LLVM OpenMP* Runtime Library
 All Classes Functions Variables Typedefs Enumerations Enumerator Groups Pages
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 
23 // Store the real or imagined machine hierarchy here
24 static hierarchy_info machine_hierarchy;
25 
26 void __kmp_cleanup_hierarchy() { machine_hierarchy.fini(); }
27 
28 void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar) {
29  kmp_uint32 depth;
30  // The test below is true if affinity is available, but set to "none". Need to
31  // init on first use of hierarchical barrier.
32  if (TCR_1(machine_hierarchy.uninitialized))
33  machine_hierarchy.init(NULL, nproc);
34 
35  // Adjust the hierarchy in case num threads exceeds original
36  if (nproc > machine_hierarchy.base_num_threads)
37  machine_hierarchy.resize(nproc);
38 
39  depth = machine_hierarchy.depth;
40  KMP_DEBUG_ASSERT(depth > 0);
41 
42  thr_bar->depth = depth;
43  thr_bar->base_leaf_kids = (kmp_uint8)machine_hierarchy.numPerLevel[0] - 1;
44  thr_bar->skip_per_level = machine_hierarchy.skipPerLevel;
45 }
46 
47 #if KMP_AFFINITY_SUPPORTED
48 
49 bool KMPAffinity::picked_api = false;
50 
51 void *KMPAffinity::Mask::operator new(size_t n) { return __kmp_allocate(n); }
52 void *KMPAffinity::Mask::operator new[](size_t n) { return __kmp_allocate(n); }
53 void KMPAffinity::Mask::operator delete(void *p) { __kmp_free(p); }
54 void KMPAffinity::Mask::operator delete[](void *p) { __kmp_free(p); }
55 void *KMPAffinity::operator new(size_t n) { return __kmp_allocate(n); }
56 void KMPAffinity::operator delete(void *p) { __kmp_free(p); }
57 
58 void KMPAffinity::pick_api() {
59  KMPAffinity *affinity_dispatch;
60  if (picked_api)
61  return;
62 #if KMP_USE_HWLOC
63  // Only use Hwloc if affinity isn't explicitly disabled and
64  // user requests Hwloc topology method
65  if (__kmp_affinity_top_method == affinity_top_method_hwloc &&
66  __kmp_affinity_type != affinity_disabled) {
67  affinity_dispatch = new KMPHwlocAffinity();
68  } else
69 #endif
70  {
71  affinity_dispatch = new KMPNativeAffinity();
72  }
73  __kmp_affinity_dispatch = affinity_dispatch;
74  picked_api = true;
75 }
76 
77 void KMPAffinity::destroy_api() {
78  if (__kmp_affinity_dispatch != NULL) {
79  delete __kmp_affinity_dispatch;
80  __kmp_affinity_dispatch = NULL;
81  picked_api = false;
82  }
83 }
84 
85 #define KMP_ADVANCE_SCAN(scan) \
86  while (*scan != '\0') { \
87  scan++; \
88  }
89 
90 // Print the affinity mask to the character array in a pretty format.
91 // The format is a comma separated list of non-negative integers or integer
92 // ranges: e.g., 1,2,3-5,7,9-15
93 // The format can also be the string "{<empty>}" if no bits are set in mask
94 char *__kmp_affinity_print_mask(char *buf, int buf_len,
95  kmp_affin_mask_t *mask) {
96  int start = 0, finish = 0, previous = 0;
97  bool first_range;
98  KMP_ASSERT(buf);
99  KMP_ASSERT(buf_len >= 40);
100  KMP_ASSERT(mask);
101  char *scan = buf;
102  char *end = buf + buf_len - 1;
103 
104  // Check for empty set.
105  if (mask->begin() == mask->end()) {
106  KMP_SNPRINTF(scan, end - scan + 1, "{<empty>}");
107  KMP_ADVANCE_SCAN(scan);
108  KMP_ASSERT(scan <= end);
109  return buf;
110  }
111 
112  first_range = true;
113  start = mask->begin();
114  while (1) {
115  // Find next range
116  // [start, previous] is inclusive range of contiguous bits in mask
117  for (finish = mask->next(start), previous = start;
118  finish == previous + 1 && finish != mask->end();
119  finish = mask->next(finish)) {
120  previous = finish;
121  }
122 
123  // The first range does not need a comma printed before it, but the rest
124  // of the ranges do need a comma beforehand
125  if (!first_range) {
126  KMP_SNPRINTF(scan, end - scan + 1, "%s", ",");
127  KMP_ADVANCE_SCAN(scan);
128  } else {
129  first_range = false;
130  }
131  // Range with three or more contiguous bits in the affinity mask
132  if (previous - start > 1) {
133  KMP_SNPRINTF(scan, end - scan + 1, "%d-%d", static_cast<int>(start),
134  static_cast<int>(previous));
135  } else {
136  // Range with one or two contiguous bits in the affinity mask
137  KMP_SNPRINTF(scan, end - scan + 1, "%d", static_cast<int>(start));
138  KMP_ADVANCE_SCAN(scan);
139  if (previous - start > 0) {
140  KMP_SNPRINTF(scan, end - scan + 1, ",%d", static_cast<int>(previous));
141  }
142  }
143  KMP_ADVANCE_SCAN(scan);
144  // Start over with new start point
145  start = finish;
146  if (start == mask->end())
147  break;
148  // Check for overflow
149  if (end - scan < 2)
150  break;
151  }
152 
153  // Check for overflow
154  KMP_ASSERT(scan <= end);
155  return buf;
156 }
157 #undef KMP_ADVANCE_SCAN
158 
159 // Print the affinity mask to the string buffer object in a pretty format
160 // The format is a comma separated list of non-negative integers or integer
161 // ranges: e.g., 1,2,3-5,7,9-15
162 // The format can also be the string "{<empty>}" if no bits are set in mask
163 kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
164  kmp_affin_mask_t *mask) {
165  int start = 0, finish = 0, previous = 0;
166  bool first_range;
167  KMP_ASSERT(buf);
168  KMP_ASSERT(mask);
169 
170  __kmp_str_buf_clear(buf);
171 
172  // Check for empty set.
173  if (mask->begin() == mask->end()) {
174  __kmp_str_buf_print(buf, "%s", "{<empty>}");
175  return buf;
176  }
177 
178  first_range = true;
179  start = mask->begin();
180  while (1) {
181  // Find next range
182  // [start, previous] is inclusive range of contiguous bits in mask
183  for (finish = mask->next(start), previous = start;
184  finish == previous + 1 && finish != mask->end();
185  finish = mask->next(finish)) {
186  previous = finish;
187  }
188 
189  // The first range does not need a comma printed before it, but the rest
190  // of the ranges do need a comma beforehand
191  if (!first_range) {
192  __kmp_str_buf_print(buf, "%s", ",");
193  } else {
194  first_range = false;
195  }
196  // Range with three or more contiguous bits in the affinity mask
197  if (previous - start > 1) {
198  __kmp_str_buf_print(buf, "%d-%d", static_cast<int>(start),
199  static_cast<int>(previous));
200  } else {
201  // Range with one or two contiguous bits in the affinity mask
202  __kmp_str_buf_print(buf, "%d", static_cast<int>(start));
203  if (previous - start > 0) {
204  __kmp_str_buf_print(buf, ",%d", static_cast<int>(previous));
205  }
206  }
207  // Start over with new start point
208  start = finish;
209  if (start == mask->end())
210  break;
211  }
212  return buf;
213 }
214 
215 void __kmp_affinity_entire_machine_mask(kmp_affin_mask_t *mask) {
216  KMP_CPU_ZERO(mask);
217 
218 #if KMP_GROUP_AFFINITY
219 
220  if (__kmp_num_proc_groups > 1) {
221  int group;
222  KMP_DEBUG_ASSERT(__kmp_GetActiveProcessorCount != NULL);
223  for (group = 0; group < __kmp_num_proc_groups; group++) {
224  int i;
225  int num = __kmp_GetActiveProcessorCount(group);
226  for (i = 0; i < num; i++) {
227  KMP_CPU_SET(i + group * (CHAR_BIT * sizeof(DWORD_PTR)), mask);
228  }
229  }
230  } else
231 
232 #endif /* KMP_GROUP_AFFINITY */
233 
234  {
235  int proc;
236  for (proc = 0; proc < __kmp_xproc; proc++) {
237  KMP_CPU_SET(proc, mask);
238  }
239  }
240 }
241 
242 // When sorting by labels, __kmp_affinity_assign_child_nums() must first be
243 // called to renumber the labels from [0..n] and place them into the child_num
244 // vector of the address object. This is done in case the labels used for
245 // the children at one node of the hierarchy differ from those used for
246 // another node at the same level. Example: suppose the machine has 2 nodes
247 // with 2 packages each. The first node contains packages 601 and 602, and
248 // second node contains packages 603 and 604. If we try to sort the table
249 // for "scatter" affinity, the table will still be sorted 601, 602, 603, 604
250 // because we are paying attention to the labels themselves, not the ordinal
251 // child numbers. By using the child numbers in the sort, the result is
252 // {0,0}=601, {0,1}=603, {1,0}=602, {1,1}=604.
253 static void __kmp_affinity_assign_child_nums(AddrUnsPair *address2os,
254  int numAddrs) {
255  KMP_DEBUG_ASSERT(numAddrs > 0);
256  int depth = address2os->first.depth;
257  unsigned *counts = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
258  unsigned *lastLabel = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
259  int labCt;
260  for (labCt = 0; labCt < depth; labCt++) {
261  address2os[0].first.childNums[labCt] = counts[labCt] = 0;
262  lastLabel[labCt] = address2os[0].first.labels[labCt];
263  }
264  int i;
265  for (i = 1; i < numAddrs; i++) {
266  for (labCt = 0; labCt < depth; labCt++) {
267  if (address2os[i].first.labels[labCt] != lastLabel[labCt]) {
268  int labCt2;
269  for (labCt2 = labCt + 1; labCt2 < depth; labCt2++) {
270  counts[labCt2] = 0;
271  lastLabel[labCt2] = address2os[i].first.labels[labCt2];
272  }
273  counts[labCt]++;
274  lastLabel[labCt] = address2os[i].first.labels[labCt];
275  break;
276  }
277  }
278  for (labCt = 0; labCt < depth; labCt++) {
279  address2os[i].first.childNums[labCt] = counts[labCt];
280  }
281  for (; labCt < (int)Address::maxDepth; labCt++) {
282  address2os[i].first.childNums[labCt] = 0;
283  }
284  }
285  __kmp_free(lastLabel);
286  __kmp_free(counts);
287 }
288 
289 // All of the __kmp_affinity_create_*_map() routines should set
290 // __kmp_affinity_masks to a vector of affinity mask objects of length
291 // __kmp_affinity_num_masks, if __kmp_affinity_type != affinity_none, and return
292 // the number of levels in the machine topology tree (zero if
293 // __kmp_affinity_type == affinity_none).
294 //
295 // All of the __kmp_affinity_create_*_map() routines should set
296 // *__kmp_affin_fullMask to the affinity mask for the initialization thread.
297 // They need to save and restore the mask, and it could be needed later, so
298 // saving it is just an optimization to avoid calling kmp_get_system_affinity()
299 // again.
300 kmp_affin_mask_t *__kmp_affin_fullMask = NULL;
301 
302 static int nCoresPerPkg, nPackages;
303 static int __kmp_nThreadsPerCore;
304 #ifndef KMP_DFLT_NTH_CORES
305 static int __kmp_ncores;
306 #endif
307 static int *__kmp_pu_os_idx = NULL;
308 
309 // __kmp_affinity_uniform_topology() doesn't work when called from
310 // places which support arbitrarily many levels in the machine topology
311 // map, i.e. the non-default cases in __kmp_affinity_create_cpuinfo_map()
312 // __kmp_affinity_create_x2apicid_map().
313 inline static bool __kmp_affinity_uniform_topology() {
314  return __kmp_avail_proc == (__kmp_nThreadsPerCore * nCoresPerPkg * nPackages);
315 }
316 
317 // Print out the detailed machine topology map, i.e. the physical locations
318 // of each OS proc.
319 static void __kmp_affinity_print_topology(AddrUnsPair *address2os, int len,
320  int depth, int pkgLevel,
321  int coreLevel, int threadLevel) {
322  int proc;
323 
324  KMP_INFORM(OSProcToPhysicalThreadMap, "KMP_AFFINITY");
325  for (proc = 0; proc < len; proc++) {
326  int level;
327  kmp_str_buf_t buf;
328  __kmp_str_buf_init(&buf);
329  for (level = 0; level < depth; level++) {
330  if (level == threadLevel) {
331  __kmp_str_buf_print(&buf, "%s ", KMP_I18N_STR(Thread));
332  } else if (level == coreLevel) {
333  __kmp_str_buf_print(&buf, "%s ", KMP_I18N_STR(Core));
334  } else if (level == pkgLevel) {
335  __kmp_str_buf_print(&buf, "%s ", KMP_I18N_STR(Package));
336  } else if (level > pkgLevel) {
337  __kmp_str_buf_print(&buf, "%s_%d ", KMP_I18N_STR(Node),
338  level - pkgLevel - 1);
339  } else {
340  __kmp_str_buf_print(&buf, "L%d ", level);
341  }
342  __kmp_str_buf_print(&buf, "%d ", address2os[proc].first.labels[level]);
343  }
344  KMP_INFORM(OSProcMapToPack, "KMP_AFFINITY", address2os[proc].second,
345  buf.str);
346  __kmp_str_buf_free(&buf);
347  }
348 }
349 
350 #if KMP_USE_HWLOC
351 
352 static void __kmp_affinity_print_hwloc_tp(AddrUnsPair *addrP, int len,
353  int depth, int *levels) {
354  int proc;
355  kmp_str_buf_t buf;
356  __kmp_str_buf_init(&buf);
357  KMP_INFORM(OSProcToPhysicalThreadMap, "KMP_AFFINITY");
358  for (proc = 0; proc < len; proc++) {
359  __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Package),
360  addrP[proc].first.labels[0]);
361  if (depth > 1) {
362  int level = 1; // iterate over levels
363  int label = 1; // iterate over labels
364  if (__kmp_numa_detected)
365  // node level follows package
366  if (levels[level++] > 0)
367  __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Node),
368  addrP[proc].first.labels[label++]);
369  if (__kmp_tile_depth > 0)
370  // tile level follows node if any, or package
371  if (levels[level++] > 0)
372  __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Tile),
373  addrP[proc].first.labels[label++]);
374  if (levels[level++] > 0)
375  // core level follows
376  __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Core),
377  addrP[proc].first.labels[label++]);
378  if (levels[level++] > 0)
379  // thread level is the latest
380  __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Thread),
381  addrP[proc].first.labels[label++]);
382  KMP_DEBUG_ASSERT(label == depth);
383  }
384  KMP_INFORM(OSProcMapToPack, "KMP_AFFINITY", addrP[proc].second, buf.str);
385  __kmp_str_buf_clear(&buf);
386  }
387  __kmp_str_buf_free(&buf);
388 }
389 
390 static int nNodePerPkg, nTilePerPkg, nTilePerNode, nCorePerNode, nCorePerTile;
391 
392 // This function removes the topology levels that are radix 1 and don't offer
393 // further information about the topology. The most common example is when you
394 // have one thread context per core, we don't want the extra thread context
395 // level if it offers no unique labels. So they are removed.
396 // return value: the new depth of address2os
397 static int __kmp_affinity_remove_radix_one_levels(AddrUnsPair *addrP, int nTh,
398  int depth, int *levels) {
399  int level;
400  int i;
401  int radix1_detected;
402  int new_depth = depth;
403  for (level = depth - 1; level > 0; --level) {
404  // Detect if this level is radix 1
405  radix1_detected = 1;
406  for (i = 1; i < nTh; ++i) {
407  if (addrP[0].first.labels[level] != addrP[i].first.labels[level]) {
408  // There are differing label values for this level so it stays
409  radix1_detected = 0;
410  break;
411  }
412  }
413  if (!radix1_detected)
414  continue;
415  // Radix 1 was detected
416  --new_depth;
417  levels[level] = -1; // mark level as not present in address2os array
418  if (level == new_depth) {
419  // "turn off" deepest level, just decrement the depth that removes
420  // the level from address2os array
421  for (i = 0; i < nTh; ++i) {
422  addrP[i].first.depth--;
423  }
424  } else {
425  // For other levels, we move labels over and also reduce the depth
426  int j;
427  for (j = level; j < new_depth; ++j) {
428  for (i = 0; i < nTh; ++i) {
429  addrP[i].first.labels[j] = addrP[i].first.labels[j + 1];
430  addrP[i].first.depth--;
431  }
432  levels[j + 1] -= 1;
433  }
434  }
435  }
436  return new_depth;
437 }
438 
439 // Returns the number of objects of type 'type' below 'obj' within the topology
440 // tree structure. e.g., if obj is a HWLOC_OBJ_PACKAGE object, and type is
441 // HWLOC_OBJ_PU, then this will return the number of PU's under the SOCKET
442 // object.
443 static int __kmp_hwloc_get_nobjs_under_obj(hwloc_obj_t obj,
444  hwloc_obj_type_t type) {
445  int retval = 0;
446  hwloc_obj_t first;
447  for (first = hwloc_get_obj_below_by_type(__kmp_hwloc_topology, obj->type,
448  obj->logical_index, type, 0);
449  first != NULL &&
450  hwloc_get_ancestor_obj_by_type(__kmp_hwloc_topology, obj->type, first) ==
451  obj;
452  first = hwloc_get_next_obj_by_type(__kmp_hwloc_topology, first->type,
453  first)) {
454  ++retval;
455  }
456  return retval;
457 }
458 
459 static int __kmp_hwloc_count_children_by_depth(hwloc_topology_t t,
460  hwloc_obj_t o,
461  kmp_hwloc_depth_t depth,
462  hwloc_obj_t *f) {
463  if (o->depth == depth) {
464  if (*f == NULL)
465  *f = o; // output first descendant found
466  return 1;
467  }
468  int sum = 0;
469  for (unsigned i = 0; i < o->arity; i++)
470  sum += __kmp_hwloc_count_children_by_depth(t, o->children[i], depth, f);
471  return sum; // will be 0 if no one found (as PU arity is 0)
472 }
473 
474 static int __kmp_hwloc_count_children_by_type(hwloc_topology_t t, hwloc_obj_t o,
475  hwloc_obj_type_t type,
476  hwloc_obj_t *f) {
477  if (!hwloc_compare_types(o->type, type)) {
478  if (*f == NULL)
479  *f = o; // output first descendant found
480  return 1;
481  }
482  int sum = 0;
483  for (unsigned i = 0; i < o->arity; i++)
484  sum += __kmp_hwloc_count_children_by_type(t, o->children[i], type, f);
485  return sum; // will be 0 if no one found (as PU arity is 0)
486 }
487 
488 static int __kmp_hwloc_process_obj_core_pu(AddrUnsPair *addrPair,
489  int &nActiveThreads,
490  int &num_active_cores,
491  hwloc_obj_t obj, int depth,
492  int *labels) {
493  hwloc_obj_t core = NULL;
494  hwloc_topology_t &tp = __kmp_hwloc_topology;
495  int NC = __kmp_hwloc_count_children_by_type(tp, obj, HWLOC_OBJ_CORE, &core);
496  for (int core_id = 0; core_id < NC; ++core_id, core = core->next_cousin) {
497  hwloc_obj_t pu = NULL;
498  KMP_DEBUG_ASSERT(core != NULL);
499  int num_active_threads = 0;
500  int NT = __kmp_hwloc_count_children_by_type(tp, core, HWLOC_OBJ_PU, &pu);
501  // int NT = core->arity; pu = core->first_child; // faster?
502  for (int pu_id = 0; pu_id < NT; ++pu_id, pu = pu->next_cousin) {
503  KMP_DEBUG_ASSERT(pu != NULL);
504  if (!KMP_CPU_ISSET(pu->os_index, __kmp_affin_fullMask))
505  continue; // skip inactive (inaccessible) unit
506  Address addr(depth + 2);
507  KA_TRACE(20, ("Hwloc inserting %d (%d) %d (%d) %d (%d) into address2os\n",
508  obj->os_index, obj->logical_index, core->os_index,
509  core->logical_index, pu->os_index, pu->logical_index));
510  for (int i = 0; i < depth; ++i)
511  addr.labels[i] = labels[i]; // package, etc.
512  addr.labels[depth] = core_id; // core
513  addr.labels[depth + 1] = pu_id; // pu
514  addrPair[nActiveThreads] = AddrUnsPair(addr, pu->os_index);
515  __kmp_pu_os_idx[nActiveThreads] = pu->os_index;
516  nActiveThreads++;
517  ++num_active_threads; // count active threads per core
518  }
519  if (num_active_threads) { // were there any active threads on the core?
520  ++__kmp_ncores; // count total active cores
521  ++num_active_cores; // count active cores per socket
522  if (num_active_threads > __kmp_nThreadsPerCore)
523  __kmp_nThreadsPerCore = num_active_threads; // calc maximum
524  }
525  }
526  return 0;
527 }
528 
529 // Check if NUMA node detected below the package,
530 // and if tile object is detected and return its depth
531 static int __kmp_hwloc_check_numa() {
532  hwloc_topology_t &tp = __kmp_hwloc_topology;
533  hwloc_obj_t hT, hC, hL, hN, hS; // hwloc objects (pointers to)
534  int depth, l2cache_depth, package_depth;
535 
536  // Get some PU
537  hT = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PU, 0);
538  if (hT == NULL) // something has gone wrong
539  return 1;
540 
541  // check NUMA node below PACKAGE
542  hN = hwloc_get_ancestor_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hT);
543  hS = hwloc_get_ancestor_obj_by_type(tp, HWLOC_OBJ_PACKAGE, hT);
544  KMP_DEBUG_ASSERT(hS != NULL);
545  if (hN != NULL && hN->depth > hS->depth) {
546  __kmp_numa_detected = TRUE; // socket includes node(s)
547  if (__kmp_affinity_gran == affinity_gran_node) {
548  __kmp_affinity_gran = affinity_gran_numa;
549  }
550  }
551 
552  package_depth = hwloc_get_type_depth(tp, HWLOC_OBJ_PACKAGE);
553  l2cache_depth = hwloc_get_cache_type_depth(tp, 2, HWLOC_OBJ_CACHE_UNIFIED);
554  // check tile, get object by depth because of multiple caches possible
555  depth = (l2cache_depth < package_depth) ? package_depth : l2cache_depth;
556  hL = hwloc_get_ancestor_obj_by_depth(tp, depth, hT);
557  hC = NULL; // not used, but reset it here just in case
558  if (hL != NULL &&
559  __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE, &hC) > 1)
560  __kmp_tile_depth = depth; // tile consists of multiple cores
561  return 0;
562 }
563 
564 static int __kmp_affinity_create_hwloc_map(AddrUnsPair **address2os,
565  kmp_i18n_id_t *const msg_id) {
566  hwloc_topology_t &tp = __kmp_hwloc_topology; // shortcut of a long name
567  *address2os = NULL;
568  *msg_id = kmp_i18n_null;
569 
570  // Save the affinity mask for the current thread.
571  kmp_affin_mask_t *oldMask;
572  KMP_CPU_ALLOC(oldMask);
573  __kmp_get_system_affinity(oldMask, TRUE);
574  __kmp_hwloc_check_numa();
575 
576  if (!KMP_AFFINITY_CAPABLE()) {
577  // Hack to try and infer the machine topology using only the data
578  // available from cpuid on the current thread, and __kmp_xproc.
579  KMP_ASSERT(__kmp_affinity_type == affinity_none);
580 
581  nCoresPerPkg = __kmp_hwloc_get_nobjs_under_obj(
582  hwloc_get_obj_by_type(tp, HWLOC_OBJ_PACKAGE, 0), HWLOC_OBJ_CORE);
583  __kmp_nThreadsPerCore = __kmp_hwloc_get_nobjs_under_obj(
584  hwloc_get_obj_by_type(tp, HWLOC_OBJ_CORE, 0), HWLOC_OBJ_PU);
585  __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
586  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
587  if (__kmp_affinity_verbose) {
588  KMP_INFORM(AffNotCapableUseLocCpuidL11, "KMP_AFFINITY");
589  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
590  if (__kmp_affinity_uniform_topology()) {
591  KMP_INFORM(Uniform, "KMP_AFFINITY");
592  } else {
593  KMP_INFORM(NonUniform, "KMP_AFFINITY");
594  }
595  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
596  __kmp_nThreadsPerCore, __kmp_ncores);
597  }
598  KMP_CPU_FREE(oldMask);
599  return 0;
600  }
601 
602  int depth = 3;
603  int levels[5] = {0, 1, 2, 3, 4}; // package, [node,] [tile,] core, thread
604  int labels[3] = {0}; // package [,node] [,tile] - head of lables array
605  if (__kmp_numa_detected)
606  ++depth;
607  if (__kmp_tile_depth)
608  ++depth;
609 
610  // Allocate the data structure to be returned.
611  AddrUnsPair *retval =
612  (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * __kmp_avail_proc);
613  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
614  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
615 
616  // When affinity is off, this routine will still be called to set
617  // __kmp_ncores, as well as __kmp_nThreadsPerCore,
618  // nCoresPerPkg, & nPackages. Make sure all these vars are set
619  // correctly, and return if affinity is not enabled.
620 
621  hwloc_obj_t socket, node, tile;
622  int nActiveThreads = 0;
623  int socket_id = 0;
624  // re-calculate globals to count only accessible resources
625  __kmp_ncores = nPackages = nCoresPerPkg = __kmp_nThreadsPerCore = 0;
626  nNodePerPkg = nTilePerPkg = nTilePerNode = nCorePerNode = nCorePerTile = 0;
627  for (socket = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PACKAGE, 0); socket != NULL;
628  socket = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PACKAGE, socket),
629  socket_id++) {
630  labels[0] = socket_id;
631  if (__kmp_numa_detected) {
632  int NN;
633  int n_active_nodes = 0;
634  node = NULL;
635  NN = __kmp_hwloc_count_children_by_type(tp, socket, HWLOC_OBJ_NUMANODE,
636  &node);
637  for (int node_id = 0; node_id < NN; ++node_id, node = node->next_cousin) {
638  labels[1] = node_id;
639  if (__kmp_tile_depth) {
640  // NUMA + tiles
641  int NT;
642  int n_active_tiles = 0;
643  tile = NULL;
644  NT = __kmp_hwloc_count_children_by_depth(tp, node, __kmp_tile_depth,
645  &tile);
646  for (int tl_id = 0; tl_id < NT; ++tl_id, tile = tile->next_cousin) {
647  labels[2] = tl_id;
648  int n_active_cores = 0;
649  __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads,
650  n_active_cores, tile, 3, labels);
651  if (n_active_cores) { // were there any active cores on the socket?
652  ++n_active_tiles; // count active tiles per node
653  if (n_active_cores > nCorePerTile)
654  nCorePerTile = n_active_cores; // calc maximum
655  }
656  }
657  if (n_active_tiles) { // were there any active tiles on the socket?
658  ++n_active_nodes; // count active nodes per package
659  if (n_active_tiles > nTilePerNode)
660  nTilePerNode = n_active_tiles; // calc maximum
661  }
662  } else {
663  // NUMA, no tiles
664  int n_active_cores = 0;
665  __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads,
666  n_active_cores, node, 2, labels);
667  if (n_active_cores) { // were there any active cores on the socket?
668  ++n_active_nodes; // count active nodes per package
669  if (n_active_cores > nCorePerNode)
670  nCorePerNode = n_active_cores; // calc maximum
671  }
672  }
673  }
674  if (n_active_nodes) { // were there any active nodes on the socket?
675  ++nPackages; // count total active packages
676  if (n_active_nodes > nNodePerPkg)
677  nNodePerPkg = n_active_nodes; // calc maximum
678  }
679  } else {
680  if (__kmp_tile_depth) {
681  // no NUMA, tiles
682  int NT;
683  int n_active_tiles = 0;
684  tile = NULL;
685  NT = __kmp_hwloc_count_children_by_depth(tp, socket, __kmp_tile_depth,
686  &tile);
687  for (int tl_id = 0; tl_id < NT; ++tl_id, tile = tile->next_cousin) {
688  labels[1] = tl_id;
689  int n_active_cores = 0;
690  __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads,
691  n_active_cores, tile, 2, labels);
692  if (n_active_cores) { // were there any active cores on the socket?
693  ++n_active_tiles; // count active tiles per package
694  if (n_active_cores > nCorePerTile)
695  nCorePerTile = n_active_cores; // calc maximum
696  }
697  }
698  if (n_active_tiles) { // were there any active tiles on the socket?
699  ++nPackages; // count total active packages
700  if (n_active_tiles > nTilePerPkg)
701  nTilePerPkg = n_active_tiles; // calc maximum
702  }
703  } else {
704  // no NUMA, no tiles
705  int n_active_cores = 0;
706  __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads, n_active_cores,
707  socket, 1, labels);
708  if (n_active_cores) { // were there any active cores on the socket?
709  ++nPackages; // count total active packages
710  if (n_active_cores > nCoresPerPkg)
711  nCoresPerPkg = n_active_cores; // calc maximum
712  }
713  }
714  }
715  }
716 
717  // If there's only one thread context to bind to, return now.
718  KMP_DEBUG_ASSERT(nActiveThreads == __kmp_avail_proc);
719  KMP_ASSERT(nActiveThreads > 0);
720  if (nActiveThreads == 1) {
721  __kmp_ncores = nPackages = 1;
722  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
723  if (__kmp_affinity_verbose) {
724  char buf[KMP_AFFIN_MASK_PRINT_LEN];
725  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
726 
727  KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
728  if (__kmp_affinity_respect_mask) {
729  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
730  } else {
731  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
732  }
733  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
734  KMP_INFORM(Uniform, "KMP_AFFINITY");
735  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
736  __kmp_nThreadsPerCore, __kmp_ncores);
737  }
738 
739  if (__kmp_affinity_type == affinity_none) {
740  __kmp_free(retval);
741  KMP_CPU_FREE(oldMask);
742  return 0;
743  }
744 
745  // Form an Address object which only includes the package level.
746  Address addr(1);
747  addr.labels[0] = retval[0].first.labels[0];
748  retval[0].first = addr;
749 
750  if (__kmp_affinity_gran_levels < 0) {
751  __kmp_affinity_gran_levels = 0;
752  }
753 
754  if (__kmp_affinity_verbose) {
755  __kmp_affinity_print_topology(retval, 1, 1, 0, -1, -1);
756  }
757 
758  *address2os = retval;
759  KMP_CPU_FREE(oldMask);
760  return 1;
761  }
762 
763  // Sort the table by physical Id.
764  qsort(retval, nActiveThreads, sizeof(*retval),
765  __kmp_affinity_cmp_Address_labels);
766 
767  // Check to see if the machine topology is uniform
768  int nPUs = nPackages * __kmp_nThreadsPerCore;
769  if (__kmp_numa_detected) {
770  if (__kmp_tile_depth) { // NUMA + tiles
771  nPUs *= (nNodePerPkg * nTilePerNode * nCorePerTile);
772  } else { // NUMA, no tiles
773  nPUs *= (nNodePerPkg * nCorePerNode);
774  }
775  } else {
776  if (__kmp_tile_depth) { // no NUMA, tiles
777  nPUs *= (nTilePerPkg * nCorePerTile);
778  } else { // no NUMA, no tiles
779  nPUs *= nCoresPerPkg;
780  }
781  }
782  unsigned uniform = (nPUs == nActiveThreads);
783 
784  // Print the machine topology summary.
785  if (__kmp_affinity_verbose) {
786  char mask[KMP_AFFIN_MASK_PRINT_LEN];
787  __kmp_affinity_print_mask(mask, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
788  if (__kmp_affinity_respect_mask) {
789  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", mask);
790  } else {
791  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", mask);
792  }
793  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
794  if (uniform) {
795  KMP_INFORM(Uniform, "KMP_AFFINITY");
796  } else {
797  KMP_INFORM(NonUniform, "KMP_AFFINITY");
798  }
799  if (__kmp_numa_detected) {
800  if (__kmp_tile_depth) { // NUMA + tiles
801  KMP_INFORM(TopologyExtraNoTi, "KMP_AFFINITY", nPackages, nNodePerPkg,
802  nTilePerNode, nCorePerTile, __kmp_nThreadsPerCore,
803  __kmp_ncores);
804  } else { // NUMA, no tiles
805  KMP_INFORM(TopologyExtraNode, "KMP_AFFINITY", nPackages, nNodePerPkg,
806  nCorePerNode, __kmp_nThreadsPerCore, __kmp_ncores);
807  nPUs *= (nNodePerPkg * nCorePerNode);
808  }
809  } else {
810  if (__kmp_tile_depth) { // no NUMA, tiles
811  KMP_INFORM(TopologyExtraTile, "KMP_AFFINITY", nPackages, nTilePerPkg,
812  nCorePerTile, __kmp_nThreadsPerCore, __kmp_ncores);
813  } else { // no NUMA, no tiles
814  kmp_str_buf_t buf;
815  __kmp_str_buf_init(&buf);
816  __kmp_str_buf_print(&buf, "%d", nPackages);
817  KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, nCoresPerPkg,
818  __kmp_nThreadsPerCore, __kmp_ncores);
819  __kmp_str_buf_free(&buf);
820  }
821  }
822  }
823 
824  if (__kmp_affinity_type == affinity_none) {
825  __kmp_free(retval);
826  KMP_CPU_FREE(oldMask);
827  return 0;
828  }
829 
830  int depth_full = depth; // number of levels before compressing
831  // Find any levels with radiix 1, and remove them from the map
832  // (except for the package level).
833  depth = __kmp_affinity_remove_radix_one_levels(retval, nActiveThreads, depth,
834  levels);
835  KMP_DEBUG_ASSERT(__kmp_affinity_gran != affinity_gran_default);
836  if (__kmp_affinity_gran_levels < 0) {
837  // Set the granularity level based on what levels are modeled
838  // in the machine topology map.
839  __kmp_affinity_gran_levels = 0; // lowest level (e.g. fine)
840  if (__kmp_affinity_gran > affinity_gran_thread) {
841  for (int i = 1; i <= depth_full; ++i) {
842  if (__kmp_affinity_gran <= i) // only count deeper levels
843  break;
844  if (levels[depth_full - i] > 0)
845  __kmp_affinity_gran_levels++;
846  }
847  }
848  if (__kmp_affinity_gran > affinity_gran_package)
849  __kmp_affinity_gran_levels++; // e.g. granularity = group
850  }
851 
852  if (__kmp_affinity_verbose)
853  __kmp_affinity_print_hwloc_tp(retval, nActiveThreads, depth, levels);
854 
855  KMP_CPU_FREE(oldMask);
856  *address2os = retval;
857  return depth;
858 }
859 #endif // KMP_USE_HWLOC
860 
861 // If we don't know how to retrieve the machine's processor topology, or
862 // encounter an error in doing so, this routine is called to form a "flat"
863 // mapping of os thread id's <-> processor id's.
864 static int __kmp_affinity_create_flat_map(AddrUnsPair **address2os,
865  kmp_i18n_id_t *const msg_id) {
866  *address2os = NULL;
867  *msg_id = kmp_i18n_null;
868 
869  // Even if __kmp_affinity_type == affinity_none, this routine might still
870  // called to set __kmp_ncores, as well as
871  // __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
872  if (!KMP_AFFINITY_CAPABLE()) {
873  KMP_ASSERT(__kmp_affinity_type == affinity_none);
874  __kmp_ncores = nPackages = __kmp_xproc;
875  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
876  if (__kmp_affinity_verbose) {
877  KMP_INFORM(AffFlatTopology, "KMP_AFFINITY");
878  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
879  KMP_INFORM(Uniform, "KMP_AFFINITY");
880  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
881  __kmp_nThreadsPerCore, __kmp_ncores);
882  }
883  return 0;
884  }
885 
886  // When affinity is off, this routine will still be called to set
887  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
888  // Make sure all these vars are set correctly, and return now if affinity is
889  // not enabled.
890  __kmp_ncores = nPackages = __kmp_avail_proc;
891  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
892  if (__kmp_affinity_verbose) {
893  char buf[KMP_AFFIN_MASK_PRINT_LEN];
894  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
895  __kmp_affin_fullMask);
896 
897  KMP_INFORM(AffCapableUseFlat, "KMP_AFFINITY");
898  if (__kmp_affinity_respect_mask) {
899  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
900  } else {
901  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
902  }
903  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
904  KMP_INFORM(Uniform, "KMP_AFFINITY");
905  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
906  __kmp_nThreadsPerCore, __kmp_ncores);
907  }
908  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
909  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
910  if (__kmp_affinity_type == affinity_none) {
911  int avail_ct = 0;
912  int i;
913  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
914  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask))
915  continue;
916  __kmp_pu_os_idx[avail_ct++] = i; // suppose indices are flat
917  }
918  return 0;
919  }
920 
921  // Contruct the data structure to be returned.
922  *address2os =
923  (AddrUnsPair *)__kmp_allocate(sizeof(**address2os) * __kmp_avail_proc);
924  int avail_ct = 0;
925  int i;
926  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
927  // Skip this proc if it is not included in the machine model.
928  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
929  continue;
930  }
931  __kmp_pu_os_idx[avail_ct] = i; // suppose indices are flat
932  Address addr(1);
933  addr.labels[0] = i;
934  (*address2os)[avail_ct++] = AddrUnsPair(addr, i);
935  }
936  if (__kmp_affinity_verbose) {
937  KMP_INFORM(OSProcToPackage, "KMP_AFFINITY");
938  }
939 
940  if (__kmp_affinity_gran_levels < 0) {
941  // Only the package level is modeled in the machine topology map,
942  // so the #levels of granularity is either 0 or 1.
943  if (__kmp_affinity_gran > affinity_gran_package) {
944  __kmp_affinity_gran_levels = 1;
945  } else {
946  __kmp_affinity_gran_levels = 0;
947  }
948  }
949  return 1;
950 }
951 
952 #if KMP_GROUP_AFFINITY
953 
954 // If multiple Windows* OS processor groups exist, we can create a 2-level
955 // topology map with the groups at level 0 and the individual procs at level 1.
956 // This facilitates letting the threads float among all procs in a group,
957 // if granularity=group (the default when there are multiple groups).
958 static int __kmp_affinity_create_proc_group_map(AddrUnsPair **address2os,
959  kmp_i18n_id_t *const msg_id) {
960  *address2os = NULL;
961  *msg_id = kmp_i18n_null;
962 
963  // If we aren't affinity capable, then return now.
964  // The flat mapping will be used.
965  if (!KMP_AFFINITY_CAPABLE()) {
966  // FIXME set *msg_id
967  return -1;
968  }
969 
970  // Contruct the data structure to be returned.
971  *address2os =
972  (AddrUnsPair *)__kmp_allocate(sizeof(**address2os) * __kmp_avail_proc);
973  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
974  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
975  int avail_ct = 0;
976  int i;
977  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
978  // Skip this proc if it is not included in the machine model.
979  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
980  continue;
981  }
982  __kmp_pu_os_idx[avail_ct] = i; // suppose indices are flat
983  Address addr(2);
984  addr.labels[0] = i / (CHAR_BIT * sizeof(DWORD_PTR));
985  addr.labels[1] = i % (CHAR_BIT * sizeof(DWORD_PTR));
986  (*address2os)[avail_ct++] = AddrUnsPair(addr, i);
987 
988  if (__kmp_affinity_verbose) {
989  KMP_INFORM(AffOSProcToGroup, "KMP_AFFINITY", i, addr.labels[0],
990  addr.labels[1]);
991  }
992  }
993 
994  if (__kmp_affinity_gran_levels < 0) {
995  if (__kmp_affinity_gran == affinity_gran_group) {
996  __kmp_affinity_gran_levels = 1;
997  } else if ((__kmp_affinity_gran == affinity_gran_fine) ||
998  (__kmp_affinity_gran == affinity_gran_thread)) {
999  __kmp_affinity_gran_levels = 0;
1000  } else {
1001  const char *gran_str = NULL;
1002  if (__kmp_affinity_gran == affinity_gran_core) {
1003  gran_str = "core";
1004  } else if (__kmp_affinity_gran == affinity_gran_package) {
1005  gran_str = "package";
1006  } else if (__kmp_affinity_gran == affinity_gran_node) {
1007  gran_str = "node";
1008  } else {
1009  KMP_ASSERT(0);
1010  }
1011 
1012  // Warning: can't use affinity granularity \"gran\" with group topology
1013  // method, using "thread"
1014  __kmp_affinity_gran_levels = 0;
1015  }
1016  }
1017  return 2;
1018 }
1019 
1020 #endif /* KMP_GROUP_AFFINITY */
1021 
1022 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1023 
1024 static int __kmp_cpuid_mask_width(int count) {
1025  int r = 0;
1026 
1027  while ((1 << r) < count)
1028  ++r;
1029  return r;
1030 }
1031 
1032 class apicThreadInfo {
1033 public:
1034  unsigned osId; // param to __kmp_affinity_bind_thread
1035  unsigned apicId; // from cpuid after binding
1036  unsigned maxCoresPerPkg; // ""
1037  unsigned maxThreadsPerPkg; // ""
1038  unsigned pkgId; // inferred from above values
1039  unsigned coreId; // ""
1040  unsigned threadId; // ""
1041 };
1042 
1043 static int __kmp_affinity_cmp_apicThreadInfo_phys_id(const void *a,
1044  const void *b) {
1045  const apicThreadInfo *aa = (const apicThreadInfo *)a;
1046  const apicThreadInfo *bb = (const apicThreadInfo *)b;
1047  if (aa->pkgId < bb->pkgId)
1048  return -1;
1049  if (aa->pkgId > bb->pkgId)
1050  return 1;
1051  if (aa->coreId < bb->coreId)
1052  return -1;
1053  if (aa->coreId > bb->coreId)
1054  return 1;
1055  if (aa->threadId < bb->threadId)
1056  return -1;
1057  if (aa->threadId > bb->threadId)
1058  return 1;
1059  return 0;
1060 }
1061 
1062 // On IA-32 architecture and Intel(R) 64 architecture, we attempt to use
1063 // an algorithm which cycles through the available os threads, setting
1064 // the current thread's affinity mask to that thread, and then retrieves
1065 // the Apic Id for each thread context using the cpuid instruction.
1066 static int __kmp_affinity_create_apicid_map(AddrUnsPair **address2os,
1067  kmp_i18n_id_t *const msg_id) {
1068  kmp_cpuid buf;
1069  *address2os = NULL;
1070  *msg_id = kmp_i18n_null;
1071 
1072  // Check if cpuid leaf 4 is supported.
1073  __kmp_x86_cpuid(0, 0, &buf);
1074  if (buf.eax < 4) {
1075  *msg_id = kmp_i18n_str_NoLeaf4Support;
1076  return -1;
1077  }
1078 
1079  // The algorithm used starts by setting the affinity to each available thread
1080  // and retrieving info from the cpuid instruction, so if we are not capable of
1081  // calling __kmp_get_system_affinity() and _kmp_get_system_affinity(), then we
1082  // need to do something else - use the defaults that we calculated from
1083  // issuing cpuid without binding to each proc.
1084  if (!KMP_AFFINITY_CAPABLE()) {
1085  // Hack to try and infer the machine topology using only the data
1086  // available from cpuid on the current thread, and __kmp_xproc.
1087  KMP_ASSERT(__kmp_affinity_type == affinity_none);
1088 
1089  // Get an upper bound on the number of threads per package using cpuid(1).
1090  // On some OS/chps combinations where HT is supported by the chip but is
1091  // disabled, this value will be 2 on a single core chip. Usually, it will be
1092  // 2 if HT is enabled and 1 if HT is disabled.
1093  __kmp_x86_cpuid(1, 0, &buf);
1094  int maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
1095  if (maxThreadsPerPkg == 0) {
1096  maxThreadsPerPkg = 1;
1097  }
1098 
1099  // The num cores per pkg comes from cpuid(4). 1 must be added to the encoded
1100  // value.
1101  //
1102  // The author of cpu_count.cpp treated this only an upper bound on the
1103  // number of cores, but I haven't seen any cases where it was greater than
1104  // the actual number of cores, so we will treat it as exact in this block of
1105  // code.
1106  //
1107  // First, we need to check if cpuid(4) is supported on this chip. To see if
1108  // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n or
1109  // greater.
1110  __kmp_x86_cpuid(0, 0, &buf);
1111  if (buf.eax >= 4) {
1112  __kmp_x86_cpuid(4, 0, &buf);
1113  nCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
1114  } else {
1115  nCoresPerPkg = 1;
1116  }
1117 
1118  // There is no way to reliably tell if HT is enabled without issuing the
1119  // cpuid instruction from every thread, can correlating the cpuid info, so
1120  // if the machine is not affinity capable, we assume that HT is off. We have
1121  // seen quite a few machines where maxThreadsPerPkg is 2, yet the machine
1122  // does not support HT.
1123  //
1124  // - Older OSes are usually found on machines with older chips, which do not
1125  // support HT.
1126  // - The performance penalty for mistakenly identifying a machine as HT when
1127  // it isn't (which results in blocktime being incorrecly set to 0) is
1128  // greater than the penalty when for mistakenly identifying a machine as
1129  // being 1 thread/core when it is really HT enabled (which results in
1130  // blocktime being incorrectly set to a positive value).
1131  __kmp_ncores = __kmp_xproc;
1132  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
1133  __kmp_nThreadsPerCore = 1;
1134  if (__kmp_affinity_verbose) {
1135  KMP_INFORM(AffNotCapableUseLocCpuid, "KMP_AFFINITY");
1136  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1137  if (__kmp_affinity_uniform_topology()) {
1138  KMP_INFORM(Uniform, "KMP_AFFINITY");
1139  } else {
1140  KMP_INFORM(NonUniform, "KMP_AFFINITY");
1141  }
1142  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
1143  __kmp_nThreadsPerCore, __kmp_ncores);
1144  }
1145  return 0;
1146  }
1147 
1148  // From here on, we can assume that it is safe to call
1149  // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
1150  // __kmp_affinity_type = affinity_none.
1151 
1152  // Save the affinity mask for the current thread.
1153  kmp_affin_mask_t *oldMask;
1154  KMP_CPU_ALLOC(oldMask);
1155  KMP_ASSERT(oldMask != NULL);
1156  __kmp_get_system_affinity(oldMask, TRUE);
1157 
1158  // Run through each of the available contexts, binding the current thread
1159  // to it, and obtaining the pertinent information using the cpuid instr.
1160  //
1161  // The relevant information is:
1162  // - Apic Id: Bits 24:31 of ebx after issuing cpuid(1) - each thread context
1163  // has a uniqie Apic Id, which is of the form pkg# : core# : thread#.
1164  // - Max Threads Per Pkg: Bits 16:23 of ebx after issuing cpuid(1). The value
1165  // of this field determines the width of the core# + thread# fields in the
1166  // Apic Id. It is also an upper bound on the number of threads per
1167  // package, but it has been verified that situations happen were it is not
1168  // exact. In particular, on certain OS/chip combinations where Intel(R)
1169  // Hyper-Threading Technology is supported by the chip but has been
1170  // disabled, the value of this field will be 2 (for a single core chip).
1171  // On other OS/chip combinations supporting Intel(R) Hyper-Threading
1172  // Technology, the value of this field will be 1 when Intel(R)
1173  // Hyper-Threading Technology is disabled and 2 when it is enabled.
1174  // - Max Cores Per Pkg: Bits 26:31 of eax after issuing cpuid(4). The value
1175  // of this field (+1) determines the width of the core# field in the Apic
1176  // Id. The comments in "cpucount.cpp" say that this value is an upper
1177  // bound, but the IA-32 architecture manual says that it is exactly the
1178  // number of cores per package, and I haven't seen any case where it
1179  // wasn't.
1180  //
1181  // From this information, deduce the package Id, core Id, and thread Id,
1182  // and set the corresponding fields in the apicThreadInfo struct.
1183  unsigned i;
1184  apicThreadInfo *threadInfo = (apicThreadInfo *)__kmp_allocate(
1185  __kmp_avail_proc * sizeof(apicThreadInfo));
1186  unsigned nApics = 0;
1187  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
1188  // Skip this proc if it is not included in the machine model.
1189  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
1190  continue;
1191  }
1192  KMP_DEBUG_ASSERT((int)nApics < __kmp_avail_proc);
1193 
1194  __kmp_affinity_dispatch->bind_thread(i);
1195  threadInfo[nApics].osId = i;
1196 
1197  // The apic id and max threads per pkg come from cpuid(1).
1198  __kmp_x86_cpuid(1, 0, &buf);
1199  if (((buf.edx >> 9) & 1) == 0) {
1200  __kmp_set_system_affinity(oldMask, TRUE);
1201  __kmp_free(threadInfo);
1202  KMP_CPU_FREE(oldMask);
1203  *msg_id = kmp_i18n_str_ApicNotPresent;
1204  return -1;
1205  }
1206  threadInfo[nApics].apicId = (buf.ebx >> 24) & 0xff;
1207  threadInfo[nApics].maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
1208  if (threadInfo[nApics].maxThreadsPerPkg == 0) {
1209  threadInfo[nApics].maxThreadsPerPkg = 1;
1210  }
1211 
1212  // Max cores per pkg comes from cpuid(4). 1 must be added to the encoded
1213  // value.
1214  //
1215  // First, we need to check if cpuid(4) is supported on this chip. To see if
1216  // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n
1217  // or greater.
1218  __kmp_x86_cpuid(0, 0, &buf);
1219  if (buf.eax >= 4) {
1220  __kmp_x86_cpuid(4, 0, &buf);
1221  threadInfo[nApics].maxCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
1222  } else {
1223  threadInfo[nApics].maxCoresPerPkg = 1;
1224  }
1225 
1226  // Infer the pkgId / coreId / threadId using only the info obtained locally.
1227  int widthCT = __kmp_cpuid_mask_width(threadInfo[nApics].maxThreadsPerPkg);
1228  threadInfo[nApics].pkgId = threadInfo[nApics].apicId >> widthCT;
1229 
1230  int widthC = __kmp_cpuid_mask_width(threadInfo[nApics].maxCoresPerPkg);
1231  int widthT = widthCT - widthC;
1232  if (widthT < 0) {
1233  // I've never seen this one happen, but I suppose it could, if the cpuid
1234  // instruction on a chip was really screwed up. Make sure to restore the
1235  // affinity mask before the tail call.
1236  __kmp_set_system_affinity(oldMask, TRUE);
1237  __kmp_free(threadInfo);
1238  KMP_CPU_FREE(oldMask);
1239  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1240  return -1;
1241  }
1242 
1243  int maskC = (1 << widthC) - 1;
1244  threadInfo[nApics].coreId = (threadInfo[nApics].apicId >> widthT) & maskC;
1245 
1246  int maskT = (1 << widthT) - 1;
1247  threadInfo[nApics].threadId = threadInfo[nApics].apicId & maskT;
1248 
1249  nApics++;
1250  }
1251 
1252  // We've collected all the info we need.
1253  // Restore the old affinity mask for this thread.
1254  __kmp_set_system_affinity(oldMask, TRUE);
1255 
1256  // If there's only one thread context to bind to, form an Address object
1257  // with depth 1 and return immediately (or, if affinity is off, set
1258  // address2os to NULL and return).
1259  //
1260  // If it is configured to omit the package level when there is only a single
1261  // package, the logic at the end of this routine won't work if there is only
1262  // a single thread - it would try to form an Address object with depth 0.
1263  KMP_ASSERT(nApics > 0);
1264  if (nApics == 1) {
1265  __kmp_ncores = nPackages = 1;
1266  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
1267  if (__kmp_affinity_verbose) {
1268  char buf[KMP_AFFIN_MASK_PRINT_LEN];
1269  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
1270 
1271  KMP_INFORM(AffUseGlobCpuid, "KMP_AFFINITY");
1272  if (__kmp_affinity_respect_mask) {
1273  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
1274  } else {
1275  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
1276  }
1277  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1278  KMP_INFORM(Uniform, "KMP_AFFINITY");
1279  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
1280  __kmp_nThreadsPerCore, __kmp_ncores);
1281  }
1282 
1283  if (__kmp_affinity_type == affinity_none) {
1284  __kmp_free(threadInfo);
1285  KMP_CPU_FREE(oldMask);
1286  return 0;
1287  }
1288 
1289  *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair));
1290  Address addr(1);
1291  addr.labels[0] = threadInfo[0].pkgId;
1292  (*address2os)[0] = AddrUnsPair(addr, threadInfo[0].osId);
1293 
1294  if (__kmp_affinity_gran_levels < 0) {
1295  __kmp_affinity_gran_levels = 0;
1296  }
1297 
1298  if (__kmp_affinity_verbose) {
1299  __kmp_affinity_print_topology(*address2os, 1, 1, 0, -1, -1);
1300  }
1301 
1302  __kmp_free(threadInfo);
1303  KMP_CPU_FREE(oldMask);
1304  return 1;
1305  }
1306 
1307  // Sort the threadInfo table by physical Id.
1308  qsort(threadInfo, nApics, sizeof(*threadInfo),
1309  __kmp_affinity_cmp_apicThreadInfo_phys_id);
1310 
1311  // The table is now sorted by pkgId / coreId / threadId, but we really don't
1312  // know the radix of any of the fields. pkgId's may be sparsely assigned among
1313  // the chips on a system. Although coreId's are usually assigned
1314  // [0 .. coresPerPkg-1] and threadId's are usually assigned
1315  // [0..threadsPerCore-1], we don't want to make any such assumptions.
1316  //
1317  // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
1318  // total # packages) are at this point - we want to determine that now. We
1319  // only have an upper bound on the first two figures.
1320  //
1321  // We also perform a consistency check at this point: the values returned by
1322  // the cpuid instruction for any thread bound to a given package had better
1323  // return the same info for maxThreadsPerPkg and maxCoresPerPkg.
1324  nPackages = 1;
1325  nCoresPerPkg = 1;
1326  __kmp_nThreadsPerCore = 1;
1327  unsigned nCores = 1;
1328 
1329  unsigned pkgCt = 1; // to determine radii
1330  unsigned lastPkgId = threadInfo[0].pkgId;
1331  unsigned coreCt = 1;
1332  unsigned lastCoreId = threadInfo[0].coreId;
1333  unsigned threadCt = 1;
1334  unsigned lastThreadId = threadInfo[0].threadId;
1335 
1336  // intra-pkg consist checks
1337  unsigned prevMaxCoresPerPkg = threadInfo[0].maxCoresPerPkg;
1338  unsigned prevMaxThreadsPerPkg = threadInfo[0].maxThreadsPerPkg;
1339 
1340  for (i = 1; i < nApics; i++) {
1341  if (threadInfo[i].pkgId != lastPkgId) {
1342  nCores++;
1343  pkgCt++;
1344  lastPkgId = threadInfo[i].pkgId;
1345  if ((int)coreCt > nCoresPerPkg)
1346  nCoresPerPkg = coreCt;
1347  coreCt = 1;
1348  lastCoreId = threadInfo[i].coreId;
1349  if ((int)threadCt > __kmp_nThreadsPerCore)
1350  __kmp_nThreadsPerCore = threadCt;
1351  threadCt = 1;
1352  lastThreadId = threadInfo[i].threadId;
1353 
1354  // This is a different package, so go on to the next iteration without
1355  // doing any consistency checks. Reset the consistency check vars, though.
1356  prevMaxCoresPerPkg = threadInfo[i].maxCoresPerPkg;
1357  prevMaxThreadsPerPkg = threadInfo[i].maxThreadsPerPkg;
1358  continue;
1359  }
1360 
1361  if (threadInfo[i].coreId != lastCoreId) {
1362  nCores++;
1363  coreCt++;
1364  lastCoreId = threadInfo[i].coreId;
1365  if ((int)threadCt > __kmp_nThreadsPerCore)
1366  __kmp_nThreadsPerCore = threadCt;
1367  threadCt = 1;
1368  lastThreadId = threadInfo[i].threadId;
1369  } else if (threadInfo[i].threadId != lastThreadId) {
1370  threadCt++;
1371  lastThreadId = threadInfo[i].threadId;
1372  } else {
1373  __kmp_free(threadInfo);
1374  KMP_CPU_FREE(oldMask);
1375  *msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
1376  return -1;
1377  }
1378 
1379  // Check to make certain that the maxCoresPerPkg and maxThreadsPerPkg
1380  // fields agree between all the threads bounds to a given package.
1381  if ((prevMaxCoresPerPkg != threadInfo[i].maxCoresPerPkg) ||
1382  (prevMaxThreadsPerPkg != threadInfo[i].maxThreadsPerPkg)) {
1383  __kmp_free(threadInfo);
1384  KMP_CPU_FREE(oldMask);
1385  *msg_id = kmp_i18n_str_InconsistentCpuidInfo;
1386  return -1;
1387  }
1388  }
1389  nPackages = pkgCt;
1390  if ((int)coreCt > nCoresPerPkg)
1391  nCoresPerPkg = coreCt;
1392  if ((int)threadCt > __kmp_nThreadsPerCore)
1393  __kmp_nThreadsPerCore = threadCt;
1394 
1395  // When affinity is off, this routine will still be called to set
1396  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
1397  // Make sure all these vars are set correctly, and return now if affinity is
1398  // not enabled.
1399  __kmp_ncores = nCores;
1400  if (__kmp_affinity_verbose) {
1401  char buf[KMP_AFFIN_MASK_PRINT_LEN];
1402  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
1403 
1404  KMP_INFORM(AffUseGlobCpuid, "KMP_AFFINITY");
1405  if (__kmp_affinity_respect_mask) {
1406  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
1407  } else {
1408  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
1409  }
1410  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1411  if (__kmp_affinity_uniform_topology()) {
1412  KMP_INFORM(Uniform, "KMP_AFFINITY");
1413  } else {
1414  KMP_INFORM(NonUniform, "KMP_AFFINITY");
1415  }
1416  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
1417  __kmp_nThreadsPerCore, __kmp_ncores);
1418  }
1419  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
1420  KMP_DEBUG_ASSERT(nApics == (unsigned)__kmp_avail_proc);
1421  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
1422  for (i = 0; i < nApics; ++i) {
1423  __kmp_pu_os_idx[i] = threadInfo[i].osId;
1424  }
1425  if (__kmp_affinity_type == affinity_none) {
1426  __kmp_free(threadInfo);
1427  KMP_CPU_FREE(oldMask);
1428  return 0;
1429  }
1430 
1431  // Now that we've determined the number of packages, the number of cores per
1432  // package, and the number of threads per core, we can construct the data
1433  // structure that is to be returned.
1434  int pkgLevel = 0;
1435  int coreLevel = (nCoresPerPkg <= 1) ? -1 : 1;
1436  int threadLevel =
1437  (__kmp_nThreadsPerCore <= 1) ? -1 : ((coreLevel >= 0) ? 2 : 1);
1438  unsigned depth = (pkgLevel >= 0) + (coreLevel >= 0) + (threadLevel >= 0);
1439 
1440  KMP_ASSERT(depth > 0);
1441  *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * nApics);
1442 
1443  for (i = 0; i < nApics; ++i) {
1444  Address addr(depth);
1445  unsigned os = threadInfo[i].osId;
1446  int d = 0;
1447 
1448  if (pkgLevel >= 0) {
1449  addr.labels[d++] = threadInfo[i].pkgId;
1450  }
1451  if (coreLevel >= 0) {
1452  addr.labels[d++] = threadInfo[i].coreId;
1453  }
1454  if (threadLevel >= 0) {
1455  addr.labels[d++] = threadInfo[i].threadId;
1456  }
1457  (*address2os)[i] = AddrUnsPair(addr, os);
1458  }
1459 
1460  if (__kmp_affinity_gran_levels < 0) {
1461  // Set the granularity level based on what levels are modeled in the machine
1462  // topology map.
1463  __kmp_affinity_gran_levels = 0;
1464  if ((threadLevel >= 0) && (__kmp_affinity_gran > affinity_gran_thread)) {
1465  __kmp_affinity_gran_levels++;
1466  }
1467  if ((coreLevel >= 0) && (__kmp_affinity_gran > affinity_gran_core)) {
1468  __kmp_affinity_gran_levels++;
1469  }
1470  if ((pkgLevel >= 0) && (__kmp_affinity_gran > affinity_gran_package)) {
1471  __kmp_affinity_gran_levels++;
1472  }
1473  }
1474 
1475  if (__kmp_affinity_verbose) {
1476  __kmp_affinity_print_topology(*address2os, nApics, depth, pkgLevel,
1477  coreLevel, threadLevel);
1478  }
1479 
1480  __kmp_free(threadInfo);
1481  KMP_CPU_FREE(oldMask);
1482  return depth;
1483 }
1484 
1485 // Intel(R) microarchitecture code name Nehalem, Dunnington and later
1486 // architectures support a newer interface for specifying the x2APIC Ids,
1487 // based on cpuid leaf 11.
1488 static int __kmp_affinity_create_x2apicid_map(AddrUnsPair **address2os,
1489  kmp_i18n_id_t *const msg_id) {
1490  kmp_cpuid buf;
1491  *address2os = NULL;
1492  *msg_id = kmp_i18n_null;
1493 
1494  // Check to see if cpuid leaf 11 is supported.
1495  __kmp_x86_cpuid(0, 0, &buf);
1496  if (buf.eax < 11) {
1497  *msg_id = kmp_i18n_str_NoLeaf11Support;
1498  return -1;
1499  }
1500  __kmp_x86_cpuid(11, 0, &buf);
1501  if (buf.ebx == 0) {
1502  *msg_id = kmp_i18n_str_NoLeaf11Support;
1503  return -1;
1504  }
1505 
1506  // Find the number of levels in the machine topology. While we're at it, get
1507  // the default values for __kmp_nThreadsPerCore & nCoresPerPkg. We will try to
1508  // get more accurate values later by explicitly counting them, but get
1509  // reasonable defaults now, in case we return early.
1510  int level;
1511  int threadLevel = -1;
1512  int coreLevel = -1;
1513  int pkgLevel = -1;
1514  __kmp_nThreadsPerCore = nCoresPerPkg = nPackages = 1;
1515 
1516  for (level = 0;; level++) {
1517  if (level > 31) {
1518  // FIXME: Hack for DPD200163180
1519  //
1520  // If level is big then something went wrong -> exiting
1521  //
1522  // There could actually be 32 valid levels in the machine topology, but so
1523  // far, the only machine we have seen which does not exit this loop before
1524  // iteration 32 has fubar x2APIC settings.
1525  //
1526  // For now, just reject this case based upon loop trip count.
1527  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1528  return -1;
1529  }
1530  __kmp_x86_cpuid(11, level, &buf);
1531  if (buf.ebx == 0) {
1532  if (pkgLevel < 0) {
1533  // Will infer nPackages from __kmp_xproc
1534  pkgLevel = level;
1535  level++;
1536  }
1537  break;
1538  }
1539  int kind = (buf.ecx >> 8) & 0xff;
1540  if (kind == 1) {
1541  // SMT level
1542  threadLevel = level;
1543  coreLevel = -1;
1544  pkgLevel = -1;
1545  __kmp_nThreadsPerCore = buf.ebx & 0xffff;
1546  if (__kmp_nThreadsPerCore == 0) {
1547  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1548  return -1;
1549  }
1550  } else if (kind == 2) {
1551  // core level
1552  coreLevel = level;
1553  pkgLevel = -1;
1554  nCoresPerPkg = buf.ebx & 0xffff;
1555  if (nCoresPerPkg == 0) {
1556  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1557  return -1;
1558  }
1559  } else {
1560  if (level <= 0) {
1561  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1562  return -1;
1563  }
1564  if (pkgLevel >= 0) {
1565  continue;
1566  }
1567  pkgLevel = level;
1568  nPackages = buf.ebx & 0xffff;
1569  if (nPackages == 0) {
1570  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1571  return -1;
1572  }
1573  }
1574  }
1575  int depth = level;
1576 
1577  // In the above loop, "level" was counted from the finest level (usually
1578  // thread) to the coarsest. The caller expects that we will place the labels
1579  // in (*address2os)[].first.labels[] in the inverse order, so we need to
1580  // invert the vars saying which level means what.
1581  if (threadLevel >= 0) {
1582  threadLevel = depth - threadLevel - 1;
1583  }
1584  if (coreLevel >= 0) {
1585  coreLevel = depth - coreLevel - 1;
1586  }
1587  KMP_DEBUG_ASSERT(pkgLevel >= 0);
1588  pkgLevel = depth - pkgLevel - 1;
1589 
1590  // The algorithm used starts by setting the affinity to each available thread
1591  // and retrieving info from the cpuid instruction, so if we are not capable of
1592  // calling __kmp_get_system_affinity() and _kmp_get_system_affinity(), then we
1593  // need to do something else - use the defaults that we calculated from
1594  // issuing cpuid without binding to each proc.
1595  if (!KMP_AFFINITY_CAPABLE()) {
1596  // Hack to try and infer the machine topology using only the data
1597  // available from cpuid on the current thread, and __kmp_xproc.
1598  KMP_ASSERT(__kmp_affinity_type == affinity_none);
1599 
1600  __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
1601  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
1602  if (__kmp_affinity_verbose) {
1603  KMP_INFORM(AffNotCapableUseLocCpuidL11, "KMP_AFFINITY");
1604  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1605  if (__kmp_affinity_uniform_topology()) {
1606  KMP_INFORM(Uniform, "KMP_AFFINITY");
1607  } else {
1608  KMP_INFORM(NonUniform, "KMP_AFFINITY");
1609  }
1610  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
1611  __kmp_nThreadsPerCore, __kmp_ncores);
1612  }
1613  return 0;
1614  }
1615 
1616  // From here on, we can assume that it is safe to call
1617  // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
1618  // __kmp_affinity_type = affinity_none.
1619 
1620  // Save the affinity mask for the current thread.
1621  kmp_affin_mask_t *oldMask;
1622  KMP_CPU_ALLOC(oldMask);
1623  __kmp_get_system_affinity(oldMask, TRUE);
1624 
1625  // Allocate the data structure to be returned.
1626  AddrUnsPair *retval =
1627  (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * __kmp_avail_proc);
1628 
1629  // Run through each of the available contexts, binding the current thread
1630  // to it, and obtaining the pertinent information using the cpuid instr.
1631  unsigned int proc;
1632  int nApics = 0;
1633  KMP_CPU_SET_ITERATE(proc, __kmp_affin_fullMask) {
1634  // Skip this proc if it is not included in the machine model.
1635  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
1636  continue;
1637  }
1638  KMP_DEBUG_ASSERT(nApics < __kmp_avail_proc);
1639 
1640  __kmp_affinity_dispatch->bind_thread(proc);
1641 
1642  // Extract labels for each level in the machine topology map from Apic ID.
1643  Address addr(depth);
1644  int prev_shift = 0;
1645 
1646  for (level = 0; level < depth; level++) {
1647  __kmp_x86_cpuid(11, level, &buf);
1648  unsigned apicId = buf.edx;
1649  if (buf.ebx == 0) {
1650  if (level != depth - 1) {
1651  KMP_CPU_FREE(oldMask);
1652  *msg_id = kmp_i18n_str_InconsistentCpuidInfo;
1653  return -1;
1654  }
1655  addr.labels[depth - level - 1] = apicId >> prev_shift;
1656  level++;
1657  break;
1658  }
1659  int shift = buf.eax & 0x1f;
1660  int mask = (1 << shift) - 1;
1661  addr.labels[depth - level - 1] = (apicId & mask) >> prev_shift;
1662  prev_shift = shift;
1663  }
1664  if (level != depth) {
1665  KMP_CPU_FREE(oldMask);
1666  *msg_id = kmp_i18n_str_InconsistentCpuidInfo;
1667  return -1;
1668  }
1669 
1670  retval[nApics] = AddrUnsPair(addr, proc);
1671  nApics++;
1672  }
1673 
1674  // We've collected all the info we need.
1675  // Restore the old affinity mask for this thread.
1676  __kmp_set_system_affinity(oldMask, TRUE);
1677 
1678  // If there's only one thread context to bind to, return now.
1679  KMP_ASSERT(nApics > 0);
1680  if (nApics == 1) {
1681  __kmp_ncores = nPackages = 1;
1682  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
1683  if (__kmp_affinity_verbose) {
1684  char buf[KMP_AFFIN_MASK_PRINT_LEN];
1685  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
1686 
1687  KMP_INFORM(AffUseGlobCpuidL11, "KMP_AFFINITY");
1688  if (__kmp_affinity_respect_mask) {
1689  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
1690  } else {
1691  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
1692  }
1693  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1694  KMP_INFORM(Uniform, "KMP_AFFINITY");
1695  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
1696  __kmp_nThreadsPerCore, __kmp_ncores);
1697  }
1698 
1699  if (__kmp_affinity_type == affinity_none) {
1700  __kmp_free(retval);
1701  KMP_CPU_FREE(oldMask);
1702  return 0;
1703  }
1704 
1705  // Form an Address object which only includes the package level.
1706  Address addr(1);
1707  addr.labels[0] = retval[0].first.labels[pkgLevel];
1708  retval[0].first = addr;
1709 
1710  if (__kmp_affinity_gran_levels < 0) {
1711  __kmp_affinity_gran_levels = 0;
1712  }
1713 
1714  if (__kmp_affinity_verbose) {
1715  __kmp_affinity_print_topology(retval, 1, 1, 0, -1, -1);
1716  }
1717 
1718  *address2os = retval;
1719  KMP_CPU_FREE(oldMask);
1720  return 1;
1721  }
1722 
1723  // Sort the table by physical Id.
1724  qsort(retval, nApics, sizeof(*retval), __kmp_affinity_cmp_Address_labels);
1725 
1726  // Find the radix at each of the levels.
1727  unsigned *totals = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
1728  unsigned *counts = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
1729  unsigned *maxCt = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
1730  unsigned *last = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
1731  for (level = 0; level < depth; level++) {
1732  totals[level] = 1;
1733  maxCt[level] = 1;
1734  counts[level] = 1;
1735  last[level] = retval[0].first.labels[level];
1736  }
1737 
1738  // From here on, the iteration variable "level" runs from the finest level to
1739  // the coarsest, i.e. we iterate forward through
1740  // (*address2os)[].first.labels[] - in the previous loops, we iterated
1741  // backwards.
1742  for (proc = 1; (int)proc < nApics; proc++) {
1743  int level;
1744  for (level = 0; level < depth; level++) {
1745  if (retval[proc].first.labels[level] != last[level]) {
1746  int j;
1747  for (j = level + 1; j < depth; j++) {
1748  totals[j]++;
1749  counts[j] = 1;
1750  // The line below causes printing incorrect topology information in
1751  // case the max value for some level (maxCt[level]) is encountered
1752  // earlier than some less value while going through the array. For
1753  // example, let pkg0 has 4 cores and pkg1 has 2 cores. Then
1754  // maxCt[1] == 2
1755  // whereas it must be 4.
1756  // TODO!!! Check if it can be commented safely
1757  // maxCt[j] = 1;
1758  last[j] = retval[proc].first.labels[j];
1759  }
1760  totals[level]++;
1761  counts[level]++;
1762  if (counts[level] > maxCt[level]) {
1763  maxCt[level] = counts[level];
1764  }
1765  last[level] = retval[proc].first.labels[level];
1766  break;
1767  } else if (level == depth - 1) {
1768  __kmp_free(last);
1769  __kmp_free(maxCt);
1770  __kmp_free(counts);
1771  __kmp_free(totals);
1772  __kmp_free(retval);
1773  KMP_CPU_FREE(oldMask);
1774  *msg_id = kmp_i18n_str_x2ApicIDsNotUnique;
1775  return -1;
1776  }
1777  }
1778  }
1779 
1780  // When affinity is off, this routine will still be called to set
1781  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
1782  // Make sure all these vars are set correctly, and return if affinity is not
1783  // enabled.
1784  if (threadLevel >= 0) {
1785  __kmp_nThreadsPerCore = maxCt[threadLevel];
1786  } else {
1787  __kmp_nThreadsPerCore = 1;
1788  }
1789  nPackages = totals[pkgLevel];
1790 
1791  if (coreLevel >= 0) {
1792  __kmp_ncores = totals[coreLevel];
1793  nCoresPerPkg = maxCt[coreLevel];
1794  } else {
1795  __kmp_ncores = nPackages;
1796  nCoresPerPkg = 1;
1797  }
1798 
1799  // Check to see if the machine topology is uniform
1800  unsigned prod = maxCt[0];
1801  for (level = 1; level < depth; level++) {
1802  prod *= maxCt[level];
1803  }
1804  bool uniform = (prod == totals[level - 1]);
1805 
1806  // Print the machine topology summary.
1807  if (__kmp_affinity_verbose) {
1808  char mask[KMP_AFFIN_MASK_PRINT_LEN];
1809  __kmp_affinity_print_mask(mask, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
1810 
1811  KMP_INFORM(AffUseGlobCpuidL11, "KMP_AFFINITY");
1812  if (__kmp_affinity_respect_mask) {
1813  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", mask);
1814  } else {
1815  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", mask);
1816  }
1817  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1818  if (uniform) {
1819  KMP_INFORM(Uniform, "KMP_AFFINITY");
1820  } else {
1821  KMP_INFORM(NonUniform, "KMP_AFFINITY");
1822  }
1823 
1824  kmp_str_buf_t buf;
1825  __kmp_str_buf_init(&buf);
1826 
1827  __kmp_str_buf_print(&buf, "%d", totals[0]);
1828  for (level = 1; level <= pkgLevel; level++) {
1829  __kmp_str_buf_print(&buf, " x %d", maxCt[level]);
1830  }
1831  KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, nCoresPerPkg,
1832  __kmp_nThreadsPerCore, __kmp_ncores);
1833 
1834  __kmp_str_buf_free(&buf);
1835  }
1836  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
1837  KMP_DEBUG_ASSERT(nApics == __kmp_avail_proc);
1838  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
1839  for (proc = 0; (int)proc < nApics; ++proc) {
1840  __kmp_pu_os_idx[proc] = retval[proc].second;
1841  }
1842  if (__kmp_affinity_type == affinity_none) {
1843  __kmp_free(last);
1844  __kmp_free(maxCt);
1845  __kmp_free(counts);
1846  __kmp_free(totals);
1847  __kmp_free(retval);
1848  KMP_CPU_FREE(oldMask);
1849  return 0;
1850  }
1851 
1852  // Find any levels with radiix 1, and remove them from the map
1853  // (except for the package level).
1854  int new_depth = 0;
1855  for (level = 0; level < depth; level++) {
1856  if ((maxCt[level] == 1) && (level != pkgLevel)) {
1857  continue;
1858  }
1859  new_depth++;
1860  }
1861 
1862  // If we are removing any levels, allocate a new vector to return,
1863  // and copy the relevant information to it.
1864  if (new_depth != depth) {
1865  AddrUnsPair *new_retval =
1866  (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * nApics);
1867  for (proc = 0; (int)proc < nApics; proc++) {
1868  Address addr(new_depth);
1869  new_retval[proc] = AddrUnsPair(addr, retval[proc].second);
1870  }
1871  int new_level = 0;
1872  int newPkgLevel = -1;
1873  int newCoreLevel = -1;
1874  int newThreadLevel = -1;
1875  for (level = 0; level < depth; level++) {
1876  if ((maxCt[level] == 1) && (level != pkgLevel)) {
1877  // Remove this level. Never remove the package level
1878  continue;
1879  }
1880  if (level == pkgLevel) {
1881  newPkgLevel = new_level;
1882  }
1883  if (level == coreLevel) {
1884  newCoreLevel = new_level;
1885  }
1886  if (level == threadLevel) {
1887  newThreadLevel = new_level;
1888  }
1889  for (proc = 0; (int)proc < nApics; proc++) {
1890  new_retval[proc].first.labels[new_level] =
1891  retval[proc].first.labels[level];
1892  }
1893  new_level++;
1894  }
1895 
1896  __kmp_free(retval);
1897  retval = new_retval;
1898  depth = new_depth;
1899  pkgLevel = newPkgLevel;
1900  coreLevel = newCoreLevel;
1901  threadLevel = newThreadLevel;
1902  }
1903 
1904  if (__kmp_affinity_gran_levels < 0) {
1905  // Set the granularity level based on what levels are modeled
1906  // in the machine topology map.
1907  __kmp_affinity_gran_levels = 0;
1908  if ((threadLevel >= 0) && (__kmp_affinity_gran > affinity_gran_thread)) {
1909  __kmp_affinity_gran_levels++;
1910  }
1911  if ((coreLevel >= 0) && (__kmp_affinity_gran > affinity_gran_core)) {
1912  __kmp_affinity_gran_levels++;
1913  }
1914  if (__kmp_affinity_gran > affinity_gran_package) {
1915  __kmp_affinity_gran_levels++;
1916  }
1917  }
1918 
1919  if (__kmp_affinity_verbose) {
1920  __kmp_affinity_print_topology(retval, nApics, depth, pkgLevel, coreLevel,
1921  threadLevel);
1922  }
1923 
1924  __kmp_free(last);
1925  __kmp_free(maxCt);
1926  __kmp_free(counts);
1927  __kmp_free(totals);
1928  KMP_CPU_FREE(oldMask);
1929  *address2os = retval;
1930  return depth;
1931 }
1932 
1933 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
1934 
1935 #define osIdIndex 0
1936 #define threadIdIndex 1
1937 #define coreIdIndex 2
1938 #define pkgIdIndex 3
1939 #define nodeIdIndex 4
1940 
1941 typedef unsigned *ProcCpuInfo;
1942 static unsigned maxIndex = pkgIdIndex;
1943 
1944 static int __kmp_affinity_cmp_ProcCpuInfo_phys_id(const void *a,
1945  const void *b) {
1946  unsigned i;
1947  const unsigned *aa = *(unsigned *const *)a;
1948  const unsigned *bb = *(unsigned *const *)b;
1949  for (i = maxIndex;; i--) {
1950  if (aa[i] < bb[i])
1951  return -1;
1952  if (aa[i] > bb[i])
1953  return 1;
1954  if (i == osIdIndex)
1955  break;
1956  }
1957  return 0;
1958 }
1959 
1960 #if KMP_USE_HIER_SCHED
1961 // Set the array sizes for the hierarchy layers
1962 static void __kmp_dispatch_set_hierarchy_values() {
1963  // Set the maximum number of L1's to number of cores
1964  // Set the maximum number of L2's to to either number of cores / 2 for
1965  // Intel(R) Xeon Phi(TM) coprocessor formally codenamed Knights Landing
1966  // Or the number of cores for Intel(R) Xeon(R) processors
1967  // Set the maximum number of NUMA nodes and L3's to number of packages
1968  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1] =
1969  nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
1970  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L1 + 1] = __kmp_ncores;
1971 #if KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
1972  if (__kmp_mic_type >= mic3)
1973  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores / 2;
1974  else
1975 #endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
1976  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores;
1977  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L3 + 1] = nPackages;
1978  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_NUMA + 1] = nPackages;
1979  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_LOOP + 1] = 1;
1980  // Set the number of threads per unit
1981  // Number of hardware threads per L1/L2/L3/NUMA/LOOP
1982  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_THREAD + 1] = 1;
1983  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L1 + 1] =
1984  __kmp_nThreadsPerCore;
1985 #if KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
1986  if (__kmp_mic_type >= mic3)
1987  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
1988  2 * __kmp_nThreadsPerCore;
1989  else
1990 #endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
1991  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
1992  __kmp_nThreadsPerCore;
1993  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L3 + 1] =
1994  nCoresPerPkg * __kmp_nThreadsPerCore;
1995  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_NUMA + 1] =
1996  nCoresPerPkg * __kmp_nThreadsPerCore;
1997  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_LOOP + 1] =
1998  nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
1999 }
2000 
2001 // Return the index into the hierarchy for this tid and layer type (L1, L2, etc)
2002 // i.e., this thread's L1 or this thread's L2, etc.
2003 int __kmp_dispatch_get_index(int tid, kmp_hier_layer_e type) {
2004  int index = type + 1;
2005  int num_hw_threads = __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1];
2006  KMP_DEBUG_ASSERT(type != kmp_hier_layer_e::LAYER_LAST);
2007  if (type == kmp_hier_layer_e::LAYER_THREAD)
2008  return tid;
2009  else if (type == kmp_hier_layer_e::LAYER_LOOP)
2010  return 0;
2011  KMP_DEBUG_ASSERT(__kmp_hier_max_units[index] != 0);
2012  if (tid >= num_hw_threads)
2013  tid = tid % num_hw_threads;
2014  return (tid / __kmp_hier_threads_per[index]) % __kmp_hier_max_units[index];
2015 }
2016 
2017 // Return the number of t1's per t2
2018 int __kmp_dispatch_get_t1_per_t2(kmp_hier_layer_e t1, kmp_hier_layer_e t2) {
2019  int i1 = t1 + 1;
2020  int i2 = t2 + 1;
2021  KMP_DEBUG_ASSERT(i1 <= i2);
2022  KMP_DEBUG_ASSERT(t1 != kmp_hier_layer_e::LAYER_LAST);
2023  KMP_DEBUG_ASSERT(t2 != kmp_hier_layer_e::LAYER_LAST);
2024  KMP_DEBUG_ASSERT(__kmp_hier_threads_per[i1] != 0);
2025  // (nthreads/t2) / (nthreads/t1) = t1 / t2
2026  return __kmp_hier_threads_per[i2] / __kmp_hier_threads_per[i1];
2027 }
2028 #endif // KMP_USE_HIER_SCHED
2029 
2030 // Parse /proc/cpuinfo (or an alternate file in the same format) to obtain the
2031 // affinity map.
2032 static int __kmp_affinity_create_cpuinfo_map(AddrUnsPair **address2os,
2033  int *line,
2034  kmp_i18n_id_t *const msg_id,
2035  FILE *f) {
2036  *address2os = NULL;
2037  *msg_id = kmp_i18n_null;
2038 
2039  // Scan of the file, and count the number of "processor" (osId) fields,
2040  // and find the highest value of <n> for a node_<n> field.
2041  char buf[256];
2042  unsigned num_records = 0;
2043  while (!feof(f)) {
2044  buf[sizeof(buf) - 1] = 1;
2045  if (!fgets(buf, sizeof(buf), f)) {
2046  // Read errors presumably because of EOF
2047  break;
2048  }
2049 
2050  char s1[] = "processor";
2051  if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
2052  num_records++;
2053  continue;
2054  }
2055 
2056  // FIXME - this will match "node_<n> <garbage>"
2057  unsigned level;
2058  if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
2059  if (nodeIdIndex + level >= maxIndex) {
2060  maxIndex = nodeIdIndex + level;
2061  }
2062  continue;
2063  }
2064  }
2065 
2066  // Check for empty file / no valid processor records, or too many. The number
2067  // of records can't exceed the number of valid bits in the affinity mask.
2068  if (num_records == 0) {
2069  *line = 0;
2070  *msg_id = kmp_i18n_str_NoProcRecords;
2071  return -1;
2072  }
2073  if (num_records > (unsigned)__kmp_xproc) {
2074  *line = 0;
2075  *msg_id = kmp_i18n_str_TooManyProcRecords;
2076  return -1;
2077  }
2078 
2079  // Set the file pointer back to the begginning, so that we can scan the file
2080  // again, this time performing a full parse of the data. Allocate a vector of
2081  // ProcCpuInfo object, where we will place the data. Adding an extra element
2082  // at the end allows us to remove a lot of extra checks for termination
2083  // conditions.
2084  if (fseek(f, 0, SEEK_SET) != 0) {
2085  *line = 0;
2086  *msg_id = kmp_i18n_str_CantRewindCpuinfo;
2087  return -1;
2088  }
2089 
2090  // Allocate the array of records to store the proc info in. The dummy
2091  // element at the end makes the logic in filling them out easier to code.
2092  unsigned **threadInfo =
2093  (unsigned **)__kmp_allocate((num_records + 1) * sizeof(unsigned *));
2094  unsigned i;
2095  for (i = 0; i <= num_records; i++) {
2096  threadInfo[i] =
2097  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2098  }
2099 
2100 #define CLEANUP_THREAD_INFO \
2101  for (i = 0; i <= num_records; i++) { \
2102  __kmp_free(threadInfo[i]); \
2103  } \
2104  __kmp_free(threadInfo);
2105 
2106  // A value of UINT_MAX means that we didn't find the field
2107  unsigned __index;
2108 
2109 #define INIT_PROC_INFO(p) \
2110  for (__index = 0; __index <= maxIndex; __index++) { \
2111  (p)[__index] = UINT_MAX; \
2112  }
2113 
2114  for (i = 0; i <= num_records; i++) {
2115  INIT_PROC_INFO(threadInfo[i]);
2116  }
2117 
2118  unsigned num_avail = 0;
2119  *line = 0;
2120  while (!feof(f)) {
2121  // Create an inner scoping level, so that all the goto targets at the end of
2122  // the loop appear in an outer scoping level. This avoids warnings about
2123  // jumping past an initialization to a target in the same block.
2124  {
2125  buf[sizeof(buf) - 1] = 1;
2126  bool long_line = false;
2127  if (!fgets(buf, sizeof(buf), f)) {
2128  // Read errors presumably because of EOF
2129  // If there is valid data in threadInfo[num_avail], then fake
2130  // a blank line in ensure that the last address gets parsed.
2131  bool valid = false;
2132  for (i = 0; i <= maxIndex; i++) {
2133  if (threadInfo[num_avail][i] != UINT_MAX) {
2134  valid = true;
2135  }
2136  }
2137  if (!valid) {
2138  break;
2139  }
2140  buf[0] = 0;
2141  } else if (!buf[sizeof(buf) - 1]) {
2142  // The line is longer than the buffer. Set a flag and don't
2143  // emit an error if we were going to ignore the line, anyway.
2144  long_line = true;
2145 
2146 #define CHECK_LINE \
2147  if (long_line) { \
2148  CLEANUP_THREAD_INFO; \
2149  *msg_id = kmp_i18n_str_LongLineCpuinfo; \
2150  return -1; \
2151  }
2152  }
2153  (*line)++;
2154 
2155  char s1[] = "processor";
2156  if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
2157  CHECK_LINE;
2158  char *p = strchr(buf + sizeof(s1) - 1, ':');
2159  unsigned val;
2160  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2161  goto no_val;
2162  if (threadInfo[num_avail][osIdIndex] != UINT_MAX)
2163 #if KMP_ARCH_AARCH64
2164  // Handle the old AArch64 /proc/cpuinfo layout differently,
2165  // it contains all of the 'processor' entries listed in a
2166  // single 'Processor' section, therefore the normal looking
2167  // for duplicates in that section will always fail.
2168  num_avail++;
2169 #else
2170  goto dup_field;
2171 #endif
2172  threadInfo[num_avail][osIdIndex] = val;
2173 #if KMP_OS_LINUX && !(KMP_ARCH_X86 || KMP_ARCH_X86_64)
2174  char path[256];
2175  KMP_SNPRINTF(
2176  path, sizeof(path),
2177  "/sys/devices/system/cpu/cpu%u/topology/physical_package_id",
2178  threadInfo[num_avail][osIdIndex]);
2179  __kmp_read_from_file(path, "%u", &threadInfo[num_avail][pkgIdIndex]);
2180 
2181  KMP_SNPRINTF(path, sizeof(path),
2182  "/sys/devices/system/cpu/cpu%u/topology/core_id",
2183  threadInfo[num_avail][osIdIndex]);
2184  __kmp_read_from_file(path, "%u", &threadInfo[num_avail][coreIdIndex]);
2185  continue;
2186 #else
2187  }
2188  char s2[] = "physical id";
2189  if (strncmp(buf, s2, sizeof(s2) - 1) == 0) {
2190  CHECK_LINE;
2191  char *p = strchr(buf + sizeof(s2) - 1, ':');
2192  unsigned val;
2193  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2194  goto no_val;
2195  if (threadInfo[num_avail][pkgIdIndex] != UINT_MAX)
2196  goto dup_field;
2197  threadInfo[num_avail][pkgIdIndex] = val;
2198  continue;
2199  }
2200  char s3[] = "core id";
2201  if (strncmp(buf, s3, sizeof(s3) - 1) == 0) {
2202  CHECK_LINE;
2203  char *p = strchr(buf + sizeof(s3) - 1, ':');
2204  unsigned val;
2205  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2206  goto no_val;
2207  if (threadInfo[num_avail][coreIdIndex] != UINT_MAX)
2208  goto dup_field;
2209  threadInfo[num_avail][coreIdIndex] = val;
2210  continue;
2211 #endif // KMP_OS_LINUX && USE_SYSFS_INFO
2212  }
2213  char s4[] = "thread id";
2214  if (strncmp(buf, s4, sizeof(s4) - 1) == 0) {
2215  CHECK_LINE;
2216  char *p = strchr(buf + sizeof(s4) - 1, ':');
2217  unsigned val;
2218  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2219  goto no_val;
2220  if (threadInfo[num_avail][threadIdIndex] != UINT_MAX)
2221  goto dup_field;
2222  threadInfo[num_avail][threadIdIndex] = val;
2223  continue;
2224  }
2225  unsigned level;
2226  if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
2227  CHECK_LINE;
2228  char *p = strchr(buf + sizeof(s4) - 1, ':');
2229  unsigned val;
2230  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2231  goto no_val;
2232  KMP_ASSERT(nodeIdIndex + level <= maxIndex);
2233  if (threadInfo[num_avail][nodeIdIndex + level] != UINT_MAX)
2234  goto dup_field;
2235  threadInfo[num_avail][nodeIdIndex + level] = val;
2236  continue;
2237  }
2238 
2239  // We didn't recognize the leading token on the line. There are lots of
2240  // leading tokens that we don't recognize - if the line isn't empty, go on
2241  // to the next line.
2242  if ((*buf != 0) && (*buf != '\n')) {
2243  // If the line is longer than the buffer, read characters
2244  // until we find a newline.
2245  if (long_line) {
2246  int ch;
2247  while (((ch = fgetc(f)) != EOF) && (ch != '\n'))
2248  ;
2249  }
2250  continue;
2251  }
2252 
2253  // A newline has signalled the end of the processor record.
2254  // Check that there aren't too many procs specified.
2255  if ((int)num_avail == __kmp_xproc) {
2256  CLEANUP_THREAD_INFO;
2257  *msg_id = kmp_i18n_str_TooManyEntries;
2258  return -1;
2259  }
2260 
2261  // Check for missing fields. The osId field must be there, and we
2262  // currently require that the physical id field is specified, also.
2263  if (threadInfo[num_avail][osIdIndex] == UINT_MAX) {
2264  CLEANUP_THREAD_INFO;
2265  *msg_id = kmp_i18n_str_MissingProcField;
2266  return -1;
2267  }
2268  if (threadInfo[0][pkgIdIndex] == UINT_MAX) {
2269  CLEANUP_THREAD_INFO;
2270  *msg_id = kmp_i18n_str_MissingPhysicalIDField;
2271  return -1;
2272  }
2273 
2274  // Skip this proc if it is not included in the machine model.
2275  if (!KMP_CPU_ISSET(threadInfo[num_avail][osIdIndex],
2276  __kmp_affin_fullMask)) {
2277  INIT_PROC_INFO(threadInfo[num_avail]);
2278  continue;
2279  }
2280 
2281  // We have a successful parse of this proc's info.
2282  // Increment the counter, and prepare for the next proc.
2283  num_avail++;
2284  KMP_ASSERT(num_avail <= num_records);
2285  INIT_PROC_INFO(threadInfo[num_avail]);
2286  }
2287  continue;
2288 
2289  no_val:
2290  CLEANUP_THREAD_INFO;
2291  *msg_id = kmp_i18n_str_MissingValCpuinfo;
2292  return -1;
2293 
2294  dup_field:
2295  CLEANUP_THREAD_INFO;
2296  *msg_id = kmp_i18n_str_DuplicateFieldCpuinfo;
2297  return -1;
2298  }
2299  *line = 0;
2300 
2301 #if KMP_MIC && REDUCE_TEAM_SIZE
2302  unsigned teamSize = 0;
2303 #endif // KMP_MIC && REDUCE_TEAM_SIZE
2304 
2305  // check for num_records == __kmp_xproc ???
2306 
2307  // If there's only one thread context to bind to, form an Address object with
2308  // depth 1 and return immediately (or, if affinity is off, set address2os to
2309  // NULL and return).
2310  //
2311  // If it is configured to omit the package level when there is only a single
2312  // package, the logic at the end of this routine won't work if there is only a
2313  // single thread - it would try to form an Address object with depth 0.
2314  KMP_ASSERT(num_avail > 0);
2315  KMP_ASSERT(num_avail <= num_records);
2316  if (num_avail == 1) {
2317  __kmp_ncores = 1;
2318  __kmp_nThreadsPerCore = nCoresPerPkg = nPackages = 1;
2319  if (__kmp_affinity_verbose) {
2320  if (!KMP_AFFINITY_CAPABLE()) {
2321  KMP_INFORM(AffNotCapableUseCpuinfo, "KMP_AFFINITY");
2322  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
2323  KMP_INFORM(Uniform, "KMP_AFFINITY");
2324  } else {
2325  char buf[KMP_AFFIN_MASK_PRINT_LEN];
2326  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
2327  __kmp_affin_fullMask);
2328  KMP_INFORM(AffCapableUseCpuinfo, "KMP_AFFINITY");
2329  if (__kmp_affinity_respect_mask) {
2330  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
2331  } else {
2332  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
2333  }
2334  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
2335  KMP_INFORM(Uniform, "KMP_AFFINITY");
2336  }
2337  int index;
2338  kmp_str_buf_t buf;
2339  __kmp_str_buf_init(&buf);
2340  __kmp_str_buf_print(&buf, "1");
2341  for (index = maxIndex - 1; index > pkgIdIndex; index--) {
2342  __kmp_str_buf_print(&buf, " x 1");
2343  }
2344  KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, 1, 1, 1);
2345  __kmp_str_buf_free(&buf);
2346  }
2347 
2348  if (__kmp_affinity_type == affinity_none) {
2349  CLEANUP_THREAD_INFO;
2350  return 0;
2351  }
2352 
2353  *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair));
2354  Address addr(1);
2355  addr.labels[0] = threadInfo[0][pkgIdIndex];
2356  (*address2os)[0] = AddrUnsPair(addr, threadInfo[0][osIdIndex]);
2357 
2358  if (__kmp_affinity_gran_levels < 0) {
2359  __kmp_affinity_gran_levels = 0;
2360  }
2361 
2362  if (__kmp_affinity_verbose) {
2363  __kmp_affinity_print_topology(*address2os, 1, 1, 0, -1, -1);
2364  }
2365 
2366  CLEANUP_THREAD_INFO;
2367  return 1;
2368  }
2369 
2370  // Sort the threadInfo table by physical Id.
2371  qsort(threadInfo, num_avail, sizeof(*threadInfo),
2372  __kmp_affinity_cmp_ProcCpuInfo_phys_id);
2373 
2374  // The table is now sorted by pkgId / coreId / threadId, but we really don't
2375  // know the radix of any of the fields. pkgId's may be sparsely assigned among
2376  // the chips on a system. Although coreId's are usually assigned
2377  // [0 .. coresPerPkg-1] and threadId's are usually assigned
2378  // [0..threadsPerCore-1], we don't want to make any such assumptions.
2379  //
2380  // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
2381  // total # packages) are at this point - we want to determine that now. We
2382  // only have an upper bound on the first two figures.
2383  unsigned *counts =
2384  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2385  unsigned *maxCt =
2386  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2387  unsigned *totals =
2388  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2389  unsigned *lastId =
2390  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2391 
2392  bool assign_thread_ids = false;
2393  unsigned threadIdCt;
2394  unsigned index;
2395 
2396 restart_radix_check:
2397  threadIdCt = 0;
2398 
2399  // Initialize the counter arrays with data from threadInfo[0].
2400  if (assign_thread_ids) {
2401  if (threadInfo[0][threadIdIndex] == UINT_MAX) {
2402  threadInfo[0][threadIdIndex] = threadIdCt++;
2403  } else if (threadIdCt <= threadInfo[0][threadIdIndex]) {
2404  threadIdCt = threadInfo[0][threadIdIndex] + 1;
2405  }
2406  }
2407  for (index = 0; index <= maxIndex; index++) {
2408  counts[index] = 1;
2409  maxCt[index] = 1;
2410  totals[index] = 1;
2411  lastId[index] = threadInfo[0][index];
2412  ;
2413  }
2414 
2415  // Run through the rest of the OS procs.
2416  for (i = 1; i < num_avail; i++) {
2417  // Find the most significant index whose id differs from the id for the
2418  // previous OS proc.
2419  for (index = maxIndex; index >= threadIdIndex; index--) {
2420  if (assign_thread_ids && (index == threadIdIndex)) {
2421  // Auto-assign the thread id field if it wasn't specified.
2422  if (threadInfo[i][threadIdIndex] == UINT_MAX) {
2423  threadInfo[i][threadIdIndex] = threadIdCt++;
2424  }
2425  // Apparently the thread id field was specified for some entries and not
2426  // others. Start the thread id counter off at the next higher thread id.
2427  else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
2428  threadIdCt = threadInfo[i][threadIdIndex] + 1;
2429  }
2430  }
2431  if (threadInfo[i][index] != lastId[index]) {
2432  // Run through all indices which are less significant, and reset the
2433  // counts to 1. At all levels up to and including index, we need to
2434  // increment the totals and record the last id.
2435  unsigned index2;
2436  for (index2 = threadIdIndex; index2 < index; index2++) {
2437  totals[index2]++;
2438  if (counts[index2] > maxCt[index2]) {
2439  maxCt[index2] = counts[index2];
2440  }
2441  counts[index2] = 1;
2442  lastId[index2] = threadInfo[i][index2];
2443  }
2444  counts[index]++;
2445  totals[index]++;
2446  lastId[index] = threadInfo[i][index];
2447 
2448  if (assign_thread_ids && (index > threadIdIndex)) {
2449 
2450 #if KMP_MIC && REDUCE_TEAM_SIZE
2451  // The default team size is the total #threads in the machine
2452  // minus 1 thread for every core that has 3 or more threads.
2453  teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
2454 #endif // KMP_MIC && REDUCE_TEAM_SIZE
2455 
2456  // Restart the thread counter, as we are on a new core.
2457  threadIdCt = 0;
2458 
2459  // Auto-assign the thread id field if it wasn't specified.
2460  if (threadInfo[i][threadIdIndex] == UINT_MAX) {
2461  threadInfo[i][threadIdIndex] = threadIdCt++;
2462  }
2463 
2464  // Aparrently the thread id field was specified for some entries and
2465  // not others. Start the thread id counter off at the next higher
2466  // thread id.
2467  else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
2468  threadIdCt = threadInfo[i][threadIdIndex] + 1;
2469  }
2470  }
2471  break;
2472  }
2473  }
2474  if (index < threadIdIndex) {
2475  // If thread ids were specified, it is an error if they are not unique.
2476  // Also, check that we waven't already restarted the loop (to be safe -
2477  // shouldn't need to).
2478  if ((threadInfo[i][threadIdIndex] != UINT_MAX) || assign_thread_ids) {
2479  __kmp_free(lastId);
2480  __kmp_free(totals);
2481  __kmp_free(maxCt);
2482  __kmp_free(counts);
2483  CLEANUP_THREAD_INFO;
2484  *msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
2485  return -1;
2486  }
2487 
2488  // If the thread ids were not specified and we see entries entries that
2489  // are duplicates, start the loop over and assign the thread ids manually.
2490  assign_thread_ids = true;
2491  goto restart_radix_check;
2492  }
2493  }
2494 
2495 #if KMP_MIC && REDUCE_TEAM_SIZE
2496  // The default team size is the total #threads in the machine
2497  // minus 1 thread for every core that has 3 or more threads.
2498  teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
2499 #endif // KMP_MIC && REDUCE_TEAM_SIZE
2500 
2501  for (index = threadIdIndex; index <= maxIndex; index++) {
2502  if (counts[index] > maxCt[index]) {
2503  maxCt[index] = counts[index];
2504  }
2505  }
2506 
2507  __kmp_nThreadsPerCore = maxCt[threadIdIndex];
2508  nCoresPerPkg = maxCt[coreIdIndex];
2509  nPackages = totals[pkgIdIndex];
2510 
2511  // Check to see if the machine topology is uniform
2512  unsigned prod = totals[maxIndex];
2513  for (index = threadIdIndex; index < maxIndex; index++) {
2514  prod *= maxCt[index];
2515  }
2516  bool uniform = (prod == totals[threadIdIndex]);
2517 
2518  // When affinity is off, this routine will still be called to set
2519  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
2520  // Make sure all these vars are set correctly, and return now if affinity is
2521  // not enabled.
2522  __kmp_ncores = totals[coreIdIndex];
2523 
2524  if (__kmp_affinity_verbose) {
2525  if (!KMP_AFFINITY_CAPABLE()) {
2526  KMP_INFORM(AffNotCapableUseCpuinfo, "KMP_AFFINITY");
2527  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
2528  if (uniform) {
2529  KMP_INFORM(Uniform, "KMP_AFFINITY");
2530  } else {
2531  KMP_INFORM(NonUniform, "KMP_AFFINITY");
2532  }
2533  } else {
2534  char buf[KMP_AFFIN_MASK_PRINT_LEN];
2535  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
2536  __kmp_affin_fullMask);
2537  KMP_INFORM(AffCapableUseCpuinfo, "KMP_AFFINITY");
2538  if (__kmp_affinity_respect_mask) {
2539  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
2540  } else {
2541  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
2542  }
2543  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
2544  if (uniform) {
2545  KMP_INFORM(Uniform, "KMP_AFFINITY");
2546  } else {
2547  KMP_INFORM(NonUniform, "KMP_AFFINITY");
2548  }
2549  }
2550  kmp_str_buf_t buf;
2551  __kmp_str_buf_init(&buf);
2552 
2553  __kmp_str_buf_print(&buf, "%d", totals[maxIndex]);
2554  for (index = maxIndex - 1; index >= pkgIdIndex; index--) {
2555  __kmp_str_buf_print(&buf, " x %d", maxCt[index]);
2556  }
2557  KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, maxCt[coreIdIndex],
2558  maxCt[threadIdIndex], __kmp_ncores);
2559 
2560  __kmp_str_buf_free(&buf);
2561  }
2562 
2563 #if KMP_MIC && REDUCE_TEAM_SIZE
2564  // Set the default team size.
2565  if ((__kmp_dflt_team_nth == 0) && (teamSize > 0)) {
2566  __kmp_dflt_team_nth = teamSize;
2567  KA_TRACE(20, ("__kmp_affinity_create_cpuinfo_map: setting "
2568  "__kmp_dflt_team_nth = %d\n",
2569  __kmp_dflt_team_nth));
2570  }
2571 #endif // KMP_MIC && REDUCE_TEAM_SIZE
2572 
2573  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
2574  KMP_DEBUG_ASSERT(num_avail == (unsigned)__kmp_avail_proc);
2575  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
2576  for (i = 0; i < num_avail; ++i) { // fill the os indices
2577  __kmp_pu_os_idx[i] = threadInfo[i][osIdIndex];
2578  }
2579 
2580  if (__kmp_affinity_type == affinity_none) {
2581  __kmp_free(lastId);
2582  __kmp_free(totals);
2583  __kmp_free(maxCt);
2584  __kmp_free(counts);
2585  CLEANUP_THREAD_INFO;
2586  return 0;
2587  }
2588 
2589  // Count the number of levels which have more nodes at that level than at the
2590  // parent's level (with there being an implicit root node of the top level).
2591  // This is equivalent to saying that there is at least one node at this level
2592  // which has a sibling. These levels are in the map, and the package level is
2593  // always in the map.
2594  bool *inMap = (bool *)__kmp_allocate((maxIndex + 1) * sizeof(bool));
2595  for (index = threadIdIndex; index < maxIndex; index++) {
2596  KMP_ASSERT(totals[index] >= totals[index + 1]);
2597  inMap[index] = (totals[index] > totals[index + 1]);
2598  }
2599  inMap[maxIndex] = (totals[maxIndex] > 1);
2600  inMap[pkgIdIndex] = true;
2601 
2602  int depth = 0;
2603  for (index = threadIdIndex; index <= maxIndex; index++) {
2604  if (inMap[index]) {
2605  depth++;
2606  }
2607  }
2608  KMP_ASSERT(depth > 0);
2609 
2610  // Construct the data structure that is to be returned.
2611  *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * num_avail);
2612  int pkgLevel = -1;
2613  int coreLevel = -1;
2614  int threadLevel = -1;
2615 
2616  for (i = 0; i < num_avail; ++i) {
2617  Address addr(depth);
2618  unsigned os = threadInfo[i][osIdIndex];
2619  int src_index;
2620  int dst_index = 0;
2621 
2622  for (src_index = maxIndex; src_index >= threadIdIndex; src_index--) {
2623  if (!inMap[src_index]) {
2624  continue;
2625  }
2626  addr.labels[dst_index] = threadInfo[i][src_index];
2627  if (src_index == pkgIdIndex) {
2628  pkgLevel = dst_index;
2629  } else if (src_index == coreIdIndex) {
2630  coreLevel = dst_index;
2631  } else if (src_index == threadIdIndex) {
2632  threadLevel = dst_index;
2633  }
2634  dst_index++;
2635  }
2636  (*address2os)[i] = AddrUnsPair(addr, os);
2637  }
2638 
2639  if (__kmp_affinity_gran_levels < 0) {
2640  // Set the granularity level based on what levels are modeled
2641  // in the machine topology map.
2642  unsigned src_index;
2643  __kmp_affinity_gran_levels = 0;
2644  for (src_index = threadIdIndex; src_index <= maxIndex; src_index++) {
2645  if (!inMap[src_index]) {
2646  continue;
2647  }
2648  switch (src_index) {
2649  case threadIdIndex:
2650  if (__kmp_affinity_gran > affinity_gran_thread) {
2651  __kmp_affinity_gran_levels++;
2652  }
2653 
2654  break;
2655  case coreIdIndex:
2656  if (__kmp_affinity_gran > affinity_gran_core) {
2657  __kmp_affinity_gran_levels++;
2658  }
2659  break;
2660 
2661  case pkgIdIndex:
2662  if (__kmp_affinity_gran > affinity_gran_package) {
2663  __kmp_affinity_gran_levels++;
2664  }
2665  break;
2666  }
2667  }
2668  }
2669 
2670  if (__kmp_affinity_verbose) {
2671  __kmp_affinity_print_topology(*address2os, num_avail, depth, pkgLevel,
2672  coreLevel, threadLevel);
2673  }
2674 
2675  __kmp_free(inMap);
2676  __kmp_free(lastId);
2677  __kmp_free(totals);
2678  __kmp_free(maxCt);
2679  __kmp_free(counts);
2680  CLEANUP_THREAD_INFO;
2681  return depth;
2682 }
2683 
2684 // Create and return a table of affinity masks, indexed by OS thread ID.
2685 // This routine handles OR'ing together all the affinity masks of threads
2686 // that are sufficiently close, if granularity > fine.
2687 static kmp_affin_mask_t *__kmp_create_masks(unsigned *maxIndex,
2688  unsigned *numUnique,
2689  AddrUnsPair *address2os,
2690  unsigned numAddrs) {
2691  // First form a table of affinity masks in order of OS thread id.
2692  unsigned depth;
2693  unsigned maxOsId;
2694  unsigned i;
2695 
2696  KMP_ASSERT(numAddrs > 0);
2697  depth = address2os[0].first.depth;
2698 
2699  maxOsId = 0;
2700  for (i = numAddrs - 1;; --i) {
2701  unsigned osId = address2os[i].second;
2702  if (osId > maxOsId) {
2703  maxOsId = osId;
2704  }
2705  if (i == 0)
2706  break;
2707  }
2708  kmp_affin_mask_t *osId2Mask;
2709  KMP_CPU_ALLOC_ARRAY(osId2Mask, (maxOsId + 1));
2710 
2711  // Sort the address2os table according to physical order. Doing so will put
2712  // all threads on the same core/package/node in consecutive locations.
2713  qsort(address2os, numAddrs, sizeof(*address2os),
2714  __kmp_affinity_cmp_Address_labels);
2715 
2716  KMP_ASSERT(__kmp_affinity_gran_levels >= 0);
2717  if (__kmp_affinity_verbose && (__kmp_affinity_gran_levels > 0)) {
2718  KMP_INFORM(ThreadsMigrate, "KMP_AFFINITY", __kmp_affinity_gran_levels);
2719  }
2720  if (__kmp_affinity_gran_levels >= (int)depth) {
2721  if (__kmp_affinity_verbose ||
2722  (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) {
2723  KMP_WARNING(AffThreadsMayMigrate);
2724  }
2725  }
2726 
2727  // Run through the table, forming the masks for all threads on each core.
2728  // Threads on the same core will have identical "Address" objects, not
2729  // considering the last level, which must be the thread id. All threads on a
2730  // core will appear consecutively.
2731  unsigned unique = 0;
2732  unsigned j = 0; // index of 1st thread on core
2733  unsigned leader = 0;
2734  Address *leaderAddr = &(address2os[0].first);
2735  kmp_affin_mask_t *sum;
2736  KMP_CPU_ALLOC_ON_STACK(sum);
2737  KMP_CPU_ZERO(sum);
2738  KMP_CPU_SET(address2os[0].second, sum);
2739  for (i = 1; i < numAddrs; i++) {
2740  // If this thread is sufficiently close to the leader (within the
2741  // granularity setting), then set the bit for this os thread in the
2742  // affinity mask for this group, and go on to the next thread.
2743  if (leaderAddr->isClose(address2os[i].first, __kmp_affinity_gran_levels)) {
2744  KMP_CPU_SET(address2os[i].second, sum);
2745  continue;
2746  }
2747 
2748  // For every thread in this group, copy the mask to the thread's entry in
2749  // the osId2Mask table. Mark the first address as a leader.
2750  for (; j < i; j++) {
2751  unsigned osId = address2os[j].second;
2752  KMP_DEBUG_ASSERT(osId <= maxOsId);
2753  kmp_affin_mask_t *mask = KMP_CPU_INDEX(osId2Mask, osId);
2754  KMP_CPU_COPY(mask, sum);
2755  address2os[j].first.leader = (j == leader);
2756  }
2757  unique++;
2758 
2759  // Start a new mask.
2760  leader = i;
2761  leaderAddr = &(address2os[i].first);
2762  KMP_CPU_ZERO(sum);
2763  KMP_CPU_SET(address2os[i].second, sum);
2764  }
2765 
2766  // For every thread in last group, copy the mask to the thread's
2767  // entry in the osId2Mask table.
2768  for (; j < i; j++) {
2769  unsigned osId = address2os[j].second;
2770  KMP_DEBUG_ASSERT(osId <= maxOsId);
2771  kmp_affin_mask_t *mask = KMP_CPU_INDEX(osId2Mask, osId);
2772  KMP_CPU_COPY(mask, sum);
2773  address2os[j].first.leader = (j == leader);
2774  }
2775  unique++;
2776  KMP_CPU_FREE_FROM_STACK(sum);
2777 
2778  *maxIndex = maxOsId;
2779  *numUnique = unique;
2780  return osId2Mask;
2781 }
2782 
2783 // Stuff for the affinity proclist parsers. It's easier to declare these vars
2784 // as file-static than to try and pass them through the calling sequence of
2785 // the recursive-descent OMP_PLACES parser.
2786 static kmp_affin_mask_t *newMasks;
2787 static int numNewMasks;
2788 static int nextNewMask;
2789 
2790 #define ADD_MASK(_mask) \
2791  { \
2792  if (nextNewMask >= numNewMasks) { \
2793  int i; \
2794  numNewMasks *= 2; \
2795  kmp_affin_mask_t *temp; \
2796  KMP_CPU_INTERNAL_ALLOC_ARRAY(temp, numNewMasks); \
2797  for (i = 0; i < numNewMasks / 2; i++) { \
2798  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i); \
2799  kmp_affin_mask_t *dest = KMP_CPU_INDEX(temp, i); \
2800  KMP_CPU_COPY(dest, src); \
2801  } \
2802  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks / 2); \
2803  newMasks = temp; \
2804  } \
2805  KMP_CPU_COPY(KMP_CPU_INDEX(newMasks, nextNewMask), (_mask)); \
2806  nextNewMask++; \
2807  }
2808 
2809 #define ADD_MASK_OSID(_osId, _osId2Mask, _maxOsId) \
2810  { \
2811  if (((_osId) > _maxOsId) || \
2812  (!KMP_CPU_ISSET((_osId), KMP_CPU_INDEX((_osId2Mask), (_osId))))) { \
2813  if (__kmp_affinity_verbose || \
2814  (__kmp_affinity_warnings && \
2815  (__kmp_affinity_type != affinity_none))) { \
2816  KMP_WARNING(AffIgnoreInvalidProcID, _osId); \
2817  } \
2818  } else { \
2819  ADD_MASK(KMP_CPU_INDEX(_osId2Mask, (_osId))); \
2820  } \
2821  }
2822 
2823 // Re-parse the proclist (for the explicit affinity type), and form the list
2824 // of affinity newMasks indexed by gtid.
2825 static void __kmp_affinity_process_proclist(kmp_affin_mask_t **out_masks,
2826  unsigned int *out_numMasks,
2827  const char *proclist,
2828  kmp_affin_mask_t *osId2Mask,
2829  int maxOsId) {
2830  int i;
2831  const char *scan = proclist;
2832  const char *next = proclist;
2833 
2834  // We use malloc() for the temporary mask vector, so that we can use
2835  // realloc() to extend it.
2836  numNewMasks = 2;
2837  KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
2838  nextNewMask = 0;
2839  kmp_affin_mask_t *sumMask;
2840  KMP_CPU_ALLOC(sumMask);
2841  int setSize = 0;
2842 
2843  for (;;) {
2844  int start, end, stride;
2845 
2846  SKIP_WS(scan);
2847  next = scan;
2848  if (*next == '\0') {
2849  break;
2850  }
2851 
2852  if (*next == '{') {
2853  int num;
2854  setSize = 0;
2855  next++; // skip '{'
2856  SKIP_WS(next);
2857  scan = next;
2858 
2859  // Read the first integer in the set.
2860  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad proclist");
2861  SKIP_DIGITS(next);
2862  num = __kmp_str_to_int(scan, *next);
2863  KMP_ASSERT2(num >= 0, "bad explicit proc list");
2864 
2865  // Copy the mask for that osId to the sum (union) mask.
2866  if ((num > maxOsId) ||
2867  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
2868  if (__kmp_affinity_verbose ||
2869  (__kmp_affinity_warnings &&
2870  (__kmp_affinity_type != affinity_none))) {
2871  KMP_WARNING(AffIgnoreInvalidProcID, num);
2872  }
2873  KMP_CPU_ZERO(sumMask);
2874  } else {
2875  KMP_CPU_COPY(sumMask, KMP_CPU_INDEX(osId2Mask, num));
2876  setSize = 1;
2877  }
2878 
2879  for (;;) {
2880  // Check for end of set.
2881  SKIP_WS(next);
2882  if (*next == '}') {
2883  next++; // skip '}'
2884  break;
2885  }
2886 
2887  // Skip optional comma.
2888  if (*next == ',') {
2889  next++;
2890  }
2891  SKIP_WS(next);
2892 
2893  // Read the next integer in the set.
2894  scan = next;
2895  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
2896 
2897  SKIP_DIGITS(next);
2898  num = __kmp_str_to_int(scan, *next);
2899  KMP_ASSERT2(num >= 0, "bad explicit proc list");
2900 
2901  // Add the mask for that osId to the sum mask.
2902  if ((num > maxOsId) ||
2903  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
2904  if (__kmp_affinity_verbose ||
2905  (__kmp_affinity_warnings &&
2906  (__kmp_affinity_type != affinity_none))) {
2907  KMP_WARNING(AffIgnoreInvalidProcID, num);
2908  }
2909  } else {
2910  KMP_CPU_UNION(sumMask, KMP_CPU_INDEX(osId2Mask, num));
2911  setSize++;
2912  }
2913  }
2914  if (setSize > 0) {
2915  ADD_MASK(sumMask);
2916  }
2917 
2918  SKIP_WS(next);
2919  if (*next == ',') {
2920  next++;
2921  }
2922  scan = next;
2923  continue;
2924  }
2925 
2926  // Read the first integer.
2927  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
2928  SKIP_DIGITS(next);
2929  start = __kmp_str_to_int(scan, *next);
2930  KMP_ASSERT2(start >= 0, "bad explicit proc list");
2931  SKIP_WS(next);
2932 
2933  // If this isn't a range, then add a mask to the list and go on.
2934  if (*next != '-') {
2935  ADD_MASK_OSID(start, osId2Mask, maxOsId);
2936 
2937  // Skip optional comma.
2938  if (*next == ',') {
2939  next++;
2940  }
2941  scan = next;
2942  continue;
2943  }
2944 
2945  // This is a range. Skip over the '-' and read in the 2nd int.
2946  next++; // skip '-'
2947  SKIP_WS(next);
2948  scan = next;
2949  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
2950  SKIP_DIGITS(next);
2951  end = __kmp_str_to_int(scan, *next);
2952  KMP_ASSERT2(end >= 0, "bad explicit proc list");
2953 
2954  // Check for a stride parameter
2955  stride = 1;
2956  SKIP_WS(next);
2957  if (*next == ':') {
2958  // A stride is specified. Skip over the ':" and read the 3rd int.
2959  int sign = +1;
2960  next++; // skip ':'
2961  SKIP_WS(next);
2962  scan = next;
2963  if (*next == '-') {
2964  sign = -1;
2965  next++;
2966  SKIP_WS(next);
2967  scan = next;
2968  }
2969  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
2970  SKIP_DIGITS(next);
2971  stride = __kmp_str_to_int(scan, *next);
2972  KMP_ASSERT2(stride >= 0, "bad explicit proc list");
2973  stride *= sign;
2974  }
2975 
2976  // Do some range checks.
2977  KMP_ASSERT2(stride != 0, "bad explicit proc list");
2978  if (stride > 0) {
2979  KMP_ASSERT2(start <= end, "bad explicit proc list");
2980  } else {
2981  KMP_ASSERT2(start >= end, "bad explicit proc list");
2982  }
2983  KMP_ASSERT2((end - start) / stride <= 65536, "bad explicit proc list");
2984 
2985  // Add the mask for each OS proc # to the list.
2986  if (stride > 0) {
2987  do {
2988  ADD_MASK_OSID(start, osId2Mask, maxOsId);
2989  start += stride;
2990  } while (start <= end);
2991  } else {
2992  do {
2993  ADD_MASK_OSID(start, osId2Mask, maxOsId);
2994  start += stride;
2995  } while (start >= end);
2996  }
2997 
2998  // Skip optional comma.
2999  SKIP_WS(next);
3000  if (*next == ',') {
3001  next++;
3002  }
3003  scan = next;
3004  }
3005 
3006  *out_numMasks = nextNewMask;
3007  if (nextNewMask == 0) {
3008  *out_masks = NULL;
3009  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3010  return;
3011  }
3012  KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
3013  for (i = 0; i < nextNewMask; i++) {
3014  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
3015  kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
3016  KMP_CPU_COPY(dest, src);
3017  }
3018  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3019  KMP_CPU_FREE(sumMask);
3020 }
3021 
3022 #if OMP_40_ENABLED
3023 
3024 /*-----------------------------------------------------------------------------
3025 Re-parse the OMP_PLACES proc id list, forming the newMasks for the different
3026 places. Again, Here is the grammar:
3027 
3028 place_list := place
3029 place_list := place , place_list
3030 place := num
3031 place := place : num
3032 place := place : num : signed
3033 place := { subplacelist }
3034 place := ! place // (lowest priority)
3035 subplace_list := subplace
3036 subplace_list := subplace , subplace_list
3037 subplace := num
3038 subplace := num : num
3039 subplace := num : num : signed
3040 signed := num
3041 signed := + signed
3042 signed := - signed
3043 -----------------------------------------------------------------------------*/
3044 
3045 static void __kmp_process_subplace_list(const char **scan,
3046  kmp_affin_mask_t *osId2Mask,
3047  int maxOsId, kmp_affin_mask_t *tempMask,
3048  int *setSize) {
3049  const char *next;
3050 
3051  for (;;) {
3052  int start, count, stride, i;
3053 
3054  // Read in the starting proc id
3055  SKIP_WS(*scan);
3056  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3057  next = *scan;
3058  SKIP_DIGITS(next);
3059  start = __kmp_str_to_int(*scan, *next);
3060  KMP_ASSERT(start >= 0);
3061  *scan = next;
3062 
3063  // valid follow sets are ',' ':' and '}'
3064  SKIP_WS(*scan);
3065  if (**scan == '}' || **scan == ',') {
3066  if ((start > maxOsId) ||
3067  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3068  if (__kmp_affinity_verbose ||
3069  (__kmp_affinity_warnings &&
3070  (__kmp_affinity_type != affinity_none))) {
3071  KMP_WARNING(AffIgnoreInvalidProcID, start);
3072  }
3073  } else {
3074  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3075  (*setSize)++;
3076  }
3077  if (**scan == '}') {
3078  break;
3079  }
3080  (*scan)++; // skip ','
3081  continue;
3082  }
3083  KMP_ASSERT2(**scan == ':', "bad explicit places list");
3084  (*scan)++; // skip ':'
3085 
3086  // Read count parameter
3087  SKIP_WS(*scan);
3088  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3089  next = *scan;
3090  SKIP_DIGITS(next);
3091  count = __kmp_str_to_int(*scan, *next);
3092  KMP_ASSERT(count >= 0);
3093  *scan = next;
3094 
3095  // valid follow sets are ',' ':' and '}'
3096  SKIP_WS(*scan);
3097  if (**scan == '}' || **scan == ',') {
3098  for (i = 0; i < count; i++) {
3099  if ((start > maxOsId) ||
3100  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3101  if (__kmp_affinity_verbose ||
3102  (__kmp_affinity_warnings &&
3103  (__kmp_affinity_type != affinity_none))) {
3104  KMP_WARNING(AffIgnoreInvalidProcID, start);
3105  }
3106  break; // don't proliferate warnings for large count
3107  } else {
3108  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3109  start++;
3110  (*setSize)++;
3111  }
3112  }
3113  if (**scan == '}') {
3114  break;
3115  }
3116  (*scan)++; // skip ','
3117  continue;
3118  }
3119  KMP_ASSERT2(**scan == ':', "bad explicit places list");
3120  (*scan)++; // skip ':'
3121 
3122  // Read stride parameter
3123  int sign = +1;
3124  for (;;) {
3125  SKIP_WS(*scan);
3126  if (**scan == '+') {
3127  (*scan)++; // skip '+'
3128  continue;
3129  }
3130  if (**scan == '-') {
3131  sign *= -1;
3132  (*scan)++; // skip '-'
3133  continue;
3134  }
3135  break;
3136  }
3137  SKIP_WS(*scan);
3138  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3139  next = *scan;
3140  SKIP_DIGITS(next);
3141  stride = __kmp_str_to_int(*scan, *next);
3142  KMP_ASSERT(stride >= 0);
3143  *scan = next;
3144  stride *= sign;
3145 
3146  // valid follow sets are ',' and '}'
3147  SKIP_WS(*scan);
3148  if (**scan == '}' || **scan == ',') {
3149  for (i = 0; i < count; i++) {
3150  if ((start > maxOsId) ||
3151  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3152  if (__kmp_affinity_verbose ||
3153  (__kmp_affinity_warnings &&
3154  (__kmp_affinity_type != affinity_none))) {
3155  KMP_WARNING(AffIgnoreInvalidProcID, start);
3156  }
3157  break; // don't proliferate warnings for large count
3158  } else {
3159  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3160  start += stride;
3161  (*setSize)++;
3162  }
3163  }
3164  if (**scan == '}') {
3165  break;
3166  }
3167  (*scan)++; // skip ','
3168  continue;
3169  }
3170 
3171  KMP_ASSERT2(0, "bad explicit places list");
3172  }
3173 }
3174 
3175 static void __kmp_process_place(const char **scan, kmp_affin_mask_t *osId2Mask,
3176  int maxOsId, kmp_affin_mask_t *tempMask,
3177  int *setSize) {
3178  const char *next;
3179 
3180  // valid follow sets are '{' '!' and num
3181  SKIP_WS(*scan);
3182  if (**scan == '{') {
3183  (*scan)++; // skip '{'
3184  __kmp_process_subplace_list(scan, osId2Mask, maxOsId, tempMask, setSize);
3185  KMP_ASSERT2(**scan == '}', "bad explicit places list");
3186  (*scan)++; // skip '}'
3187  } else if (**scan == '!') {
3188  (*scan)++; // skip '!'
3189  __kmp_process_place(scan, osId2Mask, maxOsId, tempMask, setSize);
3190  KMP_CPU_COMPLEMENT(maxOsId, tempMask);
3191  } else if ((**scan >= '0') && (**scan <= '9')) {
3192  next = *scan;
3193  SKIP_DIGITS(next);
3194  int num = __kmp_str_to_int(*scan, *next);
3195  KMP_ASSERT(num >= 0);
3196  if ((num > maxOsId) ||
3197  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
3198  if (__kmp_affinity_verbose ||
3199  (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) {
3200  KMP_WARNING(AffIgnoreInvalidProcID, num);
3201  }
3202  } else {
3203  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, num));
3204  (*setSize)++;
3205  }
3206  *scan = next; // skip num
3207  } else {
3208  KMP_ASSERT2(0, "bad explicit places list");
3209  }
3210 }
3211 
3212 // static void
3213 void __kmp_affinity_process_placelist(kmp_affin_mask_t **out_masks,
3214  unsigned int *out_numMasks,
3215  const char *placelist,
3216  kmp_affin_mask_t *osId2Mask,
3217  int maxOsId) {
3218  int i, j, count, stride, sign;
3219  const char *scan = placelist;
3220  const char *next = placelist;
3221 
3222  numNewMasks = 2;
3223  KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
3224  nextNewMask = 0;
3225 
3226  // tempMask is modified based on the previous or initial
3227  // place to form the current place
3228  // previousMask contains the previous place
3229  kmp_affin_mask_t *tempMask;
3230  kmp_affin_mask_t *previousMask;
3231  KMP_CPU_ALLOC(tempMask);
3232  KMP_CPU_ZERO(tempMask);
3233  KMP_CPU_ALLOC(previousMask);
3234  KMP_CPU_ZERO(previousMask);
3235  int setSize = 0;
3236 
3237  for (;;) {
3238  __kmp_process_place(&scan, osId2Mask, maxOsId, tempMask, &setSize);
3239 
3240  // valid follow sets are ',' ':' and EOL
3241  SKIP_WS(scan);
3242  if (*scan == '\0' || *scan == ',') {
3243  if (setSize > 0) {
3244  ADD_MASK(tempMask);
3245  }
3246  KMP_CPU_ZERO(tempMask);
3247  setSize = 0;
3248  if (*scan == '\0') {
3249  break;
3250  }
3251  scan++; // skip ','
3252  continue;
3253  }
3254 
3255  KMP_ASSERT2(*scan == ':', "bad explicit places list");
3256  scan++; // skip ':'
3257 
3258  // Read count parameter
3259  SKIP_WS(scan);
3260  KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
3261  next = scan;
3262  SKIP_DIGITS(next);
3263  count = __kmp_str_to_int(scan, *next);
3264  KMP_ASSERT(count >= 0);
3265  scan = next;
3266 
3267  // valid follow sets are ',' ':' and EOL
3268  SKIP_WS(scan);
3269  if (*scan == '\0' || *scan == ',') {
3270  stride = +1;
3271  } else {
3272  KMP_ASSERT2(*scan == ':', "bad explicit places list");
3273  scan++; // skip ':'
3274 
3275  // Read stride parameter
3276  sign = +1;
3277  for (;;) {
3278  SKIP_WS(scan);
3279  if (*scan == '+') {
3280  scan++; // skip '+'
3281  continue;
3282  }
3283  if (*scan == '-') {
3284  sign *= -1;
3285  scan++; // skip '-'
3286  continue;
3287  }
3288  break;
3289  }
3290  SKIP_WS(scan);
3291  KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
3292  next = scan;
3293  SKIP_DIGITS(next);
3294  stride = __kmp_str_to_int(scan, *next);
3295  KMP_DEBUG_ASSERT(stride >= 0);
3296  scan = next;
3297  stride *= sign;
3298  }
3299 
3300  // Add places determined by initial_place : count : stride
3301  for (i = 0; i < count; i++) {
3302  if (setSize == 0) {
3303  break;
3304  }
3305  // Add the current place, then build the next place (tempMask) from that
3306  KMP_CPU_COPY(previousMask, tempMask);
3307  ADD_MASK(previousMask);
3308  KMP_CPU_ZERO(tempMask);
3309  setSize = 0;
3310  KMP_CPU_SET_ITERATE(j, previousMask) {
3311  if (!KMP_CPU_ISSET(j, previousMask)) {
3312  continue;
3313  }
3314  if ((j + stride > maxOsId) || (j + stride < 0) ||
3315  (!KMP_CPU_ISSET(j, __kmp_affin_fullMask)) ||
3316  (!KMP_CPU_ISSET(j + stride,
3317  KMP_CPU_INDEX(osId2Mask, j + stride)))) {
3318  if ((__kmp_affinity_verbose ||
3319  (__kmp_affinity_warnings &&
3320  (__kmp_affinity_type != affinity_none))) &&
3321  i < count - 1) {
3322  KMP_WARNING(AffIgnoreInvalidProcID, j + stride);
3323  }
3324  continue;
3325  }
3326  KMP_CPU_SET(j + stride, tempMask);
3327  setSize++;
3328  }
3329  }
3330  KMP_CPU_ZERO(tempMask);
3331  setSize = 0;
3332 
3333  // valid follow sets are ',' and EOL
3334  SKIP_WS(scan);
3335  if (*scan == '\0') {
3336  break;
3337  }
3338  if (*scan == ',') {
3339  scan++; // skip ','
3340  continue;
3341  }
3342 
3343  KMP_ASSERT2(0, "bad explicit places list");
3344  }
3345 
3346  *out_numMasks = nextNewMask;
3347  if (nextNewMask == 0) {
3348  *out_masks = NULL;
3349  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3350  return;
3351  }
3352  KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
3353  KMP_CPU_FREE(tempMask);
3354  KMP_CPU_FREE(previousMask);
3355  for (i = 0; i < nextNewMask; i++) {
3356  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
3357  kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
3358  KMP_CPU_COPY(dest, src);
3359  }
3360  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3361 }
3362 
3363 #endif /* OMP_40_ENABLED */
3364 
3365 #undef ADD_MASK
3366 #undef ADD_MASK_OSID
3367 
3368 #if KMP_USE_HWLOC
3369 static int __kmp_hwloc_skip_PUs_obj(hwloc_topology_t t, hwloc_obj_t o) {
3370  // skip PUs descendants of the object o
3371  int skipped = 0;
3372  hwloc_obj_t hT = NULL;
3373  int N = __kmp_hwloc_count_children_by_type(t, o, HWLOC_OBJ_PU, &hT);
3374  for (int i = 0; i < N; ++i) {
3375  KMP_DEBUG_ASSERT(hT);
3376  unsigned idx = hT->os_index;
3377  if (KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
3378  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
3379  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
3380  ++skipped;
3381  }
3382  hT = hwloc_get_next_obj_by_type(t, HWLOC_OBJ_PU, hT);
3383  }
3384  return skipped; // count number of skipped units
3385 }
3386 
3387 static int __kmp_hwloc_obj_has_PUs(hwloc_topology_t t, hwloc_obj_t o) {
3388  // check if obj has PUs present in fullMask
3389  hwloc_obj_t hT = NULL;
3390  int N = __kmp_hwloc_count_children_by_type(t, o, HWLOC_OBJ_PU, &hT);
3391  for (int i = 0; i < N; ++i) {
3392  KMP_DEBUG_ASSERT(hT);
3393  unsigned idx = hT->os_index;
3394  if (KMP_CPU_ISSET(idx, __kmp_affin_fullMask))
3395  return 1; // found PU
3396  hT = hwloc_get_next_obj_by_type(t, HWLOC_OBJ_PU, hT);
3397  }
3398  return 0; // no PUs found
3399 }
3400 #endif // KMP_USE_HWLOC
3401 
3402 static void __kmp_apply_thread_places(AddrUnsPair **pAddr, int depth) {
3403  AddrUnsPair *newAddr;
3404  if (__kmp_hws_requested == 0)
3405  goto _exit; // no topology limiting actions requested, exit
3406 #if KMP_USE_HWLOC
3407  if (__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC) {
3408  // Number of subobjects calculated dynamically, this works fine for
3409  // any non-uniform topology.
3410  // L2 cache objects are determined by depth, other objects - by type.
3411  hwloc_topology_t tp = __kmp_hwloc_topology;
3412  int nS = 0, nN = 0, nL = 0, nC = 0,
3413  nT = 0; // logical index including skipped
3414  int nCr = 0, nTr = 0; // number of requested units
3415  int nPkg = 0, nCo = 0, n_new = 0, n_old = 0, nCpP = 0, nTpC = 0; // counters
3416  hwloc_obj_t hT, hC, hL, hN, hS; // hwloc objects (pointers to)
3417  int L2depth, idx;
3418 
3419  // check support of extensions ----------------------------------
3420  int numa_support = 0, tile_support = 0;
3421  if (__kmp_pu_os_idx)
3422  hT = hwloc_get_pu_obj_by_os_index(tp,
3423  __kmp_pu_os_idx[__kmp_avail_proc - 1]);
3424  else
3425  hT = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PU, __kmp_avail_proc - 1);
3426  if (hT == NULL) { // something's gone wrong
3427  KMP_WARNING(AffHWSubsetUnsupported);
3428  goto _exit;
3429  }
3430  // check NUMA node
3431  hN = hwloc_get_ancestor_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hT);
3432  hS = hwloc_get_ancestor_obj_by_type(tp, HWLOC_OBJ_PACKAGE, hT);
3433  if (hN != NULL && hN->depth > hS->depth) {
3434  numa_support = 1; // 1 in case socket includes node(s)
3435  } else if (__kmp_hws_node.num > 0) {
3436  // don't support sockets inside NUMA node (no such HW found for testing)
3437  KMP_WARNING(AffHWSubsetUnsupported);
3438  goto _exit;
3439  }
3440  // check L2 cahce, get object by depth because of multiple caches
3441  L2depth = hwloc_get_cache_type_depth(tp, 2, HWLOC_OBJ_CACHE_UNIFIED);
3442  hL = hwloc_get_ancestor_obj_by_depth(tp, L2depth, hT);
3443  if (hL != NULL &&
3444  __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE, &hC) > 1) {
3445  tile_support = 1; // no sense to count L2 if it includes single core
3446  } else if (__kmp_hws_tile.num > 0) {
3447  if (__kmp_hws_core.num == 0) {
3448  __kmp_hws_core = __kmp_hws_tile; // replace L2 with core
3449  __kmp_hws_tile.num = 0;
3450  } else {
3451  // L2 and core are both requested, but represent same object
3452  KMP_WARNING(AffHWSubsetInvalid);
3453  goto _exit;
3454  }
3455  }
3456  // end of check of extensions -----------------------------------
3457 
3458  // fill in unset items, validate settings -----------------------
3459  if (__kmp_hws_socket.num == 0)
3460  __kmp_hws_socket.num = nPackages; // use all available sockets
3461  if (__kmp_hws_socket.offset >= nPackages) {
3462  KMP_WARNING(AffHWSubsetManySockets);
3463  goto _exit;
3464  }
3465  if (numa_support) {
3466  hN = NULL;
3467  int NN = __kmp_hwloc_count_children_by_type(tp, hS, HWLOC_OBJ_NUMANODE,
3468  &hN); // num nodes in socket
3469  if (__kmp_hws_node.num == 0)
3470  __kmp_hws_node.num = NN; // use all available nodes
3471  if (__kmp_hws_node.offset >= NN) {
3472  KMP_WARNING(AffHWSubsetManyNodes);
3473  goto _exit;
3474  }
3475  if (tile_support) {
3476  // get num tiles in node
3477  int NL = __kmp_hwloc_count_children_by_depth(tp, hN, L2depth, &hL);
3478  if (__kmp_hws_tile.num == 0) {
3479  __kmp_hws_tile.num = NL + 1;
3480  } // use all available tiles, some node may have more tiles, thus +1
3481  if (__kmp_hws_tile.offset >= NL) {
3482  KMP_WARNING(AffHWSubsetManyTiles);
3483  goto _exit;
3484  }
3485  int NC = __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE,
3486  &hC); // num cores in tile
3487  if (__kmp_hws_core.num == 0)
3488  __kmp_hws_core.num = NC; // use all available cores
3489  if (__kmp_hws_core.offset >= NC) {
3490  KMP_WARNING(AffHWSubsetManyCores);
3491  goto _exit;
3492  }
3493  } else { // tile_support
3494  int NC = __kmp_hwloc_count_children_by_type(tp, hN, HWLOC_OBJ_CORE,
3495  &hC); // num cores in node
3496  if (__kmp_hws_core.num == 0)
3497  __kmp_hws_core.num = NC; // use all available cores
3498  if (__kmp_hws_core.offset >= NC) {
3499  KMP_WARNING(AffHWSubsetManyCores);
3500  goto _exit;
3501  }
3502  } // tile_support
3503  } else { // numa_support
3504  if (tile_support) {
3505  // get num tiles in socket
3506  int NL = __kmp_hwloc_count_children_by_depth(tp, hS, L2depth, &hL);
3507  if (__kmp_hws_tile.num == 0)
3508  __kmp_hws_tile.num = NL; // use all available tiles
3509  if (__kmp_hws_tile.offset >= NL) {
3510  KMP_WARNING(AffHWSubsetManyTiles);
3511  goto _exit;
3512  }
3513  int NC = __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE,
3514  &hC); // num cores in tile
3515  if (__kmp_hws_core.num == 0)
3516  __kmp_hws_core.num = NC; // use all available cores
3517  if (__kmp_hws_core.offset >= NC) {
3518  KMP_WARNING(AffHWSubsetManyCores);
3519  goto _exit;
3520  }
3521  } else { // tile_support
3522  int NC = __kmp_hwloc_count_children_by_type(tp, hS, HWLOC_OBJ_CORE,
3523  &hC); // num cores in socket
3524  if (__kmp_hws_core.num == 0)
3525  __kmp_hws_core.num = NC; // use all available cores
3526  if (__kmp_hws_core.offset >= NC) {
3527  KMP_WARNING(AffHWSubsetManyCores);
3528  goto _exit;
3529  }
3530  } // tile_support
3531  }
3532  if (__kmp_hws_proc.num == 0)
3533  __kmp_hws_proc.num = __kmp_nThreadsPerCore; // use all available procs
3534  if (__kmp_hws_proc.offset >= __kmp_nThreadsPerCore) {
3535  KMP_WARNING(AffHWSubsetManyProcs);
3536  goto _exit;
3537  }
3538  // end of validation --------------------------------------------
3539 
3540  if (pAddr) // pAddr is NULL in case of affinity_none
3541  newAddr = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) *
3542  __kmp_avail_proc); // max size
3543  // main loop to form HW subset ----------------------------------
3544  hS = NULL;
3545  int NP = hwloc_get_nbobjs_by_type(tp, HWLOC_OBJ_PACKAGE);
3546  for (int s = 0; s < NP; ++s) {
3547  // Check Socket -----------------------------------------------
3548  hS = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PACKAGE, hS);
3549  if (!__kmp_hwloc_obj_has_PUs(tp, hS))
3550  continue; // skip socket if all PUs are out of fullMask
3551  ++nS; // only count objects those have PUs in affinity mask
3552  if (nS <= __kmp_hws_socket.offset ||
3553  nS > __kmp_hws_socket.num + __kmp_hws_socket.offset) {
3554  n_old += __kmp_hwloc_skip_PUs_obj(tp, hS); // skip socket
3555  continue; // move to next socket
3556  }
3557  nCr = 0; // count number of cores per socket
3558  // socket requested, go down the topology tree
3559  // check 4 cases: (+NUMA+Tile), (+NUMA-Tile), (-NUMA+Tile), (-NUMA-Tile)
3560  if (numa_support) {
3561  nN = 0;
3562  hN = NULL;
3563  // num nodes in current socket
3564  int NN =
3565  __kmp_hwloc_count_children_by_type(tp, hS, HWLOC_OBJ_NUMANODE, &hN);
3566  for (int n = 0; n < NN; ++n) {
3567  // Check NUMA Node ----------------------------------------
3568  if (!__kmp_hwloc_obj_has_PUs(tp, hN)) {
3569  hN = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hN);
3570  continue; // skip node if all PUs are out of fullMask
3571  }
3572  ++nN;
3573  if (nN <= __kmp_hws_node.offset ||
3574  nN > __kmp_hws_node.num + __kmp_hws_node.offset) {
3575  // skip node as not requested
3576  n_old += __kmp_hwloc_skip_PUs_obj(tp, hN); // skip node
3577  hN = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hN);
3578  continue; // move to next node
3579  }
3580  // node requested, go down the topology tree
3581  if (tile_support) {
3582  nL = 0;
3583  hL = NULL;
3584  int NL = __kmp_hwloc_count_children_by_depth(tp, hN, L2depth, &hL);
3585  for (int l = 0; l < NL; ++l) {
3586  // Check L2 (tile) ------------------------------------
3587  if (!__kmp_hwloc_obj_has_PUs(tp, hL)) {
3588  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3589  continue; // skip tile if all PUs are out of fullMask
3590  }
3591  ++nL;
3592  if (nL <= __kmp_hws_tile.offset ||
3593  nL > __kmp_hws_tile.num + __kmp_hws_tile.offset) {
3594  // skip tile as not requested
3595  n_old += __kmp_hwloc_skip_PUs_obj(tp, hL); // skip tile
3596  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3597  continue; // move to next tile
3598  }
3599  // tile requested, go down the topology tree
3600  nC = 0;
3601  hC = NULL;
3602  // num cores in current tile
3603  int NC = __kmp_hwloc_count_children_by_type(tp, hL,
3604  HWLOC_OBJ_CORE, &hC);
3605  for (int c = 0; c < NC; ++c) {
3606  // Check Core ---------------------------------------
3607  if (!__kmp_hwloc_obj_has_PUs(tp, hC)) {
3608  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3609  continue; // skip core if all PUs are out of fullMask
3610  }
3611  ++nC;
3612  if (nC <= __kmp_hws_core.offset ||
3613  nC > __kmp_hws_core.num + __kmp_hws_core.offset) {
3614  // skip node as not requested
3615  n_old += __kmp_hwloc_skip_PUs_obj(tp, hC); // skip core
3616  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3617  continue; // move to next node
3618  }
3619  // core requested, go down to PUs
3620  nT = 0;
3621  nTr = 0;
3622  hT = NULL;
3623  // num procs in current core
3624  int NT = __kmp_hwloc_count_children_by_type(tp, hC,
3625  HWLOC_OBJ_PU, &hT);
3626  for (int t = 0; t < NT; ++t) {
3627  // Check PU ---------------------------------------
3628  idx = hT->os_index;
3629  if (!KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
3630  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3631  continue; // skip PU if not in fullMask
3632  }
3633  ++nT;
3634  if (nT <= __kmp_hws_proc.offset ||
3635  nT > __kmp_hws_proc.num + __kmp_hws_proc.offset) {
3636  // skip PU
3637  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
3638  ++n_old;
3639  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
3640  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3641  continue; // move to next node
3642  }
3643  ++nTr;
3644  if (pAddr) // collect requested thread's data
3645  newAddr[n_new] = (*pAddr)[n_old];
3646  ++n_new;
3647  ++n_old;
3648  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3649  } // threads loop
3650  if (nTr > 0) {
3651  ++nCr; // num cores per socket
3652  ++nCo; // total num cores
3653  if (nTr > nTpC)
3654  nTpC = nTr; // calc max threads per core
3655  }
3656  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3657  } // cores loop
3658  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3659  } // tiles loop
3660  } else { // tile_support
3661  // no tiles, check cores
3662  nC = 0;
3663  hC = NULL;
3664  // num cores in current node
3665  int NC =
3666  __kmp_hwloc_count_children_by_type(tp, hN, HWLOC_OBJ_CORE, &hC);
3667  for (int c = 0; c < NC; ++c) {
3668  // Check Core ---------------------------------------
3669  if (!__kmp_hwloc_obj_has_PUs(tp, hC)) {
3670  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3671  continue; // skip core if all PUs are out of fullMask
3672  }
3673  ++nC;
3674  if (nC <= __kmp_hws_core.offset ||
3675  nC > __kmp_hws_core.num + __kmp_hws_core.offset) {
3676  // skip node as not requested
3677  n_old += __kmp_hwloc_skip_PUs_obj(tp, hC); // skip core
3678  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3679  continue; // move to next node
3680  }
3681  // core requested, go down to PUs
3682  nT = 0;
3683  nTr = 0;
3684  hT = NULL;
3685  int NT =
3686  __kmp_hwloc_count_children_by_type(tp, hC, HWLOC_OBJ_PU, &hT);
3687  for (int t = 0; t < NT; ++t) {
3688  // Check PU ---------------------------------------
3689  idx = hT->os_index;
3690  if (!KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
3691  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3692  continue; // skip PU if not in fullMask
3693  }
3694  ++nT;
3695  if (nT <= __kmp_hws_proc.offset ||
3696  nT > __kmp_hws_proc.num + __kmp_hws_proc.offset) {
3697  // skip PU
3698  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
3699  ++n_old;
3700  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
3701  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3702  continue; // move to next node
3703  }
3704  ++nTr;
3705  if (pAddr) // collect requested thread's data
3706  newAddr[n_new] = (*pAddr)[n_old];
3707  ++n_new;
3708  ++n_old;
3709  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3710  } // threads loop
3711  if (nTr > 0) {
3712  ++nCr; // num cores per socket
3713  ++nCo; // total num cores
3714  if (nTr > nTpC)
3715  nTpC = nTr; // calc max threads per core
3716  }
3717  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3718  } // cores loop
3719  } // tiles support
3720  hN = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hN);
3721  } // nodes loop
3722  } else { // numa_support
3723  // no NUMA support
3724  if (tile_support) {
3725  nL = 0;
3726  hL = NULL;
3727  // num tiles in current socket
3728  int NL = __kmp_hwloc_count_children_by_depth(tp, hS, L2depth, &hL);
3729  for (int l = 0; l < NL; ++l) {
3730  // Check L2 (tile) ------------------------------------
3731  if (!__kmp_hwloc_obj_has_PUs(tp, hL)) {
3732  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3733  continue; // skip tile if all PUs are out of fullMask
3734  }
3735  ++nL;
3736  if (nL <= __kmp_hws_tile.offset ||
3737  nL > __kmp_hws_tile.num + __kmp_hws_tile.offset) {
3738  // skip tile as not requested
3739  n_old += __kmp_hwloc_skip_PUs_obj(tp, hL); // skip tile
3740  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3741  continue; // move to next tile
3742  }
3743  // tile requested, go down the topology tree
3744  nC = 0;
3745  hC = NULL;
3746  // num cores per tile
3747  int NC =
3748  __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE, &hC);
3749  for (int c = 0; c < NC; ++c) {
3750  // Check Core ---------------------------------------
3751  if (!__kmp_hwloc_obj_has_PUs(tp, hC)) {
3752  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3753  continue; // skip core if all PUs are out of fullMask
3754  }
3755  ++nC;
3756  if (nC <= __kmp_hws_core.offset ||
3757  nC > __kmp_hws_core.num + __kmp_hws_core.offset) {
3758  // skip node as not requested
3759  n_old += __kmp_hwloc_skip_PUs_obj(tp, hC); // skip core
3760  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3761  continue; // move to next node
3762  }
3763  // core requested, go down to PUs
3764  nT = 0;
3765  nTr = 0;
3766  hT = NULL;
3767  // num procs per core
3768  int NT =
3769  __kmp_hwloc_count_children_by_type(tp, hC, HWLOC_OBJ_PU, &hT);
3770  for (int t = 0; t < NT; ++t) {
3771  // Check PU ---------------------------------------
3772  idx = hT->os_index;
3773  if (!KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
3774  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3775  continue; // skip PU if not in fullMask
3776  }
3777  ++nT;
3778  if (nT <= __kmp_hws_proc.offset ||
3779  nT > __kmp_hws_proc.num + __kmp_hws_proc.offset) {
3780  // skip PU
3781  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
3782  ++n_old;
3783  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
3784  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3785  continue; // move to next node
3786  }
3787  ++nTr;
3788  if (pAddr) // collect requested thread's data
3789  newAddr[n_new] = (*pAddr)[n_old];
3790  ++n_new;
3791  ++n_old;
3792  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3793  } // threads loop
3794  if (nTr > 0) {
3795  ++nCr; // num cores per socket
3796  ++nCo; // total num cores
3797  if (nTr > nTpC)
3798  nTpC = nTr; // calc max threads per core
3799  }
3800  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3801  } // cores loop
3802  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3803  } // tiles loop
3804  } else { // tile_support
3805  // no tiles, check cores
3806  nC = 0;
3807  hC = NULL;
3808  // num cores in socket
3809  int NC =
3810  __kmp_hwloc_count_children_by_type(tp, hS, HWLOC_OBJ_CORE, &hC);
3811  for (int c = 0; c < NC; ++c) {
3812  // Check Core -------------------------------------------
3813  if (!__kmp_hwloc_obj_has_PUs(tp, hC)) {
3814  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3815  continue; // skip core if all PUs are out of fullMask
3816  }
3817  ++nC;
3818  if (nC <= __kmp_hws_core.offset ||
3819  nC > __kmp_hws_core.num + __kmp_hws_core.offset) {
3820  // skip node as not requested
3821  n_old += __kmp_hwloc_skip_PUs_obj(tp, hC); // skip core
3822  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3823  continue; // move to next node
3824  }
3825  // core requested, go down to PUs
3826  nT = 0;
3827  nTr = 0;
3828  hT = NULL;
3829  // num procs per core
3830  int NT =
3831  __kmp_hwloc_count_children_by_type(tp, hC, HWLOC_OBJ_PU, &hT);
3832  for (int t = 0; t < NT; ++t) {
3833  // Check PU ---------------------------------------
3834  idx = hT->os_index;
3835  if (!KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
3836  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3837  continue; // skip PU if not in fullMask
3838  }
3839  ++nT;
3840  if (nT <= __kmp_hws_proc.offset ||
3841  nT > __kmp_hws_proc.num + __kmp_hws_proc.offset) {
3842  // skip PU
3843  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
3844  ++n_old;
3845  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
3846  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3847  continue; // move to next node
3848  }
3849  ++nTr;
3850  if (pAddr) // collect requested thread's data
3851  newAddr[n_new] = (*pAddr)[n_old];
3852  ++n_new;
3853  ++n_old;
3854  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3855  } // threads loop
3856  if (nTr > 0) {
3857  ++nCr; // num cores per socket
3858  ++nCo; // total num cores
3859  if (nTr > nTpC)
3860  nTpC = nTr; // calc max threads per core
3861  }
3862  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3863  } // cores loop
3864  } // tiles support
3865  } // numa_support
3866  if (nCr > 0) { // found cores?
3867  ++nPkg; // num sockets
3868  if (nCr > nCpP)
3869  nCpP = nCr; // calc max cores per socket
3870  }
3871  } // sockets loop
3872 
3873  // check the subset is valid
3874  KMP_DEBUG_ASSERT(n_old == __kmp_avail_proc);
3875  KMP_DEBUG_ASSERT(nPkg > 0);
3876  KMP_DEBUG_ASSERT(nCpP > 0);
3877  KMP_DEBUG_ASSERT(nTpC > 0);
3878  KMP_DEBUG_ASSERT(nCo > 0);
3879  KMP_DEBUG_ASSERT(nPkg <= nPackages);
3880  KMP_DEBUG_ASSERT(nCpP <= nCoresPerPkg);
3881  KMP_DEBUG_ASSERT(nTpC <= __kmp_nThreadsPerCore);
3882  KMP_DEBUG_ASSERT(nCo <= __kmp_ncores);
3883 
3884  nPackages = nPkg; // correct num sockets
3885  nCoresPerPkg = nCpP; // correct num cores per socket
3886  __kmp_nThreadsPerCore = nTpC; // correct num threads per core
3887  __kmp_avail_proc = n_new; // correct num procs
3888  __kmp_ncores = nCo; // correct num cores
3889  // hwloc topology method end
3890  } else
3891 #endif // KMP_USE_HWLOC
3892  {
3893  int n_old = 0, n_new = 0, proc_num = 0;
3894  if (__kmp_hws_node.num > 0 || __kmp_hws_tile.num > 0) {
3895  KMP_WARNING(AffHWSubsetNoHWLOC);
3896  goto _exit;
3897  }
3898  if (__kmp_hws_socket.num == 0)
3899  __kmp_hws_socket.num = nPackages; // use all available sockets
3900  if (__kmp_hws_core.num == 0)
3901  __kmp_hws_core.num = nCoresPerPkg; // use all available cores
3902  if (__kmp_hws_proc.num == 0 || __kmp_hws_proc.num > __kmp_nThreadsPerCore)
3903  __kmp_hws_proc.num = __kmp_nThreadsPerCore; // use all HW contexts
3904  if (!__kmp_affinity_uniform_topology()) {
3905  KMP_WARNING(AffHWSubsetNonUniform);
3906  goto _exit; // don't support non-uniform topology
3907  }
3908  if (depth > 3) {
3909  KMP_WARNING(AffHWSubsetNonThreeLevel);
3910  goto _exit; // don't support not-3-level topology
3911  }
3912  if (__kmp_hws_socket.offset + __kmp_hws_socket.num > nPackages) {
3913  KMP_WARNING(AffHWSubsetManySockets);
3914  goto _exit;
3915  }
3916  if (__kmp_hws_core.offset + __kmp_hws_core.num > nCoresPerPkg) {
3917  KMP_WARNING(AffHWSubsetManyCores);
3918  goto _exit;
3919  }
3920  // Form the requested subset
3921  if (pAddr) // pAddr is NULL in case of affinity_none
3922  newAddr = (AddrUnsPair *)__kmp_allocate(
3923  sizeof(AddrUnsPair) * __kmp_hws_socket.num * __kmp_hws_core.num *
3924  __kmp_hws_proc.num);
3925  for (int i = 0; i < nPackages; ++i) {
3926  if (i < __kmp_hws_socket.offset ||
3927  i >= __kmp_hws_socket.offset + __kmp_hws_socket.num) {
3928  // skip not-requested socket
3929  n_old += nCoresPerPkg * __kmp_nThreadsPerCore;
3930  if (__kmp_pu_os_idx != NULL) {
3931  // walk through skipped socket
3932  for (int j = 0; j < nCoresPerPkg; ++j) {
3933  for (int k = 0; k < __kmp_nThreadsPerCore; ++k) {
3934  KMP_CPU_CLR(__kmp_pu_os_idx[proc_num], __kmp_affin_fullMask);
3935  ++proc_num;
3936  }
3937  }
3938  }
3939  } else {
3940  // walk through requested socket
3941  for (int j = 0; j < nCoresPerPkg; ++j) {
3942  if (j < __kmp_hws_core.offset ||
3943  j >= __kmp_hws_core.offset +
3944  __kmp_hws_core.num) { // skip not-requested core
3945  n_old += __kmp_nThreadsPerCore;
3946  if (__kmp_pu_os_idx != NULL) {
3947  for (int k = 0; k < __kmp_nThreadsPerCore; ++k) {
3948  KMP_CPU_CLR(__kmp_pu_os_idx[proc_num], __kmp_affin_fullMask);
3949  ++proc_num;
3950  }
3951  }
3952  } else {
3953  // walk through requested core
3954  for (int k = 0; k < __kmp_nThreadsPerCore; ++k) {
3955  if (k < __kmp_hws_proc.num) {
3956  if (pAddr) // collect requested thread's data
3957  newAddr[n_new] = (*pAddr)[n_old];
3958  n_new++;
3959  } else {
3960  if (__kmp_pu_os_idx != NULL)
3961  KMP_CPU_CLR(__kmp_pu_os_idx[proc_num], __kmp_affin_fullMask);
3962  }
3963  n_old++;
3964  ++proc_num;
3965  }
3966  }
3967  }
3968  }
3969  }
3970  KMP_DEBUG_ASSERT(n_old == nPackages * nCoresPerPkg * __kmp_nThreadsPerCore);
3971  KMP_DEBUG_ASSERT(n_new ==
3972  __kmp_hws_socket.num * __kmp_hws_core.num *
3973  __kmp_hws_proc.num);
3974  nPackages = __kmp_hws_socket.num; // correct nPackages
3975  nCoresPerPkg = __kmp_hws_core.num; // correct nCoresPerPkg
3976  __kmp_nThreadsPerCore = __kmp_hws_proc.num; // correct __kmp_nThreadsPerCore
3977  __kmp_avail_proc = n_new; // correct avail_proc
3978  __kmp_ncores = nPackages * __kmp_hws_core.num; // correct ncores
3979  } // non-hwloc topology method
3980  if (pAddr) {
3981  __kmp_free(*pAddr);
3982  *pAddr = newAddr; // replace old topology with new one
3983  }
3984  if (__kmp_affinity_verbose) {
3985  char m[KMP_AFFIN_MASK_PRINT_LEN];
3986  __kmp_affinity_print_mask(m, KMP_AFFIN_MASK_PRINT_LEN,
3987  __kmp_affin_fullMask);
3988  if (__kmp_affinity_respect_mask) {
3989  KMP_INFORM(InitOSProcSetRespect, "KMP_HW_SUBSET", m);
3990  } else {
3991  KMP_INFORM(InitOSProcSetNotRespect, "KMP_HW_SUBSET", m);
3992  }
3993  KMP_INFORM(AvailableOSProc, "KMP_HW_SUBSET", __kmp_avail_proc);
3994  kmp_str_buf_t buf;
3995  __kmp_str_buf_init(&buf);
3996  __kmp_str_buf_print(&buf, "%d", nPackages);
3997  KMP_INFORM(TopologyExtra, "KMP_HW_SUBSET", buf.str, nCoresPerPkg,
3998  __kmp_nThreadsPerCore, __kmp_ncores);
3999  __kmp_str_buf_free(&buf);
4000  }
4001 _exit:
4002  if (__kmp_pu_os_idx != NULL) {
4003  __kmp_free(__kmp_pu_os_idx);
4004  __kmp_pu_os_idx = NULL;
4005  }
4006 }
4007 
4008 // This function figures out the deepest level at which there is at least one
4009 // cluster/core with more than one processing unit bound to it.
4010 static int __kmp_affinity_find_core_level(const AddrUnsPair *address2os,
4011  int nprocs, int bottom_level) {
4012  int core_level = 0;
4013 
4014  for (int i = 0; i < nprocs; i++) {
4015  for (int j = bottom_level; j > 0; j--) {
4016  if (address2os[i].first.labels[j] > 0) {
4017  if (core_level < (j - 1)) {
4018  core_level = j - 1;
4019  }
4020  }
4021  }
4022  }
4023  return core_level;
4024 }
4025 
4026 // This function counts number of clusters/cores at given level.
4027 static int __kmp_affinity_compute_ncores(const AddrUnsPair *address2os,
4028  int nprocs, int bottom_level,
4029  int core_level) {
4030  int ncores = 0;
4031  int i, j;
4032 
4033  j = bottom_level;
4034  for (i = 0; i < nprocs; i++) {
4035  for (j = bottom_level; j > core_level; j--) {
4036  if ((i + 1) < nprocs) {
4037  if (address2os[i + 1].first.labels[j] > 0) {
4038  break;
4039  }
4040  }
4041  }
4042  if (j == core_level) {
4043  ncores++;
4044  }
4045  }
4046  if (j > core_level) {
4047  // In case of ( nprocs < __kmp_avail_proc ) we may end too deep and miss one
4048  // core. May occur when called from __kmp_affinity_find_core().
4049  ncores++;
4050  }
4051  return ncores;
4052 }
4053 
4054 // This function finds to which cluster/core given processing unit is bound.
4055 static int __kmp_affinity_find_core(const AddrUnsPair *address2os, int proc,
4056  int bottom_level, int core_level) {
4057  return __kmp_affinity_compute_ncores(address2os, proc + 1, bottom_level,
4058  core_level) -
4059  1;
4060 }
4061 
4062 // This function finds maximal number of processing units bound to a
4063 // cluster/core at given level.
4064 static int __kmp_affinity_max_proc_per_core(const AddrUnsPair *address2os,
4065  int nprocs, int bottom_level,
4066  int core_level) {
4067  int maxprocpercore = 0;
4068 
4069  if (core_level < bottom_level) {
4070  for (int i = 0; i < nprocs; i++) {
4071  int percore = address2os[i].first.labels[core_level + 1] + 1;
4072 
4073  if (percore > maxprocpercore) {
4074  maxprocpercore = percore;
4075  }
4076  }
4077  } else {
4078  maxprocpercore = 1;
4079  }
4080  return maxprocpercore;
4081 }
4082 
4083 static AddrUnsPair *address2os = NULL;
4084 static int *procarr = NULL;
4085 static int __kmp_aff_depth = 0;
4086 
4087 #if KMP_USE_HIER_SCHED
4088 #define KMP_EXIT_AFF_NONE \
4089  KMP_ASSERT(__kmp_affinity_type == affinity_none); \
4090  KMP_ASSERT(address2os == NULL); \
4091  __kmp_apply_thread_places(NULL, 0); \
4092  __kmp_create_affinity_none_places(); \
4093  __kmp_dispatch_set_hierarchy_values(); \
4094  return;
4095 #else
4096 #define KMP_EXIT_AFF_NONE \
4097  KMP_ASSERT(__kmp_affinity_type == affinity_none); \
4098  KMP_ASSERT(address2os == NULL); \
4099  __kmp_apply_thread_places(NULL, 0); \
4100  __kmp_create_affinity_none_places(); \
4101  return;
4102 #endif
4103 
4104 // Create a one element mask array (set of places) which only contains the
4105 // initial process's affinity mask
4106 static void __kmp_create_affinity_none_places() {
4107  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4108  KMP_ASSERT(__kmp_affinity_type == affinity_none);
4109  __kmp_affinity_num_masks = 1;
4110  KMP_CPU_ALLOC_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks);
4111  kmp_affin_mask_t *dest = KMP_CPU_INDEX(__kmp_affinity_masks, 0);
4112  KMP_CPU_COPY(dest, __kmp_affin_fullMask);
4113 }
4114 
4115 static int __kmp_affinity_cmp_Address_child_num(const void *a, const void *b) {
4116  const Address *aa = &(((const AddrUnsPair *)a)->first);
4117  const Address *bb = &(((const AddrUnsPair *)b)->first);
4118  unsigned depth = aa->depth;
4119  unsigned i;
4120  KMP_DEBUG_ASSERT(depth == bb->depth);
4121  KMP_DEBUG_ASSERT((unsigned)__kmp_affinity_compact <= depth);
4122  KMP_DEBUG_ASSERT(__kmp_affinity_compact >= 0);
4123  for (i = 0; i < (unsigned)__kmp_affinity_compact; i++) {
4124  int j = depth - i - 1;
4125  if (aa->childNums[j] < bb->childNums[j])
4126  return -1;
4127  if (aa->childNums[j] > bb->childNums[j])
4128  return 1;
4129  }
4130  for (; i < depth; i++) {
4131  int j = i - __kmp_affinity_compact;
4132  if (aa->childNums[j] < bb->childNums[j])
4133  return -1;
4134  if (aa->childNums[j] > bb->childNums[j])
4135  return 1;
4136  }
4137  return 0;
4138 }
4139 
4140 static void __kmp_aux_affinity_initialize(void) {
4141  if (__kmp_affinity_masks != NULL) {
4142  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4143  return;
4144  }
4145 
4146  // Create the "full" mask - this defines all of the processors that we
4147  // consider to be in the machine model. If respect is set, then it is the
4148  // initialization thread's affinity mask. Otherwise, it is all processors that
4149  // we know about on the machine.
4150  if (__kmp_affin_fullMask == NULL) {
4151  KMP_CPU_ALLOC(__kmp_affin_fullMask);
4152  }
4153  if (KMP_AFFINITY_CAPABLE()) {
4154  if (__kmp_affinity_respect_mask) {
4155  __kmp_get_system_affinity(__kmp_affin_fullMask, TRUE);
4156 
4157  // Count the number of available processors.
4158  unsigned i;
4159  __kmp_avail_proc = 0;
4160  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
4161  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
4162  continue;
4163  }
4164  __kmp_avail_proc++;
4165  }
4166  if (__kmp_avail_proc > __kmp_xproc) {
4167  if (__kmp_affinity_verbose ||
4168  (__kmp_affinity_warnings &&
4169  (__kmp_affinity_type != affinity_none))) {
4170  KMP_WARNING(ErrorInitializeAffinity);
4171  }
4172  __kmp_affinity_type = affinity_none;
4173  KMP_AFFINITY_DISABLE();
4174  return;
4175  }
4176  } else {
4177  __kmp_affinity_entire_machine_mask(__kmp_affin_fullMask);
4178  __kmp_avail_proc = __kmp_xproc;
4179  }
4180  }
4181 
4182  if (__kmp_affinity_gran == affinity_gran_tile &&
4183  // check if user's request is valid
4184  __kmp_affinity_dispatch->get_api_type() == KMPAffinity::NATIVE_OS) {
4185  KMP_WARNING(AffTilesNoHWLOC, "KMP_AFFINITY");
4186  __kmp_affinity_gran = affinity_gran_package;
4187  }
4188 
4189  int depth = -1;
4190  kmp_i18n_id_t msg_id = kmp_i18n_null;
4191 
4192  // For backward compatibility, setting KMP_CPUINFO_FILE =>
4193  // KMP_TOPOLOGY_METHOD=cpuinfo
4194  if ((__kmp_cpuinfo_file != NULL) &&
4195  (__kmp_affinity_top_method == affinity_top_method_all)) {
4196  __kmp_affinity_top_method = affinity_top_method_cpuinfo;
4197  }
4198 
4199  if (__kmp_affinity_top_method == affinity_top_method_all) {
4200  // In the default code path, errors are not fatal - we just try using
4201  // another method. We only emit a warning message if affinity is on, or the
4202  // verbose flag is set, an the nowarnings flag was not set.
4203  const char *file_name = NULL;
4204  int line = 0;
4205 #if KMP_USE_HWLOC
4206  if (depth < 0 &&
4207  __kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC) {
4208  if (__kmp_affinity_verbose) {
4209  KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
4210  }
4211  if (!__kmp_hwloc_error) {
4212  depth = __kmp_affinity_create_hwloc_map(&address2os, &msg_id);
4213  if (depth == 0) {
4214  KMP_EXIT_AFF_NONE;
4215  } else if (depth < 0 && __kmp_affinity_verbose) {
4216  KMP_INFORM(AffIgnoringHwloc, "KMP_AFFINITY");
4217  }
4218  } else if (__kmp_affinity_verbose) {
4219  KMP_INFORM(AffIgnoringHwloc, "KMP_AFFINITY");
4220  }
4221  }
4222 #endif
4223 
4224 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4225 
4226  if (depth < 0) {
4227  if (__kmp_affinity_verbose) {
4228  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(Decodingx2APIC));
4229  }
4230 
4231  file_name = NULL;
4232  depth = __kmp_affinity_create_x2apicid_map(&address2os, &msg_id);
4233  if (depth == 0) {
4234  KMP_EXIT_AFF_NONE;
4235  }
4236 
4237  if (depth < 0) {
4238  if (__kmp_affinity_verbose) {
4239  if (msg_id != kmp_i18n_null) {
4240  KMP_INFORM(AffInfoStrStr, "KMP_AFFINITY",
4241  __kmp_i18n_catgets(msg_id),
4242  KMP_I18N_STR(DecodingLegacyAPIC));
4243  } else {
4244  KMP_INFORM(AffInfoStr, "KMP_AFFINITY",
4245  KMP_I18N_STR(DecodingLegacyAPIC));
4246  }
4247  }
4248 
4249  file_name = NULL;
4250  depth = __kmp_affinity_create_apicid_map(&address2os, &msg_id);
4251  if (depth == 0) {
4252  KMP_EXIT_AFF_NONE;
4253  }
4254  }
4255  }
4256 
4257 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4258 
4259 #if KMP_OS_LINUX
4260 
4261  if (depth < 0) {
4262  if (__kmp_affinity_verbose) {
4263  if (msg_id != kmp_i18n_null) {
4264  KMP_INFORM(AffStrParseFilename, "KMP_AFFINITY",
4265  __kmp_i18n_catgets(msg_id), "/proc/cpuinfo");
4266  } else {
4267  KMP_INFORM(AffParseFilename, "KMP_AFFINITY", "/proc/cpuinfo");
4268  }
4269  }
4270 
4271  FILE *f = fopen("/proc/cpuinfo", "r");
4272  if (f == NULL) {
4273  msg_id = kmp_i18n_str_CantOpenCpuinfo;
4274  } else {
4275  file_name = "/proc/cpuinfo";
4276  depth =
4277  __kmp_affinity_create_cpuinfo_map(&address2os, &line, &msg_id, f);
4278  fclose(f);
4279  if (depth == 0) {
4280  KMP_EXIT_AFF_NONE;
4281  }
4282  }
4283  }
4284 
4285 #endif /* KMP_OS_LINUX */
4286 
4287 #if KMP_GROUP_AFFINITY
4288 
4289  if ((depth < 0) && (__kmp_num_proc_groups > 1)) {
4290  if (__kmp_affinity_verbose) {
4291  KMP_INFORM(AffWindowsProcGroupMap, "KMP_AFFINITY");
4292  }
4293 
4294  depth = __kmp_affinity_create_proc_group_map(&address2os, &msg_id);
4295  KMP_ASSERT(depth != 0);
4296  }
4297 
4298 #endif /* KMP_GROUP_AFFINITY */
4299 
4300  if (depth < 0) {
4301  if (__kmp_affinity_verbose && (msg_id != kmp_i18n_null)) {
4302  if (file_name == NULL) {
4303  KMP_INFORM(UsingFlatOS, __kmp_i18n_catgets(msg_id));
4304  } else if (line == 0) {
4305  KMP_INFORM(UsingFlatOSFile, file_name, __kmp_i18n_catgets(msg_id));
4306  } else {
4307  KMP_INFORM(UsingFlatOSFileLine, file_name, line,
4308  __kmp_i18n_catgets(msg_id));
4309  }
4310  }
4311  // FIXME - print msg if msg_id = kmp_i18n_null ???
4312 
4313  file_name = "";
4314  depth = __kmp_affinity_create_flat_map(&address2os, &msg_id);
4315  if (depth == 0) {
4316  KMP_EXIT_AFF_NONE;
4317  }
4318  KMP_ASSERT(depth > 0);
4319  KMP_ASSERT(address2os != NULL);
4320  }
4321  }
4322 
4323 #if KMP_USE_HWLOC
4324  else if (__kmp_affinity_top_method == affinity_top_method_hwloc) {
4325  KMP_ASSERT(__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC);
4326  if (__kmp_affinity_verbose) {
4327  KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
4328  }
4329  depth = __kmp_affinity_create_hwloc_map(&address2os, &msg_id);
4330  if (depth == 0) {
4331  KMP_EXIT_AFF_NONE;
4332  }
4333  }
4334 #endif // KMP_USE_HWLOC
4335 
4336 // If the user has specified that a paricular topology discovery method is to be
4337 // used, then we abort if that method fails. The exception is group affinity,
4338 // which might have been implicitly set.
4339 
4340 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4341 
4342  else if (__kmp_affinity_top_method == affinity_top_method_x2apicid) {
4343  if (__kmp_affinity_verbose) {
4344  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(Decodingx2APIC));
4345  }
4346 
4347  depth = __kmp_affinity_create_x2apicid_map(&address2os, &msg_id);
4348  if (depth == 0) {
4349  KMP_EXIT_AFF_NONE;
4350  }
4351  if (depth < 0) {
4352  KMP_ASSERT(msg_id != kmp_i18n_null);
4353  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4354  }
4355  } else if (__kmp_affinity_top_method == affinity_top_method_apicid) {
4356  if (__kmp_affinity_verbose) {
4357  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(DecodingLegacyAPIC));
4358  }
4359 
4360  depth = __kmp_affinity_create_apicid_map(&address2os, &msg_id);
4361  if (depth == 0) {
4362  KMP_EXIT_AFF_NONE;
4363  }
4364  if (depth < 0) {
4365  KMP_ASSERT(msg_id != kmp_i18n_null);
4366  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4367  }
4368  }
4369 
4370 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4371 
4372  else if (__kmp_affinity_top_method == affinity_top_method_cpuinfo) {
4373  const char *filename;
4374  if (__kmp_cpuinfo_file != NULL) {
4375  filename = __kmp_cpuinfo_file;
4376  } else {
4377  filename = "/proc/cpuinfo";
4378  }
4379 
4380  if (__kmp_affinity_verbose) {
4381  KMP_INFORM(AffParseFilename, "KMP_AFFINITY", filename);
4382  }
4383 
4384  FILE *f = fopen(filename, "r");
4385  if (f == NULL) {
4386  int code = errno;
4387  if (__kmp_cpuinfo_file != NULL) {
4388  __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4389  KMP_HNT(NameComesFrom_CPUINFO_FILE), __kmp_msg_null);
4390  } else {
4391  __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4392  __kmp_msg_null);
4393  }
4394  }
4395  int line = 0;
4396  depth = __kmp_affinity_create_cpuinfo_map(&address2os, &line, &msg_id, f);
4397  fclose(f);
4398  if (depth < 0) {
4399  KMP_ASSERT(msg_id != kmp_i18n_null);
4400  if (line > 0) {
4401  KMP_FATAL(FileLineMsgExiting, filename, line,
4402  __kmp_i18n_catgets(msg_id));
4403  } else {
4404  KMP_FATAL(FileMsgExiting, filename, __kmp_i18n_catgets(msg_id));
4405  }
4406  }
4407  if (__kmp_affinity_type == affinity_none) {
4408  KMP_ASSERT(depth == 0);
4409  KMP_EXIT_AFF_NONE;
4410  }
4411  }
4412 
4413 #if KMP_GROUP_AFFINITY
4414 
4415  else if (__kmp_affinity_top_method == affinity_top_method_group) {
4416  if (__kmp_affinity_verbose) {
4417  KMP_INFORM(AffWindowsProcGroupMap, "KMP_AFFINITY");
4418  }
4419 
4420  depth = __kmp_affinity_create_proc_group_map(&address2os, &msg_id);
4421  KMP_ASSERT(depth != 0);
4422  if (depth < 0) {
4423  KMP_ASSERT(msg_id != kmp_i18n_null);
4424  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4425  }
4426  }
4427 
4428 #endif /* KMP_GROUP_AFFINITY */
4429 
4430  else if (__kmp_affinity_top_method == affinity_top_method_flat) {
4431  if (__kmp_affinity_verbose) {
4432  KMP_INFORM(AffUsingFlatOS, "KMP_AFFINITY");
4433  }
4434 
4435  depth = __kmp_affinity_create_flat_map(&address2os, &msg_id);
4436  if (depth == 0) {
4437  KMP_EXIT_AFF_NONE;
4438  }
4439  // should not fail
4440  KMP_ASSERT(depth > 0);
4441  KMP_ASSERT(address2os != NULL);
4442  }
4443 
4444 #if KMP_USE_HIER_SCHED
4445  __kmp_dispatch_set_hierarchy_values();
4446 #endif
4447 
4448  if (address2os == NULL) {
4449  if (KMP_AFFINITY_CAPABLE() &&
4450  (__kmp_affinity_verbose ||
4451  (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none)))) {
4452  KMP_WARNING(ErrorInitializeAffinity);
4453  }
4454  __kmp_affinity_type = affinity_none;
4455  __kmp_create_affinity_none_places();
4456  KMP_AFFINITY_DISABLE();
4457  return;
4458  }
4459 
4460  if (__kmp_affinity_gran == affinity_gran_tile
4461 #if KMP_USE_HWLOC
4462  && __kmp_tile_depth == 0
4463 #endif
4464  ) {
4465  // tiles requested but not detected, warn user on this
4466  KMP_WARNING(AffTilesNoTiles, "KMP_AFFINITY");
4467  }
4468 
4469  __kmp_apply_thread_places(&address2os, depth);
4470 
4471  // Create the table of masks, indexed by thread Id.
4472  unsigned maxIndex;
4473  unsigned numUnique;
4474  kmp_affin_mask_t *osId2Mask =
4475  __kmp_create_masks(&maxIndex, &numUnique, address2os, __kmp_avail_proc);
4476  if (__kmp_affinity_gran_levels == 0) {
4477  KMP_DEBUG_ASSERT((int)numUnique == __kmp_avail_proc);
4478  }
4479 
4480  // Set the childNums vector in all Address objects. This must be done before
4481  // we can sort using __kmp_affinity_cmp_Address_child_num(), which takes into
4482  // account the setting of __kmp_affinity_compact.
4483  __kmp_affinity_assign_child_nums(address2os, __kmp_avail_proc);
4484 
4485  switch (__kmp_affinity_type) {
4486 
4487  case affinity_explicit:
4488  KMP_DEBUG_ASSERT(__kmp_affinity_proclist != NULL);
4489 #if OMP_40_ENABLED
4490  if (__kmp_nested_proc_bind.bind_types[0] == proc_bind_intel)
4491 #endif
4492  {
4493  __kmp_affinity_process_proclist(
4494  &__kmp_affinity_masks, &__kmp_affinity_num_masks,
4495  __kmp_affinity_proclist, osId2Mask, maxIndex);
4496  }
4497 #if OMP_40_ENABLED
4498  else {
4499  __kmp_affinity_process_placelist(
4500  &__kmp_affinity_masks, &__kmp_affinity_num_masks,
4501  __kmp_affinity_proclist, osId2Mask, maxIndex);
4502  }
4503 #endif
4504  if (__kmp_affinity_num_masks == 0) {
4505  if (__kmp_affinity_verbose ||
4506  (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) {
4507  KMP_WARNING(AffNoValidProcID);
4508  }
4509  __kmp_affinity_type = affinity_none;
4510  __kmp_create_affinity_none_places();
4511  return;
4512  }
4513  break;
4514 
4515  // The other affinity types rely on sorting the Addresses according to some
4516  // permutation of the machine topology tree. Set __kmp_affinity_compact and
4517  // __kmp_affinity_offset appropriately, then jump to a common code fragment
4518  // to do the sort and create the array of affinity masks.
4519 
4520  case affinity_logical:
4521  __kmp_affinity_compact = 0;
4522  if (__kmp_affinity_offset) {
4523  __kmp_affinity_offset =
4524  __kmp_nThreadsPerCore * __kmp_affinity_offset % __kmp_avail_proc;
4525  }
4526  goto sortAddresses;
4527 
4528  case affinity_physical:
4529  if (__kmp_nThreadsPerCore > 1) {
4530  __kmp_affinity_compact = 1;
4531  if (__kmp_affinity_compact >= depth) {
4532  __kmp_affinity_compact = 0;
4533  }
4534  } else {
4535  __kmp_affinity_compact = 0;
4536  }
4537  if (__kmp_affinity_offset) {
4538  __kmp_affinity_offset =
4539  __kmp_nThreadsPerCore * __kmp_affinity_offset % __kmp_avail_proc;
4540  }
4541  goto sortAddresses;
4542 
4543  case affinity_scatter:
4544  if (__kmp_affinity_compact >= depth) {
4545  __kmp_affinity_compact = 0;
4546  } else {
4547  __kmp_affinity_compact = depth - 1 - __kmp_affinity_compact;
4548  }
4549  goto sortAddresses;
4550 
4551  case affinity_compact:
4552  if (__kmp_affinity_compact >= depth) {
4553  __kmp_affinity_compact = depth - 1;
4554  }
4555  goto sortAddresses;
4556 
4557  case affinity_balanced:
4558  if (depth <= 1) {
4559  if (__kmp_affinity_verbose || __kmp_affinity_warnings) {
4560  KMP_WARNING(AffBalancedNotAvail, "KMP_AFFINITY");
4561  }
4562  __kmp_affinity_type = affinity_none;
4563  __kmp_create_affinity_none_places();
4564  return;
4565  } else if (!__kmp_affinity_uniform_topology()) {
4566  // Save the depth for further usage
4567  __kmp_aff_depth = depth;
4568 
4569  int core_level = __kmp_affinity_find_core_level(
4570  address2os, __kmp_avail_proc, depth - 1);
4571  int ncores = __kmp_affinity_compute_ncores(address2os, __kmp_avail_proc,
4572  depth - 1, core_level);
4573  int maxprocpercore = __kmp_affinity_max_proc_per_core(
4574  address2os, __kmp_avail_proc, depth - 1, core_level);
4575 
4576  int nproc = ncores * maxprocpercore;
4577  if ((nproc < 2) || (nproc < __kmp_avail_proc)) {
4578  if (__kmp_affinity_verbose || __kmp_affinity_warnings) {
4579  KMP_WARNING(AffBalancedNotAvail, "KMP_AFFINITY");
4580  }
4581  __kmp_affinity_type = affinity_none;
4582  return;
4583  }
4584 
4585  procarr = (int *)__kmp_allocate(sizeof(int) * nproc);
4586  for (int i = 0; i < nproc; i++) {
4587  procarr[i] = -1;
4588  }
4589 
4590  int lastcore = -1;
4591  int inlastcore = 0;
4592  for (int i = 0; i < __kmp_avail_proc; i++) {
4593  int proc = address2os[i].second;
4594  int core =
4595  __kmp_affinity_find_core(address2os, i, depth - 1, core_level);
4596 
4597  if (core == lastcore) {
4598  inlastcore++;
4599  } else {
4600  inlastcore = 0;
4601  }
4602  lastcore = core;
4603 
4604  procarr[core * maxprocpercore + inlastcore] = proc;
4605  }
4606  }
4607  if (__kmp_affinity_compact >= depth) {
4608  __kmp_affinity_compact = depth - 1;
4609  }
4610 
4611  sortAddresses:
4612  // Allocate the gtid->affinity mask table.
4613  if (__kmp_affinity_dups) {
4614  __kmp_affinity_num_masks = __kmp_avail_proc;
4615  } else {
4616  __kmp_affinity_num_masks = numUnique;
4617  }
4618 
4619 #if OMP_40_ENABLED
4620  if ((__kmp_nested_proc_bind.bind_types[0] != proc_bind_intel) &&
4621  (__kmp_affinity_num_places > 0) &&
4622  ((unsigned)__kmp_affinity_num_places < __kmp_affinity_num_masks)) {
4623  __kmp_affinity_num_masks = __kmp_affinity_num_places;
4624  }
4625 #endif
4626 
4627  KMP_CPU_ALLOC_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks);
4628 
4629  // Sort the address2os table according to the current setting of
4630  // __kmp_affinity_compact, then fill out __kmp_affinity_masks.
4631  qsort(address2os, __kmp_avail_proc, sizeof(*address2os),
4632  __kmp_affinity_cmp_Address_child_num);
4633  {
4634  int i;
4635  unsigned j;
4636  for (i = 0, j = 0; i < __kmp_avail_proc; i++) {
4637  if ((!__kmp_affinity_dups) && (!address2os[i].first.leader)) {
4638  continue;
4639  }
4640  unsigned osId = address2os[i].second;
4641  kmp_affin_mask_t *src = KMP_CPU_INDEX(osId2Mask, osId);
4642  kmp_affin_mask_t *dest = KMP_CPU_INDEX(__kmp_affinity_masks, j);
4643  KMP_ASSERT(KMP_CPU_ISSET(osId, src));
4644  KMP_CPU_COPY(dest, src);
4645  if (++j >= __kmp_affinity_num_masks) {
4646  break;
4647  }
4648  }
4649  KMP_DEBUG_ASSERT(j == __kmp_affinity_num_masks);
4650  }
4651  break;
4652 
4653  default:
4654  KMP_ASSERT2(0, "Unexpected affinity setting");
4655  }
4656 
4657  KMP_CPU_FREE_ARRAY(osId2Mask, maxIndex + 1);
4658  machine_hierarchy.init(address2os, __kmp_avail_proc);
4659 }
4660 #undef KMP_EXIT_AFF_NONE
4661 
4662 void __kmp_affinity_initialize(void) {
4663  // Much of the code above was written assumming that if a machine was not
4664  // affinity capable, then __kmp_affinity_type == affinity_none. We now
4665  // explicitly represent this as __kmp_affinity_type == affinity_disabled.
4666  // There are too many checks for __kmp_affinity_type == affinity_none
4667  // in this code. Instead of trying to change them all, check if
4668  // __kmp_affinity_type == affinity_disabled, and if so, slam it with
4669  // affinity_none, call the real initialization routine, then restore
4670  // __kmp_affinity_type to affinity_disabled.
4671  int disabled = (__kmp_affinity_type == affinity_disabled);
4672  if (!KMP_AFFINITY_CAPABLE()) {
4673  KMP_ASSERT(disabled);
4674  }
4675  if (disabled) {
4676  __kmp_affinity_type = affinity_none;
4677  }
4678  __kmp_aux_affinity_initialize();
4679  if (disabled) {
4680  __kmp_affinity_type = affinity_disabled;
4681  }
4682 }
4683 
4684 void __kmp_affinity_uninitialize(void) {
4685  if (__kmp_affinity_masks != NULL) {
4686  KMP_CPU_FREE_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks);
4687  __kmp_affinity_masks = NULL;
4688  }
4689  if (__kmp_affin_fullMask != NULL) {
4690  KMP_CPU_FREE(__kmp_affin_fullMask);
4691  __kmp_affin_fullMask = NULL;
4692  }
4693  __kmp_affinity_num_masks = 0;
4694  __kmp_affinity_type = affinity_default;
4695 #if OMP_40_ENABLED
4696  __kmp_affinity_num_places = 0;
4697 #endif
4698  if (__kmp_affinity_proclist != NULL) {
4699  __kmp_free(__kmp_affinity_proclist);
4700  __kmp_affinity_proclist = NULL;
4701  }
4702  if (address2os != NULL) {
4703  __kmp_free(address2os);
4704  address2os = NULL;
4705  }
4706  if (procarr != NULL) {
4707  __kmp_free(procarr);
4708  procarr = NULL;
4709  }
4710 #if KMP_USE_HWLOC
4711  if (__kmp_hwloc_topology != NULL) {
4712  hwloc_topology_destroy(__kmp_hwloc_topology);
4713  __kmp_hwloc_topology = NULL;
4714  }
4715 #endif
4716  KMPAffinity::destroy_api();
4717 }
4718 
4719 void __kmp_affinity_set_init_mask(int gtid, int isa_root) {
4720  if (!KMP_AFFINITY_CAPABLE()) {
4721  return;
4722  }
4723 
4724  kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
4725  if (th->th.th_affin_mask == NULL) {
4726  KMP_CPU_ALLOC(th->th.th_affin_mask);
4727  } else {
4728  KMP_CPU_ZERO(th->th.th_affin_mask);
4729  }
4730 
4731  // Copy the thread mask to the kmp_info_t strucuture. If
4732  // __kmp_affinity_type == affinity_none, copy the "full" mask, i.e. one that
4733  // has all of the OS proc ids set, or if __kmp_affinity_respect_mask is set,
4734  // then the full mask is the same as the mask of the initialization thread.
4735  kmp_affin_mask_t *mask;
4736  int i;
4737 
4738 #if OMP_40_ENABLED
4739  if (KMP_AFFINITY_NON_PROC_BIND)
4740 #endif
4741  {
4742  if ((__kmp_affinity_type == affinity_none) ||
4743  (__kmp_affinity_type == affinity_balanced)) {
4744 #if KMP_GROUP_AFFINITY
4745  if (__kmp_num_proc_groups > 1) {
4746  return;
4747  }
4748 #endif
4749  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4750  i = 0;
4751  mask = __kmp_affin_fullMask;
4752  } else {
4753  KMP_DEBUG_ASSERT(__kmp_affinity_num_masks > 0);
4754  i = (gtid + __kmp_affinity_offset) % __kmp_affinity_num_masks;
4755  mask = KMP_CPU_INDEX(__kmp_affinity_masks, i);
4756  }
4757  }
4758 #if OMP_40_ENABLED
4759  else {
4760  if ((!isa_root) ||
4761  (__kmp_nested_proc_bind.bind_types[0] == proc_bind_false)) {
4762 #if KMP_GROUP_AFFINITY
4763  if (__kmp_num_proc_groups > 1) {
4764  return;
4765  }
4766 #endif
4767  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4768  i = KMP_PLACE_ALL;
4769  mask = __kmp_affin_fullMask;
4770  } else {
4771  // int i = some hash function or just a counter that doesn't
4772  // always start at 0. Use gtid for now.
4773  KMP_DEBUG_ASSERT(__kmp_affinity_num_masks > 0);
4774  i = (gtid + __kmp_affinity_offset) % __kmp_affinity_num_masks;
4775  mask = KMP_CPU_INDEX(__kmp_affinity_masks, i);
4776  }
4777  }
4778 #endif
4779 
4780 #if OMP_40_ENABLED
4781  th->th.th_current_place = i;
4782  if (isa_root) {
4783  th->th.th_new_place = i;
4784  th->th.th_first_place = 0;
4785  th->th.th_last_place = __kmp_affinity_num_masks - 1;
4786  } else if (KMP_AFFINITY_NON_PROC_BIND) {
4787  // When using a Non-OMP_PROC_BIND affinity method,
4788  // set all threads' place-partition-var to the entire place list
4789  th->th.th_first_place = 0;
4790  th->th.th_last_place = __kmp_affinity_num_masks - 1;
4791  }
4792 
4793  if (i == KMP_PLACE_ALL) {
4794  KA_TRACE(100, ("__kmp_affinity_set_init_mask: binding T#%d to all places\n",
4795  gtid));
4796  } else {
4797  KA_TRACE(100, ("__kmp_affinity_set_init_mask: binding T#%d to place %d\n",
4798  gtid, i));
4799  }
4800 #else
4801  if (i == -1) {
4802  KA_TRACE(
4803  100,
4804  ("__kmp_affinity_set_init_mask: binding T#%d to __kmp_affin_fullMask\n",
4805  gtid));
4806  } else {
4807  KA_TRACE(100, ("__kmp_affinity_set_init_mask: binding T#%d to mask %d\n",
4808  gtid, i));
4809  }
4810 #endif /* OMP_40_ENABLED */
4811 
4812  KMP_CPU_COPY(th->th.th_affin_mask, mask);
4813 
4814  if (__kmp_affinity_verbose
4815  /* to avoid duplicate printing (will be correctly printed on barrier) */
4816  && (__kmp_affinity_type == affinity_none ||
4817  (i != KMP_PLACE_ALL && __kmp_affinity_type != affinity_balanced))) {
4818  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4819  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4820  th->th.th_affin_mask);
4821  KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY", (kmp_int32)getpid(),
4822  __kmp_gettid(), gtid, buf);
4823  }
4824 
4825 #if KMP_OS_WINDOWS
4826  // On Windows* OS, the process affinity mask might have changed. If the user
4827  // didn't request affinity and this call fails, just continue silently.
4828  // See CQ171393.
4829  if (__kmp_affinity_type == affinity_none) {
4830  __kmp_set_system_affinity(th->th.th_affin_mask, FALSE);
4831  } else
4832 #endif
4833  __kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
4834 }
4835 
4836 #if OMP_40_ENABLED
4837 
4838 void __kmp_affinity_set_place(int gtid) {
4839  if (!KMP_AFFINITY_CAPABLE()) {
4840  return;
4841  }
4842 
4843  kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
4844 
4845  KA_TRACE(100, ("__kmp_affinity_set_place: binding T#%d to place %d (current "
4846  "place = %d)\n",
4847  gtid, th->th.th_new_place, th->th.th_current_place));
4848 
4849  // Check that the new place is within this thread's partition.
4850  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
4851  KMP_ASSERT(th->th.th_new_place >= 0);
4852  KMP_ASSERT((unsigned)th->th.th_new_place <= __kmp_affinity_num_masks);
4853  if (th->th.th_first_place <= th->th.th_last_place) {
4854  KMP_ASSERT((th->th.th_new_place >= th->th.th_first_place) &&
4855  (th->th.th_new_place <= th->th.th_last_place));
4856  } else {
4857  KMP_ASSERT((th->th.th_new_place <= th->th.th_first_place) ||
4858  (th->th.th_new_place >= th->th.th_last_place));
4859  }
4860 
4861  // Copy the thread mask to the kmp_info_t strucuture,
4862  // and set this thread's affinity.
4863  kmp_affin_mask_t *mask =
4864  KMP_CPU_INDEX(__kmp_affinity_masks, th->th.th_new_place);
4865  KMP_CPU_COPY(th->th.th_affin_mask, mask);
4866  th->th.th_current_place = th->th.th_new_place;
4867 
4868  if (__kmp_affinity_verbose) {
4869  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4870  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4871  th->th.th_affin_mask);
4872  KMP_INFORM(BoundToOSProcSet, "OMP_PROC_BIND", (kmp_int32)getpid(),
4873  __kmp_gettid(), gtid, buf);
4874  }
4875  __kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
4876 }
4877 
4878 #endif /* OMP_40_ENABLED */
4879 
4880 int __kmp_aux_set_affinity(void **mask) {
4881  int gtid;
4882  kmp_info_t *th;
4883  int retval;
4884 
4885  if (!KMP_AFFINITY_CAPABLE()) {
4886  return -1;
4887  }
4888 
4889  gtid = __kmp_entry_gtid();
4890  KA_TRACE(1000, (""); {
4891  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4892  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4893  (kmp_affin_mask_t *)(*mask));
4894  __kmp_debug_printf(
4895  "kmp_set_affinity: setting affinity mask for thread %d = %s\n", gtid,
4896  buf);
4897  });
4898 
4899  if (__kmp_env_consistency_check) {
4900  if ((mask == NULL) || (*mask == NULL)) {
4901  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4902  } else {
4903  unsigned proc;
4904  int num_procs = 0;
4905 
4906  KMP_CPU_SET_ITERATE(proc, ((kmp_affin_mask_t *)(*mask))) {
4907  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
4908  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4909  }
4910  if (!KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask))) {
4911  continue;
4912  }
4913  num_procs++;
4914  }
4915  if (num_procs == 0) {
4916  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4917  }
4918 
4919 #if KMP_GROUP_AFFINITY
4920  if (__kmp_get_proc_group((kmp_affin_mask_t *)(*mask)) < 0) {
4921  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4922  }
4923 #endif /* KMP_GROUP_AFFINITY */
4924  }
4925  }
4926 
4927  th = __kmp_threads[gtid];
4928  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
4929  retval = __kmp_set_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
4930  if (retval == 0) {
4931  KMP_CPU_COPY(th->th.th_affin_mask, (kmp_affin_mask_t *)(*mask));
4932  }
4933 
4934 #if OMP_40_ENABLED
4935  th->th.th_current_place = KMP_PLACE_UNDEFINED;
4936  th->th.th_new_place = KMP_PLACE_UNDEFINED;
4937  th->th.th_first_place = 0;
4938  th->th.th_last_place = __kmp_affinity_num_masks - 1;
4939 
4940  // Turn off 4.0 affinity for the current tread at this parallel level.
4941  th->th.th_current_task->td_icvs.proc_bind = proc_bind_false;
4942 #endif
4943 
4944  return retval;
4945 }
4946 
4947 int __kmp_aux_get_affinity(void **mask) {
4948  int gtid;
4949  int retval;
4950  kmp_info_t *th;
4951 
4952  if (!KMP_AFFINITY_CAPABLE()) {
4953  return -1;
4954  }
4955 
4956  gtid = __kmp_entry_gtid();
4957  th = __kmp_threads[gtid];
4958  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
4959 
4960  KA_TRACE(1000, (""); {
4961  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4962  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4963  th->th.th_affin_mask);
4964  __kmp_printf("kmp_get_affinity: stored affinity mask for thread %d = %s\n",
4965  gtid, buf);
4966  });
4967 
4968  if (__kmp_env_consistency_check) {
4969  if ((mask == NULL) || (*mask == NULL)) {
4970  KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity");
4971  }
4972  }
4973 
4974 #if !KMP_OS_WINDOWS
4975 
4976  retval = __kmp_get_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
4977  KA_TRACE(1000, (""); {
4978  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4979  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4980  (kmp_affin_mask_t *)(*mask));
4981  __kmp_printf("kmp_get_affinity: system affinity mask for thread %d = %s\n",
4982  gtid, buf);
4983  });
4984  return retval;
4985 
4986 #else
4987 
4988  KMP_CPU_COPY((kmp_affin_mask_t *)(*mask), th->th.th_affin_mask);
4989  return 0;
4990 
4991 #endif /* KMP_OS_WINDOWS */
4992 }
4993 
4994 int __kmp_aux_get_affinity_max_proc() {
4995  if (!KMP_AFFINITY_CAPABLE()) {
4996  return 0;
4997  }
4998 #if KMP_GROUP_AFFINITY
4999  if (__kmp_num_proc_groups > 1) {
5000  return (int)(__kmp_num_proc_groups * sizeof(DWORD_PTR) * CHAR_BIT);
5001  }
5002 #endif
5003  return __kmp_xproc;
5004 }
5005 
5006 int __kmp_aux_set_affinity_mask_proc(int proc, void **mask) {
5007  if (!KMP_AFFINITY_CAPABLE()) {
5008  return -1;
5009  }
5010 
5011  KA_TRACE(1000, (""); {
5012  int gtid = __kmp_entry_gtid();
5013  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5014  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5015  (kmp_affin_mask_t *)(*mask));
5016  __kmp_debug_printf("kmp_set_affinity_mask_proc: setting proc %d in "
5017  "affinity mask for thread %d = %s\n",
5018  proc, gtid, buf);
5019  });
5020 
5021  if (__kmp_env_consistency_check) {
5022  if ((mask == NULL) || (*mask == NULL)) {
5023  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity_mask_proc");
5024  }
5025  }
5026 
5027  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5028  return -1;
5029  }
5030  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5031  return -2;
5032  }
5033 
5034  KMP_CPU_SET(proc, (kmp_affin_mask_t *)(*mask));
5035  return 0;
5036 }
5037 
5038 int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask) {
5039  if (!KMP_AFFINITY_CAPABLE()) {
5040  return -1;
5041  }
5042 
5043  KA_TRACE(1000, (""); {
5044  int gtid = __kmp_entry_gtid();
5045  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5046  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5047  (kmp_affin_mask_t *)(*mask));
5048  __kmp_debug_printf("kmp_unset_affinity_mask_proc: unsetting proc %d in "
5049  "affinity mask for thread %d = %s\n",
5050  proc, gtid, buf);
5051  });
5052 
5053  if (__kmp_env_consistency_check) {
5054  if ((mask == NULL) || (*mask == NULL)) {
5055  KMP_FATAL(AffinityInvalidMask, "kmp_unset_affinity_mask_proc");
5056  }
5057  }
5058 
5059  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5060  return -1;
5061  }
5062  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5063  return -2;
5064  }
5065 
5066  KMP_CPU_CLR(proc, (kmp_affin_mask_t *)(*mask));
5067  return 0;
5068 }
5069 
5070 int __kmp_aux_get_affinity_mask_proc(int proc, void **mask) {
5071  if (!KMP_AFFINITY_CAPABLE()) {
5072  return -1;
5073  }
5074 
5075  KA_TRACE(1000, (""); {
5076  int gtid = __kmp_entry_gtid();
5077  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5078  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5079  (kmp_affin_mask_t *)(*mask));
5080  __kmp_debug_printf("kmp_get_affinity_mask_proc: getting proc %d in "
5081  "affinity mask for thread %d = %s\n",
5082  proc, gtid, buf);
5083  });
5084 
5085  if (__kmp_env_consistency_check) {
5086  if ((mask == NULL) || (*mask == NULL)) {
5087  KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity_mask_proc");
5088  }
5089  }
5090 
5091  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5092  return -1;
5093  }
5094  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5095  return 0;
5096  }
5097 
5098  return KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask));
5099 }
5100 
5101 // Dynamic affinity settings - Affinity balanced
5102 void __kmp_balanced_affinity(kmp_info_t *th, int nthreads) {
5103  KMP_DEBUG_ASSERT(th);
5104  bool fine_gran = true;
5105  int tid = th->th.th_info.ds.ds_tid;
5106 
5107  switch (__kmp_affinity_gran) {
5108  case affinity_gran_fine:
5109  case affinity_gran_thread:
5110  break;
5111  case affinity_gran_core:
5112  if (__kmp_nThreadsPerCore > 1) {
5113  fine_gran = false;
5114  }
5115  break;
5116  case affinity_gran_package:
5117  if (nCoresPerPkg > 1) {
5118  fine_gran = false;
5119  }
5120  break;
5121  default:
5122  fine_gran = false;
5123  }
5124 
5125  if (__kmp_affinity_uniform_topology()) {
5126  int coreID;
5127  int threadID;
5128  // Number of hyper threads per core in HT machine
5129  int __kmp_nth_per_core = __kmp_avail_proc / __kmp_ncores;
5130  // Number of cores
5131  int ncores = __kmp_ncores;
5132  if ((nPackages > 1) && (__kmp_nth_per_core <= 1)) {
5133  __kmp_nth_per_core = __kmp_avail_proc / nPackages;
5134  ncores = nPackages;
5135  }
5136  // How many threads will be bound to each core
5137  int chunk = nthreads / ncores;
5138  // How many cores will have an additional thread bound to it - "big cores"
5139  int big_cores = nthreads % ncores;
5140  // Number of threads on the big cores
5141  int big_nth = (chunk + 1) * big_cores;
5142  if (tid < big_nth) {
5143  coreID = tid / (chunk + 1);
5144  threadID = (tid % (chunk + 1)) % __kmp_nth_per_core;
5145  } else { // tid >= big_nth
5146  coreID = (tid - big_cores) / chunk;
5147  threadID = ((tid - big_cores) % chunk) % __kmp_nth_per_core;
5148  }
5149 
5150  KMP_DEBUG_ASSERT2(KMP_AFFINITY_CAPABLE(),
5151  "Illegal set affinity operation when not capable");
5152 
5153  kmp_affin_mask_t *mask = th->th.th_affin_mask;
5154  KMP_CPU_ZERO(mask);
5155 
5156  if (fine_gran) {
5157  int osID = address2os[coreID * __kmp_nth_per_core + threadID].second;
5158  KMP_CPU_SET(osID, mask);
5159  } else {
5160  for (int i = 0; i < __kmp_nth_per_core; i++) {
5161  int osID;
5162  osID = address2os[coreID * __kmp_nth_per_core + i].second;
5163  KMP_CPU_SET(osID, mask);
5164  }
5165  }
5166  if (__kmp_affinity_verbose) {
5167  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5168  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
5169  KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY", (kmp_int32)getpid(),
5170  __kmp_gettid(), tid, buf);
5171  }
5172  __kmp_set_system_affinity(mask, TRUE);
5173  } else { // Non-uniform topology
5174 
5175  kmp_affin_mask_t *mask = th->th.th_affin_mask;
5176  KMP_CPU_ZERO(mask);
5177 
5178  int core_level = __kmp_affinity_find_core_level(
5179  address2os, __kmp_avail_proc, __kmp_aff_depth - 1);
5180  int ncores = __kmp_affinity_compute_ncores(address2os, __kmp_avail_proc,
5181  __kmp_aff_depth - 1, core_level);
5182  int nth_per_core = __kmp_affinity_max_proc_per_core(
5183  address2os, __kmp_avail_proc, __kmp_aff_depth - 1, core_level);
5184 
5185  // For performance gain consider the special case nthreads ==
5186  // __kmp_avail_proc
5187  if (nthreads == __kmp_avail_proc) {
5188  if (fine_gran) {
5189  int osID = address2os[tid].second;
5190  KMP_CPU_SET(osID, mask);
5191  } else {
5192  int core = __kmp_affinity_find_core(address2os, tid,
5193  __kmp_aff_depth - 1, core_level);
5194  for (int i = 0; i < __kmp_avail_proc; i++) {
5195  int osID = address2os[i].second;
5196  if (__kmp_affinity_find_core(address2os, i, __kmp_aff_depth - 1,
5197  core_level) == core) {
5198  KMP_CPU_SET(osID, mask);
5199  }
5200  }
5201  }
5202  } else if (nthreads <= ncores) {
5203 
5204  int core = 0;
5205  for (int i = 0; i < ncores; i++) {
5206  // Check if this core from procarr[] is in the mask
5207  int in_mask = 0;
5208  for (int j = 0; j < nth_per_core; j++) {
5209  if (procarr[i * nth_per_core + j] != -1) {
5210  in_mask = 1;
5211  break;
5212  }
5213  }
5214  if (in_mask) {
5215  if (tid == core) {
5216  for (int j = 0; j < nth_per_core; j++) {
5217  int osID = procarr[i * nth_per_core + j];
5218  if (osID != -1) {
5219  KMP_CPU_SET(osID, mask);
5220  // For fine granularity it is enough to set the first available
5221  // osID for this core
5222  if (fine_gran) {
5223  break;
5224  }
5225  }
5226  }
5227  break;
5228  } else {
5229  core++;
5230  }
5231  }
5232  }
5233  } else { // nthreads > ncores
5234  // Array to save the number of processors at each core
5235  int *nproc_at_core = (int *)KMP_ALLOCA(sizeof(int) * ncores);
5236  // Array to save the number of cores with "x" available processors;
5237  int *ncores_with_x_procs =
5238  (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
5239  // Array to save the number of cores with # procs from x to nth_per_core
5240  int *ncores_with_x_to_max_procs =
5241  (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
5242 
5243  for (int i = 0; i <= nth_per_core; i++) {
5244  ncores_with_x_procs[i] = 0;
5245  ncores_with_x_to_max_procs[i] = 0;
5246  }
5247 
5248  for (int i = 0; i < ncores; i++) {
5249  int cnt = 0;
5250  for (int j = 0; j < nth_per_core; j++) {
5251  if (procarr[i * nth_per_core + j] != -1) {
5252  cnt++;
5253  }
5254  }
5255  nproc_at_core[i] = cnt;
5256  ncores_with_x_procs[cnt]++;
5257  }
5258 
5259  for (int i = 0; i <= nth_per_core; i++) {
5260  for (int j = i; j <= nth_per_core; j++) {
5261  ncores_with_x_to_max_procs[i] += ncores_with_x_procs[j];
5262  }
5263  }
5264 
5265  // Max number of processors
5266  int nproc = nth_per_core * ncores;
5267  // An array to keep number of threads per each context
5268  int *newarr = (int *)__kmp_allocate(sizeof(int) * nproc);
5269  for (int i = 0; i < nproc; i++) {
5270  newarr[i] = 0;
5271  }
5272 
5273  int nth = nthreads;
5274  int flag = 0;
5275  while (nth > 0) {
5276  for (int j = 1; j <= nth_per_core; j++) {
5277  int cnt = ncores_with_x_to_max_procs[j];
5278  for (int i = 0; i < ncores; i++) {
5279  // Skip the core with 0 processors
5280  if (nproc_at_core[i] == 0) {
5281  continue;
5282  }
5283  for (int k = 0; k < nth_per_core; k++) {
5284  if (procarr[i * nth_per_core + k] != -1) {
5285  if (newarr[i * nth_per_core + k] == 0) {
5286  newarr[i * nth_per_core + k] = 1;
5287  cnt--;
5288  nth--;
5289  break;
5290  } else {
5291  if (flag != 0) {
5292  newarr[i * nth_per_core + k]++;
5293  cnt--;
5294  nth--;
5295  break;
5296  }
5297  }
5298  }
5299  }
5300  if (cnt == 0 || nth == 0) {
5301  break;
5302  }
5303  }
5304  if (nth == 0) {
5305  break;
5306  }
5307  }
5308  flag = 1;
5309  }
5310  int sum = 0;
5311  for (int i = 0; i < nproc; i++) {
5312  sum += newarr[i];
5313  if (sum > tid) {
5314  if (fine_gran) {
5315  int osID = procarr[i];
5316  KMP_CPU_SET(osID, mask);
5317  } else {
5318  int coreID = i / nth_per_core;
5319  for (int ii = 0; ii < nth_per_core; ii++) {
5320  int osID = procarr[coreID * nth_per_core + ii];
5321  if (osID != -1) {
5322  KMP_CPU_SET(osID, mask);
5323  }
5324  }
5325  }
5326  break;
5327  }
5328  }
5329  __kmp_free(newarr);
5330  }
5331 
5332  if (__kmp_affinity_verbose) {
5333  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5334  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
5335  KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY", (kmp_int32)getpid(),
5336  __kmp_gettid(), tid, buf);
5337  }
5338  __kmp_set_system_affinity(mask, TRUE);
5339  }
5340 }
5341 
5342 #if KMP_OS_LINUX
5343 // We don't need this entry for Windows because
5344 // there is GetProcessAffinityMask() api
5345 //
5346 // The intended usage is indicated by these steps:
5347 // 1) The user gets the current affinity mask
5348 // 2) Then sets the affinity by calling this function
5349 // 3) Error check the return value
5350 // 4) Use non-OpenMP parallelization
5351 // 5) Reset the affinity to what was stored in step 1)
5352 #ifdef __cplusplus
5353 extern "C"
5354 #endif
5355  int
5356  kmp_set_thread_affinity_mask_initial()
5357 // the function returns 0 on success,
5358 // -1 if we cannot bind thread
5359 // >0 (errno) if an error happened during binding
5360 {
5361  int gtid = __kmp_get_gtid();
5362  if (gtid < 0) {
5363  // Do not touch non-omp threads
5364  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5365  "non-omp thread, returning\n"));
5366  return -1;
5367  }
5368  if (!KMP_AFFINITY_CAPABLE() || !__kmp_init_middle) {
5369  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5370  "affinity not initialized, returning\n"));
5371  return -1;
5372  }
5373  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5374  "set full mask for thread %d\n",
5375  gtid));
5376  KMP_DEBUG_ASSERT(__kmp_affin_fullMask != NULL);
5377  return __kmp_set_system_affinity(__kmp_affin_fullMask, FALSE);
5378 }
5379 #endif
5380 
5381 #endif // KMP_AFFINITY_SUPPORTED