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 
30 // The machine topology
31 kmp_topology_t *__kmp_topology = nullptr;
32 // KMP_HW_SUBSET environment variable
33 kmp_hw_subset_t *__kmp_hw_subset = nullptr;
34 
35 // Store the real or imagined machine hierarchy here
36 static hierarchy_info machine_hierarchy;
37 
38 void __kmp_cleanup_hierarchy() { machine_hierarchy.fini(); }
39 
40 void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar) {
41  kmp_uint32 depth;
42  // The test below is true if affinity is available, but set to "none". Need to
43  // init on first use of hierarchical barrier.
44  if (TCR_1(machine_hierarchy.uninitialized))
45  machine_hierarchy.init(nproc);
46 
47  // Adjust the hierarchy in case num threads exceeds original
48  if (nproc > machine_hierarchy.base_num_threads)
49  machine_hierarchy.resize(nproc);
50 
51  depth = machine_hierarchy.depth;
52  KMP_DEBUG_ASSERT(depth > 0);
53 
54  thr_bar->depth = depth;
55  __kmp_type_convert(machine_hierarchy.numPerLevel[0] - 1,
56  &(thr_bar->base_leaf_kids));
57  thr_bar->skip_per_level = machine_hierarchy.skipPerLevel;
58 }
59 
60 static int nCoresPerPkg, nPackages;
61 static int __kmp_nThreadsPerCore;
62 #ifndef KMP_DFLT_NTH_CORES
63 static int __kmp_ncores;
64 #endif
65 
66 const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural) {
67  switch (type) {
68  case KMP_HW_SOCKET:
69  return ((plural) ? KMP_I18N_STR(Sockets) : KMP_I18N_STR(Socket));
70  case KMP_HW_DIE:
71  return ((plural) ? KMP_I18N_STR(Dice) : KMP_I18N_STR(Die));
72  case KMP_HW_MODULE:
73  return ((plural) ? KMP_I18N_STR(Modules) : KMP_I18N_STR(Module));
74  case KMP_HW_TILE:
75  return ((plural) ? KMP_I18N_STR(Tiles) : KMP_I18N_STR(Tile));
76  case KMP_HW_NUMA:
77  return ((plural) ? KMP_I18N_STR(NumaDomains) : KMP_I18N_STR(NumaDomain));
78  case KMP_HW_L3:
79  return ((plural) ? KMP_I18N_STR(L3Caches) : KMP_I18N_STR(L3Cache));
80  case KMP_HW_L2:
81  return ((plural) ? KMP_I18N_STR(L2Caches) : KMP_I18N_STR(L2Cache));
82  case KMP_HW_L1:
83  return ((plural) ? KMP_I18N_STR(L1Caches) : KMP_I18N_STR(L1Cache));
84  case KMP_HW_LLC:
85  return ((plural) ? KMP_I18N_STR(LLCaches) : KMP_I18N_STR(LLCache));
86  case KMP_HW_CORE:
87  return ((plural) ? KMP_I18N_STR(Cores) : KMP_I18N_STR(Core));
88  case KMP_HW_THREAD:
89  return ((plural) ? KMP_I18N_STR(Threads) : KMP_I18N_STR(Thread));
90  case KMP_HW_PROC_GROUP:
91  return ((plural) ? KMP_I18N_STR(ProcGroups) : KMP_I18N_STR(ProcGroup));
92  }
93  return KMP_I18N_STR(Unknown);
94 }
95 
96 const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural) {
97  switch (type) {
98  case KMP_HW_SOCKET:
99  return ((plural) ? "sockets" : "socket");
100  case KMP_HW_DIE:
101  return ((plural) ? "dice" : "die");
102  case KMP_HW_MODULE:
103  return ((plural) ? "modules" : "module");
104  case KMP_HW_TILE:
105  return ((plural) ? "tiles" : "tile");
106  case KMP_HW_NUMA:
107  return ((plural) ? "numa_domains" : "numa_domain");
108  case KMP_HW_L3:
109  return ((plural) ? "l3_caches" : "l3_cache");
110  case KMP_HW_L2:
111  return ((plural) ? "l2_caches" : "l2_cache");
112  case KMP_HW_L1:
113  return ((plural) ? "l1_caches" : "l1_cache");
114  case KMP_HW_LLC:
115  return ((plural) ? "ll_caches" : "ll_cache");
116  case KMP_HW_CORE:
117  return ((plural) ? "cores" : "core");
118  case KMP_HW_THREAD:
119  return ((plural) ? "threads" : "thread");
120  case KMP_HW_PROC_GROUP:
121  return ((plural) ? "proc_groups" : "proc_group");
122  }
123  return ((plural) ? "unknowns" : "unknown");
124 }
125 
127 // kmp_hw_thread_t methods
128 int kmp_hw_thread_t::compare_ids(const void *a, const void *b) {
129  const kmp_hw_thread_t *ahwthread = (const kmp_hw_thread_t *)a;
130  const kmp_hw_thread_t *bhwthread = (const kmp_hw_thread_t *)b;
131  int depth = __kmp_topology->get_depth();
132  for (int level = 0; level < depth; ++level) {
133  if (ahwthread->ids[level] < bhwthread->ids[level])
134  return -1;
135  else if (ahwthread->ids[level] > bhwthread->ids[level])
136  return 1;
137  }
138  if (ahwthread->os_id < bhwthread->os_id)
139  return -1;
140  else if (ahwthread->os_id > bhwthread->os_id)
141  return 1;
142  return 0;
143 }
144 
145 #if KMP_AFFINITY_SUPPORTED
146 int kmp_hw_thread_t::compare_compact(const void *a, const void *b) {
147  int i;
148  const kmp_hw_thread_t *aa = (const kmp_hw_thread_t *)a;
149  const kmp_hw_thread_t *bb = (const kmp_hw_thread_t *)b;
150  int depth = __kmp_topology->get_depth();
151  KMP_DEBUG_ASSERT(__kmp_affinity_compact >= 0);
152  KMP_DEBUG_ASSERT(__kmp_affinity_compact <= depth);
153  for (i = 0; i < __kmp_affinity_compact; i++) {
154  int j = depth - i - 1;
155  if (aa->sub_ids[j] < bb->sub_ids[j])
156  return -1;
157  if (aa->sub_ids[j] > bb->sub_ids[j])
158  return 1;
159  }
160  for (; i < depth; i++) {
161  int j = i - __kmp_affinity_compact;
162  if (aa->sub_ids[j] < bb->sub_ids[j])
163  return -1;
164  if (aa->sub_ids[j] > bb->sub_ids[j])
165  return 1;
166  }
167  return 0;
168 }
169 #endif
170 
171 void kmp_hw_thread_t::print() const {
172  int depth = __kmp_topology->get_depth();
173  printf("%4d ", os_id);
174  for (int i = 0; i < depth; ++i) {
175  printf("%4d ", ids[i]);
176  }
177  printf("\n");
178 }
179 
181 // kmp_topology_t methods
182 
183 // Remove layers that don't add information to the topology.
184 // This is done by having the layer take on the id = UNKNOWN_ID (-1)
185 void kmp_topology_t::_remove_radix1_layers() {
186  int preference[KMP_HW_LAST];
187  int top_index1, top_index2;
188  // Set up preference associative array
189  preference[KMP_HW_PROC_GROUP] = 110;
190  preference[KMP_HW_SOCKET] = 100;
191  preference[KMP_HW_CORE] = 95;
192  preference[KMP_HW_THREAD] = 90;
193  preference[KMP_HW_NUMA] = 85;
194  preference[KMP_HW_DIE] = 80;
195  preference[KMP_HW_TILE] = 75;
196  preference[KMP_HW_MODULE] = 73;
197  preference[KMP_HW_L3] = 70;
198  preference[KMP_HW_L2] = 65;
199  preference[KMP_HW_L1] = 60;
200  preference[KMP_HW_LLC] = 5;
201  top_index1 = 0;
202  top_index2 = 1;
203  while (top_index1 < depth - 1 && top_index2 < depth) {
204  kmp_hw_t type1 = types[top_index1];
205  kmp_hw_t type2 = types[top_index2];
206  KMP_ASSERT_VALID_HW_TYPE(type1);
207  KMP_ASSERT_VALID_HW_TYPE(type2);
208  // Do not allow the three main topology levels (sockets, cores, threads) to
209  // be compacted down
210  if ((type1 == KMP_HW_THREAD || type1 == KMP_HW_CORE ||
211  type1 == KMP_HW_SOCKET) &&
212  (type2 == KMP_HW_THREAD || type2 == KMP_HW_CORE ||
213  type2 == KMP_HW_SOCKET)) {
214  top_index1 = top_index2++;
215  continue;
216  }
217  bool radix1 = true;
218  bool all_same = true;
219  int id1 = hw_threads[0].ids[top_index1];
220  int id2 = hw_threads[0].ids[top_index2];
221  int pref1 = preference[type1];
222  int pref2 = preference[type2];
223  for (int hwidx = 1; hwidx < num_hw_threads; ++hwidx) {
224  if (hw_threads[hwidx].ids[top_index1] == id1 &&
225  hw_threads[hwidx].ids[top_index2] != id2) {
226  radix1 = false;
227  break;
228  }
229  if (hw_threads[hwidx].ids[top_index2] != id2)
230  all_same = false;
231  id1 = hw_threads[hwidx].ids[top_index1];
232  id2 = hw_threads[hwidx].ids[top_index2];
233  }
234  if (radix1) {
235  // Select the layer to remove based on preference
236  kmp_hw_t remove_type, keep_type;
237  int remove_layer, remove_layer_ids;
238  if (pref1 > pref2) {
239  remove_type = type2;
240  remove_layer = remove_layer_ids = top_index2;
241  keep_type = type1;
242  } else {
243  remove_type = type1;
244  remove_layer = remove_layer_ids = top_index1;
245  keep_type = type2;
246  }
247  // If all the indexes for the second (deeper) layer are the same.
248  // e.g., all are zero, then make sure to keep the first layer's ids
249  if (all_same)
250  remove_layer_ids = top_index2;
251  // Remove radix one type by setting the equivalence, removing the id from
252  // the hw threads and removing the layer from types and depth
253  set_equivalent_type(remove_type, keep_type);
254  for (int idx = 0; idx < num_hw_threads; ++idx) {
255  kmp_hw_thread_t &hw_thread = hw_threads[idx];
256  for (int d = remove_layer_ids; d < depth - 1; ++d)
257  hw_thread.ids[d] = hw_thread.ids[d + 1];
258  }
259  for (int idx = remove_layer; idx < depth - 1; ++idx)
260  types[idx] = types[idx + 1];
261  depth--;
262  } else {
263  top_index1 = top_index2++;
264  }
265  }
266  KMP_ASSERT(depth > 0);
267 }
268 
269 void kmp_topology_t::_set_last_level_cache() {
270  if (get_equivalent_type(KMP_HW_L3) != KMP_HW_UNKNOWN)
271  set_equivalent_type(KMP_HW_LLC, KMP_HW_L3);
272  else if (get_equivalent_type(KMP_HW_L2) != KMP_HW_UNKNOWN)
273  set_equivalent_type(KMP_HW_LLC, KMP_HW_L2);
274 #if KMP_MIC_SUPPORTED
275  else if (__kmp_mic_type == mic3) {
276  if (get_equivalent_type(KMP_HW_L2) != KMP_HW_UNKNOWN)
277  set_equivalent_type(KMP_HW_LLC, KMP_HW_L2);
278  else if (get_equivalent_type(KMP_HW_TILE) != KMP_HW_UNKNOWN)
279  set_equivalent_type(KMP_HW_LLC, KMP_HW_TILE);
280  // L2/Tile wasn't detected so just say L1
281  else
282  set_equivalent_type(KMP_HW_LLC, KMP_HW_L1);
283  }
284 #endif
285  else if (get_equivalent_type(KMP_HW_L1) != KMP_HW_UNKNOWN)
286  set_equivalent_type(KMP_HW_LLC, KMP_HW_L1);
287  // Fallback is to set last level cache to socket or core
288  if (get_equivalent_type(KMP_HW_LLC) == KMP_HW_UNKNOWN) {
289  if (get_equivalent_type(KMP_HW_SOCKET) != KMP_HW_UNKNOWN)
290  set_equivalent_type(KMP_HW_LLC, KMP_HW_SOCKET);
291  else if (get_equivalent_type(KMP_HW_CORE) != KMP_HW_UNKNOWN)
292  set_equivalent_type(KMP_HW_LLC, KMP_HW_CORE);
293  }
294  KMP_ASSERT(get_equivalent_type(KMP_HW_LLC) != KMP_HW_UNKNOWN);
295 }
296 
297 // Gather the count of each topology layer and the ratio
298 void kmp_topology_t::_gather_enumeration_information() {
299  int previous_id[KMP_HW_LAST];
300  int max[KMP_HW_LAST];
301 
302  for (int i = 0; i < depth; ++i) {
303  previous_id[i] = kmp_hw_thread_t::UNKNOWN_ID;
304  max[i] = 0;
305  count[i] = 0;
306  ratio[i] = 0;
307  }
308  for (int i = 0; i < num_hw_threads; ++i) {
309  kmp_hw_thread_t &hw_thread = hw_threads[i];
310  for (int layer = 0; layer < depth; ++layer) {
311  int id = hw_thread.ids[layer];
312  if (id != previous_id[layer]) {
313  // Add an additional increment to each count
314  for (int l = layer; l < depth; ++l)
315  count[l]++;
316  // Keep track of topology layer ratio statistics
317  max[layer]++;
318  for (int l = layer + 1; l < depth; ++l) {
319  if (max[l] > ratio[l])
320  ratio[l] = max[l];
321  max[l] = 1;
322  }
323  break;
324  }
325  }
326  for (int layer = 0; layer < depth; ++layer) {
327  previous_id[layer] = hw_thread.ids[layer];
328  }
329  }
330  for (int layer = 0; layer < depth; ++layer) {
331  if (max[layer] > ratio[layer])
332  ratio[layer] = max[layer];
333  }
334 }
335 
336 // Find out if the topology is uniform
337 void kmp_topology_t::_discover_uniformity() {
338  int num = 1;
339  for (int level = 0; level < depth; ++level)
340  num *= ratio[level];
341  flags.uniform = (num == count[depth - 1]);
342 }
343 
344 // Set all the sub_ids for each hardware thread
345 void kmp_topology_t::_set_sub_ids() {
346  int previous_id[KMP_HW_LAST];
347  int sub_id[KMP_HW_LAST];
348 
349  for (int i = 0; i < depth; ++i) {
350  previous_id[i] = -1;
351  sub_id[i] = -1;
352  }
353  for (int i = 0; i < num_hw_threads; ++i) {
354  kmp_hw_thread_t &hw_thread = hw_threads[i];
355  // Setup the sub_id
356  for (int j = 0; j < depth; ++j) {
357  if (hw_thread.ids[j] != previous_id[j]) {
358  sub_id[j]++;
359  for (int k = j + 1; k < depth; ++k) {
360  sub_id[k] = 0;
361  }
362  break;
363  }
364  }
365  // Set previous_id
366  for (int j = 0; j < depth; ++j) {
367  previous_id[j] = hw_thread.ids[j];
368  }
369  // Set the sub_ids field
370  for (int j = 0; j < depth; ++j) {
371  hw_thread.sub_ids[j] = sub_id[j];
372  }
373  }
374 }
375 
376 void kmp_topology_t::_set_globals() {
377  // Set nCoresPerPkg, nPackages, __kmp_nThreadsPerCore, __kmp_ncores
378  int core_level, thread_level, package_level;
379  package_level = get_level(KMP_HW_SOCKET);
380 #if KMP_GROUP_AFFINITY
381  if (package_level == -1)
382  package_level = get_level(KMP_HW_PROC_GROUP);
383 #endif
384  core_level = get_level(KMP_HW_CORE);
385  thread_level = get_level(KMP_HW_THREAD);
386 
387  KMP_ASSERT(core_level != -1);
388  KMP_ASSERT(thread_level != -1);
389 
390  __kmp_nThreadsPerCore = calculate_ratio(thread_level, core_level);
391  if (package_level != -1) {
392  nCoresPerPkg = calculate_ratio(core_level, package_level);
393  nPackages = get_count(package_level);
394  } else {
395  // assume one socket
396  nCoresPerPkg = get_count(core_level);
397  nPackages = 1;
398  }
399 #ifndef KMP_DFLT_NTH_CORES
400  __kmp_ncores = get_count(core_level);
401 #endif
402 }
403 
404 kmp_topology_t *kmp_topology_t::allocate(int nproc, int ndepth,
405  const kmp_hw_t *types) {
406  kmp_topology_t *retval;
407  // Allocate all data in one large allocation
408  size_t size = sizeof(kmp_topology_t) + sizeof(kmp_hw_thread_t) * nproc +
409  sizeof(int) * ndepth * 3;
410  char *bytes = (char *)__kmp_allocate(size);
411  retval = (kmp_topology_t *)bytes;
412  if (nproc > 0) {
413  retval->hw_threads = (kmp_hw_thread_t *)(bytes + sizeof(kmp_topology_t));
414  } else {
415  retval->hw_threads = nullptr;
416  }
417  retval->num_hw_threads = nproc;
418  retval->depth = ndepth;
419  int *arr =
420  (int *)(bytes + sizeof(kmp_topology_t) + sizeof(kmp_hw_thread_t) * nproc);
421  retval->types = (kmp_hw_t *)arr;
422  retval->ratio = arr + ndepth;
423  retval->count = arr + 2 * ndepth;
424  KMP_FOREACH_HW_TYPE(type) { retval->equivalent[type] = KMP_HW_UNKNOWN; }
425  for (int i = 0; i < ndepth; ++i) {
426  retval->types[i] = types[i];
427  retval->equivalent[types[i]] = types[i];
428  }
429  return retval;
430 }
431 
432 void kmp_topology_t::deallocate(kmp_topology_t *topology) {
433  if (topology)
434  __kmp_free(topology);
435 }
436 
437 bool kmp_topology_t::check_ids() const {
438  // Assume ids have been sorted
439  if (num_hw_threads == 0)
440  return true;
441  for (int i = 1; i < num_hw_threads; ++i) {
442  kmp_hw_thread_t &current_thread = hw_threads[i];
443  kmp_hw_thread_t &previous_thread = hw_threads[i - 1];
444  bool unique = false;
445  for (int j = 0; j < depth; ++j) {
446  if (previous_thread.ids[j] != current_thread.ids[j]) {
447  unique = true;
448  break;
449  }
450  }
451  if (unique)
452  continue;
453  return false;
454  }
455  return true;
456 }
457 
458 void kmp_topology_t::dump() const {
459  printf("***********************\n");
460  printf("*** __kmp_topology: ***\n");
461  printf("***********************\n");
462  printf("* depth: %d\n", depth);
463 
464  printf("* types: ");
465  for (int i = 0; i < depth; ++i)
466  printf("%15s ", __kmp_hw_get_keyword(types[i]));
467  printf("\n");
468 
469  printf("* ratio: ");
470  for (int i = 0; i < depth; ++i) {
471  printf("%15d ", ratio[i]);
472  }
473  printf("\n");
474 
475  printf("* count: ");
476  for (int i = 0; i < depth; ++i) {
477  printf("%15d ", count[i]);
478  }
479  printf("\n");
480 
481  printf("* equivalent map:\n");
482  KMP_FOREACH_HW_TYPE(i) {
483  const char *key = __kmp_hw_get_keyword(i);
484  const char *value = __kmp_hw_get_keyword(equivalent[i]);
485  printf("%-15s -> %-15s\n", key, value);
486  }
487 
488  printf("* uniform: %s\n", (is_uniform() ? "Yes" : "No"));
489 
490  printf("* num_hw_threads: %d\n", num_hw_threads);
491  printf("* hw_threads:\n");
492  for (int i = 0; i < num_hw_threads; ++i) {
493  hw_threads[i].print();
494  }
495  printf("***********************\n");
496 }
497 
498 void kmp_topology_t::print(const char *env_var) const {
499  kmp_str_buf_t buf;
500  int print_types_depth;
501  __kmp_str_buf_init(&buf);
502  kmp_hw_t print_types[KMP_HW_LAST + 2];
503 
504  // Num Available Threads
505  KMP_INFORM(AvailableOSProc, env_var, num_hw_threads);
506 
507  // Uniform or not
508  if (is_uniform()) {
509  KMP_INFORM(Uniform, env_var);
510  } else {
511  KMP_INFORM(NonUniform, env_var);
512  }
513 
514  // Equivalent types
515  KMP_FOREACH_HW_TYPE(type) {
516  kmp_hw_t eq_type = equivalent[type];
517  if (eq_type != KMP_HW_UNKNOWN && eq_type != type) {
518  KMP_INFORM(AffEqualTopologyTypes, env_var,
519  __kmp_hw_get_catalog_string(type),
520  __kmp_hw_get_catalog_string(eq_type));
521  }
522  }
523 
524  // Quick topology
525  KMP_ASSERT(depth > 0 && depth <= (int)KMP_HW_LAST);
526  // Create a print types array that always guarantees printing
527  // the core and thread level
528  print_types_depth = 0;
529  for (int level = 0; level < depth; ++level)
530  print_types[print_types_depth++] = types[level];
531  if (equivalent[KMP_HW_CORE] != KMP_HW_CORE) {
532  // Force in the core level for quick topology
533  if (print_types[print_types_depth - 1] == KMP_HW_THREAD) {
534  // Force core before thread e.g., 1 socket X 2 threads/socket
535  // becomes 1 socket X 1 core/socket X 2 threads/socket
536  print_types[print_types_depth - 1] = KMP_HW_CORE;
537  print_types[print_types_depth++] = KMP_HW_THREAD;
538  } else {
539  print_types[print_types_depth++] = KMP_HW_CORE;
540  }
541  }
542  // Always put threads at very end of quick topology
543  if (equivalent[KMP_HW_THREAD] != KMP_HW_THREAD)
544  print_types[print_types_depth++] = KMP_HW_THREAD;
545 
546  __kmp_str_buf_clear(&buf);
547  kmp_hw_t numerator_type;
548  kmp_hw_t denominator_type = KMP_HW_UNKNOWN;
549  int core_level = get_level(KMP_HW_CORE);
550  int ncores = get_count(core_level);
551 
552  for (int plevel = 0, level = 0; plevel < print_types_depth; ++plevel) {
553  int c;
554  bool plural;
555  numerator_type = print_types[plevel];
556  KMP_ASSERT_VALID_HW_TYPE(numerator_type);
557  if (equivalent[numerator_type] != numerator_type)
558  c = 1;
559  else
560  c = get_ratio(level++);
561  plural = (c > 1);
562  if (plevel == 0) {
563  __kmp_str_buf_print(&buf, "%d %s", c,
564  __kmp_hw_get_catalog_string(numerator_type, plural));
565  } else {
566  __kmp_str_buf_print(&buf, " x %d %s/%s", c,
567  __kmp_hw_get_catalog_string(numerator_type, plural),
568  __kmp_hw_get_catalog_string(denominator_type));
569  }
570  denominator_type = numerator_type;
571  }
572  KMP_INFORM(TopologyGeneric, env_var, buf.str, ncores);
573 
574  if (num_hw_threads <= 0) {
575  __kmp_str_buf_free(&buf);
576  return;
577  }
578 
579  // Full OS proc to hardware thread map
580  KMP_INFORM(OSProcToPhysicalThreadMap, env_var);
581  for (int i = 0; i < num_hw_threads; i++) {
582  __kmp_str_buf_clear(&buf);
583  for (int level = 0; level < depth; ++level) {
584  kmp_hw_t type = types[level];
585  __kmp_str_buf_print(&buf, "%s ", __kmp_hw_get_catalog_string(type));
586  __kmp_str_buf_print(&buf, "%d ", hw_threads[i].ids[level]);
587  }
588  KMP_INFORM(OSProcMapToPack, env_var, hw_threads[i].os_id, buf.str);
589  }
590 
591  __kmp_str_buf_free(&buf);
592 }
593 
594 void kmp_topology_t::canonicalize() {
595  _remove_radix1_layers();
596  _gather_enumeration_information();
597  _discover_uniformity();
598  _set_sub_ids();
599  _set_globals();
600  _set_last_level_cache();
601 
602 #if KMP_MIC_SUPPORTED
603  // Manually Add L2 = Tile equivalence
604  if (__kmp_mic_type == mic3) {
605  if (get_level(KMP_HW_L2) != -1)
606  set_equivalent_type(KMP_HW_TILE, KMP_HW_L2);
607  else if (get_level(KMP_HW_TILE) != -1)
608  set_equivalent_type(KMP_HW_L2, KMP_HW_TILE);
609  }
610 #endif
611 
612  // Perform post canonicalization checking
613  KMP_ASSERT(depth > 0);
614  for (int level = 0; level < depth; ++level) {
615  // All counts, ratios, and types must be valid
616  KMP_ASSERT(count[level] > 0 && ratio[level] > 0);
617  KMP_ASSERT_VALID_HW_TYPE(types[level]);
618  // Detected types must point to themselves
619  KMP_ASSERT(equivalent[types[level]] == types[level]);
620  }
621 
622 #if KMP_AFFINITY_SUPPORTED
623  // Set the number of affinity granularity levels
624  if (__kmp_affinity_gran_levels < 0) {
625  kmp_hw_t gran_type = get_equivalent_type(__kmp_affinity_gran);
626  // Check if user's granularity request is valid
627  if (gran_type == KMP_HW_UNKNOWN) {
628  // First try core, then thread, then package
629  kmp_hw_t gran_types[3] = {KMP_HW_CORE, KMP_HW_THREAD, KMP_HW_SOCKET};
630  for (auto g : gran_types) {
631  if (__kmp_topology->get_equivalent_type(g) != KMP_HW_UNKNOWN) {
632  gran_type = g;
633  break;
634  }
635  }
636  KMP_ASSERT(gran_type != KMP_HW_UNKNOWN);
637  // Warn user what granularity setting will be used instead
638  KMP_WARNING(AffGranularityBad, "KMP_AFFINITY",
639  __kmp_hw_get_catalog_string(__kmp_affinity_gran),
640  __kmp_hw_get_catalog_string(gran_type));
641  __kmp_affinity_gran = gran_type;
642  }
643  __kmp_affinity_gran_levels = 0;
644  for (int i = depth - 1; i >= 0 && get_type(i) != gran_type; --i)
645  __kmp_affinity_gran_levels++;
646  }
647 #endif // KMP_AFFINITY_SUPPORTED
648 }
649 
650 // Canonicalize an explicit packages X cores/pkg X threads/core topology
651 void kmp_topology_t::canonicalize(int npackages, int ncores_per_pkg,
652  int nthreads_per_core, int ncores) {
653  int ndepth = 3;
654  depth = ndepth;
655  KMP_FOREACH_HW_TYPE(i) { equivalent[i] = KMP_HW_UNKNOWN; }
656  for (int level = 0; level < depth; ++level) {
657  count[level] = 0;
658  ratio[level] = 0;
659  }
660  count[0] = npackages;
661  count[1] = ncores;
662  count[2] = __kmp_xproc;
663  ratio[0] = npackages;
664  ratio[1] = ncores_per_pkg;
665  ratio[2] = nthreads_per_core;
666  equivalent[KMP_HW_SOCKET] = KMP_HW_SOCKET;
667  equivalent[KMP_HW_CORE] = KMP_HW_CORE;
668  equivalent[KMP_HW_THREAD] = KMP_HW_THREAD;
669  types[0] = KMP_HW_SOCKET;
670  types[1] = KMP_HW_CORE;
671  types[2] = KMP_HW_THREAD;
672  //__kmp_avail_proc = __kmp_xproc;
673  _discover_uniformity();
674 }
675 
676 // Apply the KMP_HW_SUBSET envirable to the topology
677 // Returns true if KMP_HW_SUBSET filtered any processors
678 // otherwise, returns false
679 bool kmp_topology_t::filter_hw_subset() {
680  // If KMP_HW_SUBSET wasn't requested, then do nothing.
681  if (!__kmp_hw_subset)
682  return false;
683 
684  // Check to see if KMP_HW_SUBSET is a valid subset of the detected topology
685  int hw_subset_depth = __kmp_hw_subset->get_depth();
686  kmp_hw_t specified[KMP_HW_LAST];
687  KMP_ASSERT(hw_subset_depth > 0);
688  KMP_FOREACH_HW_TYPE(i) { specified[i] = KMP_HW_UNKNOWN; }
689  for (int i = 0; i < hw_subset_depth; ++i) {
690  int max_count;
691  int num = __kmp_hw_subset->at(i).num;
692  int offset = __kmp_hw_subset->at(i).offset;
693  kmp_hw_t type = __kmp_hw_subset->at(i).type;
694  kmp_hw_t equivalent_type = equivalent[type];
695  int level = get_level(type);
696 
697  // Check to see if current layer is in detected machine topology
698  if (equivalent_type != KMP_HW_UNKNOWN) {
699  __kmp_hw_subset->at(i).type = equivalent_type;
700  } else {
701  KMP_WARNING(AffHWSubsetNotExistGeneric,
702  __kmp_hw_get_catalog_string(type));
703  return false;
704  }
705 
706  // Check to see if current layer has already been specified
707  // either directly or through an equivalent type
708  if (specified[equivalent_type] != KMP_HW_UNKNOWN) {
709  KMP_WARNING(AffHWSubsetEqvLayers, __kmp_hw_get_catalog_string(type),
710  __kmp_hw_get_catalog_string(specified[equivalent_type]));
711  return false;
712  }
713  specified[equivalent_type] = type;
714 
715  // Check to see if layers are in order
716  if (i + 1 < hw_subset_depth) {
717  kmp_hw_t next_type = get_equivalent_type(__kmp_hw_subset->at(i + 1).type);
718  if (next_type == KMP_HW_UNKNOWN) {
719  KMP_WARNING(
720  AffHWSubsetNotExistGeneric,
721  __kmp_hw_get_catalog_string(__kmp_hw_subset->at(i + 1).type));
722  return false;
723  }
724  int next_topology_level = get_level(next_type);
725  if (level > next_topology_level) {
726  KMP_WARNING(AffHWSubsetOutOfOrder, __kmp_hw_get_catalog_string(type),
727  __kmp_hw_get_catalog_string(next_type));
728  return false;
729  }
730  }
731 
732  // Check to see if each layer's num & offset parameters are valid
733  max_count = get_ratio(level);
734  if (max_count < 0 || num + offset > max_count) {
735  bool plural = (num > 1);
736  KMP_WARNING(AffHWSubsetManyGeneric,
737  __kmp_hw_get_catalog_string(type, plural));
738  return false;
739  }
740  }
741 
742  // Apply the filtered hardware subset
743  int new_index = 0;
744  for (int i = 0; i < num_hw_threads; ++i) {
745  kmp_hw_thread_t &hw_thread = hw_threads[i];
746  // Check to see if this hardware thread should be filtered
747  bool should_be_filtered = false;
748  for (int level = 0, hw_subset_index = 0;
749  level < depth && hw_subset_index < hw_subset_depth; ++level) {
750  kmp_hw_t topology_type = types[level];
751  auto hw_subset_item = __kmp_hw_subset->at(hw_subset_index);
752  kmp_hw_t hw_subset_type = hw_subset_item.type;
753  if (topology_type != hw_subset_type)
754  continue;
755  int num = hw_subset_item.num;
756  int offset = hw_subset_item.offset;
757  hw_subset_index++;
758  if (hw_thread.sub_ids[level] < offset ||
759  hw_thread.sub_ids[level] >= offset + num) {
760  should_be_filtered = true;
761  break;
762  }
763  }
764  if (!should_be_filtered) {
765  if (i != new_index)
766  hw_threads[new_index] = hw_thread;
767  new_index++;
768  } else {
769 #if KMP_AFFINITY_SUPPORTED
770  KMP_CPU_CLR(hw_thread.os_id, __kmp_affin_fullMask);
771 #endif
772  __kmp_avail_proc--;
773  }
774  }
775  KMP_DEBUG_ASSERT(new_index <= num_hw_threads);
776  num_hw_threads = new_index;
777 
778  // Post hardware subset canonicalization
779  _gather_enumeration_information();
780  _discover_uniformity();
781  _set_globals();
782  _set_last_level_cache();
783  return true;
784 }
785 
786 bool kmp_topology_t::is_close(int hwt1, int hwt2, int hw_level) const {
787  if (hw_level >= depth)
788  return true;
789  bool retval = true;
790  const kmp_hw_thread_t &t1 = hw_threads[hwt1];
791  const kmp_hw_thread_t &t2 = hw_threads[hwt2];
792  for (int i = 0; i < (depth - hw_level); ++i) {
793  if (t1.ids[i] != t2.ids[i])
794  return false;
795  }
796  return retval;
797 }
798 
800 
801 #if KMP_AFFINITY_SUPPORTED
802 class kmp_affinity_raii_t {
803  kmp_affin_mask_t *mask;
804  bool restored;
805 
806 public:
807  kmp_affinity_raii_t() : restored(false) {
808  KMP_CPU_ALLOC(mask);
809  KMP_ASSERT(mask != NULL);
810  __kmp_get_system_affinity(mask, TRUE);
811  }
812  void restore() {
813  __kmp_set_system_affinity(mask, TRUE);
814  KMP_CPU_FREE(mask);
815  restored = true;
816  }
817  ~kmp_affinity_raii_t() {
818  if (!restored) {
819  __kmp_set_system_affinity(mask, TRUE);
820  KMP_CPU_FREE(mask);
821  }
822  }
823 };
824 
825 bool KMPAffinity::picked_api = false;
826 
827 void *KMPAffinity::Mask::operator new(size_t n) { return __kmp_allocate(n); }
828 void *KMPAffinity::Mask::operator new[](size_t n) { return __kmp_allocate(n); }
829 void KMPAffinity::Mask::operator delete(void *p) { __kmp_free(p); }
830 void KMPAffinity::Mask::operator delete[](void *p) { __kmp_free(p); }
831 void *KMPAffinity::operator new(size_t n) { return __kmp_allocate(n); }
832 void KMPAffinity::operator delete(void *p) { __kmp_free(p); }
833 
834 void KMPAffinity::pick_api() {
835  KMPAffinity *affinity_dispatch;
836  if (picked_api)
837  return;
838 #if KMP_USE_HWLOC
839  // Only use Hwloc if affinity isn't explicitly disabled and
840  // user requests Hwloc topology method
841  if (__kmp_affinity_top_method == affinity_top_method_hwloc &&
842  __kmp_affinity_type != affinity_disabled) {
843  affinity_dispatch = new KMPHwlocAffinity();
844  } else
845 #endif
846  {
847  affinity_dispatch = new KMPNativeAffinity();
848  }
849  __kmp_affinity_dispatch = affinity_dispatch;
850  picked_api = true;
851 }
852 
853 void KMPAffinity::destroy_api() {
854  if (__kmp_affinity_dispatch != NULL) {
855  delete __kmp_affinity_dispatch;
856  __kmp_affinity_dispatch = NULL;
857  picked_api = false;
858  }
859 }
860 
861 #define KMP_ADVANCE_SCAN(scan) \
862  while (*scan != '\0') { \
863  scan++; \
864  }
865 
866 // Print the affinity mask to the character array in a pretty format.
867 // The format is a comma separated list of non-negative integers or integer
868 // ranges: e.g., 1,2,3-5,7,9-15
869 // The format can also be the string "{<empty>}" if no bits are set in mask
870 char *__kmp_affinity_print_mask(char *buf, int buf_len,
871  kmp_affin_mask_t *mask) {
872  int start = 0, finish = 0, previous = 0;
873  bool first_range;
874  KMP_ASSERT(buf);
875  KMP_ASSERT(buf_len >= 40);
876  KMP_ASSERT(mask);
877  char *scan = buf;
878  char *end = buf + buf_len - 1;
879 
880  // Check for empty set.
881  if (mask->begin() == mask->end()) {
882  KMP_SNPRINTF(scan, end - scan + 1, "{<empty>}");
883  KMP_ADVANCE_SCAN(scan);
884  KMP_ASSERT(scan <= end);
885  return buf;
886  }
887 
888  first_range = true;
889  start = mask->begin();
890  while (1) {
891  // Find next range
892  // [start, previous] is inclusive range of contiguous bits in mask
893  for (finish = mask->next(start), previous = start;
894  finish == previous + 1 && finish != mask->end();
895  finish = mask->next(finish)) {
896  previous = finish;
897  }
898 
899  // The first range does not need a comma printed before it, but the rest
900  // of the ranges do need a comma beforehand
901  if (!first_range) {
902  KMP_SNPRINTF(scan, end - scan + 1, "%s", ",");
903  KMP_ADVANCE_SCAN(scan);
904  } else {
905  first_range = false;
906  }
907  // Range with three or more contiguous bits in the affinity mask
908  if (previous - start > 1) {
909  KMP_SNPRINTF(scan, end - scan + 1, "%u-%u", start, previous);
910  } else {
911  // Range with one or two contiguous bits in the affinity mask
912  KMP_SNPRINTF(scan, end - scan + 1, "%u", start);
913  KMP_ADVANCE_SCAN(scan);
914  if (previous - start > 0) {
915  KMP_SNPRINTF(scan, end - scan + 1, ",%u", previous);
916  }
917  }
918  KMP_ADVANCE_SCAN(scan);
919  // Start over with new start point
920  start = finish;
921  if (start == mask->end())
922  break;
923  // Check for overflow
924  if (end - scan < 2)
925  break;
926  }
927 
928  // Check for overflow
929  KMP_ASSERT(scan <= end);
930  return buf;
931 }
932 #undef KMP_ADVANCE_SCAN
933 
934 // Print the affinity mask to the string buffer object in a pretty format
935 // The format is a comma separated list of non-negative integers or integer
936 // ranges: e.g., 1,2,3-5,7,9-15
937 // The format can also be the string "{<empty>}" if no bits are set in mask
938 kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
939  kmp_affin_mask_t *mask) {
940  int start = 0, finish = 0, previous = 0;
941  bool first_range;
942  KMP_ASSERT(buf);
943  KMP_ASSERT(mask);
944 
945  __kmp_str_buf_clear(buf);
946 
947  // Check for empty set.
948  if (mask->begin() == mask->end()) {
949  __kmp_str_buf_print(buf, "%s", "{<empty>}");
950  return buf;
951  }
952 
953  first_range = true;
954  start = mask->begin();
955  while (1) {
956  // Find next range
957  // [start, previous] is inclusive range of contiguous bits in mask
958  for (finish = mask->next(start), previous = start;
959  finish == previous + 1 && finish != mask->end();
960  finish = mask->next(finish)) {
961  previous = finish;
962  }
963 
964  // The first range does not need a comma printed before it, but the rest
965  // of the ranges do need a comma beforehand
966  if (!first_range) {
967  __kmp_str_buf_print(buf, "%s", ",");
968  } else {
969  first_range = false;
970  }
971  // Range with three or more contiguous bits in the affinity mask
972  if (previous - start > 1) {
973  __kmp_str_buf_print(buf, "%u-%u", start, previous);
974  } else {
975  // Range with one or two contiguous bits in the affinity mask
976  __kmp_str_buf_print(buf, "%u", start);
977  if (previous - start > 0) {
978  __kmp_str_buf_print(buf, ",%u", previous);
979  }
980  }
981  // Start over with new start point
982  start = finish;
983  if (start == mask->end())
984  break;
985  }
986  return buf;
987 }
988 
989 void __kmp_affinity_entire_machine_mask(kmp_affin_mask_t *mask) {
990  KMP_CPU_ZERO(mask);
991 
992 #if KMP_GROUP_AFFINITY
993 
994  if (__kmp_num_proc_groups > 1) {
995  int group;
996  KMP_DEBUG_ASSERT(__kmp_GetActiveProcessorCount != NULL);
997  for (group = 0; group < __kmp_num_proc_groups; group++) {
998  int i;
999  int num = __kmp_GetActiveProcessorCount(group);
1000  for (i = 0; i < num; i++) {
1001  KMP_CPU_SET(i + group * (CHAR_BIT * sizeof(DWORD_PTR)), mask);
1002  }
1003  }
1004  } else
1005 
1006 #endif /* KMP_GROUP_AFFINITY */
1007 
1008  {
1009  int proc;
1010  for (proc = 0; proc < __kmp_xproc; proc++) {
1011  KMP_CPU_SET(proc, mask);
1012  }
1013  }
1014 }
1015 
1016 // All of the __kmp_affinity_create_*_map() routines should allocate the
1017 // internal topology object and set the layer ids for it. Each routine
1018 // returns a boolean on whether it was successful at doing so.
1019 kmp_affin_mask_t *__kmp_affin_fullMask = NULL;
1020 
1021 #if KMP_USE_HWLOC
1022 static inline bool __kmp_hwloc_is_cache_type(hwloc_obj_t obj) {
1023 #if HWLOC_API_VERSION >= 0x00020000
1024  return hwloc_obj_type_is_cache(obj->type);
1025 #else
1026  return obj->type == HWLOC_OBJ_CACHE;
1027 #endif
1028 }
1029 
1030 // Returns KMP_HW_* type derived from HWLOC_* type
1031 static inline kmp_hw_t __kmp_hwloc_type_2_topology_type(hwloc_obj_t obj) {
1032 
1033  if (__kmp_hwloc_is_cache_type(obj)) {
1034  if (obj->attr->cache.type == HWLOC_OBJ_CACHE_INSTRUCTION)
1035  return KMP_HW_UNKNOWN;
1036  switch (obj->attr->cache.depth) {
1037  case 1:
1038  return KMP_HW_L1;
1039  case 2:
1040 #if KMP_MIC_SUPPORTED
1041  if (__kmp_mic_type == mic3) {
1042  return KMP_HW_TILE;
1043  }
1044 #endif
1045  return KMP_HW_L2;
1046  case 3:
1047  return KMP_HW_L3;
1048  }
1049  return KMP_HW_UNKNOWN;
1050  }
1051 
1052  switch (obj->type) {
1053  case HWLOC_OBJ_PACKAGE:
1054  return KMP_HW_SOCKET;
1055  case HWLOC_OBJ_NUMANODE:
1056  return KMP_HW_NUMA;
1057  case HWLOC_OBJ_CORE:
1058  return KMP_HW_CORE;
1059  case HWLOC_OBJ_PU:
1060  return KMP_HW_THREAD;
1061  case HWLOC_OBJ_GROUP:
1062  if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_DIE)
1063  return KMP_HW_DIE;
1064  else if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_TILE)
1065  return KMP_HW_TILE;
1066  else if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_MODULE)
1067  return KMP_HW_MODULE;
1068  else if (obj->attr->group.kind == HWLOC_GROUP_KIND_WINDOWS_PROCESSOR_GROUP)
1069  return KMP_HW_PROC_GROUP;
1070  return KMP_HW_UNKNOWN;
1071 #if HWLOC_API_VERSION >= 0x00020100
1072  case HWLOC_OBJ_DIE:
1073  return KMP_HW_DIE;
1074 #endif
1075  }
1076  return KMP_HW_UNKNOWN;
1077 }
1078 
1079 // Returns the number of objects of type 'type' below 'obj' within the topology
1080 // tree structure. e.g., if obj is a HWLOC_OBJ_PACKAGE object, and type is
1081 // HWLOC_OBJ_PU, then this will return the number of PU's under the SOCKET
1082 // object.
1083 static int __kmp_hwloc_get_nobjs_under_obj(hwloc_obj_t obj,
1084  hwloc_obj_type_t type) {
1085  int retval = 0;
1086  hwloc_obj_t first;
1087  for (first = hwloc_get_obj_below_by_type(__kmp_hwloc_topology, obj->type,
1088  obj->logical_index, type, 0);
1089  first != NULL && hwloc_get_ancestor_obj_by_type(__kmp_hwloc_topology,
1090  obj->type, first) == obj;
1091  first = hwloc_get_next_obj_by_type(__kmp_hwloc_topology, first->type,
1092  first)) {
1093  ++retval;
1094  }
1095  return retval;
1096 }
1097 
1098 // This gets the sub_id for a lower object under a higher object in the
1099 // topology tree
1100 static int __kmp_hwloc_get_sub_id(hwloc_topology_t t, hwloc_obj_t higher,
1101  hwloc_obj_t lower) {
1102  hwloc_obj_t obj;
1103  hwloc_obj_type_t ltype = lower->type;
1104  int lindex = lower->logical_index - 1;
1105  int sub_id = 0;
1106  // Get the previous lower object
1107  obj = hwloc_get_obj_by_type(t, ltype, lindex);
1108  while (obj && lindex >= 0 &&
1109  hwloc_bitmap_isincluded(obj->cpuset, higher->cpuset)) {
1110  if (obj->userdata) {
1111  sub_id = (int)(RCAST(kmp_intptr_t, obj->userdata));
1112  break;
1113  }
1114  sub_id++;
1115  lindex--;
1116  obj = hwloc_get_obj_by_type(t, ltype, lindex);
1117  }
1118  // store sub_id + 1 so that 0 is differed from NULL
1119  lower->userdata = RCAST(void *, sub_id + 1);
1120  return sub_id;
1121 }
1122 
1123 static bool __kmp_affinity_create_hwloc_map(kmp_i18n_id_t *const msg_id) {
1124  kmp_hw_t type;
1125  int hw_thread_index, sub_id;
1126  int depth;
1127  hwloc_obj_t pu, obj, root, prev;
1128  kmp_hw_t types[KMP_HW_LAST];
1129  hwloc_obj_type_t hwloc_types[KMP_HW_LAST];
1130 
1131  hwloc_topology_t tp = __kmp_hwloc_topology;
1132  *msg_id = kmp_i18n_null;
1133  if (__kmp_affinity_verbose) {
1134  KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
1135  }
1136 
1137  if (!KMP_AFFINITY_CAPABLE()) {
1138  // Hack to try and infer the machine topology using only the data
1139  // available from hwloc on the current thread, and __kmp_xproc.
1140  KMP_ASSERT(__kmp_affinity_type == affinity_none);
1141  // hwloc only guarantees existance of PU object, so check PACKAGE and CORE
1142  hwloc_obj_t o = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PACKAGE, 0);
1143  if (o != NULL)
1144  nCoresPerPkg = __kmp_hwloc_get_nobjs_under_obj(o, HWLOC_OBJ_CORE);
1145  else
1146  nCoresPerPkg = 1; // no PACKAGE found
1147  o = hwloc_get_obj_by_type(tp, HWLOC_OBJ_CORE, 0);
1148  if (o != NULL)
1149  __kmp_nThreadsPerCore = __kmp_hwloc_get_nobjs_under_obj(o, HWLOC_OBJ_PU);
1150  else
1151  __kmp_nThreadsPerCore = 1; // no CORE found
1152  __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
1153  if (nCoresPerPkg == 0)
1154  nCoresPerPkg = 1; // to prevent possible division by 0
1155  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
1156  return true;
1157  }
1158 
1159  root = hwloc_get_root_obj(tp);
1160 
1161  // Figure out the depth and types in the topology
1162  depth = 0;
1163  pu = hwloc_get_pu_obj_by_os_index(tp, __kmp_affin_fullMask->begin());
1164  KMP_ASSERT(pu);
1165  obj = pu;
1166  types[depth] = KMP_HW_THREAD;
1167  hwloc_types[depth] = obj->type;
1168  depth++;
1169  while (obj != root && obj != NULL) {
1170  obj = obj->parent;
1171 #if HWLOC_API_VERSION >= 0x00020000
1172  if (obj->memory_arity) {
1173  hwloc_obj_t memory;
1174  for (memory = obj->memory_first_child; memory;
1175  memory = hwloc_get_next_child(tp, obj, memory)) {
1176  if (memory->type == HWLOC_OBJ_NUMANODE)
1177  break;
1178  }
1179  if (memory && memory->type == HWLOC_OBJ_NUMANODE) {
1180  types[depth] = KMP_HW_NUMA;
1181  hwloc_types[depth] = memory->type;
1182  depth++;
1183  }
1184  }
1185 #endif
1186  type = __kmp_hwloc_type_2_topology_type(obj);
1187  if (type != KMP_HW_UNKNOWN) {
1188  types[depth] = type;
1189  hwloc_types[depth] = obj->type;
1190  depth++;
1191  }
1192  }
1193  KMP_ASSERT(depth > 0);
1194 
1195  // Get the order for the types correct
1196  for (int i = 0, j = depth - 1; i < j; ++i, --j) {
1197  hwloc_obj_type_t hwloc_temp = hwloc_types[i];
1198  kmp_hw_t temp = types[i];
1199  types[i] = types[j];
1200  types[j] = temp;
1201  hwloc_types[i] = hwloc_types[j];
1202  hwloc_types[j] = hwloc_temp;
1203  }
1204 
1205  // Allocate the data structure to be returned.
1206  __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
1207 
1208  hw_thread_index = 0;
1209  pu = NULL;
1210  while (pu = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, pu)) {
1211  int index = depth - 1;
1212  bool included = KMP_CPU_ISSET(pu->os_index, __kmp_affin_fullMask);
1213  kmp_hw_thread_t &hw_thread = __kmp_topology->at(hw_thread_index);
1214  if (included) {
1215  hw_thread.clear();
1216  hw_thread.ids[index] = pu->logical_index;
1217  hw_thread.os_id = pu->os_index;
1218  index--;
1219  }
1220  obj = pu;
1221  prev = obj;
1222  while (obj != root && obj != NULL) {
1223  obj = obj->parent;
1224 #if HWLOC_API_VERSION >= 0x00020000
1225  // NUMA Nodes are handled differently since they are not within the
1226  // parent/child structure anymore. They are separate children
1227  // of obj (memory_first_child points to first memory child)
1228  if (obj->memory_arity) {
1229  hwloc_obj_t memory;
1230  for (memory = obj->memory_first_child; memory;
1231  memory = hwloc_get_next_child(tp, obj, memory)) {
1232  if (memory->type == HWLOC_OBJ_NUMANODE)
1233  break;
1234  }
1235  if (memory && memory->type == HWLOC_OBJ_NUMANODE) {
1236  sub_id = __kmp_hwloc_get_sub_id(tp, memory, prev);
1237  if (included) {
1238  hw_thread.ids[index] = memory->logical_index;
1239  hw_thread.ids[index + 1] = sub_id;
1240  index--;
1241  }
1242  prev = memory;
1243  }
1244  prev = obj;
1245  }
1246 #endif
1247  type = __kmp_hwloc_type_2_topology_type(obj);
1248  if (type != KMP_HW_UNKNOWN) {
1249  sub_id = __kmp_hwloc_get_sub_id(tp, obj, prev);
1250  if (included) {
1251  hw_thread.ids[index] = obj->logical_index;
1252  hw_thread.ids[index + 1] = sub_id;
1253  index--;
1254  }
1255  prev = obj;
1256  }
1257  }
1258  if (included)
1259  hw_thread_index++;
1260  }
1261  __kmp_topology->sort_ids();
1262  return true;
1263 }
1264 #endif // KMP_USE_HWLOC
1265 
1266 // If we don't know how to retrieve the machine's processor topology, or
1267 // encounter an error in doing so, this routine is called to form a "flat"
1268 // mapping of os thread id's <-> processor id's.
1269 static bool __kmp_affinity_create_flat_map(kmp_i18n_id_t *const msg_id) {
1270  *msg_id = kmp_i18n_null;
1271  int depth = 3;
1272  kmp_hw_t types[] = {KMP_HW_SOCKET, KMP_HW_CORE, KMP_HW_THREAD};
1273 
1274  if (__kmp_affinity_verbose) {
1275  KMP_INFORM(UsingFlatOS, "KMP_AFFINITY");
1276  }
1277 
1278  // Even if __kmp_affinity_type == affinity_none, this routine might still
1279  // called to set __kmp_ncores, as well as
1280  // __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
1281  if (!KMP_AFFINITY_CAPABLE()) {
1282  KMP_ASSERT(__kmp_affinity_type == affinity_none);
1283  __kmp_ncores = nPackages = __kmp_xproc;
1284  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
1285  return true;
1286  }
1287 
1288  // When affinity is off, this routine will still be called to set
1289  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
1290  // Make sure all these vars are set correctly, and return now if affinity is
1291  // not enabled.
1292  __kmp_ncores = nPackages = __kmp_avail_proc;
1293  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
1294 
1295  // Construct the data structure to be returned.
1296  __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
1297  int avail_ct = 0;
1298  int i;
1299  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
1300  // Skip this proc if it is not included in the machine model.
1301  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
1302  continue;
1303  }
1304  kmp_hw_thread_t &hw_thread = __kmp_topology->at(avail_ct);
1305  hw_thread.clear();
1306  hw_thread.os_id = i;
1307  hw_thread.ids[0] = i;
1308  hw_thread.ids[1] = 0;
1309  hw_thread.ids[2] = 0;
1310  avail_ct++;
1311  }
1312  if (__kmp_affinity_verbose) {
1313  KMP_INFORM(OSProcToPackage, "KMP_AFFINITY");
1314  }
1315  return true;
1316 }
1317 
1318 #if KMP_GROUP_AFFINITY
1319 // If multiple Windows* OS processor groups exist, we can create a 2-level
1320 // topology map with the groups at level 0 and the individual procs at level 1.
1321 // This facilitates letting the threads float among all procs in a group,
1322 // if granularity=group (the default when there are multiple groups).
1323 static bool __kmp_affinity_create_proc_group_map(kmp_i18n_id_t *const msg_id) {
1324  *msg_id = kmp_i18n_null;
1325  int depth = 3;
1326  kmp_hw_t types[] = {KMP_HW_PROC_GROUP, KMP_HW_CORE, KMP_HW_THREAD};
1327  const static size_t BITS_PER_GROUP = CHAR_BIT * sizeof(DWORD_PTR);
1328 
1329  if (__kmp_affinity_verbose) {
1330  KMP_INFORM(AffWindowsProcGroupMap, "KMP_AFFINITY");
1331  }
1332 
1333  // If we aren't affinity capable, then use flat topology
1334  if (!KMP_AFFINITY_CAPABLE()) {
1335  KMP_ASSERT(__kmp_affinity_type == affinity_none);
1336  nPackages = __kmp_num_proc_groups;
1337  __kmp_nThreadsPerCore = 1;
1338  __kmp_ncores = __kmp_xproc;
1339  nCoresPerPkg = nPackages / __kmp_ncores;
1340  return true;
1341  }
1342 
1343  // Construct the data structure to be returned.
1344  __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
1345  int avail_ct = 0;
1346  int i;
1347  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
1348  // Skip this proc if it is not included in the machine model.
1349  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
1350  continue;
1351  }
1352  kmp_hw_thread_t &hw_thread = __kmp_topology->at(avail_ct++);
1353  hw_thread.clear();
1354  hw_thread.os_id = i;
1355  hw_thread.ids[0] = i / BITS_PER_GROUP;
1356  hw_thread.ids[1] = hw_thread.ids[2] = i % BITS_PER_GROUP;
1357  }
1358  return true;
1359 }
1360 #endif /* KMP_GROUP_AFFINITY */
1361 
1362 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1363 
1364 template <kmp_uint32 LSB, kmp_uint32 MSB>
1365 static inline unsigned __kmp_extract_bits(kmp_uint32 v) {
1366  const kmp_uint32 SHIFT_LEFT = sizeof(kmp_uint32) * 8 - 1 - MSB;
1367  const kmp_uint32 SHIFT_RIGHT = LSB;
1368  kmp_uint32 retval = v;
1369  retval <<= SHIFT_LEFT;
1370  retval >>= (SHIFT_LEFT + SHIFT_RIGHT);
1371  return retval;
1372 }
1373 
1374 static int __kmp_cpuid_mask_width(int count) {
1375  int r = 0;
1376 
1377  while ((1 << r) < count)
1378  ++r;
1379  return r;
1380 }
1381 
1382 class apicThreadInfo {
1383 public:
1384  unsigned osId; // param to __kmp_affinity_bind_thread
1385  unsigned apicId; // from cpuid after binding
1386  unsigned maxCoresPerPkg; // ""
1387  unsigned maxThreadsPerPkg; // ""
1388  unsigned pkgId; // inferred from above values
1389  unsigned coreId; // ""
1390  unsigned threadId; // ""
1391 };
1392 
1393 static int __kmp_affinity_cmp_apicThreadInfo_phys_id(const void *a,
1394  const void *b) {
1395  const apicThreadInfo *aa = (const apicThreadInfo *)a;
1396  const apicThreadInfo *bb = (const apicThreadInfo *)b;
1397  if (aa->pkgId < bb->pkgId)
1398  return -1;
1399  if (aa->pkgId > bb->pkgId)
1400  return 1;
1401  if (aa->coreId < bb->coreId)
1402  return -1;
1403  if (aa->coreId > bb->coreId)
1404  return 1;
1405  if (aa->threadId < bb->threadId)
1406  return -1;
1407  if (aa->threadId > bb->threadId)
1408  return 1;
1409  return 0;
1410 }
1411 
1412 class kmp_cache_info_t {
1413 public:
1414  struct info_t {
1415  unsigned level, mask;
1416  };
1417  kmp_cache_info_t() : depth(0) { get_leaf4_levels(); }
1418  size_t get_depth() const { return depth; }
1419  info_t &operator[](size_t index) { return table[index]; }
1420  const info_t &operator[](size_t index) const { return table[index]; }
1421 
1422  static kmp_hw_t get_topology_type(unsigned level) {
1423  KMP_DEBUG_ASSERT(level >= 1 && level <= MAX_CACHE_LEVEL);
1424  switch (level) {
1425  case 1:
1426  return KMP_HW_L1;
1427  case 2:
1428  return KMP_HW_L2;
1429  case 3:
1430  return KMP_HW_L3;
1431  }
1432  return KMP_HW_UNKNOWN;
1433  }
1434 
1435 private:
1436  static const int MAX_CACHE_LEVEL = 3;
1437 
1438  size_t depth;
1439  info_t table[MAX_CACHE_LEVEL];
1440 
1441  void get_leaf4_levels() {
1442  unsigned level = 0;
1443  while (depth < MAX_CACHE_LEVEL) {
1444  unsigned cache_type, max_threads_sharing;
1445  unsigned cache_level, cache_mask_width;
1446  kmp_cpuid buf2;
1447  __kmp_x86_cpuid(4, level, &buf2);
1448  cache_type = __kmp_extract_bits<0, 4>(buf2.eax);
1449  if (!cache_type)
1450  break;
1451  // Skip instruction caches
1452  if (cache_type == 2) {
1453  level++;
1454  continue;
1455  }
1456  max_threads_sharing = __kmp_extract_bits<14, 25>(buf2.eax) + 1;
1457  cache_mask_width = __kmp_cpuid_mask_width(max_threads_sharing);
1458  cache_level = __kmp_extract_bits<5, 7>(buf2.eax);
1459  table[depth].level = cache_level;
1460  table[depth].mask = ((-1) << cache_mask_width);
1461  depth++;
1462  level++;
1463  }
1464  }
1465 };
1466 
1467 // On IA-32 architecture and Intel(R) 64 architecture, we attempt to use
1468 // an algorithm which cycles through the available os threads, setting
1469 // the current thread's affinity mask to that thread, and then retrieves
1470 // the Apic Id for each thread context using the cpuid instruction.
1471 static bool __kmp_affinity_create_apicid_map(kmp_i18n_id_t *const msg_id) {
1472  kmp_cpuid buf;
1473  *msg_id = kmp_i18n_null;
1474 
1475  if (__kmp_affinity_verbose) {
1476  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(DecodingLegacyAPIC));
1477  }
1478 
1479  // Check if cpuid leaf 4 is supported.
1480  __kmp_x86_cpuid(0, 0, &buf);
1481  if (buf.eax < 4) {
1482  *msg_id = kmp_i18n_str_NoLeaf4Support;
1483  return false;
1484  }
1485 
1486  // The algorithm used starts by setting the affinity to each available thread
1487  // and retrieving info from the cpuid instruction, so if we are not capable of
1488  // calling __kmp_get_system_affinity() and _kmp_get_system_affinity(), then we
1489  // need to do something else - use the defaults that we calculated from
1490  // issuing cpuid without binding to each proc.
1491  if (!KMP_AFFINITY_CAPABLE()) {
1492  // Hack to try and infer the machine topology using only the data
1493  // available from cpuid on the current thread, and __kmp_xproc.
1494  KMP_ASSERT(__kmp_affinity_type == affinity_none);
1495 
1496  // Get an upper bound on the number of threads per package using cpuid(1).
1497  // On some OS/chps combinations where HT is supported by the chip but is
1498  // disabled, this value will be 2 on a single core chip. Usually, it will be
1499  // 2 if HT is enabled and 1 if HT is disabled.
1500  __kmp_x86_cpuid(1, 0, &buf);
1501  int maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
1502  if (maxThreadsPerPkg == 0) {
1503  maxThreadsPerPkg = 1;
1504  }
1505 
1506  // The num cores per pkg comes from cpuid(4). 1 must be added to the encoded
1507  // value.
1508  //
1509  // The author of cpu_count.cpp treated this only an upper bound on the
1510  // number of cores, but I haven't seen any cases where it was greater than
1511  // the actual number of cores, so we will treat it as exact in this block of
1512  // code.
1513  //
1514  // First, we need to check if cpuid(4) is supported on this chip. To see if
1515  // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n or
1516  // greater.
1517  __kmp_x86_cpuid(0, 0, &buf);
1518  if (buf.eax >= 4) {
1519  __kmp_x86_cpuid(4, 0, &buf);
1520  nCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
1521  } else {
1522  nCoresPerPkg = 1;
1523  }
1524 
1525  // There is no way to reliably tell if HT is enabled without issuing the
1526  // cpuid instruction from every thread, can correlating the cpuid info, so
1527  // if the machine is not affinity capable, we assume that HT is off. We have
1528  // seen quite a few machines where maxThreadsPerPkg is 2, yet the machine
1529  // does not support HT.
1530  //
1531  // - Older OSes are usually found on machines with older chips, which do not
1532  // support HT.
1533  // - The performance penalty for mistakenly identifying a machine as HT when
1534  // it isn't (which results in blocktime being incorrectly set to 0) is
1535  // greater than the penalty when for mistakenly identifying a machine as
1536  // being 1 thread/core when it is really HT enabled (which results in
1537  // blocktime being incorrectly set to a positive value).
1538  __kmp_ncores = __kmp_xproc;
1539  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
1540  __kmp_nThreadsPerCore = 1;
1541  return true;
1542  }
1543 
1544  // From here on, we can assume that it is safe to call
1545  // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
1546  // __kmp_affinity_type = affinity_none.
1547 
1548  // Save the affinity mask for the current thread.
1549  kmp_affinity_raii_t previous_affinity;
1550 
1551  // Run through each of the available contexts, binding the current thread
1552  // to it, and obtaining the pertinent information using the cpuid instr.
1553  //
1554  // The relevant information is:
1555  // - Apic Id: Bits 24:31 of ebx after issuing cpuid(1) - each thread context
1556  // has a uniqie Apic Id, which is of the form pkg# : core# : thread#.
1557  // - Max Threads Per Pkg: Bits 16:23 of ebx after issuing cpuid(1). The value
1558  // of this field determines the width of the core# + thread# fields in the
1559  // Apic Id. It is also an upper bound on the number of threads per
1560  // package, but it has been verified that situations happen were it is not
1561  // exact. In particular, on certain OS/chip combinations where Intel(R)
1562  // Hyper-Threading Technology is supported by the chip but has been
1563  // disabled, the value of this field will be 2 (for a single core chip).
1564  // On other OS/chip combinations supporting Intel(R) Hyper-Threading
1565  // Technology, the value of this field will be 1 when Intel(R)
1566  // Hyper-Threading Technology is disabled and 2 when it is enabled.
1567  // - Max Cores Per Pkg: Bits 26:31 of eax after issuing cpuid(4). The value
1568  // of this field (+1) determines the width of the core# field in the Apic
1569  // Id. The comments in "cpucount.cpp" say that this value is an upper
1570  // bound, but the IA-32 architecture manual says that it is exactly the
1571  // number of cores per package, and I haven't seen any case where it
1572  // wasn't.
1573  //
1574  // From this information, deduce the package Id, core Id, and thread Id,
1575  // and set the corresponding fields in the apicThreadInfo struct.
1576  unsigned i;
1577  apicThreadInfo *threadInfo = (apicThreadInfo *)__kmp_allocate(
1578  __kmp_avail_proc * sizeof(apicThreadInfo));
1579  unsigned nApics = 0;
1580  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
1581  // Skip this proc if it is not included in the machine model.
1582  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
1583  continue;
1584  }
1585  KMP_DEBUG_ASSERT((int)nApics < __kmp_avail_proc);
1586 
1587  __kmp_affinity_dispatch->bind_thread(i);
1588  threadInfo[nApics].osId = i;
1589 
1590  // The apic id and max threads per pkg come from cpuid(1).
1591  __kmp_x86_cpuid(1, 0, &buf);
1592  if (((buf.edx >> 9) & 1) == 0) {
1593  __kmp_free(threadInfo);
1594  *msg_id = kmp_i18n_str_ApicNotPresent;
1595  return false;
1596  }
1597  threadInfo[nApics].apicId = (buf.ebx >> 24) & 0xff;
1598  threadInfo[nApics].maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
1599  if (threadInfo[nApics].maxThreadsPerPkg == 0) {
1600  threadInfo[nApics].maxThreadsPerPkg = 1;
1601  }
1602 
1603  // Max cores per pkg comes from cpuid(4). 1 must be added to the encoded
1604  // value.
1605  //
1606  // First, we need to check if cpuid(4) is supported on this chip. To see if
1607  // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n
1608  // or greater.
1609  __kmp_x86_cpuid(0, 0, &buf);
1610  if (buf.eax >= 4) {
1611  __kmp_x86_cpuid(4, 0, &buf);
1612  threadInfo[nApics].maxCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
1613  } else {
1614  threadInfo[nApics].maxCoresPerPkg = 1;
1615  }
1616 
1617  // Infer the pkgId / coreId / threadId using only the info obtained locally.
1618  int widthCT = __kmp_cpuid_mask_width(threadInfo[nApics].maxThreadsPerPkg);
1619  threadInfo[nApics].pkgId = threadInfo[nApics].apicId >> widthCT;
1620 
1621  int widthC = __kmp_cpuid_mask_width(threadInfo[nApics].maxCoresPerPkg);
1622  int widthT = widthCT - widthC;
1623  if (widthT < 0) {
1624  // I've never seen this one happen, but I suppose it could, if the cpuid
1625  // instruction on a chip was really screwed up. Make sure to restore the
1626  // affinity mask before the tail call.
1627  __kmp_free(threadInfo);
1628  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1629  return false;
1630  }
1631 
1632  int maskC = (1 << widthC) - 1;
1633  threadInfo[nApics].coreId = (threadInfo[nApics].apicId >> widthT) & maskC;
1634 
1635  int maskT = (1 << widthT) - 1;
1636  threadInfo[nApics].threadId = threadInfo[nApics].apicId & maskT;
1637 
1638  nApics++;
1639  }
1640 
1641  // We've collected all the info we need.
1642  // Restore the old affinity mask for this thread.
1643  previous_affinity.restore();
1644 
1645  // Sort the threadInfo table by physical Id.
1646  qsort(threadInfo, nApics, sizeof(*threadInfo),
1647  __kmp_affinity_cmp_apicThreadInfo_phys_id);
1648 
1649  // The table is now sorted by pkgId / coreId / threadId, but we really don't
1650  // know the radix of any of the fields. pkgId's may be sparsely assigned among
1651  // the chips on a system. Although coreId's are usually assigned
1652  // [0 .. coresPerPkg-1] and threadId's are usually assigned
1653  // [0..threadsPerCore-1], we don't want to make any such assumptions.
1654  //
1655  // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
1656  // total # packages) are at this point - we want to determine that now. We
1657  // only have an upper bound on the first two figures.
1658  //
1659  // We also perform a consistency check at this point: the values returned by
1660  // the cpuid instruction for any thread bound to a given package had better
1661  // return the same info for maxThreadsPerPkg and maxCoresPerPkg.
1662  nPackages = 1;
1663  nCoresPerPkg = 1;
1664  __kmp_nThreadsPerCore = 1;
1665  unsigned nCores = 1;
1666 
1667  unsigned pkgCt = 1; // to determine radii
1668  unsigned lastPkgId = threadInfo[0].pkgId;
1669  unsigned coreCt = 1;
1670  unsigned lastCoreId = threadInfo[0].coreId;
1671  unsigned threadCt = 1;
1672  unsigned lastThreadId = threadInfo[0].threadId;
1673 
1674  // intra-pkg consist checks
1675  unsigned prevMaxCoresPerPkg = threadInfo[0].maxCoresPerPkg;
1676  unsigned prevMaxThreadsPerPkg = threadInfo[0].maxThreadsPerPkg;
1677 
1678  for (i = 1; i < nApics; i++) {
1679  if (threadInfo[i].pkgId != lastPkgId) {
1680  nCores++;
1681  pkgCt++;
1682  lastPkgId = threadInfo[i].pkgId;
1683  if ((int)coreCt > nCoresPerPkg)
1684  nCoresPerPkg = coreCt;
1685  coreCt = 1;
1686  lastCoreId = threadInfo[i].coreId;
1687  if ((int)threadCt > __kmp_nThreadsPerCore)
1688  __kmp_nThreadsPerCore = threadCt;
1689  threadCt = 1;
1690  lastThreadId = threadInfo[i].threadId;
1691 
1692  // This is a different package, so go on to the next iteration without
1693  // doing any consistency checks. Reset the consistency check vars, though.
1694  prevMaxCoresPerPkg = threadInfo[i].maxCoresPerPkg;
1695  prevMaxThreadsPerPkg = threadInfo[i].maxThreadsPerPkg;
1696  continue;
1697  }
1698 
1699  if (threadInfo[i].coreId != lastCoreId) {
1700  nCores++;
1701  coreCt++;
1702  lastCoreId = threadInfo[i].coreId;
1703  if ((int)threadCt > __kmp_nThreadsPerCore)
1704  __kmp_nThreadsPerCore = threadCt;
1705  threadCt = 1;
1706  lastThreadId = threadInfo[i].threadId;
1707  } else if (threadInfo[i].threadId != lastThreadId) {
1708  threadCt++;
1709  lastThreadId = threadInfo[i].threadId;
1710  } else {
1711  __kmp_free(threadInfo);
1712  *msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
1713  return false;
1714  }
1715 
1716  // Check to make certain that the maxCoresPerPkg and maxThreadsPerPkg
1717  // fields agree between all the threads bounds to a given package.
1718  if ((prevMaxCoresPerPkg != threadInfo[i].maxCoresPerPkg) ||
1719  (prevMaxThreadsPerPkg != threadInfo[i].maxThreadsPerPkg)) {
1720  __kmp_free(threadInfo);
1721  *msg_id = kmp_i18n_str_InconsistentCpuidInfo;
1722  return false;
1723  }
1724  }
1725  // When affinity is off, this routine will still be called to set
1726  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
1727  // Make sure all these vars are set correctly
1728  nPackages = pkgCt;
1729  if ((int)coreCt > nCoresPerPkg)
1730  nCoresPerPkg = coreCt;
1731  if ((int)threadCt > __kmp_nThreadsPerCore)
1732  __kmp_nThreadsPerCore = threadCt;
1733  __kmp_ncores = nCores;
1734  KMP_DEBUG_ASSERT(nApics == (unsigned)__kmp_avail_proc);
1735 
1736  // Now that we've determined the number of packages, the number of cores per
1737  // package, and the number of threads per core, we can construct the data
1738  // structure that is to be returned.
1739  int idx = 0;
1740  int pkgLevel = 0;
1741  int coreLevel = 1;
1742  int threadLevel = 2;
1743  //(__kmp_nThreadsPerCore <= 1) ? -1 : ((coreLevel >= 0) ? 2 : 1);
1744  int depth = (pkgLevel >= 0) + (coreLevel >= 0) + (threadLevel >= 0);
1745  kmp_hw_t types[3];
1746  if (pkgLevel >= 0)
1747  types[idx++] = KMP_HW_SOCKET;
1748  if (coreLevel >= 0)
1749  types[idx++] = KMP_HW_CORE;
1750  if (threadLevel >= 0)
1751  types[idx++] = KMP_HW_THREAD;
1752 
1753  KMP_ASSERT(depth > 0);
1754  __kmp_topology = kmp_topology_t::allocate(nApics, depth, types);
1755 
1756  for (i = 0; i < nApics; ++i) {
1757  idx = 0;
1758  unsigned os = threadInfo[i].osId;
1759  kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
1760  hw_thread.clear();
1761 
1762  if (pkgLevel >= 0) {
1763  hw_thread.ids[idx++] = threadInfo[i].pkgId;
1764  }
1765  if (coreLevel >= 0) {
1766  hw_thread.ids[idx++] = threadInfo[i].coreId;
1767  }
1768  if (threadLevel >= 0) {
1769  hw_thread.ids[idx++] = threadInfo[i].threadId;
1770  }
1771  hw_thread.os_id = os;
1772  }
1773 
1774  __kmp_free(threadInfo);
1775  __kmp_topology->sort_ids();
1776  if (!__kmp_topology->check_ids()) {
1777  kmp_topology_t::deallocate(__kmp_topology);
1778  __kmp_topology = nullptr;
1779  *msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
1780  return false;
1781  }
1782  return true;
1783 }
1784 
1785 // Intel(R) microarchitecture code name Nehalem, Dunnington and later
1786 // architectures support a newer interface for specifying the x2APIC Ids,
1787 // based on CPUID.B or CPUID.1F
1788 /*
1789  * CPUID.B or 1F, Input ECX (sub leaf # aka level number)
1790  Bits Bits Bits Bits
1791  31-16 15-8 7-4 4-0
1792 ---+-----------+--------------+-------------+-----------------+
1793 EAX| reserved | reserved | reserved | Bits to Shift |
1794 ---+-----------|--------------+-------------+-----------------|
1795 EBX| reserved | Num logical processors at level (16 bits) |
1796 ---+-----------|--------------+-------------------------------|
1797 ECX| reserved | Level Type | Level Number (8 bits) |
1798 ---+-----------+--------------+-------------------------------|
1799 EDX| X2APIC ID (32 bits) |
1800 ---+----------------------------------------------------------+
1801 */
1802 
1803 enum {
1804  INTEL_LEVEL_TYPE_INVALID = 0, // Package level
1805  INTEL_LEVEL_TYPE_SMT = 1,
1806  INTEL_LEVEL_TYPE_CORE = 2,
1807  INTEL_LEVEL_TYPE_TILE = 3,
1808  INTEL_LEVEL_TYPE_MODULE = 4,
1809  INTEL_LEVEL_TYPE_DIE = 5,
1810  INTEL_LEVEL_TYPE_LAST = 6,
1811 };
1812 
1813 struct cpuid_level_info_t {
1814  unsigned level_type, mask, mask_width, nitems, cache_mask;
1815 };
1816 
1817 static kmp_hw_t __kmp_intel_type_2_topology_type(int intel_type) {
1818  switch (intel_type) {
1819  case INTEL_LEVEL_TYPE_INVALID:
1820  return KMP_HW_SOCKET;
1821  case INTEL_LEVEL_TYPE_SMT:
1822  return KMP_HW_THREAD;
1823  case INTEL_LEVEL_TYPE_CORE:
1824  return KMP_HW_CORE;
1825  case INTEL_LEVEL_TYPE_TILE:
1826  return KMP_HW_TILE;
1827  case INTEL_LEVEL_TYPE_MODULE:
1828  return KMP_HW_MODULE;
1829  case INTEL_LEVEL_TYPE_DIE:
1830  return KMP_HW_DIE;
1831  }
1832  return KMP_HW_UNKNOWN;
1833 }
1834 
1835 // This function takes the topology leaf, a levels array to store the levels
1836 // detected and a bitmap of the known levels.
1837 // Returns the number of levels in the topology
1838 static unsigned
1839 __kmp_x2apicid_get_levels(int leaf,
1840  cpuid_level_info_t levels[INTEL_LEVEL_TYPE_LAST],
1841  kmp_uint64 known_levels) {
1842  unsigned level, levels_index;
1843  unsigned level_type, mask_width, nitems;
1844  kmp_cpuid buf;
1845 
1846  // New algorithm has known topology layers act as highest unknown topology
1847  // layers when unknown topology layers exist.
1848  // e.g., Suppose layers were SMT <X> CORE <Y> <Z> PACKAGE, where <X> <Y> <Z>
1849  // are unknown topology layers, Then SMT will take the characteristics of
1850  // (SMT x <X>) and CORE will take the characteristics of (CORE x <Y> x <Z>).
1851  // This eliminates unknown portions of the topology while still keeping the
1852  // correct structure.
1853  level = levels_index = 0;
1854  do {
1855  __kmp_x86_cpuid(leaf, level, &buf);
1856  level_type = __kmp_extract_bits<8, 15>(buf.ecx);
1857  mask_width = __kmp_extract_bits<0, 4>(buf.eax);
1858  nitems = __kmp_extract_bits<0, 15>(buf.ebx);
1859  if (level_type != INTEL_LEVEL_TYPE_INVALID && nitems == 0)
1860  return 0;
1861 
1862  if (known_levels & (1ull << level_type)) {
1863  // Add a new level to the topology
1864  KMP_ASSERT(levels_index < INTEL_LEVEL_TYPE_LAST);
1865  levels[levels_index].level_type = level_type;
1866  levels[levels_index].mask_width = mask_width;
1867  levels[levels_index].nitems = nitems;
1868  levels_index++;
1869  } else {
1870  // If it is an unknown level, then logically move the previous layer up
1871  if (levels_index > 0) {
1872  levels[levels_index - 1].mask_width = mask_width;
1873  levels[levels_index - 1].nitems = nitems;
1874  }
1875  }
1876  level++;
1877  } while (level_type != INTEL_LEVEL_TYPE_INVALID);
1878 
1879  // Set the masks to & with apicid
1880  for (unsigned i = 0; i < levels_index; ++i) {
1881  if (levels[i].level_type != INTEL_LEVEL_TYPE_INVALID) {
1882  levels[i].mask = ~((-1) << levels[i].mask_width);
1883  levels[i].cache_mask = (-1) << levels[i].mask_width;
1884  for (unsigned j = 0; j < i; ++j)
1885  levels[i].mask ^= levels[j].mask;
1886  } else {
1887  KMP_DEBUG_ASSERT(levels_index > 0);
1888  levels[i].mask = (-1) << levels[i - 1].mask_width;
1889  levels[i].cache_mask = 0;
1890  }
1891  }
1892  return levels_index;
1893 }
1894 
1895 static bool __kmp_affinity_create_x2apicid_map(kmp_i18n_id_t *const msg_id) {
1896 
1897  cpuid_level_info_t levels[INTEL_LEVEL_TYPE_LAST];
1898  kmp_hw_t types[INTEL_LEVEL_TYPE_LAST];
1899  unsigned levels_index;
1900  kmp_cpuid buf;
1901  kmp_uint64 known_levels;
1902  int topology_leaf, highest_leaf, apic_id;
1903  int num_leaves;
1904  static int leaves[] = {0, 0};
1905 
1906  kmp_i18n_id_t leaf_message_id;
1907 
1908  KMP_BUILD_ASSERT(sizeof(known_levels) * CHAR_BIT > KMP_HW_LAST);
1909 
1910  *msg_id = kmp_i18n_null;
1911  if (__kmp_affinity_verbose) {
1912  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(Decodingx2APIC));
1913  }
1914 
1915  // Figure out the known topology levels
1916  known_levels = 0ull;
1917  for (int i = 0; i < INTEL_LEVEL_TYPE_LAST; ++i) {
1918  if (__kmp_intel_type_2_topology_type(i) != KMP_HW_UNKNOWN) {
1919  known_levels |= (1ull << i);
1920  }
1921  }
1922 
1923  // Get the highest cpuid leaf supported
1924  __kmp_x86_cpuid(0, 0, &buf);
1925  highest_leaf = buf.eax;
1926 
1927  // If a specific topology method was requested, only allow that specific leaf
1928  // otherwise, try both leaves 31 and 11 in that order
1929  num_leaves = 0;
1930  if (__kmp_affinity_top_method == affinity_top_method_x2apicid) {
1931  num_leaves = 1;
1932  leaves[0] = 11;
1933  leaf_message_id = kmp_i18n_str_NoLeaf11Support;
1934  } else if (__kmp_affinity_top_method == affinity_top_method_x2apicid_1f) {
1935  num_leaves = 1;
1936  leaves[0] = 31;
1937  leaf_message_id = kmp_i18n_str_NoLeaf31Support;
1938  } else {
1939  num_leaves = 2;
1940  leaves[0] = 31;
1941  leaves[1] = 11;
1942  leaf_message_id = kmp_i18n_str_NoLeaf11Support;
1943  }
1944 
1945  // Check to see if cpuid leaf 31 or 11 is supported.
1946  __kmp_nThreadsPerCore = nCoresPerPkg = nPackages = 1;
1947  topology_leaf = -1;
1948  for (int i = 0; i < num_leaves; ++i) {
1949  int leaf = leaves[i];
1950  if (highest_leaf < leaf)
1951  continue;
1952  __kmp_x86_cpuid(leaf, 0, &buf);
1953  if (buf.ebx == 0)
1954  continue;
1955  topology_leaf = leaf;
1956  levels_index = __kmp_x2apicid_get_levels(leaf, levels, known_levels);
1957  if (levels_index == 0)
1958  continue;
1959  break;
1960  }
1961  if (topology_leaf == -1 || levels_index == 0) {
1962  *msg_id = leaf_message_id;
1963  return false;
1964  }
1965  KMP_ASSERT(levels_index <= INTEL_LEVEL_TYPE_LAST);
1966 
1967  // The algorithm used starts by setting the affinity to each available thread
1968  // and retrieving info from the cpuid instruction, so if we are not capable of
1969  // calling __kmp_get_system_affinity() and __kmp_get_system_affinity(), then
1970  // we need to do something else - use the defaults that we calculated from
1971  // issuing cpuid without binding to each proc.
1972  if (!KMP_AFFINITY_CAPABLE()) {
1973  // Hack to try and infer the machine topology using only the data
1974  // available from cpuid on the current thread, and __kmp_xproc.
1975  KMP_ASSERT(__kmp_affinity_type == affinity_none);
1976  for (unsigned i = 0; i < levels_index; ++i) {
1977  if (levels[i].level_type == INTEL_LEVEL_TYPE_SMT) {
1978  __kmp_nThreadsPerCore = levels[i].nitems;
1979  } else if (levels[i].level_type == INTEL_LEVEL_TYPE_CORE) {
1980  nCoresPerPkg = levels[i].nitems;
1981  }
1982  }
1983  __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
1984  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
1985  return true;
1986  }
1987 
1988  // Allocate the data structure to be returned.
1989  int depth = levels_index;
1990  for (int i = depth - 1, j = 0; i >= 0; --i, ++j)
1991  types[j] = __kmp_intel_type_2_topology_type(levels[i].level_type);
1992  __kmp_topology =
1993  kmp_topology_t::allocate(__kmp_avail_proc, levels_index, types);
1994 
1995  // Insert equivalent cache types if they exist
1996  kmp_cache_info_t cache_info;
1997  for (size_t i = 0; i < cache_info.get_depth(); ++i) {
1998  const kmp_cache_info_t::info_t &info = cache_info[i];
1999  unsigned cache_mask = info.mask;
2000  unsigned cache_level = info.level;
2001  for (unsigned j = 0; j < levels_index; ++j) {
2002  unsigned hw_cache_mask = levels[j].cache_mask;
2003  kmp_hw_t cache_type = kmp_cache_info_t::get_topology_type(cache_level);
2004  if (hw_cache_mask == cache_mask && j < levels_index - 1) {
2005  kmp_hw_t type =
2006  __kmp_intel_type_2_topology_type(levels[j + 1].level_type);
2007  __kmp_topology->set_equivalent_type(cache_type, type);
2008  }
2009  }
2010  }
2011 
2012  // From here on, we can assume that it is safe to call
2013  // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
2014  // __kmp_affinity_type = affinity_none.
2015 
2016  // Save the affinity mask for the current thread.
2017  kmp_affinity_raii_t previous_affinity;
2018 
2019  // Run through each of the available contexts, binding the current thread
2020  // to it, and obtaining the pertinent information using the cpuid instr.
2021  unsigned int proc;
2022  int hw_thread_index = 0;
2023  KMP_CPU_SET_ITERATE(proc, __kmp_affin_fullMask) {
2024  cpuid_level_info_t my_levels[INTEL_LEVEL_TYPE_LAST];
2025  unsigned my_levels_index;
2026 
2027  // Skip this proc if it is not included in the machine model.
2028  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
2029  continue;
2030  }
2031  KMP_DEBUG_ASSERT(hw_thread_index < __kmp_avail_proc);
2032 
2033  __kmp_affinity_dispatch->bind_thread(proc);
2034 
2035  // New algorithm
2036  __kmp_x86_cpuid(topology_leaf, 0, &buf);
2037  apic_id = buf.edx;
2038  kmp_hw_thread_t &hw_thread = __kmp_topology->at(hw_thread_index);
2039  my_levels_index =
2040  __kmp_x2apicid_get_levels(topology_leaf, my_levels, known_levels);
2041  if (my_levels_index == 0 || my_levels_index != levels_index) {
2042  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
2043  return false;
2044  }
2045  hw_thread.clear();
2046  hw_thread.os_id = proc;
2047  // Put in topology information
2048  for (unsigned j = 0, idx = depth - 1; j < my_levels_index; ++j, --idx) {
2049  hw_thread.ids[idx] = apic_id & my_levels[j].mask;
2050  if (j > 0) {
2051  hw_thread.ids[idx] >>= my_levels[j - 1].mask_width;
2052  }
2053  }
2054  hw_thread_index++;
2055  }
2056  KMP_ASSERT(hw_thread_index > 0);
2057  __kmp_topology->sort_ids();
2058  if (!__kmp_topology->check_ids()) {
2059  kmp_topology_t::deallocate(__kmp_topology);
2060  __kmp_topology = nullptr;
2061  *msg_id = kmp_i18n_str_x2ApicIDsNotUnique;
2062  return false;
2063  }
2064  return true;
2065 }
2066 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2067 
2068 #define osIdIndex 0
2069 #define threadIdIndex 1
2070 #define coreIdIndex 2
2071 #define pkgIdIndex 3
2072 #define nodeIdIndex 4
2073 
2074 typedef unsigned *ProcCpuInfo;
2075 static unsigned maxIndex = pkgIdIndex;
2076 
2077 static int __kmp_affinity_cmp_ProcCpuInfo_phys_id(const void *a,
2078  const void *b) {
2079  unsigned i;
2080  const unsigned *aa = *(unsigned *const *)a;
2081  const unsigned *bb = *(unsigned *const *)b;
2082  for (i = maxIndex;; i--) {
2083  if (aa[i] < bb[i])
2084  return -1;
2085  if (aa[i] > bb[i])
2086  return 1;
2087  if (i == osIdIndex)
2088  break;
2089  }
2090  return 0;
2091 }
2092 
2093 #if KMP_USE_HIER_SCHED
2094 // Set the array sizes for the hierarchy layers
2095 static void __kmp_dispatch_set_hierarchy_values() {
2096  // Set the maximum number of L1's to number of cores
2097  // Set the maximum number of L2's to to either number of cores / 2 for
2098  // Intel(R) Xeon Phi(TM) coprocessor formally codenamed Knights Landing
2099  // Or the number of cores for Intel(R) Xeon(R) processors
2100  // Set the maximum number of NUMA nodes and L3's to number of packages
2101  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1] =
2102  nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
2103  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L1 + 1] = __kmp_ncores;
2104 #if KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_WINDOWS) && \
2105  KMP_MIC_SUPPORTED
2106  if (__kmp_mic_type >= mic3)
2107  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores / 2;
2108  else
2109 #endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
2110  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores;
2111  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L3 + 1] = nPackages;
2112  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_NUMA + 1] = nPackages;
2113  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_LOOP + 1] = 1;
2114  // Set the number of threads per unit
2115  // Number of hardware threads per L1/L2/L3/NUMA/LOOP
2116  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_THREAD + 1] = 1;
2117  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L1 + 1] =
2118  __kmp_nThreadsPerCore;
2119 #if KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_WINDOWS) && \
2120  KMP_MIC_SUPPORTED
2121  if (__kmp_mic_type >= mic3)
2122  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
2123  2 * __kmp_nThreadsPerCore;
2124  else
2125 #endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
2126  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
2127  __kmp_nThreadsPerCore;
2128  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L3 + 1] =
2129  nCoresPerPkg * __kmp_nThreadsPerCore;
2130  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_NUMA + 1] =
2131  nCoresPerPkg * __kmp_nThreadsPerCore;
2132  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_LOOP + 1] =
2133  nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
2134 }
2135 
2136 // Return the index into the hierarchy for this tid and layer type (L1, L2, etc)
2137 // i.e., this thread's L1 or this thread's L2, etc.
2138 int __kmp_dispatch_get_index(int tid, kmp_hier_layer_e type) {
2139  int index = type + 1;
2140  int num_hw_threads = __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1];
2141  KMP_DEBUG_ASSERT(type != kmp_hier_layer_e::LAYER_LAST);
2142  if (type == kmp_hier_layer_e::LAYER_THREAD)
2143  return tid;
2144  else if (type == kmp_hier_layer_e::LAYER_LOOP)
2145  return 0;
2146  KMP_DEBUG_ASSERT(__kmp_hier_max_units[index] != 0);
2147  if (tid >= num_hw_threads)
2148  tid = tid % num_hw_threads;
2149  return (tid / __kmp_hier_threads_per[index]) % __kmp_hier_max_units[index];
2150 }
2151 
2152 // Return the number of t1's per t2
2153 int __kmp_dispatch_get_t1_per_t2(kmp_hier_layer_e t1, kmp_hier_layer_e t2) {
2154  int i1 = t1 + 1;
2155  int i2 = t2 + 1;
2156  KMP_DEBUG_ASSERT(i1 <= i2);
2157  KMP_DEBUG_ASSERT(t1 != kmp_hier_layer_e::LAYER_LAST);
2158  KMP_DEBUG_ASSERT(t2 != kmp_hier_layer_e::LAYER_LAST);
2159  KMP_DEBUG_ASSERT(__kmp_hier_threads_per[i1] != 0);
2160  // (nthreads/t2) / (nthreads/t1) = t1 / t2
2161  return __kmp_hier_threads_per[i2] / __kmp_hier_threads_per[i1];
2162 }
2163 #endif // KMP_USE_HIER_SCHED
2164 
2165 static inline const char *__kmp_cpuinfo_get_filename() {
2166  const char *filename;
2167  if (__kmp_cpuinfo_file != nullptr)
2168  filename = __kmp_cpuinfo_file;
2169  else
2170  filename = "/proc/cpuinfo";
2171  return filename;
2172 }
2173 
2174 static inline const char *__kmp_cpuinfo_get_envvar() {
2175  const char *envvar = nullptr;
2176  if (__kmp_cpuinfo_file != nullptr)
2177  envvar = "KMP_CPUINFO_FILE";
2178  return envvar;
2179 }
2180 
2181 // Parse /proc/cpuinfo (or an alternate file in the same format) to obtain the
2182 // affinity map.
2183 static bool __kmp_affinity_create_cpuinfo_map(int *line,
2184  kmp_i18n_id_t *const msg_id) {
2185  const char *filename = __kmp_cpuinfo_get_filename();
2186  const char *envvar = __kmp_cpuinfo_get_envvar();
2187  *msg_id = kmp_i18n_null;
2188 
2189  if (__kmp_affinity_verbose) {
2190  KMP_INFORM(AffParseFilename, "KMP_AFFINITY", filename);
2191  }
2192 
2193  kmp_safe_raii_file_t f(filename, "r", envvar);
2194 
2195  // Scan of the file, and count the number of "processor" (osId) fields,
2196  // and find the highest value of <n> for a node_<n> field.
2197  char buf[256];
2198  unsigned num_records = 0;
2199  while (!feof(f)) {
2200  buf[sizeof(buf) - 1] = 1;
2201  if (!fgets(buf, sizeof(buf), f)) {
2202  // Read errors presumably because of EOF
2203  break;
2204  }
2205 
2206  char s1[] = "processor";
2207  if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
2208  num_records++;
2209  continue;
2210  }
2211 
2212  // FIXME - this will match "node_<n> <garbage>"
2213  unsigned level;
2214  if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
2215  // validate the input fisrt:
2216  if (level > (unsigned)__kmp_xproc) { // level is too big
2217  level = __kmp_xproc;
2218  }
2219  if (nodeIdIndex + level >= maxIndex) {
2220  maxIndex = nodeIdIndex + level;
2221  }
2222  continue;
2223  }
2224  }
2225 
2226  // Check for empty file / no valid processor records, or too many. The number
2227  // of records can't exceed the number of valid bits in the affinity mask.
2228  if (num_records == 0) {
2229  *msg_id = kmp_i18n_str_NoProcRecords;
2230  return false;
2231  }
2232  if (num_records > (unsigned)__kmp_xproc) {
2233  *msg_id = kmp_i18n_str_TooManyProcRecords;
2234  return false;
2235  }
2236 
2237  // Set the file pointer back to the beginning, so that we can scan the file
2238  // again, this time performing a full parse of the data. Allocate a vector of
2239  // ProcCpuInfo object, where we will place the data. Adding an extra element
2240  // at the end allows us to remove a lot of extra checks for termination
2241  // conditions.
2242  if (fseek(f, 0, SEEK_SET) != 0) {
2243  *msg_id = kmp_i18n_str_CantRewindCpuinfo;
2244  return false;
2245  }
2246 
2247  // Allocate the array of records to store the proc info in. The dummy
2248  // element at the end makes the logic in filling them out easier to code.
2249  unsigned **threadInfo =
2250  (unsigned **)__kmp_allocate((num_records + 1) * sizeof(unsigned *));
2251  unsigned i;
2252  for (i = 0; i <= num_records; i++) {
2253  threadInfo[i] =
2254  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2255  }
2256 
2257 #define CLEANUP_THREAD_INFO \
2258  for (i = 0; i <= num_records; i++) { \
2259  __kmp_free(threadInfo[i]); \
2260  } \
2261  __kmp_free(threadInfo);
2262 
2263  // A value of UINT_MAX means that we didn't find the field
2264  unsigned __index;
2265 
2266 #define INIT_PROC_INFO(p) \
2267  for (__index = 0; __index <= maxIndex; __index++) { \
2268  (p)[__index] = UINT_MAX; \
2269  }
2270 
2271  for (i = 0; i <= num_records; i++) {
2272  INIT_PROC_INFO(threadInfo[i]);
2273  }
2274 
2275  unsigned num_avail = 0;
2276  *line = 0;
2277  while (!feof(f)) {
2278  // Create an inner scoping level, so that all the goto targets at the end of
2279  // the loop appear in an outer scoping level. This avoids warnings about
2280  // jumping past an initialization to a target in the same block.
2281  {
2282  buf[sizeof(buf) - 1] = 1;
2283  bool long_line = false;
2284  if (!fgets(buf, sizeof(buf), f)) {
2285  // Read errors presumably because of EOF
2286  // If there is valid data in threadInfo[num_avail], then fake
2287  // a blank line in ensure that the last address gets parsed.
2288  bool valid = false;
2289  for (i = 0; i <= maxIndex; i++) {
2290  if (threadInfo[num_avail][i] != UINT_MAX) {
2291  valid = true;
2292  }
2293  }
2294  if (!valid) {
2295  break;
2296  }
2297  buf[0] = 0;
2298  } else if (!buf[sizeof(buf) - 1]) {
2299  // The line is longer than the buffer. Set a flag and don't
2300  // emit an error if we were going to ignore the line, anyway.
2301  long_line = true;
2302 
2303 #define CHECK_LINE \
2304  if (long_line) { \
2305  CLEANUP_THREAD_INFO; \
2306  *msg_id = kmp_i18n_str_LongLineCpuinfo; \
2307  return false; \
2308  }
2309  }
2310  (*line)++;
2311 
2312  char s1[] = "processor";
2313  if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
2314  CHECK_LINE;
2315  char *p = strchr(buf + sizeof(s1) - 1, ':');
2316  unsigned val;
2317  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2318  goto no_val;
2319  if (threadInfo[num_avail][osIdIndex] != UINT_MAX)
2320 #if KMP_ARCH_AARCH64
2321  // Handle the old AArch64 /proc/cpuinfo layout differently,
2322  // it contains all of the 'processor' entries listed in a
2323  // single 'Processor' section, therefore the normal looking
2324  // for duplicates in that section will always fail.
2325  num_avail++;
2326 #else
2327  goto dup_field;
2328 #endif
2329  threadInfo[num_avail][osIdIndex] = val;
2330 #if KMP_OS_LINUX && !(KMP_ARCH_X86 || KMP_ARCH_X86_64)
2331  char path[256];
2332  KMP_SNPRINTF(
2333  path, sizeof(path),
2334  "/sys/devices/system/cpu/cpu%u/topology/physical_package_id",
2335  threadInfo[num_avail][osIdIndex]);
2336  __kmp_read_from_file(path, "%u", &threadInfo[num_avail][pkgIdIndex]);
2337 
2338  KMP_SNPRINTF(path, sizeof(path),
2339  "/sys/devices/system/cpu/cpu%u/topology/core_id",
2340  threadInfo[num_avail][osIdIndex]);
2341  __kmp_read_from_file(path, "%u", &threadInfo[num_avail][coreIdIndex]);
2342  continue;
2343 #else
2344  }
2345  char s2[] = "physical id";
2346  if (strncmp(buf, s2, sizeof(s2) - 1) == 0) {
2347  CHECK_LINE;
2348  char *p = strchr(buf + sizeof(s2) - 1, ':');
2349  unsigned val;
2350  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2351  goto no_val;
2352  if (threadInfo[num_avail][pkgIdIndex] != UINT_MAX)
2353  goto dup_field;
2354  threadInfo[num_avail][pkgIdIndex] = val;
2355  continue;
2356  }
2357  char s3[] = "core id";
2358  if (strncmp(buf, s3, sizeof(s3) - 1) == 0) {
2359  CHECK_LINE;
2360  char *p = strchr(buf + sizeof(s3) - 1, ':');
2361  unsigned val;
2362  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2363  goto no_val;
2364  if (threadInfo[num_avail][coreIdIndex] != UINT_MAX)
2365  goto dup_field;
2366  threadInfo[num_avail][coreIdIndex] = val;
2367  continue;
2368 #endif // KMP_OS_LINUX && USE_SYSFS_INFO
2369  }
2370  char s4[] = "thread id";
2371  if (strncmp(buf, s4, sizeof(s4) - 1) == 0) {
2372  CHECK_LINE;
2373  char *p = strchr(buf + sizeof(s4) - 1, ':');
2374  unsigned val;
2375  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2376  goto no_val;
2377  if (threadInfo[num_avail][threadIdIndex] != UINT_MAX)
2378  goto dup_field;
2379  threadInfo[num_avail][threadIdIndex] = val;
2380  continue;
2381  }
2382  unsigned level;
2383  if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
2384  CHECK_LINE;
2385  char *p = strchr(buf + sizeof(s4) - 1, ':');
2386  unsigned val;
2387  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2388  goto no_val;
2389  // validate the input before using level:
2390  if (level > (unsigned)__kmp_xproc) { // level is too big
2391  level = __kmp_xproc;
2392  }
2393  if (threadInfo[num_avail][nodeIdIndex + level] != UINT_MAX)
2394  goto dup_field;
2395  threadInfo[num_avail][nodeIdIndex + level] = val;
2396  continue;
2397  }
2398 
2399  // We didn't recognize the leading token on the line. There are lots of
2400  // leading tokens that we don't recognize - if the line isn't empty, go on
2401  // to the next line.
2402  if ((*buf != 0) && (*buf != '\n')) {
2403  // If the line is longer than the buffer, read characters
2404  // until we find a newline.
2405  if (long_line) {
2406  int ch;
2407  while (((ch = fgetc(f)) != EOF) && (ch != '\n'))
2408  ;
2409  }
2410  continue;
2411  }
2412 
2413  // A newline has signalled the end of the processor record.
2414  // Check that there aren't too many procs specified.
2415  if ((int)num_avail == __kmp_xproc) {
2416  CLEANUP_THREAD_INFO;
2417  *msg_id = kmp_i18n_str_TooManyEntries;
2418  return false;
2419  }
2420 
2421  // Check for missing fields. The osId field must be there, and we
2422  // currently require that the physical id field is specified, also.
2423  if (threadInfo[num_avail][osIdIndex] == UINT_MAX) {
2424  CLEANUP_THREAD_INFO;
2425  *msg_id = kmp_i18n_str_MissingProcField;
2426  return false;
2427  }
2428  if (threadInfo[0][pkgIdIndex] == UINT_MAX) {
2429  CLEANUP_THREAD_INFO;
2430  *msg_id = kmp_i18n_str_MissingPhysicalIDField;
2431  return false;
2432  }
2433 
2434  // Skip this proc if it is not included in the machine model.
2435  if (!KMP_CPU_ISSET(threadInfo[num_avail][osIdIndex],
2436  __kmp_affin_fullMask)) {
2437  INIT_PROC_INFO(threadInfo[num_avail]);
2438  continue;
2439  }
2440 
2441  // We have a successful parse of this proc's info.
2442  // Increment the counter, and prepare for the next proc.
2443  num_avail++;
2444  KMP_ASSERT(num_avail <= num_records);
2445  INIT_PROC_INFO(threadInfo[num_avail]);
2446  }
2447  continue;
2448 
2449  no_val:
2450  CLEANUP_THREAD_INFO;
2451  *msg_id = kmp_i18n_str_MissingValCpuinfo;
2452  return false;
2453 
2454  dup_field:
2455  CLEANUP_THREAD_INFO;
2456  *msg_id = kmp_i18n_str_DuplicateFieldCpuinfo;
2457  return false;
2458  }
2459  *line = 0;
2460 
2461 #if KMP_MIC && REDUCE_TEAM_SIZE
2462  unsigned teamSize = 0;
2463 #endif // KMP_MIC && REDUCE_TEAM_SIZE
2464 
2465  // check for num_records == __kmp_xproc ???
2466 
2467  // If it is configured to omit the package level when there is only a single
2468  // package, the logic at the end of this routine won't work if there is only a
2469  // single thread
2470  KMP_ASSERT(num_avail > 0);
2471  KMP_ASSERT(num_avail <= num_records);
2472 
2473  // Sort the threadInfo table by physical Id.
2474  qsort(threadInfo, num_avail, sizeof(*threadInfo),
2475  __kmp_affinity_cmp_ProcCpuInfo_phys_id);
2476 
2477  // The table is now sorted by pkgId / coreId / threadId, but we really don't
2478  // know the radix of any of the fields. pkgId's may be sparsely assigned among
2479  // the chips on a system. Although coreId's are usually assigned
2480  // [0 .. coresPerPkg-1] and threadId's are usually assigned
2481  // [0..threadsPerCore-1], we don't want to make any such assumptions.
2482  //
2483  // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
2484  // total # packages) are at this point - we want to determine that now. We
2485  // only have an upper bound on the first two figures.
2486  unsigned *counts =
2487  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2488  unsigned *maxCt =
2489  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2490  unsigned *totals =
2491  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2492  unsigned *lastId =
2493  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2494 
2495  bool assign_thread_ids = false;
2496  unsigned threadIdCt;
2497  unsigned index;
2498 
2499 restart_radix_check:
2500  threadIdCt = 0;
2501 
2502  // Initialize the counter arrays with data from threadInfo[0].
2503  if (assign_thread_ids) {
2504  if (threadInfo[0][threadIdIndex] == UINT_MAX) {
2505  threadInfo[0][threadIdIndex] = threadIdCt++;
2506  } else if (threadIdCt <= threadInfo[0][threadIdIndex]) {
2507  threadIdCt = threadInfo[0][threadIdIndex] + 1;
2508  }
2509  }
2510  for (index = 0; index <= maxIndex; index++) {
2511  counts[index] = 1;
2512  maxCt[index] = 1;
2513  totals[index] = 1;
2514  lastId[index] = threadInfo[0][index];
2515  ;
2516  }
2517 
2518  // Run through the rest of the OS procs.
2519  for (i = 1; i < num_avail; i++) {
2520  // Find the most significant index whose id differs from the id for the
2521  // previous OS proc.
2522  for (index = maxIndex; index >= threadIdIndex; index--) {
2523  if (assign_thread_ids && (index == threadIdIndex)) {
2524  // Auto-assign the thread id field if it wasn't specified.
2525  if (threadInfo[i][threadIdIndex] == UINT_MAX) {
2526  threadInfo[i][threadIdIndex] = threadIdCt++;
2527  }
2528  // Apparently the thread id field was specified for some entries and not
2529  // others. Start the thread id counter off at the next higher thread id.
2530  else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
2531  threadIdCt = threadInfo[i][threadIdIndex] + 1;
2532  }
2533  }
2534  if (threadInfo[i][index] != lastId[index]) {
2535  // Run through all indices which are less significant, and reset the
2536  // counts to 1. At all levels up to and including index, we need to
2537  // increment the totals and record the last id.
2538  unsigned index2;
2539  for (index2 = threadIdIndex; index2 < index; index2++) {
2540  totals[index2]++;
2541  if (counts[index2] > maxCt[index2]) {
2542  maxCt[index2] = counts[index2];
2543  }
2544  counts[index2] = 1;
2545  lastId[index2] = threadInfo[i][index2];
2546  }
2547  counts[index]++;
2548  totals[index]++;
2549  lastId[index] = threadInfo[i][index];
2550 
2551  if (assign_thread_ids && (index > threadIdIndex)) {
2552 
2553 #if KMP_MIC && REDUCE_TEAM_SIZE
2554  // The default team size is the total #threads in the machine
2555  // minus 1 thread for every core that has 3 or more threads.
2556  teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
2557 #endif // KMP_MIC && REDUCE_TEAM_SIZE
2558 
2559  // Restart the thread counter, as we are on a new core.
2560  threadIdCt = 0;
2561 
2562  // Auto-assign the thread id field if it wasn't specified.
2563  if (threadInfo[i][threadIdIndex] == UINT_MAX) {
2564  threadInfo[i][threadIdIndex] = threadIdCt++;
2565  }
2566 
2567  // Apparently the thread id field was specified for some entries and
2568  // not others. Start the thread id counter off at the next higher
2569  // thread id.
2570  else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
2571  threadIdCt = threadInfo[i][threadIdIndex] + 1;
2572  }
2573  }
2574  break;
2575  }
2576  }
2577  if (index < threadIdIndex) {
2578  // If thread ids were specified, it is an error if they are not unique.
2579  // Also, check that we waven't already restarted the loop (to be safe -
2580  // shouldn't need to).
2581  if ((threadInfo[i][threadIdIndex] != UINT_MAX) || assign_thread_ids) {
2582  __kmp_free(lastId);
2583  __kmp_free(totals);
2584  __kmp_free(maxCt);
2585  __kmp_free(counts);
2586  CLEANUP_THREAD_INFO;
2587  *msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
2588  return false;
2589  }
2590 
2591  // If the thread ids were not specified and we see entries entries that
2592  // are duplicates, start the loop over and assign the thread ids manually.
2593  assign_thread_ids = true;
2594  goto restart_radix_check;
2595  }
2596  }
2597 
2598 #if KMP_MIC && REDUCE_TEAM_SIZE
2599  // The default team size is the total #threads in the machine
2600  // minus 1 thread for every core that has 3 or more threads.
2601  teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
2602 #endif // KMP_MIC && REDUCE_TEAM_SIZE
2603 
2604  for (index = threadIdIndex; index <= maxIndex; index++) {
2605  if (counts[index] > maxCt[index]) {
2606  maxCt[index] = counts[index];
2607  }
2608  }
2609 
2610  __kmp_nThreadsPerCore = maxCt[threadIdIndex];
2611  nCoresPerPkg = maxCt[coreIdIndex];
2612  nPackages = totals[pkgIdIndex];
2613 
2614  // When affinity is off, this routine will still be called to set
2615  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
2616  // Make sure all these vars are set correctly, and return now if affinity is
2617  // not enabled.
2618  __kmp_ncores = totals[coreIdIndex];
2619  if (!KMP_AFFINITY_CAPABLE()) {
2620  KMP_ASSERT(__kmp_affinity_type == affinity_none);
2621  return true;
2622  }
2623 
2624 #if KMP_MIC && REDUCE_TEAM_SIZE
2625  // Set the default team size.
2626  if ((__kmp_dflt_team_nth == 0) && (teamSize > 0)) {
2627  __kmp_dflt_team_nth = teamSize;
2628  KA_TRACE(20, ("__kmp_affinity_create_cpuinfo_map: setting "
2629  "__kmp_dflt_team_nth = %d\n",
2630  __kmp_dflt_team_nth));
2631  }
2632 #endif // KMP_MIC && REDUCE_TEAM_SIZE
2633 
2634  KMP_DEBUG_ASSERT(num_avail == (unsigned)__kmp_avail_proc);
2635 
2636  // Count the number of levels which have more nodes at that level than at the
2637  // parent's level (with there being an implicit root node of the top level).
2638  // This is equivalent to saying that there is at least one node at this level
2639  // which has a sibling. These levels are in the map, and the package level is
2640  // always in the map.
2641  bool *inMap = (bool *)__kmp_allocate((maxIndex + 1) * sizeof(bool));
2642  for (index = threadIdIndex; index < maxIndex; index++) {
2643  KMP_ASSERT(totals[index] >= totals[index + 1]);
2644  inMap[index] = (totals[index] > totals[index + 1]);
2645  }
2646  inMap[maxIndex] = (totals[maxIndex] > 1);
2647  inMap[pkgIdIndex] = true;
2648  inMap[coreIdIndex] = true;
2649  inMap[threadIdIndex] = true;
2650 
2651  int depth = 0;
2652  int idx = 0;
2653  kmp_hw_t types[KMP_HW_LAST];
2654  int pkgLevel = -1;
2655  int coreLevel = -1;
2656  int threadLevel = -1;
2657  for (index = threadIdIndex; index <= maxIndex; index++) {
2658  if (inMap[index]) {
2659  depth++;
2660  }
2661  }
2662  if (inMap[pkgIdIndex]) {
2663  pkgLevel = idx;
2664  types[idx++] = KMP_HW_SOCKET;
2665  }
2666  if (inMap[coreIdIndex]) {
2667  coreLevel = idx;
2668  types[idx++] = KMP_HW_CORE;
2669  }
2670  if (inMap[threadIdIndex]) {
2671  threadLevel = idx;
2672  types[idx++] = KMP_HW_THREAD;
2673  }
2674  KMP_ASSERT(depth > 0);
2675 
2676  // Construct the data structure that is to be returned.
2677  __kmp_topology = kmp_topology_t::allocate(num_avail, depth, types);
2678 
2679  for (i = 0; i < num_avail; ++i) {
2680  unsigned os = threadInfo[i][osIdIndex];
2681  int src_index;
2682  int dst_index = 0;
2683  kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
2684  hw_thread.clear();
2685  hw_thread.os_id = os;
2686 
2687  idx = 0;
2688  for (src_index = maxIndex; src_index >= threadIdIndex; src_index--) {
2689  if (!inMap[src_index]) {
2690  continue;
2691  }
2692  if (src_index == pkgIdIndex) {
2693  hw_thread.ids[pkgLevel] = threadInfo[i][src_index];
2694  } else if (src_index == coreIdIndex) {
2695  hw_thread.ids[coreLevel] = threadInfo[i][src_index];
2696  } else if (src_index == threadIdIndex) {
2697  hw_thread.ids[threadLevel] = threadInfo[i][src_index];
2698  }
2699  dst_index++;
2700  }
2701  }
2702 
2703  __kmp_free(inMap);
2704  __kmp_free(lastId);
2705  __kmp_free(totals);
2706  __kmp_free(maxCt);
2707  __kmp_free(counts);
2708  CLEANUP_THREAD_INFO;
2709  __kmp_topology->sort_ids();
2710  if (!__kmp_topology->check_ids()) {
2711  kmp_topology_t::deallocate(__kmp_topology);
2712  __kmp_topology = nullptr;
2713  *msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
2714  return false;
2715  }
2716  return true;
2717 }
2718 
2719 // Create and return a table of affinity masks, indexed by OS thread ID.
2720 // This routine handles OR'ing together all the affinity masks of threads
2721 // that are sufficiently close, if granularity > fine.
2722 static kmp_affin_mask_t *__kmp_create_masks(unsigned *maxIndex,
2723  unsigned *numUnique) {
2724  // First form a table of affinity masks in order of OS thread id.
2725  int maxOsId;
2726  int i;
2727  int numAddrs = __kmp_topology->get_num_hw_threads();
2728  int depth = __kmp_topology->get_depth();
2729  KMP_ASSERT(numAddrs);
2730  KMP_ASSERT(depth);
2731 
2732  maxOsId = 0;
2733  for (i = numAddrs - 1;; --i) {
2734  int osId = __kmp_topology->at(i).os_id;
2735  if (osId > maxOsId) {
2736  maxOsId = osId;
2737  }
2738  if (i == 0)
2739  break;
2740  }
2741  kmp_affin_mask_t *osId2Mask;
2742  KMP_CPU_ALLOC_ARRAY(osId2Mask, (maxOsId + 1));
2743  KMP_ASSERT(__kmp_affinity_gran_levels >= 0);
2744  if (__kmp_affinity_verbose && (__kmp_affinity_gran_levels > 0)) {
2745  KMP_INFORM(ThreadsMigrate, "KMP_AFFINITY", __kmp_affinity_gran_levels);
2746  }
2747  if (__kmp_affinity_gran_levels >= (int)depth) {
2748  if (__kmp_affinity_verbose ||
2749  (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) {
2750  KMP_WARNING(AffThreadsMayMigrate);
2751  }
2752  }
2753 
2754  // Run through the table, forming the masks for all threads on each core.
2755  // Threads on the same core will have identical kmp_hw_thread_t objects, not
2756  // considering the last level, which must be the thread id. All threads on a
2757  // core will appear consecutively.
2758  int unique = 0;
2759  int j = 0; // index of 1st thread on core
2760  int leader = 0;
2761  kmp_affin_mask_t *sum;
2762  KMP_CPU_ALLOC_ON_STACK(sum);
2763  KMP_CPU_ZERO(sum);
2764  KMP_CPU_SET(__kmp_topology->at(0).os_id, sum);
2765  for (i = 1; i < numAddrs; i++) {
2766  // If this thread is sufficiently close to the leader (within the
2767  // granularity setting), then set the bit for this os thread in the
2768  // affinity mask for this group, and go on to the next thread.
2769  if (__kmp_topology->is_close(leader, i, __kmp_affinity_gran_levels)) {
2770  KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
2771  continue;
2772  }
2773 
2774  // For every thread in this group, copy the mask to the thread's entry in
2775  // the osId2Mask table. Mark the first address as a leader.
2776  for (; j < i; j++) {
2777  int osId = __kmp_topology->at(j).os_id;
2778  KMP_DEBUG_ASSERT(osId <= maxOsId);
2779  kmp_affin_mask_t *mask = KMP_CPU_INDEX(osId2Mask, osId);
2780  KMP_CPU_COPY(mask, sum);
2781  __kmp_topology->at(j).leader = (j == leader);
2782  }
2783  unique++;
2784 
2785  // Start a new mask.
2786  leader = i;
2787  KMP_CPU_ZERO(sum);
2788  KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
2789  }
2790 
2791  // For every thread in last group, copy the mask to the thread's
2792  // entry in the osId2Mask table.
2793  for (; j < i; j++) {
2794  int osId = __kmp_topology->at(j).os_id;
2795  KMP_DEBUG_ASSERT(osId <= maxOsId);
2796  kmp_affin_mask_t *mask = KMP_CPU_INDEX(osId2Mask, osId);
2797  KMP_CPU_COPY(mask, sum);
2798  __kmp_topology->at(j).leader = (j == leader);
2799  }
2800  unique++;
2801  KMP_CPU_FREE_FROM_STACK(sum);
2802 
2803  *maxIndex = maxOsId;
2804  *numUnique = unique;
2805  return osId2Mask;
2806 }
2807 
2808 // Stuff for the affinity proclist parsers. It's easier to declare these vars
2809 // as file-static than to try and pass them through the calling sequence of
2810 // the recursive-descent OMP_PLACES parser.
2811 static kmp_affin_mask_t *newMasks;
2812 static int numNewMasks;
2813 static int nextNewMask;
2814 
2815 #define ADD_MASK(_mask) \
2816  { \
2817  if (nextNewMask >= numNewMasks) { \
2818  int i; \
2819  numNewMasks *= 2; \
2820  kmp_affin_mask_t *temp; \
2821  KMP_CPU_INTERNAL_ALLOC_ARRAY(temp, numNewMasks); \
2822  for (i = 0; i < numNewMasks / 2; i++) { \
2823  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i); \
2824  kmp_affin_mask_t *dest = KMP_CPU_INDEX(temp, i); \
2825  KMP_CPU_COPY(dest, src); \
2826  } \
2827  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks / 2); \
2828  newMasks = temp; \
2829  } \
2830  KMP_CPU_COPY(KMP_CPU_INDEX(newMasks, nextNewMask), (_mask)); \
2831  nextNewMask++; \
2832  }
2833 
2834 #define ADD_MASK_OSID(_osId, _osId2Mask, _maxOsId) \
2835  { \
2836  if (((_osId) > _maxOsId) || \
2837  (!KMP_CPU_ISSET((_osId), KMP_CPU_INDEX((_osId2Mask), (_osId))))) { \
2838  if (__kmp_affinity_verbose || \
2839  (__kmp_affinity_warnings && \
2840  (__kmp_affinity_type != affinity_none))) { \
2841  KMP_WARNING(AffIgnoreInvalidProcID, _osId); \
2842  } \
2843  } else { \
2844  ADD_MASK(KMP_CPU_INDEX(_osId2Mask, (_osId))); \
2845  } \
2846  }
2847 
2848 // Re-parse the proclist (for the explicit affinity type), and form the list
2849 // of affinity newMasks indexed by gtid.
2850 static void __kmp_affinity_process_proclist(kmp_affin_mask_t **out_masks,
2851  unsigned int *out_numMasks,
2852  const char *proclist,
2853  kmp_affin_mask_t *osId2Mask,
2854  int maxOsId) {
2855  int i;
2856  const char *scan = proclist;
2857  const char *next = proclist;
2858 
2859  // We use malloc() for the temporary mask vector, so that we can use
2860  // realloc() to extend it.
2861  numNewMasks = 2;
2862  KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
2863  nextNewMask = 0;
2864  kmp_affin_mask_t *sumMask;
2865  KMP_CPU_ALLOC(sumMask);
2866  int setSize = 0;
2867 
2868  for (;;) {
2869  int start, end, stride;
2870 
2871  SKIP_WS(scan);
2872  next = scan;
2873  if (*next == '\0') {
2874  break;
2875  }
2876 
2877  if (*next == '{') {
2878  int num;
2879  setSize = 0;
2880  next++; // skip '{'
2881  SKIP_WS(next);
2882  scan = next;
2883 
2884  // Read the first integer in the set.
2885  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad proclist");
2886  SKIP_DIGITS(next);
2887  num = __kmp_str_to_int(scan, *next);
2888  KMP_ASSERT2(num >= 0, "bad explicit proc list");
2889 
2890  // Copy the mask for that osId to the sum (union) mask.
2891  if ((num > maxOsId) ||
2892  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
2893  if (__kmp_affinity_verbose ||
2894  (__kmp_affinity_warnings &&
2895  (__kmp_affinity_type != affinity_none))) {
2896  KMP_WARNING(AffIgnoreInvalidProcID, num);
2897  }
2898  KMP_CPU_ZERO(sumMask);
2899  } else {
2900  KMP_CPU_COPY(sumMask, KMP_CPU_INDEX(osId2Mask, num));
2901  setSize = 1;
2902  }
2903 
2904  for (;;) {
2905  // Check for end of set.
2906  SKIP_WS(next);
2907  if (*next == '}') {
2908  next++; // skip '}'
2909  break;
2910  }
2911 
2912  // Skip optional comma.
2913  if (*next == ',') {
2914  next++;
2915  }
2916  SKIP_WS(next);
2917 
2918  // Read the next integer in the set.
2919  scan = next;
2920  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
2921 
2922  SKIP_DIGITS(next);
2923  num = __kmp_str_to_int(scan, *next);
2924  KMP_ASSERT2(num >= 0, "bad explicit proc list");
2925 
2926  // Add the mask for that osId to the sum mask.
2927  if ((num > maxOsId) ||
2928  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
2929  if (__kmp_affinity_verbose ||
2930  (__kmp_affinity_warnings &&
2931  (__kmp_affinity_type != affinity_none))) {
2932  KMP_WARNING(AffIgnoreInvalidProcID, num);
2933  }
2934  } else {
2935  KMP_CPU_UNION(sumMask, KMP_CPU_INDEX(osId2Mask, num));
2936  setSize++;
2937  }
2938  }
2939  if (setSize > 0) {
2940  ADD_MASK(sumMask);
2941  }
2942 
2943  SKIP_WS(next);
2944  if (*next == ',') {
2945  next++;
2946  }
2947  scan = next;
2948  continue;
2949  }
2950 
2951  // Read the first integer.
2952  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
2953  SKIP_DIGITS(next);
2954  start = __kmp_str_to_int(scan, *next);
2955  KMP_ASSERT2(start >= 0, "bad explicit proc list");
2956  SKIP_WS(next);
2957 
2958  // If this isn't a range, then add a mask to the list and go on.
2959  if (*next != '-') {
2960  ADD_MASK_OSID(start, osId2Mask, maxOsId);
2961 
2962  // Skip optional comma.
2963  if (*next == ',') {
2964  next++;
2965  }
2966  scan = next;
2967  continue;
2968  }
2969 
2970  // This is a range. Skip over the '-' and read in the 2nd int.
2971  next++; // skip '-'
2972  SKIP_WS(next);
2973  scan = next;
2974  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
2975  SKIP_DIGITS(next);
2976  end = __kmp_str_to_int(scan, *next);
2977  KMP_ASSERT2(end >= 0, "bad explicit proc list");
2978 
2979  // Check for a stride parameter
2980  stride = 1;
2981  SKIP_WS(next);
2982  if (*next == ':') {
2983  // A stride is specified. Skip over the ':" and read the 3rd int.
2984  int sign = +1;
2985  next++; // skip ':'
2986  SKIP_WS(next);
2987  scan = next;
2988  if (*next == '-') {
2989  sign = -1;
2990  next++;
2991  SKIP_WS(next);
2992  scan = next;
2993  }
2994  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
2995  SKIP_DIGITS(next);
2996  stride = __kmp_str_to_int(scan, *next);
2997  KMP_ASSERT2(stride >= 0, "bad explicit proc list");
2998  stride *= sign;
2999  }
3000 
3001  // Do some range checks.
3002  KMP_ASSERT2(stride != 0, "bad explicit proc list");
3003  if (stride > 0) {
3004  KMP_ASSERT2(start <= end, "bad explicit proc list");
3005  } else {
3006  KMP_ASSERT2(start >= end, "bad explicit proc list");
3007  }
3008  KMP_ASSERT2((end - start) / stride <= 65536, "bad explicit proc list");
3009 
3010  // Add the mask for each OS proc # to the list.
3011  if (stride > 0) {
3012  do {
3013  ADD_MASK_OSID(start, osId2Mask, maxOsId);
3014  start += stride;
3015  } while (start <= end);
3016  } else {
3017  do {
3018  ADD_MASK_OSID(start, osId2Mask, maxOsId);
3019  start += stride;
3020  } while (start >= end);
3021  }
3022 
3023  // Skip optional comma.
3024  SKIP_WS(next);
3025  if (*next == ',') {
3026  next++;
3027  }
3028  scan = next;
3029  }
3030 
3031  *out_numMasks = nextNewMask;
3032  if (nextNewMask == 0) {
3033  *out_masks = NULL;
3034  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3035  return;
3036  }
3037  KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
3038  for (i = 0; i < nextNewMask; i++) {
3039  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
3040  kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
3041  KMP_CPU_COPY(dest, src);
3042  }
3043  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3044  KMP_CPU_FREE(sumMask);
3045 }
3046 
3047 /*-----------------------------------------------------------------------------
3048 Re-parse the OMP_PLACES proc id list, forming the newMasks for the different
3049 places. Again, Here is the grammar:
3050 
3051 place_list := place
3052 place_list := place , place_list
3053 place := num
3054 place := place : num
3055 place := place : num : signed
3056 place := { subplacelist }
3057 place := ! place // (lowest priority)
3058 subplace_list := subplace
3059 subplace_list := subplace , subplace_list
3060 subplace := num
3061 subplace := num : num
3062 subplace := num : num : signed
3063 signed := num
3064 signed := + signed
3065 signed := - signed
3066 -----------------------------------------------------------------------------*/
3067 static void __kmp_process_subplace_list(const char **scan,
3068  kmp_affin_mask_t *osId2Mask,
3069  int maxOsId, kmp_affin_mask_t *tempMask,
3070  int *setSize) {
3071  const char *next;
3072 
3073  for (;;) {
3074  int start, count, stride, i;
3075 
3076  // Read in the starting proc id
3077  SKIP_WS(*scan);
3078  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3079  next = *scan;
3080  SKIP_DIGITS(next);
3081  start = __kmp_str_to_int(*scan, *next);
3082  KMP_ASSERT(start >= 0);
3083  *scan = next;
3084 
3085  // valid follow sets are ',' ':' and '}'
3086  SKIP_WS(*scan);
3087  if (**scan == '}' || **scan == ',') {
3088  if ((start > maxOsId) ||
3089  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3090  if (__kmp_affinity_verbose ||
3091  (__kmp_affinity_warnings &&
3092  (__kmp_affinity_type != affinity_none))) {
3093  KMP_WARNING(AffIgnoreInvalidProcID, start);
3094  }
3095  } else {
3096  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3097  (*setSize)++;
3098  }
3099  if (**scan == '}') {
3100  break;
3101  }
3102  (*scan)++; // skip ','
3103  continue;
3104  }
3105  KMP_ASSERT2(**scan == ':', "bad explicit places list");
3106  (*scan)++; // skip ':'
3107 
3108  // Read count parameter
3109  SKIP_WS(*scan);
3110  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3111  next = *scan;
3112  SKIP_DIGITS(next);
3113  count = __kmp_str_to_int(*scan, *next);
3114  KMP_ASSERT(count >= 0);
3115  *scan = next;
3116 
3117  // valid follow sets are ',' ':' and '}'
3118  SKIP_WS(*scan);
3119  if (**scan == '}' || **scan == ',') {
3120  for (i = 0; i < count; i++) {
3121  if ((start > maxOsId) ||
3122  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3123  if (__kmp_affinity_verbose ||
3124  (__kmp_affinity_warnings &&
3125  (__kmp_affinity_type != affinity_none))) {
3126  KMP_WARNING(AffIgnoreInvalidProcID, start);
3127  }
3128  break; // don't proliferate warnings for large count
3129  } else {
3130  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3131  start++;
3132  (*setSize)++;
3133  }
3134  }
3135  if (**scan == '}') {
3136  break;
3137  }
3138  (*scan)++; // skip ','
3139  continue;
3140  }
3141  KMP_ASSERT2(**scan == ':', "bad explicit places list");
3142  (*scan)++; // skip ':'
3143 
3144  // Read stride parameter
3145  int sign = +1;
3146  for (;;) {
3147  SKIP_WS(*scan);
3148  if (**scan == '+') {
3149  (*scan)++; // skip '+'
3150  continue;
3151  }
3152  if (**scan == '-') {
3153  sign *= -1;
3154  (*scan)++; // skip '-'
3155  continue;
3156  }
3157  break;
3158  }
3159  SKIP_WS(*scan);
3160  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3161  next = *scan;
3162  SKIP_DIGITS(next);
3163  stride = __kmp_str_to_int(*scan, *next);
3164  KMP_ASSERT(stride >= 0);
3165  *scan = next;
3166  stride *= sign;
3167 
3168  // valid follow sets are ',' and '}'
3169  SKIP_WS(*scan);
3170  if (**scan == '}' || **scan == ',') {
3171  for (i = 0; i < count; i++) {
3172  if ((start > maxOsId) ||
3173  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3174  if (__kmp_affinity_verbose ||
3175  (__kmp_affinity_warnings &&
3176  (__kmp_affinity_type != affinity_none))) {
3177  KMP_WARNING(AffIgnoreInvalidProcID, start);
3178  }
3179  break; // don't proliferate warnings for large count
3180  } else {
3181  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3182  start += stride;
3183  (*setSize)++;
3184  }
3185  }
3186  if (**scan == '}') {
3187  break;
3188  }
3189  (*scan)++; // skip ','
3190  continue;
3191  }
3192 
3193  KMP_ASSERT2(0, "bad explicit places list");
3194  }
3195 }
3196 
3197 static void __kmp_process_place(const char **scan, kmp_affin_mask_t *osId2Mask,
3198  int maxOsId, kmp_affin_mask_t *tempMask,
3199  int *setSize) {
3200  const char *next;
3201 
3202  // valid follow sets are '{' '!' and num
3203  SKIP_WS(*scan);
3204  if (**scan == '{') {
3205  (*scan)++; // skip '{'
3206  __kmp_process_subplace_list(scan, osId2Mask, maxOsId, tempMask, setSize);
3207  KMP_ASSERT2(**scan == '}', "bad explicit places list");
3208  (*scan)++; // skip '}'
3209  } else if (**scan == '!') {
3210  (*scan)++; // skip '!'
3211  __kmp_process_place(scan, osId2Mask, maxOsId, tempMask, setSize);
3212  KMP_CPU_COMPLEMENT(maxOsId, tempMask);
3213  } else if ((**scan >= '0') && (**scan <= '9')) {
3214  next = *scan;
3215  SKIP_DIGITS(next);
3216  int num = __kmp_str_to_int(*scan, *next);
3217  KMP_ASSERT(num >= 0);
3218  if ((num > maxOsId) ||
3219  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
3220  if (__kmp_affinity_verbose ||
3221  (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) {
3222  KMP_WARNING(AffIgnoreInvalidProcID, num);
3223  }
3224  } else {
3225  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, num));
3226  (*setSize)++;
3227  }
3228  *scan = next; // skip num
3229  } else {
3230  KMP_ASSERT2(0, "bad explicit places list");
3231  }
3232 }
3233 
3234 // static void
3235 void __kmp_affinity_process_placelist(kmp_affin_mask_t **out_masks,
3236  unsigned int *out_numMasks,
3237  const char *placelist,
3238  kmp_affin_mask_t *osId2Mask,
3239  int maxOsId) {
3240  int i, j, count, stride, sign;
3241  const char *scan = placelist;
3242  const char *next = placelist;
3243 
3244  numNewMasks = 2;
3245  KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
3246  nextNewMask = 0;
3247 
3248  // tempMask is modified based on the previous or initial
3249  // place to form the current place
3250  // previousMask contains the previous place
3251  kmp_affin_mask_t *tempMask;
3252  kmp_affin_mask_t *previousMask;
3253  KMP_CPU_ALLOC(tempMask);
3254  KMP_CPU_ZERO(tempMask);
3255  KMP_CPU_ALLOC(previousMask);
3256  KMP_CPU_ZERO(previousMask);
3257  int setSize = 0;
3258 
3259  for (;;) {
3260  __kmp_process_place(&scan, osId2Mask, maxOsId, tempMask, &setSize);
3261 
3262  // valid follow sets are ',' ':' and EOL
3263  SKIP_WS(scan);
3264  if (*scan == '\0' || *scan == ',') {
3265  if (setSize > 0) {
3266  ADD_MASK(tempMask);
3267  }
3268  KMP_CPU_ZERO(tempMask);
3269  setSize = 0;
3270  if (*scan == '\0') {
3271  break;
3272  }
3273  scan++; // skip ','
3274  continue;
3275  }
3276 
3277  KMP_ASSERT2(*scan == ':', "bad explicit places list");
3278  scan++; // skip ':'
3279 
3280  // Read count parameter
3281  SKIP_WS(scan);
3282  KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
3283  next = scan;
3284  SKIP_DIGITS(next);
3285  count = __kmp_str_to_int(scan, *next);
3286  KMP_ASSERT(count >= 0);
3287  scan = next;
3288 
3289  // valid follow sets are ',' ':' and EOL
3290  SKIP_WS(scan);
3291  if (*scan == '\0' || *scan == ',') {
3292  stride = +1;
3293  } else {
3294  KMP_ASSERT2(*scan == ':', "bad explicit places list");
3295  scan++; // skip ':'
3296 
3297  // Read stride parameter
3298  sign = +1;
3299  for (;;) {
3300  SKIP_WS(scan);
3301  if (*scan == '+') {
3302  scan++; // skip '+'
3303  continue;
3304  }
3305  if (*scan == '-') {
3306  sign *= -1;
3307  scan++; // skip '-'
3308  continue;
3309  }
3310  break;
3311  }
3312  SKIP_WS(scan);
3313  KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
3314  next = scan;
3315  SKIP_DIGITS(next);
3316  stride = __kmp_str_to_int(scan, *next);
3317  KMP_DEBUG_ASSERT(stride >= 0);
3318  scan = next;
3319  stride *= sign;
3320  }
3321 
3322  // Add places determined by initial_place : count : stride
3323  for (i = 0; i < count; i++) {
3324  if (setSize == 0) {
3325  break;
3326  }
3327  // Add the current place, then build the next place (tempMask) from that
3328  KMP_CPU_COPY(previousMask, tempMask);
3329  ADD_MASK(previousMask);
3330  KMP_CPU_ZERO(tempMask);
3331  setSize = 0;
3332  KMP_CPU_SET_ITERATE(j, previousMask) {
3333  if (!KMP_CPU_ISSET(j, previousMask)) {
3334  continue;
3335  }
3336  if ((j + stride > maxOsId) || (j + stride < 0) ||
3337  (!KMP_CPU_ISSET(j, __kmp_affin_fullMask)) ||
3338  (!KMP_CPU_ISSET(j + stride,
3339  KMP_CPU_INDEX(osId2Mask, j + stride)))) {
3340  if ((__kmp_affinity_verbose ||
3341  (__kmp_affinity_warnings &&
3342  (__kmp_affinity_type != affinity_none))) &&
3343  i < count - 1) {
3344  KMP_WARNING(AffIgnoreInvalidProcID, j + stride);
3345  }
3346  continue;
3347  }
3348  KMP_CPU_SET(j + stride, tempMask);
3349  setSize++;
3350  }
3351  }
3352  KMP_CPU_ZERO(tempMask);
3353  setSize = 0;
3354 
3355  // valid follow sets are ',' and EOL
3356  SKIP_WS(scan);
3357  if (*scan == '\0') {
3358  break;
3359  }
3360  if (*scan == ',') {
3361  scan++; // skip ','
3362  continue;
3363  }
3364 
3365  KMP_ASSERT2(0, "bad explicit places list");
3366  }
3367 
3368  *out_numMasks = nextNewMask;
3369  if (nextNewMask == 0) {
3370  *out_masks = NULL;
3371  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3372  return;
3373  }
3374  KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
3375  KMP_CPU_FREE(tempMask);
3376  KMP_CPU_FREE(previousMask);
3377  for (i = 0; i < nextNewMask; i++) {
3378  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
3379  kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
3380  KMP_CPU_COPY(dest, src);
3381  }
3382  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3383 }
3384 
3385 #undef ADD_MASK
3386 #undef ADD_MASK_OSID
3387 
3388 // This function figures out the deepest level at which there is at least one
3389 // cluster/core with more than one processing unit bound to it.
3390 static int __kmp_affinity_find_core_level(int nprocs, int bottom_level) {
3391  int core_level = 0;
3392 
3393  for (int i = 0; i < nprocs; i++) {
3394  const kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
3395  for (int j = bottom_level; j > 0; j--) {
3396  if (hw_thread.ids[j] > 0) {
3397  if (core_level < (j - 1)) {
3398  core_level = j - 1;
3399  }
3400  }
3401  }
3402  }
3403  return core_level;
3404 }
3405 
3406 // This function counts number of clusters/cores at given level.
3407 static int __kmp_affinity_compute_ncores(int nprocs, int bottom_level,
3408  int core_level) {
3409  return __kmp_topology->get_count(core_level);
3410 }
3411 // This function finds to which cluster/core given processing unit is bound.
3412 static int __kmp_affinity_find_core(int proc, int bottom_level,
3413  int core_level) {
3414  int core = 0;
3415  KMP_DEBUG_ASSERT(proc >= 0 && proc < __kmp_topology->get_num_hw_threads());
3416  for (int i = 0; i <= proc; ++i) {
3417  if (i + 1 <= proc) {
3418  for (int j = 0; j <= core_level; ++j) {
3419  if (__kmp_topology->at(i + 1).sub_ids[j] !=
3420  __kmp_topology->at(i).sub_ids[j]) {
3421  core++;
3422  break;
3423  }
3424  }
3425  }
3426  }
3427  return core;
3428 }
3429 
3430 // This function finds maximal number of processing units bound to a
3431 // cluster/core at given level.
3432 static int __kmp_affinity_max_proc_per_core(int nprocs, int bottom_level,
3433  int core_level) {
3434  if (core_level >= bottom_level)
3435  return 1;
3436  int thread_level = __kmp_topology->get_level(KMP_HW_THREAD);
3437  return __kmp_topology->calculate_ratio(thread_level, core_level);
3438 }
3439 
3440 static int *procarr = NULL;
3441 static int __kmp_aff_depth = 0;
3442 
3443 // Create a one element mask array (set of places) which only contains the
3444 // initial process's affinity mask
3445 static void __kmp_create_affinity_none_places() {
3446  KMP_ASSERT(__kmp_affin_fullMask != NULL);
3447  KMP_ASSERT(__kmp_affinity_type == affinity_none);
3448  __kmp_affinity_num_masks = 1;
3449  KMP_CPU_ALLOC_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks);
3450  kmp_affin_mask_t *dest = KMP_CPU_INDEX(__kmp_affinity_masks, 0);
3451  KMP_CPU_COPY(dest, __kmp_affin_fullMask);
3452 }
3453 
3454 static void __kmp_aux_affinity_initialize(void) {
3455  if (__kmp_affinity_masks != NULL) {
3456  KMP_ASSERT(__kmp_affin_fullMask != NULL);
3457  return;
3458  }
3459 
3460  // Create the "full" mask - this defines all of the processors that we
3461  // consider to be in the machine model. If respect is set, then it is the
3462  // initialization thread's affinity mask. Otherwise, it is all processors that
3463  // we know about on the machine.
3464  if (__kmp_affin_fullMask == NULL) {
3465  KMP_CPU_ALLOC(__kmp_affin_fullMask);
3466  }
3467  if (KMP_AFFINITY_CAPABLE()) {
3468  __kmp_get_system_affinity(__kmp_affin_fullMask, TRUE);
3469  if (__kmp_affinity_respect_mask) {
3470  // Count the number of available processors.
3471  unsigned i;
3472  __kmp_avail_proc = 0;
3473  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
3474  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
3475  continue;
3476  }
3477  __kmp_avail_proc++;
3478  }
3479  if (__kmp_avail_proc > __kmp_xproc) {
3480  if (__kmp_affinity_verbose ||
3481  (__kmp_affinity_warnings &&
3482  (__kmp_affinity_type != affinity_none))) {
3483  KMP_WARNING(ErrorInitializeAffinity);
3484  }
3485  __kmp_affinity_type = affinity_none;
3486  KMP_AFFINITY_DISABLE();
3487  return;
3488  }
3489 
3490  if (__kmp_affinity_verbose) {
3491  char buf[KMP_AFFIN_MASK_PRINT_LEN];
3492  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
3493  __kmp_affin_fullMask);
3494  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
3495  }
3496  } else {
3497  if (__kmp_affinity_verbose) {
3498  char buf[KMP_AFFIN_MASK_PRINT_LEN];
3499  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
3500  __kmp_affin_fullMask);
3501  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
3502  }
3503  __kmp_affinity_entire_machine_mask(__kmp_affin_fullMask);
3504  __kmp_avail_proc = __kmp_xproc;
3505 #if KMP_OS_WINDOWS
3506  // Set the process affinity mask since threads' affinity
3507  // masks must be subset of process mask in Windows* OS
3508  __kmp_affin_fullMask->set_process_affinity(true);
3509 #endif
3510  }
3511  }
3512 
3513  kmp_i18n_id_t msg_id = kmp_i18n_null;
3514 
3515  // For backward compatibility, setting KMP_CPUINFO_FILE =>
3516  // KMP_TOPOLOGY_METHOD=cpuinfo
3517  if ((__kmp_cpuinfo_file != NULL) &&
3518  (__kmp_affinity_top_method == affinity_top_method_all)) {
3519  __kmp_affinity_top_method = affinity_top_method_cpuinfo;
3520  }
3521 
3522  bool success = false;
3523  if (__kmp_affinity_top_method == affinity_top_method_all) {
3524 // In the default code path, errors are not fatal - we just try using
3525 // another method. We only emit a warning message if affinity is on, or the
3526 // verbose flag is set, an the nowarnings flag was not set.
3527 #if KMP_USE_HWLOC
3528  if (!success &&
3529  __kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC) {
3530  if (!__kmp_hwloc_error) {
3531  success = __kmp_affinity_create_hwloc_map(&msg_id);
3532  if (!success && __kmp_affinity_verbose) {
3533  KMP_INFORM(AffIgnoringHwloc, "KMP_AFFINITY");
3534  }
3535  } else if (__kmp_affinity_verbose) {
3536  KMP_INFORM(AffIgnoringHwloc, "KMP_AFFINITY");
3537  }
3538  }
3539 #endif
3540 
3541 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3542  if (!success) {
3543  success = __kmp_affinity_create_x2apicid_map(&msg_id);
3544  if (!success && __kmp_affinity_verbose && msg_id != kmp_i18n_null) {
3545  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", __kmp_i18n_catgets(msg_id));
3546  }
3547  }
3548  if (!success) {
3549  success = __kmp_affinity_create_apicid_map(&msg_id);
3550  if (!success && __kmp_affinity_verbose && msg_id != kmp_i18n_null) {
3551  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", __kmp_i18n_catgets(msg_id));
3552  }
3553  }
3554 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3555 
3556 #if KMP_OS_LINUX
3557  if (!success) {
3558  int line = 0;
3559  success = __kmp_affinity_create_cpuinfo_map(&line, &msg_id);
3560  if (!success && __kmp_affinity_verbose && msg_id != kmp_i18n_null) {
3561  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", __kmp_i18n_catgets(msg_id));
3562  }
3563  }
3564 #endif /* KMP_OS_LINUX */
3565 
3566 #if KMP_GROUP_AFFINITY
3567  if (!success && (__kmp_num_proc_groups > 1)) {
3568  success = __kmp_affinity_create_proc_group_map(&msg_id);
3569  if (!success && __kmp_affinity_verbose && msg_id != kmp_i18n_null) {
3570  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", __kmp_i18n_catgets(msg_id));
3571  }
3572  }
3573 #endif /* KMP_GROUP_AFFINITY */
3574 
3575  if (!success) {
3576  success = __kmp_affinity_create_flat_map(&msg_id);
3577  if (!success && __kmp_affinity_verbose && msg_id != kmp_i18n_null) {
3578  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", __kmp_i18n_catgets(msg_id));
3579  }
3580  KMP_ASSERT(success);
3581  }
3582  }
3583 
3584 // If the user has specified that a paricular topology discovery method is to be
3585 // used, then we abort if that method fails. The exception is group affinity,
3586 // which might have been implicitly set.
3587 #if KMP_USE_HWLOC
3588  else if (__kmp_affinity_top_method == affinity_top_method_hwloc) {
3589  KMP_ASSERT(__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC);
3590  success = __kmp_affinity_create_hwloc_map(&msg_id);
3591  if (!success) {
3592  KMP_ASSERT(msg_id != kmp_i18n_null);
3593  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
3594  }
3595  }
3596 #endif // KMP_USE_HWLOC
3597 
3598 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3599  else if (__kmp_affinity_top_method == affinity_top_method_x2apicid ||
3600  __kmp_affinity_top_method == affinity_top_method_x2apicid_1f) {
3601  success = __kmp_affinity_create_x2apicid_map(&msg_id);
3602  if (!success) {
3603  KMP_ASSERT(msg_id != kmp_i18n_null);
3604  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
3605  }
3606  } else if (__kmp_affinity_top_method == affinity_top_method_apicid) {
3607  success = __kmp_affinity_create_apicid_map(&msg_id);
3608  if (!success) {
3609  KMP_ASSERT(msg_id != kmp_i18n_null);
3610  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
3611  }
3612  }
3613 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3614 
3615  else if (__kmp_affinity_top_method == affinity_top_method_cpuinfo) {
3616  int line = 0;
3617  success = __kmp_affinity_create_cpuinfo_map(&line, &msg_id);
3618  if (!success) {
3619  KMP_ASSERT(msg_id != kmp_i18n_null);
3620  const char *filename = __kmp_cpuinfo_get_filename();
3621  if (line > 0) {
3622  KMP_FATAL(FileLineMsgExiting, filename, line,
3623  __kmp_i18n_catgets(msg_id));
3624  } else {
3625  KMP_FATAL(FileMsgExiting, filename, __kmp_i18n_catgets(msg_id));
3626  }
3627  }
3628  }
3629 
3630 #if KMP_GROUP_AFFINITY
3631  else if (__kmp_affinity_top_method == affinity_top_method_group) {
3632  success = __kmp_affinity_create_proc_group_map(&msg_id);
3633  KMP_ASSERT(success);
3634  if (!success) {
3635  KMP_ASSERT(msg_id != kmp_i18n_null);
3636  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
3637  }
3638  }
3639 #endif /* KMP_GROUP_AFFINITY */
3640 
3641  else if (__kmp_affinity_top_method == affinity_top_method_flat) {
3642  success = __kmp_affinity_create_flat_map(&msg_id);
3643  // should not fail
3644  KMP_ASSERT(success);
3645  }
3646 
3647  // Early exit if topology could not be created
3648  if (!__kmp_topology) {
3649  if (KMP_AFFINITY_CAPABLE() &&
3650  (__kmp_affinity_verbose ||
3651  (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none)))) {
3652  KMP_WARNING(ErrorInitializeAffinity);
3653  }
3654  if (nPackages > 0 && nCoresPerPkg > 0 && __kmp_nThreadsPerCore > 0 &&
3655  __kmp_ncores > 0) {
3656  __kmp_topology = kmp_topology_t::allocate(0, 0, NULL);
3657  __kmp_topology->canonicalize(nPackages, nCoresPerPkg,
3658  __kmp_nThreadsPerCore, __kmp_ncores);
3659  if (__kmp_affinity_verbose) {
3660  __kmp_topology->print("KMP_AFFINITY");
3661  }
3662  }
3663  __kmp_affinity_type = affinity_none;
3664  __kmp_create_affinity_none_places();
3665 #if KMP_USE_HIER_SCHED
3666  __kmp_dispatch_set_hierarchy_values();
3667 #endif
3668  KMP_AFFINITY_DISABLE();
3669  return;
3670  }
3671 
3672  // Canonicalize, print (if requested), apply KMP_HW_SUBSET, and
3673  // initialize other data structures which depend on the topology
3674  __kmp_topology->canonicalize();
3675  if (__kmp_affinity_verbose)
3676  __kmp_topology->print("KMP_AFFINITY");
3677  bool filtered = __kmp_topology->filter_hw_subset();
3678  if (filtered && __kmp_affinity_verbose)
3679  __kmp_topology->print("KMP_HW_SUBSET");
3680  machine_hierarchy.init(__kmp_topology->get_num_hw_threads());
3681  KMP_ASSERT(__kmp_avail_proc == __kmp_topology->get_num_hw_threads());
3682  // If KMP_AFFINITY=none, then only create the single "none" place
3683  // which is the process's initial affinity mask or the number of
3684  // hardware threads depending on respect,norespect
3685  if (__kmp_affinity_type == affinity_none) {
3686  __kmp_create_affinity_none_places();
3687 #if KMP_USE_HIER_SCHED
3688  __kmp_dispatch_set_hierarchy_values();
3689 #endif
3690  return;
3691  }
3692  int depth = __kmp_topology->get_depth();
3693 
3694  // Create the table of masks, indexed by thread Id.
3695  unsigned maxIndex;
3696  unsigned numUnique;
3697  kmp_affin_mask_t *osId2Mask = __kmp_create_masks(&maxIndex, &numUnique);
3698  if (__kmp_affinity_gran_levels == 0) {
3699  KMP_DEBUG_ASSERT((int)numUnique == __kmp_avail_proc);
3700  }
3701 
3702  switch (__kmp_affinity_type) {
3703 
3704  case affinity_explicit:
3705  KMP_DEBUG_ASSERT(__kmp_affinity_proclist != NULL);
3706  if (__kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) {
3707  __kmp_affinity_process_proclist(
3708  &__kmp_affinity_masks, &__kmp_affinity_num_masks,
3709  __kmp_affinity_proclist, osId2Mask, maxIndex);
3710  } else {
3711  __kmp_affinity_process_placelist(
3712  &__kmp_affinity_masks, &__kmp_affinity_num_masks,
3713  __kmp_affinity_proclist, osId2Mask, maxIndex);
3714  }
3715  if (__kmp_affinity_num_masks == 0) {
3716  if (__kmp_affinity_verbose ||
3717  (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) {
3718  KMP_WARNING(AffNoValidProcID);
3719  }
3720  __kmp_affinity_type = affinity_none;
3721  __kmp_create_affinity_none_places();
3722  return;
3723  }
3724  break;
3725 
3726  // The other affinity types rely on sorting the hardware threads according to
3727  // some permutation of the machine topology tree. Set __kmp_affinity_compact
3728  // and __kmp_affinity_offset appropriately, then jump to a common code
3729  // fragment to do the sort and create the array of affinity masks.
3730  case affinity_logical:
3731  __kmp_affinity_compact = 0;
3732  if (__kmp_affinity_offset) {
3733  __kmp_affinity_offset =
3734  __kmp_nThreadsPerCore * __kmp_affinity_offset % __kmp_avail_proc;
3735  }
3736  goto sortTopology;
3737 
3738  case affinity_physical:
3739  if (__kmp_nThreadsPerCore > 1) {
3740  __kmp_affinity_compact = 1;
3741  if (__kmp_affinity_compact >= depth) {
3742  __kmp_affinity_compact = 0;
3743  }
3744  } else {
3745  __kmp_affinity_compact = 0;
3746  }
3747  if (__kmp_affinity_offset) {
3748  __kmp_affinity_offset =
3749  __kmp_nThreadsPerCore * __kmp_affinity_offset % __kmp_avail_proc;
3750  }
3751  goto sortTopology;
3752 
3753  case affinity_scatter:
3754  if (__kmp_affinity_compact >= depth) {
3755  __kmp_affinity_compact = 0;
3756  } else {
3757  __kmp_affinity_compact = depth - 1 - __kmp_affinity_compact;
3758  }
3759  goto sortTopology;
3760 
3761  case affinity_compact:
3762  if (__kmp_affinity_compact >= depth) {
3763  __kmp_affinity_compact = depth - 1;
3764  }
3765  goto sortTopology;
3766 
3767  case affinity_balanced:
3768  if (depth <= 1) {
3769  if (__kmp_affinity_verbose || __kmp_affinity_warnings) {
3770  KMP_WARNING(AffBalancedNotAvail, "KMP_AFFINITY");
3771  }
3772  __kmp_affinity_type = affinity_none;
3773  __kmp_create_affinity_none_places();
3774  return;
3775  } else if (!__kmp_topology->is_uniform()) {
3776  // Save the depth for further usage
3777  __kmp_aff_depth = depth;
3778 
3779  int core_level =
3780  __kmp_affinity_find_core_level(__kmp_avail_proc, depth - 1);
3781  int ncores = __kmp_affinity_compute_ncores(__kmp_avail_proc, depth - 1,
3782  core_level);
3783  int maxprocpercore = __kmp_affinity_max_proc_per_core(
3784  __kmp_avail_proc, depth - 1, core_level);
3785 
3786  int nproc = ncores * maxprocpercore;
3787  if ((nproc < 2) || (nproc < __kmp_avail_proc)) {
3788  if (__kmp_affinity_verbose || __kmp_affinity_warnings) {
3789  KMP_WARNING(AffBalancedNotAvail, "KMP_AFFINITY");
3790  }
3791  __kmp_affinity_type = affinity_none;
3792  return;
3793  }
3794 
3795  procarr = (int *)__kmp_allocate(sizeof(int) * nproc);
3796  for (int i = 0; i < nproc; i++) {
3797  procarr[i] = -1;
3798  }
3799 
3800  int lastcore = -1;
3801  int inlastcore = 0;
3802  for (int i = 0; i < __kmp_avail_proc; i++) {
3803  int proc = __kmp_topology->at(i).os_id;
3804  int core = __kmp_affinity_find_core(i, depth - 1, core_level);
3805 
3806  if (core == lastcore) {
3807  inlastcore++;
3808  } else {
3809  inlastcore = 0;
3810  }
3811  lastcore = core;
3812 
3813  procarr[core * maxprocpercore + inlastcore] = proc;
3814  }
3815  }
3816  if (__kmp_affinity_compact >= depth) {
3817  __kmp_affinity_compact = depth - 1;
3818  }
3819 
3820  sortTopology:
3821  // Allocate the gtid->affinity mask table.
3822  if (__kmp_affinity_dups) {
3823  __kmp_affinity_num_masks = __kmp_avail_proc;
3824  } else {
3825  __kmp_affinity_num_masks = numUnique;
3826  }
3827 
3828  if ((__kmp_nested_proc_bind.bind_types[0] != proc_bind_intel) &&
3829  (__kmp_affinity_num_places > 0) &&
3830  ((unsigned)__kmp_affinity_num_places < __kmp_affinity_num_masks)) {
3831  __kmp_affinity_num_masks = __kmp_affinity_num_places;
3832  }
3833 
3834  KMP_CPU_ALLOC_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks);
3835 
3836  // Sort the topology table according to the current setting of
3837  // __kmp_affinity_compact, then fill out __kmp_affinity_masks.
3838  __kmp_topology->sort_compact();
3839  {
3840  int i;
3841  unsigned j;
3842  int num_hw_threads = __kmp_topology->get_num_hw_threads();
3843  for (i = 0, j = 0; i < num_hw_threads; i++) {
3844  if ((!__kmp_affinity_dups) && (!__kmp_topology->at(i).leader)) {
3845  continue;
3846  }
3847  int osId = __kmp_topology->at(i).os_id;
3848 
3849  kmp_affin_mask_t *src = KMP_CPU_INDEX(osId2Mask, osId);
3850  kmp_affin_mask_t *dest = KMP_CPU_INDEX(__kmp_affinity_masks, j);
3851  KMP_ASSERT(KMP_CPU_ISSET(osId, src));
3852  KMP_CPU_COPY(dest, src);
3853  if (++j >= __kmp_affinity_num_masks) {
3854  break;
3855  }
3856  }
3857  KMP_DEBUG_ASSERT(j == __kmp_affinity_num_masks);
3858  }
3859  // Sort the topology back using ids
3860  __kmp_topology->sort_ids();
3861  break;
3862 
3863  default:
3864  KMP_ASSERT2(0, "Unexpected affinity setting");
3865  }
3866 
3867  KMP_CPU_FREE_ARRAY(osId2Mask, maxIndex + 1);
3868 }
3869 
3870 void __kmp_affinity_initialize(void) {
3871  // Much of the code above was written assuming that if a machine was not
3872  // affinity capable, then __kmp_affinity_type == affinity_none. We now
3873  // explicitly represent this as __kmp_affinity_type == affinity_disabled.
3874  // There are too many checks for __kmp_affinity_type == affinity_none
3875  // in this code. Instead of trying to change them all, check if
3876  // __kmp_affinity_type == affinity_disabled, and if so, slam it with
3877  // affinity_none, call the real initialization routine, then restore
3878  // __kmp_affinity_type to affinity_disabled.
3879  int disabled = (__kmp_affinity_type == affinity_disabled);
3880  if (!KMP_AFFINITY_CAPABLE()) {
3881  KMP_ASSERT(disabled);
3882  }
3883  if (disabled) {
3884  __kmp_affinity_type = affinity_none;
3885  }
3886  __kmp_aux_affinity_initialize();
3887  if (disabled) {
3888  __kmp_affinity_type = affinity_disabled;
3889  }
3890 }
3891 
3892 void __kmp_affinity_uninitialize(void) {
3893  if (__kmp_affinity_masks != NULL) {
3894  KMP_CPU_FREE_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks);
3895  __kmp_affinity_masks = NULL;
3896  }
3897  if (__kmp_affin_fullMask != NULL) {
3898  KMP_CPU_FREE(__kmp_affin_fullMask);
3899  __kmp_affin_fullMask = NULL;
3900  }
3901  __kmp_affinity_num_masks = 0;
3902  __kmp_affinity_type = affinity_default;
3903  __kmp_affinity_num_places = 0;
3904  if (__kmp_affinity_proclist != NULL) {
3905  __kmp_free(__kmp_affinity_proclist);
3906  __kmp_affinity_proclist = NULL;
3907  }
3908  if (procarr != NULL) {
3909  __kmp_free(procarr);
3910  procarr = NULL;
3911  }
3912 #if KMP_USE_HWLOC
3913  if (__kmp_hwloc_topology != NULL) {
3914  hwloc_topology_destroy(__kmp_hwloc_topology);
3915  __kmp_hwloc_topology = NULL;
3916  }
3917 #endif
3918  if (__kmp_hw_subset) {
3919  kmp_hw_subset_t::deallocate(__kmp_hw_subset);
3920  __kmp_hw_subset = nullptr;
3921  }
3922  if (__kmp_topology) {
3923  kmp_topology_t::deallocate(__kmp_topology);
3924  __kmp_topology = nullptr;
3925  }
3926  KMPAffinity::destroy_api();
3927 }
3928 
3929 void __kmp_affinity_set_init_mask(int gtid, int isa_root) {
3930  if (!KMP_AFFINITY_CAPABLE()) {
3931  return;
3932  }
3933 
3934  kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
3935  if (th->th.th_affin_mask == NULL) {
3936  KMP_CPU_ALLOC(th->th.th_affin_mask);
3937  } else {
3938  KMP_CPU_ZERO(th->th.th_affin_mask);
3939  }
3940 
3941  // Copy the thread mask to the kmp_info_t structure. If
3942  // __kmp_affinity_type == affinity_none, copy the "full" mask, i.e. one that
3943  // has all of the OS proc ids set, or if __kmp_affinity_respect_mask is set,
3944  // then the full mask is the same as the mask of the initialization thread.
3945  kmp_affin_mask_t *mask;
3946  int i;
3947 
3948  if (KMP_AFFINITY_NON_PROC_BIND) {
3949  if ((__kmp_affinity_type == affinity_none) ||
3950  (__kmp_affinity_type == affinity_balanced) ||
3951  KMP_HIDDEN_HELPER_THREAD(gtid)) {
3952 #if KMP_GROUP_AFFINITY
3953  if (__kmp_num_proc_groups > 1) {
3954  return;
3955  }
3956 #endif
3957  KMP_ASSERT(__kmp_affin_fullMask != NULL);
3958  i = 0;
3959  mask = __kmp_affin_fullMask;
3960  } else {
3961  int mask_idx = __kmp_adjust_gtid_for_hidden_helpers(gtid);
3962  KMP_DEBUG_ASSERT(__kmp_affinity_num_masks > 0);
3963  i = (mask_idx + __kmp_affinity_offset) % __kmp_affinity_num_masks;
3964  mask = KMP_CPU_INDEX(__kmp_affinity_masks, i);
3965  }
3966  } else {
3967  if ((!isa_root) || KMP_HIDDEN_HELPER_THREAD(gtid) ||
3968  (__kmp_nested_proc_bind.bind_types[0] == proc_bind_false)) {
3969 #if KMP_GROUP_AFFINITY
3970  if (__kmp_num_proc_groups > 1) {
3971  return;
3972  }
3973 #endif
3974  KMP_ASSERT(__kmp_affin_fullMask != NULL);
3975  i = KMP_PLACE_ALL;
3976  mask = __kmp_affin_fullMask;
3977  } else {
3978  // int i = some hash function or just a counter that doesn't
3979  // always start at 0. Use adjusted gtid for now.
3980  int mask_idx = __kmp_adjust_gtid_for_hidden_helpers(gtid);
3981  KMP_DEBUG_ASSERT(__kmp_affinity_num_masks > 0);
3982  i = (mask_idx + __kmp_affinity_offset) % __kmp_affinity_num_masks;
3983  mask = KMP_CPU_INDEX(__kmp_affinity_masks, i);
3984  }
3985  }
3986 
3987  th->th.th_current_place = i;
3988  if (isa_root || KMP_HIDDEN_HELPER_THREAD(gtid)) {
3989  th->th.th_new_place = i;
3990  th->th.th_first_place = 0;
3991  th->th.th_last_place = __kmp_affinity_num_masks - 1;
3992  } else if (KMP_AFFINITY_NON_PROC_BIND) {
3993  // When using a Non-OMP_PROC_BIND affinity method,
3994  // set all threads' place-partition-var to the entire place list
3995  th->th.th_first_place = 0;
3996  th->th.th_last_place = __kmp_affinity_num_masks - 1;
3997  }
3998 
3999  if (i == KMP_PLACE_ALL) {
4000  KA_TRACE(100, ("__kmp_affinity_set_init_mask: binding T#%d to all places\n",
4001  gtid));
4002  } else {
4003  KA_TRACE(100, ("__kmp_affinity_set_init_mask: binding T#%d to place %d\n",
4004  gtid, i));
4005  }
4006 
4007  KMP_CPU_COPY(th->th.th_affin_mask, mask);
4008 
4009  if (__kmp_affinity_verbose && !KMP_HIDDEN_HELPER_THREAD(gtid)
4010  /* to avoid duplicate printing (will be correctly printed on barrier) */
4011  && (__kmp_affinity_type == affinity_none ||
4012  (i != KMP_PLACE_ALL && __kmp_affinity_type != affinity_balanced))) {
4013  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4014  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4015  th->th.th_affin_mask);
4016  KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY", (kmp_int32)getpid(),
4017  __kmp_gettid(), gtid, buf);
4018  }
4019 
4020 #if KMP_DEBUG
4021  // Hidden helper thread affinity only printed for debug builds
4022  if (__kmp_affinity_verbose && KMP_HIDDEN_HELPER_THREAD(gtid)) {
4023  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4024  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4025  th->th.th_affin_mask);
4026  KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY (hidden helper thread)",
4027  (kmp_int32)getpid(), __kmp_gettid(), gtid, buf);
4028  }
4029 #endif
4030 
4031 #if KMP_OS_WINDOWS
4032  // On Windows* OS, the process affinity mask might have changed. If the user
4033  // didn't request affinity and this call fails, just continue silently.
4034  // See CQ171393.
4035  if (__kmp_affinity_type == affinity_none) {
4036  __kmp_set_system_affinity(th->th.th_affin_mask, FALSE);
4037  } else
4038 #endif
4039  __kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
4040 }
4041 
4042 void __kmp_affinity_set_place(int gtid) {
4043  if (!KMP_AFFINITY_CAPABLE()) {
4044  return;
4045  }
4046 
4047  kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
4048 
4049  KA_TRACE(100, ("__kmp_affinity_set_place: binding T#%d to place %d (current "
4050  "place = %d)\n",
4051  gtid, th->th.th_new_place, th->th.th_current_place));
4052 
4053  // Check that the new place is within this thread's partition.
4054  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
4055  KMP_ASSERT(th->th.th_new_place >= 0);
4056  KMP_ASSERT((unsigned)th->th.th_new_place <= __kmp_affinity_num_masks);
4057  if (th->th.th_first_place <= th->th.th_last_place) {
4058  KMP_ASSERT((th->th.th_new_place >= th->th.th_first_place) &&
4059  (th->th.th_new_place <= th->th.th_last_place));
4060  } else {
4061  KMP_ASSERT((th->th.th_new_place <= th->th.th_first_place) ||
4062  (th->th.th_new_place >= th->th.th_last_place));
4063  }
4064 
4065  // Copy the thread mask to the kmp_info_t structure,
4066  // and set this thread's affinity.
4067  kmp_affin_mask_t *mask =
4068  KMP_CPU_INDEX(__kmp_affinity_masks, th->th.th_new_place);
4069  KMP_CPU_COPY(th->th.th_affin_mask, mask);
4070  th->th.th_current_place = th->th.th_new_place;
4071 
4072  if (__kmp_affinity_verbose) {
4073  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4074  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4075  th->th.th_affin_mask);
4076  KMP_INFORM(BoundToOSProcSet, "OMP_PROC_BIND", (kmp_int32)getpid(),
4077  __kmp_gettid(), gtid, buf);
4078  }
4079  __kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
4080 }
4081 
4082 int __kmp_aux_set_affinity(void **mask) {
4083  int gtid;
4084  kmp_info_t *th;
4085  int retval;
4086 
4087  if (!KMP_AFFINITY_CAPABLE()) {
4088  return -1;
4089  }
4090 
4091  gtid = __kmp_entry_gtid();
4092  KA_TRACE(
4093  1000, (""); {
4094  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4095  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4096  (kmp_affin_mask_t *)(*mask));
4097  __kmp_debug_printf(
4098  "kmp_set_affinity: setting affinity mask for thread %d = %s\n",
4099  gtid, buf);
4100  });
4101 
4102  if (__kmp_env_consistency_check) {
4103  if ((mask == NULL) || (*mask == NULL)) {
4104  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4105  } else {
4106  unsigned proc;
4107  int num_procs = 0;
4108 
4109  KMP_CPU_SET_ITERATE(proc, ((kmp_affin_mask_t *)(*mask))) {
4110  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
4111  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4112  }
4113  if (!KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask))) {
4114  continue;
4115  }
4116  num_procs++;
4117  }
4118  if (num_procs == 0) {
4119  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4120  }
4121 
4122 #if KMP_GROUP_AFFINITY
4123  if (__kmp_get_proc_group((kmp_affin_mask_t *)(*mask)) < 0) {
4124  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4125  }
4126 #endif /* KMP_GROUP_AFFINITY */
4127  }
4128  }
4129 
4130  th = __kmp_threads[gtid];
4131  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
4132  retval = __kmp_set_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
4133  if (retval == 0) {
4134  KMP_CPU_COPY(th->th.th_affin_mask, (kmp_affin_mask_t *)(*mask));
4135  }
4136 
4137  th->th.th_current_place = KMP_PLACE_UNDEFINED;
4138  th->th.th_new_place = KMP_PLACE_UNDEFINED;
4139  th->th.th_first_place = 0;
4140  th->th.th_last_place = __kmp_affinity_num_masks - 1;
4141 
4142  // Turn off 4.0 affinity for the current tread at this parallel level.
4143  th->th.th_current_task->td_icvs.proc_bind = proc_bind_false;
4144 
4145  return retval;
4146 }
4147 
4148 int __kmp_aux_get_affinity(void **mask) {
4149  int gtid;
4150  int retval;
4151 #if KMP_OS_WINDOWS || KMP_DEBUG
4152  kmp_info_t *th;
4153 #endif
4154  if (!KMP_AFFINITY_CAPABLE()) {
4155  return -1;
4156  }
4157 
4158  gtid = __kmp_entry_gtid();
4159 #if KMP_OS_WINDOWS || KMP_DEBUG
4160  th = __kmp_threads[gtid];
4161 #else
4162  (void)gtid; // unused variable
4163 #endif
4164  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
4165 
4166  KA_TRACE(
4167  1000, (""); {
4168  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4169  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4170  th->th.th_affin_mask);
4171  __kmp_printf(
4172  "kmp_get_affinity: stored affinity mask for thread %d = %s\n", gtid,
4173  buf);
4174  });
4175 
4176  if (__kmp_env_consistency_check) {
4177  if ((mask == NULL) || (*mask == NULL)) {
4178  KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity");
4179  }
4180  }
4181 
4182 #if !KMP_OS_WINDOWS
4183 
4184  retval = __kmp_get_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
4185  KA_TRACE(
4186  1000, (""); {
4187  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4188  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4189  (kmp_affin_mask_t *)(*mask));
4190  __kmp_printf(
4191  "kmp_get_affinity: system affinity mask for thread %d = %s\n", gtid,
4192  buf);
4193  });
4194  return retval;
4195 
4196 #else
4197  (void)retval;
4198 
4199  KMP_CPU_COPY((kmp_affin_mask_t *)(*mask), th->th.th_affin_mask);
4200  return 0;
4201 
4202 #endif /* KMP_OS_WINDOWS */
4203 }
4204 
4205 int __kmp_aux_get_affinity_max_proc() {
4206  if (!KMP_AFFINITY_CAPABLE()) {
4207  return 0;
4208  }
4209 #if KMP_GROUP_AFFINITY
4210  if (__kmp_num_proc_groups > 1) {
4211  return (int)(__kmp_num_proc_groups * sizeof(DWORD_PTR) * CHAR_BIT);
4212  }
4213 #endif
4214  return __kmp_xproc;
4215 }
4216 
4217 int __kmp_aux_set_affinity_mask_proc(int proc, void **mask) {
4218  if (!KMP_AFFINITY_CAPABLE()) {
4219  return -1;
4220  }
4221 
4222  KA_TRACE(
4223  1000, (""); {
4224  int gtid = __kmp_entry_gtid();
4225  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4226  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4227  (kmp_affin_mask_t *)(*mask));
4228  __kmp_debug_printf("kmp_set_affinity_mask_proc: setting proc %d in "
4229  "affinity mask for thread %d = %s\n",
4230  proc, gtid, buf);
4231  });
4232 
4233  if (__kmp_env_consistency_check) {
4234  if ((mask == NULL) || (*mask == NULL)) {
4235  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity_mask_proc");
4236  }
4237  }
4238 
4239  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
4240  return -1;
4241  }
4242  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
4243  return -2;
4244  }
4245 
4246  KMP_CPU_SET(proc, (kmp_affin_mask_t *)(*mask));
4247  return 0;
4248 }
4249 
4250 int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask) {
4251  if (!KMP_AFFINITY_CAPABLE()) {
4252  return -1;
4253  }
4254 
4255  KA_TRACE(
4256  1000, (""); {
4257  int gtid = __kmp_entry_gtid();
4258  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4259  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4260  (kmp_affin_mask_t *)(*mask));
4261  __kmp_debug_printf("kmp_unset_affinity_mask_proc: unsetting proc %d in "
4262  "affinity mask for thread %d = %s\n",
4263  proc, gtid, buf);
4264  });
4265 
4266  if (__kmp_env_consistency_check) {
4267  if ((mask == NULL) || (*mask == NULL)) {
4268  KMP_FATAL(AffinityInvalidMask, "kmp_unset_affinity_mask_proc");
4269  }
4270  }
4271 
4272  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
4273  return -1;
4274  }
4275  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
4276  return -2;
4277  }
4278 
4279  KMP_CPU_CLR(proc, (kmp_affin_mask_t *)(*mask));
4280  return 0;
4281 }
4282 
4283 int __kmp_aux_get_affinity_mask_proc(int proc, void **mask) {
4284  if (!KMP_AFFINITY_CAPABLE()) {
4285  return -1;
4286  }
4287 
4288  KA_TRACE(
4289  1000, (""); {
4290  int gtid = __kmp_entry_gtid();
4291  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4292  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4293  (kmp_affin_mask_t *)(*mask));
4294  __kmp_debug_printf("kmp_get_affinity_mask_proc: getting proc %d in "
4295  "affinity mask for thread %d = %s\n",
4296  proc, gtid, buf);
4297  });
4298 
4299  if (__kmp_env_consistency_check) {
4300  if ((mask == NULL) || (*mask == NULL)) {
4301  KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity_mask_proc");
4302  }
4303  }
4304 
4305  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
4306  return -1;
4307  }
4308  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
4309  return 0;
4310  }
4311 
4312  return KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask));
4313 }
4314 
4315 // Dynamic affinity settings - Affinity balanced
4316 void __kmp_balanced_affinity(kmp_info_t *th, int nthreads) {
4317  KMP_DEBUG_ASSERT(th);
4318  bool fine_gran = true;
4319  int tid = th->th.th_info.ds.ds_tid;
4320 
4321  // Do not perform balanced affinity for the hidden helper threads
4322  if (KMP_HIDDEN_HELPER_THREAD(__kmp_gtid_from_thread(th)))
4323  return;
4324 
4325  switch (__kmp_affinity_gran) {
4326  case KMP_HW_THREAD:
4327  break;
4328  case KMP_HW_CORE:
4329  if (__kmp_nThreadsPerCore > 1) {
4330  fine_gran = false;
4331  }
4332  break;
4333  case KMP_HW_SOCKET:
4334  if (nCoresPerPkg > 1) {
4335  fine_gran = false;
4336  }
4337  break;
4338  default:
4339  fine_gran = false;
4340  }
4341 
4342  if (__kmp_topology->is_uniform()) {
4343  int coreID;
4344  int threadID;
4345  // Number of hyper threads per core in HT machine
4346  int __kmp_nth_per_core = __kmp_avail_proc / __kmp_ncores;
4347  // Number of cores
4348  int ncores = __kmp_ncores;
4349  if ((nPackages > 1) && (__kmp_nth_per_core <= 1)) {
4350  __kmp_nth_per_core = __kmp_avail_proc / nPackages;
4351  ncores = nPackages;
4352  }
4353  // How many threads will be bound to each core
4354  int chunk = nthreads / ncores;
4355  // How many cores will have an additional thread bound to it - "big cores"
4356  int big_cores = nthreads % ncores;
4357  // Number of threads on the big cores
4358  int big_nth = (chunk + 1) * big_cores;
4359  if (tid < big_nth) {
4360  coreID = tid / (chunk + 1);
4361  threadID = (tid % (chunk + 1)) % __kmp_nth_per_core;
4362  } else { // tid >= big_nth
4363  coreID = (tid - big_cores) / chunk;
4364  threadID = ((tid - big_cores) % chunk) % __kmp_nth_per_core;
4365  }
4366  KMP_DEBUG_ASSERT2(KMP_AFFINITY_CAPABLE(),
4367  "Illegal set affinity operation when not capable");
4368 
4369  kmp_affin_mask_t *mask = th->th.th_affin_mask;
4370  KMP_CPU_ZERO(mask);
4371 
4372  if (fine_gran) {
4373  int osID =
4374  __kmp_topology->at(coreID * __kmp_nth_per_core + threadID).os_id;
4375  KMP_CPU_SET(osID, mask);
4376  } else {
4377  for (int i = 0; i < __kmp_nth_per_core; i++) {
4378  int osID;
4379  osID = __kmp_topology->at(coreID * __kmp_nth_per_core + i).os_id;
4380  KMP_CPU_SET(osID, mask);
4381  }
4382  }
4383  if (__kmp_affinity_verbose) {
4384  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4385  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
4386  KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY", (kmp_int32)getpid(),
4387  __kmp_gettid(), tid, buf);
4388  }
4389  __kmp_set_system_affinity(mask, TRUE);
4390  } else { // Non-uniform topology
4391 
4392  kmp_affin_mask_t *mask = th->th.th_affin_mask;
4393  KMP_CPU_ZERO(mask);
4394 
4395  int core_level =
4396  __kmp_affinity_find_core_level(__kmp_avail_proc, __kmp_aff_depth - 1);
4397  int ncores = __kmp_affinity_compute_ncores(__kmp_avail_proc,
4398  __kmp_aff_depth - 1, core_level);
4399  int nth_per_core = __kmp_affinity_max_proc_per_core(
4400  __kmp_avail_proc, __kmp_aff_depth - 1, core_level);
4401 
4402  // For performance gain consider the special case nthreads ==
4403  // __kmp_avail_proc
4404  if (nthreads == __kmp_avail_proc) {
4405  if (fine_gran) {
4406  int osID = __kmp_topology->at(tid).os_id;
4407  KMP_CPU_SET(osID, mask);
4408  } else {
4409  int core =
4410  __kmp_affinity_find_core(tid, __kmp_aff_depth - 1, core_level);
4411  for (int i = 0; i < __kmp_avail_proc; i++) {
4412  int osID = __kmp_topology->at(i).os_id;
4413  if (__kmp_affinity_find_core(i, __kmp_aff_depth - 1, core_level) ==
4414  core) {
4415  KMP_CPU_SET(osID, mask);
4416  }
4417  }
4418  }
4419  } else if (nthreads <= ncores) {
4420 
4421  int core = 0;
4422  for (int i = 0; i < ncores; i++) {
4423  // Check if this core from procarr[] is in the mask
4424  int in_mask = 0;
4425  for (int j = 0; j < nth_per_core; j++) {
4426  if (procarr[i * nth_per_core + j] != -1) {
4427  in_mask = 1;
4428  break;
4429  }
4430  }
4431  if (in_mask) {
4432  if (tid == core) {
4433  for (int j = 0; j < nth_per_core; j++) {
4434  int osID = procarr[i * nth_per_core + j];
4435  if (osID != -1) {
4436  KMP_CPU_SET(osID, mask);
4437  // For fine granularity it is enough to set the first available
4438  // osID for this core
4439  if (fine_gran) {
4440  break;
4441  }
4442  }
4443  }
4444  break;
4445  } else {
4446  core++;
4447  }
4448  }
4449  }
4450  } else { // nthreads > ncores
4451  // Array to save the number of processors at each core
4452  int *nproc_at_core = (int *)KMP_ALLOCA(sizeof(int) * ncores);
4453  // Array to save the number of cores with "x" available processors;
4454  int *ncores_with_x_procs =
4455  (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
4456  // Array to save the number of cores with # procs from x to nth_per_core
4457  int *ncores_with_x_to_max_procs =
4458  (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
4459 
4460  for (int i = 0; i <= nth_per_core; i++) {
4461  ncores_with_x_procs[i] = 0;
4462  ncores_with_x_to_max_procs[i] = 0;
4463  }
4464 
4465  for (int i = 0; i < ncores; i++) {
4466  int cnt = 0;
4467  for (int j = 0; j < nth_per_core; j++) {
4468  if (procarr[i * nth_per_core + j] != -1) {
4469  cnt++;
4470  }
4471  }
4472  nproc_at_core[i] = cnt;
4473  ncores_with_x_procs[cnt]++;
4474  }
4475 
4476  for (int i = 0; i <= nth_per_core; i++) {
4477  for (int j = i; j <= nth_per_core; j++) {
4478  ncores_with_x_to_max_procs[i] += ncores_with_x_procs[j];
4479  }
4480  }
4481 
4482  // Max number of processors
4483  int nproc = nth_per_core * ncores;
4484  // An array to keep number of threads per each context
4485  int *newarr = (int *)__kmp_allocate(sizeof(int) * nproc);
4486  for (int i = 0; i < nproc; i++) {
4487  newarr[i] = 0;
4488  }
4489 
4490  int nth = nthreads;
4491  int flag = 0;
4492  while (nth > 0) {
4493  for (int j = 1; j <= nth_per_core; j++) {
4494  int cnt = ncores_with_x_to_max_procs[j];
4495  for (int i = 0; i < ncores; i++) {
4496  // Skip the core with 0 processors
4497  if (nproc_at_core[i] == 0) {
4498  continue;
4499  }
4500  for (int k = 0; k < nth_per_core; k++) {
4501  if (procarr[i * nth_per_core + k] != -1) {
4502  if (newarr[i * nth_per_core + k] == 0) {
4503  newarr[i * nth_per_core + k] = 1;
4504  cnt--;
4505  nth--;
4506  break;
4507  } else {
4508  if (flag != 0) {
4509  newarr[i * nth_per_core + k]++;
4510  cnt--;
4511  nth--;
4512  break;
4513  }
4514  }
4515  }
4516  }
4517  if (cnt == 0 || nth == 0) {
4518  break;
4519  }
4520  }
4521  if (nth == 0) {
4522  break;
4523  }
4524  }
4525  flag = 1;
4526  }
4527  int sum = 0;
4528  for (int i = 0; i < nproc; i++) {
4529  sum += newarr[i];
4530  if (sum > tid) {
4531  if (fine_gran) {
4532  int osID = procarr[i];
4533  KMP_CPU_SET(osID, mask);
4534  } else {
4535  int coreID = i / nth_per_core;
4536  for (int ii = 0; ii < nth_per_core; ii++) {
4537  int osID = procarr[coreID * nth_per_core + ii];
4538  if (osID != -1) {
4539  KMP_CPU_SET(osID, mask);
4540  }
4541  }
4542  }
4543  break;
4544  }
4545  }
4546  __kmp_free(newarr);
4547  }
4548 
4549  if (__kmp_affinity_verbose) {
4550  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4551  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
4552  KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY", (kmp_int32)getpid(),
4553  __kmp_gettid(), tid, buf);
4554  }
4555  __kmp_set_system_affinity(mask, TRUE);
4556  }
4557 }
4558 
4559 #if KMP_OS_LINUX || KMP_OS_FREEBSD
4560 // We don't need this entry for Windows because
4561 // there is GetProcessAffinityMask() api
4562 //
4563 // The intended usage is indicated by these steps:
4564 // 1) The user gets the current affinity mask
4565 // 2) Then sets the affinity by calling this function
4566 // 3) Error check the return value
4567 // 4) Use non-OpenMP parallelization
4568 // 5) Reset the affinity to what was stored in step 1)
4569 #ifdef __cplusplus
4570 extern "C"
4571 #endif
4572  int
4573  kmp_set_thread_affinity_mask_initial()
4574 // the function returns 0 on success,
4575 // -1 if we cannot bind thread
4576 // >0 (errno) if an error happened during binding
4577 {
4578  int gtid = __kmp_get_gtid();
4579  if (gtid < 0) {
4580  // Do not touch non-omp threads
4581  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
4582  "non-omp thread, returning\n"));
4583  return -1;
4584  }
4585  if (!KMP_AFFINITY_CAPABLE() || !__kmp_init_middle) {
4586  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
4587  "affinity not initialized, returning\n"));
4588  return -1;
4589  }
4590  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
4591  "set full mask for thread %d\n",
4592  gtid));
4593  KMP_DEBUG_ASSERT(__kmp_affin_fullMask != NULL);
4594  return __kmp_set_system_affinity(__kmp_affin_fullMask, FALSE);
4595 }
4596 #endif
4597 
4598 #endif // KMP_AFFINITY_SUPPORTED