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  return false;
1400  }
1401 
1402  // Apply the filter
1403  restrict_to_mask(filtered_mask);
1404  return true;
1405 }
1406 
1407 bool kmp_topology_t::is_close(int hwt1, int hwt2,
1408  const kmp_affinity_t &stgs) const {
1409  int hw_level = stgs.gran_levels;
1410  if (hw_level >= depth)
1411  return true;
1412  bool retval = true;
1413  const kmp_hw_thread_t &t1 = hw_threads[hwt1];
1414  const kmp_hw_thread_t &t2 = hw_threads[hwt2];
1415  if (stgs.flags.core_types_gran)
1416  return t1.attrs.get_core_type() == t2.attrs.get_core_type();
1417  if (stgs.flags.core_effs_gran)
1418  return t1.attrs.get_core_eff() == t2.attrs.get_core_eff();
1419  for (int i = 0; i < (depth - hw_level); ++i) {
1420  if (t1.ids[i] != t2.ids[i])
1421  return false;
1422  }
1423  return retval;
1424 }
1425 
1427 
1428 bool KMPAffinity::picked_api = false;
1429 
1430 void *KMPAffinity::Mask::operator new(size_t n) { return __kmp_allocate(n); }
1431 void *KMPAffinity::Mask::operator new[](size_t n) { return __kmp_allocate(n); }
1432 void KMPAffinity::Mask::operator delete(void *p) { __kmp_free(p); }
1433 void KMPAffinity::Mask::operator delete[](void *p) { __kmp_free(p); }
1434 void *KMPAffinity::operator new(size_t n) { return __kmp_allocate(n); }
1435 void KMPAffinity::operator delete(void *p) { __kmp_free(p); }
1436 
1437 void KMPAffinity::pick_api() {
1438  KMPAffinity *affinity_dispatch;
1439  if (picked_api)
1440  return;
1441 #if KMP_USE_HWLOC
1442  // Only use Hwloc if affinity isn't explicitly disabled and
1443  // user requests Hwloc topology method
1444  if (__kmp_affinity_top_method == affinity_top_method_hwloc &&
1445  __kmp_affinity.type != affinity_disabled) {
1446  affinity_dispatch = new KMPHwlocAffinity();
1447  } else
1448 #endif
1449  {
1450  affinity_dispatch = new KMPNativeAffinity();
1451  }
1452  __kmp_affinity_dispatch = affinity_dispatch;
1453  picked_api = true;
1454 }
1455 
1456 void KMPAffinity::destroy_api() {
1457  if (__kmp_affinity_dispatch != NULL) {
1458  delete __kmp_affinity_dispatch;
1459  __kmp_affinity_dispatch = NULL;
1460  picked_api = false;
1461  }
1462 }
1463 
1464 #define KMP_ADVANCE_SCAN(scan) \
1465  while (*scan != '\0') { \
1466  scan++; \
1467  }
1468 
1469 // Print the affinity mask to the character array in a pretty format.
1470 // The format is a comma separated list of non-negative integers or integer
1471 // ranges: e.g., 1,2,3-5,7,9-15
1472 // The format can also be the string "{<empty>}" if no bits are set in mask
1473 char *__kmp_affinity_print_mask(char *buf, int buf_len,
1474  kmp_affin_mask_t *mask) {
1475  int start = 0, finish = 0, previous = 0;
1476  bool first_range;
1477  KMP_ASSERT(buf);
1478  KMP_ASSERT(buf_len >= 40);
1479  KMP_ASSERT(mask);
1480  char *scan = buf;
1481  char *end = buf + buf_len - 1;
1482 
1483  // Check for empty set.
1484  if (mask->begin() == mask->end()) {
1485  KMP_SNPRINTF(scan, end - scan + 1, "{<empty>}");
1486  KMP_ADVANCE_SCAN(scan);
1487  KMP_ASSERT(scan <= end);
1488  return buf;
1489  }
1490 
1491  first_range = true;
1492  start = mask->begin();
1493  while (1) {
1494  // Find next range
1495  // [start, previous] is inclusive range of contiguous bits in mask
1496  for (finish = mask->next(start), previous = start;
1497  finish == previous + 1 && finish != mask->end();
1498  finish = mask->next(finish)) {
1499  previous = finish;
1500  }
1501 
1502  // The first range does not need a comma printed before it, but the rest
1503  // of the ranges do need a comma beforehand
1504  if (!first_range) {
1505  KMP_SNPRINTF(scan, end - scan + 1, "%s", ",");
1506  KMP_ADVANCE_SCAN(scan);
1507  } else {
1508  first_range = false;
1509  }
1510  // Range with three or more contiguous bits in the affinity mask
1511  if (previous - start > 1) {
1512  KMP_SNPRINTF(scan, end - scan + 1, "%u-%u", start, previous);
1513  } else {
1514  // Range with one or two contiguous bits in the affinity mask
1515  KMP_SNPRINTF(scan, end - scan + 1, "%u", start);
1516  KMP_ADVANCE_SCAN(scan);
1517  if (previous - start > 0) {
1518  KMP_SNPRINTF(scan, end - scan + 1, ",%u", previous);
1519  }
1520  }
1521  KMP_ADVANCE_SCAN(scan);
1522  // Start over with new start point
1523  start = finish;
1524  if (start == mask->end())
1525  break;
1526  // Check for overflow
1527  if (end - scan < 2)
1528  break;
1529  }
1530 
1531  // Check for overflow
1532  KMP_ASSERT(scan <= end);
1533  return buf;
1534 }
1535 #undef KMP_ADVANCE_SCAN
1536 
1537 // Print the affinity mask to the string buffer object in a pretty format
1538 // The format is a comma separated list of non-negative integers or integer
1539 // ranges: e.g., 1,2,3-5,7,9-15
1540 // The format can also be the string "{<empty>}" if no bits are set in mask
1541 kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
1542  kmp_affin_mask_t *mask) {
1543  int start = 0, finish = 0, previous = 0;
1544  bool first_range;
1545  KMP_ASSERT(buf);
1546  KMP_ASSERT(mask);
1547 
1548  __kmp_str_buf_clear(buf);
1549 
1550  // Check for empty set.
1551  if (mask->begin() == mask->end()) {
1552  __kmp_str_buf_print(buf, "%s", "{<empty>}");
1553  return buf;
1554  }
1555 
1556  first_range = true;
1557  start = mask->begin();
1558  while (1) {
1559  // Find next range
1560  // [start, previous] is inclusive range of contiguous bits in mask
1561  for (finish = mask->next(start), previous = start;
1562  finish == previous + 1 && finish != mask->end();
1563  finish = mask->next(finish)) {
1564  previous = finish;
1565  }
1566 
1567  // The first range does not need a comma printed before it, but the rest
1568  // of the ranges do need a comma beforehand
1569  if (!first_range) {
1570  __kmp_str_buf_print(buf, "%s", ",");
1571  } else {
1572  first_range = false;
1573  }
1574  // Range with three or more contiguous bits in the affinity mask
1575  if (previous - start > 1) {
1576  __kmp_str_buf_print(buf, "%u-%u", start, previous);
1577  } else {
1578  // Range with one or two contiguous bits in the affinity mask
1579  __kmp_str_buf_print(buf, "%u", start);
1580  if (previous - start > 0) {
1581  __kmp_str_buf_print(buf, ",%u", previous);
1582  }
1583  }
1584  // Start over with new start point
1585  start = finish;
1586  if (start == mask->end())
1587  break;
1588  }
1589  return buf;
1590 }
1591 
1592 static kmp_affin_mask_t *__kmp_parse_cpu_list(const char *path) {
1593  kmp_affin_mask_t *mask;
1594  KMP_CPU_ALLOC(mask);
1595  KMP_CPU_ZERO(mask);
1596 #if KMP_OS_LINUX
1597  int n, begin_cpu, end_cpu;
1598  kmp_safe_raii_file_t file;
1599  auto skip_ws = [](FILE *f) {
1600  int c;
1601  do {
1602  c = fgetc(f);
1603  } while (isspace(c));
1604  if (c != EOF)
1605  ungetc(c, f);
1606  };
1607  // File contains CSV of integer ranges representing the CPUs
1608  // e.g., 1,2,4-7,9,11-15
1609  int status = file.try_open(path, "r");
1610  if (status != 0)
1611  return mask;
1612  while (!feof(file)) {
1613  skip_ws(file);
1614  n = fscanf(file, "%d", &begin_cpu);
1615  if (n != 1)
1616  break;
1617  skip_ws(file);
1618  int c = fgetc(file);
1619  if (c == EOF || c == ',') {
1620  // Just single CPU
1621  end_cpu = begin_cpu;
1622  } else if (c == '-') {
1623  // Range of CPUs
1624  skip_ws(file);
1625  n = fscanf(file, "%d", &end_cpu);
1626  if (n != 1)
1627  break;
1628  skip_ws(file);
1629  c = fgetc(file); // skip ','
1630  } else {
1631  // Syntax problem
1632  break;
1633  }
1634  // Ensure a valid range of CPUs
1635  if (begin_cpu < 0 || begin_cpu >= __kmp_xproc || end_cpu < 0 ||
1636  end_cpu >= __kmp_xproc || begin_cpu > end_cpu) {
1637  continue;
1638  }
1639  // Insert [begin_cpu, end_cpu] into mask
1640  for (int cpu = begin_cpu; cpu <= end_cpu; ++cpu) {
1641  KMP_CPU_SET(cpu, mask);
1642  }
1643  }
1644 #endif
1645  return mask;
1646 }
1647 
1648 // Return (possibly empty) affinity mask representing the offline CPUs
1649 // Caller must free the mask
1650 kmp_affin_mask_t *__kmp_affinity_get_offline_cpus() {
1651  return __kmp_parse_cpu_list("/sys/devices/system/cpu/offline");
1652 }
1653 
1654 // Return the number of available procs
1655 int __kmp_affinity_entire_machine_mask(kmp_affin_mask_t *mask) {
1656  int avail_proc = 0;
1657  KMP_CPU_ZERO(mask);
1658 
1659 #if KMP_GROUP_AFFINITY
1660 
1661  if (__kmp_num_proc_groups > 1) {
1662  int group;
1663  KMP_DEBUG_ASSERT(__kmp_GetActiveProcessorCount != NULL);
1664  for (group = 0; group < __kmp_num_proc_groups; group++) {
1665  int i;
1666  int num = __kmp_GetActiveProcessorCount(group);
1667  for (i = 0; i < num; i++) {
1668  KMP_CPU_SET(i + group * (CHAR_BIT * sizeof(DWORD_PTR)), mask);
1669  avail_proc++;
1670  }
1671  }
1672  } else
1673 
1674 #endif /* KMP_GROUP_AFFINITY */
1675 
1676  {
1677  int proc;
1678  kmp_affin_mask_t *offline_cpus = __kmp_affinity_get_offline_cpus();
1679  for (proc = 0; proc < __kmp_xproc; proc++) {
1680  // Skip offline CPUs
1681  if (KMP_CPU_ISSET(proc, offline_cpus))
1682  continue;
1683  KMP_CPU_SET(proc, mask);
1684  avail_proc++;
1685  }
1686  KMP_CPU_FREE(offline_cpus);
1687  }
1688 
1689  return avail_proc;
1690 }
1691 
1692 // All of the __kmp_affinity_create_*_map() routines should allocate the
1693 // internal topology object and set the layer ids for it. Each routine
1694 // returns a boolean on whether it was successful at doing so.
1695 kmp_affin_mask_t *__kmp_affin_fullMask = NULL;
1696 // Original mask is a subset of full mask in multiple processor groups topology
1697 kmp_affin_mask_t *__kmp_affin_origMask = NULL;
1698 
1699 #if KMP_USE_HWLOC
1700 static inline bool __kmp_hwloc_is_cache_type(hwloc_obj_t obj) {
1701 #if HWLOC_API_VERSION >= 0x00020000
1702  return hwloc_obj_type_is_cache(obj->type);
1703 #else
1704  return obj->type == HWLOC_OBJ_CACHE;
1705 #endif
1706 }
1707 
1708 // Returns KMP_HW_* type derived from HWLOC_* type
1709 static inline kmp_hw_t __kmp_hwloc_type_2_topology_type(hwloc_obj_t obj) {
1710 
1711  if (__kmp_hwloc_is_cache_type(obj)) {
1712  if (obj->attr->cache.type == HWLOC_OBJ_CACHE_INSTRUCTION)
1713  return KMP_HW_UNKNOWN;
1714  switch (obj->attr->cache.depth) {
1715  case 1:
1716  return KMP_HW_L1;
1717  case 2:
1718 #if KMP_MIC_SUPPORTED
1719  if (__kmp_mic_type == mic3) {
1720  return KMP_HW_TILE;
1721  }
1722 #endif
1723  return KMP_HW_L2;
1724  case 3:
1725  return KMP_HW_L3;
1726  }
1727  return KMP_HW_UNKNOWN;
1728  }
1729 
1730  switch (obj->type) {
1731  case HWLOC_OBJ_PACKAGE:
1732  return KMP_HW_SOCKET;
1733  case HWLOC_OBJ_NUMANODE:
1734  return KMP_HW_NUMA;
1735  case HWLOC_OBJ_CORE:
1736  return KMP_HW_CORE;
1737  case HWLOC_OBJ_PU:
1738  return KMP_HW_THREAD;
1739  case HWLOC_OBJ_GROUP:
1740 #if HWLOC_API_VERSION >= 0x00020000
1741  if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_DIE)
1742  return KMP_HW_DIE;
1743  else if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_TILE)
1744  return KMP_HW_TILE;
1745  else if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_MODULE)
1746  return KMP_HW_MODULE;
1747  else if (obj->attr->group.kind == HWLOC_GROUP_KIND_WINDOWS_PROCESSOR_GROUP)
1748  return KMP_HW_PROC_GROUP;
1749 #endif
1750  return KMP_HW_UNKNOWN;
1751 #if HWLOC_API_VERSION >= 0x00020100
1752  case HWLOC_OBJ_DIE:
1753  return KMP_HW_DIE;
1754 #endif
1755  }
1756  return KMP_HW_UNKNOWN;
1757 }
1758 
1759 // Returns the number of objects of type 'type' below 'obj' within the topology
1760 // tree structure. e.g., if obj is a HWLOC_OBJ_PACKAGE object, and type is
1761 // HWLOC_OBJ_PU, then this will return the number of PU's under the SOCKET
1762 // object.
1763 static int __kmp_hwloc_get_nobjs_under_obj(hwloc_obj_t obj,
1764  hwloc_obj_type_t type) {
1765  int retval = 0;
1766  hwloc_obj_t first;
1767  for (first = hwloc_get_obj_below_by_type(__kmp_hwloc_topology, obj->type,
1768  obj->logical_index, type, 0);
1769  first != NULL && hwloc_get_ancestor_obj_by_type(__kmp_hwloc_topology,
1770  obj->type, first) == obj;
1771  first = hwloc_get_next_obj_by_type(__kmp_hwloc_topology, first->type,
1772  first)) {
1773  ++retval;
1774  }
1775  return retval;
1776 }
1777 
1778 // This gets the sub_id for a lower object under a higher object in the
1779 // topology tree
1780 static int __kmp_hwloc_get_sub_id(hwloc_topology_t t, hwloc_obj_t higher,
1781  hwloc_obj_t lower) {
1782  hwloc_obj_t obj;
1783  hwloc_obj_type_t ltype = lower->type;
1784  int lindex = lower->logical_index - 1;
1785  int sub_id = 0;
1786  // Get the previous lower object
1787  obj = hwloc_get_obj_by_type(t, ltype, lindex);
1788  while (obj && lindex >= 0 &&
1789  hwloc_bitmap_isincluded(obj->cpuset, higher->cpuset)) {
1790  if (obj->userdata) {
1791  sub_id = (int)(RCAST(kmp_intptr_t, obj->userdata));
1792  break;
1793  }
1794  sub_id++;
1795  lindex--;
1796  obj = hwloc_get_obj_by_type(t, ltype, lindex);
1797  }
1798  // store sub_id + 1 so that 0 is differed from NULL
1799  lower->userdata = RCAST(void *, sub_id + 1);
1800  return sub_id;
1801 }
1802 
1803 static bool __kmp_affinity_create_hwloc_map(kmp_i18n_id_t *const msg_id) {
1804  kmp_hw_t type;
1805  int hw_thread_index, sub_id;
1806  int depth;
1807  hwloc_obj_t pu, obj, root, prev;
1808  kmp_hw_t types[KMP_HW_LAST];
1809  hwloc_obj_type_t hwloc_types[KMP_HW_LAST];
1810 
1811  hwloc_topology_t tp = __kmp_hwloc_topology;
1812  *msg_id = kmp_i18n_null;
1813  if (__kmp_affinity.flags.verbose) {
1814  KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
1815  }
1816 
1817  if (!KMP_AFFINITY_CAPABLE()) {
1818  // Hack to try and infer the machine topology using only the data
1819  // available from hwloc on the current thread, and __kmp_xproc.
1820  KMP_ASSERT(__kmp_affinity.type == affinity_none);
1821  // hwloc only guarantees existance of PU object, so check PACKAGE and CORE
1822  hwloc_obj_t o = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PACKAGE, 0);
1823  if (o != NULL)
1824  nCoresPerPkg = __kmp_hwloc_get_nobjs_under_obj(o, HWLOC_OBJ_CORE);
1825  else
1826  nCoresPerPkg = 1; // no PACKAGE found
1827  o = hwloc_get_obj_by_type(tp, HWLOC_OBJ_CORE, 0);
1828  if (o != NULL)
1829  __kmp_nThreadsPerCore = __kmp_hwloc_get_nobjs_under_obj(o, HWLOC_OBJ_PU);
1830  else
1831  __kmp_nThreadsPerCore = 1; // no CORE found
1832  if (__kmp_nThreadsPerCore == 0)
1833  __kmp_nThreadsPerCore = 1;
1834  __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
1835  if (nCoresPerPkg == 0)
1836  nCoresPerPkg = 1; // to prevent possible division by 0
1837  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
1838  return true;
1839  }
1840 
1841 #if HWLOC_API_VERSION >= 0x00020400
1842  // Handle multiple types of cores if they exist on the system
1843  int nr_cpu_kinds = hwloc_cpukinds_get_nr(tp, 0);
1844 
1845  typedef struct kmp_hwloc_cpukinds_info_t {
1846  int efficiency;
1847  kmp_hw_core_type_t core_type;
1848  hwloc_bitmap_t mask;
1849  } kmp_hwloc_cpukinds_info_t;
1850  kmp_hwloc_cpukinds_info_t *cpukinds = nullptr;
1851 
1852  if (nr_cpu_kinds > 0) {
1853  unsigned nr_infos;
1854  struct hwloc_info_s *infos;
1855  cpukinds = (kmp_hwloc_cpukinds_info_t *)__kmp_allocate(
1856  sizeof(kmp_hwloc_cpukinds_info_t) * nr_cpu_kinds);
1857  for (unsigned idx = 0; idx < (unsigned)nr_cpu_kinds; ++idx) {
1858  cpukinds[idx].efficiency = -1;
1859  cpukinds[idx].core_type = KMP_HW_CORE_TYPE_UNKNOWN;
1860  cpukinds[idx].mask = hwloc_bitmap_alloc();
1861  if (hwloc_cpukinds_get_info(tp, idx, cpukinds[idx].mask,
1862  &cpukinds[idx].efficiency, &nr_infos, &infos,
1863  0) == 0) {
1864  for (unsigned i = 0; i < nr_infos; ++i) {
1865  if (__kmp_str_match("CoreType", 8, infos[i].name)) {
1866 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1867  if (__kmp_str_match("IntelAtom", 9, infos[i].value)) {
1868  cpukinds[idx].core_type = KMP_HW_CORE_TYPE_ATOM;
1869  break;
1870  } else if (__kmp_str_match("IntelCore", 9, infos[i].value)) {
1871  cpukinds[idx].core_type = KMP_HW_CORE_TYPE_CORE;
1872  break;
1873  }
1874 #endif
1875  }
1876  }
1877  }
1878  }
1879  }
1880 #endif
1881 
1882  root = hwloc_get_root_obj(tp);
1883 
1884  // Figure out the depth and types in the topology
1885  depth = 0;
1886  obj = hwloc_get_pu_obj_by_os_index(tp, __kmp_affin_fullMask->begin());
1887  while (obj && obj != root) {
1888 #if HWLOC_API_VERSION >= 0x00020000
1889  if (obj->memory_arity) {
1890  hwloc_obj_t memory;
1891  for (memory = obj->memory_first_child; memory;
1892  memory = hwloc_get_next_child(tp, obj, memory)) {
1893  if (memory->type == HWLOC_OBJ_NUMANODE)
1894  break;
1895  }
1896  if (memory && memory->type == HWLOC_OBJ_NUMANODE) {
1897  types[depth] = KMP_HW_NUMA;
1898  hwloc_types[depth] = memory->type;
1899  depth++;
1900  }
1901  }
1902 #endif
1903  type = __kmp_hwloc_type_2_topology_type(obj);
1904  if (type != KMP_HW_UNKNOWN) {
1905  types[depth] = type;
1906  hwloc_types[depth] = obj->type;
1907  depth++;
1908  }
1909  obj = obj->parent;
1910  }
1911  KMP_ASSERT(depth > 0);
1912 
1913  // Get the order for the types correct
1914  for (int i = 0, j = depth - 1; i < j; ++i, --j) {
1915  hwloc_obj_type_t hwloc_temp = hwloc_types[i];
1916  kmp_hw_t temp = types[i];
1917  types[i] = types[j];
1918  types[j] = temp;
1919  hwloc_types[i] = hwloc_types[j];
1920  hwloc_types[j] = hwloc_temp;
1921  }
1922 
1923  // Allocate the data structure to be returned.
1924  __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
1925 
1926  hw_thread_index = 0;
1927  pu = NULL;
1928  while ((pu = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, pu))) {
1929  int index = depth - 1;
1930  bool included = KMP_CPU_ISSET(pu->os_index, __kmp_affin_fullMask);
1931  kmp_hw_thread_t &hw_thread = __kmp_topology->at(hw_thread_index);
1932  if (included) {
1933  hw_thread.clear();
1934  hw_thread.ids[index] = pu->logical_index;
1935  hw_thread.os_id = pu->os_index;
1936  hw_thread.original_idx = hw_thread_index;
1937  // If multiple core types, then set that attribute for the hardware thread
1938 #if HWLOC_API_VERSION >= 0x00020400
1939  if (cpukinds) {
1940  int cpukind_index = -1;
1941  for (int i = 0; i < nr_cpu_kinds; ++i) {
1942  if (hwloc_bitmap_isset(cpukinds[i].mask, hw_thread.os_id)) {
1943  cpukind_index = i;
1944  break;
1945  }
1946  }
1947  if (cpukind_index >= 0) {
1948  hw_thread.attrs.set_core_type(cpukinds[cpukind_index].core_type);
1949  hw_thread.attrs.set_core_eff(cpukinds[cpukind_index].efficiency);
1950  }
1951  }
1952 #endif
1953  index--;
1954  }
1955  obj = pu;
1956  prev = obj;
1957  while (obj != root && obj != NULL) {
1958  obj = obj->parent;
1959 #if HWLOC_API_VERSION >= 0x00020000
1960  // NUMA Nodes are handled differently since they are not within the
1961  // parent/child structure anymore. They are separate children
1962  // of obj (memory_first_child points to first memory child)
1963  if (obj->memory_arity) {
1964  hwloc_obj_t memory;
1965  for (memory = obj->memory_first_child; memory;
1966  memory = hwloc_get_next_child(tp, obj, memory)) {
1967  if (memory->type == HWLOC_OBJ_NUMANODE)
1968  break;
1969  }
1970  if (memory && memory->type == HWLOC_OBJ_NUMANODE) {
1971  sub_id = __kmp_hwloc_get_sub_id(tp, memory, prev);
1972  if (included) {
1973  hw_thread.ids[index] = memory->logical_index;
1974  hw_thread.ids[index + 1] = sub_id;
1975  index--;
1976  }
1977  }
1978  prev = obj;
1979  }
1980 #endif
1981  type = __kmp_hwloc_type_2_topology_type(obj);
1982  if (type != KMP_HW_UNKNOWN) {
1983  sub_id = __kmp_hwloc_get_sub_id(tp, obj, prev);
1984  if (included) {
1985  hw_thread.ids[index] = obj->logical_index;
1986  hw_thread.ids[index + 1] = sub_id;
1987  index--;
1988  }
1989  prev = obj;
1990  }
1991  }
1992  if (included)
1993  hw_thread_index++;
1994  }
1995 
1996 #if HWLOC_API_VERSION >= 0x00020400
1997  // Free the core types information
1998  if (cpukinds) {
1999  for (int idx = 0; idx < nr_cpu_kinds; ++idx)
2000  hwloc_bitmap_free(cpukinds[idx].mask);
2001  __kmp_free(cpukinds);
2002  }
2003 #endif
2004  __kmp_topology->sort_ids();
2005  return true;
2006 }
2007 #endif // KMP_USE_HWLOC
2008 
2009 // If we don't know how to retrieve the machine's processor topology, or
2010 // encounter an error in doing so, this routine is called to form a "flat"
2011 // mapping of os thread id's <-> processor id's.
2012 static bool __kmp_affinity_create_flat_map(kmp_i18n_id_t *const msg_id) {
2013  *msg_id = kmp_i18n_null;
2014  int depth = 3;
2015  kmp_hw_t types[] = {KMP_HW_SOCKET, KMP_HW_CORE, KMP_HW_THREAD};
2016 
2017  if (__kmp_affinity.flags.verbose) {
2018  KMP_INFORM(UsingFlatOS, "KMP_AFFINITY");
2019  }
2020 
2021  // Even if __kmp_affinity.type == affinity_none, this routine might still
2022  // be called to set __kmp_ncores, as well as
2023  // __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
2024  if (!KMP_AFFINITY_CAPABLE()) {
2025  KMP_ASSERT(__kmp_affinity.type == affinity_none);
2026  __kmp_ncores = nPackages = __kmp_xproc;
2027  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
2028  return true;
2029  }
2030 
2031  // When affinity is off, this routine will still be called to set
2032  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
2033  // Make sure all these vars are set correctly, and return now if affinity is
2034  // not enabled.
2035  __kmp_ncores = nPackages = __kmp_avail_proc;
2036  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
2037 
2038  // Construct the data structure to be returned.
2039  __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
2040  int avail_ct = 0;
2041  int i;
2042  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
2043  // Skip this proc if it is not included in the machine model.
2044  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
2045  continue;
2046  }
2047  kmp_hw_thread_t &hw_thread = __kmp_topology->at(avail_ct);
2048  hw_thread.clear();
2049  hw_thread.os_id = i;
2050  hw_thread.original_idx = avail_ct;
2051  hw_thread.ids[0] = i;
2052  hw_thread.ids[1] = 0;
2053  hw_thread.ids[2] = 0;
2054  avail_ct++;
2055  }
2056  if (__kmp_affinity.flags.verbose) {
2057  KMP_INFORM(OSProcToPackage, "KMP_AFFINITY");
2058  }
2059  return true;
2060 }
2061 
2062 #if KMP_GROUP_AFFINITY
2063 // If multiple Windows* OS processor groups exist, we can create a 2-level
2064 // topology map with the groups at level 0 and the individual procs at level 1.
2065 // This facilitates letting the threads float among all procs in a group,
2066 // if granularity=group (the default when there are multiple groups).
2067 static bool __kmp_affinity_create_proc_group_map(kmp_i18n_id_t *const msg_id) {
2068  *msg_id = kmp_i18n_null;
2069  int depth = 3;
2070  kmp_hw_t types[] = {KMP_HW_PROC_GROUP, KMP_HW_CORE, KMP_HW_THREAD};
2071  const static size_t BITS_PER_GROUP = CHAR_BIT * sizeof(DWORD_PTR);
2072 
2073  if (__kmp_affinity.flags.verbose) {
2074  KMP_INFORM(AffWindowsProcGroupMap, "KMP_AFFINITY");
2075  }
2076 
2077  // If we aren't affinity capable, then use flat topology
2078  if (!KMP_AFFINITY_CAPABLE()) {
2079  KMP_ASSERT(__kmp_affinity.type == affinity_none);
2080  nPackages = __kmp_num_proc_groups;
2081  __kmp_nThreadsPerCore = 1;
2082  __kmp_ncores = __kmp_xproc;
2083  nCoresPerPkg = nPackages / __kmp_ncores;
2084  return true;
2085  }
2086 
2087  // Construct the data structure to be returned.
2088  __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
2089  int avail_ct = 0;
2090  int i;
2091  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
2092  // Skip this proc if it is not included in the machine model.
2093  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
2094  continue;
2095  }
2096  kmp_hw_thread_t &hw_thread = __kmp_topology->at(avail_ct);
2097  hw_thread.clear();
2098  hw_thread.os_id = i;
2099  hw_thread.original_idx = avail_ct;
2100  hw_thread.ids[0] = i / BITS_PER_GROUP;
2101  hw_thread.ids[1] = hw_thread.ids[2] = i % BITS_PER_GROUP;
2102  avail_ct++;
2103  }
2104  return true;
2105 }
2106 #endif /* KMP_GROUP_AFFINITY */
2107 
2108 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2109 
2110 template <kmp_uint32 LSB, kmp_uint32 MSB>
2111 static inline unsigned __kmp_extract_bits(kmp_uint32 v) {
2112  const kmp_uint32 SHIFT_LEFT = sizeof(kmp_uint32) * 8 - 1 - MSB;
2113  const kmp_uint32 SHIFT_RIGHT = LSB;
2114  kmp_uint32 retval = v;
2115  retval <<= SHIFT_LEFT;
2116  retval >>= (SHIFT_LEFT + SHIFT_RIGHT);
2117  return retval;
2118 }
2119 
2120 static int __kmp_cpuid_mask_width(int count) {
2121  int r = 0;
2122 
2123  while ((1 << r) < count)
2124  ++r;
2125  return r;
2126 }
2127 
2128 class apicThreadInfo {
2129 public:
2130  unsigned osId; // param to __kmp_affinity_bind_thread
2131  unsigned apicId; // from cpuid after binding
2132  unsigned maxCoresPerPkg; // ""
2133  unsigned maxThreadsPerPkg; // ""
2134  unsigned pkgId; // inferred from above values
2135  unsigned coreId; // ""
2136  unsigned threadId; // ""
2137 };
2138 
2139 static int __kmp_affinity_cmp_apicThreadInfo_phys_id(const void *a,
2140  const void *b) {
2141  const apicThreadInfo *aa = (const apicThreadInfo *)a;
2142  const apicThreadInfo *bb = (const apicThreadInfo *)b;
2143  if (aa->pkgId < bb->pkgId)
2144  return -1;
2145  if (aa->pkgId > bb->pkgId)
2146  return 1;
2147  if (aa->coreId < bb->coreId)
2148  return -1;
2149  if (aa->coreId > bb->coreId)
2150  return 1;
2151  if (aa->threadId < bb->threadId)
2152  return -1;
2153  if (aa->threadId > bb->threadId)
2154  return 1;
2155  return 0;
2156 }
2157 
2158 class cpuid_cache_info_t {
2159 public:
2160  struct info_t {
2161  unsigned level = 0;
2162  unsigned mask = 0;
2163  bool operator==(const info_t &rhs) const {
2164  return level == rhs.level && mask == rhs.mask;
2165  }
2166  bool operator!=(const info_t &rhs) const { return !operator==(rhs); }
2167  };
2168  cpuid_cache_info_t() : depth(0) {
2169  table[MAX_CACHE_LEVEL].level = 0;
2170  table[MAX_CACHE_LEVEL].mask = 0;
2171  }
2172  size_t get_depth() const { return depth; }
2173  info_t &operator[](size_t index) { return table[index]; }
2174  const info_t &operator[](size_t index) const { return table[index]; }
2175  bool operator==(const cpuid_cache_info_t &rhs) const {
2176  if (rhs.depth != depth)
2177  return false;
2178  for (size_t i = 0; i < depth; ++i)
2179  if (table[i] != rhs.table[i])
2180  return false;
2181  return true;
2182  }
2183  bool operator!=(const cpuid_cache_info_t &rhs) const {
2184  return !operator==(rhs);
2185  }
2186  // Get cache information assocaited with L1, L2, L3 cache, etc.
2187  // If level does not exist, then return the "NULL" level (level 0)
2188  const info_t &get_level(unsigned level) const {
2189  for (size_t i = 0; i < depth; ++i) {
2190  if (table[i].level == level)
2191  return table[i];
2192  }
2193  return table[MAX_CACHE_LEVEL];
2194  }
2195 
2196  static kmp_hw_t get_topology_type(unsigned level) {
2197  KMP_DEBUG_ASSERT(level >= 1 && level <= MAX_CACHE_LEVEL);
2198  switch (level) {
2199  case 1:
2200  return KMP_HW_L1;
2201  case 2:
2202  return KMP_HW_L2;
2203  case 3:
2204  return KMP_HW_L3;
2205  }
2206  return KMP_HW_UNKNOWN;
2207  }
2208  void get_leaf4_levels() {
2209  unsigned level = 0;
2210  while (depth < MAX_CACHE_LEVEL) {
2211  unsigned cache_type, max_threads_sharing;
2212  unsigned cache_level, cache_mask_width;
2213  kmp_cpuid buf2;
2214  __kmp_x86_cpuid(4, level, &buf2);
2215  cache_type = __kmp_extract_bits<0, 4>(buf2.eax);
2216  if (!cache_type)
2217  break;
2218  // Skip instruction caches
2219  if (cache_type == 2) {
2220  level++;
2221  continue;
2222  }
2223  max_threads_sharing = __kmp_extract_bits<14, 25>(buf2.eax) + 1;
2224  cache_mask_width = __kmp_cpuid_mask_width(max_threads_sharing);
2225  cache_level = __kmp_extract_bits<5, 7>(buf2.eax);
2226  table[depth].level = cache_level;
2227  table[depth].mask = ((-1) << cache_mask_width);
2228  depth++;
2229  level++;
2230  }
2231  }
2232  static const int MAX_CACHE_LEVEL = 3;
2233 
2234 private:
2235  size_t depth;
2236  info_t table[MAX_CACHE_LEVEL + 1];
2237 };
2238 
2239 // On IA-32 architecture and Intel(R) 64 architecture, we attempt to use
2240 // an algorithm which cycles through the available os threads, setting
2241 // the current thread's affinity mask to that thread, and then retrieves
2242 // the Apic Id for each thread context using the cpuid instruction.
2243 static bool __kmp_affinity_create_apicid_map(kmp_i18n_id_t *const msg_id) {
2244  kmp_cpuid buf;
2245  *msg_id = kmp_i18n_null;
2246 
2247  if (__kmp_affinity.flags.verbose) {
2248  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(DecodingLegacyAPIC));
2249  }
2250 
2251  // Check if cpuid leaf 4 is supported.
2252  __kmp_x86_cpuid(0, 0, &buf);
2253  if (buf.eax < 4) {
2254  *msg_id = kmp_i18n_str_NoLeaf4Support;
2255  return false;
2256  }
2257 
2258  // The algorithm used starts by setting the affinity to each available thread
2259  // and retrieving info from the cpuid instruction, so if we are not capable of
2260  // calling __kmp_get_system_affinity() and _kmp_get_system_affinity(), then we
2261  // need to do something else - use the defaults that we calculated from
2262  // issuing cpuid without binding to each proc.
2263  if (!KMP_AFFINITY_CAPABLE()) {
2264  // Hack to try and infer the machine topology using only the data
2265  // available from cpuid on the current thread, and __kmp_xproc.
2266  KMP_ASSERT(__kmp_affinity.type == affinity_none);
2267 
2268  // Get an upper bound on the number of threads per package using cpuid(1).
2269  // On some OS/chps combinations where HT is supported by the chip but is
2270  // disabled, this value will be 2 on a single core chip. Usually, it will be
2271  // 2 if HT is enabled and 1 if HT is disabled.
2272  __kmp_x86_cpuid(1, 0, &buf);
2273  int maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
2274  if (maxThreadsPerPkg == 0) {
2275  maxThreadsPerPkg = 1;
2276  }
2277 
2278  // The num cores per pkg comes from cpuid(4). 1 must be added to the encoded
2279  // value.
2280  //
2281  // The author of cpu_count.cpp treated this only an upper bound on the
2282  // number of cores, but I haven't seen any cases where it was greater than
2283  // the actual number of cores, so we will treat it as exact in this block of
2284  // code.
2285  //
2286  // First, we need to check if cpuid(4) is supported on this chip. To see if
2287  // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n or
2288  // greater.
2289  __kmp_x86_cpuid(0, 0, &buf);
2290  if (buf.eax >= 4) {
2291  __kmp_x86_cpuid(4, 0, &buf);
2292  nCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
2293  } else {
2294  nCoresPerPkg = 1;
2295  }
2296 
2297  // There is no way to reliably tell if HT is enabled without issuing the
2298  // cpuid instruction from every thread, can correlating the cpuid info, so
2299  // if the machine is not affinity capable, we assume that HT is off. We have
2300  // seen quite a few machines where maxThreadsPerPkg is 2, yet the machine
2301  // does not support HT.
2302  //
2303  // - Older OSes are usually found on machines with older chips, which do not
2304  // support HT.
2305  // - The performance penalty for mistakenly identifying a machine as HT when
2306  // it isn't (which results in blocktime being incorrectly set to 0) is
2307  // greater than the penalty when for mistakenly identifying a machine as
2308  // being 1 thread/core when it is really HT enabled (which results in
2309  // blocktime being incorrectly set to a positive value).
2310  __kmp_ncores = __kmp_xproc;
2311  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
2312  __kmp_nThreadsPerCore = 1;
2313  return true;
2314  }
2315 
2316  // From here on, we can assume that it is safe to call
2317  // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
2318  // __kmp_affinity.type = affinity_none.
2319 
2320  // Save the affinity mask for the current thread.
2321  kmp_affinity_raii_t previous_affinity;
2322 
2323  // Run through each of the available contexts, binding the current thread
2324  // to it, and obtaining the pertinent information using the cpuid instr.
2325  //
2326  // The relevant information is:
2327  // - Apic Id: Bits 24:31 of ebx after issuing cpuid(1) - each thread context
2328  // has a uniqie Apic Id, which is of the form pkg# : core# : thread#.
2329  // - Max Threads Per Pkg: Bits 16:23 of ebx after issuing cpuid(1). The value
2330  // of this field determines the width of the core# + thread# fields in the
2331  // Apic Id. It is also an upper bound on the number of threads per
2332  // package, but it has been verified that situations happen were it is not
2333  // exact. In particular, on certain OS/chip combinations where Intel(R)
2334  // Hyper-Threading Technology is supported by the chip but has been
2335  // disabled, the value of this field will be 2 (for a single core chip).
2336  // On other OS/chip combinations supporting Intel(R) Hyper-Threading
2337  // Technology, the value of this field will be 1 when Intel(R)
2338  // Hyper-Threading Technology is disabled and 2 when it is enabled.
2339  // - Max Cores Per Pkg: Bits 26:31 of eax after issuing cpuid(4). The value
2340  // of this field (+1) determines the width of the core# field in the Apic
2341  // Id. The comments in "cpucount.cpp" say that this value is an upper
2342  // bound, but the IA-32 architecture manual says that it is exactly the
2343  // number of cores per package, and I haven't seen any case where it
2344  // wasn't.
2345  //
2346  // From this information, deduce the package Id, core Id, and thread Id,
2347  // and set the corresponding fields in the apicThreadInfo struct.
2348  unsigned i;
2349  apicThreadInfo *threadInfo = (apicThreadInfo *)__kmp_allocate(
2350  __kmp_avail_proc * sizeof(apicThreadInfo));
2351  unsigned nApics = 0;
2352  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
2353  // Skip this proc if it is not included in the machine model.
2354  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
2355  continue;
2356  }
2357  KMP_DEBUG_ASSERT((int)nApics < __kmp_avail_proc);
2358 
2359  __kmp_affinity_dispatch->bind_thread(i);
2360  threadInfo[nApics].osId = i;
2361 
2362  // The apic id and max threads per pkg come from cpuid(1).
2363  __kmp_x86_cpuid(1, 0, &buf);
2364  if (((buf.edx >> 9) & 1) == 0) {
2365  __kmp_free(threadInfo);
2366  *msg_id = kmp_i18n_str_ApicNotPresent;
2367  return false;
2368  }
2369  threadInfo[nApics].apicId = (buf.ebx >> 24) & 0xff;
2370  threadInfo[nApics].maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
2371  if (threadInfo[nApics].maxThreadsPerPkg == 0) {
2372  threadInfo[nApics].maxThreadsPerPkg = 1;
2373  }
2374 
2375  // Max cores per pkg comes from cpuid(4). 1 must be added to the encoded
2376  // value.
2377  //
2378  // First, we need to check if cpuid(4) is supported on this chip. To see if
2379  // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n
2380  // or greater.
2381  __kmp_x86_cpuid(0, 0, &buf);
2382  if (buf.eax >= 4) {
2383  __kmp_x86_cpuid(4, 0, &buf);
2384  threadInfo[nApics].maxCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
2385  } else {
2386  threadInfo[nApics].maxCoresPerPkg = 1;
2387  }
2388 
2389  // Infer the pkgId / coreId / threadId using only the info obtained locally.
2390  int widthCT = __kmp_cpuid_mask_width(threadInfo[nApics].maxThreadsPerPkg);
2391  threadInfo[nApics].pkgId = threadInfo[nApics].apicId >> widthCT;
2392 
2393  int widthC = __kmp_cpuid_mask_width(threadInfo[nApics].maxCoresPerPkg);
2394  int widthT = widthCT - widthC;
2395  if (widthT < 0) {
2396  // I've never seen this one happen, but I suppose it could, if the cpuid
2397  // instruction on a chip was really screwed up. Make sure to restore the
2398  // affinity mask before the tail call.
2399  __kmp_free(threadInfo);
2400  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
2401  return false;
2402  }
2403 
2404  int maskC = (1 << widthC) - 1;
2405  threadInfo[nApics].coreId = (threadInfo[nApics].apicId >> widthT) & maskC;
2406 
2407  int maskT = (1 << widthT) - 1;
2408  threadInfo[nApics].threadId = threadInfo[nApics].apicId & maskT;
2409 
2410  nApics++;
2411  }
2412 
2413  // We've collected all the info we need.
2414  // Restore the old affinity mask for this thread.
2415  previous_affinity.restore();
2416 
2417  // Sort the threadInfo table by physical Id.
2418  qsort(threadInfo, nApics, sizeof(*threadInfo),
2419  __kmp_affinity_cmp_apicThreadInfo_phys_id);
2420 
2421  // The table is now sorted by pkgId / coreId / threadId, but we really don't
2422  // know the radix of any of the fields. pkgId's may be sparsely assigned among
2423  // the chips on a system. Although coreId's are usually assigned
2424  // [0 .. coresPerPkg-1] and threadId's are usually assigned
2425  // [0..threadsPerCore-1], we don't want to make any such assumptions.
2426  //
2427  // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
2428  // total # packages) are at this point - we want to determine that now. We
2429  // only have an upper bound on the first two figures.
2430  //
2431  // We also perform a consistency check at this point: the values returned by
2432  // the cpuid instruction for any thread bound to a given package had better
2433  // return the same info for maxThreadsPerPkg and maxCoresPerPkg.
2434  nPackages = 1;
2435  nCoresPerPkg = 1;
2436  __kmp_nThreadsPerCore = 1;
2437  unsigned nCores = 1;
2438 
2439  unsigned pkgCt = 1; // to determine radii
2440  unsigned lastPkgId = threadInfo[0].pkgId;
2441  unsigned coreCt = 1;
2442  unsigned lastCoreId = threadInfo[0].coreId;
2443  unsigned threadCt = 1;
2444  unsigned lastThreadId = threadInfo[0].threadId;
2445 
2446  // intra-pkg consist checks
2447  unsigned prevMaxCoresPerPkg = threadInfo[0].maxCoresPerPkg;
2448  unsigned prevMaxThreadsPerPkg = threadInfo[0].maxThreadsPerPkg;
2449 
2450  for (i = 1; i < nApics; i++) {
2451  if (threadInfo[i].pkgId != lastPkgId) {
2452  nCores++;
2453  pkgCt++;
2454  lastPkgId = threadInfo[i].pkgId;
2455  if ((int)coreCt > nCoresPerPkg)
2456  nCoresPerPkg = coreCt;
2457  coreCt = 1;
2458  lastCoreId = threadInfo[i].coreId;
2459  if ((int)threadCt > __kmp_nThreadsPerCore)
2460  __kmp_nThreadsPerCore = threadCt;
2461  threadCt = 1;
2462  lastThreadId = threadInfo[i].threadId;
2463 
2464  // This is a different package, so go on to the next iteration without
2465  // doing any consistency checks. Reset the consistency check vars, though.
2466  prevMaxCoresPerPkg = threadInfo[i].maxCoresPerPkg;
2467  prevMaxThreadsPerPkg = threadInfo[i].maxThreadsPerPkg;
2468  continue;
2469  }
2470 
2471  if (threadInfo[i].coreId != lastCoreId) {
2472  nCores++;
2473  coreCt++;
2474  lastCoreId = threadInfo[i].coreId;
2475  if ((int)threadCt > __kmp_nThreadsPerCore)
2476  __kmp_nThreadsPerCore = threadCt;
2477  threadCt = 1;
2478  lastThreadId = threadInfo[i].threadId;
2479  } else if (threadInfo[i].threadId != lastThreadId) {
2480  threadCt++;
2481  lastThreadId = threadInfo[i].threadId;
2482  } else {
2483  __kmp_free(threadInfo);
2484  *msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
2485  return false;
2486  }
2487 
2488  // Check to make certain that the maxCoresPerPkg and maxThreadsPerPkg
2489  // fields agree between all the threads bounds to a given package.
2490  if ((prevMaxCoresPerPkg != threadInfo[i].maxCoresPerPkg) ||
2491  (prevMaxThreadsPerPkg != threadInfo[i].maxThreadsPerPkg)) {
2492  __kmp_free(threadInfo);
2493  *msg_id = kmp_i18n_str_InconsistentCpuidInfo;
2494  return false;
2495  }
2496  }
2497  // When affinity is off, this routine will still be called to set
2498  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
2499  // Make sure all these vars are set correctly
2500  nPackages = pkgCt;
2501  if ((int)coreCt > nCoresPerPkg)
2502  nCoresPerPkg = coreCt;
2503  if ((int)threadCt > __kmp_nThreadsPerCore)
2504  __kmp_nThreadsPerCore = threadCt;
2505  __kmp_ncores = nCores;
2506  KMP_DEBUG_ASSERT(nApics == (unsigned)__kmp_avail_proc);
2507 
2508  // Now that we've determined the number of packages, the number of cores per
2509  // package, and the number of threads per core, we can construct the data
2510  // structure that is to be returned.
2511  int idx = 0;
2512  int pkgLevel = 0;
2513  int coreLevel = 1;
2514  int threadLevel = 2;
2515  //(__kmp_nThreadsPerCore <= 1) ? -1 : ((coreLevel >= 0) ? 2 : 1);
2516  int depth = (pkgLevel >= 0) + (coreLevel >= 0) + (threadLevel >= 0);
2517  kmp_hw_t types[3];
2518  if (pkgLevel >= 0)
2519  types[idx++] = KMP_HW_SOCKET;
2520  if (coreLevel >= 0)
2521  types[idx++] = KMP_HW_CORE;
2522  if (threadLevel >= 0)
2523  types[idx++] = KMP_HW_THREAD;
2524 
2525  KMP_ASSERT(depth > 0);
2526  __kmp_topology = kmp_topology_t::allocate(nApics, depth, types);
2527 
2528  for (i = 0; i < nApics; ++i) {
2529  idx = 0;
2530  unsigned os = threadInfo[i].osId;
2531  kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
2532  hw_thread.clear();
2533 
2534  if (pkgLevel >= 0) {
2535  hw_thread.ids[idx++] = threadInfo[i].pkgId;
2536  }
2537  if (coreLevel >= 0) {
2538  hw_thread.ids[idx++] = threadInfo[i].coreId;
2539  }
2540  if (threadLevel >= 0) {
2541  hw_thread.ids[idx++] = threadInfo[i].threadId;
2542  }
2543  hw_thread.os_id = os;
2544  hw_thread.original_idx = i;
2545  }
2546 
2547  __kmp_free(threadInfo);
2548  __kmp_topology->sort_ids();
2549  if (!__kmp_topology->check_ids()) {
2550  kmp_topology_t::deallocate(__kmp_topology);
2551  __kmp_topology = nullptr;
2552  *msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
2553  return false;
2554  }
2555  return true;
2556 }
2557 
2558 // Hybrid cpu detection using CPUID.1A
2559 // Thread should be pinned to processor already
2560 static void __kmp_get_hybrid_info(kmp_hw_core_type_t *type, int *efficiency,
2561  unsigned *native_model_id) {
2562  kmp_cpuid buf;
2563  __kmp_x86_cpuid(0x1a, 0, &buf);
2564  *type = (kmp_hw_core_type_t)__kmp_extract_bits<24, 31>(buf.eax);
2565  switch (*type) {
2566  case KMP_HW_CORE_TYPE_ATOM:
2567  *efficiency = 0;
2568  break;
2569  case KMP_HW_CORE_TYPE_CORE:
2570  *efficiency = 1;
2571  break;
2572  default:
2573  *efficiency = 0;
2574  }
2575  *native_model_id = __kmp_extract_bits<0, 23>(buf.eax);
2576 }
2577 
2578 // Intel(R) microarchitecture code name Nehalem, Dunnington and later
2579 // architectures support a newer interface for specifying the x2APIC Ids,
2580 // based on CPUID.B or CPUID.1F
2581 /*
2582  * CPUID.B or 1F, Input ECX (sub leaf # aka level number)
2583  Bits Bits Bits Bits
2584  31-16 15-8 7-4 4-0
2585 ---+-----------+--------------+-------------+-----------------+
2586 EAX| reserved | reserved | reserved | Bits to Shift |
2587 ---+-----------|--------------+-------------+-----------------|
2588 EBX| reserved | Num logical processors at level (16 bits) |
2589 ---+-----------|--------------+-------------------------------|
2590 ECX| reserved | Level Type | Level Number (8 bits) |
2591 ---+-----------+--------------+-------------------------------|
2592 EDX| X2APIC ID (32 bits) |
2593 ---+----------------------------------------------------------+
2594 */
2595 
2596 enum {
2597  INTEL_LEVEL_TYPE_INVALID = 0, // Package level
2598  INTEL_LEVEL_TYPE_SMT = 1,
2599  INTEL_LEVEL_TYPE_CORE = 2,
2600  INTEL_LEVEL_TYPE_MODULE = 3,
2601  INTEL_LEVEL_TYPE_TILE = 4,
2602  INTEL_LEVEL_TYPE_DIE = 5,
2603  INTEL_LEVEL_TYPE_LAST = 6,
2604 };
2605 KMP_BUILD_ASSERT(INTEL_LEVEL_TYPE_LAST < sizeof(unsigned) * CHAR_BIT);
2606 #define KMP_LEAF_1F_KNOWN_LEVELS ((1u << INTEL_LEVEL_TYPE_LAST) - 1u)
2607 
2608 static kmp_hw_t __kmp_intel_type_2_topology_type(int intel_type) {
2609  switch (intel_type) {
2610  case INTEL_LEVEL_TYPE_INVALID:
2611  return KMP_HW_SOCKET;
2612  case INTEL_LEVEL_TYPE_SMT:
2613  return KMP_HW_THREAD;
2614  case INTEL_LEVEL_TYPE_CORE:
2615  return KMP_HW_CORE;
2616  case INTEL_LEVEL_TYPE_TILE:
2617  return KMP_HW_TILE;
2618  case INTEL_LEVEL_TYPE_MODULE:
2619  return KMP_HW_MODULE;
2620  case INTEL_LEVEL_TYPE_DIE:
2621  return KMP_HW_DIE;
2622  }
2623  return KMP_HW_UNKNOWN;
2624 }
2625 
2626 static int __kmp_topology_type_2_intel_type(kmp_hw_t type) {
2627  switch (type) {
2628  case KMP_HW_SOCKET:
2629  return INTEL_LEVEL_TYPE_INVALID;
2630  case KMP_HW_THREAD:
2631  return INTEL_LEVEL_TYPE_SMT;
2632  case KMP_HW_CORE:
2633  return INTEL_LEVEL_TYPE_CORE;
2634  case KMP_HW_TILE:
2635  return INTEL_LEVEL_TYPE_TILE;
2636  case KMP_HW_MODULE:
2637  return INTEL_LEVEL_TYPE_MODULE;
2638  case KMP_HW_DIE:
2639  return INTEL_LEVEL_TYPE_DIE;
2640  default:
2641  return INTEL_LEVEL_TYPE_INVALID;
2642  }
2643 }
2644 
2645 struct cpuid_level_info_t {
2646  unsigned level_type, mask, mask_width, nitems, cache_mask;
2647 };
2648 
2649 class cpuid_topo_desc_t {
2650  unsigned desc = 0;
2651 
2652 public:
2653  void clear() { desc = 0; }
2654  bool contains(int intel_type) const {
2655  KMP_DEBUG_ASSERT(intel_type >= 0 && intel_type < INTEL_LEVEL_TYPE_LAST);
2656  if ((1u << intel_type) & desc)
2657  return true;
2658  return false;
2659  }
2660  bool contains_topology_type(kmp_hw_t type) const {
2661  KMP_DEBUG_ASSERT(type >= 0 && type < KMP_HW_LAST);
2662  int intel_type = __kmp_topology_type_2_intel_type(type);
2663  return contains(intel_type);
2664  }
2665  bool contains(cpuid_topo_desc_t rhs) const {
2666  return ((desc | rhs.desc) == desc);
2667  }
2668  void add(int intel_type) { desc |= (1u << intel_type); }
2669  void add(cpuid_topo_desc_t rhs) { desc |= rhs.desc; }
2670 };
2671 
2672 struct cpuid_proc_info_t {
2673  // Topology info
2674  int os_id;
2675  unsigned apic_id;
2676  unsigned depth;
2677  // Hybrid info
2678  unsigned native_model_id;
2679  int efficiency;
2680  kmp_hw_core_type_t type;
2681  cpuid_topo_desc_t description;
2682 
2683  cpuid_level_info_t levels[INTEL_LEVEL_TYPE_LAST];
2684 };
2685 
2686 // This function takes the topology leaf, an info pointer to store the levels
2687 // detected, and writable descriptors for the total topology.
2688 // Returns whether total types, depth, or description were modified.
2689 static bool __kmp_x2apicid_get_levels(int leaf, cpuid_proc_info_t *info,
2690  kmp_hw_t total_types[KMP_HW_LAST],
2691  int *total_depth,
2692  cpuid_topo_desc_t *total_description) {
2693  unsigned level, levels_index;
2694  unsigned level_type, mask_width, nitems;
2695  kmp_cpuid buf;
2696  cpuid_level_info_t(&levels)[INTEL_LEVEL_TYPE_LAST] = info->levels;
2697  bool retval = false;
2698 
2699  // New algorithm has known topology layers act as highest unknown topology
2700  // layers when unknown topology layers exist.
2701  // e.g., Suppose layers were SMT <X> CORE <Y> <Z> PACKAGE, where <X> <Y> <Z>
2702  // are unknown topology layers, Then SMT will take the characteristics of
2703  // (SMT x <X>) and CORE will take the characteristics of (CORE x <Y> x <Z>).
2704  // This eliminates unknown portions of the topology while still keeping the
2705  // correct structure.
2706  level = levels_index = 0;
2707  do {
2708  __kmp_x86_cpuid(leaf, level, &buf);
2709  level_type = __kmp_extract_bits<8, 15>(buf.ecx);
2710  mask_width = __kmp_extract_bits<0, 4>(buf.eax);
2711  nitems = __kmp_extract_bits<0, 15>(buf.ebx);
2712  if (level_type != INTEL_LEVEL_TYPE_INVALID && nitems == 0) {
2713  info->depth = 0;
2714  return retval;
2715  }
2716 
2717  if (KMP_LEAF_1F_KNOWN_LEVELS & (1u << level_type)) {
2718  // Add a new level to the topology
2719  KMP_ASSERT(levels_index < INTEL_LEVEL_TYPE_LAST);
2720  levels[levels_index].level_type = level_type;
2721  levels[levels_index].mask_width = mask_width;
2722  levels[levels_index].nitems = nitems;
2723  levels_index++;
2724  } else {
2725  // If it is an unknown level, then logically move the previous layer up
2726  if (levels_index > 0) {
2727  levels[levels_index - 1].mask_width = mask_width;
2728  levels[levels_index - 1].nitems = nitems;
2729  }
2730  }
2731  level++;
2732  } while (level_type != INTEL_LEVEL_TYPE_INVALID);
2733  KMP_ASSERT(levels_index <= INTEL_LEVEL_TYPE_LAST);
2734  info->description.clear();
2735  info->depth = levels_index;
2736 
2737  // If types, depth, and total_description are uninitialized,
2738  // then initialize them now
2739  if (*total_depth == 0) {
2740  *total_depth = info->depth;
2741  total_description->clear();
2742  for (int i = *total_depth - 1, j = 0; i >= 0; --i, ++j) {
2743  total_types[j] =
2744  __kmp_intel_type_2_topology_type(info->levels[i].level_type);
2745  total_description->add(info->levels[i].level_type);
2746  }
2747  retval = true;
2748  }
2749 
2750  // Ensure the INTEL_LEVEL_TYPE_INVALID (Socket) layer isn't first
2751  if (levels_index == 0 || levels[0].level_type == INTEL_LEVEL_TYPE_INVALID)
2752  return 0;
2753 
2754  // Set the masks to & with apicid
2755  for (unsigned i = 0; i < levels_index; ++i) {
2756  if (levels[i].level_type != INTEL_LEVEL_TYPE_INVALID) {
2757  levels[i].mask = ~((-1) << levels[i].mask_width);
2758  levels[i].cache_mask = (-1) << levels[i].mask_width;
2759  for (unsigned j = 0; j < i; ++j)
2760  levels[i].mask ^= levels[j].mask;
2761  } else {
2762  KMP_DEBUG_ASSERT(i > 0);
2763  levels[i].mask = (-1) << levels[i - 1].mask_width;
2764  levels[i].cache_mask = 0;
2765  }
2766  info->description.add(info->levels[i].level_type);
2767  }
2768 
2769  // If this processor has level type not on other processors, then make
2770  // sure to include it in total types, depth, and description.
2771  // One assumption here is that the first type, i.e. socket, is known.
2772  // Another assumption is that types array is always large enough to fit any
2773  // new layers since its length is KMP_HW_LAST.
2774  if (!total_description->contains(info->description)) {
2775  for (int i = info->depth - 1, j = 0; i >= 0; --i, ++j) {
2776  // If this level is known already, then skip it.
2777  if (total_description->contains(levels[i].level_type))
2778  continue;
2779  // Unknown level, insert before last known level
2780  kmp_hw_t curr_type =
2781  __kmp_intel_type_2_topology_type(levels[i].level_type);
2782  KMP_ASSERT(j != 0 && "Bad APIC Id information");
2783  // Move over all known levels to make room for new level
2784  for (int k = info->depth - 1; k >= j; --k) {
2785  KMP_DEBUG_ASSERT(k + 1 < KMP_HW_LAST);
2786  total_types[k + 1] = total_types[k];
2787  }
2788  // Insert new level
2789  total_types[j] = curr_type;
2790  (*total_depth)++;
2791  }
2792  total_description->add(info->description);
2793  retval = true;
2794  }
2795  return retval;
2796 }
2797 
2798 static bool __kmp_affinity_create_x2apicid_map(kmp_i18n_id_t *const msg_id) {
2799 
2800  kmp_hw_t types[INTEL_LEVEL_TYPE_LAST];
2801  kmp_cpuid buf;
2802  int topology_leaf, highest_leaf;
2803  int num_leaves;
2804  int depth = 0;
2805  cpuid_topo_desc_t total_description;
2806  static int leaves[] = {0, 0};
2807 
2808  // If affinity is disabled, __kmp_avail_proc may be zero
2809  int ninfos = (__kmp_avail_proc > 0 ? __kmp_avail_proc : 1);
2810  cpuid_proc_info_t *proc_info = (cpuid_proc_info_t *)__kmp_allocate(
2811  (sizeof(cpuid_proc_info_t) + sizeof(cpuid_cache_info_t)) * ninfos);
2812  cpuid_cache_info_t *cache_info = (cpuid_cache_info_t *)(proc_info + ninfos);
2813 
2814  kmp_i18n_id_t leaf_message_id;
2815 
2816  *msg_id = kmp_i18n_null;
2817  if (__kmp_affinity.flags.verbose) {
2818  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(Decodingx2APIC));
2819  }
2820 
2821  // Get the highest cpuid leaf supported
2822  __kmp_x86_cpuid(0, 0, &buf);
2823  highest_leaf = buf.eax;
2824 
2825  // If a specific topology method was requested, only allow that specific leaf
2826  // otherwise, try both leaves 31 and 11 in that order
2827  num_leaves = 0;
2828  if (__kmp_affinity_top_method == affinity_top_method_x2apicid) {
2829  num_leaves = 1;
2830  leaves[0] = 11;
2831  leaf_message_id = kmp_i18n_str_NoLeaf11Support;
2832  } else if (__kmp_affinity_top_method == affinity_top_method_x2apicid_1f) {
2833  num_leaves = 1;
2834  leaves[0] = 31;
2835  leaf_message_id = kmp_i18n_str_NoLeaf31Support;
2836  } else {
2837  num_leaves = 2;
2838  leaves[0] = 31;
2839  leaves[1] = 11;
2840  leaf_message_id = kmp_i18n_str_NoLeaf11Support;
2841  }
2842 
2843  // Check to see if cpuid leaf 31 or 11 is supported.
2844  __kmp_nThreadsPerCore = nCoresPerPkg = nPackages = 1;
2845  topology_leaf = -1;
2846  for (int i = 0; i < num_leaves; ++i) {
2847  int leaf = leaves[i];
2848  if (highest_leaf < leaf)
2849  continue;
2850  __kmp_x86_cpuid(leaf, 0, &buf);
2851  if (buf.ebx == 0)
2852  continue;
2853  topology_leaf = leaf;
2854  __kmp_x2apicid_get_levels(leaf, &proc_info[0], types, &depth,
2855  &total_description);
2856  if (depth == 0)
2857  continue;
2858  break;
2859  }
2860  if (topology_leaf == -1 || depth == 0) {
2861  *msg_id = leaf_message_id;
2862  __kmp_free(proc_info);
2863  return false;
2864  }
2865  KMP_ASSERT(depth <= INTEL_LEVEL_TYPE_LAST);
2866 
2867  // The algorithm used starts by setting the affinity to each available thread
2868  // and retrieving info from the cpuid instruction, so if we are not capable of
2869  // calling __kmp_get_system_affinity() and __kmp_get_system_affinity(), then
2870  // we need to do something else - use the defaults that we calculated from
2871  // issuing cpuid without binding to each proc.
2872  if (!KMP_AFFINITY_CAPABLE()) {
2873  // Hack to try and infer the machine topology using only the data
2874  // available from cpuid on the current thread, and __kmp_xproc.
2875  KMP_ASSERT(__kmp_affinity.type == affinity_none);
2876  for (int i = 0; i < depth; ++i) {
2877  if (proc_info[0].levels[i].level_type == INTEL_LEVEL_TYPE_SMT) {
2878  __kmp_nThreadsPerCore = proc_info[0].levels[i].nitems;
2879  } else if (proc_info[0].levels[i].level_type == INTEL_LEVEL_TYPE_CORE) {
2880  nCoresPerPkg = proc_info[0].levels[i].nitems;
2881  }
2882  }
2883  __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
2884  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
2885  __kmp_free(proc_info);
2886  return true;
2887  }
2888 
2889  // From here on, we can assume that it is safe to call
2890  // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
2891  // __kmp_affinity.type = affinity_none.
2892 
2893  // Save the affinity mask for the current thread.
2894  kmp_affinity_raii_t previous_affinity;
2895 
2896  // Run through each of the available contexts, binding the current thread
2897  // to it, and obtaining the pertinent information using the cpuid instr.
2898  unsigned int proc;
2899  int hw_thread_index = 0;
2900  bool uniform_caches = true;
2901 
2902  KMP_CPU_SET_ITERATE(proc, __kmp_affin_fullMask) {
2903  // Skip this proc if it is not included in the machine model.
2904  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
2905  continue;
2906  }
2907  KMP_DEBUG_ASSERT(hw_thread_index < __kmp_avail_proc);
2908 
2909  // Gather topology information
2910  __kmp_affinity_dispatch->bind_thread(proc);
2911  __kmp_x86_cpuid(topology_leaf, 0, &buf);
2912  proc_info[hw_thread_index].os_id = proc;
2913  proc_info[hw_thread_index].apic_id = buf.edx;
2914  __kmp_x2apicid_get_levels(topology_leaf, &proc_info[hw_thread_index], types,
2915  &depth, &total_description);
2916  if (proc_info[hw_thread_index].depth == 0) {
2917  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
2918  __kmp_free(proc_info);
2919  return false;
2920  }
2921  // Gather cache information and insert afterwards
2922  cache_info[hw_thread_index].get_leaf4_levels();
2923  if (uniform_caches && hw_thread_index > 0)
2924  if (cache_info[0] != cache_info[hw_thread_index])
2925  uniform_caches = false;
2926  // Hybrid information
2927  if (__kmp_is_hybrid_cpu() && highest_leaf >= 0x1a) {
2928  __kmp_get_hybrid_info(&proc_info[hw_thread_index].type,
2929  &proc_info[hw_thread_index].efficiency,
2930  &proc_info[hw_thread_index].native_model_id);
2931  }
2932  hw_thread_index++;
2933  }
2934  KMP_ASSERT(hw_thread_index > 0);
2935  previous_affinity.restore();
2936 
2937  // Allocate the data structure to be returned.
2938  __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
2939 
2940  // Create topology Ids and hybrid types in __kmp_topology
2941  for (int i = 0; i < __kmp_topology->get_num_hw_threads(); ++i) {
2942  kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
2943  hw_thread.clear();
2944  hw_thread.os_id = proc_info[i].os_id;
2945  hw_thread.original_idx = i;
2946  unsigned apic_id = proc_info[i].apic_id;
2947  // Put in topology information
2948  for (int j = 0, idx = depth - 1; j < depth; ++j, --idx) {
2949  if (!(proc_info[i].description.contains_topology_type(
2950  __kmp_topology->get_type(j)))) {
2951  hw_thread.ids[idx] = kmp_hw_thread_t::UNKNOWN_ID;
2952  } else {
2953  hw_thread.ids[idx] = apic_id & proc_info[i].levels[j].mask;
2954  if (j > 0) {
2955  hw_thread.ids[idx] >>= proc_info[i].levels[j - 1].mask_width;
2956  }
2957  }
2958  }
2959  hw_thread.attrs.set_core_type(proc_info[i].type);
2960  hw_thread.attrs.set_core_eff(proc_info[i].efficiency);
2961  }
2962 
2963  __kmp_topology->sort_ids();
2964 
2965  // Change Ids to logical Ids
2966  for (int j = 0; j < depth - 1; ++j) {
2967  int new_id = 0;
2968  int prev_id = __kmp_topology->at(0).ids[j];
2969  int curr_id = __kmp_topology->at(0).ids[j + 1];
2970  __kmp_topology->at(0).ids[j + 1] = new_id;
2971  for (int i = 1; i < __kmp_topology->get_num_hw_threads(); ++i) {
2972  kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
2973  if (hw_thread.ids[j] == prev_id && hw_thread.ids[j + 1] == curr_id) {
2974  hw_thread.ids[j + 1] = new_id;
2975  } else if (hw_thread.ids[j] == prev_id &&
2976  hw_thread.ids[j + 1] != curr_id) {
2977  curr_id = hw_thread.ids[j + 1];
2978  hw_thread.ids[j + 1] = ++new_id;
2979  } else {
2980  prev_id = hw_thread.ids[j];
2981  curr_id = hw_thread.ids[j + 1];
2982  hw_thread.ids[j + 1] = ++new_id;
2983  }
2984  }
2985  }
2986 
2987  // First check for easy cache placement. This occurs when caches are
2988  // equivalent to a layer in the CPUID leaf 0xb or 0x1f topology.
2989  if (uniform_caches) {
2990  for (size_t i = 0; i < cache_info[0].get_depth(); ++i) {
2991  unsigned cache_mask = cache_info[0][i].mask;
2992  unsigned cache_level = cache_info[0][i].level;
2993  KMP_ASSERT(cache_level <= cpuid_cache_info_t::MAX_CACHE_LEVEL);
2994  kmp_hw_t cache_type = cpuid_cache_info_t::get_topology_type(cache_level);
2995  __kmp_topology->set_equivalent_type(cache_type, cache_type);
2996  for (int j = 0; j < depth; ++j) {
2997  unsigned hw_cache_mask = proc_info[0].levels[j].cache_mask;
2998  if (hw_cache_mask == cache_mask && j < depth - 1) {
2999  kmp_hw_t type = __kmp_intel_type_2_topology_type(
3000  proc_info[0].levels[j + 1].level_type);
3001  __kmp_topology->set_equivalent_type(cache_type, type);
3002  }
3003  }
3004  }
3005  } else {
3006  // If caches are non-uniform, then record which caches exist.
3007  for (int i = 0; i < __kmp_topology->get_num_hw_threads(); ++i) {
3008  for (size_t j = 0; j < cache_info[i].get_depth(); ++j) {
3009  unsigned cache_level = cache_info[i][j].level;
3010  kmp_hw_t cache_type =
3011  cpuid_cache_info_t::get_topology_type(cache_level);
3012  if (__kmp_topology->get_equivalent_type(cache_type) == KMP_HW_UNKNOWN)
3013  __kmp_topology->set_equivalent_type(cache_type, cache_type);
3014  }
3015  }
3016  }
3017 
3018  // See if any cache level needs to be added manually through cache Ids
3019  bool unresolved_cache_levels = false;
3020  for (unsigned level = 1; level <= cpuid_cache_info_t::MAX_CACHE_LEVEL;
3021  ++level) {
3022  kmp_hw_t cache_type = cpuid_cache_info_t::get_topology_type(level);
3023  // This also filters out caches which may not be in the topology
3024  // since the equivalent type might be KMP_HW_UNKNOWN.
3025  if (__kmp_topology->get_equivalent_type(cache_type) == cache_type) {
3026  unresolved_cache_levels = true;
3027  break;
3028  }
3029  }
3030 
3031  // Insert unresolved cache layers into machine topology using cache Ids
3032  if (unresolved_cache_levels) {
3033  int num_hw_threads = __kmp_topology->get_num_hw_threads();
3034  int *ids = (int *)__kmp_allocate(sizeof(int) * num_hw_threads);
3035  for (unsigned l = 1; l <= cpuid_cache_info_t::MAX_CACHE_LEVEL; ++l) {
3036  kmp_hw_t cache_type = cpuid_cache_info_t::get_topology_type(l);
3037  if (__kmp_topology->get_equivalent_type(cache_type) != cache_type)
3038  continue;
3039  for (int i = 0; i < num_hw_threads; ++i) {
3040  int original_idx = __kmp_topology->at(i).original_idx;
3041  ids[i] = kmp_hw_thread_t::UNKNOWN_ID;
3042  const cpuid_cache_info_t::info_t &info =
3043  cache_info[original_idx].get_level(l);
3044  // if cache level not in topology for this processor, then skip
3045  if (info.level == 0)
3046  continue;
3047  ids[i] = info.mask & proc_info[original_idx].apic_id;
3048  }
3049  __kmp_topology->insert_layer(cache_type, ids);
3050  }
3051  }
3052 
3053  if (!__kmp_topology->check_ids()) {
3054  kmp_topology_t::deallocate(__kmp_topology);
3055  __kmp_topology = nullptr;
3056  *msg_id = kmp_i18n_str_x2ApicIDsNotUnique;
3057  __kmp_free(proc_info);
3058  return false;
3059  }
3060  __kmp_free(proc_info);
3061  return true;
3062 }
3063 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3064 
3065 #define osIdIndex 0
3066 #define threadIdIndex 1
3067 #define coreIdIndex 2
3068 #define pkgIdIndex 3
3069 #define nodeIdIndex 4
3070 
3071 typedef unsigned *ProcCpuInfo;
3072 static unsigned maxIndex = pkgIdIndex;
3073 
3074 static int __kmp_affinity_cmp_ProcCpuInfo_phys_id(const void *a,
3075  const void *b) {
3076  unsigned i;
3077  const unsigned *aa = *(unsigned *const *)a;
3078  const unsigned *bb = *(unsigned *const *)b;
3079  for (i = maxIndex;; i--) {
3080  if (aa[i] < bb[i])
3081  return -1;
3082  if (aa[i] > bb[i])
3083  return 1;
3084  if (i == osIdIndex)
3085  break;
3086  }
3087  return 0;
3088 }
3089 
3090 #if KMP_USE_HIER_SCHED
3091 // Set the array sizes for the hierarchy layers
3092 static void __kmp_dispatch_set_hierarchy_values() {
3093  // Set the maximum number of L1's to number of cores
3094  // Set the maximum number of L2's to either number of cores / 2 for
3095  // Intel(R) Xeon Phi(TM) coprocessor formally codenamed Knights Landing
3096  // Or the number of cores for Intel(R) Xeon(R) processors
3097  // Set the maximum number of NUMA nodes and L3's to number of packages
3098  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1] =
3099  nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
3100  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L1 + 1] = __kmp_ncores;
3101 #if KMP_ARCH_X86_64 && \
3102  (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_DRAGONFLY || \
3103  KMP_OS_WINDOWS) && \
3104  KMP_MIC_SUPPORTED
3105  if (__kmp_mic_type >= mic3)
3106  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores / 2;
3107  else
3108 #endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
3109  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores;
3110  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L3 + 1] = nPackages;
3111  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_NUMA + 1] = nPackages;
3112  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_LOOP + 1] = 1;
3113  // Set the number of threads per unit
3114  // Number of hardware threads per L1/L2/L3/NUMA/LOOP
3115  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_THREAD + 1] = 1;
3116  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L1 + 1] =
3117  __kmp_nThreadsPerCore;
3118 #if KMP_ARCH_X86_64 && \
3119  (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_DRAGONFLY || \
3120  KMP_OS_WINDOWS) && \
3121  KMP_MIC_SUPPORTED
3122  if (__kmp_mic_type >= mic3)
3123  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
3124  2 * __kmp_nThreadsPerCore;
3125  else
3126 #endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
3127  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
3128  __kmp_nThreadsPerCore;
3129  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L3 + 1] =
3130  nCoresPerPkg * __kmp_nThreadsPerCore;
3131  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_NUMA + 1] =
3132  nCoresPerPkg * __kmp_nThreadsPerCore;
3133  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_LOOP + 1] =
3134  nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
3135 }
3136 
3137 // Return the index into the hierarchy for this tid and layer type (L1, L2, etc)
3138 // i.e., this thread's L1 or this thread's L2, etc.
3139 int __kmp_dispatch_get_index(int tid, kmp_hier_layer_e type) {
3140  int index = type + 1;
3141  int num_hw_threads = __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1];
3142  KMP_DEBUG_ASSERT(type != kmp_hier_layer_e::LAYER_LAST);
3143  if (type == kmp_hier_layer_e::LAYER_THREAD)
3144  return tid;
3145  else if (type == kmp_hier_layer_e::LAYER_LOOP)
3146  return 0;
3147  KMP_DEBUG_ASSERT(__kmp_hier_max_units[index] != 0);
3148  if (tid >= num_hw_threads)
3149  tid = tid % num_hw_threads;
3150  return (tid / __kmp_hier_threads_per[index]) % __kmp_hier_max_units[index];
3151 }
3152 
3153 // Return the number of t1's per t2
3154 int __kmp_dispatch_get_t1_per_t2(kmp_hier_layer_e t1, kmp_hier_layer_e t2) {
3155  int i1 = t1 + 1;
3156  int i2 = t2 + 1;
3157  KMP_DEBUG_ASSERT(i1 <= i2);
3158  KMP_DEBUG_ASSERT(t1 != kmp_hier_layer_e::LAYER_LAST);
3159  KMP_DEBUG_ASSERT(t2 != kmp_hier_layer_e::LAYER_LAST);
3160  KMP_DEBUG_ASSERT(__kmp_hier_threads_per[i1] != 0);
3161  // (nthreads/t2) / (nthreads/t1) = t1 / t2
3162  return __kmp_hier_threads_per[i2] / __kmp_hier_threads_per[i1];
3163 }
3164 #endif // KMP_USE_HIER_SCHED
3165 
3166 static inline const char *__kmp_cpuinfo_get_filename() {
3167  const char *filename;
3168  if (__kmp_cpuinfo_file != nullptr)
3169  filename = __kmp_cpuinfo_file;
3170  else
3171  filename = "/proc/cpuinfo";
3172  return filename;
3173 }
3174 
3175 static inline const char *__kmp_cpuinfo_get_envvar() {
3176  const char *envvar = nullptr;
3177  if (__kmp_cpuinfo_file != nullptr)
3178  envvar = "KMP_CPUINFO_FILE";
3179  return envvar;
3180 }
3181 
3182 static bool __kmp_package_id_from_core_siblings_list(unsigned **threadInfo,
3183  unsigned num_avail,
3184  unsigned idx) {
3185  if (!KMP_AFFINITY_CAPABLE())
3186  return false;
3187 
3188  char path[256];
3189  KMP_SNPRINTF(path, sizeof(path),
3190  "/sys/devices/system/cpu/cpu%u/topology/core_siblings_list",
3191  threadInfo[idx][osIdIndex]);
3192  kmp_affin_mask_t *siblings = __kmp_parse_cpu_list(path);
3193  for (unsigned i = 0; i < num_avail; ++i) {
3194  unsigned cpu_id = threadInfo[i][osIdIndex];
3195  KMP_ASSERT(cpu_id < __kmp_affin_mask_size * CHAR_BIT);
3196  if (!KMP_CPU_ISSET(cpu_id, siblings))
3197  continue;
3198  if (threadInfo[i][pkgIdIndex] == UINT_MAX) {
3199  // Arbitrarily pick the first index we encounter, it only matters that
3200  // the value is the same for all siblings.
3201  threadInfo[i][pkgIdIndex] = idx;
3202  } else if (threadInfo[i][pkgIdIndex] != idx) {
3203  // Contradictory sibling lists.
3204  KMP_CPU_FREE(siblings);
3205  return false;
3206  }
3207  }
3208  KMP_ASSERT(threadInfo[idx][pkgIdIndex] != UINT_MAX);
3209  KMP_CPU_FREE(siblings);
3210  return true;
3211 }
3212 
3213 // Parse /proc/cpuinfo (or an alternate file in the same format) to obtain the
3214 // affinity map. On AIX, the map is obtained through system SRAD (Scheduler
3215 // Resource Allocation Domain).
3216 static bool __kmp_affinity_create_cpuinfo_map(int *line,
3217  kmp_i18n_id_t *const msg_id) {
3218  *msg_id = kmp_i18n_null;
3219 
3220 #if KMP_OS_AIX
3221  unsigned num_records = __kmp_xproc;
3222 #else
3223  const char *filename = __kmp_cpuinfo_get_filename();
3224  const char *envvar = __kmp_cpuinfo_get_envvar();
3225 
3226  if (__kmp_affinity.flags.verbose) {
3227  KMP_INFORM(AffParseFilename, "KMP_AFFINITY", filename);
3228  }
3229 
3230  kmp_safe_raii_file_t f(filename, "r", envvar);
3231 
3232  // Scan of the file, and count the number of "processor" (osId) fields,
3233  // and find the highest value of <n> for a node_<n> field.
3234  char buf[256];
3235  unsigned num_records = 0;
3236  while (!feof(f)) {
3237  buf[sizeof(buf) - 1] = 1;
3238  if (!fgets(buf, sizeof(buf), f)) {
3239  // Read errors presumably because of EOF
3240  break;
3241  }
3242 
3243  char s1[] = "processor";
3244  if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
3245  num_records++;
3246  continue;
3247  }
3248 
3249  // FIXME - this will match "node_<n> <garbage>"
3250  unsigned level;
3251  if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
3252  // validate the input fisrt:
3253  if (level > (unsigned)__kmp_xproc) { // level is too big
3254  level = __kmp_xproc;
3255  }
3256  if (nodeIdIndex + level >= maxIndex) {
3257  maxIndex = nodeIdIndex + level;
3258  }
3259  continue;
3260  }
3261  }
3262 
3263  // Check for empty file / no valid processor records, or too many. The number
3264  // of records can't exceed the number of valid bits in the affinity mask.
3265  if (num_records == 0) {
3266  *msg_id = kmp_i18n_str_NoProcRecords;
3267  return false;
3268  }
3269  if (num_records > (unsigned)__kmp_xproc) {
3270  *msg_id = kmp_i18n_str_TooManyProcRecords;
3271  return false;
3272  }
3273 
3274  // Set the file pointer back to the beginning, so that we can scan the file
3275  // again, this time performing a full parse of the data. Allocate a vector of
3276  // ProcCpuInfo object, where we will place the data. Adding an extra element
3277  // at the end allows us to remove a lot of extra checks for termination
3278  // conditions.
3279  if (fseek(f, 0, SEEK_SET) != 0) {
3280  *msg_id = kmp_i18n_str_CantRewindCpuinfo;
3281  return false;
3282  }
3283 #endif // KMP_OS_AIX
3284 
3285  // Allocate the array of records to store the proc info in. The dummy
3286  // element at the end makes the logic in filling them out easier to code.
3287  unsigned **threadInfo =
3288  (unsigned **)__kmp_allocate((num_records + 1) * sizeof(unsigned *));
3289  unsigned i;
3290  for (i = 0; i <= num_records; i++) {
3291  threadInfo[i] =
3292  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3293  }
3294 
3295 #define CLEANUP_THREAD_INFO \
3296  for (i = 0; i <= num_records; i++) { \
3297  __kmp_free(threadInfo[i]); \
3298  } \
3299  __kmp_free(threadInfo);
3300 
3301  // A value of UINT_MAX means that we didn't find the field
3302  unsigned __index;
3303 
3304 #define INIT_PROC_INFO(p) \
3305  for (__index = 0; __index <= maxIndex; __index++) { \
3306  (p)[__index] = UINT_MAX; \
3307  }
3308 
3309  for (i = 0; i <= num_records; i++) {
3310  INIT_PROC_INFO(threadInfo[i]);
3311  }
3312 
3313 #if KMP_OS_AIX
3314  int smt_threads;
3315  lpar_info_format1_t cpuinfo;
3316  unsigned num_avail = __kmp_xproc;
3317 
3318  if (__kmp_affinity.flags.verbose)
3319  KMP_INFORM(AffParseFilename, "KMP_AFFINITY", "system info for topology");
3320 
3321  // Get the number of SMT threads per core.
3322  smt_threads = syssmt(GET_NUMBER_SMT_SETS, 0, 0, NULL);
3323 
3324  // Allocate a resource set containing available system resourses.
3325  rsethandle_t sys_rset = rs_alloc(RS_SYSTEM);
3326  if (sys_rset == NULL) {
3327  CLEANUP_THREAD_INFO;
3328  *msg_id = kmp_i18n_str_UnknownTopology;
3329  return false;
3330  }
3331  // Allocate a resource set for the SRAD info.
3332  rsethandle_t srad = rs_alloc(RS_EMPTY);
3333  if (srad == NULL) {
3334  rs_free(sys_rset);
3335  CLEANUP_THREAD_INFO;
3336  *msg_id = kmp_i18n_str_UnknownTopology;
3337  return false;
3338  }
3339 
3340  // Get the SRAD system detail level.
3341  int sradsdl = rs_getinfo(NULL, R_SRADSDL, 0);
3342  if (sradsdl < 0) {
3343  rs_free(sys_rset);
3344  rs_free(srad);
3345  CLEANUP_THREAD_INFO;
3346  *msg_id = kmp_i18n_str_UnknownTopology;
3347  return false;
3348  }
3349  // Get the number of RADs at that SRAD SDL.
3350  int num_rads = rs_numrads(sys_rset, sradsdl, 0);
3351  if (num_rads < 0) {
3352  rs_free(sys_rset);
3353  rs_free(srad);
3354  CLEANUP_THREAD_INFO;
3355  *msg_id = kmp_i18n_str_UnknownTopology;
3356  return false;
3357  }
3358 
3359  // Get the maximum number of procs that may be contained in a resource set.
3360  int max_procs = rs_getinfo(NULL, R_MAXPROCS, 0);
3361  if (max_procs < 0) {
3362  rs_free(sys_rset);
3363  rs_free(srad);
3364  CLEANUP_THREAD_INFO;
3365  *msg_id = kmp_i18n_str_UnknownTopology;
3366  return false;
3367  }
3368 
3369  int cur_rad = 0;
3370  int num_set = 0;
3371  for (int srad_idx = 0; cur_rad < num_rads && srad_idx < VMI_MAXRADS;
3372  ++srad_idx) {
3373  // Check if the SRAD is available in the RSET.
3374  if (rs_getrad(sys_rset, srad, sradsdl, srad_idx, 0) < 0)
3375  continue;
3376 
3377  for (int cpu = 0; cpu < max_procs; cpu++) {
3378  // Set the info for the cpu if it is in the SRAD.
3379  if (rs_op(RS_TESTRESOURCE, srad, NULL, R_PROCS, cpu)) {
3380  threadInfo[cpu][osIdIndex] = cpu;
3381  threadInfo[cpu][pkgIdIndex] = cur_rad;
3382  threadInfo[cpu][coreIdIndex] = cpu / smt_threads;
3383  ++num_set;
3384  if (num_set >= num_avail) {
3385  // Done if all available CPUs have been set.
3386  break;
3387  }
3388  }
3389  }
3390  ++cur_rad;
3391  }
3392  rs_free(sys_rset);
3393  rs_free(srad);
3394 
3395  // The topology is already sorted.
3396 
3397 #else // !KMP_OS_AIX
3398  unsigned num_avail = 0;
3399  *line = 0;
3400 #if KMP_ARCH_S390X
3401  bool reading_s390x_sys_info = true;
3402 #endif
3403  while (!feof(f)) {
3404  // Create an inner scoping level, so that all the goto targets at the end of
3405  // the loop appear in an outer scoping level. This avoids warnings about
3406  // jumping past an initialization to a target in the same block.
3407  {
3408  buf[sizeof(buf) - 1] = 1;
3409  bool long_line = false;
3410  if (!fgets(buf, sizeof(buf), f)) {
3411  // Read errors presumably because of EOF
3412  // If there is valid data in threadInfo[num_avail], then fake
3413  // a blank line in ensure that the last address gets parsed.
3414  bool valid = false;
3415  for (i = 0; i <= maxIndex; i++) {
3416  if (threadInfo[num_avail][i] != UINT_MAX) {
3417  valid = true;
3418  }
3419  }
3420  if (!valid) {
3421  break;
3422  }
3423  buf[0] = 0;
3424  } else if (!buf[sizeof(buf) - 1]) {
3425  // The line is longer than the buffer. Set a flag and don't
3426  // emit an error if we were going to ignore the line, anyway.
3427  long_line = true;
3428 
3429 #define CHECK_LINE \
3430  if (long_line) { \
3431  CLEANUP_THREAD_INFO; \
3432  *msg_id = kmp_i18n_str_LongLineCpuinfo; \
3433  return false; \
3434  }
3435  }
3436  (*line)++;
3437 
3438 #if KMP_ARCH_LOONGARCH64
3439  // The parsing logic of /proc/cpuinfo in this function highly depends on
3440  // the blank lines between each processor info block. But on LoongArch a
3441  // blank line exists before the first processor info block (i.e. after the
3442  // "system type" line). This blank line was added because the "system
3443  // type" line is unrelated to any of the CPUs. We must skip this line so
3444  // that the original logic works on LoongArch.
3445  if (*buf == '\n' && *line == 2)
3446  continue;
3447 #endif
3448 #if KMP_ARCH_S390X
3449  // s390x /proc/cpuinfo starts with a variable number of lines containing
3450  // the overall system information. Skip them.
3451  if (reading_s390x_sys_info) {
3452  if (*buf == '\n')
3453  reading_s390x_sys_info = false;
3454  continue;
3455  }
3456 #endif
3457 
3458 #if KMP_ARCH_S390X
3459  char s1[] = "cpu number";
3460 #else
3461  char s1[] = "processor";
3462 #endif
3463  if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
3464  CHECK_LINE;
3465  char *p = strchr(buf + sizeof(s1) - 1, ':');
3466  unsigned val;
3467  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3468  goto no_val;
3469  if (threadInfo[num_avail][osIdIndex] != UINT_MAX)
3470 #if KMP_ARCH_AARCH64
3471  // Handle the old AArch64 /proc/cpuinfo layout differently,
3472  // it contains all of the 'processor' entries listed in a
3473  // single 'Processor' section, therefore the normal looking
3474  // for duplicates in that section will always fail.
3475  num_avail++;
3476 #else
3477  goto dup_field;
3478 #endif
3479  threadInfo[num_avail][osIdIndex] = val;
3480 #if KMP_OS_LINUX && !(KMP_ARCH_X86 || KMP_ARCH_X86_64)
3481  char path[256];
3482  KMP_SNPRINTF(
3483  path, sizeof(path),
3484  "/sys/devices/system/cpu/cpu%u/topology/physical_package_id",
3485  threadInfo[num_avail][osIdIndex]);
3486  __kmp_read_from_file(path, "%u", &threadInfo[num_avail][pkgIdIndex]);
3487 
3488 #if KMP_ARCH_S390X
3489  // Disambiguate physical_package_id.
3490  unsigned book_id;
3491  KMP_SNPRINTF(path, sizeof(path),
3492  "/sys/devices/system/cpu/cpu%u/topology/book_id",
3493  threadInfo[num_avail][osIdIndex]);
3494  __kmp_read_from_file(path, "%u", &book_id);
3495  threadInfo[num_avail][pkgIdIndex] |= (book_id << 8);
3496 
3497  unsigned drawer_id;
3498  KMP_SNPRINTF(path, sizeof(path),
3499  "/sys/devices/system/cpu/cpu%u/topology/drawer_id",
3500  threadInfo[num_avail][osIdIndex]);
3501  __kmp_read_from_file(path, "%u", &drawer_id);
3502  threadInfo[num_avail][pkgIdIndex] |= (drawer_id << 16);
3503 #endif
3504 
3505  KMP_SNPRINTF(path, sizeof(path),
3506  "/sys/devices/system/cpu/cpu%u/topology/core_id",
3507  threadInfo[num_avail][osIdIndex]);
3508  __kmp_read_from_file(path, "%u", &threadInfo[num_avail][coreIdIndex]);
3509  continue;
3510 #else
3511  }
3512  char s2[] = "physical id";
3513  if (strncmp(buf, s2, sizeof(s2) - 1) == 0) {
3514  CHECK_LINE;
3515  char *p = strchr(buf + sizeof(s2) - 1, ':');
3516  unsigned val;
3517  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3518  goto no_val;
3519  if (threadInfo[num_avail][pkgIdIndex] != UINT_MAX)
3520  goto dup_field;
3521  threadInfo[num_avail][pkgIdIndex] = val;
3522  continue;
3523  }
3524  char s3[] = "core id";
3525  if (strncmp(buf, s3, sizeof(s3) - 1) == 0) {
3526  CHECK_LINE;
3527  char *p = strchr(buf + sizeof(s3) - 1, ':');
3528  unsigned val;
3529  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3530  goto no_val;
3531  if (threadInfo[num_avail][coreIdIndex] != UINT_MAX)
3532  goto dup_field;
3533  threadInfo[num_avail][coreIdIndex] = val;
3534  continue;
3535 #endif // KMP_OS_LINUX && USE_SYSFS_INFO
3536  }
3537  char s4[] = "thread id";
3538  if (strncmp(buf, s4, sizeof(s4) - 1) == 0) {
3539  CHECK_LINE;
3540  char *p = strchr(buf + sizeof(s4) - 1, ':');
3541  unsigned val;
3542  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3543  goto no_val;
3544  if (threadInfo[num_avail][threadIdIndex] != UINT_MAX)
3545  goto dup_field;
3546  threadInfo[num_avail][threadIdIndex] = val;
3547  continue;
3548  }
3549  unsigned level;
3550  if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
3551  CHECK_LINE;
3552  char *p = strchr(buf + sizeof(s4) - 1, ':');
3553  unsigned val;
3554  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3555  goto no_val;
3556  // validate the input before using level:
3557  if (level > (unsigned)__kmp_xproc) { // level is too big
3558  level = __kmp_xproc;
3559  }
3560  if (threadInfo[num_avail][nodeIdIndex + level] != UINT_MAX)
3561  goto dup_field;
3562  threadInfo[num_avail][nodeIdIndex + level] = val;
3563  continue;
3564  }
3565 
3566  // We didn't recognize the leading token on the line. There are lots of
3567  // leading tokens that we don't recognize - if the line isn't empty, go on
3568  // to the next line.
3569  if ((*buf != 0) && (*buf != '\n')) {
3570  // If the line is longer than the buffer, read characters
3571  // until we find a newline.
3572  if (long_line) {
3573  int ch;
3574  while (((ch = fgetc(f)) != EOF) && (ch != '\n'))
3575  ;
3576  }
3577  continue;
3578  }
3579 
3580  // A newline has signalled the end of the processor record.
3581  // Check that there aren't too many procs specified.
3582  if ((int)num_avail == __kmp_xproc) {
3583  CLEANUP_THREAD_INFO;
3584  *msg_id = kmp_i18n_str_TooManyEntries;
3585  return false;
3586  }
3587 
3588  // Check for missing fields. The osId field must be there. The physical
3589  // id field will be checked later.
3590  if (threadInfo[num_avail][osIdIndex] == UINT_MAX) {
3591  CLEANUP_THREAD_INFO;
3592  *msg_id = kmp_i18n_str_MissingProcField;
3593  return false;
3594  }
3595 
3596  // Skip this proc if it is not included in the machine model.
3597  if (KMP_AFFINITY_CAPABLE() &&
3598  !KMP_CPU_ISSET(threadInfo[num_avail][osIdIndex],
3599  __kmp_affin_fullMask)) {
3600  INIT_PROC_INFO(threadInfo[num_avail]);
3601  continue;
3602  }
3603 
3604  // We have a successful parse of this proc's info.
3605  // Increment the counter, and prepare for the next proc.
3606  num_avail++;
3607  KMP_ASSERT(num_avail <= num_records);
3608  INIT_PROC_INFO(threadInfo[num_avail]);
3609  }
3610  continue;
3611 
3612  no_val:
3613  CLEANUP_THREAD_INFO;
3614  *msg_id = kmp_i18n_str_MissingValCpuinfo;
3615  return false;
3616 
3617  dup_field:
3618  CLEANUP_THREAD_INFO;
3619  *msg_id = kmp_i18n_str_DuplicateFieldCpuinfo;
3620  return false;
3621  }
3622  *line = 0;
3623 
3624  // At least on powerpc, Linux may return -1 for physical_package_id. Try
3625  // to reconstruct topology from core_siblings_list in that case.
3626  for (i = 0; i < num_avail; ++i) {
3627  if (threadInfo[i][pkgIdIndex] == UINT_MAX) {
3628  if (!__kmp_package_id_from_core_siblings_list(threadInfo, num_avail, i)) {
3629  CLEANUP_THREAD_INFO;
3630  *msg_id = kmp_i18n_str_MissingPhysicalIDField;
3631  return false;
3632  }
3633  }
3634  }
3635 
3636 #if KMP_MIC && REDUCE_TEAM_SIZE
3637  unsigned teamSize = 0;
3638 #endif // KMP_MIC && REDUCE_TEAM_SIZE
3639 
3640  // check for num_records == __kmp_xproc ???
3641 
3642  // If it is configured to omit the package level when there is only a single
3643  // package, the logic at the end of this routine won't work if there is only a
3644  // single thread
3645  KMP_ASSERT(num_avail > 0);
3646  KMP_ASSERT(num_avail <= num_records);
3647 
3648  // Sort the threadInfo table by physical Id.
3649  qsort(threadInfo, num_avail, sizeof(*threadInfo),
3650  __kmp_affinity_cmp_ProcCpuInfo_phys_id);
3651 
3652 #endif // KMP_OS_AIX
3653 
3654  // The table is now sorted by pkgId / coreId / threadId, but we really don't
3655  // know the radix of any of the fields. pkgId's may be sparsely assigned among
3656  // the chips on a system. Although coreId's are usually assigned
3657  // [0 .. coresPerPkg-1] and threadId's are usually assigned
3658  // [0..threadsPerCore-1], we don't want to make any such assumptions.
3659  //
3660  // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
3661  // total # packages) are at this point - we want to determine that now. We
3662  // only have an upper bound on the first two figures.
3663  unsigned *counts =
3664  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3665  unsigned *maxCt =
3666  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3667  unsigned *totals =
3668  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3669  unsigned *lastId =
3670  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3671 
3672  bool assign_thread_ids = false;
3673  unsigned threadIdCt;
3674  unsigned index;
3675 
3676 restart_radix_check:
3677  threadIdCt = 0;
3678 
3679  // Initialize the counter arrays with data from threadInfo[0].
3680  if (assign_thread_ids) {
3681  if (threadInfo[0][threadIdIndex] == UINT_MAX) {
3682  threadInfo[0][threadIdIndex] = threadIdCt++;
3683  } else if (threadIdCt <= threadInfo[0][threadIdIndex]) {
3684  threadIdCt = threadInfo[0][threadIdIndex] + 1;
3685  }
3686  }
3687  for (index = 0; index <= maxIndex; index++) {
3688  counts[index] = 1;
3689  maxCt[index] = 1;
3690  totals[index] = 1;
3691  lastId[index] = threadInfo[0][index];
3692  ;
3693  }
3694 
3695  // Run through the rest of the OS procs.
3696  for (i = 1; i < num_avail; i++) {
3697  // Find the most significant index whose id differs from the id for the
3698  // previous OS proc.
3699  for (index = maxIndex; index >= threadIdIndex; index--) {
3700  if (assign_thread_ids && (index == threadIdIndex)) {
3701  // Auto-assign the thread id field if it wasn't specified.
3702  if (threadInfo[i][threadIdIndex] == UINT_MAX) {
3703  threadInfo[i][threadIdIndex] = threadIdCt++;
3704  }
3705  // Apparently the thread id field was specified for some entries and not
3706  // others. Start the thread id counter off at the next higher thread id.
3707  else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
3708  threadIdCt = threadInfo[i][threadIdIndex] + 1;
3709  }
3710  }
3711  if (threadInfo[i][index] != lastId[index]) {
3712  // Run through all indices which are less significant, and reset the
3713  // counts to 1. At all levels up to and including index, we need to
3714  // increment the totals and record the last id.
3715  unsigned index2;
3716  for (index2 = threadIdIndex; index2 < index; index2++) {
3717  totals[index2]++;
3718  if (counts[index2] > maxCt[index2]) {
3719  maxCt[index2] = counts[index2];
3720  }
3721  counts[index2] = 1;
3722  lastId[index2] = threadInfo[i][index2];
3723  }
3724  counts[index]++;
3725  totals[index]++;
3726  lastId[index] = threadInfo[i][index];
3727 
3728  if (assign_thread_ids && (index > threadIdIndex)) {
3729 
3730 #if KMP_MIC && REDUCE_TEAM_SIZE
3731  // The default team size is the total #threads in the machine
3732  // minus 1 thread for every core that has 3 or more threads.
3733  teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
3734 #endif // KMP_MIC && REDUCE_TEAM_SIZE
3735 
3736  // Restart the thread counter, as we are on a new core.
3737  threadIdCt = 0;
3738 
3739  // Auto-assign the thread id field if it wasn't specified.
3740  if (threadInfo[i][threadIdIndex] == UINT_MAX) {
3741  threadInfo[i][threadIdIndex] = threadIdCt++;
3742  }
3743 
3744  // Apparently the thread id field was specified for some entries and
3745  // not others. Start the thread id counter off at the next higher
3746  // thread id.
3747  else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
3748  threadIdCt = threadInfo[i][threadIdIndex] + 1;
3749  }
3750  }
3751  break;
3752  }
3753  }
3754  if (index < threadIdIndex) {
3755  // If thread ids were specified, it is an error if they are not unique.
3756  // Also, check that we waven't already restarted the loop (to be safe -
3757  // shouldn't need to).
3758  if ((threadInfo[i][threadIdIndex] != UINT_MAX) || assign_thread_ids) {
3759  __kmp_free(lastId);
3760  __kmp_free(totals);
3761  __kmp_free(maxCt);
3762  __kmp_free(counts);
3763  CLEANUP_THREAD_INFO;
3764  *msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
3765  return false;
3766  }
3767 
3768  // If the thread ids were not specified and we see entries that
3769  // are duplicates, start the loop over and assign the thread ids manually.
3770  assign_thread_ids = true;
3771  goto restart_radix_check;
3772  }
3773  }
3774 
3775 #if KMP_MIC && REDUCE_TEAM_SIZE
3776  // The default team size is the total #threads in the machine
3777  // minus 1 thread for every core that has 3 or more threads.
3778  teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
3779 #endif // KMP_MIC && REDUCE_TEAM_SIZE
3780 
3781  for (index = threadIdIndex; index <= maxIndex; index++) {
3782  if (counts[index] > maxCt[index]) {
3783  maxCt[index] = counts[index];
3784  }
3785  }
3786 
3787  __kmp_nThreadsPerCore = maxCt[threadIdIndex];
3788  nCoresPerPkg = maxCt[coreIdIndex];
3789  nPackages = totals[pkgIdIndex];
3790 
3791  // When affinity is off, this routine will still be called to set
3792  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
3793  // Make sure all these vars are set correctly, and return now if affinity is
3794  // not enabled.
3795  __kmp_ncores = totals[coreIdIndex];
3796  if (!KMP_AFFINITY_CAPABLE()) {
3797  KMP_ASSERT(__kmp_affinity.type == affinity_none);
3798  return true;
3799  }
3800 
3801 #if KMP_MIC && REDUCE_TEAM_SIZE
3802  // Set the default team size.
3803  if ((__kmp_dflt_team_nth == 0) && (teamSize > 0)) {
3804  __kmp_dflt_team_nth = teamSize;
3805  KA_TRACE(20, ("__kmp_affinity_create_cpuinfo_map: setting "
3806  "__kmp_dflt_team_nth = %d\n",
3807  __kmp_dflt_team_nth));
3808  }
3809 #endif // KMP_MIC && REDUCE_TEAM_SIZE
3810 
3811  KMP_DEBUG_ASSERT(num_avail == (unsigned)__kmp_avail_proc);
3812 
3813  // Count the number of levels which have more nodes at that level than at the
3814  // parent's level (with there being an implicit root node of the top level).
3815  // This is equivalent to saying that there is at least one node at this level
3816  // which has a sibling. These levels are in the map, and the package level is
3817  // always in the map.
3818  bool *inMap = (bool *)__kmp_allocate((maxIndex + 1) * sizeof(bool));
3819  for (index = threadIdIndex; index < maxIndex; index++) {
3820  KMP_ASSERT(totals[index] >= totals[index + 1]);
3821  inMap[index] = (totals[index] > totals[index + 1]);
3822  }
3823  inMap[maxIndex] = (totals[maxIndex] > 1);
3824  inMap[pkgIdIndex] = true;
3825  inMap[coreIdIndex] = true;
3826  inMap[threadIdIndex] = true;
3827 
3828  int depth = 0;
3829  int idx = 0;
3830  kmp_hw_t types[KMP_HW_LAST];
3831  int pkgLevel = -1;
3832  int coreLevel = -1;
3833  int threadLevel = -1;
3834  for (index = threadIdIndex; index <= maxIndex; index++) {
3835  if (inMap[index]) {
3836  depth++;
3837  }
3838  }
3839  if (inMap[pkgIdIndex]) {
3840  pkgLevel = idx;
3841  types[idx++] = KMP_HW_SOCKET;
3842  }
3843  if (inMap[coreIdIndex]) {
3844  coreLevel = idx;
3845  types[idx++] = KMP_HW_CORE;
3846  }
3847  if (inMap[threadIdIndex]) {
3848  threadLevel = idx;
3849  types[idx++] = KMP_HW_THREAD;
3850  }
3851  KMP_ASSERT(depth > 0);
3852 
3853  // Construct the data structure that is to be returned.
3854  __kmp_topology = kmp_topology_t::allocate(num_avail, depth, types);
3855 
3856  for (i = 0; i < num_avail; ++i) {
3857  unsigned os = threadInfo[i][osIdIndex];
3858  int src_index;
3859  kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
3860  hw_thread.clear();
3861  hw_thread.os_id = os;
3862  hw_thread.original_idx = i;
3863 
3864  idx = 0;
3865  for (src_index = maxIndex; src_index >= threadIdIndex; src_index--) {
3866  if (!inMap[src_index]) {
3867  continue;
3868  }
3869  if (src_index == pkgIdIndex) {
3870  hw_thread.ids[pkgLevel] = threadInfo[i][src_index];
3871  } else if (src_index == coreIdIndex) {
3872  hw_thread.ids[coreLevel] = threadInfo[i][src_index];
3873  } else if (src_index == threadIdIndex) {
3874  hw_thread.ids[threadLevel] = threadInfo[i][src_index];
3875  }
3876  }
3877  }
3878 
3879  __kmp_free(inMap);
3880  __kmp_free(lastId);
3881  __kmp_free(totals);
3882  __kmp_free(maxCt);
3883  __kmp_free(counts);
3884  CLEANUP_THREAD_INFO;
3885  __kmp_topology->sort_ids();
3886 
3887  int tlevel = __kmp_topology->get_level(KMP_HW_THREAD);
3888  if (tlevel > 0) {
3889  // If the thread level does not have ids, then put them in.
3890  if (__kmp_topology->at(0).ids[tlevel] == kmp_hw_thread_t::UNKNOWN_ID) {
3891  __kmp_topology->at(0).ids[tlevel] = 0;
3892  }
3893  for (int i = 1; i < __kmp_topology->get_num_hw_threads(); ++i) {
3894  kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
3895  if (hw_thread.ids[tlevel] != kmp_hw_thread_t::UNKNOWN_ID)
3896  continue;
3897  kmp_hw_thread_t &prev_hw_thread = __kmp_topology->at(i - 1);
3898  // Check if socket, core, anything above thread level changed.
3899  // If the ids did change, then restart thread id at 0
3900  // Otherwise, set thread id to prev thread's id + 1
3901  for (int j = 0; j < tlevel; ++j) {
3902  if (hw_thread.ids[j] != prev_hw_thread.ids[j]) {
3903  hw_thread.ids[tlevel] = 0;
3904  break;
3905  }
3906  }
3907  if (hw_thread.ids[tlevel] == kmp_hw_thread_t::UNKNOWN_ID)
3908  hw_thread.ids[tlevel] = prev_hw_thread.ids[tlevel] + 1;
3909  }
3910  }
3911 
3912  if (!__kmp_topology->check_ids()) {
3913  kmp_topology_t::deallocate(__kmp_topology);
3914  __kmp_topology = nullptr;
3915  *msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
3916  return false;
3917  }
3918  return true;
3919 }
3920 
3921 // Create and return a table of affinity masks, indexed by OS thread ID.
3922 // This routine handles OR'ing together all the affinity masks of threads
3923 // that are sufficiently close, if granularity > fine.
3924 template <typename FindNextFunctionType>
3925 static void __kmp_create_os_id_masks(unsigned *numUnique,
3926  kmp_affinity_t &affinity,
3927  FindNextFunctionType find_next) {
3928  // First form a table of affinity masks in order of OS thread id.
3929  int maxOsId;
3930  int i;
3931  int numAddrs = __kmp_topology->get_num_hw_threads();
3932  int depth = __kmp_topology->get_depth();
3933  const char *env_var = __kmp_get_affinity_env_var(affinity);
3934  KMP_ASSERT(numAddrs);
3935  KMP_ASSERT(depth);
3936 
3937  i = find_next(-1);
3938  // If could not find HW thread location that satisfies find_next conditions,
3939  // then return and fallback to increment find_next.
3940  if (i >= numAddrs)
3941  return;
3942 
3943  maxOsId = 0;
3944  for (i = numAddrs - 1;; --i) {
3945  int osId = __kmp_topology->at(i).os_id;
3946  if (osId > maxOsId) {
3947  maxOsId = osId;
3948  }
3949  if (i == 0)
3950  break;
3951  }
3952  affinity.num_os_id_masks = maxOsId + 1;
3953  KMP_CPU_ALLOC_ARRAY(affinity.os_id_masks, affinity.num_os_id_masks);
3954  KMP_ASSERT(affinity.gran_levels >= 0);
3955  if (affinity.flags.verbose && (affinity.gran_levels > 0)) {
3956  KMP_INFORM(ThreadsMigrate, env_var, affinity.gran_levels);
3957  }
3958  if (affinity.gran_levels >= (int)depth) {
3959  KMP_AFF_WARNING(affinity, AffThreadsMayMigrate);
3960  }
3961 
3962  // Run through the table, forming the masks for all threads on each core.
3963  // Threads on the same core will have identical kmp_hw_thread_t objects, not
3964  // considering the last level, which must be the thread id. All threads on a
3965  // core will appear consecutively.
3966  int unique = 0;
3967  int j = 0; // index of 1st thread on core
3968  int leader = 0;
3969  kmp_affin_mask_t *sum;
3970  KMP_CPU_ALLOC_ON_STACK(sum);
3971  KMP_CPU_ZERO(sum);
3972 
3973  i = j = leader = find_next(-1);
3974  KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
3975  kmp_full_mask_modifier_t full_mask;
3976  for (i = find_next(i); i < numAddrs; i = find_next(i)) {
3977  // If this thread is sufficiently close to the leader (within the
3978  // granularity setting), then set the bit for this os thread in the
3979  // affinity mask for this group, and go on to the next thread.
3980  if (__kmp_topology->is_close(leader, i, affinity)) {
3981  KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
3982  continue;
3983  }
3984 
3985  // For every thread in this group, copy the mask to the thread's entry in
3986  // the OS Id mask table. Mark the first address as a leader.
3987  for (; j < i; j = find_next(j)) {
3988  int osId = __kmp_topology->at(j).os_id;
3989  KMP_DEBUG_ASSERT(osId <= maxOsId);
3990  kmp_affin_mask_t *mask = KMP_CPU_INDEX(affinity.os_id_masks, osId);
3991  KMP_CPU_COPY(mask, sum);
3992  __kmp_topology->at(j).leader = (j == leader);
3993  }
3994  unique++;
3995 
3996  // Start a new mask.
3997  leader = i;
3998  full_mask.include(sum);
3999  KMP_CPU_ZERO(sum);
4000  KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
4001  }
4002 
4003  // For every thread in last group, copy the mask to the thread's
4004  // entry in the OS Id mask table.
4005  for (; j < i; j = find_next(j)) {
4006  int osId = __kmp_topology->at(j).os_id;
4007  KMP_DEBUG_ASSERT(osId <= maxOsId);
4008  kmp_affin_mask_t *mask = KMP_CPU_INDEX(affinity.os_id_masks, osId);
4009  KMP_CPU_COPY(mask, sum);
4010  __kmp_topology->at(j).leader = (j == leader);
4011  }
4012  full_mask.include(sum);
4013  unique++;
4014  KMP_CPU_FREE_FROM_STACK(sum);
4015 
4016  // See if the OS Id mask table further restricts or changes the full mask
4017  if (full_mask.restrict_to_mask() && affinity.flags.verbose) {
4018  __kmp_topology->print(env_var);
4019  }
4020 
4021  *numUnique = unique;
4022 }
4023 
4024 // Stuff for the affinity proclist parsers. It's easier to declare these vars
4025 // as file-static than to try and pass them through the calling sequence of
4026 // the recursive-descent OMP_PLACES parser.
4027 static kmp_affin_mask_t *newMasks;
4028 static int numNewMasks;
4029 static int nextNewMask;
4030 
4031 #define ADD_MASK(_mask) \
4032  { \
4033  if (nextNewMask >= numNewMasks) { \
4034  int i; \
4035  numNewMasks *= 2; \
4036  kmp_affin_mask_t *temp; \
4037  KMP_CPU_INTERNAL_ALLOC_ARRAY(temp, numNewMasks); \
4038  for (i = 0; i < numNewMasks / 2; i++) { \
4039  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i); \
4040  kmp_affin_mask_t *dest = KMP_CPU_INDEX(temp, i); \
4041  KMP_CPU_COPY(dest, src); \
4042  } \
4043  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks / 2); \
4044  newMasks = temp; \
4045  } \
4046  KMP_CPU_COPY(KMP_CPU_INDEX(newMasks, nextNewMask), (_mask)); \
4047  nextNewMask++; \
4048  }
4049 
4050 #define ADD_MASK_OSID(_osId, _osId2Mask, _maxOsId) \
4051  { \
4052  if (((_osId) > _maxOsId) || \
4053  (!KMP_CPU_ISSET((_osId), KMP_CPU_INDEX((_osId2Mask), (_osId))))) { \
4054  KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, _osId); \
4055  } else { \
4056  ADD_MASK(KMP_CPU_INDEX(_osId2Mask, (_osId))); \
4057  } \
4058  }
4059 
4060 // Re-parse the proclist (for the explicit affinity type), and form the list
4061 // of affinity newMasks indexed by gtid.
4062 static void __kmp_affinity_process_proclist(kmp_affinity_t &affinity) {
4063  int i;
4064  kmp_affin_mask_t **out_masks = &affinity.masks;
4065  unsigned *out_numMasks = &affinity.num_masks;
4066  const char *proclist = affinity.proclist;
4067  kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
4068  int maxOsId = affinity.num_os_id_masks - 1;
4069  const char *scan = proclist;
4070  const char *next = proclist;
4071 
4072  // We use malloc() for the temporary mask vector, so that we can use
4073  // realloc() to extend it.
4074  numNewMasks = 2;
4075  KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
4076  nextNewMask = 0;
4077  kmp_affin_mask_t *sumMask;
4078  KMP_CPU_ALLOC(sumMask);
4079  int setSize = 0;
4080 
4081  for (;;) {
4082  int start, end, stride;
4083 
4084  SKIP_WS(scan);
4085  next = scan;
4086  if (*next == '\0') {
4087  break;
4088  }
4089 
4090  if (*next == '{') {
4091  int num;
4092  setSize = 0;
4093  next++; // skip '{'
4094  SKIP_WS(next);
4095  scan = next;
4096 
4097  // Read the first integer in the set.
4098  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad proclist");
4099  SKIP_DIGITS(next);
4100  num = __kmp_str_to_int(scan, *next);
4101  KMP_ASSERT2(num >= 0, "bad explicit proc list");
4102 
4103  // Copy the mask for that osId to the sum (union) mask.
4104  if ((num > maxOsId) ||
4105  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
4106  KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, num);
4107  KMP_CPU_ZERO(sumMask);
4108  } else {
4109  KMP_CPU_COPY(sumMask, KMP_CPU_INDEX(osId2Mask, num));
4110  setSize = 1;
4111  }
4112 
4113  for (;;) {
4114  // Check for end of set.
4115  SKIP_WS(next);
4116  if (*next == '}') {
4117  next++; // skip '}'
4118  break;
4119  }
4120 
4121  // Skip optional comma.
4122  if (*next == ',') {
4123  next++;
4124  }
4125  SKIP_WS(next);
4126 
4127  // Read the next integer in the set.
4128  scan = next;
4129  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
4130 
4131  SKIP_DIGITS(next);
4132  num = __kmp_str_to_int(scan, *next);
4133  KMP_ASSERT2(num >= 0, "bad explicit proc list");
4134 
4135  // Add the mask for that osId to the sum mask.
4136  if ((num > maxOsId) ||
4137  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
4138  KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, num);
4139  } else {
4140  KMP_CPU_UNION(sumMask, KMP_CPU_INDEX(osId2Mask, num));
4141  setSize++;
4142  }
4143  }
4144  if (setSize > 0) {
4145  ADD_MASK(sumMask);
4146  }
4147 
4148  SKIP_WS(next);
4149  if (*next == ',') {
4150  next++;
4151  }
4152  scan = next;
4153  continue;
4154  }
4155 
4156  // Read the first integer.
4157  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
4158  SKIP_DIGITS(next);
4159  start = __kmp_str_to_int(scan, *next);
4160  KMP_ASSERT2(start >= 0, "bad explicit proc list");
4161  SKIP_WS(next);
4162 
4163  // If this isn't a range, then add a mask to the list and go on.
4164  if (*next != '-') {
4165  ADD_MASK_OSID(start, osId2Mask, maxOsId);
4166 
4167  // Skip optional comma.
4168  if (*next == ',') {
4169  next++;
4170  }
4171  scan = next;
4172  continue;
4173  }
4174 
4175  // This is a range. Skip over the '-' and read in the 2nd int.
4176  next++; // skip '-'
4177  SKIP_WS(next);
4178  scan = next;
4179  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
4180  SKIP_DIGITS(next);
4181  end = __kmp_str_to_int(scan, *next);
4182  KMP_ASSERT2(end >= 0, "bad explicit proc list");
4183 
4184  // Check for a stride parameter
4185  stride = 1;
4186  SKIP_WS(next);
4187  if (*next == ':') {
4188  // A stride is specified. Skip over the ':" and read the 3rd int.
4189  int sign = +1;
4190  next++; // skip ':'
4191  SKIP_WS(next);
4192  scan = next;
4193  if (*next == '-') {
4194  sign = -1;
4195  next++;
4196  SKIP_WS(next);
4197  scan = next;
4198  }
4199  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
4200  SKIP_DIGITS(next);
4201  stride = __kmp_str_to_int(scan, *next);
4202  KMP_ASSERT2(stride >= 0, "bad explicit proc list");
4203  stride *= sign;
4204  }
4205 
4206  // Do some range checks.
4207  KMP_ASSERT2(stride != 0, "bad explicit proc list");
4208  if (stride > 0) {
4209  KMP_ASSERT2(start <= end, "bad explicit proc list");
4210  } else {
4211  KMP_ASSERT2(start >= end, "bad explicit proc list");
4212  }
4213  KMP_ASSERT2((end - start) / stride <= 65536, "bad explicit proc list");
4214 
4215  // Add the mask for each OS proc # to the list.
4216  if (stride > 0) {
4217  do {
4218  ADD_MASK_OSID(start, osId2Mask, maxOsId);
4219  start += stride;
4220  } while (start <= end);
4221  } else {
4222  do {
4223  ADD_MASK_OSID(start, osId2Mask, maxOsId);
4224  start += stride;
4225  } while (start >= end);
4226  }
4227 
4228  // Skip optional comma.
4229  SKIP_WS(next);
4230  if (*next == ',') {
4231  next++;
4232  }
4233  scan = next;
4234  }
4235 
4236  *out_numMasks = nextNewMask;
4237  if (nextNewMask == 0) {
4238  *out_masks = NULL;
4239  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
4240  return;
4241  }
4242  KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
4243  for (i = 0; i < nextNewMask; i++) {
4244  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
4245  kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
4246  KMP_CPU_COPY(dest, src);
4247  }
4248  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
4249  KMP_CPU_FREE(sumMask);
4250 }
4251 
4252 /*-----------------------------------------------------------------------------
4253 Re-parse the OMP_PLACES proc id list, forming the newMasks for the different
4254 places. Again, Here is the grammar:
4255 
4256 place_list := place
4257 place_list := place , place_list
4258 place := num
4259 place := place : num
4260 place := place : num : signed
4261 place := { subplacelist }
4262 place := ! place // (lowest priority)
4263 subplace_list := subplace
4264 subplace_list := subplace , subplace_list
4265 subplace := num
4266 subplace := num : num
4267 subplace := num : num : signed
4268 signed := num
4269 signed := + signed
4270 signed := - signed
4271 -----------------------------------------------------------------------------*/
4272 static void __kmp_process_subplace_list(const char **scan,
4273  kmp_affinity_t &affinity, int maxOsId,
4274  kmp_affin_mask_t *tempMask,
4275  int *setSize) {
4276  const char *next;
4277  kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
4278 
4279  for (;;) {
4280  int start, count, stride, i;
4281 
4282  // Read in the starting proc id
4283  SKIP_WS(*scan);
4284  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
4285  next = *scan;
4286  SKIP_DIGITS(next);
4287  start = __kmp_str_to_int(*scan, *next);
4288  KMP_ASSERT(start >= 0);
4289  *scan = next;
4290 
4291  // valid follow sets are ',' ':' and '}'
4292  SKIP_WS(*scan);
4293  if (**scan == '}' || **scan == ',') {
4294  if ((start > maxOsId) ||
4295  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
4296  KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, start);
4297  } else {
4298  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
4299  (*setSize)++;
4300  }
4301  if (**scan == '}') {
4302  break;
4303  }
4304  (*scan)++; // skip ','
4305  continue;
4306  }
4307  KMP_ASSERT2(**scan == ':', "bad explicit places list");
4308  (*scan)++; // skip ':'
4309 
4310  // Read count parameter
4311  SKIP_WS(*scan);
4312  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
4313  next = *scan;
4314  SKIP_DIGITS(next);
4315  count = __kmp_str_to_int(*scan, *next);
4316  KMP_ASSERT(count >= 0);
4317  *scan = next;
4318 
4319  // valid follow sets are ',' ':' and '}'
4320  SKIP_WS(*scan);
4321  if (**scan == '}' || **scan == ',') {
4322  for (i = 0; i < count; i++) {
4323  if ((start > maxOsId) ||
4324  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
4325  KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, start);
4326  break; // don't proliferate warnings for large count
4327  } else {
4328  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
4329  start++;
4330  (*setSize)++;
4331  }
4332  }
4333  if (**scan == '}') {
4334  break;
4335  }
4336  (*scan)++; // skip ','
4337  continue;
4338  }
4339  KMP_ASSERT2(**scan == ':', "bad explicit places list");
4340  (*scan)++; // skip ':'
4341 
4342  // Read stride parameter
4343  int sign = +1;
4344  for (;;) {
4345  SKIP_WS(*scan);
4346  if (**scan == '+') {
4347  (*scan)++; // skip '+'
4348  continue;
4349  }
4350  if (**scan == '-') {
4351  sign *= -1;
4352  (*scan)++; // skip '-'
4353  continue;
4354  }
4355  break;
4356  }
4357  SKIP_WS(*scan);
4358  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
4359  next = *scan;
4360  SKIP_DIGITS(next);
4361  stride = __kmp_str_to_int(*scan, *next);
4362  KMP_ASSERT(stride >= 0);
4363  *scan = next;
4364  stride *= sign;
4365 
4366  // valid follow sets are ',' and '}'
4367  SKIP_WS(*scan);
4368  if (**scan == '}' || **scan == ',') {
4369  for (i = 0; i < count; i++) {
4370  if ((start > maxOsId) ||
4371  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
4372  KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, start);
4373  break; // don't proliferate warnings for large count
4374  } else {
4375  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
4376  start += stride;
4377  (*setSize)++;
4378  }
4379  }
4380  if (**scan == '}') {
4381  break;
4382  }
4383  (*scan)++; // skip ','
4384  continue;
4385  }
4386 
4387  KMP_ASSERT2(0, "bad explicit places list");
4388  }
4389 }
4390 
4391 static void __kmp_process_place(const char **scan, kmp_affinity_t &affinity,
4392  int maxOsId, kmp_affin_mask_t *tempMask,
4393  int *setSize) {
4394  const char *next;
4395  kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
4396 
4397  // valid follow sets are '{' '!' and num
4398  SKIP_WS(*scan);
4399  if (**scan == '{') {
4400  (*scan)++; // skip '{'
4401  __kmp_process_subplace_list(scan, affinity, maxOsId, tempMask, setSize);
4402  KMP_ASSERT2(**scan == '}', "bad explicit places list");
4403  (*scan)++; // skip '}'
4404  } else if (**scan == '!') {
4405  (*scan)++; // skip '!'
4406  __kmp_process_place(scan, affinity, maxOsId, tempMask, setSize);
4407  KMP_CPU_COMPLEMENT(maxOsId, tempMask);
4408  } else if ((**scan >= '0') && (**scan <= '9')) {
4409  next = *scan;
4410  SKIP_DIGITS(next);
4411  int num = __kmp_str_to_int(*scan, *next);
4412  KMP_ASSERT(num >= 0);
4413  if ((num > maxOsId) ||
4414  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
4415  KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, num);
4416  } else {
4417  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, num));
4418  (*setSize)++;
4419  }
4420  *scan = next; // skip num
4421  } else {
4422  KMP_ASSERT2(0, "bad explicit places list");
4423  }
4424 }
4425 
4426 // static void
4427 void __kmp_affinity_process_placelist(kmp_affinity_t &affinity) {
4428  int i, j, count, stride, sign;
4429  kmp_affin_mask_t **out_masks = &affinity.masks;
4430  unsigned *out_numMasks = &affinity.num_masks;
4431  const char *placelist = affinity.proclist;
4432  kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
4433  int maxOsId = affinity.num_os_id_masks - 1;
4434  const char *scan = placelist;
4435  const char *next = placelist;
4436 
4437  numNewMasks = 2;
4438  KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
4439  nextNewMask = 0;
4440 
4441  // tempMask is modified based on the previous or initial
4442  // place to form the current place
4443  // previousMask contains the previous place
4444  kmp_affin_mask_t *tempMask;
4445  kmp_affin_mask_t *previousMask;
4446  KMP_CPU_ALLOC(tempMask);
4447  KMP_CPU_ZERO(tempMask);
4448  KMP_CPU_ALLOC(previousMask);
4449  KMP_CPU_ZERO(previousMask);
4450  int setSize = 0;
4451 
4452  for (;;) {
4453  __kmp_process_place(&scan, affinity, maxOsId, tempMask, &setSize);
4454 
4455  // valid follow sets are ',' ':' and EOL
4456  SKIP_WS(scan);
4457  if (*scan == '\0' || *scan == ',') {
4458  if (setSize > 0) {
4459  ADD_MASK(tempMask);
4460  }
4461  KMP_CPU_ZERO(tempMask);
4462  setSize = 0;
4463  if (*scan == '\0') {
4464  break;
4465  }
4466  scan++; // skip ','
4467  continue;
4468  }
4469 
4470  KMP_ASSERT2(*scan == ':', "bad explicit places list");
4471  scan++; // skip ':'
4472 
4473  // Read count parameter
4474  SKIP_WS(scan);
4475  KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
4476  next = scan;
4477  SKIP_DIGITS(next);
4478  count = __kmp_str_to_int(scan, *next);
4479  KMP_ASSERT(count >= 0);
4480  scan = next;
4481 
4482  // valid follow sets are ',' ':' and EOL
4483  SKIP_WS(scan);
4484  if (*scan == '\0' || *scan == ',') {
4485  stride = +1;
4486  } else {
4487  KMP_ASSERT2(*scan == ':', "bad explicit places list");
4488  scan++; // skip ':'
4489 
4490  // Read stride parameter
4491  sign = +1;
4492  for (;;) {
4493  SKIP_WS(scan);
4494  if (*scan == '+') {
4495  scan++; // skip '+'
4496  continue;
4497  }
4498  if (*scan == '-') {
4499  sign *= -1;
4500  scan++; // skip '-'
4501  continue;
4502  }
4503  break;
4504  }
4505  SKIP_WS(scan);
4506  KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
4507  next = scan;
4508  SKIP_DIGITS(next);
4509  stride = __kmp_str_to_int(scan, *next);
4510  KMP_DEBUG_ASSERT(stride >= 0);
4511  scan = next;
4512  stride *= sign;
4513  }
4514 
4515  // Add places determined by initial_place : count : stride
4516  for (i = 0; i < count; i++) {
4517  if (setSize == 0) {
4518  break;
4519  }
4520  // Add the current place, then build the next place (tempMask) from that
4521  KMP_CPU_COPY(previousMask, tempMask);
4522  ADD_MASK(previousMask);
4523  KMP_CPU_ZERO(tempMask);
4524  setSize = 0;
4525  KMP_CPU_SET_ITERATE(j, previousMask) {
4526  if (!KMP_CPU_ISSET(j, previousMask)) {
4527  continue;
4528  }
4529  if ((j + stride > maxOsId) || (j + stride < 0) ||
4530  (!KMP_CPU_ISSET(j, __kmp_affin_fullMask)) ||
4531  (!KMP_CPU_ISSET(j + stride,
4532  KMP_CPU_INDEX(osId2Mask, j + stride)))) {
4533  if (i < count - 1) {
4534  KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, j + stride);
4535  }
4536  continue;
4537  }
4538  KMP_CPU_SET(j + stride, tempMask);
4539  setSize++;
4540  }
4541  }
4542  KMP_CPU_ZERO(tempMask);
4543  setSize = 0;
4544 
4545  // valid follow sets are ',' and EOL
4546  SKIP_WS(scan);
4547  if (*scan == '\0') {
4548  break;
4549  }
4550  if (*scan == ',') {
4551  scan++; // skip ','
4552  continue;
4553  }
4554 
4555  KMP_ASSERT2(0, "bad explicit places list");
4556  }
4557 
4558  *out_numMasks = nextNewMask;
4559  if (nextNewMask == 0) {
4560  *out_masks = NULL;
4561  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
4562  return;
4563  }
4564  KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
4565  KMP_CPU_FREE(tempMask);
4566  KMP_CPU_FREE(previousMask);
4567  for (i = 0; i < nextNewMask; i++) {
4568  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
4569  kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
4570  KMP_CPU_COPY(dest, src);
4571  }
4572  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
4573 }
4574 
4575 #undef ADD_MASK
4576 #undef ADD_MASK_OSID
4577 
4578 // This function figures out the deepest level at which there is at least one
4579 // cluster/core with more than one processing unit bound to it.
4580 static int __kmp_affinity_find_core_level(int nprocs, int bottom_level) {
4581  int core_level = 0;
4582 
4583  for (int i = 0; i < nprocs; i++) {
4584  const kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
4585  for (int j = bottom_level; j > 0; j--) {
4586  if (hw_thread.ids[j] > 0) {
4587  if (core_level < (j - 1)) {
4588  core_level = j - 1;
4589  }
4590  }
4591  }
4592  }
4593  return core_level;
4594 }
4595 
4596 // This function counts number of clusters/cores at given level.
4597 static int __kmp_affinity_compute_ncores(int nprocs, int bottom_level,
4598  int core_level) {
4599  return __kmp_topology->get_count(core_level);
4600 }
4601 // This function finds to which cluster/core given processing unit is bound.
4602 static int __kmp_affinity_find_core(int proc, int bottom_level,
4603  int core_level) {
4604  int core = 0;
4605  KMP_DEBUG_ASSERT(proc >= 0 && proc < __kmp_topology->get_num_hw_threads());
4606  for (int i = 0; i <= proc; ++i) {
4607  if (i + 1 <= proc) {
4608  for (int j = 0; j <= core_level; ++j) {
4609  if (__kmp_topology->at(i + 1).sub_ids[j] !=
4610  __kmp_topology->at(i).sub_ids[j]) {
4611  core++;
4612  break;
4613  }
4614  }
4615  }
4616  }
4617  return core;
4618 }
4619 
4620 // This function finds maximal number of processing units bound to a
4621 // cluster/core at given level.
4622 static int __kmp_affinity_max_proc_per_core(int nprocs, int bottom_level,
4623  int core_level) {
4624  if (core_level >= bottom_level)
4625  return 1;
4626  int thread_level = __kmp_topology->get_level(KMP_HW_THREAD);
4627  return __kmp_topology->calculate_ratio(thread_level, core_level);
4628 }
4629 
4630 static int *procarr = NULL;
4631 static int __kmp_aff_depth = 0;
4632 static int *__kmp_osid_to_hwthread_map = NULL;
4633 
4634 static void __kmp_affinity_get_mask_topology_info(const kmp_affin_mask_t *mask,
4635  kmp_affinity_ids_t &ids,
4636  kmp_affinity_attrs_t &attrs) {
4637  if (!KMP_AFFINITY_CAPABLE())
4638  return;
4639 
4640  // Initiailze ids and attrs thread data
4641  for (int i = 0; i < KMP_HW_LAST; ++i)
4642  ids.ids[i] = kmp_hw_thread_t::UNKNOWN_ID;
4643  attrs = KMP_AFFINITY_ATTRS_UNKNOWN;
4644 
4645  // Iterate through each os id within the mask and determine
4646  // the topology id and attribute information
4647  int cpu;
4648  int depth = __kmp_topology->get_depth();
4649  KMP_CPU_SET_ITERATE(cpu, mask) {
4650  int osid_idx = __kmp_osid_to_hwthread_map[cpu];
4651  ids.os_id = cpu;
4652  const kmp_hw_thread_t &hw_thread = __kmp_topology->at(osid_idx);
4653  for (int level = 0; level < depth; ++level) {
4654  kmp_hw_t type = __kmp_topology->get_type(level);
4655  int id = hw_thread.sub_ids[level];
4656  if (ids.ids[type] == kmp_hw_thread_t::UNKNOWN_ID || ids.ids[type] == id) {
4657  ids.ids[type] = id;
4658  } else {
4659  // This mask spans across multiple topology units, set it as such
4660  // and mark every level below as such as well.
4661  ids.ids[type] = kmp_hw_thread_t::MULTIPLE_ID;
4662  for (; level < depth; ++level) {
4663  kmp_hw_t type = __kmp_topology->get_type(level);
4664  ids.ids[type] = kmp_hw_thread_t::MULTIPLE_ID;
4665  }
4666  }
4667  }
4668  if (!attrs.valid) {
4669  attrs.core_type = hw_thread.attrs.get_core_type();
4670  attrs.core_eff = hw_thread.attrs.get_core_eff();
4671  attrs.valid = 1;
4672  } else {
4673  // This mask spans across multiple attributes, set it as such
4674  if (attrs.core_type != hw_thread.attrs.get_core_type())
4675  attrs.core_type = KMP_HW_CORE_TYPE_UNKNOWN;
4676  if (attrs.core_eff != hw_thread.attrs.get_core_eff())
4677  attrs.core_eff = kmp_hw_attr_t::UNKNOWN_CORE_EFF;
4678  }
4679  }
4680 }
4681 
4682 static void __kmp_affinity_get_thread_topology_info(kmp_info_t *th) {
4683  if (!KMP_AFFINITY_CAPABLE())
4684  return;
4685  const kmp_affin_mask_t *mask = th->th.th_affin_mask;
4686  kmp_affinity_ids_t &ids = th->th.th_topology_ids;
4687  kmp_affinity_attrs_t &attrs = th->th.th_topology_attrs;
4688  __kmp_affinity_get_mask_topology_info(mask, ids, attrs);
4689 }
4690 
4691 // Assign the topology information to each place in the place list
4692 // A thread can then grab not only its affinity mask, but the topology
4693 // information associated with that mask. e.g., Which socket is a thread on
4694 static void __kmp_affinity_get_topology_info(kmp_affinity_t &affinity) {
4695  if (!KMP_AFFINITY_CAPABLE())
4696  return;
4697  if (affinity.type != affinity_none) {
4698  KMP_ASSERT(affinity.num_os_id_masks);
4699  KMP_ASSERT(affinity.os_id_masks);
4700  }
4701  KMP_ASSERT(affinity.num_masks);
4702  KMP_ASSERT(affinity.masks);
4703  KMP_ASSERT(__kmp_affin_fullMask);
4704 
4705  int max_cpu = __kmp_affin_fullMask->get_max_cpu();
4706  int num_hw_threads = __kmp_topology->get_num_hw_threads();
4707 
4708  // Allocate thread topology information
4709  if (!affinity.ids) {
4710  affinity.ids = (kmp_affinity_ids_t *)__kmp_allocate(
4711  sizeof(kmp_affinity_ids_t) * affinity.num_masks);
4712  }
4713  if (!affinity.attrs) {
4714  affinity.attrs = (kmp_affinity_attrs_t *)__kmp_allocate(
4715  sizeof(kmp_affinity_attrs_t) * affinity.num_masks);
4716  }
4717  if (!__kmp_osid_to_hwthread_map) {
4718  // Want the +1 because max_cpu should be valid index into map
4719  __kmp_osid_to_hwthread_map =
4720  (int *)__kmp_allocate(sizeof(int) * (max_cpu + 1));
4721  }
4722 
4723  // Create the OS proc to hardware thread map
4724  for (int hw_thread = 0; hw_thread < num_hw_threads; ++hw_thread) {
4725  int os_id = __kmp_topology->at(hw_thread).os_id;
4726  if (KMP_CPU_ISSET(os_id, __kmp_affin_fullMask))
4727  __kmp_osid_to_hwthread_map[os_id] = hw_thread;
4728  }
4729 
4730  for (unsigned i = 0; i < affinity.num_masks; ++i) {
4731  kmp_affinity_ids_t &ids = affinity.ids[i];
4732  kmp_affinity_attrs_t &attrs = affinity.attrs[i];
4733  kmp_affin_mask_t *mask = KMP_CPU_INDEX(affinity.masks, i);
4734  __kmp_affinity_get_mask_topology_info(mask, ids, attrs);
4735  }
4736 }
4737 
4738 // Called when __kmp_topology is ready
4739 static void __kmp_aux_affinity_initialize_other_data(kmp_affinity_t &affinity) {
4740  // Initialize other data structures which depend on the topology
4741  if (__kmp_topology && __kmp_topology->get_num_hw_threads()) {
4742  machine_hierarchy.init(__kmp_topology->get_num_hw_threads());
4743  __kmp_affinity_get_topology_info(affinity);
4744 #if KMP_WEIGHTED_ITERATIONS_SUPPORTED
4745  __kmp_first_osid_with_ecore = __kmp_get_first_osid_with_ecore();
4746 #endif
4747  }
4748 }
4749 
4750 // Create a one element mask array (set of places) which only contains the
4751 // initial process's affinity mask
4752 static void __kmp_create_affinity_none_places(kmp_affinity_t &affinity) {
4753  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4754  KMP_ASSERT(affinity.type == affinity_none);
4755  KMP_ASSERT(__kmp_avail_proc == __kmp_topology->get_num_hw_threads());
4756  affinity.num_masks = 1;
4757  KMP_CPU_ALLOC_ARRAY(affinity.masks, affinity.num_masks);
4758  kmp_affin_mask_t *dest = KMP_CPU_INDEX(affinity.masks, 0);
4759  KMP_CPU_COPY(dest, __kmp_affin_fullMask);
4760  __kmp_aux_affinity_initialize_other_data(affinity);
4761 }
4762 
4763 static void __kmp_aux_affinity_initialize_masks(kmp_affinity_t &affinity) {
4764  // Create the "full" mask - this defines all of the processors that we
4765  // consider to be in the machine model. If respect is set, then it is the
4766  // initialization thread's affinity mask. Otherwise, it is all processors that
4767  // we know about on the machine.
4768  int verbose = affinity.flags.verbose;
4769  const char *env_var = affinity.env_var;
4770 
4771  // Already initialized
4772  if (__kmp_affin_fullMask && __kmp_affin_origMask)
4773  return;
4774 
4775  if (__kmp_affin_fullMask == NULL) {
4776  KMP_CPU_ALLOC(__kmp_affin_fullMask);
4777  }
4778  if (__kmp_affin_origMask == NULL) {
4779  KMP_CPU_ALLOC(__kmp_affin_origMask);
4780  }
4781  if (KMP_AFFINITY_CAPABLE()) {
4782  __kmp_get_system_affinity(__kmp_affin_fullMask, TRUE);
4783  // Make a copy before possible expanding to the entire machine mask
4784  __kmp_affin_origMask->copy(__kmp_affin_fullMask);
4785  if (affinity.flags.respect) {
4786  // Count the number of available processors.
4787  unsigned i;
4788  __kmp_avail_proc = 0;
4789  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
4790  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
4791  continue;
4792  }
4793  __kmp_avail_proc++;
4794  }
4795  if (__kmp_avail_proc > __kmp_xproc) {
4796  KMP_AFF_WARNING(affinity, ErrorInitializeAffinity);
4797  affinity.type = affinity_none;
4798  KMP_AFFINITY_DISABLE();
4799  return;
4800  }
4801 
4802  if (verbose) {
4803  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4804  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4805  __kmp_affin_fullMask);
4806  KMP_INFORM(InitOSProcSetRespect, env_var, buf);
4807  }
4808  } else {
4809  if (verbose) {
4810  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4811  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4812  __kmp_affin_fullMask);
4813  KMP_INFORM(InitOSProcSetNotRespect, env_var, buf);
4814  }
4815  __kmp_avail_proc =
4816  __kmp_affinity_entire_machine_mask(__kmp_affin_fullMask);
4817 #if KMP_OS_WINDOWS
4818  if (__kmp_num_proc_groups <= 1) {
4819  // Copy expanded full mask if topology has single processor group
4820  __kmp_affin_origMask->copy(__kmp_affin_fullMask);
4821  }
4822  // Set the process affinity mask since threads' affinity
4823  // masks must be subset of process mask in Windows* OS
4824  __kmp_affin_fullMask->set_process_affinity(true);
4825 #endif
4826  }
4827  }
4828 }
4829 
4830 static bool __kmp_aux_affinity_initialize_topology(kmp_affinity_t &affinity) {
4831  bool success = false;
4832  const char *env_var = affinity.env_var;
4833  kmp_i18n_id_t msg_id = kmp_i18n_null;
4834  int verbose = affinity.flags.verbose;
4835 
4836  // For backward compatibility, setting KMP_CPUINFO_FILE =>
4837  // KMP_TOPOLOGY_METHOD=cpuinfo
4838  if ((__kmp_cpuinfo_file != NULL) &&
4839  (__kmp_affinity_top_method == affinity_top_method_all)) {
4840  __kmp_affinity_top_method = affinity_top_method_cpuinfo;
4841  }
4842 
4843  if (__kmp_affinity_top_method == affinity_top_method_all) {
4844 // In the default code path, errors are not fatal - we just try using
4845 // another method. We only emit a warning message if affinity is on, or the
4846 // verbose flag is set, an the nowarnings flag was not set.
4847 #if KMP_USE_HWLOC
4848  if (!success &&
4849  __kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC) {
4850  if (!__kmp_hwloc_error) {
4851  success = __kmp_affinity_create_hwloc_map(&msg_id);
4852  if (!success && verbose) {
4853  KMP_INFORM(AffIgnoringHwloc, env_var);
4854  }
4855  } else if (verbose) {
4856  KMP_INFORM(AffIgnoringHwloc, env_var);
4857  }
4858  }
4859 #endif
4860 
4861 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4862  if (!success) {
4863  success = __kmp_affinity_create_x2apicid_map(&msg_id);
4864  if (!success && verbose && msg_id != kmp_i18n_null) {
4865  KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4866  }
4867  }
4868  if (!success) {
4869  success = __kmp_affinity_create_apicid_map(&msg_id);
4870  if (!success && verbose && msg_id != kmp_i18n_null) {
4871  KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4872  }
4873  }
4874 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4875 
4876 #if KMP_OS_LINUX || KMP_OS_AIX
4877  if (!success) {
4878  int line = 0;
4879  success = __kmp_affinity_create_cpuinfo_map(&line, &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_OS_LINUX */
4885 
4886 #if KMP_GROUP_AFFINITY
4887  if (!success && (__kmp_num_proc_groups > 1)) {
4888  success = __kmp_affinity_create_proc_group_map(&msg_id);
4889  if (!success && verbose && msg_id != kmp_i18n_null) {
4890  KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4891  }
4892  }
4893 #endif /* KMP_GROUP_AFFINITY */
4894 
4895  if (!success) {
4896  success = __kmp_affinity_create_flat_map(&msg_id);
4897  if (!success && verbose && msg_id != kmp_i18n_null) {
4898  KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4899  }
4900  KMP_ASSERT(success);
4901  }
4902  }
4903 
4904 // If the user has specified that a paricular topology discovery method is to be
4905 // used, then we abort if that method fails. The exception is group affinity,
4906 // which might have been implicitly set.
4907 #if KMP_USE_HWLOC
4908  else if (__kmp_affinity_top_method == affinity_top_method_hwloc) {
4909  KMP_ASSERT(__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC);
4910  success = __kmp_affinity_create_hwloc_map(&msg_id);
4911  if (!success) {
4912  KMP_ASSERT(msg_id != kmp_i18n_null);
4913  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4914  }
4915  }
4916 #endif // KMP_USE_HWLOC
4917 
4918 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4919  else if (__kmp_affinity_top_method == affinity_top_method_x2apicid ||
4920  __kmp_affinity_top_method == affinity_top_method_x2apicid_1f) {
4921  success = __kmp_affinity_create_x2apicid_map(&msg_id);
4922  if (!success) {
4923  KMP_ASSERT(msg_id != kmp_i18n_null);
4924  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4925  }
4926  } else if (__kmp_affinity_top_method == affinity_top_method_apicid) {
4927  success = __kmp_affinity_create_apicid_map(&msg_id);
4928  if (!success) {
4929  KMP_ASSERT(msg_id != kmp_i18n_null);
4930  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4931  }
4932  }
4933 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4934 
4935  else if (__kmp_affinity_top_method == affinity_top_method_cpuinfo) {
4936  int line = 0;
4937  success = __kmp_affinity_create_cpuinfo_map(&line, &msg_id);
4938  if (!success) {
4939  KMP_ASSERT(msg_id != kmp_i18n_null);
4940  const char *filename = __kmp_cpuinfo_get_filename();
4941  if (line > 0) {
4942  KMP_FATAL(FileLineMsgExiting, filename, line,
4943  __kmp_i18n_catgets(msg_id));
4944  } else {
4945  KMP_FATAL(FileMsgExiting, filename, __kmp_i18n_catgets(msg_id));
4946  }
4947  }
4948  }
4949 
4950 #if KMP_GROUP_AFFINITY
4951  else if (__kmp_affinity_top_method == affinity_top_method_group) {
4952  success = __kmp_affinity_create_proc_group_map(&msg_id);
4953  KMP_ASSERT(success);
4954  if (!success) {
4955  KMP_ASSERT(msg_id != kmp_i18n_null);
4956  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4957  }
4958  }
4959 #endif /* KMP_GROUP_AFFINITY */
4960 
4961  else if (__kmp_affinity_top_method == affinity_top_method_flat) {
4962  success = __kmp_affinity_create_flat_map(&msg_id);
4963  // should not fail
4964  KMP_ASSERT(success);
4965  }
4966 
4967  // Early exit if topology could not be created
4968  if (!__kmp_topology) {
4969  if (KMP_AFFINITY_CAPABLE()) {
4970  KMP_AFF_WARNING(affinity, ErrorInitializeAffinity);
4971  }
4972  if (nPackages > 0 && nCoresPerPkg > 0 && __kmp_nThreadsPerCore > 0 &&
4973  __kmp_ncores > 0) {
4974  __kmp_topology = kmp_topology_t::allocate(0, 0, NULL);
4975  __kmp_topology->canonicalize(nPackages, nCoresPerPkg,
4976  __kmp_nThreadsPerCore, __kmp_ncores);
4977  if (verbose) {
4978  __kmp_topology->print(env_var);
4979  }
4980  }
4981  return false;
4982  }
4983 
4984  // Canonicalize, print (if requested), apply KMP_HW_SUBSET
4985  __kmp_topology->canonicalize();
4986  if (verbose)
4987  __kmp_topology->print(env_var);
4988  bool filtered = __kmp_topology->filter_hw_subset();
4989  if (filtered && verbose)
4990  __kmp_topology->print("KMP_HW_SUBSET");
4991  return success;
4992 }
4993 
4994 static void __kmp_aux_affinity_initialize(kmp_affinity_t &affinity) {
4995  bool is_regular_affinity = (&affinity == &__kmp_affinity);
4996  bool is_hidden_helper_affinity = (&affinity == &__kmp_hh_affinity);
4997  const char *env_var = __kmp_get_affinity_env_var(affinity);
4998 
4999  if (affinity.flags.initialized) {
5000  KMP_ASSERT(__kmp_affin_fullMask != NULL);
5001  return;
5002  }
5003 
5004  if (is_regular_affinity && (!__kmp_affin_fullMask || !__kmp_affin_origMask))
5005  __kmp_aux_affinity_initialize_masks(affinity);
5006 
5007  if (is_regular_affinity && !__kmp_topology) {
5008  bool success = __kmp_aux_affinity_initialize_topology(affinity);
5009  if (success) {
5010  KMP_ASSERT(__kmp_avail_proc == __kmp_topology->get_num_hw_threads());
5011  } else {
5012  affinity.type = affinity_none;
5013  KMP_AFFINITY_DISABLE();
5014  }
5015  }
5016 
5017  // If KMP_AFFINITY=none, then only create the single "none" place
5018  // which is the process's initial affinity mask or the number of
5019  // hardware threads depending on respect,norespect
5020  if (affinity.type == affinity_none) {
5021  __kmp_create_affinity_none_places(affinity);
5022 #if KMP_USE_HIER_SCHED
5023  __kmp_dispatch_set_hierarchy_values();
5024 #endif
5025  affinity.flags.initialized = TRUE;
5026  return;
5027  }
5028 
5029  __kmp_topology->set_granularity(affinity);
5030  int depth = __kmp_topology->get_depth();
5031 
5032  // Create the table of masks, indexed by thread Id.
5033  unsigned numUnique = 0;
5034  int numAddrs = __kmp_topology->get_num_hw_threads();
5035  // If OMP_PLACES=cores:<attribute> specified, then attempt
5036  // to make OS Id mask table using those attributes
5037  if (affinity.core_attr_gran.valid) {
5038  __kmp_create_os_id_masks(&numUnique, affinity, [&](int idx) {
5039  KMP_ASSERT(idx >= -1);
5040  for (int i = idx + 1; i < numAddrs; ++i)
5041  if (__kmp_topology->at(i).attrs.contains(affinity.core_attr_gran))
5042  return i;
5043  return numAddrs;
5044  });
5045  if (!affinity.os_id_masks) {
5046  const char *core_attribute;
5047  if (affinity.core_attr_gran.core_eff != kmp_hw_attr_t::UNKNOWN_CORE_EFF)
5048  core_attribute = "core_efficiency";
5049  else
5050  core_attribute = "core_type";
5051  KMP_AFF_WARNING(affinity, AffIgnoringNotAvailable, env_var,
5052  core_attribute,
5053  __kmp_hw_get_catalog_string(KMP_HW_CORE, /*plural=*/true))
5054  }
5055  }
5056  // If core attributes did not work, or none were specified,
5057  // then make OS Id mask table using typical incremental way with
5058  // checking for validity of each id at granularity level specified.
5059  if (!affinity.os_id_masks) {
5060  int gran = affinity.gran_levels;
5061  int gran_level = depth - 1 - affinity.gran_levels;
5062  if (gran >= 0 && gran_level >= 0 && gran_level < depth) {
5063  __kmp_create_os_id_masks(
5064  &numUnique, affinity, [depth, numAddrs, &affinity](int idx) {
5065  KMP_ASSERT(idx >= -1);
5066  int gran = affinity.gran_levels;
5067  int gran_level = depth - 1 - affinity.gran_levels;
5068  for (int i = idx + 1; i < numAddrs; ++i)
5069  if ((gran >= depth) ||
5070  (gran < depth && __kmp_topology->at(i).ids[gran_level] !=
5071  kmp_hw_thread_t::UNKNOWN_ID))
5072  return i;
5073  return numAddrs;
5074  });
5075  }
5076  }
5077  // Final attempt to make OS Id mask table using typical incremental way.
5078  if (!affinity.os_id_masks) {
5079  __kmp_create_os_id_masks(&numUnique, affinity, [](int idx) {
5080  KMP_ASSERT(idx >= -1);
5081  return idx + 1;
5082  });
5083  }
5084 
5085  switch (affinity.type) {
5086 
5087  case affinity_explicit:
5088  KMP_DEBUG_ASSERT(affinity.proclist != NULL);
5089  if (is_hidden_helper_affinity ||
5090  __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) {
5091  __kmp_affinity_process_proclist(affinity);
5092  } else {
5093  __kmp_affinity_process_placelist(affinity);
5094  }
5095  if (affinity.num_masks == 0) {
5096  KMP_AFF_WARNING(affinity, AffNoValidProcID);
5097  affinity.type = affinity_none;
5098  __kmp_create_affinity_none_places(affinity);
5099  affinity.flags.initialized = TRUE;
5100  return;
5101  }
5102  break;
5103 
5104  // The other affinity types rely on sorting the hardware threads according to
5105  // some permutation of the machine topology tree. Set affinity.compact
5106  // and affinity.offset appropriately, then jump to a common code
5107  // fragment to do the sort and create the array of affinity masks.
5108  case affinity_logical:
5109  affinity.compact = 0;
5110  if (affinity.offset) {
5111  affinity.offset =
5112  __kmp_nThreadsPerCore * affinity.offset % __kmp_avail_proc;
5113  }
5114  goto sortTopology;
5115 
5116  case affinity_physical:
5117  if (__kmp_nThreadsPerCore > 1) {
5118  affinity.compact = 1;
5119  if (affinity.compact >= depth) {
5120  affinity.compact = 0;
5121  }
5122  } else {
5123  affinity.compact = 0;
5124  }
5125  if (affinity.offset) {
5126  affinity.offset =
5127  __kmp_nThreadsPerCore * affinity.offset % __kmp_avail_proc;
5128  }
5129  goto sortTopology;
5130 
5131  case affinity_scatter:
5132  if (affinity.compact >= depth) {
5133  affinity.compact = 0;
5134  } else {
5135  affinity.compact = depth - 1 - affinity.compact;
5136  }
5137  goto sortTopology;
5138 
5139  case affinity_compact:
5140  if (affinity.compact >= depth) {
5141  affinity.compact = depth - 1;
5142  }
5143  goto sortTopology;
5144 
5145  case affinity_balanced:
5146  if (depth <= 1 || is_hidden_helper_affinity) {
5147  KMP_AFF_WARNING(affinity, AffBalancedNotAvail, env_var);
5148  affinity.type = affinity_none;
5149  __kmp_create_affinity_none_places(affinity);
5150  affinity.flags.initialized = TRUE;
5151  return;
5152  } else if (!__kmp_topology->is_uniform()) {
5153  // Save the depth for further usage
5154  __kmp_aff_depth = depth;
5155 
5156  int core_level =
5157  __kmp_affinity_find_core_level(__kmp_avail_proc, depth - 1);
5158  int ncores = __kmp_affinity_compute_ncores(__kmp_avail_proc, depth - 1,
5159  core_level);
5160  int maxprocpercore = __kmp_affinity_max_proc_per_core(
5161  __kmp_avail_proc, depth - 1, core_level);
5162 
5163  int nproc = ncores * maxprocpercore;
5164  if ((nproc < 2) || (nproc < __kmp_avail_proc)) {
5165  KMP_AFF_WARNING(affinity, AffBalancedNotAvail, env_var);
5166  affinity.type = affinity_none;
5167  __kmp_create_affinity_none_places(affinity);
5168  affinity.flags.initialized = TRUE;
5169  return;
5170  }
5171 
5172  procarr = (int *)__kmp_allocate(sizeof(int) * nproc);
5173  for (int i = 0; i < nproc; i++) {
5174  procarr[i] = -1;
5175  }
5176 
5177  int lastcore = -1;
5178  int inlastcore = 0;
5179  for (int i = 0; i < __kmp_avail_proc; i++) {
5180  int proc = __kmp_topology->at(i).os_id;
5181  int core = __kmp_affinity_find_core(i, depth - 1, core_level);
5182 
5183  if (core == lastcore) {
5184  inlastcore++;
5185  } else {
5186  inlastcore = 0;
5187  }
5188  lastcore = core;
5189 
5190  procarr[core * maxprocpercore + inlastcore] = proc;
5191  }
5192  }
5193  if (affinity.compact >= depth) {
5194  affinity.compact = depth - 1;
5195  }
5196 
5197  sortTopology:
5198  // Allocate the gtid->affinity mask table.
5199  if (affinity.flags.dups) {
5200  affinity.num_masks = __kmp_avail_proc;
5201  } else {
5202  affinity.num_masks = numUnique;
5203  }
5204 
5205  if ((__kmp_nested_proc_bind.bind_types[0] != proc_bind_intel) &&
5206  (__kmp_affinity_num_places > 0) &&
5207  ((unsigned)__kmp_affinity_num_places < affinity.num_masks) &&
5208  !is_hidden_helper_affinity) {
5209  affinity.num_masks = __kmp_affinity_num_places;
5210  }
5211 
5212  KMP_CPU_ALLOC_ARRAY(affinity.masks, affinity.num_masks);
5213 
5214  // Sort the topology table according to the current setting of
5215  // affinity.compact, then fill out affinity.masks.
5216  __kmp_topology->sort_compact(affinity);
5217  {
5218  int i;
5219  unsigned j;
5220  int num_hw_threads = __kmp_topology->get_num_hw_threads();
5221  kmp_full_mask_modifier_t full_mask;
5222  for (i = 0, j = 0; i < num_hw_threads; i++) {
5223  if ((!affinity.flags.dups) && (!__kmp_topology->at(i).leader)) {
5224  continue;
5225  }
5226  int osId = __kmp_topology->at(i).os_id;
5227 
5228  kmp_affin_mask_t *src = KMP_CPU_INDEX(affinity.os_id_masks, osId);
5229  if (KMP_CPU_ISEMPTY(src))
5230  continue;
5231  kmp_affin_mask_t *dest = KMP_CPU_INDEX(affinity.masks, j);
5232  KMP_ASSERT(KMP_CPU_ISSET(osId, src));
5233  KMP_CPU_COPY(dest, src);
5234  full_mask.include(src);
5235  if (++j >= affinity.num_masks) {
5236  break;
5237  }
5238  }
5239  KMP_DEBUG_ASSERT(j == affinity.num_masks);
5240  // See if the places list further restricts or changes the full mask
5241  if (full_mask.restrict_to_mask() && affinity.flags.verbose) {
5242  __kmp_topology->print(env_var);
5243  }
5244  }
5245  // Sort the topology back using ids
5246  __kmp_topology->sort_ids();
5247  break;
5248 
5249  default:
5250  KMP_ASSERT2(0, "Unexpected affinity setting");
5251  }
5252  __kmp_aux_affinity_initialize_other_data(affinity);
5253  affinity.flags.initialized = TRUE;
5254 }
5255 
5256 void __kmp_affinity_initialize(kmp_affinity_t &affinity) {
5257  // Much of the code above was written assuming that if a machine was not
5258  // affinity capable, then affinity type == affinity_none.
5259  // We now explicitly represent this as affinity type == affinity_disabled.
5260  // There are too many checks for affinity type == affinity_none in this code.
5261  // Instead of trying to change them all, check if
5262  // affinity type == affinity_disabled, and if so, slam it with affinity_none,
5263  // call the real initialization routine, then restore affinity type to
5264  // affinity_disabled.
5265  int disabled = (affinity.type == affinity_disabled);
5266  if (!KMP_AFFINITY_CAPABLE())
5267  KMP_ASSERT(disabled);
5268  if (disabled)
5269  affinity.type = affinity_none;
5270  __kmp_aux_affinity_initialize(affinity);
5271  if (disabled)
5272  affinity.type = affinity_disabled;
5273 }
5274 
5275 void __kmp_affinity_uninitialize(void) {
5276  for (kmp_affinity_t *affinity : __kmp_affinities) {
5277  if (affinity->masks != NULL)
5278  KMP_CPU_FREE_ARRAY(affinity->masks, affinity->num_masks);
5279  if (affinity->os_id_masks != NULL)
5280  KMP_CPU_FREE_ARRAY(affinity->os_id_masks, affinity->num_os_id_masks);
5281  if (affinity->proclist != NULL)
5282  __kmp_free(affinity->proclist);
5283  if (affinity->ids != NULL)
5284  __kmp_free(affinity->ids);
5285  if (affinity->attrs != NULL)
5286  __kmp_free(affinity->attrs);
5287  *affinity = KMP_AFFINITY_INIT(affinity->env_var);
5288  }
5289  if (__kmp_affin_origMask != NULL) {
5290  if (KMP_AFFINITY_CAPABLE()) {
5291 #if KMP_OS_AIX
5292  // Uninitialize by unbinding the thread.
5293  bindprocessor(BINDTHREAD, thread_self(), PROCESSOR_CLASS_ANY);
5294 #else
5295  __kmp_set_system_affinity(__kmp_affin_origMask, FALSE);
5296 #endif
5297  }
5298  KMP_CPU_FREE(__kmp_affin_origMask);
5299  __kmp_affin_origMask = NULL;
5300  }
5301  __kmp_affinity_num_places = 0;
5302  if (procarr != NULL) {
5303  __kmp_free(procarr);
5304  procarr = NULL;
5305  }
5306  if (__kmp_osid_to_hwthread_map) {
5307  __kmp_free(__kmp_osid_to_hwthread_map);
5308  __kmp_osid_to_hwthread_map = NULL;
5309  }
5310 #if KMP_USE_HWLOC
5311  if (__kmp_hwloc_topology != NULL) {
5312  hwloc_topology_destroy(__kmp_hwloc_topology);
5313  __kmp_hwloc_topology = NULL;
5314  }
5315 #endif
5316  if (__kmp_hw_subset) {
5317  kmp_hw_subset_t::deallocate(__kmp_hw_subset);
5318  __kmp_hw_subset = nullptr;
5319  }
5320  if (__kmp_topology) {
5321  kmp_topology_t::deallocate(__kmp_topology);
5322  __kmp_topology = nullptr;
5323  }
5324  KMPAffinity::destroy_api();
5325 }
5326 
5327 static void __kmp_select_mask_by_gtid(int gtid, const kmp_affinity_t *affinity,
5328  int *place, kmp_affin_mask_t **mask) {
5329  int mask_idx;
5330  bool is_hidden_helper = KMP_HIDDEN_HELPER_THREAD(gtid);
5331  if (is_hidden_helper)
5332  // The first gtid is the regular primary thread, the second gtid is the main
5333  // thread of hidden team which does not participate in task execution.
5334  mask_idx = gtid - 2;
5335  else
5336  mask_idx = __kmp_adjust_gtid_for_hidden_helpers(gtid);
5337  KMP_DEBUG_ASSERT(affinity->num_masks > 0);
5338  *place = (mask_idx + affinity->offset) % affinity->num_masks;
5339  *mask = KMP_CPU_INDEX(affinity->masks, *place);
5340 }
5341 
5342 // This function initializes the per-thread data concerning affinity including
5343 // the mask and topology information
5344 void __kmp_affinity_set_init_mask(int gtid, int isa_root) {
5345 
5346  kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
5347 
5348  // Set the thread topology information to default of unknown
5349  for (int id = 0; id < KMP_HW_LAST; ++id)
5350  th->th.th_topology_ids.ids[id] = kmp_hw_thread_t::UNKNOWN_ID;
5351  th->th.th_topology_attrs = KMP_AFFINITY_ATTRS_UNKNOWN;
5352 
5353  if (!KMP_AFFINITY_CAPABLE()) {
5354  return;
5355  }
5356 
5357  if (th->th.th_affin_mask == NULL) {
5358  KMP_CPU_ALLOC(th->th.th_affin_mask);
5359  } else {
5360  KMP_CPU_ZERO(th->th.th_affin_mask);
5361  }
5362 
5363  // Copy the thread mask to the kmp_info_t structure. If
5364  // __kmp_affinity.type == affinity_none, copy the "full" mask, i.e.
5365  // one that has all of the OS proc ids set, or if
5366  // __kmp_affinity.flags.respect is set, then the full mask is the
5367  // same as the mask of the initialization thread.
5368  kmp_affin_mask_t *mask;
5369  int i;
5370  const kmp_affinity_t *affinity;
5371  bool is_hidden_helper = KMP_HIDDEN_HELPER_THREAD(gtid);
5372 
5373  if (is_hidden_helper)
5374  affinity = &__kmp_hh_affinity;
5375  else
5376  affinity = &__kmp_affinity;
5377 
5378  if (KMP_AFFINITY_NON_PROC_BIND || is_hidden_helper) {
5379  if ((affinity->type == affinity_none) ||
5380  (affinity->type == affinity_balanced) ||
5381  KMP_HIDDEN_HELPER_MAIN_THREAD(gtid)) {
5382 #if KMP_GROUP_AFFINITY
5383  if (__kmp_num_proc_groups > 1) {
5384  return;
5385  }
5386 #endif
5387  KMP_ASSERT(__kmp_affin_fullMask != NULL);
5388  i = 0;
5389  mask = __kmp_affin_fullMask;
5390  } else {
5391  __kmp_select_mask_by_gtid(gtid, affinity, &i, &mask);
5392  }
5393  } else {
5394  if (!isa_root || __kmp_nested_proc_bind.bind_types[0] == proc_bind_false) {
5395 #if KMP_GROUP_AFFINITY
5396  if (__kmp_num_proc_groups > 1) {
5397  return;
5398  }
5399 #endif
5400  KMP_ASSERT(__kmp_affin_fullMask != NULL);
5401  i = KMP_PLACE_ALL;
5402  mask = __kmp_affin_fullMask;
5403  } else {
5404  __kmp_select_mask_by_gtid(gtid, affinity, &i, &mask);
5405  }
5406  }
5407 
5408  th->th.th_current_place = i;
5409  if (isa_root && !is_hidden_helper) {
5410  th->th.th_new_place = i;
5411  th->th.th_first_place = 0;
5412  th->th.th_last_place = affinity->num_masks - 1;
5413  } else if (KMP_AFFINITY_NON_PROC_BIND) {
5414  // When using a Non-OMP_PROC_BIND affinity method,
5415  // set all threads' place-partition-var to the entire place list
5416  th->th.th_first_place = 0;
5417  th->th.th_last_place = affinity->num_masks - 1;
5418  }
5419  // Copy topology information associated with the place
5420  if (i >= 0) {
5421  th->th.th_topology_ids = __kmp_affinity.ids[i];
5422  th->th.th_topology_attrs = __kmp_affinity.attrs[i];
5423  }
5424 
5425  if (i == KMP_PLACE_ALL) {
5426  KA_TRACE(100, ("__kmp_affinity_set_init_mask: setting T#%d to all places\n",
5427  gtid));
5428  } else {
5429  KA_TRACE(100, ("__kmp_affinity_set_init_mask: setting T#%d to place %d\n",
5430  gtid, i));
5431  }
5432 
5433  KMP_CPU_COPY(th->th.th_affin_mask, mask);
5434 }
5435 
5436 void __kmp_affinity_bind_init_mask(int gtid) {
5437  if (!KMP_AFFINITY_CAPABLE()) {
5438  return;
5439  }
5440  kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
5441  const kmp_affinity_t *affinity;
5442  const char *env_var;
5443  bool is_hidden_helper = KMP_HIDDEN_HELPER_THREAD(gtid);
5444 
5445  if (is_hidden_helper)
5446  affinity = &__kmp_hh_affinity;
5447  else
5448  affinity = &__kmp_affinity;
5449  env_var = __kmp_get_affinity_env_var(*affinity, /*for_binding=*/true);
5450  /* to avoid duplicate printing (will be correctly printed on barrier) */
5451  if (affinity->flags.verbose && (affinity->type == affinity_none ||
5452  (th->th.th_current_place != KMP_PLACE_ALL &&
5453  affinity->type != affinity_balanced)) &&
5454  !KMP_HIDDEN_HELPER_MAIN_THREAD(gtid)) {
5455  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5456  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5457  th->th.th_affin_mask);
5458  KMP_INFORM(BoundToOSProcSet, env_var, (kmp_int32)getpid(), __kmp_gettid(),
5459  gtid, buf);
5460  }
5461 
5462 #if KMP_OS_WINDOWS
5463  // On Windows* OS, the process affinity mask might have changed. If the user
5464  // didn't request affinity and this call fails, just continue silently.
5465  // See CQ171393.
5466  if (affinity->type == affinity_none) {
5467  __kmp_set_system_affinity(th->th.th_affin_mask, FALSE);
5468  } else
5469 #endif
5470 #ifndef KMP_OS_AIX
5471  // Do not set the full mask as the init mask on AIX.
5472  __kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
5473 #endif
5474 }
5475 
5476 void __kmp_affinity_bind_place(int gtid) {
5477  // Hidden helper threads should not be affected by OMP_PLACES/OMP_PROC_BIND
5478  if (!KMP_AFFINITY_CAPABLE() || KMP_HIDDEN_HELPER_THREAD(gtid)) {
5479  return;
5480  }
5481 
5482  kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
5483 
5484  KA_TRACE(100, ("__kmp_affinity_bind_place: binding T#%d to place %d (current "
5485  "place = %d)\n",
5486  gtid, th->th.th_new_place, th->th.th_current_place));
5487 
5488  // Check that the new place is within this thread's partition.
5489  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
5490  KMP_ASSERT(th->th.th_new_place >= 0);
5491  KMP_ASSERT((unsigned)th->th.th_new_place <= __kmp_affinity.num_masks);
5492  if (th->th.th_first_place <= th->th.th_last_place) {
5493  KMP_ASSERT((th->th.th_new_place >= th->th.th_first_place) &&
5494  (th->th.th_new_place <= th->th.th_last_place));
5495  } else {
5496  KMP_ASSERT((th->th.th_new_place <= th->th.th_first_place) ||
5497  (th->th.th_new_place >= th->th.th_last_place));
5498  }
5499 
5500  // Copy the thread mask to the kmp_info_t structure,
5501  // and set this thread's affinity.
5502  kmp_affin_mask_t *mask =
5503  KMP_CPU_INDEX(__kmp_affinity.masks, th->th.th_new_place);
5504  KMP_CPU_COPY(th->th.th_affin_mask, mask);
5505  th->th.th_current_place = th->th.th_new_place;
5506 
5507  if (__kmp_affinity.flags.verbose) {
5508  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5509  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5510  th->th.th_affin_mask);
5511  KMP_INFORM(BoundToOSProcSet, "OMP_PROC_BIND", (kmp_int32)getpid(),
5512  __kmp_gettid(), gtid, buf);
5513  }
5514  __kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
5515 }
5516 
5517 int __kmp_aux_set_affinity(void **mask) {
5518  int gtid;
5519  kmp_info_t *th;
5520  int retval;
5521 
5522  if (!KMP_AFFINITY_CAPABLE()) {
5523  return -1;
5524  }
5525 
5526  gtid = __kmp_entry_gtid();
5527  KA_TRACE(
5528  1000, (""); {
5529  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5530  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5531  (kmp_affin_mask_t *)(*mask));
5532  __kmp_debug_printf(
5533  "kmp_set_affinity: setting affinity mask for thread %d = %s\n",
5534  gtid, buf);
5535  });
5536 
5537  if (__kmp_env_consistency_check) {
5538  if ((mask == NULL) || (*mask == NULL)) {
5539  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
5540  } else {
5541  unsigned proc;
5542  int num_procs = 0;
5543 
5544  KMP_CPU_SET_ITERATE(proc, ((kmp_affin_mask_t *)(*mask))) {
5545  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5546  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
5547  }
5548  if (!KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask))) {
5549  continue;
5550  }
5551  num_procs++;
5552  }
5553  if (num_procs == 0) {
5554  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
5555  }
5556 
5557 #if KMP_GROUP_AFFINITY
5558  if (__kmp_get_proc_group((kmp_affin_mask_t *)(*mask)) < 0) {
5559  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
5560  }
5561 #endif /* KMP_GROUP_AFFINITY */
5562  }
5563  }
5564 
5565  th = __kmp_threads[gtid];
5566  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
5567  retval = __kmp_set_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
5568  if (retval == 0) {
5569  KMP_CPU_COPY(th->th.th_affin_mask, (kmp_affin_mask_t *)(*mask));
5570  }
5571 
5572  th->th.th_current_place = KMP_PLACE_UNDEFINED;
5573  th->th.th_new_place = KMP_PLACE_UNDEFINED;
5574  th->th.th_first_place = 0;
5575  th->th.th_last_place = __kmp_affinity.num_masks - 1;
5576 
5577  // Turn off 4.0 affinity for the current tread at this parallel level.
5578  th->th.th_current_task->td_icvs.proc_bind = proc_bind_false;
5579 
5580  return retval;
5581 }
5582 
5583 int __kmp_aux_get_affinity(void **mask) {
5584  int gtid;
5585  int retval;
5586 #if KMP_OS_WINDOWS || KMP_OS_AIX || KMP_DEBUG
5587  kmp_info_t *th;
5588 #endif
5589  if (!KMP_AFFINITY_CAPABLE()) {
5590  return -1;
5591  }
5592 
5593  gtid = __kmp_entry_gtid();
5594 #if KMP_OS_WINDOWS || KMP_OS_AIX || KMP_DEBUG
5595  th = __kmp_threads[gtid];
5596 #else
5597  (void)gtid; // unused variable
5598 #endif
5599  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
5600 
5601  KA_TRACE(
5602  1000, (""); {
5603  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5604  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5605  th->th.th_affin_mask);
5606  __kmp_printf(
5607  "kmp_get_affinity: stored affinity mask for thread %d = %s\n", gtid,
5608  buf);
5609  });
5610 
5611  if (__kmp_env_consistency_check) {
5612  if ((mask == NULL) || (*mask == NULL)) {
5613  KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity");
5614  }
5615  }
5616 
5617 #if !KMP_OS_WINDOWS && !KMP_OS_AIX
5618 
5619  retval = __kmp_get_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
5620  KA_TRACE(
5621  1000, (""); {
5622  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5623  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5624  (kmp_affin_mask_t *)(*mask));
5625  __kmp_printf(
5626  "kmp_get_affinity: system affinity mask for thread %d = %s\n", gtid,
5627  buf);
5628  });
5629  return retval;
5630 
5631 #else
5632  (void)retval;
5633 
5634  KMP_CPU_COPY((kmp_affin_mask_t *)(*mask), th->th.th_affin_mask);
5635  return 0;
5636 
5637 #endif /* !KMP_OS_WINDOWS && !KMP_OS_AIX */
5638 }
5639 
5640 int __kmp_aux_get_affinity_max_proc() {
5641  if (!KMP_AFFINITY_CAPABLE()) {
5642  return 0;
5643  }
5644 #if KMP_GROUP_AFFINITY
5645  if (__kmp_num_proc_groups > 1) {
5646  return (int)(__kmp_num_proc_groups * sizeof(DWORD_PTR) * CHAR_BIT);
5647  }
5648 #endif
5649  return __kmp_xproc;
5650 }
5651 
5652 int __kmp_aux_set_affinity_mask_proc(int proc, void **mask) {
5653  if (!KMP_AFFINITY_CAPABLE()) {
5654  return -1;
5655  }
5656 
5657  KA_TRACE(
5658  1000, (""); {
5659  int gtid = __kmp_entry_gtid();
5660  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5661  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5662  (kmp_affin_mask_t *)(*mask));
5663  __kmp_debug_printf("kmp_set_affinity_mask_proc: setting proc %d in "
5664  "affinity mask for thread %d = %s\n",
5665  proc, gtid, buf);
5666  });
5667 
5668  if (__kmp_env_consistency_check) {
5669  if ((mask == NULL) || (*mask == NULL)) {
5670  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity_mask_proc");
5671  }
5672  }
5673 
5674  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5675  return -1;
5676  }
5677  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5678  return -2;
5679  }
5680 
5681  KMP_CPU_SET(proc, (kmp_affin_mask_t *)(*mask));
5682  return 0;
5683 }
5684 
5685 int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask) {
5686  if (!KMP_AFFINITY_CAPABLE()) {
5687  return -1;
5688  }
5689 
5690  KA_TRACE(
5691  1000, (""); {
5692  int gtid = __kmp_entry_gtid();
5693  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5694  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5695  (kmp_affin_mask_t *)(*mask));
5696  __kmp_debug_printf("kmp_unset_affinity_mask_proc: unsetting proc %d in "
5697  "affinity mask for thread %d = %s\n",
5698  proc, gtid, buf);
5699  });
5700 
5701  if (__kmp_env_consistency_check) {
5702  if ((mask == NULL) || (*mask == NULL)) {
5703  KMP_FATAL(AffinityInvalidMask, "kmp_unset_affinity_mask_proc");
5704  }
5705  }
5706 
5707  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5708  return -1;
5709  }
5710  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5711  return -2;
5712  }
5713 
5714  KMP_CPU_CLR(proc, (kmp_affin_mask_t *)(*mask));
5715  return 0;
5716 }
5717 
5718 int __kmp_aux_get_affinity_mask_proc(int proc, void **mask) {
5719  if (!KMP_AFFINITY_CAPABLE()) {
5720  return -1;
5721  }
5722 
5723  KA_TRACE(
5724  1000, (""); {
5725  int gtid = __kmp_entry_gtid();
5726  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5727  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5728  (kmp_affin_mask_t *)(*mask));
5729  __kmp_debug_printf("kmp_get_affinity_mask_proc: getting proc %d in "
5730  "affinity mask for thread %d = %s\n",
5731  proc, gtid, buf);
5732  });
5733 
5734  if (__kmp_env_consistency_check) {
5735  if ((mask == NULL) || (*mask == NULL)) {
5736  KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity_mask_proc");
5737  }
5738  }
5739 
5740  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5741  return -1;
5742  }
5743  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5744  return 0;
5745  }
5746 
5747  return KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask));
5748 }
5749 
5750 #if KMP_WEIGHTED_ITERATIONS_SUPPORTED
5751 // Returns first os proc id with ATOM core
5752 int __kmp_get_first_osid_with_ecore(void) {
5753  int low = 0;
5754  int high = __kmp_topology->get_num_hw_threads() - 1;
5755  int mid = 0;
5756  while (high - low > 1) {
5757  mid = (high + low) / 2;
5758  if (__kmp_topology->at(mid).attrs.get_core_type() ==
5759  KMP_HW_CORE_TYPE_CORE) {
5760  low = mid + 1;
5761  } else {
5762  high = mid;
5763  }
5764  }
5765  if (__kmp_topology->at(mid).attrs.get_core_type() == KMP_HW_CORE_TYPE_ATOM) {
5766  return mid;
5767  }
5768  return -1;
5769 }
5770 #endif
5771 
5772 // Dynamic affinity settings - Affinity balanced
5773 void __kmp_balanced_affinity(kmp_info_t *th, int nthreads) {
5774  KMP_DEBUG_ASSERT(th);
5775  bool fine_gran = true;
5776  int tid = th->th.th_info.ds.ds_tid;
5777  const char *env_var = "KMP_AFFINITY";
5778 
5779  // Do not perform balanced affinity for the hidden helper threads
5780  if (KMP_HIDDEN_HELPER_THREAD(__kmp_gtid_from_thread(th)))
5781  return;
5782 
5783  switch (__kmp_affinity.gran) {
5784  case KMP_HW_THREAD:
5785  break;
5786  case KMP_HW_CORE:
5787  if (__kmp_nThreadsPerCore > 1) {
5788  fine_gran = false;
5789  }
5790  break;
5791  case KMP_HW_SOCKET:
5792  if (nCoresPerPkg > 1) {
5793  fine_gran = false;
5794  }
5795  break;
5796  default:
5797  fine_gran = false;
5798  }
5799 
5800  if (__kmp_topology->is_uniform()) {
5801  int coreID;
5802  int threadID;
5803  // Number of hyper threads per core in HT machine
5804  int __kmp_nth_per_core = __kmp_avail_proc / __kmp_ncores;
5805  // Number of cores
5806  int ncores = __kmp_ncores;
5807  if ((nPackages > 1) && (__kmp_nth_per_core <= 1)) {
5808  __kmp_nth_per_core = __kmp_avail_proc / nPackages;
5809  ncores = nPackages;
5810  }
5811  // How many threads will be bound to each core
5812  int chunk = nthreads / ncores;
5813  // How many cores will have an additional thread bound to it - "big cores"
5814  int big_cores = nthreads % ncores;
5815  // Number of threads on the big cores
5816  int big_nth = (chunk + 1) * big_cores;
5817  if (tid < big_nth) {
5818  coreID = tid / (chunk + 1);
5819  threadID = (tid % (chunk + 1)) % __kmp_nth_per_core;
5820  } else { // tid >= big_nth
5821  coreID = (tid - big_cores) / chunk;
5822  threadID = ((tid - big_cores) % chunk) % __kmp_nth_per_core;
5823  }
5824  KMP_DEBUG_ASSERT2(KMP_AFFINITY_CAPABLE(),
5825  "Illegal set affinity operation when not capable");
5826 
5827  kmp_affin_mask_t *mask = th->th.th_affin_mask;
5828  KMP_CPU_ZERO(mask);
5829 
5830  if (fine_gran) {
5831  int osID =
5832  __kmp_topology->at(coreID * __kmp_nth_per_core + threadID).os_id;
5833  KMP_CPU_SET(osID, mask);
5834  } else {
5835  for (int i = 0; i < __kmp_nth_per_core; i++) {
5836  int osID;
5837  osID = __kmp_topology->at(coreID * __kmp_nth_per_core + i).os_id;
5838  KMP_CPU_SET(osID, mask);
5839  }
5840  }
5841  if (__kmp_affinity.flags.verbose) {
5842  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5843  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
5844  KMP_INFORM(BoundToOSProcSet, env_var, (kmp_int32)getpid(), __kmp_gettid(),
5845  tid, buf);
5846  }
5847  __kmp_affinity_get_thread_topology_info(th);
5848  __kmp_set_system_affinity(mask, TRUE);
5849  } else { // Non-uniform topology
5850 
5851  kmp_affin_mask_t *mask = th->th.th_affin_mask;
5852  KMP_CPU_ZERO(mask);
5853 
5854  int core_level =
5855  __kmp_affinity_find_core_level(__kmp_avail_proc, __kmp_aff_depth - 1);
5856  int ncores = __kmp_affinity_compute_ncores(__kmp_avail_proc,
5857  __kmp_aff_depth - 1, core_level);
5858  int nth_per_core = __kmp_affinity_max_proc_per_core(
5859  __kmp_avail_proc, __kmp_aff_depth - 1, core_level);
5860 
5861  // For performance gain consider the special case nthreads ==
5862  // __kmp_avail_proc
5863  if (nthreads == __kmp_avail_proc) {
5864  if (fine_gran) {
5865  int osID = __kmp_topology->at(tid).os_id;
5866  KMP_CPU_SET(osID, mask);
5867  } else {
5868  int core =
5869  __kmp_affinity_find_core(tid, __kmp_aff_depth - 1, core_level);
5870  for (int i = 0; i < __kmp_avail_proc; i++) {
5871  int osID = __kmp_topology->at(i).os_id;
5872  if (__kmp_affinity_find_core(i, __kmp_aff_depth - 1, core_level) ==
5873  core) {
5874  KMP_CPU_SET(osID, mask);
5875  }
5876  }
5877  }
5878  } else if (nthreads <= ncores) {
5879 
5880  int core = 0;
5881  for (int i = 0; i < ncores; i++) {
5882  // Check if this core from procarr[] is in the mask
5883  int in_mask = 0;
5884  for (int j = 0; j < nth_per_core; j++) {
5885  if (procarr[i * nth_per_core + j] != -1) {
5886  in_mask = 1;
5887  break;
5888  }
5889  }
5890  if (in_mask) {
5891  if (tid == core) {
5892  for (int j = 0; j < nth_per_core; j++) {
5893  int osID = procarr[i * nth_per_core + j];
5894  if (osID != -1) {
5895  KMP_CPU_SET(osID, mask);
5896  // For fine granularity it is enough to set the first available
5897  // osID for this core
5898  if (fine_gran) {
5899  break;
5900  }
5901  }
5902  }
5903  break;
5904  } else {
5905  core++;
5906  }
5907  }
5908  }
5909  } else { // nthreads > ncores
5910  // Array to save the number of processors at each core
5911  int *nproc_at_core = (int *)KMP_ALLOCA(sizeof(int) * ncores);
5912  // Array to save the number of cores with "x" available processors;
5913  int *ncores_with_x_procs =
5914  (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
5915  // Array to save the number of cores with # procs from x to nth_per_core
5916  int *ncores_with_x_to_max_procs =
5917  (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
5918 
5919  for (int i = 0; i <= nth_per_core; i++) {
5920  ncores_with_x_procs[i] = 0;
5921  ncores_with_x_to_max_procs[i] = 0;
5922  }
5923 
5924  for (int i = 0; i < ncores; i++) {
5925  int cnt = 0;
5926  for (int j = 0; j < nth_per_core; j++) {
5927  if (procarr[i * nth_per_core + j] != -1) {
5928  cnt++;
5929  }
5930  }
5931  nproc_at_core[i] = cnt;
5932  ncores_with_x_procs[cnt]++;
5933  }
5934 
5935  for (int i = 0; i <= nth_per_core; i++) {
5936  for (int j = i; j <= nth_per_core; j++) {
5937  ncores_with_x_to_max_procs[i] += ncores_with_x_procs[j];
5938  }
5939  }
5940 
5941  // Max number of processors
5942  int nproc = nth_per_core * ncores;
5943  // An array to keep number of threads per each context
5944  int *newarr = (int *)__kmp_allocate(sizeof(int) * nproc);
5945  for (int i = 0; i < nproc; i++) {
5946  newarr[i] = 0;
5947  }
5948 
5949  int nth = nthreads;
5950  int flag = 0;
5951  while (nth > 0) {
5952  for (int j = 1; j <= nth_per_core; j++) {
5953  int cnt = ncores_with_x_to_max_procs[j];
5954  for (int i = 0; i < ncores; i++) {
5955  // Skip the core with 0 processors
5956  if (nproc_at_core[i] == 0) {
5957  continue;
5958  }
5959  for (int k = 0; k < nth_per_core; k++) {
5960  if (procarr[i * nth_per_core + k] != -1) {
5961  if (newarr[i * nth_per_core + k] == 0) {
5962  newarr[i * nth_per_core + k] = 1;
5963  cnt--;
5964  nth--;
5965  break;
5966  } else {
5967  if (flag != 0) {
5968  newarr[i * nth_per_core + k]++;
5969  cnt--;
5970  nth--;
5971  break;
5972  }
5973  }
5974  }
5975  }
5976  if (cnt == 0 || nth == 0) {
5977  break;
5978  }
5979  }
5980  if (nth == 0) {
5981  break;
5982  }
5983  }
5984  flag = 1;
5985  }
5986  int sum = 0;
5987  for (int i = 0; i < nproc; i++) {
5988  sum += newarr[i];
5989  if (sum > tid) {
5990  if (fine_gran) {
5991  int osID = procarr[i];
5992  KMP_CPU_SET(osID, mask);
5993  } else {
5994  int coreID = i / nth_per_core;
5995  for (int ii = 0; ii < nth_per_core; ii++) {
5996  int osID = procarr[coreID * nth_per_core + ii];
5997  if (osID != -1) {
5998  KMP_CPU_SET(osID, mask);
5999  }
6000  }
6001  }
6002  break;
6003  }
6004  }
6005  __kmp_free(newarr);
6006  }
6007 
6008  if (__kmp_affinity.flags.verbose) {
6009  char buf[KMP_AFFIN_MASK_PRINT_LEN];
6010  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
6011  KMP_INFORM(BoundToOSProcSet, env_var, (kmp_int32)getpid(), __kmp_gettid(),
6012  tid, buf);
6013  }
6014  __kmp_affinity_get_thread_topology_info(th);
6015  __kmp_set_system_affinity(mask, TRUE);
6016  }
6017 }
6018 
6019 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_DRAGONFLY || \
6020  KMP_OS_AIX
6021 // We don't need this entry for Windows because
6022 // there is GetProcessAffinityMask() api
6023 //
6024 // The intended usage is indicated by these steps:
6025 // 1) The user gets the current affinity mask
6026 // 2) Then sets the affinity by calling this function
6027 // 3) Error check the return value
6028 // 4) Use non-OpenMP parallelization
6029 // 5) Reset the affinity to what was stored in step 1)
6030 #ifdef __cplusplus
6031 extern "C"
6032 #endif
6033  int
6034  kmp_set_thread_affinity_mask_initial()
6035 // the function returns 0 on success,
6036 // -1 if we cannot bind thread
6037 // >0 (errno) if an error happened during binding
6038 {
6039  int gtid = __kmp_get_gtid();
6040  if (gtid < 0) {
6041  // Do not touch non-omp threads
6042  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
6043  "non-omp thread, returning\n"));
6044  return -1;
6045  }
6046  if (!KMP_AFFINITY_CAPABLE() || !__kmp_init_middle) {
6047  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
6048  "affinity not initialized, returning\n"));
6049  return -1;
6050  }
6051  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
6052  "set full mask for thread %d\n",
6053  gtid));
6054  KMP_DEBUG_ASSERT(__kmp_affin_fullMask != NULL);
6055 #if KMP_OS_AIX
6056  return bindprocessor(BINDTHREAD, thread_self(), PROCESSOR_CLASS_ANY);
6057 #else
6058  return __kmp_set_system_affinity(__kmp_affin_fullMask, FALSE);
6059 #endif
6060 }
6061 #endif
6062 
6063 #endif // KMP_AFFINITY_SUPPORTED
int try_open(const char *filename, const char *mode)
Definition: kmp.h:4759