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 void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar) {
42  kmp_uint32 depth;
43  // The test below is true if affinity is available, but set to "none". Need to
44  // init on first use of hierarchical barrier.
45  if (TCR_1(machine_hierarchy.uninitialized))
46  machine_hierarchy.init(nproc);
47 
48  // Adjust the hierarchy in case num threads exceeds original
49  if (nproc > machine_hierarchy.base_num_threads)
50  machine_hierarchy.resize(nproc);
51 
52  depth = machine_hierarchy.depth;
53  KMP_DEBUG_ASSERT(depth > 0);
54 
55  thr_bar->depth = depth;
56  __kmp_type_convert(machine_hierarchy.numPerLevel[0] - 1,
57  &(thr_bar->base_leaf_kids));
58  thr_bar->skip_per_level = machine_hierarchy.skipPerLevel;
59 }
60 
61 static int nCoresPerPkg, nPackages;
62 static int __kmp_nThreadsPerCore;
63 #ifndef KMP_DFLT_NTH_CORES
64 static int __kmp_ncores;
65 #endif
66 
67 const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural) {
68  switch (type) {
69  case KMP_HW_SOCKET:
70  return ((plural) ? KMP_I18N_STR(Sockets) : KMP_I18N_STR(Socket));
71  case KMP_HW_DIE:
72  return ((plural) ? KMP_I18N_STR(Dice) : KMP_I18N_STR(Die));
73  case KMP_HW_MODULE:
74  return ((plural) ? KMP_I18N_STR(Modules) : KMP_I18N_STR(Module));
75  case KMP_HW_TILE:
76  return ((plural) ? KMP_I18N_STR(Tiles) : KMP_I18N_STR(Tile));
77  case KMP_HW_NUMA:
78  return ((plural) ? KMP_I18N_STR(NumaDomains) : KMP_I18N_STR(NumaDomain));
79  case KMP_HW_L3:
80  return ((plural) ? KMP_I18N_STR(L3Caches) : KMP_I18N_STR(L3Cache));
81  case KMP_HW_L2:
82  return ((plural) ? KMP_I18N_STR(L2Caches) : KMP_I18N_STR(L2Cache));
83  case KMP_HW_L1:
84  return ((plural) ? KMP_I18N_STR(L1Caches) : KMP_I18N_STR(L1Cache));
85  case KMP_HW_LLC:
86  return ((plural) ? KMP_I18N_STR(LLCaches) : KMP_I18N_STR(LLCache));
87  case KMP_HW_CORE:
88  return ((plural) ? KMP_I18N_STR(Cores) : KMP_I18N_STR(Core));
89  case KMP_HW_THREAD:
90  return ((plural) ? KMP_I18N_STR(Threads) : KMP_I18N_STR(Thread));
91  case KMP_HW_PROC_GROUP:
92  return ((plural) ? KMP_I18N_STR(ProcGroups) : KMP_I18N_STR(ProcGroup));
93  }
94  return KMP_I18N_STR(Unknown);
95 }
96 
97 const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural) {
98  switch (type) {
99  case KMP_HW_SOCKET:
100  return ((plural) ? "sockets" : "socket");
101  case KMP_HW_DIE:
102  return ((plural) ? "dice" : "die");
103  case KMP_HW_MODULE:
104  return ((plural) ? "modules" : "module");
105  case KMP_HW_TILE:
106  return ((plural) ? "tiles" : "tile");
107  case KMP_HW_NUMA:
108  return ((plural) ? "numa_domains" : "numa_domain");
109  case KMP_HW_L3:
110  return ((plural) ? "l3_caches" : "l3_cache");
111  case KMP_HW_L2:
112  return ((plural) ? "l2_caches" : "l2_cache");
113  case KMP_HW_L1:
114  return ((plural) ? "l1_caches" : "l1_cache");
115  case KMP_HW_LLC:
116  return ((plural) ? "ll_caches" : "ll_cache");
117  case KMP_HW_CORE:
118  return ((plural) ? "cores" : "core");
119  case KMP_HW_THREAD:
120  return ((plural) ? "threads" : "thread");
121  case KMP_HW_PROC_GROUP:
122  return ((plural) ? "proc_groups" : "proc_group");
123  }
124  return ((plural) ? "unknowns" : "unknown");
125 }
126 
127 const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type) {
128  switch (type) {
129  case KMP_HW_CORE_TYPE_UNKNOWN:
130  return "unknown";
131 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
132  case KMP_HW_CORE_TYPE_ATOM:
133  return "Intel Atom(R) processor";
134  case KMP_HW_CORE_TYPE_CORE:
135  return "Intel(R) Core(TM) processor";
136 #endif
137  }
138  return "unknown";
139 }
140 
141 #if KMP_AFFINITY_SUPPORTED
142 // If affinity is supported, check the affinity
143 // verbose and warning flags before printing warning
144 #define KMP_AFF_WARNING(...) \
145  if (__kmp_affinity_verbose || \
146  (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) { \
147  KMP_WARNING(__VA_ARGS__); \
148  }
149 #else
150 #define KMP_AFF_WARNING KMP_WARNING
151 #endif
152 
154 // kmp_hw_thread_t methods
155 int kmp_hw_thread_t::compare_ids(const void *a, const void *b) {
156  const kmp_hw_thread_t *ahwthread = (const kmp_hw_thread_t *)a;
157  const kmp_hw_thread_t *bhwthread = (const kmp_hw_thread_t *)b;
158  int depth = __kmp_topology->get_depth();
159  for (int level = 0; level < depth; ++level) {
160  if (ahwthread->ids[level] < bhwthread->ids[level])
161  return -1;
162  else if (ahwthread->ids[level] > bhwthread->ids[level])
163  return 1;
164  }
165  if (ahwthread->os_id < bhwthread->os_id)
166  return -1;
167  else if (ahwthread->os_id > bhwthread->os_id)
168  return 1;
169  return 0;
170 }
171 
172 #if KMP_AFFINITY_SUPPORTED
173 int kmp_hw_thread_t::compare_compact(const void *a, const void *b) {
174  int i;
175  const kmp_hw_thread_t *aa = (const kmp_hw_thread_t *)a;
176  const kmp_hw_thread_t *bb = (const kmp_hw_thread_t *)b;
177  int depth = __kmp_topology->get_depth();
178  KMP_DEBUG_ASSERT(__kmp_affinity_compact >= 0);
179  KMP_DEBUG_ASSERT(__kmp_affinity_compact <= depth);
180  for (i = 0; i < __kmp_affinity_compact; i++) {
181  int j = depth - i - 1;
182  if (aa->sub_ids[j] < bb->sub_ids[j])
183  return -1;
184  if (aa->sub_ids[j] > bb->sub_ids[j])
185  return 1;
186  }
187  for (; i < depth; i++) {
188  int j = i - __kmp_affinity_compact;
189  if (aa->sub_ids[j] < bb->sub_ids[j])
190  return -1;
191  if (aa->sub_ids[j] > bb->sub_ids[j])
192  return 1;
193  }
194  return 0;
195 }
196 #endif
197 
198 void kmp_hw_thread_t::print() const {
199  int depth = __kmp_topology->get_depth();
200  printf("%4d ", os_id);
201  for (int i = 0; i < depth; ++i) {
202  printf("%4d ", ids[i]);
203  }
204  if (attrs) {
205  if (attrs.is_core_type_valid())
206  printf(" (%s)", __kmp_hw_get_core_type_string(attrs.get_core_type()));
207  if (attrs.is_core_eff_valid())
208  printf(" (eff=%d)", attrs.get_core_eff());
209  }
210  printf("\n");
211 }
212 
214 // kmp_topology_t methods
215 
216 // Add a layer to the topology based on the ids. Assume the topology
217 // is perfectly nested (i.e., so no object has more than one parent)
218 void kmp_topology_t::_insert_layer(kmp_hw_t type, const int *ids) {
219  // Figure out where the layer should go by comparing the ids of the current
220  // layers with the new ids
221  int target_layer;
222  int previous_id = kmp_hw_thread_t::UNKNOWN_ID;
223  int previous_new_id = kmp_hw_thread_t::UNKNOWN_ID;
224 
225  // Start from the highest layer and work down to find target layer
226  // If new layer is equal to another layer then put the new layer above
227  for (target_layer = 0; target_layer < depth; ++target_layer) {
228  bool layers_equal = true;
229  bool strictly_above_target_layer = false;
230  for (int i = 0; i < num_hw_threads; ++i) {
231  int id = hw_threads[i].ids[target_layer];
232  int new_id = ids[i];
233  if (id != previous_id && new_id == previous_new_id) {
234  // Found the layer we are strictly above
235  strictly_above_target_layer = true;
236  layers_equal = false;
237  break;
238  } else if (id == previous_id && new_id != previous_new_id) {
239  // Found a layer we are below. Move to next layer and check.
240  layers_equal = false;
241  break;
242  }
243  previous_id = id;
244  previous_new_id = new_id;
245  }
246  if (strictly_above_target_layer || layers_equal)
247  break;
248  }
249 
250  // Found the layer we are above. Now move everything to accommodate the new
251  // layer. And put the new ids and type into the topology.
252  for (int i = depth - 1, j = depth; i >= target_layer; --i, --j)
253  types[j] = types[i];
254  types[target_layer] = type;
255  for (int k = 0; k < num_hw_threads; ++k) {
256  for (int i = depth - 1, j = depth; i >= target_layer; --i, --j)
257  hw_threads[k].ids[j] = hw_threads[k].ids[i];
258  hw_threads[k].ids[target_layer] = ids[k];
259  }
260  equivalent[type] = type;
261  depth++;
262 }
263 
264 #if KMP_GROUP_AFFINITY
265 // Insert the Windows Processor Group structure into the topology
266 void kmp_topology_t::_insert_windows_proc_groups() {
267  // Do not insert the processor group structure for a single group
268  if (__kmp_num_proc_groups == 1)
269  return;
270  kmp_affin_mask_t *mask;
271  int *ids = (int *)__kmp_allocate(sizeof(int) * num_hw_threads);
272  KMP_CPU_ALLOC(mask);
273  for (int i = 0; i < num_hw_threads; ++i) {
274  KMP_CPU_ZERO(mask);
275  KMP_CPU_SET(hw_threads[i].os_id, mask);
276  ids[i] = __kmp_get_proc_group(mask);
277  }
278  KMP_CPU_FREE(mask);
279  _insert_layer(KMP_HW_PROC_GROUP, ids);
280  __kmp_free(ids);
281 }
282 #endif
283 
284 // Remove layers that don't add information to the topology.
285 // This is done by having the layer take on the id = UNKNOWN_ID (-1)
286 void kmp_topology_t::_remove_radix1_layers() {
287  int preference[KMP_HW_LAST];
288  int top_index1, top_index2;
289  // Set up preference associative array
290  preference[KMP_HW_SOCKET] = 110;
291  preference[KMP_HW_PROC_GROUP] = 100;
292  preference[KMP_HW_CORE] = 95;
293  preference[KMP_HW_THREAD] = 90;
294  preference[KMP_HW_NUMA] = 85;
295  preference[KMP_HW_DIE] = 80;
296  preference[KMP_HW_TILE] = 75;
297  preference[KMP_HW_MODULE] = 73;
298  preference[KMP_HW_L3] = 70;
299  preference[KMP_HW_L2] = 65;
300  preference[KMP_HW_L1] = 60;
301  preference[KMP_HW_LLC] = 5;
302  top_index1 = 0;
303  top_index2 = 1;
304  while (top_index1 < depth - 1 && top_index2 < depth) {
305  kmp_hw_t type1 = types[top_index1];
306  kmp_hw_t type2 = types[top_index2];
307  KMP_ASSERT_VALID_HW_TYPE(type1);
308  KMP_ASSERT_VALID_HW_TYPE(type2);
309  // Do not allow the three main topology levels (sockets, cores, threads) to
310  // be compacted down
311  if ((type1 == KMP_HW_THREAD || type1 == KMP_HW_CORE ||
312  type1 == KMP_HW_SOCKET) &&
313  (type2 == KMP_HW_THREAD || type2 == KMP_HW_CORE ||
314  type2 == KMP_HW_SOCKET)) {
315  top_index1 = top_index2++;
316  continue;
317  }
318  bool radix1 = true;
319  bool all_same = true;
320  int id1 = hw_threads[0].ids[top_index1];
321  int id2 = hw_threads[0].ids[top_index2];
322  int pref1 = preference[type1];
323  int pref2 = preference[type2];
324  for (int hwidx = 1; hwidx < num_hw_threads; ++hwidx) {
325  if (hw_threads[hwidx].ids[top_index1] == id1 &&
326  hw_threads[hwidx].ids[top_index2] != id2) {
327  radix1 = false;
328  break;
329  }
330  if (hw_threads[hwidx].ids[top_index2] != id2)
331  all_same = false;
332  id1 = hw_threads[hwidx].ids[top_index1];
333  id2 = hw_threads[hwidx].ids[top_index2];
334  }
335  if (radix1) {
336  // Select the layer to remove based on preference
337  kmp_hw_t remove_type, keep_type;
338  int remove_layer, remove_layer_ids;
339  if (pref1 > pref2) {
340  remove_type = type2;
341  remove_layer = remove_layer_ids = top_index2;
342  keep_type = type1;
343  } else {
344  remove_type = type1;
345  remove_layer = remove_layer_ids = top_index1;
346  keep_type = type2;
347  }
348  // If all the indexes for the second (deeper) layer are the same.
349  // e.g., all are zero, then make sure to keep the first layer's ids
350  if (all_same)
351  remove_layer_ids = top_index2;
352  // Remove radix one type by setting the equivalence, removing the id from
353  // the hw threads and removing the layer from types and depth
354  set_equivalent_type(remove_type, keep_type);
355  for (int idx = 0; idx < num_hw_threads; ++idx) {
356  kmp_hw_thread_t &hw_thread = hw_threads[idx];
357  for (int d = remove_layer_ids; d < depth - 1; ++d)
358  hw_thread.ids[d] = hw_thread.ids[d + 1];
359  }
360  for (int idx = remove_layer; idx < depth - 1; ++idx)
361  types[idx] = types[idx + 1];
362  depth--;
363  } else {
364  top_index1 = top_index2++;
365  }
366  }
367  KMP_ASSERT(depth > 0);
368 }
369 
370 void kmp_topology_t::_set_last_level_cache() {
371  if (get_equivalent_type(KMP_HW_L3) != KMP_HW_UNKNOWN)
372  set_equivalent_type(KMP_HW_LLC, KMP_HW_L3);
373  else if (get_equivalent_type(KMP_HW_L2) != KMP_HW_UNKNOWN)
374  set_equivalent_type(KMP_HW_LLC, KMP_HW_L2);
375 #if KMP_MIC_SUPPORTED
376  else if (__kmp_mic_type == mic3) {
377  if (get_equivalent_type(KMP_HW_L2) != KMP_HW_UNKNOWN)
378  set_equivalent_type(KMP_HW_LLC, KMP_HW_L2);
379  else if (get_equivalent_type(KMP_HW_TILE) != KMP_HW_UNKNOWN)
380  set_equivalent_type(KMP_HW_LLC, KMP_HW_TILE);
381  // L2/Tile wasn't detected so just say L1
382  else
383  set_equivalent_type(KMP_HW_LLC, KMP_HW_L1);
384  }
385 #endif
386  else if (get_equivalent_type(KMP_HW_L1) != KMP_HW_UNKNOWN)
387  set_equivalent_type(KMP_HW_LLC, KMP_HW_L1);
388  // Fallback is to set last level cache to socket or core
389  if (get_equivalent_type(KMP_HW_LLC) == KMP_HW_UNKNOWN) {
390  if (get_equivalent_type(KMP_HW_SOCKET) != KMP_HW_UNKNOWN)
391  set_equivalent_type(KMP_HW_LLC, KMP_HW_SOCKET);
392  else if (get_equivalent_type(KMP_HW_CORE) != KMP_HW_UNKNOWN)
393  set_equivalent_type(KMP_HW_LLC, KMP_HW_CORE);
394  }
395  KMP_ASSERT(get_equivalent_type(KMP_HW_LLC) != KMP_HW_UNKNOWN);
396 }
397 
398 // Gather the count of each topology layer and the ratio
399 void kmp_topology_t::_gather_enumeration_information() {
400  int previous_id[KMP_HW_LAST];
401  int max[KMP_HW_LAST];
402 
403  for (int i = 0; i < depth; ++i) {
404  previous_id[i] = kmp_hw_thread_t::UNKNOWN_ID;
405  max[i] = 0;
406  count[i] = 0;
407  ratio[i] = 0;
408  }
409  int core_level = get_level(KMP_HW_CORE);
410  for (int i = 0; i < num_hw_threads; ++i) {
411  kmp_hw_thread_t &hw_thread = hw_threads[i];
412  for (int layer = 0; layer < depth; ++layer) {
413  int id = hw_thread.ids[layer];
414  if (id != previous_id[layer]) {
415  // Add an additional increment to each count
416  for (int l = layer; l < depth; ++l)
417  count[l]++;
418  // Keep track of topology layer ratio statistics
419  max[layer]++;
420  for (int l = layer + 1; l < depth; ++l) {
421  if (max[l] > ratio[l])
422  ratio[l] = max[l];
423  max[l] = 1;
424  }
425  // Figure out the number of different core types
426  // and efficiencies for hybrid CPUs
427  if (__kmp_is_hybrid_cpu() && core_level >= 0 && layer <= core_level) {
428  if (hw_thread.attrs.is_core_eff_valid() &&
429  hw_thread.attrs.core_eff >= num_core_efficiencies) {
430  // Because efficiencies can range from 0 to max efficiency - 1,
431  // the number of efficiencies is max efficiency + 1
432  num_core_efficiencies = hw_thread.attrs.core_eff + 1;
433  }
434  if (hw_thread.attrs.is_core_type_valid()) {
435  bool found = false;
436  for (int j = 0; j < num_core_types; ++j) {
437  if (hw_thread.attrs.get_core_type() == core_types[j]) {
438  found = true;
439  break;
440  }
441  }
442  if (!found) {
443  KMP_ASSERT(num_core_types < KMP_HW_MAX_NUM_CORE_TYPES);
444  core_types[num_core_types++] = hw_thread.attrs.get_core_type();
445  }
446  }
447  }
448  break;
449  }
450  }
451  for (int layer = 0; layer < depth; ++layer) {
452  previous_id[layer] = hw_thread.ids[layer];
453  }
454  }
455  for (int layer = 0; layer < depth; ++layer) {
456  if (max[layer] > ratio[layer])
457  ratio[layer] = max[layer];
458  }
459 }
460 
461 int kmp_topology_t::_get_ncores_with_attr(const kmp_hw_attr_t &attr,
462  int above_level,
463  bool find_all) const {
464  int current, current_max;
465  int previous_id[KMP_HW_LAST];
466  for (int i = 0; i < depth; ++i)
467  previous_id[i] = kmp_hw_thread_t::UNKNOWN_ID;
468  int core_level = get_level(KMP_HW_CORE);
469  if (find_all)
470  above_level = -1;
471  KMP_ASSERT(above_level < core_level);
472  current_max = 0;
473  current = 0;
474  for (int i = 0; i < num_hw_threads; ++i) {
475  kmp_hw_thread_t &hw_thread = hw_threads[i];
476  if (!find_all && hw_thread.ids[above_level] != previous_id[above_level]) {
477  if (current > current_max)
478  current_max = current;
479  current = hw_thread.attrs.contains(attr);
480  } else {
481  for (int level = above_level + 1; level <= core_level; ++level) {
482  if (hw_thread.ids[level] != previous_id[level]) {
483  if (hw_thread.attrs.contains(attr))
484  current++;
485  break;
486  }
487  }
488  }
489  for (int level = 0; level < depth; ++level)
490  previous_id[level] = hw_thread.ids[level];
491  }
492  if (current > current_max)
493  current_max = current;
494  return current_max;
495 }
496 
497 // Find out if the topology is uniform
498 void kmp_topology_t::_discover_uniformity() {
499  int num = 1;
500  for (int level = 0; level < depth; ++level)
501  num *= ratio[level];
502  flags.uniform = (num == count[depth - 1]);
503 }
504 
505 // Set all the sub_ids for each hardware thread
506 void kmp_topology_t::_set_sub_ids() {
507  int previous_id[KMP_HW_LAST];
508  int sub_id[KMP_HW_LAST];
509 
510  for (int i = 0; i < depth; ++i) {
511  previous_id[i] = -1;
512  sub_id[i] = -1;
513  }
514  for (int i = 0; i < num_hw_threads; ++i) {
515  kmp_hw_thread_t &hw_thread = hw_threads[i];
516  // Setup the sub_id
517  for (int j = 0; j < depth; ++j) {
518  if (hw_thread.ids[j] != previous_id[j]) {
519  sub_id[j]++;
520  for (int k = j + 1; k < depth; ++k) {
521  sub_id[k] = 0;
522  }
523  break;
524  }
525  }
526  // Set previous_id
527  for (int j = 0; j < depth; ++j) {
528  previous_id[j] = hw_thread.ids[j];
529  }
530  // Set the sub_ids field
531  for (int j = 0; j < depth; ++j) {
532  hw_thread.sub_ids[j] = sub_id[j];
533  }
534  }
535 }
536 
537 void kmp_topology_t::_set_globals() {
538  // Set nCoresPerPkg, nPackages, __kmp_nThreadsPerCore, __kmp_ncores
539  int core_level, thread_level, package_level;
540  package_level = get_level(KMP_HW_SOCKET);
541 #if KMP_GROUP_AFFINITY
542  if (package_level == -1)
543  package_level = get_level(KMP_HW_PROC_GROUP);
544 #endif
545  core_level = get_level(KMP_HW_CORE);
546  thread_level = get_level(KMP_HW_THREAD);
547 
548  KMP_ASSERT(core_level != -1);
549  KMP_ASSERT(thread_level != -1);
550 
551  __kmp_nThreadsPerCore = calculate_ratio(thread_level, core_level);
552  if (package_level != -1) {
553  nCoresPerPkg = calculate_ratio(core_level, package_level);
554  nPackages = get_count(package_level);
555  } else {
556  // assume one socket
557  nCoresPerPkg = get_count(core_level);
558  nPackages = 1;
559  }
560 #ifndef KMP_DFLT_NTH_CORES
561  __kmp_ncores = get_count(core_level);
562 #endif
563 }
564 
565 kmp_topology_t *kmp_topology_t::allocate(int nproc, int ndepth,
566  const kmp_hw_t *types) {
567  kmp_topology_t *retval;
568  // Allocate all data in one large allocation
569  size_t size = sizeof(kmp_topology_t) + sizeof(kmp_hw_thread_t) * nproc +
570  sizeof(int) * (size_t)KMP_HW_LAST * 3;
571  char *bytes = (char *)__kmp_allocate(size);
572  retval = (kmp_topology_t *)bytes;
573  if (nproc > 0) {
574  retval->hw_threads = (kmp_hw_thread_t *)(bytes + sizeof(kmp_topology_t));
575  } else {
576  retval->hw_threads = nullptr;
577  }
578  retval->num_hw_threads = nproc;
579  retval->depth = ndepth;
580  int *arr =
581  (int *)(bytes + sizeof(kmp_topology_t) + sizeof(kmp_hw_thread_t) * nproc);
582  retval->types = (kmp_hw_t *)arr;
583  retval->ratio = arr + (size_t)KMP_HW_LAST;
584  retval->count = arr + 2 * (size_t)KMP_HW_LAST;
585  retval->num_core_efficiencies = 0;
586  retval->num_core_types = 0;
587  for (int i = 0; i < KMP_HW_MAX_NUM_CORE_TYPES; ++i)
588  retval->core_types[i] = KMP_HW_CORE_TYPE_UNKNOWN;
589  KMP_FOREACH_HW_TYPE(type) { retval->equivalent[type] = KMP_HW_UNKNOWN; }
590  for (int i = 0; i < ndepth; ++i) {
591  retval->types[i] = types[i];
592  retval->equivalent[types[i]] = types[i];
593  }
594  return retval;
595 }
596 
597 void kmp_topology_t::deallocate(kmp_topology_t *topology) {
598  if (topology)
599  __kmp_free(topology);
600 }
601 
602 bool kmp_topology_t::check_ids() const {
603  // Assume ids have been sorted
604  if (num_hw_threads == 0)
605  return true;
606  for (int i = 1; i < num_hw_threads; ++i) {
607  kmp_hw_thread_t &current_thread = hw_threads[i];
608  kmp_hw_thread_t &previous_thread = hw_threads[i - 1];
609  bool unique = false;
610  for (int j = 0; j < depth; ++j) {
611  if (previous_thread.ids[j] != current_thread.ids[j]) {
612  unique = true;
613  break;
614  }
615  }
616  if (unique)
617  continue;
618  return false;
619  }
620  return true;
621 }
622 
623 void kmp_topology_t::dump() const {
624  printf("***********************\n");
625  printf("*** __kmp_topology: ***\n");
626  printf("***********************\n");
627  printf("* depth: %d\n", depth);
628 
629  printf("* types: ");
630  for (int i = 0; i < depth; ++i)
631  printf("%15s ", __kmp_hw_get_keyword(types[i]));
632  printf("\n");
633 
634  printf("* ratio: ");
635  for (int i = 0; i < depth; ++i) {
636  printf("%15d ", ratio[i]);
637  }
638  printf("\n");
639 
640  printf("* count: ");
641  for (int i = 0; i < depth; ++i) {
642  printf("%15d ", count[i]);
643  }
644  printf("\n");
645 
646  printf("* num_core_eff: %d\n", num_core_efficiencies);
647  printf("* num_core_types: %d\n", num_core_types);
648  printf("* core_types: ");
649  for (int i = 0; i < num_core_types; ++i)
650  printf("%3d ", core_types[i]);
651  printf("\n");
652 
653  printf("* equivalent map:\n");
654  KMP_FOREACH_HW_TYPE(i) {
655  const char *key = __kmp_hw_get_keyword(i);
656  const char *value = __kmp_hw_get_keyword(equivalent[i]);
657  printf("%-15s -> %-15s\n", key, value);
658  }
659 
660  printf("* uniform: %s\n", (is_uniform() ? "Yes" : "No"));
661 
662  printf("* num_hw_threads: %d\n", num_hw_threads);
663  printf("* hw_threads:\n");
664  for (int i = 0; i < num_hw_threads; ++i) {
665  hw_threads[i].print();
666  }
667  printf("***********************\n");
668 }
669 
670 void kmp_topology_t::print(const char *env_var) const {
671  kmp_str_buf_t buf;
672  int print_types_depth;
673  __kmp_str_buf_init(&buf);
674  kmp_hw_t print_types[KMP_HW_LAST + 2];
675 
676  // Num Available Threads
677  KMP_INFORM(AvailableOSProc, env_var, num_hw_threads);
678 
679  // Uniform or not
680  if (is_uniform()) {
681  KMP_INFORM(Uniform, env_var);
682  } else {
683  KMP_INFORM(NonUniform, env_var);
684  }
685 
686  // Equivalent types
687  KMP_FOREACH_HW_TYPE(type) {
688  kmp_hw_t eq_type = equivalent[type];
689  if (eq_type != KMP_HW_UNKNOWN && eq_type != type) {
690  KMP_INFORM(AffEqualTopologyTypes, env_var,
691  __kmp_hw_get_catalog_string(type),
692  __kmp_hw_get_catalog_string(eq_type));
693  }
694  }
695 
696  // Quick topology
697  KMP_ASSERT(depth > 0 && depth <= (int)KMP_HW_LAST);
698  // Create a print types array that always guarantees printing
699  // the core and thread level
700  print_types_depth = 0;
701  for (int level = 0; level < depth; ++level)
702  print_types[print_types_depth++] = types[level];
703  if (equivalent[KMP_HW_CORE] != KMP_HW_CORE) {
704  // Force in the core level for quick topology
705  if (print_types[print_types_depth - 1] == KMP_HW_THREAD) {
706  // Force core before thread e.g., 1 socket X 2 threads/socket
707  // becomes 1 socket X 1 core/socket X 2 threads/socket
708  print_types[print_types_depth - 1] = KMP_HW_CORE;
709  print_types[print_types_depth++] = KMP_HW_THREAD;
710  } else {
711  print_types[print_types_depth++] = KMP_HW_CORE;
712  }
713  }
714  // Always put threads at very end of quick topology
715  if (equivalent[KMP_HW_THREAD] != KMP_HW_THREAD)
716  print_types[print_types_depth++] = KMP_HW_THREAD;
717 
718  __kmp_str_buf_clear(&buf);
719  kmp_hw_t numerator_type;
720  kmp_hw_t denominator_type = KMP_HW_UNKNOWN;
721  int core_level = get_level(KMP_HW_CORE);
722  int ncores = get_count(core_level);
723 
724  for (int plevel = 0, level = 0; plevel < print_types_depth; ++plevel) {
725  int c;
726  bool plural;
727  numerator_type = print_types[plevel];
728  KMP_ASSERT_VALID_HW_TYPE(numerator_type);
729  if (equivalent[numerator_type] != numerator_type)
730  c = 1;
731  else
732  c = get_ratio(level++);
733  plural = (c > 1);
734  if (plevel == 0) {
735  __kmp_str_buf_print(&buf, "%d %s", c,
736  __kmp_hw_get_catalog_string(numerator_type, plural));
737  } else {
738  __kmp_str_buf_print(&buf, " x %d %s/%s", c,
739  __kmp_hw_get_catalog_string(numerator_type, plural),
740  __kmp_hw_get_catalog_string(denominator_type));
741  }
742  denominator_type = numerator_type;
743  }
744  KMP_INFORM(TopologyGeneric, env_var, buf.str, ncores);
745 
746  // Hybrid topology information
747  if (__kmp_is_hybrid_cpu()) {
748  for (int i = 0; i < num_core_types; ++i) {
749  kmp_hw_core_type_t core_type = core_types[i];
750  kmp_hw_attr_t attr;
751  attr.clear();
752  attr.set_core_type(core_type);
753  int ncores = get_ncores_with_attr(attr);
754  if (ncores > 0) {
755  KMP_INFORM(TopologyHybrid, env_var, ncores,
756  __kmp_hw_get_core_type_string(core_type));
757  KMP_ASSERT(num_core_efficiencies <= KMP_HW_MAX_NUM_CORE_EFFS)
758  for (int eff = 0; eff < num_core_efficiencies; ++eff) {
759  attr.set_core_eff(eff);
760  int ncores_with_eff = get_ncores_with_attr(attr);
761  if (ncores_with_eff > 0) {
762  KMP_INFORM(TopologyHybridCoreEff, env_var, ncores_with_eff, eff);
763  }
764  }
765  }
766  }
767  }
768 
769  if (num_hw_threads <= 0) {
770  __kmp_str_buf_free(&buf);
771  return;
772  }
773 
774  // Full OS proc to hardware thread map
775  KMP_INFORM(OSProcToPhysicalThreadMap, env_var);
776  for (int i = 0; i < num_hw_threads; i++) {
777  __kmp_str_buf_clear(&buf);
778  for (int level = 0; level < depth; ++level) {
779  kmp_hw_t type = types[level];
780  __kmp_str_buf_print(&buf, "%s ", __kmp_hw_get_catalog_string(type));
781  __kmp_str_buf_print(&buf, "%d ", hw_threads[i].ids[level]);
782  }
783  if (__kmp_is_hybrid_cpu())
784  __kmp_str_buf_print(
785  &buf, "(%s)",
786  __kmp_hw_get_core_type_string(hw_threads[i].attrs.get_core_type()));
787  KMP_INFORM(OSProcMapToPack, env_var, hw_threads[i].os_id, buf.str);
788  }
789 
790  __kmp_str_buf_free(&buf);
791 }
792 
793 void kmp_topology_t::canonicalize() {
794 #if KMP_GROUP_AFFINITY
795  _insert_windows_proc_groups();
796 #endif
797  _remove_radix1_layers();
798  _gather_enumeration_information();
799  _discover_uniformity();
800  _set_sub_ids();
801  _set_globals();
802  _set_last_level_cache();
803 
804 #if KMP_MIC_SUPPORTED
805  // Manually Add L2 = Tile equivalence
806  if (__kmp_mic_type == mic3) {
807  if (get_level(KMP_HW_L2) != -1)
808  set_equivalent_type(KMP_HW_TILE, KMP_HW_L2);
809  else if (get_level(KMP_HW_TILE) != -1)
810  set_equivalent_type(KMP_HW_L2, KMP_HW_TILE);
811  }
812 #endif
813 
814  // Perform post canonicalization checking
815  KMP_ASSERT(depth > 0);
816  for (int level = 0; level < depth; ++level) {
817  // All counts, ratios, and types must be valid
818  KMP_ASSERT(count[level] > 0 && ratio[level] > 0);
819  KMP_ASSERT_VALID_HW_TYPE(types[level]);
820  // Detected types must point to themselves
821  KMP_ASSERT(equivalent[types[level]] == types[level]);
822  }
823 
824 #if KMP_AFFINITY_SUPPORTED
825  // Set the number of affinity granularity levels
826  if (__kmp_affinity_gran_levels < 0) {
827  kmp_hw_t gran_type = get_equivalent_type(__kmp_affinity_gran);
828  // Check if user's granularity request is valid
829  if (gran_type == KMP_HW_UNKNOWN) {
830  // First try core, then thread, then package
831  kmp_hw_t gran_types[3] = {KMP_HW_CORE, KMP_HW_THREAD, KMP_HW_SOCKET};
832  for (auto g : gran_types) {
833  if (get_equivalent_type(g) != KMP_HW_UNKNOWN) {
834  gran_type = g;
835  break;
836  }
837  }
838  KMP_ASSERT(gran_type != KMP_HW_UNKNOWN);
839  // Warn user what granularity setting will be used instead
840  KMP_AFF_WARNING(AffGranularityBad, "KMP_AFFINITY",
841  __kmp_hw_get_catalog_string(__kmp_affinity_gran),
842  __kmp_hw_get_catalog_string(gran_type));
843  __kmp_affinity_gran = gran_type;
844  }
845 #if KMP_GROUP_AFFINITY
846  // If more than one processor group exists, and the level of
847  // granularity specified by the user is too coarse, then the
848  // granularity must be adjusted "down" to processor group affinity
849  // because threads can only exist within one processor group.
850  // For example, if a user sets granularity=socket and there are two
851  // processor groups that cover a socket, then the runtime must
852  // restrict the granularity down to the processor group level.
853  if (__kmp_num_proc_groups > 1) {
854  int gran_depth = get_level(gran_type);
855  int proc_group_depth = get_level(KMP_HW_PROC_GROUP);
856  if (gran_depth >= 0 && proc_group_depth >= 0 &&
857  gran_depth < proc_group_depth) {
858  KMP_AFF_WARNING(AffGranTooCoarseProcGroup, "KMP_AFFINITY",
859  __kmp_hw_get_catalog_string(__kmp_affinity_gran));
860  __kmp_affinity_gran = gran_type = KMP_HW_PROC_GROUP;
861  }
862  }
863 #endif
864  __kmp_affinity_gran_levels = 0;
865  for (int i = depth - 1; i >= 0 && get_type(i) != gran_type; --i)
866  __kmp_affinity_gran_levels++;
867  }
868 #endif // KMP_AFFINITY_SUPPORTED
869 }
870 
871 // Canonicalize an explicit packages X cores/pkg X threads/core topology
872 void kmp_topology_t::canonicalize(int npackages, int ncores_per_pkg,
873  int nthreads_per_core, int ncores) {
874  int ndepth = 3;
875  depth = ndepth;
876  KMP_FOREACH_HW_TYPE(i) { equivalent[i] = KMP_HW_UNKNOWN; }
877  for (int level = 0; level < depth; ++level) {
878  count[level] = 0;
879  ratio[level] = 0;
880  }
881  count[0] = npackages;
882  count[1] = ncores;
883  count[2] = __kmp_xproc;
884  ratio[0] = npackages;
885  ratio[1] = ncores_per_pkg;
886  ratio[2] = nthreads_per_core;
887  equivalent[KMP_HW_SOCKET] = KMP_HW_SOCKET;
888  equivalent[KMP_HW_CORE] = KMP_HW_CORE;
889  equivalent[KMP_HW_THREAD] = KMP_HW_THREAD;
890  types[0] = KMP_HW_SOCKET;
891  types[1] = KMP_HW_CORE;
892  types[2] = KMP_HW_THREAD;
893  //__kmp_avail_proc = __kmp_xproc;
894  _discover_uniformity();
895 }
896 
897 // Represents running sub IDs for a single core attribute where
898 // attribute values have SIZE possibilities.
899 template <size_t SIZE, typename IndexFunc> struct kmp_sub_ids_t {
900  int last_level; // last level in topology to consider for sub_ids
901  int sub_id[SIZE]; // The sub ID for a given attribute value
902  int prev_sub_id[KMP_HW_LAST];
903  IndexFunc indexer;
904 
905 public:
906  kmp_sub_ids_t(int last_level) : last_level(last_level) {
907  KMP_ASSERT(last_level < KMP_HW_LAST);
908  for (size_t i = 0; i < SIZE; ++i)
909  sub_id[i] = -1;
910  for (size_t i = 0; i < KMP_HW_LAST; ++i)
911  prev_sub_id[i] = -1;
912  }
913  void update(const kmp_hw_thread_t &hw_thread) {
914  int idx = indexer(hw_thread);
915  KMP_ASSERT(idx < (int)SIZE);
916  for (int level = 0; level <= last_level; ++level) {
917  if (hw_thread.sub_ids[level] != prev_sub_id[level]) {
918  if (level < last_level)
919  sub_id[idx] = -1;
920  sub_id[idx]++;
921  break;
922  }
923  }
924  for (int level = 0; level <= last_level; ++level)
925  prev_sub_id[level] = hw_thread.sub_ids[level];
926  }
927  int get_sub_id(const kmp_hw_thread_t &hw_thread) const {
928  return sub_id[indexer(hw_thread)];
929  }
930 };
931 
932 static kmp_str_buf_t *
933 __kmp_hw_get_catalog_core_string(const kmp_hw_attr_t &attr, kmp_str_buf_t *buf,
934  bool plural) {
935  __kmp_str_buf_init(buf);
936  if (attr.is_core_type_valid())
937  __kmp_str_buf_print(buf, "%s %s",
938  __kmp_hw_get_core_type_string(attr.get_core_type()),
939  __kmp_hw_get_catalog_string(KMP_HW_CORE, plural));
940  else
941  __kmp_str_buf_print(buf, "%s eff=%d",
942  __kmp_hw_get_catalog_string(KMP_HW_CORE, plural),
943  attr.get_core_eff());
944  return buf;
945 }
946 
947 // Apply the KMP_HW_SUBSET envirable to the topology
948 // Returns true if KMP_HW_SUBSET filtered any processors
949 // otherwise, returns false
950 bool kmp_topology_t::filter_hw_subset() {
951  // If KMP_HW_SUBSET wasn't requested, then do nothing.
952  if (!__kmp_hw_subset)
953  return false;
954 
955  // First, sort the KMP_HW_SUBSET items by the machine topology
956  __kmp_hw_subset->sort();
957 
958  // Check to see if KMP_HW_SUBSET is a valid subset of the detected topology
959  bool using_core_types = false;
960  bool using_core_effs = false;
961  int hw_subset_depth = __kmp_hw_subset->get_depth();
962  kmp_hw_t specified[KMP_HW_LAST];
963  int *topology_levels = (int *)KMP_ALLOCA(sizeof(int) * hw_subset_depth);
964  KMP_ASSERT(hw_subset_depth > 0);
965  KMP_FOREACH_HW_TYPE(i) { specified[i] = KMP_HW_UNKNOWN; }
966  int core_level = get_level(KMP_HW_CORE);
967  for (int i = 0; i < hw_subset_depth; ++i) {
968  int max_count;
969  const kmp_hw_subset_t::item_t &item = __kmp_hw_subset->at(i);
970  int num = item.num[0];
971  int offset = item.offset[0];
972  kmp_hw_t type = item.type;
973  kmp_hw_t equivalent_type = equivalent[type];
974  int level = get_level(type);
975  topology_levels[i] = level;
976 
977  // Check to see if current layer is in detected machine topology
978  if (equivalent_type != KMP_HW_UNKNOWN) {
979  __kmp_hw_subset->at(i).type = equivalent_type;
980  } else {
981  KMP_AFF_WARNING(AffHWSubsetNotExistGeneric,
982  __kmp_hw_get_catalog_string(type));
983  return false;
984  }
985 
986  // Check to see if current layer has already been
987  // specified either directly or through an equivalent type
988  if (specified[equivalent_type] != KMP_HW_UNKNOWN) {
989  KMP_AFF_WARNING(AffHWSubsetEqvLayers, __kmp_hw_get_catalog_string(type),
990  __kmp_hw_get_catalog_string(specified[equivalent_type]));
991  return false;
992  }
993  specified[equivalent_type] = type;
994 
995  // Check to see if each layer's num & offset parameters are valid
996  max_count = get_ratio(level);
997  if (max_count < 0 ||
998  (num != kmp_hw_subset_t::USE_ALL && num + offset > max_count)) {
999  bool plural = (num > 1);
1000  KMP_AFF_WARNING(AffHWSubsetManyGeneric,
1001  __kmp_hw_get_catalog_string(type, plural));
1002  return false;
1003  }
1004 
1005  // Check to see if core attributes are consistent
1006  if (core_level == level) {
1007  // Determine which core attributes are specified
1008  for (int j = 0; j < item.num_attrs; ++j) {
1009  if (item.attr[j].is_core_type_valid())
1010  using_core_types = true;
1011  if (item.attr[j].is_core_eff_valid())
1012  using_core_effs = true;
1013  }
1014 
1015  // Check if using a single core attribute on non-hybrid arch.
1016  // Do not ignore all of KMP_HW_SUBSET, just ignore the attribute.
1017  //
1018  // Check if using multiple core attributes on non-hyrbid arch.
1019  // Ignore all of KMP_HW_SUBSET if this is the case.
1020  if ((using_core_effs || using_core_types) && !__kmp_is_hybrid_cpu()) {
1021  if (item.num_attrs == 1) {
1022  if (using_core_effs) {
1023  KMP_AFF_WARNING(AffHWSubsetIgnoringAttr, "efficiency");
1024  } else {
1025  KMP_AFF_WARNING(AffHWSubsetIgnoringAttr, "core_type");
1026  }
1027  using_core_effs = false;
1028  using_core_types = false;
1029  } else {
1030  KMP_AFF_WARNING(AffHWSubsetAttrsNonHybrid);
1031  return false;
1032  }
1033  }
1034 
1035  // Check if using both core types and core efficiencies together
1036  if (using_core_types && using_core_effs) {
1037  KMP_AFF_WARNING(AffHWSubsetIncompat, "core_type", "efficiency");
1038  return false;
1039  }
1040 
1041  // Check that core efficiency values are valid
1042  if (using_core_effs) {
1043  for (int j = 0; j < item.num_attrs; ++j) {
1044  if (item.attr[j].is_core_eff_valid()) {
1045  int core_eff = item.attr[j].get_core_eff();
1046  if (core_eff < 0 || core_eff >= num_core_efficiencies) {
1047  kmp_str_buf_t buf;
1048  __kmp_str_buf_init(&buf);
1049  __kmp_str_buf_print(&buf, "%d", item.attr[j].get_core_eff());
1050  __kmp_msg(kmp_ms_warning,
1051  KMP_MSG(AffHWSubsetAttrInvalid, "efficiency", buf.str),
1052  KMP_HNT(ValidValuesRange, 0, num_core_efficiencies - 1),
1053  __kmp_msg_null);
1054  __kmp_str_buf_free(&buf);
1055  return false;
1056  }
1057  }
1058  }
1059  }
1060 
1061  // Check that the number of requested cores with attributes is valid
1062  if (using_core_types || using_core_effs) {
1063  for (int j = 0; j < item.num_attrs; ++j) {
1064  int num = item.num[j];
1065  int offset = item.offset[j];
1066  int level_above = core_level - 1;
1067  if (level_above >= 0) {
1068  max_count = get_ncores_with_attr_per(item.attr[j], level_above);
1069  if (max_count <= 0 ||
1070  (num != kmp_hw_subset_t::USE_ALL && num + offset > max_count)) {
1071  kmp_str_buf_t buf;
1072  __kmp_hw_get_catalog_core_string(item.attr[j], &buf, num > 0);
1073  KMP_AFF_WARNING(AffHWSubsetManyGeneric, buf.str);
1074  __kmp_str_buf_free(&buf);
1075  return false;
1076  }
1077  }
1078  }
1079  }
1080 
1081  if ((using_core_types || using_core_effs) && item.num_attrs > 1) {
1082  for (int j = 0; j < item.num_attrs; ++j) {
1083  // Ambiguous use of specific core attribute + generic core
1084  // e.g., 4c & 3c:intel_core or 4c & 3c:eff1
1085  if (!item.attr[j]) {
1086  kmp_hw_attr_t other_attr;
1087  for (int k = 0; k < item.num_attrs; ++k) {
1088  if (item.attr[k] != item.attr[j]) {
1089  other_attr = item.attr[k];
1090  break;
1091  }
1092  }
1093  kmp_str_buf_t buf;
1094  __kmp_hw_get_catalog_core_string(other_attr, &buf, item.num[j] > 0);
1095  KMP_AFF_WARNING(AffHWSubsetIncompat,
1096  __kmp_hw_get_catalog_string(KMP_HW_CORE), buf.str);
1097  __kmp_str_buf_free(&buf);
1098  return false;
1099  }
1100  // Allow specifying a specific core type or core eff exactly once
1101  for (int k = 0; k < j; ++k) {
1102  if (!item.attr[j] || !item.attr[k])
1103  continue;
1104  if (item.attr[k] == item.attr[j]) {
1105  kmp_str_buf_t buf;
1106  __kmp_hw_get_catalog_core_string(item.attr[j], &buf,
1107  item.num[j] > 0);
1108  KMP_AFF_WARNING(AffHWSubsetAttrRepeat, buf.str);
1109  __kmp_str_buf_free(&buf);
1110  return false;
1111  }
1112  }
1113  }
1114  }
1115  }
1116  }
1117 
1118  struct core_type_indexer {
1119  int operator()(const kmp_hw_thread_t &t) const {
1120  switch (t.attrs.get_core_type()) {
1121 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1122  case KMP_HW_CORE_TYPE_ATOM:
1123  return 1;
1124  case KMP_HW_CORE_TYPE_CORE:
1125  return 2;
1126 #endif
1127  case KMP_HW_CORE_TYPE_UNKNOWN:
1128  return 0;
1129  }
1130  KMP_ASSERT(0);
1131  return 0;
1132  }
1133  };
1134  struct core_eff_indexer {
1135  int operator()(const kmp_hw_thread_t &t) const {
1136  return t.attrs.get_core_eff();
1137  }
1138  };
1139 
1140  kmp_sub_ids_t<KMP_HW_MAX_NUM_CORE_TYPES, core_type_indexer> core_type_sub_ids(
1141  core_level);
1142  kmp_sub_ids_t<KMP_HW_MAX_NUM_CORE_EFFS, core_eff_indexer> core_eff_sub_ids(
1143  core_level);
1144 
1145  // Determine which hardware threads should be filtered.
1146  int num_filtered = 0;
1147  bool *filtered = (bool *)__kmp_allocate(sizeof(bool) * num_hw_threads);
1148  for (int i = 0; i < num_hw_threads; ++i) {
1149  kmp_hw_thread_t &hw_thread = hw_threads[i];
1150  // Update type_sub_id
1151  if (using_core_types)
1152  core_type_sub_ids.update(hw_thread);
1153  if (using_core_effs)
1154  core_eff_sub_ids.update(hw_thread);
1155 
1156  // Check to see if this hardware thread should be filtered
1157  bool should_be_filtered = false;
1158  for (int hw_subset_index = 0; hw_subset_index < hw_subset_depth;
1159  ++hw_subset_index) {
1160  const auto &hw_subset_item = __kmp_hw_subset->at(hw_subset_index);
1161  int level = topology_levels[hw_subset_index];
1162  if (level == -1)
1163  continue;
1164  if ((using_core_effs || using_core_types) && level == core_level) {
1165  // Look for the core attribute in KMP_HW_SUBSET which corresponds
1166  // to this hardware thread's core attribute. Use this num,offset plus
1167  // the running sub_id for the particular core attribute of this hardware
1168  // thread to determine if the hardware thread should be filtered or not.
1169  int attr_idx;
1170  kmp_hw_core_type_t core_type = hw_thread.attrs.get_core_type();
1171  int core_eff = hw_thread.attrs.get_core_eff();
1172  for (attr_idx = 0; attr_idx < hw_subset_item.num_attrs; ++attr_idx) {
1173  if (using_core_types &&
1174  hw_subset_item.attr[attr_idx].get_core_type() == core_type)
1175  break;
1176  if (using_core_effs &&
1177  hw_subset_item.attr[attr_idx].get_core_eff() == core_eff)
1178  break;
1179  }
1180  // This core attribute isn't in the KMP_HW_SUBSET so always filter it.
1181  if (attr_idx == hw_subset_item.num_attrs) {
1182  should_be_filtered = true;
1183  break;
1184  }
1185  int sub_id;
1186  int num = hw_subset_item.num[attr_idx];
1187  int offset = hw_subset_item.offset[attr_idx];
1188  if (using_core_types)
1189  sub_id = core_type_sub_ids.get_sub_id(hw_thread);
1190  else
1191  sub_id = core_eff_sub_ids.get_sub_id(hw_thread);
1192  if (sub_id < offset ||
1193  (num != kmp_hw_subset_t::USE_ALL && sub_id >= offset + num)) {
1194  should_be_filtered = true;
1195  break;
1196  }
1197  } else {
1198  int num = hw_subset_item.num[0];
1199  int offset = hw_subset_item.offset[0];
1200  if (hw_thread.sub_ids[level] < offset ||
1201  (num != kmp_hw_subset_t::USE_ALL &&
1202  hw_thread.sub_ids[level] >= offset + num)) {
1203  should_be_filtered = true;
1204  break;
1205  }
1206  }
1207  }
1208  // Collect filtering information
1209  filtered[i] = should_be_filtered;
1210  if (should_be_filtered)
1211  num_filtered++;
1212  }
1213 
1214  // One last check that we shouldn't allow filtering entire machine
1215  if (num_filtered == num_hw_threads) {
1216  KMP_AFF_WARNING(AffHWSubsetAllFiltered);
1217  __kmp_free(filtered);
1218  return false;
1219  }
1220 
1221  // Apply the filter
1222  int new_index = 0;
1223  for (int i = 0; i < num_hw_threads; ++i) {
1224  if (!filtered[i]) {
1225  if (i != new_index)
1226  hw_threads[new_index] = hw_threads[i];
1227  new_index++;
1228  } else {
1229 #if KMP_AFFINITY_SUPPORTED
1230  KMP_CPU_CLR(hw_threads[i].os_id, __kmp_affin_fullMask);
1231 #endif
1232  __kmp_avail_proc--;
1233  }
1234  }
1235 
1236  KMP_DEBUG_ASSERT(new_index <= num_hw_threads);
1237  num_hw_threads = new_index;
1238 
1239  // Post hardware subset canonicalization
1240  _gather_enumeration_information();
1241  _discover_uniformity();
1242  _set_globals();
1243  _set_last_level_cache();
1244  __kmp_free(filtered);
1245  return true;
1246 }
1247 
1248 bool kmp_topology_t::is_close(int hwt1, int hwt2, int hw_level) const {
1249  if (hw_level >= depth)
1250  return true;
1251  bool retval = true;
1252  const kmp_hw_thread_t &t1 = hw_threads[hwt1];
1253  const kmp_hw_thread_t &t2 = hw_threads[hwt2];
1254  for (int i = 0; i < (depth - hw_level); ++i) {
1255  if (t1.ids[i] != t2.ids[i])
1256  return false;
1257  }
1258  return retval;
1259 }
1260 
1262 
1263 #if KMP_AFFINITY_SUPPORTED
1264 class kmp_affinity_raii_t {
1265  kmp_affin_mask_t *mask;
1266  bool restored;
1267 
1268 public:
1269  kmp_affinity_raii_t() : restored(false) {
1270  KMP_CPU_ALLOC(mask);
1271  KMP_ASSERT(mask != NULL);
1272  __kmp_get_system_affinity(mask, TRUE);
1273  }
1274  void restore() {
1275  __kmp_set_system_affinity(mask, TRUE);
1276  KMP_CPU_FREE(mask);
1277  restored = true;
1278  }
1279  ~kmp_affinity_raii_t() {
1280  if (!restored) {
1281  __kmp_set_system_affinity(mask, TRUE);
1282  KMP_CPU_FREE(mask);
1283  }
1284  }
1285 };
1286 
1287 bool KMPAffinity::picked_api = false;
1288 
1289 void *KMPAffinity::Mask::operator new(size_t n) { return __kmp_allocate(n); }
1290 void *KMPAffinity::Mask::operator new[](size_t n) { return __kmp_allocate(n); }
1291 void KMPAffinity::Mask::operator delete(void *p) { __kmp_free(p); }
1292 void KMPAffinity::Mask::operator delete[](void *p) { __kmp_free(p); }
1293 void *KMPAffinity::operator new(size_t n) { return __kmp_allocate(n); }
1294 void KMPAffinity::operator delete(void *p) { __kmp_free(p); }
1295 
1296 void KMPAffinity::pick_api() {
1297  KMPAffinity *affinity_dispatch;
1298  if (picked_api)
1299  return;
1300 #if KMP_USE_HWLOC
1301  // Only use Hwloc if affinity isn't explicitly disabled and
1302  // user requests Hwloc topology method
1303  if (__kmp_affinity_top_method == affinity_top_method_hwloc &&
1304  __kmp_affinity_type != affinity_disabled) {
1305  affinity_dispatch = new KMPHwlocAffinity();
1306  } else
1307 #endif
1308  {
1309  affinity_dispatch = new KMPNativeAffinity();
1310  }
1311  __kmp_affinity_dispatch = affinity_dispatch;
1312  picked_api = true;
1313 }
1314 
1315 void KMPAffinity::destroy_api() {
1316  if (__kmp_affinity_dispatch != NULL) {
1317  delete __kmp_affinity_dispatch;
1318  __kmp_affinity_dispatch = NULL;
1319  picked_api = false;
1320  }
1321 }
1322 
1323 #define KMP_ADVANCE_SCAN(scan) \
1324  while (*scan != '\0') { \
1325  scan++; \
1326  }
1327 
1328 // Print the affinity mask to the character array in a pretty format.
1329 // The format is a comma separated list of non-negative integers or integer
1330 // ranges: e.g., 1,2,3-5,7,9-15
1331 // The format can also be the string "{<empty>}" if no bits are set in mask
1332 char *__kmp_affinity_print_mask(char *buf, int buf_len,
1333  kmp_affin_mask_t *mask) {
1334  int start = 0, finish = 0, previous = 0;
1335  bool first_range;
1336  KMP_ASSERT(buf);
1337  KMP_ASSERT(buf_len >= 40);
1338  KMP_ASSERT(mask);
1339  char *scan = buf;
1340  char *end = buf + buf_len - 1;
1341 
1342  // Check for empty set.
1343  if (mask->begin() == mask->end()) {
1344  KMP_SNPRINTF(scan, end - scan + 1, "{<empty>}");
1345  KMP_ADVANCE_SCAN(scan);
1346  KMP_ASSERT(scan <= end);
1347  return buf;
1348  }
1349 
1350  first_range = true;
1351  start = mask->begin();
1352  while (1) {
1353  // Find next range
1354  // [start, previous] is inclusive range of contiguous bits in mask
1355  for (finish = mask->next(start), previous = start;
1356  finish == previous + 1 && finish != mask->end();
1357  finish = mask->next(finish)) {
1358  previous = finish;
1359  }
1360 
1361  // The first range does not need a comma printed before it, but the rest
1362  // of the ranges do need a comma beforehand
1363  if (!first_range) {
1364  KMP_SNPRINTF(scan, end - scan + 1, "%s", ",");
1365  KMP_ADVANCE_SCAN(scan);
1366  } else {
1367  first_range = false;
1368  }
1369  // Range with three or more contiguous bits in the affinity mask
1370  if (previous - start > 1) {
1371  KMP_SNPRINTF(scan, end - scan + 1, "%u-%u", start, previous);
1372  } else {
1373  // Range with one or two contiguous bits in the affinity mask
1374  KMP_SNPRINTF(scan, end - scan + 1, "%u", start);
1375  KMP_ADVANCE_SCAN(scan);
1376  if (previous - start > 0) {
1377  KMP_SNPRINTF(scan, end - scan + 1, ",%u", previous);
1378  }
1379  }
1380  KMP_ADVANCE_SCAN(scan);
1381  // Start over with new start point
1382  start = finish;
1383  if (start == mask->end())
1384  break;
1385  // Check for overflow
1386  if (end - scan < 2)
1387  break;
1388  }
1389 
1390  // Check for overflow
1391  KMP_ASSERT(scan <= end);
1392  return buf;
1393 }
1394 #undef KMP_ADVANCE_SCAN
1395 
1396 // Print the affinity mask to the string buffer object in a pretty format
1397 // The format is a comma separated list of non-negative integers or integer
1398 // ranges: e.g., 1,2,3-5,7,9-15
1399 // The format can also be the string "{<empty>}" if no bits are set in mask
1400 kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
1401  kmp_affin_mask_t *mask) {
1402  int start = 0, finish = 0, previous = 0;
1403  bool first_range;
1404  KMP_ASSERT(buf);
1405  KMP_ASSERT(mask);
1406 
1407  __kmp_str_buf_clear(buf);
1408 
1409  // Check for empty set.
1410  if (mask->begin() == mask->end()) {
1411  __kmp_str_buf_print(buf, "%s", "{<empty>}");
1412  return buf;
1413  }
1414 
1415  first_range = true;
1416  start = mask->begin();
1417  while (1) {
1418  // Find next range
1419  // [start, previous] is inclusive range of contiguous bits in mask
1420  for (finish = mask->next(start), previous = start;
1421  finish == previous + 1 && finish != mask->end();
1422  finish = mask->next(finish)) {
1423  previous = finish;
1424  }
1425 
1426  // The first range does not need a comma printed before it, but the rest
1427  // of the ranges do need a comma beforehand
1428  if (!first_range) {
1429  __kmp_str_buf_print(buf, "%s", ",");
1430  } else {
1431  first_range = false;
1432  }
1433  // Range with three or more contiguous bits in the affinity mask
1434  if (previous - start > 1) {
1435  __kmp_str_buf_print(buf, "%u-%u", start, previous);
1436  } else {
1437  // Range with one or two contiguous bits in the affinity mask
1438  __kmp_str_buf_print(buf, "%u", start);
1439  if (previous - start > 0) {
1440  __kmp_str_buf_print(buf, ",%u", previous);
1441  }
1442  }
1443  // Start over with new start point
1444  start = finish;
1445  if (start == mask->end())
1446  break;
1447  }
1448  return buf;
1449 }
1450 
1451 // Return (possibly empty) affinity mask representing the offline CPUs
1452 // Caller must free the mask
1453 kmp_affin_mask_t *__kmp_affinity_get_offline_cpus() {
1454  kmp_affin_mask_t *offline;
1455  KMP_CPU_ALLOC(offline);
1456  KMP_CPU_ZERO(offline);
1457 #if KMP_OS_LINUX
1458  int n, begin_cpu, end_cpu;
1459  kmp_safe_raii_file_t offline_file;
1460  auto skip_ws = [](FILE *f) {
1461  int c;
1462  do {
1463  c = fgetc(f);
1464  } while (isspace(c));
1465  if (c != EOF)
1466  ungetc(c, f);
1467  };
1468  // File contains CSV of integer ranges representing the offline CPUs
1469  // e.g., 1,2,4-7,9,11-15
1470  int status = offline_file.try_open("/sys/devices/system/cpu/offline", "r");
1471  if (status != 0)
1472  return offline;
1473  while (!feof(offline_file)) {
1474  skip_ws(offline_file);
1475  n = fscanf(offline_file, "%d", &begin_cpu);
1476  if (n != 1)
1477  break;
1478  skip_ws(offline_file);
1479  int c = fgetc(offline_file);
1480  if (c == EOF || c == ',') {
1481  // Just single CPU
1482  end_cpu = begin_cpu;
1483  } else if (c == '-') {
1484  // Range of CPUs
1485  skip_ws(offline_file);
1486  n = fscanf(offline_file, "%d", &end_cpu);
1487  if (n != 1)
1488  break;
1489  skip_ws(offline_file);
1490  c = fgetc(offline_file); // skip ','
1491  } else {
1492  // Syntax problem
1493  break;
1494  }
1495  // Ensure a valid range of CPUs
1496  if (begin_cpu < 0 || begin_cpu >= __kmp_xproc || end_cpu < 0 ||
1497  end_cpu >= __kmp_xproc || begin_cpu > end_cpu) {
1498  continue;
1499  }
1500  // Insert [begin_cpu, end_cpu] into offline mask
1501  for (int cpu = begin_cpu; cpu <= end_cpu; ++cpu) {
1502  KMP_CPU_SET(cpu, offline);
1503  }
1504  }
1505 #endif
1506  return offline;
1507 }
1508 
1509 // Return the number of available procs
1510 int __kmp_affinity_entire_machine_mask(kmp_affin_mask_t *mask) {
1511  int avail_proc = 0;
1512  KMP_CPU_ZERO(mask);
1513 
1514 #if KMP_GROUP_AFFINITY
1515 
1516  if (__kmp_num_proc_groups > 1) {
1517  int group;
1518  KMP_DEBUG_ASSERT(__kmp_GetActiveProcessorCount != NULL);
1519  for (group = 0; group < __kmp_num_proc_groups; group++) {
1520  int i;
1521  int num = __kmp_GetActiveProcessorCount(group);
1522  for (i = 0; i < num; i++) {
1523  KMP_CPU_SET(i + group * (CHAR_BIT * sizeof(DWORD_PTR)), mask);
1524  avail_proc++;
1525  }
1526  }
1527  } else
1528 
1529 #endif /* KMP_GROUP_AFFINITY */
1530 
1531  {
1532  int proc;
1533  kmp_affin_mask_t *offline_cpus = __kmp_affinity_get_offline_cpus();
1534  for (proc = 0; proc < __kmp_xproc; proc++) {
1535  // Skip offline CPUs
1536  if (KMP_CPU_ISSET(proc, offline_cpus))
1537  continue;
1538  KMP_CPU_SET(proc, mask);
1539  avail_proc++;
1540  }
1541  KMP_CPU_FREE(offline_cpus);
1542  }
1543 
1544  return avail_proc;
1545 }
1546 
1547 // All of the __kmp_affinity_create_*_map() routines should allocate the
1548 // internal topology object and set the layer ids for it. Each routine
1549 // returns a boolean on whether it was successful at doing so.
1550 kmp_affin_mask_t *__kmp_affin_fullMask = NULL;
1551 // Original mask is a subset of full mask in multiple processor groups topology
1552 kmp_affin_mask_t *__kmp_affin_origMask = NULL;
1553 
1554 #if KMP_USE_HWLOC
1555 static inline bool __kmp_hwloc_is_cache_type(hwloc_obj_t obj) {
1556 #if HWLOC_API_VERSION >= 0x00020000
1557  return hwloc_obj_type_is_cache(obj->type);
1558 #else
1559  return obj->type == HWLOC_OBJ_CACHE;
1560 #endif
1561 }
1562 
1563 // Returns KMP_HW_* type derived from HWLOC_* type
1564 static inline kmp_hw_t __kmp_hwloc_type_2_topology_type(hwloc_obj_t obj) {
1565 
1566  if (__kmp_hwloc_is_cache_type(obj)) {
1567  if (obj->attr->cache.type == HWLOC_OBJ_CACHE_INSTRUCTION)
1568  return KMP_HW_UNKNOWN;
1569  switch (obj->attr->cache.depth) {
1570  case 1:
1571  return KMP_HW_L1;
1572  case 2:
1573 #if KMP_MIC_SUPPORTED
1574  if (__kmp_mic_type == mic3) {
1575  return KMP_HW_TILE;
1576  }
1577 #endif
1578  return KMP_HW_L2;
1579  case 3:
1580  return KMP_HW_L3;
1581  }
1582  return KMP_HW_UNKNOWN;
1583  }
1584 
1585  switch (obj->type) {
1586  case HWLOC_OBJ_PACKAGE:
1587  return KMP_HW_SOCKET;
1588  case HWLOC_OBJ_NUMANODE:
1589  return KMP_HW_NUMA;
1590  case HWLOC_OBJ_CORE:
1591  return KMP_HW_CORE;
1592  case HWLOC_OBJ_PU:
1593  return KMP_HW_THREAD;
1594  case HWLOC_OBJ_GROUP:
1595  if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_DIE)
1596  return KMP_HW_DIE;
1597  else if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_TILE)
1598  return KMP_HW_TILE;
1599  else if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_MODULE)
1600  return KMP_HW_MODULE;
1601  else if (obj->attr->group.kind == HWLOC_GROUP_KIND_WINDOWS_PROCESSOR_GROUP)
1602  return KMP_HW_PROC_GROUP;
1603  return KMP_HW_UNKNOWN;
1604 #if HWLOC_API_VERSION >= 0x00020100
1605  case HWLOC_OBJ_DIE:
1606  return KMP_HW_DIE;
1607 #endif
1608  }
1609  return KMP_HW_UNKNOWN;
1610 }
1611 
1612 // Returns the number of objects of type 'type' below 'obj' within the topology
1613 // tree structure. e.g., if obj is a HWLOC_OBJ_PACKAGE object, and type is
1614 // HWLOC_OBJ_PU, then this will return the number of PU's under the SOCKET
1615 // object.
1616 static int __kmp_hwloc_get_nobjs_under_obj(hwloc_obj_t obj,
1617  hwloc_obj_type_t type) {
1618  int retval = 0;
1619  hwloc_obj_t first;
1620  for (first = hwloc_get_obj_below_by_type(__kmp_hwloc_topology, obj->type,
1621  obj->logical_index, type, 0);
1622  first != NULL && hwloc_get_ancestor_obj_by_type(__kmp_hwloc_topology,
1623  obj->type, first) == obj;
1624  first = hwloc_get_next_obj_by_type(__kmp_hwloc_topology, first->type,
1625  first)) {
1626  ++retval;
1627  }
1628  return retval;
1629 }
1630 
1631 // This gets the sub_id for a lower object under a higher object in the
1632 // topology tree
1633 static int __kmp_hwloc_get_sub_id(hwloc_topology_t t, hwloc_obj_t higher,
1634  hwloc_obj_t lower) {
1635  hwloc_obj_t obj;
1636  hwloc_obj_type_t ltype = lower->type;
1637  int lindex = lower->logical_index - 1;
1638  int sub_id = 0;
1639  // Get the previous lower object
1640  obj = hwloc_get_obj_by_type(t, ltype, lindex);
1641  while (obj && lindex >= 0 &&
1642  hwloc_bitmap_isincluded(obj->cpuset, higher->cpuset)) {
1643  if (obj->userdata) {
1644  sub_id = (int)(RCAST(kmp_intptr_t, obj->userdata));
1645  break;
1646  }
1647  sub_id++;
1648  lindex--;
1649  obj = hwloc_get_obj_by_type(t, ltype, lindex);
1650  }
1651  // store sub_id + 1 so that 0 is differed from NULL
1652  lower->userdata = RCAST(void *, sub_id + 1);
1653  return sub_id;
1654 }
1655 
1656 static bool __kmp_affinity_create_hwloc_map(kmp_i18n_id_t *const msg_id) {
1657  kmp_hw_t type;
1658  int hw_thread_index, sub_id;
1659  int depth;
1660  hwloc_obj_t pu, obj, root, prev;
1661  kmp_hw_t types[KMP_HW_LAST];
1662  hwloc_obj_type_t hwloc_types[KMP_HW_LAST];
1663 
1664  hwloc_topology_t tp = __kmp_hwloc_topology;
1665  *msg_id = kmp_i18n_null;
1666  if (__kmp_affinity_verbose) {
1667  KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
1668  }
1669 
1670  if (!KMP_AFFINITY_CAPABLE()) {
1671  // Hack to try and infer the machine topology using only the data
1672  // available from hwloc on the current thread, and __kmp_xproc.
1673  KMP_ASSERT(__kmp_affinity_type == affinity_none);
1674  // hwloc only guarantees existance of PU object, so check PACKAGE and CORE
1675  hwloc_obj_t o = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PACKAGE, 0);
1676  if (o != NULL)
1677  nCoresPerPkg = __kmp_hwloc_get_nobjs_under_obj(o, HWLOC_OBJ_CORE);
1678  else
1679  nCoresPerPkg = 1; // no PACKAGE found
1680  o = hwloc_get_obj_by_type(tp, HWLOC_OBJ_CORE, 0);
1681  if (o != NULL)
1682  __kmp_nThreadsPerCore = __kmp_hwloc_get_nobjs_under_obj(o, HWLOC_OBJ_PU);
1683  else
1684  __kmp_nThreadsPerCore = 1; // no CORE found
1685  __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
1686  if (nCoresPerPkg == 0)
1687  nCoresPerPkg = 1; // to prevent possible division by 0
1688  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
1689  return true;
1690  }
1691 
1692  // Handle multiple types of cores if they exist on the system
1693  int nr_cpu_kinds = hwloc_cpukinds_get_nr(tp, 0);
1694 
1695  typedef struct kmp_hwloc_cpukinds_info_t {
1696  int efficiency;
1697  kmp_hw_core_type_t core_type;
1698  hwloc_bitmap_t mask;
1699  } kmp_hwloc_cpukinds_info_t;
1700  kmp_hwloc_cpukinds_info_t *cpukinds = nullptr;
1701 
1702  if (nr_cpu_kinds > 0) {
1703  unsigned nr_infos;
1704  struct hwloc_info_s *infos;
1705  cpukinds = (kmp_hwloc_cpukinds_info_t *)__kmp_allocate(
1706  sizeof(kmp_hwloc_cpukinds_info_t) * nr_cpu_kinds);
1707  for (unsigned idx = 0; idx < (unsigned)nr_cpu_kinds; ++idx) {
1708  cpukinds[idx].efficiency = -1;
1709  cpukinds[idx].core_type = KMP_HW_CORE_TYPE_UNKNOWN;
1710  cpukinds[idx].mask = hwloc_bitmap_alloc();
1711  if (hwloc_cpukinds_get_info(tp, idx, cpukinds[idx].mask,
1712  &cpukinds[idx].efficiency, &nr_infos, &infos,
1713  0) == 0) {
1714  for (unsigned i = 0; i < nr_infos; ++i) {
1715  if (__kmp_str_match("CoreType", 8, infos[i].name)) {
1716 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1717  if (__kmp_str_match("IntelAtom", 9, infos[i].value)) {
1718  cpukinds[idx].core_type = KMP_HW_CORE_TYPE_ATOM;
1719  break;
1720  } else if (__kmp_str_match("IntelCore", 9, infos[i].value)) {
1721  cpukinds[idx].core_type = KMP_HW_CORE_TYPE_CORE;
1722  break;
1723  }
1724 #endif
1725  }
1726  }
1727  }
1728  }
1729  }
1730 
1731  root = hwloc_get_root_obj(tp);
1732 
1733  // Figure out the depth and types in the topology
1734  depth = 0;
1735  pu = hwloc_get_pu_obj_by_os_index(tp, __kmp_affin_fullMask->begin());
1736  KMP_ASSERT(pu);
1737  obj = pu;
1738  types[depth] = KMP_HW_THREAD;
1739  hwloc_types[depth] = obj->type;
1740  depth++;
1741  while (obj != root && obj != NULL) {
1742  obj = obj->parent;
1743 #if HWLOC_API_VERSION >= 0x00020000
1744  if (obj->memory_arity) {
1745  hwloc_obj_t memory;
1746  for (memory = obj->memory_first_child; memory;
1747  memory = hwloc_get_next_child(tp, obj, memory)) {
1748  if (memory->type == HWLOC_OBJ_NUMANODE)
1749  break;
1750  }
1751  if (memory && memory->type == HWLOC_OBJ_NUMANODE) {
1752  types[depth] = KMP_HW_NUMA;
1753  hwloc_types[depth] = memory->type;
1754  depth++;
1755  }
1756  }
1757 #endif
1758  type = __kmp_hwloc_type_2_topology_type(obj);
1759  if (type != KMP_HW_UNKNOWN) {
1760  types[depth] = type;
1761  hwloc_types[depth] = obj->type;
1762  depth++;
1763  }
1764  }
1765  KMP_ASSERT(depth > 0);
1766 
1767  // Get the order for the types correct
1768  for (int i = 0, j = depth - 1; i < j; ++i, --j) {
1769  hwloc_obj_type_t hwloc_temp = hwloc_types[i];
1770  kmp_hw_t temp = types[i];
1771  types[i] = types[j];
1772  types[j] = temp;
1773  hwloc_types[i] = hwloc_types[j];
1774  hwloc_types[j] = hwloc_temp;
1775  }
1776 
1777  // Allocate the data structure to be returned.
1778  __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
1779 
1780  hw_thread_index = 0;
1781  pu = NULL;
1782  while ((pu = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, pu))) {
1783  int index = depth - 1;
1784  bool included = KMP_CPU_ISSET(pu->os_index, __kmp_affin_fullMask);
1785  kmp_hw_thread_t &hw_thread = __kmp_topology->at(hw_thread_index);
1786  if (included) {
1787  hw_thread.clear();
1788  hw_thread.ids[index] = pu->logical_index;
1789  hw_thread.os_id = pu->os_index;
1790  // If multiple core types, then set that attribute for the hardware thread
1791  if (cpukinds) {
1792  int cpukind_index = -1;
1793  for (int i = 0; i < nr_cpu_kinds; ++i) {
1794  if (hwloc_bitmap_isset(cpukinds[i].mask, hw_thread.os_id)) {
1795  cpukind_index = i;
1796  break;
1797  }
1798  }
1799  if (cpukind_index >= 0) {
1800  hw_thread.attrs.set_core_type(cpukinds[cpukind_index].core_type);
1801  hw_thread.attrs.set_core_eff(cpukinds[cpukind_index].efficiency);
1802  }
1803  }
1804  index--;
1805  }
1806  obj = pu;
1807  prev = obj;
1808  while (obj != root && obj != NULL) {
1809  obj = obj->parent;
1810 #if HWLOC_API_VERSION >= 0x00020000
1811  // NUMA Nodes are handled differently since they are not within the
1812  // parent/child structure anymore. They are separate children
1813  // of obj (memory_first_child points to first memory child)
1814  if (obj->memory_arity) {
1815  hwloc_obj_t memory;
1816  for (memory = obj->memory_first_child; memory;
1817  memory = hwloc_get_next_child(tp, obj, memory)) {
1818  if (memory->type == HWLOC_OBJ_NUMANODE)
1819  break;
1820  }
1821  if (memory && memory->type == HWLOC_OBJ_NUMANODE) {
1822  sub_id = __kmp_hwloc_get_sub_id(tp, memory, prev);
1823  if (included) {
1824  hw_thread.ids[index] = memory->logical_index;
1825  hw_thread.ids[index + 1] = sub_id;
1826  index--;
1827  }
1828  prev = memory;
1829  }
1830  prev = obj;
1831  }
1832 #endif
1833  type = __kmp_hwloc_type_2_topology_type(obj);
1834  if (type != KMP_HW_UNKNOWN) {
1835  sub_id = __kmp_hwloc_get_sub_id(tp, obj, prev);
1836  if (included) {
1837  hw_thread.ids[index] = obj->logical_index;
1838  hw_thread.ids[index + 1] = sub_id;
1839  index--;
1840  }
1841  prev = obj;
1842  }
1843  }
1844  if (included)
1845  hw_thread_index++;
1846  }
1847 
1848  // Free the core types information
1849  if (cpukinds) {
1850  for (int idx = 0; idx < nr_cpu_kinds; ++idx)
1851  hwloc_bitmap_free(cpukinds[idx].mask);
1852  __kmp_free(cpukinds);
1853  }
1854  __kmp_topology->sort_ids();
1855  return true;
1856 }
1857 #endif // KMP_USE_HWLOC
1858 
1859 // If we don't know how to retrieve the machine's processor topology, or
1860 // encounter an error in doing so, this routine is called to form a "flat"
1861 // mapping of os thread id's <-> processor id's.
1862 static bool __kmp_affinity_create_flat_map(kmp_i18n_id_t *const msg_id) {
1863  *msg_id = kmp_i18n_null;
1864  int depth = 3;
1865  kmp_hw_t types[] = {KMP_HW_SOCKET, KMP_HW_CORE, KMP_HW_THREAD};
1866 
1867  if (__kmp_affinity_verbose) {
1868  KMP_INFORM(UsingFlatOS, "KMP_AFFINITY");
1869  }
1870 
1871  // Even if __kmp_affinity_type == affinity_none, this routine might still
1872  // called to set __kmp_ncores, as well as
1873  // __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
1874  if (!KMP_AFFINITY_CAPABLE()) {
1875  KMP_ASSERT(__kmp_affinity_type == affinity_none);
1876  __kmp_ncores = nPackages = __kmp_xproc;
1877  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
1878  return true;
1879  }
1880 
1881  // When affinity is off, this routine will still be called to set
1882  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
1883  // Make sure all these vars are set correctly, and return now if affinity is
1884  // not enabled.
1885  __kmp_ncores = nPackages = __kmp_avail_proc;
1886  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
1887 
1888  // Construct the data structure to be returned.
1889  __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
1890  int avail_ct = 0;
1891  int i;
1892  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
1893  // Skip this proc if it is not included in the machine model.
1894  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
1895  continue;
1896  }
1897  kmp_hw_thread_t &hw_thread = __kmp_topology->at(avail_ct);
1898  hw_thread.clear();
1899  hw_thread.os_id = i;
1900  hw_thread.ids[0] = i;
1901  hw_thread.ids[1] = 0;
1902  hw_thread.ids[2] = 0;
1903  avail_ct++;
1904  }
1905  if (__kmp_affinity_verbose) {
1906  KMP_INFORM(OSProcToPackage, "KMP_AFFINITY");
1907  }
1908  return true;
1909 }
1910 
1911 #if KMP_GROUP_AFFINITY
1912 // If multiple Windows* OS processor groups exist, we can create a 2-level
1913 // topology map with the groups at level 0 and the individual procs at level 1.
1914 // This facilitates letting the threads float among all procs in a group,
1915 // if granularity=group (the default when there are multiple groups).
1916 static bool __kmp_affinity_create_proc_group_map(kmp_i18n_id_t *const msg_id) {
1917  *msg_id = kmp_i18n_null;
1918  int depth = 3;
1919  kmp_hw_t types[] = {KMP_HW_PROC_GROUP, KMP_HW_CORE, KMP_HW_THREAD};
1920  const static size_t BITS_PER_GROUP = CHAR_BIT * sizeof(DWORD_PTR);
1921 
1922  if (__kmp_affinity_verbose) {
1923  KMP_INFORM(AffWindowsProcGroupMap, "KMP_AFFINITY");
1924  }
1925 
1926  // If we aren't affinity capable, then use flat topology
1927  if (!KMP_AFFINITY_CAPABLE()) {
1928  KMP_ASSERT(__kmp_affinity_type == affinity_none);
1929  nPackages = __kmp_num_proc_groups;
1930  __kmp_nThreadsPerCore = 1;
1931  __kmp_ncores = __kmp_xproc;
1932  nCoresPerPkg = nPackages / __kmp_ncores;
1933  return true;
1934  }
1935 
1936  // Construct the data structure to be returned.
1937  __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
1938  int avail_ct = 0;
1939  int i;
1940  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
1941  // Skip this proc if it is not included in the machine model.
1942  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
1943  continue;
1944  }
1945  kmp_hw_thread_t &hw_thread = __kmp_topology->at(avail_ct++);
1946  hw_thread.clear();
1947  hw_thread.os_id = i;
1948  hw_thread.ids[0] = i / BITS_PER_GROUP;
1949  hw_thread.ids[1] = hw_thread.ids[2] = i % BITS_PER_GROUP;
1950  }
1951  return true;
1952 }
1953 #endif /* KMP_GROUP_AFFINITY */
1954 
1955 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1956 
1957 template <kmp_uint32 LSB, kmp_uint32 MSB>
1958 static inline unsigned __kmp_extract_bits(kmp_uint32 v) {
1959  const kmp_uint32 SHIFT_LEFT = sizeof(kmp_uint32) * 8 - 1 - MSB;
1960  const kmp_uint32 SHIFT_RIGHT = LSB;
1961  kmp_uint32 retval = v;
1962  retval <<= SHIFT_LEFT;
1963  retval >>= (SHIFT_LEFT + SHIFT_RIGHT);
1964  return retval;
1965 }
1966 
1967 static int __kmp_cpuid_mask_width(int count) {
1968  int r = 0;
1969 
1970  while ((1 << r) < count)
1971  ++r;
1972  return r;
1973 }
1974 
1975 class apicThreadInfo {
1976 public:
1977  unsigned osId; // param to __kmp_affinity_bind_thread
1978  unsigned apicId; // from cpuid after binding
1979  unsigned maxCoresPerPkg; // ""
1980  unsigned maxThreadsPerPkg; // ""
1981  unsigned pkgId; // inferred from above values
1982  unsigned coreId; // ""
1983  unsigned threadId; // ""
1984 };
1985 
1986 static int __kmp_affinity_cmp_apicThreadInfo_phys_id(const void *a,
1987  const void *b) {
1988  const apicThreadInfo *aa = (const apicThreadInfo *)a;
1989  const apicThreadInfo *bb = (const apicThreadInfo *)b;
1990  if (aa->pkgId < bb->pkgId)
1991  return -1;
1992  if (aa->pkgId > bb->pkgId)
1993  return 1;
1994  if (aa->coreId < bb->coreId)
1995  return -1;
1996  if (aa->coreId > bb->coreId)
1997  return 1;
1998  if (aa->threadId < bb->threadId)
1999  return -1;
2000  if (aa->threadId > bb->threadId)
2001  return 1;
2002  return 0;
2003 }
2004 
2005 class kmp_cache_info_t {
2006 public:
2007  struct info_t {
2008  unsigned level, mask;
2009  };
2010  kmp_cache_info_t() : depth(0) { get_leaf4_levels(); }
2011  size_t get_depth() const { return depth; }
2012  info_t &operator[](size_t index) { return table[index]; }
2013  const info_t &operator[](size_t index) const { return table[index]; }
2014 
2015  static kmp_hw_t get_topology_type(unsigned level) {
2016  KMP_DEBUG_ASSERT(level >= 1 && level <= MAX_CACHE_LEVEL);
2017  switch (level) {
2018  case 1:
2019  return KMP_HW_L1;
2020  case 2:
2021  return KMP_HW_L2;
2022  case 3:
2023  return KMP_HW_L3;
2024  }
2025  return KMP_HW_UNKNOWN;
2026  }
2027 
2028 private:
2029  static const int MAX_CACHE_LEVEL = 3;
2030 
2031  size_t depth;
2032  info_t table[MAX_CACHE_LEVEL];
2033 
2034  void get_leaf4_levels() {
2035  unsigned level = 0;
2036  while (depth < MAX_CACHE_LEVEL) {
2037  unsigned cache_type, max_threads_sharing;
2038  unsigned cache_level, cache_mask_width;
2039  kmp_cpuid buf2;
2040  __kmp_x86_cpuid(4, level, &buf2);
2041  cache_type = __kmp_extract_bits<0, 4>(buf2.eax);
2042  if (!cache_type)
2043  break;
2044  // Skip instruction caches
2045  if (cache_type == 2) {
2046  level++;
2047  continue;
2048  }
2049  max_threads_sharing = __kmp_extract_bits<14, 25>(buf2.eax) + 1;
2050  cache_mask_width = __kmp_cpuid_mask_width(max_threads_sharing);
2051  cache_level = __kmp_extract_bits<5, 7>(buf2.eax);
2052  table[depth].level = cache_level;
2053  table[depth].mask = ((-1) << cache_mask_width);
2054  depth++;
2055  level++;
2056  }
2057  }
2058 };
2059 
2060 // On IA-32 architecture and Intel(R) 64 architecture, we attempt to use
2061 // an algorithm which cycles through the available os threads, setting
2062 // the current thread's affinity mask to that thread, and then retrieves
2063 // the Apic Id for each thread context using the cpuid instruction.
2064 static bool __kmp_affinity_create_apicid_map(kmp_i18n_id_t *const msg_id) {
2065  kmp_cpuid buf;
2066  *msg_id = kmp_i18n_null;
2067 
2068  if (__kmp_affinity_verbose) {
2069  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(DecodingLegacyAPIC));
2070  }
2071 
2072  // Check if cpuid leaf 4 is supported.
2073  __kmp_x86_cpuid(0, 0, &buf);
2074  if (buf.eax < 4) {
2075  *msg_id = kmp_i18n_str_NoLeaf4Support;
2076  return false;
2077  }
2078 
2079  // The algorithm used starts by setting the affinity to each available thread
2080  // and retrieving info from the cpuid instruction, so if we are not capable of
2081  // calling __kmp_get_system_affinity() and _kmp_get_system_affinity(), then we
2082  // need to do something else - use the defaults that we calculated from
2083  // issuing cpuid without binding to each proc.
2084  if (!KMP_AFFINITY_CAPABLE()) {
2085  // Hack to try and infer the machine topology using only the data
2086  // available from cpuid on the current thread, and __kmp_xproc.
2087  KMP_ASSERT(__kmp_affinity_type == affinity_none);
2088 
2089  // Get an upper bound on the number of threads per package using cpuid(1).
2090  // On some OS/chps combinations where HT is supported by the chip but is
2091  // disabled, this value will be 2 on a single core chip. Usually, it will be
2092  // 2 if HT is enabled and 1 if HT is disabled.
2093  __kmp_x86_cpuid(1, 0, &buf);
2094  int maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
2095  if (maxThreadsPerPkg == 0) {
2096  maxThreadsPerPkg = 1;
2097  }
2098 
2099  // The num cores per pkg comes from cpuid(4). 1 must be added to the encoded
2100  // value.
2101  //
2102  // The author of cpu_count.cpp treated this only an upper bound on the
2103  // number of cores, but I haven't seen any cases where it was greater than
2104  // the actual number of cores, so we will treat it as exact in this block of
2105  // code.
2106  //
2107  // First, we need to check if cpuid(4) is supported on this chip. To see if
2108  // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n or
2109  // greater.
2110  __kmp_x86_cpuid(0, 0, &buf);
2111  if (buf.eax >= 4) {
2112  __kmp_x86_cpuid(4, 0, &buf);
2113  nCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
2114  } else {
2115  nCoresPerPkg = 1;
2116  }
2117 
2118  // There is no way to reliably tell if HT is enabled without issuing the
2119  // cpuid instruction from every thread, can correlating the cpuid info, so
2120  // if the machine is not affinity capable, we assume that HT is off. We have
2121  // seen quite a few machines where maxThreadsPerPkg is 2, yet the machine
2122  // does not support HT.
2123  //
2124  // - Older OSes are usually found on machines with older chips, which do not
2125  // support HT.
2126  // - The performance penalty for mistakenly identifying a machine as HT when
2127  // it isn't (which results in blocktime being incorrectly set to 0) is
2128  // greater than the penalty when for mistakenly identifying a machine as
2129  // being 1 thread/core when it is really HT enabled (which results in
2130  // blocktime being incorrectly set to a positive value).
2131  __kmp_ncores = __kmp_xproc;
2132  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
2133  __kmp_nThreadsPerCore = 1;
2134  return true;
2135  }
2136 
2137  // From here on, we can assume that it is safe to call
2138  // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
2139  // __kmp_affinity_type = affinity_none.
2140 
2141  // Save the affinity mask for the current thread.
2142  kmp_affinity_raii_t previous_affinity;
2143 
2144  // Run through each of the available contexts, binding the current thread
2145  // to it, and obtaining the pertinent information using the cpuid instr.
2146  //
2147  // The relevant information is:
2148  // - Apic Id: Bits 24:31 of ebx after issuing cpuid(1) - each thread context
2149  // has a uniqie Apic Id, which is of the form pkg# : core# : thread#.
2150  // - Max Threads Per Pkg: Bits 16:23 of ebx after issuing cpuid(1). The value
2151  // of this field determines the width of the core# + thread# fields in the
2152  // Apic Id. It is also an upper bound on the number of threads per
2153  // package, but it has been verified that situations happen were it is not
2154  // exact. In particular, on certain OS/chip combinations where Intel(R)
2155  // Hyper-Threading Technology is supported by the chip but has been
2156  // disabled, the value of this field will be 2 (for a single core chip).
2157  // On other OS/chip combinations supporting Intel(R) Hyper-Threading
2158  // Technology, the value of this field will be 1 when Intel(R)
2159  // Hyper-Threading Technology is disabled and 2 when it is enabled.
2160  // - Max Cores Per Pkg: Bits 26:31 of eax after issuing cpuid(4). The value
2161  // of this field (+1) determines the width of the core# field in the Apic
2162  // Id. The comments in "cpucount.cpp" say that this value is an upper
2163  // bound, but the IA-32 architecture manual says that it is exactly the
2164  // number of cores per package, and I haven't seen any case where it
2165  // wasn't.
2166  //
2167  // From this information, deduce the package Id, core Id, and thread Id,
2168  // and set the corresponding fields in the apicThreadInfo struct.
2169  unsigned i;
2170  apicThreadInfo *threadInfo = (apicThreadInfo *)__kmp_allocate(
2171  __kmp_avail_proc * sizeof(apicThreadInfo));
2172  unsigned nApics = 0;
2173  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
2174  // Skip this proc if it is not included in the machine model.
2175  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
2176  continue;
2177  }
2178  KMP_DEBUG_ASSERT((int)nApics < __kmp_avail_proc);
2179 
2180  __kmp_affinity_dispatch->bind_thread(i);
2181  threadInfo[nApics].osId = i;
2182 
2183  // The apic id and max threads per pkg come from cpuid(1).
2184  __kmp_x86_cpuid(1, 0, &buf);
2185  if (((buf.edx >> 9) & 1) == 0) {
2186  __kmp_free(threadInfo);
2187  *msg_id = kmp_i18n_str_ApicNotPresent;
2188  return false;
2189  }
2190  threadInfo[nApics].apicId = (buf.ebx >> 24) & 0xff;
2191  threadInfo[nApics].maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
2192  if (threadInfo[nApics].maxThreadsPerPkg == 0) {
2193  threadInfo[nApics].maxThreadsPerPkg = 1;
2194  }
2195 
2196  // Max cores per pkg comes from cpuid(4). 1 must be added to the encoded
2197  // value.
2198  //
2199  // First, we need to check if cpuid(4) is supported on this chip. To see if
2200  // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n
2201  // or greater.
2202  __kmp_x86_cpuid(0, 0, &buf);
2203  if (buf.eax >= 4) {
2204  __kmp_x86_cpuid(4, 0, &buf);
2205  threadInfo[nApics].maxCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
2206  } else {
2207  threadInfo[nApics].maxCoresPerPkg = 1;
2208  }
2209 
2210  // Infer the pkgId / coreId / threadId using only the info obtained locally.
2211  int widthCT = __kmp_cpuid_mask_width(threadInfo[nApics].maxThreadsPerPkg);
2212  threadInfo[nApics].pkgId = threadInfo[nApics].apicId >> widthCT;
2213 
2214  int widthC = __kmp_cpuid_mask_width(threadInfo[nApics].maxCoresPerPkg);
2215  int widthT = widthCT - widthC;
2216  if (widthT < 0) {
2217  // I've never seen this one happen, but I suppose it could, if the cpuid
2218  // instruction on a chip was really screwed up. Make sure to restore the
2219  // affinity mask before the tail call.
2220  __kmp_free(threadInfo);
2221  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
2222  return false;
2223  }
2224 
2225  int maskC = (1 << widthC) - 1;
2226  threadInfo[nApics].coreId = (threadInfo[nApics].apicId >> widthT) & maskC;
2227 
2228  int maskT = (1 << widthT) - 1;
2229  threadInfo[nApics].threadId = threadInfo[nApics].apicId & maskT;
2230 
2231  nApics++;
2232  }
2233 
2234  // We've collected all the info we need.
2235  // Restore the old affinity mask for this thread.
2236  previous_affinity.restore();
2237 
2238  // Sort the threadInfo table by physical Id.
2239  qsort(threadInfo, nApics, sizeof(*threadInfo),
2240  __kmp_affinity_cmp_apicThreadInfo_phys_id);
2241 
2242  // The table is now sorted by pkgId / coreId / threadId, but we really don't
2243  // know the radix of any of the fields. pkgId's may be sparsely assigned among
2244  // the chips on a system. Although coreId's are usually assigned
2245  // [0 .. coresPerPkg-1] and threadId's are usually assigned
2246  // [0..threadsPerCore-1], we don't want to make any such assumptions.
2247  //
2248  // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
2249  // total # packages) are at this point - we want to determine that now. We
2250  // only have an upper bound on the first two figures.
2251  //
2252  // We also perform a consistency check at this point: the values returned by
2253  // the cpuid instruction for any thread bound to a given package had better
2254  // return the same info for maxThreadsPerPkg and maxCoresPerPkg.
2255  nPackages = 1;
2256  nCoresPerPkg = 1;
2257  __kmp_nThreadsPerCore = 1;
2258  unsigned nCores = 1;
2259 
2260  unsigned pkgCt = 1; // to determine radii
2261  unsigned lastPkgId = threadInfo[0].pkgId;
2262  unsigned coreCt = 1;
2263  unsigned lastCoreId = threadInfo[0].coreId;
2264  unsigned threadCt = 1;
2265  unsigned lastThreadId = threadInfo[0].threadId;
2266 
2267  // intra-pkg consist checks
2268  unsigned prevMaxCoresPerPkg = threadInfo[0].maxCoresPerPkg;
2269  unsigned prevMaxThreadsPerPkg = threadInfo[0].maxThreadsPerPkg;
2270 
2271  for (i = 1; i < nApics; i++) {
2272  if (threadInfo[i].pkgId != lastPkgId) {
2273  nCores++;
2274  pkgCt++;
2275  lastPkgId = threadInfo[i].pkgId;
2276  if ((int)coreCt > nCoresPerPkg)
2277  nCoresPerPkg = coreCt;
2278  coreCt = 1;
2279  lastCoreId = threadInfo[i].coreId;
2280  if ((int)threadCt > __kmp_nThreadsPerCore)
2281  __kmp_nThreadsPerCore = threadCt;
2282  threadCt = 1;
2283  lastThreadId = threadInfo[i].threadId;
2284 
2285  // This is a different package, so go on to the next iteration without
2286  // doing any consistency checks. Reset the consistency check vars, though.
2287  prevMaxCoresPerPkg = threadInfo[i].maxCoresPerPkg;
2288  prevMaxThreadsPerPkg = threadInfo[i].maxThreadsPerPkg;
2289  continue;
2290  }
2291 
2292  if (threadInfo[i].coreId != lastCoreId) {
2293  nCores++;
2294  coreCt++;
2295  lastCoreId = threadInfo[i].coreId;
2296  if ((int)threadCt > __kmp_nThreadsPerCore)
2297  __kmp_nThreadsPerCore = threadCt;
2298  threadCt = 1;
2299  lastThreadId = threadInfo[i].threadId;
2300  } else if (threadInfo[i].threadId != lastThreadId) {
2301  threadCt++;
2302  lastThreadId = threadInfo[i].threadId;
2303  } else {
2304  __kmp_free(threadInfo);
2305  *msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
2306  return false;
2307  }
2308 
2309  // Check to make certain that the maxCoresPerPkg and maxThreadsPerPkg
2310  // fields agree between all the threads bounds to a given package.
2311  if ((prevMaxCoresPerPkg != threadInfo[i].maxCoresPerPkg) ||
2312  (prevMaxThreadsPerPkg != threadInfo[i].maxThreadsPerPkg)) {
2313  __kmp_free(threadInfo);
2314  *msg_id = kmp_i18n_str_InconsistentCpuidInfo;
2315  return false;
2316  }
2317  }
2318  // When affinity is off, this routine will still be called to set
2319  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
2320  // Make sure all these vars are set correctly
2321  nPackages = pkgCt;
2322  if ((int)coreCt > nCoresPerPkg)
2323  nCoresPerPkg = coreCt;
2324  if ((int)threadCt > __kmp_nThreadsPerCore)
2325  __kmp_nThreadsPerCore = threadCt;
2326  __kmp_ncores = nCores;
2327  KMP_DEBUG_ASSERT(nApics == (unsigned)__kmp_avail_proc);
2328 
2329  // Now that we've determined the number of packages, the number of cores per
2330  // package, and the number of threads per core, we can construct the data
2331  // structure that is to be returned.
2332  int idx = 0;
2333  int pkgLevel = 0;
2334  int coreLevel = 1;
2335  int threadLevel = 2;
2336  //(__kmp_nThreadsPerCore <= 1) ? -1 : ((coreLevel >= 0) ? 2 : 1);
2337  int depth = (pkgLevel >= 0) + (coreLevel >= 0) + (threadLevel >= 0);
2338  kmp_hw_t types[3];
2339  if (pkgLevel >= 0)
2340  types[idx++] = KMP_HW_SOCKET;
2341  if (coreLevel >= 0)
2342  types[idx++] = KMP_HW_CORE;
2343  if (threadLevel >= 0)
2344  types[idx++] = KMP_HW_THREAD;
2345 
2346  KMP_ASSERT(depth > 0);
2347  __kmp_topology = kmp_topology_t::allocate(nApics, depth, types);
2348 
2349  for (i = 0; i < nApics; ++i) {
2350  idx = 0;
2351  unsigned os = threadInfo[i].osId;
2352  kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
2353  hw_thread.clear();
2354 
2355  if (pkgLevel >= 0) {
2356  hw_thread.ids[idx++] = threadInfo[i].pkgId;
2357  }
2358  if (coreLevel >= 0) {
2359  hw_thread.ids[idx++] = threadInfo[i].coreId;
2360  }
2361  if (threadLevel >= 0) {
2362  hw_thread.ids[idx++] = threadInfo[i].threadId;
2363  }
2364  hw_thread.os_id = os;
2365  }
2366 
2367  __kmp_free(threadInfo);
2368  __kmp_topology->sort_ids();
2369  if (!__kmp_topology->check_ids()) {
2370  kmp_topology_t::deallocate(__kmp_topology);
2371  __kmp_topology = nullptr;
2372  *msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
2373  return false;
2374  }
2375  return true;
2376 }
2377 
2378 // Hybrid cpu detection using CPUID.1A
2379 // Thread should be pinned to processor already
2380 static void __kmp_get_hybrid_info(kmp_hw_core_type_t *type, int *efficiency,
2381  unsigned *native_model_id) {
2382  kmp_cpuid buf;
2383  __kmp_x86_cpuid(0x1a, 0, &buf);
2384  *type = (kmp_hw_core_type_t)__kmp_extract_bits<24, 31>(buf.eax);
2385  switch (*type) {
2386  case KMP_HW_CORE_TYPE_ATOM:
2387  *efficiency = 0;
2388  break;
2389  case KMP_HW_CORE_TYPE_CORE:
2390  *efficiency = 1;
2391  break;
2392  default:
2393  *efficiency = 0;
2394  }
2395  *native_model_id = __kmp_extract_bits<0, 23>(buf.eax);
2396 }
2397 
2398 // Intel(R) microarchitecture code name Nehalem, Dunnington and later
2399 // architectures support a newer interface for specifying the x2APIC Ids,
2400 // based on CPUID.B or CPUID.1F
2401 /*
2402  * CPUID.B or 1F, Input ECX (sub leaf # aka level number)
2403  Bits Bits Bits Bits
2404  31-16 15-8 7-4 4-0
2405 ---+-----------+--------------+-------------+-----------------+
2406 EAX| reserved | reserved | reserved | Bits to Shift |
2407 ---+-----------|--------------+-------------+-----------------|
2408 EBX| reserved | Num logical processors at level (16 bits) |
2409 ---+-----------|--------------+-------------------------------|
2410 ECX| reserved | Level Type | Level Number (8 bits) |
2411 ---+-----------+--------------+-------------------------------|
2412 EDX| X2APIC ID (32 bits) |
2413 ---+----------------------------------------------------------+
2414 */
2415 
2416 enum {
2417  INTEL_LEVEL_TYPE_INVALID = 0, // Package level
2418  INTEL_LEVEL_TYPE_SMT = 1,
2419  INTEL_LEVEL_TYPE_CORE = 2,
2420  INTEL_LEVEL_TYPE_TILE = 3,
2421  INTEL_LEVEL_TYPE_MODULE = 4,
2422  INTEL_LEVEL_TYPE_DIE = 5,
2423  INTEL_LEVEL_TYPE_LAST = 6,
2424 };
2425 
2426 struct cpuid_level_info_t {
2427  unsigned level_type, mask, mask_width, nitems, cache_mask;
2428 };
2429 
2430 static kmp_hw_t __kmp_intel_type_2_topology_type(int intel_type) {
2431  switch (intel_type) {
2432  case INTEL_LEVEL_TYPE_INVALID:
2433  return KMP_HW_SOCKET;
2434  case INTEL_LEVEL_TYPE_SMT:
2435  return KMP_HW_THREAD;
2436  case INTEL_LEVEL_TYPE_CORE:
2437  return KMP_HW_CORE;
2438  case INTEL_LEVEL_TYPE_TILE:
2439  return KMP_HW_TILE;
2440  case INTEL_LEVEL_TYPE_MODULE:
2441  return KMP_HW_MODULE;
2442  case INTEL_LEVEL_TYPE_DIE:
2443  return KMP_HW_DIE;
2444  }
2445  return KMP_HW_UNKNOWN;
2446 }
2447 
2448 // This function takes the topology leaf, a levels array to store the levels
2449 // detected and a bitmap of the known levels.
2450 // Returns the number of levels in the topology
2451 static unsigned
2452 __kmp_x2apicid_get_levels(int leaf,
2453  cpuid_level_info_t levels[INTEL_LEVEL_TYPE_LAST],
2454  kmp_uint64 known_levels) {
2455  unsigned level, levels_index;
2456  unsigned level_type, mask_width, nitems;
2457  kmp_cpuid buf;
2458 
2459  // New algorithm has known topology layers act as highest unknown topology
2460  // layers when unknown topology layers exist.
2461  // e.g., Suppose layers were SMT <X> CORE <Y> <Z> PACKAGE, where <X> <Y> <Z>
2462  // are unknown topology layers, Then SMT will take the characteristics of
2463  // (SMT x <X>) and CORE will take the characteristics of (CORE x <Y> x <Z>).
2464  // This eliminates unknown portions of the topology while still keeping the
2465  // correct structure.
2466  level = levels_index = 0;
2467  do {
2468  __kmp_x86_cpuid(leaf, level, &buf);
2469  level_type = __kmp_extract_bits<8, 15>(buf.ecx);
2470  mask_width = __kmp_extract_bits<0, 4>(buf.eax);
2471  nitems = __kmp_extract_bits<0, 15>(buf.ebx);
2472  if (level_type != INTEL_LEVEL_TYPE_INVALID && nitems == 0)
2473  return 0;
2474 
2475  if (known_levels & (1ull << level_type)) {
2476  // Add a new level to the topology
2477  KMP_ASSERT(levels_index < INTEL_LEVEL_TYPE_LAST);
2478  levels[levels_index].level_type = level_type;
2479  levels[levels_index].mask_width = mask_width;
2480  levels[levels_index].nitems = nitems;
2481  levels_index++;
2482  } else {
2483  // If it is an unknown level, then logically move the previous layer up
2484  if (levels_index > 0) {
2485  levels[levels_index - 1].mask_width = mask_width;
2486  levels[levels_index - 1].nitems = nitems;
2487  }
2488  }
2489  level++;
2490  } while (level_type != INTEL_LEVEL_TYPE_INVALID);
2491 
2492  // Set the masks to & with apicid
2493  for (unsigned i = 0; i < levels_index; ++i) {
2494  if (levels[i].level_type != INTEL_LEVEL_TYPE_INVALID) {
2495  levels[i].mask = ~((-1) << levels[i].mask_width);
2496  levels[i].cache_mask = (-1) << levels[i].mask_width;
2497  for (unsigned j = 0; j < i; ++j)
2498  levels[i].mask ^= levels[j].mask;
2499  } else {
2500  KMP_DEBUG_ASSERT(levels_index > 0);
2501  levels[i].mask = (-1) << levels[i - 1].mask_width;
2502  levels[i].cache_mask = 0;
2503  }
2504  }
2505  return levels_index;
2506 }
2507 
2508 static bool __kmp_affinity_create_x2apicid_map(kmp_i18n_id_t *const msg_id) {
2509 
2510  cpuid_level_info_t levels[INTEL_LEVEL_TYPE_LAST];
2511  kmp_hw_t types[INTEL_LEVEL_TYPE_LAST];
2512  unsigned levels_index;
2513  kmp_cpuid buf;
2514  kmp_uint64 known_levels;
2515  int topology_leaf, highest_leaf, apic_id;
2516  int num_leaves;
2517  static int leaves[] = {0, 0};
2518 
2519  kmp_i18n_id_t leaf_message_id;
2520 
2521  KMP_BUILD_ASSERT(sizeof(known_levels) * CHAR_BIT > KMP_HW_LAST);
2522 
2523  *msg_id = kmp_i18n_null;
2524  if (__kmp_affinity_verbose) {
2525  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(Decodingx2APIC));
2526  }
2527 
2528  // Figure out the known topology levels
2529  known_levels = 0ull;
2530  for (int i = 0; i < INTEL_LEVEL_TYPE_LAST; ++i) {
2531  if (__kmp_intel_type_2_topology_type(i) != KMP_HW_UNKNOWN) {
2532  known_levels |= (1ull << i);
2533  }
2534  }
2535 
2536  // Get the highest cpuid leaf supported
2537  __kmp_x86_cpuid(0, 0, &buf);
2538  highest_leaf = buf.eax;
2539 
2540  // If a specific topology method was requested, only allow that specific leaf
2541  // otherwise, try both leaves 31 and 11 in that order
2542  num_leaves = 0;
2543  if (__kmp_affinity_top_method == affinity_top_method_x2apicid) {
2544  num_leaves = 1;
2545  leaves[0] = 11;
2546  leaf_message_id = kmp_i18n_str_NoLeaf11Support;
2547  } else if (__kmp_affinity_top_method == affinity_top_method_x2apicid_1f) {
2548  num_leaves = 1;
2549  leaves[0] = 31;
2550  leaf_message_id = kmp_i18n_str_NoLeaf31Support;
2551  } else {
2552  num_leaves = 2;
2553  leaves[0] = 31;
2554  leaves[1] = 11;
2555  leaf_message_id = kmp_i18n_str_NoLeaf11Support;
2556  }
2557 
2558  // Check to see if cpuid leaf 31 or 11 is supported.
2559  __kmp_nThreadsPerCore = nCoresPerPkg = nPackages = 1;
2560  topology_leaf = -1;
2561  for (int i = 0; i < num_leaves; ++i) {
2562  int leaf = leaves[i];
2563  if (highest_leaf < leaf)
2564  continue;
2565  __kmp_x86_cpuid(leaf, 0, &buf);
2566  if (buf.ebx == 0)
2567  continue;
2568  topology_leaf = leaf;
2569  levels_index = __kmp_x2apicid_get_levels(leaf, levels, known_levels);
2570  if (levels_index == 0)
2571  continue;
2572  break;
2573  }
2574  if (topology_leaf == -1 || levels_index == 0) {
2575  *msg_id = leaf_message_id;
2576  return false;
2577  }
2578  KMP_ASSERT(levels_index <= INTEL_LEVEL_TYPE_LAST);
2579 
2580  // The algorithm used starts by setting the affinity to each available thread
2581  // and retrieving info from the cpuid instruction, so if we are not capable of
2582  // calling __kmp_get_system_affinity() and __kmp_get_system_affinity(), then
2583  // we need to do something else - use the defaults that we calculated from
2584  // issuing cpuid without binding to each proc.
2585  if (!KMP_AFFINITY_CAPABLE()) {
2586  // Hack to try and infer the machine topology using only the data
2587  // available from cpuid on the current thread, and __kmp_xproc.
2588  KMP_ASSERT(__kmp_affinity_type == affinity_none);
2589  for (unsigned i = 0; i < levels_index; ++i) {
2590  if (levels[i].level_type == INTEL_LEVEL_TYPE_SMT) {
2591  __kmp_nThreadsPerCore = levels[i].nitems;
2592  } else if (levels[i].level_type == INTEL_LEVEL_TYPE_CORE) {
2593  nCoresPerPkg = levels[i].nitems;
2594  }
2595  }
2596  __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
2597  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
2598  return true;
2599  }
2600 
2601  // Allocate the data structure to be returned.
2602  int depth = levels_index;
2603  for (int i = depth - 1, j = 0; i >= 0; --i, ++j)
2604  types[j] = __kmp_intel_type_2_topology_type(levels[i].level_type);
2605  __kmp_topology =
2606  kmp_topology_t::allocate(__kmp_avail_proc, levels_index, types);
2607 
2608  // Insert equivalent cache types if they exist
2609  kmp_cache_info_t cache_info;
2610  for (size_t i = 0; i < cache_info.get_depth(); ++i) {
2611  const kmp_cache_info_t::info_t &info = cache_info[i];
2612  unsigned cache_mask = info.mask;
2613  unsigned cache_level = info.level;
2614  for (unsigned j = 0; j < levels_index; ++j) {
2615  unsigned hw_cache_mask = levels[j].cache_mask;
2616  kmp_hw_t cache_type = kmp_cache_info_t::get_topology_type(cache_level);
2617  if (hw_cache_mask == cache_mask && j < levels_index - 1) {
2618  kmp_hw_t type =
2619  __kmp_intel_type_2_topology_type(levels[j + 1].level_type);
2620  __kmp_topology->set_equivalent_type(cache_type, type);
2621  }
2622  }
2623  }
2624 
2625  // From here on, we can assume that it is safe to call
2626  // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
2627  // __kmp_affinity_type = affinity_none.
2628 
2629  // Save the affinity mask for the current thread.
2630  kmp_affinity_raii_t previous_affinity;
2631 
2632  // Run through each of the available contexts, binding the current thread
2633  // to it, and obtaining the pertinent information using the cpuid instr.
2634  unsigned int proc;
2635  int hw_thread_index = 0;
2636  KMP_CPU_SET_ITERATE(proc, __kmp_affin_fullMask) {
2637  cpuid_level_info_t my_levels[INTEL_LEVEL_TYPE_LAST];
2638  unsigned my_levels_index;
2639 
2640  // Skip this proc if it is not included in the machine model.
2641  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
2642  continue;
2643  }
2644  KMP_DEBUG_ASSERT(hw_thread_index < __kmp_avail_proc);
2645 
2646  __kmp_affinity_dispatch->bind_thread(proc);
2647 
2648  // New algorithm
2649  __kmp_x86_cpuid(topology_leaf, 0, &buf);
2650  apic_id = buf.edx;
2651  kmp_hw_thread_t &hw_thread = __kmp_topology->at(hw_thread_index);
2652  my_levels_index =
2653  __kmp_x2apicid_get_levels(topology_leaf, my_levels, known_levels);
2654  if (my_levels_index == 0 || my_levels_index != levels_index) {
2655  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
2656  return false;
2657  }
2658  hw_thread.clear();
2659  hw_thread.os_id = proc;
2660  // Put in topology information
2661  for (unsigned j = 0, idx = depth - 1; j < my_levels_index; ++j, --idx) {
2662  hw_thread.ids[idx] = apic_id & my_levels[j].mask;
2663  if (j > 0) {
2664  hw_thread.ids[idx] >>= my_levels[j - 1].mask_width;
2665  }
2666  }
2667  // Hybrid information
2668  if (__kmp_is_hybrid_cpu() && highest_leaf >= 0x1a) {
2669  kmp_hw_core_type_t type;
2670  unsigned native_model_id;
2671  int efficiency;
2672  __kmp_get_hybrid_info(&type, &efficiency, &native_model_id);
2673  hw_thread.attrs.set_core_type(type);
2674  hw_thread.attrs.set_core_eff(efficiency);
2675  }
2676  hw_thread_index++;
2677  }
2678  KMP_ASSERT(hw_thread_index > 0);
2679  __kmp_topology->sort_ids();
2680  if (!__kmp_topology->check_ids()) {
2681  kmp_topology_t::deallocate(__kmp_topology);
2682  __kmp_topology = nullptr;
2683  *msg_id = kmp_i18n_str_x2ApicIDsNotUnique;
2684  return false;
2685  }
2686  return true;
2687 }
2688 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2689 
2690 #define osIdIndex 0
2691 #define threadIdIndex 1
2692 #define coreIdIndex 2
2693 #define pkgIdIndex 3
2694 #define nodeIdIndex 4
2695 
2696 typedef unsigned *ProcCpuInfo;
2697 static unsigned maxIndex = pkgIdIndex;
2698 
2699 static int __kmp_affinity_cmp_ProcCpuInfo_phys_id(const void *a,
2700  const void *b) {
2701  unsigned i;
2702  const unsigned *aa = *(unsigned *const *)a;
2703  const unsigned *bb = *(unsigned *const *)b;
2704  for (i = maxIndex;; i--) {
2705  if (aa[i] < bb[i])
2706  return -1;
2707  if (aa[i] > bb[i])
2708  return 1;
2709  if (i == osIdIndex)
2710  break;
2711  }
2712  return 0;
2713 }
2714 
2715 #if KMP_USE_HIER_SCHED
2716 // Set the array sizes for the hierarchy layers
2717 static void __kmp_dispatch_set_hierarchy_values() {
2718  // Set the maximum number of L1's to number of cores
2719  // Set the maximum number of L2's to to either number of cores / 2 for
2720  // Intel(R) Xeon Phi(TM) coprocessor formally codenamed Knights Landing
2721  // Or the number of cores for Intel(R) Xeon(R) processors
2722  // Set the maximum number of NUMA nodes and L3's to number of packages
2723  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1] =
2724  nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
2725  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L1 + 1] = __kmp_ncores;
2726 #if KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_WINDOWS) && \
2727  KMP_MIC_SUPPORTED
2728  if (__kmp_mic_type >= mic3)
2729  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores / 2;
2730  else
2731 #endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
2732  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores;
2733  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L3 + 1] = nPackages;
2734  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_NUMA + 1] = nPackages;
2735  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_LOOP + 1] = 1;
2736  // Set the number of threads per unit
2737  // Number of hardware threads per L1/L2/L3/NUMA/LOOP
2738  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_THREAD + 1] = 1;
2739  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L1 + 1] =
2740  __kmp_nThreadsPerCore;
2741 #if KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_WINDOWS) && \
2742  KMP_MIC_SUPPORTED
2743  if (__kmp_mic_type >= mic3)
2744  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
2745  2 * __kmp_nThreadsPerCore;
2746  else
2747 #endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
2748  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
2749  __kmp_nThreadsPerCore;
2750  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L3 + 1] =
2751  nCoresPerPkg * __kmp_nThreadsPerCore;
2752  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_NUMA + 1] =
2753  nCoresPerPkg * __kmp_nThreadsPerCore;
2754  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_LOOP + 1] =
2755  nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
2756 }
2757 
2758 // Return the index into the hierarchy for this tid and layer type (L1, L2, etc)
2759 // i.e., this thread's L1 or this thread's L2, etc.
2760 int __kmp_dispatch_get_index(int tid, kmp_hier_layer_e type) {
2761  int index = type + 1;
2762  int num_hw_threads = __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1];
2763  KMP_DEBUG_ASSERT(type != kmp_hier_layer_e::LAYER_LAST);
2764  if (type == kmp_hier_layer_e::LAYER_THREAD)
2765  return tid;
2766  else if (type == kmp_hier_layer_e::LAYER_LOOP)
2767  return 0;
2768  KMP_DEBUG_ASSERT(__kmp_hier_max_units[index] != 0);
2769  if (tid >= num_hw_threads)
2770  tid = tid % num_hw_threads;
2771  return (tid / __kmp_hier_threads_per[index]) % __kmp_hier_max_units[index];
2772 }
2773 
2774 // Return the number of t1's per t2
2775 int __kmp_dispatch_get_t1_per_t2(kmp_hier_layer_e t1, kmp_hier_layer_e t2) {
2776  int i1 = t1 + 1;
2777  int i2 = t2 + 1;
2778  KMP_DEBUG_ASSERT(i1 <= i2);
2779  KMP_DEBUG_ASSERT(t1 != kmp_hier_layer_e::LAYER_LAST);
2780  KMP_DEBUG_ASSERT(t2 != kmp_hier_layer_e::LAYER_LAST);
2781  KMP_DEBUG_ASSERT(__kmp_hier_threads_per[i1] != 0);
2782  // (nthreads/t2) / (nthreads/t1) = t1 / t2
2783  return __kmp_hier_threads_per[i2] / __kmp_hier_threads_per[i1];
2784 }
2785 #endif // KMP_USE_HIER_SCHED
2786 
2787 static inline const char *__kmp_cpuinfo_get_filename() {
2788  const char *filename;
2789  if (__kmp_cpuinfo_file != nullptr)
2790  filename = __kmp_cpuinfo_file;
2791  else
2792  filename = "/proc/cpuinfo";
2793  return filename;
2794 }
2795 
2796 static inline const char *__kmp_cpuinfo_get_envvar() {
2797  const char *envvar = nullptr;
2798  if (__kmp_cpuinfo_file != nullptr)
2799  envvar = "KMP_CPUINFO_FILE";
2800  return envvar;
2801 }
2802 
2803 // Parse /proc/cpuinfo (or an alternate file in the same format) to obtain the
2804 // affinity map.
2805 static bool __kmp_affinity_create_cpuinfo_map(int *line,
2806  kmp_i18n_id_t *const msg_id) {
2807  const char *filename = __kmp_cpuinfo_get_filename();
2808  const char *envvar = __kmp_cpuinfo_get_envvar();
2809  *msg_id = kmp_i18n_null;
2810 
2811  if (__kmp_affinity_verbose) {
2812  KMP_INFORM(AffParseFilename, "KMP_AFFINITY", filename);
2813  }
2814 
2815  kmp_safe_raii_file_t f(filename, "r", envvar);
2816 
2817  // Scan of the file, and count the number of "processor" (osId) fields,
2818  // and find the highest value of <n> for a node_<n> field.
2819  char buf[256];
2820  unsigned num_records = 0;
2821  while (!feof(f)) {
2822  buf[sizeof(buf) - 1] = 1;
2823  if (!fgets(buf, sizeof(buf), f)) {
2824  // Read errors presumably because of EOF
2825  break;
2826  }
2827 
2828  char s1[] = "processor";
2829  if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
2830  num_records++;
2831  continue;
2832  }
2833 
2834  // FIXME - this will match "node_<n> <garbage>"
2835  unsigned level;
2836  if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
2837  // validate the input fisrt:
2838  if (level > (unsigned)__kmp_xproc) { // level is too big
2839  level = __kmp_xproc;
2840  }
2841  if (nodeIdIndex + level >= maxIndex) {
2842  maxIndex = nodeIdIndex + level;
2843  }
2844  continue;
2845  }
2846  }
2847 
2848  // Check for empty file / no valid processor records, or too many. The number
2849  // of records can't exceed the number of valid bits in the affinity mask.
2850  if (num_records == 0) {
2851  *msg_id = kmp_i18n_str_NoProcRecords;
2852  return false;
2853  }
2854  if (num_records > (unsigned)__kmp_xproc) {
2855  *msg_id = kmp_i18n_str_TooManyProcRecords;
2856  return false;
2857  }
2858 
2859  // Set the file pointer back to the beginning, so that we can scan the file
2860  // again, this time performing a full parse of the data. Allocate a vector of
2861  // ProcCpuInfo object, where we will place the data. Adding an extra element
2862  // at the end allows us to remove a lot of extra checks for termination
2863  // conditions.
2864  if (fseek(f, 0, SEEK_SET) != 0) {
2865  *msg_id = kmp_i18n_str_CantRewindCpuinfo;
2866  return false;
2867  }
2868 
2869  // Allocate the array of records to store the proc info in. The dummy
2870  // element at the end makes the logic in filling them out easier to code.
2871  unsigned **threadInfo =
2872  (unsigned **)__kmp_allocate((num_records + 1) * sizeof(unsigned *));
2873  unsigned i;
2874  for (i = 0; i <= num_records; i++) {
2875  threadInfo[i] =
2876  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2877  }
2878 
2879 #define CLEANUP_THREAD_INFO \
2880  for (i = 0; i <= num_records; i++) { \
2881  __kmp_free(threadInfo[i]); \
2882  } \
2883  __kmp_free(threadInfo);
2884 
2885  // A value of UINT_MAX means that we didn't find the field
2886  unsigned __index;
2887 
2888 #define INIT_PROC_INFO(p) \
2889  for (__index = 0; __index <= maxIndex; __index++) { \
2890  (p)[__index] = UINT_MAX; \
2891  }
2892 
2893  for (i = 0; i <= num_records; i++) {
2894  INIT_PROC_INFO(threadInfo[i]);
2895  }
2896 
2897  unsigned num_avail = 0;
2898  *line = 0;
2899  while (!feof(f)) {
2900  // Create an inner scoping level, so that all the goto targets at the end of
2901  // the loop appear in an outer scoping level. This avoids warnings about
2902  // jumping past an initialization to a target in the same block.
2903  {
2904  buf[sizeof(buf) - 1] = 1;
2905  bool long_line = false;
2906  if (!fgets(buf, sizeof(buf), f)) {
2907  // Read errors presumably because of EOF
2908  // If there is valid data in threadInfo[num_avail], then fake
2909  // a blank line in ensure that the last address gets parsed.
2910  bool valid = false;
2911  for (i = 0; i <= maxIndex; i++) {
2912  if (threadInfo[num_avail][i] != UINT_MAX) {
2913  valid = true;
2914  }
2915  }
2916  if (!valid) {
2917  break;
2918  }
2919  buf[0] = 0;
2920  } else if (!buf[sizeof(buf) - 1]) {
2921  // The line is longer than the buffer. Set a flag and don't
2922  // emit an error if we were going to ignore the line, anyway.
2923  long_line = true;
2924 
2925 #define CHECK_LINE \
2926  if (long_line) { \
2927  CLEANUP_THREAD_INFO; \
2928  *msg_id = kmp_i18n_str_LongLineCpuinfo; \
2929  return false; \
2930  }
2931  }
2932  (*line)++;
2933 
2934  char s1[] = "processor";
2935  if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
2936  CHECK_LINE;
2937  char *p = strchr(buf + sizeof(s1) - 1, ':');
2938  unsigned val;
2939  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2940  goto no_val;
2941  if (threadInfo[num_avail][osIdIndex] != UINT_MAX)
2942 #if KMP_ARCH_AARCH64
2943  // Handle the old AArch64 /proc/cpuinfo layout differently,
2944  // it contains all of the 'processor' entries listed in a
2945  // single 'Processor' section, therefore the normal looking
2946  // for duplicates in that section will always fail.
2947  num_avail++;
2948 #else
2949  goto dup_field;
2950 #endif
2951  threadInfo[num_avail][osIdIndex] = val;
2952 #if KMP_OS_LINUX && !(KMP_ARCH_X86 || KMP_ARCH_X86_64)
2953  char path[256];
2954  KMP_SNPRINTF(
2955  path, sizeof(path),
2956  "/sys/devices/system/cpu/cpu%u/topology/physical_package_id",
2957  threadInfo[num_avail][osIdIndex]);
2958  __kmp_read_from_file(path, "%u", &threadInfo[num_avail][pkgIdIndex]);
2959 
2960  KMP_SNPRINTF(path, sizeof(path),
2961  "/sys/devices/system/cpu/cpu%u/topology/core_id",
2962  threadInfo[num_avail][osIdIndex]);
2963  __kmp_read_from_file(path, "%u", &threadInfo[num_avail][coreIdIndex]);
2964  continue;
2965 #else
2966  }
2967  char s2[] = "physical id";
2968  if (strncmp(buf, s2, sizeof(s2) - 1) == 0) {
2969  CHECK_LINE;
2970  char *p = strchr(buf + sizeof(s2) - 1, ':');
2971  unsigned val;
2972  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2973  goto no_val;
2974  if (threadInfo[num_avail][pkgIdIndex] != UINT_MAX)
2975  goto dup_field;
2976  threadInfo[num_avail][pkgIdIndex] = val;
2977  continue;
2978  }
2979  char s3[] = "core id";
2980  if (strncmp(buf, s3, sizeof(s3) - 1) == 0) {
2981  CHECK_LINE;
2982  char *p = strchr(buf + sizeof(s3) - 1, ':');
2983  unsigned val;
2984  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2985  goto no_val;
2986  if (threadInfo[num_avail][coreIdIndex] != UINT_MAX)
2987  goto dup_field;
2988  threadInfo[num_avail][coreIdIndex] = val;
2989  continue;
2990 #endif // KMP_OS_LINUX && USE_SYSFS_INFO
2991  }
2992  char s4[] = "thread id";
2993  if (strncmp(buf, s4, sizeof(s4) - 1) == 0) {
2994  CHECK_LINE;
2995  char *p = strchr(buf + sizeof(s4) - 1, ':');
2996  unsigned val;
2997  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2998  goto no_val;
2999  if (threadInfo[num_avail][threadIdIndex] != UINT_MAX)
3000  goto dup_field;
3001  threadInfo[num_avail][threadIdIndex] = val;
3002  continue;
3003  }
3004  unsigned level;
3005  if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
3006  CHECK_LINE;
3007  char *p = strchr(buf + sizeof(s4) - 1, ':');
3008  unsigned val;
3009  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
3010  goto no_val;
3011  // validate the input before using level:
3012  if (level > (unsigned)__kmp_xproc) { // level is too big
3013  level = __kmp_xproc;
3014  }
3015  if (threadInfo[num_avail][nodeIdIndex + level] != UINT_MAX)
3016  goto dup_field;
3017  threadInfo[num_avail][nodeIdIndex + level] = val;
3018  continue;
3019  }
3020 
3021  // We didn't recognize the leading token on the line. There are lots of
3022  // leading tokens that we don't recognize - if the line isn't empty, go on
3023  // to the next line.
3024  if ((*buf != 0) && (*buf != '\n')) {
3025  // If the line is longer than the buffer, read characters
3026  // until we find a newline.
3027  if (long_line) {
3028  int ch;
3029  while (((ch = fgetc(f)) != EOF) && (ch != '\n'))
3030  ;
3031  }
3032  continue;
3033  }
3034 
3035  // A newline has signalled the end of the processor record.
3036  // Check that there aren't too many procs specified.
3037  if ((int)num_avail == __kmp_xproc) {
3038  CLEANUP_THREAD_INFO;
3039  *msg_id = kmp_i18n_str_TooManyEntries;
3040  return false;
3041  }
3042 
3043  // Check for missing fields. The osId field must be there, and we
3044  // currently require that the physical id field is specified, also.
3045  if (threadInfo[num_avail][osIdIndex] == UINT_MAX) {
3046  CLEANUP_THREAD_INFO;
3047  *msg_id = kmp_i18n_str_MissingProcField;
3048  return false;
3049  }
3050  if (threadInfo[0][pkgIdIndex] == UINT_MAX) {
3051  CLEANUP_THREAD_INFO;
3052  *msg_id = kmp_i18n_str_MissingPhysicalIDField;
3053  return false;
3054  }
3055 
3056  // Skip this proc if it is not included in the machine model.
3057  if (!KMP_CPU_ISSET(threadInfo[num_avail][osIdIndex],
3058  __kmp_affin_fullMask)) {
3059  INIT_PROC_INFO(threadInfo[num_avail]);
3060  continue;
3061  }
3062 
3063  // We have a successful parse of this proc's info.
3064  // Increment the counter, and prepare for the next proc.
3065  num_avail++;
3066  KMP_ASSERT(num_avail <= num_records);
3067  INIT_PROC_INFO(threadInfo[num_avail]);
3068  }
3069  continue;
3070 
3071  no_val:
3072  CLEANUP_THREAD_INFO;
3073  *msg_id = kmp_i18n_str_MissingValCpuinfo;
3074  return false;
3075 
3076  dup_field:
3077  CLEANUP_THREAD_INFO;
3078  *msg_id = kmp_i18n_str_DuplicateFieldCpuinfo;
3079  return false;
3080  }
3081  *line = 0;
3082 
3083 #if KMP_MIC && REDUCE_TEAM_SIZE
3084  unsigned teamSize = 0;
3085 #endif // KMP_MIC && REDUCE_TEAM_SIZE
3086 
3087  // check for num_records == __kmp_xproc ???
3088 
3089  // If it is configured to omit the package level when there is only a single
3090  // package, the logic at the end of this routine won't work if there is only a
3091  // single thread
3092  KMP_ASSERT(num_avail > 0);
3093  KMP_ASSERT(num_avail <= num_records);
3094 
3095  // Sort the threadInfo table by physical Id.
3096  qsort(threadInfo, num_avail, sizeof(*threadInfo),
3097  __kmp_affinity_cmp_ProcCpuInfo_phys_id);
3098 
3099  // The table is now sorted by pkgId / coreId / threadId, but we really don't
3100  // know the radix of any of the fields. pkgId's may be sparsely assigned among
3101  // the chips on a system. Although coreId's are usually assigned
3102  // [0 .. coresPerPkg-1] and threadId's are usually assigned
3103  // [0..threadsPerCore-1], we don't want to make any such assumptions.
3104  //
3105  // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
3106  // total # packages) are at this point - we want to determine that now. We
3107  // only have an upper bound on the first two figures.
3108  unsigned *counts =
3109  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3110  unsigned *maxCt =
3111  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3112  unsigned *totals =
3113  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3114  unsigned *lastId =
3115  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3116 
3117  bool assign_thread_ids = false;
3118  unsigned threadIdCt;
3119  unsigned index;
3120 
3121 restart_radix_check:
3122  threadIdCt = 0;
3123 
3124  // Initialize the counter arrays with data from threadInfo[0].
3125  if (assign_thread_ids) {
3126  if (threadInfo[0][threadIdIndex] == UINT_MAX) {
3127  threadInfo[0][threadIdIndex] = threadIdCt++;
3128  } else if (threadIdCt <= threadInfo[0][threadIdIndex]) {
3129  threadIdCt = threadInfo[0][threadIdIndex] + 1;
3130  }
3131  }
3132  for (index = 0; index <= maxIndex; index++) {
3133  counts[index] = 1;
3134  maxCt[index] = 1;
3135  totals[index] = 1;
3136  lastId[index] = threadInfo[0][index];
3137  ;
3138  }
3139 
3140  // Run through the rest of the OS procs.
3141  for (i = 1; i < num_avail; i++) {
3142  // Find the most significant index whose id differs from the id for the
3143  // previous OS proc.
3144  for (index = maxIndex; index >= threadIdIndex; index--) {
3145  if (assign_thread_ids && (index == threadIdIndex)) {
3146  // Auto-assign the thread id field if it wasn't specified.
3147  if (threadInfo[i][threadIdIndex] == UINT_MAX) {
3148  threadInfo[i][threadIdIndex] = threadIdCt++;
3149  }
3150  // Apparently the thread id field was specified for some entries and not
3151  // others. Start the thread id counter off at the next higher thread id.
3152  else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
3153  threadIdCt = threadInfo[i][threadIdIndex] + 1;
3154  }
3155  }
3156  if (threadInfo[i][index] != lastId[index]) {
3157  // Run through all indices which are less significant, and reset the
3158  // counts to 1. At all levels up to and including index, we need to
3159  // increment the totals and record the last id.
3160  unsigned index2;
3161  for (index2 = threadIdIndex; index2 < index; index2++) {
3162  totals[index2]++;
3163  if (counts[index2] > maxCt[index2]) {
3164  maxCt[index2] = counts[index2];
3165  }
3166  counts[index2] = 1;
3167  lastId[index2] = threadInfo[i][index2];
3168  }
3169  counts[index]++;
3170  totals[index]++;
3171  lastId[index] = threadInfo[i][index];
3172 
3173  if (assign_thread_ids && (index > threadIdIndex)) {
3174 
3175 #if KMP_MIC && REDUCE_TEAM_SIZE
3176  // The default team size is the total #threads in the machine
3177  // minus 1 thread for every core that has 3 or more threads.
3178  teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
3179 #endif // KMP_MIC && REDUCE_TEAM_SIZE
3180 
3181  // Restart the thread counter, as we are on a new core.
3182  threadIdCt = 0;
3183 
3184  // Auto-assign the thread id field if it wasn't specified.
3185  if (threadInfo[i][threadIdIndex] == UINT_MAX) {
3186  threadInfo[i][threadIdIndex] = threadIdCt++;
3187  }
3188 
3189  // Apparently the thread id field was specified for some entries and
3190  // not others. Start the thread id counter off at the next higher
3191  // thread id.
3192  else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
3193  threadIdCt = threadInfo[i][threadIdIndex] + 1;
3194  }
3195  }
3196  break;
3197  }
3198  }
3199  if (index < threadIdIndex) {
3200  // If thread ids were specified, it is an error if they are not unique.
3201  // Also, check that we waven't already restarted the loop (to be safe -
3202  // shouldn't need to).
3203  if ((threadInfo[i][threadIdIndex] != UINT_MAX) || assign_thread_ids) {
3204  __kmp_free(lastId);
3205  __kmp_free(totals);
3206  __kmp_free(maxCt);
3207  __kmp_free(counts);
3208  CLEANUP_THREAD_INFO;
3209  *msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
3210  return false;
3211  }
3212 
3213  // If the thread ids were not specified and we see entries entries that
3214  // are duplicates, start the loop over and assign the thread ids manually.
3215  assign_thread_ids = true;
3216  goto restart_radix_check;
3217  }
3218  }
3219 
3220 #if KMP_MIC && REDUCE_TEAM_SIZE
3221  // The default team size is the total #threads in the machine
3222  // minus 1 thread for every core that has 3 or more threads.
3223  teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
3224 #endif // KMP_MIC && REDUCE_TEAM_SIZE
3225 
3226  for (index = threadIdIndex; index <= maxIndex; index++) {
3227  if (counts[index] > maxCt[index]) {
3228  maxCt[index] = counts[index];
3229  }
3230  }
3231 
3232  __kmp_nThreadsPerCore = maxCt[threadIdIndex];
3233  nCoresPerPkg = maxCt[coreIdIndex];
3234  nPackages = totals[pkgIdIndex];
3235 
3236  // When affinity is off, this routine will still be called to set
3237  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
3238  // Make sure all these vars are set correctly, and return now if affinity is
3239  // not enabled.
3240  __kmp_ncores = totals[coreIdIndex];
3241  if (!KMP_AFFINITY_CAPABLE()) {
3242  KMP_ASSERT(__kmp_affinity_type == affinity_none);
3243  return true;
3244  }
3245 
3246 #if KMP_MIC && REDUCE_TEAM_SIZE
3247  // Set the default team size.
3248  if ((__kmp_dflt_team_nth == 0) && (teamSize > 0)) {
3249  __kmp_dflt_team_nth = teamSize;
3250  KA_TRACE(20, ("__kmp_affinity_create_cpuinfo_map: setting "
3251  "__kmp_dflt_team_nth = %d\n",
3252  __kmp_dflt_team_nth));
3253  }
3254 #endif // KMP_MIC && REDUCE_TEAM_SIZE
3255 
3256  KMP_DEBUG_ASSERT(num_avail == (unsigned)__kmp_avail_proc);
3257 
3258  // Count the number of levels which have more nodes at that level than at the
3259  // parent's level (with there being an implicit root node of the top level).
3260  // This is equivalent to saying that there is at least one node at this level
3261  // which has a sibling. These levels are in the map, and the package level is
3262  // always in the map.
3263  bool *inMap = (bool *)__kmp_allocate((maxIndex + 1) * sizeof(bool));
3264  for (index = threadIdIndex; index < maxIndex; index++) {
3265  KMP_ASSERT(totals[index] >= totals[index + 1]);
3266  inMap[index] = (totals[index] > totals[index + 1]);
3267  }
3268  inMap[maxIndex] = (totals[maxIndex] > 1);
3269  inMap[pkgIdIndex] = true;
3270  inMap[coreIdIndex] = true;
3271  inMap[threadIdIndex] = true;
3272 
3273  int depth = 0;
3274  int idx = 0;
3275  kmp_hw_t types[KMP_HW_LAST];
3276  int pkgLevel = -1;
3277  int coreLevel = -1;
3278  int threadLevel = -1;
3279  for (index = threadIdIndex; index <= maxIndex; index++) {
3280  if (inMap[index]) {
3281  depth++;
3282  }
3283  }
3284  if (inMap[pkgIdIndex]) {
3285  pkgLevel = idx;
3286  types[idx++] = KMP_HW_SOCKET;
3287  }
3288  if (inMap[coreIdIndex]) {
3289  coreLevel = idx;
3290  types[idx++] = KMP_HW_CORE;
3291  }
3292  if (inMap[threadIdIndex]) {
3293  threadLevel = idx;
3294  types[idx++] = KMP_HW_THREAD;
3295  }
3296  KMP_ASSERT(depth > 0);
3297 
3298  // Construct the data structure that is to be returned.
3299  __kmp_topology = kmp_topology_t::allocate(num_avail, depth, types);
3300 
3301  for (i = 0; i < num_avail; ++i) {
3302  unsigned os = threadInfo[i][osIdIndex];
3303  int src_index;
3304  int dst_index = 0;
3305  kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
3306  hw_thread.clear();
3307  hw_thread.os_id = os;
3308 
3309  idx = 0;
3310  for (src_index = maxIndex; src_index >= threadIdIndex; src_index--) {
3311  if (!inMap[src_index]) {
3312  continue;
3313  }
3314  if (src_index == pkgIdIndex) {
3315  hw_thread.ids[pkgLevel] = threadInfo[i][src_index];
3316  } else if (src_index == coreIdIndex) {
3317  hw_thread.ids[coreLevel] = threadInfo[i][src_index];
3318  } else if (src_index == threadIdIndex) {
3319  hw_thread.ids[threadLevel] = threadInfo[i][src_index];
3320  }
3321  dst_index++;
3322  }
3323  }
3324 
3325  __kmp_free(inMap);
3326  __kmp_free(lastId);
3327  __kmp_free(totals);
3328  __kmp_free(maxCt);
3329  __kmp_free(counts);
3330  CLEANUP_THREAD_INFO;
3331  __kmp_topology->sort_ids();
3332  if (!__kmp_topology->check_ids()) {
3333  kmp_topology_t::deallocate(__kmp_topology);
3334  __kmp_topology = nullptr;
3335  *msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
3336  return false;
3337  }
3338  return true;
3339 }
3340 
3341 // Create and return a table of affinity masks, indexed by OS thread ID.
3342 // This routine handles OR'ing together all the affinity masks of threads
3343 // that are sufficiently close, if granularity > fine.
3344 static kmp_affin_mask_t *__kmp_create_masks(unsigned *maxIndex,
3345  unsigned *numUnique) {
3346  // First form a table of affinity masks in order of OS thread id.
3347  int maxOsId;
3348  int i;
3349  int numAddrs = __kmp_topology->get_num_hw_threads();
3350  int depth = __kmp_topology->get_depth();
3351  KMP_ASSERT(numAddrs);
3352  KMP_ASSERT(depth);
3353 
3354  maxOsId = 0;
3355  for (i = numAddrs - 1;; --i) {
3356  int osId = __kmp_topology->at(i).os_id;
3357  if (osId > maxOsId) {
3358  maxOsId = osId;
3359  }
3360  if (i == 0)
3361  break;
3362  }
3363  kmp_affin_mask_t *osId2Mask;
3364  KMP_CPU_ALLOC_ARRAY(osId2Mask, (maxOsId + 1));
3365  KMP_ASSERT(__kmp_affinity_gran_levels >= 0);
3366  if (__kmp_affinity_verbose && (__kmp_affinity_gran_levels > 0)) {
3367  KMP_INFORM(ThreadsMigrate, "KMP_AFFINITY", __kmp_affinity_gran_levels);
3368  }
3369  if (__kmp_affinity_gran_levels >= (int)depth) {
3370  KMP_AFF_WARNING(AffThreadsMayMigrate);
3371  }
3372 
3373  // Run through the table, forming the masks for all threads on each core.
3374  // Threads on the same core will have identical kmp_hw_thread_t objects, not
3375  // considering the last level, which must be the thread id. All threads on a
3376  // core will appear consecutively.
3377  int unique = 0;
3378  int j = 0; // index of 1st thread on core
3379  int leader = 0;
3380  kmp_affin_mask_t *sum;
3381  KMP_CPU_ALLOC_ON_STACK(sum);
3382  KMP_CPU_ZERO(sum);
3383  KMP_CPU_SET(__kmp_topology->at(0).os_id, sum);
3384  for (i = 1; i < numAddrs; i++) {
3385  // If this thread is sufficiently close to the leader (within the
3386  // granularity setting), then set the bit for this os thread in the
3387  // affinity mask for this group, and go on to the next thread.
3388  if (__kmp_topology->is_close(leader, i, __kmp_affinity_gran_levels)) {
3389  KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
3390  continue;
3391  }
3392 
3393  // For every thread in this group, copy the mask to the thread's entry in
3394  // the osId2Mask table. Mark the first address as a leader.
3395  for (; j < i; j++) {
3396  int osId = __kmp_topology->at(j).os_id;
3397  KMP_DEBUG_ASSERT(osId <= maxOsId);
3398  kmp_affin_mask_t *mask = KMP_CPU_INDEX(osId2Mask, osId);
3399  KMP_CPU_COPY(mask, sum);
3400  __kmp_topology->at(j).leader = (j == leader);
3401  }
3402  unique++;
3403 
3404  // Start a new mask.
3405  leader = i;
3406  KMP_CPU_ZERO(sum);
3407  KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
3408  }
3409 
3410  // For every thread in last group, copy the mask to the thread's
3411  // entry in the osId2Mask table.
3412  for (; j < i; j++) {
3413  int osId = __kmp_topology->at(j).os_id;
3414  KMP_DEBUG_ASSERT(osId <= maxOsId);
3415  kmp_affin_mask_t *mask = KMP_CPU_INDEX(osId2Mask, osId);
3416  KMP_CPU_COPY(mask, sum);
3417  __kmp_topology->at(j).leader = (j == leader);
3418  }
3419  unique++;
3420  KMP_CPU_FREE_FROM_STACK(sum);
3421 
3422  *maxIndex = maxOsId;
3423  *numUnique = unique;
3424  return osId2Mask;
3425 }
3426 
3427 // Stuff for the affinity proclist parsers. It's easier to declare these vars
3428 // as file-static than to try and pass them through the calling sequence of
3429 // the recursive-descent OMP_PLACES parser.
3430 static kmp_affin_mask_t *newMasks;
3431 static int numNewMasks;
3432 static int nextNewMask;
3433 
3434 #define ADD_MASK(_mask) \
3435  { \
3436  if (nextNewMask >= numNewMasks) { \
3437  int i; \
3438  numNewMasks *= 2; \
3439  kmp_affin_mask_t *temp; \
3440  KMP_CPU_INTERNAL_ALLOC_ARRAY(temp, numNewMasks); \
3441  for (i = 0; i < numNewMasks / 2; i++) { \
3442  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i); \
3443  kmp_affin_mask_t *dest = KMP_CPU_INDEX(temp, i); \
3444  KMP_CPU_COPY(dest, src); \
3445  } \
3446  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks / 2); \
3447  newMasks = temp; \
3448  } \
3449  KMP_CPU_COPY(KMP_CPU_INDEX(newMasks, nextNewMask), (_mask)); \
3450  nextNewMask++; \
3451  }
3452 
3453 #define ADD_MASK_OSID(_osId, _osId2Mask, _maxOsId) \
3454  { \
3455  if (((_osId) > _maxOsId) || \
3456  (!KMP_CPU_ISSET((_osId), KMP_CPU_INDEX((_osId2Mask), (_osId))))) { \
3457  KMP_AFF_WARNING(AffIgnoreInvalidProcID, _osId); \
3458  } else { \
3459  ADD_MASK(KMP_CPU_INDEX(_osId2Mask, (_osId))); \
3460  } \
3461  }
3462 
3463 // Re-parse the proclist (for the explicit affinity type), and form the list
3464 // of affinity newMasks indexed by gtid.
3465 static void __kmp_affinity_process_proclist(kmp_affin_mask_t **out_masks,
3466  unsigned int *out_numMasks,
3467  const char *proclist,
3468  kmp_affin_mask_t *osId2Mask,
3469  int maxOsId) {
3470  int i;
3471  const char *scan = proclist;
3472  const char *next = proclist;
3473 
3474  // We use malloc() for the temporary mask vector, so that we can use
3475  // realloc() to extend it.
3476  numNewMasks = 2;
3477  KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
3478  nextNewMask = 0;
3479  kmp_affin_mask_t *sumMask;
3480  KMP_CPU_ALLOC(sumMask);
3481  int setSize = 0;
3482 
3483  for (;;) {
3484  int start, end, stride;
3485 
3486  SKIP_WS(scan);
3487  next = scan;
3488  if (*next == '\0') {
3489  break;
3490  }
3491 
3492  if (*next == '{') {
3493  int num;
3494  setSize = 0;
3495  next++; // skip '{'
3496  SKIP_WS(next);
3497  scan = next;
3498 
3499  // Read the first integer in the set.
3500  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad proclist");
3501  SKIP_DIGITS(next);
3502  num = __kmp_str_to_int(scan, *next);
3503  KMP_ASSERT2(num >= 0, "bad explicit proc list");
3504 
3505  // Copy the mask for that osId to the sum (union) mask.
3506  if ((num > maxOsId) ||
3507  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
3508  KMP_AFF_WARNING(AffIgnoreInvalidProcID, num);
3509  KMP_CPU_ZERO(sumMask);
3510  } else {
3511  KMP_CPU_COPY(sumMask, KMP_CPU_INDEX(osId2Mask, num));
3512  setSize = 1;
3513  }
3514 
3515  for (;;) {
3516  // Check for end of set.
3517  SKIP_WS(next);
3518  if (*next == '}') {
3519  next++; // skip '}'
3520  break;
3521  }
3522 
3523  // Skip optional comma.
3524  if (*next == ',') {
3525  next++;
3526  }
3527  SKIP_WS(next);
3528 
3529  // Read the next integer in the set.
3530  scan = next;
3531  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
3532 
3533  SKIP_DIGITS(next);
3534  num = __kmp_str_to_int(scan, *next);
3535  KMP_ASSERT2(num >= 0, "bad explicit proc list");
3536 
3537  // Add the mask for that osId to the sum mask.
3538  if ((num > maxOsId) ||
3539  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
3540  KMP_AFF_WARNING(AffIgnoreInvalidProcID, num);
3541  } else {
3542  KMP_CPU_UNION(sumMask, KMP_CPU_INDEX(osId2Mask, num));
3543  setSize++;
3544  }
3545  }
3546  if (setSize > 0) {
3547  ADD_MASK(sumMask);
3548  }
3549 
3550  SKIP_WS(next);
3551  if (*next == ',') {
3552  next++;
3553  }
3554  scan = next;
3555  continue;
3556  }
3557 
3558  // Read the first integer.
3559  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
3560  SKIP_DIGITS(next);
3561  start = __kmp_str_to_int(scan, *next);
3562  KMP_ASSERT2(start >= 0, "bad explicit proc list");
3563  SKIP_WS(next);
3564 
3565  // If this isn't a range, then add a mask to the list and go on.
3566  if (*next != '-') {
3567  ADD_MASK_OSID(start, osId2Mask, maxOsId);
3568 
3569  // Skip optional comma.
3570  if (*next == ',') {
3571  next++;
3572  }
3573  scan = next;
3574  continue;
3575  }
3576 
3577  // This is a range. Skip over the '-' and read in the 2nd int.
3578  next++; // skip '-'
3579  SKIP_WS(next);
3580  scan = next;
3581  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
3582  SKIP_DIGITS(next);
3583  end = __kmp_str_to_int(scan, *next);
3584  KMP_ASSERT2(end >= 0, "bad explicit proc list");
3585 
3586  // Check for a stride parameter
3587  stride = 1;
3588  SKIP_WS(next);
3589  if (*next == ':') {
3590  // A stride is specified. Skip over the ':" and read the 3rd int.
3591  int sign = +1;
3592  next++; // skip ':'
3593  SKIP_WS(next);
3594  scan = next;
3595  if (*next == '-') {
3596  sign = -1;
3597  next++;
3598  SKIP_WS(next);
3599  scan = next;
3600  }
3601  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
3602  SKIP_DIGITS(next);
3603  stride = __kmp_str_to_int(scan, *next);
3604  KMP_ASSERT2(stride >= 0, "bad explicit proc list");
3605  stride *= sign;
3606  }
3607 
3608  // Do some range checks.
3609  KMP_ASSERT2(stride != 0, "bad explicit proc list");
3610  if (stride > 0) {
3611  KMP_ASSERT2(start <= end, "bad explicit proc list");
3612  } else {
3613  KMP_ASSERT2(start >= end, "bad explicit proc list");
3614  }
3615  KMP_ASSERT2((end - start) / stride <= 65536, "bad explicit proc list");
3616 
3617  // Add the mask for each OS proc # to the list.
3618  if (stride > 0) {
3619  do {
3620  ADD_MASK_OSID(start, osId2Mask, maxOsId);
3621  start += stride;
3622  } while (start <= end);
3623  } else {
3624  do {
3625  ADD_MASK_OSID(start, osId2Mask, maxOsId);
3626  start += stride;
3627  } while (start >= end);
3628  }
3629 
3630  // Skip optional comma.
3631  SKIP_WS(next);
3632  if (*next == ',') {
3633  next++;
3634  }
3635  scan = next;
3636  }
3637 
3638  *out_numMasks = nextNewMask;
3639  if (nextNewMask == 0) {
3640  *out_masks = NULL;
3641  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3642  return;
3643  }
3644  KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
3645  for (i = 0; i < nextNewMask; i++) {
3646  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
3647  kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
3648  KMP_CPU_COPY(dest, src);
3649  }
3650  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3651  KMP_CPU_FREE(sumMask);
3652 }
3653 
3654 /*-----------------------------------------------------------------------------
3655 Re-parse the OMP_PLACES proc id list, forming the newMasks for the different
3656 places. Again, Here is the grammar:
3657 
3658 place_list := place
3659 place_list := place , place_list
3660 place := num
3661 place := place : num
3662 place := place : num : signed
3663 place := { subplacelist }
3664 place := ! place // (lowest priority)
3665 subplace_list := subplace
3666 subplace_list := subplace , subplace_list
3667 subplace := num
3668 subplace := num : num
3669 subplace := num : num : signed
3670 signed := num
3671 signed := + signed
3672 signed := - signed
3673 -----------------------------------------------------------------------------*/
3674 static void __kmp_process_subplace_list(const char **scan,
3675  kmp_affin_mask_t *osId2Mask,
3676  int maxOsId, kmp_affin_mask_t *tempMask,
3677  int *setSize) {
3678  const char *next;
3679 
3680  for (;;) {
3681  int start, count, stride, i;
3682 
3683  // Read in the starting proc id
3684  SKIP_WS(*scan);
3685  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3686  next = *scan;
3687  SKIP_DIGITS(next);
3688  start = __kmp_str_to_int(*scan, *next);
3689  KMP_ASSERT(start >= 0);
3690  *scan = next;
3691 
3692  // valid follow sets are ',' ':' and '}'
3693  SKIP_WS(*scan);
3694  if (**scan == '}' || **scan == ',') {
3695  if ((start > maxOsId) ||
3696  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3697  KMP_AFF_WARNING(AffIgnoreInvalidProcID, start);
3698  } else {
3699  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3700  (*setSize)++;
3701  }
3702  if (**scan == '}') {
3703  break;
3704  }
3705  (*scan)++; // skip ','
3706  continue;
3707  }
3708  KMP_ASSERT2(**scan == ':', "bad explicit places list");
3709  (*scan)++; // skip ':'
3710 
3711  // Read count parameter
3712  SKIP_WS(*scan);
3713  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3714  next = *scan;
3715  SKIP_DIGITS(next);
3716  count = __kmp_str_to_int(*scan, *next);
3717  KMP_ASSERT(count >= 0);
3718  *scan = next;
3719 
3720  // valid follow sets are ',' ':' and '}'
3721  SKIP_WS(*scan);
3722  if (**scan == '}' || **scan == ',') {
3723  for (i = 0; i < count; i++) {
3724  if ((start > maxOsId) ||
3725  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3726  KMP_AFF_WARNING(AffIgnoreInvalidProcID, start);
3727  break; // don't proliferate warnings for large count
3728  } else {
3729  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3730  start++;
3731  (*setSize)++;
3732  }
3733  }
3734  if (**scan == '}') {
3735  break;
3736  }
3737  (*scan)++; // skip ','
3738  continue;
3739  }
3740  KMP_ASSERT2(**scan == ':', "bad explicit places list");
3741  (*scan)++; // skip ':'
3742 
3743  // Read stride parameter
3744  int sign = +1;
3745  for (;;) {
3746  SKIP_WS(*scan);
3747  if (**scan == '+') {
3748  (*scan)++; // skip '+'
3749  continue;
3750  }
3751  if (**scan == '-') {
3752  sign *= -1;
3753  (*scan)++; // skip '-'
3754  continue;
3755  }
3756  break;
3757  }
3758  SKIP_WS(*scan);
3759  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3760  next = *scan;
3761  SKIP_DIGITS(next);
3762  stride = __kmp_str_to_int(*scan, *next);
3763  KMP_ASSERT(stride >= 0);
3764  *scan = next;
3765  stride *= sign;
3766 
3767  // valid follow sets are ',' and '}'
3768  SKIP_WS(*scan);
3769  if (**scan == '}' || **scan == ',') {
3770  for (i = 0; i < count; i++) {
3771  if ((start > maxOsId) ||
3772  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3773  KMP_AFF_WARNING(AffIgnoreInvalidProcID, start);
3774  break; // don't proliferate warnings for large count
3775  } else {
3776  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3777  start += stride;
3778  (*setSize)++;
3779  }
3780  }
3781  if (**scan == '}') {
3782  break;
3783  }
3784  (*scan)++; // skip ','
3785  continue;
3786  }
3787 
3788  KMP_ASSERT2(0, "bad explicit places list");
3789  }
3790 }
3791 
3792 static void __kmp_process_place(const char **scan, kmp_affin_mask_t *osId2Mask,
3793  int maxOsId, kmp_affin_mask_t *tempMask,
3794  int *setSize) {
3795  const char *next;
3796 
3797  // valid follow sets are '{' '!' and num
3798  SKIP_WS(*scan);
3799  if (**scan == '{') {
3800  (*scan)++; // skip '{'
3801  __kmp_process_subplace_list(scan, osId2Mask, maxOsId, tempMask, setSize);
3802  KMP_ASSERT2(**scan == '}', "bad explicit places list");
3803  (*scan)++; // skip '}'
3804  } else if (**scan == '!') {
3805  (*scan)++; // skip '!'
3806  __kmp_process_place(scan, osId2Mask, maxOsId, tempMask, setSize);
3807  KMP_CPU_COMPLEMENT(maxOsId, tempMask);
3808  } else if ((**scan >= '0') && (**scan <= '9')) {
3809  next = *scan;
3810  SKIP_DIGITS(next);
3811  int num = __kmp_str_to_int(*scan, *next);
3812  KMP_ASSERT(num >= 0);
3813  if ((num > maxOsId) ||
3814  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
3815  KMP_AFF_WARNING(AffIgnoreInvalidProcID, num);
3816  } else {
3817  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, num));
3818  (*setSize)++;
3819  }
3820  *scan = next; // skip num
3821  } else {
3822  KMP_ASSERT2(0, "bad explicit places list");
3823  }
3824 }
3825 
3826 // static void
3827 void __kmp_affinity_process_placelist(kmp_affin_mask_t **out_masks,
3828  unsigned int *out_numMasks,
3829  const char *placelist,
3830  kmp_affin_mask_t *osId2Mask,
3831  int maxOsId) {
3832  int i, j, count, stride, sign;
3833  const char *scan = placelist;
3834  const char *next = placelist;
3835 
3836  numNewMasks = 2;
3837  KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
3838  nextNewMask = 0;
3839 
3840  // tempMask is modified based on the previous or initial
3841  // place to form the current place
3842  // previousMask contains the previous place
3843  kmp_affin_mask_t *tempMask;
3844  kmp_affin_mask_t *previousMask;
3845  KMP_CPU_ALLOC(tempMask);
3846  KMP_CPU_ZERO(tempMask);
3847  KMP_CPU_ALLOC(previousMask);
3848  KMP_CPU_ZERO(previousMask);
3849  int setSize = 0;
3850 
3851  for (;;) {
3852  __kmp_process_place(&scan, osId2Mask, maxOsId, tempMask, &setSize);
3853 
3854  // valid follow sets are ',' ':' and EOL
3855  SKIP_WS(scan);
3856  if (*scan == '\0' || *scan == ',') {
3857  if (setSize > 0) {
3858  ADD_MASK(tempMask);
3859  }
3860  KMP_CPU_ZERO(tempMask);
3861  setSize = 0;
3862  if (*scan == '\0') {
3863  break;
3864  }
3865  scan++; // skip ','
3866  continue;
3867  }
3868 
3869  KMP_ASSERT2(*scan == ':', "bad explicit places list");
3870  scan++; // skip ':'
3871 
3872  // Read count parameter
3873  SKIP_WS(scan);
3874  KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
3875  next = scan;
3876  SKIP_DIGITS(next);
3877  count = __kmp_str_to_int(scan, *next);
3878  KMP_ASSERT(count >= 0);
3879  scan = next;
3880 
3881  // valid follow sets are ',' ':' and EOL
3882  SKIP_WS(scan);
3883  if (*scan == '\0' || *scan == ',') {
3884  stride = +1;
3885  } else {
3886  KMP_ASSERT2(*scan == ':', "bad explicit places list");
3887  scan++; // skip ':'
3888 
3889  // Read stride parameter
3890  sign = +1;
3891  for (;;) {
3892  SKIP_WS(scan);
3893  if (*scan == '+') {
3894  scan++; // skip '+'
3895  continue;
3896  }
3897  if (*scan == '-') {
3898  sign *= -1;
3899  scan++; // skip '-'
3900  continue;
3901  }
3902  break;
3903  }
3904  SKIP_WS(scan);
3905  KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
3906  next = scan;
3907  SKIP_DIGITS(next);
3908  stride = __kmp_str_to_int(scan, *next);
3909  KMP_DEBUG_ASSERT(stride >= 0);
3910  scan = next;
3911  stride *= sign;
3912  }
3913 
3914  // Add places determined by initial_place : count : stride
3915  for (i = 0; i < count; i++) {
3916  if (setSize == 0) {
3917  break;
3918  }
3919  // Add the current place, then build the next place (tempMask) from that
3920  KMP_CPU_COPY(previousMask, tempMask);
3921  ADD_MASK(previousMask);
3922  KMP_CPU_ZERO(tempMask);
3923  setSize = 0;
3924  KMP_CPU_SET_ITERATE(j, previousMask) {
3925  if (!KMP_CPU_ISSET(j, previousMask)) {
3926  continue;
3927  }
3928  if ((j + stride > maxOsId) || (j + stride < 0) ||
3929  (!KMP_CPU_ISSET(j, __kmp_affin_fullMask)) ||
3930  (!KMP_CPU_ISSET(j + stride,
3931  KMP_CPU_INDEX(osId2Mask, j + stride)))) {
3932  if (i < count - 1) {
3933  KMP_AFF_WARNING(AffIgnoreInvalidProcID, j + stride);
3934  }
3935  continue;
3936  }
3937  KMP_CPU_SET(j + stride, tempMask);
3938  setSize++;
3939  }
3940  }
3941  KMP_CPU_ZERO(tempMask);
3942  setSize = 0;
3943 
3944  // valid follow sets are ',' and EOL
3945  SKIP_WS(scan);
3946  if (*scan == '\0') {
3947  break;
3948  }
3949  if (*scan == ',') {
3950  scan++; // skip ','
3951  continue;
3952  }
3953 
3954  KMP_ASSERT2(0, "bad explicit places list");
3955  }
3956 
3957  *out_numMasks = nextNewMask;
3958  if (nextNewMask == 0) {
3959  *out_masks = NULL;
3960  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3961  return;
3962  }
3963  KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
3964  KMP_CPU_FREE(tempMask);
3965  KMP_CPU_FREE(previousMask);
3966  for (i = 0; i < nextNewMask; i++) {
3967  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
3968  kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
3969  KMP_CPU_COPY(dest, src);
3970  }
3971  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3972 }
3973 
3974 #undef ADD_MASK
3975 #undef ADD_MASK_OSID
3976 
3977 // This function figures out the deepest level at which there is at least one
3978 // cluster/core with more than one processing unit bound to it.
3979 static int __kmp_affinity_find_core_level(int nprocs, int bottom_level) {
3980  int core_level = 0;
3981 
3982  for (int i = 0; i < nprocs; i++) {
3983  const kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
3984  for (int j = bottom_level; j > 0; j--) {
3985  if (hw_thread.ids[j] > 0) {
3986  if (core_level < (j - 1)) {
3987  core_level = j - 1;
3988  }
3989  }
3990  }
3991  }
3992  return core_level;
3993 }
3994 
3995 // This function counts number of clusters/cores at given level.
3996 static int __kmp_affinity_compute_ncores(int nprocs, int bottom_level,
3997  int core_level) {
3998  return __kmp_topology->get_count(core_level);
3999 }
4000 // This function finds to which cluster/core given processing unit is bound.
4001 static int __kmp_affinity_find_core(int proc, int bottom_level,
4002  int core_level) {
4003  int core = 0;
4004  KMP_DEBUG_ASSERT(proc >= 0 && proc < __kmp_topology->get_num_hw_threads());
4005  for (int i = 0; i <= proc; ++i) {
4006  if (i + 1 <= proc) {
4007  for (int j = 0; j <= core_level; ++j) {
4008  if (__kmp_topology->at(i + 1).sub_ids[j] !=
4009  __kmp_topology->at(i).sub_ids[j]) {
4010  core++;
4011  break;
4012  }
4013  }
4014  }
4015  }
4016  return core;
4017 }
4018 
4019 // This function finds maximal number of processing units bound to a
4020 // cluster/core at given level.
4021 static int __kmp_affinity_max_proc_per_core(int nprocs, int bottom_level,
4022  int core_level) {
4023  if (core_level >= bottom_level)
4024  return 1;
4025  int thread_level = __kmp_topology->get_level(KMP_HW_THREAD);
4026  return __kmp_topology->calculate_ratio(thread_level, core_level);
4027 }
4028 
4029 static int *procarr = NULL;
4030 static int __kmp_aff_depth = 0;
4031 
4032 // Create a one element mask array (set of places) which only contains the
4033 // initial process's affinity mask
4034 static void __kmp_create_affinity_none_places() {
4035  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4036  KMP_ASSERT(__kmp_affinity_type == affinity_none);
4037  __kmp_affinity_num_masks = 1;
4038  KMP_CPU_ALLOC_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks);
4039  kmp_affin_mask_t *dest = KMP_CPU_INDEX(__kmp_affinity_masks, 0);
4040  KMP_CPU_COPY(dest, __kmp_affin_fullMask);
4041 }
4042 
4043 static void __kmp_aux_affinity_initialize(void) {
4044  if (__kmp_affinity_masks != NULL) {
4045  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4046  return;
4047  }
4048 
4049  // Create the "full" mask - this defines all of the processors that we
4050  // consider to be in the machine model. If respect is set, then it is the
4051  // initialization thread's affinity mask. Otherwise, it is all processors that
4052  // we know about on the machine.
4053  if (__kmp_affin_fullMask == NULL) {
4054  KMP_CPU_ALLOC(__kmp_affin_fullMask);
4055  }
4056  if (__kmp_affin_origMask == NULL) {
4057  KMP_CPU_ALLOC(__kmp_affin_origMask);
4058  }
4059  if (KMP_AFFINITY_CAPABLE()) {
4060  __kmp_get_system_affinity(__kmp_affin_fullMask, TRUE);
4061  // Make a copy before possible expanding to the entire machine mask
4062  __kmp_affin_origMask->copy(__kmp_affin_fullMask);
4063  if (__kmp_affinity_respect_mask) {
4064  // Count the number of available processors.
4065  unsigned i;
4066  __kmp_avail_proc = 0;
4067  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
4068  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
4069  continue;
4070  }
4071  __kmp_avail_proc++;
4072  }
4073  if (__kmp_avail_proc > __kmp_xproc) {
4074  KMP_AFF_WARNING(ErrorInitializeAffinity);
4075  __kmp_affinity_type = affinity_none;
4076  KMP_AFFINITY_DISABLE();
4077  return;
4078  }
4079 
4080  if (__kmp_affinity_verbose) {
4081  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4082  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4083  __kmp_affin_fullMask);
4084  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
4085  }
4086  } else {
4087  if (__kmp_affinity_verbose) {
4088  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4089  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4090  __kmp_affin_fullMask);
4091  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
4092  }
4093  __kmp_avail_proc =
4094  __kmp_affinity_entire_machine_mask(__kmp_affin_fullMask);
4095 #if KMP_OS_WINDOWS
4096  if (__kmp_num_proc_groups <= 1) {
4097  // Copy expanded full mask if topology has single processor group
4098  __kmp_affin_origMask->copy(__kmp_affin_fullMask);
4099  }
4100  // Set the process affinity mask since threads' affinity
4101  // masks must be subset of process mask in Windows* OS
4102  __kmp_affin_fullMask->set_process_affinity(true);
4103 #endif
4104  }
4105  }
4106 
4107  kmp_i18n_id_t msg_id = kmp_i18n_null;
4108 
4109  // For backward compatibility, setting KMP_CPUINFO_FILE =>
4110  // KMP_TOPOLOGY_METHOD=cpuinfo
4111  if ((__kmp_cpuinfo_file != NULL) &&
4112  (__kmp_affinity_top_method == affinity_top_method_all)) {
4113  __kmp_affinity_top_method = affinity_top_method_cpuinfo;
4114  }
4115 
4116  bool success = false;
4117  if (__kmp_affinity_top_method == affinity_top_method_all) {
4118 // In the default code path, errors are not fatal - we just try using
4119 // another method. We only emit a warning message if affinity is on, or the
4120 // verbose flag is set, an the nowarnings flag was not set.
4121 #if KMP_USE_HWLOC
4122  if (!success &&
4123  __kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC) {
4124  if (!__kmp_hwloc_error) {
4125  success = __kmp_affinity_create_hwloc_map(&msg_id);
4126  if (!success && __kmp_affinity_verbose) {
4127  KMP_INFORM(AffIgnoringHwloc, "KMP_AFFINITY");
4128  }
4129  } else if (__kmp_affinity_verbose) {
4130  KMP_INFORM(AffIgnoringHwloc, "KMP_AFFINITY");
4131  }
4132  }
4133 #endif
4134 
4135 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4136  if (!success) {
4137  success = __kmp_affinity_create_x2apicid_map(&msg_id);
4138  if (!success && __kmp_affinity_verbose && msg_id != kmp_i18n_null) {
4139  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", __kmp_i18n_catgets(msg_id));
4140  }
4141  }
4142  if (!success) {
4143  success = __kmp_affinity_create_apicid_map(&msg_id);
4144  if (!success && __kmp_affinity_verbose && msg_id != kmp_i18n_null) {
4145  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", __kmp_i18n_catgets(msg_id));
4146  }
4147  }
4148 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4149 
4150 #if KMP_OS_LINUX
4151  if (!success) {
4152  int line = 0;
4153  success = __kmp_affinity_create_cpuinfo_map(&line, &msg_id);
4154  if (!success && __kmp_affinity_verbose && msg_id != kmp_i18n_null) {
4155  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", __kmp_i18n_catgets(msg_id));
4156  }
4157  }
4158 #endif /* KMP_OS_LINUX */
4159 
4160 #if KMP_GROUP_AFFINITY
4161  if (!success && (__kmp_num_proc_groups > 1)) {
4162  success = __kmp_affinity_create_proc_group_map(&msg_id);
4163  if (!success && __kmp_affinity_verbose && msg_id != kmp_i18n_null) {
4164  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", __kmp_i18n_catgets(msg_id));
4165  }
4166  }
4167 #endif /* KMP_GROUP_AFFINITY */
4168 
4169  if (!success) {
4170  success = __kmp_affinity_create_flat_map(&msg_id);
4171  if (!success && __kmp_affinity_verbose && msg_id != kmp_i18n_null) {
4172  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", __kmp_i18n_catgets(msg_id));
4173  }
4174  KMP_ASSERT(success);
4175  }
4176  }
4177 
4178 // If the user has specified that a paricular topology discovery method is to be
4179 // used, then we abort if that method fails. The exception is group affinity,
4180 // which might have been implicitly set.
4181 #if KMP_USE_HWLOC
4182  else if (__kmp_affinity_top_method == affinity_top_method_hwloc) {
4183  KMP_ASSERT(__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC);
4184  success = __kmp_affinity_create_hwloc_map(&msg_id);
4185  if (!success) {
4186  KMP_ASSERT(msg_id != kmp_i18n_null);
4187  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4188  }
4189  }
4190 #endif // KMP_USE_HWLOC
4191 
4192 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4193  else if (__kmp_affinity_top_method == affinity_top_method_x2apicid ||
4194  __kmp_affinity_top_method == affinity_top_method_x2apicid_1f) {
4195  success = __kmp_affinity_create_x2apicid_map(&msg_id);
4196  if (!success) {
4197  KMP_ASSERT(msg_id != kmp_i18n_null);
4198  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4199  }
4200  } else if (__kmp_affinity_top_method == affinity_top_method_apicid) {
4201  success = __kmp_affinity_create_apicid_map(&msg_id);
4202  if (!success) {
4203  KMP_ASSERT(msg_id != kmp_i18n_null);
4204  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4205  }
4206  }
4207 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4208 
4209  else if (__kmp_affinity_top_method == affinity_top_method_cpuinfo) {
4210  int line = 0;
4211  success = __kmp_affinity_create_cpuinfo_map(&line, &msg_id);
4212  if (!success) {
4213  KMP_ASSERT(msg_id != kmp_i18n_null);
4214  const char *filename = __kmp_cpuinfo_get_filename();
4215  if (line > 0) {
4216  KMP_FATAL(FileLineMsgExiting, filename, line,
4217  __kmp_i18n_catgets(msg_id));
4218  } else {
4219  KMP_FATAL(FileMsgExiting, filename, __kmp_i18n_catgets(msg_id));
4220  }
4221  }
4222  }
4223 
4224 #if KMP_GROUP_AFFINITY
4225  else if (__kmp_affinity_top_method == affinity_top_method_group) {
4226  success = __kmp_affinity_create_proc_group_map(&msg_id);
4227  KMP_ASSERT(success);
4228  if (!success) {
4229  KMP_ASSERT(msg_id != kmp_i18n_null);
4230  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4231  }
4232  }
4233 #endif /* KMP_GROUP_AFFINITY */
4234 
4235  else if (__kmp_affinity_top_method == affinity_top_method_flat) {
4236  success = __kmp_affinity_create_flat_map(&msg_id);
4237  // should not fail
4238  KMP_ASSERT(success);
4239  }
4240 
4241  // Early exit if topology could not be created
4242  if (!__kmp_topology) {
4243  if (KMP_AFFINITY_CAPABLE()) {
4244  KMP_AFF_WARNING(ErrorInitializeAffinity);
4245  }
4246  if (nPackages > 0 && nCoresPerPkg > 0 && __kmp_nThreadsPerCore > 0 &&
4247  __kmp_ncores > 0) {
4248  __kmp_topology = kmp_topology_t::allocate(0, 0, NULL);
4249  __kmp_topology->canonicalize(nPackages, nCoresPerPkg,
4250  __kmp_nThreadsPerCore, __kmp_ncores);
4251  if (__kmp_affinity_verbose) {
4252  __kmp_topology->print("KMP_AFFINITY");
4253  }
4254  }
4255  __kmp_affinity_type = affinity_none;
4256  __kmp_create_affinity_none_places();
4257 #if KMP_USE_HIER_SCHED
4258  __kmp_dispatch_set_hierarchy_values();
4259 #endif
4260  KMP_AFFINITY_DISABLE();
4261  return;
4262  }
4263 
4264  // Canonicalize, print (if requested), apply KMP_HW_SUBSET, and
4265  // initialize other data structures which depend on the topology
4266  __kmp_topology->canonicalize();
4267  if (__kmp_affinity_verbose)
4268  __kmp_topology->print("KMP_AFFINITY");
4269  bool filtered = __kmp_topology->filter_hw_subset();
4270  if (filtered) {
4271 #if KMP_OS_WINDOWS
4272  // Copy filtered full mask if topology has single processor group
4273  if (__kmp_num_proc_groups <= 1)
4274 #endif
4275  __kmp_affin_origMask->copy(__kmp_affin_fullMask);
4276  }
4277  if (filtered && __kmp_affinity_verbose)
4278  __kmp_topology->print("KMP_HW_SUBSET");
4279  machine_hierarchy.init(__kmp_topology->get_num_hw_threads());
4280  KMP_ASSERT(__kmp_avail_proc == __kmp_topology->get_num_hw_threads());
4281  // If KMP_AFFINITY=none, then only create the single "none" place
4282  // which is the process's initial affinity mask or the number of
4283  // hardware threads depending on respect,norespect
4284  if (__kmp_affinity_type == affinity_none) {
4285  __kmp_create_affinity_none_places();
4286 #if KMP_USE_HIER_SCHED
4287  __kmp_dispatch_set_hierarchy_values();
4288 #endif
4289  return;
4290  }
4291  int depth = __kmp_topology->get_depth();
4292 
4293  // Create the table of masks, indexed by thread Id.
4294  unsigned maxIndex;
4295  unsigned numUnique;
4296  kmp_affin_mask_t *osId2Mask = __kmp_create_masks(&maxIndex, &numUnique);
4297  if (__kmp_affinity_gran_levels == 0) {
4298  KMP_DEBUG_ASSERT((int)numUnique == __kmp_avail_proc);
4299  }
4300 
4301  switch (__kmp_affinity_type) {
4302 
4303  case affinity_explicit:
4304  KMP_DEBUG_ASSERT(__kmp_affinity_proclist != NULL);
4305  if (__kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) {
4306  __kmp_affinity_process_proclist(
4307  &__kmp_affinity_masks, &__kmp_affinity_num_masks,
4308  __kmp_affinity_proclist, osId2Mask, maxIndex);
4309  } else {
4310  __kmp_affinity_process_placelist(
4311  &__kmp_affinity_masks, &__kmp_affinity_num_masks,
4312  __kmp_affinity_proclist, osId2Mask, maxIndex);
4313  }
4314  if (__kmp_affinity_num_masks == 0) {
4315  KMP_AFF_WARNING(AffNoValidProcID);
4316  __kmp_affinity_type = affinity_none;
4317  __kmp_create_affinity_none_places();
4318  return;
4319  }
4320  break;
4321 
4322  // The other affinity types rely on sorting the hardware threads according to
4323  // some permutation of the machine topology tree. Set __kmp_affinity_compact
4324  // and __kmp_affinity_offset appropriately, then jump to a common code
4325  // fragment to do the sort and create the array of affinity masks.
4326  case affinity_logical:
4327  __kmp_affinity_compact = 0;
4328  if (__kmp_affinity_offset) {
4329  __kmp_affinity_offset =
4330  __kmp_nThreadsPerCore * __kmp_affinity_offset % __kmp_avail_proc;
4331  }
4332  goto sortTopology;
4333 
4334  case affinity_physical:
4335  if (__kmp_nThreadsPerCore > 1) {
4336  __kmp_affinity_compact = 1;
4337  if (__kmp_affinity_compact >= depth) {
4338  __kmp_affinity_compact = 0;
4339  }
4340  } else {
4341  __kmp_affinity_compact = 0;
4342  }
4343  if (__kmp_affinity_offset) {
4344  __kmp_affinity_offset =
4345  __kmp_nThreadsPerCore * __kmp_affinity_offset % __kmp_avail_proc;
4346  }
4347  goto sortTopology;
4348 
4349  case affinity_scatter:
4350  if (__kmp_affinity_compact >= depth) {
4351  __kmp_affinity_compact = 0;
4352  } else {
4353  __kmp_affinity_compact = depth - 1 - __kmp_affinity_compact;
4354  }
4355  goto sortTopology;
4356 
4357  case affinity_compact:
4358  if (__kmp_affinity_compact >= depth) {
4359  __kmp_affinity_compact = depth - 1;
4360  }
4361  goto sortTopology;
4362 
4363  case affinity_balanced:
4364  if (depth <= 1) {
4365  KMP_AFF_WARNING(AffBalancedNotAvail, "KMP_AFFINITY");
4366  __kmp_affinity_type = affinity_none;
4367  __kmp_create_affinity_none_places();
4368  return;
4369  } else if (!__kmp_topology->is_uniform()) {
4370  // Save the depth for further usage
4371  __kmp_aff_depth = depth;
4372 
4373  int core_level =
4374  __kmp_affinity_find_core_level(__kmp_avail_proc, depth - 1);
4375  int ncores = __kmp_affinity_compute_ncores(__kmp_avail_proc, depth - 1,
4376  core_level);
4377  int maxprocpercore = __kmp_affinity_max_proc_per_core(
4378  __kmp_avail_proc, depth - 1, core_level);
4379 
4380  int nproc = ncores * maxprocpercore;
4381  if ((nproc < 2) || (nproc < __kmp_avail_proc)) {
4382  KMP_AFF_WARNING(AffBalancedNotAvail, "KMP_AFFINITY");
4383  __kmp_affinity_type = affinity_none;
4384  return;
4385  }
4386 
4387  procarr = (int *)__kmp_allocate(sizeof(int) * nproc);
4388  for (int i = 0; i < nproc; i++) {
4389  procarr[i] = -1;
4390  }
4391 
4392  int lastcore = -1;
4393  int inlastcore = 0;
4394  for (int i = 0; i < __kmp_avail_proc; i++) {
4395  int proc = __kmp_topology->at(i).os_id;
4396  int core = __kmp_affinity_find_core(i, depth - 1, core_level);
4397 
4398  if (core == lastcore) {
4399  inlastcore++;
4400  } else {
4401  inlastcore = 0;
4402  }
4403  lastcore = core;
4404 
4405  procarr[core * maxprocpercore + inlastcore] = proc;
4406  }
4407  }
4408  if (__kmp_affinity_compact >= depth) {
4409  __kmp_affinity_compact = depth - 1;
4410  }
4411 
4412  sortTopology:
4413  // Allocate the gtid->affinity mask table.
4414  if (__kmp_affinity_dups) {
4415  __kmp_affinity_num_masks = __kmp_avail_proc;
4416  } else {
4417  __kmp_affinity_num_masks = numUnique;
4418  }
4419 
4420  if ((__kmp_nested_proc_bind.bind_types[0] != proc_bind_intel) &&
4421  (__kmp_affinity_num_places > 0) &&
4422  ((unsigned)__kmp_affinity_num_places < __kmp_affinity_num_masks)) {
4423  __kmp_affinity_num_masks = __kmp_affinity_num_places;
4424  }
4425 
4426  KMP_CPU_ALLOC_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks);
4427 
4428  // Sort the topology table according to the current setting of
4429  // __kmp_affinity_compact, then fill out __kmp_affinity_masks.
4430  __kmp_topology->sort_compact();
4431  {
4432  int i;
4433  unsigned j;
4434  int num_hw_threads = __kmp_topology->get_num_hw_threads();
4435  for (i = 0, j = 0; i < num_hw_threads; i++) {
4436  if ((!__kmp_affinity_dups) && (!__kmp_topology->at(i).leader)) {
4437  continue;
4438  }
4439  int osId = __kmp_topology->at(i).os_id;
4440 
4441  kmp_affin_mask_t *src = KMP_CPU_INDEX(osId2Mask, osId);
4442  kmp_affin_mask_t *dest = KMP_CPU_INDEX(__kmp_affinity_masks, j);
4443  KMP_ASSERT(KMP_CPU_ISSET(osId, src));
4444  KMP_CPU_COPY(dest, src);
4445  if (++j >= __kmp_affinity_num_masks) {
4446  break;
4447  }
4448  }
4449  KMP_DEBUG_ASSERT(j == __kmp_affinity_num_masks);
4450  }
4451  // Sort the topology back using ids
4452  __kmp_topology->sort_ids();
4453  break;
4454 
4455  default:
4456  KMP_ASSERT2(0, "Unexpected affinity setting");
4457  }
4458 
4459  KMP_CPU_FREE_ARRAY(osId2Mask, maxIndex + 1);
4460 }
4461 
4462 void __kmp_affinity_initialize(void) {
4463  // Much of the code above was written assuming that if a machine was not
4464  // affinity capable, then __kmp_affinity_type == affinity_none. We now
4465  // explicitly represent this as __kmp_affinity_type == affinity_disabled.
4466  // There are too many checks for __kmp_affinity_type == affinity_none
4467  // in this code. Instead of trying to change them all, check if
4468  // __kmp_affinity_type == affinity_disabled, and if so, slam it with
4469  // affinity_none, call the real initialization routine, then restore
4470  // __kmp_affinity_type to affinity_disabled.
4471  int disabled = (__kmp_affinity_type == affinity_disabled);
4472  if (!KMP_AFFINITY_CAPABLE()) {
4473  KMP_ASSERT(disabled);
4474  }
4475  if (disabled) {
4476  __kmp_affinity_type = affinity_none;
4477  }
4478  __kmp_aux_affinity_initialize();
4479  if (disabled) {
4480  __kmp_affinity_type = affinity_disabled;
4481  }
4482 }
4483 
4484 void __kmp_affinity_uninitialize(void) {
4485  if (__kmp_affinity_masks != NULL) {
4486  KMP_CPU_FREE_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks);
4487  __kmp_affinity_masks = NULL;
4488  }
4489  if (__kmp_affin_fullMask != NULL) {
4490  KMP_CPU_FREE(__kmp_affin_fullMask);
4491  __kmp_affin_fullMask = NULL;
4492  }
4493  if (__kmp_affin_origMask != NULL) {
4494  KMP_CPU_FREE(__kmp_affin_origMask);
4495  __kmp_affin_origMask = NULL;
4496  }
4497  __kmp_affinity_num_masks = 0;
4498  __kmp_affinity_type = affinity_default;
4499  __kmp_affinity_num_places = 0;
4500  if (__kmp_affinity_proclist != NULL) {
4501  __kmp_free(__kmp_affinity_proclist);
4502  __kmp_affinity_proclist = NULL;
4503  }
4504  if (procarr != NULL) {
4505  __kmp_free(procarr);
4506  procarr = NULL;
4507  }
4508 #if KMP_USE_HWLOC
4509  if (__kmp_hwloc_topology != NULL) {
4510  hwloc_topology_destroy(__kmp_hwloc_topology);
4511  __kmp_hwloc_topology = NULL;
4512  }
4513 #endif
4514  if (__kmp_hw_subset) {
4515  kmp_hw_subset_t::deallocate(__kmp_hw_subset);
4516  __kmp_hw_subset = nullptr;
4517  }
4518  if (__kmp_topology) {
4519  kmp_topology_t::deallocate(__kmp_topology);
4520  __kmp_topology = nullptr;
4521  }
4522  KMPAffinity::destroy_api();
4523 }
4524 
4525 void __kmp_affinity_set_init_mask(int gtid, int isa_root) {
4526  if (!KMP_AFFINITY_CAPABLE()) {
4527  return;
4528  }
4529 
4530  kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
4531  if (th->th.th_affin_mask == NULL) {
4532  KMP_CPU_ALLOC(th->th.th_affin_mask);
4533  } else {
4534  KMP_CPU_ZERO(th->th.th_affin_mask);
4535  }
4536 
4537  // Copy the thread mask to the kmp_info_t structure. If
4538  // __kmp_affinity_type == affinity_none, copy the "full" mask, i.e. one that
4539  // has all of the OS proc ids set, or if __kmp_affinity_respect_mask is set,
4540  // then the full mask is the same as the mask of the initialization thread.
4541  kmp_affin_mask_t *mask;
4542  int i;
4543 
4544  if (KMP_AFFINITY_NON_PROC_BIND) {
4545  if ((__kmp_affinity_type == affinity_none) ||
4546  (__kmp_affinity_type == affinity_balanced) ||
4547  KMP_HIDDEN_HELPER_THREAD(gtid)) {
4548 #if KMP_GROUP_AFFINITY
4549  if (__kmp_num_proc_groups > 1) {
4550  return;
4551  }
4552 #endif
4553  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4554  i = 0;
4555  mask = __kmp_affin_fullMask;
4556  } else {
4557  int mask_idx = __kmp_adjust_gtid_for_hidden_helpers(gtid);
4558  KMP_DEBUG_ASSERT(__kmp_affinity_num_masks > 0);
4559  i = (mask_idx + __kmp_affinity_offset) % __kmp_affinity_num_masks;
4560  mask = KMP_CPU_INDEX(__kmp_affinity_masks, i);
4561  }
4562  } else {
4563  if ((!isa_root) || KMP_HIDDEN_HELPER_THREAD(gtid) ||
4564  (__kmp_nested_proc_bind.bind_types[0] == proc_bind_false)) {
4565 #if KMP_GROUP_AFFINITY
4566  if (__kmp_num_proc_groups > 1) {
4567  return;
4568  }
4569 #endif
4570  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4571  i = KMP_PLACE_ALL;
4572  mask = __kmp_affin_fullMask;
4573  } else {
4574  // int i = some hash function or just a counter that doesn't
4575  // always start at 0. Use adjusted gtid for now.
4576  int mask_idx = __kmp_adjust_gtid_for_hidden_helpers(gtid);
4577  KMP_DEBUG_ASSERT(__kmp_affinity_num_masks > 0);
4578  i = (mask_idx + __kmp_affinity_offset) % __kmp_affinity_num_masks;
4579  mask = KMP_CPU_INDEX(__kmp_affinity_masks, i);
4580  }
4581  }
4582 
4583  th->th.th_current_place = i;
4584  if (isa_root || KMP_HIDDEN_HELPER_THREAD(gtid)) {
4585  th->th.th_new_place = i;
4586  th->th.th_first_place = 0;
4587  th->th.th_last_place = __kmp_affinity_num_masks - 1;
4588  } else if (KMP_AFFINITY_NON_PROC_BIND) {
4589  // When using a Non-OMP_PROC_BIND affinity method,
4590  // set all threads' place-partition-var to the entire place list
4591  th->th.th_first_place = 0;
4592  th->th.th_last_place = __kmp_affinity_num_masks - 1;
4593  }
4594 
4595  if (i == KMP_PLACE_ALL) {
4596  KA_TRACE(100, ("__kmp_affinity_set_init_mask: binding T#%d to all places\n",
4597  gtid));
4598  } else {
4599  KA_TRACE(100, ("__kmp_affinity_set_init_mask: binding T#%d to place %d\n",
4600  gtid, i));
4601  }
4602 
4603  KMP_CPU_COPY(th->th.th_affin_mask, mask);
4604 
4605  if (__kmp_affinity_verbose && !KMP_HIDDEN_HELPER_THREAD(gtid)
4606  /* to avoid duplicate printing (will be correctly printed on barrier) */
4607  && (__kmp_affinity_type == affinity_none ||
4608  (i != KMP_PLACE_ALL && __kmp_affinity_type != affinity_balanced))) {
4609  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4610  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4611  th->th.th_affin_mask);
4612  KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY", (kmp_int32)getpid(),
4613  __kmp_gettid(), gtid, buf);
4614  }
4615 
4616 #if KMP_DEBUG
4617  // Hidden helper thread affinity only printed for debug builds
4618  if (__kmp_affinity_verbose && KMP_HIDDEN_HELPER_THREAD(gtid)) {
4619  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4620  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4621  th->th.th_affin_mask);
4622  KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY (hidden helper thread)",
4623  (kmp_int32)getpid(), __kmp_gettid(), gtid, buf);
4624  }
4625 #endif
4626 
4627 #if KMP_OS_WINDOWS
4628  // On Windows* OS, the process affinity mask might have changed. If the user
4629  // didn't request affinity and this call fails, just continue silently.
4630  // See CQ171393.
4631  if (__kmp_affinity_type == affinity_none) {
4632  __kmp_set_system_affinity(th->th.th_affin_mask, FALSE);
4633  } else
4634 #endif
4635  __kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
4636 }
4637 
4638 void __kmp_affinity_set_place(int gtid) {
4639  if (!KMP_AFFINITY_CAPABLE()) {
4640  return;
4641  }
4642 
4643  kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
4644 
4645  KA_TRACE(100, ("__kmp_affinity_set_place: binding T#%d to place %d (current "
4646  "place = %d)\n",
4647  gtid, th->th.th_new_place, th->th.th_current_place));
4648 
4649  // Check that the new place is within this thread's partition.
4650  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
4651  KMP_ASSERT(th->th.th_new_place >= 0);
4652  KMP_ASSERT((unsigned)th->th.th_new_place <= __kmp_affinity_num_masks);
4653  if (th->th.th_first_place <= th->th.th_last_place) {
4654  KMP_ASSERT((th->th.th_new_place >= th->th.th_first_place) &&
4655  (th->th.th_new_place <= th->th.th_last_place));
4656  } else {
4657  KMP_ASSERT((th->th.th_new_place <= th->th.th_first_place) ||
4658  (th->th.th_new_place >= th->th.th_last_place));
4659  }
4660 
4661  // Copy the thread mask to the kmp_info_t structure,
4662  // and set this thread's affinity.
4663  kmp_affin_mask_t *mask =
4664  KMP_CPU_INDEX(__kmp_affinity_masks, th->th.th_new_place);
4665  KMP_CPU_COPY(th->th.th_affin_mask, mask);
4666  th->th.th_current_place = th->th.th_new_place;
4667 
4668  if (__kmp_affinity_verbose) {
4669  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4670  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4671  th->th.th_affin_mask);
4672  KMP_INFORM(BoundToOSProcSet, "OMP_PROC_BIND", (kmp_int32)getpid(),
4673  __kmp_gettid(), gtid, buf);
4674  }
4675  __kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
4676 }
4677 
4678 int __kmp_aux_set_affinity(void **mask) {
4679  int gtid;
4680  kmp_info_t *th;
4681  int retval;
4682 
4683  if (!KMP_AFFINITY_CAPABLE()) {
4684  return -1;
4685  }
4686 
4687  gtid = __kmp_entry_gtid();
4688  KA_TRACE(
4689  1000, (""); {
4690  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4691  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4692  (kmp_affin_mask_t *)(*mask));
4693  __kmp_debug_printf(
4694  "kmp_set_affinity: setting affinity mask for thread %d = %s\n",
4695  gtid, buf);
4696  });
4697 
4698  if (__kmp_env_consistency_check) {
4699  if ((mask == NULL) || (*mask == NULL)) {
4700  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4701  } else {
4702  unsigned proc;
4703  int num_procs = 0;
4704 
4705  KMP_CPU_SET_ITERATE(proc, ((kmp_affin_mask_t *)(*mask))) {
4706  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
4707  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4708  }
4709  if (!KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask))) {
4710  continue;
4711  }
4712  num_procs++;
4713  }
4714  if (num_procs == 0) {
4715  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4716  }
4717 
4718 #if KMP_GROUP_AFFINITY
4719  if (__kmp_get_proc_group((kmp_affin_mask_t *)(*mask)) < 0) {
4720  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4721  }
4722 #endif /* KMP_GROUP_AFFINITY */
4723  }
4724  }
4725 
4726  th = __kmp_threads[gtid];
4727  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
4728  retval = __kmp_set_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
4729  if (retval == 0) {
4730  KMP_CPU_COPY(th->th.th_affin_mask, (kmp_affin_mask_t *)(*mask));
4731  }
4732 
4733  th->th.th_current_place = KMP_PLACE_UNDEFINED;
4734  th->th.th_new_place = KMP_PLACE_UNDEFINED;
4735  th->th.th_first_place = 0;
4736  th->th.th_last_place = __kmp_affinity_num_masks - 1;
4737 
4738  // Turn off 4.0 affinity for the current tread at this parallel level.
4739  th->th.th_current_task->td_icvs.proc_bind = proc_bind_false;
4740 
4741  return retval;
4742 }
4743 
4744 int __kmp_aux_get_affinity(void **mask) {
4745  int gtid;
4746  int retval;
4747 #if KMP_OS_WINDOWS || KMP_DEBUG
4748  kmp_info_t *th;
4749 #endif
4750  if (!KMP_AFFINITY_CAPABLE()) {
4751  return -1;
4752  }
4753 
4754  gtid = __kmp_entry_gtid();
4755 #if KMP_OS_WINDOWS || KMP_DEBUG
4756  th = __kmp_threads[gtid];
4757 #else
4758  (void)gtid; // unused variable
4759 #endif
4760  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
4761 
4762  KA_TRACE(
4763  1000, (""); {
4764  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4765  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4766  th->th.th_affin_mask);
4767  __kmp_printf(
4768  "kmp_get_affinity: stored affinity mask for thread %d = %s\n", gtid,
4769  buf);
4770  });
4771 
4772  if (__kmp_env_consistency_check) {
4773  if ((mask == NULL) || (*mask == NULL)) {
4774  KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity");
4775  }
4776  }
4777 
4778 #if !KMP_OS_WINDOWS
4779 
4780  retval = __kmp_get_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
4781  KA_TRACE(
4782  1000, (""); {
4783  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4784  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4785  (kmp_affin_mask_t *)(*mask));
4786  __kmp_printf(
4787  "kmp_get_affinity: system affinity mask for thread %d = %s\n", gtid,
4788  buf);
4789  });
4790  return retval;
4791 
4792 #else
4793  (void)retval;
4794 
4795  KMP_CPU_COPY((kmp_affin_mask_t *)(*mask), th->th.th_affin_mask);
4796  return 0;
4797 
4798 #endif /* KMP_OS_WINDOWS */
4799 }
4800 
4801 int __kmp_aux_get_affinity_max_proc() {
4802  if (!KMP_AFFINITY_CAPABLE()) {
4803  return 0;
4804  }
4805 #if KMP_GROUP_AFFINITY
4806  if (__kmp_num_proc_groups > 1) {
4807  return (int)(__kmp_num_proc_groups * sizeof(DWORD_PTR) * CHAR_BIT);
4808  }
4809 #endif
4810  return __kmp_xproc;
4811 }
4812 
4813 int __kmp_aux_set_affinity_mask_proc(int proc, void **mask) {
4814  if (!KMP_AFFINITY_CAPABLE()) {
4815  return -1;
4816  }
4817 
4818  KA_TRACE(
4819  1000, (""); {
4820  int gtid = __kmp_entry_gtid();
4821  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4822  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4823  (kmp_affin_mask_t *)(*mask));
4824  __kmp_debug_printf("kmp_set_affinity_mask_proc: setting proc %d in "
4825  "affinity mask for thread %d = %s\n",
4826  proc, gtid, buf);
4827  });
4828 
4829  if (__kmp_env_consistency_check) {
4830  if ((mask == NULL) || (*mask == NULL)) {
4831  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity_mask_proc");
4832  }
4833  }
4834 
4835  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
4836  return -1;
4837  }
4838  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
4839  return -2;
4840  }
4841 
4842  KMP_CPU_SET(proc, (kmp_affin_mask_t *)(*mask));
4843  return 0;
4844 }
4845 
4846 int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask) {
4847  if (!KMP_AFFINITY_CAPABLE()) {
4848  return -1;
4849  }
4850 
4851  KA_TRACE(
4852  1000, (""); {
4853  int gtid = __kmp_entry_gtid();
4854  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4855  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4856  (kmp_affin_mask_t *)(*mask));
4857  __kmp_debug_printf("kmp_unset_affinity_mask_proc: unsetting proc %d in "
4858  "affinity mask for thread %d = %s\n",
4859  proc, gtid, buf);
4860  });
4861 
4862  if (__kmp_env_consistency_check) {
4863  if ((mask == NULL) || (*mask == NULL)) {
4864  KMP_FATAL(AffinityInvalidMask, "kmp_unset_affinity_mask_proc");
4865  }
4866  }
4867 
4868  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
4869  return -1;
4870  }
4871  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
4872  return -2;
4873  }
4874 
4875  KMP_CPU_CLR(proc, (kmp_affin_mask_t *)(*mask));
4876  return 0;
4877 }
4878 
4879 int __kmp_aux_get_affinity_mask_proc(int proc, void **mask) {
4880  if (!KMP_AFFINITY_CAPABLE()) {
4881  return -1;
4882  }
4883 
4884  KA_TRACE(
4885  1000, (""); {
4886  int gtid = __kmp_entry_gtid();
4887  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4888  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4889  (kmp_affin_mask_t *)(*mask));
4890  __kmp_debug_printf("kmp_get_affinity_mask_proc: getting proc %d in "
4891  "affinity mask for thread %d = %s\n",
4892  proc, gtid, buf);
4893  });
4894 
4895  if (__kmp_env_consistency_check) {
4896  if ((mask == NULL) || (*mask == NULL)) {
4897  KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity_mask_proc");
4898  }
4899  }
4900 
4901  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
4902  return -1;
4903  }
4904  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
4905  return 0;
4906  }
4907 
4908  return KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask));
4909 }
4910 
4911 // Dynamic affinity settings - Affinity balanced
4912 void __kmp_balanced_affinity(kmp_info_t *th, int nthreads) {
4913  KMP_DEBUG_ASSERT(th);
4914  bool fine_gran = true;
4915  int tid = th->th.th_info.ds.ds_tid;
4916 
4917  // Do not perform balanced affinity for the hidden helper threads
4918  if (KMP_HIDDEN_HELPER_THREAD(__kmp_gtid_from_thread(th)))
4919  return;
4920 
4921  switch (__kmp_affinity_gran) {
4922  case KMP_HW_THREAD:
4923  break;
4924  case KMP_HW_CORE:
4925  if (__kmp_nThreadsPerCore > 1) {
4926  fine_gran = false;
4927  }
4928  break;
4929  case KMP_HW_SOCKET:
4930  if (nCoresPerPkg > 1) {
4931  fine_gran = false;
4932  }
4933  break;
4934  default:
4935  fine_gran = false;
4936  }
4937 
4938  if (__kmp_topology->is_uniform()) {
4939  int coreID;
4940  int threadID;
4941  // Number of hyper threads per core in HT machine
4942  int __kmp_nth_per_core = __kmp_avail_proc / __kmp_ncores;
4943  // Number of cores
4944  int ncores = __kmp_ncores;
4945  if ((nPackages > 1) && (__kmp_nth_per_core <= 1)) {
4946  __kmp_nth_per_core = __kmp_avail_proc / nPackages;
4947  ncores = nPackages;
4948  }
4949  // How many threads will be bound to each core
4950  int chunk = nthreads / ncores;
4951  // How many cores will have an additional thread bound to it - "big cores"
4952  int big_cores = nthreads % ncores;
4953  // Number of threads on the big cores
4954  int big_nth = (chunk + 1) * big_cores;
4955  if (tid < big_nth) {
4956  coreID = tid / (chunk + 1);
4957  threadID = (tid % (chunk + 1)) % __kmp_nth_per_core;
4958  } else { // tid >= big_nth
4959  coreID = (tid - big_cores) / chunk;
4960  threadID = ((tid - big_cores) % chunk) % __kmp_nth_per_core;
4961  }
4962  KMP_DEBUG_ASSERT2(KMP_AFFINITY_CAPABLE(),
4963  "Illegal set affinity operation when not capable");
4964 
4965  kmp_affin_mask_t *mask = th->th.th_affin_mask;
4966  KMP_CPU_ZERO(mask);
4967 
4968  if (fine_gran) {
4969  int osID =
4970  __kmp_topology->at(coreID * __kmp_nth_per_core + threadID).os_id;
4971  KMP_CPU_SET(osID, mask);
4972  } else {
4973  for (int i = 0; i < __kmp_nth_per_core; i++) {
4974  int osID;
4975  osID = __kmp_topology->at(coreID * __kmp_nth_per_core + i).os_id;
4976  KMP_CPU_SET(osID, mask);
4977  }
4978  }
4979  if (__kmp_affinity_verbose) {
4980  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4981  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
4982  KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY", (kmp_int32)getpid(),
4983  __kmp_gettid(), tid, buf);
4984  }
4985  __kmp_set_system_affinity(mask, TRUE);
4986  } else { // Non-uniform topology
4987 
4988  kmp_affin_mask_t *mask = th->th.th_affin_mask;
4989  KMP_CPU_ZERO(mask);
4990 
4991  int core_level =
4992  __kmp_affinity_find_core_level(__kmp_avail_proc, __kmp_aff_depth - 1);
4993  int ncores = __kmp_affinity_compute_ncores(__kmp_avail_proc,
4994  __kmp_aff_depth - 1, core_level);
4995  int nth_per_core = __kmp_affinity_max_proc_per_core(
4996  __kmp_avail_proc, __kmp_aff_depth - 1, core_level);
4997 
4998  // For performance gain consider the special case nthreads ==
4999  // __kmp_avail_proc
5000  if (nthreads == __kmp_avail_proc) {
5001  if (fine_gran) {
5002  int osID = __kmp_topology->at(tid).os_id;
5003  KMP_CPU_SET(osID, mask);
5004  } else {
5005  int core =
5006  __kmp_affinity_find_core(tid, __kmp_aff_depth - 1, core_level);
5007  for (int i = 0; i < __kmp_avail_proc; i++) {
5008  int osID = __kmp_topology->at(i).os_id;
5009  if (__kmp_affinity_find_core(i, __kmp_aff_depth - 1, core_level) ==
5010  core) {
5011  KMP_CPU_SET(osID, mask);
5012  }
5013  }
5014  }
5015  } else if (nthreads <= ncores) {
5016 
5017  int core = 0;
5018  for (int i = 0; i < ncores; i++) {
5019  // Check if this core from procarr[] is in the mask
5020  int in_mask = 0;
5021  for (int j = 0; j < nth_per_core; j++) {
5022  if (procarr[i * nth_per_core + j] != -1) {
5023  in_mask = 1;
5024  break;
5025  }
5026  }
5027  if (in_mask) {
5028  if (tid == core) {
5029  for (int j = 0; j < nth_per_core; j++) {
5030  int osID = procarr[i * nth_per_core + j];
5031  if (osID != -1) {
5032  KMP_CPU_SET(osID, mask);
5033  // For fine granularity it is enough to set the first available
5034  // osID for this core
5035  if (fine_gran) {
5036  break;
5037  }
5038  }
5039  }
5040  break;
5041  } else {
5042  core++;
5043  }
5044  }
5045  }
5046  } else { // nthreads > ncores
5047  // Array to save the number of processors at each core
5048  int *nproc_at_core = (int *)KMP_ALLOCA(sizeof(int) * ncores);
5049  // Array to save the number of cores with "x" available processors;
5050  int *ncores_with_x_procs =
5051  (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
5052  // Array to save the number of cores with # procs from x to nth_per_core
5053  int *ncores_with_x_to_max_procs =
5054  (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
5055 
5056  for (int i = 0; i <= nth_per_core; i++) {
5057  ncores_with_x_procs[i] = 0;
5058  ncores_with_x_to_max_procs[i] = 0;
5059  }
5060 
5061  for (int i = 0; i < ncores; i++) {
5062  int cnt = 0;
5063  for (int j = 0; j < nth_per_core; j++) {
5064  if (procarr[i * nth_per_core + j] != -1) {
5065  cnt++;
5066  }
5067  }
5068  nproc_at_core[i] = cnt;
5069  ncores_with_x_procs[cnt]++;
5070  }
5071 
5072  for (int i = 0; i <= nth_per_core; i++) {
5073  for (int j = i; j <= nth_per_core; j++) {
5074  ncores_with_x_to_max_procs[i] += ncores_with_x_procs[j];
5075  }
5076  }
5077 
5078  // Max number of processors
5079  int nproc = nth_per_core * ncores;
5080  // An array to keep number of threads per each context
5081  int *newarr = (int *)__kmp_allocate(sizeof(int) * nproc);
5082  for (int i = 0; i < nproc; i++) {
5083  newarr[i] = 0;
5084  }
5085 
5086  int nth = nthreads;
5087  int flag = 0;
5088  while (nth > 0) {
5089  for (int j = 1; j <= nth_per_core; j++) {
5090  int cnt = ncores_with_x_to_max_procs[j];
5091  for (int i = 0; i < ncores; i++) {
5092  // Skip the core with 0 processors
5093  if (nproc_at_core[i] == 0) {
5094  continue;
5095  }
5096  for (int k = 0; k < nth_per_core; k++) {
5097  if (procarr[i * nth_per_core + k] != -1) {
5098  if (newarr[i * nth_per_core + k] == 0) {
5099  newarr[i * nth_per_core + k] = 1;
5100  cnt--;
5101  nth--;
5102  break;
5103  } else {
5104  if (flag != 0) {
5105  newarr[i * nth_per_core + k]++;
5106  cnt--;
5107  nth--;
5108  break;
5109  }
5110  }
5111  }
5112  }
5113  if (cnt == 0 || nth == 0) {
5114  break;
5115  }
5116  }
5117  if (nth == 0) {
5118  break;
5119  }
5120  }
5121  flag = 1;
5122  }
5123  int sum = 0;
5124  for (int i = 0; i < nproc; i++) {
5125  sum += newarr[i];
5126  if (sum > tid) {
5127  if (fine_gran) {
5128  int osID = procarr[i];
5129  KMP_CPU_SET(osID, mask);
5130  } else {
5131  int coreID = i / nth_per_core;
5132  for (int ii = 0; ii < nth_per_core; ii++) {
5133  int osID = procarr[coreID * nth_per_core + ii];
5134  if (osID != -1) {
5135  KMP_CPU_SET(osID, mask);
5136  }
5137  }
5138  }
5139  break;
5140  }
5141  }
5142  __kmp_free(newarr);
5143  }
5144 
5145  if (__kmp_affinity_verbose) {
5146  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5147  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
5148  KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY", (kmp_int32)getpid(),
5149  __kmp_gettid(), tid, buf);
5150  }
5151  __kmp_set_system_affinity(mask, TRUE);
5152  }
5153 }
5154 
5155 #if KMP_OS_LINUX || KMP_OS_FREEBSD
5156 // We don't need this entry for Windows because
5157 // there is GetProcessAffinityMask() api
5158 //
5159 // The intended usage is indicated by these steps:
5160 // 1) The user gets the current affinity mask
5161 // 2) Then sets the affinity by calling this function
5162 // 3) Error check the return value
5163 // 4) Use non-OpenMP parallelization
5164 // 5) Reset the affinity to what was stored in step 1)
5165 #ifdef __cplusplus
5166 extern "C"
5167 #endif
5168  int
5169  kmp_set_thread_affinity_mask_initial()
5170 // the function returns 0 on success,
5171 // -1 if we cannot bind thread
5172 // >0 (errno) if an error happened during binding
5173 {
5174  int gtid = __kmp_get_gtid();
5175  if (gtid < 0) {
5176  // Do not touch non-omp threads
5177  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5178  "non-omp thread, returning\n"));
5179  return -1;
5180  }
5181  if (!KMP_AFFINITY_CAPABLE() || !__kmp_init_middle) {
5182  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5183  "affinity not initialized, returning\n"));
5184  return -1;
5185  }
5186  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5187  "set full mask for thread %d\n",
5188  gtid));
5189  KMP_DEBUG_ASSERT(__kmp_affin_fullMask != NULL);
5190  return __kmp_set_system_affinity(__kmp_affin_fullMask, FALSE);
5191 }
5192 #endif
5193 
5194 #endif // KMP_AFFINITY_SUPPORTED
int try_open(const char *filename, const char *mode)
Definition: kmp.h:4381