LLVM OpenMP* Runtime Library
kmp_affinity.cpp
1 /*
2  * kmp_affinity.cpp -- affinity management
3  */
4 
5 //===----------------------------------------------------------------------===//
6 //
7 // The LLVM Compiler Infrastructure
8 //
9 // This file is dual licensed under the MIT and the University of Illinois Open
10 // Source Licenses. See LICENSE.txt for details.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "kmp.h"
15 #include "kmp_affinity.h"
16 #include "kmp_i18n.h"
17 #include "kmp_io.h"
18 #include "kmp_str.h"
19 #include "kmp_wrapper_getpid.h"
20 #if KMP_USE_HIER_SCHED
21 #include "kmp_dispatch_hier.h"
22 #endif
23 
24 // Store the real or imagined machine hierarchy here
25 static hierarchy_info machine_hierarchy;
26 
27 void __kmp_cleanup_hierarchy() { machine_hierarchy.fini(); }
28 
29 void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar) {
30  kmp_uint32 depth;
31  // The test below is true if affinity is available, but set to "none". Need to
32  // init on first use of hierarchical barrier.
33  if (TCR_1(machine_hierarchy.uninitialized))
34  machine_hierarchy.init(NULL, nproc);
35 
36  // Adjust the hierarchy in case num threads exceeds original
37  if (nproc > machine_hierarchy.base_num_threads)
38  machine_hierarchy.resize(nproc);
39 
40  depth = machine_hierarchy.depth;
41  KMP_DEBUG_ASSERT(depth > 0);
42 
43  thr_bar->depth = depth;
44  thr_bar->base_leaf_kids = (kmp_uint8)machine_hierarchy.numPerLevel[0] - 1;
45  thr_bar->skip_per_level = machine_hierarchy.skipPerLevel;
46 }
47 
48 #if KMP_AFFINITY_SUPPORTED
49 
50 bool KMPAffinity::picked_api = false;
51 
52 void *KMPAffinity::Mask::operator new(size_t n) { return __kmp_allocate(n); }
53 void *KMPAffinity::Mask::operator new[](size_t n) { return __kmp_allocate(n); }
54 void KMPAffinity::Mask::operator delete(void *p) { __kmp_free(p); }
55 void KMPAffinity::Mask::operator delete[](void *p) { __kmp_free(p); }
56 void *KMPAffinity::operator new(size_t n) { return __kmp_allocate(n); }
57 void KMPAffinity::operator delete(void *p) { __kmp_free(p); }
58 
59 void KMPAffinity::pick_api() {
60  KMPAffinity *affinity_dispatch;
61  if (picked_api)
62  return;
63 #if KMP_USE_HWLOC
64  // Only use Hwloc if affinity isn't explicitly disabled and
65  // user requests Hwloc topology method
66  if (__kmp_affinity_top_method == affinity_top_method_hwloc &&
67  __kmp_affinity_type != affinity_disabled) {
68  affinity_dispatch = new KMPHwlocAffinity();
69  } else
70 #endif
71  {
72  affinity_dispatch = new KMPNativeAffinity();
73  }
74  __kmp_affinity_dispatch = affinity_dispatch;
75  picked_api = true;
76 }
77 
78 void KMPAffinity::destroy_api() {
79  if (__kmp_affinity_dispatch != NULL) {
80  delete __kmp_affinity_dispatch;
81  __kmp_affinity_dispatch = NULL;
82  picked_api = false;
83  }
84 }
85 
86 #define KMP_ADVANCE_SCAN(scan) \
87  while (*scan != '\0') { \
88  scan++; \
89  }
90 
91 // Print the affinity mask to the character array in a pretty format.
92 // The format is a comma separated list of non-negative integers or integer
93 // ranges: e.g., 1,2,3-5,7,9-15
94 // The format can also be the string "{<empty>}" if no bits are set in mask
95 char *__kmp_affinity_print_mask(char *buf, int buf_len,
96  kmp_affin_mask_t *mask) {
97  int start = 0, finish = 0, previous = 0;
98  bool first_range;
99  KMP_ASSERT(buf);
100  KMP_ASSERT(buf_len >= 40);
101  KMP_ASSERT(mask);
102  char *scan = buf;
103  char *end = buf + buf_len - 1;
104 
105  // Check for empty set.
106  if (mask->begin() == mask->end()) {
107  KMP_SNPRINTF(scan, end - scan + 1, "{<empty>}");
108  KMP_ADVANCE_SCAN(scan);
109  KMP_ASSERT(scan <= end);
110  return buf;
111  }
112 
113  first_range = true;
114  start = mask->begin();
115  while (1) {
116  // Find next range
117  // [start, previous] is inclusive range of contiguous bits in mask
118  for (finish = mask->next(start), previous = start;
119  finish == previous + 1 && finish != mask->end();
120  finish = mask->next(finish)) {
121  previous = finish;
122  }
123 
124  // The first range does not need a comma printed before it, but the rest
125  // of the ranges do need a comma beforehand
126  if (!first_range) {
127  KMP_SNPRINTF(scan, end - scan + 1, "%s", ",");
128  KMP_ADVANCE_SCAN(scan);
129  } else {
130  first_range = false;
131  }
132  // Range with three or more contiguous bits in the affinity mask
133  if (previous - start > 1) {
134  KMP_SNPRINTF(scan, end - scan + 1, "%d-%d", static_cast<int>(start),
135  static_cast<int>(previous));
136  } else {
137  // Range with one or two contiguous bits in the affinity mask
138  KMP_SNPRINTF(scan, end - scan + 1, "%d", static_cast<int>(start));
139  KMP_ADVANCE_SCAN(scan);
140  if (previous - start > 0) {
141  KMP_SNPRINTF(scan, end - scan + 1, ",%d", static_cast<int>(previous));
142  }
143  }
144  KMP_ADVANCE_SCAN(scan);
145  // Start over with new start point
146  start = finish;
147  if (start == mask->end())
148  break;
149  // Check for overflow
150  if (end - scan < 2)
151  break;
152  }
153 
154  // Check for overflow
155  KMP_ASSERT(scan <= end);
156  return buf;
157 }
158 #undef KMP_ADVANCE_SCAN
159 
160 // Print the affinity mask to the string buffer object in a pretty format
161 // The format is a comma separated list of non-negative integers or integer
162 // ranges: e.g., 1,2,3-5,7,9-15
163 // The format can also be the string "{<empty>}" if no bits are set in mask
164 kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
165  kmp_affin_mask_t *mask) {
166  int start = 0, finish = 0, previous = 0;
167  bool first_range;
168  KMP_ASSERT(buf);
169  KMP_ASSERT(mask);
170 
171  __kmp_str_buf_clear(buf);
172 
173  // Check for empty set.
174  if (mask->begin() == mask->end()) {
175  __kmp_str_buf_print(buf, "%s", "{<empty>}");
176  return buf;
177  }
178 
179  first_range = true;
180  start = mask->begin();
181  while (1) {
182  // Find next range
183  // [start, previous] is inclusive range of contiguous bits in mask
184  for (finish = mask->next(start), previous = start;
185  finish == previous + 1 && finish != mask->end();
186  finish = mask->next(finish)) {
187  previous = finish;
188  }
189 
190  // The first range does not need a comma printed before it, but the rest
191  // of the ranges do need a comma beforehand
192  if (!first_range) {
193  __kmp_str_buf_print(buf, "%s", ",");
194  } else {
195  first_range = false;
196  }
197  // Range with three or more contiguous bits in the affinity mask
198  if (previous - start > 1) {
199  __kmp_str_buf_print(buf, "%d-%d", static_cast<int>(start),
200  static_cast<int>(previous));
201  } else {
202  // Range with one or two contiguous bits in the affinity mask
203  __kmp_str_buf_print(buf, "%d", static_cast<int>(start));
204  if (previous - start > 0) {
205  __kmp_str_buf_print(buf, ",%d", static_cast<int>(previous));
206  }
207  }
208  // Start over with new start point
209  start = finish;
210  if (start == mask->end())
211  break;
212  }
213  return buf;
214 }
215 
216 void __kmp_affinity_entire_machine_mask(kmp_affin_mask_t *mask) {
217  KMP_CPU_ZERO(mask);
218 
219 #if KMP_GROUP_AFFINITY
220 
221  if (__kmp_num_proc_groups > 1) {
222  int group;
223  KMP_DEBUG_ASSERT(__kmp_GetActiveProcessorCount != NULL);
224  for (group = 0; group < __kmp_num_proc_groups; group++) {
225  int i;
226  int num = __kmp_GetActiveProcessorCount(group);
227  for (i = 0; i < num; i++) {
228  KMP_CPU_SET(i + group * (CHAR_BIT * sizeof(DWORD_PTR)), mask);
229  }
230  }
231  } else
232 
233 #endif /* KMP_GROUP_AFFINITY */
234 
235  {
236  int proc;
237  for (proc = 0; proc < __kmp_xproc; proc++) {
238  KMP_CPU_SET(proc, mask);
239  }
240  }
241 }
242 
243 // When sorting by labels, __kmp_affinity_assign_child_nums() must first be
244 // called to renumber the labels from [0..n] and place them into the child_num
245 // vector of the address object. This is done in case the labels used for
246 // the children at one node of the hierarchy differ from those used for
247 // another node at the same level. Example: suppose the machine has 2 nodes
248 // with 2 packages each. The first node contains packages 601 and 602, and
249 // second node contains packages 603 and 604. If we try to sort the table
250 // for "scatter" affinity, the table will still be sorted 601, 602, 603, 604
251 // because we are paying attention to the labels themselves, not the ordinal
252 // child numbers. By using the child numbers in the sort, the result is
253 // {0,0}=601, {0,1}=603, {1,0}=602, {1,1}=604.
254 static void __kmp_affinity_assign_child_nums(AddrUnsPair *address2os,
255  int numAddrs) {
256  KMP_DEBUG_ASSERT(numAddrs > 0);
257  int depth = address2os->first.depth;
258  unsigned *counts = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
259  unsigned *lastLabel = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
260  int labCt;
261  for (labCt = 0; labCt < depth; labCt++) {
262  address2os[0].first.childNums[labCt] = counts[labCt] = 0;
263  lastLabel[labCt] = address2os[0].first.labels[labCt];
264  }
265  int i;
266  for (i = 1; i < numAddrs; i++) {
267  for (labCt = 0; labCt < depth; labCt++) {
268  if (address2os[i].first.labels[labCt] != lastLabel[labCt]) {
269  int labCt2;
270  for (labCt2 = labCt + 1; labCt2 < depth; labCt2++) {
271  counts[labCt2] = 0;
272  lastLabel[labCt2] = address2os[i].first.labels[labCt2];
273  }
274  counts[labCt]++;
275  lastLabel[labCt] = address2os[i].first.labels[labCt];
276  break;
277  }
278  }
279  for (labCt = 0; labCt < depth; labCt++) {
280  address2os[i].first.childNums[labCt] = counts[labCt];
281  }
282  for (; labCt < (int)Address::maxDepth; labCt++) {
283  address2os[i].first.childNums[labCt] = 0;
284  }
285  }
286  __kmp_free(lastLabel);
287  __kmp_free(counts);
288 }
289 
290 // All of the __kmp_affinity_create_*_map() routines should set
291 // __kmp_affinity_masks to a vector of affinity mask objects of length
292 // __kmp_affinity_num_masks, if __kmp_affinity_type != affinity_none, and return
293 // the number of levels in the machine topology tree (zero if
294 // __kmp_affinity_type == affinity_none).
295 //
296 // All of the __kmp_affinity_create_*_map() routines should set
297 // *__kmp_affin_fullMask to the affinity mask for the initialization thread.
298 // They need to save and restore the mask, and it could be needed later, so
299 // saving it is just an optimization to avoid calling kmp_get_system_affinity()
300 // again.
301 kmp_affin_mask_t *__kmp_affin_fullMask = NULL;
302 
303 static int nCoresPerPkg, nPackages;
304 static int __kmp_nThreadsPerCore;
305 #ifndef KMP_DFLT_NTH_CORES
306 static int __kmp_ncores;
307 #endif
308 static int *__kmp_pu_os_idx = NULL;
309 
310 // __kmp_affinity_uniform_topology() doesn't work when called from
311 // places which support arbitrarily many levels in the machine topology
312 // map, i.e. the non-default cases in __kmp_affinity_create_cpuinfo_map()
313 // __kmp_affinity_create_x2apicid_map().
314 inline static bool __kmp_affinity_uniform_topology() {
315  return __kmp_avail_proc == (__kmp_nThreadsPerCore * nCoresPerPkg * nPackages);
316 }
317 
318 // Print out the detailed machine topology map, i.e. the physical locations
319 // of each OS proc.
320 static void __kmp_affinity_print_topology(AddrUnsPair *address2os, int len,
321  int depth, int pkgLevel,
322  int coreLevel, int threadLevel) {
323  int proc;
324 
325  KMP_INFORM(OSProcToPhysicalThreadMap, "KMP_AFFINITY");
326  for (proc = 0; proc < len; proc++) {
327  int level;
328  kmp_str_buf_t buf;
329  __kmp_str_buf_init(&buf);
330  for (level = 0; level < depth; level++) {
331  if (level == threadLevel) {
332  __kmp_str_buf_print(&buf, "%s ", KMP_I18N_STR(Thread));
333  } else if (level == coreLevel) {
334  __kmp_str_buf_print(&buf, "%s ", KMP_I18N_STR(Core));
335  } else if (level == pkgLevel) {
336  __kmp_str_buf_print(&buf, "%s ", KMP_I18N_STR(Package));
337  } else if (level > pkgLevel) {
338  __kmp_str_buf_print(&buf, "%s_%d ", KMP_I18N_STR(Node),
339  level - pkgLevel - 1);
340  } else {
341  __kmp_str_buf_print(&buf, "L%d ", level);
342  }
343  __kmp_str_buf_print(&buf, "%d ", address2os[proc].first.labels[level]);
344  }
345  KMP_INFORM(OSProcMapToPack, "KMP_AFFINITY", address2os[proc].second,
346  buf.str);
347  __kmp_str_buf_free(&buf);
348  }
349 }
350 
351 #if KMP_USE_HWLOC
352 
353 static void __kmp_affinity_print_hwloc_tp(AddrUnsPair *addrP, int len,
354  int depth, int *levels) {
355  int proc;
356  kmp_str_buf_t buf;
357  __kmp_str_buf_init(&buf);
358  KMP_INFORM(OSProcToPhysicalThreadMap, "KMP_AFFINITY");
359  for (proc = 0; proc < len; proc++) {
360  __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Package),
361  addrP[proc].first.labels[0]);
362  if (depth > 1) {
363  int level = 1; // iterate over levels
364  int label = 1; // iterate over labels
365  if (__kmp_numa_detected)
366  // node level follows package
367  if (levels[level++] > 0)
368  __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Node),
369  addrP[proc].first.labels[label++]);
370  if (__kmp_tile_depth > 0)
371  // tile level follows node if any, or package
372  if (levels[level++] > 0)
373  __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Tile),
374  addrP[proc].first.labels[label++]);
375  if (levels[level++] > 0)
376  // core level follows
377  __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Core),
378  addrP[proc].first.labels[label++]);
379  if (levels[level++] > 0)
380  // thread level is the latest
381  __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Thread),
382  addrP[proc].first.labels[label++]);
383  KMP_DEBUG_ASSERT(label == depth);
384  }
385  KMP_INFORM(OSProcMapToPack, "KMP_AFFINITY", addrP[proc].second, buf.str);
386  __kmp_str_buf_clear(&buf);
387  }
388  __kmp_str_buf_free(&buf);
389 }
390 
391 static int nNodePerPkg, nTilePerPkg, nTilePerNode, nCorePerNode, nCorePerTile;
392 
393 // This function removes the topology levels that are radix 1 and don't offer
394 // further information about the topology. The most common example is when you
395 // have one thread context per core, we don't want the extra thread context
396 // level if it offers no unique labels. So they are removed.
397 // return value: the new depth of address2os
398 static int __kmp_affinity_remove_radix_one_levels(AddrUnsPair *addrP, int nTh,
399  int depth, int *levels) {
400  int level;
401  int i;
402  int radix1_detected;
403  int new_depth = depth;
404  for (level = depth - 1; level > 0; --level) {
405  // Detect if this level is radix 1
406  radix1_detected = 1;
407  for (i = 1; i < nTh; ++i) {
408  if (addrP[0].first.labels[level] != addrP[i].first.labels[level]) {
409  // There are differing label values for this level so it stays
410  radix1_detected = 0;
411  break;
412  }
413  }
414  if (!radix1_detected)
415  continue;
416  // Radix 1 was detected
417  --new_depth;
418  levels[level] = -1; // mark level as not present in address2os array
419  if (level == new_depth) {
420  // "turn off" deepest level, just decrement the depth that removes
421  // the level from address2os array
422  for (i = 0; i < nTh; ++i) {
423  addrP[i].first.depth--;
424  }
425  } else {
426  // For other levels, we move labels over and also reduce the depth
427  int j;
428  for (j = level; j < new_depth; ++j) {
429  for (i = 0; i < nTh; ++i) {
430  addrP[i].first.labels[j] = addrP[i].first.labels[j + 1];
431  addrP[i].first.depth--;
432  }
433  levels[j + 1] -= 1;
434  }
435  }
436  }
437  return new_depth;
438 }
439 
440 // Returns the number of objects of type 'type' below 'obj' within the topology
441 // tree structure. e.g., if obj is a HWLOC_OBJ_PACKAGE object, and type is
442 // HWLOC_OBJ_PU, then this will return the number of PU's under the SOCKET
443 // object.
444 static int __kmp_hwloc_get_nobjs_under_obj(hwloc_obj_t obj,
445  hwloc_obj_type_t type) {
446  int retval = 0;
447  hwloc_obj_t first;
448  for (first = hwloc_get_obj_below_by_type(__kmp_hwloc_topology, obj->type,
449  obj->logical_index, type, 0);
450  first != NULL &&
451  hwloc_get_ancestor_obj_by_type(__kmp_hwloc_topology, obj->type, first) ==
452  obj;
453  first = hwloc_get_next_obj_by_type(__kmp_hwloc_topology, first->type,
454  first)) {
455  ++retval;
456  }
457  return retval;
458 }
459 
460 static int __kmp_hwloc_count_children_by_depth(hwloc_topology_t t,
461  hwloc_obj_t o, unsigned 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;
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  // check tile, get object by depth because of multiple caches possible
553  depth = hwloc_get_cache_type_depth(tp, 2, HWLOC_OBJ_CACHE_UNIFIED);
554  hL = hwloc_get_ancestor_obj_by_depth(tp, depth, hT);
555  hC = NULL; // not used, but reset it here just in case
556  if (hL != NULL &&
557  __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE, &hC) > 1)
558  __kmp_tile_depth = depth; // tile consists of multiple cores
559  return 0;
560 }
561 
562 static int __kmp_affinity_create_hwloc_map(AddrUnsPair **address2os,
563  kmp_i18n_id_t *const msg_id) {
564  hwloc_topology_t &tp = __kmp_hwloc_topology; // shortcut of a long name
565  *address2os = NULL;
566  *msg_id = kmp_i18n_null;
567 
568  // Save the affinity mask for the current thread.
569  kmp_affin_mask_t *oldMask;
570  KMP_CPU_ALLOC(oldMask);
571  __kmp_get_system_affinity(oldMask, TRUE);
572  __kmp_hwloc_check_numa();
573 
574  if (!KMP_AFFINITY_CAPABLE()) {
575  // Hack to try and infer the machine topology using only the data
576  // available from cpuid on the current thread, and __kmp_xproc.
577  KMP_ASSERT(__kmp_affinity_type == affinity_none);
578 
579  nCoresPerPkg = __kmp_hwloc_get_nobjs_under_obj(
580  hwloc_get_obj_by_type(tp, HWLOC_OBJ_PACKAGE, 0), HWLOC_OBJ_CORE);
581  __kmp_nThreadsPerCore = __kmp_hwloc_get_nobjs_under_obj(
582  hwloc_get_obj_by_type(tp, HWLOC_OBJ_CORE, 0), HWLOC_OBJ_PU);
583  __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
584  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
585  if (__kmp_affinity_verbose) {
586  KMP_INFORM(AffNotCapableUseLocCpuidL11, "KMP_AFFINITY");
587  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
588  if (__kmp_affinity_uniform_topology()) {
589  KMP_INFORM(Uniform, "KMP_AFFINITY");
590  } else {
591  KMP_INFORM(NonUniform, "KMP_AFFINITY");
592  }
593  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
594  __kmp_nThreadsPerCore, __kmp_ncores);
595  }
596  KMP_CPU_FREE(oldMask);
597  return 0;
598  }
599 
600  int depth = 3;
601  int levels[5] = {0, 1, 2, 3, 4}; // package, [node,] [tile,] core, thread
602  int labels[3] = {0}; // package [,node] [,tile] - head of lables array
603  if (__kmp_numa_detected)
604  ++depth;
605  if (__kmp_tile_depth)
606  ++depth;
607 
608  // Allocate the data structure to be returned.
609  AddrUnsPair *retval =
610  (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * __kmp_avail_proc);
611  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
612  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
613 
614  // When affinity is off, this routine will still be called to set
615  // __kmp_ncores, as well as __kmp_nThreadsPerCore,
616  // nCoresPerPkg, & nPackages. Make sure all these vars are set
617  // correctly, and return if affinity is not enabled.
618 
619  hwloc_obj_t socket, node, tile;
620  int nActiveThreads = 0;
621  int socket_id = 0;
622  // re-calculate globals to count only accessible resources
623  __kmp_ncores = nPackages = nCoresPerPkg = __kmp_nThreadsPerCore = 0;
624  nNodePerPkg = nTilePerPkg = nTilePerNode = nCorePerNode = nCorePerTile = 0;
625  for (socket = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PACKAGE, 0); socket != NULL;
626  socket = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PACKAGE, socket),
627  socket_id++) {
628  labels[0] = socket_id;
629  if (__kmp_numa_detected) {
630  int NN;
631  int n_active_nodes = 0;
632  node = NULL;
633  NN = __kmp_hwloc_count_children_by_type(tp, socket, HWLOC_OBJ_NUMANODE,
634  &node);
635  for (int node_id = 0; node_id < NN; ++node_id, node = node->next_cousin) {
636  labels[1] = node_id;
637  if (__kmp_tile_depth) {
638  // NUMA + tiles
639  int NT;
640  int n_active_tiles = 0;
641  tile = NULL;
642  NT = __kmp_hwloc_count_children_by_depth(tp, node, __kmp_tile_depth,
643  &tile);
644  for (int tl_id = 0; tl_id < NT; ++tl_id, tile = tile->next_cousin) {
645  labels[2] = tl_id;
646  int n_active_cores = 0;
647  __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads,
648  n_active_cores, tile, 3, labels);
649  if (n_active_cores) { // were there any active cores on the socket?
650  ++n_active_tiles; // count active tiles per node
651  if (n_active_cores > nCorePerTile)
652  nCorePerTile = n_active_cores; // calc maximum
653  }
654  }
655  if (n_active_tiles) { // were there any active tiles on the socket?
656  ++n_active_nodes; // count active nodes per package
657  if (n_active_tiles > nTilePerNode)
658  nTilePerNode = n_active_tiles; // calc maximum
659  }
660  } else {
661  // NUMA, no tiles
662  int n_active_cores = 0;
663  __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads,
664  n_active_cores, node, 2, labels);
665  if (n_active_cores) { // were there any active cores on the socket?
666  ++n_active_nodes; // count active nodes per package
667  if (n_active_cores > nCorePerNode)
668  nCorePerNode = n_active_cores; // calc maximum
669  }
670  }
671  }
672  if (n_active_nodes) { // were there any active nodes on the socket?
673  ++nPackages; // count total active packages
674  if (n_active_nodes > nNodePerPkg)
675  nNodePerPkg = n_active_nodes; // calc maximum
676  }
677  } else {
678  if (__kmp_tile_depth) {
679  // no NUMA, tiles
680  int NT;
681  int n_active_tiles = 0;
682  tile = NULL;
683  NT = __kmp_hwloc_count_children_by_depth(tp, socket, __kmp_tile_depth,
684  &tile);
685  for (int tl_id = 0; tl_id < NT; ++tl_id, tile = tile->next_cousin) {
686  labels[1] = tl_id;
687  int n_active_cores = 0;
688  __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads,
689  n_active_cores, tile, 2, labels);
690  if (n_active_cores) { // were there any active cores on the socket?
691  ++n_active_tiles; // count active tiles per package
692  if (n_active_cores > nCorePerTile)
693  nCorePerTile = n_active_cores; // calc maximum
694  }
695  }
696  if (n_active_tiles) { // were there any active tiles on the socket?
697  ++nPackages; // count total active packages
698  if (n_active_tiles > nTilePerPkg)
699  nTilePerPkg = n_active_tiles; // calc maximum
700  }
701  } else {
702  // no NUMA, no tiles
703  int n_active_cores = 0;
704  __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads, n_active_cores,
705  socket, 1, labels);
706  if (n_active_cores) { // were there any active cores on the socket?
707  ++nPackages; // count total active packages
708  if (n_active_cores > nCoresPerPkg)
709  nCoresPerPkg = n_active_cores; // calc maximum
710  }
711  }
712  }
713  }
714 
715  // If there's only one thread context to bind to, return now.
716  KMP_DEBUG_ASSERT(nActiveThreads == __kmp_avail_proc);
717  KMP_ASSERT(nActiveThreads > 0);
718  if (nActiveThreads == 1) {
719  __kmp_ncores = nPackages = 1;
720  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
721  if (__kmp_affinity_verbose) {
722  char buf[KMP_AFFIN_MASK_PRINT_LEN];
723  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
724 
725  KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
726  if (__kmp_affinity_respect_mask) {
727  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
728  } else {
729  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
730  }
731  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
732  KMP_INFORM(Uniform, "KMP_AFFINITY");
733  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
734  __kmp_nThreadsPerCore, __kmp_ncores);
735  }
736 
737  if (__kmp_affinity_type == affinity_none) {
738  __kmp_free(retval);
739  KMP_CPU_FREE(oldMask);
740  return 0;
741  }
742 
743  // Form an Address object which only includes the package level.
744  Address addr(1);
745  addr.labels[0] = retval[0].first.labels[0];
746  retval[0].first = addr;
747 
748  if (__kmp_affinity_gran_levels < 0) {
749  __kmp_affinity_gran_levels = 0;
750  }
751 
752  if (__kmp_affinity_verbose) {
753  __kmp_affinity_print_topology(retval, 1, 1, 0, -1, -1);
754  }
755 
756  *address2os = retval;
757  KMP_CPU_FREE(oldMask);
758  return 1;
759  }
760 
761  // Sort the table by physical Id.
762  qsort(retval, nActiveThreads, sizeof(*retval),
763  __kmp_affinity_cmp_Address_labels);
764 
765  // Check to see if the machine topology is uniform
766  int nPUs = nPackages * __kmp_nThreadsPerCore;
767  if (__kmp_numa_detected) {
768  if (__kmp_tile_depth) { // NUMA + tiles
769  nPUs *= (nNodePerPkg * nTilePerNode * nCorePerTile);
770  } else { // NUMA, no tiles
771  nPUs *= (nNodePerPkg * nCorePerNode);
772  }
773  } else {
774  if (__kmp_tile_depth) { // no NUMA, tiles
775  nPUs *= (nTilePerPkg * nCorePerTile);
776  } else { // no NUMA, no tiles
777  nPUs *= nCoresPerPkg;
778  }
779  }
780  unsigned uniform = (nPUs == nActiveThreads);
781 
782  // Print the machine topology summary.
783  if (__kmp_affinity_verbose) {
784  char mask[KMP_AFFIN_MASK_PRINT_LEN];
785  __kmp_affinity_print_mask(mask, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
786  if (__kmp_affinity_respect_mask) {
787  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", mask);
788  } else {
789  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", mask);
790  }
791  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
792  if (uniform) {
793  KMP_INFORM(Uniform, "KMP_AFFINITY");
794  } else {
795  KMP_INFORM(NonUniform, "KMP_AFFINITY");
796  }
797  if (__kmp_numa_detected) {
798  if (__kmp_tile_depth) { // NUMA + tiles
799  KMP_INFORM(TopologyExtraNoTi, "KMP_AFFINITY", nPackages, nNodePerPkg,
800  nTilePerNode, nCorePerTile, __kmp_nThreadsPerCore,
801  __kmp_ncores);
802  } else { // NUMA, no tiles
803  KMP_INFORM(TopologyExtraNode, "KMP_AFFINITY", nPackages, nNodePerPkg,
804  nCorePerNode, __kmp_nThreadsPerCore, __kmp_ncores);
805  nPUs *= (nNodePerPkg * nCorePerNode);
806  }
807  } else {
808  if (__kmp_tile_depth) { // no NUMA, tiles
809  KMP_INFORM(TopologyExtraTile, "KMP_AFFINITY", nPackages, nTilePerPkg,
810  nCorePerTile, __kmp_nThreadsPerCore, __kmp_ncores);
811  } else { // no NUMA, no tiles
812  kmp_str_buf_t buf;
813  __kmp_str_buf_init(&buf);
814  __kmp_str_buf_print(&buf, "%d", nPackages);
815  KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, nCoresPerPkg,
816  __kmp_nThreadsPerCore, __kmp_ncores);
817  __kmp_str_buf_free(&buf);
818  }
819  }
820  }
821 
822  if (__kmp_affinity_type == affinity_none) {
823  __kmp_free(retval);
824  KMP_CPU_FREE(oldMask);
825  return 0;
826  }
827 
828  int depth_full = depth; // number of levels before compressing
829  // Find any levels with radiix 1, and remove them from the map
830  // (except for the package level).
831  depth = __kmp_affinity_remove_radix_one_levels(retval, nActiveThreads, depth,
832  levels);
833  KMP_DEBUG_ASSERT(__kmp_affinity_gran != affinity_gran_default);
834  if (__kmp_affinity_gran_levels < 0) {
835  // Set the granularity level based on what levels are modeled
836  // in the machine topology map.
837  __kmp_affinity_gran_levels = 0; // lowest level (e.g. fine)
838  if (__kmp_affinity_gran > affinity_gran_thread) {
839  for (int i = 1; i <= depth_full; ++i) {
840  if (__kmp_affinity_gran <= i) // only count deeper levels
841  break;
842  if (levels[depth_full - i] > 0)
843  __kmp_affinity_gran_levels++;
844  }
845  }
846  if (__kmp_affinity_gran > affinity_gran_package)
847  __kmp_affinity_gran_levels++; // e.g. granularity = group
848  }
849 
850  if (__kmp_affinity_verbose)
851  __kmp_affinity_print_hwloc_tp(retval, nActiveThreads, depth, levels);
852 
853  KMP_CPU_FREE(oldMask);
854  *address2os = retval;
855  return depth;
856 }
857 #endif // KMP_USE_HWLOC
858 
859 // If we don't know how to retrieve the machine's processor topology, or
860 // encounter an error in doing so, this routine is called to form a "flat"
861 // mapping of os thread id's <-> processor id's.
862 static int __kmp_affinity_create_flat_map(AddrUnsPair **address2os,
863  kmp_i18n_id_t *const msg_id) {
864  *address2os = NULL;
865  *msg_id = kmp_i18n_null;
866 
867  // Even if __kmp_affinity_type == affinity_none, this routine might still
868  // called to set __kmp_ncores, as well as
869  // __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
870  if (!KMP_AFFINITY_CAPABLE()) {
871  KMP_ASSERT(__kmp_affinity_type == affinity_none);
872  __kmp_ncores = nPackages = __kmp_xproc;
873  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
874  if (__kmp_affinity_verbose) {
875  KMP_INFORM(AffFlatTopology, "KMP_AFFINITY");
876  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
877  KMP_INFORM(Uniform, "KMP_AFFINITY");
878  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
879  __kmp_nThreadsPerCore, __kmp_ncores);
880  }
881  return 0;
882  }
883 
884  // When affinity is off, this routine will still be called to set
885  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
886  // Make sure all these vars are set correctly, and return now if affinity is
887  // not enabled.
888  __kmp_ncores = nPackages = __kmp_avail_proc;
889  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
890  if (__kmp_affinity_verbose) {
891  char buf[KMP_AFFIN_MASK_PRINT_LEN];
892  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
893  __kmp_affin_fullMask);
894 
895  KMP_INFORM(AffCapableUseFlat, "KMP_AFFINITY");
896  if (__kmp_affinity_respect_mask) {
897  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
898  } else {
899  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
900  }
901  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
902  KMP_INFORM(Uniform, "KMP_AFFINITY");
903  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
904  __kmp_nThreadsPerCore, __kmp_ncores);
905  }
906  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
907  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
908  if (__kmp_affinity_type == affinity_none) {
909  int avail_ct = 0;
910  int i;
911  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
912  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask))
913  continue;
914  __kmp_pu_os_idx[avail_ct++] = i; // suppose indices are flat
915  }
916  return 0;
917  }
918 
919  // Contruct the data structure to be returned.
920  *address2os =
921  (AddrUnsPair *)__kmp_allocate(sizeof(**address2os) * __kmp_avail_proc);
922  int avail_ct = 0;
923  int i;
924  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
925  // Skip this proc if it is not included in the machine model.
926  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
927  continue;
928  }
929  __kmp_pu_os_idx[avail_ct] = i; // suppose indices are flat
930  Address addr(1);
931  addr.labels[0] = i;
932  (*address2os)[avail_ct++] = AddrUnsPair(addr, i);
933  }
934  if (__kmp_affinity_verbose) {
935  KMP_INFORM(OSProcToPackage, "KMP_AFFINITY");
936  }
937 
938  if (__kmp_affinity_gran_levels < 0) {
939  // Only the package level is modeled in the machine topology map,
940  // so the #levels of granularity is either 0 or 1.
941  if (__kmp_affinity_gran > affinity_gran_package) {
942  __kmp_affinity_gran_levels = 1;
943  } else {
944  __kmp_affinity_gran_levels = 0;
945  }
946  }
947  return 1;
948 }
949 
950 #if KMP_GROUP_AFFINITY
951 
952 // If multiple Windows* OS processor groups exist, we can create a 2-level
953 // topology map with the groups at level 0 and the individual procs at level 1.
954 // This facilitates letting the threads float among all procs in a group,
955 // if granularity=group (the default when there are multiple groups).
956 static int __kmp_affinity_create_proc_group_map(AddrUnsPair **address2os,
957  kmp_i18n_id_t *const msg_id) {
958  *address2os = NULL;
959  *msg_id = kmp_i18n_null;
960 
961  // If we aren't affinity capable, then return now.
962  // The flat mapping will be used.
963  if (!KMP_AFFINITY_CAPABLE()) {
964  // FIXME set *msg_id
965  return -1;
966  }
967 
968  // Contruct the data structure to be returned.
969  *address2os =
970  (AddrUnsPair *)__kmp_allocate(sizeof(**address2os) * __kmp_avail_proc);
971  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
972  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
973  int avail_ct = 0;
974  int i;
975  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
976  // Skip this proc if it is not included in the machine model.
977  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
978  continue;
979  }
980  __kmp_pu_os_idx[avail_ct] = i; // suppose indices are flat
981  Address addr(2);
982  addr.labels[0] = i / (CHAR_BIT * sizeof(DWORD_PTR));
983  addr.labels[1] = i % (CHAR_BIT * sizeof(DWORD_PTR));
984  (*address2os)[avail_ct++] = AddrUnsPair(addr, i);
985 
986  if (__kmp_affinity_verbose) {
987  KMP_INFORM(AffOSProcToGroup, "KMP_AFFINITY", i, addr.labels[0],
988  addr.labels[1]);
989  }
990  }
991 
992  if (__kmp_affinity_gran_levels < 0) {
993  if (__kmp_affinity_gran == affinity_gran_group) {
994  __kmp_affinity_gran_levels = 1;
995  } else if ((__kmp_affinity_gran == affinity_gran_fine) ||
996  (__kmp_affinity_gran == affinity_gran_thread)) {
997  __kmp_affinity_gran_levels = 0;
998  } else {
999  const char *gran_str = NULL;
1000  if (__kmp_affinity_gran == affinity_gran_core) {
1001  gran_str = "core";
1002  } else if (__kmp_affinity_gran == affinity_gran_package) {
1003  gran_str = "package";
1004  } else if (__kmp_affinity_gran == affinity_gran_node) {
1005  gran_str = "node";
1006  } else {
1007  KMP_ASSERT(0);
1008  }
1009 
1010  // Warning: can't use affinity granularity \"gran\" with group topology
1011  // method, using "thread"
1012  __kmp_affinity_gran_levels = 0;
1013  }
1014  }
1015  return 2;
1016 }
1017 
1018 #endif /* KMP_GROUP_AFFINITY */
1019 
1020 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1021 
1022 static int __kmp_cpuid_mask_width(int count) {
1023  int r = 0;
1024 
1025  while ((1 << r) < count)
1026  ++r;
1027  return r;
1028 }
1029 
1030 class apicThreadInfo {
1031 public:
1032  unsigned osId; // param to __kmp_affinity_bind_thread
1033  unsigned apicId; // from cpuid after binding
1034  unsigned maxCoresPerPkg; // ""
1035  unsigned maxThreadsPerPkg; // ""
1036  unsigned pkgId; // inferred from above values
1037  unsigned coreId; // ""
1038  unsigned threadId; // ""
1039 };
1040 
1041 static int __kmp_affinity_cmp_apicThreadInfo_phys_id(const void *a,
1042  const void *b) {
1043  const apicThreadInfo *aa = (const apicThreadInfo *)a;
1044  const apicThreadInfo *bb = (const apicThreadInfo *)b;
1045  if (aa->pkgId < bb->pkgId)
1046  return -1;
1047  if (aa->pkgId > bb->pkgId)
1048  return 1;
1049  if (aa->coreId < bb->coreId)
1050  return -1;
1051  if (aa->coreId > bb->coreId)
1052  return 1;
1053  if (aa->threadId < bb->threadId)
1054  return -1;
1055  if (aa->threadId > bb->threadId)
1056  return 1;
1057  return 0;
1058 }
1059 
1060 // On IA-32 architecture and Intel(R) 64 architecture, we attempt to use
1061 // an algorithm which cycles through the available os threads, setting
1062 // the current thread's affinity mask to that thread, and then retrieves
1063 // the Apic Id for each thread context using the cpuid instruction.
1064 static int __kmp_affinity_create_apicid_map(AddrUnsPair **address2os,
1065  kmp_i18n_id_t *const msg_id) {
1066  kmp_cpuid buf;
1067  *address2os = NULL;
1068  *msg_id = kmp_i18n_null;
1069 
1070  // Check if cpuid leaf 4 is supported.
1071  __kmp_x86_cpuid(0, 0, &buf);
1072  if (buf.eax < 4) {
1073  *msg_id = kmp_i18n_str_NoLeaf4Support;
1074  return -1;
1075  }
1076 
1077  // The algorithm used starts by setting the affinity to each available thread
1078  // and retrieving info from the cpuid instruction, so if we are not capable of
1079  // calling __kmp_get_system_affinity() and _kmp_get_system_affinity(), then we
1080  // need to do something else - use the defaults that we calculated from
1081  // issuing cpuid without binding to each proc.
1082  if (!KMP_AFFINITY_CAPABLE()) {
1083  // Hack to try and infer the machine topology using only the data
1084  // available from cpuid on the current thread, and __kmp_xproc.
1085  KMP_ASSERT(__kmp_affinity_type == affinity_none);
1086 
1087  // Get an upper bound on the number of threads per package using cpuid(1).
1088  // On some OS/chps combinations where HT is supported by the chip but is
1089  // disabled, this value will be 2 on a single core chip. Usually, it will be
1090  // 2 if HT is enabled and 1 if HT is disabled.
1091  __kmp_x86_cpuid(1, 0, &buf);
1092  int maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
1093  if (maxThreadsPerPkg == 0) {
1094  maxThreadsPerPkg = 1;
1095  }
1096 
1097  // The num cores per pkg comes from cpuid(4). 1 must be added to the encoded
1098  // value.
1099  //
1100  // The author of cpu_count.cpp treated this only an upper bound on the
1101  // number of cores, but I haven't seen any cases where it was greater than
1102  // the actual number of cores, so we will treat it as exact in this block of
1103  // code.
1104  //
1105  // First, we need to check if cpuid(4) is supported on this chip. To see if
1106  // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n or
1107  // greater.
1108  __kmp_x86_cpuid(0, 0, &buf);
1109  if (buf.eax >= 4) {
1110  __kmp_x86_cpuid(4, 0, &buf);
1111  nCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
1112  } else {
1113  nCoresPerPkg = 1;
1114  }
1115 
1116  // There is no way to reliably tell if HT is enabled without issuing the
1117  // cpuid instruction from every thread, can correlating the cpuid info, so
1118  // if the machine is not affinity capable, we assume that HT is off. We have
1119  // seen quite a few machines where maxThreadsPerPkg is 2, yet the machine
1120  // does not support HT.
1121  //
1122  // - Older OSes are usually found on machines with older chips, which do not
1123  // support HT.
1124  // - The performance penalty for mistakenly identifying a machine as HT when
1125  // it isn't (which results in blocktime being incorrecly set to 0) is
1126  // greater than the penalty when for mistakenly identifying a machine as
1127  // being 1 thread/core when it is really HT enabled (which results in
1128  // blocktime being incorrectly set to a positive value).
1129  __kmp_ncores = __kmp_xproc;
1130  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
1131  __kmp_nThreadsPerCore = 1;
1132  if (__kmp_affinity_verbose) {
1133  KMP_INFORM(AffNotCapableUseLocCpuid, "KMP_AFFINITY");
1134  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1135  if (__kmp_affinity_uniform_topology()) {
1136  KMP_INFORM(Uniform, "KMP_AFFINITY");
1137  } else {
1138  KMP_INFORM(NonUniform, "KMP_AFFINITY");
1139  }
1140  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
1141  __kmp_nThreadsPerCore, __kmp_ncores);
1142  }
1143  return 0;
1144  }
1145 
1146  // From here on, we can assume that it is safe to call
1147  // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
1148  // __kmp_affinity_type = affinity_none.
1149 
1150  // Save the affinity mask for the current thread.
1151  kmp_affin_mask_t *oldMask;
1152  KMP_CPU_ALLOC(oldMask);
1153  KMP_ASSERT(oldMask != NULL);
1154  __kmp_get_system_affinity(oldMask, TRUE);
1155 
1156  // Run through each of the available contexts, binding the current thread
1157  // to it, and obtaining the pertinent information using the cpuid instr.
1158  //
1159  // The relevant information is:
1160  // - Apic Id: Bits 24:31 of ebx after issuing cpuid(1) - each thread context
1161  // has a uniqie Apic Id, which is of the form pkg# : core# : thread#.
1162  // - Max Threads Per Pkg: Bits 16:23 of ebx after issuing cpuid(1). The value
1163  // of this field determines the width of the core# + thread# fields in the
1164  // Apic Id. It is also an upper bound on the number of threads per
1165  // package, but it has been verified that situations happen were it is not
1166  // exact. In particular, on certain OS/chip combinations where Intel(R)
1167  // Hyper-Threading Technology is supported by the chip but has been
1168  // disabled, the value of this field will be 2 (for a single core chip).
1169  // On other OS/chip combinations supporting Intel(R) Hyper-Threading
1170  // Technology, the value of this field will be 1 when Intel(R)
1171  // Hyper-Threading Technology is disabled and 2 when it is enabled.
1172  // - Max Cores Per Pkg: Bits 26:31 of eax after issuing cpuid(4). The value
1173  // of this field (+1) determines the width of the core# field in the Apic
1174  // Id. The comments in "cpucount.cpp" say that this value is an upper
1175  // bound, but the IA-32 architecture manual says that it is exactly the
1176  // number of cores per package, and I haven't seen any case where it
1177  // wasn't.
1178  //
1179  // From this information, deduce the package Id, core Id, and thread Id,
1180  // and set the corresponding fields in the apicThreadInfo struct.
1181  unsigned i;
1182  apicThreadInfo *threadInfo = (apicThreadInfo *)__kmp_allocate(
1183  __kmp_avail_proc * sizeof(apicThreadInfo));
1184  unsigned nApics = 0;
1185  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
1186  // Skip this proc if it is not included in the machine model.
1187  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
1188  continue;
1189  }
1190  KMP_DEBUG_ASSERT((int)nApics < __kmp_avail_proc);
1191 
1192  __kmp_affinity_dispatch->bind_thread(i);
1193  threadInfo[nApics].osId = i;
1194 
1195  // The apic id and max threads per pkg come from cpuid(1).
1196  __kmp_x86_cpuid(1, 0, &buf);
1197  if (((buf.edx >> 9) & 1) == 0) {
1198  __kmp_set_system_affinity(oldMask, TRUE);
1199  __kmp_free(threadInfo);
1200  KMP_CPU_FREE(oldMask);
1201  *msg_id = kmp_i18n_str_ApicNotPresent;
1202  return -1;
1203  }
1204  threadInfo[nApics].apicId = (buf.ebx >> 24) & 0xff;
1205  threadInfo[nApics].maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
1206  if (threadInfo[nApics].maxThreadsPerPkg == 0) {
1207  threadInfo[nApics].maxThreadsPerPkg = 1;
1208  }
1209 
1210  // Max cores per pkg comes from cpuid(4). 1 must be added to the encoded
1211  // value.
1212  //
1213  // First, we need to check if cpuid(4) is supported on this chip. To see if
1214  // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n
1215  // or greater.
1216  __kmp_x86_cpuid(0, 0, &buf);
1217  if (buf.eax >= 4) {
1218  __kmp_x86_cpuid(4, 0, &buf);
1219  threadInfo[nApics].maxCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
1220  } else {
1221  threadInfo[nApics].maxCoresPerPkg = 1;
1222  }
1223 
1224  // Infer the pkgId / coreId / threadId using only the info obtained locally.
1225  int widthCT = __kmp_cpuid_mask_width(threadInfo[nApics].maxThreadsPerPkg);
1226  threadInfo[nApics].pkgId = threadInfo[nApics].apicId >> widthCT;
1227 
1228  int widthC = __kmp_cpuid_mask_width(threadInfo[nApics].maxCoresPerPkg);
1229  int widthT = widthCT - widthC;
1230  if (widthT < 0) {
1231  // I've never seen this one happen, but I suppose it could, if the cpuid
1232  // instruction on a chip was really screwed up. Make sure to restore the
1233  // affinity mask before the tail call.
1234  __kmp_set_system_affinity(oldMask, TRUE);
1235  __kmp_free(threadInfo);
1236  KMP_CPU_FREE(oldMask);
1237  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1238  return -1;
1239  }
1240 
1241  int maskC = (1 << widthC) - 1;
1242  threadInfo[nApics].coreId = (threadInfo[nApics].apicId >> widthT) & maskC;
1243 
1244  int maskT = (1 << widthT) - 1;
1245  threadInfo[nApics].threadId = threadInfo[nApics].apicId & maskT;
1246 
1247  nApics++;
1248  }
1249 
1250  // We've collected all the info we need.
1251  // Restore the old affinity mask for this thread.
1252  __kmp_set_system_affinity(oldMask, TRUE);
1253 
1254  // If there's only one thread context to bind to, form an Address object
1255  // with depth 1 and return immediately (or, if affinity is off, set
1256  // address2os to NULL and return).
1257  //
1258  // If it is configured to omit the package level when there is only a single
1259  // package, the logic at the end of this routine won't work if there is only
1260  // a single thread - it would try to form an Address object with depth 0.
1261  KMP_ASSERT(nApics > 0);
1262  if (nApics == 1) {
1263  __kmp_ncores = nPackages = 1;
1264  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
1265  if (__kmp_affinity_verbose) {
1266  char buf[KMP_AFFIN_MASK_PRINT_LEN];
1267  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
1268 
1269  KMP_INFORM(AffUseGlobCpuid, "KMP_AFFINITY");
1270  if (__kmp_affinity_respect_mask) {
1271  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
1272  } else {
1273  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
1274  }
1275  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1276  KMP_INFORM(Uniform, "KMP_AFFINITY");
1277  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
1278  __kmp_nThreadsPerCore, __kmp_ncores);
1279  }
1280 
1281  if (__kmp_affinity_type == affinity_none) {
1282  __kmp_free(threadInfo);
1283  KMP_CPU_FREE(oldMask);
1284  return 0;
1285  }
1286 
1287  *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair));
1288  Address addr(1);
1289  addr.labels[0] = threadInfo[0].pkgId;
1290  (*address2os)[0] = AddrUnsPair(addr, threadInfo[0].osId);
1291 
1292  if (__kmp_affinity_gran_levels < 0) {
1293  __kmp_affinity_gran_levels = 0;
1294  }
1295 
1296  if (__kmp_affinity_verbose) {
1297  __kmp_affinity_print_topology(*address2os, 1, 1, 0, -1, -1);
1298  }
1299 
1300  __kmp_free(threadInfo);
1301  KMP_CPU_FREE(oldMask);
1302  return 1;
1303  }
1304 
1305  // Sort the threadInfo table by physical Id.
1306  qsort(threadInfo, nApics, sizeof(*threadInfo),
1307  __kmp_affinity_cmp_apicThreadInfo_phys_id);
1308 
1309  // The table is now sorted by pkgId / coreId / threadId, but we really don't
1310  // know the radix of any of the fields. pkgId's may be sparsely assigned among
1311  // the chips on a system. Although coreId's are usually assigned
1312  // [0 .. coresPerPkg-1] and threadId's are usually assigned
1313  // [0..threadsPerCore-1], we don't want to make any such assumptions.
1314  //
1315  // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
1316  // total # packages) are at this point - we want to determine that now. We
1317  // only have an upper bound on the first two figures.
1318  //
1319  // We also perform a consistency check at this point: the values returned by
1320  // the cpuid instruction for any thread bound to a given package had better
1321  // return the same info for maxThreadsPerPkg and maxCoresPerPkg.
1322  nPackages = 1;
1323  nCoresPerPkg = 1;
1324  __kmp_nThreadsPerCore = 1;
1325  unsigned nCores = 1;
1326 
1327  unsigned pkgCt = 1; // to determine radii
1328  unsigned lastPkgId = threadInfo[0].pkgId;
1329  unsigned coreCt = 1;
1330  unsigned lastCoreId = threadInfo[0].coreId;
1331  unsigned threadCt = 1;
1332  unsigned lastThreadId = threadInfo[0].threadId;
1333 
1334  // intra-pkg consist checks
1335  unsigned prevMaxCoresPerPkg = threadInfo[0].maxCoresPerPkg;
1336  unsigned prevMaxThreadsPerPkg = threadInfo[0].maxThreadsPerPkg;
1337 
1338  for (i = 1; i < nApics; i++) {
1339  if (threadInfo[i].pkgId != lastPkgId) {
1340  nCores++;
1341  pkgCt++;
1342  lastPkgId = threadInfo[i].pkgId;
1343  if ((int)coreCt > nCoresPerPkg)
1344  nCoresPerPkg = coreCt;
1345  coreCt = 1;
1346  lastCoreId = threadInfo[i].coreId;
1347  if ((int)threadCt > __kmp_nThreadsPerCore)
1348  __kmp_nThreadsPerCore = threadCt;
1349  threadCt = 1;
1350  lastThreadId = threadInfo[i].threadId;
1351 
1352  // This is a different package, so go on to the next iteration without
1353  // doing any consistency checks. Reset the consistency check vars, though.
1354  prevMaxCoresPerPkg = threadInfo[i].maxCoresPerPkg;
1355  prevMaxThreadsPerPkg = threadInfo[i].maxThreadsPerPkg;
1356  continue;
1357  }
1358 
1359  if (threadInfo[i].coreId != lastCoreId) {
1360  nCores++;
1361  coreCt++;
1362  lastCoreId = threadInfo[i].coreId;
1363  if ((int)threadCt > __kmp_nThreadsPerCore)
1364  __kmp_nThreadsPerCore = threadCt;
1365  threadCt = 1;
1366  lastThreadId = threadInfo[i].threadId;
1367  } else if (threadInfo[i].threadId != lastThreadId) {
1368  threadCt++;
1369  lastThreadId = threadInfo[i].threadId;
1370  } else {
1371  __kmp_free(threadInfo);
1372  KMP_CPU_FREE(oldMask);
1373  *msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
1374  return -1;
1375  }
1376 
1377  // Check to make certain that the maxCoresPerPkg and maxThreadsPerPkg
1378  // fields agree between all the threads bounds to a given package.
1379  if ((prevMaxCoresPerPkg != threadInfo[i].maxCoresPerPkg) ||
1380  (prevMaxThreadsPerPkg != threadInfo[i].maxThreadsPerPkg)) {
1381  __kmp_free(threadInfo);
1382  KMP_CPU_FREE(oldMask);
1383  *msg_id = kmp_i18n_str_InconsistentCpuidInfo;
1384  return -1;
1385  }
1386  }
1387  nPackages = pkgCt;
1388  if ((int)coreCt > nCoresPerPkg)
1389  nCoresPerPkg = coreCt;
1390  if ((int)threadCt > __kmp_nThreadsPerCore)
1391  __kmp_nThreadsPerCore = threadCt;
1392 
1393  // When affinity is off, this routine will still be called to set
1394  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
1395  // Make sure all these vars are set correctly, and return now if affinity is
1396  // not enabled.
1397  __kmp_ncores = nCores;
1398  if (__kmp_affinity_verbose) {
1399  char buf[KMP_AFFIN_MASK_PRINT_LEN];
1400  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
1401 
1402  KMP_INFORM(AffUseGlobCpuid, "KMP_AFFINITY");
1403  if (__kmp_affinity_respect_mask) {
1404  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
1405  } else {
1406  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
1407  }
1408  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1409  if (__kmp_affinity_uniform_topology()) {
1410  KMP_INFORM(Uniform, "KMP_AFFINITY");
1411  } else {
1412  KMP_INFORM(NonUniform, "KMP_AFFINITY");
1413  }
1414  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
1415  __kmp_nThreadsPerCore, __kmp_ncores);
1416  }
1417  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
1418  KMP_DEBUG_ASSERT(nApics == (unsigned)__kmp_avail_proc);
1419  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
1420  for (i = 0; i < nApics; ++i) {
1421  __kmp_pu_os_idx[i] = threadInfo[i].osId;
1422  }
1423  if (__kmp_affinity_type == affinity_none) {
1424  __kmp_free(threadInfo);
1425  KMP_CPU_FREE(oldMask);
1426  return 0;
1427  }
1428 
1429  // Now that we've determined the number of packages, the number of cores per
1430  // package, and the number of threads per core, we can construct the data
1431  // structure that is to be returned.
1432  int pkgLevel = 0;
1433  int coreLevel = (nCoresPerPkg <= 1) ? -1 : 1;
1434  int threadLevel =
1435  (__kmp_nThreadsPerCore <= 1) ? -1 : ((coreLevel >= 0) ? 2 : 1);
1436  unsigned depth = (pkgLevel >= 0) + (coreLevel >= 0) + (threadLevel >= 0);
1437 
1438  KMP_ASSERT(depth > 0);
1439  *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * nApics);
1440 
1441  for (i = 0; i < nApics; ++i) {
1442  Address addr(depth);
1443  unsigned os = threadInfo[i].osId;
1444  int d = 0;
1445 
1446  if (pkgLevel >= 0) {
1447  addr.labels[d++] = threadInfo[i].pkgId;
1448  }
1449  if (coreLevel >= 0) {
1450  addr.labels[d++] = threadInfo[i].coreId;
1451  }
1452  if (threadLevel >= 0) {
1453  addr.labels[d++] = threadInfo[i].threadId;
1454  }
1455  (*address2os)[i] = AddrUnsPair(addr, os);
1456  }
1457 
1458  if (__kmp_affinity_gran_levels < 0) {
1459  // Set the granularity level based on what levels are modeled in the machine
1460  // topology map.
1461  __kmp_affinity_gran_levels = 0;
1462  if ((threadLevel >= 0) && (__kmp_affinity_gran > affinity_gran_thread)) {
1463  __kmp_affinity_gran_levels++;
1464  }
1465  if ((coreLevel >= 0) && (__kmp_affinity_gran > affinity_gran_core)) {
1466  __kmp_affinity_gran_levels++;
1467  }
1468  if ((pkgLevel >= 0) && (__kmp_affinity_gran > affinity_gran_package)) {
1469  __kmp_affinity_gran_levels++;
1470  }
1471  }
1472 
1473  if (__kmp_affinity_verbose) {
1474  __kmp_affinity_print_topology(*address2os, nApics, depth, pkgLevel,
1475  coreLevel, threadLevel);
1476  }
1477 
1478  __kmp_free(threadInfo);
1479  KMP_CPU_FREE(oldMask);
1480  return depth;
1481 }
1482 
1483 // Intel(R) microarchitecture code name Nehalem, Dunnington and later
1484 // architectures support a newer interface for specifying the x2APIC Ids,
1485 // based on cpuid leaf 11.
1486 static int __kmp_affinity_create_x2apicid_map(AddrUnsPair **address2os,
1487  kmp_i18n_id_t *const msg_id) {
1488  kmp_cpuid buf;
1489  *address2os = NULL;
1490  *msg_id = kmp_i18n_null;
1491 
1492  // Check to see if cpuid leaf 11 is supported.
1493  __kmp_x86_cpuid(0, 0, &buf);
1494  if (buf.eax < 11) {
1495  *msg_id = kmp_i18n_str_NoLeaf11Support;
1496  return -1;
1497  }
1498  __kmp_x86_cpuid(11, 0, &buf);
1499  if (buf.ebx == 0) {
1500  *msg_id = kmp_i18n_str_NoLeaf11Support;
1501  return -1;
1502  }
1503 
1504  // Find the number of levels in the machine topology. While we're at it, get
1505  // the default values for __kmp_nThreadsPerCore & nCoresPerPkg. We will try to
1506  // get more accurate values later by explicitly counting them, but get
1507  // reasonable defaults now, in case we return early.
1508  int level;
1509  int threadLevel = -1;
1510  int coreLevel = -1;
1511  int pkgLevel = -1;
1512  __kmp_nThreadsPerCore = nCoresPerPkg = nPackages = 1;
1513 
1514  for (level = 0;; level++) {
1515  if (level > 31) {
1516  // FIXME: Hack for DPD200163180
1517  //
1518  // If level is big then something went wrong -> exiting
1519  //
1520  // There could actually be 32 valid levels in the machine topology, but so
1521  // far, the only machine we have seen which does not exit this loop before
1522  // iteration 32 has fubar x2APIC settings.
1523  //
1524  // For now, just reject this case based upon loop trip count.
1525  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1526  return -1;
1527  }
1528  __kmp_x86_cpuid(11, level, &buf);
1529  if (buf.ebx == 0) {
1530  if (pkgLevel < 0) {
1531  // Will infer nPackages from __kmp_xproc
1532  pkgLevel = level;
1533  level++;
1534  }
1535  break;
1536  }
1537  int kind = (buf.ecx >> 8) & 0xff;
1538  if (kind == 1) {
1539  // SMT level
1540  threadLevel = level;
1541  coreLevel = -1;
1542  pkgLevel = -1;
1543  __kmp_nThreadsPerCore = buf.ebx & 0xffff;
1544  if (__kmp_nThreadsPerCore == 0) {
1545  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1546  return -1;
1547  }
1548  } else if (kind == 2) {
1549  // core level
1550  coreLevel = level;
1551  pkgLevel = -1;
1552  nCoresPerPkg = buf.ebx & 0xffff;
1553  if (nCoresPerPkg == 0) {
1554  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1555  return -1;
1556  }
1557  } else {
1558  if (level <= 0) {
1559  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1560  return -1;
1561  }
1562  if (pkgLevel >= 0) {
1563  continue;
1564  }
1565  pkgLevel = level;
1566  nPackages = buf.ebx & 0xffff;
1567  if (nPackages == 0) {
1568  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1569  return -1;
1570  }
1571  }
1572  }
1573  int depth = level;
1574 
1575  // In the above loop, "level" was counted from the finest level (usually
1576  // thread) to the coarsest. The caller expects that we will place the labels
1577  // in (*address2os)[].first.labels[] in the inverse order, so we need to
1578  // invert the vars saying which level means what.
1579  if (threadLevel >= 0) {
1580  threadLevel = depth - threadLevel - 1;
1581  }
1582  if (coreLevel >= 0) {
1583  coreLevel = depth - coreLevel - 1;
1584  }
1585  KMP_DEBUG_ASSERT(pkgLevel >= 0);
1586  pkgLevel = depth - pkgLevel - 1;
1587 
1588  // The algorithm used starts by setting the affinity to each available thread
1589  // and retrieving info from the cpuid instruction, so if we are not capable of
1590  // calling __kmp_get_system_affinity() and _kmp_get_system_affinity(), then we
1591  // need to do something else - use the defaults that we calculated from
1592  // issuing cpuid without binding to each proc.
1593  if (!KMP_AFFINITY_CAPABLE()) {
1594  // Hack to try and infer the machine topology using only the data
1595  // available from cpuid on the current thread, and __kmp_xproc.
1596  KMP_ASSERT(__kmp_affinity_type == affinity_none);
1597 
1598  __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
1599  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
1600  if (__kmp_affinity_verbose) {
1601  KMP_INFORM(AffNotCapableUseLocCpuidL11, "KMP_AFFINITY");
1602  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1603  if (__kmp_affinity_uniform_topology()) {
1604  KMP_INFORM(Uniform, "KMP_AFFINITY");
1605  } else {
1606  KMP_INFORM(NonUniform, "KMP_AFFINITY");
1607  }
1608  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
1609  __kmp_nThreadsPerCore, __kmp_ncores);
1610  }
1611  return 0;
1612  }
1613 
1614  // From here on, we can assume that it is safe to call
1615  // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
1616  // __kmp_affinity_type = affinity_none.
1617 
1618  // Save the affinity mask for the current thread.
1619  kmp_affin_mask_t *oldMask;
1620  KMP_CPU_ALLOC(oldMask);
1621  __kmp_get_system_affinity(oldMask, TRUE);
1622 
1623  // Allocate the data structure to be returned.
1624  AddrUnsPair *retval =
1625  (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * __kmp_avail_proc);
1626 
1627  // Run through each of the available contexts, binding the current thread
1628  // to it, and obtaining the pertinent information using the cpuid instr.
1629  unsigned int proc;
1630  int nApics = 0;
1631  KMP_CPU_SET_ITERATE(proc, __kmp_affin_fullMask) {
1632  // Skip this proc if it is not included in the machine model.
1633  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
1634  continue;
1635  }
1636  KMP_DEBUG_ASSERT(nApics < __kmp_avail_proc);
1637 
1638  __kmp_affinity_dispatch->bind_thread(proc);
1639 
1640  // Extract labels for each level in the machine topology map from Apic ID.
1641  Address addr(depth);
1642  int prev_shift = 0;
1643 
1644  for (level = 0; level < depth; level++) {
1645  __kmp_x86_cpuid(11, level, &buf);
1646  unsigned apicId = buf.edx;
1647  if (buf.ebx == 0) {
1648  if (level != depth - 1) {
1649  KMP_CPU_FREE(oldMask);
1650  *msg_id = kmp_i18n_str_InconsistentCpuidInfo;
1651  return -1;
1652  }
1653  addr.labels[depth - level - 1] = apicId >> prev_shift;
1654  level++;
1655  break;
1656  }
1657  int shift = buf.eax & 0x1f;
1658  int mask = (1 << shift) - 1;
1659  addr.labels[depth - level - 1] = (apicId & mask) >> prev_shift;
1660  prev_shift = shift;
1661  }
1662  if (level != depth) {
1663  KMP_CPU_FREE(oldMask);
1664  *msg_id = kmp_i18n_str_InconsistentCpuidInfo;
1665  return -1;
1666  }
1667 
1668  retval[nApics] = AddrUnsPair(addr, proc);
1669  nApics++;
1670  }
1671 
1672  // We've collected all the info we need.
1673  // Restore the old affinity mask for this thread.
1674  __kmp_set_system_affinity(oldMask, TRUE);
1675 
1676  // If there's only one thread context to bind to, return now.
1677  KMP_ASSERT(nApics > 0);
1678  if (nApics == 1) {
1679  __kmp_ncores = nPackages = 1;
1680  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
1681  if (__kmp_affinity_verbose) {
1682  char buf[KMP_AFFIN_MASK_PRINT_LEN];
1683  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
1684 
1685  KMP_INFORM(AffUseGlobCpuidL11, "KMP_AFFINITY");
1686  if (__kmp_affinity_respect_mask) {
1687  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
1688  } else {
1689  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
1690  }
1691  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1692  KMP_INFORM(Uniform, "KMP_AFFINITY");
1693  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
1694  __kmp_nThreadsPerCore, __kmp_ncores);
1695  }
1696 
1697  if (__kmp_affinity_type == affinity_none) {
1698  __kmp_free(retval);
1699  KMP_CPU_FREE(oldMask);
1700  return 0;
1701  }
1702 
1703  // Form an Address object which only includes the package level.
1704  Address addr(1);
1705  addr.labels[0] = retval[0].first.labels[pkgLevel];
1706  retval[0].first = addr;
1707 
1708  if (__kmp_affinity_gran_levels < 0) {
1709  __kmp_affinity_gran_levels = 0;
1710  }
1711 
1712  if (__kmp_affinity_verbose) {
1713  __kmp_affinity_print_topology(retval, 1, 1, 0, -1, -1);
1714  }
1715 
1716  *address2os = retval;
1717  KMP_CPU_FREE(oldMask);
1718  return 1;
1719  }
1720 
1721  // Sort the table by physical Id.
1722  qsort(retval, nApics, sizeof(*retval), __kmp_affinity_cmp_Address_labels);
1723 
1724  // Find the radix at each of the levels.
1725  unsigned *totals = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
1726  unsigned *counts = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
1727  unsigned *maxCt = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
1728  unsigned *last = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
1729  for (level = 0; level < depth; level++) {
1730  totals[level] = 1;
1731  maxCt[level] = 1;
1732  counts[level] = 1;
1733  last[level] = retval[0].first.labels[level];
1734  }
1735 
1736  // From here on, the iteration variable "level" runs from the finest level to
1737  // the coarsest, i.e. we iterate forward through
1738  // (*address2os)[].first.labels[] - in the previous loops, we iterated
1739  // backwards.
1740  for (proc = 1; (int)proc < nApics; proc++) {
1741  int level;
1742  for (level = 0; level < depth; level++) {
1743  if (retval[proc].first.labels[level] != last[level]) {
1744  int j;
1745  for (j = level + 1; j < depth; j++) {
1746  totals[j]++;
1747  counts[j] = 1;
1748  // The line below causes printing incorrect topology information in
1749  // case the max value for some level (maxCt[level]) is encountered
1750  // earlier than some less value while going through the array. For
1751  // example, let pkg0 has 4 cores and pkg1 has 2 cores. Then
1752  // maxCt[1] == 2
1753  // whereas it must be 4.
1754  // TODO!!! Check if it can be commented safely
1755  // maxCt[j] = 1;
1756  last[j] = retval[proc].first.labels[j];
1757  }
1758  totals[level]++;
1759  counts[level]++;
1760  if (counts[level] > maxCt[level]) {
1761  maxCt[level] = counts[level];
1762  }
1763  last[level] = retval[proc].first.labels[level];
1764  break;
1765  } else if (level == depth - 1) {
1766  __kmp_free(last);
1767  __kmp_free(maxCt);
1768  __kmp_free(counts);
1769  __kmp_free(totals);
1770  __kmp_free(retval);
1771  KMP_CPU_FREE(oldMask);
1772  *msg_id = kmp_i18n_str_x2ApicIDsNotUnique;
1773  return -1;
1774  }
1775  }
1776  }
1777 
1778  // When affinity is off, this routine will still be called to set
1779  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
1780  // Make sure all these vars are set correctly, and return if affinity is not
1781  // enabled.
1782  if (threadLevel >= 0) {
1783  __kmp_nThreadsPerCore = maxCt[threadLevel];
1784  } else {
1785  __kmp_nThreadsPerCore = 1;
1786  }
1787  nPackages = totals[pkgLevel];
1788 
1789  if (coreLevel >= 0) {
1790  __kmp_ncores = totals[coreLevel];
1791  nCoresPerPkg = maxCt[coreLevel];
1792  } else {
1793  __kmp_ncores = nPackages;
1794  nCoresPerPkg = 1;
1795  }
1796 
1797  // Check to see if the machine topology is uniform
1798  unsigned prod = maxCt[0];
1799  for (level = 1; level < depth; level++) {
1800  prod *= maxCt[level];
1801  }
1802  bool uniform = (prod == totals[level - 1]);
1803 
1804  // Print the machine topology summary.
1805  if (__kmp_affinity_verbose) {
1806  char mask[KMP_AFFIN_MASK_PRINT_LEN];
1807  __kmp_affinity_print_mask(mask, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
1808 
1809  KMP_INFORM(AffUseGlobCpuidL11, "KMP_AFFINITY");
1810  if (__kmp_affinity_respect_mask) {
1811  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", mask);
1812  } else {
1813  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", mask);
1814  }
1815  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1816  if (uniform) {
1817  KMP_INFORM(Uniform, "KMP_AFFINITY");
1818  } else {
1819  KMP_INFORM(NonUniform, "KMP_AFFINITY");
1820  }
1821 
1822  kmp_str_buf_t buf;
1823  __kmp_str_buf_init(&buf);
1824 
1825  __kmp_str_buf_print(&buf, "%d", totals[0]);
1826  for (level = 1; level <= pkgLevel; level++) {
1827  __kmp_str_buf_print(&buf, " x %d", maxCt[level]);
1828  }
1829  KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, nCoresPerPkg,
1830  __kmp_nThreadsPerCore, __kmp_ncores);
1831 
1832  __kmp_str_buf_free(&buf);
1833  }
1834  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
1835  KMP_DEBUG_ASSERT(nApics == __kmp_avail_proc);
1836  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
1837  for (proc = 0; (int)proc < nApics; ++proc) {
1838  __kmp_pu_os_idx[proc] = retval[proc].second;
1839  }
1840  if (__kmp_affinity_type == affinity_none) {
1841  __kmp_free(last);
1842  __kmp_free(maxCt);
1843  __kmp_free(counts);
1844  __kmp_free(totals);
1845  __kmp_free(retval);
1846  KMP_CPU_FREE(oldMask);
1847  return 0;
1848  }
1849 
1850  // Find any levels with radiix 1, and remove them from the map
1851  // (except for the package level).
1852  int new_depth = 0;
1853  for (level = 0; level < depth; level++) {
1854  if ((maxCt[level] == 1) && (level != pkgLevel)) {
1855  continue;
1856  }
1857  new_depth++;
1858  }
1859 
1860  // If we are removing any levels, allocate a new vector to return,
1861  // and copy the relevant information to it.
1862  if (new_depth != depth) {
1863  AddrUnsPair *new_retval =
1864  (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * nApics);
1865  for (proc = 0; (int)proc < nApics; proc++) {
1866  Address addr(new_depth);
1867  new_retval[proc] = AddrUnsPair(addr, retval[proc].second);
1868  }
1869  int new_level = 0;
1870  int newPkgLevel = -1;
1871  int newCoreLevel = -1;
1872  int newThreadLevel = -1;
1873  for (level = 0; level < depth; level++) {
1874  if ((maxCt[level] == 1) && (level != pkgLevel)) {
1875  // Remove this level. Never remove the package level
1876  continue;
1877  }
1878  if (level == pkgLevel) {
1879  newPkgLevel = new_level;
1880  }
1881  if (level == coreLevel) {
1882  newCoreLevel = new_level;
1883  }
1884  if (level == threadLevel) {
1885  newThreadLevel = new_level;
1886  }
1887  for (proc = 0; (int)proc < nApics; proc++) {
1888  new_retval[proc].first.labels[new_level] =
1889  retval[proc].first.labels[level];
1890  }
1891  new_level++;
1892  }
1893 
1894  __kmp_free(retval);
1895  retval = new_retval;
1896  depth = new_depth;
1897  pkgLevel = newPkgLevel;
1898  coreLevel = newCoreLevel;
1899  threadLevel = newThreadLevel;
1900  }
1901 
1902  if (__kmp_affinity_gran_levels < 0) {
1903  // Set the granularity level based on what levels are modeled
1904  // in the machine topology map.
1905  __kmp_affinity_gran_levels = 0;
1906  if ((threadLevel >= 0) && (__kmp_affinity_gran > affinity_gran_thread)) {
1907  __kmp_affinity_gran_levels++;
1908  }
1909  if ((coreLevel >= 0) && (__kmp_affinity_gran > affinity_gran_core)) {
1910  __kmp_affinity_gran_levels++;
1911  }
1912  if (__kmp_affinity_gran > affinity_gran_package) {
1913  __kmp_affinity_gran_levels++;
1914  }
1915  }
1916 
1917  if (__kmp_affinity_verbose) {
1918  __kmp_affinity_print_topology(retval, nApics, depth, pkgLevel, coreLevel,
1919  threadLevel);
1920  }
1921 
1922  __kmp_free(last);
1923  __kmp_free(maxCt);
1924  __kmp_free(counts);
1925  __kmp_free(totals);
1926  KMP_CPU_FREE(oldMask);
1927  *address2os = retval;
1928  return depth;
1929 }
1930 
1931 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
1932 
1933 #define osIdIndex 0
1934 #define threadIdIndex 1
1935 #define coreIdIndex 2
1936 #define pkgIdIndex 3
1937 #define nodeIdIndex 4
1938 
1939 typedef unsigned *ProcCpuInfo;
1940 static unsigned maxIndex = pkgIdIndex;
1941 
1942 static int __kmp_affinity_cmp_ProcCpuInfo_phys_id(const void *a,
1943  const void *b) {
1944  unsigned i;
1945  const unsigned *aa = *(unsigned *const *)a;
1946  const unsigned *bb = *(unsigned *const *)b;
1947  for (i = maxIndex;; i--) {
1948  if (aa[i] < bb[i])
1949  return -1;
1950  if (aa[i] > bb[i])
1951  return 1;
1952  if (i == osIdIndex)
1953  break;
1954  }
1955  return 0;
1956 }
1957 
1958 #if KMP_USE_HIER_SCHED
1959 // Set the array sizes for the hierarchy layers
1960 static void __kmp_dispatch_set_hierarchy_values() {
1961  // Set the maximum number of L1's to number of cores
1962  // Set the maximum number of L2's to to either number of cores / 2 for
1963  // Intel(R) Xeon Phi(TM) coprocessor formally codenamed Knights Landing
1964  // Or the number of cores for Intel(R) Xeon(R) processors
1965  // Set the maximum number of NUMA nodes and L3's to number of packages
1966  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1] =
1967  nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
1968  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L1 + 1] = __kmp_ncores;
1969 #if KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
1970  if (__kmp_mic_type >= mic3)
1971  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores / 2;
1972  else
1973 #endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
1974  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores;
1975  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L3 + 1] = nPackages;
1976  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_NUMA + 1] = nPackages;
1977  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_LOOP + 1] = 1;
1978  // Set the number of threads per unit
1979  // Number of hardware threads per L1/L2/L3/NUMA/LOOP
1980  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_THREAD + 1] = 1;
1981  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L1 + 1] =
1982  __kmp_nThreadsPerCore;
1983 #if KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
1984  if (__kmp_mic_type >= mic3)
1985  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
1986  2 * __kmp_nThreadsPerCore;
1987  else
1988 #endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
1989  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
1990  __kmp_nThreadsPerCore;
1991  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L3 + 1] =
1992  nCoresPerPkg * __kmp_nThreadsPerCore;
1993  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_NUMA + 1] =
1994  nCoresPerPkg * __kmp_nThreadsPerCore;
1995  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_LOOP + 1] =
1996  nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
1997 }
1998 
1999 // Return the index into the hierarchy for this tid and layer type (L1, L2, etc)
2000 // i.e., this thread's L1 or this thread's L2, etc.
2001 int __kmp_dispatch_get_index(int tid, kmp_hier_layer_e type) {
2002  int index = type + 1;
2003  int num_hw_threads = __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1];
2004  KMP_DEBUG_ASSERT(type != kmp_hier_layer_e::LAYER_LAST);
2005  if (type == kmp_hier_layer_e::LAYER_THREAD)
2006  return tid;
2007  else if (type == kmp_hier_layer_e::LAYER_LOOP)
2008  return 0;
2009  KMP_DEBUG_ASSERT(__kmp_hier_max_units[index] != 0);
2010  if (tid >= num_hw_threads)
2011  tid = tid % num_hw_threads;
2012  return (tid / __kmp_hier_threads_per[index]) % __kmp_hier_max_units[index];
2013 }
2014 
2015 // Return the number of t1's per t2
2016 int __kmp_dispatch_get_t1_per_t2(kmp_hier_layer_e t1, kmp_hier_layer_e t2) {
2017  int i1 = t1 + 1;
2018  int i2 = t2 + 1;
2019  KMP_DEBUG_ASSERT(i1 <= i2);
2020  KMP_DEBUG_ASSERT(t1 != kmp_hier_layer_e::LAYER_LAST);
2021  KMP_DEBUG_ASSERT(t2 != kmp_hier_layer_e::LAYER_LAST);
2022  KMP_DEBUG_ASSERT(__kmp_hier_threads_per[i1] != 0);
2023  // (nthreads/t2) / (nthreads/t1) = t1 / t2
2024  return __kmp_hier_threads_per[i2] / __kmp_hier_threads_per[i1];
2025 }
2026 #endif // KMP_USE_HIER_SCHED
2027 
2028 // Parse /proc/cpuinfo (or an alternate file in the same format) to obtain the
2029 // affinity map.
2030 static int __kmp_affinity_create_cpuinfo_map(AddrUnsPair **address2os,
2031  int *line,
2032  kmp_i18n_id_t *const msg_id,
2033  FILE *f) {
2034  *address2os = NULL;
2035  *msg_id = kmp_i18n_null;
2036 
2037  // Scan of the file, and count the number of "processor" (osId) fields,
2038  // and find the highest value of <n> for a node_<n> field.
2039  char buf[256];
2040  unsigned num_records = 0;
2041  while (!feof(f)) {
2042  buf[sizeof(buf) - 1] = 1;
2043  if (!fgets(buf, sizeof(buf), f)) {
2044  // Read errors presumably because of EOF
2045  break;
2046  }
2047 
2048  char s1[] = "processor";
2049  if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
2050  num_records++;
2051  continue;
2052  }
2053 
2054  // FIXME - this will match "node_<n> <garbage>"
2055  unsigned level;
2056  if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
2057  if (nodeIdIndex + level >= maxIndex) {
2058  maxIndex = nodeIdIndex + level;
2059  }
2060  continue;
2061  }
2062  }
2063 
2064  // Check for empty file / no valid processor records, or too many. The number
2065  // of records can't exceed the number of valid bits in the affinity mask.
2066  if (num_records == 0) {
2067  *line = 0;
2068  *msg_id = kmp_i18n_str_NoProcRecords;
2069  return -1;
2070  }
2071  if (num_records > (unsigned)__kmp_xproc) {
2072  *line = 0;
2073  *msg_id = kmp_i18n_str_TooManyProcRecords;
2074  return -1;
2075  }
2076 
2077  // Set the file pointer back to the begginning, so that we can scan the file
2078  // again, this time performing a full parse of the data. Allocate a vector of
2079  // ProcCpuInfo object, where we will place the data. Adding an extra element
2080  // at the end allows us to remove a lot of extra checks for termination
2081  // conditions.
2082  if (fseek(f, 0, SEEK_SET) != 0) {
2083  *line = 0;
2084  *msg_id = kmp_i18n_str_CantRewindCpuinfo;
2085  return -1;
2086  }
2087 
2088  // Allocate the array of records to store the proc info in. The dummy
2089  // element at the end makes the logic in filling them out easier to code.
2090  unsigned **threadInfo =
2091  (unsigned **)__kmp_allocate((num_records + 1) * sizeof(unsigned *));
2092  unsigned i;
2093  for (i = 0; i <= num_records; i++) {
2094  threadInfo[i] =
2095  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2096  }
2097 
2098 #define CLEANUP_THREAD_INFO \
2099  for (i = 0; i <= num_records; i++) { \
2100  __kmp_free(threadInfo[i]); \
2101  } \
2102  __kmp_free(threadInfo);
2103 
2104  // A value of UINT_MAX means that we didn't find the field
2105  unsigned __index;
2106 
2107 #define INIT_PROC_INFO(p) \
2108  for (__index = 0; __index <= maxIndex; __index++) { \
2109  (p)[__index] = UINT_MAX; \
2110  }
2111 
2112  for (i = 0; i <= num_records; i++) {
2113  INIT_PROC_INFO(threadInfo[i]);
2114  }
2115 
2116  unsigned num_avail = 0;
2117  *line = 0;
2118  while (!feof(f)) {
2119  // Create an inner scoping level, so that all the goto targets at the end of
2120  // the loop appear in an outer scoping level. This avoids warnings about
2121  // jumping past an initialization to a target in the same block.
2122  {
2123  buf[sizeof(buf) - 1] = 1;
2124  bool long_line = false;
2125  if (!fgets(buf, sizeof(buf), f)) {
2126  // Read errors presumably because of EOF
2127  // If there is valid data in threadInfo[num_avail], then fake
2128  // a blank line in ensure that the last address gets parsed.
2129  bool valid = false;
2130  for (i = 0; i <= maxIndex; i++) {
2131  if (threadInfo[num_avail][i] != UINT_MAX) {
2132  valid = true;
2133  }
2134  }
2135  if (!valid) {
2136  break;
2137  }
2138  buf[0] = 0;
2139  } else if (!buf[sizeof(buf) - 1]) {
2140  // The line is longer than the buffer. Set a flag and don't
2141  // emit an error if we were going to ignore the line, anyway.
2142  long_line = true;
2143 
2144 #define CHECK_LINE \
2145  if (long_line) { \
2146  CLEANUP_THREAD_INFO; \
2147  *msg_id = kmp_i18n_str_LongLineCpuinfo; \
2148  return -1; \
2149  }
2150  }
2151  (*line)++;
2152 
2153  char s1[] = "processor";
2154  if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
2155  CHECK_LINE;
2156  char *p = strchr(buf + sizeof(s1) - 1, ':');
2157  unsigned val;
2158  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2159  goto no_val;
2160  if (threadInfo[num_avail][osIdIndex] != UINT_MAX)
2161 #if KMP_ARCH_AARCH64
2162  // Handle the old AArch64 /proc/cpuinfo layout differently,
2163  // it contains all of the 'processor' entries listed in a
2164  // single 'Processor' section, therefore the normal looking
2165  // for duplicates in that section will always fail.
2166  num_avail++;
2167 #else
2168  goto dup_field;
2169 #endif
2170  threadInfo[num_avail][osIdIndex] = val;
2171 #if KMP_OS_LINUX && !(KMP_ARCH_X86 || KMP_ARCH_X86_64)
2172  char path[256];
2173  KMP_SNPRINTF(
2174  path, sizeof(path),
2175  "/sys/devices/system/cpu/cpu%u/topology/physical_package_id",
2176  threadInfo[num_avail][osIdIndex]);
2177  __kmp_read_from_file(path, "%u", &threadInfo[num_avail][pkgIdIndex]);
2178 
2179  KMP_SNPRINTF(path, sizeof(path),
2180  "/sys/devices/system/cpu/cpu%u/topology/core_id",
2181  threadInfo[num_avail][osIdIndex]);
2182  __kmp_read_from_file(path, "%u", &threadInfo[num_avail][coreIdIndex]);
2183  continue;
2184 #else
2185  }
2186  char s2[] = "physical id";
2187  if (strncmp(buf, s2, sizeof(s2) - 1) == 0) {
2188  CHECK_LINE;
2189  char *p = strchr(buf + sizeof(s2) - 1, ':');
2190  unsigned val;
2191  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2192  goto no_val;
2193  if (threadInfo[num_avail][pkgIdIndex] != UINT_MAX)
2194  goto dup_field;
2195  threadInfo[num_avail][pkgIdIndex] = val;
2196  continue;
2197  }
2198  char s3[] = "core id";
2199  if (strncmp(buf, s3, sizeof(s3) - 1) == 0) {
2200  CHECK_LINE;
2201  char *p = strchr(buf + sizeof(s3) - 1, ':');
2202  unsigned val;
2203  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2204  goto no_val;
2205  if (threadInfo[num_avail][coreIdIndex] != UINT_MAX)
2206  goto dup_field;
2207  threadInfo[num_avail][coreIdIndex] = val;
2208  continue;
2209 #endif // KMP_OS_LINUX && USE_SYSFS_INFO
2210  }
2211  char s4[] = "thread id";
2212  if (strncmp(buf, s4, sizeof(s4) - 1) == 0) {
2213  CHECK_LINE;
2214  char *p = strchr(buf + sizeof(s4) - 1, ':');
2215  unsigned val;
2216  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2217  goto no_val;
2218  if (threadInfo[num_avail][threadIdIndex] != UINT_MAX)
2219  goto dup_field;
2220  threadInfo[num_avail][threadIdIndex] = val;
2221  continue;
2222  }
2223  unsigned level;
2224  if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
2225  CHECK_LINE;
2226  char *p = strchr(buf + sizeof(s4) - 1, ':');
2227  unsigned val;
2228  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2229  goto no_val;
2230  KMP_ASSERT(nodeIdIndex + level <= maxIndex);
2231  if (threadInfo[num_avail][nodeIdIndex + level] != UINT_MAX)
2232  goto dup_field;
2233  threadInfo[num_avail][nodeIdIndex + level] = val;
2234  continue;
2235  }
2236 
2237  // We didn't recognize the leading token on the line. There are lots of
2238  // leading tokens that we don't recognize - if the line isn't empty, go on
2239  // to the next line.
2240  if ((*buf != 0) && (*buf != '\n')) {
2241  // If the line is longer than the buffer, read characters
2242  // until we find a newline.
2243  if (long_line) {
2244  int ch;
2245  while (((ch = fgetc(f)) != EOF) && (ch != '\n'))
2246  ;
2247  }
2248  continue;
2249  }
2250 
2251  // A newline has signalled the end of the processor record.
2252  // Check that there aren't too many procs specified.
2253  if ((int)num_avail == __kmp_xproc) {
2254  CLEANUP_THREAD_INFO;
2255  *msg_id = kmp_i18n_str_TooManyEntries;
2256  return -1;
2257  }
2258 
2259  // Check for missing fields. The osId field must be there, and we
2260  // currently require that the physical id field is specified, also.
2261  if (threadInfo[num_avail][osIdIndex] == UINT_MAX) {
2262  CLEANUP_THREAD_INFO;
2263  *msg_id = kmp_i18n_str_MissingProcField;
2264  return -1;
2265  }
2266  if (threadInfo[0][pkgIdIndex] == UINT_MAX) {
2267  CLEANUP_THREAD_INFO;
2268  *msg_id = kmp_i18n_str_MissingPhysicalIDField;
2269  return -1;
2270  }
2271 
2272  // Skip this proc if it is not included in the machine model.
2273  if (!KMP_CPU_ISSET(threadInfo[num_avail][osIdIndex],
2274  __kmp_affin_fullMask)) {
2275  INIT_PROC_INFO(threadInfo[num_avail]);
2276  continue;
2277  }
2278 
2279  // We have a successful parse of this proc's info.
2280  // Increment the counter, and prepare for the next proc.
2281  num_avail++;
2282  KMP_ASSERT(num_avail <= num_records);
2283  INIT_PROC_INFO(threadInfo[num_avail]);
2284  }
2285  continue;
2286 
2287  no_val:
2288  CLEANUP_THREAD_INFO;
2289  *msg_id = kmp_i18n_str_MissingValCpuinfo;
2290  return -1;
2291 
2292  dup_field:
2293  CLEANUP_THREAD_INFO;
2294  *msg_id = kmp_i18n_str_DuplicateFieldCpuinfo;
2295  return -1;
2296  }
2297  *line = 0;
2298 
2299 #if KMP_MIC && REDUCE_TEAM_SIZE
2300  unsigned teamSize = 0;
2301 #endif // KMP_MIC && REDUCE_TEAM_SIZE
2302 
2303  // check for num_records == __kmp_xproc ???
2304 
2305  // If there's only one thread context to bind to, form an Address object with
2306  // depth 1 and return immediately (or, if affinity is off, set address2os to
2307  // NULL and return).
2308  //
2309  // If it is configured to omit the package level when there is only a single
2310  // package, the logic at the end of this routine won't work if there is only a
2311  // single thread - it would try to form an Address object with depth 0.
2312  KMP_ASSERT(num_avail > 0);
2313  KMP_ASSERT(num_avail <= num_records);
2314  if (num_avail == 1) {
2315  __kmp_ncores = 1;
2316  __kmp_nThreadsPerCore = nCoresPerPkg = nPackages = 1;
2317  if (__kmp_affinity_verbose) {
2318  if (!KMP_AFFINITY_CAPABLE()) {
2319  KMP_INFORM(AffNotCapableUseCpuinfo, "KMP_AFFINITY");
2320  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
2321  KMP_INFORM(Uniform, "KMP_AFFINITY");
2322  } else {
2323  char buf[KMP_AFFIN_MASK_PRINT_LEN];
2324  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
2325  __kmp_affin_fullMask);
2326  KMP_INFORM(AffCapableUseCpuinfo, "KMP_AFFINITY");
2327  if (__kmp_affinity_respect_mask) {
2328  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
2329  } else {
2330  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
2331  }
2332  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
2333  KMP_INFORM(Uniform, "KMP_AFFINITY");
2334  }
2335  int index;
2336  kmp_str_buf_t buf;
2337  __kmp_str_buf_init(&buf);
2338  __kmp_str_buf_print(&buf, "1");
2339  for (index = maxIndex - 1; index > pkgIdIndex; index--) {
2340  __kmp_str_buf_print(&buf, " x 1");
2341  }
2342  KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, 1, 1, 1);
2343  __kmp_str_buf_free(&buf);
2344  }
2345 
2346  if (__kmp_affinity_type == affinity_none) {
2347  CLEANUP_THREAD_INFO;
2348  return 0;
2349  }
2350 
2351  *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair));
2352  Address addr(1);
2353  addr.labels[0] = threadInfo[0][pkgIdIndex];
2354  (*address2os)[0] = AddrUnsPair(addr, threadInfo[0][osIdIndex]);
2355 
2356  if (__kmp_affinity_gran_levels < 0) {
2357  __kmp_affinity_gran_levels = 0;
2358  }
2359 
2360  if (__kmp_affinity_verbose) {
2361  __kmp_affinity_print_topology(*address2os, 1, 1, 0, -1, -1);
2362  }
2363 
2364  CLEANUP_THREAD_INFO;
2365  return 1;
2366  }
2367 
2368  // Sort the threadInfo table by physical Id.
2369  qsort(threadInfo, num_avail, sizeof(*threadInfo),
2370  __kmp_affinity_cmp_ProcCpuInfo_phys_id);
2371 
2372  // The table is now sorted by pkgId / coreId / threadId, but we really don't
2373  // know the radix of any of the fields. pkgId's may be sparsely assigned among
2374  // the chips on a system. Although coreId's are usually assigned
2375  // [0 .. coresPerPkg-1] and threadId's are usually assigned
2376  // [0..threadsPerCore-1], we don't want to make any such assumptions.
2377  //
2378  // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
2379  // total # packages) are at this point - we want to determine that now. We
2380  // only have an upper bound on the first two figures.
2381  unsigned *counts =
2382  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2383  unsigned *maxCt =
2384  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2385  unsigned *totals =
2386  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2387  unsigned *lastId =
2388  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2389 
2390  bool assign_thread_ids = false;
2391  unsigned threadIdCt;
2392  unsigned index;
2393 
2394 restart_radix_check:
2395  threadIdCt = 0;
2396 
2397  // Initialize the counter arrays with data from threadInfo[0].
2398  if (assign_thread_ids) {
2399  if (threadInfo[0][threadIdIndex] == UINT_MAX) {
2400  threadInfo[0][threadIdIndex] = threadIdCt++;
2401  } else if (threadIdCt <= threadInfo[0][threadIdIndex]) {
2402  threadIdCt = threadInfo[0][threadIdIndex] + 1;
2403  }
2404  }
2405  for (index = 0; index <= maxIndex; index++) {
2406  counts[index] = 1;
2407  maxCt[index] = 1;
2408  totals[index] = 1;
2409  lastId[index] = threadInfo[0][index];
2410  ;
2411  }
2412 
2413  // Run through the rest of the OS procs.
2414  for (i = 1; i < num_avail; i++) {
2415  // Find the most significant index whose id differs from the id for the
2416  // previous OS proc.
2417  for (index = maxIndex; index >= threadIdIndex; index--) {
2418  if (assign_thread_ids && (index == threadIdIndex)) {
2419  // Auto-assign the thread id field if it wasn't specified.
2420  if (threadInfo[i][threadIdIndex] == UINT_MAX) {
2421  threadInfo[i][threadIdIndex] = threadIdCt++;
2422  }
2423  // Apparently the thread id field was specified for some entries and not
2424  // others. Start the thread id counter off at the next higher thread id.
2425  else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
2426  threadIdCt = threadInfo[i][threadIdIndex] + 1;
2427  }
2428  }
2429  if (threadInfo[i][index] != lastId[index]) {
2430  // Run through all indices which are less significant, and reset the
2431  // counts to 1. At all levels up to and including index, we need to
2432  // increment the totals and record the last id.
2433  unsigned index2;
2434  for (index2 = threadIdIndex; index2 < index; index2++) {
2435  totals[index2]++;
2436  if (counts[index2] > maxCt[index2]) {
2437  maxCt[index2] = counts[index2];
2438  }
2439  counts[index2] = 1;
2440  lastId[index2] = threadInfo[i][index2];
2441  }
2442  counts[index]++;
2443  totals[index]++;
2444  lastId[index] = threadInfo[i][index];
2445 
2446  if (assign_thread_ids && (index > threadIdIndex)) {
2447 
2448 #if KMP_MIC && REDUCE_TEAM_SIZE
2449  // The default team size is the total #threads in the machine
2450  // minus 1 thread for every core that has 3 or more threads.
2451  teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
2452 #endif // KMP_MIC && REDUCE_TEAM_SIZE
2453 
2454  // Restart the thread counter, as we are on a new core.
2455  threadIdCt = 0;
2456 
2457  // Auto-assign the thread id field if it wasn't specified.
2458  if (threadInfo[i][threadIdIndex] == UINT_MAX) {
2459  threadInfo[i][threadIdIndex] = threadIdCt++;
2460  }
2461 
2462  // Aparrently the thread id field was specified for some entries and
2463  // not others. Start the thread id counter off at the next higher
2464  // thread id.
2465  else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
2466  threadIdCt = threadInfo[i][threadIdIndex] + 1;
2467  }
2468  }
2469  break;
2470  }
2471  }
2472  if (index < threadIdIndex) {
2473  // If thread ids were specified, it is an error if they are not unique.
2474  // Also, check that we waven't already restarted the loop (to be safe -
2475  // shouldn't need to).
2476  if ((threadInfo[i][threadIdIndex] != UINT_MAX) || assign_thread_ids) {
2477  __kmp_free(lastId);
2478  __kmp_free(totals);
2479  __kmp_free(maxCt);
2480  __kmp_free(counts);
2481  CLEANUP_THREAD_INFO;
2482  *msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
2483  return -1;
2484  }
2485 
2486  // If the thread ids were not specified and we see entries entries that
2487  // are duplicates, start the loop over and assign the thread ids manually.
2488  assign_thread_ids = true;
2489  goto restart_radix_check;
2490  }
2491  }
2492 
2493 #if KMP_MIC && REDUCE_TEAM_SIZE
2494  // The default team size is the total #threads in the machine
2495  // minus 1 thread for every core that has 3 or more threads.
2496  teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
2497 #endif // KMP_MIC && REDUCE_TEAM_SIZE
2498 
2499  for (index = threadIdIndex; index <= maxIndex; index++) {
2500  if (counts[index] > maxCt[index]) {
2501  maxCt[index] = counts[index];
2502  }
2503  }
2504 
2505  __kmp_nThreadsPerCore = maxCt[threadIdIndex];
2506  nCoresPerPkg = maxCt[coreIdIndex];
2507  nPackages = totals[pkgIdIndex];
2508 
2509  // Check to see if the machine topology is uniform
2510  unsigned prod = totals[maxIndex];
2511  for (index = threadIdIndex; index < maxIndex; index++) {
2512  prod *= maxCt[index];
2513  }
2514  bool uniform = (prod == totals[threadIdIndex]);
2515 
2516  // When affinity is off, this routine will still be called to set
2517  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
2518  // Make sure all these vars are set correctly, and return now if affinity is
2519  // not enabled.
2520  __kmp_ncores = totals[coreIdIndex];
2521 
2522  if (__kmp_affinity_verbose) {
2523  if (!KMP_AFFINITY_CAPABLE()) {
2524  KMP_INFORM(AffNotCapableUseCpuinfo, "KMP_AFFINITY");
2525  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
2526  if (uniform) {
2527  KMP_INFORM(Uniform, "KMP_AFFINITY");
2528  } else {
2529  KMP_INFORM(NonUniform, "KMP_AFFINITY");
2530  }
2531  } else {
2532  char buf[KMP_AFFIN_MASK_PRINT_LEN];
2533  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
2534  __kmp_affin_fullMask);
2535  KMP_INFORM(AffCapableUseCpuinfo, "KMP_AFFINITY");
2536  if (__kmp_affinity_respect_mask) {
2537  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
2538  } else {
2539  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
2540  }
2541  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
2542  if (uniform) {
2543  KMP_INFORM(Uniform, "KMP_AFFINITY");
2544  } else {
2545  KMP_INFORM(NonUniform, "KMP_AFFINITY");
2546  }
2547  }
2548  kmp_str_buf_t buf;
2549  __kmp_str_buf_init(&buf);
2550 
2551  __kmp_str_buf_print(&buf, "%d", totals[maxIndex]);
2552  for (index = maxIndex - 1; index >= pkgIdIndex; index--) {
2553  __kmp_str_buf_print(&buf, " x %d", maxCt[index]);
2554  }
2555  KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, maxCt[coreIdIndex],
2556  maxCt[threadIdIndex], __kmp_ncores);
2557 
2558  __kmp_str_buf_free(&buf);
2559  }
2560 
2561 #if KMP_MIC && REDUCE_TEAM_SIZE
2562  // Set the default team size.
2563  if ((__kmp_dflt_team_nth == 0) && (teamSize > 0)) {
2564  __kmp_dflt_team_nth = teamSize;
2565  KA_TRACE(20, ("__kmp_affinity_create_cpuinfo_map: setting "
2566  "__kmp_dflt_team_nth = %d\n",
2567  __kmp_dflt_team_nth));
2568  }
2569 #endif // KMP_MIC && REDUCE_TEAM_SIZE
2570 
2571  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
2572  KMP_DEBUG_ASSERT(num_avail == (unsigned)__kmp_avail_proc);
2573  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
2574  for (i = 0; i < num_avail; ++i) { // fill the os indices
2575  __kmp_pu_os_idx[i] = threadInfo[i][osIdIndex];
2576  }
2577 
2578  if (__kmp_affinity_type == affinity_none) {
2579  __kmp_free(lastId);
2580  __kmp_free(totals);
2581  __kmp_free(maxCt);
2582  __kmp_free(counts);
2583  CLEANUP_THREAD_INFO;
2584  return 0;
2585  }
2586 
2587  // Count the number of levels which have more nodes at that level than at the
2588  // parent's level (with there being an implicit root node of the top level).
2589  // This is equivalent to saying that there is at least one node at this level
2590  // which has a sibling. These levels are in the map, and the package level is
2591  // always in the map.
2592  bool *inMap = (bool *)__kmp_allocate((maxIndex + 1) * sizeof(bool));
2593  for (index = threadIdIndex; index < maxIndex; index++) {
2594  KMP_ASSERT(totals[index] >= totals[index + 1]);
2595  inMap[index] = (totals[index] > totals[index + 1]);
2596  }
2597  inMap[maxIndex] = (totals[maxIndex] > 1);
2598  inMap[pkgIdIndex] = true;
2599 
2600  int depth = 0;
2601  for (index = threadIdIndex; index <= maxIndex; index++) {
2602  if (inMap[index]) {
2603  depth++;
2604  }
2605  }
2606  KMP_ASSERT(depth > 0);
2607 
2608  // Construct the data structure that is to be returned.
2609  *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * num_avail);
2610  int pkgLevel = -1;
2611  int coreLevel = -1;
2612  int threadLevel = -1;
2613 
2614  for (i = 0; i < num_avail; ++i) {
2615  Address addr(depth);
2616  unsigned os = threadInfo[i][osIdIndex];
2617  int src_index;
2618  int dst_index = 0;
2619 
2620  for (src_index = maxIndex; src_index >= threadIdIndex; src_index--) {
2621  if (!inMap[src_index]) {
2622  continue;
2623  }
2624  addr.labels[dst_index] = threadInfo[i][src_index];
2625  if (src_index == pkgIdIndex) {
2626  pkgLevel = dst_index;
2627  } else if (src_index == coreIdIndex) {
2628  coreLevel = dst_index;
2629  } else if (src_index == threadIdIndex) {
2630  threadLevel = dst_index;
2631  }
2632  dst_index++;
2633  }
2634  (*address2os)[i] = AddrUnsPair(addr, os);
2635  }
2636 
2637  if (__kmp_affinity_gran_levels < 0) {
2638  // Set the granularity level based on what levels are modeled
2639  // in the machine topology map.
2640  unsigned src_index;
2641  __kmp_affinity_gran_levels = 0;
2642  for (src_index = threadIdIndex; src_index <= maxIndex; src_index++) {
2643  if (!inMap[src_index]) {
2644  continue;
2645  }
2646  switch (src_index) {
2647  case threadIdIndex:
2648  if (__kmp_affinity_gran > affinity_gran_thread) {
2649  __kmp_affinity_gran_levels++;
2650  }
2651 
2652  break;
2653  case coreIdIndex:
2654  if (__kmp_affinity_gran > affinity_gran_core) {
2655  __kmp_affinity_gran_levels++;
2656  }
2657  break;
2658 
2659  case pkgIdIndex:
2660  if (__kmp_affinity_gran > affinity_gran_package) {
2661  __kmp_affinity_gran_levels++;
2662  }
2663  break;
2664  }
2665  }
2666  }
2667 
2668  if (__kmp_affinity_verbose) {
2669  __kmp_affinity_print_topology(*address2os, num_avail, depth, pkgLevel,
2670  coreLevel, threadLevel);
2671  }
2672 
2673  __kmp_free(inMap);
2674  __kmp_free(lastId);
2675  __kmp_free(totals);
2676  __kmp_free(maxCt);
2677  __kmp_free(counts);
2678  CLEANUP_THREAD_INFO;
2679  return depth;
2680 }
2681 
2682 // Create and return a table of affinity masks, indexed by OS thread ID.
2683 // This routine handles OR'ing together all the affinity masks of threads
2684 // that are sufficiently close, if granularity > fine.
2685 static kmp_affin_mask_t *__kmp_create_masks(unsigned *maxIndex,
2686  unsigned *numUnique,
2687  AddrUnsPair *address2os,
2688  unsigned numAddrs) {
2689  // First form a table of affinity masks in order of OS thread id.
2690  unsigned depth;
2691  unsigned maxOsId;
2692  unsigned i;
2693 
2694  KMP_ASSERT(numAddrs > 0);
2695  depth = address2os[0].first.depth;
2696 
2697  maxOsId = 0;
2698  for (i = numAddrs - 1;; --i) {
2699  unsigned osId = address2os[i].second;
2700  if (osId > maxOsId) {
2701  maxOsId = osId;
2702  }
2703  if (i == 0)
2704  break;
2705  }
2706  kmp_affin_mask_t *osId2Mask;
2707  KMP_CPU_ALLOC_ARRAY(osId2Mask, (maxOsId + 1));
2708 
2709  // Sort the address2os table according to physical order. Doing so will put
2710  // all threads on the same core/package/node in consecutive locations.
2711  qsort(address2os, numAddrs, sizeof(*address2os),
2712  __kmp_affinity_cmp_Address_labels);
2713 
2714  KMP_ASSERT(__kmp_affinity_gran_levels >= 0);
2715  if (__kmp_affinity_verbose && (__kmp_affinity_gran_levels > 0)) {
2716  KMP_INFORM(ThreadsMigrate, "KMP_AFFINITY", __kmp_affinity_gran_levels);
2717  }
2718  if (__kmp_affinity_gran_levels >= (int)depth) {
2719  if (__kmp_affinity_verbose ||
2720  (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) {
2721  KMP_WARNING(AffThreadsMayMigrate);
2722  }
2723  }
2724 
2725  // Run through the table, forming the masks for all threads on each core.
2726  // Threads on the same core will have identical "Address" objects, not
2727  // considering the last level, which must be the thread id. All threads on a
2728  // core will appear consecutively.
2729  unsigned unique = 0;
2730  unsigned j = 0; // index of 1st thread on core
2731  unsigned leader = 0;
2732  Address *leaderAddr = &(address2os[0].first);
2733  kmp_affin_mask_t *sum;
2734  KMP_CPU_ALLOC_ON_STACK(sum);
2735  KMP_CPU_ZERO(sum);
2736  KMP_CPU_SET(address2os[0].second, sum);
2737  for (i = 1; i < numAddrs; i++) {
2738  // If this thread is sufficiently close to the leader (within the
2739  // granularity setting), then set the bit for this os thread in the
2740  // affinity mask for this group, and go on to the next thread.
2741  if (leaderAddr->isClose(address2os[i].first, __kmp_affinity_gran_levels)) {
2742  KMP_CPU_SET(address2os[i].second, sum);
2743  continue;
2744  }
2745 
2746  // For every thread in this group, copy the mask to the thread's entry in
2747  // the osId2Mask table. Mark the first address as a leader.
2748  for (; j < i; j++) {
2749  unsigned osId = address2os[j].second;
2750  KMP_DEBUG_ASSERT(osId <= maxOsId);
2751  kmp_affin_mask_t *mask = KMP_CPU_INDEX(osId2Mask, osId);
2752  KMP_CPU_COPY(mask, sum);
2753  address2os[j].first.leader = (j == leader);
2754  }
2755  unique++;
2756 
2757  // Start a new mask.
2758  leader = i;
2759  leaderAddr = &(address2os[i].first);
2760  KMP_CPU_ZERO(sum);
2761  KMP_CPU_SET(address2os[i].second, sum);
2762  }
2763 
2764  // For every thread in last group, copy the mask to the thread's
2765  // entry in the osId2Mask table.
2766  for (; j < i; j++) {
2767  unsigned osId = address2os[j].second;
2768  KMP_DEBUG_ASSERT(osId <= maxOsId);
2769  kmp_affin_mask_t *mask = KMP_CPU_INDEX(osId2Mask, osId);
2770  KMP_CPU_COPY(mask, sum);
2771  address2os[j].first.leader = (j == leader);
2772  }
2773  unique++;
2774  KMP_CPU_FREE_FROM_STACK(sum);
2775 
2776  *maxIndex = maxOsId;
2777  *numUnique = unique;
2778  return osId2Mask;
2779 }
2780 
2781 // Stuff for the affinity proclist parsers. It's easier to declare these vars
2782 // as file-static than to try and pass them through the calling sequence of
2783 // the recursive-descent OMP_PLACES parser.
2784 static kmp_affin_mask_t *newMasks;
2785 static int numNewMasks;
2786 static int nextNewMask;
2787 
2788 #define ADD_MASK(_mask) \
2789  { \
2790  if (nextNewMask >= numNewMasks) { \
2791  int i; \
2792  numNewMasks *= 2; \
2793  kmp_affin_mask_t *temp; \
2794  KMP_CPU_INTERNAL_ALLOC_ARRAY(temp, numNewMasks); \
2795  for (i = 0; i < numNewMasks / 2; i++) { \
2796  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i); \
2797  kmp_affin_mask_t *dest = KMP_CPU_INDEX(temp, i); \
2798  KMP_CPU_COPY(dest, src); \
2799  } \
2800  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks / 2); \
2801  newMasks = temp; \
2802  } \
2803  KMP_CPU_COPY(KMP_CPU_INDEX(newMasks, nextNewMask), (_mask)); \
2804  nextNewMask++; \
2805  }
2806 
2807 #define ADD_MASK_OSID(_osId, _osId2Mask, _maxOsId) \
2808  { \
2809  if (((_osId) > _maxOsId) || \
2810  (!KMP_CPU_ISSET((_osId), KMP_CPU_INDEX((_osId2Mask), (_osId))))) { \
2811  if (__kmp_affinity_verbose || \
2812  (__kmp_affinity_warnings && \
2813  (__kmp_affinity_type != affinity_none))) { \
2814  KMP_WARNING(AffIgnoreInvalidProcID, _osId); \
2815  } \
2816  } else { \
2817  ADD_MASK(KMP_CPU_INDEX(_osId2Mask, (_osId))); \
2818  } \
2819  }
2820 
2821 // Re-parse the proclist (for the explicit affinity type), and form the list
2822 // of affinity newMasks indexed by gtid.
2823 static void __kmp_affinity_process_proclist(kmp_affin_mask_t **out_masks,
2824  unsigned int *out_numMasks,
2825  const char *proclist,
2826  kmp_affin_mask_t *osId2Mask,
2827  int maxOsId) {
2828  int i;
2829  const char *scan = proclist;
2830  const char *next = proclist;
2831 
2832  // We use malloc() for the temporary mask vector, so that we can use
2833  // realloc() to extend it.
2834  numNewMasks = 2;
2835  KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
2836  nextNewMask = 0;
2837  kmp_affin_mask_t *sumMask;
2838  KMP_CPU_ALLOC(sumMask);
2839  int setSize = 0;
2840 
2841  for (;;) {
2842  int start, end, stride;
2843 
2844  SKIP_WS(scan);
2845  next = scan;
2846  if (*next == '\0') {
2847  break;
2848  }
2849 
2850  if (*next == '{') {
2851  int num;
2852  setSize = 0;
2853  next++; // skip '{'
2854  SKIP_WS(next);
2855  scan = next;
2856 
2857  // Read the first integer in the set.
2858  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad proclist");
2859  SKIP_DIGITS(next);
2860  num = __kmp_str_to_int(scan, *next);
2861  KMP_ASSERT2(num >= 0, "bad explicit proc list");
2862 
2863  // Copy the mask for that osId to the sum (union) mask.
2864  if ((num > maxOsId) ||
2865  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
2866  if (__kmp_affinity_verbose ||
2867  (__kmp_affinity_warnings &&
2868  (__kmp_affinity_type != affinity_none))) {
2869  KMP_WARNING(AffIgnoreInvalidProcID, num);
2870  }
2871  KMP_CPU_ZERO(sumMask);
2872  } else {
2873  KMP_CPU_COPY(sumMask, KMP_CPU_INDEX(osId2Mask, num));
2874  setSize = 1;
2875  }
2876 
2877  for (;;) {
2878  // Check for end of set.
2879  SKIP_WS(next);
2880  if (*next == '}') {
2881  next++; // skip '}'
2882  break;
2883  }
2884 
2885  // Skip optional comma.
2886  if (*next == ',') {
2887  next++;
2888  }
2889  SKIP_WS(next);
2890 
2891  // Read the next integer in the set.
2892  scan = next;
2893  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
2894 
2895  SKIP_DIGITS(next);
2896  num = __kmp_str_to_int(scan, *next);
2897  KMP_ASSERT2(num >= 0, "bad explicit proc list");
2898 
2899  // Add the mask for that osId to the sum mask.
2900  if ((num > maxOsId) ||
2901  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
2902  if (__kmp_affinity_verbose ||
2903  (__kmp_affinity_warnings &&
2904  (__kmp_affinity_type != affinity_none))) {
2905  KMP_WARNING(AffIgnoreInvalidProcID, num);
2906  }
2907  } else {
2908  KMP_CPU_UNION(sumMask, KMP_CPU_INDEX(osId2Mask, num));
2909  setSize++;
2910  }
2911  }
2912  if (setSize > 0) {
2913  ADD_MASK(sumMask);
2914  }
2915 
2916  SKIP_WS(next);
2917  if (*next == ',') {
2918  next++;
2919  }
2920  scan = next;
2921  continue;
2922  }
2923 
2924  // Read the first integer.
2925  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
2926  SKIP_DIGITS(next);
2927  start = __kmp_str_to_int(scan, *next);
2928  KMP_ASSERT2(start >= 0, "bad explicit proc list");
2929  SKIP_WS(next);
2930 
2931  // If this isn't a range, then add a mask to the list and go on.
2932  if (*next != '-') {
2933  ADD_MASK_OSID(start, osId2Mask, maxOsId);
2934 
2935  // Skip optional comma.
2936  if (*next == ',') {
2937  next++;
2938  }
2939  scan = next;
2940  continue;
2941  }
2942 
2943  // This is a range. Skip over the '-' and read in the 2nd int.
2944  next++; // skip '-'
2945  SKIP_WS(next);
2946  scan = next;
2947  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
2948  SKIP_DIGITS(next);
2949  end = __kmp_str_to_int(scan, *next);
2950  KMP_ASSERT2(end >= 0, "bad explicit proc list");
2951 
2952  // Check for a stride parameter
2953  stride = 1;
2954  SKIP_WS(next);
2955  if (*next == ':') {
2956  // A stride is specified. Skip over the ':" and read the 3rd int.
2957  int sign = +1;
2958  next++; // skip ':'
2959  SKIP_WS(next);
2960  scan = next;
2961  if (*next == '-') {
2962  sign = -1;
2963  next++;
2964  SKIP_WS(next);
2965  scan = next;
2966  }
2967  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
2968  SKIP_DIGITS(next);
2969  stride = __kmp_str_to_int(scan, *next);
2970  KMP_ASSERT2(stride >= 0, "bad explicit proc list");
2971  stride *= sign;
2972  }
2973 
2974  // Do some range checks.
2975  KMP_ASSERT2(stride != 0, "bad explicit proc list");
2976  if (stride > 0) {
2977  KMP_ASSERT2(start <= end, "bad explicit proc list");
2978  } else {
2979  KMP_ASSERT2(start >= end, "bad explicit proc list");
2980  }
2981  KMP_ASSERT2((end - start) / stride <= 65536, "bad explicit proc list");
2982 
2983  // Add the mask for each OS proc # to the list.
2984  if (stride > 0) {
2985  do {
2986  ADD_MASK_OSID(start, osId2Mask, maxOsId);
2987  start += stride;
2988  } while (start <= end);
2989  } else {
2990  do {
2991  ADD_MASK_OSID(start, osId2Mask, maxOsId);
2992  start += stride;
2993  } while (start >= end);
2994  }
2995 
2996  // Skip optional comma.
2997  SKIP_WS(next);
2998  if (*next == ',') {
2999  next++;
3000  }
3001  scan = next;
3002  }
3003 
3004  *out_numMasks = nextNewMask;
3005  if (nextNewMask == 0) {
3006  *out_masks = NULL;
3007  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3008  return;
3009  }
3010  KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
3011  for (i = 0; i < nextNewMask; i++) {
3012  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
3013  kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
3014  KMP_CPU_COPY(dest, src);
3015  }
3016  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3017  KMP_CPU_FREE(sumMask);
3018 }
3019 
3020 #if OMP_40_ENABLED
3021 
3022 /*-----------------------------------------------------------------------------
3023 Re-parse the OMP_PLACES proc id list, forming the newMasks for the different
3024 places. Again, Here is the grammar:
3025 
3026 place_list := place
3027 place_list := place , place_list
3028 place := num
3029 place := place : num
3030 place := place : num : signed
3031 place := { subplacelist }
3032 place := ! place // (lowest priority)
3033 subplace_list := subplace
3034 subplace_list := subplace , subplace_list
3035 subplace := num
3036 subplace := num : num
3037 subplace := num : num : signed
3038 signed := num
3039 signed := + signed
3040 signed := - signed
3041 -----------------------------------------------------------------------------*/
3042 
3043 static void __kmp_process_subplace_list(const char **scan,
3044  kmp_affin_mask_t *osId2Mask,
3045  int maxOsId, kmp_affin_mask_t *tempMask,
3046  int *setSize) {
3047  const char *next;
3048 
3049  for (;;) {
3050  int start, count, stride, i;
3051 
3052  // Read in the starting proc id
3053  SKIP_WS(*scan);
3054  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3055  next = *scan;
3056  SKIP_DIGITS(next);
3057  start = __kmp_str_to_int(*scan, *next);
3058  KMP_ASSERT(start >= 0);
3059  *scan = next;
3060 
3061  // valid follow sets are ',' ':' and '}'
3062  SKIP_WS(*scan);
3063  if (**scan == '}' || **scan == ',') {
3064  if ((start > maxOsId) ||
3065  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3066  if (__kmp_affinity_verbose ||
3067  (__kmp_affinity_warnings &&
3068  (__kmp_affinity_type != affinity_none))) {
3069  KMP_WARNING(AffIgnoreInvalidProcID, start);
3070  }
3071  } else {
3072  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3073  (*setSize)++;
3074  }
3075  if (**scan == '}') {
3076  break;
3077  }
3078  (*scan)++; // skip ','
3079  continue;
3080  }
3081  KMP_ASSERT2(**scan == ':', "bad explicit places list");
3082  (*scan)++; // skip ':'
3083 
3084  // Read count parameter
3085  SKIP_WS(*scan);
3086  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3087  next = *scan;
3088  SKIP_DIGITS(next);
3089  count = __kmp_str_to_int(*scan, *next);
3090  KMP_ASSERT(count >= 0);
3091  *scan = next;
3092 
3093  // valid follow sets are ',' ':' and '}'
3094  SKIP_WS(*scan);
3095  if (**scan == '}' || **scan == ',') {
3096  for (i = 0; i < count; i++) {
3097  if ((start > maxOsId) ||
3098  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3099  if (__kmp_affinity_verbose ||
3100  (__kmp_affinity_warnings &&
3101  (__kmp_affinity_type != affinity_none))) {
3102  KMP_WARNING(AffIgnoreInvalidProcID, start);
3103  }
3104  break; // don't proliferate warnings for large count
3105  } else {
3106  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3107  start++;
3108  (*setSize)++;
3109  }
3110  }
3111  if (**scan == '}') {
3112  break;
3113  }
3114  (*scan)++; // skip ','
3115  continue;
3116  }
3117  KMP_ASSERT2(**scan == ':', "bad explicit places list");
3118  (*scan)++; // skip ':'
3119 
3120  // Read stride parameter
3121  int sign = +1;
3122  for (;;) {
3123  SKIP_WS(*scan);
3124  if (**scan == '+') {
3125  (*scan)++; // skip '+'
3126  continue;
3127  }
3128  if (**scan == '-') {
3129  sign *= -1;
3130  (*scan)++; // skip '-'
3131  continue;
3132  }
3133  break;
3134  }
3135  SKIP_WS(*scan);
3136  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3137  next = *scan;
3138  SKIP_DIGITS(next);
3139  stride = __kmp_str_to_int(*scan, *next);
3140  KMP_ASSERT(stride >= 0);
3141  *scan = next;
3142  stride *= sign;
3143 
3144  // valid follow sets are ',' and '}'
3145  SKIP_WS(*scan);
3146  if (**scan == '}' || **scan == ',') {
3147  for (i = 0; i < count; i++) {
3148  if ((start > maxOsId) ||
3149  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3150  if (__kmp_affinity_verbose ||
3151  (__kmp_affinity_warnings &&
3152  (__kmp_affinity_type != affinity_none))) {
3153  KMP_WARNING(AffIgnoreInvalidProcID, start);
3154  }
3155  break; // don't proliferate warnings for large count
3156  } else {
3157  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3158  start += stride;
3159  (*setSize)++;
3160  }
3161  }
3162  if (**scan == '}') {
3163  break;
3164  }
3165  (*scan)++; // skip ','
3166  continue;
3167  }
3168 
3169  KMP_ASSERT2(0, "bad explicit places list");
3170  }
3171 }
3172 
3173 static void __kmp_process_place(const char **scan, kmp_affin_mask_t *osId2Mask,
3174  int maxOsId, kmp_affin_mask_t *tempMask,
3175  int *setSize) {
3176  const char *next;
3177 
3178  // valid follow sets are '{' '!' and num
3179  SKIP_WS(*scan);
3180  if (**scan == '{') {
3181  (*scan)++; // skip '{'
3182  __kmp_process_subplace_list(scan, osId2Mask, maxOsId, tempMask, setSize);
3183  KMP_ASSERT2(**scan == '}', "bad explicit places list");
3184  (*scan)++; // skip '}'
3185  } else if (**scan == '!') {
3186  (*scan)++; // skip '!'
3187  __kmp_process_place(scan, osId2Mask, maxOsId, tempMask, setSize);
3188  KMP_CPU_COMPLEMENT(maxOsId, tempMask);
3189  } else if ((**scan >= '0') && (**scan <= '9')) {
3190  next = *scan;
3191  SKIP_DIGITS(next);
3192  int num = __kmp_str_to_int(*scan, *next);
3193  KMP_ASSERT(num >= 0);
3194  if ((num > maxOsId) ||
3195  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
3196  if (__kmp_affinity_verbose ||
3197  (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) {
3198  KMP_WARNING(AffIgnoreInvalidProcID, num);
3199  }
3200  } else {
3201  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, num));
3202  (*setSize)++;
3203  }
3204  *scan = next; // skip num
3205  } else {
3206  KMP_ASSERT2(0, "bad explicit places list");
3207  }
3208 }
3209 
3210 // static void
3211 void __kmp_affinity_process_placelist(kmp_affin_mask_t **out_masks,
3212  unsigned int *out_numMasks,
3213  const char *placelist,
3214  kmp_affin_mask_t *osId2Mask,
3215  int maxOsId) {
3216  int i, j, count, stride, sign;
3217  const char *scan = placelist;
3218  const char *next = placelist;
3219 
3220  numNewMasks = 2;
3221  KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
3222  nextNewMask = 0;
3223 
3224  // tempMask is modified based on the previous or initial
3225  // place to form the current place
3226  // previousMask contains the previous place
3227  kmp_affin_mask_t *tempMask;
3228  kmp_affin_mask_t *previousMask;
3229  KMP_CPU_ALLOC(tempMask);
3230  KMP_CPU_ZERO(tempMask);
3231  KMP_CPU_ALLOC(previousMask);
3232  KMP_CPU_ZERO(previousMask);
3233  int setSize = 0;
3234 
3235  for (;;) {
3236  __kmp_process_place(&scan, osId2Mask, maxOsId, tempMask, &setSize);
3237 
3238  // valid follow sets are ',' ':' and EOL
3239  SKIP_WS(scan);
3240  if (*scan == '\0' || *scan == ',') {
3241  if (setSize > 0) {
3242  ADD_MASK(tempMask);
3243  }
3244  KMP_CPU_ZERO(tempMask);
3245  setSize = 0;
3246  if (*scan == '\0') {
3247  break;
3248  }
3249  scan++; // skip ','
3250  continue;
3251  }
3252 
3253  KMP_ASSERT2(*scan == ':', "bad explicit places list");
3254  scan++; // skip ':'
3255 
3256  // Read count parameter
3257  SKIP_WS(scan);
3258  KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
3259  next = scan;
3260  SKIP_DIGITS(next);
3261  count = __kmp_str_to_int(scan, *next);
3262  KMP_ASSERT(count >= 0);
3263  scan = next;
3264 
3265  // valid follow sets are ',' ':' and EOL
3266  SKIP_WS(scan);
3267  if (*scan == '\0' || *scan == ',') {
3268  stride = +1;
3269  } else {
3270  KMP_ASSERT2(*scan == ':', "bad explicit places list");
3271  scan++; // skip ':'
3272 
3273  // Read stride parameter
3274  sign = +1;
3275  for (;;) {
3276  SKIP_WS(scan);
3277  if (*scan == '+') {
3278  scan++; // skip '+'
3279  continue;
3280  }
3281  if (*scan == '-') {
3282  sign *= -1;
3283  scan++; // skip '-'
3284  continue;
3285  }
3286  break;
3287  }
3288  SKIP_WS(scan);
3289  KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
3290  next = scan;
3291  SKIP_DIGITS(next);
3292  stride = __kmp_str_to_int(scan, *next);
3293  KMP_DEBUG_ASSERT(stride >= 0);
3294  scan = next;
3295  stride *= sign;
3296  }
3297 
3298  // Add places determined by initial_place : count : stride
3299  for (i = 0; i < count; i++) {
3300  if (setSize == 0) {
3301  break;
3302  }
3303  // Add the current place, then build the next place (tempMask) from that
3304  KMP_CPU_COPY(previousMask, tempMask);
3305  ADD_MASK(previousMask);
3306  KMP_CPU_ZERO(tempMask);
3307  setSize = 0;
3308  KMP_CPU_SET_ITERATE(j, previousMask) {
3309  if (!KMP_CPU_ISSET(j, previousMask)) {
3310  continue;
3311  }
3312  if ((j + stride > maxOsId) || (j + stride < 0) ||
3313  (!KMP_CPU_ISSET(j, __kmp_affin_fullMask)) ||
3314  (!KMP_CPU_ISSET(j + stride,
3315  KMP_CPU_INDEX(osId2Mask, j + stride)))) {
3316  if ((__kmp_affinity_verbose ||
3317  (__kmp_affinity_warnings &&
3318  (__kmp_affinity_type != affinity_none))) &&
3319  i < count - 1) {
3320  KMP_WARNING(AffIgnoreInvalidProcID, j + stride);
3321  }
3322  continue;
3323  }
3324  KMP_CPU_SET(j + stride, tempMask);
3325  setSize++;
3326  }
3327  }
3328  KMP_CPU_ZERO(tempMask);
3329  setSize = 0;
3330 
3331  // valid follow sets are ',' and EOL
3332  SKIP_WS(scan);
3333  if (*scan == '\0') {
3334  break;
3335  }
3336  if (*scan == ',') {
3337  scan++; // skip ','
3338  continue;
3339  }
3340 
3341  KMP_ASSERT2(0, "bad explicit places list");
3342  }
3343 
3344  *out_numMasks = nextNewMask;
3345  if (nextNewMask == 0) {
3346  *out_masks = NULL;
3347  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3348  return;
3349  }
3350  KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
3351  KMP_CPU_FREE(tempMask);
3352  KMP_CPU_FREE(previousMask);
3353  for (i = 0; i < nextNewMask; i++) {
3354  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
3355  kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
3356  KMP_CPU_COPY(dest, src);
3357  }
3358  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3359 }
3360 
3361 #endif /* OMP_40_ENABLED */
3362 
3363 #undef ADD_MASK
3364 #undef ADD_MASK_OSID
3365 
3366 #if KMP_USE_HWLOC
3367 static int __kmp_hwloc_skip_PUs_obj(hwloc_topology_t t, hwloc_obj_t o) {
3368  // skip PUs descendants of the object o
3369  int skipped = 0;
3370  hwloc_obj_t hT = NULL;
3371  int N = __kmp_hwloc_count_children_by_type(t, o, HWLOC_OBJ_PU, &hT);
3372  for (int i = 0; i < N; ++i) {
3373  KMP_DEBUG_ASSERT(hT);
3374  unsigned idx = hT->os_index;
3375  if (KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
3376  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
3377  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
3378  ++skipped;
3379  }
3380  hT = hwloc_get_next_obj_by_type(t, HWLOC_OBJ_PU, hT);
3381  }
3382  return skipped; // count number of skipped units
3383 }
3384 
3385 static int __kmp_hwloc_obj_has_PUs(hwloc_topology_t t, hwloc_obj_t o) {
3386  // check if obj has PUs present in fullMask
3387  hwloc_obj_t hT = NULL;
3388  int N = __kmp_hwloc_count_children_by_type(t, o, HWLOC_OBJ_PU, &hT);
3389  for (int i = 0; i < N; ++i) {
3390  KMP_DEBUG_ASSERT(hT);
3391  unsigned idx = hT->os_index;
3392  if (KMP_CPU_ISSET(idx, __kmp_affin_fullMask))
3393  return 1; // found PU
3394  hT = hwloc_get_next_obj_by_type(t, HWLOC_OBJ_PU, hT);
3395  }
3396  return 0; // no PUs found
3397 }
3398 #endif // KMP_USE_HWLOC
3399 
3400 static void __kmp_apply_thread_places(AddrUnsPair **pAddr, int depth) {
3401  AddrUnsPair *newAddr;
3402  if (__kmp_hws_requested == 0)
3403  goto _exit; // no topology limiting actions requested, exit
3404 #if KMP_USE_HWLOC
3405  if (__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC) {
3406  // Number of subobjects calculated dynamically, this works fine for
3407  // any non-uniform topology.
3408  // L2 cache objects are determined by depth, other objects - by type.
3409  hwloc_topology_t tp = __kmp_hwloc_topology;
3410  int nS = 0, nN = 0, nL = 0, nC = 0,
3411  nT = 0; // logical index including skipped
3412  int nCr = 0, nTr = 0; // number of requested units
3413  int nPkg = 0, nCo = 0, n_new = 0, n_old = 0, nCpP = 0, nTpC = 0; // counters
3414  hwloc_obj_t hT, hC, hL, hN, hS; // hwloc objects (pointers to)
3415  int L2depth, idx;
3416 
3417  // check support of extensions ----------------------------------
3418  int numa_support = 0, tile_support = 0;
3419  if (__kmp_pu_os_idx)
3420  hT = hwloc_get_pu_obj_by_os_index(tp,
3421  __kmp_pu_os_idx[__kmp_avail_proc - 1]);
3422  else
3423  hT = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PU, __kmp_avail_proc - 1);
3424  if (hT == NULL) { // something's gone wrong
3425  KMP_WARNING(AffHWSubsetUnsupported);
3426  goto _exit;
3427  }
3428  // check NUMA node
3429  hN = hwloc_get_ancestor_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hT);
3430  hS = hwloc_get_ancestor_obj_by_type(tp, HWLOC_OBJ_PACKAGE, hT);
3431  if (hN != NULL && hN->depth > hS->depth) {
3432  numa_support = 1; // 1 in case socket includes node(s)
3433  } else if (__kmp_hws_node.num > 0) {
3434  // don't support sockets inside NUMA node (no such HW found for testing)
3435  KMP_WARNING(AffHWSubsetUnsupported);
3436  goto _exit;
3437  }
3438  // check L2 cahce, get object by depth because of multiple caches
3439  L2depth = hwloc_get_cache_type_depth(tp, 2, HWLOC_OBJ_CACHE_UNIFIED);
3440  hL = hwloc_get_ancestor_obj_by_depth(tp, L2depth, hT);
3441  if (hL != NULL &&
3442  __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE, &hC) > 1) {
3443  tile_support = 1; // no sense to count L2 if it includes single core
3444  } else if (__kmp_hws_tile.num > 0) {
3445  if (__kmp_hws_core.num == 0) {
3446  __kmp_hws_core = __kmp_hws_tile; // replace L2 with core
3447  __kmp_hws_tile.num = 0;
3448  } else {
3449  // L2 and core are both requested, but represent same object
3450  KMP_WARNING(AffHWSubsetInvalid);
3451  goto _exit;
3452  }
3453  }
3454  // end of check of extensions -----------------------------------
3455 
3456  // fill in unset items, validate settings -----------------------
3457  if (__kmp_hws_socket.num == 0)
3458  __kmp_hws_socket.num = nPackages; // use all available sockets
3459  if (__kmp_hws_socket.offset >= nPackages) {
3460  KMP_WARNING(AffHWSubsetManySockets);
3461  goto _exit;
3462  }
3463  if (numa_support) {
3464  hN = NULL;
3465  int NN = __kmp_hwloc_count_children_by_type(tp, hS, HWLOC_OBJ_NUMANODE,
3466  &hN); // num nodes in socket
3467  if (__kmp_hws_node.num == 0)
3468  __kmp_hws_node.num = NN; // use all available nodes
3469  if (__kmp_hws_node.offset >= NN) {
3470  KMP_WARNING(AffHWSubsetManyNodes);
3471  goto _exit;
3472  }
3473  if (tile_support) {
3474  // get num tiles in node
3475  int NL = __kmp_hwloc_count_children_by_depth(tp, hN, L2depth, &hL);
3476  if (__kmp_hws_tile.num == 0) {
3477  __kmp_hws_tile.num = NL + 1;
3478  } // use all available tiles, some node may have more tiles, thus +1
3479  if (__kmp_hws_tile.offset >= NL) {
3480  KMP_WARNING(AffHWSubsetManyTiles);
3481  goto _exit;
3482  }
3483  int NC = __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE,
3484  &hC); // num cores in tile
3485  if (__kmp_hws_core.num == 0)
3486  __kmp_hws_core.num = NC; // use all available cores
3487  if (__kmp_hws_core.offset >= NC) {
3488  KMP_WARNING(AffHWSubsetManyCores);
3489  goto _exit;
3490  }
3491  } else { // tile_support
3492  int NC = __kmp_hwloc_count_children_by_type(tp, hN, HWLOC_OBJ_CORE,
3493  &hC); // num cores in node
3494  if (__kmp_hws_core.num == 0)
3495  __kmp_hws_core.num = NC; // use all available cores
3496  if (__kmp_hws_core.offset >= NC) {
3497  KMP_WARNING(AffHWSubsetManyCores);
3498  goto _exit;
3499  }
3500  } // tile_support
3501  } else { // numa_support
3502  if (tile_support) {
3503  // get num tiles in socket
3504  int NL = __kmp_hwloc_count_children_by_depth(tp, hS, L2depth, &hL);
3505  if (__kmp_hws_tile.num == 0)
3506  __kmp_hws_tile.num = NL; // use all available tiles
3507  if (__kmp_hws_tile.offset >= NL) {
3508  KMP_WARNING(AffHWSubsetManyTiles);
3509  goto _exit;
3510  }
3511  int NC = __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE,
3512  &hC); // num cores in tile
3513  if (__kmp_hws_core.num == 0)
3514  __kmp_hws_core.num = NC; // use all available cores
3515  if (__kmp_hws_core.offset >= NC) {
3516  KMP_WARNING(AffHWSubsetManyCores);
3517  goto _exit;
3518  }
3519  } else { // tile_support
3520  int NC = __kmp_hwloc_count_children_by_type(tp, hS, HWLOC_OBJ_CORE,
3521  &hC); // num cores in socket
3522  if (__kmp_hws_core.num == 0)
3523  __kmp_hws_core.num = NC; // use all available cores
3524  if (__kmp_hws_core.offset >= NC) {
3525  KMP_WARNING(AffHWSubsetManyCores);
3526  goto _exit;
3527  }
3528  } // tile_support
3529  }
3530  if (__kmp_hws_proc.num == 0)
3531  __kmp_hws_proc.num = __kmp_nThreadsPerCore; // use all available procs
3532  if (__kmp_hws_proc.offset >= __kmp_nThreadsPerCore) {
3533  KMP_WARNING(AffHWSubsetManyProcs);
3534  goto _exit;
3535  }
3536  // end of validation --------------------------------------------
3537 
3538  if (pAddr) // pAddr is NULL in case of affinity_none
3539  newAddr = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) *
3540  __kmp_avail_proc); // max size
3541  // main loop to form HW subset ----------------------------------
3542  hS = NULL;
3543  int NP = hwloc_get_nbobjs_by_type(tp, HWLOC_OBJ_PACKAGE);
3544  for (int s = 0; s < NP; ++s) {
3545  // Check Socket -----------------------------------------------
3546  hS = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PACKAGE, hS);
3547  if (!__kmp_hwloc_obj_has_PUs(tp, hS))
3548  continue; // skip socket if all PUs are out of fullMask
3549  ++nS; // only count objects those have PUs in affinity mask
3550  if (nS <= __kmp_hws_socket.offset ||
3551  nS > __kmp_hws_socket.num + __kmp_hws_socket.offset) {
3552  n_old += __kmp_hwloc_skip_PUs_obj(tp, hS); // skip socket
3553  continue; // move to next socket
3554  }
3555  nCr = 0; // count number of cores per socket
3556  // socket requested, go down the topology tree
3557  // check 4 cases: (+NUMA+Tile), (+NUMA-Tile), (-NUMA+Tile), (-NUMA-Tile)
3558  if (numa_support) {
3559  nN = 0;
3560  hN = NULL;
3561  // num nodes in current socket
3562  int NN =
3563  __kmp_hwloc_count_children_by_type(tp, hS, HWLOC_OBJ_NUMANODE, &hN);
3564  for (int n = 0; n < NN; ++n) {
3565  // Check NUMA Node ----------------------------------------
3566  if (!__kmp_hwloc_obj_has_PUs(tp, hN)) {
3567  hN = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hN);
3568  continue; // skip node if all PUs are out of fullMask
3569  }
3570  ++nN;
3571  if (nN <= __kmp_hws_node.offset ||
3572  nN > __kmp_hws_node.num + __kmp_hws_node.offset) {
3573  // skip node as not requested
3574  n_old += __kmp_hwloc_skip_PUs_obj(tp, hN); // skip node
3575  hN = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hN);
3576  continue; // move to next node
3577  }
3578  // node requested, go down the topology tree
3579  if (tile_support) {
3580  nL = 0;
3581  hL = NULL;
3582  int NL = __kmp_hwloc_count_children_by_depth(tp, hN, L2depth, &hL);
3583  for (int l = 0; l < NL; ++l) {
3584  // Check L2 (tile) ------------------------------------
3585  if (!__kmp_hwloc_obj_has_PUs(tp, hL)) {
3586  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3587  continue; // skip tile if all PUs are out of fullMask
3588  }
3589  ++nL;
3590  if (nL <= __kmp_hws_tile.offset ||
3591  nL > __kmp_hws_tile.num + __kmp_hws_tile.offset) {
3592  // skip tile as not requested
3593  n_old += __kmp_hwloc_skip_PUs_obj(tp, hL); // skip tile
3594  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3595  continue; // move to next tile
3596  }
3597  // tile requested, go down the topology tree
3598  nC = 0;
3599  hC = NULL;
3600  // num cores in current tile
3601  int NC = __kmp_hwloc_count_children_by_type(tp, hL,
3602  HWLOC_OBJ_CORE, &hC);
3603  for (int c = 0; c < NC; ++c) {
3604  // Check Core ---------------------------------------
3605  if (!__kmp_hwloc_obj_has_PUs(tp, hC)) {
3606  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3607  continue; // skip core if all PUs are out of fullMask
3608  }
3609  ++nC;
3610  if (nC <= __kmp_hws_core.offset ||
3611  nC > __kmp_hws_core.num + __kmp_hws_core.offset) {
3612  // skip node as not requested
3613  n_old += __kmp_hwloc_skip_PUs_obj(tp, hC); // skip core
3614  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3615  continue; // move to next node
3616  }
3617  // core requested, go down to PUs
3618  nT = 0;
3619  nTr = 0;
3620  hT = NULL;
3621  // num procs in current core
3622  int NT = __kmp_hwloc_count_children_by_type(tp, hC,
3623  HWLOC_OBJ_PU, &hT);
3624  for (int t = 0; t < NT; ++t) {
3625  // Check PU ---------------------------------------
3626  idx = hT->os_index;
3627  if (!KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
3628  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3629  continue; // skip PU if not in fullMask
3630  }
3631  ++nT;
3632  if (nT <= __kmp_hws_proc.offset ||
3633  nT > __kmp_hws_proc.num + __kmp_hws_proc.offset) {
3634  // skip PU
3635  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
3636  ++n_old;
3637  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
3638  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3639  continue; // move to next node
3640  }
3641  ++nTr;
3642  if (pAddr) // collect requested thread's data
3643  newAddr[n_new] = (*pAddr)[n_old];
3644  ++n_new;
3645  ++n_old;
3646  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3647  } // threads loop
3648  if (nTr > 0) {
3649  ++nCr; // num cores per socket
3650  ++nCo; // total num cores
3651  if (nTr > nTpC)
3652  nTpC = nTr; // calc max threads per core
3653  }
3654  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3655  } // cores loop
3656  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3657  } // tiles loop
3658  } else { // tile_support
3659  // no tiles, check cores
3660  nC = 0;
3661  hC = NULL;
3662  // num cores in current node
3663  int NC =
3664  __kmp_hwloc_count_children_by_type(tp, hN, HWLOC_OBJ_CORE, &hC);
3665  for (int c = 0; c < NC; ++c) {
3666  // Check Core ---------------------------------------
3667  if (!__kmp_hwloc_obj_has_PUs(tp, hC)) {
3668  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3669  continue; // skip core if all PUs are out of fullMask
3670  }
3671  ++nC;
3672  if (nC <= __kmp_hws_core.offset ||
3673  nC > __kmp_hws_core.num + __kmp_hws_core.offset) {
3674  // skip node as not requested
3675  n_old += __kmp_hwloc_skip_PUs_obj(tp, hC); // skip core
3676  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3677  continue; // move to next node
3678  }
3679  // core requested, go down to PUs
3680  nT = 0;
3681  nTr = 0;
3682  hT = NULL;
3683  int NT =
3684  __kmp_hwloc_count_children_by_type(tp, hC, HWLOC_OBJ_PU, &hT);
3685  for (int t = 0; t < NT; ++t) {
3686  // Check PU ---------------------------------------
3687  idx = hT->os_index;
3688  if (!KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
3689  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3690  continue; // skip PU if not in fullMask
3691  }
3692  ++nT;
3693  if (nT <= __kmp_hws_proc.offset ||
3694  nT > __kmp_hws_proc.num + __kmp_hws_proc.offset) {
3695  // skip PU
3696  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
3697  ++n_old;
3698  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
3699  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3700  continue; // move to next node
3701  }
3702  ++nTr;
3703  if (pAddr) // collect requested thread's data
3704  newAddr[n_new] = (*pAddr)[n_old];
3705  ++n_new;
3706  ++n_old;
3707  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3708  } // threads loop
3709  if (nTr > 0) {
3710  ++nCr; // num cores per socket
3711  ++nCo; // total num cores
3712  if (nTr > nTpC)
3713  nTpC = nTr; // calc max threads per core
3714  }
3715  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3716  } // cores loop
3717  } // tiles support
3718  hN = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hN);
3719  } // nodes loop
3720  } else { // numa_support
3721  // no NUMA support
3722  if (tile_support) {
3723  nL = 0;
3724  hL = NULL;
3725  // num tiles in current socket
3726  int NL = __kmp_hwloc_count_children_by_depth(tp, hS, L2depth, &hL);
3727  for (int l = 0; l < NL; ++l) {
3728  // Check L2 (tile) ------------------------------------
3729  if (!__kmp_hwloc_obj_has_PUs(tp, hL)) {
3730  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3731  continue; // skip tile if all PUs are out of fullMask
3732  }
3733  ++nL;
3734  if (nL <= __kmp_hws_tile.offset ||
3735  nL > __kmp_hws_tile.num + __kmp_hws_tile.offset) {
3736  // skip tile as not requested
3737  n_old += __kmp_hwloc_skip_PUs_obj(tp, hL); // skip tile
3738  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3739  continue; // move to next tile
3740  }
3741  // tile requested, go down the topology tree
3742  nC = 0;
3743  hC = NULL;
3744  // num cores per tile
3745  int NC =
3746  __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE, &hC);
3747  for (int c = 0; c < NC; ++c) {
3748  // Check Core ---------------------------------------
3749  if (!__kmp_hwloc_obj_has_PUs(tp, hC)) {
3750  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3751  continue; // skip core if all PUs are out of fullMask
3752  }
3753  ++nC;
3754  if (nC <= __kmp_hws_core.offset ||
3755  nC > __kmp_hws_core.num + __kmp_hws_core.offset) {
3756  // skip node as not requested
3757  n_old += __kmp_hwloc_skip_PUs_obj(tp, hC); // skip core
3758  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3759  continue; // move to next node
3760  }
3761  // core requested, go down to PUs
3762  nT = 0;
3763  nTr = 0;
3764  hT = NULL;
3765  // num procs per core
3766  int NT =
3767  __kmp_hwloc_count_children_by_type(tp, hC, HWLOC_OBJ_PU, &hT);
3768  for (int t = 0; t < NT; ++t) {
3769  // Check PU ---------------------------------------
3770  idx = hT->os_index;
3771  if (!KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
3772  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3773  continue; // skip PU if not in fullMask
3774  }
3775  ++nT;
3776  if (nT <= __kmp_hws_proc.offset ||
3777  nT > __kmp_hws_proc.num + __kmp_hws_proc.offset) {
3778  // skip PU
3779  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
3780  ++n_old;
3781  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
3782  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3783  continue; // move to next node
3784  }
3785  ++nTr;
3786  if (pAddr) // collect requested thread's data
3787  newAddr[n_new] = (*pAddr)[n_old];
3788  ++n_new;
3789  ++n_old;
3790  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3791  } // threads loop
3792  if (nTr > 0) {
3793  ++nCr; // num cores per socket
3794  ++nCo; // total num cores
3795  if (nTr > nTpC)
3796  nTpC = nTr; // calc max threads per core
3797  }
3798  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3799  } // cores loop
3800  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3801  } // tiles loop
3802  } else { // tile_support
3803  // no tiles, check cores
3804  nC = 0;
3805  hC = NULL;
3806  // num cores in socket
3807  int NC =
3808  __kmp_hwloc_count_children_by_type(tp, hS, HWLOC_OBJ_CORE, &hC);
3809  for (int c = 0; c < NC; ++c) {
3810  // Check Core -------------------------------------------
3811  if (!__kmp_hwloc_obj_has_PUs(tp, hC)) {
3812  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3813  continue; // skip core if all PUs are out of fullMask
3814  }
3815  ++nC;
3816  if (nC <= __kmp_hws_core.offset ||
3817  nC > __kmp_hws_core.num + __kmp_hws_core.offset) {
3818  // skip node as not requested
3819  n_old += __kmp_hwloc_skip_PUs_obj(tp, hC); // skip core
3820  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3821  continue; // move to next node
3822  }
3823  // core requested, go down to PUs
3824  nT = 0;
3825  nTr = 0;
3826  hT = NULL;
3827  // num procs per core
3828  int NT =
3829  __kmp_hwloc_count_children_by_type(tp, hC, HWLOC_OBJ_PU, &hT);
3830  for (int t = 0; t < NT; ++t) {
3831  // Check PU ---------------------------------------
3832  idx = hT->os_index;
3833  if (!KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
3834  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3835  continue; // skip PU if not in fullMask
3836  }
3837  ++nT;
3838  if (nT <= __kmp_hws_proc.offset ||
3839  nT > __kmp_hws_proc.num + __kmp_hws_proc.offset) {
3840  // skip PU
3841  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
3842  ++n_old;
3843  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
3844  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3845  continue; // move to next node
3846  }
3847  ++nTr;
3848  if (pAddr) // collect requested thread's data
3849  newAddr[n_new] = (*pAddr)[n_old];
3850  ++n_new;
3851  ++n_old;
3852  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3853  } // threads loop
3854  if (nTr > 0) {
3855  ++nCr; // num cores per socket
3856  ++nCo; // total num cores
3857  if (nTr > nTpC)
3858  nTpC = nTr; // calc max threads per core
3859  }
3860  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3861  } // cores loop
3862  } // tiles support
3863  } // numa_support
3864  if (nCr > 0) { // found cores?
3865  ++nPkg; // num sockets
3866  if (nCr > nCpP)
3867  nCpP = nCr; // calc max cores per socket
3868  }
3869  } // sockets loop
3870 
3871  // check the subset is valid
3872  KMP_DEBUG_ASSERT(n_old == __kmp_avail_proc);
3873  KMP_DEBUG_ASSERT(nPkg > 0);
3874  KMP_DEBUG_ASSERT(nCpP > 0);
3875  KMP_DEBUG_ASSERT(nTpC > 0);
3876  KMP_DEBUG_ASSERT(nCo > 0);
3877  KMP_DEBUG_ASSERT(nPkg <= nPackages);
3878  KMP_DEBUG_ASSERT(nCpP <= nCoresPerPkg);
3879  KMP_DEBUG_ASSERT(nTpC <= __kmp_nThreadsPerCore);
3880  KMP_DEBUG_ASSERT(nCo <= __kmp_ncores);
3881 
3882  nPackages = nPkg; // correct num sockets
3883  nCoresPerPkg = nCpP; // correct num cores per socket
3884  __kmp_nThreadsPerCore = nTpC; // correct num threads per core
3885  __kmp_avail_proc = n_new; // correct num procs
3886  __kmp_ncores = nCo; // correct num cores
3887  // hwloc topology method end
3888  } else
3889 #endif // KMP_USE_HWLOC
3890  {
3891  int n_old = 0, n_new = 0, proc_num = 0;
3892  if (__kmp_hws_node.num > 0 || __kmp_hws_tile.num > 0) {
3893  KMP_WARNING(AffHWSubsetNoHWLOC);
3894  goto _exit;
3895  }
3896  if (__kmp_hws_socket.num == 0)
3897  __kmp_hws_socket.num = nPackages; // use all available sockets
3898  if (__kmp_hws_core.num == 0)
3899  __kmp_hws_core.num = nCoresPerPkg; // use all available cores
3900  if (__kmp_hws_proc.num == 0 || __kmp_hws_proc.num > __kmp_nThreadsPerCore)
3901  __kmp_hws_proc.num = __kmp_nThreadsPerCore; // use all HW contexts
3902  if (!__kmp_affinity_uniform_topology()) {
3903  KMP_WARNING(AffHWSubsetNonUniform);
3904  goto _exit; // don't support non-uniform topology
3905  }
3906  if (depth > 3) {
3907  KMP_WARNING(AffHWSubsetNonThreeLevel);
3908  goto _exit; // don't support not-3-level topology
3909  }
3910  if (__kmp_hws_socket.offset + __kmp_hws_socket.num > nPackages) {
3911  KMP_WARNING(AffHWSubsetManySockets);
3912  goto _exit;
3913  }
3914  if (__kmp_hws_core.offset + __kmp_hws_core.num > nCoresPerPkg) {
3915  KMP_WARNING(AffHWSubsetManyCores);
3916  goto _exit;
3917  }
3918  // Form the requested subset
3919  if (pAddr) // pAddr is NULL in case of affinity_none
3920  newAddr = (AddrUnsPair *)__kmp_allocate(
3921  sizeof(AddrUnsPair) * __kmp_hws_socket.num * __kmp_hws_core.num *
3922  __kmp_hws_proc.num);
3923  for (int i = 0; i < nPackages; ++i) {
3924  if (i < __kmp_hws_socket.offset ||
3925  i >= __kmp_hws_socket.offset + __kmp_hws_socket.num) {
3926  // skip not-requested socket
3927  n_old += nCoresPerPkg * __kmp_nThreadsPerCore;
3928  if (__kmp_pu_os_idx != NULL) {
3929  // walk through skipped socket
3930  for (int j = 0; j < nCoresPerPkg; ++j) {
3931  for (int k = 0; k < __kmp_nThreadsPerCore; ++k) {
3932  KMP_CPU_CLR(__kmp_pu_os_idx[proc_num], __kmp_affin_fullMask);
3933  ++proc_num;
3934  }
3935  }
3936  }
3937  } else {
3938  // walk through requested socket
3939  for (int j = 0; j < nCoresPerPkg; ++j) {
3940  if (j < __kmp_hws_core.offset ||
3941  j >= __kmp_hws_core.offset +
3942  __kmp_hws_core.num) { // skip not-requested core
3943  n_old += __kmp_nThreadsPerCore;
3944  if (__kmp_pu_os_idx != NULL) {
3945  for (int k = 0; k < __kmp_nThreadsPerCore; ++k) {
3946  KMP_CPU_CLR(__kmp_pu_os_idx[proc_num], __kmp_affin_fullMask);
3947  ++proc_num;
3948  }
3949  }
3950  } else {
3951  // walk through requested core
3952  for (int k = 0; k < __kmp_nThreadsPerCore; ++k) {
3953  if (k < __kmp_hws_proc.num) {
3954  if (pAddr) // collect requested thread's data
3955  newAddr[n_new] = (*pAddr)[n_old];
3956  n_new++;
3957  } else {
3958  if (__kmp_pu_os_idx != NULL)
3959  KMP_CPU_CLR(__kmp_pu_os_idx[proc_num], __kmp_affin_fullMask);
3960  }
3961  n_old++;
3962  ++proc_num;
3963  }
3964  }
3965  }
3966  }
3967  }
3968  KMP_DEBUG_ASSERT(n_old == nPackages * nCoresPerPkg * __kmp_nThreadsPerCore);
3969  KMP_DEBUG_ASSERT(n_new ==
3970  __kmp_hws_socket.num * __kmp_hws_core.num *
3971  __kmp_hws_proc.num);
3972  nPackages = __kmp_hws_socket.num; // correct nPackages
3973  nCoresPerPkg = __kmp_hws_core.num; // correct nCoresPerPkg
3974  __kmp_nThreadsPerCore = __kmp_hws_proc.num; // correct __kmp_nThreadsPerCore
3975  __kmp_avail_proc = n_new; // correct avail_proc
3976  __kmp_ncores = nPackages * __kmp_hws_core.num; // correct ncores
3977  } // non-hwloc topology method
3978  if (pAddr) {
3979  __kmp_free(*pAddr);
3980  *pAddr = newAddr; // replace old topology with new one
3981  }
3982  if (__kmp_affinity_verbose) {
3983  char m[KMP_AFFIN_MASK_PRINT_LEN];
3984  __kmp_affinity_print_mask(m, KMP_AFFIN_MASK_PRINT_LEN,
3985  __kmp_affin_fullMask);
3986  if (__kmp_affinity_respect_mask) {
3987  KMP_INFORM(InitOSProcSetRespect, "KMP_HW_SUBSET", m);
3988  } else {
3989  KMP_INFORM(InitOSProcSetNotRespect, "KMP_HW_SUBSET", m);
3990  }
3991  KMP_INFORM(AvailableOSProc, "KMP_HW_SUBSET", __kmp_avail_proc);
3992  kmp_str_buf_t buf;
3993  __kmp_str_buf_init(&buf);
3994  __kmp_str_buf_print(&buf, "%d", nPackages);
3995  KMP_INFORM(TopologyExtra, "KMP_HW_SUBSET", buf.str, nCoresPerPkg,
3996  __kmp_nThreadsPerCore, __kmp_ncores);
3997  __kmp_str_buf_free(&buf);
3998  }
3999 _exit:
4000  if (__kmp_pu_os_idx != NULL) {
4001  __kmp_free(__kmp_pu_os_idx);
4002  __kmp_pu_os_idx = NULL;
4003  }
4004 }
4005 
4006 // This function figures out the deepest level at which there is at least one
4007 // cluster/core with more than one processing unit bound to it.
4008 static int __kmp_affinity_find_core_level(const AddrUnsPair *address2os,
4009  int nprocs, int bottom_level) {
4010  int core_level = 0;
4011 
4012  for (int i = 0; i < nprocs; i++) {
4013  for (int j = bottom_level; j > 0; j--) {
4014  if (address2os[i].first.labels[j] > 0) {
4015  if (core_level < (j - 1)) {
4016  core_level = j - 1;
4017  }
4018  }
4019  }
4020  }
4021  return core_level;
4022 }
4023 
4024 // This function counts number of clusters/cores at given level.
4025 static int __kmp_affinity_compute_ncores(const AddrUnsPair *address2os,
4026  int nprocs, int bottom_level,
4027  int core_level) {
4028  int ncores = 0;
4029  int i, j;
4030 
4031  j = bottom_level;
4032  for (i = 0; i < nprocs; i++) {
4033  for (j = bottom_level; j > core_level; j--) {
4034  if ((i + 1) < nprocs) {
4035  if (address2os[i + 1].first.labels[j] > 0) {
4036  break;
4037  }
4038  }
4039  }
4040  if (j == core_level) {
4041  ncores++;
4042  }
4043  }
4044  if (j > core_level) {
4045  // In case of ( nprocs < __kmp_avail_proc ) we may end too deep and miss one
4046  // core. May occur when called from __kmp_affinity_find_core().
4047  ncores++;
4048  }
4049  return ncores;
4050 }
4051 
4052 // This function finds to which cluster/core given processing unit is bound.
4053 static int __kmp_affinity_find_core(const AddrUnsPair *address2os, int proc,
4054  int bottom_level, int core_level) {
4055  return __kmp_affinity_compute_ncores(address2os, proc + 1, bottom_level,
4056  core_level) -
4057  1;
4058 }
4059 
4060 // This function finds maximal number of processing units bound to a
4061 // cluster/core at given level.
4062 static int __kmp_affinity_max_proc_per_core(const AddrUnsPair *address2os,
4063  int nprocs, int bottom_level,
4064  int core_level) {
4065  int maxprocpercore = 0;
4066 
4067  if (core_level < bottom_level) {
4068  for (int i = 0; i < nprocs; i++) {
4069  int percore = address2os[i].first.labels[core_level + 1] + 1;
4070 
4071  if (percore > maxprocpercore) {
4072  maxprocpercore = percore;
4073  }
4074  }
4075  } else {
4076  maxprocpercore = 1;
4077  }
4078  return maxprocpercore;
4079 }
4080 
4081 static AddrUnsPair *address2os = NULL;
4082 static int *procarr = NULL;
4083 static int __kmp_aff_depth = 0;
4084 
4085 #if KMP_USE_HIER_SCHED
4086 #define KMP_EXIT_AFF_NONE \
4087  KMP_ASSERT(__kmp_affinity_type == affinity_none); \
4088  KMP_ASSERT(address2os == NULL); \
4089  __kmp_apply_thread_places(NULL, 0); \
4090  __kmp_create_affinity_none_places(); \
4091  __kmp_dispatch_set_hierarchy_values(); \
4092  return;
4093 #else
4094 #define KMP_EXIT_AFF_NONE \
4095  KMP_ASSERT(__kmp_affinity_type == affinity_none); \
4096  KMP_ASSERT(address2os == NULL); \
4097  __kmp_apply_thread_places(NULL, 0); \
4098  __kmp_create_affinity_none_places(); \
4099  return;
4100 #endif
4101 
4102 // Create a one element mask array (set of places) which only contains the
4103 // initial process's affinity mask
4104 static void __kmp_create_affinity_none_places() {
4105  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4106  KMP_ASSERT(__kmp_affinity_type == affinity_none);
4107  __kmp_affinity_num_masks = 1;
4108  KMP_CPU_ALLOC_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks);
4109  kmp_affin_mask_t *dest = KMP_CPU_INDEX(__kmp_affinity_masks, 0);
4110  KMP_CPU_COPY(dest, __kmp_affin_fullMask);
4111 }
4112 
4113 static int __kmp_affinity_cmp_Address_child_num(const void *a, const void *b) {
4114  const Address *aa = &(((const AddrUnsPair *)a)->first);
4115  const Address *bb = &(((const AddrUnsPair *)b)->first);
4116  unsigned depth = aa->depth;
4117  unsigned i;
4118  KMP_DEBUG_ASSERT(depth == bb->depth);
4119  KMP_DEBUG_ASSERT((unsigned)__kmp_affinity_compact <= depth);
4120  KMP_DEBUG_ASSERT(__kmp_affinity_compact >= 0);
4121  for (i = 0; i < (unsigned)__kmp_affinity_compact; i++) {
4122  int j = depth - i - 1;
4123  if (aa->childNums[j] < bb->childNums[j])
4124  return -1;
4125  if (aa->childNums[j] > bb->childNums[j])
4126  return 1;
4127  }
4128  for (; i < depth; i++) {
4129  int j = i - __kmp_affinity_compact;
4130  if (aa->childNums[j] < bb->childNums[j])
4131  return -1;
4132  if (aa->childNums[j] > bb->childNums[j])
4133  return 1;
4134  }
4135  return 0;
4136 }
4137 
4138 static void __kmp_aux_affinity_initialize(void) {
4139  if (__kmp_affinity_masks != NULL) {
4140  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4141  return;
4142  }
4143 
4144  // Create the "full" mask - this defines all of the processors that we
4145  // consider to be in the machine model. If respect is set, then it is the
4146  // initialization thread's affinity mask. Otherwise, it is all processors that
4147  // we know about on the machine.
4148  if (__kmp_affin_fullMask == NULL) {
4149  KMP_CPU_ALLOC(__kmp_affin_fullMask);
4150  }
4151  if (KMP_AFFINITY_CAPABLE()) {
4152  if (__kmp_affinity_respect_mask) {
4153  __kmp_get_system_affinity(__kmp_affin_fullMask, TRUE);
4154 
4155  // Count the number of available processors.
4156  unsigned i;
4157  __kmp_avail_proc = 0;
4158  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
4159  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
4160  continue;
4161  }
4162  __kmp_avail_proc++;
4163  }
4164  if (__kmp_avail_proc > __kmp_xproc) {
4165  if (__kmp_affinity_verbose ||
4166  (__kmp_affinity_warnings &&
4167  (__kmp_affinity_type != affinity_none))) {
4168  KMP_WARNING(ErrorInitializeAffinity);
4169  }
4170  __kmp_affinity_type = affinity_none;
4171  KMP_AFFINITY_DISABLE();
4172  return;
4173  }
4174  } else {
4175  __kmp_affinity_entire_machine_mask(__kmp_affin_fullMask);
4176  __kmp_avail_proc = __kmp_xproc;
4177  }
4178  }
4179 
4180  if (__kmp_affinity_gran == affinity_gran_tile &&
4181  // check if user's request is valid
4182  __kmp_affinity_dispatch->get_api_type() == KMPAffinity::NATIVE_OS) {
4183  KMP_WARNING(AffTilesNoHWLOC, "KMP_AFFINITY");
4184  __kmp_affinity_gran = affinity_gran_package;
4185  }
4186 
4187  int depth = -1;
4188  kmp_i18n_id_t msg_id = kmp_i18n_null;
4189 
4190  // For backward compatibility, setting KMP_CPUINFO_FILE =>
4191  // KMP_TOPOLOGY_METHOD=cpuinfo
4192  if ((__kmp_cpuinfo_file != NULL) &&
4193  (__kmp_affinity_top_method == affinity_top_method_all)) {
4194  __kmp_affinity_top_method = affinity_top_method_cpuinfo;
4195  }
4196 
4197  if (__kmp_affinity_top_method == affinity_top_method_all) {
4198  // In the default code path, errors are not fatal - we just try using
4199  // another method. We only emit a warning message if affinity is on, or the
4200  // verbose flag is set, an the nowarnings flag was not set.
4201  const char *file_name = NULL;
4202  int line = 0;
4203 #if KMP_USE_HWLOC
4204  if (depth < 0 &&
4205  __kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC) {
4206  if (__kmp_affinity_verbose) {
4207  KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
4208  }
4209  if (!__kmp_hwloc_error) {
4210  depth = __kmp_affinity_create_hwloc_map(&address2os, &msg_id);
4211  if (depth == 0) {
4212  KMP_EXIT_AFF_NONE;
4213  } else if (depth < 0 && __kmp_affinity_verbose) {
4214  KMP_INFORM(AffIgnoringHwloc, "KMP_AFFINITY");
4215  }
4216  } else if (__kmp_affinity_verbose) {
4217  KMP_INFORM(AffIgnoringHwloc, "KMP_AFFINITY");
4218  }
4219  }
4220 #endif
4221 
4222 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4223 
4224  if (depth < 0) {
4225  if (__kmp_affinity_verbose) {
4226  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(Decodingx2APIC));
4227  }
4228 
4229  file_name = NULL;
4230  depth = __kmp_affinity_create_x2apicid_map(&address2os, &msg_id);
4231  if (depth == 0) {
4232  KMP_EXIT_AFF_NONE;
4233  }
4234 
4235  if (depth < 0) {
4236  if (__kmp_affinity_verbose) {
4237  if (msg_id != kmp_i18n_null) {
4238  KMP_INFORM(AffInfoStrStr, "KMP_AFFINITY",
4239  __kmp_i18n_catgets(msg_id),
4240  KMP_I18N_STR(DecodingLegacyAPIC));
4241  } else {
4242  KMP_INFORM(AffInfoStr, "KMP_AFFINITY",
4243  KMP_I18N_STR(DecodingLegacyAPIC));
4244  }
4245  }
4246 
4247  file_name = NULL;
4248  depth = __kmp_affinity_create_apicid_map(&address2os, &msg_id);
4249  if (depth == 0) {
4250  KMP_EXIT_AFF_NONE;
4251  }
4252  }
4253  }
4254 
4255 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4256 
4257 #if KMP_OS_LINUX
4258 
4259  if (depth < 0) {
4260  if (__kmp_affinity_verbose) {
4261  if (msg_id != kmp_i18n_null) {
4262  KMP_INFORM(AffStrParseFilename, "KMP_AFFINITY",
4263  __kmp_i18n_catgets(msg_id), "/proc/cpuinfo");
4264  } else {
4265  KMP_INFORM(AffParseFilename, "KMP_AFFINITY", "/proc/cpuinfo");
4266  }
4267  }
4268 
4269  FILE *f = fopen("/proc/cpuinfo", "r");
4270  if (f == NULL) {
4271  msg_id = kmp_i18n_str_CantOpenCpuinfo;
4272  } else {
4273  file_name = "/proc/cpuinfo";
4274  depth =
4275  __kmp_affinity_create_cpuinfo_map(&address2os, &line, &msg_id, f);
4276  fclose(f);
4277  if (depth == 0) {
4278  KMP_EXIT_AFF_NONE;
4279  }
4280  }
4281  }
4282 
4283 #endif /* KMP_OS_LINUX */
4284 
4285 #if KMP_GROUP_AFFINITY
4286 
4287  if ((depth < 0) && (__kmp_num_proc_groups > 1)) {
4288  if (__kmp_affinity_verbose) {
4289  KMP_INFORM(AffWindowsProcGroupMap, "KMP_AFFINITY");
4290  }
4291 
4292  depth = __kmp_affinity_create_proc_group_map(&address2os, &msg_id);
4293  KMP_ASSERT(depth != 0);
4294  }
4295 
4296 #endif /* KMP_GROUP_AFFINITY */
4297 
4298  if (depth < 0) {
4299  if (__kmp_affinity_verbose && (msg_id != kmp_i18n_null)) {
4300  if (file_name == NULL) {
4301  KMP_INFORM(UsingFlatOS, __kmp_i18n_catgets(msg_id));
4302  } else if (line == 0) {
4303  KMP_INFORM(UsingFlatOSFile, file_name, __kmp_i18n_catgets(msg_id));
4304  } else {
4305  KMP_INFORM(UsingFlatOSFileLine, file_name, line,
4306  __kmp_i18n_catgets(msg_id));
4307  }
4308  }
4309  // FIXME - print msg if msg_id = kmp_i18n_null ???
4310 
4311  file_name = "";
4312  depth = __kmp_affinity_create_flat_map(&address2os, &msg_id);
4313  if (depth == 0) {
4314  KMP_EXIT_AFF_NONE;
4315  }
4316  KMP_ASSERT(depth > 0);
4317  KMP_ASSERT(address2os != NULL);
4318  }
4319  }
4320 
4321 #if KMP_USE_HWLOC
4322  else if (__kmp_affinity_top_method == affinity_top_method_hwloc) {
4323  KMP_ASSERT(__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC);
4324  if (__kmp_affinity_verbose) {
4325  KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
4326  }
4327  depth = __kmp_affinity_create_hwloc_map(&address2os, &msg_id);
4328  if (depth == 0) {
4329  KMP_EXIT_AFF_NONE;
4330  }
4331  }
4332 #endif // KMP_USE_HWLOC
4333 
4334 // If the user has specified that a paricular topology discovery method is to be
4335 // used, then we abort if that method fails. The exception is group affinity,
4336 // which might have been implicitly set.
4337 
4338 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4339 
4340  else if (__kmp_affinity_top_method == affinity_top_method_x2apicid) {
4341  if (__kmp_affinity_verbose) {
4342  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(Decodingx2APIC));
4343  }
4344 
4345  depth = __kmp_affinity_create_x2apicid_map(&address2os, &msg_id);
4346  if (depth == 0) {
4347  KMP_EXIT_AFF_NONE;
4348  }
4349  if (depth < 0) {
4350  KMP_ASSERT(msg_id != kmp_i18n_null);
4351  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4352  }
4353  } else if (__kmp_affinity_top_method == affinity_top_method_apicid) {
4354  if (__kmp_affinity_verbose) {
4355  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(DecodingLegacyAPIC));
4356  }
4357 
4358  depth = __kmp_affinity_create_apicid_map(&address2os, &msg_id);
4359  if (depth == 0) {
4360  KMP_EXIT_AFF_NONE;
4361  }
4362  if (depth < 0) {
4363  KMP_ASSERT(msg_id != kmp_i18n_null);
4364  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4365  }
4366  }
4367 
4368 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4369 
4370  else if (__kmp_affinity_top_method == affinity_top_method_cpuinfo) {
4371  const char *filename;
4372  if (__kmp_cpuinfo_file != NULL) {
4373  filename = __kmp_cpuinfo_file;
4374  } else {
4375  filename = "/proc/cpuinfo";
4376  }
4377 
4378  if (__kmp_affinity_verbose) {
4379  KMP_INFORM(AffParseFilename, "KMP_AFFINITY", filename);
4380  }
4381 
4382  FILE *f = fopen(filename, "r");
4383  if (f == NULL) {
4384  int code = errno;
4385  if (__kmp_cpuinfo_file != NULL) {
4386  __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4387  KMP_HNT(NameComesFrom_CPUINFO_FILE), __kmp_msg_null);
4388  } else {
4389  __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4390  __kmp_msg_null);
4391  }
4392  }
4393  int line = 0;
4394  depth = __kmp_affinity_create_cpuinfo_map(&address2os, &line, &msg_id, f);
4395  fclose(f);
4396  if (depth < 0) {
4397  KMP_ASSERT(msg_id != kmp_i18n_null);
4398  if (line > 0) {
4399  KMP_FATAL(FileLineMsgExiting, filename, line,
4400  __kmp_i18n_catgets(msg_id));
4401  } else {
4402  KMP_FATAL(FileMsgExiting, filename, __kmp_i18n_catgets(msg_id));
4403  }
4404  }
4405  if (__kmp_affinity_type == affinity_none) {
4406  KMP_ASSERT(depth == 0);
4407  KMP_EXIT_AFF_NONE;
4408  }
4409  }
4410 
4411 #if KMP_GROUP_AFFINITY
4412 
4413  else if (__kmp_affinity_top_method == affinity_top_method_group) {
4414  if (__kmp_affinity_verbose) {
4415  KMP_INFORM(AffWindowsProcGroupMap, "KMP_AFFINITY");
4416  }
4417 
4418  depth = __kmp_affinity_create_proc_group_map(&address2os, &msg_id);
4419  KMP_ASSERT(depth != 0);
4420  if (depth < 0) {
4421  KMP_ASSERT(msg_id != kmp_i18n_null);
4422  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4423  }
4424  }
4425 
4426 #endif /* KMP_GROUP_AFFINITY */
4427 
4428  else if (__kmp_affinity_top_method == affinity_top_method_flat) {
4429  if (__kmp_affinity_verbose) {
4430  KMP_INFORM(AffUsingFlatOS, "KMP_AFFINITY");
4431  }
4432 
4433  depth = __kmp_affinity_create_flat_map(&address2os, &msg_id);
4434  if (depth == 0) {
4435  KMP_EXIT_AFF_NONE;
4436  }
4437  // should not fail
4438  KMP_ASSERT(depth > 0);
4439  KMP_ASSERT(address2os != NULL);
4440  }
4441 
4442 #if KMP_USE_HIER_SCHED
4443  __kmp_dispatch_set_hierarchy_values();
4444 #endif
4445 
4446  if (address2os == NULL) {
4447  if (KMP_AFFINITY_CAPABLE() &&
4448  (__kmp_affinity_verbose ||
4449  (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none)))) {
4450  KMP_WARNING(ErrorInitializeAffinity);
4451  }
4452  __kmp_affinity_type = affinity_none;
4453  __kmp_create_affinity_none_places();
4454  KMP_AFFINITY_DISABLE();
4455  return;
4456  }
4457 
4458  if (__kmp_affinity_gran == affinity_gran_tile
4459 #if KMP_USE_HWLOC
4460  && __kmp_tile_depth == 0
4461 #endif
4462  ) {
4463  // tiles requested but not detected, warn user on this
4464  KMP_WARNING(AffTilesNoTiles, "KMP_AFFINITY");
4465  }
4466 
4467  __kmp_apply_thread_places(&address2os, depth);
4468 
4469  // Create the table of masks, indexed by thread Id.
4470  unsigned maxIndex;
4471  unsigned numUnique;
4472  kmp_affin_mask_t *osId2Mask =
4473  __kmp_create_masks(&maxIndex, &numUnique, address2os, __kmp_avail_proc);
4474  if (__kmp_affinity_gran_levels == 0) {
4475  KMP_DEBUG_ASSERT((int)numUnique == __kmp_avail_proc);
4476  }
4477 
4478  // Set the childNums vector in all Address objects. This must be done before
4479  // we can sort using __kmp_affinity_cmp_Address_child_num(), which takes into
4480  // account the setting of __kmp_affinity_compact.
4481  __kmp_affinity_assign_child_nums(address2os, __kmp_avail_proc);
4482 
4483  switch (__kmp_affinity_type) {
4484 
4485  case affinity_explicit:
4486  KMP_DEBUG_ASSERT(__kmp_affinity_proclist != NULL);
4487 #if OMP_40_ENABLED
4488  if (__kmp_nested_proc_bind.bind_types[0] == proc_bind_intel)
4489 #endif
4490  {
4491  __kmp_affinity_process_proclist(
4492  &__kmp_affinity_masks, &__kmp_affinity_num_masks,
4493  __kmp_affinity_proclist, osId2Mask, maxIndex);
4494  }
4495 #if OMP_40_ENABLED
4496  else {
4497  __kmp_affinity_process_placelist(
4498  &__kmp_affinity_masks, &__kmp_affinity_num_masks,
4499  __kmp_affinity_proclist, osId2Mask, maxIndex);
4500  }
4501 #endif
4502  if (__kmp_affinity_num_masks == 0) {
4503  if (__kmp_affinity_verbose ||
4504  (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) {
4505  KMP_WARNING(AffNoValidProcID);
4506  }
4507  __kmp_affinity_type = affinity_none;
4508  __kmp_create_affinity_none_places();
4509  return;
4510  }
4511  break;
4512 
4513  // The other affinity types rely on sorting the Addresses according to some
4514  // permutation of the machine topology tree. Set __kmp_affinity_compact and
4515  // __kmp_affinity_offset appropriately, then jump to a common code fragment
4516  // to do the sort and create the array of affinity masks.
4517 
4518  case affinity_logical:
4519  __kmp_affinity_compact = 0;
4520  if (__kmp_affinity_offset) {
4521  __kmp_affinity_offset =
4522  __kmp_nThreadsPerCore * __kmp_affinity_offset % __kmp_avail_proc;
4523  }
4524  goto sortAddresses;
4525 
4526  case affinity_physical:
4527  if (__kmp_nThreadsPerCore > 1) {
4528  __kmp_affinity_compact = 1;
4529  if (__kmp_affinity_compact >= depth) {
4530  __kmp_affinity_compact = 0;
4531  }
4532  } else {
4533  __kmp_affinity_compact = 0;
4534  }
4535  if (__kmp_affinity_offset) {
4536  __kmp_affinity_offset =
4537  __kmp_nThreadsPerCore * __kmp_affinity_offset % __kmp_avail_proc;
4538  }
4539  goto sortAddresses;
4540 
4541  case affinity_scatter:
4542  if (__kmp_affinity_compact >= depth) {
4543  __kmp_affinity_compact = 0;
4544  } else {
4545  __kmp_affinity_compact = depth - 1 - __kmp_affinity_compact;
4546  }
4547  goto sortAddresses;
4548 
4549  case affinity_compact:
4550  if (__kmp_affinity_compact >= depth) {
4551  __kmp_affinity_compact = depth - 1;
4552  }
4553  goto sortAddresses;
4554 
4555  case affinity_balanced:
4556  if (depth <= 1) {
4557  if (__kmp_affinity_verbose || __kmp_affinity_warnings) {
4558  KMP_WARNING(AffBalancedNotAvail, "KMP_AFFINITY");
4559  }
4560  __kmp_affinity_type = affinity_none;
4561  __kmp_create_affinity_none_places();
4562  return;
4563  } else if (!__kmp_affinity_uniform_topology()) {
4564  // Save the depth for further usage
4565  __kmp_aff_depth = depth;
4566 
4567  int core_level = __kmp_affinity_find_core_level(
4568  address2os, __kmp_avail_proc, depth - 1);
4569  int ncores = __kmp_affinity_compute_ncores(address2os, __kmp_avail_proc,
4570  depth - 1, core_level);
4571  int maxprocpercore = __kmp_affinity_max_proc_per_core(
4572  address2os, __kmp_avail_proc, depth - 1, core_level);
4573 
4574  int nproc = ncores * maxprocpercore;
4575  if ((nproc < 2) || (nproc < __kmp_avail_proc)) {
4576  if (__kmp_affinity_verbose || __kmp_affinity_warnings) {
4577  KMP_WARNING(AffBalancedNotAvail, "KMP_AFFINITY");
4578  }
4579  __kmp_affinity_type = affinity_none;
4580  return;
4581  }
4582 
4583  procarr = (int *)__kmp_allocate(sizeof(int) * nproc);
4584  for (int i = 0; i < nproc; i++) {
4585  procarr[i] = -1;
4586  }
4587 
4588  int lastcore = -1;
4589  int inlastcore = 0;
4590  for (int i = 0; i < __kmp_avail_proc; i++) {
4591  int proc = address2os[i].second;
4592  int core =
4593  __kmp_affinity_find_core(address2os, i, depth - 1, core_level);
4594 
4595  if (core == lastcore) {
4596  inlastcore++;
4597  } else {
4598  inlastcore = 0;
4599  }
4600  lastcore = core;
4601 
4602  procarr[core * maxprocpercore + inlastcore] = proc;
4603  }
4604  }
4605  if (__kmp_affinity_compact >= depth) {
4606  __kmp_affinity_compact = depth - 1;
4607  }
4608 
4609  sortAddresses:
4610  // Allocate the gtid->affinity mask table.
4611  if (__kmp_affinity_dups) {
4612  __kmp_affinity_num_masks = __kmp_avail_proc;
4613  } else {
4614  __kmp_affinity_num_masks = numUnique;
4615  }
4616 
4617 #if OMP_40_ENABLED
4618  if ((__kmp_nested_proc_bind.bind_types[0] != proc_bind_intel) &&
4619  (__kmp_affinity_num_places > 0) &&
4620  ((unsigned)__kmp_affinity_num_places < __kmp_affinity_num_masks)) {
4621  __kmp_affinity_num_masks = __kmp_affinity_num_places;
4622  }
4623 #endif
4624 
4625  KMP_CPU_ALLOC_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks);
4626 
4627  // Sort the address2os table according to the current setting of
4628  // __kmp_affinity_compact, then fill out __kmp_affinity_masks.
4629  qsort(address2os, __kmp_avail_proc, sizeof(*address2os),
4630  __kmp_affinity_cmp_Address_child_num);
4631  {
4632  int i;
4633  unsigned j;
4634  for (i = 0, j = 0; i < __kmp_avail_proc; i++) {
4635  if ((!__kmp_affinity_dups) && (!address2os[i].first.leader)) {
4636  continue;
4637  }
4638  unsigned osId = address2os[i].second;
4639  kmp_affin_mask_t *src = KMP_CPU_INDEX(osId2Mask, osId);
4640  kmp_affin_mask_t *dest = KMP_CPU_INDEX(__kmp_affinity_masks, j);
4641  KMP_ASSERT(KMP_CPU_ISSET(osId, src));
4642  KMP_CPU_COPY(dest, src);
4643  if (++j >= __kmp_affinity_num_masks) {
4644  break;
4645  }
4646  }
4647  KMP_DEBUG_ASSERT(j == __kmp_affinity_num_masks);
4648  }
4649  break;
4650 
4651  default:
4652  KMP_ASSERT2(0, "Unexpected affinity setting");
4653  }
4654 
4655  KMP_CPU_FREE_ARRAY(osId2Mask, maxIndex + 1);
4656  machine_hierarchy.init(address2os, __kmp_avail_proc);
4657 }
4658 #undef KMP_EXIT_AFF_NONE
4659 
4660 void __kmp_affinity_initialize(void) {
4661  // Much of the code above was written assumming that if a machine was not
4662  // affinity capable, then __kmp_affinity_type == affinity_none. We now
4663  // explicitly represent this as __kmp_affinity_type == affinity_disabled.
4664  // There are too many checks for __kmp_affinity_type == affinity_none
4665  // in this code. Instead of trying to change them all, check if
4666  // __kmp_affinity_type == affinity_disabled, and if so, slam it with
4667  // affinity_none, call the real initialization routine, then restore
4668  // __kmp_affinity_type to affinity_disabled.
4669  int disabled = (__kmp_affinity_type == affinity_disabled);
4670  if (!KMP_AFFINITY_CAPABLE()) {
4671  KMP_ASSERT(disabled);
4672  }
4673  if (disabled) {
4674  __kmp_affinity_type = affinity_none;
4675  }
4676  __kmp_aux_affinity_initialize();
4677  if (disabled) {
4678  __kmp_affinity_type = affinity_disabled;
4679  }
4680 }
4681 
4682 void __kmp_affinity_uninitialize(void) {
4683  if (__kmp_affinity_masks != NULL) {
4684  KMP_CPU_FREE_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks);
4685  __kmp_affinity_masks = NULL;
4686  }
4687  if (__kmp_affin_fullMask != NULL) {
4688  KMP_CPU_FREE(__kmp_affin_fullMask);
4689  __kmp_affin_fullMask = NULL;
4690  }
4691  __kmp_affinity_num_masks = 0;
4692  __kmp_affinity_type = affinity_default;
4693 #if OMP_40_ENABLED
4694  __kmp_affinity_num_places = 0;
4695 #endif
4696  if (__kmp_affinity_proclist != NULL) {
4697  __kmp_free(__kmp_affinity_proclist);
4698  __kmp_affinity_proclist = NULL;
4699  }
4700  if (address2os != NULL) {
4701  __kmp_free(address2os);
4702  address2os = NULL;
4703  }
4704  if (procarr != NULL) {
4705  __kmp_free(procarr);
4706  procarr = NULL;
4707  }
4708 #if KMP_USE_HWLOC
4709  if (__kmp_hwloc_topology != NULL) {
4710  hwloc_topology_destroy(__kmp_hwloc_topology);
4711  __kmp_hwloc_topology = NULL;
4712  }
4713 #endif
4714  KMPAffinity::destroy_api();
4715 }
4716 
4717 void __kmp_affinity_set_init_mask(int gtid, int isa_root) {
4718  if (!KMP_AFFINITY_CAPABLE()) {
4719  return;
4720  }
4721 
4722  kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
4723  if (th->th.th_affin_mask == NULL) {
4724  KMP_CPU_ALLOC(th->th.th_affin_mask);
4725  } else {
4726  KMP_CPU_ZERO(th->th.th_affin_mask);
4727  }
4728 
4729  // Copy the thread mask to the kmp_info_t strucuture. If
4730  // __kmp_affinity_type == affinity_none, copy the "full" mask, i.e. one that
4731  // has all of the OS proc ids set, or if __kmp_affinity_respect_mask is set,
4732  // then the full mask is the same as the mask of the initialization thread.
4733  kmp_affin_mask_t *mask;
4734  int i;
4735 
4736 #if OMP_40_ENABLED
4737  if (KMP_AFFINITY_NON_PROC_BIND)
4738 #endif
4739  {
4740  if ((__kmp_affinity_type == affinity_none) ||
4741  (__kmp_affinity_type == affinity_balanced)) {
4742 #if KMP_GROUP_AFFINITY
4743  if (__kmp_num_proc_groups > 1) {
4744  return;
4745  }
4746 #endif
4747  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4748  i = 0;
4749  mask = __kmp_affin_fullMask;
4750  } else {
4751  KMP_DEBUG_ASSERT(__kmp_affinity_num_masks > 0);
4752  i = (gtid + __kmp_affinity_offset) % __kmp_affinity_num_masks;
4753  mask = KMP_CPU_INDEX(__kmp_affinity_masks, i);
4754  }
4755  }
4756 #if OMP_40_ENABLED
4757  else {
4758  if ((!isa_root) ||
4759  (__kmp_nested_proc_bind.bind_types[0] == proc_bind_false)) {
4760 #if KMP_GROUP_AFFINITY
4761  if (__kmp_num_proc_groups > 1) {
4762  return;
4763  }
4764 #endif
4765  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4766  i = KMP_PLACE_ALL;
4767  mask = __kmp_affin_fullMask;
4768  } else {
4769  // int i = some hash function or just a counter that doesn't
4770  // always start at 0. Use gtid for now.
4771  KMP_DEBUG_ASSERT(__kmp_affinity_num_masks > 0);
4772  i = (gtid + __kmp_affinity_offset) % __kmp_affinity_num_masks;
4773  mask = KMP_CPU_INDEX(__kmp_affinity_masks, i);
4774  }
4775  }
4776 #endif
4777 
4778 #if OMP_40_ENABLED
4779  th->th.th_current_place = i;
4780  if (isa_root) {
4781  th->th.th_new_place = i;
4782  th->th.th_first_place = 0;
4783  th->th.th_last_place = __kmp_affinity_num_masks - 1;
4784  } else if (KMP_AFFINITY_NON_PROC_BIND) {
4785  // When using a Non-OMP_PROC_BIND affinity method,
4786  // set all threads' place-partition-var to the entire place list
4787  th->th.th_first_place = 0;
4788  th->th.th_last_place = __kmp_affinity_num_masks - 1;
4789  }
4790 
4791  if (i == KMP_PLACE_ALL) {
4792  KA_TRACE(100, ("__kmp_affinity_set_init_mask: binding T#%d to all places\n",
4793  gtid));
4794  } else {
4795  KA_TRACE(100, ("__kmp_affinity_set_init_mask: binding T#%d to place %d\n",
4796  gtid, i));
4797  }
4798 #else
4799  if (i == -1) {
4800  KA_TRACE(
4801  100,
4802  ("__kmp_affinity_set_init_mask: binding T#%d to __kmp_affin_fullMask\n",
4803  gtid));
4804  } else {
4805  KA_TRACE(100, ("__kmp_affinity_set_init_mask: binding T#%d to mask %d\n",
4806  gtid, i));
4807  }
4808 #endif /* OMP_40_ENABLED */
4809 
4810  KMP_CPU_COPY(th->th.th_affin_mask, mask);
4811 
4812  if (__kmp_affinity_verbose
4813  /* to avoid duplicate printing (will be correctly printed on barrier) */
4814  && (__kmp_affinity_type == affinity_none ||
4815  (i != KMP_PLACE_ALL && __kmp_affinity_type != affinity_balanced))) {
4816  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4817  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4818  th->th.th_affin_mask);
4819  KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY", (kmp_int32)getpid(),
4820  __kmp_gettid(), gtid, buf);
4821  }
4822 
4823 #if KMP_OS_WINDOWS
4824  // On Windows* OS, the process affinity mask might have changed. If the user
4825  // didn't request affinity and this call fails, just continue silently.
4826  // See CQ171393.
4827  if (__kmp_affinity_type == affinity_none) {
4828  __kmp_set_system_affinity(th->th.th_affin_mask, FALSE);
4829  } else
4830 #endif
4831  __kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
4832 }
4833 
4834 #if OMP_40_ENABLED
4835 
4836 void __kmp_affinity_set_place(int gtid) {
4837  if (!KMP_AFFINITY_CAPABLE()) {
4838  return;
4839  }
4840 
4841  kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
4842 
4843  KA_TRACE(100, ("__kmp_affinity_set_place: binding T#%d to place %d (current "
4844  "place = %d)\n",
4845  gtid, th->th.th_new_place, th->th.th_current_place));
4846 
4847  // Check that the new place is within this thread's partition.
4848  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
4849  KMP_ASSERT(th->th.th_new_place >= 0);
4850  KMP_ASSERT((unsigned)th->th.th_new_place <= __kmp_affinity_num_masks);
4851  if (th->th.th_first_place <= th->th.th_last_place) {
4852  KMP_ASSERT((th->th.th_new_place >= th->th.th_first_place) &&
4853  (th->th.th_new_place <= th->th.th_last_place));
4854  } else {
4855  KMP_ASSERT((th->th.th_new_place <= th->th.th_first_place) ||
4856  (th->th.th_new_place >= th->th.th_last_place));
4857  }
4858 
4859  // Copy the thread mask to the kmp_info_t strucuture,
4860  // and set this thread's affinity.
4861  kmp_affin_mask_t *mask =
4862  KMP_CPU_INDEX(__kmp_affinity_masks, th->th.th_new_place);
4863  KMP_CPU_COPY(th->th.th_affin_mask, mask);
4864  th->th.th_current_place = th->th.th_new_place;
4865 
4866  if (__kmp_affinity_verbose) {
4867  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4868  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4869  th->th.th_affin_mask);
4870  KMP_INFORM(BoundToOSProcSet, "OMP_PROC_BIND", (kmp_int32)getpid(),
4871  __kmp_gettid(), gtid, buf);
4872  }
4873  __kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
4874 }
4875 
4876 #endif /* OMP_40_ENABLED */
4877 
4878 int __kmp_aux_set_affinity(void **mask) {
4879  int gtid;
4880  kmp_info_t *th;
4881  int retval;
4882 
4883  if (!KMP_AFFINITY_CAPABLE()) {
4884  return -1;
4885  }
4886 
4887  gtid = __kmp_entry_gtid();
4888  KA_TRACE(1000, ; {
4889  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4890  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4891  (kmp_affin_mask_t *)(*mask));
4892  __kmp_debug_printf(
4893  "kmp_set_affinity: setting affinity mask for thread %d = %s\n", gtid,
4894  buf);
4895  });
4896 
4897  if (__kmp_env_consistency_check) {
4898  if ((mask == NULL) || (*mask == NULL)) {
4899  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4900  } else {
4901  unsigned proc;
4902  int num_procs = 0;
4903 
4904  KMP_CPU_SET_ITERATE(proc, ((kmp_affin_mask_t *)(*mask))) {
4905  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
4906  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4907  }
4908  if (!KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask))) {
4909  continue;
4910  }
4911  num_procs++;
4912  }
4913  if (num_procs == 0) {
4914  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4915  }
4916 
4917 #if KMP_GROUP_AFFINITY
4918  if (__kmp_get_proc_group((kmp_affin_mask_t *)(*mask)) < 0) {
4919  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4920  }
4921 #endif /* KMP_GROUP_AFFINITY */
4922  }
4923  }
4924 
4925  th = __kmp_threads[gtid];
4926  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
4927  retval = __kmp_set_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
4928  if (retval == 0) {
4929  KMP_CPU_COPY(th->th.th_affin_mask, (kmp_affin_mask_t *)(*mask));
4930  }
4931 
4932 #if OMP_40_ENABLED
4933  th->th.th_current_place = KMP_PLACE_UNDEFINED;
4934  th->th.th_new_place = KMP_PLACE_UNDEFINED;
4935  th->th.th_first_place = 0;
4936  th->th.th_last_place = __kmp_affinity_num_masks - 1;
4937 
4938  // Turn off 4.0 affinity for the current tread at this parallel level.
4939  th->th.th_current_task->td_icvs.proc_bind = proc_bind_false;
4940 #endif
4941 
4942  return retval;
4943 }
4944 
4945 int __kmp_aux_get_affinity(void **mask) {
4946  int gtid;
4947  int retval;
4948  kmp_info_t *th;
4949 
4950  if (!KMP_AFFINITY_CAPABLE()) {
4951  return -1;
4952  }
4953 
4954  gtid = __kmp_entry_gtid();
4955  th = __kmp_threads[gtid];
4956  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
4957 
4958  KA_TRACE(1000, ; {
4959  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4960  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4961  th->th.th_affin_mask);
4962  __kmp_printf("kmp_get_affinity: stored affinity mask for thread %d = %s\n",
4963  gtid, buf);
4964  });
4965 
4966  if (__kmp_env_consistency_check) {
4967  if ((mask == NULL) || (*mask == NULL)) {
4968  KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity");
4969  }
4970  }
4971 
4972 #if !KMP_OS_WINDOWS
4973 
4974  retval = __kmp_get_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
4975  KA_TRACE(1000, ; {
4976  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4977  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4978  (kmp_affin_mask_t *)(*mask));
4979  __kmp_printf("kmp_get_affinity: system affinity mask for thread %d = %s\n",
4980  gtid, buf);
4981  });
4982  return retval;
4983 
4984 #else
4985 
4986  KMP_CPU_COPY((kmp_affin_mask_t *)(*mask), th->th.th_affin_mask);
4987  return 0;
4988 
4989 #endif /* KMP_OS_WINDOWS */
4990 }
4991 
4992 int __kmp_aux_get_affinity_max_proc() {
4993  if (!KMP_AFFINITY_CAPABLE()) {
4994  return 0;
4995  }
4996 #if KMP_GROUP_AFFINITY
4997  if (__kmp_num_proc_groups > 1) {
4998  return (int)(__kmp_num_proc_groups * sizeof(DWORD_PTR) * CHAR_BIT);
4999  }
5000 #endif
5001  return __kmp_xproc;
5002 }
5003 
5004 int __kmp_aux_set_affinity_mask_proc(int proc, void **mask) {
5005  if (!KMP_AFFINITY_CAPABLE()) {
5006  return -1;
5007  }
5008 
5009  KA_TRACE(1000, ; {
5010  int gtid = __kmp_entry_gtid();
5011  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5012  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5013  (kmp_affin_mask_t *)(*mask));
5014  __kmp_debug_printf("kmp_set_affinity_mask_proc: setting proc %d in "
5015  "affinity mask for thread %d = %s\n",
5016  proc, gtid, buf);
5017  });
5018 
5019  if (__kmp_env_consistency_check) {
5020  if ((mask == NULL) || (*mask == NULL)) {
5021  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity_mask_proc");
5022  }
5023  }
5024 
5025  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5026  return -1;
5027  }
5028  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5029  return -2;
5030  }
5031 
5032  KMP_CPU_SET(proc, (kmp_affin_mask_t *)(*mask));
5033  return 0;
5034 }
5035 
5036 int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask) {
5037  if (!KMP_AFFINITY_CAPABLE()) {
5038  return -1;
5039  }
5040 
5041  KA_TRACE(1000, ; {
5042  int gtid = __kmp_entry_gtid();
5043  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5044  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5045  (kmp_affin_mask_t *)(*mask));
5046  __kmp_debug_printf("kmp_unset_affinity_mask_proc: unsetting proc %d in "
5047  "affinity mask for thread %d = %s\n",
5048  proc, gtid, buf);
5049  });
5050 
5051  if (__kmp_env_consistency_check) {
5052  if ((mask == NULL) || (*mask == NULL)) {
5053  KMP_FATAL(AffinityInvalidMask, "kmp_unset_affinity_mask_proc");
5054  }
5055  }
5056 
5057  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5058  return -1;
5059  }
5060  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5061  return -2;
5062  }
5063 
5064  KMP_CPU_CLR(proc, (kmp_affin_mask_t *)(*mask));
5065  return 0;
5066 }
5067 
5068 int __kmp_aux_get_affinity_mask_proc(int proc, void **mask) {
5069  if (!KMP_AFFINITY_CAPABLE()) {
5070  return -1;
5071  }
5072 
5073  KA_TRACE(1000, ; {
5074  int gtid = __kmp_entry_gtid();
5075  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5076  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5077  (kmp_affin_mask_t *)(*mask));
5078  __kmp_debug_printf("kmp_get_affinity_mask_proc: getting proc %d in "
5079  "affinity mask for thread %d = %s\n",
5080  proc, gtid, buf);
5081  });
5082 
5083  if (__kmp_env_consistency_check) {
5084  if ((mask == NULL) || (*mask == NULL)) {
5085  KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity_mask_proc");
5086  }
5087  }
5088 
5089  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5090  return -1;
5091  }
5092  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5093  return 0;
5094  }
5095 
5096  return KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask));
5097 }
5098 
5099 // Dynamic affinity settings - Affinity balanced
5100 void __kmp_balanced_affinity(kmp_info_t *th, int nthreads) {
5101  KMP_DEBUG_ASSERT(th);
5102  bool fine_gran = true;
5103  int tid = th->th.th_info.ds.ds_tid;
5104 
5105  switch (__kmp_affinity_gran) {
5106  case affinity_gran_fine:
5107  case affinity_gran_thread:
5108  break;
5109  case affinity_gran_core:
5110  if (__kmp_nThreadsPerCore > 1) {
5111  fine_gran = false;
5112  }
5113  break;
5114  case affinity_gran_package:
5115  if (nCoresPerPkg > 1) {
5116  fine_gran = false;
5117  }
5118  break;
5119  default:
5120  fine_gran = false;
5121  }
5122 
5123  if (__kmp_affinity_uniform_topology()) {
5124  int coreID;
5125  int threadID;
5126  // Number of hyper threads per core in HT machine
5127  int __kmp_nth_per_core = __kmp_avail_proc / __kmp_ncores;
5128  // Number of cores
5129  int ncores = __kmp_ncores;
5130  if ((nPackages > 1) && (__kmp_nth_per_core <= 1)) {
5131  __kmp_nth_per_core = __kmp_avail_proc / nPackages;
5132  ncores = nPackages;
5133  }
5134  // How many threads will be bound to each core
5135  int chunk = nthreads / ncores;
5136  // How many cores will have an additional thread bound to it - "big cores"
5137  int big_cores = nthreads % ncores;
5138  // Number of threads on the big cores
5139  int big_nth = (chunk + 1) * big_cores;
5140  if (tid < big_nth) {
5141  coreID = tid / (chunk + 1);
5142  threadID = (tid % (chunk + 1)) % __kmp_nth_per_core;
5143  } else { // tid >= big_nth
5144  coreID = (tid - big_cores) / chunk;
5145  threadID = ((tid - big_cores) % chunk) % __kmp_nth_per_core;
5146  }
5147 
5148  KMP_DEBUG_ASSERT2(KMP_AFFINITY_CAPABLE(),
5149  "Illegal set affinity operation when not capable");
5150 
5151  kmp_affin_mask_t *mask = th->th.th_affin_mask;
5152  KMP_CPU_ZERO(mask);
5153 
5154  if (fine_gran) {
5155  int osID = address2os[coreID * __kmp_nth_per_core + threadID].second;
5156  KMP_CPU_SET(osID, mask);
5157  } else {
5158  for (int i = 0; i < __kmp_nth_per_core; i++) {
5159  int osID;
5160  osID = address2os[coreID * __kmp_nth_per_core + i].second;
5161  KMP_CPU_SET(osID, mask);
5162  }
5163  }
5164  if (__kmp_affinity_verbose) {
5165  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5166  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
5167  KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY", (kmp_int32)getpid(),
5168  __kmp_gettid(), tid, buf);
5169  }
5170  __kmp_set_system_affinity(mask, TRUE);
5171  } else { // Non-uniform topology
5172 
5173  kmp_affin_mask_t *mask = th->th.th_affin_mask;
5174  KMP_CPU_ZERO(mask);
5175 
5176  int core_level = __kmp_affinity_find_core_level(
5177  address2os, __kmp_avail_proc, __kmp_aff_depth - 1);
5178  int ncores = __kmp_affinity_compute_ncores(address2os, __kmp_avail_proc,
5179  __kmp_aff_depth - 1, core_level);
5180  int nth_per_core = __kmp_affinity_max_proc_per_core(
5181  address2os, __kmp_avail_proc, __kmp_aff_depth - 1, core_level);
5182 
5183  // For performance gain consider the special case nthreads ==
5184  // __kmp_avail_proc
5185  if (nthreads == __kmp_avail_proc) {
5186  if (fine_gran) {
5187  int osID = address2os[tid].second;
5188  KMP_CPU_SET(osID, mask);
5189  } else {
5190  int core = __kmp_affinity_find_core(address2os, tid,
5191  __kmp_aff_depth - 1, core_level);
5192  for (int i = 0; i < __kmp_avail_proc; i++) {
5193  int osID = address2os[i].second;
5194  if (__kmp_affinity_find_core(address2os, i, __kmp_aff_depth - 1,
5195  core_level) == core) {
5196  KMP_CPU_SET(osID, mask);
5197  }
5198  }
5199  }
5200  } else if (nthreads <= ncores) {
5201 
5202  int core = 0;
5203  for (int i = 0; i < ncores; i++) {
5204  // Check if this core from procarr[] is in the mask
5205  int in_mask = 0;
5206  for (int j = 0; j < nth_per_core; j++) {
5207  if (procarr[i * nth_per_core + j] != -1) {
5208  in_mask = 1;
5209  break;
5210  }
5211  }
5212  if (in_mask) {
5213  if (tid == core) {
5214  for (int j = 0; j < nth_per_core; j++) {
5215  int osID = procarr[i * nth_per_core + j];
5216  if (osID != -1) {
5217  KMP_CPU_SET(osID, mask);
5218  // For fine granularity it is enough to set the first available
5219  // osID for this core
5220  if (fine_gran) {
5221  break;
5222  }
5223  }
5224  }
5225  break;
5226  } else {
5227  core++;
5228  }
5229  }
5230  }
5231  } else { // nthreads > ncores
5232  // Array to save the number of processors at each core
5233  int *nproc_at_core = (int *)KMP_ALLOCA(sizeof(int) * ncores);
5234  // Array to save the number of cores with "x" available processors;
5235  int *ncores_with_x_procs =
5236  (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
5237  // Array to save the number of cores with # procs from x to nth_per_core
5238  int *ncores_with_x_to_max_procs =
5239  (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
5240 
5241  for (int i = 0; i <= nth_per_core; i++) {
5242  ncores_with_x_procs[i] = 0;
5243  ncores_with_x_to_max_procs[i] = 0;
5244  }
5245 
5246  for (int i = 0; i < ncores; i++) {
5247  int cnt = 0;
5248  for (int j = 0; j < nth_per_core; j++) {
5249  if (procarr[i * nth_per_core + j] != -1) {
5250  cnt++;
5251  }
5252  }
5253  nproc_at_core[i] = cnt;
5254  ncores_with_x_procs[cnt]++;
5255  }
5256 
5257  for (int i = 0; i <= nth_per_core; i++) {
5258  for (int j = i; j <= nth_per_core; j++) {
5259  ncores_with_x_to_max_procs[i] += ncores_with_x_procs[j];
5260  }
5261  }
5262 
5263  // Max number of processors
5264  int nproc = nth_per_core * ncores;
5265  // An array to keep number of threads per each context
5266  int *newarr = (int *)__kmp_allocate(sizeof(int) * nproc);
5267  for (int i = 0; i < nproc; i++) {
5268  newarr[i] = 0;
5269  }
5270 
5271  int nth = nthreads;
5272  int flag = 0;
5273  while (nth > 0) {
5274  for (int j = 1; j <= nth_per_core; j++) {
5275  int cnt = ncores_with_x_to_max_procs[j];
5276  for (int i = 0; i < ncores; i++) {
5277  // Skip the core with 0 processors
5278  if (nproc_at_core[i] == 0) {
5279  continue;
5280  }
5281  for (int k = 0; k < nth_per_core; k++) {
5282  if (procarr[i * nth_per_core + k] != -1) {
5283  if (newarr[i * nth_per_core + k] == 0) {
5284  newarr[i * nth_per_core + k] = 1;
5285  cnt--;
5286  nth--;
5287  break;
5288  } else {
5289  if (flag != 0) {
5290  newarr[i * nth_per_core + k]++;
5291  cnt--;
5292  nth--;
5293  break;
5294  }
5295  }
5296  }
5297  }
5298  if (cnt == 0 || nth == 0) {
5299  break;
5300  }
5301  }
5302  if (nth == 0) {
5303  break;
5304  }
5305  }
5306  flag = 1;
5307  }
5308  int sum = 0;
5309  for (int i = 0; i < nproc; i++) {
5310  sum += newarr[i];
5311  if (sum > tid) {
5312  if (fine_gran) {
5313  int osID = procarr[i];
5314  KMP_CPU_SET(osID, mask);
5315  } else {
5316  int coreID = i / nth_per_core;
5317  for (int ii = 0; ii < nth_per_core; ii++) {
5318  int osID = procarr[coreID * nth_per_core + ii];
5319  if (osID != -1) {
5320  KMP_CPU_SET(osID, mask);
5321  }
5322  }
5323  }
5324  break;
5325  }
5326  }
5327  __kmp_free(newarr);
5328  }
5329 
5330  if (__kmp_affinity_verbose) {
5331  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5332  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
5333  KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY", (kmp_int32)getpid(),
5334  __kmp_gettid(), tid, buf);
5335  }
5336  __kmp_set_system_affinity(mask, TRUE);
5337  }
5338 }
5339 
5340 #if KMP_OS_LINUX
5341 // We don't need this entry for Windows because
5342 // there is GetProcessAffinityMask() api
5343 //
5344 // The intended usage is indicated by these steps:
5345 // 1) The user gets the current affinity mask
5346 // 2) Then sets the affinity by calling this function
5347 // 3) Error check the return value
5348 // 4) Use non-OpenMP parallelization
5349 // 5) Reset the affinity to what was stored in step 1)
5350 #ifdef __cplusplus
5351 extern "C"
5352 #endif
5353  int
5354  kmp_set_thread_affinity_mask_initial()
5355 // the function returns 0 on success,
5356 // -1 if we cannot bind thread
5357 // >0 (errno) if an error happened during binding
5358 {
5359  int gtid = __kmp_get_gtid();
5360  if (gtid < 0) {
5361  // Do not touch non-omp threads
5362  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5363  "non-omp thread, returning\n"));
5364  return -1;
5365  }
5366  if (!KMP_AFFINITY_CAPABLE() || !__kmp_init_middle) {
5367  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5368  "affinity not initialized, returning\n"));
5369  return -1;
5370  }
5371  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5372  "set full mask for thread %d\n",
5373  gtid));
5374  KMP_DEBUG_ASSERT(__kmp_affin_fullMask != NULL);
5375  return __kmp_set_system_affinity(__kmp_affin_fullMask, FALSE);
5376 }
5377 #endif
5378 
5379 #endif // KMP_AFFINITY_SUPPORTED