LLVM OpenMP* Runtime Library
kmp_affinity.cpp
1 /*
2  * kmp_affinity.cpp -- affinity management
3  */
4 
5 //===----------------------------------------------------------------------===//
6 //
7 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
8 // See https://llvm.org/LICENSE.txt for license information.
9 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "kmp.h"
14 #include "kmp_affinity.h"
15 #include "kmp_i18n.h"
16 #include "kmp_io.h"
17 #include "kmp_str.h"
18 #include "kmp_wrapper_getpid.h"
19 #if KMP_USE_HIER_SCHED
20 #include "kmp_dispatch_hier.h"
21 #endif
22 #if KMP_USE_HWLOC
23 // Copied from hwloc
24 #define HWLOC_GROUP_KIND_INTEL_MODULE 102
25 #define HWLOC_GROUP_KIND_INTEL_TILE 103
26 #define HWLOC_GROUP_KIND_INTEL_DIE 104
27 #define HWLOC_GROUP_KIND_WINDOWS_PROCESSOR_GROUP 220
28 #endif
29 #include <ctype.h>
30 
31 // The machine topology
32 kmp_topology_t *__kmp_topology = nullptr;
33 // KMP_HW_SUBSET environment variable
34 kmp_hw_subset_t *__kmp_hw_subset = nullptr;
35 
36 // Store the real or imagined machine hierarchy here
37 static hierarchy_info machine_hierarchy;
38 
39 void __kmp_cleanup_hierarchy() { machine_hierarchy.fini(); }
40 
41 #if KMP_AFFINITY_SUPPORTED
42 // Helper class to see if place lists further restrict the fullMask
43 class kmp_full_mask_modifier_t {
44  kmp_affin_mask_t *mask;
45 
46 public:
47  kmp_full_mask_modifier_t() {
48  KMP_CPU_ALLOC(mask);
49  KMP_CPU_ZERO(mask);
50  }
51  ~kmp_full_mask_modifier_t() {
52  KMP_CPU_FREE(mask);
53  mask = nullptr;
54  }
55  void include(const kmp_affin_mask_t *other) { KMP_CPU_UNION(mask, other); }
56  // If the new full mask is different from the current full mask,
57  // then switch them. Returns true if full mask was affected, false otherwise.
58  bool restrict_to_mask() {
59  // See if the new mask further restricts or changes the full mask
60  if (KMP_CPU_EQUAL(__kmp_affin_fullMask, mask) || KMP_CPU_ISEMPTY(mask))
61  return false;
62  return __kmp_topology->restrict_to_mask(mask);
63  }
64 };
65 
66 static inline const char *
67 __kmp_get_affinity_env_var(const kmp_affinity_t &affinity,
68  bool for_binding = false) {
69  if (affinity.flags.omp_places) {
70  if (for_binding)
71  return "OMP_PROC_BIND";
72  return "OMP_PLACES";
73  }
74  return affinity.env_var;
75 }
76 #endif // KMP_AFFINITY_SUPPORTED
77 
78 void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar) {
79  kmp_uint32 depth;
80  // The test below is true if affinity is available, but set to "none". Need to
81  // init on first use of hierarchical barrier.
82  if (TCR_1(machine_hierarchy.uninitialized))
83  machine_hierarchy.init(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 
104 const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural) {
105  switch (type) {
106  case KMP_HW_SOCKET:
107  return ((plural) ? KMP_I18N_STR(Sockets) : KMP_I18N_STR(Socket));
108  case KMP_HW_DIE:
109  return ((plural) ? KMP_I18N_STR(Dice) : KMP_I18N_STR(Die));
110  case KMP_HW_MODULE:
111  return ((plural) ? KMP_I18N_STR(Modules) : KMP_I18N_STR(Module));
112  case KMP_HW_TILE:
113  return ((plural) ? KMP_I18N_STR(Tiles) : KMP_I18N_STR(Tile));
114  case KMP_HW_NUMA:
115  return ((plural) ? KMP_I18N_STR(NumaDomains) : KMP_I18N_STR(NumaDomain));
116  case KMP_HW_L3:
117  return ((plural) ? KMP_I18N_STR(L3Caches) : KMP_I18N_STR(L3Cache));
118  case KMP_HW_L2:
119  return ((plural) ? KMP_I18N_STR(L2Caches) : KMP_I18N_STR(L2Cache));
120  case KMP_HW_L1:
121  return ((plural) ? KMP_I18N_STR(L1Caches) : KMP_I18N_STR(L1Cache));
122  case KMP_HW_LLC:
123  return ((plural) ? KMP_I18N_STR(LLCaches) : KMP_I18N_STR(LLCache));
124  case KMP_HW_CORE:
125  return ((plural) ? KMP_I18N_STR(Cores) : KMP_I18N_STR(Core));
126  case KMP_HW_THREAD:
127  return ((plural) ? KMP_I18N_STR(Threads) : KMP_I18N_STR(Thread));
128  case KMP_HW_PROC_GROUP:
129  return ((plural) ? KMP_I18N_STR(ProcGroups) : KMP_I18N_STR(ProcGroup));
130  case KMP_HW_UNKNOWN:
131  case KMP_HW_LAST:
132  return KMP_I18N_STR(Unknown);
133  }
134  KMP_ASSERT2(false, "Unhandled kmp_hw_t enumeration");
135  KMP_BUILTIN_UNREACHABLE;
136 }
137 
138 const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural) {
139  switch (type) {
140  case KMP_HW_SOCKET:
141  return ((plural) ? "sockets" : "socket");
142  case KMP_HW_DIE:
143  return ((plural) ? "dice" : "die");
144  case KMP_HW_MODULE:
145  return ((plural) ? "modules" : "module");
146  case KMP_HW_TILE:
147  return ((plural) ? "tiles" : "tile");
148  case KMP_HW_NUMA:
149  return ((plural) ? "numa_domains" : "numa_domain");
150  case KMP_HW_L3:
151  return ((plural) ? "l3_caches" : "l3_cache");
152  case KMP_HW_L2:
153  return ((plural) ? "l2_caches" : "l2_cache");
154  case KMP_HW_L1:
155  return ((plural) ? "l1_caches" : "l1_cache");
156  case KMP_HW_LLC:
157  return ((plural) ? "ll_caches" : "ll_cache");
158  case KMP_HW_CORE:
159  return ((plural) ? "cores" : "core");
160  case KMP_HW_THREAD:
161  return ((plural) ? "threads" : "thread");
162  case KMP_HW_PROC_GROUP:
163  return ((plural) ? "proc_groups" : "proc_group");
164  case KMP_HW_UNKNOWN:
165  case KMP_HW_LAST:
166  return ((plural) ? "unknowns" : "unknown");
167  }
168  KMP_ASSERT2(false, "Unhandled kmp_hw_t enumeration");
169  KMP_BUILTIN_UNREACHABLE;
170 }
171 
172 const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type) {
173  switch (type) {
174  case KMP_HW_CORE_TYPE_UNKNOWN:
175  case KMP_HW_MAX_NUM_CORE_TYPES:
176  return "unknown";
177 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
178  case KMP_HW_CORE_TYPE_ATOM:
179  return "Intel Atom(R) processor";
180  case KMP_HW_CORE_TYPE_CORE:
181  return "Intel(R) Core(TM) processor";
182 #endif
183  }
184  KMP_ASSERT2(false, "Unhandled kmp_hw_core_type_t enumeration");
185  KMP_BUILTIN_UNREACHABLE;
186 }
187 
188 #if KMP_AFFINITY_SUPPORTED
189 // If affinity is supported, check the affinity
190 // verbose and warning flags before printing warning
191 #define KMP_AFF_WARNING(s, ...) \
192  if (s.flags.verbose || (s.flags.warnings && (s.type != affinity_none))) { \
193  KMP_WARNING(__VA_ARGS__); \
194  }
195 #else
196 #define KMP_AFF_WARNING(s, ...) KMP_WARNING(__VA_ARGS__)
197 #endif
198 
200 // kmp_hw_thread_t methods
201 int kmp_hw_thread_t::compare_ids(const void *a, const void *b) {
202  const kmp_hw_thread_t *ahwthread = (const kmp_hw_thread_t *)a;
203  const kmp_hw_thread_t *bhwthread = (const kmp_hw_thread_t *)b;
204  int depth = __kmp_topology->get_depth();
205  for (int level = 0; level < depth; ++level) {
206  // Reverse sort (higher efficiencies earlier in list) cores by core
207  // efficiency if available.
208  if (__kmp_is_hybrid_cpu() &&
209  __kmp_topology->get_type(level) == KMP_HW_CORE &&
210  ahwthread->attrs.is_core_eff_valid() &&
211  bhwthread->attrs.is_core_eff_valid()) {
212  if (ahwthread->attrs.get_core_eff() < bhwthread->attrs.get_core_eff())
213  return 1;
214  if (ahwthread->attrs.get_core_eff() > bhwthread->attrs.get_core_eff())
215  return -1;
216  }
217  if (ahwthread->ids[level] == bhwthread->ids[level])
218  continue;
219  // If the hardware id is unknown for this level, then place hardware thread
220  // further down in the sorted list as it should take last priority
221  if (ahwthread->ids[level] == UNKNOWN_ID)
222  return 1;
223  else if (bhwthread->ids[level] == UNKNOWN_ID)
224  return -1;
225  else if (ahwthread->ids[level] < bhwthread->ids[level])
226  return -1;
227  else if (ahwthread->ids[level] > bhwthread->ids[level])
228  return 1;
229  }
230  if (ahwthread->os_id < bhwthread->os_id)
231  return -1;
232  else if (ahwthread->os_id > bhwthread->os_id)
233  return 1;
234  return 0;
235 }
236 
237 #if KMP_AFFINITY_SUPPORTED
238 int kmp_hw_thread_t::compare_compact(const void *a, const void *b) {
239  int i;
240  const kmp_hw_thread_t *aa = (const kmp_hw_thread_t *)a;
241  const kmp_hw_thread_t *bb = (const kmp_hw_thread_t *)b;
242  int depth = __kmp_topology->get_depth();
243  int compact = __kmp_topology->compact;
244  KMP_DEBUG_ASSERT(compact >= 0);
245  KMP_DEBUG_ASSERT(compact <= depth);
246  for (i = 0; i < compact; i++) {
247  int j = depth - i - 1;
248  if (aa->sub_ids[j] < bb->sub_ids[j])
249  return -1;
250  if (aa->sub_ids[j] > bb->sub_ids[j])
251  return 1;
252  }
253  for (; i < depth; i++) {
254  int j = i - compact;
255  if (aa->sub_ids[j] < bb->sub_ids[j])
256  return -1;
257  if (aa->sub_ids[j] > bb->sub_ids[j])
258  return 1;
259  }
260  return 0;
261 }
262 #endif
263 
264 void kmp_hw_thread_t::print() const {
265  int depth = __kmp_topology->get_depth();
266  printf("%4d ", os_id);
267  for (int i = 0; i < depth; ++i) {
268  printf("%4d (%d) ", ids[i], sub_ids[i]);
269  }
270  if (attrs) {
271  if (attrs.is_core_type_valid())
272  printf(" (%s)", __kmp_hw_get_core_type_string(attrs.get_core_type()));
273  if (attrs.is_core_eff_valid())
274  printf(" (eff=%d)", attrs.get_core_eff());
275  }
276  if (leader)
277  printf(" (leader)");
278  printf("\n");
279 }
280 
282 // kmp_topology_t methods
283 
284 // Add a layer to the topology based on the ids. Assume the topology
285 // is perfectly nested (i.e., so no object has more than one parent)
286 void kmp_topology_t::insert_layer(kmp_hw_t type, const int *ids) {
287  // Figure out where the layer should go by comparing the ids of the current
288  // layers with the new ids
289  int target_layer;
290  int previous_id = kmp_hw_thread_t::UNKNOWN_ID;
291  int previous_new_id = kmp_hw_thread_t::UNKNOWN_ID;
292 
293  // Start from the highest layer and work down to find target layer
294  // If new layer is equal to another layer then put the new layer above
295  for (target_layer = 0; target_layer < depth; ++target_layer) {
296  bool layers_equal = true;
297  bool strictly_above_target_layer = false;
298  for (int i = 0; i < num_hw_threads; ++i) {
299  int id = hw_threads[i].ids[target_layer];
300  int new_id = ids[i];
301  if (id != previous_id && new_id == previous_new_id) {
302  // Found the layer we are strictly above
303  strictly_above_target_layer = true;
304  layers_equal = false;
305  break;
306  } else if (id == previous_id && new_id != previous_new_id) {
307  // Found a layer we are below. Move to next layer and check.
308  layers_equal = false;
309  break;
310  }
311  previous_id = id;
312  previous_new_id = new_id;
313  }
314  if (strictly_above_target_layer || layers_equal)
315  break;
316  }
317 
318  // Found the layer we are above. Now move everything to accommodate the new
319  // layer. And put the new ids and type into the topology.
320  for (int i = depth - 1, j = depth; i >= target_layer; --i, --j)
321  types[j] = types[i];
322  types[target_layer] = type;
323  for (int k = 0; k < num_hw_threads; ++k) {
324  for (int i = depth - 1, j = depth; i >= target_layer; --i, --j)
325  hw_threads[k].ids[j] = hw_threads[k].ids[i];
326  hw_threads[k].ids[target_layer] = ids[k];
327  }
328  equivalent[type] = type;
329  depth++;
330 }
331 
332 #if KMP_GROUP_AFFINITY
333 // Insert the Windows Processor Group structure into the topology
334 void kmp_topology_t::_insert_windows_proc_groups() {
335  // Do not insert the processor group structure for a single group
336  if (__kmp_num_proc_groups == 1)
337  return;
338  kmp_affin_mask_t *mask;
339  int *ids = (int *)__kmp_allocate(sizeof(int) * num_hw_threads);
340  KMP_CPU_ALLOC(mask);
341  for (int i = 0; i < num_hw_threads; ++i) {
342  KMP_CPU_ZERO(mask);
343  KMP_CPU_SET(hw_threads[i].os_id, mask);
344  ids[i] = __kmp_get_proc_group(mask);
345  }
346  KMP_CPU_FREE(mask);
347  insert_layer(KMP_HW_PROC_GROUP, ids);
348  __kmp_free(ids);
349 
350  // sort topology after adding proc groups
351  __kmp_topology->sort_ids();
352 }
353 #endif
354 
355 // Remove layers that don't add information to the topology.
356 // This is done by having the layer take on the id = UNKNOWN_ID (-1)
357 void kmp_topology_t::_remove_radix1_layers() {
358  int preference[KMP_HW_LAST];
359  int top_index1, top_index2;
360  // Set up preference associative array
361  preference[KMP_HW_SOCKET] = 110;
362  preference[KMP_HW_PROC_GROUP] = 100;
363  preference[KMP_HW_CORE] = 95;
364  preference[KMP_HW_THREAD] = 90;
365  preference[KMP_HW_NUMA] = 85;
366  preference[KMP_HW_DIE] = 80;
367  preference[KMP_HW_TILE] = 75;
368  preference[KMP_HW_MODULE] = 73;
369  preference[KMP_HW_L3] = 70;
370  preference[KMP_HW_L2] = 65;
371  preference[KMP_HW_L1] = 60;
372  preference[KMP_HW_LLC] = 5;
373  top_index1 = 0;
374  top_index2 = 1;
375  while (top_index1 < depth - 1 && top_index2 < depth) {
376  kmp_hw_t type1 = types[top_index1];
377  kmp_hw_t type2 = types[top_index2];
378  KMP_ASSERT_VALID_HW_TYPE(type1);
379  KMP_ASSERT_VALID_HW_TYPE(type2);
380  // Do not allow the three main topology levels (sockets, cores, threads) to
381  // be compacted down
382  if ((type1 == KMP_HW_THREAD || type1 == KMP_HW_CORE ||
383  type1 == KMP_HW_SOCKET) &&
384  (type2 == KMP_HW_THREAD || type2 == KMP_HW_CORE ||
385  type2 == KMP_HW_SOCKET)) {
386  top_index1 = top_index2++;
387  continue;
388  }
389  bool radix1 = true;
390  bool all_same = true;
391  int id1 = hw_threads[0].ids[top_index1];
392  int id2 = hw_threads[0].ids[top_index2];
393  int pref1 = preference[type1];
394  int pref2 = preference[type2];
395  for (int hwidx = 1; hwidx < num_hw_threads; ++hwidx) {
396  if (hw_threads[hwidx].ids[top_index1] == id1 &&
397  hw_threads[hwidx].ids[top_index2] != id2) {
398  radix1 = false;
399  break;
400  }
401  if (hw_threads[hwidx].ids[top_index2] != id2)
402  all_same = false;
403  id1 = hw_threads[hwidx].ids[top_index1];
404  id2 = hw_threads[hwidx].ids[top_index2];
405  }
406  if (radix1) {
407  // Select the layer to remove based on preference
408  kmp_hw_t remove_type, keep_type;
409  int remove_layer, remove_layer_ids;
410  if (pref1 > pref2) {
411  remove_type = type2;
412  remove_layer = remove_layer_ids = top_index2;
413  keep_type = type1;
414  } else {
415  remove_type = type1;
416  remove_layer = remove_layer_ids = top_index1;
417  keep_type = type2;
418  }
419  // If all the indexes for the second (deeper) layer are the same.
420  // e.g., all are zero, then make sure to keep the first layer's ids
421  if (all_same)
422  remove_layer_ids = top_index2;
423  // Remove radix one type by setting the equivalence, removing the id from
424  // the hw threads and removing the layer from types and depth
425  set_equivalent_type(remove_type, keep_type);
426  for (int idx = 0; idx < num_hw_threads; ++idx) {
427  kmp_hw_thread_t &hw_thread = hw_threads[idx];
428  for (int d = remove_layer_ids; d < depth - 1; ++d)
429  hw_thread.ids[d] = hw_thread.ids[d + 1];
430  }
431  for (int idx = remove_layer; idx < depth - 1; ++idx)
432  types[idx] = types[idx + 1];
433  depth--;
434  } else {
435  top_index1 = top_index2++;
436  }
437  }
438  KMP_ASSERT(depth > 0);
439 }
440 
441 void kmp_topology_t::_set_last_level_cache() {
442  if (get_equivalent_type(KMP_HW_L3) != KMP_HW_UNKNOWN)
443  set_equivalent_type(KMP_HW_LLC, KMP_HW_L3);
444  else if (get_equivalent_type(KMP_HW_L2) != KMP_HW_UNKNOWN)
445  set_equivalent_type(KMP_HW_LLC, KMP_HW_L2);
446 #if KMP_MIC_SUPPORTED
447  else if (__kmp_mic_type == mic3) {
448  if (get_equivalent_type(KMP_HW_L2) != KMP_HW_UNKNOWN)
449  set_equivalent_type(KMP_HW_LLC, KMP_HW_L2);
450  else if (get_equivalent_type(KMP_HW_TILE) != KMP_HW_UNKNOWN)
451  set_equivalent_type(KMP_HW_LLC, KMP_HW_TILE);
452  // L2/Tile wasn't detected so just say L1
453  else
454  set_equivalent_type(KMP_HW_LLC, KMP_HW_L1);
455  }
456 #endif
457  else if (get_equivalent_type(KMP_HW_L1) != KMP_HW_UNKNOWN)
458  set_equivalent_type(KMP_HW_LLC, KMP_HW_L1);
459  // Fallback is to set last level cache to socket or core
460  if (get_equivalent_type(KMP_HW_LLC) == KMP_HW_UNKNOWN) {
461  if (get_equivalent_type(KMP_HW_SOCKET) != KMP_HW_UNKNOWN)
462  set_equivalent_type(KMP_HW_LLC, KMP_HW_SOCKET);
463  else if (get_equivalent_type(KMP_HW_CORE) != KMP_HW_UNKNOWN)
464  set_equivalent_type(KMP_HW_LLC, KMP_HW_CORE);
465  }
466  KMP_ASSERT(get_equivalent_type(KMP_HW_LLC) != KMP_HW_UNKNOWN);
467 }
468 
469 // Gather the count of each topology layer and the ratio
470 void kmp_topology_t::_gather_enumeration_information() {
471  int previous_id[KMP_HW_LAST];
472  int max[KMP_HW_LAST];
473 
474  for (int i = 0; i < depth; ++i) {
475  previous_id[i] = kmp_hw_thread_t::UNKNOWN_ID;
476  max[i] = 0;
477  count[i] = 0;
478  ratio[i] = 0;
479  }
480  int core_level = get_level(KMP_HW_CORE);
481  for (int i = 0; i < num_hw_threads; ++i) {
482  kmp_hw_thread_t &hw_thread = hw_threads[i];
483  for (int layer = 0; layer < depth; ++layer) {
484  int id = hw_thread.ids[layer];
485  if (id != previous_id[layer]) {
486  // Add an additional increment to each count
487  for (int l = layer; l < depth; ++l) {
488  if (hw_thread.ids[l] != kmp_hw_thread_t::UNKNOWN_ID)
489  count[l]++;
490  }
491  // Keep track of topology layer ratio statistics
492  if (hw_thread.ids[layer] != kmp_hw_thread_t::UNKNOWN_ID)
493  max[layer]++;
494  for (int l = layer + 1; l < depth; ++l) {
495  if (max[l] > ratio[l])
496  ratio[l] = max[l];
497  max[l] = 1;
498  }
499  // Figure out the number of different core types
500  // and efficiencies for hybrid CPUs
501  if (__kmp_is_hybrid_cpu() && core_level >= 0 && layer <= core_level) {
502  if (hw_thread.attrs.is_core_eff_valid() &&
503  hw_thread.attrs.core_eff >= num_core_efficiencies) {
504  // Because efficiencies can range from 0 to max efficiency - 1,
505  // the number of efficiencies is max efficiency + 1
506  num_core_efficiencies = hw_thread.attrs.core_eff + 1;
507  }
508  if (hw_thread.attrs.is_core_type_valid()) {
509  bool found = false;
510  for (int j = 0; j < num_core_types; ++j) {
511  if (hw_thread.attrs.get_core_type() == core_types[j]) {
512  found = true;
513  break;
514  }
515  }
516  if (!found) {
517  KMP_ASSERT(num_core_types < KMP_HW_MAX_NUM_CORE_TYPES);
518  core_types[num_core_types++] = hw_thread.attrs.get_core_type();
519  }
520  }
521  }
522  break;
523  }
524  }
525  for (int layer = 0; layer < depth; ++layer) {
526  previous_id[layer] = hw_thread.ids[layer];
527  }
528  }
529  for (int layer = 0; layer < depth; ++layer) {
530  if (max[layer] > ratio[layer])
531  ratio[layer] = max[layer];
532  }
533 }
534 
535 int kmp_topology_t::_get_ncores_with_attr(const kmp_hw_attr_t &attr,
536  int above_level,
537  bool find_all) const {
538  int current, current_max;
539  int previous_id[KMP_HW_LAST];
540  for (int i = 0; i < depth; ++i)
541  previous_id[i] = kmp_hw_thread_t::UNKNOWN_ID;
542  int core_level = get_level(KMP_HW_CORE);
543  if (find_all)
544  above_level = -1;
545  KMP_ASSERT(above_level < core_level);
546  current_max = 0;
547  current = 0;
548  for (int i = 0; i < num_hw_threads; ++i) {
549  kmp_hw_thread_t &hw_thread = hw_threads[i];
550  if (!find_all && hw_thread.ids[above_level] != previous_id[above_level]) {
551  if (current > current_max)
552  current_max = current;
553  current = hw_thread.attrs.contains(attr);
554  } else {
555  for (int level = above_level + 1; level <= core_level; ++level) {
556  if (hw_thread.ids[level] != previous_id[level]) {
557  if (hw_thread.attrs.contains(attr))
558  current++;
559  break;
560  }
561  }
562  }
563  for (int level = 0; level < depth; ++level)
564  previous_id[level] = hw_thread.ids[level];
565  }
566  if (current > current_max)
567  current_max = current;
568  return current_max;
569 }
570 
571 // Find out if the topology is uniform
572 void kmp_topology_t::_discover_uniformity() {
573  int num = 1;
574  for (int level = 0; level < depth; ++level)
575  num *= ratio[level];
576  flags.uniform = (num == count[depth - 1]);
577 }
578 
579 // Set all the sub_ids for each hardware thread
580 void kmp_topology_t::_set_sub_ids() {
581  int previous_id[KMP_HW_LAST];
582  int sub_id[KMP_HW_LAST];
583 
584  for (int i = 0; i < depth; ++i) {
585  previous_id[i] = -1;
586  sub_id[i] = -1;
587  }
588  for (int i = 0; i < num_hw_threads; ++i) {
589  kmp_hw_thread_t &hw_thread = hw_threads[i];
590  // Setup the sub_id
591  for (int j = 0; j < depth; ++j) {
592  if (hw_thread.ids[j] != previous_id[j]) {
593  sub_id[j]++;
594  for (int k = j + 1; k < depth; ++k) {
595  sub_id[k] = 0;
596  }
597  break;
598  }
599  }
600  // Set previous_id
601  for (int j = 0; j < depth; ++j) {
602  previous_id[j] = hw_thread.ids[j];
603  }
604  // Set the sub_ids field
605  for (int j = 0; j < depth; ++j) {
606  hw_thread.sub_ids[j] = sub_id[j];
607  }
608  }
609 }
610 
611 void kmp_topology_t::_set_globals() {
612  // Set nCoresPerPkg, nPackages, __kmp_nThreadsPerCore, __kmp_ncores
613  int core_level, thread_level, package_level;
614  package_level = get_level(KMP_HW_SOCKET);
615 #if KMP_GROUP_AFFINITY
616  if (package_level == -1)
617  package_level = get_level(KMP_HW_PROC_GROUP);
618 #endif
619  core_level = get_level(KMP_HW_CORE);
620  thread_level = get_level(KMP_HW_THREAD);
621 
622  KMP_ASSERT(core_level != -1);
623  KMP_ASSERT(thread_level != -1);
624 
625  __kmp_nThreadsPerCore = calculate_ratio(thread_level, core_level);
626  if (package_level != -1) {
627  nCoresPerPkg = calculate_ratio(core_level, package_level);
628  nPackages = get_count(package_level);
629  } else {
630  // assume one socket
631  nCoresPerPkg = get_count(core_level);
632  nPackages = 1;
633  }
634 #ifndef KMP_DFLT_NTH_CORES
635  __kmp_ncores = get_count(core_level);
636 #endif
637 }
638 
639 kmp_topology_t *kmp_topology_t::allocate(int nproc, int ndepth,
640  const kmp_hw_t *types) {
641  kmp_topology_t *retval;
642  // Allocate all data in one large allocation
643  size_t size = sizeof(kmp_topology_t) + sizeof(kmp_hw_thread_t) * nproc +
644  sizeof(int) * (size_t)KMP_HW_LAST * 3;
645  char *bytes = (char *)__kmp_allocate(size);
646  retval = (kmp_topology_t *)bytes;
647  if (nproc > 0) {
648  retval->hw_threads = (kmp_hw_thread_t *)(bytes + sizeof(kmp_topology_t));
649  } else {
650  retval->hw_threads = nullptr;
651  }
652  retval->num_hw_threads = nproc;
653  retval->depth = ndepth;
654  int *arr =
655  (int *)(bytes + sizeof(kmp_topology_t) + sizeof(kmp_hw_thread_t) * nproc);
656  retval->types = (kmp_hw_t *)arr;
657  retval->ratio = arr + (size_t)KMP_HW_LAST;
658  retval->count = arr + 2 * (size_t)KMP_HW_LAST;
659  retval->num_core_efficiencies = 0;
660  retval->num_core_types = 0;
661  retval->compact = 0;
662  for (int i = 0; i < KMP_HW_MAX_NUM_CORE_TYPES; ++i)
663  retval->core_types[i] = KMP_HW_CORE_TYPE_UNKNOWN;
664  KMP_FOREACH_HW_TYPE(type) { retval->equivalent[type] = KMP_HW_UNKNOWN; }
665  for (int i = 0; i < ndepth; ++i) {
666  retval->types[i] = types[i];
667  retval->equivalent[types[i]] = types[i];
668  }
669  return retval;
670 }
671 
672 void kmp_topology_t::deallocate(kmp_topology_t *topology) {
673  if (topology)
674  __kmp_free(topology);
675 }
676 
677 bool kmp_topology_t::check_ids() const {
678  // Assume ids have been sorted
679  if (num_hw_threads == 0)
680  return true;
681  for (int i = 1; i < num_hw_threads; ++i) {
682  kmp_hw_thread_t &current_thread = hw_threads[i];
683  kmp_hw_thread_t &previous_thread = hw_threads[i - 1];
684  bool unique = false;
685  for (int j = 0; j < depth; ++j) {
686  if (previous_thread.ids[j] != current_thread.ids[j]) {
687  unique = true;
688  break;
689  }
690  }
691  if (unique)
692  continue;
693  return false;
694  }
695  return true;
696 }
697 
698 void kmp_topology_t::dump() const {
699  printf("***********************\n");
700  printf("*** __kmp_topology: ***\n");
701  printf("***********************\n");
702  printf("* depth: %d\n", depth);
703 
704  printf("* types: ");
705  for (int i = 0; i < depth; ++i)
706  printf("%15s ", __kmp_hw_get_keyword(types[i]));
707  printf("\n");
708 
709  printf("* ratio: ");
710  for (int i = 0; i < depth; ++i) {
711  printf("%15d ", ratio[i]);
712  }
713  printf("\n");
714 
715  printf("* count: ");
716  for (int i = 0; i < depth; ++i) {
717  printf("%15d ", count[i]);
718  }
719  printf("\n");
720 
721  printf("* num_core_eff: %d\n", num_core_efficiencies);
722  printf("* num_core_types: %d\n", num_core_types);
723  printf("* core_types: ");
724  for (int i = 0; i < num_core_types; ++i)
725  printf("%3d ", core_types[i]);
726  printf("\n");
727 
728  printf("* equivalent map:\n");
729  KMP_FOREACH_HW_TYPE(i) {
730  const char *key = __kmp_hw_get_keyword(i);
731  const char *value = __kmp_hw_get_keyword(equivalent[i]);
732  printf("%-15s -> %-15s\n", key, value);
733  }
734 
735  printf("* uniform: %s\n", (is_uniform() ? "Yes" : "No"));
736 
737  printf("* num_hw_threads: %d\n", num_hw_threads);
738  printf("* hw_threads:\n");
739  for (int i = 0; i < num_hw_threads; ++i) {
740  hw_threads[i].print();
741  }
742  printf("***********************\n");
743 }
744 
745 void kmp_topology_t::print(const char *env_var) const {
746  kmp_str_buf_t buf;
747  int print_types_depth;
748  __kmp_str_buf_init(&buf);
749  kmp_hw_t print_types[KMP_HW_LAST + 2];
750 
751  // Num Available Threads
752  if (num_hw_threads) {
753  KMP_INFORM(AvailableOSProc, env_var, num_hw_threads);
754  } else {
755  KMP_INFORM(AvailableOSProc, env_var, __kmp_xproc);
756  }
757 
758  // Uniform or not
759  if (is_uniform()) {
760  KMP_INFORM(Uniform, env_var);
761  } else {
762  KMP_INFORM(NonUniform, env_var);
763  }
764 
765  // Equivalent types
766  KMP_FOREACH_HW_TYPE(type) {
767  kmp_hw_t eq_type = equivalent[type];
768  if (eq_type != KMP_HW_UNKNOWN && eq_type != type) {
769  KMP_INFORM(AffEqualTopologyTypes, env_var,
770  __kmp_hw_get_catalog_string(type),
771  __kmp_hw_get_catalog_string(eq_type));
772  }
773  }
774 
775  // Quick topology
776  KMP_ASSERT(depth > 0 && depth <= (int)KMP_HW_LAST);
777  // Create a print types array that always guarantees printing
778  // the core and thread level
779  print_types_depth = 0;
780  for (int level = 0; level < depth; ++level)
781  print_types[print_types_depth++] = types[level];
782  if (equivalent[KMP_HW_CORE] != KMP_HW_CORE) {
783  // Force in the core level for quick topology
784  if (print_types[print_types_depth - 1] == KMP_HW_THREAD) {
785  // Force core before thread e.g., 1 socket X 2 threads/socket
786  // becomes 1 socket X 1 core/socket X 2 threads/socket
787  print_types[print_types_depth - 1] = KMP_HW_CORE;
788  print_types[print_types_depth++] = KMP_HW_THREAD;
789  } else {
790  print_types[print_types_depth++] = KMP_HW_CORE;
791  }
792  }
793  // Always put threads at very end of quick topology
794  if (equivalent[KMP_HW_THREAD] != KMP_HW_THREAD)
795  print_types[print_types_depth++] = KMP_HW_THREAD;
796 
797  __kmp_str_buf_clear(&buf);
798  kmp_hw_t numerator_type;
799  kmp_hw_t denominator_type = KMP_HW_UNKNOWN;
800  int core_level = get_level(KMP_HW_CORE);
801  int ncores = get_count(core_level);
802 
803  for (int plevel = 0, level = 0; plevel < print_types_depth; ++plevel) {
804  int c;
805  bool plural;
806  numerator_type = print_types[plevel];
807  KMP_ASSERT_VALID_HW_TYPE(numerator_type);
808  if (equivalent[numerator_type] != numerator_type)
809  c = 1;
810  else
811  c = get_ratio(level++);
812  plural = (c > 1);
813  if (plevel == 0) {
814  __kmp_str_buf_print(&buf, "%d %s", c,
815  __kmp_hw_get_catalog_string(numerator_type, plural));
816  } else {
817  __kmp_str_buf_print(&buf, " x %d %s/%s", c,
818  __kmp_hw_get_catalog_string(numerator_type, plural),
819  __kmp_hw_get_catalog_string(denominator_type));
820  }
821  denominator_type = numerator_type;
822  }
823  KMP_INFORM(TopologyGeneric, env_var, buf.str, ncores);
824 
825  // Hybrid topology information
826  if (__kmp_is_hybrid_cpu()) {
827  for (int i = 0; i < num_core_types; ++i) {
828  kmp_hw_core_type_t core_type = core_types[i];
829  kmp_hw_attr_t attr;
830  attr.clear();
831  attr.set_core_type(core_type);
832  int ncores = get_ncores_with_attr(attr);
833  if (ncores > 0) {
834  KMP_INFORM(TopologyHybrid, env_var, ncores,
835  __kmp_hw_get_core_type_string(core_type));
836  KMP_ASSERT(num_core_efficiencies <= KMP_HW_MAX_NUM_CORE_EFFS)
837  for (int eff = 0; eff < num_core_efficiencies; ++eff) {
838  attr.set_core_eff(eff);
839  int ncores_with_eff = get_ncores_with_attr(attr);
840  if (ncores_with_eff > 0) {
841  KMP_INFORM(TopologyHybridCoreEff, env_var, ncores_with_eff, eff);
842  }
843  }
844  }
845  }
846  }
847 
848  if (num_hw_threads <= 0) {
849  __kmp_str_buf_free(&buf);
850  return;
851  }
852 
853  // Full OS proc to hardware thread map
854  KMP_INFORM(OSProcToPhysicalThreadMap, env_var);
855  for (int i = 0; i < num_hw_threads; i++) {
856  __kmp_str_buf_clear(&buf);
857  for (int level = 0; level < depth; ++level) {
858  if (hw_threads[i].ids[level] == kmp_hw_thread_t::UNKNOWN_ID)
859  continue;
860  kmp_hw_t type = types[level];
861  __kmp_str_buf_print(&buf, "%s ", __kmp_hw_get_catalog_string(type));
862  __kmp_str_buf_print(&buf, "%d ", hw_threads[i].ids[level]);
863  }
864  if (__kmp_is_hybrid_cpu())
865  __kmp_str_buf_print(
866  &buf, "(%s)",
867  __kmp_hw_get_core_type_string(hw_threads[i].attrs.get_core_type()));
868  KMP_INFORM(OSProcMapToPack, env_var, hw_threads[i].os_id, buf.str);
869  }
870 
871  __kmp_str_buf_free(&buf);
872 }
873 
874 #if KMP_AFFINITY_SUPPORTED
875 void kmp_topology_t::set_granularity(kmp_affinity_t &affinity) const {
876  const char *env_var = __kmp_get_affinity_env_var(affinity);
877  // If requested hybrid CPU attributes for granularity (either OMP_PLACES or
878  // KMP_AFFINITY), but none exist, then reset granularity and have below method
879  // select a granularity and warn user.
880  if (!__kmp_is_hybrid_cpu()) {
881  if (affinity.core_attr_gran.valid) {
882  // OMP_PLACES with cores:<attribute> but non-hybrid arch, use cores
883  // instead
884  KMP_AFF_WARNING(
885  affinity, AffIgnoringNonHybrid, env_var,
886  __kmp_hw_get_catalog_string(KMP_HW_CORE, /*plural=*/true));
887  affinity.gran = KMP_HW_CORE;
888  affinity.gran_levels = -1;
889  affinity.core_attr_gran = KMP_AFFINITY_ATTRS_UNKNOWN;
890  affinity.flags.core_types_gran = affinity.flags.core_effs_gran = 0;
891  } else if (affinity.flags.core_types_gran ||
892  affinity.flags.core_effs_gran) {
893  // OMP_PLACES=core_types|core_effs but non-hybrid, use cores instead
894  if (affinity.flags.omp_places) {
895  KMP_AFF_WARNING(
896  affinity, AffIgnoringNonHybrid, env_var,
897  __kmp_hw_get_catalog_string(KMP_HW_CORE, /*plural=*/true));
898  } else {
899  // KMP_AFFINITY=granularity=core_type|core_eff,...
900  KMP_AFF_WARNING(affinity, AffGranularityBad, env_var,
901  "Intel(R) Hybrid Technology core attribute",
902  __kmp_hw_get_catalog_string(KMP_HW_CORE));
903  }
904  affinity.gran = KMP_HW_CORE;
905  affinity.gran_levels = -1;
906  affinity.core_attr_gran = KMP_AFFINITY_ATTRS_UNKNOWN;
907  affinity.flags.core_types_gran = affinity.flags.core_effs_gran = 0;
908  }
909  }
910  // Set the number of affinity granularity levels
911  if (affinity.gran_levels < 0) {
912  kmp_hw_t gran_type = get_equivalent_type(affinity.gran);
913  // Check if user's granularity request is valid
914  if (gran_type == KMP_HW_UNKNOWN) {
915  // First try core, then thread, then package
916  kmp_hw_t gran_types[3] = {KMP_HW_CORE, KMP_HW_THREAD, KMP_HW_SOCKET};
917  for (auto g : gran_types) {
918  if (get_equivalent_type(g) != KMP_HW_UNKNOWN) {
919  gran_type = g;
920  break;
921  }
922  }
923  KMP_ASSERT(gran_type != KMP_HW_UNKNOWN);
924  // Warn user what granularity setting will be used instead
925  KMP_AFF_WARNING(affinity, AffGranularityBad, env_var,
926  __kmp_hw_get_catalog_string(affinity.gran),
927  __kmp_hw_get_catalog_string(gran_type));
928  affinity.gran = gran_type;
929  }
930 #if KMP_GROUP_AFFINITY
931  // If more than one processor group exists, and the level of
932  // granularity specified by the user is too coarse, then the
933  // granularity must be adjusted "down" to processor group affinity
934  // because threads can only exist within one processor group.
935  // For example, if a user sets granularity=socket and there are two
936  // processor groups that cover a socket, then the runtime must
937  // restrict the granularity down to the processor group level.
938  if (__kmp_num_proc_groups > 1) {
939  int gran_depth = get_level(gran_type);
940  int proc_group_depth = get_level(KMP_HW_PROC_GROUP);
941  if (gran_depth >= 0 && proc_group_depth >= 0 &&
942  gran_depth < proc_group_depth) {
943  KMP_AFF_WARNING(affinity, AffGranTooCoarseProcGroup, env_var,
944  __kmp_hw_get_catalog_string(affinity.gran));
945  affinity.gran = gran_type = KMP_HW_PROC_GROUP;
946  }
947  }
948 #endif
949  affinity.gran_levels = 0;
950  for (int i = depth - 1; i >= 0 && get_type(i) != gran_type; --i)
951  affinity.gran_levels++;
952  }
953 }
954 #endif
955 
956 void kmp_topology_t::canonicalize() {
957 #if KMP_GROUP_AFFINITY
958  _insert_windows_proc_groups();
959 #endif
960  _remove_radix1_layers();
961  _gather_enumeration_information();
962  _discover_uniformity();
963  _set_sub_ids();
964  _set_globals();
965  _set_last_level_cache();
966 
967 #if KMP_MIC_SUPPORTED
968  // Manually Add L2 = Tile equivalence
969  if (__kmp_mic_type == mic3) {
970  if (get_level(KMP_HW_L2) != -1)
971  set_equivalent_type(KMP_HW_TILE, KMP_HW_L2);
972  else if (get_level(KMP_HW_TILE) != -1)
973  set_equivalent_type(KMP_HW_L2, KMP_HW_TILE);
974  }
975 #endif
976 
977  // Perform post canonicalization checking
978  KMP_ASSERT(depth > 0);
979  for (int level = 0; level < depth; ++level) {
980  // All counts, ratios, and types must be valid
981  KMP_ASSERT(count[level] > 0 && ratio[level] > 0);
982  KMP_ASSERT_VALID_HW_TYPE(types[level]);
983  // Detected types must point to themselves
984  KMP_ASSERT(equivalent[types[level]] == types[level]);
985  }
986 }
987 
988 // Canonicalize an explicit packages X cores/pkg X threads/core topology
989 void kmp_topology_t::canonicalize(int npackages, int ncores_per_pkg,
990  int nthreads_per_core, int ncores) {
991  int ndepth = 3;
992  depth = ndepth;
993  KMP_FOREACH_HW_TYPE(i) { equivalent[i] = KMP_HW_UNKNOWN; }
994  for (int level = 0; level < depth; ++level) {
995  count[level] = 0;
996  ratio[level] = 0;
997  }
998  count[0] = npackages;
999  count[1] = ncores;
1000  count[2] = __kmp_xproc;
1001  ratio[0] = npackages;
1002  ratio[1] = ncores_per_pkg;
1003  ratio[2] = nthreads_per_core;
1004  equivalent[KMP_HW_SOCKET] = KMP_HW_SOCKET;
1005  equivalent[KMP_HW_CORE] = KMP_HW_CORE;
1006  equivalent[KMP_HW_THREAD] = KMP_HW_THREAD;
1007  types[0] = KMP_HW_SOCKET;
1008  types[1] = KMP_HW_CORE;
1009  types[2] = KMP_HW_THREAD;
1010  //__kmp_avail_proc = __kmp_xproc;
1011  _discover_uniformity();
1012 }
1013 
1014 #if KMP_AFFINITY_SUPPORTED
1015 static kmp_str_buf_t *
1016 __kmp_hw_get_catalog_core_string(const kmp_hw_attr_t &attr, kmp_str_buf_t *buf,
1017  bool plural) {
1018  __kmp_str_buf_init(buf);
1019  if (attr.is_core_type_valid())
1020  __kmp_str_buf_print(buf, "%s %s",
1021  __kmp_hw_get_core_type_string(attr.get_core_type()),
1022  __kmp_hw_get_catalog_string(KMP_HW_CORE, plural));
1023  else
1024  __kmp_str_buf_print(buf, "%s eff=%d",
1025  __kmp_hw_get_catalog_string(KMP_HW_CORE, plural),
1026  attr.get_core_eff());
1027  return buf;
1028 }
1029 
1030 bool kmp_topology_t::restrict_to_mask(const kmp_affin_mask_t *mask) {
1031  // Apply the filter
1032  bool affected;
1033  int new_index = 0;
1034  for (int i = 0; i < num_hw_threads; ++i) {
1035  int os_id = hw_threads[i].os_id;
1036  if (KMP_CPU_ISSET(os_id, mask)) {
1037  if (i != new_index)
1038  hw_threads[new_index] = hw_threads[i];
1039  new_index++;
1040  } else {
1041  KMP_CPU_CLR(os_id, __kmp_affin_fullMask);
1042  __kmp_avail_proc--;
1043  }
1044  }
1045 
1046  KMP_DEBUG_ASSERT(new_index <= num_hw_threads);
1047  affected = (num_hw_threads != new_index);
1048  num_hw_threads = new_index;
1049 
1050  // Post hardware subset canonicalization
1051  if (affected) {
1052  _gather_enumeration_information();
1053  _discover_uniformity();
1054  _set_globals();
1055  _set_last_level_cache();
1056 #if KMP_OS_WINDOWS
1057  // Copy filtered full mask if topology has single processor group
1058  if (__kmp_num_proc_groups <= 1)
1059 #endif
1060  __kmp_affin_origMask->copy(__kmp_affin_fullMask);
1061  }
1062  return affected;
1063 }
1064 
1065 // Apply the KMP_HW_SUBSET envirable to the topology
1066 // Returns true if KMP_HW_SUBSET filtered any processors
1067 // otherwise, returns false
1068 bool kmp_topology_t::filter_hw_subset() {
1069  // If KMP_HW_SUBSET wasn't requested, then do nothing.
1070  if (!__kmp_hw_subset)
1071  return false;
1072 
1073  // First, sort the KMP_HW_SUBSET items by the machine topology
1074  __kmp_hw_subset->sort();
1075 
1076  __kmp_hw_subset->canonicalize(__kmp_topology);
1077 
1078  // Check to see if KMP_HW_SUBSET is a valid subset of the detected topology
1079  bool using_core_types = false;
1080  bool using_core_effs = false;
1081  bool is_absolute = __kmp_hw_subset->is_absolute();
1082  int hw_subset_depth = __kmp_hw_subset->get_depth();
1083  kmp_hw_t specified[KMP_HW_LAST];
1084  int *topology_levels = (int *)KMP_ALLOCA(sizeof(int) * hw_subset_depth);
1085  KMP_ASSERT(hw_subset_depth > 0);
1086  KMP_FOREACH_HW_TYPE(i) { specified[i] = KMP_HW_UNKNOWN; }
1087  int core_level = get_level(KMP_HW_CORE);
1088  for (int i = 0; i < hw_subset_depth; ++i) {
1089  int max_count;
1090  const kmp_hw_subset_t::item_t &item = __kmp_hw_subset->at(i);
1091  int num = item.num[0];
1092  int offset = item.offset[0];
1093  kmp_hw_t type = item.type;
1094  kmp_hw_t equivalent_type = equivalent[type];
1095  int level = get_level(type);
1096  topology_levels[i] = level;
1097 
1098  // Check to see if current layer is in detected machine topology
1099  if (equivalent_type != KMP_HW_UNKNOWN) {
1100  __kmp_hw_subset->at(i).type = equivalent_type;
1101  } else {
1102  KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetNotExistGeneric,
1103  __kmp_hw_get_catalog_string(type));
1104  return false;
1105  }
1106 
1107  // Check to see if current layer has already been
1108  // specified either directly or through an equivalent type
1109  if (specified[equivalent_type] != KMP_HW_UNKNOWN) {
1110  KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetEqvLayers,
1111  __kmp_hw_get_catalog_string(type),
1112  __kmp_hw_get_catalog_string(specified[equivalent_type]));
1113  return false;
1114  }
1115  specified[equivalent_type] = type;
1116 
1117  // Check to see if each layer's num & offset parameters are valid
1118  max_count = get_ratio(level);
1119  if (!is_absolute) {
1120  if (max_count < 0 ||
1121  (num != kmp_hw_subset_t::USE_ALL && num + offset > max_count)) {
1122  bool plural = (num > 1);
1123  KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetManyGeneric,
1124  __kmp_hw_get_catalog_string(type, plural));
1125  return false;
1126  }
1127  }
1128 
1129  // Check to see if core attributes are consistent
1130  if (core_level == level) {
1131  // Determine which core attributes are specified
1132  for (int j = 0; j < item.num_attrs; ++j) {
1133  if (item.attr[j].is_core_type_valid())
1134  using_core_types = true;
1135  if (item.attr[j].is_core_eff_valid())
1136  using_core_effs = true;
1137  }
1138 
1139  // Check if using a single core attribute on non-hybrid arch.
1140  // Do not ignore all of KMP_HW_SUBSET, just ignore the attribute.
1141  //
1142  // Check if using multiple core attributes on non-hyrbid arch.
1143  // Ignore all of KMP_HW_SUBSET if this is the case.
1144  if ((using_core_effs || using_core_types) && !__kmp_is_hybrid_cpu()) {
1145  if (item.num_attrs == 1) {
1146  if (using_core_effs) {
1147  KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetIgnoringAttr,
1148  "efficiency");
1149  } else {
1150  KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetIgnoringAttr,
1151  "core_type");
1152  }
1153  using_core_effs = false;
1154  using_core_types = false;
1155  } else {
1156  KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetAttrsNonHybrid);
1157  return false;
1158  }
1159  }
1160 
1161  // Check if using both core types and core efficiencies together
1162  if (using_core_types && using_core_effs) {
1163  KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetIncompat, "core_type",
1164  "efficiency");
1165  return false;
1166  }
1167 
1168  // Check that core efficiency values are valid
1169  if (using_core_effs) {
1170  for (int j = 0; j < item.num_attrs; ++j) {
1171  if (item.attr[j].is_core_eff_valid()) {
1172  int core_eff = item.attr[j].get_core_eff();
1173  if (core_eff < 0 || core_eff >= num_core_efficiencies) {
1174  kmp_str_buf_t buf;
1175  __kmp_str_buf_init(&buf);
1176  __kmp_str_buf_print(&buf, "%d", item.attr[j].get_core_eff());
1177  __kmp_msg(kmp_ms_warning,
1178  KMP_MSG(AffHWSubsetAttrInvalid, "efficiency", buf.str),
1179  KMP_HNT(ValidValuesRange, 0, num_core_efficiencies - 1),
1180  __kmp_msg_null);
1181  __kmp_str_buf_free(&buf);
1182  return false;
1183  }
1184  }
1185  }
1186  }
1187 
1188  // Check that the number of requested cores with attributes is valid
1189  if ((using_core_types || using_core_effs) && !is_absolute) {
1190  for (int j = 0; j < item.num_attrs; ++j) {
1191  int num = item.num[j];
1192  int offset = item.offset[j];
1193  int level_above = core_level - 1;
1194  if (level_above >= 0) {
1195  max_count = get_ncores_with_attr_per(item.attr[j], level_above);
1196  if (max_count <= 0 ||
1197  (num != kmp_hw_subset_t::USE_ALL && num + offset > max_count)) {
1198  kmp_str_buf_t buf;
1199  __kmp_hw_get_catalog_core_string(item.attr[j], &buf, num > 0);
1200  KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetManyGeneric, buf.str);
1201  __kmp_str_buf_free(&buf);
1202  return false;
1203  }
1204  }
1205  }
1206  }
1207 
1208  if ((using_core_types || using_core_effs) && item.num_attrs > 1) {
1209  for (int j = 0; j < item.num_attrs; ++j) {
1210  // Ambiguous use of specific core attribute + generic core
1211  // e.g., 4c & 3c:intel_core or 4c & 3c:eff1
1212  if (!item.attr[j]) {
1213  kmp_hw_attr_t other_attr;
1214  for (int k = 0; k < item.num_attrs; ++k) {
1215  if (item.attr[k] != item.attr[j]) {
1216  other_attr = item.attr[k];
1217  break;
1218  }
1219  }
1220  kmp_str_buf_t buf;
1221  __kmp_hw_get_catalog_core_string(other_attr, &buf, item.num[j] > 0);
1222  KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetIncompat,
1223  __kmp_hw_get_catalog_string(KMP_HW_CORE), buf.str);
1224  __kmp_str_buf_free(&buf);
1225  return false;
1226  }
1227  // Allow specifying a specific core type or core eff exactly once
1228  for (int k = 0; k < j; ++k) {
1229  if (!item.attr[j] || !item.attr[k])
1230  continue;
1231  if (item.attr[k] == item.attr[j]) {
1232  kmp_str_buf_t buf;
1233  __kmp_hw_get_catalog_core_string(item.attr[j], &buf,
1234  item.num[j] > 0);
1235  KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetAttrRepeat, buf.str);
1236  __kmp_str_buf_free(&buf);
1237  return false;
1238  }
1239  }
1240  }
1241  }
1242  }
1243  }
1244 
1245  // For keeping track of sub_ids for an absolute KMP_HW_SUBSET
1246  // or core attributes (core type or efficiency)
1247  int prev_sub_ids[KMP_HW_LAST];
1248  int abs_sub_ids[KMP_HW_LAST];
1249  int core_eff_sub_ids[KMP_HW_MAX_NUM_CORE_EFFS];
1250  int core_type_sub_ids[KMP_HW_MAX_NUM_CORE_TYPES];
1251  for (size_t i = 0; i < KMP_HW_LAST; ++i) {
1252  abs_sub_ids[i] = -1;
1253  prev_sub_ids[i] = -1;
1254  }
1255  for (size_t i = 0; i < KMP_HW_MAX_NUM_CORE_EFFS; ++i)
1256  core_eff_sub_ids[i] = -1;
1257  for (size_t i = 0; i < KMP_HW_MAX_NUM_CORE_TYPES; ++i)
1258  core_type_sub_ids[i] = -1;
1259 
1260  // Determine which hardware threads should be filtered.
1261 
1262  // Helpful to determine if a topology layer is targeted by an absolute subset
1263  auto is_targeted = [&](int level) {
1264  if (is_absolute) {
1265  for (int i = 0; i < hw_subset_depth; ++i)
1266  if (topology_levels[i] == level)
1267  return true;
1268  return false;
1269  }
1270  // If not absolute KMP_HW_SUBSET, then every layer is seen as targeted
1271  return true;
1272  };
1273 
1274  // Helpful to index into core type sub Ids array
1275  auto get_core_type_index = [](const kmp_hw_thread_t &t) {
1276  switch (t.attrs.get_core_type()) {
1277  case KMP_HW_CORE_TYPE_UNKNOWN:
1278  case KMP_HW_MAX_NUM_CORE_TYPES:
1279  return 0;
1280 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1281  case KMP_HW_CORE_TYPE_ATOM:
1282  return 1;
1283  case KMP_HW_CORE_TYPE_CORE:
1284  return 2;
1285 #endif
1286  }
1287  KMP_ASSERT2(false, "Unhandled kmp_hw_thread_t enumeration");
1288  KMP_BUILTIN_UNREACHABLE;
1289  };
1290 
1291  // Helpful to index into core efficiencies sub Ids array
1292  auto get_core_eff_index = [](const kmp_hw_thread_t &t) {
1293  return t.attrs.get_core_eff();
1294  };
1295 
1296  int num_filtered = 0;
1297  kmp_affin_mask_t *filtered_mask;
1298  KMP_CPU_ALLOC(filtered_mask);
1299  KMP_CPU_COPY(filtered_mask, __kmp_affin_fullMask);
1300  for (int i = 0; i < num_hw_threads; ++i) {
1301  kmp_hw_thread_t &hw_thread = hw_threads[i];
1302 
1303  // Figure out the absolute sub ids and core eff/type sub ids
1304  if (is_absolute || using_core_effs || using_core_types) {
1305  for (int level = 0; level < get_depth(); ++level) {
1306  if (hw_thread.sub_ids[level] != prev_sub_ids[level]) {
1307  bool found_targeted = false;
1308  for (int j = level; j < get_depth(); ++j) {
1309  bool targeted = is_targeted(j);
1310  if (!found_targeted && targeted) {
1311  found_targeted = true;
1312  abs_sub_ids[j]++;
1313  if (j == core_level && using_core_effs)
1314  core_eff_sub_ids[get_core_eff_index(hw_thread)]++;
1315  if (j == core_level && using_core_types)
1316  core_type_sub_ids[get_core_type_index(hw_thread)]++;
1317  } else if (targeted) {
1318  abs_sub_ids[j] = 0;
1319  if (j == core_level && using_core_effs)
1320  core_eff_sub_ids[get_core_eff_index(hw_thread)] = 0;
1321  if (j == core_level && using_core_types)
1322  core_type_sub_ids[get_core_type_index(hw_thread)] = 0;
1323  }
1324  }
1325  break;
1326  }
1327  }
1328  for (int level = 0; level < get_depth(); ++level)
1329  prev_sub_ids[level] = hw_thread.sub_ids[level];
1330  }
1331 
1332  // Check to see if this hardware thread should be filtered
1333  bool should_be_filtered = false;
1334  for (int hw_subset_index = 0; hw_subset_index < hw_subset_depth;
1335  ++hw_subset_index) {
1336  const auto &hw_subset_item = __kmp_hw_subset->at(hw_subset_index);
1337  int level = topology_levels[hw_subset_index];
1338  if (level == -1)
1339  continue;
1340  if ((using_core_effs || using_core_types) && level == core_level) {
1341  // Look for the core attribute in KMP_HW_SUBSET which corresponds
1342  // to this hardware thread's core attribute. Use this num,offset plus
1343  // the running sub_id for the particular core attribute of this hardware
1344  // thread to determine if the hardware thread should be filtered or not.
1345  int attr_idx;
1346  kmp_hw_core_type_t core_type = hw_thread.attrs.get_core_type();
1347  int core_eff = hw_thread.attrs.get_core_eff();
1348  for (attr_idx = 0; attr_idx < hw_subset_item.num_attrs; ++attr_idx) {
1349  if (using_core_types &&
1350  hw_subset_item.attr[attr_idx].get_core_type() == core_type)
1351  break;
1352  if (using_core_effs &&
1353  hw_subset_item.attr[attr_idx].get_core_eff() == core_eff)
1354  break;
1355  }
1356  // This core attribute isn't in the KMP_HW_SUBSET so always filter it.
1357  if (attr_idx == hw_subset_item.num_attrs) {
1358  should_be_filtered = true;
1359  break;
1360  }
1361  int sub_id;
1362  int num = hw_subset_item.num[attr_idx];
1363  int offset = hw_subset_item.offset[attr_idx];
1364  if (using_core_types)
1365  sub_id = core_type_sub_ids[get_core_type_index(hw_thread)];
1366  else
1367  sub_id = core_eff_sub_ids[get_core_eff_index(hw_thread)];
1368  if (sub_id < offset ||
1369  (num != kmp_hw_subset_t::USE_ALL && sub_id >= offset + num)) {
1370  should_be_filtered = true;
1371  break;
1372  }
1373  } else {
1374  int sub_id;
1375  int num = hw_subset_item.num[0];
1376  int offset = hw_subset_item.offset[0];
1377  if (is_absolute)
1378  sub_id = abs_sub_ids[level];
1379  else
1380  sub_id = hw_thread.sub_ids[level];
1381  if (hw_thread.ids[level] == kmp_hw_thread_t::UNKNOWN_ID ||
1382  sub_id < offset ||
1383  (num != kmp_hw_subset_t::USE_ALL && sub_id >= offset + num)) {
1384  should_be_filtered = true;
1385  break;
1386  }
1387  }
1388  }
1389  // Collect filtering information
1390  if (should_be_filtered) {
1391  KMP_CPU_CLR(hw_thread.os_id, filtered_mask);
1392  num_filtered++;
1393  }
1394  }
1395 
1396  // One last check that we shouldn't allow filtering entire machine
1397  if (num_filtered == num_hw_threads) {
1398  KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetAllFiltered);
1399  KMP_CPU_FREE(filtered_mask);
1400  return false;
1401  }
1402 
1403  // Apply the filter
1404  restrict_to_mask(filtered_mask);
1405  KMP_CPU_FREE(filtered_mask);
1406  return true;
1407 }
1408 
1409 bool kmp_topology_t::is_close(int hwt1, int hwt2,
1410  const kmp_affinity_t &stgs) const {
1411  int hw_level = stgs.gran_levels;
1412  if (hw_level >= depth)
1413  return true;
1414  bool retval = true;
1415  const kmp_hw_thread_t &t1 = hw_threads[hwt1];
1416  const kmp_hw_thread_t &t2 = hw_threads[hwt2];
1417  if (stgs.flags.core_types_gran)
1418  return t1.attrs.get_core_type() == t2.attrs.get_core_type();
1419  if (stgs.flags.core_effs_gran)
1420  return t1.attrs.get_core_eff() == t2.attrs.get_core_eff();
1421  for (int i = 0; i < (depth - hw_level); ++i) {
1422  if (t1.ids[i] != t2.ids[i])
1423  return false;
1424  }
1425  return retval;
1426 }
1427 
1429 
1430 bool KMPAffinity::picked_api = false;
1431 
1432 void *KMPAffinity::Mask::operator new(size_t n) { return __kmp_allocate(n); }
1433 void *KMPAffinity::Mask::operator new[](size_t n) { return __kmp_allocate(n); }
1434 void KMPAffinity::Mask::operator delete(void *p) { __kmp_free(p); }
1435 void KMPAffinity::Mask::operator delete[](void *p) { __kmp_free(p); }
1436 void *KMPAffinity::operator new(size_t n) { return __kmp_allocate(n); }
1437 void KMPAffinity::operator delete(void *p) { __kmp_free(p); }
1438 
1439 void KMPAffinity::pick_api() {
1440  KMPAffinity *affinity_dispatch;
1441  if (picked_api)
1442  return;
1443 #if KMP_USE_HWLOC
1444  // Only use Hwloc if affinity isn't explicitly disabled and
1445  // user requests Hwloc topology method
1446  if (__kmp_affinity_top_method == affinity_top_method_hwloc &&
1447  __kmp_affinity.type != affinity_disabled) {
1448  affinity_dispatch = new KMPHwlocAffinity();
1449  __kmp_hwloc_available = true;
1450  } else
1451 #endif
1452  {
1453  affinity_dispatch = new KMPNativeAffinity();
1454  }
1455  __kmp_affinity_dispatch = affinity_dispatch;
1456  picked_api = true;
1457 }
1458 
1459 void KMPAffinity::destroy_api() {
1460  if (__kmp_affinity_dispatch != NULL) {
1461  delete __kmp_affinity_dispatch;
1462  __kmp_affinity_dispatch = NULL;
1463  picked_api = false;
1464  }
1465 }
1466 
1467 #define KMP_ADVANCE_SCAN(scan) \
1468  while (*scan != '\0') { \
1469  scan++; \
1470  }
1471 
1472 // Print the affinity mask to the character array in a pretty format.
1473 // The format is a comma separated list of non-negative integers or integer
1474 // ranges: e.g., 1,2,3-5,7,9-15
1475 // The format can also be the string "{<empty>}" if no bits are set in mask
1476 char *__kmp_affinity_print_mask(char *buf, int buf_len,
1477  kmp_affin_mask_t *mask) {
1478  int start = 0, finish = 0, previous = 0;
1479  bool first_range;
1480  KMP_ASSERT(buf);
1481  KMP_ASSERT(buf_len >= 40);
1482  KMP_ASSERT(mask);
1483  char *scan = buf;
1484  char *end = buf + buf_len - 1;
1485 
1486  // Check for empty set.
1487  if (mask->begin() == mask->end()) {
1488  KMP_SNPRINTF(scan, end - scan + 1, "{<empty>}");
1489  KMP_ADVANCE_SCAN(scan);
1490  KMP_ASSERT(scan <= end);
1491  return buf;
1492  }
1493 
1494  first_range = true;
1495  start = mask->begin();
1496  while (1) {
1497  // Find next range
1498  // [start, previous] is inclusive range of contiguous bits in mask
1499  for (finish = mask->next(start), previous = start;
1500  finish == previous + 1 && finish != mask->end();
1501  finish = mask->next(finish)) {
1502  previous = finish;
1503  }
1504 
1505  // The first range does not need a comma printed before it, but the rest
1506  // of the ranges do need a comma beforehand
1507  if (!first_range) {
1508  KMP_SNPRINTF(scan, end - scan + 1, "%s", ",");
1509  KMP_ADVANCE_SCAN(scan);
1510  } else {
1511  first_range = false;
1512  }
1513  // Range with three or more contiguous bits in the affinity mask
1514  if (previous - start > 1) {
1515  KMP_SNPRINTF(scan, end - scan + 1, "%u-%u", start, previous);
1516  } else {
1517  // Range with one or two contiguous bits in the affinity mask
1518  KMP_SNPRINTF(scan, end - scan + 1, "%u", start);
1519  KMP_ADVANCE_SCAN(scan);
1520  if (previous - start > 0) {
1521  KMP_SNPRINTF(scan, end - scan + 1, ",%u", previous);
1522  }
1523  }
1524  KMP_ADVANCE_SCAN(scan);
1525  // Start over with new start point
1526  start = finish;
1527  if (start == mask->end())
1528  break;
1529  // Check for overflow
1530  if (end - scan < 2)
1531  break;
1532  }
1533 
1534  // Check for overflow
1535  KMP_ASSERT(scan <= end);
1536  return buf;
1537 }
1538 #undef KMP_ADVANCE_SCAN
1539 
1540 // Print the affinity mask to the string buffer object in a pretty format
1541 // The format is a comma separated list of non-negative integers or integer
1542 // ranges: e.g., 1,2,3-5,7,9-15
1543 // The format can also be the string "{<empty>}" if no bits are set in mask
1544 kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
1545  kmp_affin_mask_t *mask) {
1546  int start = 0, finish = 0, previous = 0;
1547  bool first_range;
1548  KMP_ASSERT(buf);
1549  KMP_ASSERT(mask);
1550 
1551  __kmp_str_buf_clear(buf);
1552 
1553  // Check for empty set.
1554  if (mask->begin() == mask->end()) {
1555  __kmp_str_buf_print(buf, "%s", "{<empty>}");
1556  return buf;
1557  }
1558 
1559  first_range = true;
1560  start = mask->begin();
1561  while (1) {
1562  // Find next range
1563  // [start, previous] is inclusive range of contiguous bits in mask
1564  for (finish = mask->next(start), previous = start;
1565  finish == previous + 1 && finish != mask->end();
1566  finish = mask->next(finish)) {
1567  previous = finish;
1568  }
1569 
1570  // The first range does not need a comma printed before it, but the rest
1571  // of the ranges do need a comma beforehand
1572  if (!first_range) {
1573  __kmp_str_buf_print(buf, "%s", ",");
1574  } else {
1575  first_range = false;
1576  }
1577  // Range with three or more contiguous bits in the affinity mask
1578  if (previous - start > 1) {
1579  __kmp_str_buf_print(buf, "%u-%u", start, previous);
1580  } else {
1581  // Range with one or two contiguous bits in the affinity mask
1582  __kmp_str_buf_print(buf, "%u", start);
1583  if (previous - start > 0) {
1584  __kmp_str_buf_print(buf, ",%u", previous);
1585  }
1586  }
1587  // Start over with new start point
1588  start = finish;
1589  if (start == mask->end())
1590  break;
1591  }
1592  return buf;
1593 }
1594 
1595 static kmp_affin_mask_t *__kmp_parse_cpu_list(const char *path) {
1596  kmp_affin_mask_t *mask;
1597  KMP_CPU_ALLOC(mask);
1598  KMP_CPU_ZERO(mask);
1599 #if KMP_OS_LINUX
1600  int n, begin_cpu, end_cpu;
1601  kmp_safe_raii_file_t file;
1602  auto skip_ws = [](FILE *f) {
1603  int c;
1604  do {
1605  c = fgetc(f);
1606  } while (isspace(c));
1607  if (c != EOF)
1608  ungetc(c, f);
1609  };
1610  // File contains CSV of integer ranges representing the CPUs
1611  // e.g., 1,2,4-7,9,11-15
1612  int status = file.try_open(path, "r");
1613  if (status != 0)
1614  return mask;
1615  while (!feof(file)) {
1616  skip_ws(file);
1617  n = fscanf(file, "%d", &begin_cpu);
1618  if (n != 1)
1619  break;
1620  skip_ws(file);
1621  int c = fgetc(file);
1622  if (c == EOF || c == ',') {
1623  // Just single CPU
1624  end_cpu = begin_cpu;
1625  } else if (c == '-') {
1626  // Range of CPUs
1627  skip_ws(file);
1628  n = fscanf(file, "%d", &end_cpu);
1629  if (n != 1)
1630  break;
1631  skip_ws(file);
1632  c = fgetc(file); // skip ','
1633  } else {
1634  // Syntax problem
1635  break;
1636  }
1637  // Ensure a valid range of CPUs
1638  if (begin_cpu < 0 || begin_cpu >= __kmp_xproc || end_cpu < 0 ||
1639  end_cpu >= __kmp_xproc || begin_cpu > end_cpu) {
1640  continue;
1641  }
1642  // Insert [begin_cpu, end_cpu] into mask
1643  for (int cpu = begin_cpu; cpu <= end_cpu; ++cpu) {
1644  KMP_CPU_SET(cpu, mask);
1645  }
1646  }
1647 #endif
1648  return mask;
1649 }
1650 
1651 // Return (possibly empty) affinity mask representing the offline CPUs
1652 // Caller must free the mask
1653 kmp_affin_mask_t *__kmp_affinity_get_offline_cpus() {
1654  return __kmp_parse_cpu_list("/sys/devices/system/cpu/offline");
1655 }
1656 
1657 // Return the number of available procs
1658 int __kmp_affinity_entire_machine_mask(kmp_affin_mask_t *mask) {
1659  int avail_proc = 0;
1660  KMP_CPU_ZERO(mask);
1661 
1662 #if KMP_GROUP_AFFINITY
1663 
1664  if (__kmp_num_proc_groups > 1) {
1665  int group;
1666  KMP_DEBUG_ASSERT(__kmp_GetActiveProcessorCount != NULL);
1667  for (group = 0; group < __kmp_num_proc_groups; group++) {
1668  int i;
1669  int num = __kmp_GetActiveProcessorCount(group);
1670  for (i = 0; i < num; i++) {
1671  KMP_CPU_SET(i + group * (CHAR_BIT * sizeof(DWORD_PTR)), mask);
1672  avail_proc++;
1673  }
1674  }
1675  } else
1676 
1677 #endif /* KMP_GROUP_AFFINITY */
1678 
1679  {
1680  int proc;
1681  kmp_affin_mask_t *offline_cpus = __kmp_affinity_get_offline_cpus();
1682  for (proc = 0; proc < __kmp_xproc; proc++) {
1683  // Skip offline CPUs
1684  if (KMP_CPU_ISSET(proc, offline_cpus))
1685  continue;
1686  KMP_CPU_SET(proc, mask);
1687  avail_proc++;
1688  }
1689  KMP_CPU_FREE(offline_cpus);
1690  }
1691 
1692  return avail_proc;
1693 }
1694 
1695 // All of the __kmp_affinity_create_*_map() routines should allocate the
1696 // internal topology object and set the layer ids for it. Each routine
1697 // returns a boolean on whether it was successful at doing so.
1698 kmp_affin_mask_t *__kmp_affin_fullMask = NULL;
1699 // Original mask is a subset of full mask in multiple processor groups topology
1700 kmp_affin_mask_t *__kmp_affin_origMask = NULL;
1701 
1702 #if KMP_USE_HWLOC
1703 static inline bool __kmp_hwloc_is_cache_type(hwloc_obj_t obj) {
1704 #if HWLOC_API_VERSION >= 0x00020000
1705  return hwloc_obj_type_is_cache(obj->type);
1706 #else
1707  return obj->type == HWLOC_OBJ_CACHE;
1708 #endif
1709 }
1710 
1711 // Returns KMP_HW_* type derived from HWLOC_* type
1712 static inline kmp_hw_t __kmp_hwloc_type_2_topology_type(hwloc_obj_t obj) {
1713 
1714  if (__kmp_hwloc_is_cache_type(obj)) {
1715  if (obj->attr->cache.type == HWLOC_OBJ_CACHE_INSTRUCTION)
1716  return KMP_HW_UNKNOWN;
1717  switch (obj->attr->cache.depth) {
1718  case 1:
1719  return KMP_HW_L1;
1720  case 2:
1721 #if KMP_MIC_SUPPORTED
1722  if (__kmp_mic_type == mic3) {
1723  return KMP_HW_TILE;
1724  }
1725 #endif
1726  return KMP_HW_L2;
1727  case 3:
1728  return KMP_HW_L3;
1729  }
1730  return KMP_HW_UNKNOWN;
1731  }
1732 
1733  switch (obj->type) {
1734  case HWLOC_OBJ_PACKAGE:
1735  return KMP_HW_SOCKET;
1736  case HWLOC_OBJ_NUMANODE:
1737  return KMP_HW_NUMA;
1738  case HWLOC_OBJ_CORE:
1739  return KMP_HW_CORE;
1740  case HWLOC_OBJ_PU:
1741  return KMP_HW_THREAD;
1742  case HWLOC_OBJ_GROUP:
1743 #if HWLOC_API_VERSION >= 0x00020000
1744  if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_DIE)
1745  return KMP_HW_DIE;
1746  else if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_TILE)
1747  return KMP_HW_TILE;
1748  else if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_MODULE)
1749  return KMP_HW_MODULE;
1750  else if (obj->attr->group.kind == HWLOC_GROUP_KIND_WINDOWS_PROCESSOR_GROUP)
1751  return KMP_HW_PROC_GROUP;
1752 #endif
1753  return KMP_HW_UNKNOWN;
1754 #if HWLOC_API_VERSION >= 0x00020100
1755  case HWLOC_OBJ_DIE:
1756  return KMP_HW_DIE;
1757 #endif
1758  }
1759  return KMP_HW_UNKNOWN;
1760 }
1761 
1762 // Returns the number of objects of type 'type' below 'obj' within the topology
1763 // tree structure. e.g., if obj is a HWLOC_OBJ_PACKAGE object, and type is
1764 // HWLOC_OBJ_PU, then this will return the number of PU's under the SOCKET
1765 // object.
1766 static int __kmp_hwloc_get_nobjs_under_obj(hwloc_obj_t obj,
1767  hwloc_obj_type_t type) {
1768  int retval = 0;
1769  hwloc_obj_t first;
1770  for (first = hwloc_get_obj_below_by_type(__kmp_hwloc_topology, obj->type,
1771  obj->logical_index, type, 0);
1772  first != NULL && hwloc_get_ancestor_obj_by_type(__kmp_hwloc_topology,
1773  obj->type, first) == obj;
1774  first = hwloc_get_next_obj_by_type(__kmp_hwloc_topology, first->type,
1775  first)) {
1776  ++retval;
1777  }
1778  return retval;
1779 }
1780 
1781 // This gets the sub_id for a lower object under a higher object in the
1782 // topology tree
1783 static int __kmp_hwloc_get_sub_id(hwloc_topology_t t, hwloc_obj_t higher,
1784  hwloc_obj_t lower) {
1785  hwloc_obj_t obj;
1786  hwloc_obj_type_t ltype = lower->type;
1787  int lindex = lower->logical_index - 1;
1788  int sub_id = 0;
1789  // Get the previous lower object
1790  obj = hwloc_get_obj_by_type(t, ltype, lindex);
1791  while (obj && lindex >= 0 &&
1792  hwloc_bitmap_isincluded(obj->cpuset, higher->cpuset)) {
1793  if (obj->userdata) {
1794  sub_id = (int)(RCAST(kmp_intptr_t, obj->userdata));
1795  break;
1796  }
1797  sub_id++;
1798  lindex--;
1799  obj = hwloc_get_obj_by_type(t, ltype, lindex);
1800  }
1801  // store sub_id + 1 so that 0 is differed from NULL
1802  lower->userdata = RCAST(void *, sub_id + 1);
1803  return sub_id;
1804 }
1805 
1806 static bool __kmp_affinity_create_hwloc_map(kmp_i18n_id_t *const msg_id) {
1807  kmp_hw_t type;
1808  int hw_thread_index, sub_id;
1809  int depth;
1810  hwloc_obj_t pu, obj, root, prev;
1811  kmp_hw_t types[KMP_HW_LAST];
1812  hwloc_obj_type_t hwloc_types[KMP_HW_LAST];
1813 
1814  hwloc_topology_t tp = __kmp_hwloc_topology;
1815  *msg_id = kmp_i18n_null;
1816  if (__kmp_affinity.flags.verbose) {
1817  KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
1818  }
1819 
1820  if (!KMP_AFFINITY_CAPABLE()) {
1821  // Hack to try and infer the machine topology using only the data
1822  // available from hwloc on the current thread, and __kmp_xproc.
1823  KMP_ASSERT(__kmp_affinity.type == affinity_none);
1824  // hwloc only guarantees existance of PU object, so check PACKAGE and CORE
1825  hwloc_obj_t o = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PACKAGE, 0);
1826  if (o != NULL)
1827  nCoresPerPkg = __kmp_hwloc_get_nobjs_under_obj(o, HWLOC_OBJ_CORE);
1828  else
1829  nCoresPerPkg = 1; // no PACKAGE found
1830  o = hwloc_get_obj_by_type(tp, HWLOC_OBJ_CORE, 0);
1831  if (o != NULL)
1832  __kmp_nThreadsPerCore = __kmp_hwloc_get_nobjs_under_obj(o, HWLOC_OBJ_PU);
1833  else
1834  __kmp_nThreadsPerCore = 1; // no CORE found
1835  if (__kmp_nThreadsPerCore == 0)
1836  __kmp_nThreadsPerCore = 1;
1837  __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
1838  if (nCoresPerPkg == 0)
1839  nCoresPerPkg = 1; // to prevent possible division by 0
1840  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
1841  return true;
1842  }
1843 
1844 #if HWLOC_API_VERSION >= 0x00020400
1845  // Handle multiple types of cores if they exist on the system
1846  int nr_cpu_kinds = hwloc_cpukinds_get_nr(tp, 0);
1847 
1848  typedef struct kmp_hwloc_cpukinds_info_t {
1849  int efficiency;
1850  kmp_hw_core_type_t core_type;
1851  hwloc_bitmap_t mask;
1852  } kmp_hwloc_cpukinds_info_t;
1853  kmp_hwloc_cpukinds_info_t *cpukinds = nullptr;
1854 
1855  if (nr_cpu_kinds > 0) {
1856  unsigned nr_infos;
1857  struct hwloc_info_s *infos;
1858  cpukinds = (kmp_hwloc_cpukinds_info_t *)__kmp_allocate(
1859  sizeof(kmp_hwloc_cpukinds_info_t) * nr_cpu_kinds);
1860  for (unsigned idx = 0; idx < (unsigned)nr_cpu_kinds; ++idx) {
1861  cpukinds[idx].efficiency = -1;
1862  cpukinds[idx].core_type = KMP_HW_CORE_TYPE_UNKNOWN;
1863  cpukinds[idx].mask = hwloc_bitmap_alloc();
1864  if (hwloc_cpukinds_get_info(tp, idx, cpukinds[idx].mask,
1865  &cpukinds[idx].efficiency, &nr_infos, &infos,
1866  0) == 0) {
1867  for (unsigned i = 0; i < nr_infos; ++i) {
1868  if (__kmp_str_match("CoreType", 8, infos[i].name)) {
1869 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1870  if (__kmp_str_match("IntelAtom", 9, infos[i].value)) {
1871  cpukinds[idx].core_type = KMP_HW_CORE_TYPE_ATOM;
1872  break;
1873  } else if (__kmp_str_match("IntelCore", 9, infos[i].value)) {
1874  cpukinds[idx].core_type = KMP_HW_CORE_TYPE_CORE;
1875  break;
1876  }
1877 #endif
1878  }
1879  }
1880  }
1881  }
1882  }
1883 #endif
1884 
1885  root = hwloc_get_root_obj(tp);
1886 
1887  // Figure out the depth and types in the topology
1888  depth = 0;
1889  obj = hwloc_get_pu_obj_by_os_index(tp, __kmp_affin_fullMask->begin());
1890  while (obj && obj != root) {
1891 #if HWLOC_API_VERSION >= 0x00020000
1892  if (obj->memory_arity) {
1893  hwloc_obj_t memory;
1894  for (memory = obj->memory_first_child; memory;
1895  memory = hwloc_get_next_child(tp, obj, memory)) {
1896  if (memory->type == HWLOC_OBJ_NUMANODE)
1897  break;
1898  }
1899  if (memory && memory->type == HWLOC_OBJ_NUMANODE) {
1900  types[depth] = KMP_HW_NUMA;
1901  hwloc_types[depth] = memory->type;
1902  depth++;
1903  }
1904  }
1905 #endif
1906  type = __kmp_hwloc_type_2_topology_type(obj);
1907  if (type != KMP_HW_UNKNOWN) {
1908  types[depth] = type;
1909  hwloc_types[depth] = obj->type;
1910  depth++;
1911  }
1912  obj = obj->parent;
1913  }
1914  KMP_ASSERT(depth > 0);
1915 
1916  // Get the order for the types correct
1917  for (int i = 0, j = depth - 1; i < j; ++i, --j) {
1918  hwloc_obj_type_t hwloc_temp = hwloc_types[i];
1919  kmp_hw_t temp = types[i];
1920  types[i] = types[j];
1921  types[j] = temp;
1922  hwloc_types[i] = hwloc_types[j];
1923  hwloc_types[j] = hwloc_temp;
1924  }
1925 
1926  // Allocate the data structure to be returned.
1927  __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
1928 
1929  hw_thread_index = 0;
1930  pu = NULL;
1931  while ((pu = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, pu))) {
1932  int index = depth - 1;
1933  bool included = KMP_CPU_ISSET(pu->os_index, __kmp_affin_fullMask);
1934  kmp_hw_thread_t &hw_thread = __kmp_topology->at(hw_thread_index);
1935  if (included) {
1936  hw_thread.clear();
1937  hw_thread.ids[index] = pu->logical_index;
1938  hw_thread.os_id = pu->os_index;
1939  hw_thread.original_idx = hw_thread_index;
1940  // If multiple core types, then set that attribute for the hardware thread
1941 #if HWLOC_API_VERSION >= 0x00020400
1942  if (cpukinds) {
1943  int cpukind_index = -1;
1944  for (int i = 0; i < nr_cpu_kinds; ++i) {
1945  if (hwloc_bitmap_isset(cpukinds[i].mask, hw_thread.os_id)) {
1946  cpukind_index = i;
1947  break;
1948  }
1949  }
1950  if (cpukind_index >= 0) {
1951  hw_thread.attrs.set_core_type(cpukinds[cpukind_index].core_type);
1952  hw_thread.attrs.set_core_eff(cpukinds[cpukind_index].efficiency);
1953  }
1954  }
1955 #endif
1956  index--;
1957  }
1958  obj = pu;
1959  prev = obj;
1960  while (obj != root && obj != NULL) {
1961  obj = obj->parent;
1962 #if HWLOC_API_VERSION >= 0x00020000
1963  // NUMA Nodes are handled differently since they are not within the
1964  // parent/child structure anymore. They are separate children
1965  // of obj (memory_first_child points to first memory child)
1966  if (obj->memory_arity) {
1967  hwloc_obj_t memory;
1968  for (memory = obj->memory_first_child; memory;
1969  memory = hwloc_get_next_child(tp, obj, memory)) {
1970  if (memory->type == HWLOC_OBJ_NUMANODE)
1971  break;
1972  }
1973  if (memory && memory->type == HWLOC_OBJ_NUMANODE) {
1974  sub_id = __kmp_hwloc_get_sub_id(tp, memory, prev);
1975  if (included) {
1976  hw_thread.ids[index] = memory->logical_index;
1977  hw_thread.ids[index + 1] = sub_id;
1978  index--;
1979  }
1980  }
1981  prev = obj;
1982  }
1983 #endif
1984  type = __kmp_hwloc_type_2_topology_type(obj);
1985  if (type != KMP_HW_UNKNOWN) {
1986  sub_id = __kmp_hwloc_get_sub_id(tp, obj, prev);
1987  if (included) {
1988  hw_thread.ids[index] = obj->logical_index;
1989  hw_thread.ids[index + 1] = sub_id;
1990  index--;
1991  }
1992  prev = obj;
1993  }
1994  }
1995  if (included)
1996  hw_thread_index++;
1997  }
1998 
1999 #if HWLOC_API_VERSION >= 0x00020400
2000  // Free the core types information
2001  if (cpukinds) {
2002  for (int idx = 0; idx < nr_cpu_kinds; ++idx)
2003  hwloc_bitmap_free(cpukinds[idx].mask);
2004  __kmp_free(cpukinds);
2005  }
2006 #endif
2007  __kmp_topology->sort_ids();
2008  return true;
2009 }
2010 #endif // KMP_USE_HWLOC
2011 
2012 // If we don't know how to retrieve the machine's processor topology, or
2013 // encounter an error in doing so, this routine is called to form a "flat"
2014 // mapping of os thread id's <-> processor id's.
2015 static bool __kmp_affinity_create_flat_map(kmp_i18n_id_t *const msg_id) {
2016  *msg_id = kmp_i18n_null;
2017  int depth = 3;
2018  kmp_hw_t types[] = {KMP_HW_SOCKET, KMP_HW_CORE, KMP_HW_THREAD};
2019 
2020  if (__kmp_affinity.flags.verbose) {
2021  KMP_INFORM(UsingFlatOS, "KMP_AFFINITY");
2022  }
2023 
2024  // Even if __kmp_affinity.type == affinity_none, this routine might still
2025  // be called to set __kmp_ncores, as well as
2026  // __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
2027  if (!KMP_AFFINITY_CAPABLE()) {
2028  KMP_ASSERT(__kmp_affinity.type == affinity_none);
2029  __kmp_ncores = nPackages = __kmp_xproc;
2030  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
2031  return true;
2032  }
2033 
2034  // When affinity is off, this routine will still be called to set
2035  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
2036  // Make sure all these vars are set correctly, and return now if affinity is
2037  // not enabled.
2038  __kmp_ncores = nPackages = __kmp_avail_proc;
2039  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
2040 
2041  // Construct the data structure to be returned.
2042  __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
2043  int avail_ct = 0;
2044  int i;
2045  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
2046  // Skip this proc if it is not included in the machine model.
2047  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
2048  continue;
2049  }
2050  kmp_hw_thread_t &hw_thread = __kmp_topology->at(avail_ct);
2051  hw_thread.clear();
2052  hw_thread.os_id = i;
2053  hw_thread.original_idx = avail_ct;
2054  hw_thread.ids[0] = i;
2055  hw_thread.ids[1] = 0;
2056  hw_thread.ids[2] = 0;
2057  avail_ct++;
2058  }
2059  if (__kmp_affinity.flags.verbose) {
2060  KMP_INFORM(OSProcToPackage, "KMP_AFFINITY");
2061  }
2062  return true;
2063 }
2064 
2065 #if KMP_GROUP_AFFINITY
2066 // If multiple Windows* OS processor groups exist, we can create a 2-level
2067 // topology map with the groups at level 0 and the individual procs at level 1.
2068 // This facilitates letting the threads float among all procs in a group,
2069 // if granularity=group (the default when there are multiple groups).
2070 static bool __kmp_affinity_create_proc_group_map(kmp_i18n_id_t *const msg_id) {
2071  *msg_id = kmp_i18n_null;
2072  int depth = 3;
2073  kmp_hw_t types[] = {KMP_HW_PROC_GROUP, KMP_HW_CORE, KMP_HW_THREAD};
2074  const static size_t BITS_PER_GROUP = CHAR_BIT * sizeof(DWORD_PTR);
2075 
2076  if (__kmp_affinity.flags.verbose) {
2077  KMP_INFORM(AffWindowsProcGroupMap, "KMP_AFFINITY");
2078  }
2079 
2080  // If we aren't affinity capable, then use flat topology
2081  if (!KMP_AFFINITY_CAPABLE()) {
2082  KMP_ASSERT(__kmp_affinity.type == affinity_none);
2083  nPackages = __kmp_num_proc_groups;
2084  __kmp_nThreadsPerCore = 1;
2085  __kmp_ncores = __kmp_xproc;
2086  nCoresPerPkg = nPackages / __kmp_ncores;
2087  return true;
2088  }
2089 
2090  // Construct the data structure to be returned.
2091  __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
2092  int avail_ct = 0;
2093  int i;
2094  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
2095  // Skip this proc if it is not included in the machine model.
2096  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
2097  continue;
2098  }
2099  kmp_hw_thread_t &hw_thread = __kmp_topology->at(avail_ct);
2100  hw_thread.clear();
2101  hw_thread.os_id = i;
2102  hw_thread.original_idx = avail_ct;
2103  hw_thread.ids[0] = i / BITS_PER_GROUP;
2104  hw_thread.ids[1] = hw_thread.ids[2] = i % BITS_PER_GROUP;
2105  avail_ct++;
2106  }
2107  return true;
2108 }
2109 #endif /* KMP_GROUP_AFFINITY */
2110 
2111 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2112 
2113 template <kmp_uint32 LSB, kmp_uint32 MSB>
2114 static inline unsigned __kmp_extract_bits(kmp_uint32 v) {
2115  const kmp_uint32 SHIFT_LEFT = sizeof(kmp_uint32) * 8 - 1 - MSB;
2116  const kmp_uint32 SHIFT_RIGHT = LSB;
2117  kmp_uint32 retval = v;
2118  retval <<= SHIFT_LEFT;
2119  retval >>= (SHIFT_LEFT + SHIFT_RIGHT);
2120  return retval;
2121 }
2122 
2123 static int __kmp_cpuid_mask_width(int count) {
2124  int r = 0;
2125 
2126  while ((1 << r) < count)
2127  ++r;
2128  return r;
2129 }
2130 
2131 class apicThreadInfo {
2132 public:
2133  unsigned osId; // param to __kmp_affinity_bind_thread
2134  unsigned apicId; // from cpuid after binding
2135  unsigned maxCoresPerPkg; // ""
2136  unsigned maxThreadsPerPkg; // ""
2137  unsigned pkgId; // inferred from above values
2138  unsigned coreId; // ""
2139  unsigned threadId; // ""
2140 };
2141 
2142 static int __kmp_affinity_cmp_apicThreadInfo_phys_id(const void *a,
2143  const void *b) {
2144  const apicThreadInfo *aa = (const apicThreadInfo *)a;
2145  const apicThreadInfo *bb = (const apicThreadInfo *)b;
2146  if (aa->pkgId < bb->pkgId)
2147  return -1;
2148  if (aa->pkgId > bb->pkgId)
2149  return 1;
2150  if (aa->coreId < bb->coreId)
2151  return -1;
2152  if (aa->coreId > bb->coreId)
2153  return 1;
2154  if (aa->threadId < bb->threadId)
2155  return -1;
2156  if (aa->threadId > bb->threadId)
2157  return 1;
2158  return 0;
2159 }
2160 
2161 class cpuid_cache_info_t {
2162 public:
2163  struct info_t {
2164  unsigned level = 0;
2165  unsigned mask = 0;
2166  bool operator==(const info_t &rhs) const {
2167  return level == rhs.level && mask == rhs.mask;
2168  }
2169  bool operator!=(const info_t &rhs) const { return !operator==(rhs); }
2170  };
2171  cpuid_cache_info_t() : depth(0) {
2172  table[MAX_CACHE_LEVEL].level = 0;
2173  table[MAX_CACHE_LEVEL].mask = 0;
2174  }
2175  size_t get_depth() const { return depth; }
2176  info_t &operator[](size_t index) { return table[index]; }
2177  const info_t &operator[](size_t index) const { return table[index]; }
2178  bool operator==(const cpuid_cache_info_t &rhs) const {
2179  if (rhs.depth != depth)
2180  return false;
2181  for (size_t i = 0; i < depth; ++i)
2182  if (table[i] != rhs.table[i])
2183  return false;
2184  return true;
2185  }
2186  bool operator!=(const cpuid_cache_info_t &rhs) const {
2187  return !operator==(rhs);
2188  }
2189  // Get cache information assocaited with L1, L2, L3 cache, etc.
2190  // If level does not exist, then return the "NULL" level (level 0)
2191  const info_t &get_level(unsigned level) const {
2192  for (size_t i = 0; i < depth; ++i) {
2193  if (table[i].level == level)
2194  return table[i];
2195  }
2196  return table[MAX_CACHE_LEVEL];
2197  }
2198 
2199  static kmp_hw_t get_topology_type(unsigned level) {
2200  KMP_DEBUG_ASSERT(level >= 1 && level <= MAX_CACHE_LEVEL);
2201  switch (level) {
2202  case 1:
2203  return KMP_HW_L1;
2204  case 2:
2205  return KMP_HW_L2;
2206  case 3:
2207  return KMP_HW_L3;
2208  }
2209  return KMP_HW_UNKNOWN;
2210  }
2211  void get_leaf4_levels() {
2212  unsigned level = 0;
2213  while (depth < MAX_CACHE_LEVEL) {
2214  unsigned cache_type, max_threads_sharing;
2215  unsigned cache_level, cache_mask_width;
2216  kmp_cpuid buf2;
2217  __kmp_x86_cpuid(4, level, &buf2);
2218  cache_type = __kmp_extract_bits<0, 4>(buf2.eax);
2219  if (!cache_type)
2220  break;
2221  // Skip instruction caches
2222  if (cache_type == 2) {
2223  level++;
2224  continue;
2225  }
2226  max_threads_sharing = __kmp_extract_bits<14, 25>(buf2.eax) + 1;
2227  cache_mask_width = __kmp_cpuid_mask_width(max_threads_sharing);
2228  cache_level = __kmp_extract_bits<5, 7>(buf2.eax);
2229  table[depth].level = cache_level;
2230  table[depth].mask = ((0xffffffffu) << cache_mask_width);
2231  depth++;
2232  level++;
2233  }
2234  }
2235  static const int MAX_CACHE_LEVEL = 3;
2236 
2237 private:
2238  size_t depth;
2239  info_t table[MAX_CACHE_LEVEL + 1];
2240 };
2241 
2242 // On IA-32 architecture and Intel(R) 64 architecture, we attempt to use
2243 // an algorithm which cycles through the available os threads, setting
2244 // the current thread's affinity mask to that thread, and then retrieves
2245 // the Apic Id for each thread context using the cpuid instruction.
2246 static bool __kmp_affinity_create_apicid_map(kmp_i18n_id_t *const msg_id) {
2247  kmp_cpuid buf;
2248  *msg_id = kmp_i18n_null;
2249 
2250  if (__kmp_affinity.flags.verbose) {
2251  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(DecodingLegacyAPIC));
2252  }
2253 
2254  // Check if cpuid leaf 4 is supported.
2255  __kmp_x86_cpuid(0, 0, &buf);
2256  if (buf.eax < 4) {
2257  *msg_id = kmp_i18n_str_NoLeaf4Support;
2258  return false;
2259  }
2260 
2261  // The algorithm used starts by setting the affinity to each available thread
2262  // and retrieving info from the cpuid instruction, so if we are not capable of
2263  // calling __kmp_get_system_affinity() and _kmp_get_system_affinity(), then we
2264  // need to do something else - use the defaults that we calculated from
2265  // issuing cpuid without binding to each proc.
2266  if (!KMP_AFFINITY_CAPABLE()) {
2267  // Hack to try and infer the machine topology using only the data
2268  // available from cpuid on the current thread, and __kmp_xproc.
2269  KMP_ASSERT(__kmp_affinity.type == affinity_none);
2270 
2271  // Get an upper bound on the number of threads per package using cpuid(1).
2272  // On some OS/chps combinations where HT is supported by the chip but is
2273  // disabled, this value will be 2 on a single core chip. Usually, it will be
2274  // 2 if HT is enabled and 1 if HT is disabled.
2275  __kmp_x86_cpuid(1, 0, &buf);
2276  int maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
2277  if (maxThreadsPerPkg == 0) {
2278  maxThreadsPerPkg = 1;
2279  }
2280 
2281  // The num cores per pkg comes from cpuid(4). 1 must be added to the encoded
2282  // value.
2283  //
2284  // The author of cpu_count.cpp treated this only an upper bound on the
2285  // number of cores, but I haven't seen any cases where it was greater than
2286  // the actual number of cores, so we will treat it as exact in this block of
2287  // code.
2288  //
2289  // First, we need to check if cpuid(4) is supported on this chip. To see if
2290  // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n or
2291  // greater.
2292  __kmp_x86_cpuid(0, 0, &buf);
2293  if (buf.eax >= 4) {
2294  __kmp_x86_cpuid(4, 0, &buf);
2295  nCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
2296  } else {
2297  nCoresPerPkg = 1;
2298  }
2299 
2300  // There is no way to reliably tell if HT is enabled without issuing the
2301  // cpuid instruction from every thread, can correlating the cpuid info, so
2302  // if the machine is not affinity capable, we assume that HT is off. We have
2303  // seen quite a few machines where maxThreadsPerPkg is 2, yet the machine
2304  // does not support HT.
2305  //
2306  // - Older OSes are usually found on machines with older chips, which do not
2307  // support HT.
2308  // - The performance penalty for mistakenly identifying a machine as HT when
2309  // it isn't (which results in blocktime being incorrectly set to 0) is
2310  // greater than the penalty when for mistakenly identifying a machine as
2311  // being 1 thread/core when it is really HT enabled (which results in
2312  // blocktime being incorrectly set to a positive value).
2313  __kmp_ncores = __kmp_xproc;
2314  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
2315  __kmp_nThreadsPerCore = 1;
2316  return true;
2317  }
2318 
2319  // From here on, we can assume that it is safe to call
2320  // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
2321  // __kmp_affinity.type = affinity_none.
2322 
2323  // Save the affinity mask for the current thread.
2324  kmp_affinity_raii_t previous_affinity;
2325 
2326  // Run through each of the available contexts, binding the current thread
2327  // to it, and obtaining the pertinent information using the cpuid instr.
2328  //
2329  // The relevant information is:
2330  // - Apic Id: Bits 24:31 of ebx after issuing cpuid(1) - each thread context
2331  // has a uniqie Apic Id, which is of the form pkg# : core# : thread#.
2332  // - Max Threads Per Pkg: Bits 16:23 of ebx after issuing cpuid(1). The value
2333  // of this field determines the width of the core# + thread# fields in the
2334  // Apic Id. It is also an upper bound on the number of threads per
2335  // package, but it has been verified that situations happen were it is not
2336  // exact. In particular, on certain OS/chip combinations where Intel(R)
2337  // Hyper-Threading Technology is supported by the chip but has been
2338  // disabled, the value of this field will be 2 (for a single core chip).
2339  // On other OS/chip combinations supporting Intel(R) Hyper-Threading
2340  // Technology, the value of this field will be 1 when Intel(R)
2341  // Hyper-Threading Technology is disabled and 2 when it is enabled.
2342  // - Max Cores Per Pkg: Bits 26:31 of eax after issuing cpuid(4). The value
2343  // of this field (+1) determines the width of the core# field in the Apic
2344  // Id. The comments in "cpucount.cpp" say that this value is an upper
2345  // bound, but the IA-32 architecture manual says that it is exactly the
2346  // number of cores per package, and I haven't seen any case where it
2347  // wasn't.
2348  //
2349  // From this information, deduce the package Id, core Id, and thread Id,
2350  // and set the corresponding fields in the apicThreadInfo struct.
2351  unsigned i;
2352  apicThreadInfo *threadInfo = (apicThreadInfo *)__kmp_allocate(
2353  __kmp_avail_proc * sizeof(apicThreadInfo));
2354  unsigned nApics = 0;
2355  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
2356  // Skip this proc if it is not included in the machine model.
2357  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
2358  continue;
2359  }
2360  KMP_DEBUG_ASSERT((int)nApics < __kmp_avail_proc);
2361 
2362  __kmp_affinity_dispatch->bind_thread(i);
2363  threadInfo[nApics].osId = i;
2364 
2365  // The apic id and max threads per pkg come from cpuid(1).
2366  __kmp_x86_cpuid(1, 0, &buf);
2367  if (((buf.edx >> 9) & 1) == 0) {
2368  __kmp_free(threadInfo);
2369  *msg_id = kmp_i18n_str_ApicNotPresent;
2370  return false;
2371  }
2372  threadInfo[nApics].apicId = (buf.ebx >> 24) & 0xff;
2373  threadInfo[nApics].maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
2374  if (threadInfo[nApics].maxThreadsPerPkg == 0) {
2375  threadInfo[nApics].maxThreadsPerPkg = 1;
2376  }
2377 
2378  // Max cores per pkg comes from cpuid(4). 1 must be added to the encoded
2379  // value.
2380  //
2381  // First, we need to check if cpuid(4) is supported on this chip. To see if
2382  // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n
2383  // or greater.
2384  __kmp_x86_cpuid(0, 0, &buf);
2385  if (buf.eax >= 4) {
2386  __kmp_x86_cpuid(4, 0, &buf);
2387  threadInfo[nApics].maxCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
2388  } else {
2389  threadInfo[nApics].maxCoresPerPkg = 1;
2390  }
2391 
2392  // Infer the pkgId / coreId / threadId using only the info obtained locally.
2393  int widthCT = __kmp_cpuid_mask_width(threadInfo[nApics].maxThreadsPerPkg);
2394  threadInfo[nApics].pkgId = threadInfo[nApics].apicId >> widthCT;
2395 
2396  int widthC = __kmp_cpuid_mask_width(threadInfo[nApics].maxCoresPerPkg);
2397  int widthT = widthCT - widthC;
2398  if (widthT < 0) {
2399  // I've never seen this one happen, but I suppose it could, if the cpuid
2400  // instruction on a chip was really screwed up. Make sure to restore the
2401  // affinity mask before the tail call.
2402  __kmp_free(threadInfo);
2403  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
2404  return false;
2405  }
2406 
2407  int maskC = (1 << widthC) - 1;
2408  threadInfo[nApics].coreId = (threadInfo[nApics].apicId >> widthT) & maskC;
2409 
2410  int maskT = (1 << widthT) - 1;
2411  threadInfo[nApics].threadId = threadInfo[nApics].apicId & maskT;
2412 
2413  nApics++;
2414  }
2415 
2416  // We've collected all the info we need.
2417  // Restore the old affinity mask for this thread.
2418  previous_affinity.restore();
2419 
2420  // Sort the threadInfo table by physical Id.
2421  qsort(threadInfo, nApics, sizeof(*threadInfo),
2422  __kmp_affinity_cmp_apicThreadInfo_phys_id);
2423 
2424  // The table is now sorted by pkgId / coreId / threadId, but we really don't
2425  // know the radix of any of the fields. pkgId's may be sparsely assigned among
2426  // the chips on a system. Although coreId's are usually assigned
2427  // [0 .. coresPerPkg-1] and threadId's are usually assigned
2428  // [0..threadsPerCore-1], we don't want to make any such assumptions.
2429  //
2430  // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
2431  // total # packages) are at this point - we want to determine that now. We
2432  // only have an upper bound on the first two figures.
2433  //
2434  // We also perform a consistency check at this point: the values returned by
2435  // the cpuid instruction for any thread bound to a given package had better
2436  // return the same info for maxThreadsPerPkg and maxCoresPerPkg.
2437  nPackages = 1;
2438  nCoresPerPkg = 1;
2439  __kmp_nThreadsPerCore = 1;
2440  unsigned nCores = 1;
2441 
2442  unsigned pkgCt = 1; // to determine radii
2443  unsigned lastPkgId = threadInfo[0].pkgId;
2444  unsigned coreCt = 1;
2445  unsigned lastCoreId = threadInfo[0].coreId;
2446  unsigned threadCt = 1;
2447  unsigned lastThreadId = threadInfo[0].threadId;
2448 
2449  // intra-pkg consist checks
2450  unsigned prevMaxCoresPerPkg = threadInfo[0].maxCoresPerPkg;
2451  unsigned prevMaxThreadsPerPkg = threadInfo[0].maxThreadsPerPkg;
2452 
2453  for (i = 1; i < nApics; i++) {
2454  if (threadInfo[i].pkgId != lastPkgId) {
2455  nCores++;
2456  pkgCt++;
2457  lastPkgId = threadInfo[i].pkgId;
2458  if ((int)coreCt > nCoresPerPkg)
2459  nCoresPerPkg = coreCt;
2460  coreCt = 1;
2461  lastCoreId = threadInfo[i].coreId;
2462  if ((int)threadCt > __kmp_nThreadsPerCore)
2463  __kmp_nThreadsPerCore = threadCt;
2464  threadCt = 1;
2465  lastThreadId = threadInfo[i].threadId;
2466 
2467  // This is a different package, so go on to the next iteration without
2468  // doing any consistency checks. Reset the consistency check vars, though.
2469  prevMaxCoresPerPkg = threadInfo[i].maxCoresPerPkg;
2470  prevMaxThreadsPerPkg = threadInfo[i].maxThreadsPerPkg;
2471  continue;
2472  }
2473 
2474  if (threadInfo[i].coreId != lastCoreId) {
2475  nCores++;
2476  coreCt++;
2477  lastCoreId = threadInfo[i].coreId;
2478  if ((int)threadCt > __kmp_nThreadsPerCore)
2479  __kmp_nThreadsPerCore = threadCt;
2480  threadCt = 1;
2481  lastThreadId = threadInfo[i].threadId;
2482  } else if (threadInfo[i].threadId != lastThreadId) {
2483  threadCt++;
2484  lastThreadId = threadInfo[i].threadId;
2485  } else {
2486  __kmp_free(threadInfo);
2487  *msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
2488  return false;
2489  }
2490 
2491  // Check to make certain that the maxCoresPerPkg and maxThreadsPerPkg
2492  // fields agree between all the threads bounds to a given package.
2493  if ((prevMaxCoresPerPkg != threadInfo[i].maxCoresPerPkg) ||
2494  (prevMaxThreadsPerPkg != threadInfo[i].maxThreadsPerPkg)) {
2495  __kmp_free(threadInfo);
2496  *msg_id = kmp_i18n_str_InconsistentCpuidInfo;
2497  return false;
2498  }
2499  }
2500  // When affinity is off, this routine will still be called to set
2501  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
2502  // Make sure all these vars are set correctly
2503  nPackages = pkgCt;
2504  if ((int)coreCt > nCoresPerPkg)
2505  nCoresPerPkg = coreCt;
2506  if ((int)threadCt > __kmp_nThreadsPerCore)
2507  __kmp_nThreadsPerCore = threadCt;
2508  __kmp_ncores = nCores;
2509  KMP_DEBUG_ASSERT(nApics == (unsigned)__kmp_avail_proc);
2510 
2511  // Now that we've determined the number of packages, the number of cores per
2512  // package, and the number of threads per core, we can construct the data
2513  // structure that is to be returned.
2514  int idx = 0;
2515  int pkgLevel = 0;
2516  int coreLevel = 1;
2517  int threadLevel = 2;
2518  //(__kmp_nThreadsPerCore <= 1) ? -1 : ((coreLevel >= 0) ? 2 : 1);
2519  int depth = (pkgLevel >= 0) + (coreLevel >= 0) + (threadLevel >= 0);
2520  kmp_hw_t types[3];
2521  if (pkgLevel >= 0)
2522  types[idx++] = KMP_HW_SOCKET;
2523  if (coreLevel >= 0)
2524  types[idx++] = KMP_HW_CORE;
2525  if (threadLevel >= 0)
2526  types[idx++] = KMP_HW_THREAD;
2527 
2528  KMP_ASSERT(depth > 0);
2529  __kmp_topology = kmp_topology_t::allocate(nApics, depth, types);
2530 
2531  for (i = 0; i < nApics; ++i) {
2532  idx = 0;
2533  unsigned os = threadInfo[i].osId;
2534  kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
2535  hw_thread.clear();
2536 
2537  if (pkgLevel >= 0) {
2538  hw_thread.ids[idx++] = threadInfo[i].pkgId;
2539  }
2540  if (coreLevel >= 0) {
2541  hw_thread.ids[idx++] = threadInfo[i].coreId;
2542  }
2543  if (threadLevel >= 0) {
2544  hw_thread.ids[idx++] = threadInfo[i].threadId;
2545  }
2546  hw_thread.os_id = os;
2547  hw_thread.original_idx = i;
2548  }
2549 
2550  __kmp_free(threadInfo);
2551  __kmp_topology->sort_ids();
2552  if (!__kmp_topology->check_ids()) {
2553  kmp_topology_t::deallocate(__kmp_topology);
2554  __kmp_topology = nullptr;
2555  *msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
2556  return false;
2557  }
2558  return true;
2559 }
2560 
2561 // Hybrid cpu detection using CPUID.1A
2562 // Thread should be pinned to processor already
2563 static void __kmp_get_hybrid_info(kmp_hw_core_type_t *type, int *efficiency,
2564  unsigned *native_model_id) {
2565  kmp_cpuid buf;
2566  __kmp_x86_cpuid(0x1a, 0, &buf);
2567  *type = (kmp_hw_core_type_t)__kmp_extract_bits<24, 31>(buf.eax);
2568  switch (*type) {
2569  case KMP_HW_CORE_TYPE_ATOM:
2570  *efficiency = 0;
2571  break;
2572  case KMP_HW_CORE_TYPE_CORE:
2573  *efficiency = 1;
2574  break;
2575  default:
2576  *efficiency = 0;
2577  }
2578  *native_model_id = __kmp_extract_bits<0, 23>(buf.eax);
2579 }
2580 
2581 // Intel(R) microarchitecture code name Nehalem, Dunnington and later
2582 // architectures support a newer interface for specifying the x2APIC Ids,
2583 // based on CPUID.B or CPUID.1F
2584 /*
2585  * CPUID.B or 1F, Input ECX (sub leaf # aka level number)
2586  Bits Bits Bits Bits
2587  31-16 15-8 7-4 4-0
2588 ---+-----------+--------------+-------------+-----------------+
2589 EAX| reserved | reserved | reserved | Bits to Shift |
2590 ---+-----------|--------------+-------------+-----------------|
2591 EBX| reserved | Num logical processors at level (16 bits) |
2592 ---+-----------|--------------+-------------------------------|
2593 ECX| reserved | Level Type | Level Number (8 bits) |
2594 ---+-----------+--------------+-------------------------------|
2595 EDX| X2APIC ID (32 bits) |
2596 ---+----------------------------------------------------------+
2597 */
2598 
2599 enum {
2600  INTEL_LEVEL_TYPE_INVALID = 0, // Package level
2601  INTEL_LEVEL_TYPE_SMT = 1,
2602  INTEL_LEVEL_TYPE_CORE = 2,
2603  INTEL_LEVEL_TYPE_MODULE = 3,
2604  INTEL_LEVEL_TYPE_TILE = 4,
2605  INTEL_LEVEL_TYPE_DIE = 5,
2606  INTEL_LEVEL_TYPE_LAST = 6,
2607 };
2608 KMP_BUILD_ASSERT(INTEL_LEVEL_TYPE_LAST < sizeof(unsigned) * CHAR_BIT);
2609 #define KMP_LEAF_1F_KNOWN_LEVELS ((1u << INTEL_LEVEL_TYPE_LAST) - 1u)
2610 
2611 static kmp_hw_t __kmp_intel_type_2_topology_type(int intel_type) {
2612  switch (intel_type) {
2613  case INTEL_LEVEL_TYPE_INVALID:
2614  return KMP_HW_SOCKET;
2615  case INTEL_LEVEL_TYPE_SMT:
2616  return KMP_HW_THREAD;
2617  case INTEL_LEVEL_TYPE_CORE:
2618  return KMP_HW_CORE;
2619  case INTEL_LEVEL_TYPE_TILE:
2620  return KMP_HW_TILE;
2621  case INTEL_LEVEL_TYPE_MODULE:
2622  return KMP_HW_MODULE;
2623  case INTEL_LEVEL_TYPE_DIE:
2624  return KMP_HW_DIE;
2625  }
2626  return KMP_HW_UNKNOWN;
2627 }
2628 
2629 static int __kmp_topology_type_2_intel_type(kmp_hw_t type) {
2630  switch (type) {
2631  case KMP_HW_SOCKET:
2632  return INTEL_LEVEL_TYPE_INVALID;
2633  case KMP_HW_THREAD:
2634  return INTEL_LEVEL_TYPE_SMT;
2635  case KMP_HW_CORE:
2636  return INTEL_LEVEL_TYPE_CORE;
2637  case KMP_HW_TILE:
2638  return INTEL_LEVEL_TYPE_TILE;
2639  case KMP_HW_MODULE:
2640  return INTEL_LEVEL_TYPE_MODULE;
2641  case KMP_HW_DIE:
2642  return INTEL_LEVEL_TYPE_DIE;
2643  default:
2644  return INTEL_LEVEL_TYPE_INVALID;
2645  }
2646 }
2647 
2648 struct cpuid_level_info_t {
2649  unsigned level_type, mask, mask_width, nitems, cache_mask;
2650 };
2651 
2652 class cpuid_topo_desc_t {
2653  unsigned desc = 0;
2654 
2655 public:
2656  void clear() { desc = 0; }
2657  bool contains(int intel_type) const {
2658  KMP_DEBUG_ASSERT(intel_type >= 0 && intel_type < INTEL_LEVEL_TYPE_LAST);
2659  if ((1u << intel_type) & desc)
2660  return true;
2661  return false;
2662  }
2663  bool contains_topology_type(kmp_hw_t type) const {
2664  KMP_DEBUG_ASSERT(type >= 0 && type < KMP_HW_LAST);
2665  int intel_type = __kmp_topology_type_2_intel_type(type);
2666  return contains(intel_type);
2667  }
2668  bool contains(cpuid_topo_desc_t rhs) const {
2669  return ((desc | rhs.desc) == desc);
2670  }
2671  void add(int intel_type) { desc |= (1u << intel_type); }
2672  void add(cpuid_topo_desc_t rhs) { desc |= rhs.desc; }
2673 };
2674 
2675 struct cpuid_proc_info_t {
2676  // Topology info
2677  int os_id;
2678  unsigned apic_id;
2679  unsigned depth;
2680  // Hybrid info
2681  unsigned native_model_id;
2682  int efficiency;
2683  kmp_hw_core_type_t type;
2684  cpuid_topo_desc_t description;
2685 
2686  cpuid_level_info_t levels[INTEL_LEVEL_TYPE_LAST];
2687 };
2688 
2689 // This function takes the topology leaf, an info pointer to store the levels
2690 // detected, and writable descriptors for the total topology.
2691 // Returns whether total types, depth, or description were modified.
2692 static bool __kmp_x2apicid_get_levels(int leaf, cpuid_proc_info_t *info,
2693  kmp_hw_t total_types[KMP_HW_LAST],
2694  int *total_depth,
2695  cpuid_topo_desc_t *total_description) {
2696  unsigned level, levels_index;
2697  unsigned level_type, mask_width, nitems;
2698  kmp_cpuid buf;
2699  cpuid_level_info_t(&levels)[INTEL_LEVEL_TYPE_LAST] = info->levels;
2700  bool retval = false;
2701 
2702  // New algorithm has known topology layers act as highest unknown topology
2703  // layers when unknown topology layers exist.
2704  // e.g., Suppose layers were SMT <X> CORE <Y> <Z> PACKAGE, where <X> <Y> <Z>
2705  // are unknown topology layers, Then SMT will take the characteristics of
2706  // (SMT x <X>) and CORE will take the characteristics of (CORE x <Y> x <Z>).
2707  // This eliminates unknown portions of the topology while still keeping the
2708  // correct structure.
2709  level = levels_index = 0;
2710  do {
2711  __kmp_x86_cpuid(leaf, level, &buf);
2712  level_type = __kmp_extract_bits<8, 15>(buf.ecx);
2713  mask_width = __kmp_extract_bits<0, 4>(buf.eax);
2714  nitems = __kmp_extract_bits<0, 15>(buf.ebx);
2715  if (level_type != INTEL_LEVEL_TYPE_INVALID && nitems == 0) {
2716  info->depth = 0;
2717  return retval;
2718  }
2719 
2720  if (KMP_LEAF_1F_KNOWN_LEVELS & (1u << level_type)) {
2721  // Add a new level to the topology
2722  KMP_ASSERT(levels_index < INTEL_LEVEL_TYPE_LAST);
2723  levels[levels_index].level_type = level_type;
2724  levels[levels_index].mask_width = mask_width;
2725  levels[levels_index].nitems = nitems;
2726  levels_index++;
2727  } else {
2728  // If it is an unknown level, then logically move the previous layer up
2729  if (levels_index > 0) {
2730  levels[levels_index - 1].mask_width = mask_width;
2731  levels[levels_index - 1].nitems = nitems;
2732  }
2733  }
2734  level++;
2735  } while (level_type != INTEL_LEVEL_TYPE_INVALID);
2736  KMP_ASSERT(levels_index <= INTEL_LEVEL_TYPE_LAST);
2737  info->description.clear();
2738  info->depth = levels_index;
2739 
2740  // If types, depth, and total_description are uninitialized,
2741  // then initialize them now
2742  if (*total_depth == 0) {
2743  *total_depth = info->depth;
2744  total_description->clear();
2745  for (int i = *total_depth - 1, j = 0; i >= 0; --i, ++j) {
2746  total_types[j] =
2747  __kmp_intel_type_2_topology_type(info->levels[i].level_type);
2748  total_description->add(info->levels[i].level_type);
2749  }
2750  retval = true;
2751  }
2752 
2753  // Ensure the INTEL_LEVEL_TYPE_INVALID (Socket) layer isn't first
2754  if (levels_index == 0 || levels[0].level_type == INTEL_LEVEL_TYPE_INVALID)
2755  return 0;
2756 
2757  // Set the masks to & with apicid
2758  for (unsigned i = 0; i < levels_index; ++i) {
2759  if (levels[i].level_type != INTEL_LEVEL_TYPE_INVALID) {
2760  levels[i].mask = ~((0xffffffffu) << levels[i].mask_width);
2761  levels[i].cache_mask = (0xffffffffu) << levels[i].mask_width;
2762  for (unsigned j = 0; j < i; ++j)
2763  levels[i].mask ^= levels[j].mask;
2764  } else {
2765  KMP_DEBUG_ASSERT(i > 0);
2766  levels[i].mask = (0xffffffffu) << levels[i - 1].mask_width;
2767  levels[i].cache_mask = 0;
2768  }
2769  info->description.add(info->levels[i].level_type);
2770  }
2771 
2772  // If this processor has level type not on other processors, then make
2773  // sure to include it in total types, depth, and description.
2774  // One assumption here is that the first type, i.e. socket, is known.
2775  // Another assumption is that types array is always large enough to fit any
2776  // new layers since its length is KMP_HW_LAST.
2777  if (!total_description->contains(info->description)) {
2778  for (int i = info->depth - 1, j = 0; i >= 0; --i, ++j) {
2779  // If this level is known already, then skip it.
2780  if (total_description->contains(levels[i].level_type))
2781  continue;
2782  // Unknown level, insert before last known level
2783  kmp_hw_t curr_type =
2784  __kmp_intel_type_2_topology_type(levels[i].level_type);
2785  KMP_ASSERT(j != 0 && "Bad APIC Id information");
2786  // Move over all known levels to make room for new level
2787  for (int k = info->depth - 1; k >= j; --k) {
2788  KMP_DEBUG_ASSERT(k + 1 < KMP_HW_LAST);
2789  total_types[k + 1] = total_types[k];
2790  }
2791  // Insert new level
2792  total_types[j] = curr_type;
2793  (*total_depth)++;
2794  }
2795  total_description->add(info->description);
2796  retval = true;
2797  }
2798  return retval;
2799 }
2800 
2801 static bool __kmp_affinity_create_x2apicid_map(kmp_i18n_id_t *const msg_id) {
2802 
2803  kmp_hw_t types[INTEL_LEVEL_TYPE_LAST];
2804  kmp_cpuid buf;
2805  int topology_leaf, highest_leaf;
2806  int num_leaves;
2807  int depth = 0;
2808  cpuid_topo_desc_t total_description;
2809  static int leaves[] = {0, 0};
2810 
2811  // If affinity is disabled, __kmp_avail_proc may be zero
2812  int ninfos = (__kmp_avail_proc > 0 ? __kmp_avail_proc : 1);
2813  cpuid_proc_info_t *proc_info = (cpuid_proc_info_t *)__kmp_allocate(
2814  (sizeof(cpuid_proc_info_t) + sizeof(cpuid_cache_info_t)) * ninfos);
2815  cpuid_cache_info_t *cache_info = (cpuid_cache_info_t *)(proc_info + ninfos);
2816 
2817  kmp_i18n_id_t leaf_message_id;
2818 
2819  *msg_id = kmp_i18n_null;
2820  if (__kmp_affinity.flags.verbose) {
2821  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(Decodingx2APIC));
2822  }
2823 
2824  // Get the highest cpuid leaf supported
2825  __kmp_x86_cpuid(0, 0, &buf);
2826  highest_leaf = buf.eax;
2827 
2828  // If a specific topology method was requested, only allow that specific leaf
2829  // otherwise, try both leaves 31 and 11 in that order
2830  num_leaves = 0;
2831  if (__kmp_affinity_top_method == affinity_top_method_x2apicid) {
2832  num_leaves = 1;
2833  leaves[0] = 11;
2834  leaf_message_id = kmp_i18n_str_NoLeaf11Support;
2835  } else if (__kmp_affinity_top_method == affinity_top_method_x2apicid_1f) {
2836  num_leaves = 1;
2837  leaves[0] = 31;
2838  leaf_message_id = kmp_i18n_str_NoLeaf31Support;
2839  } else {
2840  num_leaves = 2;
2841  leaves[0] = 31;
2842  leaves[1] = 11;
2843  leaf_message_id = kmp_i18n_str_NoLeaf11Support;
2844  }
2845 
2846  // Check to see if cpuid leaf 31 or 11 is supported.
2847  __kmp_nThreadsPerCore = nCoresPerPkg = nPackages = 1;
2848  topology_leaf = -1;
2849  for (int i = 0; i < num_leaves; ++i) {
2850  int leaf = leaves[i];
2851  if (highest_leaf < leaf)
2852  continue;
2853  __kmp_x86_cpuid(leaf, 0, &buf);
2854  if (buf.ebx == 0)
2855  continue;
2856  topology_leaf = leaf;
2857  __kmp_x2apicid_get_levels(leaf, &proc_info[0], types, &depth,
2858  &total_description);
2859  if (depth == 0)
2860  continue;
2861  break;
2862  }
2863  if (topology_leaf == -1 || depth == 0) {
2864  *msg_id = leaf_message_id;
2865  __kmp_free(proc_info);
2866  return false;
2867  }
2868  KMP_ASSERT(depth <= INTEL_LEVEL_TYPE_LAST);
2869 
2870  // The algorithm used starts by setting the affinity to each available thread
2871  // and retrieving info from the cpuid instruction, so if we are not capable of
2872  // calling __kmp_get_system_affinity() and __kmp_get_system_affinity(), then
2873  // we need to do something else - use the defaults that we calculated from
2874  // issuing cpuid without binding to each proc.
2875  if (!KMP_AFFINITY_CAPABLE()) {
2876  // Hack to try and infer the machine topology using only the data
2877  // available from cpuid on the current thread, and __kmp_xproc.
2878  KMP_ASSERT(__kmp_affinity.type == affinity_none);
2879  for (int i = 0; i < depth; ++i) {
2880  if (proc_info[0].levels[i].level_type == INTEL_LEVEL_TYPE_SMT) {
2881  __kmp_nThreadsPerCore = proc_info[0].levels[i].nitems;
2882  } else if (proc_info[0].levels[i].level_type == INTEL_LEVEL_TYPE_CORE) {
2883  nCoresPerPkg = proc_info[0].levels[i].nitems;
2884  }
2885  }
2886  __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
2887  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
2888  __kmp_free(proc_info);
2889  return true;
2890  }
2891 
2892  // From here on, we can assume that it is safe to call
2893  // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
2894  // __kmp_affinity.type = affinity_none.
2895 
2896  // Save the affinity mask for the current thread.
2897  kmp_affinity_raii_t previous_affinity;
2898 
2899  // Run through each of the available contexts, binding the current thread
2900  // to it, and obtaining the pertinent information using the cpuid instr.
2901  unsigned int proc;
2902  int hw_thread_index = 0;
2903  bool uniform_caches = true;
2904 
2905  KMP_CPU_SET_ITERATE(proc, __kmp_affin_fullMask) {
2906  // Skip this proc if it is not included in the machine model.
2907  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
2908  continue;
2909  }
2910  KMP_DEBUG_ASSERT(hw_thread_index < __kmp_avail_proc);
2911 
2912  // Gather topology information
2913  __kmp_affinity_dispatch->bind_thread(proc);
2914  __kmp_x86_cpuid(topology_leaf, 0, &buf);
2915  proc_info[hw_thread_index].os_id = proc;
2916  proc_info[hw_thread_index].apic_id = buf.edx;
2917  __kmp_x2apicid_get_levels(topology_leaf, &proc_info[hw_thread_index], types,
2918  &depth, &total_description);
2919  if (proc_info[hw_thread_index].depth == 0) {
2920  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
2921  __kmp_free(proc_info);
2922  return false;
2923  }
2924  // Gather cache information and insert afterwards
2925  cache_info[hw_thread_index].get_leaf4_levels();
2926  if (uniform_caches && hw_thread_index > 0)
2927  if (cache_info[0] != cache_info[hw_thread_index])
2928  uniform_caches = false;
2929  // Hybrid information
2930  if (__kmp_is_hybrid_cpu() && highest_leaf >= 0x1a) {
2931  __kmp_get_hybrid_info(&proc_info[hw_thread_index].type,
2932  &proc_info[hw_thread_index].efficiency,
2933  &proc_info[hw_thread_index].native_model_id);
2934  }
2935  hw_thread_index++;
2936  }
2937  KMP_ASSERT(hw_thread_index > 0);
2938  previous_affinity.restore();
2939 
2940  // Allocate the data structure to be returned.
2941  __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
2942 
2943  // Create topology Ids and hybrid types in __kmp_topology
2944  for (int i = 0; i < __kmp_topology->get_num_hw_threads(); ++i) {
2945  kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
2946  hw_thread.clear();
2947  hw_thread.os_id = proc_info[i].os_id;
2948  hw_thread.original_idx = i;
2949  unsigned apic_id = proc_info[i].apic_id;
2950  // Put in topology information
2951  for (int j = 0, idx = depth - 1; j < depth; ++j, --idx) {
2952  if (!(proc_info[i].description.contains_topology_type(
2953  __kmp_topology->get_type(j)))) {
2954  hw_thread.ids[idx] = kmp_hw_thread_t::UNKNOWN_ID;
2955  } else {
2956  hw_thread.ids[idx] = apic_id & proc_info[i].levels[j].mask;
2957  if (j > 0) {
2958  hw_thread.ids[idx] >>= proc_info[i].levels[j - 1].mask_width;
2959  }
2960  }
2961  }
2962  hw_thread.attrs.set_core_type(proc_info[i].type);
2963  hw_thread.attrs.set_core_eff(proc_info[i].efficiency);
2964  }
2965 
2966  __kmp_topology->sort_ids();
2967 
2968  // Change Ids to logical Ids
2969  for (int j = 0; j < depth - 1; ++j) {
2970  int new_id = 0;
2971  int prev_id = __kmp_topology->at(0).ids[j];
2972  int curr_id = __kmp_topology->at(0).ids[j + 1];
2973  __kmp_topology->at(0).ids[j + 1] = new_id;
2974  for (int i = 1; i < __kmp_topology->get_num_hw_threads(); ++i) {
2975  kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
2976  if (hw_thread.ids[j] == prev_id && hw_thread.ids[j + 1] == curr_id) {
2977  hw_thread.ids[j + 1] = new_id;
2978  } else if (hw_thread.ids[j] == prev_id &&
2979  hw_thread.ids[j + 1] != curr_id) {
2980  curr_id = hw_thread.ids[j + 1];
2981  hw_thread.ids[j + 1] = ++new_id;
2982  } else {
2983  prev_id = hw_thread.ids[j];
2984  curr_id = hw_thread.ids[j + 1];
2985  hw_thread.ids[j + 1] = ++new_id;
2986  }
2987  }
2988  }
2989 
2990  // First check for easy cache placement. This occurs when caches are
2991  // equivalent to a layer in the CPUID leaf 0xb or 0x1f topology.
2992  if (uniform_caches) {
2993  for (size_t i = 0; i < cache_info[0].get_depth(); ++i) {
2994  unsigned cache_mask = cache_info[0][i].mask;
2995  unsigned cache_level = cache_info[0][i].level;
2996  KMP_ASSERT(cache_level <= cpuid_cache_info_t::MAX_CACHE_LEVEL);
2997  kmp_hw_t cache_type = cpuid_cache_info_t::get_topology_type(cache_level);
2998  __kmp_topology->set_equivalent_type(cache_type, cache_type);
2999  for (int j = 0; j < depth; ++j) {
3000  unsigned hw_cache_mask = proc_info[0].levels[j].cache_mask;
3001  if (hw_cache_mask == cache_mask && j < depth - 1) {
3002  kmp_hw_t type = __kmp_intel_type_2_topology_type(
3003  proc_info[0].levels[j + 1].level_type);
3004  __kmp_topology->set_equivalent_type(cache_type, type);
3005  }
3006  }
3007  }
3008  } else {
3009  // If caches are non-uniform, then record which caches exist.
3010  for (int i = 0; i < __kmp_topology->get_num_hw_threads(); ++i) {
3011  for (size_t j = 0; j < cache_info[i].get_depth(); ++j) {
3012  unsigned cache_level = cache_info[i][j].level;
3013  kmp_hw_t cache_type =
3014  cpuid_cache_info_t::get_topology_type(cache_level);
3015  if (__kmp_topology->get_equivalent_type(cache_type) == KMP_HW_UNKNOWN)
3016  __kmp_topology->set_equivalent_type(cache_type, cache_type);
3017  }
3018  }
3019  }
3020 
3021  // See if any cache level needs to be added manually through cache Ids
3022  bool unresolved_cache_levels = false;
3023  for (unsigned level = 1; level <= cpuid_cache_info_t::MAX_CACHE_LEVEL;
3024  ++level) {
3025  kmp_hw_t cache_type = cpuid_cache_info_t::get_topology_type(level);
3026  // This also filters out caches which may not be in the topology
3027  // since the equivalent type might be KMP_HW_UNKNOWN.
3028  if (__kmp_topology->get_equivalent_type(cache_type) == cache_type) {
3029  unresolved_cache_levels = true;
3030  break;
3031  }
3032  }
3033 
3034  // Insert unresolved cache layers into machine topology using cache Ids
3035  if (unresolved_cache_levels) {
3036  int num_hw_threads = __kmp_topology->get_num_hw_threads();
3037  int *ids = (int *)__kmp_allocate(sizeof(int) * num_hw_threads);
3038  for (unsigned l = 1; l <= cpuid_cache_info_t::MAX_CACHE_LEVEL; ++l) {
3039  kmp_hw_t cache_type = cpuid_cache_info_t::get_topology_type(l);
3040  if (__kmp_topology->get_equivalent_type(cache_type) != cache_type)
3041  continue;
3042  for (int i = 0; i < num_hw_threads; ++i) {
3043  int original_idx = __kmp_topology->at(i).original_idx;
3044  ids[i] = kmp_hw_thread_t::UNKNOWN_ID;
3045  const cpuid_cache_info_t::info_t &info =
3046  cache_info[original_idx].get_level(l);
3047  // if cache level not in topology for this processor, then skip
3048  if (info.level == 0)
3049  continue;
3050  ids[i] = info.mask & proc_info[original_idx].apic_id;
3051  }
3052  __kmp_topology->insert_layer(cache_type, ids);
3053  }
3054  }
3055 
3056  if (!__kmp_topology->check_ids()) {
3057  kmp_topology_t::deallocate(__kmp_topology);
3058  __kmp_topology = nullptr;
3059  *msg_id = kmp_i18n_str_x2ApicIDsNotUnique;
3060  __kmp_free(proc_info);
3061  return false;
3062  }
3063  __kmp_free(proc_info);
3064  return true;
3065 }
3066 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3067 
3068 #define osIdIndex 0
3069 #define threadIdIndex 1
3070 #define coreIdIndex 2
3071 #define pkgIdIndex 3
3072 #define nodeIdIndex 4
3073 
3074 typedef unsigned *ProcCpuInfo;
3075 static unsigned maxIndex = pkgIdIndex;
3076 
3077 static int __kmp_affinity_cmp_ProcCpuInfo_phys_id(const void *a,
3078  const void *b) {
3079  unsigned i;
3080  const unsigned *aa = *(unsigned *const *)a;
3081  const unsigned *bb = *(unsigned *const *)b;
3082  for (i = maxIndex;; i--) {
3083  if (aa[i] < bb[i])
3084  return -1;
3085  if (aa[i] > bb[i])
3086  return 1;
3087  if (i == osIdIndex)
3088  break;
3089  }
3090  return 0;
3091 }
3092 
3093 #if KMP_USE_HIER_SCHED
3094 // Set the array sizes for the hierarchy layers
3095 static void __kmp_dispatch_set_hierarchy_values() {
3096  // Set the maximum number of L1's to number of cores
3097  // Set the maximum number of L2's to either number of cores / 2 for
3098  // Intel(R) Xeon Phi(TM) coprocessor formally codenamed Knights Landing
3099  // Or the number of cores for Intel(R) Xeon(R) processors
3100  // Set the maximum number of NUMA nodes and L3's to number of packages
3101  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1] =
3102  nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
3103  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L1 + 1] = __kmp_ncores;
3104 #if KMP_ARCH_X86_64 && \
3105  (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_DRAGONFLY || \
3106  KMP_OS_WINDOWS) && \
3107  KMP_MIC_SUPPORTED
3108  if (__kmp_mic_type >= mic3)
3109  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores / 2;
3110  else
3111 #endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
3112  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores;
3113  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L3 + 1] = nPackages;
3114  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_NUMA + 1] = nPackages;
3115  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_LOOP + 1] = 1;
3116  // Set the number of threads per unit
3117  // Number of hardware threads per L1/L2/L3/NUMA/LOOP
3118  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_THREAD + 1] = 1;
3119  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L1 + 1] =
3120  __kmp_nThreadsPerCore;
3121 #if KMP_ARCH_X86_64 && \
3122  (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_DRAGONFLY || \
3123  KMP_OS_WINDOWS) && \
3124  KMP_MIC_SUPPORTED
3125  if (__kmp_mic_type >= mic3)
3126  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
3127  2 * __kmp_nThreadsPerCore;
3128  else
3129 #endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
3130  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
3131  __kmp_nThreadsPerCore;
3132  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L3 + 1] =
3133  nCoresPerPkg * __kmp_nThreadsPerCore;
3134  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_NUMA + 1] =
3135  nCoresPerPkg * __kmp_nThreadsPerCore;
3136  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_LOOP + 1] =
3137  nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
3138 }
3139 
3140 // Return the index into the hierarchy for this tid and layer type (L1, L2, etc)
3141 // i.e., this thread's L1 or this thread's L2, etc.
3142 int __kmp_dispatch_get_index(int tid, kmp_hier_layer_e type) {
3143  int index = type + 1;
3144  int num_hw_threads = __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1];
3145  KMP_DEBUG_ASSERT(type != kmp_hier_layer_e::LAYER_LAST);
3146  if (type == kmp_hier_layer_e::LAYER_THREAD)
3147  return tid;
3148  else if (type == kmp_hier_layer_e::LAYER_LOOP)
3149  return 0;
3150  KMP_DEBUG_ASSERT(__kmp_hier_max_units[index] != 0);
3151  if (tid >= num_hw_threads)
3152  tid = tid % num_hw_threads;
3153  return (tid / __kmp_hier_threads_per[index]) % __kmp_hier_max_units[index];
3154 }
3155 
3156 // Return the number of t1's per t2
3157 int __kmp_dispatch_get_t1_per_t2(kmp_hier_layer_e t1, kmp_hier_layer_e t2) {
3158  int i1 = t1 + 1;
3159  int i2 = t2 + 1;
3160  KMP_DEBUG_ASSERT(i1 <= i2);
3161  KMP_DEBUG_ASSERT(t1 != kmp_hier_layer_e::LAYER_LAST);
3162  KMP_DEBUG_ASSERT(t2 != kmp_hier_layer_e::LAYER_LAST);
3163  KMP_DEBUG_ASSERT(__kmp_hier_threads_per[i1] != 0);
3164  // (nthreads/t2) / (nthreads/t1) = t1 / t2
3165  return __kmp_hier_threads_per[i2] / __kmp_hier_threads_per[i1];
3166 }
3167 #endif // KMP_USE_HIER_SCHED
3168 
3169 static inline const char *__kmp_cpuinfo_get_filename() {
3170  const char *filename;
3171  if (__kmp_cpuinfo_file != nullptr)
3172  filename = __kmp_cpuinfo_file;
3173  else
3174  filename = "/proc/cpuinfo";
3175  return filename;
3176 }
3177 
3178 static inline const char *__kmp_cpuinfo_get_envvar() {
3179  const char *envvar = nullptr;
3180  if (__kmp_cpuinfo_file != nullptr)
3181  envvar = "KMP_CPUINFO_FILE";
3182  return envvar;
3183 }
3184 
3185 static bool __kmp_package_id_from_core_siblings_list(unsigned **threadInfo,
3186  unsigned num_avail,
3187  unsigned idx) {
3188  if (!KMP_AFFINITY_CAPABLE())
3189  return false;
3190 
3191  char path[256];
3192  KMP_SNPRINTF(path, sizeof(path),
3193  "/sys/devices/system/cpu/cpu%u/topology/core_siblings_list",
3194  threadInfo[idx][osIdIndex]);
3195  kmp_affin_mask_t *siblings = __kmp_parse_cpu_list(path);
3196  for (unsigned i = 0; i < num_avail; ++i) {
3197  unsigned cpu_id = threadInfo[i][osIdIndex];
3198  KMP_ASSERT(cpu_id < __kmp_affin_mask_size * CHAR_BIT);
3199  if (!KMP_CPU_ISSET(cpu_id, siblings))
3200  continue;
3201  if (threadInfo[i][pkgIdIndex] == UINT_MAX) {
3202  // Arbitrarily pick the first index we encounter, it only matters that
3203  // the value is the same for all siblings.
3204  threadInfo[i][pkgIdIndex] = idx;
3205  } else if (threadInfo[i][pkgIdIndex] != idx) {
3206  // Contradictory sibling lists.
3207  KMP_CPU_FREE(siblings);
3208  return false;
3209  }
3210  }
3211  KMP_ASSERT(threadInfo[idx][pkgIdIndex] != UINT_MAX);
3212  KMP_CPU_FREE(siblings);
3213  return true;
3214 }
3215 
3216 // Parse /proc/cpuinfo (or an alternate file in the same format) to obtain the
3217 // affinity map. On AIX, the map is obtained through system SRAD (Scheduler
3218 // Resource Allocation Domain).
3219 static bool __kmp_affinity_create_cpuinfo_map(int *line,
3220  kmp_i18n_id_t *const msg_id) {
3221  *msg_id = kmp_i18n_null;
3222 
3223 #if KMP_OS_AIX
3224  unsigned num_records = __kmp_xproc;
3225 #else
3226  const char *filename = __kmp_cpuinfo_get_filename();
3227  const char *envvar = __kmp_cpuinfo_get_envvar();
3228 
3229  if (__kmp_affinity.flags.verbose) {
3230  KMP_INFORM(AffParseFilename, "KMP_AFFINITY", filename);
3231  }
3232 
3233  kmp_safe_raii_file_t f(filename, "r", envvar);
3234 
3235  // Scan of the file, and count the number of "processor" (osId) fields,
3236  // and find the highest value of <n> for a node_<n> field.
3237  char buf[256];
3238  unsigned num_records = 0;
3239  while (!feof(f)) {
3240  buf[sizeof(buf) - 1] = 1;
3241  if (!fgets(buf, sizeof(buf), f)) {
3242  // Read errors presumably because of EOF
3243  break;
3244  }
3245 
3246  char s1[] = "processor";
3247  if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
3248  num_records++;
3249  continue;
3250  }
3251 
3252  // FIXME - this will match "node_<n> <garbage>"
3253  unsigned level;
3254  if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
3255  // validate the input fisrt:
3256  if (level > (unsigned)__kmp_xproc) { // level is too big
3257  level = __kmp_xproc;
3258  }
3259  if (nodeIdIndex + level >= maxIndex) {
3260  maxIndex = nodeIdIndex + level;
3261  }
3262  continue;
3263  }
3264  }
3265 
3266  // Check for empty file / no valid processor records, or too many. The number
3267  // of records can't exceed the number of valid bits in the affinity mask.
3268  if (num_records == 0) {
3269  *msg_id = kmp_i18n_str_NoProcRecords;
3270  return false;
3271  }
3272  if (num_records > (unsigned)__kmp_xproc) {
3273  *msg_id = kmp_i18n_str_TooManyProcRecords;
3274  return false;
3275  }
3276 
3277  // Set the file pointer back to the beginning, so that we can scan the file
3278  // again, this time performing a full parse of the data. Allocate a vector of
3279  // ProcCpuInfo object, where we will place the data. Adding an extra element
3280  // at the end allows us to remove a lot of extra checks for termination
3281  // conditions.
3282  if (fseek(f, 0, SEEK_SET) != 0) {
3283  *msg_id = kmp_i18n_str_CantRewindCpuinfo;
3284  return false;
3285  }
3286 #endif // KMP_OS_AIX
3287 
3288  // Allocate the array of records to store the proc info in. The dummy
3289  // element at the end makes the logic in filling them out easier to code.
3290  unsigned **threadInfo =
3291  (unsigned **)__kmp_allocate((num_records + 1) * sizeof(unsigned *));
3292  unsigned i;
3293  for (i = 0; i <= num_records; i++) {
3294  threadInfo[i] =
3295  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3296  }
3297 
3298 #define CLEANUP_THREAD_INFO \
3299  for (i = 0; i <= num_records; i++) { \
3300  __kmp_free(threadInfo[i]); \
3301  } \
3302  __kmp_free(threadInfo);
3303 
3304  // A value of UINT_MAX means that we didn't find the field
3305  unsigned __index;
3306 
3307 #define INIT_PROC_INFO(p) \
3308  for (__index = 0; __index <= maxIndex; __index++) { \
3309  (p)[__index] = UINT_MAX; \
3310  }
3311 
3312  for (i = 0; i <= num_records; i++) {
3313  INIT_PROC_INFO(threadInfo[i]);
3314  }
3315 
3316 #if KMP_OS_AIX
3317  int smt_threads;
3318  lpar_info_format1_t cpuinfo;
3319  unsigned num_avail = __kmp_xproc;
3320 
3321  if (__kmp_affinity.flags.verbose)
3322  KMP_INFORM(AffParseFilename, "KMP_AFFINITY", "system info for topology");
3323 
3324  // Get the number of SMT threads per core.
3325  smt_threads = syssmt(GET_NUMBER_SMT_SETS, 0, 0, NULL);
3326 
3327  // Allocate a resource set containing available system resourses.
3328  rsethandle_t sys_rset = rs_alloc(RS_SYSTEM);
3329  if (sys_rset == NULL) {
3330  CLEANUP_THREAD_INFO;
3331  *msg_id = kmp_i18n_str_UnknownTopology;
3332  return false;
3333  }
3334  // Allocate a resource set for the SRAD info.
3335  rsethandle_t srad = rs_alloc(RS_EMPTY);
3336  if (srad == NULL) {
3337  rs_free(sys_rset);
3338  CLEANUP_THREAD_INFO;
3339  *msg_id = kmp_i18n_str_UnknownTopology;
3340  return false;
3341  }
3342 
3343  // Get the SRAD system detail level.
3344  int sradsdl = rs_getinfo(NULL, R_SRADSDL, 0);
3345  if (sradsdl < 0) {
3346  rs_free(sys_rset);
3347  rs_free(srad);
3348  CLEANUP_THREAD_INFO;
3349  *msg_id = kmp_i18n_str_UnknownTopology;
3350  return false;
3351  }
3352  // Get the number of RADs at that SRAD SDL.
3353  int num_rads = rs_numrads(sys_rset, sradsdl, 0);
3354  if (num_rads < 0) {
3355  rs_free(sys_rset);
3356  rs_free(srad);
3357  CLEANUP_THREAD_INFO;
3358  *msg_id = kmp_i18n_str_UnknownTopology;
3359  return false;
3360  }
3361 
3362  // Get the maximum number of procs that may be contained in a resource set.
3363  int max_procs = rs_getinfo(NULL, R_MAXPROCS, 0);
3364  if (max_procs < 0) {
3365  rs_free(sys_rset);
3366  rs_free(srad);
3367  CLEANUP_THREAD_INFO;
3368  *msg_id = kmp_i18n_str_UnknownTopology;
3369  return false;
3370  }
3371 
3372  int cur_rad = 0;
3373  int num_set = 0;
3374  for (int srad_idx = 0; cur_rad < num_rads && srad_idx < VMI_MAXRADS;
3375  ++srad_idx) {
3376  // Check if the SRAD is available in the RSET.
3377  if (rs_getrad(sys_rset, srad, sradsdl, srad_idx, 0) < 0)
3378  continue;
3379 
3380  for (int cpu = 0; cpu < max_procs; cpu++) {
3381  // Set the info for the cpu if it is in the SRAD.
3382  if (rs_op(RS_TESTRESOURCE, srad, NULL, R_PROCS, cpu)) {
3383  threadInfo[cpu][osIdIndex] = cpu;
3384  threadInfo[cpu][pkgIdIndex] = cur_rad;
3385  threadInfo[cpu][coreIdIndex] = cpu / smt_threads;
3386  ++num_set;
3387  if (num_set >= num_avail) {
3388  // Done if all available CPUs have been set.
3389  break;
3390  }
3391  }
3392  }
3393  ++cur_rad;
3394  }
3395  rs_free(sys_rset);
3396  rs_free(srad);
3397 
3398  // The topology is already sorted.
3399 
3400 #else // !KMP_OS_AIX
3401  unsigned num_avail = 0;
3402  *line = 0;
3403 #if KMP_ARCH_S390X
3404  bool reading_s390x_sys_info = true;
3405 #endif
3406  while (!feof(f)) {
3407  // Create an inner scoping level, so that all the goto targets at the end of
3408  // the loop appear in an outer scoping level. This avoids warnings about
3409  // jumping past an initialization to a target in the same block.
3410  {
3411  buf[sizeof(buf) - 1] = 1;
3412  bool long_line = false;
3413  if (!fgets(buf, sizeof(buf), f)) {
3414  // Read errors presumably because of EOF
3415  // If there is valid data in threadInfo[num_avail], then fake
3416  // a blank line in ensure that the last address gets parsed.
3417  bool valid = false;
3418  for (i = 0; i <= maxIndex; i++) {
3419  if (threadInfo[num_avail][i] != UINT_MAX) {
3420  valid = true;
3421  }
3422  }
3423  if (!valid) {
3424  break;
3425  }
3426  buf[0] = 0;
3427  } else if (!buf[sizeof(buf) - 1]) {
3428  // The line is longer than the buffer. Set a flag and don't
3429  // emit an error if we were going to ignore the line, anyway.
3430  long_line = true;
3431 
3432 #define CHECK_LINE \
3433  if (long_line) { \
3434  CLEANUP_THREAD_INFO; \
3435  *msg_id = kmp_i18n_str_LongLineCpuinfo; \
3436  return false; \
3437  }
3438  }
3439  (*line)++;
3440 
3441 #if KMP_ARCH_LOONGARCH64
3442  // The parsing logic of /proc/cpuinfo in this function highly depends on
3443  // the blank lines between each processor info block. But on LoongArch a
3444  // blank line exists before the first processor info block (i.e. after the
3445  // "system type" line). This blank line was added because the "system
3446  // type" line is unrelated to any of the CPUs. We must skip this line so
3447  // that the original logic works on LoongArch.
3448  if (*buf == '\n' && *line == 2)
3449  continue;
3450 #endif
3451 #if KMP_ARCH_S390X
3452  // s390x /proc/cpuinfo starts with a variable number of lines containing
3453  // the overall system information. Skip them.
3454  if (reading_s390x_sys_info) {
3455  if (*buf == '\n')
3456  reading_s390x_sys_info = false;
3457  continue;
3458  }
3459 #endif
3460 
3461 #if KMP_ARCH_S390X
3462  char s1[] = "cpu number";
3463 #else
3464  char s1[] = "processor";
3465 #endif
3466  if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
3467  CHECK_LINE;
3468  char *p = strchr(buf + sizeof(s1) - 1, ':');
3469  unsigned val;
3470  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3471  goto no_val;
3472  if (threadInfo[num_avail][osIdIndex] != UINT_MAX)
3473 #if KMP_ARCH_AARCH64
3474  // Handle the old AArch64 /proc/cpuinfo layout differently,
3475  // it contains all of the 'processor' entries listed in a
3476  // single 'Processor' section, therefore the normal looking
3477  // for duplicates in that section will always fail.
3478  num_avail++;
3479 #else
3480  goto dup_field;
3481 #endif
3482  threadInfo[num_avail][osIdIndex] = val;
3483 #if KMP_OS_LINUX && !(KMP_ARCH_X86 || KMP_ARCH_X86_64)
3484  char path[256];
3485  KMP_SNPRINTF(
3486  path, sizeof(path),
3487  "/sys/devices/system/cpu/cpu%u/topology/physical_package_id",
3488  threadInfo[num_avail][osIdIndex]);
3489  __kmp_read_from_file(path, "%u", &threadInfo[num_avail][pkgIdIndex]);
3490 
3491 #if KMP_ARCH_S390X
3492  // Disambiguate physical_package_id.
3493  unsigned book_id;
3494  KMP_SNPRINTF(path, sizeof(path),
3495  "/sys/devices/system/cpu/cpu%u/topology/book_id",
3496  threadInfo[num_avail][osIdIndex]);
3497  __kmp_read_from_file(path, "%u", &book_id);
3498  threadInfo[num_avail][pkgIdIndex] |= (book_id << 8);
3499 
3500  unsigned drawer_id;
3501  KMP_SNPRINTF(path, sizeof(path),
3502  "/sys/devices/system/cpu/cpu%u/topology/drawer_id",
3503  threadInfo[num_avail][osIdIndex]);
3504  __kmp_read_from_file(path, "%u", &drawer_id);
3505  threadInfo[num_avail][pkgIdIndex] |= (drawer_id << 16);
3506 #endif
3507 
3508  KMP_SNPRINTF(path, sizeof(path),
3509  "/sys/devices/system/cpu/cpu%u/topology/core_id",
3510  threadInfo[num_avail][osIdIndex]);
3511  __kmp_read_from_file(path, "%u", &threadInfo[num_avail][coreIdIndex]);
3512  continue;
3513 #else
3514  }
3515  char s2[] = "physical id";
3516  if (strncmp(buf, s2, sizeof(s2) - 1) == 0) {
3517  CHECK_LINE;
3518  char *p = strchr(buf + sizeof(s2) - 1, ':');
3519  unsigned val;
3520  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3521  goto no_val;
3522  if (threadInfo[num_avail][pkgIdIndex] != UINT_MAX)
3523  goto dup_field;
3524  threadInfo[num_avail][pkgIdIndex] = val;
3525  continue;
3526  }
3527  char s3[] = "core id";
3528  if (strncmp(buf, s3, sizeof(s3) - 1) == 0) {
3529  CHECK_LINE;
3530  char *p = strchr(buf + sizeof(s3) - 1, ':');
3531  unsigned val;
3532  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3533  goto no_val;
3534  if (threadInfo[num_avail][coreIdIndex] != UINT_MAX)
3535  goto dup_field;
3536  threadInfo[num_avail][coreIdIndex] = val;
3537  continue;
3538 #endif // KMP_OS_LINUX && USE_SYSFS_INFO
3539  }
3540  char s4[] = "thread id";
3541  if (strncmp(buf, s4, sizeof(s4) - 1) == 0) {
3542  CHECK_LINE;
3543  char *p = strchr(buf + sizeof(s4) - 1, ':');
3544  unsigned val;
3545  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3546  goto no_val;
3547  if (threadInfo[num_avail][threadIdIndex] != UINT_MAX)
3548  goto dup_field;
3549  threadInfo[num_avail][threadIdIndex] = val;
3550  continue;
3551  }
3552  unsigned level;
3553  if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
3554  CHECK_LINE;
3555  char *p = strchr(buf + sizeof(s4) - 1, ':');
3556  unsigned val;
3557  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3558  goto no_val;
3559  // validate the input before using level:
3560  if (level > (unsigned)__kmp_xproc) { // level is too big
3561  level = __kmp_xproc;
3562  }
3563  if (threadInfo[num_avail][nodeIdIndex + level] != UINT_MAX)
3564  goto dup_field;
3565  threadInfo[num_avail][nodeIdIndex + level] = val;
3566  continue;
3567  }
3568 
3569  // We didn't recognize the leading token on the line. There are lots of
3570  // leading tokens that we don't recognize - if the line isn't empty, go on
3571  // to the next line.
3572  if ((*buf != 0) && (*buf != '\n')) {
3573  // If the line is longer than the buffer, read characters
3574  // until we find a newline.
3575  if (long_line) {
3576  int ch;
3577  while (((ch = fgetc(f)) != EOF) && (ch != '\n'))
3578  ;
3579  }
3580  continue;
3581  }
3582 
3583  // A newline has signalled the end of the processor record.
3584  // Check that there aren't too many procs specified.
3585  if ((int)num_avail == __kmp_xproc) {
3586  CLEANUP_THREAD_INFO;
3587  *msg_id = kmp_i18n_str_TooManyEntries;
3588  return false;
3589  }
3590 
3591  // Check for missing fields. The osId field must be there. The physical
3592  // id field will be checked later.
3593  if (threadInfo[num_avail][osIdIndex] == UINT_MAX) {
3594  CLEANUP_THREAD_INFO;
3595  *msg_id = kmp_i18n_str_MissingProcField;
3596  return false;
3597  }
3598 
3599  // Skip this proc if it is not included in the machine model.
3600  if (KMP_AFFINITY_CAPABLE() &&
3601  !KMP_CPU_ISSET(threadInfo[num_avail][osIdIndex],
3602  __kmp_affin_fullMask)) {
3603  INIT_PROC_INFO(threadInfo[num_avail]);
3604  continue;
3605  }
3606 
3607  // We have a successful parse of this proc's info.
3608  // Increment the counter, and prepare for the next proc.
3609  num_avail++;
3610  KMP_ASSERT(num_avail <= num_records);
3611  INIT_PROC_INFO(threadInfo[num_avail]);
3612  }
3613  continue;
3614 
3615  no_val:
3616  CLEANUP_THREAD_INFO;
3617  *msg_id = kmp_i18n_str_MissingValCpuinfo;
3618  return false;
3619 
3620  dup_field:
3621  CLEANUP_THREAD_INFO;
3622  *msg_id = kmp_i18n_str_DuplicateFieldCpuinfo;
3623  return false;
3624  }
3625  *line = 0;
3626 
3627  // At least on powerpc, Linux may return -1 for physical_package_id. Try
3628  // to reconstruct topology from core_siblings_list in that case.
3629  for (i = 0; i < num_avail; ++i) {
3630  if (threadInfo[i][pkgIdIndex] == UINT_MAX) {
3631  if (!__kmp_package_id_from_core_siblings_list(threadInfo, num_avail, i)) {
3632  CLEANUP_THREAD_INFO;
3633  *msg_id = kmp_i18n_str_MissingPhysicalIDField;
3634  return false;
3635  }
3636  }
3637  }
3638 
3639 #if KMP_MIC && REDUCE_TEAM_SIZE
3640  unsigned teamSize = 0;
3641 #endif // KMP_MIC && REDUCE_TEAM_SIZE
3642 
3643  // check for num_records == __kmp_xproc ???
3644 
3645  // If it is configured to omit the package level when there is only a single
3646  // package, the logic at the end of this routine won't work if there is only a
3647  // single thread
3648  KMP_ASSERT(num_avail > 0);
3649  KMP_ASSERT(num_avail <= num_records);
3650 
3651  // Sort the threadInfo table by physical Id.
3652  qsort(threadInfo, num_avail, sizeof(*threadInfo),
3653  __kmp_affinity_cmp_ProcCpuInfo_phys_id);
3654 
3655 #endif // KMP_OS_AIX
3656 
3657  // The table is now sorted by pkgId / coreId / threadId, but we really don't
3658  // know the radix of any of the fields. pkgId's may be sparsely assigned among
3659  // the chips on a system. Although coreId's are usually assigned
3660  // [0 .. coresPerPkg-1] and threadId's are usually assigned
3661  // [0..threadsPerCore-1], we don't want to make any such assumptions.
3662  //
3663  // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
3664  // total # packages) are at this point - we want to determine that now. We
3665  // only have an upper bound on the first two figures.
3666  unsigned *counts =
3667  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3668  unsigned *maxCt =
3669  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3670  unsigned *totals =
3671  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3672  unsigned *lastId =
3673  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3674 
3675  bool assign_thread_ids = false;
3676  unsigned threadIdCt;
3677  unsigned index;
3678 
3679 restart_radix_check:
3680  threadIdCt = 0;
3681 
3682  // Initialize the counter arrays with data from threadInfo[0].
3683  if (assign_thread_ids) {
3684  if (threadInfo[0][threadIdIndex] == UINT_MAX) {
3685  threadInfo[0][threadIdIndex] = threadIdCt++;
3686  } else if (threadIdCt <= threadInfo[0][threadIdIndex]) {
3687  threadIdCt = threadInfo[0][threadIdIndex] + 1;
3688  }
3689  }
3690  for (index = 0; index <= maxIndex; index++) {
3691  counts[index] = 1;
3692  maxCt[index] = 1;
3693  totals[index] = 1;
3694  lastId[index] = threadInfo[0][index];
3695  ;
3696  }
3697 
3698  // Run through the rest of the OS procs.
3699  for (i = 1; i < num_avail; i++) {
3700  // Find the most significant index whose id differs from the id for the
3701  // previous OS proc.
3702  for (index = maxIndex; index >= threadIdIndex; index--) {
3703  if (assign_thread_ids && (index == threadIdIndex)) {
3704  // Auto-assign the thread id field if it wasn't specified.
3705  if (threadInfo[i][threadIdIndex] == UINT_MAX) {
3706  threadInfo[i][threadIdIndex] = threadIdCt++;
3707  }
3708  // Apparently the thread id field was specified for some entries and not
3709  // others. Start the thread id counter off at the next higher thread id.
3710  else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
3711  threadIdCt = threadInfo[i][threadIdIndex] + 1;
3712  }
3713  }
3714  if (threadInfo[i][index] != lastId[index]) {
3715  // Run through all indices which are less significant, and reset the
3716  // counts to 1. At all levels up to and including index, we need to
3717  // increment the totals and record the last id.
3718  unsigned index2;
3719  for (index2 = threadIdIndex; index2 < index; index2++) {
3720  totals[index2]++;
3721  if (counts[index2] > maxCt[index2]) {
3722  maxCt[index2] = counts[index2];
3723  }
3724  counts[index2] = 1;
3725  lastId[index2] = threadInfo[i][index2];
3726  }
3727  counts[index]++;
3728  totals[index]++;
3729  lastId[index] = threadInfo[i][index];
3730 
3731  if (assign_thread_ids && (index > threadIdIndex)) {
3732 
3733 #if KMP_MIC && REDUCE_TEAM_SIZE
3734  // The default team size is the total #threads in the machine
3735  // minus 1 thread for every core that has 3 or more threads.
3736  teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
3737 #endif // KMP_MIC && REDUCE_TEAM_SIZE
3738 
3739  // Restart the thread counter, as we are on a new core.
3740  threadIdCt = 0;
3741 
3742  // Auto-assign the thread id field if it wasn't specified.
3743  if (threadInfo[i][threadIdIndex] == UINT_MAX) {
3744  threadInfo[i][threadIdIndex] = threadIdCt++;
3745  }
3746 
3747  // Apparently the thread id field was specified for some entries and
3748  // not others. Start the thread id counter off at the next higher
3749  // thread id.
3750  else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
3751  threadIdCt = threadInfo[i][threadIdIndex] + 1;
3752  }
3753  }
3754  break;
3755  }
3756  }
3757  if (index < threadIdIndex) {
3758  // If thread ids were specified, it is an error if they are not unique.
3759  // Also, check that we waven't already restarted the loop (to be safe -
3760  // shouldn't need to).
3761  if ((threadInfo[i][threadIdIndex] != UINT_MAX) || assign_thread_ids) {
3762  __kmp_free(lastId);
3763  __kmp_free(totals);
3764  __kmp_free(maxCt);
3765  __kmp_free(counts);
3766  CLEANUP_THREAD_INFO;
3767  *msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
3768  return false;
3769  }
3770 
3771  // If the thread ids were not specified and we see entries that
3772  // are duplicates, start the loop over and assign the thread ids manually.
3773  assign_thread_ids = true;
3774  goto restart_radix_check;
3775  }
3776  }
3777 
3778 #if KMP_MIC && REDUCE_TEAM_SIZE
3779  // The default team size is the total #threads in the machine
3780  // minus 1 thread for every core that has 3 or more threads.
3781  teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
3782 #endif // KMP_MIC && REDUCE_TEAM_SIZE
3783 
3784  for (index = threadIdIndex; index <= maxIndex; index++) {
3785  if (counts[index] > maxCt[index]) {
3786  maxCt[index] = counts[index];
3787  }
3788  }
3789 
3790  __kmp_nThreadsPerCore = maxCt[threadIdIndex];
3791  nCoresPerPkg = maxCt[coreIdIndex];
3792  nPackages = totals[pkgIdIndex];
3793 
3794  // When affinity is off, this routine will still be called to set
3795  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
3796  // Make sure all these vars are set correctly, and return now if affinity is
3797  // not enabled.
3798  __kmp_ncores = totals[coreIdIndex];
3799  if (!KMP_AFFINITY_CAPABLE()) {
3800  KMP_ASSERT(__kmp_affinity.type == affinity_none);
3801  return true;
3802  }
3803 
3804 #if KMP_MIC && REDUCE_TEAM_SIZE
3805  // Set the default team size.
3806  if ((__kmp_dflt_team_nth == 0) && (teamSize > 0)) {
3807  __kmp_dflt_team_nth = teamSize;
3808  KA_TRACE(20, ("__kmp_affinity_create_cpuinfo_map: setting "
3809  "__kmp_dflt_team_nth = %d\n",
3810  __kmp_dflt_team_nth));
3811  }
3812 #endif // KMP_MIC && REDUCE_TEAM_SIZE
3813 
3814  KMP_DEBUG_ASSERT(num_avail == (unsigned)__kmp_avail_proc);
3815 
3816  // Count the number of levels which have more nodes at that level than at the
3817  // parent's level (with there being an implicit root node of the top level).
3818  // This is equivalent to saying that there is at least one node at this level
3819  // which has a sibling. These levels are in the map, and the package level is
3820  // always in the map.
3821  bool *inMap = (bool *)__kmp_allocate((maxIndex + 1) * sizeof(bool));
3822  for (index = threadIdIndex; index < maxIndex; index++) {
3823  KMP_ASSERT(totals[index] >= totals[index + 1]);
3824  inMap[index] = (totals[index] > totals[index + 1]);
3825  }
3826  inMap[maxIndex] = (totals[maxIndex] > 1);
3827  inMap[pkgIdIndex] = true;
3828  inMap[coreIdIndex] = true;
3829  inMap[threadIdIndex] = true;
3830 
3831  int depth = 0;
3832  int idx = 0;
3833  kmp_hw_t types[KMP_HW_LAST];
3834  int pkgLevel = -1;
3835  int coreLevel = -1;
3836  int threadLevel = -1;
3837  for (index = threadIdIndex; index <= maxIndex; index++) {
3838  if (inMap[index]) {
3839  depth++;
3840  }
3841  }
3842  if (inMap[pkgIdIndex]) {
3843  pkgLevel = idx;
3844  types[idx++] = KMP_HW_SOCKET;
3845  }
3846  if (inMap[coreIdIndex]) {
3847  coreLevel = idx;
3848  types[idx++] = KMP_HW_CORE;
3849  }
3850  if (inMap[threadIdIndex]) {
3851  threadLevel = idx;
3852  types[idx++] = KMP_HW_THREAD;
3853  }
3854  KMP_ASSERT(depth > 0);
3855 
3856  // Construct the data structure that is to be returned.
3857  __kmp_topology = kmp_topology_t::allocate(num_avail, depth, types);
3858 
3859  for (i = 0; i < num_avail; ++i) {
3860  unsigned os = threadInfo[i][osIdIndex];
3861  int src_index;
3862  kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
3863  hw_thread.clear();
3864  hw_thread.os_id = os;
3865  hw_thread.original_idx = i;
3866 
3867  idx = 0;
3868  for (src_index = maxIndex; src_index >= threadIdIndex; src_index--) {
3869  if (!inMap[src_index]) {
3870  continue;
3871  }
3872  if (src_index == pkgIdIndex) {
3873  hw_thread.ids[pkgLevel] = threadInfo[i][src_index];
3874  } else if (src_index == coreIdIndex) {
3875  hw_thread.ids[coreLevel] = threadInfo[i][src_index];
3876  } else if (src_index == threadIdIndex) {
3877  hw_thread.ids[threadLevel] = threadInfo[i][src_index];
3878  }
3879  }
3880  }
3881 
3882  __kmp_free(inMap);
3883  __kmp_free(lastId);
3884  __kmp_free(totals);
3885  __kmp_free(maxCt);
3886  __kmp_free(counts);
3887  CLEANUP_THREAD_INFO;
3888  __kmp_topology->sort_ids();
3889 
3890  int tlevel = __kmp_topology->get_level(KMP_HW_THREAD);
3891  if (tlevel > 0) {
3892  // If the thread level does not have ids, then put them in.
3893  if (__kmp_topology->at(0).ids[tlevel] == kmp_hw_thread_t::UNKNOWN_ID) {
3894  __kmp_topology->at(0).ids[tlevel] = 0;
3895  }
3896  for (int i = 1; i < __kmp_topology->get_num_hw_threads(); ++i) {
3897  kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
3898  if (hw_thread.ids[tlevel] != kmp_hw_thread_t::UNKNOWN_ID)
3899  continue;
3900  kmp_hw_thread_t &prev_hw_thread = __kmp_topology->at(i - 1);
3901  // Check if socket, core, anything above thread level changed.
3902  // If the ids did change, then restart thread id at 0
3903  // Otherwise, set thread id to prev thread's id + 1
3904  for (int j = 0; j < tlevel; ++j) {
3905  if (hw_thread.ids[j] != prev_hw_thread.ids[j]) {
3906  hw_thread.ids[tlevel] = 0;
3907  break;
3908  }
3909  }
3910  if (hw_thread.ids[tlevel] == kmp_hw_thread_t::UNKNOWN_ID)
3911  hw_thread.ids[tlevel] = prev_hw_thread.ids[tlevel] + 1;
3912  }
3913  }
3914 
3915  if (!__kmp_topology->check_ids()) {
3916  kmp_topology_t::deallocate(__kmp_topology);
3917  __kmp_topology = nullptr;
3918  *msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
3919  return false;
3920  }
3921  return true;
3922 }
3923 
3924 // Create and return a table of affinity masks, indexed by OS thread ID.
3925 // This routine handles OR'ing together all the affinity masks of threads
3926 // that are sufficiently close, if granularity > fine.
3927 template <typename FindNextFunctionType>
3928 static void __kmp_create_os_id_masks(unsigned *numUnique,
3929  kmp_affinity_t &affinity,
3930  FindNextFunctionType find_next) {
3931  // First form a table of affinity masks in order of OS thread id.
3932  int maxOsId;
3933  int i;
3934  int numAddrs = __kmp_topology->get_num_hw_threads();
3935  int depth = __kmp_topology->get_depth();
3936  const char *env_var = __kmp_get_affinity_env_var(affinity);
3937  KMP_ASSERT(numAddrs);
3938  KMP_ASSERT(depth);
3939 
3940  i = find_next(-1);
3941  // If could not find HW thread location that satisfies find_next conditions,
3942  // then return and fallback to increment find_next.
3943  if (i >= numAddrs)
3944  return;
3945 
3946  maxOsId = 0;
3947  for (i = numAddrs - 1;; --i) {
3948  int osId = __kmp_topology->at(i).os_id;
3949  if (osId > maxOsId) {
3950  maxOsId = osId;
3951  }
3952  if (i == 0)
3953  break;
3954  }
3955  affinity.num_os_id_masks = maxOsId + 1;
3956  KMP_CPU_ALLOC_ARRAY(affinity.os_id_masks, affinity.num_os_id_masks);
3957  KMP_ASSERT(affinity.gran_levels >= 0);
3958  if (affinity.flags.verbose && (affinity.gran_levels > 0)) {
3959  KMP_INFORM(ThreadsMigrate, env_var, affinity.gran_levels);
3960  }
3961  if (affinity.gran_levels >= (int)depth) {
3962  KMP_AFF_WARNING(affinity, AffThreadsMayMigrate);
3963  }
3964 
3965  // Run through the table, forming the masks for all threads on each core.
3966  // Threads on the same core will have identical kmp_hw_thread_t objects, not
3967  // considering the last level, which must be the thread id. All threads on a
3968  // core will appear consecutively.
3969  int unique = 0;
3970  int j = 0; // index of 1st thread on core
3971  int leader = 0;
3972  kmp_affin_mask_t *sum;
3973  KMP_CPU_ALLOC_ON_STACK(sum);
3974  KMP_CPU_ZERO(sum);
3975 
3976  i = j = leader = find_next(-1);
3977  KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
3978  kmp_full_mask_modifier_t full_mask;
3979  for (i = find_next(i); i < numAddrs; i = find_next(i)) {
3980  // If this thread is sufficiently close to the leader (within the
3981  // granularity setting), then set the bit for this os thread in the
3982  // affinity mask for this group, and go on to the next thread.
3983  if (__kmp_topology->is_close(leader, i, affinity)) {
3984  KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
3985  continue;
3986  }
3987 
3988  // For every thread in this group, copy the mask to the thread's entry in
3989  // the OS Id mask table. Mark the first address as a leader.
3990  for (; j < i; j = find_next(j)) {
3991  int osId = __kmp_topology->at(j).os_id;
3992  KMP_DEBUG_ASSERT(osId <= maxOsId);
3993  kmp_affin_mask_t *mask = KMP_CPU_INDEX(affinity.os_id_masks, osId);
3994  KMP_CPU_COPY(mask, sum);
3995  __kmp_topology->at(j).leader = (j == leader);
3996  }
3997  unique++;
3998 
3999  // Start a new mask.
4000  leader = i;
4001  full_mask.include(sum);
4002  KMP_CPU_ZERO(sum);
4003  KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
4004  }
4005 
4006  // For every thread in last group, copy the mask to the thread's
4007  // entry in the OS Id mask table.
4008  for (; j < i; j = find_next(j)) {
4009  int osId = __kmp_topology->at(j).os_id;
4010  KMP_DEBUG_ASSERT(osId <= maxOsId);
4011  kmp_affin_mask_t *mask = KMP_CPU_INDEX(affinity.os_id_masks, osId);
4012  KMP_CPU_COPY(mask, sum);
4013  __kmp_topology->at(j).leader = (j == leader);
4014  }
4015  full_mask.include(sum);
4016  unique++;
4017  KMP_CPU_FREE_FROM_STACK(sum);
4018 
4019  // See if the OS Id mask table further restricts or changes the full mask
4020  if (full_mask.restrict_to_mask() && affinity.flags.verbose) {
4021  __kmp_topology->print(env_var);
4022  }
4023 
4024  *numUnique = unique;
4025 }
4026 
4027 // Stuff for the affinity proclist parsers. It's easier to declare these vars
4028 // as file-static than to try and pass them through the calling sequence of
4029 // the recursive-descent OMP_PLACES parser.
4030 static kmp_affin_mask_t *newMasks;
4031 static int numNewMasks;
4032 static int nextNewMask;
4033 
4034 #define ADD_MASK(_mask) \
4035  { \
4036  if (nextNewMask >= numNewMasks) { \
4037  int i; \
4038  numNewMasks *= 2; \
4039  kmp_affin_mask_t *temp; \
4040  KMP_CPU_INTERNAL_ALLOC_ARRAY(temp, numNewMasks); \
4041  for (i = 0; i < numNewMasks / 2; i++) { \
4042  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i); \
4043  kmp_affin_mask_t *dest = KMP_CPU_INDEX(temp, i); \
4044  KMP_CPU_COPY(dest, src); \
4045  } \
4046  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks / 2); \
4047  newMasks = temp; \
4048  } \
4049  KMP_CPU_COPY(KMP_CPU_INDEX(newMasks, nextNewMask), (_mask)); \
4050  nextNewMask++; \
4051  }
4052 
4053 #define ADD_MASK_OSID(_osId, _osId2Mask, _maxOsId) \
4054  { \
4055  if (((_osId) > _maxOsId) || \
4056  (!KMP_CPU_ISSET((_osId), KMP_CPU_INDEX((_osId2Mask), (_osId))))) { \
4057  KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, _osId); \
4058  } else { \
4059  ADD_MASK(KMP_CPU_INDEX(_osId2Mask, (_osId))); \
4060  } \
4061  }
4062 
4063 // Re-parse the proclist (for the explicit affinity type), and form the list
4064 // of affinity newMasks indexed by gtid.
4065 static void __kmp_affinity_process_proclist(kmp_affinity_t &affinity) {
4066  int i;
4067  kmp_affin_mask_t **out_masks = &affinity.masks;
4068  unsigned *out_numMasks = &affinity.num_masks;
4069  const char *proclist = affinity.proclist;
4070  kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
4071  int maxOsId = affinity.num_os_id_masks - 1;
4072  const char *scan = proclist;
4073  const char *next = proclist;
4074 
4075  // We use malloc() for the temporary mask vector, so that we can use
4076  // realloc() to extend it.
4077  numNewMasks = 2;
4078  KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
4079  nextNewMask = 0;
4080  kmp_affin_mask_t *sumMask;
4081  KMP_CPU_ALLOC(sumMask);
4082  int setSize = 0;
4083 
4084  for (;;) {
4085  int start, end, stride;
4086 
4087  SKIP_WS(scan);
4088  next = scan;
4089  if (*next == '\0') {
4090  break;
4091  }
4092 
4093  if (*next == '{') {
4094  int num;
4095  setSize = 0;
4096  next++; // skip '{'
4097  SKIP_WS(next);
4098  scan = next;
4099 
4100  // Read the first integer in the set.
4101  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad proclist");
4102  SKIP_DIGITS(next);
4103  num = __kmp_str_to_int(scan, *next);
4104  KMP_ASSERT2(num >= 0, "bad explicit proc list");
4105 
4106  // Copy the mask for that osId to the sum (union) mask.
4107  if ((num > maxOsId) ||
4108  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
4109  KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, num);
4110  KMP_CPU_ZERO(sumMask);
4111  } else {
4112  KMP_CPU_COPY(sumMask, KMP_CPU_INDEX(osId2Mask, num));
4113  setSize = 1;
4114  }
4115 
4116  for (;;) {
4117  // Check for end of set.
4118  SKIP_WS(next);
4119  if (*next == '}') {
4120  next++; // skip '}'
4121  break;
4122  }
4123 
4124  // Skip optional comma.
4125  if (*next == ',') {
4126  next++;
4127  }
4128  SKIP_WS(next);
4129 
4130  // Read the next integer in the set.
4131  scan = next;
4132  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
4133 
4134  SKIP_DIGITS(next);
4135  num = __kmp_str_to_int(scan, *next);
4136  KMP_ASSERT2(num >= 0, "bad explicit proc list");
4137 
4138  // Add the mask for that osId to the sum mask.
4139  if ((num > maxOsId) ||
4140  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
4141  KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, num);
4142  } else {
4143  KMP_CPU_UNION(sumMask, KMP_CPU_INDEX(osId2Mask, num));
4144  setSize++;
4145  }
4146  }
4147  if (setSize > 0) {
4148  ADD_MASK(sumMask);
4149  }
4150 
4151  SKIP_WS(next);
4152  if (*next == ',') {
4153  next++;
4154  }
4155  scan = next;
4156  continue;
4157  }
4158 
4159  // Read the first integer.
4160  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
4161  SKIP_DIGITS(next);
4162  start = __kmp_str_to_int(scan, *next);
4163  KMP_ASSERT2(start >= 0, "bad explicit proc list");
4164  SKIP_WS(next);
4165 
4166  // If this isn't a range, then add a mask to the list and go on.
4167  if (*next != '-') {
4168  ADD_MASK_OSID(start, osId2Mask, maxOsId);
4169 
4170  // Skip optional comma.
4171  if (*next == ',') {
4172  next++;
4173  }
4174  scan = next;
4175  continue;
4176  }
4177 
4178  // This is a range. Skip over the '-' and read in the 2nd int.
4179  next++; // skip '-'
4180  SKIP_WS(next);
4181  scan = next;
4182  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
4183  SKIP_DIGITS(next);
4184  end = __kmp_str_to_int(scan, *next);
4185  KMP_ASSERT2(end >= 0, "bad explicit proc list");
4186 
4187  // Check for a stride parameter
4188  stride = 1;
4189  SKIP_WS(next);
4190  if (*next == ':') {
4191  // A stride is specified. Skip over the ':" and read the 3rd int.
4192  int sign = +1;
4193  next++; // skip ':'
4194  SKIP_WS(next);
4195  scan = next;
4196  if (*next == '-') {
4197  sign = -1;
4198  next++;
4199  SKIP_WS(next);
4200  scan = next;
4201  }
4202  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
4203  SKIP_DIGITS(next);
4204  stride = __kmp_str_to_int(scan, *next);
4205  KMP_ASSERT2(stride >= 0, "bad explicit proc list");
4206  stride *= sign;
4207  }
4208 
4209  // Do some range checks.
4210  KMP_ASSERT2(stride != 0, "bad explicit proc list");
4211  if (stride > 0) {
4212  KMP_ASSERT2(start <= end, "bad explicit proc list");
4213  } else {
4214  KMP_ASSERT2(start >= end, "bad explicit proc list");
4215  }
4216  KMP_ASSERT2((end - start) / stride <= 65536, "bad explicit proc list");
4217 
4218  // Add the mask for each OS proc # to the list.
4219  if (stride > 0) {
4220  do {
4221  ADD_MASK_OSID(start, osId2Mask, maxOsId);
4222  // Prevent possible overflow calculation
4223  if (end - start < stride)
4224  break;
4225  start += stride;
4226  } while (start <= end);
4227  } else {
4228  do {
4229  ADD_MASK_OSID(start, osId2Mask, maxOsId);
4230  start += stride;
4231  } while (start >= end);
4232  }
4233 
4234  // Skip optional comma.
4235  SKIP_WS(next);
4236  if (*next == ',') {
4237  next++;
4238  }
4239  scan = next;
4240  }
4241 
4242  *out_numMasks = nextNewMask;
4243  if (nextNewMask == 0) {
4244  *out_masks = NULL;
4245  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
4246  KMP_CPU_FREE(sumMask);
4247  return;
4248  }
4249  KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
4250  for (i = 0; i < nextNewMask; i++) {
4251  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
4252  kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
4253  KMP_CPU_COPY(dest, src);
4254  }
4255  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
4256  KMP_CPU_FREE(sumMask);
4257 }
4258 
4259 /*-----------------------------------------------------------------------------
4260 Re-parse the OMP_PLACES proc id list, forming the newMasks for the different
4261 places. Again, Here is the grammar:
4262 
4263 place_list := place
4264 place_list := place , place_list
4265 place := num
4266 place := place : num
4267 place := place : num : signed
4268 place := { subplacelist }
4269 place := ! place // (lowest priority)
4270 subplace_list := subplace
4271 subplace_list := subplace , subplace_list
4272 subplace := num
4273 subplace := num : num
4274 subplace := num : num : signed
4275 signed := num
4276 signed := + signed
4277 signed := - signed
4278 -----------------------------------------------------------------------------*/
4279 static void __kmp_process_subplace_list(const char **scan,
4280  kmp_affinity_t &affinity, int maxOsId,
4281  kmp_affin_mask_t *tempMask,
4282  int *setSize) {
4283  const char *next;
4284  kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
4285 
4286  for (;;) {
4287  int start, count, stride, i;
4288 
4289  // Read in the starting proc id
4290  SKIP_WS(*scan);
4291  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
4292  next = *scan;
4293  SKIP_DIGITS(next);
4294  start = __kmp_str_to_int(*scan, *next);
4295  KMP_ASSERT(start >= 0);
4296  *scan = next;
4297 
4298  // valid follow sets are ',' ':' and '}'
4299  SKIP_WS(*scan);
4300  if (**scan == '}' || **scan == ',') {
4301  if ((start > maxOsId) ||
4302  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
4303  KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, start);
4304  } else {
4305  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
4306  (*setSize)++;
4307  }
4308  if (**scan == '}') {
4309  break;
4310  }
4311  (*scan)++; // skip ','
4312  continue;
4313  }
4314  KMP_ASSERT2(**scan == ':', "bad explicit places list");
4315  (*scan)++; // skip ':'
4316 
4317  // Read count parameter
4318  SKIP_WS(*scan);
4319  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
4320  next = *scan;
4321  SKIP_DIGITS(next);
4322  count = __kmp_str_to_int(*scan, *next);
4323  KMP_ASSERT(count >= 0);
4324  *scan = next;
4325 
4326  // valid follow sets are ',' ':' and '}'
4327  SKIP_WS(*scan);
4328  if (**scan == '}' || **scan == ',') {
4329  for (i = 0; i < count; i++) {
4330  if ((start > maxOsId) ||
4331  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
4332  KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, start);
4333  break; // don't proliferate warnings for large count
4334  } else {
4335  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
4336  start++;
4337  (*setSize)++;
4338  }
4339  }
4340  if (**scan == '}') {
4341  break;
4342  }
4343  (*scan)++; // skip ','
4344  continue;
4345  }
4346  KMP_ASSERT2(**scan == ':', "bad explicit places list");
4347  (*scan)++; // skip ':'
4348 
4349  // Read stride parameter
4350  int sign = +1;
4351  for (;;) {
4352  SKIP_WS(*scan);
4353  if (**scan == '+') {
4354  (*scan)++; // skip '+'
4355  continue;
4356  }
4357  if (**scan == '-') {
4358  sign *= -1;
4359  (*scan)++; // skip '-'
4360  continue;
4361  }
4362  break;
4363  }
4364  SKIP_WS(*scan);
4365  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
4366  next = *scan;
4367  SKIP_DIGITS(next);
4368  stride = __kmp_str_to_int(*scan, *next);
4369  KMP_ASSERT(stride >= 0);
4370  *scan = next;
4371  stride *= sign;
4372 
4373  // valid follow sets are ',' and '}'
4374  SKIP_WS(*scan);
4375  if (**scan == '}' || **scan == ',') {
4376  for (i = 0; i < count; i++) {
4377  if ((start > maxOsId) ||
4378  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
4379  KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, start);
4380  break; // don't proliferate warnings for large count
4381  } else {
4382  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
4383  start += stride;
4384  (*setSize)++;
4385  }
4386  }
4387  if (**scan == '}') {
4388  break;
4389  }
4390  (*scan)++; // skip ','
4391  continue;
4392  }
4393 
4394  KMP_ASSERT2(0, "bad explicit places list");
4395  }
4396 }
4397 
4398 static void __kmp_process_place(const char **scan, kmp_affinity_t &affinity,
4399  int maxOsId, kmp_affin_mask_t *tempMask,
4400  int *setSize) {
4401  const char *next;
4402  kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
4403 
4404  // valid follow sets are '{' '!' and num
4405  SKIP_WS(*scan);
4406  if (**scan == '{') {
4407  (*scan)++; // skip '{'
4408  __kmp_process_subplace_list(scan, affinity, maxOsId, tempMask, setSize);
4409  KMP_ASSERT2(**scan == '}', "bad explicit places list");
4410  (*scan)++; // skip '}'
4411  } else if (**scan == '!') {
4412  (*scan)++; // skip '!'
4413  __kmp_process_place(scan, affinity, maxOsId, tempMask, setSize);
4414  KMP_CPU_COMPLEMENT(maxOsId, tempMask);
4415  KMP_CPU_AND(tempMask, __kmp_affin_fullMask);
4416  } else if ((**scan >= '0') && (**scan <= '9')) {
4417  next = *scan;
4418  SKIP_DIGITS(next);
4419  int num = __kmp_str_to_int(*scan, *next);
4420  KMP_ASSERT(num >= 0);
4421  if ((num > maxOsId) ||
4422  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
4423  KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, num);
4424  } else {
4425  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, num));
4426  (*setSize)++;
4427  }
4428  *scan = next; // skip num
4429  } else {
4430  KMP_ASSERT2(0, "bad explicit places list");
4431  }
4432 }
4433 
4434 // static void
4435 void __kmp_affinity_process_placelist(kmp_affinity_t &affinity) {
4436  int i, j, count, stride, sign;
4437  kmp_affin_mask_t **out_masks = &affinity.masks;
4438  unsigned *out_numMasks = &affinity.num_masks;
4439  const char *placelist = affinity.proclist;
4440  kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
4441  int maxOsId = affinity.num_os_id_masks - 1;
4442  const char *scan = placelist;
4443  const char *next = placelist;
4444 
4445  numNewMasks = 2;
4446  KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
4447  nextNewMask = 0;
4448 
4449  // tempMask is modified based on the previous or initial
4450  // place to form the current place
4451  // previousMask contains the previous place
4452  kmp_affin_mask_t *tempMask;
4453  kmp_affin_mask_t *previousMask;
4454  KMP_CPU_ALLOC(tempMask);
4455  KMP_CPU_ZERO(tempMask);
4456  KMP_CPU_ALLOC(previousMask);
4457  KMP_CPU_ZERO(previousMask);
4458  int setSize = 0;
4459 
4460  for (;;) {
4461  __kmp_process_place(&scan, affinity, maxOsId, tempMask, &setSize);
4462 
4463  // valid follow sets are ',' ':' and EOL
4464  SKIP_WS(scan);
4465  if (*scan == '\0' || *scan == ',') {
4466  if (setSize > 0) {
4467  ADD_MASK(tempMask);
4468  }
4469  KMP_CPU_ZERO(tempMask);
4470  setSize = 0;
4471  if (*scan == '\0') {
4472  break;
4473  }
4474  scan++; // skip ','
4475  continue;
4476  }
4477 
4478  KMP_ASSERT2(*scan == ':', "bad explicit places list");
4479  scan++; // skip ':'
4480 
4481  // Read count parameter
4482  SKIP_WS(scan);
4483  KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
4484  next = scan;
4485  SKIP_DIGITS(next);
4486  count = __kmp_str_to_int(scan, *next);
4487  KMP_ASSERT(count >= 0);
4488  scan = next;
4489 
4490  // valid follow sets are ',' ':' and EOL
4491  SKIP_WS(scan);
4492  if (*scan == '\0' || *scan == ',') {
4493  stride = +1;
4494  } else {
4495  KMP_ASSERT2(*scan == ':', "bad explicit places list");
4496  scan++; // skip ':'
4497 
4498  // Read stride parameter
4499  sign = +1;
4500  for (;;) {
4501  SKIP_WS(scan);
4502  if (*scan == '+') {
4503  scan++; // skip '+'
4504  continue;
4505  }
4506  if (*scan == '-') {
4507  sign *= -1;
4508  scan++; // skip '-'
4509  continue;
4510  }
4511  break;
4512  }
4513  SKIP_WS(scan);
4514  KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
4515  next = scan;
4516  SKIP_DIGITS(next);
4517  stride = __kmp_str_to_int(scan, *next);
4518  KMP_DEBUG_ASSERT(stride >= 0);
4519  scan = next;
4520  stride *= sign;
4521  }
4522 
4523  // Add places determined by initial_place : count : stride
4524  for (i = 0; i < count; i++) {
4525  if (setSize == 0) {
4526  break;
4527  }
4528  // Add the current place, then build the next place (tempMask) from that
4529  KMP_CPU_COPY(previousMask, tempMask);
4530  ADD_MASK(previousMask);
4531  KMP_CPU_ZERO(tempMask);
4532  setSize = 0;
4533  KMP_CPU_SET_ITERATE(j, previousMask) {
4534  if (!KMP_CPU_ISSET(j, previousMask)) {
4535  continue;
4536  }
4537  if ((j + stride > maxOsId) || (j + stride < 0) ||
4538  (!KMP_CPU_ISSET(j, __kmp_affin_fullMask)) ||
4539  (!KMP_CPU_ISSET(j + stride,
4540  KMP_CPU_INDEX(osId2Mask, j + stride)))) {
4541  if (i < count - 1) {
4542  KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, j + stride);
4543  }
4544  continue;
4545  }
4546  KMP_CPU_SET(j + stride, tempMask);
4547  setSize++;
4548  }
4549  }
4550  KMP_CPU_ZERO(tempMask);
4551  setSize = 0;
4552 
4553  // valid follow sets are ',' and EOL
4554  SKIP_WS(scan);
4555  if (*scan == '\0') {
4556  break;
4557  }
4558  if (*scan == ',') {
4559  scan++; // skip ','
4560  continue;
4561  }
4562 
4563  KMP_ASSERT2(0, "bad explicit places list");
4564  }
4565 
4566  *out_numMasks = nextNewMask;
4567  if (nextNewMask == 0) {
4568  *out_masks = NULL;
4569  KMP_CPU_FREE(tempMask);
4570  KMP_CPU_FREE(previousMask);
4571  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
4572  return;
4573  }
4574  KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
4575  KMP_CPU_FREE(tempMask);
4576  KMP_CPU_FREE(previousMask);
4577  for (i = 0; i < nextNewMask; i++) {
4578  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
4579  kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
4580  KMP_CPU_COPY(dest, src);
4581  }
4582  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
4583 }
4584 
4585 #undef ADD_MASK
4586 #undef ADD_MASK_OSID
4587 
4588 // This function figures out the deepest level at which there is at least one
4589 // cluster/core with more than one processing unit bound to it.
4590 static int __kmp_affinity_find_core_level(int nprocs, int bottom_level) {
4591  int core_level = 0;
4592 
4593  for (int i = 0; i < nprocs; i++) {
4594  const kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
4595  for (int j = bottom_level; j > 0; j--) {
4596  if (hw_thread.ids[j] > 0) {
4597  if (core_level < (j - 1)) {
4598  core_level = j - 1;
4599  }
4600  }
4601  }
4602  }
4603  return core_level;
4604 }
4605 
4606 // This function counts number of clusters/cores at given level.
4607 static int __kmp_affinity_compute_ncores(int nprocs, int bottom_level,
4608  int core_level) {
4609  return __kmp_topology->get_count(core_level);
4610 }
4611 // This function finds to which cluster/core given processing unit is bound.
4612 static int __kmp_affinity_find_core(int proc, int bottom_level,
4613  int core_level) {
4614  int core = 0;
4615  KMP_DEBUG_ASSERT(proc >= 0 && proc < __kmp_topology->get_num_hw_threads());
4616  for (int i = 0; i <= proc; ++i) {
4617  if (i + 1 <= proc) {
4618  for (int j = 0; j <= core_level; ++j) {
4619  if (__kmp_topology->at(i + 1).sub_ids[j] !=
4620  __kmp_topology->at(i).sub_ids[j]) {
4621  core++;
4622  break;
4623  }
4624  }
4625  }
4626  }
4627  return core;
4628 }
4629 
4630 // This function finds maximal number of processing units bound to a
4631 // cluster/core at given level.
4632 static int __kmp_affinity_max_proc_per_core(int nprocs, int bottom_level,
4633  int core_level) {
4634  if (core_level >= bottom_level)
4635  return 1;
4636  int thread_level = __kmp_topology->get_level(KMP_HW_THREAD);
4637  return __kmp_topology->calculate_ratio(thread_level, core_level);
4638 }
4639 
4640 static int *procarr = NULL;
4641 static int __kmp_aff_depth = 0;
4642 static int *__kmp_osid_to_hwthread_map = NULL;
4643 
4644 static void __kmp_affinity_get_mask_topology_info(const kmp_affin_mask_t *mask,
4645  kmp_affinity_ids_t &ids,
4646  kmp_affinity_attrs_t &attrs) {
4647  if (!KMP_AFFINITY_CAPABLE())
4648  return;
4649 
4650  // Initiailze ids and attrs thread data
4651  for (int i = 0; i < KMP_HW_LAST; ++i)
4652  ids.ids[i] = kmp_hw_thread_t::UNKNOWN_ID;
4653  attrs = KMP_AFFINITY_ATTRS_UNKNOWN;
4654 
4655  // Iterate through each os id within the mask and determine
4656  // the topology id and attribute information
4657  int cpu;
4658  int depth = __kmp_topology->get_depth();
4659  KMP_CPU_SET_ITERATE(cpu, mask) {
4660  int osid_idx = __kmp_osid_to_hwthread_map[cpu];
4661  ids.os_id = cpu;
4662  const kmp_hw_thread_t &hw_thread = __kmp_topology->at(osid_idx);
4663  for (int level = 0; level < depth; ++level) {
4664  kmp_hw_t type = __kmp_topology->get_type(level);
4665  int id = hw_thread.sub_ids[level];
4666  if (ids.ids[type] == kmp_hw_thread_t::UNKNOWN_ID || ids.ids[type] == id) {
4667  ids.ids[type] = id;
4668  } else {
4669  // This mask spans across multiple topology units, set it as such
4670  // and mark every level below as such as well.
4671  ids.ids[type] = kmp_hw_thread_t::MULTIPLE_ID;
4672  for (; level < depth; ++level) {
4673  kmp_hw_t type = __kmp_topology->get_type(level);
4674  ids.ids[type] = kmp_hw_thread_t::MULTIPLE_ID;
4675  }
4676  }
4677  }
4678  if (!attrs.valid) {
4679  attrs.core_type = hw_thread.attrs.get_core_type();
4680  attrs.core_eff = hw_thread.attrs.get_core_eff();
4681  attrs.valid = 1;
4682  } else {
4683  // This mask spans across multiple attributes, set it as such
4684  if (attrs.core_type != hw_thread.attrs.get_core_type())
4685  attrs.core_type = KMP_HW_CORE_TYPE_UNKNOWN;
4686  if (attrs.core_eff != hw_thread.attrs.get_core_eff())
4687  attrs.core_eff = kmp_hw_attr_t::UNKNOWN_CORE_EFF;
4688  }
4689  }
4690 }
4691 
4692 static void __kmp_affinity_get_thread_topology_info(kmp_info_t *th) {
4693  if (!KMP_AFFINITY_CAPABLE())
4694  return;
4695  const kmp_affin_mask_t *mask = th->th.th_affin_mask;
4696  kmp_affinity_ids_t &ids = th->th.th_topology_ids;
4697  kmp_affinity_attrs_t &attrs = th->th.th_topology_attrs;
4698  __kmp_affinity_get_mask_topology_info(mask, ids, attrs);
4699 }
4700 
4701 // Assign the topology information to each place in the place list
4702 // A thread can then grab not only its affinity mask, but the topology
4703 // information associated with that mask. e.g., Which socket is a thread on
4704 static void __kmp_affinity_get_topology_info(kmp_affinity_t &affinity) {
4705  if (!KMP_AFFINITY_CAPABLE())
4706  return;
4707  if (affinity.type != affinity_none) {
4708  KMP_ASSERT(affinity.num_os_id_masks);
4709  KMP_ASSERT(affinity.os_id_masks);
4710  }
4711  KMP_ASSERT(affinity.num_masks);
4712  KMP_ASSERT(affinity.masks);
4713  KMP_ASSERT(__kmp_affin_fullMask);
4714 
4715  int max_cpu = __kmp_affin_fullMask->get_max_cpu();
4716  int num_hw_threads = __kmp_topology->get_num_hw_threads();
4717 
4718  // Allocate thread topology information
4719  if (!affinity.ids) {
4720  affinity.ids = (kmp_affinity_ids_t *)__kmp_allocate(
4721  sizeof(kmp_affinity_ids_t) * affinity.num_masks);
4722  }
4723  if (!affinity.attrs) {
4724  affinity.attrs = (kmp_affinity_attrs_t *)__kmp_allocate(
4725  sizeof(kmp_affinity_attrs_t) * affinity.num_masks);
4726  }
4727  if (!__kmp_osid_to_hwthread_map) {
4728  // Want the +1 because max_cpu should be valid index into map
4729  __kmp_osid_to_hwthread_map =
4730  (int *)__kmp_allocate(sizeof(int) * (max_cpu + 1));
4731  }
4732 
4733  // Create the OS proc to hardware thread map
4734  for (int hw_thread = 0; hw_thread < num_hw_threads; ++hw_thread) {
4735  int os_id = __kmp_topology->at(hw_thread).os_id;
4736  if (KMP_CPU_ISSET(os_id, __kmp_affin_fullMask))
4737  __kmp_osid_to_hwthread_map[os_id] = hw_thread;
4738  }
4739 
4740  for (unsigned i = 0; i < affinity.num_masks; ++i) {
4741  kmp_affinity_ids_t &ids = affinity.ids[i];
4742  kmp_affinity_attrs_t &attrs = affinity.attrs[i];
4743  kmp_affin_mask_t *mask = KMP_CPU_INDEX(affinity.masks, i);
4744  __kmp_affinity_get_mask_topology_info(mask, ids, attrs);
4745  }
4746 }
4747 
4748 // Called when __kmp_topology is ready
4749 static void __kmp_aux_affinity_initialize_other_data(kmp_affinity_t &affinity) {
4750  // Initialize other data structures which depend on the topology
4751  if (__kmp_topology && __kmp_topology->get_num_hw_threads()) {
4752  machine_hierarchy.init(__kmp_topology->get_num_hw_threads());
4753  __kmp_affinity_get_topology_info(affinity);
4754 #if KMP_WEIGHTED_ITERATIONS_SUPPORTED
4755  __kmp_first_osid_with_ecore = __kmp_get_first_osid_with_ecore();
4756 #endif
4757  }
4758 }
4759 
4760 // Create a one element mask array (set of places) which only contains the
4761 // initial process's affinity mask
4762 static void __kmp_create_affinity_none_places(kmp_affinity_t &affinity) {
4763  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4764  KMP_ASSERT(affinity.type == affinity_none);
4765  KMP_ASSERT(__kmp_avail_proc == __kmp_topology->get_num_hw_threads());
4766  affinity.num_masks = 1;
4767  KMP_CPU_ALLOC_ARRAY(affinity.masks, affinity.num_masks);
4768  kmp_affin_mask_t *dest = KMP_CPU_INDEX(affinity.masks, 0);
4769  KMP_CPU_COPY(dest, __kmp_affin_fullMask);
4770  __kmp_aux_affinity_initialize_other_data(affinity);
4771 }
4772 
4773 static void __kmp_aux_affinity_initialize_masks(kmp_affinity_t &affinity) {
4774  // Create the "full" mask - this defines all of the processors that we
4775  // consider to be in the machine model. If respect is set, then it is the
4776  // initialization thread's affinity mask. Otherwise, it is all processors that
4777  // we know about on the machine.
4778  int verbose = affinity.flags.verbose;
4779  const char *env_var = affinity.env_var;
4780 
4781  // Already initialized
4782  if (__kmp_affin_fullMask && __kmp_affin_origMask)
4783  return;
4784 
4785  if (__kmp_affin_fullMask == NULL) {
4786  KMP_CPU_ALLOC(__kmp_affin_fullMask);
4787  }
4788  if (__kmp_affin_origMask == NULL) {
4789  KMP_CPU_ALLOC(__kmp_affin_origMask);
4790  }
4791  if (KMP_AFFINITY_CAPABLE()) {
4792  __kmp_get_system_affinity(__kmp_affin_fullMask, TRUE);
4793  // Make a copy before possible expanding to the entire machine mask
4794  __kmp_affin_origMask->copy(__kmp_affin_fullMask);
4795  if (affinity.flags.respect) {
4796  // Count the number of available processors.
4797  unsigned i;
4798  __kmp_avail_proc = 0;
4799  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
4800  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
4801  continue;
4802  }
4803  __kmp_avail_proc++;
4804  }
4805  if (__kmp_avail_proc > __kmp_xproc) {
4806  KMP_AFF_WARNING(affinity, ErrorInitializeAffinity);
4807  affinity.type = affinity_none;
4808  KMP_AFFINITY_DISABLE();
4809  return;
4810  }
4811 
4812  if (verbose) {
4813  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4814  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4815  __kmp_affin_fullMask);
4816  KMP_INFORM(InitOSProcSetRespect, env_var, buf);
4817  }
4818  } else {
4819  if (verbose) {
4820  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4821  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4822  __kmp_affin_fullMask);
4823  KMP_INFORM(InitOSProcSetNotRespect, env_var, buf);
4824  }
4825  __kmp_avail_proc =
4826  __kmp_affinity_entire_machine_mask(__kmp_affin_fullMask);
4827 #if KMP_OS_WINDOWS
4828  if (__kmp_num_proc_groups <= 1) {
4829  // Copy expanded full mask if topology has single processor group
4830  __kmp_affin_origMask->copy(__kmp_affin_fullMask);
4831  }
4832  // Set the process affinity mask since threads' affinity
4833  // masks must be subset of process mask in Windows* OS
4834  __kmp_affin_fullMask->set_process_affinity(true);
4835 #endif
4836  }
4837  }
4838 }
4839 
4840 static bool __kmp_aux_affinity_initialize_topology(kmp_affinity_t &affinity) {
4841  bool success = false;
4842  const char *env_var = affinity.env_var;
4843  kmp_i18n_id_t msg_id = kmp_i18n_null;
4844  int verbose = affinity.flags.verbose;
4845 
4846  // For backward compatibility, setting KMP_CPUINFO_FILE =>
4847  // KMP_TOPOLOGY_METHOD=cpuinfo
4848  if ((__kmp_cpuinfo_file != NULL) &&
4849  (__kmp_affinity_top_method == affinity_top_method_all)) {
4850  __kmp_affinity_top_method = affinity_top_method_cpuinfo;
4851  }
4852 
4853  if (__kmp_affinity_top_method == affinity_top_method_all) {
4854 // In the default code path, errors are not fatal - we just try using
4855 // another method. We only emit a warning message if affinity is on, or the
4856 // verbose flag is set, an the nowarnings flag was not set.
4857 #if KMP_USE_HWLOC
4858  if (!success &&
4859  __kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC) {
4860  if (!__kmp_hwloc_error) {
4861  success = __kmp_affinity_create_hwloc_map(&msg_id);
4862  if (!success && verbose) {
4863  KMP_INFORM(AffIgnoringHwloc, env_var);
4864  }
4865  } else if (verbose) {
4866  KMP_INFORM(AffIgnoringHwloc, env_var);
4867  }
4868  }
4869 #endif
4870 
4871 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4872  if (!success) {
4873  success = __kmp_affinity_create_x2apicid_map(&msg_id);
4874  if (!success && verbose && msg_id != kmp_i18n_null) {
4875  KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4876  }
4877  }
4878  if (!success) {
4879  success = __kmp_affinity_create_apicid_map(&msg_id);
4880  if (!success && verbose && msg_id != kmp_i18n_null) {
4881  KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4882  }
4883  }
4884 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4885 
4886 #if KMP_OS_LINUX || KMP_OS_AIX
4887  if (!success) {
4888  int line = 0;
4889  success = __kmp_affinity_create_cpuinfo_map(&line, &msg_id);
4890  if (!success && verbose && msg_id != kmp_i18n_null) {
4891  KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4892  }
4893  }
4894 #endif /* KMP_OS_LINUX */
4895 
4896 #if KMP_GROUP_AFFINITY
4897  if (!success && (__kmp_num_proc_groups > 1)) {
4898  success = __kmp_affinity_create_proc_group_map(&msg_id);
4899  if (!success && verbose && msg_id != kmp_i18n_null) {
4900  KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4901  }
4902  }
4903 #endif /* KMP_GROUP_AFFINITY */
4904 
4905  if (!success) {
4906  success = __kmp_affinity_create_flat_map(&msg_id);
4907  if (!success && verbose && msg_id != kmp_i18n_null) {
4908  KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4909  }
4910  KMP_ASSERT(success);
4911  }
4912  }
4913 
4914 // If the user has specified that a paricular topology discovery method is to be
4915 // used, then we abort if that method fails. The exception is group affinity,
4916 // which might have been implicitly set.
4917 #if KMP_USE_HWLOC
4918  else if (__kmp_affinity_top_method == affinity_top_method_hwloc) {
4919  KMP_ASSERT(__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC);
4920  success = __kmp_affinity_create_hwloc_map(&msg_id);
4921  if (!success) {
4922  KMP_ASSERT(msg_id != kmp_i18n_null);
4923  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4924  }
4925  }
4926 #endif // KMP_USE_HWLOC
4927 
4928 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4929  else if (__kmp_affinity_top_method == affinity_top_method_x2apicid ||
4930  __kmp_affinity_top_method == affinity_top_method_x2apicid_1f) {
4931  success = __kmp_affinity_create_x2apicid_map(&msg_id);
4932  if (!success) {
4933  KMP_ASSERT(msg_id != kmp_i18n_null);
4934  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4935  }
4936  } else if (__kmp_affinity_top_method == affinity_top_method_apicid) {
4937  success = __kmp_affinity_create_apicid_map(&msg_id);
4938  if (!success) {
4939  KMP_ASSERT(msg_id != kmp_i18n_null);
4940  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4941  }
4942  }
4943 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4944 
4945  else if (__kmp_affinity_top_method == affinity_top_method_cpuinfo) {
4946  int line = 0;
4947  success = __kmp_affinity_create_cpuinfo_map(&line, &msg_id);
4948  if (!success) {
4949  KMP_ASSERT(msg_id != kmp_i18n_null);
4950  const char *filename = __kmp_cpuinfo_get_filename();
4951  if (line > 0) {
4952  KMP_FATAL(FileLineMsgExiting, filename, line,
4953  __kmp_i18n_catgets(msg_id));
4954  } else {
4955  KMP_FATAL(FileMsgExiting, filename, __kmp_i18n_catgets(msg_id));
4956  }
4957  }
4958  }
4959 
4960 #if KMP_GROUP_AFFINITY
4961  else if (__kmp_affinity_top_method == affinity_top_method_group) {
4962  success = __kmp_affinity_create_proc_group_map(&msg_id);
4963  KMP_ASSERT(success);
4964  if (!success) {
4965  KMP_ASSERT(msg_id != kmp_i18n_null);
4966  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4967  }
4968  }
4969 #endif /* KMP_GROUP_AFFINITY */
4970 
4971  else if (__kmp_affinity_top_method == affinity_top_method_flat) {
4972  success = __kmp_affinity_create_flat_map(&msg_id);
4973  // should not fail
4974  KMP_ASSERT(success);
4975  }
4976 
4977  // Early exit if topology could not be created
4978  if (!__kmp_topology) {
4979  if (KMP_AFFINITY_CAPABLE()) {
4980  KMP_AFF_WARNING(affinity, ErrorInitializeAffinity);
4981  }
4982  if (nPackages > 0 && nCoresPerPkg > 0 && __kmp_nThreadsPerCore > 0 &&
4983  __kmp_ncores > 0) {
4984  __kmp_topology = kmp_topology_t::allocate(0, 0, NULL);
4985  __kmp_topology->canonicalize(nPackages, nCoresPerPkg,
4986  __kmp_nThreadsPerCore, __kmp_ncores);
4987  if (verbose) {
4988  __kmp_topology->print(env_var);
4989  }
4990  }
4991  return false;
4992  }
4993 
4994  // Canonicalize, print (if requested), apply KMP_HW_SUBSET
4995  __kmp_topology->canonicalize();
4996  if (verbose)
4997  __kmp_topology->print(env_var);
4998  bool filtered = __kmp_topology->filter_hw_subset();
4999  if (filtered && verbose)
5000  __kmp_topology->print("KMP_HW_SUBSET");
5001  return success;
5002 }
5003 
5004 static void __kmp_aux_affinity_initialize(kmp_affinity_t &affinity) {
5005  bool is_regular_affinity = (&affinity == &__kmp_affinity);
5006  bool is_hidden_helper_affinity = (&affinity == &__kmp_hh_affinity);
5007  const char *env_var = __kmp_get_affinity_env_var(affinity);
5008 
5009  if (affinity.flags.initialized) {
5010  KMP_ASSERT(__kmp_affin_fullMask != NULL);
5011  return;
5012  }
5013 
5014  if (is_regular_affinity && (!__kmp_affin_fullMask || !__kmp_affin_origMask))
5015  __kmp_aux_affinity_initialize_masks(affinity);
5016 
5017  if (is_regular_affinity && !__kmp_topology) {
5018  bool success = __kmp_aux_affinity_initialize_topology(affinity);
5019  if (success) {
5020  KMP_ASSERT(__kmp_avail_proc == __kmp_topology->get_num_hw_threads());
5021  } else {
5022  affinity.type = affinity_none;
5023  KMP_AFFINITY_DISABLE();
5024  }
5025  }
5026 
5027  // If KMP_AFFINITY=none, then only create the single "none" place
5028  // which is the process's initial affinity mask or the number of
5029  // hardware threads depending on respect,norespect
5030  if (affinity.type == affinity_none) {
5031  __kmp_create_affinity_none_places(affinity);
5032 #if KMP_USE_HIER_SCHED
5033  __kmp_dispatch_set_hierarchy_values();
5034 #endif
5035  affinity.flags.initialized = TRUE;
5036  return;
5037  }
5038 
5039  __kmp_topology->set_granularity(affinity);
5040  int depth = __kmp_topology->get_depth();
5041 
5042  // Create the table of masks, indexed by thread Id.
5043  unsigned numUnique = 0;
5044  int numAddrs = __kmp_topology->get_num_hw_threads();
5045  // If OMP_PLACES=cores:<attribute> specified, then attempt
5046  // to make OS Id mask table using those attributes
5047  if (affinity.core_attr_gran.valid) {
5048  __kmp_create_os_id_masks(&numUnique, affinity, [&](int idx) {
5049  KMP_ASSERT(idx >= -1);
5050  for (int i = idx + 1; i < numAddrs; ++i)
5051  if (__kmp_topology->at(i).attrs.contains(affinity.core_attr_gran))
5052  return i;
5053  return numAddrs;
5054  });
5055  if (!affinity.os_id_masks) {
5056  const char *core_attribute;
5057  if (affinity.core_attr_gran.core_eff != kmp_hw_attr_t::UNKNOWN_CORE_EFF)
5058  core_attribute = "core_efficiency";
5059  else
5060  core_attribute = "core_type";
5061  KMP_AFF_WARNING(affinity, AffIgnoringNotAvailable, env_var,
5062  core_attribute,
5063  __kmp_hw_get_catalog_string(KMP_HW_CORE, /*plural=*/true))
5064  }
5065  }
5066  // If core attributes did not work, or none were specified,
5067  // then make OS Id mask table using typical incremental way with
5068  // checking for validity of each id at granularity level specified.
5069  if (!affinity.os_id_masks) {
5070  int gran = affinity.gran_levels;
5071  int gran_level = depth - 1 - affinity.gran_levels;
5072  if (gran >= 0 && gran_level >= 0 && gran_level < depth) {
5073  __kmp_create_os_id_masks(
5074  &numUnique, affinity, [depth, numAddrs, &affinity](int idx) {
5075  KMP_ASSERT(idx >= -1);
5076  int gran = affinity.gran_levels;
5077  int gran_level = depth - 1 - affinity.gran_levels;
5078  for (int i = idx + 1; i < numAddrs; ++i)
5079  if ((gran >= depth) ||
5080  (gran < depth && __kmp_topology->at(i).ids[gran_level] !=
5081  kmp_hw_thread_t::UNKNOWN_ID))
5082  return i;
5083  return numAddrs;
5084  });
5085  }
5086  }
5087  // Final attempt to make OS Id mask table using typical incremental way.
5088  if (!affinity.os_id_masks) {
5089  __kmp_create_os_id_masks(&numUnique, affinity, [](int idx) {
5090  KMP_ASSERT(idx >= -1);
5091  return idx + 1;
5092  });
5093  }
5094 
5095  switch (affinity.type) {
5096 
5097  case affinity_explicit:
5098  KMP_DEBUG_ASSERT(affinity.proclist != NULL);
5099  if (is_hidden_helper_affinity ||
5100  __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) {
5101  __kmp_affinity_process_proclist(affinity);
5102  } else {
5103  __kmp_affinity_process_placelist(affinity);
5104  }
5105  if (affinity.num_masks == 0) {
5106  KMP_AFF_WARNING(affinity, AffNoValidProcID);
5107  affinity.type = affinity_none;
5108  __kmp_create_affinity_none_places(affinity);
5109  affinity.flags.initialized = TRUE;
5110  return;
5111  }
5112  break;
5113 
5114  // The other affinity types rely on sorting the hardware threads according to
5115  // some permutation of the machine topology tree. Set affinity.compact
5116  // and affinity.offset appropriately, then jump to a common code
5117  // fragment to do the sort and create the array of affinity masks.
5118  case affinity_logical:
5119  affinity.compact = 0;
5120  if (affinity.offset) {
5121  affinity.offset =
5122  __kmp_nThreadsPerCore * affinity.offset % __kmp_avail_proc;
5123  }
5124  goto sortTopology;
5125 
5126  case affinity_physical:
5127  if (__kmp_nThreadsPerCore > 1) {
5128  affinity.compact = 1;
5129  if (affinity.compact >= depth) {
5130  affinity.compact = 0;
5131  }
5132  } else {
5133  affinity.compact = 0;
5134  }
5135  if (affinity.offset) {
5136  affinity.offset =
5137  __kmp_nThreadsPerCore * affinity.offset % __kmp_avail_proc;
5138  }
5139  goto sortTopology;
5140 
5141  case affinity_scatter:
5142  if (affinity.compact >= depth) {
5143  affinity.compact = 0;
5144  } else {
5145  affinity.compact = depth - 1 - affinity.compact;
5146  }
5147  goto sortTopology;
5148 
5149  case affinity_compact:
5150  if (affinity.compact >= depth) {
5151  affinity.compact = depth - 1;
5152  }
5153  goto sortTopology;
5154 
5155  case affinity_balanced:
5156  if (depth <= 1 || is_hidden_helper_affinity) {
5157  KMP_AFF_WARNING(affinity, AffBalancedNotAvail, env_var);
5158  affinity.type = affinity_none;
5159  __kmp_create_affinity_none_places(affinity);
5160  affinity.flags.initialized = TRUE;
5161  return;
5162  } else if (!__kmp_topology->is_uniform()) {
5163  // Save the depth for further usage
5164  __kmp_aff_depth = depth;
5165 
5166  int core_level =
5167  __kmp_affinity_find_core_level(__kmp_avail_proc, depth - 1);
5168  int ncores = __kmp_affinity_compute_ncores(__kmp_avail_proc, depth - 1,
5169  core_level);
5170  int maxprocpercore = __kmp_affinity_max_proc_per_core(
5171  __kmp_avail_proc, depth - 1, core_level);
5172 
5173  int nproc = ncores * maxprocpercore;
5174  if ((nproc < 2) || (nproc < __kmp_avail_proc)) {
5175  KMP_AFF_WARNING(affinity, AffBalancedNotAvail, env_var);
5176  affinity.type = affinity_none;
5177  __kmp_create_affinity_none_places(affinity);
5178  affinity.flags.initialized = TRUE;
5179  return;
5180  }
5181 
5182  procarr = (int *)__kmp_allocate(sizeof(int) * nproc);
5183  for (int i = 0; i < nproc; i++) {
5184  procarr[i] = -1;
5185  }
5186 
5187  int lastcore = -1;
5188  int inlastcore = 0;
5189  for (int i = 0; i < __kmp_avail_proc; i++) {
5190  int proc = __kmp_topology->at(i).os_id;
5191  int core = __kmp_affinity_find_core(i, depth - 1, core_level);
5192 
5193  if (core == lastcore) {
5194  inlastcore++;
5195  } else {
5196  inlastcore = 0;
5197  }
5198  lastcore = core;
5199 
5200  procarr[core * maxprocpercore + inlastcore] = proc;
5201  }
5202  }
5203  if (affinity.compact >= depth) {
5204  affinity.compact = depth - 1;
5205  }
5206 
5207  sortTopology:
5208  // Allocate the gtid->affinity mask table.
5209  if (affinity.flags.dups) {
5210  affinity.num_masks = __kmp_avail_proc;
5211  } else {
5212  affinity.num_masks = numUnique;
5213  }
5214 
5215  if ((__kmp_nested_proc_bind.bind_types[0] != proc_bind_intel) &&
5216  (__kmp_affinity_num_places > 0) &&
5217  ((unsigned)__kmp_affinity_num_places < affinity.num_masks) &&
5218  !is_hidden_helper_affinity) {
5219  affinity.num_masks = __kmp_affinity_num_places;
5220  }
5221 
5222  KMP_CPU_ALLOC_ARRAY(affinity.masks, affinity.num_masks);
5223 
5224  // Sort the topology table according to the current setting of
5225  // affinity.compact, then fill out affinity.masks.
5226  __kmp_topology->sort_compact(affinity);
5227  {
5228  int i;
5229  unsigned j;
5230  int num_hw_threads = __kmp_topology->get_num_hw_threads();
5231  kmp_full_mask_modifier_t full_mask;
5232  for (i = 0, j = 0; i < num_hw_threads; i++) {
5233  if ((!affinity.flags.dups) && (!__kmp_topology->at(i).leader)) {
5234  continue;
5235  }
5236  int osId = __kmp_topology->at(i).os_id;
5237 
5238  kmp_affin_mask_t *src = KMP_CPU_INDEX(affinity.os_id_masks, osId);
5239  if (KMP_CPU_ISEMPTY(src))
5240  continue;
5241  kmp_affin_mask_t *dest = KMP_CPU_INDEX(affinity.masks, j);
5242  KMP_ASSERT(KMP_CPU_ISSET(osId, src));
5243  KMP_CPU_COPY(dest, src);
5244  full_mask.include(src);
5245  if (++j >= affinity.num_masks) {
5246  break;
5247  }
5248  }
5249  KMP_DEBUG_ASSERT(j == affinity.num_masks);
5250  // See if the places list further restricts or changes the full mask
5251  if (full_mask.restrict_to_mask() && affinity.flags.verbose) {
5252  __kmp_topology->print(env_var);
5253  }
5254  }
5255  // Sort the topology back using ids
5256  __kmp_topology->sort_ids();
5257  break;
5258 
5259  default:
5260  KMP_ASSERT2(0, "Unexpected affinity setting");
5261  }
5262  __kmp_aux_affinity_initialize_other_data(affinity);
5263  affinity.flags.initialized = TRUE;
5264 }
5265 
5266 void __kmp_affinity_initialize(kmp_affinity_t &affinity) {
5267  // Much of the code above was written assuming that if a machine was not
5268  // affinity capable, then affinity type == affinity_none.
5269  // We now explicitly represent this as affinity type == affinity_disabled.
5270  // There are too many checks for affinity type == affinity_none in this code.
5271  // Instead of trying to change them all, check if
5272  // affinity type == affinity_disabled, and if so, slam it with affinity_none,
5273  // call the real initialization routine, then restore affinity type to
5274  // affinity_disabled.
5275  int disabled = (affinity.type == affinity_disabled);
5276  if (!KMP_AFFINITY_CAPABLE())
5277  KMP_ASSERT(disabled);
5278  if (disabled)
5279  affinity.type = affinity_none;
5280  __kmp_aux_affinity_initialize(affinity);
5281  if (disabled)
5282  affinity.type = affinity_disabled;
5283 }
5284 
5285 void __kmp_affinity_uninitialize(void) {
5286  for (kmp_affinity_t *affinity : __kmp_affinities) {
5287  if (affinity->masks != NULL)
5288  KMP_CPU_FREE_ARRAY(affinity->masks, affinity->num_masks);
5289  if (affinity->os_id_masks != NULL)
5290  KMP_CPU_FREE_ARRAY(affinity->os_id_masks, affinity->num_os_id_masks);
5291  if (affinity->proclist != NULL)
5292  KMP_INTERNAL_FREE(affinity->proclist);
5293  if (affinity->ids != NULL)
5294  __kmp_free(affinity->ids);
5295  if (affinity->attrs != NULL)
5296  __kmp_free(affinity->attrs);
5297  *affinity = KMP_AFFINITY_INIT(affinity->env_var);
5298  }
5299  if (__kmp_affin_fullMask != NULL) {
5300  KMP_CPU_FREE(__kmp_affin_fullMask);
5301  __kmp_affin_fullMask = NULL;
5302  }
5303  __kmp_avail_proc = 0;
5304  if (__kmp_affin_origMask != NULL) {
5305  if (KMP_AFFINITY_CAPABLE()) {
5306 #if KMP_OS_AIX
5307  // Uninitialize by unbinding the thread.
5308  bindprocessor(BINDTHREAD, thread_self(), PROCESSOR_CLASS_ANY);
5309 #else
5310  __kmp_set_system_affinity(__kmp_affin_origMask, FALSE);
5311 #endif
5312  }
5313  KMP_CPU_FREE(__kmp_affin_origMask);
5314  __kmp_affin_origMask = NULL;
5315  }
5316  __kmp_affinity_num_places = 0;
5317  if (procarr != NULL) {
5318  __kmp_free(procarr);
5319  procarr = NULL;
5320  }
5321  if (__kmp_osid_to_hwthread_map) {
5322  __kmp_free(__kmp_osid_to_hwthread_map);
5323  __kmp_osid_to_hwthread_map = NULL;
5324  }
5325 #if KMP_USE_HWLOC
5326  if (__kmp_hwloc_topology != NULL) {
5327  hwloc_topology_destroy(__kmp_hwloc_topology);
5328  __kmp_hwloc_topology = NULL;
5329  }
5330 #endif
5331  if (__kmp_hw_subset) {
5332  kmp_hw_subset_t::deallocate(__kmp_hw_subset);
5333  __kmp_hw_subset = nullptr;
5334  }
5335  if (__kmp_topology) {
5336  kmp_topology_t::deallocate(__kmp_topology);
5337  __kmp_topology = nullptr;
5338  }
5339  KMPAffinity::destroy_api();
5340 }
5341 
5342 static void __kmp_select_mask_by_gtid(int gtid, const kmp_affinity_t *affinity,
5343  int *place, kmp_affin_mask_t **mask) {
5344  int mask_idx;
5345  bool is_hidden_helper = KMP_HIDDEN_HELPER_THREAD(gtid);
5346  if (is_hidden_helper)
5347  // The first gtid is the regular primary thread, the second gtid is the main
5348  // thread of hidden team which does not participate in task execution.
5349  mask_idx = gtid - 2;
5350  else
5351  mask_idx = __kmp_adjust_gtid_for_hidden_helpers(gtid);
5352  KMP_DEBUG_ASSERT(affinity->num_masks > 0);
5353  *place = (mask_idx + affinity->offset) % affinity->num_masks;
5354  *mask = KMP_CPU_INDEX(affinity->masks, *place);
5355 }
5356 
5357 // This function initializes the per-thread data concerning affinity including
5358 // the mask and topology information
5359 void __kmp_affinity_set_init_mask(int gtid, int isa_root) {
5360 
5361  kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
5362 
5363  // Set the thread topology information to default of unknown
5364  for (int id = 0; id < KMP_HW_LAST; ++id)
5365  th->th.th_topology_ids.ids[id] = kmp_hw_thread_t::UNKNOWN_ID;
5366  th->th.th_topology_attrs = KMP_AFFINITY_ATTRS_UNKNOWN;
5367 
5368  if (!KMP_AFFINITY_CAPABLE()) {
5369  return;
5370  }
5371 
5372  if (th->th.th_affin_mask == NULL) {
5373  KMP_CPU_ALLOC(th->th.th_affin_mask);
5374  } else {
5375  KMP_CPU_ZERO(th->th.th_affin_mask);
5376  }
5377 
5378  // Copy the thread mask to the kmp_info_t structure. If
5379  // __kmp_affinity.type == affinity_none, copy the "full" mask, i.e.
5380  // one that has all of the OS proc ids set, or if
5381  // __kmp_affinity.flags.respect is set, then the full mask is the
5382  // same as the mask of the initialization thread.
5383  kmp_affin_mask_t *mask;
5384  int i;
5385  const kmp_affinity_t *affinity;
5386  bool is_hidden_helper = KMP_HIDDEN_HELPER_THREAD(gtid);
5387 
5388  if (is_hidden_helper)
5389  affinity = &__kmp_hh_affinity;
5390  else
5391  affinity = &__kmp_affinity;
5392 
5393  if (KMP_AFFINITY_NON_PROC_BIND || is_hidden_helper) {
5394  if ((affinity->type == affinity_none) ||
5395  (affinity->type == affinity_balanced) ||
5396  KMP_HIDDEN_HELPER_MAIN_THREAD(gtid)) {
5397 #if KMP_GROUP_AFFINITY
5398  if (__kmp_num_proc_groups > 1) {
5399  return;
5400  }
5401 #endif
5402  KMP_ASSERT(__kmp_affin_fullMask != NULL);
5403  i = 0;
5404  mask = __kmp_affin_fullMask;
5405  } else {
5406  __kmp_select_mask_by_gtid(gtid, affinity, &i, &mask);
5407  }
5408  } else {
5409  if (!isa_root || __kmp_nested_proc_bind.bind_types[0] == proc_bind_false) {
5410 #if KMP_GROUP_AFFINITY
5411  if (__kmp_num_proc_groups > 1) {
5412  return;
5413  }
5414 #endif
5415  KMP_ASSERT(__kmp_affin_fullMask != NULL);
5416  i = KMP_PLACE_ALL;
5417  mask = __kmp_affin_fullMask;
5418  } else {
5419  __kmp_select_mask_by_gtid(gtid, affinity, &i, &mask);
5420  }
5421  }
5422 
5423  th->th.th_current_place = i;
5424  if (isa_root && !is_hidden_helper) {
5425  th->th.th_new_place = i;
5426  th->th.th_first_place = 0;
5427  th->th.th_last_place = affinity->num_masks - 1;
5428  } else if (KMP_AFFINITY_NON_PROC_BIND) {
5429  // When using a Non-OMP_PROC_BIND affinity method,
5430  // set all threads' place-partition-var to the entire place list
5431  th->th.th_first_place = 0;
5432  th->th.th_last_place = affinity->num_masks - 1;
5433  }
5434  // Copy topology information associated with the place
5435  if (i >= 0) {
5436  th->th.th_topology_ids = __kmp_affinity.ids[i];
5437  th->th.th_topology_attrs = __kmp_affinity.attrs[i];
5438  }
5439 
5440  if (i == KMP_PLACE_ALL) {
5441  KA_TRACE(100, ("__kmp_affinity_set_init_mask: setting T#%d to all places\n",
5442  gtid));
5443  } else {
5444  KA_TRACE(100, ("__kmp_affinity_set_init_mask: setting T#%d to place %d\n",
5445  gtid, i));
5446  }
5447 
5448  KMP_CPU_COPY(th->th.th_affin_mask, mask);
5449 }
5450 
5451 void __kmp_affinity_bind_init_mask(int gtid) {
5452  if (!KMP_AFFINITY_CAPABLE()) {
5453  return;
5454  }
5455  kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
5456  const kmp_affinity_t *affinity;
5457  const char *env_var;
5458  bool is_hidden_helper = KMP_HIDDEN_HELPER_THREAD(gtid);
5459 
5460  if (is_hidden_helper)
5461  affinity = &__kmp_hh_affinity;
5462  else
5463  affinity = &__kmp_affinity;
5464  env_var = __kmp_get_affinity_env_var(*affinity, /*for_binding=*/true);
5465  /* to avoid duplicate printing (will be correctly printed on barrier) */
5466  if (affinity->flags.verbose && (affinity->type == affinity_none ||
5467  (th->th.th_current_place != KMP_PLACE_ALL &&
5468  affinity->type != affinity_balanced)) &&
5469  !KMP_HIDDEN_HELPER_MAIN_THREAD(gtid)) {
5470  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5471  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5472  th->th.th_affin_mask);
5473  KMP_INFORM(BoundToOSProcSet, env_var, (kmp_int32)getpid(), __kmp_gettid(),
5474  gtid, buf);
5475  }
5476 
5477 #if KMP_OS_WINDOWS
5478  // On Windows* OS, the process affinity mask might have changed. If the user
5479  // didn't request affinity and this call fails, just continue silently.
5480  // See CQ171393.
5481  if (affinity->type == affinity_none) {
5482  __kmp_set_system_affinity(th->th.th_affin_mask, FALSE);
5483  } else
5484 #endif
5485 #if !KMP_OS_AIX
5486  // Do not set the full mask as the init mask on AIX.
5487  __kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
5488 #endif
5489 }
5490 
5491 void __kmp_affinity_bind_place(int gtid) {
5492  // Hidden helper threads should not be affected by OMP_PLACES/OMP_PROC_BIND
5493  if (!KMP_AFFINITY_CAPABLE() || KMP_HIDDEN_HELPER_THREAD(gtid)) {
5494  return;
5495  }
5496 
5497  kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
5498 
5499  KA_TRACE(100, ("__kmp_affinity_bind_place: binding T#%d to place %d (current "
5500  "place = %d)\n",
5501  gtid, th->th.th_new_place, th->th.th_current_place));
5502 
5503  // Check that the new place is within this thread's partition.
5504  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
5505  KMP_ASSERT(th->th.th_new_place >= 0);
5506  KMP_ASSERT((unsigned)th->th.th_new_place <= __kmp_affinity.num_masks);
5507  if (th->th.th_first_place <= th->th.th_last_place) {
5508  KMP_ASSERT((th->th.th_new_place >= th->th.th_first_place) &&
5509  (th->th.th_new_place <= th->th.th_last_place));
5510  } else {
5511  KMP_ASSERT((th->th.th_new_place <= th->th.th_first_place) ||
5512  (th->th.th_new_place >= th->th.th_last_place));
5513  }
5514 
5515  // Copy the thread mask to the kmp_info_t structure,
5516  // and set this thread's affinity.
5517  kmp_affin_mask_t *mask =
5518  KMP_CPU_INDEX(__kmp_affinity.masks, th->th.th_new_place);
5519  KMP_CPU_COPY(th->th.th_affin_mask, mask);
5520  th->th.th_current_place = th->th.th_new_place;
5521 
5522  if (__kmp_affinity.flags.verbose) {
5523  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5524  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5525  th->th.th_affin_mask);
5526  KMP_INFORM(BoundToOSProcSet, "OMP_PROC_BIND", (kmp_int32)getpid(),
5527  __kmp_gettid(), gtid, buf);
5528  }
5529  __kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
5530 }
5531 
5532 int __kmp_aux_set_affinity(void **mask) {
5533  int gtid;
5534  kmp_info_t *th;
5535  int retval;
5536 
5537  if (!KMP_AFFINITY_CAPABLE()) {
5538  return -1;
5539  }
5540 
5541  gtid = __kmp_entry_gtid();
5542  KA_TRACE(
5543  1000, (""); {
5544  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5545  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5546  (kmp_affin_mask_t *)(*mask));
5547  __kmp_debug_printf(
5548  "kmp_set_affinity: setting affinity mask for thread %d = %s\n",
5549  gtid, buf);
5550  });
5551 
5552  if (__kmp_env_consistency_check) {
5553  if ((mask == NULL) || (*mask == NULL)) {
5554  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
5555  } else {
5556  unsigned proc;
5557  int num_procs = 0;
5558 
5559  KMP_CPU_SET_ITERATE(proc, ((kmp_affin_mask_t *)(*mask))) {
5560  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5561  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
5562  }
5563  if (!KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask))) {
5564  continue;
5565  }
5566  num_procs++;
5567  }
5568  if (num_procs == 0) {
5569  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
5570  }
5571 
5572 #if KMP_GROUP_AFFINITY
5573  if (__kmp_get_proc_group((kmp_affin_mask_t *)(*mask)) < 0) {
5574  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
5575  }
5576 #endif /* KMP_GROUP_AFFINITY */
5577  }
5578  }
5579 
5580  th = __kmp_threads[gtid];
5581  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
5582  retval = __kmp_set_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
5583  if (retval == 0) {
5584  KMP_CPU_COPY(th->th.th_affin_mask, (kmp_affin_mask_t *)(*mask));
5585  }
5586 
5587  th->th.th_current_place = KMP_PLACE_UNDEFINED;
5588  th->th.th_new_place = KMP_PLACE_UNDEFINED;
5589  th->th.th_first_place = 0;
5590  th->th.th_last_place = __kmp_affinity.num_masks - 1;
5591 
5592  // Turn off 4.0 affinity for the current tread at this parallel level.
5593  th->th.th_current_task->td_icvs.proc_bind = proc_bind_false;
5594 
5595  return retval;
5596 }
5597 
5598 int __kmp_aux_get_affinity(void **mask) {
5599  int gtid;
5600  int retval;
5601 #if KMP_OS_WINDOWS || KMP_OS_AIX || KMP_DEBUG
5602  kmp_info_t *th;
5603 #endif
5604  if (!KMP_AFFINITY_CAPABLE()) {
5605  return -1;
5606  }
5607 
5608  gtid = __kmp_entry_gtid();
5609 #if KMP_OS_WINDOWS || KMP_OS_AIX || KMP_DEBUG
5610  th = __kmp_threads[gtid];
5611 #else
5612  (void)gtid; // unused variable
5613 #endif
5614  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
5615 
5616  KA_TRACE(
5617  1000, (""); {
5618  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5619  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5620  th->th.th_affin_mask);
5621  __kmp_printf(
5622  "kmp_get_affinity: stored affinity mask for thread %d = %s\n", gtid,
5623  buf);
5624  });
5625 
5626  if (__kmp_env_consistency_check) {
5627  if ((mask == NULL) || (*mask == NULL)) {
5628  KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity");
5629  }
5630  }
5631 
5632 #if !KMP_OS_WINDOWS && !KMP_OS_AIX
5633 
5634  retval = __kmp_get_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
5635  KA_TRACE(
5636  1000, (""); {
5637  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5638  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5639  (kmp_affin_mask_t *)(*mask));
5640  __kmp_printf(
5641  "kmp_get_affinity: system affinity mask for thread %d = %s\n", gtid,
5642  buf);
5643  });
5644  return retval;
5645 
5646 #else
5647  (void)retval;
5648 
5649  KMP_CPU_COPY((kmp_affin_mask_t *)(*mask), th->th.th_affin_mask);
5650  return 0;
5651 
5652 #endif /* !KMP_OS_WINDOWS && !KMP_OS_AIX */
5653 }
5654 
5655 int __kmp_aux_get_affinity_max_proc() {
5656  if (!KMP_AFFINITY_CAPABLE()) {
5657  return 0;
5658  }
5659 #if KMP_GROUP_AFFINITY
5660  if (__kmp_num_proc_groups > 1) {
5661  return (int)(__kmp_num_proc_groups * sizeof(DWORD_PTR) * CHAR_BIT);
5662  }
5663 #endif
5664  return __kmp_xproc;
5665 }
5666 
5667 int __kmp_aux_set_affinity_mask_proc(int proc, void **mask) {
5668  if (!KMP_AFFINITY_CAPABLE()) {
5669  return -1;
5670  }
5671 
5672  KA_TRACE(
5673  1000, (""); {
5674  int gtid = __kmp_entry_gtid();
5675  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5676  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5677  (kmp_affin_mask_t *)(*mask));
5678  __kmp_debug_printf("kmp_set_affinity_mask_proc: setting proc %d in "
5679  "affinity mask for thread %d = %s\n",
5680  proc, gtid, buf);
5681  });
5682 
5683  if (__kmp_env_consistency_check) {
5684  if ((mask == NULL) || (*mask == NULL)) {
5685  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity_mask_proc");
5686  }
5687  }
5688 
5689  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5690  return -1;
5691  }
5692  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5693  return -2;
5694  }
5695 
5696  KMP_CPU_SET(proc, (kmp_affin_mask_t *)(*mask));
5697  return 0;
5698 }
5699 
5700 int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask) {
5701  if (!KMP_AFFINITY_CAPABLE()) {
5702  return -1;
5703  }
5704 
5705  KA_TRACE(
5706  1000, (""); {
5707  int gtid = __kmp_entry_gtid();
5708  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5709  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5710  (kmp_affin_mask_t *)(*mask));
5711  __kmp_debug_printf("kmp_unset_affinity_mask_proc: unsetting proc %d in "
5712  "affinity mask for thread %d = %s\n",
5713  proc, gtid, buf);
5714  });
5715 
5716  if (__kmp_env_consistency_check) {
5717  if ((mask == NULL) || (*mask == NULL)) {
5718  KMP_FATAL(AffinityInvalidMask, "kmp_unset_affinity_mask_proc");
5719  }
5720  }
5721 
5722  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5723  return -1;
5724  }
5725  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5726  return -2;
5727  }
5728 
5729  KMP_CPU_CLR(proc, (kmp_affin_mask_t *)(*mask));
5730  return 0;
5731 }
5732 
5733 int __kmp_aux_get_affinity_mask_proc(int proc, void **mask) {
5734  if (!KMP_AFFINITY_CAPABLE()) {
5735  return -1;
5736  }
5737 
5738  KA_TRACE(
5739  1000, (""); {
5740  int gtid = __kmp_entry_gtid();
5741  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5742  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5743  (kmp_affin_mask_t *)(*mask));
5744  __kmp_debug_printf("kmp_get_affinity_mask_proc: getting proc %d in "
5745  "affinity mask for thread %d = %s\n",
5746  proc, gtid, buf);
5747  });
5748 
5749  if (__kmp_env_consistency_check) {
5750  if ((mask == NULL) || (*mask == NULL)) {
5751  KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity_mask_proc");
5752  }
5753  }
5754 
5755  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5756  return -1;
5757  }
5758  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5759  return 0;
5760  }
5761 
5762  return KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask));
5763 }
5764 
5765 #if KMP_WEIGHTED_ITERATIONS_SUPPORTED
5766 // Returns first os proc id with ATOM core
5767 int __kmp_get_first_osid_with_ecore(void) {
5768  int low = 0;
5769  int high = __kmp_topology->get_num_hw_threads() - 1;
5770  int mid = 0;
5771  while (high - low > 1) {
5772  mid = (high + low) / 2;
5773  if (__kmp_topology->at(mid).attrs.get_core_type() ==
5774  KMP_HW_CORE_TYPE_CORE) {
5775  low = mid + 1;
5776  } else {
5777  high = mid;
5778  }
5779  }
5780  if (__kmp_topology->at(mid).attrs.get_core_type() == KMP_HW_CORE_TYPE_ATOM) {
5781  return mid;
5782  }
5783  return -1;
5784 }
5785 #endif
5786 
5787 // Dynamic affinity settings - Affinity balanced
5788 void __kmp_balanced_affinity(kmp_info_t *th, int nthreads) {
5789  KMP_DEBUG_ASSERT(th);
5790  bool fine_gran = true;
5791  int tid = th->th.th_info.ds.ds_tid;
5792  const char *env_var = "KMP_AFFINITY";
5793 
5794  // Do not perform balanced affinity for the hidden helper threads
5795  if (KMP_HIDDEN_HELPER_THREAD(__kmp_gtid_from_thread(th)))
5796  return;
5797 
5798  switch (__kmp_affinity.gran) {
5799  case KMP_HW_THREAD:
5800  break;
5801  case KMP_HW_CORE:
5802  if (__kmp_nThreadsPerCore > 1) {
5803  fine_gran = false;
5804  }
5805  break;
5806  case KMP_HW_SOCKET:
5807  if (nCoresPerPkg > 1) {
5808  fine_gran = false;
5809  }
5810  break;
5811  default:
5812  fine_gran = false;
5813  }
5814 
5815  if (__kmp_topology->is_uniform()) {
5816  int coreID;
5817  int threadID;
5818  // Number of hyper threads per core in HT machine
5819  int __kmp_nth_per_core = __kmp_avail_proc / __kmp_ncores;
5820  // Number of cores
5821  int ncores = __kmp_ncores;
5822  if ((nPackages > 1) && (__kmp_nth_per_core <= 1)) {
5823  __kmp_nth_per_core = __kmp_avail_proc / nPackages;
5824  ncores = nPackages;
5825  }
5826  // How many threads will be bound to each core
5827  int chunk = nthreads / ncores;
5828  // How many cores will have an additional thread bound to it - "big cores"
5829  int big_cores = nthreads % ncores;
5830  // Number of threads on the big cores
5831  int big_nth = (chunk + 1) * big_cores;
5832  if (tid < big_nth) {
5833  coreID = tid / (chunk + 1);
5834  threadID = (tid % (chunk + 1)) % __kmp_nth_per_core;
5835  } else { // tid >= big_nth
5836  coreID = (tid - big_cores) / chunk;
5837  threadID = ((tid - big_cores) % chunk) % __kmp_nth_per_core;
5838  }
5839  KMP_DEBUG_ASSERT2(KMP_AFFINITY_CAPABLE(),
5840  "Illegal set affinity operation when not capable");
5841 
5842  kmp_affin_mask_t *mask = th->th.th_affin_mask;
5843  KMP_CPU_ZERO(mask);
5844 
5845  if (fine_gran) {
5846  int osID =
5847  __kmp_topology->at(coreID * __kmp_nth_per_core + threadID).os_id;
5848  KMP_CPU_SET(osID, mask);
5849  } else {
5850  for (int i = 0; i < __kmp_nth_per_core; i++) {
5851  int osID;
5852  osID = __kmp_topology->at(coreID * __kmp_nth_per_core + i).os_id;
5853  KMP_CPU_SET(osID, mask);
5854  }
5855  }
5856  if (__kmp_affinity.flags.verbose) {
5857  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5858  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
5859  KMP_INFORM(BoundToOSProcSet, env_var, (kmp_int32)getpid(), __kmp_gettid(),
5860  tid, buf);
5861  }
5862  __kmp_affinity_get_thread_topology_info(th);
5863  __kmp_set_system_affinity(mask, TRUE);
5864  } else { // Non-uniform topology
5865 
5866  kmp_affin_mask_t *mask = th->th.th_affin_mask;
5867  KMP_CPU_ZERO(mask);
5868 
5869  int core_level =
5870  __kmp_affinity_find_core_level(__kmp_avail_proc, __kmp_aff_depth - 1);
5871  int ncores = __kmp_affinity_compute_ncores(__kmp_avail_proc,
5872  __kmp_aff_depth - 1, core_level);
5873  int nth_per_core = __kmp_affinity_max_proc_per_core(
5874  __kmp_avail_proc, __kmp_aff_depth - 1, core_level);
5875 
5876  // For performance gain consider the special case nthreads ==
5877  // __kmp_avail_proc
5878  if (nthreads == __kmp_avail_proc) {
5879  if (fine_gran) {
5880  int osID = __kmp_topology->at(tid).os_id;
5881  KMP_CPU_SET(osID, mask);
5882  } else {
5883  int core =
5884  __kmp_affinity_find_core(tid, __kmp_aff_depth - 1, core_level);
5885  for (int i = 0; i < __kmp_avail_proc; i++) {
5886  int osID = __kmp_topology->at(i).os_id;
5887  if (__kmp_affinity_find_core(i, __kmp_aff_depth - 1, core_level) ==
5888  core) {
5889  KMP_CPU_SET(osID, mask);
5890  }
5891  }
5892  }
5893  } else if (nthreads <= ncores) {
5894 
5895  int core = 0;
5896  for (int i = 0; i < ncores; i++) {
5897  // Check if this core from procarr[] is in the mask
5898  int in_mask = 0;
5899  for (int j = 0; j < nth_per_core; j++) {
5900  if (procarr[i * nth_per_core + j] != -1) {
5901  in_mask = 1;
5902  break;
5903  }
5904  }
5905  if (in_mask) {
5906  if (tid == core) {
5907  for (int j = 0; j < nth_per_core; j++) {
5908  int osID = procarr[i * nth_per_core + j];
5909  if (osID != -1) {
5910  KMP_CPU_SET(osID, mask);
5911  // For fine granularity it is enough to set the first available
5912  // osID for this core
5913  if (fine_gran) {
5914  break;
5915  }
5916  }
5917  }
5918  break;
5919  } else {
5920  core++;
5921  }
5922  }
5923  }
5924  } else { // nthreads > ncores
5925  // Array to save the number of processors at each core
5926  int *nproc_at_core = (int *)KMP_ALLOCA(sizeof(int) * ncores);
5927  // Array to save the number of cores with "x" available processors;
5928  int *ncores_with_x_procs =
5929  (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
5930  // Array to save the number of cores with # procs from x to nth_per_core
5931  int *ncores_with_x_to_max_procs =
5932  (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
5933 
5934  for (int i = 0; i <= nth_per_core; i++) {
5935  ncores_with_x_procs[i] = 0;
5936  ncores_with_x_to_max_procs[i] = 0;
5937  }
5938 
5939  for (int i = 0; i < ncores; i++) {
5940  int cnt = 0;
5941  for (int j = 0; j < nth_per_core; j++) {
5942  if (procarr[i * nth_per_core + j] != -1) {
5943  cnt++;
5944  }
5945  }
5946  nproc_at_core[i] = cnt;
5947  ncores_with_x_procs[cnt]++;
5948  }
5949 
5950  for (int i = 0; i <= nth_per_core; i++) {
5951  for (int j = i; j <= nth_per_core; j++) {
5952  ncores_with_x_to_max_procs[i] += ncores_with_x_procs[j];
5953  }
5954  }
5955 
5956  // Max number of processors
5957  int nproc = nth_per_core * ncores;
5958  // An array to keep number of threads per each context
5959  int *newarr = (int *)__kmp_allocate(sizeof(int) * nproc);
5960  for (int i = 0; i < nproc; i++) {
5961  newarr[i] = 0;
5962  }
5963 
5964  int nth = nthreads;
5965  int flag = 0;
5966  while (nth > 0) {
5967  for (int j = 1; j <= nth_per_core; j++) {
5968  int cnt = ncores_with_x_to_max_procs[j];
5969  for (int i = 0; i < ncores; i++) {
5970  // Skip the core with 0 processors
5971  if (nproc_at_core[i] == 0) {
5972  continue;
5973  }
5974  for (int k = 0; k < nth_per_core; k++) {
5975  if (procarr[i * nth_per_core + k] != -1) {
5976  if (newarr[i * nth_per_core + k] == 0) {
5977  newarr[i * nth_per_core + k] = 1;
5978  cnt--;
5979  nth--;
5980  break;
5981  } else {
5982  if (flag != 0) {
5983  newarr[i * nth_per_core + k]++;
5984  cnt--;
5985  nth--;
5986  break;
5987  }
5988  }
5989  }
5990  }
5991  if (cnt == 0 || nth == 0) {
5992  break;
5993  }
5994  }
5995  if (nth == 0) {
5996  break;
5997  }
5998  }
5999  flag = 1;
6000  }
6001  int sum = 0;
6002  for (int i = 0; i < nproc; i++) {
6003  sum += newarr[i];
6004  if (sum > tid) {
6005  if (fine_gran) {
6006  int osID = procarr[i];
6007  KMP_CPU_SET(osID, mask);
6008  } else {
6009  int coreID = i / nth_per_core;
6010  for (int ii = 0; ii < nth_per_core; ii++) {
6011  int osID = procarr[coreID * nth_per_core + ii];
6012  if (osID != -1) {
6013  KMP_CPU_SET(osID, mask);
6014  }
6015  }
6016  }
6017  break;
6018  }
6019  }
6020  __kmp_free(newarr);
6021  }
6022 
6023  if (__kmp_affinity.flags.verbose) {
6024  char buf[KMP_AFFIN_MASK_PRINT_LEN];
6025  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
6026  KMP_INFORM(BoundToOSProcSet, env_var, (kmp_int32)getpid(), __kmp_gettid(),
6027  tid, buf);
6028  }
6029  __kmp_affinity_get_thread_topology_info(th);
6030  __kmp_set_system_affinity(mask, TRUE);
6031  }
6032 }
6033 
6034 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_DRAGONFLY || \
6035  KMP_OS_AIX
6036 // We don't need this entry for Windows because
6037 // there is GetProcessAffinityMask() api
6038 //
6039 // The intended usage is indicated by these steps:
6040 // 1) The user gets the current affinity mask
6041 // 2) Then sets the affinity by calling this function
6042 // 3) Error check the return value
6043 // 4) Use non-OpenMP parallelization
6044 // 5) Reset the affinity to what was stored in step 1)
6045 #ifdef __cplusplus
6046 extern "C"
6047 #endif
6048  int
6049  kmp_set_thread_affinity_mask_initial()
6050 // the function returns 0 on success,
6051 // -1 if we cannot bind thread
6052 // >0 (errno) if an error happened during binding
6053 {
6054  int gtid = __kmp_get_gtid();
6055  if (gtid < 0) {
6056  // Do not touch non-omp threads
6057  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
6058  "non-omp thread, returning\n"));
6059  return -1;
6060  }
6061  if (!KMP_AFFINITY_CAPABLE() || !__kmp_init_middle) {
6062  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
6063  "affinity not initialized, returning\n"));
6064  return -1;
6065  }
6066  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
6067  "set full mask for thread %d\n",
6068  gtid));
6069  KMP_DEBUG_ASSERT(__kmp_affin_fullMask != NULL);
6070 #if KMP_OS_AIX
6071  return bindprocessor(BINDTHREAD, thread_self(), PROCESSOR_CLASS_ANY);
6072 #else
6073  return __kmp_set_system_affinity(__kmp_affin_fullMask, FALSE);
6074 #endif
6075 }
6076 #endif
6077 
6078 #endif // KMP_AFFINITY_SUPPORTED
int try_open(const char *filename, const char *mode)
Definition: kmp.h:4750