.eax >= 4) {
2213 __kmp_x86_cpuid(4, 0, &buf);
2214 nCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
2215 } else {
2216 nCoresPerPkg = 1;
2217 }
2218
2219 // There is no way to reliably tell if HT is enabled without issuing the
2220 // cpuid instruction from every thread, can correlating the cpuid info, so
2221 // if the machine is not affinity capable, we assume that HT is off. We have
2222 // seen quite a few machines where maxThreadsPerPkg is 2, yet the machine
2223 // does not support HT.
2224 //
2225 // - Older OSes are usually found on machines with older chips, which do not
2226 // support HT.
2227 // - The performance penalty for mistakenly identifying a machine as HT when
2228 // it isn't (which results in blocktime being incorrectly set to 0) is
2229 // greater than the penalty when for mistakenly identifying a machine as
2230 // being 1 thread/core when it is really HT enabled (which results in
2231 // blocktime being incorrectly set to a positive value).
2232 __kmp_ncores = __kmp_xproc;
2233 nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
2234 __kmp_nThreadsPerCore = 1;
2235 return true;
2236 }
2237
2238 // From here on, we can assume that it is safe to call
2239 // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
2240 // __kmp_affinity.type = affinity_none.
2241
2242 // Save the affinity mask for the current thread.
2243 kmp_affinity_raii_t previous_affinity;
2244
2245 // Run through each of the available contexts, binding the current thread
2246 // to it, and obtaining the pertinent information using the cpuid instr.
2247 //
2248 // The relevant information is:
2249 // - Apic Id: Bits 24:31 of ebx after issuing cpuid(1) - each thread context
2250 // has a uniqie Apic Id, which is of the form pkg# : core# : thread#.
2251 // - Max Threads Per Pkg: Bits 16:23 of ebx after issuing cpuid(1). The value
2252 // of this field determines the width of the core# + thread# fields in the
2253 // Apic Id. It is also an upper bound on the number of threads per
2254 // package, but it has been verified that situations happen were it is not
2255 // exact. In particular, on certain OS/chip combinations where Intel(R)
2256 // Hyper-Threading Technology is supported by the chip but has been
2257 // disabled, the value of this field will be 2 (for a single core chip).
2258 // On other OS/chip combinations supporting Intel(R) Hyper-Threading
2259 // Technology, the value of this field will be 1 when Intel(R)
2260 // Hyper-Threading Technology is disabled and 2 when it is enabled.
2261 // - Max Cores Per Pkg: Bits 26:31 of eax after issuing cpuid(4). The value
2262 // of this field (+1) determines the width of the core# field in the Apic
2263 // Id. The comments in "cpucount.cpp" say that this value is an upper
2264 // bound, but the IA-32 architecture manual says that it is exactly the
2265 // number of cores per package, and I haven't seen any case where it
2266 // wasn't.
2267 //
2268 // From this information, deduce the package Id, core Id, and thread Id,
2269 // and set the corresponding fields in the apicThreadInfo struct.
2270 unsigned i;
2271 apicThreadInfo *threadInfo = (apicThreadInfo *)__kmp_allocate(
2272 __kmp_avail_proc * sizeof(apicThreadInfo));
2273 unsigned nApics = 0;
2274 KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
2275 // Skip this proc if it is not included in the machine model.
2276 if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
2277 continue;
2278 }
2279 KMP_DEBUG_ASSERT((int)nApics < __kmp_avail_proc);
2280
2281 __kmp_affinity_dispatch->bind_thread(i);
2282 threadInfo[nApics].osId = i;
2283
2284 // The apic id and max threads per pkg come from cpuid(1).
2285 __kmp_x86_cpuid(1, 0, &buf);
2286 if (((buf.edx >> 9) & 1) == 0) {
2287 __kmp_free(threadInfo);
2288 *msg_id = kmp_i18n_str_ApicNotPresent;
2289 return false;
2290 }
2291 threadInfo[nApics].apicId = (buf.ebx >> 24) & 0xff;
2292 threadInfo[nApics].maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
2293 if (threadInfo[nApics].maxThreadsPerPkg == 0) {
2294 threadInfo[nApics].maxThreadsPerPkg = 1;
2295 }
2296
2297 // Max cores per pkg comes from cpuid(4). 1 must be added to the encoded
2298 // value.
2299 //
2300 // First, we need to check if cpuid(4) is supported on this chip. To see if
2301 // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n
2302 // or greater.
2303 __kmp_x86_cpuid(0, 0, &buf);
2304 if (buf.eax >= 4) {
2305 __kmp_x86_cpuid(4, 0, &buf);
2306 threadInfo[nApics].maxCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
2307 } else {
2308 threadInfo[nApics].maxCoresPerPkg = 1;
2309 }
2310
2311 // Infer the pkgId / coreId / threadId using only the info obtained locally.
2312 int widthCT = __kmp_cpuid_mask_width(threadInfo[nApics].maxThreadsPerPkg);
2313 threadInfo[nApics].pkgId = threadInfo[nApics].apicId >> widthCT;
2314
2315 int widthC = __kmp_cpuid_mask_width(threadInfo[nApics].maxCoresPerPkg);
2316 int widthT = widthCT - widthC;
2317 if (widthT < 0) {
2318 // I've never seen this one happen, but I suppose it could, if the cpuid
2319 // instruction on a chip was really screwed up. Make sure to restore the
2320 // affinity mask before the tail call.
2321 __kmp_free(threadInfo);
2322 *msg_id = kmp_i18n_str_InvalidCpuidInfo;
2323 return false;
2324 }
2325
2326 int maskC = (1 << widthC) - 1;
2327 threadInfo[nApics].coreId = (threadInfo[nApics].apicId >> widthT) & maskC;
2328
2329 int maskT = (1 << widthT) - 1;
2330 threadInfo[nApics].threadId = threadInfo[nApics].apicId & maskT;
2331
2332 nApics++;
2333 }
2334
2335 // We've collected all the info we need.
2336 // Restore the old affinity mask for this thread.
2337 previous_affinity.restore();
2338
2339 // Sort the threadInfo table by physical Id.
2340 qsort(threadInfo, nApics, sizeof(*threadInfo),
2341 __kmp_affinity_cmp_apicThreadInfo_phys_id);
2342
2343 // The table is now sorted by pkgId / coreId / threadId, but we really don't
2344 // know the radix of any of the fields. pkgId's may be sparsely assigned among
2345 // the chips on a system. Although coreId's are usually assigned
2346 // [0 .. coresPerPkg-1] and threadId's are usually assigned
2347 // [0..threadsPerCore-1], we don't want to make any such assumptions.
2348 //
2349 // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
2350 // total # packages) are at this point - we want to determine that now. We
2351 // only have an upper bound on the first two figures.
2352 //
2353 // We also perform a consistency check at this point: the values returned by
2354 // the cpuid instruction for any thread bound to a given package had better
2355 // return the same info for maxThreadsPerPkg and maxCoresPerPkg.
2356 nPackages = 1;
2357 nCoresPerPkg = 1;
2358 __kmp_nThreadsPerCore = 1;
2359 unsigned nCores = 1;
2360
2361 unsigned pkgCt = 1; // to determine radii
2362 unsigned lastPkgId = threadInfo[0].pkgId;
2363 unsigned coreCt = 1;
2364 unsigned lastCoreId = threadInfo[0].coreId;
2365 unsigned threadCt = 1;
2366 unsigned lastThreadId = threadInfo[0].threadId;
2367
2368 // intra-pkg consist checks
2369 unsigned prevMaxCoresPerPkg = threadInfo[0].maxCoresPerPkg;
2370 unsigned prevMaxThreadsPerPkg = threadInfo[0].maxThreadsPerPkg;
2371
2372 for (i = 1; i < nApics; i++) {
2373 if (threadInfo[i].pkgId != lastPkgId) {
2374 nCores++;
2375 pkgCt++;
2376 lastPkgId = threadInfo[i].pkgId;
2377 if ((int)coreCt > nCoresPerPkg)
2378 nCoresPerPkg = coreCt;
2379 coreCt = 1;
2380 lastCoreId = threadInfo[i].coreId;
2381 if ((int)threadCt > __kmp_nThreadsPerCore)
2382 __kmp_nThreadsPerCore = threadCt;
2383 threadCt = 1;
2384 lastThreadId = threadInfo[i].threadId;
2385
2386 // This is a different package, so go on to the next iteration without
2387 // doing any consistency checks. Reset the consistency check vars, though.
2388 prevMaxCoresPerPkg = threadInfo[i].maxCoresPerPkg;
2389 prevMaxThreadsPerPkg = threadInfo[i].maxThreadsPerPkg;
2390 continue;
2391 }
2392
2393 if (threadInfo[i].coreId != lastCoreId) {
2394 nCores++;
2395 coreCt++;
2396 lastCoreId = threadInfo[i].coreId;
2397 if ((int)threadCt > __kmp_nThreadsPerCore)
2398 __kmp_nThreadsPerCore = threadCt;
2399 threadCt = 1;
2400 lastThreadId = threadInfo[i].threadId;
2401 } else if (threadInfo[i].threadId != lastThreadId) {
2402 threadCt++;
2403 lastThreadId = threadInfo[i].threadId;
2404 } else {
2405 __kmp_free(threadInfo);
2406 *msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
2407 return false;
2408 }
2409
2410 // Check to make certain that the maxCoresPerPkg and maxThreadsPerPkg
2411 // fields agree between all the threads bounds to a given package.
2412 if ((prevMaxCoresPerPkg != threadInfo[i].maxCoresPerPkg) ||
2413 (prevMaxThreadsPerPkg != threadInfo[i].maxThreadsPerPkg)) {
2414 __kmp_free(threadInfo);
2415 *msg_id = kmp_i18n_str_InconsistentCpuidInfo;
2416 return false;
2417 }
2418 }
2419 // When affinity is off, this routine will still be called to set
2420 // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
2421 // Make sure all these vars are set correctly
2422 nPackages = pkgCt;
2423 if ((int)coreCt > nCoresPerPkg)
2424 nCoresPerPkg = coreCt;
2425 if ((int)threadCt > __kmp_nThreadsPerCore)
2426 __kmp_nThreadsPerCore = threadCt;
2427 __kmp_ncores = nCores;
2428 KMP_DEBUG_ASSERT(nApics == (unsigned)__kmp_avail_proc);
2429
2430 // Now that we've determined the number of packages, the number of cores per
2431 // package, and the number of threads per core, we can construct the data
2432 // structure that is to be returned.
2433 int idx = 0;
2434 int pkgLevel = 0;
2435 int coreLevel = 1;
2436 int threadLevel = 2;
2437 //(__kmp_nThreadsPerCore <= 1) ? -1 : ((coreLevel >= 0) ? 2 : 1);
2438 int depth = (pkgLevel >= 0) + (coreLevel >= 0) + (threadLevel >= 0);
2439 kmp_hw_t types[3];
2440 if (pkgLevel >= 0)
2441 types[idx++] = KMP_HW_SOCKET;
2442 if (coreLevel >= 0)
2443 types[idx++] = KMP_HW_CORE;
2444 if (threadLevel >= 0)
2445 types[idx++] = KMP_HW_THREAD;
2446
2447 KMP_ASSERT(depth > 0);
2448 __kmp_topology = kmp_topology_t::allocate(nApics, depth, types);
2449
2450 for (i = 0; i < nApics; ++i) {
2451 idx = 0;
2452 unsigned os = threadInfo[i].osId;
2453 kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
2454 hw_thread.clear();
2455
2456 if (pkgLevel >= 0) {
2457 hw_thread.ids[idx++] = threadInfo[i].pkgId;
2458 }
2459 if (coreLevel >= 0) {
2460 hw_thread.ids[idx++] = threadInfo[i].coreId;
2461 }
2462 if (threadLevel >= 0) {
2463 hw_thread.ids[idx++] = threadInfo[i].threadId;
2464 }
2465 hw_thread.os_id = os;
2466 }
2467
2468 __kmp_free(threadInfo);
2469 __kmp_topology->sort_ids();
2470 if (!__kmp_topology->check_ids()) {
2471 kmp_topology_t::deallocate(__kmp_topology);
2472 __kmp_topology = nullptr;
2473 *msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
2474 return false;
2475 }
2476 return true;
2477 }
2478
2479 // Hybrid cpu detection using CPUID.1A
2480 // Thread should be pinned to processor already
__kmp_get_hybrid_info(kmp_hw_core_type_t * type,int * efficiency,unsigned * native_model_id)2481 static void __kmp_get_hybrid_info(kmp_hw_core_type_t *type, int *efficiency,
2482 unsigned *native_model_id) {
2483 kmp_cpuid buf;
2484 __kmp_x86_cpuid(0x1a, 0, &buf);
2485 *type = (kmp_hw_core_type_t)__kmp_extract_bits<24, 31>(buf.eax);
2486 switch (*type) {
2487 case KMP_HW_CORE_TYPE_ATOM:
2488 *efficiency = 0;
2489 break;
2490 case KMP_HW_CORE_TYPE_CORE:
2491 *efficiency = 1;
2492 break;
2493 default:
2494 *efficiency = 0;
2495 }
2496 *native_model_id = __kmp_extract_bits<0, 23>(buf.eax);
2497 }
2498
2499 // Intel(R) microarchitecture code name Nehalem, Dunnington and later
2500 // architectures support a newer interface for specifying the x2APIC Ids,
2501 // based on CPUID.B or CPUID.1F
2502 /*
2503 * CPUID.B or 1F, Input ECX (sub leaf # aka level number)
2504 Bits Bits Bits Bits
2505 31-16 15-8 7-4 4-0
2506 ---+-----------+--------------+-------------+-----------------+
2507 EAX| reserved | reserved | reserved | Bits to Shift |
2508 ---+-----------|--------------+-------------+-----------------|
2509 EBX| reserved | Num logical processors at level (16 bits) |
2510 ---+-----------|--------------+-------------------------------|
2511 ECX| reserved | Level Type | Level Number (8 bits) |
2512 ---+-----------+--------------+-------------------------------|
2513 EDX| X2APIC ID (32 bits) |
2514 ---+----------------------------------------------------------+
2515 */
2516
2517 enum {
2518 INTEL_LEVEL_TYPE_INVALID = 0, // Package level
2519 INTEL_LEVEL_TYPE_SMT = 1,
2520 INTEL_LEVEL_TYPE_CORE = 2,
2521 INTEL_LEVEL_TYPE_MODULE = 3,
2522 INTEL_LEVEL_TYPE_TILE = 4,
2523 INTEL_LEVEL_TYPE_DIE = 5,
2524 INTEL_LEVEL_TYPE_LAST = 6,
2525 };
2526
2527 struct cpuid_level_info_t {
2528 unsigned level_type, mask, mask_width, nitems, cache_mask;
2529 };
2530
__kmp_intel_type_2_topology_type(int intel_type)2531 static kmp_hw_t __kmp_intel_type_2_topology_type(int intel_type) {
2532 switch (intel_type) {
2533 case INTEL_LEVEL_TYPE_INVALID:
2534 return KMP_HW_SOCKET;
2535 case INTEL_LEVEL_TYPE_SMT:
2536 return KMP_HW_THREAD;
2537 case INTEL_LEVEL_TYPE_CORE:
2538 return KMP_HW_CORE;
2539 case INTEL_LEVEL_TYPE_TILE:
2540 return KMP_HW_TILE;
2541 case INTEL_LEVEL_TYPE_MODULE:
2542 return KMP_HW_MODULE;
2543 case INTEL_LEVEL_TYPE_DIE:
2544 return KMP_HW_DIE;
2545 }
2546 return KMP_HW_UNKNOWN;
2547 }
2548
2549 // This function takes the topology leaf, a levels array to store the levels
2550 // detected and a bitmap of the known levels.
2551 // Returns the number of levels in the topology
2552 static unsigned
__kmp_x2apicid_get_levels(int leaf,cpuid_level_info_t levels[INTEL_LEVEL_TYPE_LAST],kmp_uint64 known_levels)2553 __kmp_x2apicid_get_levels(int leaf,
2554 cpuid_level_info_t levels[INTEL_LEVEL_TYPE_LAST],
2555 kmp_uint64 known_levels) {
2556 unsigned level, levels_index;
2557 unsigned level_type, mask_width, nitems;
2558 kmp_cpuid buf;
2559
2560 // New algorithm has known topology layers act as highest unknown topology
2561 // layers when unknown topology layers exist.
2562 // e.g., Suppose layers were SMT <X> CORE <Y> <Z> PACKAGE, where <X> <Y> <Z>
2563 // are unknown topology layers, Then SMT will take the characteristics of
2564 // (SMT x <X>) and CORE will take the characteristics of (CORE x <Y> x <Z>).
2565 // This eliminates unknown portions of the topology while still keeping the
2566 // correct structure.
2567 level = levels_index = 0;
2568 do {
2569 __kmp_x86_cpuid(leaf, level, &buf);
2570 level_type = __kmp_extract_bits<8, 15>(buf.ecx);
2571 mask_width = __kmp_extract_bits<0, 4>(buf.eax);
2572 nitems = __kmp_extract_bits<0, 15>(buf.ebx);
2573 if (level_type != INTEL_LEVEL_TYPE_INVALID && nitems == 0)
2574 return 0;
2575
2576 if (known_levels & (1ull << level_type)) {
2577 // Add a new level to the topology
2578 KMP_ASSERT(levels_index < INTEL_LEVEL_TYPE_LAST);
2579 levels[levels_index].level_type = level_type;
2580 levels[levels_index].mask_width = mask_width;
2581 levels[levels_index].nitems = nitems;
2582 levels_index++;
2583 } else {
2584 // If it is an unknown level, then logically move the previous layer up
2585 if (levels_index > 0) {
2586 levels[levels_index - 1].mask_width = mask_width;
2587 levels[levels_index - 1].nitems = nitems;
2588 }
2589 }
2590 level++;
2591 } while (level_type != INTEL_LEVEL_TYPE_INVALID);
2592
2593 // Ensure the INTEL_LEVEL_TYPE_INVALID (Socket) layer isn't first
2594 if (levels_index == 0 || levels[0].level_type == INTEL_LEVEL_TYPE_INVALID)
2595 return 0;
2596
2597 // Set the masks to & with apicid
2598 for (unsigned i = 0; i < levels_index; ++i) {
2599 if (levels[i].level_type != INTEL_LEVEL_TYPE_INVALID) {
2600 levels[i].mask = ~((-1) << levels[i].mask_width);
2601 levels[i].cache_mask = (-1) << levels[i].mask_width;
2602 for (unsigned j = 0; j < i; ++j)
2603 levels[i].mask ^= levels[j].mask;
2604 } else {
2605 KMP_DEBUG_ASSERT(i > 0);
2606 levels[i].mask = (-1) << levels[i - 1].mask_width;
2607 levels[i].cache_mask = 0;
2608 }
2609 }
2610 return levels_index;
2611 }
2612
__kmp_affinity_create_x2apicid_map(kmp_i18n_id_t * const msg_id)2613 static bool __kmp_affinity_create_x2apicid_map(kmp_i18n_id_t *const msg_id) {
2614
2615 cpuid_level_info_t levels[INTEL_LEVEL_TYPE_LAST];
2616 kmp_hw_t types[INTEL_LEVEL_TYPE_LAST];
2617 unsigned levels_index;
2618 kmp_cpuid buf;
2619 kmp_uint64 known_levels;
2620 int topology_leaf, highest_leaf, apic_id;
2621 int num_leaves;
2622 static int leaves[] = {0, 0};
2623
2624 kmp_i18n_id_t leaf_message_id;
2625
2626 KMP_BUILD_ASSERT(sizeof(known_levels) * CHAR_BIT > KMP_HW_LAST);
2627
2628 *msg_id = kmp_i18n_null;
2629 if (__kmp_affinity.flags.verbose) {
2630 KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(Decodingx2APIC));
2631 }
2632
2633 // Figure out the known topology levels
2634 known_levels = 0ull;
2635 for (int i = 0; i < INTEL_LEVEL_TYPE_LAST; ++i) {
2636 if (__kmp_intel_type_2_topology_type(i) != KMP_HW_UNKNOWN) {
2637 known_levels |= (1ull << i);
2638 }
2639 }
2640
2641 // Get the highest cpuid leaf supported
2642 __kmp_x86_cpuid(0, 0, &buf);
2643 highest_leaf = buf.eax;
2644
2645 // If a specific topology method was requested, only allow that specific leaf
2646 // otherwise, try both leaves 31 and 11 in that order
2647 num_leaves = 0;
2648 if (__kmp_affinity_top_method == affinity_top_method_x2apicid) {
2649 num_leaves = 1;
2650 leaves[0] = 11;
2651 leaf_message_id = kmp_i18n_str_NoLeaf11Support;
2652 } else if (__kmp_affinity_top_method == affinity_top_method_x2apicid_1f) {
2653 num_leaves = 1;
2654 leaves[0] = 31;
2655 leaf_message_id = kmp_i18n_str_NoLeaf31Support;
2656 } else {
2657 num_leaves = 2;
2658 leaves[0] = 31;
2659 leaves[1] = 11;
2660 leaf_message_id = kmp_i18n_str_NoLeaf11Support;
2661 }
2662
2663 // Check to see if cpuid leaf 31 or 11 is supported.
2664 __kmp_nThreadsPerCore = nCoresPerPkg = nPackages = 1;
2665 topology_leaf = -1;
2666 for (int i = 0; i < num_leaves; ++i) {
2667 int leaf = leaves[i];
2668 if (highest_leaf < leaf)
2669 continue;
2670 __kmp_x86_cpuid(leaf, 0, &buf);
2671 if (buf.ebx == 0)
2672 continue;
2673 topology_leaf = leaf;
2674 levels_index = __kmp_x2apicid_get_levels(leaf, levels, known_levels);
2675 if (levels_index == 0)
2676 continue;
2677 break;
2678 }
2679 if (topology_leaf == -1 || levels_index == 0) {
2680 *msg_id = leaf_message_id;
2681 return false;
2682 }
2683 KMP_ASSERT(levels_index <= INTEL_LEVEL_TYPE_LAST);
2684
2685 // The algorithm used starts by setting the affinity to each available thread
2686 // and retrieving info from the cpuid instruction, so if we are not capable of
2687 // calling __kmp_get_system_affinity() and __kmp_get_system_affinity(), then
2688 // we need to do something else - use the defaults that we calculated from
2689 // issuing cpuid without binding to each proc.
2690 if (!KMP_AFFINITY_CAPABLE()) {
2691 // Hack to try and infer the machine topology using only the data
2692 // available from cpuid on the current thread, and __kmp_xproc.
2693 KMP_ASSERT(__kmp_affinity.type == affinity_none);
2694 for (unsigned i = 0; i < levels_index; ++i) {
2695 if (levels[i].level_type == INTEL_LEVEL_TYPE_SMT) {
2696 __kmp_nThreadsPerCore = levels[i].nitems;
2697 } else if (levels[i].level_type == INTEL_LEVEL_TYPE_CORE) {
2698 nCoresPerPkg = levels[i].nitems;
2699 }
2700 }
2701 __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
2702 nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
2703 return true;
2704 }
2705
2706 // Allocate the data structure to be returned.
2707 int depth = levels_index;
2708 for (int i = depth - 1, j = 0; i >= 0; --i, ++j)
2709 types[j] = __kmp_intel_type_2_topology_type(levels[i].level_type);
2710 __kmp_topology =
2711 kmp_topology_t::allocate(__kmp_avail_proc, levels_index, types);
2712
2713 // Insert equivalent cache types if they exist
2714 kmp_cache_info_t cache_info;
2715 for (size_t i = 0; i < cache_info.get_depth(); ++i) {
2716 const kmp_cache_info_t::info_t &info = cache_info[i];
2717 unsigned cache_mask = info.mask;
2718 unsigned cache_level = info.level;
2719 for (unsigned j = 0; j < levels_index; ++j) {
2720 unsigned hw_cache_mask = levels[j].cache_mask;
2721 kmp_hw_t cache_type = kmp_cache_info_t::get_topology_type(cache_level);
2722 if (hw_cache_mask == cache_mask && j < levels_index - 1) {
2723 kmp_hw_t type =
2724 __kmp_intel_type_2_topology_type(levels[j + 1].level_type);
2725 __kmp_topology->set_equivalent_type(cache_type, type);
2726 }
2727 }
2728 }
2729
2730 // From here on, we can assume that it is safe to call
2731 // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
2732 // __kmp_affinity.type = affinity_none.
2733
2734 // Save the affinity mask for the current thread.
2735 kmp_affinity_raii_t previous_affinity;
2736
2737 // Run through each of the available contexts, binding the current thread
2738 // to it, and obtaining the pertinent information using the cpuid instr.
2739 unsigned int proc;
2740 int hw_thread_index = 0;
2741 KMP_CPU_SET_ITERATE(proc, __kmp_affin_fullMask) {
2742 cpuid_level_info_t my_levels[INTEL_LEVEL_TYPE_LAST];
2743 unsigned my_levels_index;
2744
2745 // Skip this proc if it is not included in the machine model.
2746 if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
2747 continue;
2748 }
2749 KMP_DEBUG_ASSERT(hw_thread_index < __kmp_avail_proc);
2750
2751 __kmp_affinity_dispatch->bind_thread(proc);
2752
2753 // New algorithm
2754 __kmp_x86_cpuid(topology_leaf, 0, &buf);
2755 apic_id = buf.edx;
2756 kmp_hw_thread_t &hw_thread = __kmp_topology->at(hw_thread_index);
2757 my_levels_index =
2758 __kmp_x2apicid_get_levels(topology_leaf, my_levels, known_levels);
2759 if (my_levels_index == 0 || my_levels_index != levels_index) {
2760 *msg_id = kmp_i18n_str_InvalidCpuidInfo;
2761 return false;
2762 }
2763 hw_thread.clear();
2764 hw_thread.os_id = proc;
2765 // Put in topology information
2766 for (unsigned j = 0, idx = depth - 1; j < my_levels_index; ++j, --idx) {
2767 hw_thread.ids[idx] = apic_id & my_levels[j].mask;
2768 if (j > 0) {
2769 hw_thread.ids[idx] >>= my_levels[j - 1].mask_width;
2770 }
2771 }
2772 // Hybrid information
2773 if (__kmp_is_hybrid_cpu() && highest_leaf >= 0x1a) {
2774 kmp_hw_core_type_t type;
2775 unsigned native_model_id;
2776 int efficiency;
2777 __kmp_get_hybrid_info(&type, &efficiency, &native_model_id);
2778 hw_thread.attrs.set_core_type(type);
2779 hw_thread.attrs.set_core_eff(efficiency);
2780 }
2781 hw_thread_index++;
2782 }
2783 KMP_ASSERT(hw_thread_index > 0);
2784 __kmp_topology->sort_ids();
2785 if (!__kmp_topology->check_ids()) {
2786 kmp_topology_t::deallocate(__kmp_topology);
2787 __kmp_topology = nullptr;
2788 *msg_id = kmp_i18n_str_x2ApicIDsNotUnique;
2789 return false;
2790 }
2791 return true;
2792 }
2793 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2794
2795 #define osIdIndex 0
2796 #define threadIdIndex 1
2797 #define coreIdIndex 2
2798 #define pkgIdIndex 3
2799 #define nodeIdIndex 4
2800
2801 typedef unsigned *ProcCpuInfo;
2802 static unsigned maxIndex = pkgIdIndex;
2803
__kmp_affinity_cmp_ProcCpuInfo_phys_id(const void * a,const void * b)2804 static int __kmp_affinity_cmp_ProcCpuInfo_phys_id(const void *a,
2805 const void *b) {
2806 unsigned i;
2807 const unsigned *aa = *(unsigned *const *)a;
2808 const unsigned *bb = *(unsigned *const *)b;
2809 for (i = maxIndex;; i--) {
2810 if (aa[i] < bb[i])
2811 return -1;
2812 if (aa[i] > bb[i])
2813 return 1;
2814 if (i == osIdIndex)
2815 break;
2816 }
2817 return 0;
2818 }
2819
2820 #if KMP_USE_HIER_SCHED
2821 // Set the array sizes for the hierarchy layers
__kmp_dispatch_set_hierarchy_values()2822 static void __kmp_dispatch_set_hierarchy_values() {
2823 // Set the maximum number of L1's to number of cores
2824 // Set the maximum number of L2's to either number of cores / 2 for
2825 // Intel(R) Xeon Phi(TM) coprocessor formally codenamed Knights Landing
2826 // Or the number of cores for Intel(R) Xeon(R) processors
2827 // Set the maximum number of NUMA nodes and L3's to number of packages
2828 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1] =
2829 nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
2830 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L1 + 1] = __kmp_ncores;
2831 #if KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_WINDOWS) && \
2832 KMP_MIC_SUPPORTED
2833 if (__kmp_mic_type >= mic3)
2834 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores / 2;
2835 else
2836 #endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
2837 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores;
2838 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L3 + 1] = nPackages;
2839 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_NUMA + 1] = nPackages;
2840 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_LOOP + 1] = 1;
2841 // Set the number of threads per unit
2842 // Number of hardware threads per L1/L2/L3/NUMA/LOOP
2843 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_THREAD + 1] = 1;
2844 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L1 + 1] =
2845 __kmp_nThreadsPerCore;
2846 #if KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_WINDOWS) && \
2847 KMP_MIC_SUPPORTED
2848 if (__kmp_mic_type >= mic3)
2849 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
2850 2 * __kmp_nThreadsPerCore;
2851 else
2852 #endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
2853 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
2854 __kmp_nThreadsPerCore;
2855 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L3 + 1] =
2856 nCoresPerPkg * __kmp_nThreadsPerCore;
2857 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_NUMA + 1] =
2858 nCoresPerPkg * __kmp_nThreadsPerCore;
2859 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_LOOP + 1] =
2860 nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
2861 }
2862
2863 // Return the index into the hierarchy for this tid and layer type (L1, L2, etc)
2864 // i.e., this thread's L1 or this thread's L2, etc.
__kmp_dispatch_get_index(int tid,kmp_hier_layer_e type)2865 int __kmp_dispatch_get_index(int tid, kmp_hier_layer_e type) {
2866 int index = type + 1;
2867 int num_hw_threads = __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1];
2868 KMP_DEBUG_ASSERT(type != kmp_hier_layer_e::LAYER_LAST);
2869 if (type == kmp_hier_layer_e::LAYER_THREAD)
2870 return tid;
2871 else if (type == kmp_hier_layer_e::LAYER_LOOP)
2872 return 0;
2873 KMP_DEBUG_ASSERT(__kmp_hier_max_units[index] != 0);
2874 if (tid >= num_hw_threads)
2875 tid = tid % num_hw_threads;
2876 return (tid / __kmp_hier_threads_per[index]) % __kmp_hier_max_units[index];
2877 }
2878
2879 // Return the number of t1's per t2
__kmp_dispatch_get_t1_per_t2(kmp_hier_layer_e t1,kmp_hier_layer_e t2)2880 int __kmp_dispatch_get_t1_per_t2(kmp_hier_layer_e t1, kmp_hier_layer_e t2) {
2881 int i1 = t1 + 1;
2882 int i2 = t2 + 1;
2883 KMP_DEBUG_ASSERT(i1 <= i2);
2884 KMP_DEBUG_ASSERT(t1 != kmp_hier_layer_e::LAYER_LAST);
2885 KMP_DEBUG_ASSERT(t2 != kmp_hier_layer_e::LAYER_LAST);
2886 KMP_DEBUG_ASSERT(__kmp_hier_threads_per[i1] != 0);
2887 // (nthreads/t2) / (nthreads/t1) = t1 / t2
2888 return __kmp_hier_threads_per[i2] / __kmp_hier_threads_per[i1];
2889 }
2890 #endif // KMP_USE_HIER_SCHED
2891
__kmp_cpuinfo_get_filename()2892 static inline const char *__kmp_cpuinfo_get_filename() {
2893 const char *filename;
2894 if (__kmp_cpuinfo_file != nullptr)
2895 filename = __kmp_cpuinfo_file;
2896 else
2897 filename = "/proc/cpuinfo";
2898 return filename;
2899 }
2900
__kmp_cpuinfo_get_envvar()2901 static inline const char *__kmp_cpuinfo_get_envvar() {
2902 const char *envvar = nullptr;
2903 if (__kmp_cpuinfo_file != nullptr)
2904 envvar = "KMP_CPUINFO_FILE";
2905 return envvar;
2906 }
2907
2908 // Parse /proc/cpuinfo (or an alternate file in the same format) to obtain the
2909 // affinity map. On AIX, the map is obtained through system SRAD (Scheduler
2910 // Resource Allocation Domain).
__kmp_affinity_create_cpuinfo_map(int * line,kmp_i18n_id_t * const msg_id)2911 static bool __kmp_affinity_create_cpuinfo_map(int *line,
2912 kmp_i18n_id_t *const msg_id) {
2913 *msg_id = kmp_i18n_null;
2914
2915 #if KMP_OS_AIX
2916 unsigned num_records = __kmp_xproc;
2917 #else
2918 const char *filename = __kmp_cpuinfo_get_filename();
2919 const char *envvar = __kmp_cpuinfo_get_envvar();
2920
2921 if (__kmp_affinity.flags.verbose) {
2922 KMP_INFORM(AffParseFilename, "KMP_AFFINITY", filename);
2923 }
2924
2925 kmp_safe_raii_file_t f(filename, "r", envvar);
2926
2927 // Scan of the file, and count the number of "processor" (osId) fields,
2928 // and find the highest value of <n> for a node_<n> field.
2929 char buf[256];
2930 unsigned num_records = 0;
2931 while (!feof(f)) {
2932 buf[sizeof(buf) - 1] = 1;
2933 if (!fgets(buf, sizeof(buf), f)) {
2934 // Read errors presumably because of EOF
2935 break;
2936 }
2937
2938 char s1[] = "processor";
2939 if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
2940 num_records++;
2941 continue;
2942 }
2943
2944 // FIXME - this will match "node_<n> <garbage>"
2945 unsigned level;
2946 if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
2947 // validate the input fisrt:
2948 if (level > (unsigned)__kmp_xproc) { // level is too big
2949 level = __kmp_xproc;
2950 }
2951 if (nodeIdIndex + level >= maxIndex) {
2952 maxIndex = nodeIdIndex + level;
2953 }
2954 continue;
2955 }
2956 }
2957
2958 // Check for empty file / no valid processor records, or too many. The number
2959 // of records can't exceed the number of valid bits in the affinity mask.
2960 if (num_records == 0) {
2961 *msg_id = kmp_i18n_str_NoProcRecords;
2962 return false;
2963 }
2964 if (num_records > (unsigned)__kmp_xproc) {
2965 *msg_id = kmp_i18n_str_TooManyProcRecords;
2966 return false;
2967 }
2968
2969 // Set the file pointer back to the beginning, so that we can scan the file
2970 // again, this time performing a full parse of the data. Allocate a vector of
2971 // ProcCpuInfo object, where we will place the data. Adding an extra element
2972 // at the end allows us to remove a lot of extra checks for termination
2973 // conditions.
2974 if (fseek(f, 0, SEEK_SET) != 0) {
2975 *msg_id = kmp_i18n_str_CantRewindCpuinfo;
2976 return false;
2977 }
2978 #endif // KMP_OS_AIX
2979
2980 // Allocate the array of records to store the proc info in. The dummy
2981 // element at the end makes the logic in filling them out easier to code.
2982 unsigned **threadInfo =
2983 (unsigned **)__kmp_allocate((num_records + 1) * sizeof(unsigned *));
2984 unsigned i;
2985 for (i = 0; i <= num_records; i++) {
2986 threadInfo[i] =
2987 (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2988 }
2989
2990 #define CLEANUP_THREAD_INFO \
2991 for (i = 0; i <= num_records; i++) { \
2992 __kmp_free(threadInfo[i]); \
2993 } \
2994 __kmp_free(threadInfo);
2995
2996 // A value of UINT_MAX means that we didn't find the field
2997 unsigned __index;
2998
2999 #define INIT_PROC_INFO(p) \
3000 for (__index = 0; __index <= maxIndex; __index++) { \
3001 (p)[__index] = UINT_MAX; \
3002 }
3003
3004 for (i = 0; i <= num_records; i++) {
3005 INIT_PROC_INFO(threadInfo[i]);
3006 }
3007
3008 #if KMP_OS_AIX
3009 int smt_threads;
3010 lpar_info_format1_t cpuinfo;
3011 unsigned num_avail = __kmp_xproc;
3012
3013 if (__kmp_affinity.flags.verbose)
3014 KMP_INFORM(AffParseFilename, "KMP_AFFINITY", "system info for topology");
3015
3016 // Get the number of SMT threads per core.
3017 int retval =
3018 lpar_get_info(LPAR_INFO_FORMAT1, &cpuinfo, sizeof(lpar_info_format1_t));
3019 if (!retval)
3020 smt_threads = cpuinfo.smt_threads;
3021 else {
3022 CLEANUP_THREAD_INFO;
3023 *msg_id = kmp_i18n_str_UnknownTopology;
3024 return false;
3025 }
3026
3027 // Allocate a resource set containing available system resourses.
3028 rsethandle_t sys_rset = rs_alloc(RS_SYSTEM);
3029 if (sys_rset == NULL) {
3030 CLEANUP_THREAD_INFO;
3031 *msg_id = kmp_i18n_str_UnknownTopology;
3032 return false;
3033 }
3034 // Allocate a resource set for the SRAD info.
3035 rsethandle_t srad = rs_alloc(RS_EMPTY);
3036 if (srad == NULL) {
3037 rs_free(sys_rset);
3038 CLEANUP_THREAD_INFO;
3039 *msg_id = kmp_i18n_str_UnknownTopology;
3040 return false;
3041 }
3042
3043 // Get the SRAD system detail level.
3044 int sradsdl = rs_getinfo(NULL, R_SRADSDL, 0);
3045 if (sradsdl < 0) {
3046 rs_free(sys_rset);
3047 rs_free(srad);
3048 CLEANUP_THREAD_INFO;
3049 *msg_id = kmp_i18n_str_UnknownTopology;
3050 return false;
3051 }
3052 // Get the number of RADs at that SRAD SDL.
3053 int num_rads = rs_numrads(sys_rset, sradsdl, 0);
3054 if (num_rads < 0) {
3055 rs_free(sys_rset);
3056 rs_free(srad);
3057 CLEANUP_THREAD_INFO;
3058 *msg_id = kmp_i18n_str_UnknownTopology;
3059 return false;
3060 }
3061
3062 // Get the maximum number of procs that may be contained in a resource set.
3063 int max_procs = rs_getinfo(NULL, R_MAXPROCS, 0);
3064 if (max_procs < 0) {
3065 rs_free(sys_rset);
3066 rs_free(srad);
3067 CLEANUP_THREAD_INFO;
3068 *msg_id = kmp_i18n_str_UnknownTopology;
3069 return false;
3070 }
3071
3072 int cur_rad = 0;
3073 int num_set = 0;
3074 for (int srad_idx = 0; cur_rad < num_rads && srad_idx < VMI_MAXRADS;
3075 ++srad_idx) {
3076 // Check if the SRAD is available in the RSET.
3077 if (rs_getrad(sys_rset, srad, sradsdl, srad_idx, 0) < 0)
3078 continue;
3079
3080 for (int cpu = 0; cpu < max_procs; cpu++) {
3081 // Set the info for the cpu if it is in the SRAD.
3082 if (rs_op(RS_TESTRESOURCE, srad, NULL, R_PROCS, cpu)) {
3083 threadInfo[cpu][osIdIndex] = cpu;
3084 threadInfo[cpu][pkgIdIndex] = cur_rad;
3085 threadInfo[cpu][coreIdIndex] = cpu / smt_threads;
3086 ++num_set;
3087 if (num_set >= num_avail) {
3088 // Done if all available CPUs have been set.
3089 break;
3090 }
3091 }
3092 }
3093 ++cur_rad;
3094 }
3095 rs_free(sys_rset);
3096 rs_free(srad);
3097
3098 // The topology is already sorted.
3099
3100 #else // !KMP_OS_AIX
3101 unsigned num_avail = 0;
3102 *line = 0;
3103 #if KMP_ARCH_S390X
3104 bool reading_s390x_sys_info = true;
3105 #endif
3106 while (!feof(f)) {
3107 // Create an inner scoping level, so that all the goto targets at the end of
3108 // the loop appear in an outer scoping level. This avoids warnings about
3109 // jumping past an initialization to a target in the same block.
3110 {
3111 buf[sizeof(buf) - 1] = 1;
3112 bool long_line = false;
3113 if (!fgets(buf, sizeof(buf), f)) {
3114 // Read errors presumably because of EOF
3115 // If there is valid data in threadInfo[num_avail], then fake
3116 // a blank line in ensure that the last address gets parsed.
3117 bool valid = false;
3118 for (i = 0; i <= maxIndex; i++) {
3119 if (threadInfo[num_avail][i] != UINT_MAX) {
3120 valid = true;
3121 }
3122 }
3123 if (!valid) {
3124 break;
3125 }
3126 buf[0] = 0;
3127 } else if (!buf[sizeof(buf) - 1]) {
3128 // The line is longer than the buffer. Set a flag and don't
3129 // emit an error if we were going to ignore the line, anyway.
3130 long_line = true;
3131
3132 #define CHECK_LINE \
3133 if (long_line) { \
3134 CLEANUP_THREAD_INFO; \
3135 *msg_id = kmp_i18n_str_LongLineCpuinfo; \
3136 return false; \
3137 }
3138 }
3139 (*line)++;
3140
3141 #if KMP_ARCH_LOONGARCH64
3142 // The parsing logic of /proc/cpuinfo in this function highly depends on
3143 // the blank lines between each processor info block. But on LoongArch a
3144 // blank line exists before the first processor info block (i.e. after the
3145 // "system type" line). This blank line was added because the "system
3146 // type" line is unrelated to any of the CPUs. We must skip this line so
3147 // that the original logic works on LoongArch.
3148 if (*buf == '\n' && *line == 2)
3149 continue;
3150 #endif
3151 #if KMP_ARCH_S390X
3152 // s390x /proc/cpuinfo starts with a variable number of lines containing
3153 // the overall system information. Skip them.
3154 if (reading_s390x_sys_info) {
3155 if (*buf == '\n')
3156 reading_s390x_sys_info = false;
3157 continue;
3158 }
3159 #endif
3160
3161 #if KMP_ARCH_S390X
3162 char s1[] = "cpu number";
3163 #else
3164 char s1[] = "processor";
3165 #endif
3166 if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
3167 CHECK_LINE;
3168 char *p = strchr(buf + sizeof(s1) - 1, ':');
3169 unsigned val;
3170 if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3171 goto no_val;
3172 if (threadInfo[num_avail][osIdIndex] != UINT_MAX)
3173 #if KMP_ARCH_AARCH64
3174 // Handle the old AArch64 /proc/cpuinfo layout differently,
3175 // it contains all of the 'processor' entries listed in a
3176 // single 'Processor' section, therefore the normal looking
3177 // for duplicates in that section will always fail.
3178 num_avail++;
3179 #else
3180 goto dup_field;
3181 #endif
3182 threadInfo[num_avail][osIdIndex] = val;
3183 #if KMP_OS_LINUX && !(KMP_ARCH_X86 || KMP_ARCH_X86_64)
3184 char path[256];
3185 KMP_SNPRINTF(
3186 path, sizeof(path),
3187 "/sys/devices/system/cpu/cpu%u/topology/physical_package_id",
3188 threadInfo[num_avail][osIdIndex]);
3189 __kmp_read_from_file(path, "%u", &threadInfo[num_avail][pkgIdIndex]);
3190
3191 #if KMP_ARCH_S390X
3192 // Disambiguate physical_package_id.
3193 unsigned book_id;
3194 KMP_SNPRINTF(path, sizeof(path),
3195 "/sys/devices/system/cpu/cpu%u/topology/book_id",
3196 threadInfo[num_avail][osIdIndex]);
3197 __kmp_read_from_file(path, "%u", &book_id);
3198 threadInfo[num_avail][pkgIdIndex] |= (book_id << 8);
3199
3200 unsigned drawer_id;
3201 KMP_SNPRINTF(path, sizeof(path),
3202 "/sys/devices/system/cpu/cpu%u/topology/drawer_id",
3203 threadInfo[num_avail][osIdIndex]);
3204 __kmp_read_from_file(path, "%u", &drawer_id);
3205 threadInfo[num_avail][pkgIdIndex] |= (drawer_id << 16);
3206 #endif
3207
3208 KMP_SNPRINTF(path, sizeof(path),
3209 "/sys/devices/system/cpu/cpu%u/topology/core_id",
3210 threadInfo[num_avail][osIdIndex]);
3211 __kmp_read_from_file(path, "%u", &threadInfo[num_avail][coreIdIndex]);
3212 continue;
3213 #else
3214 }
3215 char s2[] = "physical id";
3216 if (strncmp(buf, s2, sizeof(s2) - 1) == 0) {
3217 CHECK_LINE;
3218 char *p = strchr(buf + sizeof(s2) - 1, ':');
3219 unsigned val;
3220 if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3221 goto no_val;
3222 if (threadInfo[num_avail][pkgIdIndex] != UINT_MAX)
3223 goto dup_field;
3224 threadInfo[num_avail][pkgIdIndex] = val;
3225 continue;
3226 }
3227 char s3[] = "core id";
3228 if (strncmp(buf, s3, sizeof(s3) - 1) == 0) {
3229 CHECK_LINE;
3230 char *p = strchr(buf + sizeof(s3) - 1, ':');
3231 unsigned val;
3232 if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3233 goto no_val;
3234 if (threadInfo[num_avail][coreIdIndex] != UINT_MAX)
3235 goto dup_field;
3236 threadInfo[num_avail][coreIdIndex] = val;
3237 continue;
3238 #endif // KMP_OS_LINUX && USE_SYSFS_INFO
3239 }
3240 char s4[] = "thread id";
3241 if (strncmp(buf, s4, sizeof(s4) - 1) == 0) {
3242 CHECK_LINE;
3243 char *p = strchr(buf + sizeof(s4) - 1, ':');
3244 unsigned val;
3245 if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3246 goto no_val;
3247 if (threadInfo[num_avail][threadIdIndex] != UINT_MAX)
3248 goto dup_field;
3249 threadInfo[num_avail][threadIdIndex] = val;
3250 continue;
3251 }
3252 unsigned level;
3253 if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
3254 CHECK_LINE;
3255 char *p = strchr(buf + sizeof(s4) - 1, ':');
3256 unsigned val;
3257 if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3258 goto no_val;
3259 // validate the input before using level:
3260 if (level > (unsigned)__kmp_xproc) { // level is too big
3261 level = __kmp_xproc;
3262 }
3263 if (threadInfo[num_avail][nodeIdIndex + level] != UINT_MAX)
3264 goto dup_field;
3265 threadInfo[num_avail][nodeIdIndex + level] = val;
3266 continue;
3267 }
3268
3269 // We didn't recognize the leading token on the line. There are lots of
3270 // leading tokens that we don't recognize - if the line isn't empty, go on
3271 // to the next line.
3272 if ((*buf != 0) && (*buf != '\n')) {
3273 // If the line is longer than the buffer, read characters
3274 // until we find a newline.
3275 if (long_line) {
3276 int ch;
3277 while (((ch = fgetc(f)) != EOF) && (ch != '\n'))
3278 ;
3279 }
3280 continue;
3281 }
3282
3283 // A newline has signalled the end of the processor record.
3284 // Check that there aren't too many procs specified.
3285 if ((int)num_avail == __kmp_xproc) {
3286 CLEANUP_THREAD_INFO;
3287 *msg_id = kmp_i18n_str_TooManyEntries;
3288 return false;
3289 }
3290
3291 // Check for missing fields. The osId field must be there, and we
3292 // currently require that the physical id field is specified, also.
3293 if (threadInfo[num_avail][osIdIndex] == UINT_MAX) {
3294 CLEANUP_THREAD_INFO;
3295 *msg_id = kmp_i18n_str_MissingProcField;
3296 return false;
3297 }
3298 if (threadInfo[0][pkgIdIndex] == UINT_MAX) {
3299 CLEANUP_THREAD_INFO;
3300 *msg_id = kmp_i18n_str_MissingPhysicalIDField;
3301 return false;
3302 }
3303
3304 // Skip this proc if it is not included in the machine model.
3305 if (KMP_AFFINITY_CAPABLE() &&
3306 !KMP_CPU_ISSET(threadInfo[num_avail][osIdIndex],
3307 __kmp_affin_fullMask)) {
3308 INIT_PROC_INFO(threadInfo[num_avail]);
3309 continue;
3310 }
3311
3312 // We have a successful parse of this proc's info.
3313 // Increment the counter, and prepare for the next proc.
3314 num_avail++;
3315 KMP_ASSERT(num_avail <= num_records);
3316 INIT_PROC_INFO(threadInfo[num_avail]);
3317 }
3318 continue;
3319
3320 no_val:
3321 CLEANUP_THREAD_INFO;
3322 *msg_id = kmp_i18n_str_MissingValCpuinfo;
3323 return false;
3324
3325 dup_field:
3326 CLEANUP_THREAD_INFO;
3327 *msg_id = kmp_i18n_str_DuplicateFieldCpuinfo;
3328 return false;
3329 }
3330 *line = 0;
3331
3332 #if KMP_MIC && REDUCE_TEAM_SIZE
3333 unsigned teamSize = 0;
3334 #endif // KMP_MIC && REDUCE_TEAM_SIZE
3335
3336 // check for num_records == __kmp_xproc ???
3337
3338 // If it is configured to omit the package level when there is only a single
3339 // package, the logic at the end of this routine won't work if there is only a
3340 // single thread
3341 KMP_ASSERT(num_avail > 0);
3342 KMP_ASSERT(num_avail <= num_records);
3343
3344 // Sort the threadInfo table by physical Id.
3345 qsort(threadInfo, num_avail, sizeof(*threadInfo),
3346 __kmp_affinity_cmp_ProcCpuInfo_phys_id);
3347
3348 #endif // KMP_OS_AIX
3349
3350 // The table is now sorted by pkgId / coreId / threadId, but we really don't
3351 // know the radix of any of the fields. pkgId's may be sparsely assigned among
3352 // the chips on a system. Although coreId's are usually assigned
3353 // [0 .. coresPerPkg-1] and threadId's are usually assigned
3354 // [0..threadsPerCore-1], we don't want to make any such assumptions.
3355 //
3356 // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
3357 // total # packages) are at this point - we want to determine that now. We
3358 // only have an upper bound on the first two figures.
3359 unsigned *counts =
3360 (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3361 unsigned *maxCt =
3362 (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3363 unsigned *totals =
3364 (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3365 unsigned *lastId =
3366 (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3367
3368 bool assign_thread_ids = false;
3369 unsigned threadIdCt;
3370 unsigned index;
3371
3372 restart_radix_check:
3373 threadIdCt = 0;
3374
3375 // Initialize the counter arrays with data from threadInfo[0].
3376 if (assign_thread_ids) {
3377 if (threadInfo[0][threadIdIndex] == UINT_MAX) {
3378 threadInfo[0][threadIdIndex] = threadIdCt++;
3379 } else if (threadIdCt <= threadInfo[0][threadIdIndex]) {
3380 threadIdCt = threadInfo[0][threadIdIndex] + 1;
3381 }
3382 }
3383 for (index = 0; index <= maxIndex; index++) {
3384 counts[index] = 1;
3385 maxCt[index] = 1;
3386 totals[index] = 1;
3387 lastId[index] = threadInfo[0][index];
3388 ;
3389 }
3390
3391 // Run through the rest of the OS procs.
3392 for (i = 1; i < num_avail; i++) {
3393 // Find the most significant index whose id differs from the id for the
3394 // previous OS proc.
3395 for (index = maxIndex; index >= threadIdIndex; index--) {
3396 if (assign_thread_ids && (index == threadIdIndex)) {
3397 // Auto-assign the thread id field if it wasn't specified.
3398 if (threadInfo[i][threadIdIndex] == UINT_MAX) {
3399 threadInfo[i][threadIdIndex] = threadIdCt++;
3400 }
3401 // Apparently the thread id field was specified for some entries and not
3402 // others. Start the thread id counter off at the next higher thread id.
3403 else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
3404 threadIdCt = threadInfo[i][threadIdIndex] + 1;
3405 }
3406 }
3407 if (threadInfo[i][index] != lastId[index]) {
3408 // Run through all indices which are less significant, and reset the
3409 // counts to 1. At all levels up to and including index, we need to
3410 // increment the totals and record the last id.
3411 unsigned index2;
3412 for (index2 = threadIdIndex; index2 < index; index2++) {
3413 totals[index2]++;
3414 if (counts[index2] > maxCt[index2]) {
3415 maxCt[index2] = counts[index2];
3416 }
3417 counts[index2] = 1;
3418 lastId[index2] = threadInfo[i][index2];
3419 }
3420 counts[index]++;
3421 totals[index]++;
3422 lastId[index] = threadInfo[i][index];
3423
3424 if (assign_thread_ids && (index > threadIdIndex)) {
3425
3426 #if KMP_MIC && REDUCE_TEAM_SIZE
3427 // The default team size is the total #threads in the machine
3428 // minus 1 thread for every core that has 3 or more threads.
3429 teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
3430 #endif // KMP_MIC && REDUCE_TEAM_SIZE
3431
3432 // Restart the thread counter, as we are on a new core.
3433 threadIdCt = 0;
3434
3435 // Auto-assign the thread id field if it wasn't specified.
3436 if (threadInfo[i][threadIdIndex] == UINT_MAX) {
3437 threadInfo[i][threadIdIndex] = threadIdCt++;
3438 }
3439
3440 // Apparently the thread id field was specified for some entries and
3441 // not others. Start the thread id counter off at the next higher
3442 // thread id.
3443 else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
3444 threadIdCt = threadInfo[i][threadIdIndex] + 1;
3445 }
3446 }
3447 break;
3448 }
3449 }
3450 if (index < threadIdIndex) {
3451 // If thread ids were specified, it is an error if they are not unique.
3452 // Also, check that we waven't already restarted the loop (to be safe -
3453 // shouldn't need to).
3454 if ((threadInfo[i][threadIdIndex] != UINT_MAX) || assign_thread_ids) {
3455 __kmp_free(lastId);
3456 __kmp_free(totals);
3457 __kmp_free(maxCt);
3458 __kmp_free(counts);
3459 CLEANUP_THREAD_INFO;
3460 *msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
3461 return false;
3462 }
3463
3464 // If the thread ids were not specified and we see entries that
3465 // are duplicates, start the loop over and assign the thread ids manually.
3466 assign_thread_ids = true;
3467 goto restart_radix_check;
3468 }
3469 }
3470
3471 #if KMP_MIC && REDUCE_TEAM_SIZE
3472 // The default team size is the total #threads in the machine
3473 // minus 1 thread for every core that has 3 or more threads.
3474 teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
3475 #endif // KMP_MIC && REDUCE_TEAM_SIZE
3476
3477 for (index = threadIdIndex; index <= maxIndex; index++) {
3478 if (counts[index] > maxCt[index]) {
3479 maxCt[index] = counts[index];
3480 }
3481 }
3482
3483 __kmp_nThreadsPerCore = maxCt[threadIdIndex];
3484 nCoresPerPkg = maxCt[coreIdIndex];
3485 nPackages = totals[pkgIdIndex];
3486
3487 // When affinity is off, this routine will still be called to set
3488 // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
3489 // Make sure all these vars are set correctly, and return now if affinity is
3490 // not enabled.
3491 __kmp_ncores = totals[coreIdIndex];
3492 if (!KMP_AFFINITY_CAPABLE()) {
3493 KMP_ASSERT(__kmp_affinity.type == affinity_none);
3494 return true;
3495 }
3496
3497 #if KMP_MIC && REDUCE_TEAM_SIZE
3498 // Set the default team size.
3499 if ((__kmp_dflt_team_nth == 0) && (teamSize > 0)) {
3500 __kmp_dflt_team_nth = teamSize;
3501 KA_TRACE(20, ("__kmp_affinity_create_cpuinfo_map: setting "
3502 "__kmp_dflt_team_nth = %d\n",
3503 __kmp_dflt_team_nth));
3504 }
3505 #endif // KMP_MIC && REDUCE_TEAM_SIZE
3506
3507 KMP_DEBUG_ASSERT(num_avail == (unsigned)__kmp_avail_proc);
3508
3509 // Count the number of levels which have more nodes at that level than at the
3510 // parent's level (with there being an implicit root node of the top level).
3511 // This is equivalent to saying that there is at least one node at this level
3512 // which has a sibling. These levels are in the map, and the package level is
3513 // always in the map.
3514 bool *inMap = (bool *)__kmp_allocate((maxIndex + 1) * sizeof(bool));
3515 for (index = threadIdIndex; index < maxIndex; index++) {
3516 KMP_ASSERT(totals[index] >= totals[index + 1]);
3517 inMap[index] = (totals[index] > totals[index + 1]);
3518 }
3519 inMap[maxIndex] = (totals[maxIndex] > 1);
3520 inMap[pkgIdIndex] = true;
3521 inMap[coreIdIndex] = true;
3522 inMap[threadIdIndex] = true;
3523
3524 int depth = 0;
3525 int idx = 0;
3526 kmp_hw_t types[KMP_HW_LAST];
3527 int pkgLevel = -1;
3528 int coreLevel = -1;
3529 int threadLevel = -1;
3530 for (index = threadIdIndex; index <= maxIndex; index++) {
3531 if (inMap[index]) {
3532 depth++;
3533 }
3534 }
3535 if (inMap[pkgIdIndex]) {
3536 pkgLevel = idx;
3537 types[idx++] = KMP_HW_SOCKET;
3538 }
3539 if (inMap[coreIdIndex]) {
3540 coreLevel = idx;
3541 types[idx++] = KMP_HW_CORE;
3542 }
3543 if (inMap[threadIdIndex]) {
3544 threadLevel = idx;
3545 types[idx++] = KMP_HW_THREAD;
3546 }
3547 KMP_ASSERT(depth > 0);
3548
3549 // Construct the data structure that is to be returned.
3550 __kmp_topology = kmp_topology_t::allocate(num_avail, depth, types);
3551
3552 for (i = 0; i < num_avail; ++i) {
3553 unsigned os = threadInfo[i][osIdIndex];
3554 int src_index;
3555 kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
3556 hw_thread.clear();
3557 hw_thread.os_id = os;
3558
3559 idx = 0;
3560 for (src_index = maxIndex; src_index >= threadIdIndex; src_index--) {
3561 if (!inMap[src_index]) {
3562 continue;
3563 }
3564 if (src_index == pkgIdIndex) {
3565 hw_thread.ids[pkgLevel] = threadInfo[i][src_index];
3566 } else if (src_index == coreIdIndex) {
3567 hw_thread.ids[coreLevel] = threadInfo[i][src_index];
3568 } else if (src_index == threadIdIndex) {
3569 hw_thread.ids[threadLevel] = threadInfo[i][src_index];
3570 }
3571 }
3572 }
3573
3574 __kmp_free(inMap);
3575 __kmp_free(lastId);
3576 __kmp_free(totals);
3577 __kmp_free(maxCt);
3578 __kmp_free(counts);
3579 CLEANUP_THREAD_INFO;
3580 __kmp_topology->sort_ids();
3581 if (!__kmp_topology->check_ids()) {
3582 kmp_topology_t::deallocate(__kmp_topology);
3583 __kmp_topology = nullptr;
3584 *msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
3585 return false;
3586 }
3587 return true;
3588 }
3589
3590 // Create and return a table of affinity masks, indexed by OS thread ID.
3591 // This routine handles OR'ing together all the affinity masks of threads
3592 // that are sufficiently close, if granularity > fine.
3593 template <typename FindNextFunctionType>
__kmp_create_os_id_masks(unsigned * numUnique,kmp_affinity_t & affinity,FindNextFunctionType find_next)3594 static void __kmp_create_os_id_masks(unsigned *numUnique,
3595 kmp_affinity_t &affinity,
3596 FindNextFunctionType find_next) {
3597 // First form a table of affinity masks in order of OS thread id.
3598 int maxOsId;
3599 int i;
3600 int numAddrs = __kmp_topology->get_num_hw_threads();
3601 int depth = __kmp_topology->get_depth();
3602 const char *env_var = __kmp_get_affinity_env_var(affinity);
3603 KMP_ASSERT(numAddrs);
3604 KMP_ASSERT(depth);
3605
3606 i = find_next(-1);
3607 // If could not find HW thread location with attributes, then return and
3608 // fallback to increment find_next and disregard core attributes.
3609 if (i >= numAddrs)
3610 return;
3611
3612 maxOsId = 0;
3613 for (i = numAddrs - 1;; --i) {
3614 int osId = __kmp_topology->at(i).os_id;
3615 if (osId > maxOsId) {
3616 maxOsId = osId;
3617 }
3618 if (i == 0)
3619 break;
3620 }
3621 affinity.num_os_id_masks = maxOsId + 1;
3622 KMP_CPU_ALLOC_ARRAY(affinity.os_id_masks, affinity.num_os_id_masks);
3623 KMP_ASSERT(affinity.gran_levels >= 0);
3624 if (affinity.flags.verbose && (affinity.gran_levels > 0)) {
3625 KMP_INFORM(ThreadsMigrate, env_var, affinity.gran_levels);
3626 }
3627 if (affinity.gran_levels >= (int)depth) {
3628 KMP_AFF_WARNING(affinity, AffThreadsMayMigrate);
3629 }
3630
3631 // Run through the table, forming the masks for all threads on each core.
3632 // Threads on the same core will have identical kmp_hw_thread_t objects, not
3633 // considering the last level, which must be the thread id. All threads on a
3634 // core will appear consecutively.
3635 int unique = 0;
3636 int j = 0; // index of 1st thread on core
3637 int leader = 0;
3638 kmp_affin_mask_t *sum;
3639 KMP_CPU_ALLOC_ON_STACK(sum);
3640 KMP_CPU_ZERO(sum);
3641
3642 i = j = leader = find_next(-1);
3643 KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
3644 kmp_full_mask_modifier_t full_mask;
3645 for (i = find_next(i); i < numAddrs; i = find_next(i)) {
3646 // If this thread is sufficiently close to the leader (within the
3647 // granularity setting), then set the bit for this os thread in the
3648 // affinity mask for this group, and go on to the next thread.
3649 if (__kmp_topology->is_close(leader, i, affinity)) {
3650 KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
3651 continue;
3652 }
3653
3654 // For every thread in this group, copy the mask to the thread's entry in
3655 // the OS Id mask table. Mark the first address as a leader.
3656 for (; j < i; j = find_next(j)) {
3657 int osId = __kmp_topology->at(j).os_id;
3658 KMP_DEBUG_ASSERT(osId <= maxOsId);
3659 kmp_affin_mask_t *mask = KMP_CPU_INDEX(affinity.os_id_masks, osId);
3660 KMP_CPU_COPY(mask, sum);
3661 __kmp_topology->at(j).leader = (j == leader);
3662 }
3663 unique++;
3664
3665 // Start a new mask.
3666 leader = i;
3667 full_mask.include(sum);
3668 KMP_CPU_ZERO(sum);
3669 KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
3670 }
3671
3672 // For every thread in last group, copy the mask to the thread's
3673 // entry in the OS Id mask table.
3674 for (; j < i; j = find_next(j)) {
3675 int osId = __kmp_topology->at(j).os_id;
3676 KMP_DEBUG_ASSERT(osId <= maxOsId);
3677 kmp_affin_mask_t *mask = KMP_CPU_INDEX(affinity.os_id_masks, osId);
3678 KMP_CPU_COPY(mask, sum);
3679 __kmp_topology->at(j).leader = (j == leader);
3680 }
3681 full_mask.include(sum);
3682 unique++;
3683 KMP_CPU_FREE_FROM_STACK(sum);
3684
3685 // See if the OS Id mask table further restricts or changes the full mask
3686 if (full_mask.restrict_to_mask() && affinity.flags.verbose) {
3687 __kmp_topology->print(env_var);
3688 }
3689
3690 *numUnique = unique;
3691 }
3692
3693 // Stuff for the affinity proclist parsers. It's easier to declare these vars
3694 // as file-static than to try and pass them through the calling sequence of
3695 // the recursive-descent OMP_PLACES parser.
3696 static kmp_affin_mask_t *newMasks;
3697 static int numNewMasks;
3698 static int nextNewMask;
3699
3700 #define ADD_MASK(_mask) \
3701 { \
3702 if (nextNewMask >= numNewMasks) { \
3703 int i; \
3704 numNewMasks *= 2; \
3705 kmp_affin_mask_t *temp; \
3706 KMP_CPU_INTERNAL_ALLOC_ARRAY(temp, numNewMasks); \
3707 for (i = 0; i < numNewMasks / 2; i++) { \
3708 kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i); \
3709 kmp_affin_mask_t *dest = KMP_CPU_INDEX(temp, i); \
3710 KMP_CPU_COPY(dest, src); \
3711 } \
3712 KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks / 2); \
3713 newMasks = temp; \
3714 } \
3715 KMP_CPU_COPY(KMP_CPU_INDEX(newMasks, nextNewMask), (_mask)); \
3716 nextNewMask++; \
3717 }
3718
3719 #define ADD_MASK_OSID(_osId, _osId2Mask, _maxOsId) \
3720 { \
3721 if (((_osId) > _maxOsId) || \
3722 (!KMP_CPU_ISSET((_osId), KMP_CPU_INDEX((_osId2Mask), (_osId))))) { \
3723 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, _osId); \
3724 } else { \
3725 ADD_MASK(KMP_CPU_INDEX(_osId2Mask, (_osId))); \
3726 } \
3727 }
3728
3729 // Re-parse the proclist (for the explicit affinity type), and form the list
3730 // of affinity newMasks indexed by gtid.
__kmp_affinity_process_proclist(kmp_affinity_t & affinity)3731 static void __kmp_affinity_process_proclist(kmp_affinity_t &affinity) {
3732 int i;
3733 kmp_affin_mask_t **out_masks = &affinity.masks;
3734 unsigned *out_numMasks = &affinity.num_masks;
3735 const char *proclist = affinity.proclist;
3736 kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
3737 int maxOsId = affinity.num_os_id_masks - 1;
3738 const char *scan = proclist;
3739 const char *next = proclist;
3740
3741 // We use malloc() for the temporary mask vector, so that we can use
3742 // realloc() to extend it.
3743 numNewMasks = 2;
3744 KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
3745 nextNewMask = 0;
3746 kmp_affin_mask_t *sumMask;
3747 KMP_CPU_ALLOC(sumMask);
3748 int setSize = 0;
3749
3750 for (;;) {
3751 int start, end, stride;
3752
3753 SKIP_WS(scan);
3754 next = scan;
3755 if (*next == '\0') {
3756 break;
3757 }
3758
3759 if (*next == '{') {
3760 int num;
3761 setSize = 0;
3762 next++; // skip '{'
3763 SKIP_WS(next);
3764 scan = next;
3765
3766 // Read the first integer in the set.
3767 KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad proclist");
3768 SKIP_DIGITS(next);
3769 num = __kmp_str_to_int(scan, *next);
3770 KMP_ASSERT2(num >= 0, "bad explicit proc list");
3771
3772 // Copy the mask for that osId to the sum (union) mask.
3773 if ((num > maxOsId) ||
3774 (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
3775 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, num);
3776 KMP_CPU_ZERO(sumMask);
3777 } else {
3778 KMP_CPU_COPY(sumMask, KMP_CPU_INDEX(osId2Mask, num));
3779 setSize = 1;
3780 }
3781
3782 for (;;) {
3783 // Check for end of set.
3784 SKIP_WS(next);
3785 if (*next == '}') {
3786 next++; // skip '}'
3787 break;
3788 }
3789
3790 // Skip optional comma.
3791 if (*next == ',') {
3792 next++;
3793 }
3794 SKIP_WS(next);
3795
3796 // Read the next integer in the set.
3797 scan = next;
3798 KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
3799
3800 SKIP_DIGITS(next);
3801 num = __kmp_str_to_int(scan, *next);
3802 KMP_ASSERT2(num >= 0, "bad explicit proc list");
3803
3804 // Add the mask for that osId to the sum mask.
3805 if ((num > maxOsId) ||
3806 (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
3807 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, num);
3808 } else {
3809 KMP_CPU_UNION(sumMask, KMP_CPU_INDEX(osId2Mask, num));
3810 setSize++;
3811 }
3812 }
3813 if (setSize > 0) {
3814 ADD_MASK(sumMask);
3815 }
3816
3817 SKIP_WS(next);
3818 if (*next == ',') {
3819 next++;
3820 }
3821 scan = next;
3822 continue;
3823 }
3824
3825 // Read the first integer.
3826 KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
3827 SKIP_DIGITS(next);
3828 start = __kmp_str_to_int(scan, *next);
3829 KMP_ASSERT2(start >= 0, "bad explicit proc list");
3830 SKIP_WS(next);
3831
3832 // If this isn't a range, then add a mask to the list and go on.
3833 if (*next != '-') {
3834 ADD_MASK_OSID(start, osId2Mask, maxOsId);
3835
3836 // Skip optional comma.
3837 if (*next == ',') {
3838 next++;
3839 }
3840 scan = next;
3841 continue;
3842 }
3843
3844 // This is a range. Skip over the '-' and read in the 2nd int.
3845 next++; // skip '-'
3846 SKIP_WS(next);
3847 scan = next;
3848 KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
3849 SKIP_DIGITS(next);
3850 end = __kmp_str_to_int(scan, *next);
3851 KMP_ASSERT2(end >= 0, "bad explicit proc list");
3852
3853 // Check for a stride parameter
3854 stride = 1;
3855 SKIP_WS(next);
3856 if (*next == ':') {
3857 // A stride is specified. Skip over the ':" and read the 3rd int.
3858 int sign = +1;
3859 next++; // skip ':'
3860 SKIP_WS(next);
3861 scan = next;
3862 if (*next == '-') {
3863 sign = -1;
3864 next++;
3865 SKIP_WS(next);
3866 scan = next;
3867 }
3868 KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
3869 SKIP_DIGITS(next);
3870 stride = __kmp_str_to_int(scan, *next);
3871 KMP_ASSERT2(stride >= 0, "bad explicit proc list");
3872 stride *= sign;
3873 }
3874
3875 // Do some range checks.
3876 KMP_ASSERT2(stride != 0, "bad explicit proc list");
3877 if (stride > 0) {
3878 KMP_ASSERT2(start <= end, "bad explicit proc list");
3879 } else {
3880 KMP_ASSERT2(start >= end, "bad explicit proc list");
3881 }
3882 KMP_ASSERT2((end - start) / stride <= 65536, "bad explicit proc list");
3883
3884 // Add the mask for each OS proc # to the list.
3885 if (stride > 0) {
3886 do {
3887 ADD_MASK_OSID(start, osId2Mask, maxOsId);
3888 start += stride;
3889 } while (start <= end);
3890 } else {
3891 do {
3892 ADD_MASK_OSID(start, osId2Mask, maxOsId);
3893 start += stride;
3894 } while (start >= end);
3895 }
3896
3897 // Skip optional comma.
3898 SKIP_WS(next);
3899 if (*next == ',') {
3900 next++;
3901 }
3902 scan = next;
3903 }
3904
3905 *out_numMasks = nextNewMask;
3906 if (nextNewMask == 0) {
3907 *out_masks = NULL;
3908 KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3909 return;
3910 }
3911 KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
3912 for (i = 0; i < nextNewMask; i++) {
3913 kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
3914 kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
3915 KMP_CPU_COPY(dest, src);
3916 }
3917 KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3918 KMP_CPU_FREE(sumMask);
3919 }
3920
3921 /*-----------------------------------------------------------------------------
3922 Re-parse the OMP_PLACES proc id list, forming the newMasks for the different
3923 places. Again, Here is the grammar:
3924
3925 place_list := place
3926 place_list := place , place_list
3927 place := num
3928 place := place : num
3929 place := place : num : signed
3930 place := { subplacelist }
3931 place := ! place // (lowest priority)
3932 subplace_list := subplace
3933 subplace_list := subplace , subplace_list
3934 subplace := num
3935 subplace := num : num
3936 subplace := num : num : signed
3937 signed := num
3938 signed := + signed
3939 signed := - signed
3940 -----------------------------------------------------------------------------*/
__kmp_process_subplace_list(const char ** scan,kmp_affinity_t & affinity,int maxOsId,kmp_affin_mask_t * tempMask,int * setSize)3941 static void __kmp_process_subplace_list(const char **scan,
3942 kmp_affinity_t &affinity, int maxOsId,
3943 kmp_affin_mask_t *tempMask,
3944 int *setSize) {
3945 const char *next;
3946 kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
3947
3948 for (;;) {
3949 int start, count, stride, i;
3950
3951 // Read in the starting proc id
3952 SKIP_WS(*scan);
3953 KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3954 next = *scan;
3955 SKIP_DIGITS(next);
3956 start = __kmp_str_to_int(*scan, *next);
3957 KMP_ASSERT(start >= 0);
3958 *scan = next;
3959
3960 // valid follow sets are ',' ':' and '}'
3961 SKIP_WS(*scan);
3962 if (**scan == '}' || **scan == ',') {
3963 if ((start > maxOsId) ||
3964 (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3965 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, start);
3966 } else {
3967 KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3968 (*setSize)++;
3969 }
3970 if (**scan == '}') {
3971 break;
3972 }
3973 (*scan)++; // skip ','
3974 continue;
3975 }
3976 KMP_ASSERT2(**scan == ':', "bad explicit places list");
3977 (*scan)++; // skip ':'
3978
3979 // Read count parameter
3980 SKIP_WS(*scan);
3981 KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3982 next = *scan;
3983 SKIP_DIGITS(next);
3984 count = __kmp_str_to_int(*scan, *next);
3985 KMP_ASSERT(count >= 0);
3986 *scan = next;
3987
3988 // valid follow sets are ',' ':' and '}'
3989 SKIP_WS(*scan);
3990 if (**scan == '}' || **scan == ',') {
3991 for (i = 0; i < count; i++) {
3992 if ((start > maxOsId) ||
3993 (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3994 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, start);
3995 break; // don't proliferate warnings for large count
3996 } else {
3997 KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3998 start++;
3999 (*setSize)++;
4000 }
4001 }
4002 if (**scan == '}') {
4003 break;
4004 }
4005 (*scan)++; // skip ','
4006 continue;
4007 }
4008 KMP_ASSERT2(**scan == ':', "bad explicit places list");
4009 (*scan)++; // skip ':'
4010
4011 // Read stride parameter
4012 int sign = +1;
4013 for (;;) {
4014 SKIP_WS(*scan);
4015 if (**scan == '+') {
4016 (*scan)++; // skip '+'
4017 continue;
4018 }
4019 if (**scan == '-') {
4020 sign *= -1;
4021 (*scan)++; // skip '-'
4022 continue;
4023 }
4024 break;
4025 }
4026 SKIP_WS(*scan);
4027 KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
4028 next = *scan;
4029 SKIP_DIGITS(next);
4030 stride = __kmp_str_to_int(*scan, *next);
4031 KMP_ASSERT(stride >= 0);
4032 *scan = next;
4033 stride *= sign;
4034
4035 // valid follow sets are ',' and '}'
4036 SKIP_WS(*scan);
4037 if (**scan == '}' || **scan == ',') {
4038 for (i = 0; i < count; i++) {
4039 if ((start > maxOsId) ||
4040 (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
4041 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, start);
4042 break; // don't proliferate warnings for large count
4043 } else {
4044 KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
4045 start += stride;
4046 (*setSize)++;
4047 }
4048 }
4049 if (**scan == '}') {
4050 break;
4051 }
4052 (*scan)++; // skip ','
4053 continue;
4054 }
4055
4056 KMP_ASSERT2(0, "bad explicit places list");
4057 }
4058 }
4059
__kmp_process_place(const char ** scan,kmp_affinity_t & affinity,int maxOsId,kmp_affin_mask_t * tempMask,int * setSize)4060 static void __kmp_process_place(const char **scan, kmp_affinity_t &affinity,
4061 int maxOsId, kmp_affin_mask_t *tempMask,
4062 int *setSize) {
4063 const char *next;
4064 kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
4065
4066 // valid follow sets are '{' '!' and num
4067 SKIP_WS(*scan);
4068 if (**scan == '{') {
4069 (*scan)++; // skip '{'
4070 __kmp_process_subplace_list(scan, affinity, maxOsId, tempMask, setSize);
4071 KMP_ASSERT2(**scan == '}', "bad explicit places list");
4072 (*scan)++; // skip '}'
4073 } else if (**scan == '!') {
4074 (*scan)++; // skip '!'
4075 __kmp_process_place(scan, affinity, maxOsId, tempMask, setSize);
4076 KMP_CPU_COMPLEMENT(maxOsId, tempMask);
4077 } else if ((**scan >= '0') && (**scan <= '9')) {
4078 next = *scan;
4079 SKIP_DIGITS(next);
4080 int num = __kmp_str_to_int(*scan, *next);
4081 KMP_ASSERT(num >= 0);
4082 if ((num > maxOsId) ||
4083 (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
4084 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, num);
4085 } else {
4086 KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, num));
4087 (*setSize)++;
4088 }
4089 *scan = next; // skip num
4090 } else {
4091 KMP_ASSERT2(0, "bad explicit places list");
4092 }
4093 }
4094
4095 // static void
__kmp_affinity_process_placelist(kmp_affinity_t & affinity)4096 void __kmp_affinity_process_placelist(kmp_affinity_t &affinity) {
4097 int i, j, count, stride, sign;
4098 kmp_affin_mask_t **out_masks = &affinity.masks;
4099 unsigned *out_numMasks = &affinity.num_masks;
4100 const char *placelist = affinity.proclist;
4101 kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
4102 int maxOsId = affinity.num_os_id_masks - 1;
4103 const char *scan = placelist;
4104 const char *next = placelist;
4105
4106 numNewMasks = 2;
4107 KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
4108 nextNewMask = 0;
4109
4110 // tempMask is modified based on the previous or initial
4111 // place to form the current place
4112 // previousMask contains the previous place
4113 kmp_affin_mask_t *tempMask;
4114 kmp_affin_mask_t *previousMask;
4115 KMP_CPU_ALLOC(tempMask);
4116 KMP_CPU_ZERO(tempMask);
4117 KMP_CPU_ALLOC(previousMask);
4118 KMP_CPU_ZERO(previousMask);
4119 int setSize = 0;
4120
4121 for (;;) {
4122 __kmp_process_place(&scan, affinity, maxOsId, tempMask, &setSize);
4123
4124 // valid follow sets are ',' ':' and EOL
4125 SKIP_WS(scan);
4126 if (*scan == '\0' || *scan == ',') {
4127 if (setSize > 0) {
4128 ADD_MASK(tempMask);
4129 }
4130 KMP_CPU_ZERO(tempMask);
4131 setSize = 0;
4132 if (*scan == '\0') {
4133 break;
4134 }
4135 scan++; // skip ','
4136 continue;
4137 }
4138
4139 KMP_ASSERT2(*scan == ':', "bad explicit places list");
4140 scan++; // skip ':'
4141
4142 // Read count parameter
4143 SKIP_WS(scan);
4144 KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
4145 next = scan;
4146 SKIP_DIGITS(next);
4147 count = __kmp_str_to_int(scan, *next);
4148 KMP_ASSERT(count >= 0);
4149 scan = next;
4150
4151 // valid follow sets are ',' ':' and EOL
4152 SKIP_WS(scan);
4153 if (*scan == '\0' || *scan == ',') {
4154 stride = +1;
4155 } else {
4156 KMP_ASSERT2(*scan == ':', "bad explicit places list");
4157 scan++; // skip ':'
4158
4159 // Read stride parameter
4160 sign = +1;
4161 for (;;) {
4162 SKIP_WS(scan);
4163 if (*scan == '+') {
4164 scan++; // skip '+'
4165 continue;
4166 }
4167 if (*scan == '-') {
4168 sign *= -1;
4169 scan++; // skip '-'
4170 continue;
4171 }
4172 break;
4173 }
4174 SKIP_WS(scan);
4175 KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
4176 next = scan;
4177 SKIP_DIGITS(next);
4178 stride = __kmp_str_to_int(scan, *next);
4179 KMP_DEBUG_ASSERT(stride >= 0);
4180 scan = next;
4181 stride *= sign;
4182 }
4183
4184 // Add places determined by initial_place : count : stride
4185 for (i = 0; i < count; i++) {
4186 if (setSize == 0) {
4187 break;
4188 }
4189 // Add the current place, then build the next place (tempMask) from that
4190 KMP_CPU_COPY(previousMask, tempMask);
4191 ADD_MASK(previousMask);
4192 KMP_CPU_ZERO(tempMask);
4193 setSize = 0;
4194 KMP_CPU_SET_ITERATE(j, previousMask) {
4195 if (!KMP_CPU_ISSET(j, previousMask)) {
4196 continue;
4197 }
4198 if ((j + stride > maxOsId) || (j + stride < 0) ||
4199 (!KMP_CPU_ISSET(j, __kmp_affin_fullMask)) ||
4200 (!KMP_CPU_ISSET(j + stride,
4201 KMP_CPU_INDEX(osId2Mask, j + stride)))) {
4202 if (i < count - 1) {
4203 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, j + stride);
4204 }
4205 continue;
4206 }
4207 KMP_CPU_SET(j + stride, tempMask);
4208 setSize++;
4209 }
4210 }
4211 KMP_CPU_ZERO(tempMask);
4212 setSize = 0;
4213
4214 // valid follow sets are ',' and EOL
4215 SKIP_WS(scan);
4216 if (*scan == '\0') {
4217 break;
4218 }
4219 if (*scan == ',') {
4220 scan++; // skip ','
4221 continue;
4222 }
4223
4224 KMP_ASSERT2(0, "bad explicit places list");
4225 }
4226
4227 *out_numMasks = nextNewMask;
4228 if (nextNewMask == 0) {
4229 *out_masks = NULL;
4230 KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
4231 return;
4232 }
4233 KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
4234 KMP_CPU_FREE(tempMask);
4235 KMP_CPU_FREE(previousMask);
4236 for (i = 0; i < nextNewMask; i++) {
4237 kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
4238 kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
4239 KMP_CPU_COPY(dest, src);
4240 }
4241 KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
4242 }
4243
4244 #undef ADD_MASK
4245 #undef ADD_MASK_OSID
4246
4247 // This function figures out the deepest level at which there is at least one
4248 // cluster/core with more than one processing unit bound to it.
__kmp_affinity_find_core_level(int nprocs,int bottom_level)4249 static int __kmp_affinity_find_core_level(int nprocs, int bottom_level) {
4250 int core_level = 0;
4251
4252 for (int i = 0; i < nprocs; i++) {
4253 const kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
4254 for (int j = bottom_level; j > 0; j--) {
4255 if (hw_thread.ids[j] > 0) {
4256 if (core_level < (j - 1)) {
4257 core_level = j - 1;
4258 }
4259 }
4260 }
4261 }
4262 return core_level;
4263 }
4264
4265 // This function counts number of clusters/cores at given level.
__kmp_affinity_compute_ncores(int nprocs,int bottom_level,int core_level)4266 static int __kmp_affinity_compute_ncores(int nprocs, int bottom_level,
4267 int core_level) {
4268 return __kmp_topology->get_count(core_level);
4269 }
4270 // This function finds to which cluster/core given processing unit is bound.
__kmp_affinity_find_core(int proc,int bottom_level,int core_level)4271 static int __kmp_affinity_find_core(int proc, int bottom_level,
4272 int core_level) {
4273 int core = 0;
4274 KMP_DEBUG_ASSERT(proc >= 0 && proc < __kmp_topology->get_num_hw_threads());
4275 for (int i = 0; i <= proc; ++i) {
4276 if (i + 1 <= proc) {
4277 for (int j = 0; j <= core_level; ++j) {
4278 if (__kmp_topology->at(i + 1).sub_ids[j] !=
4279 __kmp_topology->at(i).sub_ids[j]) {
4280 core++;
4281 break;
4282 }
4283 }
4284 }
4285 }
4286 return core;
4287 }
4288
4289 // This function finds maximal number of processing units bound to a
4290 // cluster/core at given level.
__kmp_affinity_max_proc_per_core(int nprocs,int bottom_level,int core_level)4291 static int __kmp_affinity_max_proc_per_core(int nprocs, int bottom_level,
4292 int core_level) {
4293 if (core_level >= bottom_level)
4294 return 1;
4295 int thread_level = __kmp_topology->get_level(KMP_HW_THREAD);
4296 return __kmp_topology->calculate_ratio(thread_level, core_level);
4297 }
4298
4299 static int *procarr = NULL;
4300 static int __kmp_aff_depth = 0;
4301 static int *__kmp_osid_to_hwthread_map = NULL;
4302
__kmp_affinity_get_mask_topology_info(const kmp_affin_mask_t * mask,kmp_affinity_ids_t & ids,kmp_affinity_attrs_t & attrs)4303 static void __kmp_affinity_get_mask_topology_info(const kmp_affin_mask_t *mask,
4304 kmp_affinity_ids_t &ids,
4305 kmp_affinity_attrs_t &attrs) {
4306 if (!KMP_AFFINITY_CAPABLE())
4307 return;
4308
4309 // Initiailze ids and attrs thread data
4310 for (int i = 0; i < KMP_HW_LAST; ++i)
4311 ids.ids[i] = kmp_hw_thread_t::UNKNOWN_ID;
4312 attrs = KMP_AFFINITY_ATTRS_UNKNOWN;
4313
4314 // Iterate through each os id within the mask and determine
4315 // the topology id and attribute information
4316 int cpu;
4317 int depth = __kmp_topology->get_depth();
4318 KMP_CPU_SET_ITERATE(cpu, mask) {
4319 int osid_idx = __kmp_osid_to_hwthread_map[cpu];
4320 ids.os_id = cpu;
4321 const kmp_hw_thread_t &hw_thread = __kmp_topology->at(osid_idx);
4322 for (int level = 0; level < depth; ++level) {
4323 kmp_hw_t type = __kmp_topology->get_type(level);
4324 int id = hw_thread.sub_ids[level];
4325 if (ids.ids[type] == kmp_hw_thread_t::UNKNOWN_ID || ids.ids[type] == id) {
4326 ids.ids[type] = id;
4327 } else {
4328 // This mask spans across multiple topology units, set it as such
4329 // and mark every level below as such as well.
4330 ids.ids[type] = kmp_hw_thread_t::MULTIPLE_ID;
4331 for (; level < depth; ++level) {
4332 kmp_hw_t type = __kmp_topology->get_type(level);
4333 ids.ids[type] = kmp_hw_thread_t::MULTIPLE_ID;
4334 }
4335 }
4336 }
4337 if (!attrs.valid) {
4338 attrs.core_type = hw_thread.attrs.get_core_type();
4339 attrs.core_eff = hw_thread.attrs.get_core_eff();
4340 attrs.valid = 1;
4341 } else {
4342 // This mask spans across multiple attributes, set it as such
4343 if (attrs.core_type != hw_thread.attrs.get_core_type())
4344 attrs.core_type = KMP_HW_CORE_TYPE_UNKNOWN;
4345 if (attrs.core_eff != hw_thread.attrs.get_core_eff())
4346 attrs.core_eff = kmp_hw_attr_t::UNKNOWN_CORE_EFF;
4347 }
4348 }
4349 }
4350
__kmp_affinity_get_thread_topology_info(kmp_info_t * th)4351 static void __kmp_affinity_get_thread_topology_info(kmp_info_t *th) {
4352 if (!KMP_AFFINITY_CAPABLE())
4353 return;
4354 const kmp_affin_mask_t *mask = th->th.th_affin_mask;
4355 kmp_affinity_ids_t &ids = th->th.th_topology_ids;
4356 kmp_affinity_attrs_t &attrs = th->th.th_topology_attrs;
4357 __kmp_affinity_get_mask_topology_info(mask, ids, attrs);
4358 }
4359
4360 // Assign the topology information to each place in the place list
4361 // A thread can then grab not only its affinity mask, but the topology
4362 // information associated with that mask. e.g., Which socket is a thread on
__kmp_affinity_get_topology_info(kmp_affinity_t & affinity)4363 static void __kmp_affinity_get_topology_info(kmp_affinity_t &affinity) {
4364 if (!KMP_AFFINITY_CAPABLE())
4365 return;
4366 if (affinity.type != affinity_none) {
4367 KMP_ASSERT(affinity.num_os_id_masks);
4368 KMP_ASSERT(affinity.os_id_masks);
4369 }
4370 KMP_ASSERT(affinity.num_masks);
4371 KMP_ASSERT(affinity.masks);
4372 KMP_ASSERT(__kmp_affin_fullMask);
4373
4374 int max_cpu = __kmp_affin_fullMask->get_max_cpu();
4375 int num_hw_threads = __kmp_topology->get_num_hw_threads();
4376
4377 // Allocate thread topology information
4378 if (!affinity.ids) {
4379 affinity.ids = (kmp_affinity_ids_t *)__kmp_allocate(
4380 sizeof(kmp_affinity_ids_t) * affinity.num_masks);
4381 }
4382 if (!affinity.attrs) {
4383 affinity.attrs = (kmp_affinity_attrs_t *)__kmp_allocate(
4384 sizeof(kmp_affinity_attrs_t) * affinity.num_masks);
4385 }
4386 if (!__kmp_osid_to_hwthread_map) {
4387 // Want the +1 because max_cpu should be valid index into map
4388 __kmp_osid_to_hwthread_map =
4389 (int *)__kmp_allocate(sizeof(int) * (max_cpu + 1));
4390 }
4391
4392 // Create the OS proc to hardware thread map
4393 for (int hw_thread = 0; hw_thread < num_hw_threads; ++hw_thread) {
4394 int os_id = __kmp_topology->at(hw_thread).os_id;
4395 if (KMP_CPU_ISSET(os_id, __kmp_affin_fullMask))
4396 __kmp_osid_to_hwthread_map[os_id] = hw_thread;
4397 }
4398
4399 for (unsigned i = 0; i < affinity.num_masks; ++i) {
4400 kmp_affinity_ids_t &ids = affinity.ids[i];
4401 kmp_affinity_attrs_t &attrs = affinity.attrs[i];
4402 kmp_affin_mask_t *mask = KMP_CPU_INDEX(affinity.masks, i);
4403 __kmp_affinity_get_mask_topology_info(mask, ids, attrs);
4404 }
4405 }
4406
4407 // Called when __kmp_topology is ready
__kmp_aux_affinity_initialize_other_data(kmp_affinity_t & affinity)4408 static void __kmp_aux_affinity_initialize_other_data(kmp_affinity_t &affinity) {
4409 // Initialize other data structures which depend on the topology
4410 if (__kmp_topology && __kmp_topology->get_num_hw_threads()) {
4411 machine_hierarchy.init(__kmp_topology->get_num_hw_threads());
4412 __kmp_affinity_get_topology_info(affinity);
4413 #if KMP_WEIGHTED_ITERATIONS_SUPPORTED
4414 __kmp_first_osid_with_ecore = __kmp_get_first_osid_with_ecore();
4415 #endif
4416 }
4417 }
4418
4419 // Create a one element mask array (set of places) which only contains the
4420 // initial process's affinity mask
__kmp_create_affinity_none_places(kmp_affinity_t & affinity)4421 static void __kmp_create_affinity_none_places(kmp_affinity_t &affinity) {
4422 KMP_ASSERT(__kmp_affin_fullMask != NULL);
4423 KMP_ASSERT(affinity.type == affinity_none);
4424 KMP_ASSERT(__kmp_avail_proc == __kmp_topology->get_num_hw_threads());
4425 affinity.num_masks = 1;
4426 KMP_CPU_ALLOC_ARRAY(affinity.masks, affinity.num_masks);
4427 kmp_affin_mask_t *dest = KMP_CPU_INDEX(affinity.masks, 0);
4428 KMP_CPU_COPY(dest, __kmp_affin_fullMask);
4429 __kmp_aux_affinity_initialize_other_data(affinity);
4430 }
4431
__kmp_aux_affinity_initialize_masks(kmp_affinity_t & affinity)4432 static void __kmp_aux_affinity_initialize_masks(kmp_affinity_t &affinity) {
4433 // Create the "full" mask - this defines all of the processors that we
4434 // consider to be in the machine model. If respect is set, then it is the
4435 // initialization thread's affinity mask. Otherwise, it is all processors that
4436 // we know about on the machine.
4437 int verbose = affinity.flags.verbose;
4438 const char *env_var = affinity.env_var;
4439
4440 // Already initialized
4441 if (__kmp_affin_fullMask && __kmp_affin_origMask)
4442 return;
4443
4444 if (__kmp_affin_fullMask == NULL) {
4445 KMP_CPU_ALLOC(__kmp_affin_fullMask);
4446 }
4447 if (__kmp_affin_origMask == NULL) {
4448 KMP_CPU_ALLOC(__kmp_affin_origMask);
4449 }
4450 if (KMP_AFFINITY_CAPABLE()) {
4451 __kmp_get_system_affinity(__kmp_affin_fullMask, TRUE);
4452 // Make a copy before possible expanding to the entire machine mask
4453 __kmp_affin_origMask->copy(__kmp_affin_fullMask);
4454 if (affinity.flags.respect) {
4455 // Count the number of available processors.
4456 unsigned i;
4457 __kmp_avail_proc = 0;
4458 KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
4459 if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
4460 continue;
4461 }
4462 __kmp_avail_proc++;
4463 }
4464 if (__kmp_avail_proc > __kmp_xproc) {
4465 KMP_AFF_WARNING(affinity, ErrorInitializeAffinity);
4466 affinity.type = affinity_none;
4467 KMP_AFFINITY_DISABLE();
4468 return;
4469 }
4470
4471 if (verbose) {
4472 char buf[KMP_AFFIN_MASK_PRINT_LEN];
4473 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4474 __kmp_affin_fullMask);
4475 KMP_INFORM(InitOSProcSetRespect, env_var, buf);
4476 }
4477 } else {
4478 if (verbose) {
4479 char buf[KMP_AFFIN_MASK_PRINT_LEN];
4480 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4481 __kmp_affin_fullMask);
4482 KMP_INFORM(InitOSProcSetNotRespect, env_var, buf);
4483 }
4484 __kmp_avail_proc =
4485 __kmp_affinity_entire_machine_mask(__kmp_affin_fullMask);
4486 #if KMP_OS_WINDOWS
4487 if (__kmp_num_proc_groups <= 1) {
4488 // Copy expanded full mask if topology has single processor group
4489 __kmp_affin_origMask->copy(__kmp_affin_fullMask);
4490 }
4491 // Set the process affinity mask since threads' affinity
4492 // masks must be subset of process mask in Windows* OS
4493 __kmp_affin_fullMask->set_process_affinity(true);
4494 #endif
4495 }
4496 }
4497 }
4498
__kmp_aux_affinity_initialize_topology(kmp_affinity_t & affinity)4499 static bool __kmp_aux_affinity_initialize_topology(kmp_affinity_t &affinity) {
4500 bool success = false;
4501 const char *env_var = affinity.env_var;
4502 kmp_i18n_id_t msg_id = kmp_i18n_null;
4503 int verbose = affinity.flags.verbose;
4504
4505 // For backward compatibility, setting KMP_CPUINFO_FILE =>
4506 // KMP_TOPOLOGY_METHOD=cpuinfo
4507 if ((__kmp_cpuinfo_file != NULL) &&
4508 (__kmp_affinity_top_method == affinity_top_method_all)) {
4509 __kmp_affinity_top_method = affinity_top_method_cpuinfo;
4510 }
4511
4512 if (__kmp_affinity_top_method == affinity_top_method_all) {
4513 // In the default code path, errors are not fatal - we just try using
4514 // another method. We only emit a warning message if affinity is on, or the
4515 // verbose flag is set, an the nowarnings flag was not set.
4516 #if KMP_USE_HWLOC
4517 if (!success &&
4518 __kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC) {
4519 if (!__kmp_hwloc_error) {
4520 success = __kmp_affinity_create_hwloc_map(&msg_id);
4521 if (!success && verbose) {
4522 KMP_INFORM(AffIgnoringHwloc, env_var);
4523 }
4524 } else if (verbose) {
4525 KMP_INFORM(AffIgnoringHwloc, env_var);
4526 }
4527 }
4528 #endif
4529
4530 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4531 if (!success) {
4532 success = __kmp_affinity_create_x2apicid_map(&msg_id);
4533 if (!success && verbose && msg_id != kmp_i18n_null) {
4534 KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4535 }
4536 }
4537 if (!success) {
4538 success = __kmp_affinity_create_apicid_map(&msg_id);
4539 if (!success && verbose && msg_id != kmp_i18n_null) {
4540 KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4541 }
4542 }
4543 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4544
4545 #if KMP_OS_LINUX || KMP_OS_AIX
4546 if (!success) {
4547 int line = 0;
4548 success = __kmp_affinity_create_cpuinfo_map(&line, &msg_id);
4549 if (!success && verbose && msg_id != kmp_i18n_null) {
4550 KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4551 }
4552 }
4553 #endif /* KMP_OS_LINUX */
4554
4555 #if KMP_GROUP_AFFINITY
4556 if (!success && (__kmp_num_proc_groups > 1)) {
4557 success = __kmp_affinity_create_proc_group_map(&msg_id);
4558 if (!success && verbose && msg_id != kmp_i18n_null) {
4559 KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4560 }
4561 }
4562 #endif /* KMP_GROUP_AFFINITY */
4563
4564 if (!success) {
4565 success = __kmp_affinity_create_flat_map(&msg_id);
4566 if (!success && verbose && msg_id != kmp_i18n_null) {
4567 KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4568 }
4569 KMP_ASSERT(success);
4570 }
4571 }
4572
4573 // If the user has specified that a paricular topology discovery method is to be
4574 // used, then we abort if that method fails. The exception is group affinity,
4575 // which might have been implicitly set.
4576 #if KMP_USE_HWLOC
4577 else if (__kmp_affinity_top_method == affinity_top_method_hwloc) {
4578 KMP_ASSERT(__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC);
4579 success = __kmp_affinity_create_hwloc_map(&msg_id);
4580 if (!success) {
4581 KMP_ASSERT(msg_id != kmp_i18n_null);
4582 KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4583 }
4584 }
4585 #endif // KMP_USE_HWLOC
4586
4587 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4588 else if (__kmp_affinity_top_method == affinity_top_method_x2apicid ||
4589 __kmp_affinity_top_method == affinity_top_method_x2apicid_1f) {
4590 success = __kmp_affinity_create_x2apicid_map(&msg_id);
4591 if (!success) {
4592 KMP_ASSERT(msg_id != kmp_i18n_null);
4593 KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4594 }
4595 } else if (__kmp_affinity_top_method == affinity_top_method_apicid) {
4596 success = __kmp_affinity_create_apicid_map(&msg_id);
4597 if (!success) {
4598 KMP_ASSERT(msg_id != kmp_i18n_null);
4599 KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4600 }
4601 }
4602 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4603
4604 else if (__kmp_affinity_top_method == affinity_top_method_cpuinfo) {
4605 int line = 0;
4606 success = __kmp_affinity_create_cpuinfo_map(&line, &msg_id);
4607 if (!success) {
4608 KMP_ASSERT(msg_id != kmp_i18n_null);
4609 const char *filename = __kmp_cpuinfo_get_filename();
4610 if (line > 0) {
4611 KMP_FATAL(FileLineMsgExiting, filename, line,
4612 __kmp_i18n_catgets(msg_id));
4613 } else {
4614 KMP_FATAL(FileMsgExiting, filename, __kmp_i18n_catgets(msg_id));
4615 }
4616 }
4617 }
4618
4619 #if KMP_GROUP_AFFINITY
4620 else if (__kmp_affinity_top_method == affinity_top_method_group) {
4621 success = __kmp_affinity_create_proc_group_map(&msg_id);
4622 KMP_ASSERT(success);
4623 if (!success) {
4624 KMP_ASSERT(msg_id != kmp_i18n_null);
4625 KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4626 }
4627 }
4628 #endif /* KMP_GROUP_AFFINITY */
4629
4630 else if (__kmp_affinity_top_method == affinity_top_method_flat) {
4631 success = __kmp_affinity_create_flat_map(&msg_id);
4632 // should not fail
4633 KMP_ASSERT(success);
4634 }
4635
4636 // Early exit if topology could not be created
4637 if (!__kmp_topology) {
4638 if (KMP_AFFINITY_CAPABLE()) {
4639 KMP_AFF_WARNING(affinity, ErrorInitializeAffinity);
4640 }
4641 if (nPackages > 0 && nCoresPerPkg > 0 && __kmp_nThreadsPerCore > 0 &&
4642 __kmp_ncores > 0) {
4643 __kmp_topology = kmp_topology_t::allocate(0, 0, NULL);
4644 __kmp_topology->canonicalize(nPackages, nCoresPerPkg,
4645 __kmp_nThreadsPerCore, __kmp_ncores);
4646 if (verbose) {
4647 __kmp_topology->print(env_var);
4648 }
4649 }
4650 return false;
4651 }
4652
4653 // Canonicalize, print (if requested), apply KMP_HW_SUBSET
4654 __kmp_topology->canonicalize();
4655 if (verbose)
4656 __kmp_topology->print(env_var);
4657 bool filtered = __kmp_topology->filter_hw_subset();
4658 if (filtered && verbose)
4659 __kmp_topology->print("KMP_HW_SUBSET");
4660 return success;
4661 }
4662
__kmp_aux_affinity_initialize(kmp_affinity_t & affinity)4663 static void __kmp_aux_affinity_initialize(kmp_affinity_t &affinity) {
4664 bool is_regular_affinity = (&affinity == &__kmp_affinity);
4665 bool is_hidden_helper_affinity = (&affinity == &__kmp_hh_affinity);
4666 const char *env_var = __kmp_get_affinity_env_var(affinity);
4667
4668 if (affinity.flags.initialized) {
4669 KMP_ASSERT(__kmp_affin_fullMask != NULL);
4670 return;
4671 }
4672
4673 if (is_regular_affinity && (!__kmp_affin_fullMask || !__kmp_affin_origMask))
4674 __kmp_aux_affinity_initialize_masks(affinity);
4675
4676 if (is_regular_affinity && !__kmp_topology) {
4677 bool success = __kmp_aux_affinity_initialize_topology(affinity);
4678 if (success) {
4679 KMP_ASSERT(__kmp_avail_proc == __kmp_topology->get_num_hw_threads());
4680 } else {
4681 affinity.type = affinity_none;
4682 KMP_AFFINITY_DISABLE();
4683 }
4684 }
4685
4686 // If KMP_AFFINITY=none, then only create the single "none" place
4687 // which is the process's initial affinity mask or the number of
4688 // hardware threads depending on respect,norespect
4689 if (affinity.type == affinity_none) {
4690 __kmp_create_affinity_none_places(affinity);
4691 #if KMP_USE_HIER_SCHED
4692 __kmp_dispatch_set_hierarchy_values();
4693 #endif
4694 affinity.flags.initialized = TRUE;
4695 return;
4696 }
4697
4698 __kmp_topology->set_granularity(affinity);
4699 int depth = __kmp_topology->get_depth();
4700
4701 // Create the table of masks, indexed by thread Id.
4702 unsigned numUnique;
4703 int numAddrs = __kmp_topology->get_num_hw_threads();
4704 // If OMP_PLACES=cores:<attribute> specified, then attempt
4705 // to make OS Id mask table using those attributes
4706 if (affinity.core_attr_gran.valid) {
4707 __kmp_create_os_id_masks(&numUnique, affinity, [&](int idx) {
4708 KMP_ASSERT(idx >= -1);
4709 for (int i = idx + 1; i < numAddrs; ++i)
4710 if (__kmp_topology->at(i).attrs.contains(affinity.core_attr_gran))
4711 return i;
4712 return numAddrs;
4713 });
4714 if (!affinity.os_id_masks) {
4715 const char *core_attribute;
4716 if (affinity.core_attr_gran.core_eff != kmp_hw_attr_t::UNKNOWN_CORE_EFF)
4717 core_attribute = "core_efficiency";
4718 else
4719 core_attribute = "core_type";
4720 KMP_AFF_WARNING(affinity, AffIgnoringNotAvailable, env_var,
4721 core_attribute,
4722 __kmp_hw_get_catalog_string(KMP_HW_CORE, /*plural=*/true))
4723 }
4724 }
4725 // If core attributes did not work, or none were specified,
4726 // then make OS Id mask table using typical incremental way.
4727 if (!affinity.os_id_masks) {
4728 __kmp_create_os_id_masks(&numUnique, affinity, [](int idx) {
4729 KMP_ASSERT(idx >= -1);
4730 return idx + 1;
4731 });
4732 }
4733 if (affinity.gran_levels == 0) {
4734 KMP_DEBUG_ASSERT((int)numUnique == __kmp_avail_proc);
4735 }
4736
4737 switch (affinity.type) {
4738
4739 case affinity_explicit:
4740 KMP_DEBUG_ASSERT(affinity.proclist != NULL);
4741 if (is_hidden_helper_affinity ||
4742 __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) {
4743 __kmp_affinity_process_proclist(affinity);
4744 } else {
4745 __kmp_affinity_process_placelist(affinity);
4746 }
4747 if (affinity.num_masks == 0) {
4748 KMP_AFF_WARNING(affinity, AffNoValidProcID);
4749 affinity.type = affinity_none;
4750 __kmp_create_affinity_none_places(affinity);
4751 affinity.flags.initialized = TRUE;
4752 return;
4753 }
4754 break;
4755
4756 // The other affinity types rely on sorting the hardware threads according to
4757 // some permutation of the machine topology tree. Set affinity.compact
4758 // and affinity.offset appropriately, then jump to a common code
4759 // fragment to do the sort and create the array of affinity masks.
4760 case affinity_logical:
4761 affinity.compact = 0;
4762 if (affinity.offset) {
4763 affinity.offset =
4764 __kmp_nThreadsPerCore * affinity.offset % __kmp_avail_proc;
4765 }
4766 goto sortTopology;
4767
4768 case affinity_physical:
4769 if (__kmp_nThreadsPerCore > 1) {
4770 affinity.compact = 1;
4771 if (affinity.compact >= depth) {
4772 affinity.compact = 0;
4773 }
4774 } else {
4775 affinity.compact = 0;
4776 }
4777 if (affinity.offset) {
4778 affinity.offset =
4779 __kmp_nThreadsPerCore * affinity.offset % __kmp_avail_proc;
4780 }
4781 goto sortTopology;
4782
4783 case affinity_scatter:
4784 if (affinity.compact >= depth) {
4785 affinity.compact = 0;
4786 } else {
4787 affinity.compact = depth - 1 - affinity.compact;
4788 }
4789 goto sortTopology;
4790
4791 case affinity_compact:
4792 if (affinity.compact >= depth) {
4793 affinity.compact = depth - 1;
4794 }
4795 goto sortTopology;
4796
4797 case affinity_balanced:
4798 if (depth <= 1 || is_hidden_helper_affinity) {
4799 KMP_AFF_WARNING(affinity, AffBalancedNotAvail, env_var);
4800 affinity.type = affinity_none;
4801 __kmp_create_affinity_none_places(affinity);
4802 affinity.flags.initialized = TRUE;
4803 return;
4804 } else if (!__kmp_topology->is_uniform()) {
4805 // Save the depth for further usage
4806 __kmp_aff_depth = depth;
4807
4808 int core_level =
4809 __kmp_affinity_find_core_level(__kmp_avail_proc, depth - 1);
4810 int ncores = __kmp_affinity_compute_ncores(__kmp_avail_proc, depth - 1,
4811 core_level);
4812 int maxprocpercore = __kmp_affinity_max_proc_per_core(
4813 __kmp_avail_proc, depth - 1, core_level);
4814
4815 int nproc = ncores * maxprocpercore;
4816 if ((nproc < 2) || (nproc < __kmp_avail_proc)) {
4817 KMP_AFF_WARNING(affinity, AffBalancedNotAvail, env_var);
4818 affinity.type = affinity_none;
4819 __kmp_create_affinity_none_places(affinity);
4820 affinity.flags.initialized = TRUE;
4821 return;
4822 }
4823
4824 procarr = (int *)__kmp_allocate(sizeof(int) * nproc);
4825 for (int i = 0; i < nproc; i++) {
4826 procarr[i] = -1;
4827 }
4828
4829 int lastcore = -1;
4830 int inlastcore = 0;
4831 for (int i = 0; i < __kmp_avail_proc; i++) {
4832 int proc = __kmp_topology->at(i).os_id;
4833 int core = __kmp_affinity_find_core(i, depth - 1, core_level);
4834
4835 if (core == lastcore) {
4836 inlastcore++;
4837 } else {
4838 inlastcore = 0;
4839 }
4840 lastcore = core;
4841
4842 procarr[core * maxprocpercore + inlastcore] = proc;
4843 }
4844 }
4845 if (affinity.compact >= depth) {
4846 affinity.compact = depth - 1;
4847 }
4848
4849 sortTopology:
4850 // Allocate the gtid->affinity mask table.
4851 if (affinity.flags.dups) {
4852 affinity.num_masks = __kmp_avail_proc;
4853 } else {
4854 affinity.num_masks = numUnique;
4855 }
4856
4857 if ((__kmp_nested_proc_bind.bind_types[0] != proc_bind_intel) &&
4858 (__kmp_affinity_num_places > 0) &&
4859 ((unsigned)__kmp_affinity_num_places < affinity.num_masks) &&
4860 !is_hidden_helper_affinity) {
4861 affinity.num_masks = __kmp_affinity_num_places;
4862 }
4863
4864 KMP_CPU_ALLOC_ARRAY(affinity.masks, affinity.num_masks);
4865
4866 // Sort the topology table according to the current setting of
4867 // affinity.compact, then fill out affinity.masks.
4868 __kmp_topology->sort_compact(affinity);
4869 {
4870 int i;
4871 unsigned j;
4872 int num_hw_threads = __kmp_topology->get_num_hw_threads();
4873 kmp_full_mask_modifier_t full_mask;
4874 for (i = 0, j = 0; i < num_hw_threads; i++) {
4875 if ((!affinity.flags.dups) && (!__kmp_topology->at(i).leader)) {
4876 continue;
4877 }
4878 int osId = __kmp_topology->at(i).os_id;
4879
4880 kmp_affin_mask_t *src = KMP_CPU_INDEX(affinity.os_id_masks, osId);
4881 kmp_affin_mask_t *dest = KMP_CPU_INDEX(affinity.masks, j);
4882 KMP_ASSERT(KMP_CPU_ISSET(osId, src));
4883 KMP_CPU_COPY(dest, src);
4884 full_mask.include(src);
4885 if (++j >= affinity.num_masks) {
4886 break;
4887 }
4888 }
4889 KMP_DEBUG_ASSERT(j == affinity.num_masks);
4890 // See if the places list further restricts or changes the full mask
4891 if (full_mask.restrict_to_mask() && affinity.flags.verbose) {
4892 __kmp_topology->print(env_var);
4893 }
4894 }
4895 // Sort the topology back using ids
4896 __kmp_topology->sort_ids();
4897 break;
4898
4899 default:
4900 KMP_ASSERT2(0, "Unexpected affinity setting");
4901 }
4902 __kmp_aux_affinity_initialize_other_data(affinity);
4903 affinity.flags.initialized = TRUE;
4904 }
4905
__kmp_affinity_initialize(kmp_affinity_t & affinity)4906 void __kmp_affinity_initialize(kmp_affinity_t &affinity) {
4907 // Much of the code above was written assuming that if a machine was not
4908 // affinity capable, then affinity type == affinity_none.
4909 // We now explicitly represent this as affinity type == affinity_disabled.
4910 // There are too many checks for affinity type == affinity_none in this code.
4911 // Instead of trying to change them all, check if
4912 // affinity type == affinity_disabled, and if so, slam it with affinity_none,
4913 // call the real initialization routine, then restore affinity type to
4914 // affinity_disabled.
4915 int disabled = (affinity.type == affinity_disabled);
4916 if (!KMP_AFFINITY_CAPABLE())
4917 KMP_ASSERT(disabled);
4918 if (disabled)
4919 affinity.type = affinity_none;
4920 __kmp_aux_affinity_initialize(affinity);
4921 if (disabled)
4922 affinity.type = affinity_disabled;
4923 }
4924
__kmp_affinity_uninitialize(void)4925 void __kmp_affinity_uninitialize(void) {
4926 for (kmp_affinity_t *affinity : __kmp_affinities) {
4927 if (affinity->masks != NULL)
4928 KMP_CPU_FREE_ARRAY(affinity->masks, affinity->num_masks);
4929 if (affinity->os_id_masks != NULL)
4930 KMP_CPU_FREE_ARRAY(affinity->os_id_masks, affinity->num_os_id_masks);
4931 if (affinity->proclist != NULL)
4932 __kmp_free(affinity->proclist);
4933 if (affinity->ids != NULL)
4934 __kmp_free(affinity->ids);
4935 if (affinity->attrs != NULL)
4936 __kmp_free(affinity->attrs);
4937 *affinity = KMP_AFFINITY_INIT(affinity->env_var);
4938 }
4939 if (__kmp_affin_origMask != NULL) {
4940 if (KMP_AFFINITY_CAPABLE()) {
4941 #if KMP_OS_AIX
4942 // Uninitialize by unbinding the thread.
4943 bindprocessor(BINDTHREAD, thread_self(), PROCESSOR_CLASS_ANY);
4944 #else
4945 __kmp_set_system_affinity(__kmp_affin_origMask, FALSE);
4946 #endif
4947 }
4948 KMP_CPU_FREE(__kmp_affin_origMask);
4949 __kmp_affin_origMask = NULL;
4950 }
4951 __kmp_affinity_num_places = 0;
4952 if (procarr != NULL) {
4953 __kmp_free(procarr);
4954 procarr = NULL;
4955 }
4956 if (__kmp_osid_to_hwthread_map) {
4957 __kmp_free(__kmp_osid_to_hwthread_map);
4958 __kmp_osid_to_hwthread_map = NULL;
4959 }
4960 #if KMP_USE_HWLOC
4961 if (__kmp_hwloc_topology != NULL) {
4962 hwloc_topology_destroy(__kmp_hwloc_topology);
4963 __kmp_hwloc_topology = NULL;
4964 }
4965 #endif
4966 if (__kmp_hw_subset) {
4967 kmp_hw_subset_t::deallocate(__kmp_hw_subset);
4968 __kmp_hw_subset = nullptr;
4969 }
4970 if (__kmp_topology) {
4971 kmp_topology_t::deallocate(__kmp_topology);
4972 __kmp_topology = nullptr;
4973 }
4974 KMPAffinity::destroy_api();
4975 }
4976
__kmp_select_mask_by_gtid(int gtid,const kmp_affinity_t * affinity,int * place,kmp_affin_mask_t ** mask)4977 static void __kmp_select_mask_by_gtid(int gtid, const kmp_affinity_t *affinity,
4978 int *place, kmp_affin_mask_t **mask) {
4979 int mask_idx;
4980 bool is_hidden_helper = KMP_HIDDEN_HELPER_THREAD(gtid);
4981 if (is_hidden_helper)
4982 // The first gtid is the regular primary thread, the second gtid is the main
4983 // thread of hidden team which does not participate in task execution.
4984 mask_idx = gtid - 2;
4985 else
4986 mask_idx = __kmp_adjust_gtid_for_hidden_helpers(gtid);
4987 KMP_DEBUG_ASSERT(affinity->num_masks > 0);
4988 *place = (mask_idx + affinity->offset) % affinity->num_masks;
4989 *mask = KMP_CPU_INDEX(affinity->masks, *place);
4990 }
4991
4992 // This function initializes the per-thread data concerning affinity including
4993 // the mask and topology information
__kmp_affinity_set_init_mask(int gtid,int isa_root)4994 void __kmp_affinity_set_init_mask(int gtid, int isa_root) {
4995
4996 kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
4997
4998 // Set the thread topology information to default of unknown
4999 for (int id = 0; id < KMP_HW_LAST; ++id)
5000 th->th.th_topology_ids.ids[id] = kmp_hw_thread_t::UNKNOWN_ID;
5001 th->th.th_topology_attrs = KMP_AFFINITY_ATTRS_UNKNOWN;
5002
5003 if (!KMP_AFFINITY_CAPABLE()) {
5004 return;
5005 }
5006
5007 if (th->th.th_affin_mask == NULL) {
5008 KMP_CPU_ALLOC(th->th.th_affin_mask);
5009 } else {
5010 KMP_CPU_ZERO(th->th.th_affin_mask);
5011 }
5012
5013 // Copy the thread mask to the kmp_info_t structure. If
5014 // __kmp_affinity.type == affinity_none, copy the "full" mask, i.e.
5015 // one that has all of the OS proc ids set, or if
5016 // __kmp_affinity.flags.respect is set, then the full mask is the
5017 // same as the mask of the initialization thread.
5018 kmp_affin_mask_t *mask;
5019 int i;
5020 const kmp_affinity_t *affinity;
5021 bool is_hidden_helper = KMP_HIDDEN_HELPER_THREAD(gtid);
5022
5023 if (is_hidden_helper)
5024 affinity = &__kmp_hh_affinity;
5025 else
5026 affinity = &__kmp_affinity;
5027
5028 if (KMP_AFFINITY_NON_PROC_BIND || is_hidden_helper) {
5029 if ((affinity->type == affinity_none) ||
5030 (affinity->type == affinity_balanced) ||
5031 KMP_HIDDEN_HELPER_MAIN_THREAD(gtid)) {
5032 #if KMP_GROUP_AFFINITY
5033 if (__kmp_num_proc_groups > 1) {
5034 return;
5035 }
5036 #endif
5037 KMP_ASSERT(__kmp_affin_fullMask != NULL);
5038 i = 0;
5039 mask = __kmp_affin_fullMask;
5040 } else {
5041 __kmp_select_mask_by_gtid(gtid, affinity, &i, &mask);
5042 }
5043 } else {
5044 if (!isa_root || __kmp_nested_proc_bind.bind_types[0] == proc_bind_false) {
5045 #if KMP_GROUP_AFFINITY
5046 if (__kmp_num_proc_groups > 1) {
5047 return;
5048 }
5049 #endif
5050 KMP_ASSERT(__kmp_affin_fullMask != NULL);
5051 i = KMP_PLACE_ALL;
5052 mask = __kmp_affin_fullMask;
5053 } else {
5054 __kmp_select_mask_by_gtid(gtid, affinity, &i, &mask);
5055 }
5056 }
5057
5058 th->th.th_current_place = i;
5059 if (isa_root && !is_hidden_helper) {
5060 th->th.th_new_place = i;
5061 th->th.th_first_place = 0;
5062 th->th.th_last_place = affinity->num_masks - 1;
5063 } else if (KMP_AFFINITY_NON_PROC_BIND) {
5064 // When using a Non-OMP_PROC_BIND affinity method,
5065 // set all threads' place-partition-var to the entire place list
5066 th->th.th_first_place = 0;
5067 th->th.th_last_place = affinity->num_masks - 1;
5068 }
5069 // Copy topology information associated with the place
5070 if (i >= 0) {
5071 th->th.th_topology_ids = __kmp_affinity.ids[i];
5072 th->th.th_topology_attrs = __kmp_affinity.attrs[i];
5073 }
5074
5075 if (i == KMP_PLACE_ALL) {
5076 KA_TRACE(100, ("__kmp_affinity_set_init_mask: setting T#%d to all places\n",
5077 gtid));
5078 } else {
5079 KA_TRACE(100, ("__kmp_affinity_set_init_mask: setting T#%d to place %d\n",
5080 gtid, i));
5081 }
5082
5083 KMP_CPU_COPY(th->th.th_affin_mask, mask);
5084 }
5085
__kmp_affinity_bind_init_mask(int gtid)5086 void __kmp_affinity_bind_init_mask(int gtid) {
5087 if (!KMP_AFFINITY_CAPABLE()) {
5088 return;
5089 }
5090 kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
5091 const kmp_affinity_t *affinity;
5092 const char *env_var;
5093 bool is_hidden_helper = KMP_HIDDEN_HELPER_THREAD(gtid);
5094
5095 if (is_hidden_helper)
5096 affinity = &__kmp_hh_affinity;
5097 else
5098 affinity = &__kmp_affinity;
5099 env_var = __kmp_get_affinity_env_var(*affinity, /*for_binding=*/true);
5100 /* to avoid duplicate printing (will be correctly printed on barrier) */
5101 if (affinity->flags.verbose && (affinity->type == affinity_none ||
5102 (th->th.th_current_place != KMP_PLACE_ALL &&
5103 affinity->type != affinity_balanced)) &&
5104 !KMP_HIDDEN_HELPER_MAIN_THREAD(gtid)) {
5105 char buf[KMP_AFFIN_MASK_PRINT_LEN];
5106 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5107 th->th.th_affin_mask);
5108 KMP_INFORM(BoundToOSProcSet, env_var, (kmp_int32)getpid(), __kmp_gettid(),
5109 gtid, buf);
5110 }
5111
5112 #if KMP_OS_WINDOWS
5113 // On Windows* OS, the process affinity mask might have changed. If the user
5114 // didn't request affinity and this call fails, just continue silently.
5115 // See CQ171393.
5116 if (affinity->type == affinity_none) {
5117 __kmp_set_system_affinity(th->th.th_affin_mask, FALSE);
5118 } else
5119 #endif
5120 #ifndef KMP_OS_AIX
5121 // Do not set the full mask as the init mask on AIX.
5122 __kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
5123 #endif
5124 }
5125
__kmp_affinity_bind_place(int gtid)5126 void __kmp_affinity_bind_place(int gtid) {
5127 // Hidden helper threads should not be affected by OMP_PLACES/OMP_PROC_BIND
5128 if (!KMP_AFFINITY_CAPABLE() || KMP_HIDDEN_HELPER_THREAD(gtid)) {
5129 return;
5130 }
5131
5132 kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
5133
5134 KA_TRACE(100, ("__kmp_affinity_bind_place: binding T#%d to place %d (current "
5135 "place = %d)\n",
5136 gtid, th->th.th_new_place, th->th.th_current_place));
5137
5138 // Check that the new place is within this thread's partition.
5139 KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
5140 KMP_ASSERT(th->th.th_new_place >= 0);
5141 KMP_ASSERT((unsigned)th->th.th_new_place <= __kmp_affinity.num_masks);
5142 if (th->th.th_first_place <= th->th.th_last_place) {
5143 KMP_ASSERT((th->th.th_new_place >= th->th.th_first_place) &&
5144 (th->th.th_new_place <= th->th.th_last_place));
5145 } else {
5146 KMP_ASSERT((th->th.th_new_place <= th->th.th_first_place) ||
5147 (th->th.th_new_place >= th->th.th_last_place));
5148 }
5149
5150 // Copy the thread mask to the kmp_info_t structure,
5151 // and set this thread's affinity.
5152 kmp_affin_mask_t *mask =
5153 KMP_CPU_INDEX(__kmp_affinity.masks, th->th.th_new_place);
5154 KMP_CPU_COPY(th->th.th_affin_mask, mask);
5155 th->th.th_current_place = th->th.th_new_place;
5156
5157 if (__kmp_affinity.flags.verbose) {
5158 char buf[KMP_AFFIN_MASK_PRINT_LEN];
5159 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5160 th->th.th_affin_mask);
5161 KMP_INFORM(BoundToOSProcSet, "OMP_PROC_BIND", (kmp_int32)getpid(),
5162 __kmp_gettid(), gtid, buf);
5163 }
5164 __kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
5165 }
5166
__kmp_aux_set_affinity(void ** mask)5167 int __kmp_aux_set_affinity(void **mask) {
5168 int gtid;
5169 kmp_info_t *th;
5170 int retval;
5171
5172 if (!KMP_AFFINITY_CAPABLE()) {
5173 return -1;
5174 }
5175
5176 gtid = __kmp_entry_gtid();
5177 KA_TRACE(
5178 1000, (""); {
5179 char buf[KMP_AFFIN_MASK_PRINT_LEN];
5180 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5181 (kmp_affin_mask_t *)(*mask));
5182 __kmp_debug_printf(
5183 "kmp_set_affinity: setting affinity mask for thread %d = %s\n",
5184 gtid, buf);
5185 });
5186
5187 if (__kmp_env_consistency_check) {
5188 if ((mask == NULL) || (*mask == NULL)) {
5189 KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
5190 } else {
5191 unsigned proc;
5192 int num_procs = 0;
5193
5194 KMP_CPU_SET_ITERATE(proc, ((kmp_affin_mask_t *)(*mask))) {
5195 if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5196 KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
5197 }
5198 if (!KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask))) {
5199 continue;
5200 }
5201 num_procs++;
5202 }
5203 if (num_procs == 0) {
5204 KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
5205 }
5206
5207 #if KMP_GROUP_AFFINITY
5208 if (__kmp_get_proc_group((kmp_affin_mask_t *)(*mask)) < 0) {
5209 KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
5210 }
5211 #endif /* KMP_GROUP_AFFINITY */
5212 }
5213 }
5214
5215 th = __kmp_threads[gtid];
5216 KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
5217 retval = __kmp_set_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
5218 if (retval == 0) {
5219 KMP_CPU_COPY(th->th.th_affin_mask, (kmp_affin_mask_t *)(*mask));
5220 }
5221
5222 th->th.th_current_place = KMP_PLACE_UNDEFINED;
5223 th->th.th_new_place = KMP_PLACE_UNDEFINED;
5224 th->th.th_first_place = 0;
5225 th->th.th_last_place = __kmp_affinity.num_masks - 1;
5226
5227 // Turn off 4.0 affinity for the current tread at this parallel level.
5228 th->th.th_current_task->td_icvs.proc_bind = proc_bind_false;
5229
5230 return retval;
5231 }
5232
__kmp_aux_get_affinity(void ** mask)5233 int __kmp_aux_get_affinity(void **mask) {
5234 int gtid;
5235 int retval;
5236 #if KMP_OS_WINDOWS || KMP_OS_AIX || KMP_DEBUG
5237 kmp_info_t *th;
5238 #endif
5239 if (!KMP_AFFINITY_CAPABLE()) {
5240 return -1;
5241 }
5242
5243 gtid = __kmp_entry_gtid();
5244 #if KMP_OS_WINDOWS || KMP_OS_AIX || KMP_DEBUG
5245 th = __kmp_threads[gtid];
5246 #else
5247 (void)gtid; // unused variable
5248 #endif
5249 KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
5250
5251 KA_TRACE(
5252 1000, (""); {
5253 char buf[KMP_AFFIN_MASK_PRINT_LEN];
5254 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5255 th->th.th_affin_mask);
5256 __kmp_printf(
5257 "kmp_get_affinity: stored affinity mask for thread %d = %s\n", gtid,
5258 buf);
5259 });
5260
5261 if (__kmp_env_consistency_check) {
5262 if ((mask == NULL) || (*mask == NULL)) {
5263 KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity");
5264 }
5265 }
5266
5267 #if !KMP_OS_WINDOWS && !KMP_OS_AIX
5268
5269 retval = __kmp_get_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
5270 KA_TRACE(
5271 1000, (""); {
5272 char buf[KMP_AFFIN_MASK_PRINT_LEN];
5273 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5274 (kmp_affin_mask_t *)(*mask));
5275 __kmp_printf(
5276 "kmp_get_affinity: system affinity mask for thread %d = %s\n", gtid,
5277 buf);
5278 });
5279 return retval;
5280
5281 #else
5282 (void)retval;
5283
5284 KMP_CPU_COPY((kmp_affin_mask_t *)(*mask), th->th.th_affin_mask);
5285 return 0;
5286
5287 #endif /* !KMP_OS_WINDOWS && !KMP_OS_AIX */
5288 }
5289
__kmp_aux_get_affinity_max_proc()5290 int __kmp_aux_get_affinity_max_proc() {
5291 if (!KMP_AFFINITY_CAPABLE()) {
5292 return 0;
5293 }
5294 #if KMP_GROUP_AFFINITY
5295 if (__kmp_num_proc_groups > 1) {
5296 return (int)(__kmp_num_proc_groups * sizeof(DWORD_PTR) * CHAR_BIT);
5297 }
5298 #endif
5299 return __kmp_xproc;
5300 }
5301
__kmp_aux_set_affinity_mask_proc(int proc,void ** mask)5302 int __kmp_aux_set_affinity_mask_proc(int proc, void **mask) {
5303 if (!KMP_AFFINITY_CAPABLE()) {
5304 return -1;
5305 }
5306
5307 KA_TRACE(
5308 1000, (""); {
5309 int gtid = __kmp_entry_gtid();
5310 char buf[KMP_AFFIN_MASK_PRINT_LEN];
5311 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5312 (kmp_affin_mask_t *)(*mask));
5313 __kmp_debug_printf("kmp_set_affinity_mask_proc: setting proc %d in "
5314 "affinity mask for thread %d = %s\n",
5315 proc, gtid, buf);
5316 });
5317
5318 if (__kmp_env_consistency_check) {
5319 if ((mask == NULL) || (*mask == NULL)) {
5320 KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity_mask_proc");
5321 }
5322 }
5323
5324 if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5325 return -1;
5326 }
5327 if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5328 return -2;
5329 }
5330
5331 KMP_CPU_SET(proc, (kmp_affin_mask_t *)(*mask));
5332 return 0;
5333 }
5334
__kmp_aux_unset_affinity_mask_proc(int proc,void ** mask)5335 int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask) {
5336 if (!KMP_AFFINITY_CAPABLE()) {
5337 return -1;
5338 }
5339
5340 KA_TRACE(
5341 1000, (""); {
5342 int gtid = __kmp_entry_gtid();
5343 char buf[KMP_AFFIN_MASK_PRINT_LEN];
5344 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5345 (kmp_affin_mask_t *)(*mask));
5346 __kmp_debug_printf("kmp_unset_affinity_mask_proc: unsetting proc %d in "
5347 "affinity mask for thread %d = %s\n",
5348 proc, gtid, buf);
5349 });
5350
5351 if (__kmp_env_consistency_check) {
5352 if ((mask == NULL) || (*mask == NULL)) {
5353 KMP_FATAL(AffinityInvalidMask, "kmp_unset_affinity_mask_proc");
5354 }
5355 }
5356
5357 if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5358 return -1;
5359 }
5360 if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5361 return -2;
5362 }
5363
5364 KMP_CPU_CLR(proc, (kmp_affin_mask_t *)(*mask));
5365 return 0;
5366 }
5367
__kmp_aux_get_affinity_mask_proc(int proc,void ** mask)5368 int __kmp_aux_get_affinity_mask_proc(int proc, void **mask) {
5369 if (!KMP_AFFINITY_CAPABLE()) {
5370 return -1;
5371 }
5372
5373 KA_TRACE(
5374 1000, (""); {
5375 int gtid = __kmp_entry_gtid();
5376 char buf[KMP_AFFIN_MASK_PRINT_LEN];
5377 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5378 (kmp_affin_mask_t *)(*mask));
5379 __kmp_debug_printf("kmp_get_affinity_mask_proc: getting proc %d in "
5380 "affinity mask for thread %d = %s\n",
5381 proc, gtid, buf);
5382 });
5383
5384 if (__kmp_env_consistency_check) {
5385 if ((mask == NULL) || (*mask == NULL)) {
5386 KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity_mask_proc");
5387 }
5388 }
5389
5390 if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5391 return -1;
5392 }
5393 if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5394 return 0;
5395 }
5396
5397 return KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask));
5398 }
5399
5400 #if KMP_WEIGHTED_ITERATIONS_SUPPORTED
5401 // Returns first os proc id with ATOM core
__kmp_get_first_osid_with_ecore(void)5402 int __kmp_get_first_osid_with_ecore(void) {
5403 int low = 0;
5404 int high = __kmp_topology->get_num_hw_threads() - 1;
5405 int mid = 0;
5406 while (high - low > 1) {
5407 mid = (high + low) / 2;
5408 if (__kmp_topology->at(mid).attrs.get_core_type() ==
5409 KMP_HW_CORE_TYPE_CORE) {
5410 low = mid + 1;
5411 } else {
5412 high = mid;
5413 }
5414 }
5415 if (__kmp_topology->at(mid).attrs.get_core_type() == KMP_HW_CORE_TYPE_ATOM) {
5416 return mid;
5417 }
5418 return -1;
5419 }
5420 #endif
5421
5422 // Dynamic affinity settings - Affinity balanced
__kmp_balanced_affinity(kmp_info_t * th,int nthreads)5423 void __kmp_balanced_affinity(kmp_info_t *th, int nthreads) {
5424 KMP_DEBUG_ASSERT(th);
5425 bool fine_gran = true;
5426 int tid = th->th.th_info.ds.ds_tid;
5427 const char *env_var = "KMP_AFFINITY";
5428
5429 // Do not perform balanced affinity for the hidden helper threads
5430 if (KMP_HIDDEN_HELPER_THREAD(__kmp_gtid_from_thread(th)))
5431 return;
5432
5433 switch (__kmp_affinity.gran) {
5434 case KMP_HW_THREAD:
5435 break;
5436 case KMP_HW_CORE:
5437 if (__kmp_nThreadsPerCore > 1) {
5438 fine_gran = false;
5439 }
5440 break;
5441 case KMP_HW_SOCKET:
5442 if (nCoresPerPkg > 1) {
5443 fine_gran = false;
5444 }
5445 break;
5446 default:
5447 fine_gran = false;
5448 }
5449
5450 if (__kmp_topology->is_uniform()) {
5451 int coreID;
5452 int threadID;
5453 // Number of hyper threads per core in HT machine
5454 int __kmp_nth_per_core = __kmp_avail_proc / __kmp_ncores;
5455 // Number of cores
5456 int ncores = __kmp_ncores;
5457 if ((nPackages > 1) && (__kmp_nth_per_core <= 1)) {
5458 __kmp_nth_per_core = __kmp_avail_proc / nPackages;
5459 ncores = nPackages;
5460 }
5461 // How many threads will be bound to each core
5462 int chunk = nthreads / ncores;
5463 // How many cores will have an additional thread bound to it - "big cores"
5464 int big_cores = nthreads % ncores;
5465 // Number of threads on the big cores
5466 int big_nth = (chunk + 1) * big_cores;
5467 if (tid < big_nth) {
5468 coreID = tid / (chunk + 1);
5469 threadID = (tid % (chunk + 1)) % __kmp_nth_per_core;
5470 } else { // tid >= big_nth
5471 coreID = (tid - big_cores) / chunk;
5472 threadID = ((tid - big_cores) % chunk) % __kmp_nth_per_core;
5473 }
5474 KMP_DEBUG_ASSERT2(KMP_AFFINITY_CAPABLE(),
5475 "Illegal set affinity operation when not capable");
5476
5477 kmp_affin_mask_t *mask = th->th.th_affin_mask;
5478 KMP_CPU_ZERO(mask);
5479
5480 if (fine_gran) {
5481 int osID =
5482 __kmp_topology->at(coreID * __kmp_nth_per_core + threadID).os_id;
5483 KMP_CPU_SET(osID, mask);
5484 } else {
5485 for (int i = 0; i < __kmp_nth_per_core; i++) {
5486 int osID;
5487 osID = __kmp_topology->at(coreID * __kmp_nth_per_core + i).os_id;
5488 KMP_CPU_SET(osID, mask);
5489 }
5490 }
5491 if (__kmp_affinity.flags.verbose) {
5492 char buf[KMP_AFFIN_MASK_PRINT_LEN];
5493 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
5494 KMP_INFORM(BoundToOSProcSet, env_var, (kmp_int32)getpid(), __kmp_gettid(),
5495 tid, buf);
5496 }
5497 __kmp_affinity_get_thread_topology_info(th);
5498 __kmp_set_system_affinity(mask, TRUE);
5499 } else { // Non-uniform topology
5500
5501 kmp_affin_mask_t *mask = th->th.th_affin_mask;
5502 KMP_CPU_ZERO(mask);
5503
5504 int core_level =
5505 __kmp_affinity_find_core_level(__kmp_avail_proc, __kmp_aff_depth - 1);
5506 int ncores = __kmp_affinity_compute_ncores(__kmp_avail_proc,
5507 __kmp_aff_depth - 1, core_level);
5508 int nth_per_core = __kmp_affinity_max_proc_per_core(
5509 __kmp_avail_proc, __kmp_aff_depth - 1, core_level);
5510
5511 // For performance gain consider the special case nthreads ==
5512 // __kmp_avail_proc
5513 if (nthreads == __kmp_avail_proc) {
5514 if (fine_gran) {
5515 int osID = __kmp_topology->at(tid).os_id;
5516 KMP_CPU_SET(osID, mask);
5517 } else {
5518 int core =
5519 __kmp_affinity_find_core(tid, __kmp_aff_depth - 1, core_level);
5520 for (int i = 0; i < __kmp_avail_proc; i++) {
5521 int osID = __kmp_topology->at(i).os_id;
5522 if (__kmp_affinity_find_core(i, __kmp_aff_depth - 1, core_level) ==
5523 core) {
5524 KMP_CPU_SET(osID, mask);
5525 }
5526 }
5527 }
5528 } else if (nthreads <= ncores) {
5529
5530 int core = 0;
5531 for (int i = 0; i < ncores; i++) {
5532 // Check if this core from procarr[] is in the mask
5533 int in_mask = 0;
5534 for (int j = 0; j < nth_per_core; j++) {
5535 if (procarr[i * nth_per_core + j] != -1) {
5536 in_mask = 1;
5537 break;
5538 }
5539 }
5540 if (in_mask) {
5541 if (tid == core) {
5542 for (int j = 0; j < nth_per_core; j++) {
5543 int osID = procarr[i * nth_per_core + j];
5544 if (osID != -1) {
5545 KMP_CPU_SET(osID, mask);
5546 // For fine granularity it is enough to set the first available
5547 // osID for this core
5548 if (fine_gran) {
5549 break;
5550 }
5551 }
5552 }
5553 break;
5554 } else {
5555 core++;
5556 }
5557 }
5558 }
5559 } else { // nthreads > ncores
5560 // Array to save the number of processors at each core
5561 int *nproc_at_core = (int *)KMP_ALLOCA(sizeof(int) * ncores);
5562 // Array to save the number of cores with "x" available processors;
5563 int *ncores_with_x_procs =
5564 (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
5565 // Array to save the number of cores with # procs from x to nth_per_core
5566 int *ncores_with_x_to_max_procs =
5567 (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
5568
5569 for (int i = 0; i <= nth_per_core; i++) {
5570 ncores_with_x_procs[i] = 0;
5571 ncores_with_x_to_max_procs[i] = 0;
5572 }
5573
5574 for (int i = 0; i < ncores; i++) {
5575 int cnt = 0;
5576 for (int j = 0; j < nth_per_core; j++) {
5577 if (procarr[i * nth_per_core + j] != -1) {
5578 cnt++;
5579 }
5580 }
5581 nproc_at_core[i] = cnt;
5582 ncores_with_x_procs[cnt]++;
5583 }
5584
5585 for (int i = 0; i <= nth_per_core; i++) {
5586 for (int j = i; j <= nth_per_core; j++) {
5587 ncores_with_x_to_max_procs[i] += ncores_with_x_procs[j];
5588 }
5589 }
5590
5591 // Max number of processors
5592 int nproc = nth_per_core * ncores;
5593 // An array to keep number of threads per each context
5594 int *newarr = (int *)__kmp_allocate(sizeof(int) * nproc);
5595 for (int i = 0; i < nproc; i++) {
5596 newarr[i] = 0;
5597 }
5598
5599 int nth = nthreads;
5600 int flag = 0;
5601 while (nth > 0) {
5602 for (int j = 1; j <= nth_per_core; j++) {
5603 int cnt = ncores_with_x_to_max_procs[j];
5604 for (int i = 0; i < ncores; i++) {
5605 // Skip the core with 0 processors
5606 if (nproc_at_core[i] == 0) {
5607 continue;
5608 }
5609 for (int k = 0; k < nth_per_core; k++) {
5610 if (procarr[i * nth_per_core + k] != -1) {
5611 if (newarr[i * nth_per_core + k] == 0) {
5612 newarr[i * nth_per_core + k] = 1;
5613 cnt--;
5614 nth--;
5615 break;
5616 } else {
5617 if (flag != 0) {
5618 newarr[i * nth_per_core + k]++;
5619 cnt--;
5620 nth--;
5621 break;
5622 }
5623 }
5624 }
5625 }
5626 if (cnt == 0 || nth == 0) {
5627 break;
5628 }
5629 }
5630 if (nth == 0) {
5631 break;
5632 }
5633 }
5634 flag = 1;
5635 }
5636 int sum = 0;
5637 for (int i = 0; i < nproc; i++) {
5638 sum += newarr[i];
5639 if (sum > tid) {
5640 if (fine_gran) {
5641 int osID = procarr[i];
5642 KMP_CPU_SET(osID, mask);
5643 } else {
5644 int coreID = i / nth_per_core;
5645 for (int ii = 0; ii < nth_per_core; ii++) {
5646 int osID = procarr[coreID * nth_per_core + ii];
5647 if (osID != -1) {
5648 KMP_CPU_SET(osID, mask);
5649 }
5650 }
5651 }
5652 break;
5653 }
5654 }
5655 __kmp_free(newarr);
5656 }
5657
5658 if (__kmp_affinity.flags.verbose) {
5659 char buf[KMP_AFFIN_MASK_PRINT_LEN];
5660 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
5661 KMP_INFORM(BoundToOSProcSet, env_var, (kmp_int32)getpid(), __kmp_gettid(),
5662 tid, buf);
5663 }
5664 __kmp_affinity_get_thread_topology_info(th);
5665 __kmp_set_system_affinity(mask, TRUE);
5666 }
5667 }
5668
5669 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_AIX
5670 // We don't need this entry for Windows because
5671 // there is GetProcessAffinityMask() api
5672 //
5673 // The intended usage is indicated by these steps:
5674 // 1) The user gets the current affinity mask
5675 // 2) Then sets the affinity by calling this function
5676 // 3) Error check the return value
5677 // 4) Use non-OpenMP parallelization
5678 // 5) Reset the affinity to what was stored in step 1)
5679 #ifdef __cplusplus
5680 extern "C"
5681 #endif
5682 int
kmp_set_thread_affinity_mask_initial()5683 kmp_set_thread_affinity_mask_initial()
5684 // the function returns 0 on success,
5685 // -1 if we cannot bind thread
5686 // >0 (errno) if an error happened during binding
5687 {
5688 int gtid = __kmp_get_gtid();
5689 if (gtid < 0) {
5690 // Do not touch non-omp threads
5691 KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5692 "non-omp thread, returning\n"));
5693 return -1;
5694 }
5695 if (!KMP_AFFINITY_CAPABLE() || !__kmp_init_middle) {
5696 KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5697 "affinity not initialized, returning\n"));
5698 return -1;
5699 }
5700 KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5701 "set full mask for thread %d\n",
5702 gtid));
5703 KMP_DEBUG_ASSERT(__kmp_affin_fullMask != NULL);
5704 #if KMP_OS_AIX
5705 return bindprocessor(BINDTHREAD, thread_self(), PROCESSOR_CLASS_ANY);
5706 #else
5707 return __kmp_set_system_affinity(__kmp_affin_fullMask, FALSE);
5708 #endif
5709 }
5710 #endif
5711
5712 #endif // KMP_AFFINITY_SUPPORTED
5713