LLVM OpenMP* Runtime Library
kmp_affinity.cpp
1 /*
2  * kmp_affinity.cpp -- affinity management
3  */
4 
5 //===----------------------------------------------------------------------===//
6 //
7 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
8 // See https://llvm.org/LICENSE.txt for license information.
9 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "kmp.h"
14 #include "kmp_affinity.h"
15 #include "kmp_i18n.h"
16 #include "kmp_io.h"
17 #include "kmp_str.h"
18 #include "kmp_wrapper_getpid.h"
19 #if KMP_USE_HIER_SCHED
20 #include "kmp_dispatch_hier.h"
21 #endif
22 
23 // Store the real or imagined machine hierarchy here
24 static hierarchy_info machine_hierarchy;
25 
26 void __kmp_cleanup_hierarchy() { machine_hierarchy.fini(); }
27 
28 void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar) {
29  kmp_uint32 depth;
30  // The test below is true if affinity is available, but set to "none". Need to
31  // init on first use of hierarchical barrier.
32  if (TCR_1(machine_hierarchy.uninitialized))
33  machine_hierarchy.init(NULL, nproc);
34 
35  // Adjust the hierarchy in case num threads exceeds original
36  if (nproc > machine_hierarchy.base_num_threads)
37  machine_hierarchy.resize(nproc);
38 
39  depth = machine_hierarchy.depth;
40  KMP_DEBUG_ASSERT(depth > 0);
41 
42  thr_bar->depth = depth;
43  thr_bar->base_leaf_kids = (kmp_uint8)machine_hierarchy.numPerLevel[0] - 1;
44  thr_bar->skip_per_level = machine_hierarchy.skipPerLevel;
45 }
46 
47 #if KMP_AFFINITY_SUPPORTED
48 
49 bool KMPAffinity::picked_api = false;
50 
51 void *KMPAffinity::Mask::operator new(size_t n) { return __kmp_allocate(n); }
52 void *KMPAffinity::Mask::operator new[](size_t n) { return __kmp_allocate(n); }
53 void KMPAffinity::Mask::operator delete(void *p) { __kmp_free(p); }
54 void KMPAffinity::Mask::operator delete[](void *p) { __kmp_free(p); }
55 void *KMPAffinity::operator new(size_t n) { return __kmp_allocate(n); }
56 void KMPAffinity::operator delete(void *p) { __kmp_free(p); }
57 
58 void KMPAffinity::pick_api() {
59  KMPAffinity *affinity_dispatch;
60  if (picked_api)
61  return;
62 #if KMP_USE_HWLOC
63  // Only use Hwloc if affinity isn't explicitly disabled and
64  // user requests Hwloc topology method
65  if (__kmp_affinity_top_method == affinity_top_method_hwloc &&
66  __kmp_affinity_type != affinity_disabled) {
67  affinity_dispatch = new KMPHwlocAffinity();
68  } else
69 #endif
70  {
71  affinity_dispatch = new KMPNativeAffinity();
72  }
73  __kmp_affinity_dispatch = affinity_dispatch;
74  picked_api = true;
75 }
76 
77 void KMPAffinity::destroy_api() {
78  if (__kmp_affinity_dispatch != NULL) {
79  delete __kmp_affinity_dispatch;
80  __kmp_affinity_dispatch = NULL;
81  picked_api = false;
82  }
83 }
84 
85 #define KMP_ADVANCE_SCAN(scan) \
86  while (*scan != '\0') { \
87  scan++; \
88  }
89 
90 // Print the affinity mask to the character array in a pretty format.
91 // The format is a comma separated list of non-negative integers or integer
92 // ranges: e.g., 1,2,3-5,7,9-15
93 // The format can also be the string "{<empty>}" if no bits are set in mask
94 char *__kmp_affinity_print_mask(char *buf, int buf_len,
95  kmp_affin_mask_t *mask) {
96  int start = 0, finish = 0, previous = 0;
97  bool first_range;
98  KMP_ASSERT(buf);
99  KMP_ASSERT(buf_len >= 40);
100  KMP_ASSERT(mask);
101  char *scan = buf;
102  char *end = buf + buf_len - 1;
103 
104  // Check for empty set.
105  if (mask->begin() == mask->end()) {
106  KMP_SNPRINTF(scan, end - scan + 1, "{<empty>}");
107  KMP_ADVANCE_SCAN(scan);
108  KMP_ASSERT(scan <= end);
109  return buf;
110  }
111 
112  first_range = true;
113  start = mask->begin();
114  while (1) {
115  // Find next range
116  // [start, previous] is inclusive range of contiguous bits in mask
117  for (finish = mask->next(start), previous = start;
118  finish == previous + 1 && finish != mask->end();
119  finish = mask->next(finish)) {
120  previous = finish;
121  }
122 
123  // The first range does not need a comma printed before it, but the rest
124  // of the ranges do need a comma beforehand
125  if (!first_range) {
126  KMP_SNPRINTF(scan, end - scan + 1, "%s", ",");
127  KMP_ADVANCE_SCAN(scan);
128  } else {
129  first_range = false;
130  }
131  // Range with three or more contiguous bits in the affinity mask
132  if (previous - start > 1) {
133  KMP_SNPRINTF(scan, end - scan + 1, "%d-%d", static_cast<int>(start),
134  static_cast<int>(previous));
135  } else {
136  // Range with one or two contiguous bits in the affinity mask
137  KMP_SNPRINTF(scan, end - scan + 1, "%d", static_cast<int>(start));
138  KMP_ADVANCE_SCAN(scan);
139  if (previous - start > 0) {
140  KMP_SNPRINTF(scan, end - scan + 1, ",%d", static_cast<int>(previous));
141  }
142  }
143  KMP_ADVANCE_SCAN(scan);
144  // Start over with new start point
145  start = finish;
146  if (start == mask->end())
147  break;
148  // Check for overflow
149  if (end - scan < 2)
150  break;
151  }
152 
153  // Check for overflow
154  KMP_ASSERT(scan <= end);
155  return buf;
156 }
157 #undef KMP_ADVANCE_SCAN
158 
159 // Print the affinity mask to the string buffer object in a pretty format
160 // The format is a comma separated list of non-negative integers or integer
161 // ranges: e.g., 1,2,3-5,7,9-15
162 // The format can also be the string "{<empty>}" if no bits are set in mask
163 kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
164  kmp_affin_mask_t *mask) {
165  int start = 0, finish = 0, previous = 0;
166  bool first_range;
167  KMP_ASSERT(buf);
168  KMP_ASSERT(mask);
169 
170  __kmp_str_buf_clear(buf);
171 
172  // Check for empty set.
173  if (mask->begin() == mask->end()) {
174  __kmp_str_buf_print(buf, "%s", "{<empty>}");
175  return buf;
176  }
177 
178  first_range = true;
179  start = mask->begin();
180  while (1) {
181  // Find next range
182  // [start, previous] is inclusive range of contiguous bits in mask
183  for (finish = mask->next(start), previous = start;
184  finish == previous + 1 && finish != mask->end();
185  finish = mask->next(finish)) {
186  previous = finish;
187  }
188 
189  // The first range does not need a comma printed before it, but the rest
190  // of the ranges do need a comma beforehand
191  if (!first_range) {
192  __kmp_str_buf_print(buf, "%s", ",");
193  } else {
194  first_range = false;
195  }
196  // Range with three or more contiguous bits in the affinity mask
197  if (previous - start > 1) {
198  __kmp_str_buf_print(buf, "%d-%d", static_cast<int>(start),
199  static_cast<int>(previous));
200  } else {
201  // Range with one or two contiguous bits in the affinity mask
202  __kmp_str_buf_print(buf, "%d", static_cast<int>(start));
203  if (previous - start > 0) {
204  __kmp_str_buf_print(buf, ",%d", static_cast<int>(previous));
205  }
206  }
207  // Start over with new start point
208  start = finish;
209  if (start == mask->end())
210  break;
211  }
212  return buf;
213 }
214 
215 void __kmp_affinity_entire_machine_mask(kmp_affin_mask_t *mask) {
216  KMP_CPU_ZERO(mask);
217 
218 #if KMP_GROUP_AFFINITY
219 
220  if (__kmp_num_proc_groups > 1) {
221  int group;
222  KMP_DEBUG_ASSERT(__kmp_GetActiveProcessorCount != NULL);
223  for (group = 0; group < __kmp_num_proc_groups; group++) {
224  int i;
225  int num = __kmp_GetActiveProcessorCount(group);
226  for (i = 0; i < num; i++) {
227  KMP_CPU_SET(i + group * (CHAR_BIT * sizeof(DWORD_PTR)), mask);
228  }
229  }
230  } else
231 
232 #endif /* KMP_GROUP_AFFINITY */
233 
234  {
235  int proc;
236  for (proc = 0; proc < __kmp_xproc; proc++) {
237  KMP_CPU_SET(proc, mask);
238  }
239  }
240 }
241 
242 // When sorting by labels, __kmp_affinity_assign_child_nums() must first be
243 // called to renumber the labels from [0..n] and place them into the child_num
244 // vector of the address object. This is done in case the labels used for
245 // the children at one node of the hierarchy differ from those used for
246 // another node at the same level. Example: suppose the machine has 2 nodes
247 // with 2 packages each. The first node contains packages 601 and 602, and
248 // second node contains packages 603 and 604. If we try to sort the table
249 // for "scatter" affinity, the table will still be sorted 601, 602, 603, 604
250 // because we are paying attention to the labels themselves, not the ordinal
251 // child numbers. By using the child numbers in the sort, the result is
252 // {0,0}=601, {0,1}=603, {1,0}=602, {1,1}=604.
253 static void __kmp_affinity_assign_child_nums(AddrUnsPair *address2os,
254  int numAddrs) {
255  KMP_DEBUG_ASSERT(numAddrs > 0);
256  int depth = address2os->first.depth;
257  unsigned *counts = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
258  unsigned *lastLabel = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
259  int labCt;
260  for (labCt = 0; labCt < depth; labCt++) {
261  address2os[0].first.childNums[labCt] = counts[labCt] = 0;
262  lastLabel[labCt] = address2os[0].first.labels[labCt];
263  }
264  int i;
265  for (i = 1; i < numAddrs; i++) {
266  for (labCt = 0; labCt < depth; labCt++) {
267  if (address2os[i].first.labels[labCt] != lastLabel[labCt]) {
268  int labCt2;
269  for (labCt2 = labCt + 1; labCt2 < depth; labCt2++) {
270  counts[labCt2] = 0;
271  lastLabel[labCt2] = address2os[i].first.labels[labCt2];
272  }
273  counts[labCt]++;
274  lastLabel[labCt] = address2os[i].first.labels[labCt];
275  break;
276  }
277  }
278  for (labCt = 0; labCt < depth; labCt++) {
279  address2os[i].first.childNums[labCt] = counts[labCt];
280  }
281  for (; labCt < (int)Address::maxDepth; labCt++) {
282  address2os[i].first.childNums[labCt] = 0;
283  }
284  }
285  __kmp_free(lastLabel);
286  __kmp_free(counts);
287 }
288 
289 // All of the __kmp_affinity_create_*_map() routines should set
290 // __kmp_affinity_masks to a vector of affinity mask objects of length
291 // __kmp_affinity_num_masks, if __kmp_affinity_type != affinity_none, and return
292 // the number of levels in the machine topology tree (zero if
293 // __kmp_affinity_type == affinity_none).
294 //
295 // All of the __kmp_affinity_create_*_map() routines should set
296 // *__kmp_affin_fullMask to the affinity mask for the initialization thread.
297 // They need to save and restore the mask, and it could be needed later, so
298 // saving it is just an optimization to avoid calling kmp_get_system_affinity()
299 // again.
300 kmp_affin_mask_t *__kmp_affin_fullMask = NULL;
301 
302 static int nCoresPerPkg, nPackages;
303 static int __kmp_nThreadsPerCore;
304 #ifndef KMP_DFLT_NTH_CORES
305 static int __kmp_ncores;
306 #endif
307 static int *__kmp_pu_os_idx = NULL;
308 
309 // __kmp_affinity_uniform_topology() doesn't work when called from
310 // places which support arbitrarily many levels in the machine topology
311 // map, i.e. the non-default cases in __kmp_affinity_create_cpuinfo_map()
312 // __kmp_affinity_create_x2apicid_map().
313 inline static bool __kmp_affinity_uniform_topology() {
314  return __kmp_avail_proc == (__kmp_nThreadsPerCore * nCoresPerPkg * nPackages);
315 }
316 
317 // Print out the detailed machine topology map, i.e. the physical locations
318 // of each OS proc.
319 static void __kmp_affinity_print_topology(AddrUnsPair *address2os, int len,
320  int depth, int pkgLevel,
321  int coreLevel, int threadLevel) {
322  int proc;
323 
324  KMP_INFORM(OSProcToPhysicalThreadMap, "KMP_AFFINITY");
325  for (proc = 0; proc < len; proc++) {
326  int level;
327  kmp_str_buf_t buf;
328  __kmp_str_buf_init(&buf);
329  for (level = 0; level < depth; level++) {
330  if (level == threadLevel) {
331  __kmp_str_buf_print(&buf, "%s ", KMP_I18N_STR(Thread));
332  } else if (level == coreLevel) {
333  __kmp_str_buf_print(&buf, "%s ", KMP_I18N_STR(Core));
334  } else if (level == pkgLevel) {
335  __kmp_str_buf_print(&buf, "%s ", KMP_I18N_STR(Package));
336  } else if (level > pkgLevel) {
337  __kmp_str_buf_print(&buf, "%s_%d ", KMP_I18N_STR(Node),
338  level - pkgLevel - 1);
339  } else {
340  __kmp_str_buf_print(&buf, "L%d ", level);
341  }
342  __kmp_str_buf_print(&buf, "%d ", address2os[proc].first.labels[level]);
343  }
344  KMP_INFORM(OSProcMapToPack, "KMP_AFFINITY", address2os[proc].second,
345  buf.str);
346  __kmp_str_buf_free(&buf);
347  }
348 }
349 
350 #if KMP_USE_HWLOC
351 
352 static void __kmp_affinity_print_hwloc_tp(AddrUnsPair *addrP, int len,
353  int depth, int *levels) {
354  int proc;
355  kmp_str_buf_t buf;
356  __kmp_str_buf_init(&buf);
357  KMP_INFORM(OSProcToPhysicalThreadMap, "KMP_AFFINITY");
358  for (proc = 0; proc < len; proc++) {
359  __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Package),
360  addrP[proc].first.labels[0]);
361  if (depth > 1) {
362  int level = 1; // iterate over levels
363  int label = 1; // iterate over labels
364  if (__kmp_numa_detected)
365  // node level follows package
366  if (levels[level++] > 0)
367  __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Node),
368  addrP[proc].first.labels[label++]);
369  if (__kmp_tile_depth > 0)
370  // tile level follows node if any, or package
371  if (levels[level++] > 0)
372  __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Tile),
373  addrP[proc].first.labels[label++]);
374  if (levels[level++] > 0)
375  // core level follows
376  __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Core),
377  addrP[proc].first.labels[label++]);
378  if (levels[level++] > 0)
379  // thread level is the latest
380  __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Thread),
381  addrP[proc].first.labels[label++]);
382  KMP_DEBUG_ASSERT(label == depth);
383  }
384  KMP_INFORM(OSProcMapToPack, "KMP_AFFINITY", addrP[proc].second, buf.str);
385  __kmp_str_buf_clear(&buf);
386  }
387  __kmp_str_buf_free(&buf);
388 }
389 
390 static int nNodePerPkg, nTilePerPkg, nTilePerNode, nCorePerNode, nCorePerTile;
391 
392 // This function removes the topology levels that are radix 1 and don't offer
393 // further information about the topology. The most common example is when you
394 // have one thread context per core, we don't want the extra thread context
395 // level if it offers no unique labels. So they are removed.
396 // return value: the new depth of address2os
397 static int __kmp_affinity_remove_radix_one_levels(AddrUnsPair *addrP, int nTh,
398  int depth, int *levels) {
399  int level;
400  int i;
401  int radix1_detected;
402  int new_depth = depth;
403  for (level = depth - 1; level > 0; --level) {
404  // Detect if this level is radix 1
405  radix1_detected = 1;
406  for (i = 1; i < nTh; ++i) {
407  if (addrP[0].first.labels[level] != addrP[i].first.labels[level]) {
408  // There are differing label values for this level so it stays
409  radix1_detected = 0;
410  break;
411  }
412  }
413  if (!radix1_detected)
414  continue;
415  // Radix 1 was detected
416  --new_depth;
417  levels[level] = -1; // mark level as not present in address2os array
418  if (level == new_depth) {
419  // "turn off" deepest level, just decrement the depth that removes
420  // the level from address2os array
421  for (i = 0; i < nTh; ++i) {
422  addrP[i].first.depth--;
423  }
424  } else {
425  // For other levels, we move labels over and also reduce the depth
426  int j;
427  for (j = level; j < new_depth; ++j) {
428  for (i = 0; i < nTh; ++i) {
429  addrP[i].first.labels[j] = addrP[i].first.labels[j + 1];
430  addrP[i].first.depth--;
431  }
432  levels[j + 1] -= 1;
433  }
434  }
435  }
436  return new_depth;
437 }
438 
439 // Returns the number of objects of type 'type' below 'obj' within the topology
440 // tree structure. e.g., if obj is a HWLOC_OBJ_PACKAGE object, and type is
441 // HWLOC_OBJ_PU, then this will return the number of PU's under the SOCKET
442 // object.
443 static int __kmp_hwloc_get_nobjs_under_obj(hwloc_obj_t obj,
444  hwloc_obj_type_t type) {
445  int retval = 0;
446  hwloc_obj_t first;
447  for (first = hwloc_get_obj_below_by_type(__kmp_hwloc_topology, obj->type,
448  obj->logical_index, type, 0);
449  first != NULL &&
450  hwloc_get_ancestor_obj_by_type(__kmp_hwloc_topology, obj->type, first) ==
451  obj;
452  first = hwloc_get_next_obj_by_type(__kmp_hwloc_topology, first->type,
453  first)) {
454  ++retval;
455  }
456  return retval;
457 }
458 
459 static int __kmp_hwloc_count_children_by_depth(hwloc_topology_t t,
460  hwloc_obj_t o,
461  kmp_hwloc_depth_t depth,
462  hwloc_obj_t *f) {
463  if (o->depth == depth) {
464  if (*f == NULL)
465  *f = o; // output first descendant found
466  return 1;
467  }
468  int sum = 0;
469  for (unsigned i = 0; i < o->arity; i++)
470  sum += __kmp_hwloc_count_children_by_depth(t, o->children[i], depth, f);
471  return sum; // will be 0 if no one found (as PU arity is 0)
472 }
473 
474 static int __kmp_hwloc_count_children_by_type(hwloc_topology_t t, hwloc_obj_t o,
475  hwloc_obj_type_t type,
476  hwloc_obj_t *f) {
477  if (!hwloc_compare_types(o->type, type)) {
478  if (*f == NULL)
479  *f = o; // output first descendant found
480  return 1;
481  }
482  int sum = 0;
483  for (unsigned i = 0; i < o->arity; i++)
484  sum += __kmp_hwloc_count_children_by_type(t, o->children[i], type, f);
485  return sum; // will be 0 if no one found (as PU arity is 0)
486 }
487 
488 static int __kmp_hwloc_process_obj_core_pu(AddrUnsPair *addrPair,
489  int &nActiveThreads,
490  int &num_active_cores,
491  hwloc_obj_t obj, int depth,
492  int *labels) {
493  hwloc_obj_t core = NULL;
494  hwloc_topology_t &tp = __kmp_hwloc_topology;
495  int NC = __kmp_hwloc_count_children_by_type(tp, obj, HWLOC_OBJ_CORE, &core);
496  for (int core_id = 0; core_id < NC; ++core_id, core = core->next_cousin) {
497  hwloc_obj_t pu = NULL;
498  KMP_DEBUG_ASSERT(core != NULL);
499  int num_active_threads = 0;
500  int NT = __kmp_hwloc_count_children_by_type(tp, core, HWLOC_OBJ_PU, &pu);
501  // int NT = core->arity; pu = core->first_child; // faster?
502  for (int pu_id = 0; pu_id < NT; ++pu_id, pu = pu->next_cousin) {
503  KMP_DEBUG_ASSERT(pu != NULL);
504  if (!KMP_CPU_ISSET(pu->os_index, __kmp_affin_fullMask))
505  continue; // skip inactive (inaccessible) unit
506  Address addr(depth + 2);
507  KA_TRACE(20, ("Hwloc inserting %d (%d) %d (%d) %d (%d) into address2os\n",
508  obj->os_index, obj->logical_index, core->os_index,
509  core->logical_index, pu->os_index, pu->logical_index));
510  for (int i = 0; i < depth; ++i)
511  addr.labels[i] = labels[i]; // package, etc.
512  addr.labels[depth] = core_id; // core
513  addr.labels[depth + 1] = pu_id; // pu
514  addrPair[nActiveThreads] = AddrUnsPair(addr, pu->os_index);
515  __kmp_pu_os_idx[nActiveThreads] = pu->os_index;
516  nActiveThreads++;
517  ++num_active_threads; // count active threads per core
518  }
519  if (num_active_threads) { // were there any active threads on the core?
520  ++__kmp_ncores; // count total active cores
521  ++num_active_cores; // count active cores per socket
522  if (num_active_threads > __kmp_nThreadsPerCore)
523  __kmp_nThreadsPerCore = num_active_threads; // calc maximum
524  }
525  }
526  return 0;
527 }
528 
529 // Check if NUMA node detected below the package,
530 // and if tile object is detected and return its depth
531 static int __kmp_hwloc_check_numa() {
532  hwloc_topology_t &tp = __kmp_hwloc_topology;
533  hwloc_obj_t hT, hC, hL, hN, hS; // hwloc objects (pointers to)
534  int depth, l2cache_depth, package_depth;
535 
536  // Get some PU
537  hT = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PU, 0);
538  if (hT == NULL) // something has gone wrong
539  return 1;
540 
541  // check NUMA node below PACKAGE
542  hN = hwloc_get_ancestor_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hT);
543  hS = hwloc_get_ancestor_obj_by_type(tp, HWLOC_OBJ_PACKAGE, hT);
544  KMP_DEBUG_ASSERT(hS != NULL);
545  if (hN != NULL && hN->depth > hS->depth) {
546  __kmp_numa_detected = TRUE; // socket includes node(s)
547  if (__kmp_affinity_gran == affinity_gran_node) {
548  __kmp_affinity_gran = affinity_gran_numa;
549  }
550  }
551 
552  package_depth = hwloc_get_type_depth(tp, HWLOC_OBJ_PACKAGE);
553  l2cache_depth = hwloc_get_cache_type_depth(tp, 2, HWLOC_OBJ_CACHE_UNIFIED);
554  // check tile, get object by depth because of multiple caches possible
555  depth = (l2cache_depth < package_depth) ? package_depth : l2cache_depth;
556  hL = hwloc_get_ancestor_obj_by_depth(tp, depth, hT);
557  hC = NULL; // not used, but reset it here just in case
558  if (hL != NULL &&
559  __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE, &hC) > 1)
560  __kmp_tile_depth = depth; // tile consists of multiple cores
561  return 0;
562 }
563 
564 static int __kmp_affinity_create_hwloc_map(AddrUnsPair **address2os,
565  kmp_i18n_id_t *const msg_id) {
566  hwloc_topology_t &tp = __kmp_hwloc_topology; // shortcut of a long name
567  *address2os = NULL;
568  *msg_id = kmp_i18n_null;
569 
570  // Save the affinity mask for the current thread.
571  kmp_affin_mask_t *oldMask;
572  KMP_CPU_ALLOC(oldMask);
573  __kmp_get_system_affinity(oldMask, TRUE);
574  __kmp_hwloc_check_numa();
575 
576  if (!KMP_AFFINITY_CAPABLE()) {
577  // Hack to try and infer the machine topology using only the data
578  // available from cpuid on the current thread, and __kmp_xproc.
579  KMP_ASSERT(__kmp_affinity_type == affinity_none);
580  // hwloc only guarantees existance of PU object, so check PACKAGE and CORE
581  hwloc_obj_t o = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PACKAGE, 0);
582  if (o != NULL)
583  nCoresPerPkg = __kmp_hwloc_get_nobjs_under_obj(o, HWLOC_OBJ_CORE);
584  else
585  nCoresPerPkg = 1; // no PACKAGE found
586  o = hwloc_get_obj_by_type(tp, HWLOC_OBJ_CORE, 0);
587  if (o != NULL)
588  __kmp_nThreadsPerCore = __kmp_hwloc_get_nobjs_under_obj(o, HWLOC_OBJ_PU);
589  else
590  __kmp_nThreadsPerCore = 1; // no CORE found
591  __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
592  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
593  if (__kmp_affinity_verbose) {
594  KMP_INFORM(AffNotCapableUseLocCpuidL11, "KMP_AFFINITY");
595  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
596  if (__kmp_affinity_uniform_topology()) {
597  KMP_INFORM(Uniform, "KMP_AFFINITY");
598  } else {
599  KMP_INFORM(NonUniform, "KMP_AFFINITY");
600  }
601  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
602  __kmp_nThreadsPerCore, __kmp_ncores);
603  }
604  KMP_CPU_FREE(oldMask);
605  return 0;
606  }
607 
608  int depth = 3;
609  int levels[5] = {0, 1, 2, 3, 4}; // package, [node,] [tile,] core, thread
610  int labels[3] = {0}; // package [,node] [,tile] - head of labels array
611  if (__kmp_numa_detected)
612  ++depth;
613  if (__kmp_tile_depth)
614  ++depth;
615 
616  // Allocate the data structure to be returned.
617  AddrUnsPair *retval =
618  (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * __kmp_avail_proc);
619  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
620  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
621 
622  // When affinity is off, this routine will still be called to set
623  // __kmp_ncores, as well as __kmp_nThreadsPerCore,
624  // nCoresPerPkg, & nPackages. Make sure all these vars are set
625  // correctly, and return if affinity is not enabled.
626 
627  hwloc_obj_t socket, node, tile;
628  int nActiveThreads = 0;
629  int socket_id = 0;
630  // re-calculate globals to count only accessible resources
631  __kmp_ncores = nPackages = nCoresPerPkg = __kmp_nThreadsPerCore = 0;
632  nNodePerPkg = nTilePerPkg = nTilePerNode = nCorePerNode = nCorePerTile = 0;
633  for (socket = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PACKAGE, 0); socket != NULL;
634  socket = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PACKAGE, socket),
635  socket_id++) {
636  labels[0] = socket_id;
637  if (__kmp_numa_detected) {
638  int NN;
639  int n_active_nodes = 0;
640  node = NULL;
641  NN = __kmp_hwloc_count_children_by_type(tp, socket, HWLOC_OBJ_NUMANODE,
642  &node);
643  for (int node_id = 0; node_id < NN; ++node_id, node = node->next_cousin) {
644  labels[1] = node_id;
645  if (__kmp_tile_depth) {
646  // NUMA + tiles
647  int NT;
648  int n_active_tiles = 0;
649  tile = NULL;
650  NT = __kmp_hwloc_count_children_by_depth(tp, node, __kmp_tile_depth,
651  &tile);
652  for (int tl_id = 0; tl_id < NT; ++tl_id, tile = tile->next_cousin) {
653  labels[2] = tl_id;
654  int n_active_cores = 0;
655  __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads,
656  n_active_cores, tile, 3, labels);
657  if (n_active_cores) { // were there any active cores on the socket?
658  ++n_active_tiles; // count active tiles per node
659  if (n_active_cores > nCorePerTile)
660  nCorePerTile = n_active_cores; // calc maximum
661  }
662  }
663  if (n_active_tiles) { // were there any active tiles on the socket?
664  ++n_active_nodes; // count active nodes per package
665  if (n_active_tiles > nTilePerNode)
666  nTilePerNode = n_active_tiles; // calc maximum
667  }
668  } else {
669  // NUMA, no tiles
670  int n_active_cores = 0;
671  __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads,
672  n_active_cores, node, 2, labels);
673  if (n_active_cores) { // were there any active cores on the socket?
674  ++n_active_nodes; // count active nodes per package
675  if (n_active_cores > nCorePerNode)
676  nCorePerNode = n_active_cores; // calc maximum
677  }
678  }
679  }
680  if (n_active_nodes) { // were there any active nodes on the socket?
681  ++nPackages; // count total active packages
682  if (n_active_nodes > nNodePerPkg)
683  nNodePerPkg = n_active_nodes; // calc maximum
684  }
685  } else {
686  if (__kmp_tile_depth) {
687  // no NUMA, tiles
688  int NT;
689  int n_active_tiles = 0;
690  tile = NULL;
691  NT = __kmp_hwloc_count_children_by_depth(tp, socket, __kmp_tile_depth,
692  &tile);
693  for (int tl_id = 0; tl_id < NT; ++tl_id, tile = tile->next_cousin) {
694  labels[1] = tl_id;
695  int n_active_cores = 0;
696  __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads,
697  n_active_cores, tile, 2, labels);
698  if (n_active_cores) { // were there any active cores on the socket?
699  ++n_active_tiles; // count active tiles per package
700  if (n_active_cores > nCorePerTile)
701  nCorePerTile = n_active_cores; // calc maximum
702  }
703  }
704  if (n_active_tiles) { // were there any active tiles on the socket?
705  ++nPackages; // count total active packages
706  if (n_active_tiles > nTilePerPkg)
707  nTilePerPkg = n_active_tiles; // calc maximum
708  }
709  } else {
710  // no NUMA, no tiles
711  int n_active_cores = 0;
712  __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads, n_active_cores,
713  socket, 1, labels);
714  if (n_active_cores) { // were there any active cores on the socket?
715  ++nPackages; // count total active packages
716  if (n_active_cores > nCoresPerPkg)
717  nCoresPerPkg = n_active_cores; // calc maximum
718  }
719  }
720  }
721  }
722 
723  // If there's only one thread context to bind to, return now.
724  KMP_DEBUG_ASSERT(nActiveThreads == __kmp_avail_proc);
725  KMP_ASSERT(nActiveThreads > 0);
726  if (nActiveThreads == 1) {
727  __kmp_ncores = nPackages = 1;
728  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
729  if (__kmp_affinity_verbose) {
730  char buf[KMP_AFFIN_MASK_PRINT_LEN];
731  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
732 
733  KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
734  if (__kmp_affinity_respect_mask) {
735  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
736  } else {
737  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
738  }
739  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
740  KMP_INFORM(Uniform, "KMP_AFFINITY");
741  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
742  __kmp_nThreadsPerCore, __kmp_ncores);
743  }
744 
745  if (__kmp_affinity_type == affinity_none) {
746  __kmp_free(retval);
747  KMP_CPU_FREE(oldMask);
748  return 0;
749  }
750 
751  // Form an Address object which only includes the package level.
752  Address addr(1);
753  addr.labels[0] = retval[0].first.labels[0];
754  retval[0].first = addr;
755 
756  if (__kmp_affinity_gran_levels < 0) {
757  __kmp_affinity_gran_levels = 0;
758  }
759 
760  if (__kmp_affinity_verbose) {
761  __kmp_affinity_print_topology(retval, 1, 1, 0, -1, -1);
762  }
763 
764  *address2os = retval;
765  KMP_CPU_FREE(oldMask);
766  return 1;
767  }
768 
769  // Sort the table by physical Id.
770  qsort(retval, nActiveThreads, sizeof(*retval),
771  __kmp_affinity_cmp_Address_labels);
772 
773  // Check to see if the machine topology is uniform
774  int nPUs = nPackages * __kmp_nThreadsPerCore;
775  if (__kmp_numa_detected) {
776  if (__kmp_tile_depth) { // NUMA + tiles
777  nPUs *= (nNodePerPkg * nTilePerNode * nCorePerTile);
778  } else { // NUMA, no tiles
779  nPUs *= (nNodePerPkg * nCorePerNode);
780  }
781  } else {
782  if (__kmp_tile_depth) { // no NUMA, tiles
783  nPUs *= (nTilePerPkg * nCorePerTile);
784  } else { // no NUMA, no tiles
785  nPUs *= nCoresPerPkg;
786  }
787  }
788  unsigned uniform = (nPUs == nActiveThreads);
789 
790  // Print the machine topology summary.
791  if (__kmp_affinity_verbose) {
792  char mask[KMP_AFFIN_MASK_PRINT_LEN];
793  __kmp_affinity_print_mask(mask, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
794  if (__kmp_affinity_respect_mask) {
795  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", mask);
796  } else {
797  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", mask);
798  }
799  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
800  if (uniform) {
801  KMP_INFORM(Uniform, "KMP_AFFINITY");
802  } else {
803  KMP_INFORM(NonUniform, "KMP_AFFINITY");
804  }
805  if (__kmp_numa_detected) {
806  if (__kmp_tile_depth) { // NUMA + tiles
807  KMP_INFORM(TopologyExtraNoTi, "KMP_AFFINITY", nPackages, nNodePerPkg,
808  nTilePerNode, nCorePerTile, __kmp_nThreadsPerCore,
809  __kmp_ncores);
810  } else { // NUMA, no tiles
811  KMP_INFORM(TopologyExtraNode, "KMP_AFFINITY", nPackages, nNodePerPkg,
812  nCorePerNode, __kmp_nThreadsPerCore, __kmp_ncores);
813  nPUs *= (nNodePerPkg * nCorePerNode);
814  }
815  } else {
816  if (__kmp_tile_depth) { // no NUMA, tiles
817  KMP_INFORM(TopologyExtraTile, "KMP_AFFINITY", nPackages, nTilePerPkg,
818  nCorePerTile, __kmp_nThreadsPerCore, __kmp_ncores);
819  } else { // no NUMA, no tiles
820  kmp_str_buf_t buf;
821  __kmp_str_buf_init(&buf);
822  __kmp_str_buf_print(&buf, "%d", nPackages);
823  KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, nCoresPerPkg,
824  __kmp_nThreadsPerCore, __kmp_ncores);
825  __kmp_str_buf_free(&buf);
826  }
827  }
828  }
829 
830  if (__kmp_affinity_type == affinity_none) {
831  __kmp_free(retval);
832  KMP_CPU_FREE(oldMask);
833  return 0;
834  }
835 
836  int depth_full = depth; // number of levels before compressing
837  // Find any levels with radix 1, and remove them from the map
838  // (except for the package level).
839  depth = __kmp_affinity_remove_radix_one_levels(retval, nActiveThreads, depth,
840  levels);
841  KMP_DEBUG_ASSERT(__kmp_affinity_gran != affinity_gran_default);
842  if (__kmp_affinity_gran_levels < 0) {
843  // Set the granularity level based on what levels are modeled
844  // in the machine topology map.
845  __kmp_affinity_gran_levels = 0; // lowest level (e.g. fine)
846  if (__kmp_affinity_gran > affinity_gran_thread) {
847  for (int i = 1; i <= depth_full; ++i) {
848  if (__kmp_affinity_gran <= i) // only count deeper levels
849  break;
850  if (levels[depth_full - i] > 0)
851  __kmp_affinity_gran_levels++;
852  }
853  }
854  if (__kmp_affinity_gran > affinity_gran_package)
855  __kmp_affinity_gran_levels++; // e.g. granularity = group
856  }
857 
858  if (__kmp_affinity_verbose)
859  __kmp_affinity_print_hwloc_tp(retval, nActiveThreads, depth, levels);
860 
861  KMP_CPU_FREE(oldMask);
862  *address2os = retval;
863  return depth;
864 }
865 #endif // KMP_USE_HWLOC
866 
867 // If we don't know how to retrieve the machine's processor topology, or
868 // encounter an error in doing so, this routine is called to form a "flat"
869 // mapping of os thread id's <-> processor id's.
870 static int __kmp_affinity_create_flat_map(AddrUnsPair **address2os,
871  kmp_i18n_id_t *const msg_id) {
872  *address2os = NULL;
873  *msg_id = kmp_i18n_null;
874 
875  // Even if __kmp_affinity_type == affinity_none, this routine might still
876  // called to set __kmp_ncores, as well as
877  // __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
878  if (!KMP_AFFINITY_CAPABLE()) {
879  KMP_ASSERT(__kmp_affinity_type == affinity_none);
880  __kmp_ncores = nPackages = __kmp_xproc;
881  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
882  if (__kmp_affinity_verbose) {
883  KMP_INFORM(AffFlatTopology, "KMP_AFFINITY");
884  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
885  KMP_INFORM(Uniform, "KMP_AFFINITY");
886  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
887  __kmp_nThreadsPerCore, __kmp_ncores);
888  }
889  return 0;
890  }
891 
892  // When affinity is off, this routine will still be called to set
893  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
894  // Make sure all these vars are set correctly, and return now if affinity is
895  // not enabled.
896  __kmp_ncores = nPackages = __kmp_avail_proc;
897  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
898  if (__kmp_affinity_verbose) {
899  char buf[KMP_AFFIN_MASK_PRINT_LEN];
900  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
901  __kmp_affin_fullMask);
902 
903  KMP_INFORM(AffCapableUseFlat, "KMP_AFFINITY");
904  if (__kmp_affinity_respect_mask) {
905  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
906  } else {
907  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
908  }
909  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
910  KMP_INFORM(Uniform, "KMP_AFFINITY");
911  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
912  __kmp_nThreadsPerCore, __kmp_ncores);
913  }
914  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
915  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
916  if (__kmp_affinity_type == affinity_none) {
917  int avail_ct = 0;
918  int i;
919  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
920  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask))
921  continue;
922  __kmp_pu_os_idx[avail_ct++] = i; // suppose indices are flat
923  }
924  return 0;
925  }
926 
927  // Construct the data structure to be returned.
928  *address2os =
929  (AddrUnsPair *)__kmp_allocate(sizeof(**address2os) * __kmp_avail_proc);
930  int avail_ct = 0;
931  int i;
932  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
933  // Skip this proc if it is not included in the machine model.
934  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
935  continue;
936  }
937  __kmp_pu_os_idx[avail_ct] = i; // suppose indices are flat
938  Address addr(1);
939  addr.labels[0] = i;
940  (*address2os)[avail_ct++] = AddrUnsPair(addr, i);
941  }
942  if (__kmp_affinity_verbose) {
943  KMP_INFORM(OSProcToPackage, "KMP_AFFINITY");
944  }
945 
946  if (__kmp_affinity_gran_levels < 0) {
947  // Only the package level is modeled in the machine topology map,
948  // so the #levels of granularity is either 0 or 1.
949  if (__kmp_affinity_gran > affinity_gran_package) {
950  __kmp_affinity_gran_levels = 1;
951  } else {
952  __kmp_affinity_gran_levels = 0;
953  }
954  }
955  return 1;
956 }
957 
958 #if KMP_GROUP_AFFINITY
959 
960 // If multiple Windows* OS processor groups exist, we can create a 2-level
961 // topology map with the groups at level 0 and the individual procs at level 1.
962 // This facilitates letting the threads float among all procs in a group,
963 // if granularity=group (the default when there are multiple groups).
964 static int __kmp_affinity_create_proc_group_map(AddrUnsPair **address2os,
965  kmp_i18n_id_t *const msg_id) {
966  *address2os = NULL;
967  *msg_id = kmp_i18n_null;
968 
969  // If we aren't affinity capable, then return now.
970  // The flat mapping will be used.
971  if (!KMP_AFFINITY_CAPABLE()) {
972  // FIXME set *msg_id
973  return -1;
974  }
975 
976  // Construct the data structure to be returned.
977  *address2os =
978  (AddrUnsPair *)__kmp_allocate(sizeof(**address2os) * __kmp_avail_proc);
979  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
980  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
981  int avail_ct = 0;
982  int i;
983  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
984  // Skip this proc if it is not included in the machine model.
985  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
986  continue;
987  }
988  __kmp_pu_os_idx[avail_ct] = i; // suppose indices are flat
989  Address addr(2);
990  addr.labels[0] = i / (CHAR_BIT * sizeof(DWORD_PTR));
991  addr.labels[1] = i % (CHAR_BIT * sizeof(DWORD_PTR));
992  (*address2os)[avail_ct++] = AddrUnsPair(addr, i);
993 
994  if (__kmp_affinity_verbose) {
995  KMP_INFORM(AffOSProcToGroup, "KMP_AFFINITY", i, addr.labels[0],
996  addr.labels[1]);
997  }
998  }
999 
1000  if (__kmp_affinity_gran_levels < 0) {
1001  if (__kmp_affinity_gran == affinity_gran_group) {
1002  __kmp_affinity_gran_levels = 1;
1003  } else if ((__kmp_affinity_gran == affinity_gran_fine) ||
1004  (__kmp_affinity_gran == affinity_gran_thread)) {
1005  __kmp_affinity_gran_levels = 0;
1006  } else {
1007  const char *gran_str = NULL;
1008  if (__kmp_affinity_gran == affinity_gran_core) {
1009  gran_str = "core";
1010  } else if (__kmp_affinity_gran == affinity_gran_package) {
1011  gran_str = "package";
1012  } else if (__kmp_affinity_gran == affinity_gran_node) {
1013  gran_str = "node";
1014  } else {
1015  KMP_ASSERT(0);
1016  }
1017 
1018  // Warning: can't use affinity granularity \"gran\" with group topology
1019  // method, using "thread"
1020  __kmp_affinity_gran_levels = 0;
1021  }
1022  }
1023  return 2;
1024 }
1025 
1026 #endif /* KMP_GROUP_AFFINITY */
1027 
1028 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1029 
1030 static int __kmp_cpuid_mask_width(int count) {
1031  int r = 0;
1032 
1033  while ((1 << r) < count)
1034  ++r;
1035  return r;
1036 }
1037 
1038 class apicThreadInfo {
1039 public:
1040  unsigned osId; // param to __kmp_affinity_bind_thread
1041  unsigned apicId; // from cpuid after binding
1042  unsigned maxCoresPerPkg; // ""
1043  unsigned maxThreadsPerPkg; // ""
1044  unsigned pkgId; // inferred from above values
1045  unsigned coreId; // ""
1046  unsigned threadId; // ""
1047 };
1048 
1049 static int __kmp_affinity_cmp_apicThreadInfo_phys_id(const void *a,
1050  const void *b) {
1051  const apicThreadInfo *aa = (const apicThreadInfo *)a;
1052  const apicThreadInfo *bb = (const apicThreadInfo *)b;
1053  if (aa->pkgId < bb->pkgId)
1054  return -1;
1055  if (aa->pkgId > bb->pkgId)
1056  return 1;
1057  if (aa->coreId < bb->coreId)
1058  return -1;
1059  if (aa->coreId > bb->coreId)
1060  return 1;
1061  if (aa->threadId < bb->threadId)
1062  return -1;
1063  if (aa->threadId > bb->threadId)
1064  return 1;
1065  return 0;
1066 }
1067 
1068 // On IA-32 architecture and Intel(R) 64 architecture, we attempt to use
1069 // an algorithm which cycles through the available os threads, setting
1070 // the current thread's affinity mask to that thread, and then retrieves
1071 // the Apic Id for each thread context using the cpuid instruction.
1072 static int __kmp_affinity_create_apicid_map(AddrUnsPair **address2os,
1073  kmp_i18n_id_t *const msg_id) {
1074  kmp_cpuid buf;
1075  *address2os = NULL;
1076  *msg_id = kmp_i18n_null;
1077 
1078  // Check if cpuid leaf 4 is supported.
1079  __kmp_x86_cpuid(0, 0, &buf);
1080  if (buf.eax < 4) {
1081  *msg_id = kmp_i18n_str_NoLeaf4Support;
1082  return -1;
1083  }
1084 
1085  // The algorithm used starts by setting the affinity to each available thread
1086  // and retrieving info from the cpuid instruction, so if we are not capable of
1087  // calling __kmp_get_system_affinity() and _kmp_get_system_affinity(), then we
1088  // need to do something else - use the defaults that we calculated from
1089  // issuing cpuid without binding to each proc.
1090  if (!KMP_AFFINITY_CAPABLE()) {
1091  // Hack to try and infer the machine topology using only the data
1092  // available from cpuid on the current thread, and __kmp_xproc.
1093  KMP_ASSERT(__kmp_affinity_type == affinity_none);
1094 
1095  // Get an upper bound on the number of threads per package using cpuid(1).
1096  // On some OS/chps combinations where HT is supported by the chip but is
1097  // disabled, this value will be 2 on a single core chip. Usually, it will be
1098  // 2 if HT is enabled and 1 if HT is disabled.
1099  __kmp_x86_cpuid(1, 0, &buf);
1100  int maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
1101  if (maxThreadsPerPkg == 0) {
1102  maxThreadsPerPkg = 1;
1103  }
1104 
1105  // The num cores per pkg comes from cpuid(4). 1 must be added to the encoded
1106  // value.
1107  //
1108  // The author of cpu_count.cpp treated this only an upper bound on the
1109  // number of cores, but I haven't seen any cases where it was greater than
1110  // the actual number of cores, so we will treat it as exact in this block of
1111  // code.
1112  //
1113  // First, we need to check if cpuid(4) is supported on this chip. To see if
1114  // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n or
1115  // greater.
1116  __kmp_x86_cpuid(0, 0, &buf);
1117  if (buf.eax >= 4) {
1118  __kmp_x86_cpuid(4, 0, &buf);
1119  nCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
1120  } else {
1121  nCoresPerPkg = 1;
1122  }
1123 
1124  // There is no way to reliably tell if HT is enabled without issuing the
1125  // cpuid instruction from every thread, can correlating the cpuid info, so
1126  // if the machine is not affinity capable, we assume that HT is off. We have
1127  // seen quite a few machines where maxThreadsPerPkg is 2, yet the machine
1128  // does not support HT.
1129  //
1130  // - Older OSes are usually found on machines with older chips, which do not
1131  // support HT.
1132  // - The performance penalty for mistakenly identifying a machine as HT when
1133  // it isn't (which results in blocktime being incorrectly set to 0) is
1134  // greater than the penalty when for mistakenly identifying a machine as
1135  // being 1 thread/core when it is really HT enabled (which results in
1136  // blocktime being incorrectly set to a positive value).
1137  __kmp_ncores = __kmp_xproc;
1138  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
1139  __kmp_nThreadsPerCore = 1;
1140  if (__kmp_affinity_verbose) {
1141  KMP_INFORM(AffNotCapableUseLocCpuid, "KMP_AFFINITY");
1142  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1143  if (__kmp_affinity_uniform_topology()) {
1144  KMP_INFORM(Uniform, "KMP_AFFINITY");
1145  } else {
1146  KMP_INFORM(NonUniform, "KMP_AFFINITY");
1147  }
1148  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
1149  __kmp_nThreadsPerCore, __kmp_ncores);
1150  }
1151  return 0;
1152  }
1153 
1154  // From here on, we can assume that it is safe to call
1155  // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
1156  // __kmp_affinity_type = affinity_none.
1157 
1158  // Save the affinity mask for the current thread.
1159  kmp_affin_mask_t *oldMask;
1160  KMP_CPU_ALLOC(oldMask);
1161  KMP_ASSERT(oldMask != NULL);
1162  __kmp_get_system_affinity(oldMask, TRUE);
1163 
1164  // Run through each of the available contexts, binding the current thread
1165  // to it, and obtaining the pertinent information using the cpuid instr.
1166  //
1167  // The relevant information is:
1168  // - Apic Id: Bits 24:31 of ebx after issuing cpuid(1) - each thread context
1169  // has a uniqie Apic Id, which is of the form pkg# : core# : thread#.
1170  // - Max Threads Per Pkg: Bits 16:23 of ebx after issuing cpuid(1). The value
1171  // of this field determines the width of the core# + thread# fields in the
1172  // Apic Id. It is also an upper bound on the number of threads per
1173  // package, but it has been verified that situations happen were it is not
1174  // exact. In particular, on certain OS/chip combinations where Intel(R)
1175  // Hyper-Threading Technology is supported by the chip but has been
1176  // disabled, the value of this field will be 2 (for a single core chip).
1177  // On other OS/chip combinations supporting Intel(R) Hyper-Threading
1178  // Technology, the value of this field will be 1 when Intel(R)
1179  // Hyper-Threading Technology is disabled and 2 when it is enabled.
1180  // - Max Cores Per Pkg: Bits 26:31 of eax after issuing cpuid(4). The value
1181  // of this field (+1) determines the width of the core# field in the Apic
1182  // Id. The comments in "cpucount.cpp" say that this value is an upper
1183  // bound, but the IA-32 architecture manual says that it is exactly the
1184  // number of cores per package, and I haven't seen any case where it
1185  // wasn't.
1186  //
1187  // From this information, deduce the package Id, core Id, and thread Id,
1188  // and set the corresponding fields in the apicThreadInfo struct.
1189  unsigned i;
1190  apicThreadInfo *threadInfo = (apicThreadInfo *)__kmp_allocate(
1191  __kmp_avail_proc * sizeof(apicThreadInfo));
1192  unsigned nApics = 0;
1193  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
1194  // Skip this proc if it is not included in the machine model.
1195  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
1196  continue;
1197  }
1198  KMP_DEBUG_ASSERT((int)nApics < __kmp_avail_proc);
1199 
1200  __kmp_affinity_dispatch->bind_thread(i);
1201  threadInfo[nApics].osId = i;
1202 
1203  // The apic id and max threads per pkg come from cpuid(1).
1204  __kmp_x86_cpuid(1, 0, &buf);
1205  if (((buf.edx >> 9) & 1) == 0) {
1206  __kmp_set_system_affinity(oldMask, TRUE);
1207  __kmp_free(threadInfo);
1208  KMP_CPU_FREE(oldMask);
1209  *msg_id = kmp_i18n_str_ApicNotPresent;
1210  return -1;
1211  }
1212  threadInfo[nApics].apicId = (buf.ebx >> 24) & 0xff;
1213  threadInfo[nApics].maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
1214  if (threadInfo[nApics].maxThreadsPerPkg == 0) {
1215  threadInfo[nApics].maxThreadsPerPkg = 1;
1216  }
1217 
1218  // Max cores per pkg comes from cpuid(4). 1 must be added to the encoded
1219  // value.
1220  //
1221  // First, we need to check if cpuid(4) is supported on this chip. To see if
1222  // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n
1223  // or greater.
1224  __kmp_x86_cpuid(0, 0, &buf);
1225  if (buf.eax >= 4) {
1226  __kmp_x86_cpuid(4, 0, &buf);
1227  threadInfo[nApics].maxCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
1228  } else {
1229  threadInfo[nApics].maxCoresPerPkg = 1;
1230  }
1231 
1232  // Infer the pkgId / coreId / threadId using only the info obtained locally.
1233  int widthCT = __kmp_cpuid_mask_width(threadInfo[nApics].maxThreadsPerPkg);
1234  threadInfo[nApics].pkgId = threadInfo[nApics].apicId >> widthCT;
1235 
1236  int widthC = __kmp_cpuid_mask_width(threadInfo[nApics].maxCoresPerPkg);
1237  int widthT = widthCT - widthC;
1238  if (widthT < 0) {
1239  // I've never seen this one happen, but I suppose it could, if the cpuid
1240  // instruction on a chip was really screwed up. Make sure to restore the
1241  // affinity mask before the tail call.
1242  __kmp_set_system_affinity(oldMask, TRUE);
1243  __kmp_free(threadInfo);
1244  KMP_CPU_FREE(oldMask);
1245  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1246  return -1;
1247  }
1248 
1249  int maskC = (1 << widthC) - 1;
1250  threadInfo[nApics].coreId = (threadInfo[nApics].apicId >> widthT) & maskC;
1251 
1252  int maskT = (1 << widthT) - 1;
1253  threadInfo[nApics].threadId = threadInfo[nApics].apicId & maskT;
1254 
1255  nApics++;
1256  }
1257 
1258  // We've collected all the info we need.
1259  // Restore the old affinity mask for this thread.
1260  __kmp_set_system_affinity(oldMask, TRUE);
1261 
1262  // If there's only one thread context to bind to, form an Address object
1263  // with depth 1 and return immediately (or, if affinity is off, set
1264  // address2os to NULL and return).
1265  //
1266  // If it is configured to omit the package level when there is only a single
1267  // package, the logic at the end of this routine won't work if there is only
1268  // a single thread - it would try to form an Address object with depth 0.
1269  KMP_ASSERT(nApics > 0);
1270  if (nApics == 1) {
1271  __kmp_ncores = nPackages = 1;
1272  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
1273  if (__kmp_affinity_verbose) {
1274  char buf[KMP_AFFIN_MASK_PRINT_LEN];
1275  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
1276 
1277  KMP_INFORM(AffUseGlobCpuid, "KMP_AFFINITY");
1278  if (__kmp_affinity_respect_mask) {
1279  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
1280  } else {
1281  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
1282  }
1283  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1284  KMP_INFORM(Uniform, "KMP_AFFINITY");
1285  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
1286  __kmp_nThreadsPerCore, __kmp_ncores);
1287  }
1288 
1289  if (__kmp_affinity_type == affinity_none) {
1290  __kmp_free(threadInfo);
1291  KMP_CPU_FREE(oldMask);
1292  return 0;
1293  }
1294 
1295  *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair));
1296  Address addr(1);
1297  addr.labels[0] = threadInfo[0].pkgId;
1298  (*address2os)[0] = AddrUnsPair(addr, threadInfo[0].osId);
1299 
1300  if (__kmp_affinity_gran_levels < 0) {
1301  __kmp_affinity_gran_levels = 0;
1302  }
1303 
1304  if (__kmp_affinity_verbose) {
1305  __kmp_affinity_print_topology(*address2os, 1, 1, 0, -1, -1);
1306  }
1307 
1308  __kmp_free(threadInfo);
1309  KMP_CPU_FREE(oldMask);
1310  return 1;
1311  }
1312 
1313  // Sort the threadInfo table by physical Id.
1314  qsort(threadInfo, nApics, sizeof(*threadInfo),
1315  __kmp_affinity_cmp_apicThreadInfo_phys_id);
1316 
1317  // The table is now sorted by pkgId / coreId / threadId, but we really don't
1318  // know the radix of any of the fields. pkgId's may be sparsely assigned among
1319  // the chips on a system. Although coreId's are usually assigned
1320  // [0 .. coresPerPkg-1] and threadId's are usually assigned
1321  // [0..threadsPerCore-1], we don't want to make any such assumptions.
1322  //
1323  // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
1324  // total # packages) are at this point - we want to determine that now. We
1325  // only have an upper bound on the first two figures.
1326  //
1327  // We also perform a consistency check at this point: the values returned by
1328  // the cpuid instruction for any thread bound to a given package had better
1329  // return the same info for maxThreadsPerPkg and maxCoresPerPkg.
1330  nPackages = 1;
1331  nCoresPerPkg = 1;
1332  __kmp_nThreadsPerCore = 1;
1333  unsigned nCores = 1;
1334 
1335  unsigned pkgCt = 1; // to determine radii
1336  unsigned lastPkgId = threadInfo[0].pkgId;
1337  unsigned coreCt = 1;
1338  unsigned lastCoreId = threadInfo[0].coreId;
1339  unsigned threadCt = 1;
1340  unsigned lastThreadId = threadInfo[0].threadId;
1341 
1342  // intra-pkg consist checks
1343  unsigned prevMaxCoresPerPkg = threadInfo[0].maxCoresPerPkg;
1344  unsigned prevMaxThreadsPerPkg = threadInfo[0].maxThreadsPerPkg;
1345 
1346  for (i = 1; i < nApics; i++) {
1347  if (threadInfo[i].pkgId != lastPkgId) {
1348  nCores++;
1349  pkgCt++;
1350  lastPkgId = threadInfo[i].pkgId;
1351  if ((int)coreCt > nCoresPerPkg)
1352  nCoresPerPkg = coreCt;
1353  coreCt = 1;
1354  lastCoreId = threadInfo[i].coreId;
1355  if ((int)threadCt > __kmp_nThreadsPerCore)
1356  __kmp_nThreadsPerCore = threadCt;
1357  threadCt = 1;
1358  lastThreadId = threadInfo[i].threadId;
1359 
1360  // This is a different package, so go on to the next iteration without
1361  // doing any consistency checks. Reset the consistency check vars, though.
1362  prevMaxCoresPerPkg = threadInfo[i].maxCoresPerPkg;
1363  prevMaxThreadsPerPkg = threadInfo[i].maxThreadsPerPkg;
1364  continue;
1365  }
1366 
1367  if (threadInfo[i].coreId != lastCoreId) {
1368  nCores++;
1369  coreCt++;
1370  lastCoreId = threadInfo[i].coreId;
1371  if ((int)threadCt > __kmp_nThreadsPerCore)
1372  __kmp_nThreadsPerCore = threadCt;
1373  threadCt = 1;
1374  lastThreadId = threadInfo[i].threadId;
1375  } else if (threadInfo[i].threadId != lastThreadId) {
1376  threadCt++;
1377  lastThreadId = threadInfo[i].threadId;
1378  } else {
1379  __kmp_free(threadInfo);
1380  KMP_CPU_FREE(oldMask);
1381  *msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
1382  return -1;
1383  }
1384 
1385  // Check to make certain that the maxCoresPerPkg and maxThreadsPerPkg
1386  // fields agree between all the threads bounds to a given package.
1387  if ((prevMaxCoresPerPkg != threadInfo[i].maxCoresPerPkg) ||
1388  (prevMaxThreadsPerPkg != threadInfo[i].maxThreadsPerPkg)) {
1389  __kmp_free(threadInfo);
1390  KMP_CPU_FREE(oldMask);
1391  *msg_id = kmp_i18n_str_InconsistentCpuidInfo;
1392  return -1;
1393  }
1394  }
1395  nPackages = pkgCt;
1396  if ((int)coreCt > nCoresPerPkg)
1397  nCoresPerPkg = coreCt;
1398  if ((int)threadCt > __kmp_nThreadsPerCore)
1399  __kmp_nThreadsPerCore = threadCt;
1400 
1401  // When affinity is off, this routine will still be called to set
1402  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
1403  // Make sure all these vars are set correctly, and return now if affinity is
1404  // not enabled.
1405  __kmp_ncores = nCores;
1406  if (__kmp_affinity_verbose) {
1407  char buf[KMP_AFFIN_MASK_PRINT_LEN];
1408  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
1409 
1410  KMP_INFORM(AffUseGlobCpuid, "KMP_AFFINITY");
1411  if (__kmp_affinity_respect_mask) {
1412  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
1413  } else {
1414  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
1415  }
1416  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1417  if (__kmp_affinity_uniform_topology()) {
1418  KMP_INFORM(Uniform, "KMP_AFFINITY");
1419  } else {
1420  KMP_INFORM(NonUniform, "KMP_AFFINITY");
1421  }
1422  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
1423  __kmp_nThreadsPerCore, __kmp_ncores);
1424  }
1425  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
1426  KMP_DEBUG_ASSERT(nApics == (unsigned)__kmp_avail_proc);
1427  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
1428  for (i = 0; i < nApics; ++i) {
1429  __kmp_pu_os_idx[i] = threadInfo[i].osId;
1430  }
1431  if (__kmp_affinity_type == affinity_none) {
1432  __kmp_free(threadInfo);
1433  KMP_CPU_FREE(oldMask);
1434  return 0;
1435  }
1436 
1437  // Now that we've determined the number of packages, the number of cores per
1438  // package, and the number of threads per core, we can construct the data
1439  // structure that is to be returned.
1440  int pkgLevel = 0;
1441  int coreLevel = (nCoresPerPkg <= 1) ? -1 : 1;
1442  int threadLevel =
1443  (__kmp_nThreadsPerCore <= 1) ? -1 : ((coreLevel >= 0) ? 2 : 1);
1444  unsigned depth = (pkgLevel >= 0) + (coreLevel >= 0) + (threadLevel >= 0);
1445 
1446  KMP_ASSERT(depth > 0);
1447  *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * nApics);
1448 
1449  for (i = 0; i < nApics; ++i) {
1450  Address addr(depth);
1451  unsigned os = threadInfo[i].osId;
1452  int d = 0;
1453 
1454  if (pkgLevel >= 0) {
1455  addr.labels[d++] = threadInfo[i].pkgId;
1456  }
1457  if (coreLevel >= 0) {
1458  addr.labels[d++] = threadInfo[i].coreId;
1459  }
1460  if (threadLevel >= 0) {
1461  addr.labels[d++] = threadInfo[i].threadId;
1462  }
1463  (*address2os)[i] = AddrUnsPair(addr, os);
1464  }
1465 
1466  if (__kmp_affinity_gran_levels < 0) {
1467  // Set the granularity level based on what levels are modeled in the machine
1468  // topology map.
1469  __kmp_affinity_gran_levels = 0;
1470  if ((threadLevel >= 0) && (__kmp_affinity_gran > affinity_gran_thread)) {
1471  __kmp_affinity_gran_levels++;
1472  }
1473  if ((coreLevel >= 0) && (__kmp_affinity_gran > affinity_gran_core)) {
1474  __kmp_affinity_gran_levels++;
1475  }
1476  if ((pkgLevel >= 0) && (__kmp_affinity_gran > affinity_gran_package)) {
1477  __kmp_affinity_gran_levels++;
1478  }
1479  }
1480 
1481  if (__kmp_affinity_verbose) {
1482  __kmp_affinity_print_topology(*address2os, nApics, depth, pkgLevel,
1483  coreLevel, threadLevel);
1484  }
1485 
1486  __kmp_free(threadInfo);
1487  KMP_CPU_FREE(oldMask);
1488  return depth;
1489 }
1490 
1491 // Intel(R) microarchitecture code name Nehalem, Dunnington and later
1492 // architectures support a newer interface for specifying the x2APIC Ids,
1493 // based on cpuid leaf 11.
1494 static int __kmp_affinity_create_x2apicid_map(AddrUnsPair **address2os,
1495  kmp_i18n_id_t *const msg_id) {
1496  kmp_cpuid buf;
1497  *address2os = NULL;
1498  *msg_id = kmp_i18n_null;
1499 
1500  // Check to see if cpuid leaf 11 is supported.
1501  __kmp_x86_cpuid(0, 0, &buf);
1502  if (buf.eax < 11) {
1503  *msg_id = kmp_i18n_str_NoLeaf11Support;
1504  return -1;
1505  }
1506  __kmp_x86_cpuid(11, 0, &buf);
1507  if (buf.ebx == 0) {
1508  *msg_id = kmp_i18n_str_NoLeaf11Support;
1509  return -1;
1510  }
1511 
1512  // Find the number of levels in the machine topology. While we're at it, get
1513  // the default values for __kmp_nThreadsPerCore & nCoresPerPkg. We will try to
1514  // get more accurate values later by explicitly counting them, but get
1515  // reasonable defaults now, in case we return early.
1516  int level;
1517  int threadLevel = -1;
1518  int coreLevel = -1;
1519  int pkgLevel = -1;
1520  __kmp_nThreadsPerCore = nCoresPerPkg = nPackages = 1;
1521 
1522  for (level = 0;; level++) {
1523  if (level > 31) {
1524  // FIXME: Hack for DPD200163180
1525  //
1526  // If level is big then something went wrong -> exiting
1527  //
1528  // There could actually be 32 valid levels in the machine topology, but so
1529  // far, the only machine we have seen which does not exit this loop before
1530  // iteration 32 has fubar x2APIC settings.
1531  //
1532  // For now, just reject this case based upon loop trip count.
1533  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1534  return -1;
1535  }
1536  __kmp_x86_cpuid(11, level, &buf);
1537  if (buf.ebx == 0) {
1538  if (pkgLevel < 0) {
1539  // Will infer nPackages from __kmp_xproc
1540  pkgLevel = level;
1541  level++;
1542  }
1543  break;
1544  }
1545  int kind = (buf.ecx >> 8) & 0xff;
1546  if (kind == 1) {
1547  // SMT level
1548  threadLevel = level;
1549  coreLevel = -1;
1550  pkgLevel = -1;
1551  __kmp_nThreadsPerCore = buf.ebx & 0xffff;
1552  if (__kmp_nThreadsPerCore == 0) {
1553  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1554  return -1;
1555  }
1556  } else if (kind == 2) {
1557  // core level
1558  coreLevel = level;
1559  pkgLevel = -1;
1560  nCoresPerPkg = buf.ebx & 0xffff;
1561  if (nCoresPerPkg == 0) {
1562  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1563  return -1;
1564  }
1565  } else {
1566  if (level <= 0) {
1567  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1568  return -1;
1569  }
1570  if (pkgLevel >= 0) {
1571  continue;
1572  }
1573  pkgLevel = level;
1574  nPackages = buf.ebx & 0xffff;
1575  if (nPackages == 0) {
1576  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1577  return -1;
1578  }
1579  }
1580  }
1581  int depth = level;
1582 
1583  // In the above loop, "level" was counted from the finest level (usually
1584  // thread) to the coarsest. The caller expects that we will place the labels
1585  // in (*address2os)[].first.labels[] in the inverse order, so we need to
1586  // invert the vars saying which level means what.
1587  if (threadLevel >= 0) {
1588  threadLevel = depth - threadLevel - 1;
1589  }
1590  if (coreLevel >= 0) {
1591  coreLevel = depth - coreLevel - 1;
1592  }
1593  KMP_DEBUG_ASSERT(pkgLevel >= 0);
1594  pkgLevel = depth - pkgLevel - 1;
1595 
1596  // The algorithm used starts by setting the affinity to each available thread
1597  // and retrieving info from the cpuid instruction, so if we are not capable of
1598  // calling __kmp_get_system_affinity() and _kmp_get_system_affinity(), then we
1599  // need to do something else - use the defaults that we calculated from
1600  // issuing cpuid without binding to each proc.
1601  if (!KMP_AFFINITY_CAPABLE()) {
1602  // Hack to try and infer the machine topology using only the data
1603  // available from cpuid on the current thread, and __kmp_xproc.
1604  KMP_ASSERT(__kmp_affinity_type == affinity_none);
1605 
1606  __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
1607  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
1608  if (__kmp_affinity_verbose) {
1609  KMP_INFORM(AffNotCapableUseLocCpuidL11, "KMP_AFFINITY");
1610  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1611  if (__kmp_affinity_uniform_topology()) {
1612  KMP_INFORM(Uniform, "KMP_AFFINITY");
1613  } else {
1614  KMP_INFORM(NonUniform, "KMP_AFFINITY");
1615  }
1616  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
1617  __kmp_nThreadsPerCore, __kmp_ncores);
1618  }
1619  return 0;
1620  }
1621 
1622  // From here on, we can assume that it is safe to call
1623  // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
1624  // __kmp_affinity_type = affinity_none.
1625 
1626  // Save the affinity mask for the current thread.
1627  kmp_affin_mask_t *oldMask;
1628  KMP_CPU_ALLOC(oldMask);
1629  __kmp_get_system_affinity(oldMask, TRUE);
1630 
1631  // Allocate the data structure to be returned.
1632  AddrUnsPair *retval =
1633  (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * __kmp_avail_proc);
1634 
1635  // Run through each of the available contexts, binding the current thread
1636  // to it, and obtaining the pertinent information using the cpuid instr.
1637  unsigned int proc;
1638  int nApics = 0;
1639  KMP_CPU_SET_ITERATE(proc, __kmp_affin_fullMask) {
1640  // Skip this proc if it is not included in the machine model.
1641  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
1642  continue;
1643  }
1644  KMP_DEBUG_ASSERT(nApics < __kmp_avail_proc);
1645 
1646  __kmp_affinity_dispatch->bind_thread(proc);
1647 
1648  // Extract labels for each level in the machine topology map from Apic ID.
1649  Address addr(depth);
1650  int prev_shift = 0;
1651 
1652  for (level = 0; level < depth; level++) {
1653  __kmp_x86_cpuid(11, level, &buf);
1654  unsigned apicId = buf.edx;
1655  if (buf.ebx == 0) {
1656  if (level != depth - 1) {
1657  KMP_CPU_FREE(oldMask);
1658  *msg_id = kmp_i18n_str_InconsistentCpuidInfo;
1659  return -1;
1660  }
1661  addr.labels[depth - level - 1] = apicId >> prev_shift;
1662  level++;
1663  break;
1664  }
1665  int shift = buf.eax & 0x1f;
1666  int mask = (1 << shift) - 1;
1667  addr.labels[depth - level - 1] = (apicId & mask) >> prev_shift;
1668  prev_shift = shift;
1669  }
1670  if (level != depth) {
1671  KMP_CPU_FREE(oldMask);
1672  *msg_id = kmp_i18n_str_InconsistentCpuidInfo;
1673  return -1;
1674  }
1675 
1676  retval[nApics] = AddrUnsPair(addr, proc);
1677  nApics++;
1678  }
1679 
1680  // We've collected all the info we need.
1681  // Restore the old affinity mask for this thread.
1682  __kmp_set_system_affinity(oldMask, TRUE);
1683 
1684  // If there's only one thread context to bind to, return now.
1685  KMP_ASSERT(nApics > 0);
1686  if (nApics == 1) {
1687  __kmp_ncores = nPackages = 1;
1688  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
1689  if (__kmp_affinity_verbose) {
1690  char buf[KMP_AFFIN_MASK_PRINT_LEN];
1691  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
1692 
1693  KMP_INFORM(AffUseGlobCpuidL11, "KMP_AFFINITY");
1694  if (__kmp_affinity_respect_mask) {
1695  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
1696  } else {
1697  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
1698  }
1699  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1700  KMP_INFORM(Uniform, "KMP_AFFINITY");
1701  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
1702  __kmp_nThreadsPerCore, __kmp_ncores);
1703  }
1704 
1705  if (__kmp_affinity_type == affinity_none) {
1706  __kmp_free(retval);
1707  KMP_CPU_FREE(oldMask);
1708  return 0;
1709  }
1710 
1711  // Form an Address object which only includes the package level.
1712  Address addr(1);
1713  addr.labels[0] = retval[0].first.labels[pkgLevel];
1714  retval[0].first = addr;
1715 
1716  if (__kmp_affinity_gran_levels < 0) {
1717  __kmp_affinity_gran_levels = 0;
1718  }
1719 
1720  if (__kmp_affinity_verbose) {
1721  __kmp_affinity_print_topology(retval, 1, 1, 0, -1, -1);
1722  }
1723 
1724  *address2os = retval;
1725  KMP_CPU_FREE(oldMask);
1726  return 1;
1727  }
1728 
1729  // Sort the table by physical Id.
1730  qsort(retval, nApics, sizeof(*retval), __kmp_affinity_cmp_Address_labels);
1731 
1732  // Find the radix at each of the levels.
1733  unsigned *totals = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
1734  unsigned *counts = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
1735  unsigned *maxCt = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
1736  unsigned *last = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
1737  for (level = 0; level < depth; level++) {
1738  totals[level] = 1;
1739  maxCt[level] = 1;
1740  counts[level] = 1;
1741  last[level] = retval[0].first.labels[level];
1742  }
1743 
1744  // From here on, the iteration variable "level" runs from the finest level to
1745  // the coarsest, i.e. we iterate forward through
1746  // (*address2os)[].first.labels[] - in the previous loops, we iterated
1747  // backwards.
1748  for (proc = 1; (int)proc < nApics; proc++) {
1749  int level;
1750  for (level = 0; level < depth; level++) {
1751  if (retval[proc].first.labels[level] != last[level]) {
1752  int j;
1753  for (j = level + 1; j < depth; j++) {
1754  totals[j]++;
1755  counts[j] = 1;
1756  // The line below causes printing incorrect topology information in
1757  // case the max value for some level (maxCt[level]) is encountered
1758  // earlier than some less value while going through the array. For
1759  // example, let pkg0 has 4 cores and pkg1 has 2 cores. Then
1760  // maxCt[1] == 2
1761  // whereas it must be 4.
1762  // TODO!!! Check if it can be commented safely
1763  // maxCt[j] = 1;
1764  last[j] = retval[proc].first.labels[j];
1765  }
1766  totals[level]++;
1767  counts[level]++;
1768  if (counts[level] > maxCt[level]) {
1769  maxCt[level] = counts[level];
1770  }
1771  last[level] = retval[proc].first.labels[level];
1772  break;
1773  } else if (level == depth - 1) {
1774  __kmp_free(last);
1775  __kmp_free(maxCt);
1776  __kmp_free(counts);
1777  __kmp_free(totals);
1778  __kmp_free(retval);
1779  KMP_CPU_FREE(oldMask);
1780  *msg_id = kmp_i18n_str_x2ApicIDsNotUnique;
1781  return -1;
1782  }
1783  }
1784  }
1785 
1786  // When affinity is off, this routine will still be called to set
1787  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
1788  // Make sure all these vars are set correctly, and return if affinity is not
1789  // enabled.
1790  if (threadLevel >= 0) {
1791  __kmp_nThreadsPerCore = maxCt[threadLevel];
1792  } else {
1793  __kmp_nThreadsPerCore = 1;
1794  }
1795  nPackages = totals[pkgLevel];
1796 
1797  if (coreLevel >= 0) {
1798  __kmp_ncores = totals[coreLevel];
1799  nCoresPerPkg = maxCt[coreLevel];
1800  } else {
1801  __kmp_ncores = nPackages;
1802  nCoresPerPkg = 1;
1803  }
1804 
1805  // Check to see if the machine topology is uniform
1806  unsigned prod = maxCt[0];
1807  for (level = 1; level < depth; level++) {
1808  prod *= maxCt[level];
1809  }
1810  bool uniform = (prod == totals[level - 1]);
1811 
1812  // Print the machine topology summary.
1813  if (__kmp_affinity_verbose) {
1814  char mask[KMP_AFFIN_MASK_PRINT_LEN];
1815  __kmp_affinity_print_mask(mask, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
1816 
1817  KMP_INFORM(AffUseGlobCpuidL11, "KMP_AFFINITY");
1818  if (__kmp_affinity_respect_mask) {
1819  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", mask);
1820  } else {
1821  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", mask);
1822  }
1823  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1824  if (uniform) {
1825  KMP_INFORM(Uniform, "KMP_AFFINITY");
1826  } else {
1827  KMP_INFORM(NonUniform, "KMP_AFFINITY");
1828  }
1829 
1830  kmp_str_buf_t buf;
1831  __kmp_str_buf_init(&buf);
1832 
1833  __kmp_str_buf_print(&buf, "%d", totals[0]);
1834  for (level = 1; level <= pkgLevel; level++) {
1835  __kmp_str_buf_print(&buf, " x %d", maxCt[level]);
1836  }
1837  KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, nCoresPerPkg,
1838  __kmp_nThreadsPerCore, __kmp_ncores);
1839 
1840  __kmp_str_buf_free(&buf);
1841  }
1842  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
1843  KMP_DEBUG_ASSERT(nApics == __kmp_avail_proc);
1844  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
1845  for (proc = 0; (int)proc < nApics; ++proc) {
1846  __kmp_pu_os_idx[proc] = retval[proc].second;
1847  }
1848  if (__kmp_affinity_type == affinity_none) {
1849  __kmp_free(last);
1850  __kmp_free(maxCt);
1851  __kmp_free(counts);
1852  __kmp_free(totals);
1853  __kmp_free(retval);
1854  KMP_CPU_FREE(oldMask);
1855  return 0;
1856  }
1857 
1858  // Find any levels with radix 1, and remove them from the map
1859  // (except for the package level).
1860  int new_depth = 0;
1861  for (level = 0; level < depth; level++) {
1862  if ((maxCt[level] == 1) && (level != pkgLevel)) {
1863  continue;
1864  }
1865  new_depth++;
1866  }
1867 
1868  // If we are removing any levels, allocate a new vector to return,
1869  // and copy the relevant information to it.
1870  if (new_depth != depth) {
1871  AddrUnsPair *new_retval =
1872  (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * nApics);
1873  for (proc = 0; (int)proc < nApics; proc++) {
1874  Address addr(new_depth);
1875  new_retval[proc] = AddrUnsPair(addr, retval[proc].second);
1876  }
1877  int new_level = 0;
1878  int newPkgLevel = -1;
1879  int newCoreLevel = -1;
1880  int newThreadLevel = -1;
1881  for (level = 0; level < depth; level++) {
1882  if ((maxCt[level] == 1) && (level != pkgLevel)) {
1883  // Remove this level. Never remove the package level
1884  continue;
1885  }
1886  if (level == pkgLevel) {
1887  newPkgLevel = new_level;
1888  }
1889  if (level == coreLevel) {
1890  newCoreLevel = new_level;
1891  }
1892  if (level == threadLevel) {
1893  newThreadLevel = new_level;
1894  }
1895  for (proc = 0; (int)proc < nApics; proc++) {
1896  new_retval[proc].first.labels[new_level] =
1897  retval[proc].first.labels[level];
1898  }
1899  new_level++;
1900  }
1901 
1902  __kmp_free(retval);
1903  retval = new_retval;
1904  depth = new_depth;
1905  pkgLevel = newPkgLevel;
1906  coreLevel = newCoreLevel;
1907  threadLevel = newThreadLevel;
1908  }
1909 
1910  if (__kmp_affinity_gran_levels < 0) {
1911  // Set the granularity level based on what levels are modeled
1912  // in the machine topology map.
1913  __kmp_affinity_gran_levels = 0;
1914  if ((threadLevel >= 0) && (__kmp_affinity_gran > affinity_gran_thread)) {
1915  __kmp_affinity_gran_levels++;
1916  }
1917  if ((coreLevel >= 0) && (__kmp_affinity_gran > affinity_gran_core)) {
1918  __kmp_affinity_gran_levels++;
1919  }
1920  if (__kmp_affinity_gran > affinity_gran_package) {
1921  __kmp_affinity_gran_levels++;
1922  }
1923  }
1924 
1925  if (__kmp_affinity_verbose) {
1926  __kmp_affinity_print_topology(retval, nApics, depth, pkgLevel, coreLevel,
1927  threadLevel);
1928  }
1929 
1930  __kmp_free(last);
1931  __kmp_free(maxCt);
1932  __kmp_free(counts);
1933  __kmp_free(totals);
1934  KMP_CPU_FREE(oldMask);
1935  *address2os = retval;
1936  return depth;
1937 }
1938 
1939 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
1940 
1941 #define osIdIndex 0
1942 #define threadIdIndex 1
1943 #define coreIdIndex 2
1944 #define pkgIdIndex 3
1945 #define nodeIdIndex 4
1946 
1947 typedef unsigned *ProcCpuInfo;
1948 static unsigned maxIndex = pkgIdIndex;
1949 
1950 static int __kmp_affinity_cmp_ProcCpuInfo_phys_id(const void *a,
1951  const void *b) {
1952  unsigned i;
1953  const unsigned *aa = *(unsigned *const *)a;
1954  const unsigned *bb = *(unsigned *const *)b;
1955  for (i = maxIndex;; i--) {
1956  if (aa[i] < bb[i])
1957  return -1;
1958  if (aa[i] > bb[i])
1959  return 1;
1960  if (i == osIdIndex)
1961  break;
1962  }
1963  return 0;
1964 }
1965 
1966 #if KMP_USE_HIER_SCHED
1967 // Set the array sizes for the hierarchy layers
1968 static void __kmp_dispatch_set_hierarchy_values() {
1969  // Set the maximum number of L1's to number of cores
1970  // Set the maximum number of L2's to to either number of cores / 2 for
1971  // Intel(R) Xeon Phi(TM) coprocessor formally codenamed Knights Landing
1972  // Or the number of cores for Intel(R) Xeon(R) processors
1973  // Set the maximum number of NUMA nodes and L3's to number of packages
1974  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1] =
1975  nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
1976  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L1 + 1] = __kmp_ncores;
1977 #if KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_WINDOWS) && \
1978  KMP_MIC_SUPPORTED
1979  if (__kmp_mic_type >= mic3)
1980  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores / 2;
1981  else
1982 #endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
1983  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores;
1984  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L3 + 1] = nPackages;
1985  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_NUMA + 1] = nPackages;
1986  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_LOOP + 1] = 1;
1987  // Set the number of threads per unit
1988  // Number of hardware threads per L1/L2/L3/NUMA/LOOP
1989  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_THREAD + 1] = 1;
1990  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L1 + 1] =
1991  __kmp_nThreadsPerCore;
1992 #if KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_WINDOWS) && \
1993  KMP_MIC_SUPPORTED
1994  if (__kmp_mic_type >= mic3)
1995  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
1996  2 * __kmp_nThreadsPerCore;
1997  else
1998 #endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
1999  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
2000  __kmp_nThreadsPerCore;
2001  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L3 + 1] =
2002  nCoresPerPkg * __kmp_nThreadsPerCore;
2003  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_NUMA + 1] =
2004  nCoresPerPkg * __kmp_nThreadsPerCore;
2005  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_LOOP + 1] =
2006  nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
2007 }
2008 
2009 // Return the index into the hierarchy for this tid and layer type (L1, L2, etc)
2010 // i.e., this thread's L1 or this thread's L2, etc.
2011 int __kmp_dispatch_get_index(int tid, kmp_hier_layer_e type) {
2012  int index = type + 1;
2013  int num_hw_threads = __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1];
2014  KMP_DEBUG_ASSERT(type != kmp_hier_layer_e::LAYER_LAST);
2015  if (type == kmp_hier_layer_e::LAYER_THREAD)
2016  return tid;
2017  else if (type == kmp_hier_layer_e::LAYER_LOOP)
2018  return 0;
2019  KMP_DEBUG_ASSERT(__kmp_hier_max_units[index] != 0);
2020  if (tid >= num_hw_threads)
2021  tid = tid % num_hw_threads;
2022  return (tid / __kmp_hier_threads_per[index]) % __kmp_hier_max_units[index];
2023 }
2024 
2025 // Return the number of t1's per t2
2026 int __kmp_dispatch_get_t1_per_t2(kmp_hier_layer_e t1, kmp_hier_layer_e t2) {
2027  int i1 = t1 + 1;
2028  int i2 = t2 + 1;
2029  KMP_DEBUG_ASSERT(i1 <= i2);
2030  KMP_DEBUG_ASSERT(t1 != kmp_hier_layer_e::LAYER_LAST);
2031  KMP_DEBUG_ASSERT(t2 != kmp_hier_layer_e::LAYER_LAST);
2032  KMP_DEBUG_ASSERT(__kmp_hier_threads_per[i1] != 0);
2033  // (nthreads/t2) / (nthreads/t1) = t1 / t2
2034  return __kmp_hier_threads_per[i2] / __kmp_hier_threads_per[i1];
2035 }
2036 #endif // KMP_USE_HIER_SCHED
2037 
2038 // Parse /proc/cpuinfo (or an alternate file in the same format) to obtain the
2039 // affinity map.
2040 static int __kmp_affinity_create_cpuinfo_map(AddrUnsPair **address2os,
2041  int *line,
2042  kmp_i18n_id_t *const msg_id,
2043  FILE *f) {
2044  *address2os = NULL;
2045  *msg_id = kmp_i18n_null;
2046 
2047  // Scan of the file, and count the number of "processor" (osId) fields,
2048  // and find the highest value of <n> for a node_<n> field.
2049  char buf[256];
2050  unsigned num_records = 0;
2051  while (!feof(f)) {
2052  buf[sizeof(buf) - 1] = 1;
2053  if (!fgets(buf, sizeof(buf), f)) {
2054  // Read errors presumably because of EOF
2055  break;
2056  }
2057 
2058  char s1[] = "processor";
2059  if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
2060  num_records++;
2061  continue;
2062  }
2063 
2064  // FIXME - this will match "node_<n> <garbage>"
2065  unsigned level;
2066  if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
2067  if (nodeIdIndex + level >= maxIndex) {
2068  maxIndex = nodeIdIndex + level;
2069  }
2070  continue;
2071  }
2072  }
2073 
2074  // Check for empty file / no valid processor records, or too many. The number
2075  // of records can't exceed the number of valid bits in the affinity mask.
2076  if (num_records == 0) {
2077  *line = 0;
2078  *msg_id = kmp_i18n_str_NoProcRecords;
2079  return -1;
2080  }
2081  if (num_records > (unsigned)__kmp_xproc) {
2082  *line = 0;
2083  *msg_id = kmp_i18n_str_TooManyProcRecords;
2084  return -1;
2085  }
2086 
2087  // Set the file pointer back to the beginning, so that we can scan the file
2088  // again, this time performing a full parse of the data. Allocate a vector of
2089  // ProcCpuInfo object, where we will place the data. Adding an extra element
2090  // at the end allows us to remove a lot of extra checks for termination
2091  // conditions.
2092  if (fseek(f, 0, SEEK_SET) != 0) {
2093  *line = 0;
2094  *msg_id = kmp_i18n_str_CantRewindCpuinfo;
2095  return -1;
2096  }
2097 
2098  // Allocate the array of records to store the proc info in. The dummy
2099  // element at the end makes the logic in filling them out easier to code.
2100  unsigned **threadInfo =
2101  (unsigned **)__kmp_allocate((num_records + 1) * sizeof(unsigned *));
2102  unsigned i;
2103  for (i = 0; i <= num_records; i++) {
2104  threadInfo[i] =
2105  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2106  }
2107 
2108 #define CLEANUP_THREAD_INFO \
2109  for (i = 0; i <= num_records; i++) { \
2110  __kmp_free(threadInfo[i]); \
2111  } \
2112  __kmp_free(threadInfo);
2113 
2114  // A value of UINT_MAX means that we didn't find the field
2115  unsigned __index;
2116 
2117 #define INIT_PROC_INFO(p) \
2118  for (__index = 0; __index <= maxIndex; __index++) { \
2119  (p)[__index] = UINT_MAX; \
2120  }
2121 
2122  for (i = 0; i <= num_records; i++) {
2123  INIT_PROC_INFO(threadInfo[i]);
2124  }
2125 
2126  unsigned num_avail = 0;
2127  *line = 0;
2128  while (!feof(f)) {
2129  // Create an inner scoping level, so that all the goto targets at the end of
2130  // the loop appear in an outer scoping level. This avoids warnings about
2131  // jumping past an initialization to a target in the same block.
2132  {
2133  buf[sizeof(buf) - 1] = 1;
2134  bool long_line = false;
2135  if (!fgets(buf, sizeof(buf), f)) {
2136  // Read errors presumably because of EOF
2137  // If there is valid data in threadInfo[num_avail], then fake
2138  // a blank line in ensure that the last address gets parsed.
2139  bool valid = false;
2140  for (i = 0; i <= maxIndex; i++) {
2141  if (threadInfo[num_avail][i] != UINT_MAX) {
2142  valid = true;
2143  }
2144  }
2145  if (!valid) {
2146  break;
2147  }
2148  buf[0] = 0;
2149  } else if (!buf[sizeof(buf) - 1]) {
2150  // The line is longer than the buffer. Set a flag and don't
2151  // emit an error if we were going to ignore the line, anyway.
2152  long_line = true;
2153 
2154 #define CHECK_LINE \
2155  if (long_line) { \
2156  CLEANUP_THREAD_INFO; \
2157  *msg_id = kmp_i18n_str_LongLineCpuinfo; \
2158  return -1; \
2159  }
2160  }
2161  (*line)++;
2162 
2163  char s1[] = "processor";
2164  if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
2165  CHECK_LINE;
2166  char *p = strchr(buf + sizeof(s1) - 1, ':');
2167  unsigned val;
2168  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2169  goto no_val;
2170  if (threadInfo[num_avail][osIdIndex] != UINT_MAX)
2171 #if KMP_ARCH_AARCH64
2172  // Handle the old AArch64 /proc/cpuinfo layout differently,
2173  // it contains all of the 'processor' entries listed in a
2174  // single 'Processor' section, therefore the normal looking
2175  // for duplicates in that section will always fail.
2176  num_avail++;
2177 #else
2178  goto dup_field;
2179 #endif
2180  threadInfo[num_avail][osIdIndex] = val;
2181 #if KMP_OS_LINUX && !(KMP_ARCH_X86 || KMP_ARCH_X86_64)
2182  char path[256];
2183  KMP_SNPRINTF(
2184  path, sizeof(path),
2185  "/sys/devices/system/cpu/cpu%u/topology/physical_package_id",
2186  threadInfo[num_avail][osIdIndex]);
2187  __kmp_read_from_file(path, "%u", &threadInfo[num_avail][pkgIdIndex]);
2188 
2189  KMP_SNPRINTF(path, sizeof(path),
2190  "/sys/devices/system/cpu/cpu%u/topology/core_id",
2191  threadInfo[num_avail][osIdIndex]);
2192  __kmp_read_from_file(path, "%u", &threadInfo[num_avail][coreIdIndex]);
2193  continue;
2194 #else
2195  }
2196  char s2[] = "physical id";
2197  if (strncmp(buf, s2, sizeof(s2) - 1) == 0) {
2198  CHECK_LINE;
2199  char *p = strchr(buf + sizeof(s2) - 1, ':');
2200  unsigned val;
2201  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2202  goto no_val;
2203  if (threadInfo[num_avail][pkgIdIndex] != UINT_MAX)
2204  goto dup_field;
2205  threadInfo[num_avail][pkgIdIndex] = val;
2206  continue;
2207  }
2208  char s3[] = "core id";
2209  if (strncmp(buf, s3, sizeof(s3) - 1) == 0) {
2210  CHECK_LINE;
2211  char *p = strchr(buf + sizeof(s3) - 1, ':');
2212  unsigned val;
2213  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2214  goto no_val;
2215  if (threadInfo[num_avail][coreIdIndex] != UINT_MAX)
2216  goto dup_field;
2217  threadInfo[num_avail][coreIdIndex] = val;
2218  continue;
2219 #endif // KMP_OS_LINUX && USE_SYSFS_INFO
2220  }
2221  char s4[] = "thread id";
2222  if (strncmp(buf, s4, sizeof(s4) - 1) == 0) {
2223  CHECK_LINE;
2224  char *p = strchr(buf + sizeof(s4) - 1, ':');
2225  unsigned val;
2226  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2227  goto no_val;
2228  if (threadInfo[num_avail][threadIdIndex] != UINT_MAX)
2229  goto dup_field;
2230  threadInfo[num_avail][threadIdIndex] = val;
2231  continue;
2232  }
2233  unsigned level;
2234  if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
2235  CHECK_LINE;
2236  char *p = strchr(buf + sizeof(s4) - 1, ':');
2237  unsigned val;
2238  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2239  goto no_val;
2240  KMP_ASSERT(nodeIdIndex + level <= maxIndex);
2241  if (threadInfo[num_avail][nodeIdIndex + level] != UINT_MAX)
2242  goto dup_field;
2243  threadInfo[num_avail][nodeIdIndex + level] = val;
2244  continue;
2245  }
2246 
2247  // We didn't recognize the leading token on the line. There are lots of
2248  // leading tokens that we don't recognize - if the line isn't empty, go on
2249  // to the next line.
2250  if ((*buf != 0) && (*buf != '\n')) {
2251  // If the line is longer than the buffer, read characters
2252  // until we find a newline.
2253  if (long_line) {
2254  int ch;
2255  while (((ch = fgetc(f)) != EOF) && (ch != '\n'))
2256  ;
2257  }
2258  continue;
2259  }
2260 
2261  // A newline has signalled the end of the processor record.
2262  // Check that there aren't too many procs specified.
2263  if ((int)num_avail == __kmp_xproc) {
2264  CLEANUP_THREAD_INFO;
2265  *msg_id = kmp_i18n_str_TooManyEntries;
2266  return -1;
2267  }
2268 
2269  // Check for missing fields. The osId field must be there, and we
2270  // currently require that the physical id field is specified, also.
2271  if (threadInfo[num_avail][osIdIndex] == UINT_MAX) {
2272  CLEANUP_THREAD_INFO;
2273  *msg_id = kmp_i18n_str_MissingProcField;
2274  return -1;
2275  }
2276  if (threadInfo[0][pkgIdIndex] == UINT_MAX) {
2277  CLEANUP_THREAD_INFO;
2278  *msg_id = kmp_i18n_str_MissingPhysicalIDField;
2279  return -1;
2280  }
2281 
2282  // Skip this proc if it is not included in the machine model.
2283  if (!KMP_CPU_ISSET(threadInfo[num_avail][osIdIndex],
2284  __kmp_affin_fullMask)) {
2285  INIT_PROC_INFO(threadInfo[num_avail]);
2286  continue;
2287  }
2288 
2289  // We have a successful parse of this proc's info.
2290  // Increment the counter, and prepare for the next proc.
2291  num_avail++;
2292  KMP_ASSERT(num_avail <= num_records);
2293  INIT_PROC_INFO(threadInfo[num_avail]);
2294  }
2295  continue;
2296 
2297  no_val:
2298  CLEANUP_THREAD_INFO;
2299  *msg_id = kmp_i18n_str_MissingValCpuinfo;
2300  return -1;
2301 
2302  dup_field:
2303  CLEANUP_THREAD_INFO;
2304  *msg_id = kmp_i18n_str_DuplicateFieldCpuinfo;
2305  return -1;
2306  }
2307  *line = 0;
2308 
2309 #if KMP_MIC && REDUCE_TEAM_SIZE
2310  unsigned teamSize = 0;
2311 #endif // KMP_MIC && REDUCE_TEAM_SIZE
2312 
2313  // check for num_records == __kmp_xproc ???
2314 
2315  // If there's only one thread context to bind to, form an Address object with
2316  // depth 1 and return immediately (or, if affinity is off, set address2os to
2317  // NULL and return).
2318  //
2319  // If it is configured to omit the package level when there is only a single
2320  // package, the logic at the end of this routine won't work if there is only a
2321  // single thread - it would try to form an Address object with depth 0.
2322  KMP_ASSERT(num_avail > 0);
2323  KMP_ASSERT(num_avail <= num_records);
2324  if (num_avail == 1) {
2325  __kmp_ncores = 1;
2326  __kmp_nThreadsPerCore = nCoresPerPkg = nPackages = 1;
2327  if (__kmp_affinity_verbose) {
2328  if (!KMP_AFFINITY_CAPABLE()) {
2329  KMP_INFORM(AffNotCapableUseCpuinfo, "KMP_AFFINITY");
2330  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
2331  KMP_INFORM(Uniform, "KMP_AFFINITY");
2332  } else {
2333  char buf[KMP_AFFIN_MASK_PRINT_LEN];
2334  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
2335  __kmp_affin_fullMask);
2336  KMP_INFORM(AffCapableUseCpuinfo, "KMP_AFFINITY");
2337  if (__kmp_affinity_respect_mask) {
2338  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
2339  } else {
2340  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
2341  }
2342  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
2343  KMP_INFORM(Uniform, "KMP_AFFINITY");
2344  }
2345  int index;
2346  kmp_str_buf_t buf;
2347  __kmp_str_buf_init(&buf);
2348  __kmp_str_buf_print(&buf, "1");
2349  for (index = maxIndex - 1; index > pkgIdIndex; index--) {
2350  __kmp_str_buf_print(&buf, " x 1");
2351  }
2352  KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, 1, 1, 1);
2353  __kmp_str_buf_free(&buf);
2354  }
2355 
2356  if (__kmp_affinity_type == affinity_none) {
2357  CLEANUP_THREAD_INFO;
2358  return 0;
2359  }
2360 
2361  *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair));
2362  Address addr(1);
2363  addr.labels[0] = threadInfo[0][pkgIdIndex];
2364  (*address2os)[0] = AddrUnsPair(addr, threadInfo[0][osIdIndex]);
2365 
2366  if (__kmp_affinity_gran_levels < 0) {
2367  __kmp_affinity_gran_levels = 0;
2368  }
2369 
2370  if (__kmp_affinity_verbose) {
2371  __kmp_affinity_print_topology(*address2os, 1, 1, 0, -1, -1);
2372  }
2373 
2374  CLEANUP_THREAD_INFO;
2375  return 1;
2376  }
2377 
2378  // Sort the threadInfo table by physical Id.
2379  qsort(threadInfo, num_avail, sizeof(*threadInfo),
2380  __kmp_affinity_cmp_ProcCpuInfo_phys_id);
2381 
2382  // The table is now sorted by pkgId / coreId / threadId, but we really don't
2383  // know the radix of any of the fields. pkgId's may be sparsely assigned among
2384  // the chips on a system. Although coreId's are usually assigned
2385  // [0 .. coresPerPkg-1] and threadId's are usually assigned
2386  // [0..threadsPerCore-1], we don't want to make any such assumptions.
2387  //
2388  // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
2389  // total # packages) are at this point - we want to determine that now. We
2390  // only have an upper bound on the first two figures.
2391  unsigned *counts =
2392  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2393  unsigned *maxCt =
2394  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2395  unsigned *totals =
2396  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2397  unsigned *lastId =
2398  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2399 
2400  bool assign_thread_ids = false;
2401  unsigned threadIdCt;
2402  unsigned index;
2403 
2404 restart_radix_check:
2405  threadIdCt = 0;
2406 
2407  // Initialize the counter arrays with data from threadInfo[0].
2408  if (assign_thread_ids) {
2409  if (threadInfo[0][threadIdIndex] == UINT_MAX) {
2410  threadInfo[0][threadIdIndex] = threadIdCt++;
2411  } else if (threadIdCt <= threadInfo[0][threadIdIndex]) {
2412  threadIdCt = threadInfo[0][threadIdIndex] + 1;
2413  }
2414  }
2415  for (index = 0; index <= maxIndex; index++) {
2416  counts[index] = 1;
2417  maxCt[index] = 1;
2418  totals[index] = 1;
2419  lastId[index] = threadInfo[0][index];
2420  ;
2421  }
2422 
2423  // Run through the rest of the OS procs.
2424  for (i = 1; i < num_avail; i++) {
2425  // Find the most significant index whose id differs from the id for the
2426  // previous OS proc.
2427  for (index = maxIndex; index >= threadIdIndex; index--) {
2428  if (assign_thread_ids && (index == threadIdIndex)) {
2429  // Auto-assign the thread id field if it wasn't specified.
2430  if (threadInfo[i][threadIdIndex] == UINT_MAX) {
2431  threadInfo[i][threadIdIndex] = threadIdCt++;
2432  }
2433  // Apparently the thread id field was specified for some entries and not
2434  // others. Start the thread id counter off at the next higher thread id.
2435  else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
2436  threadIdCt = threadInfo[i][threadIdIndex] + 1;
2437  }
2438  }
2439  if (threadInfo[i][index] != lastId[index]) {
2440  // Run through all indices which are less significant, and reset the
2441  // counts to 1. At all levels up to and including index, we need to
2442  // increment the totals and record the last id.
2443  unsigned index2;
2444  for (index2 = threadIdIndex; index2 < index; index2++) {
2445  totals[index2]++;
2446  if (counts[index2] > maxCt[index2]) {
2447  maxCt[index2] = counts[index2];
2448  }
2449  counts[index2] = 1;
2450  lastId[index2] = threadInfo[i][index2];
2451  }
2452  counts[index]++;
2453  totals[index]++;
2454  lastId[index] = threadInfo[i][index];
2455 
2456  if (assign_thread_ids && (index > threadIdIndex)) {
2457 
2458 #if KMP_MIC && REDUCE_TEAM_SIZE
2459  // The default team size is the total #threads in the machine
2460  // minus 1 thread for every core that has 3 or more threads.
2461  teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
2462 #endif // KMP_MIC && REDUCE_TEAM_SIZE
2463 
2464  // Restart the thread counter, as we are on a new core.
2465  threadIdCt = 0;
2466 
2467  // Auto-assign the thread id field if it wasn't specified.
2468  if (threadInfo[i][threadIdIndex] == UINT_MAX) {
2469  threadInfo[i][threadIdIndex] = threadIdCt++;
2470  }
2471 
2472  // Apparently the thread id field was specified for some entries and
2473  // not others. Start the thread id counter off at the next higher
2474  // thread id.
2475  else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
2476  threadIdCt = threadInfo[i][threadIdIndex] + 1;
2477  }
2478  }
2479  break;
2480  }
2481  }
2482  if (index < threadIdIndex) {
2483  // If thread ids were specified, it is an error if they are not unique.
2484  // Also, check that we waven't already restarted the loop (to be safe -
2485  // shouldn't need to).
2486  if ((threadInfo[i][threadIdIndex] != UINT_MAX) || assign_thread_ids) {
2487  __kmp_free(lastId);
2488  __kmp_free(totals);
2489  __kmp_free(maxCt);
2490  __kmp_free(counts);
2491  CLEANUP_THREAD_INFO;
2492  *msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
2493  return -1;
2494  }
2495 
2496  // If the thread ids were not specified and we see entries entries that
2497  // are duplicates, start the loop over and assign the thread ids manually.
2498  assign_thread_ids = true;
2499  goto restart_radix_check;
2500  }
2501  }
2502 
2503 #if KMP_MIC && REDUCE_TEAM_SIZE
2504  // The default team size is the total #threads in the machine
2505  // minus 1 thread for every core that has 3 or more threads.
2506  teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
2507 #endif // KMP_MIC && REDUCE_TEAM_SIZE
2508 
2509  for (index = threadIdIndex; index <= maxIndex; index++) {
2510  if (counts[index] > maxCt[index]) {
2511  maxCt[index] = counts[index];
2512  }
2513  }
2514 
2515  __kmp_nThreadsPerCore = maxCt[threadIdIndex];
2516  nCoresPerPkg = maxCt[coreIdIndex];
2517  nPackages = totals[pkgIdIndex];
2518 
2519  // Check to see if the machine topology is uniform
2520  unsigned prod = totals[maxIndex];
2521  for (index = threadIdIndex; index < maxIndex; index++) {
2522  prod *= maxCt[index];
2523  }
2524  bool uniform = (prod == totals[threadIdIndex]);
2525 
2526  // When affinity is off, this routine will still be called to set
2527  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
2528  // Make sure all these vars are set correctly, and return now if affinity is
2529  // not enabled.
2530  __kmp_ncores = totals[coreIdIndex];
2531 
2532  if (__kmp_affinity_verbose) {
2533  if (!KMP_AFFINITY_CAPABLE()) {
2534  KMP_INFORM(AffNotCapableUseCpuinfo, "KMP_AFFINITY");
2535  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
2536  if (uniform) {
2537  KMP_INFORM(Uniform, "KMP_AFFINITY");
2538  } else {
2539  KMP_INFORM(NonUniform, "KMP_AFFINITY");
2540  }
2541  } else {
2542  char buf[KMP_AFFIN_MASK_PRINT_LEN];
2543  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
2544  __kmp_affin_fullMask);
2545  KMP_INFORM(AffCapableUseCpuinfo, "KMP_AFFINITY");
2546  if (__kmp_affinity_respect_mask) {
2547  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
2548  } else {
2549  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
2550  }
2551  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
2552  if (uniform) {
2553  KMP_INFORM(Uniform, "KMP_AFFINITY");
2554  } else {
2555  KMP_INFORM(NonUniform, "KMP_AFFINITY");
2556  }
2557  }
2558  kmp_str_buf_t buf;
2559  __kmp_str_buf_init(&buf);
2560 
2561  __kmp_str_buf_print(&buf, "%d", totals[maxIndex]);
2562  for (index = maxIndex - 1; index >= pkgIdIndex; index--) {
2563  __kmp_str_buf_print(&buf, " x %d", maxCt[index]);
2564  }
2565  KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, maxCt[coreIdIndex],
2566  maxCt[threadIdIndex], __kmp_ncores);
2567 
2568  __kmp_str_buf_free(&buf);
2569  }
2570 
2571 #if KMP_MIC && REDUCE_TEAM_SIZE
2572  // Set the default team size.
2573  if ((__kmp_dflt_team_nth == 0) && (teamSize > 0)) {
2574  __kmp_dflt_team_nth = teamSize;
2575  KA_TRACE(20, ("__kmp_affinity_create_cpuinfo_map: setting "
2576  "__kmp_dflt_team_nth = %d\n",
2577  __kmp_dflt_team_nth));
2578  }
2579 #endif // KMP_MIC && REDUCE_TEAM_SIZE
2580 
2581  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
2582  KMP_DEBUG_ASSERT(num_avail == (unsigned)__kmp_avail_proc);
2583  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
2584  for (i = 0; i < num_avail; ++i) { // fill the os indices
2585  __kmp_pu_os_idx[i] = threadInfo[i][osIdIndex];
2586  }
2587 
2588  if (__kmp_affinity_type == affinity_none) {
2589  __kmp_free(lastId);
2590  __kmp_free(totals);
2591  __kmp_free(maxCt);
2592  __kmp_free(counts);
2593  CLEANUP_THREAD_INFO;
2594  return 0;
2595  }
2596 
2597  // Count the number of levels which have more nodes at that level than at the
2598  // parent's level (with there being an implicit root node of the top level).
2599  // This is equivalent to saying that there is at least one node at this level
2600  // which has a sibling. These levels are in the map, and the package level is
2601  // always in the map.
2602  bool *inMap = (bool *)__kmp_allocate((maxIndex + 1) * sizeof(bool));
2603  for (index = threadIdIndex; index < maxIndex; index++) {
2604  KMP_ASSERT(totals[index] >= totals[index + 1]);
2605  inMap[index] = (totals[index] > totals[index + 1]);
2606  }
2607  inMap[maxIndex] = (totals[maxIndex] > 1);
2608  inMap[pkgIdIndex] = true;
2609 
2610  int depth = 0;
2611  for (index = threadIdIndex; index <= maxIndex; index++) {
2612  if (inMap[index]) {
2613  depth++;
2614  }
2615  }
2616  KMP_ASSERT(depth > 0);
2617 
2618  // Construct the data structure that is to be returned.
2619  *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * num_avail);
2620  int pkgLevel = -1;
2621  int coreLevel = -1;
2622  int threadLevel = -1;
2623 
2624  for (i = 0; i < num_avail; ++i) {
2625  Address addr(depth);
2626  unsigned os = threadInfo[i][osIdIndex];
2627  int src_index;
2628  int dst_index = 0;
2629 
2630  for (src_index = maxIndex; src_index >= threadIdIndex; src_index--) {
2631  if (!inMap[src_index]) {
2632  continue;
2633  }
2634  addr.labels[dst_index] = threadInfo[i][src_index];
2635  if (src_index == pkgIdIndex) {
2636  pkgLevel = dst_index;
2637  } else if (src_index == coreIdIndex) {
2638  coreLevel = dst_index;
2639  } else if (src_index == threadIdIndex) {
2640  threadLevel = dst_index;
2641  }
2642  dst_index++;
2643  }
2644  (*address2os)[i] = AddrUnsPair(addr, os);
2645  }
2646 
2647  if (__kmp_affinity_gran_levels < 0) {
2648  // Set the granularity level based on what levels are modeled
2649  // in the machine topology map.
2650  unsigned src_index;
2651  __kmp_affinity_gran_levels = 0;
2652  for (src_index = threadIdIndex; src_index <= maxIndex; src_index++) {
2653  if (!inMap[src_index]) {
2654  continue;
2655  }
2656  switch (src_index) {
2657  case threadIdIndex:
2658  if (__kmp_affinity_gran > affinity_gran_thread) {
2659  __kmp_affinity_gran_levels++;
2660  }
2661 
2662  break;
2663  case coreIdIndex:
2664  if (__kmp_affinity_gran > affinity_gran_core) {
2665  __kmp_affinity_gran_levels++;
2666  }
2667  break;
2668 
2669  case pkgIdIndex:
2670  if (__kmp_affinity_gran > affinity_gran_package) {
2671  __kmp_affinity_gran_levels++;
2672  }
2673  break;
2674  }
2675  }
2676  }
2677 
2678  if (__kmp_affinity_verbose) {
2679  __kmp_affinity_print_topology(*address2os, num_avail, depth, pkgLevel,
2680  coreLevel, threadLevel);
2681  }
2682 
2683  __kmp_free(inMap);
2684  __kmp_free(lastId);
2685  __kmp_free(totals);
2686  __kmp_free(maxCt);
2687  __kmp_free(counts);
2688  CLEANUP_THREAD_INFO;
2689  return depth;
2690 }
2691 
2692 // Create and return a table of affinity masks, indexed by OS thread ID.
2693 // This routine handles OR'ing together all the affinity masks of threads
2694 // that are sufficiently close, if granularity > fine.
2695 static kmp_affin_mask_t *__kmp_create_masks(unsigned *maxIndex,
2696  unsigned *numUnique,
2697  AddrUnsPair *address2os,
2698  unsigned numAddrs) {
2699  // First form a table of affinity masks in order of OS thread id.
2700  unsigned depth;
2701  unsigned maxOsId;
2702  unsigned i;
2703 
2704  KMP_ASSERT(numAddrs > 0);
2705  depth = address2os[0].first.depth;
2706 
2707  maxOsId = 0;
2708  for (i = numAddrs - 1;; --i) {
2709  unsigned osId = address2os[i].second;
2710  if (osId > maxOsId) {
2711  maxOsId = osId;
2712  }
2713  if (i == 0)
2714  break;
2715  }
2716  kmp_affin_mask_t *osId2Mask;
2717  KMP_CPU_ALLOC_ARRAY(osId2Mask, (maxOsId + 1));
2718 
2719  // Sort the address2os table according to physical order. Doing so will put
2720  // all threads on the same core/package/node in consecutive locations.
2721  qsort(address2os, numAddrs, sizeof(*address2os),
2722  __kmp_affinity_cmp_Address_labels);
2723 
2724  KMP_ASSERT(__kmp_affinity_gran_levels >= 0);
2725  if (__kmp_affinity_verbose && (__kmp_affinity_gran_levels > 0)) {
2726  KMP_INFORM(ThreadsMigrate, "KMP_AFFINITY", __kmp_affinity_gran_levels);
2727  }
2728  if (__kmp_affinity_gran_levels >= (int)depth) {
2729  if (__kmp_affinity_verbose ||
2730  (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) {
2731  KMP_WARNING(AffThreadsMayMigrate);
2732  }
2733  }
2734 
2735  // Run through the table, forming the masks for all threads on each core.
2736  // Threads on the same core will have identical "Address" objects, not
2737  // considering the last level, which must be the thread id. All threads on a
2738  // core will appear consecutively.
2739  unsigned unique = 0;
2740  unsigned j = 0; // index of 1st thread on core
2741  unsigned leader = 0;
2742  Address *leaderAddr = &(address2os[0].first);
2743  kmp_affin_mask_t *sum;
2744  KMP_CPU_ALLOC_ON_STACK(sum);
2745  KMP_CPU_ZERO(sum);
2746  KMP_CPU_SET(address2os[0].second, sum);
2747  for (i = 1; i < numAddrs; i++) {
2748  // If this thread is sufficiently close to the leader (within the
2749  // granularity setting), then set the bit for this os thread in the
2750  // affinity mask for this group, and go on to the next thread.
2751  if (leaderAddr->isClose(address2os[i].first, __kmp_affinity_gran_levels)) {
2752  KMP_CPU_SET(address2os[i].second, sum);
2753  continue;
2754  }
2755 
2756  // For every thread in this group, copy the mask to the thread's entry in
2757  // the osId2Mask table. Mark the first address as a leader.
2758  for (; j < i; j++) {
2759  unsigned osId = address2os[j].second;
2760  KMP_DEBUG_ASSERT(osId <= maxOsId);
2761  kmp_affin_mask_t *mask = KMP_CPU_INDEX(osId2Mask, osId);
2762  KMP_CPU_COPY(mask, sum);
2763  address2os[j].first.leader = (j == leader);
2764  }
2765  unique++;
2766 
2767  // Start a new mask.
2768  leader = i;
2769  leaderAddr = &(address2os[i].first);
2770  KMP_CPU_ZERO(sum);
2771  KMP_CPU_SET(address2os[i].second, sum);
2772  }
2773 
2774  // For every thread in last group, copy the mask to the thread's
2775  // entry in the osId2Mask table.
2776  for (; j < i; j++) {
2777  unsigned osId = address2os[j].second;
2778  KMP_DEBUG_ASSERT(osId <= maxOsId);
2779  kmp_affin_mask_t *mask = KMP_CPU_INDEX(osId2Mask, osId);
2780  KMP_CPU_COPY(mask, sum);
2781  address2os[j].first.leader = (j == leader);
2782  }
2783  unique++;
2784  KMP_CPU_FREE_FROM_STACK(sum);
2785 
2786  *maxIndex = maxOsId;
2787  *numUnique = unique;
2788  return osId2Mask;
2789 }
2790 
2791 // Stuff for the affinity proclist parsers. It's easier to declare these vars
2792 // as file-static than to try and pass them through the calling sequence of
2793 // the recursive-descent OMP_PLACES parser.
2794 static kmp_affin_mask_t *newMasks;
2795 static int numNewMasks;
2796 static int nextNewMask;
2797 
2798 #define ADD_MASK(_mask) \
2799  { \
2800  if (nextNewMask >= numNewMasks) { \
2801  int i; \
2802  numNewMasks *= 2; \
2803  kmp_affin_mask_t *temp; \
2804  KMP_CPU_INTERNAL_ALLOC_ARRAY(temp, numNewMasks); \
2805  for (i = 0; i < numNewMasks / 2; i++) { \
2806  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i); \
2807  kmp_affin_mask_t *dest = KMP_CPU_INDEX(temp, i); \
2808  KMP_CPU_COPY(dest, src); \
2809  } \
2810  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks / 2); \
2811  newMasks = temp; \
2812  } \
2813  KMP_CPU_COPY(KMP_CPU_INDEX(newMasks, nextNewMask), (_mask)); \
2814  nextNewMask++; \
2815  }
2816 
2817 #define ADD_MASK_OSID(_osId, _osId2Mask, _maxOsId) \
2818  { \
2819  if (((_osId) > _maxOsId) || \
2820  (!KMP_CPU_ISSET((_osId), KMP_CPU_INDEX((_osId2Mask), (_osId))))) { \
2821  if (__kmp_affinity_verbose || \
2822  (__kmp_affinity_warnings && \
2823  (__kmp_affinity_type != affinity_none))) { \
2824  KMP_WARNING(AffIgnoreInvalidProcID, _osId); \
2825  } \
2826  } else { \
2827  ADD_MASK(KMP_CPU_INDEX(_osId2Mask, (_osId))); \
2828  } \
2829  }
2830 
2831 // Re-parse the proclist (for the explicit affinity type), and form the list
2832 // of affinity newMasks indexed by gtid.
2833 static void __kmp_affinity_process_proclist(kmp_affin_mask_t **out_masks,
2834  unsigned int *out_numMasks,
2835  const char *proclist,
2836  kmp_affin_mask_t *osId2Mask,
2837  int maxOsId) {
2838  int i;
2839  const char *scan = proclist;
2840  const char *next = proclist;
2841 
2842  // We use malloc() for the temporary mask vector, so that we can use
2843  // realloc() to extend it.
2844  numNewMasks = 2;
2845  KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
2846  nextNewMask = 0;
2847  kmp_affin_mask_t *sumMask;
2848  KMP_CPU_ALLOC(sumMask);
2849  int setSize = 0;
2850 
2851  for (;;) {
2852  int start, end, stride;
2853 
2854  SKIP_WS(scan);
2855  next = scan;
2856  if (*next == '\0') {
2857  break;
2858  }
2859 
2860  if (*next == '{') {
2861  int num;
2862  setSize = 0;
2863  next++; // skip '{'
2864  SKIP_WS(next);
2865  scan = next;
2866 
2867  // Read the first integer in the set.
2868  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad proclist");
2869  SKIP_DIGITS(next);
2870  num = __kmp_str_to_int(scan, *next);
2871  KMP_ASSERT2(num >= 0, "bad explicit proc list");
2872 
2873  // Copy the mask for that osId to the sum (union) mask.
2874  if ((num > maxOsId) ||
2875  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
2876  if (__kmp_affinity_verbose ||
2877  (__kmp_affinity_warnings &&
2878  (__kmp_affinity_type != affinity_none))) {
2879  KMP_WARNING(AffIgnoreInvalidProcID, num);
2880  }
2881  KMP_CPU_ZERO(sumMask);
2882  } else {
2883  KMP_CPU_COPY(sumMask, KMP_CPU_INDEX(osId2Mask, num));
2884  setSize = 1;
2885  }
2886 
2887  for (;;) {
2888  // Check for end of set.
2889  SKIP_WS(next);
2890  if (*next == '}') {
2891  next++; // skip '}'
2892  break;
2893  }
2894 
2895  // Skip optional comma.
2896  if (*next == ',') {
2897  next++;
2898  }
2899  SKIP_WS(next);
2900 
2901  // Read the next integer in the set.
2902  scan = next;
2903  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
2904 
2905  SKIP_DIGITS(next);
2906  num = __kmp_str_to_int(scan, *next);
2907  KMP_ASSERT2(num >= 0, "bad explicit proc list");
2908 
2909  // Add the mask for that osId to the sum mask.
2910  if ((num > maxOsId) ||
2911  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
2912  if (__kmp_affinity_verbose ||
2913  (__kmp_affinity_warnings &&
2914  (__kmp_affinity_type != affinity_none))) {
2915  KMP_WARNING(AffIgnoreInvalidProcID, num);
2916  }
2917  } else {
2918  KMP_CPU_UNION(sumMask, KMP_CPU_INDEX(osId2Mask, num));
2919  setSize++;
2920  }
2921  }
2922  if (setSize > 0) {
2923  ADD_MASK(sumMask);
2924  }
2925 
2926  SKIP_WS(next);
2927  if (*next == ',') {
2928  next++;
2929  }
2930  scan = next;
2931  continue;
2932  }
2933 
2934  // Read the first integer.
2935  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
2936  SKIP_DIGITS(next);
2937  start = __kmp_str_to_int(scan, *next);
2938  KMP_ASSERT2(start >= 0, "bad explicit proc list");
2939  SKIP_WS(next);
2940 
2941  // If this isn't a range, then add a mask to the list and go on.
2942  if (*next != '-') {
2943  ADD_MASK_OSID(start, osId2Mask, maxOsId);
2944 
2945  // Skip optional comma.
2946  if (*next == ',') {
2947  next++;
2948  }
2949  scan = next;
2950  continue;
2951  }
2952 
2953  // This is a range. Skip over the '-' and read in the 2nd int.
2954  next++; // skip '-'
2955  SKIP_WS(next);
2956  scan = next;
2957  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
2958  SKIP_DIGITS(next);
2959  end = __kmp_str_to_int(scan, *next);
2960  KMP_ASSERT2(end >= 0, "bad explicit proc list");
2961 
2962  // Check for a stride parameter
2963  stride = 1;
2964  SKIP_WS(next);
2965  if (*next == ':') {
2966  // A stride is specified. Skip over the ':" and read the 3rd int.
2967  int sign = +1;
2968  next++; // skip ':'
2969  SKIP_WS(next);
2970  scan = next;
2971  if (*next == '-') {
2972  sign = -1;
2973  next++;
2974  SKIP_WS(next);
2975  scan = next;
2976  }
2977  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
2978  SKIP_DIGITS(next);
2979  stride = __kmp_str_to_int(scan, *next);
2980  KMP_ASSERT2(stride >= 0, "bad explicit proc list");
2981  stride *= sign;
2982  }
2983 
2984  // Do some range checks.
2985  KMP_ASSERT2(stride != 0, "bad explicit proc list");
2986  if (stride > 0) {
2987  KMP_ASSERT2(start <= end, "bad explicit proc list");
2988  } else {
2989  KMP_ASSERT2(start >= end, "bad explicit proc list");
2990  }
2991  KMP_ASSERT2((end - start) / stride <= 65536, "bad explicit proc list");
2992 
2993  // Add the mask for each OS proc # to the list.
2994  if (stride > 0) {
2995  do {
2996  ADD_MASK_OSID(start, osId2Mask, maxOsId);
2997  start += stride;
2998  } while (start <= end);
2999  } else {
3000  do {
3001  ADD_MASK_OSID(start, osId2Mask, maxOsId);
3002  start += stride;
3003  } while (start >= end);
3004  }
3005 
3006  // Skip optional comma.
3007  SKIP_WS(next);
3008  if (*next == ',') {
3009  next++;
3010  }
3011  scan = next;
3012  }
3013 
3014  *out_numMasks = nextNewMask;
3015  if (nextNewMask == 0) {
3016  *out_masks = NULL;
3017  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3018  return;
3019  }
3020  KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
3021  for (i = 0; i < nextNewMask; i++) {
3022  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
3023  kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
3024  KMP_CPU_COPY(dest, src);
3025  }
3026  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3027  KMP_CPU_FREE(sumMask);
3028 }
3029 
3030 /*-----------------------------------------------------------------------------
3031 Re-parse the OMP_PLACES proc id list, forming the newMasks for the different
3032 places. Again, Here is the grammar:
3033 
3034 place_list := place
3035 place_list := place , place_list
3036 place := num
3037 place := place : num
3038 place := place : num : signed
3039 place := { subplacelist }
3040 place := ! place // (lowest priority)
3041 subplace_list := subplace
3042 subplace_list := subplace , subplace_list
3043 subplace := num
3044 subplace := num : num
3045 subplace := num : num : signed
3046 signed := num
3047 signed := + signed
3048 signed := - signed
3049 -----------------------------------------------------------------------------*/
3050 static void __kmp_process_subplace_list(const char **scan,
3051  kmp_affin_mask_t *osId2Mask,
3052  int maxOsId, kmp_affin_mask_t *tempMask,
3053  int *setSize) {
3054  const char *next;
3055 
3056  for (;;) {
3057  int start, count, stride, i;
3058 
3059  // Read in the starting proc id
3060  SKIP_WS(*scan);
3061  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3062  next = *scan;
3063  SKIP_DIGITS(next);
3064  start = __kmp_str_to_int(*scan, *next);
3065  KMP_ASSERT(start >= 0);
3066  *scan = next;
3067 
3068  // valid follow sets are ',' ':' and '}'
3069  SKIP_WS(*scan);
3070  if (**scan == '}' || **scan == ',') {
3071  if ((start > maxOsId) ||
3072  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3073  if (__kmp_affinity_verbose ||
3074  (__kmp_affinity_warnings &&
3075  (__kmp_affinity_type != affinity_none))) {
3076  KMP_WARNING(AffIgnoreInvalidProcID, start);
3077  }
3078  } else {
3079  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3080  (*setSize)++;
3081  }
3082  if (**scan == '}') {
3083  break;
3084  }
3085  (*scan)++; // skip ','
3086  continue;
3087  }
3088  KMP_ASSERT2(**scan == ':', "bad explicit places list");
3089  (*scan)++; // skip ':'
3090 
3091  // Read count parameter
3092  SKIP_WS(*scan);
3093  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3094  next = *scan;
3095  SKIP_DIGITS(next);
3096  count = __kmp_str_to_int(*scan, *next);
3097  KMP_ASSERT(count >= 0);
3098  *scan = next;
3099 
3100  // valid follow sets are ',' ':' and '}'
3101  SKIP_WS(*scan);
3102  if (**scan == '}' || **scan == ',') {
3103  for (i = 0; i < count; i++) {
3104  if ((start > maxOsId) ||
3105  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3106  if (__kmp_affinity_verbose ||
3107  (__kmp_affinity_warnings &&
3108  (__kmp_affinity_type != affinity_none))) {
3109  KMP_WARNING(AffIgnoreInvalidProcID, start);
3110  }
3111  break; // don't proliferate warnings for large count
3112  } else {
3113  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3114  start++;
3115  (*setSize)++;
3116  }
3117  }
3118  if (**scan == '}') {
3119  break;
3120  }
3121  (*scan)++; // skip ','
3122  continue;
3123  }
3124  KMP_ASSERT2(**scan == ':', "bad explicit places list");
3125  (*scan)++; // skip ':'
3126 
3127  // Read stride parameter
3128  int sign = +1;
3129  for (;;) {
3130  SKIP_WS(*scan);
3131  if (**scan == '+') {
3132  (*scan)++; // skip '+'
3133  continue;
3134  }
3135  if (**scan == '-') {
3136  sign *= -1;
3137  (*scan)++; // skip '-'
3138  continue;
3139  }
3140  break;
3141  }
3142  SKIP_WS(*scan);
3143  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3144  next = *scan;
3145  SKIP_DIGITS(next);
3146  stride = __kmp_str_to_int(*scan, *next);
3147  KMP_ASSERT(stride >= 0);
3148  *scan = next;
3149  stride *= sign;
3150 
3151  // valid follow sets are ',' and '}'
3152  SKIP_WS(*scan);
3153  if (**scan == '}' || **scan == ',') {
3154  for (i = 0; i < count; i++) {
3155  if ((start > maxOsId) ||
3156  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3157  if (__kmp_affinity_verbose ||
3158  (__kmp_affinity_warnings &&
3159  (__kmp_affinity_type != affinity_none))) {
3160  KMP_WARNING(AffIgnoreInvalidProcID, start);
3161  }
3162  break; // don't proliferate warnings for large count
3163  } else {
3164  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3165  start += stride;
3166  (*setSize)++;
3167  }
3168  }
3169  if (**scan == '}') {
3170  break;
3171  }
3172  (*scan)++; // skip ','
3173  continue;
3174  }
3175 
3176  KMP_ASSERT2(0, "bad explicit places list");
3177  }
3178 }
3179 
3180 static void __kmp_process_place(const char **scan, kmp_affin_mask_t *osId2Mask,
3181  int maxOsId, kmp_affin_mask_t *tempMask,
3182  int *setSize) {
3183  const char *next;
3184 
3185  // valid follow sets are '{' '!' and num
3186  SKIP_WS(*scan);
3187  if (**scan == '{') {
3188  (*scan)++; // skip '{'
3189  __kmp_process_subplace_list(scan, osId2Mask, maxOsId, tempMask, setSize);
3190  KMP_ASSERT2(**scan == '}', "bad explicit places list");
3191  (*scan)++; // skip '}'
3192  } else if (**scan == '!') {
3193  (*scan)++; // skip '!'
3194  __kmp_process_place(scan, osId2Mask, maxOsId, tempMask, setSize);
3195  KMP_CPU_COMPLEMENT(maxOsId, tempMask);
3196  } else if ((**scan >= '0') && (**scan <= '9')) {
3197  next = *scan;
3198  SKIP_DIGITS(next);
3199  int num = __kmp_str_to_int(*scan, *next);
3200  KMP_ASSERT(num >= 0);
3201  if ((num > maxOsId) ||
3202  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
3203  if (__kmp_affinity_verbose ||
3204  (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) {
3205  KMP_WARNING(AffIgnoreInvalidProcID, num);
3206  }
3207  } else {
3208  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, num));
3209  (*setSize)++;
3210  }
3211  *scan = next; // skip num
3212  } else {
3213  KMP_ASSERT2(0, "bad explicit places list");
3214  }
3215 }
3216 
3217 // static void
3218 void __kmp_affinity_process_placelist(kmp_affin_mask_t **out_masks,
3219  unsigned int *out_numMasks,
3220  const char *placelist,
3221  kmp_affin_mask_t *osId2Mask,
3222  int maxOsId) {
3223  int i, j, count, stride, sign;
3224  const char *scan = placelist;
3225  const char *next = placelist;
3226 
3227  numNewMasks = 2;
3228  KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
3229  nextNewMask = 0;
3230 
3231  // tempMask is modified based on the previous or initial
3232  // place to form the current place
3233  // previousMask contains the previous place
3234  kmp_affin_mask_t *tempMask;
3235  kmp_affin_mask_t *previousMask;
3236  KMP_CPU_ALLOC(tempMask);
3237  KMP_CPU_ZERO(tempMask);
3238  KMP_CPU_ALLOC(previousMask);
3239  KMP_CPU_ZERO(previousMask);
3240  int setSize = 0;
3241 
3242  for (;;) {
3243  __kmp_process_place(&scan, osId2Mask, maxOsId, tempMask, &setSize);
3244 
3245  // valid follow sets are ',' ':' and EOL
3246  SKIP_WS(scan);
3247  if (*scan == '\0' || *scan == ',') {
3248  if (setSize > 0) {
3249  ADD_MASK(tempMask);
3250  }
3251  KMP_CPU_ZERO(tempMask);
3252  setSize = 0;
3253  if (*scan == '\0') {
3254  break;
3255  }
3256  scan++; // skip ','
3257  continue;
3258  }
3259 
3260  KMP_ASSERT2(*scan == ':', "bad explicit places list");
3261  scan++; // skip ':'
3262 
3263  // Read count parameter
3264  SKIP_WS(scan);
3265  KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
3266  next = scan;
3267  SKIP_DIGITS(next);
3268  count = __kmp_str_to_int(scan, *next);
3269  KMP_ASSERT(count >= 0);
3270  scan = next;
3271 
3272  // valid follow sets are ',' ':' and EOL
3273  SKIP_WS(scan);
3274  if (*scan == '\0' || *scan == ',') {
3275  stride = +1;
3276  } else {
3277  KMP_ASSERT2(*scan == ':', "bad explicit places list");
3278  scan++; // skip ':'
3279 
3280  // Read stride parameter
3281  sign = +1;
3282  for (;;) {
3283  SKIP_WS(scan);
3284  if (*scan == '+') {
3285  scan++; // skip '+'
3286  continue;
3287  }
3288  if (*scan == '-') {
3289  sign *= -1;
3290  scan++; // skip '-'
3291  continue;
3292  }
3293  break;
3294  }
3295  SKIP_WS(scan);
3296  KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
3297  next = scan;
3298  SKIP_DIGITS(next);
3299  stride = __kmp_str_to_int(scan, *next);
3300  KMP_DEBUG_ASSERT(stride >= 0);
3301  scan = next;
3302  stride *= sign;
3303  }
3304 
3305  // Add places determined by initial_place : count : stride
3306  for (i = 0; i < count; i++) {
3307  if (setSize == 0) {
3308  break;
3309  }
3310  // Add the current place, then build the next place (tempMask) from that
3311  KMP_CPU_COPY(previousMask, tempMask);
3312  ADD_MASK(previousMask);
3313  KMP_CPU_ZERO(tempMask);
3314  setSize = 0;
3315  KMP_CPU_SET_ITERATE(j, previousMask) {
3316  if (!KMP_CPU_ISSET(j, previousMask)) {
3317  continue;
3318  }
3319  if ((j + stride > maxOsId) || (j + stride < 0) ||
3320  (!KMP_CPU_ISSET(j, __kmp_affin_fullMask)) ||
3321  (!KMP_CPU_ISSET(j + stride,
3322  KMP_CPU_INDEX(osId2Mask, j + stride)))) {
3323  if ((__kmp_affinity_verbose ||
3324  (__kmp_affinity_warnings &&
3325  (__kmp_affinity_type != affinity_none))) &&
3326  i < count - 1) {
3327  KMP_WARNING(AffIgnoreInvalidProcID, j + stride);
3328  }
3329  continue;
3330  }
3331  KMP_CPU_SET(j + stride, tempMask);
3332  setSize++;
3333  }
3334  }
3335  KMP_CPU_ZERO(tempMask);
3336  setSize = 0;
3337 
3338  // valid follow sets are ',' and EOL
3339  SKIP_WS(scan);
3340  if (*scan == '\0') {
3341  break;
3342  }
3343  if (*scan == ',') {
3344  scan++; // skip ','
3345  continue;
3346  }
3347 
3348  KMP_ASSERT2(0, "bad explicit places list");
3349  }
3350 
3351  *out_numMasks = nextNewMask;
3352  if (nextNewMask == 0) {
3353  *out_masks = NULL;
3354  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3355  return;
3356  }
3357  KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
3358  KMP_CPU_FREE(tempMask);
3359  KMP_CPU_FREE(previousMask);
3360  for (i = 0; i < nextNewMask; i++) {
3361  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
3362  kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
3363  KMP_CPU_COPY(dest, src);
3364  }
3365  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3366 }
3367 
3368 #undef ADD_MASK
3369 #undef ADD_MASK_OSID
3370 
3371 #if KMP_USE_HWLOC
3372 static int __kmp_hwloc_skip_PUs_obj(hwloc_topology_t t, hwloc_obj_t o) {
3373  // skip PUs descendants of the object o
3374  int skipped = 0;
3375  hwloc_obj_t hT = NULL;
3376  int N = __kmp_hwloc_count_children_by_type(t, o, HWLOC_OBJ_PU, &hT);
3377  for (int i = 0; i < N; ++i) {
3378  KMP_DEBUG_ASSERT(hT);
3379  unsigned idx = hT->os_index;
3380  if (KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
3381  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
3382  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
3383  ++skipped;
3384  }
3385  hT = hwloc_get_next_obj_by_type(t, HWLOC_OBJ_PU, hT);
3386  }
3387  return skipped; // count number of skipped units
3388 }
3389 
3390 static int __kmp_hwloc_obj_has_PUs(hwloc_topology_t t, hwloc_obj_t o) {
3391  // check if obj has PUs present in fullMask
3392  hwloc_obj_t hT = NULL;
3393  int N = __kmp_hwloc_count_children_by_type(t, o, HWLOC_OBJ_PU, &hT);
3394  for (int i = 0; i < N; ++i) {
3395  KMP_DEBUG_ASSERT(hT);
3396  unsigned idx = hT->os_index;
3397  if (KMP_CPU_ISSET(idx, __kmp_affin_fullMask))
3398  return 1; // found PU
3399  hT = hwloc_get_next_obj_by_type(t, HWLOC_OBJ_PU, hT);
3400  }
3401  return 0; // no PUs found
3402 }
3403 #endif // KMP_USE_HWLOC
3404 
3405 static void __kmp_apply_thread_places(AddrUnsPair **pAddr, int depth) {
3406  AddrUnsPair *newAddr;
3407  if (__kmp_hws_requested == 0)
3408  goto _exit; // no topology limiting actions requested, exit
3409 #if KMP_USE_HWLOC
3410  if (__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC) {
3411  // Number of subobjects calculated dynamically, this works fine for
3412  // any non-uniform topology.
3413  // L2 cache objects are determined by depth, other objects - by type.
3414  hwloc_topology_t tp = __kmp_hwloc_topology;
3415  int nS = 0, nN = 0, nL = 0, nC = 0,
3416  nT = 0; // logical index including skipped
3417  int nCr = 0, nTr = 0; // number of requested units
3418  int nPkg = 0, nCo = 0, n_new = 0, n_old = 0, nCpP = 0, nTpC = 0; // counters
3419  hwloc_obj_t hT, hC, hL, hN, hS; // hwloc objects (pointers to)
3420  int L2depth, idx;
3421 
3422  // check support of extensions ----------------------------------
3423  int numa_support = 0, tile_support = 0;
3424  if (__kmp_pu_os_idx)
3425  hT = hwloc_get_pu_obj_by_os_index(tp,
3426  __kmp_pu_os_idx[__kmp_avail_proc - 1]);
3427  else
3428  hT = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PU, __kmp_avail_proc - 1);
3429  if (hT == NULL) { // something's gone wrong
3430  KMP_WARNING(AffHWSubsetUnsupported);
3431  goto _exit;
3432  }
3433  // check NUMA node
3434  hN = hwloc_get_ancestor_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hT);
3435  hS = hwloc_get_ancestor_obj_by_type(tp, HWLOC_OBJ_PACKAGE, hT);
3436  if (hN != NULL && hN->depth > hS->depth) {
3437  numa_support = 1; // 1 in case socket includes node(s)
3438  } else if (__kmp_hws_node.num > 0) {
3439  // don't support sockets inside NUMA node (no such HW found for testing)
3440  KMP_WARNING(AffHWSubsetUnsupported);
3441  goto _exit;
3442  }
3443  // check L2 cahce, get object by depth because of multiple caches
3444  L2depth = hwloc_get_cache_type_depth(tp, 2, HWLOC_OBJ_CACHE_UNIFIED);
3445  hL = hwloc_get_ancestor_obj_by_depth(tp, L2depth, hT);
3446  if (hL != NULL &&
3447  __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE, &hC) > 1) {
3448  tile_support = 1; // no sense to count L2 if it includes single core
3449  } else if (__kmp_hws_tile.num > 0) {
3450  if (__kmp_hws_core.num == 0) {
3451  __kmp_hws_core = __kmp_hws_tile; // replace L2 with core
3452  __kmp_hws_tile.num = 0;
3453  } else {
3454  // L2 and core are both requested, but represent same object
3455  KMP_WARNING(AffHWSubsetInvalid);
3456  goto _exit;
3457  }
3458  }
3459  // end of check of extensions -----------------------------------
3460 
3461  // fill in unset items, validate settings -----------------------
3462  if (__kmp_hws_socket.num == 0)
3463  __kmp_hws_socket.num = nPackages; // use all available sockets
3464  if (__kmp_hws_socket.offset >= nPackages) {
3465  KMP_WARNING(AffHWSubsetManySockets);
3466  goto _exit;
3467  }
3468  if (numa_support) {
3469  hN = NULL;
3470  int NN = __kmp_hwloc_count_children_by_type(tp, hS, HWLOC_OBJ_NUMANODE,
3471  &hN); // num nodes in socket
3472  if (__kmp_hws_node.num == 0)
3473  __kmp_hws_node.num = NN; // use all available nodes
3474  if (__kmp_hws_node.offset >= NN) {
3475  KMP_WARNING(AffHWSubsetManyNodes);
3476  goto _exit;
3477  }
3478  if (tile_support) {
3479  // get num tiles in node
3480  int NL = __kmp_hwloc_count_children_by_depth(tp, hN, L2depth, &hL);
3481  if (__kmp_hws_tile.num == 0) {
3482  __kmp_hws_tile.num = NL + 1;
3483  } // use all available tiles, some node may have more tiles, thus +1
3484  if (__kmp_hws_tile.offset >= NL) {
3485  KMP_WARNING(AffHWSubsetManyTiles);
3486  goto _exit;
3487  }
3488  int NC = __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE,
3489  &hC); // num cores in tile
3490  if (__kmp_hws_core.num == 0)
3491  __kmp_hws_core.num = NC; // use all available cores
3492  if (__kmp_hws_core.offset >= NC) {
3493  KMP_WARNING(AffHWSubsetManyCores);
3494  goto _exit;
3495  }
3496  } else { // tile_support
3497  int NC = __kmp_hwloc_count_children_by_type(tp, hN, HWLOC_OBJ_CORE,
3498  &hC); // num cores in node
3499  if (__kmp_hws_core.num == 0)
3500  __kmp_hws_core.num = NC; // use all available cores
3501  if (__kmp_hws_core.offset >= NC) {
3502  KMP_WARNING(AffHWSubsetManyCores);
3503  goto _exit;
3504  }
3505  } // tile_support
3506  } else { // numa_support
3507  if (tile_support) {
3508  // get num tiles in socket
3509  int NL = __kmp_hwloc_count_children_by_depth(tp, hS, L2depth, &hL);
3510  if (__kmp_hws_tile.num == 0)
3511  __kmp_hws_tile.num = NL; // use all available tiles
3512  if (__kmp_hws_tile.offset >= NL) {
3513  KMP_WARNING(AffHWSubsetManyTiles);
3514  goto _exit;
3515  }
3516  int NC = __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE,
3517  &hC); // num cores in tile
3518  if (__kmp_hws_core.num == 0)
3519  __kmp_hws_core.num = NC; // use all available cores
3520  if (__kmp_hws_core.offset >= NC) {
3521  KMP_WARNING(AffHWSubsetManyCores);
3522  goto _exit;
3523  }
3524  } else { // tile_support
3525  int NC = __kmp_hwloc_count_children_by_type(tp, hS, HWLOC_OBJ_CORE,
3526  &hC); // num cores in socket
3527  if (__kmp_hws_core.num == 0)
3528  __kmp_hws_core.num = NC; // use all available cores
3529  if (__kmp_hws_core.offset >= NC) {
3530  KMP_WARNING(AffHWSubsetManyCores);
3531  goto _exit;
3532  }
3533  } // tile_support
3534  }
3535  if (__kmp_hws_proc.num == 0)
3536  __kmp_hws_proc.num = __kmp_nThreadsPerCore; // use all available procs
3537  if (__kmp_hws_proc.offset >= __kmp_nThreadsPerCore) {
3538  KMP_WARNING(AffHWSubsetManyProcs);
3539  goto _exit;
3540  }
3541  // end of validation --------------------------------------------
3542 
3543  if (pAddr) // pAddr is NULL in case of affinity_none
3544  newAddr = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) *
3545  __kmp_avail_proc); // max size
3546  // main loop to form HW subset ----------------------------------
3547  hS = NULL;
3548  int NP = hwloc_get_nbobjs_by_type(tp, HWLOC_OBJ_PACKAGE);
3549  for (int s = 0; s < NP; ++s) {
3550  // Check Socket -----------------------------------------------
3551  hS = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PACKAGE, hS);
3552  if (!__kmp_hwloc_obj_has_PUs(tp, hS))
3553  continue; // skip socket if all PUs are out of fullMask
3554  ++nS; // only count objects those have PUs in affinity mask
3555  if (nS <= __kmp_hws_socket.offset ||
3556  nS > __kmp_hws_socket.num + __kmp_hws_socket.offset) {
3557  n_old += __kmp_hwloc_skip_PUs_obj(tp, hS); // skip socket
3558  continue; // move to next socket
3559  }
3560  nCr = 0; // count number of cores per socket
3561  // socket requested, go down the topology tree
3562  // check 4 cases: (+NUMA+Tile), (+NUMA-Tile), (-NUMA+Tile), (-NUMA-Tile)
3563  if (numa_support) {
3564  nN = 0;
3565  hN = NULL;
3566  // num nodes in current socket
3567  int NN =
3568  __kmp_hwloc_count_children_by_type(tp, hS, HWLOC_OBJ_NUMANODE, &hN);
3569  for (int n = 0; n < NN; ++n) {
3570  // Check NUMA Node ----------------------------------------
3571  if (!__kmp_hwloc_obj_has_PUs(tp, hN)) {
3572  hN = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hN);
3573  continue; // skip node if all PUs are out of fullMask
3574  }
3575  ++nN;
3576  if (nN <= __kmp_hws_node.offset ||
3577  nN > __kmp_hws_node.num + __kmp_hws_node.offset) {
3578  // skip node as not requested
3579  n_old += __kmp_hwloc_skip_PUs_obj(tp, hN); // skip node
3580  hN = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hN);
3581  continue; // move to next node
3582  }
3583  // node requested, go down the topology tree
3584  if (tile_support) {
3585  nL = 0;
3586  hL = NULL;
3587  int NL = __kmp_hwloc_count_children_by_depth(tp, hN, L2depth, &hL);
3588  for (int l = 0; l < NL; ++l) {
3589  // Check L2 (tile) ------------------------------------
3590  if (!__kmp_hwloc_obj_has_PUs(tp, hL)) {
3591  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3592  continue; // skip tile if all PUs are out of fullMask
3593  }
3594  ++nL;
3595  if (nL <= __kmp_hws_tile.offset ||
3596  nL > __kmp_hws_tile.num + __kmp_hws_tile.offset) {
3597  // skip tile as not requested
3598  n_old += __kmp_hwloc_skip_PUs_obj(tp, hL); // skip tile
3599  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3600  continue; // move to next tile
3601  }
3602  // tile requested, go down the topology tree
3603  nC = 0;
3604  hC = NULL;
3605  // num cores in current tile
3606  int NC = __kmp_hwloc_count_children_by_type(tp, hL,
3607  HWLOC_OBJ_CORE, &hC);
3608  for (int c = 0; c < NC; ++c) {
3609  // Check Core ---------------------------------------
3610  if (!__kmp_hwloc_obj_has_PUs(tp, hC)) {
3611  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3612  continue; // skip core if all PUs are out of fullMask
3613  }
3614  ++nC;
3615  if (nC <= __kmp_hws_core.offset ||
3616  nC > __kmp_hws_core.num + __kmp_hws_core.offset) {
3617  // skip node as not requested
3618  n_old += __kmp_hwloc_skip_PUs_obj(tp, hC); // skip core
3619  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3620  continue; // move to next node
3621  }
3622  // core requested, go down to PUs
3623  nT = 0;
3624  nTr = 0;
3625  hT = NULL;
3626  // num procs in current core
3627  int NT = __kmp_hwloc_count_children_by_type(tp, hC,
3628  HWLOC_OBJ_PU, &hT);
3629  for (int t = 0; t < NT; ++t) {
3630  // Check PU ---------------------------------------
3631  idx = hT->os_index;
3632  if (!KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
3633  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3634  continue; // skip PU if not in fullMask
3635  }
3636  ++nT;
3637  if (nT <= __kmp_hws_proc.offset ||
3638  nT > __kmp_hws_proc.num + __kmp_hws_proc.offset) {
3639  // skip PU
3640  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
3641  ++n_old;
3642  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
3643  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3644  continue; // move to next node
3645  }
3646  ++nTr;
3647  if (pAddr) // collect requested thread's data
3648  newAddr[n_new] = (*pAddr)[n_old];
3649  ++n_new;
3650  ++n_old;
3651  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3652  } // threads loop
3653  if (nTr > 0) {
3654  ++nCr; // num cores per socket
3655  ++nCo; // total num cores
3656  if (nTr > nTpC)
3657  nTpC = nTr; // calc max threads per core
3658  }
3659  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3660  } // cores loop
3661  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3662  } // tiles loop
3663  } else { // tile_support
3664  // no tiles, check cores
3665  nC = 0;
3666  hC = NULL;
3667  // num cores in current node
3668  int NC =
3669  __kmp_hwloc_count_children_by_type(tp, hN, HWLOC_OBJ_CORE, &hC);
3670  for (int c = 0; c < NC; ++c) {
3671  // Check Core ---------------------------------------
3672  if (!__kmp_hwloc_obj_has_PUs(tp, hC)) {
3673  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3674  continue; // skip core if all PUs are out of fullMask
3675  }
3676  ++nC;
3677  if (nC <= __kmp_hws_core.offset ||
3678  nC > __kmp_hws_core.num + __kmp_hws_core.offset) {
3679  // skip node as not requested
3680  n_old += __kmp_hwloc_skip_PUs_obj(tp, hC); // skip core
3681  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3682  continue; // move to next node
3683  }
3684  // core requested, go down to PUs
3685  nT = 0;
3686  nTr = 0;
3687  hT = NULL;
3688  int NT =
3689  __kmp_hwloc_count_children_by_type(tp, hC, HWLOC_OBJ_PU, &hT);
3690  for (int t = 0; t < NT; ++t) {
3691  // Check PU ---------------------------------------
3692  idx = hT->os_index;
3693  if (!KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
3694  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3695  continue; // skip PU if not in fullMask
3696  }
3697  ++nT;
3698  if (nT <= __kmp_hws_proc.offset ||
3699  nT > __kmp_hws_proc.num + __kmp_hws_proc.offset) {
3700  // skip PU
3701  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
3702  ++n_old;
3703  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
3704  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3705  continue; // move to next node
3706  }
3707  ++nTr;
3708  if (pAddr) // collect requested thread's data
3709  newAddr[n_new] = (*pAddr)[n_old];
3710  ++n_new;
3711  ++n_old;
3712  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3713  } // threads loop
3714  if (nTr > 0) {
3715  ++nCr; // num cores per socket
3716  ++nCo; // total num cores
3717  if (nTr > nTpC)
3718  nTpC = nTr; // calc max threads per core
3719  }
3720  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3721  } // cores loop
3722  } // tiles support
3723  hN = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hN);
3724  } // nodes loop
3725  } else { // numa_support
3726  // no NUMA support
3727  if (tile_support) {
3728  nL = 0;
3729  hL = NULL;
3730  // num tiles in current socket
3731  int NL = __kmp_hwloc_count_children_by_depth(tp, hS, L2depth, &hL);
3732  for (int l = 0; l < NL; ++l) {
3733  // Check L2 (tile) ------------------------------------
3734  if (!__kmp_hwloc_obj_has_PUs(tp, hL)) {
3735  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3736  continue; // skip tile if all PUs are out of fullMask
3737  }
3738  ++nL;
3739  if (nL <= __kmp_hws_tile.offset ||
3740  nL > __kmp_hws_tile.num + __kmp_hws_tile.offset) {
3741  // skip tile as not requested
3742  n_old += __kmp_hwloc_skip_PUs_obj(tp, hL); // skip tile
3743  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3744  continue; // move to next tile
3745  }
3746  // tile requested, go down the topology tree
3747  nC = 0;
3748  hC = NULL;
3749  // num cores per tile
3750  int NC =
3751  __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE, &hC);
3752  for (int c = 0; c < NC; ++c) {
3753  // Check Core ---------------------------------------
3754  if (!__kmp_hwloc_obj_has_PUs(tp, hC)) {
3755  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3756  continue; // skip core if all PUs are out of fullMask
3757  }
3758  ++nC;
3759  if (nC <= __kmp_hws_core.offset ||
3760  nC > __kmp_hws_core.num + __kmp_hws_core.offset) {
3761  // skip node as not requested
3762  n_old += __kmp_hwloc_skip_PUs_obj(tp, hC); // skip core
3763  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3764  continue; // move to next node
3765  }
3766  // core requested, go down to PUs
3767  nT = 0;
3768  nTr = 0;
3769  hT = NULL;
3770  // num procs per core
3771  int NT =
3772  __kmp_hwloc_count_children_by_type(tp, hC, HWLOC_OBJ_PU, &hT);
3773  for (int t = 0; t < NT; ++t) {
3774  // Check PU ---------------------------------------
3775  idx = hT->os_index;
3776  if (!KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
3777  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3778  continue; // skip PU if not in fullMask
3779  }
3780  ++nT;
3781  if (nT <= __kmp_hws_proc.offset ||
3782  nT > __kmp_hws_proc.num + __kmp_hws_proc.offset) {
3783  // skip PU
3784  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
3785  ++n_old;
3786  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
3787  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3788  continue; // move to next node
3789  }
3790  ++nTr;
3791  if (pAddr) // collect requested thread's data
3792  newAddr[n_new] = (*pAddr)[n_old];
3793  ++n_new;
3794  ++n_old;
3795  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3796  } // threads loop
3797  if (nTr > 0) {
3798  ++nCr; // num cores per socket
3799  ++nCo; // total num cores
3800  if (nTr > nTpC)
3801  nTpC = nTr; // calc max threads per core
3802  }
3803  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3804  } // cores loop
3805  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3806  } // tiles loop
3807  } else { // tile_support
3808  // no tiles, check cores
3809  nC = 0;
3810  hC = NULL;
3811  // num cores in socket
3812  int NC =
3813  __kmp_hwloc_count_children_by_type(tp, hS, HWLOC_OBJ_CORE, &hC);
3814  for (int c = 0; c < NC; ++c) {
3815  // Check Core -------------------------------------------
3816  if (!__kmp_hwloc_obj_has_PUs(tp, hC)) {
3817  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3818  continue; // skip core if all PUs are out of fullMask
3819  }
3820  ++nC;
3821  if (nC <= __kmp_hws_core.offset ||
3822  nC > __kmp_hws_core.num + __kmp_hws_core.offset) {
3823  // skip node as not requested
3824  n_old += __kmp_hwloc_skip_PUs_obj(tp, hC); // skip core
3825  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3826  continue; // move to next node
3827  }
3828  // core requested, go down to PUs
3829  nT = 0;
3830  nTr = 0;
3831  hT = NULL;
3832  // num procs per core
3833  int NT =
3834  __kmp_hwloc_count_children_by_type(tp, hC, HWLOC_OBJ_PU, &hT);
3835  for (int t = 0; t < NT; ++t) {
3836  // Check PU ---------------------------------------
3837  idx = hT->os_index;
3838  if (!KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
3839  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3840  continue; // skip PU if not in fullMask
3841  }
3842  ++nT;
3843  if (nT <= __kmp_hws_proc.offset ||
3844  nT > __kmp_hws_proc.num + __kmp_hws_proc.offset) {
3845  // skip PU
3846  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
3847  ++n_old;
3848  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
3849  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3850  continue; // move to next node
3851  }
3852  ++nTr;
3853  if (pAddr) // collect requested thread's data
3854  newAddr[n_new] = (*pAddr)[n_old];
3855  ++n_new;
3856  ++n_old;
3857  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3858  } // threads loop
3859  if (nTr > 0) {
3860  ++nCr; // num cores per socket
3861  ++nCo; // total num cores
3862  if (nTr > nTpC)
3863  nTpC = nTr; // calc max threads per core
3864  }
3865  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3866  } // cores loop
3867  } // tiles support
3868  } // numa_support
3869  if (nCr > 0) { // found cores?
3870  ++nPkg; // num sockets
3871  if (nCr > nCpP)
3872  nCpP = nCr; // calc max cores per socket
3873  }
3874  } // sockets loop
3875 
3876  // check the subset is valid
3877  KMP_DEBUG_ASSERT(n_old == __kmp_avail_proc);
3878  KMP_DEBUG_ASSERT(nPkg > 0);
3879  KMP_DEBUG_ASSERT(nCpP > 0);
3880  KMP_DEBUG_ASSERT(nTpC > 0);
3881  KMP_DEBUG_ASSERT(nCo > 0);
3882  KMP_DEBUG_ASSERT(nPkg <= nPackages);
3883  KMP_DEBUG_ASSERT(nCpP <= nCoresPerPkg);
3884  KMP_DEBUG_ASSERT(nTpC <= __kmp_nThreadsPerCore);
3885  KMP_DEBUG_ASSERT(nCo <= __kmp_ncores);
3886 
3887  nPackages = nPkg; // correct num sockets
3888  nCoresPerPkg = nCpP; // correct num cores per socket
3889  __kmp_nThreadsPerCore = nTpC; // correct num threads per core
3890  __kmp_avail_proc = n_new; // correct num procs
3891  __kmp_ncores = nCo; // correct num cores
3892  // hwloc topology method end
3893  } else
3894 #endif // KMP_USE_HWLOC
3895  {
3896  int n_old = 0, n_new = 0, proc_num = 0;
3897  if (__kmp_hws_node.num > 0 || __kmp_hws_tile.num > 0) {
3898  KMP_WARNING(AffHWSubsetNoHWLOC);
3899  goto _exit;
3900  }
3901  if (__kmp_hws_socket.num == 0)
3902  __kmp_hws_socket.num = nPackages; // use all available sockets
3903  if (__kmp_hws_core.num == 0)
3904  __kmp_hws_core.num = nCoresPerPkg; // use all available cores
3905  if (__kmp_hws_proc.num == 0 || __kmp_hws_proc.num > __kmp_nThreadsPerCore)
3906  __kmp_hws_proc.num = __kmp_nThreadsPerCore; // use all HW contexts
3907  if (!__kmp_affinity_uniform_topology()) {
3908  KMP_WARNING(AffHWSubsetNonUniform);
3909  goto _exit; // don't support non-uniform topology
3910  }
3911  if (depth > 3) {
3912  KMP_WARNING(AffHWSubsetNonThreeLevel);
3913  goto _exit; // don't support not-3-level topology
3914  }
3915  if (__kmp_hws_socket.offset + __kmp_hws_socket.num > nPackages) {
3916  KMP_WARNING(AffHWSubsetManySockets);
3917  goto _exit;
3918  }
3919  if (__kmp_hws_core.offset + __kmp_hws_core.num > nCoresPerPkg) {
3920  KMP_WARNING(AffHWSubsetManyCores);
3921  goto _exit;
3922  }
3923  // Form the requested subset
3924  if (pAddr) // pAddr is NULL in case of affinity_none
3925  newAddr = (AddrUnsPair *)__kmp_allocate(
3926  sizeof(AddrUnsPair) * __kmp_hws_socket.num * __kmp_hws_core.num *
3927  __kmp_hws_proc.num);
3928  for (int i = 0; i < nPackages; ++i) {
3929  if (i < __kmp_hws_socket.offset ||
3930  i >= __kmp_hws_socket.offset + __kmp_hws_socket.num) {
3931  // skip not-requested socket
3932  n_old += nCoresPerPkg * __kmp_nThreadsPerCore;
3933  if (__kmp_pu_os_idx != NULL) {
3934  // walk through skipped socket
3935  for (int j = 0; j < nCoresPerPkg; ++j) {
3936  for (int k = 0; k < __kmp_nThreadsPerCore; ++k) {
3937  KMP_CPU_CLR(__kmp_pu_os_idx[proc_num], __kmp_affin_fullMask);
3938  ++proc_num;
3939  }
3940  }
3941  }
3942  } else {
3943  // walk through requested socket
3944  for (int j = 0; j < nCoresPerPkg; ++j) {
3945  if (j < __kmp_hws_core.offset ||
3946  j >= __kmp_hws_core.offset +
3947  __kmp_hws_core.num) { // skip not-requested core
3948  n_old += __kmp_nThreadsPerCore;
3949  if (__kmp_pu_os_idx != NULL) {
3950  for (int k = 0; k < __kmp_nThreadsPerCore; ++k) {
3951  KMP_CPU_CLR(__kmp_pu_os_idx[proc_num], __kmp_affin_fullMask);
3952  ++proc_num;
3953  }
3954  }
3955  } else {
3956  // walk through requested core
3957  for (int k = 0; k < __kmp_nThreadsPerCore; ++k) {
3958  if (k < __kmp_hws_proc.num) {
3959  if (pAddr) // collect requested thread's data
3960  newAddr[n_new] = (*pAddr)[n_old];
3961  n_new++;
3962  } else {
3963  if (__kmp_pu_os_idx != NULL)
3964  KMP_CPU_CLR(__kmp_pu_os_idx[proc_num], __kmp_affin_fullMask);
3965  }
3966  n_old++;
3967  ++proc_num;
3968  }
3969  }
3970  }
3971  }
3972  }
3973  KMP_DEBUG_ASSERT(n_old == nPackages * nCoresPerPkg * __kmp_nThreadsPerCore);
3974  KMP_DEBUG_ASSERT(n_new ==
3975  __kmp_hws_socket.num * __kmp_hws_core.num *
3976  __kmp_hws_proc.num);
3977  nPackages = __kmp_hws_socket.num; // correct nPackages
3978  nCoresPerPkg = __kmp_hws_core.num; // correct nCoresPerPkg
3979  __kmp_nThreadsPerCore = __kmp_hws_proc.num; // correct __kmp_nThreadsPerCore
3980  __kmp_avail_proc = n_new; // correct avail_proc
3981  __kmp_ncores = nPackages * __kmp_hws_core.num; // correct ncores
3982  } // non-hwloc topology method
3983  if (pAddr) {
3984  __kmp_free(*pAddr);
3985  *pAddr = newAddr; // replace old topology with new one
3986  }
3987  if (__kmp_affinity_verbose) {
3988  char m[KMP_AFFIN_MASK_PRINT_LEN];
3989  __kmp_affinity_print_mask(m, KMP_AFFIN_MASK_PRINT_LEN,
3990  __kmp_affin_fullMask);
3991  if (__kmp_affinity_respect_mask) {
3992  KMP_INFORM(InitOSProcSetRespect, "KMP_HW_SUBSET", m);
3993  } else {
3994  KMP_INFORM(InitOSProcSetNotRespect, "KMP_HW_SUBSET", m);
3995  }
3996  KMP_INFORM(AvailableOSProc, "KMP_HW_SUBSET", __kmp_avail_proc);
3997  kmp_str_buf_t buf;
3998  __kmp_str_buf_init(&buf);
3999  __kmp_str_buf_print(&buf, "%d", nPackages);
4000  KMP_INFORM(TopologyExtra, "KMP_HW_SUBSET", buf.str, nCoresPerPkg,
4001  __kmp_nThreadsPerCore, __kmp_ncores);
4002  __kmp_str_buf_free(&buf);
4003  }
4004 _exit:
4005  if (__kmp_pu_os_idx != NULL) {
4006  __kmp_free(__kmp_pu_os_idx);
4007  __kmp_pu_os_idx = NULL;
4008  }
4009 }
4010 
4011 // This function figures out the deepest level at which there is at least one
4012 // cluster/core with more than one processing unit bound to it.
4013 static int __kmp_affinity_find_core_level(const AddrUnsPair *address2os,
4014  int nprocs, int bottom_level) {
4015  int core_level = 0;
4016 
4017  for (int i = 0; i < nprocs; i++) {
4018  for (int j = bottom_level; j > 0; j--) {
4019  if (address2os[i].first.labels[j] > 0) {
4020  if (core_level < (j - 1)) {
4021  core_level = j - 1;
4022  }
4023  }
4024  }
4025  }
4026  return core_level;
4027 }
4028 
4029 // This function counts number of clusters/cores at given level.
4030 static int __kmp_affinity_compute_ncores(const AddrUnsPair *address2os,
4031  int nprocs, int bottom_level,
4032  int core_level) {
4033  int ncores = 0;
4034  int i, j;
4035 
4036  j = bottom_level;
4037  for (i = 0; i < nprocs; i++) {
4038  for (j = bottom_level; j > core_level; j--) {
4039  if ((i + 1) < nprocs) {
4040  if (address2os[i + 1].first.labels[j] > 0) {
4041  break;
4042  }
4043  }
4044  }
4045  if (j == core_level) {
4046  ncores++;
4047  }
4048  }
4049  if (j > core_level) {
4050  // In case of ( nprocs < __kmp_avail_proc ) we may end too deep and miss one
4051  // core. May occur when called from __kmp_affinity_find_core().
4052  ncores++;
4053  }
4054  return ncores;
4055 }
4056 
4057 // This function finds to which cluster/core given processing unit is bound.
4058 static int __kmp_affinity_find_core(const AddrUnsPair *address2os, int proc,
4059  int bottom_level, int core_level) {
4060  return __kmp_affinity_compute_ncores(address2os, proc + 1, bottom_level,
4061  core_level) -
4062  1;
4063 }
4064 
4065 // This function finds maximal number of processing units bound to a
4066 // cluster/core at given level.
4067 static int __kmp_affinity_max_proc_per_core(const AddrUnsPair *address2os,
4068  int nprocs, int bottom_level,
4069  int core_level) {
4070  int maxprocpercore = 0;
4071 
4072  if (core_level < bottom_level) {
4073  for (int i = 0; i < nprocs; i++) {
4074  int percore = address2os[i].first.labels[core_level + 1] + 1;
4075 
4076  if (percore > maxprocpercore) {
4077  maxprocpercore = percore;
4078  }
4079  }
4080  } else {
4081  maxprocpercore = 1;
4082  }
4083  return maxprocpercore;
4084 }
4085 
4086 static AddrUnsPair *address2os = NULL;
4087 static int *procarr = NULL;
4088 static int __kmp_aff_depth = 0;
4089 
4090 #if KMP_USE_HIER_SCHED
4091 #define KMP_EXIT_AFF_NONE \
4092  KMP_ASSERT(__kmp_affinity_type == affinity_none); \
4093  KMP_ASSERT(address2os == NULL); \
4094  __kmp_apply_thread_places(NULL, 0); \
4095  __kmp_create_affinity_none_places(); \
4096  __kmp_dispatch_set_hierarchy_values(); \
4097  return;
4098 #else
4099 #define KMP_EXIT_AFF_NONE \
4100  KMP_ASSERT(__kmp_affinity_type == affinity_none); \
4101  KMP_ASSERT(address2os == NULL); \
4102  __kmp_apply_thread_places(NULL, 0); \
4103  __kmp_create_affinity_none_places(); \
4104  return;
4105 #endif
4106 
4107 // Create a one element mask array (set of places) which only contains the
4108 // initial process's affinity mask
4109 static void __kmp_create_affinity_none_places() {
4110  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4111  KMP_ASSERT(__kmp_affinity_type == affinity_none);
4112  __kmp_affinity_num_masks = 1;
4113  KMP_CPU_ALLOC_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks);
4114  kmp_affin_mask_t *dest = KMP_CPU_INDEX(__kmp_affinity_masks, 0);
4115  KMP_CPU_COPY(dest, __kmp_affin_fullMask);
4116 }
4117 
4118 static int __kmp_affinity_cmp_Address_child_num(const void *a, const void *b) {
4119  const Address *aa = &(((const AddrUnsPair *)a)->first);
4120  const Address *bb = &(((const AddrUnsPair *)b)->first);
4121  unsigned depth = aa->depth;
4122  unsigned i;
4123  KMP_DEBUG_ASSERT(depth == bb->depth);
4124  KMP_DEBUG_ASSERT((unsigned)__kmp_affinity_compact <= depth);
4125  KMP_DEBUG_ASSERT(__kmp_affinity_compact >= 0);
4126  for (i = 0; i < (unsigned)__kmp_affinity_compact; i++) {
4127  int j = depth - i - 1;
4128  if (aa->childNums[j] < bb->childNums[j])
4129  return -1;
4130  if (aa->childNums[j] > bb->childNums[j])
4131  return 1;
4132  }
4133  for (; i < depth; i++) {
4134  int j = i - __kmp_affinity_compact;
4135  if (aa->childNums[j] < bb->childNums[j])
4136  return -1;
4137  if (aa->childNums[j] > bb->childNums[j])
4138  return 1;
4139  }
4140  return 0;
4141 }
4142 
4143 static void __kmp_aux_affinity_initialize(void) {
4144  if (__kmp_affinity_masks != NULL) {
4145  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4146  return;
4147  }
4148 
4149  // Create the "full" mask - this defines all of the processors that we
4150  // consider to be in the machine model. If respect is set, then it is the
4151  // initialization thread's affinity mask. Otherwise, it is all processors that
4152  // we know about on the machine.
4153  if (__kmp_affin_fullMask == NULL) {
4154  KMP_CPU_ALLOC(__kmp_affin_fullMask);
4155  }
4156  if (KMP_AFFINITY_CAPABLE()) {
4157  if (__kmp_affinity_respect_mask) {
4158  __kmp_get_system_affinity(__kmp_affin_fullMask, TRUE);
4159 
4160  // Count the number of available processors.
4161  unsigned i;
4162  __kmp_avail_proc = 0;
4163  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
4164  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
4165  continue;
4166  }
4167  __kmp_avail_proc++;
4168  }
4169  if (__kmp_avail_proc > __kmp_xproc) {
4170  if (__kmp_affinity_verbose ||
4171  (__kmp_affinity_warnings &&
4172  (__kmp_affinity_type != affinity_none))) {
4173  KMP_WARNING(ErrorInitializeAffinity);
4174  }
4175  __kmp_affinity_type = affinity_none;
4176  KMP_AFFINITY_DISABLE();
4177  return;
4178  }
4179  } else {
4180  __kmp_affinity_entire_machine_mask(__kmp_affin_fullMask);
4181  __kmp_avail_proc = __kmp_xproc;
4182  }
4183  }
4184 
4185  if (__kmp_affinity_gran == affinity_gran_tile &&
4186  // check if user's request is valid
4187  __kmp_affinity_dispatch->get_api_type() == KMPAffinity::NATIVE_OS) {
4188  KMP_WARNING(AffTilesNoHWLOC, "KMP_AFFINITY");
4189  __kmp_affinity_gran = affinity_gran_package;
4190  }
4191 
4192  int depth = -1;
4193  kmp_i18n_id_t msg_id = kmp_i18n_null;
4194 
4195  // For backward compatibility, setting KMP_CPUINFO_FILE =>
4196  // KMP_TOPOLOGY_METHOD=cpuinfo
4197  if ((__kmp_cpuinfo_file != NULL) &&
4198  (__kmp_affinity_top_method == affinity_top_method_all)) {
4199  __kmp_affinity_top_method = affinity_top_method_cpuinfo;
4200  }
4201 
4202  if (__kmp_affinity_top_method == affinity_top_method_all) {
4203  // In the default code path, errors are not fatal - we just try using
4204  // another method. We only emit a warning message if affinity is on, or the
4205  // verbose flag is set, and the nowarnings flag was not set.
4206  const char *file_name = NULL;
4207  int line = 0;
4208 #if KMP_USE_HWLOC
4209  if (depth < 0 &&
4210  __kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC) {
4211  if (__kmp_affinity_verbose) {
4212  KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
4213  }
4214  if (!__kmp_hwloc_error) {
4215  depth = __kmp_affinity_create_hwloc_map(&address2os, &msg_id);
4216  if (depth == 0) {
4217  KMP_EXIT_AFF_NONE;
4218  } else if (depth < 0 && __kmp_affinity_verbose) {
4219  KMP_INFORM(AffIgnoringHwloc, "KMP_AFFINITY");
4220  }
4221  } else if (__kmp_affinity_verbose) {
4222  KMP_INFORM(AffIgnoringHwloc, "KMP_AFFINITY");
4223  }
4224  }
4225 #endif
4226 
4227 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4228 
4229  if (depth < 0) {
4230  if (__kmp_affinity_verbose) {
4231  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(Decodingx2APIC));
4232  }
4233 
4234  file_name = NULL;
4235  depth = __kmp_affinity_create_x2apicid_map(&address2os, &msg_id);
4236  if (depth == 0) {
4237  KMP_EXIT_AFF_NONE;
4238  }
4239 
4240  if (depth < 0) {
4241  if (__kmp_affinity_verbose) {
4242  if (msg_id != kmp_i18n_null) {
4243  KMP_INFORM(AffInfoStrStr, "KMP_AFFINITY",
4244  __kmp_i18n_catgets(msg_id),
4245  KMP_I18N_STR(DecodingLegacyAPIC));
4246  } else {
4247  KMP_INFORM(AffInfoStr, "KMP_AFFINITY",
4248  KMP_I18N_STR(DecodingLegacyAPIC));
4249  }
4250  }
4251 
4252  file_name = NULL;
4253  depth = __kmp_affinity_create_apicid_map(&address2os, &msg_id);
4254  if (depth == 0) {
4255  KMP_EXIT_AFF_NONE;
4256  }
4257  }
4258  }
4259 
4260 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4261 
4262 #if KMP_OS_LINUX
4263 
4264  if (depth < 0) {
4265  if (__kmp_affinity_verbose) {
4266  if (msg_id != kmp_i18n_null) {
4267  KMP_INFORM(AffStrParseFilename, "KMP_AFFINITY",
4268  __kmp_i18n_catgets(msg_id), "/proc/cpuinfo");
4269  } else {
4270  KMP_INFORM(AffParseFilename, "KMP_AFFINITY", "/proc/cpuinfo");
4271  }
4272  }
4273 
4274  FILE *f = fopen("/proc/cpuinfo", "r");
4275  if (f == NULL) {
4276  msg_id = kmp_i18n_str_CantOpenCpuinfo;
4277  } else {
4278  file_name = "/proc/cpuinfo";
4279  depth =
4280  __kmp_affinity_create_cpuinfo_map(&address2os, &line, &msg_id, f);
4281  fclose(f);
4282  if (depth == 0) {
4283  KMP_EXIT_AFF_NONE;
4284  }
4285  }
4286  }
4287 
4288 #endif /* KMP_OS_LINUX */
4289 
4290 #if KMP_GROUP_AFFINITY
4291 
4292  if ((depth < 0) && (__kmp_num_proc_groups > 1)) {
4293  if (__kmp_affinity_verbose) {
4294  KMP_INFORM(AffWindowsProcGroupMap, "KMP_AFFINITY");
4295  }
4296 
4297  depth = __kmp_affinity_create_proc_group_map(&address2os, &msg_id);
4298  KMP_ASSERT(depth != 0);
4299  }
4300 
4301 #endif /* KMP_GROUP_AFFINITY */
4302 
4303  if (depth < 0) {
4304  if (__kmp_affinity_verbose && (msg_id != kmp_i18n_null)) {
4305  if (file_name == NULL) {
4306  KMP_INFORM(UsingFlatOS, __kmp_i18n_catgets(msg_id));
4307  } else if (line == 0) {
4308  KMP_INFORM(UsingFlatOSFile, file_name, __kmp_i18n_catgets(msg_id));
4309  } else {
4310  KMP_INFORM(UsingFlatOSFileLine, file_name, line,
4311  __kmp_i18n_catgets(msg_id));
4312  }
4313  }
4314  // FIXME - print msg if msg_id = kmp_i18n_null ???
4315 
4316  file_name = "";
4317  depth = __kmp_affinity_create_flat_map(&address2os, &msg_id);
4318  if (depth == 0) {
4319  KMP_EXIT_AFF_NONE;
4320  }
4321  KMP_ASSERT(depth > 0);
4322  KMP_ASSERT(address2os != NULL);
4323  }
4324  }
4325 
4326 #if KMP_USE_HWLOC
4327  else if (__kmp_affinity_top_method == affinity_top_method_hwloc) {
4328  KMP_ASSERT(__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC);
4329  if (__kmp_affinity_verbose) {
4330  KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
4331  }
4332  depth = __kmp_affinity_create_hwloc_map(&address2os, &msg_id);
4333  if (depth == 0) {
4334  KMP_EXIT_AFF_NONE;
4335  }
4336  }
4337 #endif // KMP_USE_HWLOC
4338 
4339 // If the user has specified that a particular topology discovery method is to be
4340 // used, then we abort if that method fails. The exception is group affinity,
4341 // which might have been implicitly set.
4342 
4343 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4344 
4345  else if (__kmp_affinity_top_method == affinity_top_method_x2apicid) {
4346  if (__kmp_affinity_verbose) {
4347  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(Decodingx2APIC));
4348  }
4349 
4350  depth = __kmp_affinity_create_x2apicid_map(&address2os, &msg_id);
4351  if (depth == 0) {
4352  KMP_EXIT_AFF_NONE;
4353  }
4354  if (depth < 0) {
4355  KMP_ASSERT(msg_id != kmp_i18n_null);
4356  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4357  }
4358  } else if (__kmp_affinity_top_method == affinity_top_method_apicid) {
4359  if (__kmp_affinity_verbose) {
4360  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(DecodingLegacyAPIC));
4361  }
4362 
4363  depth = __kmp_affinity_create_apicid_map(&address2os, &msg_id);
4364  if (depth == 0) {
4365  KMP_EXIT_AFF_NONE;
4366  }
4367  if (depth < 0) {
4368  KMP_ASSERT(msg_id != kmp_i18n_null);
4369  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4370  }
4371  }
4372 
4373 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4374 
4375  else if (__kmp_affinity_top_method == affinity_top_method_cpuinfo) {
4376  const char *filename;
4377  if (__kmp_cpuinfo_file != NULL) {
4378  filename = __kmp_cpuinfo_file;
4379  } else {
4380  filename = "/proc/cpuinfo";
4381  }
4382 
4383  if (__kmp_affinity_verbose) {
4384  KMP_INFORM(AffParseFilename, "KMP_AFFINITY", filename);
4385  }
4386 
4387  FILE *f = fopen(filename, "r");
4388  if (f == NULL) {
4389  int code = errno;
4390  if (__kmp_cpuinfo_file != NULL) {
4391  __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4392  KMP_HNT(NameComesFrom_CPUINFO_FILE), __kmp_msg_null);
4393  } else {
4394  __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4395  __kmp_msg_null);
4396  }
4397  }
4398  int line = 0;
4399  depth = __kmp_affinity_create_cpuinfo_map(&address2os, &line, &msg_id, f);
4400  fclose(f);
4401  if (depth < 0) {
4402  KMP_ASSERT(msg_id != kmp_i18n_null);
4403  if (line > 0) {
4404  KMP_FATAL(FileLineMsgExiting, filename, line,
4405  __kmp_i18n_catgets(msg_id));
4406  } else {
4407  KMP_FATAL(FileMsgExiting, filename, __kmp_i18n_catgets(msg_id));
4408  }
4409  }
4410  if (__kmp_affinity_type == affinity_none) {
4411  KMP_ASSERT(depth == 0);
4412  KMP_EXIT_AFF_NONE;
4413  }
4414  }
4415 
4416 #if KMP_GROUP_AFFINITY
4417 
4418  else if (__kmp_affinity_top_method == affinity_top_method_group) {
4419  if (__kmp_affinity_verbose) {
4420  KMP_INFORM(AffWindowsProcGroupMap, "KMP_AFFINITY");
4421  }
4422 
4423  depth = __kmp_affinity_create_proc_group_map(&address2os, &msg_id);
4424  KMP_ASSERT(depth != 0);
4425  if (depth < 0) {
4426  KMP_ASSERT(msg_id != kmp_i18n_null);
4427  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4428  }
4429  }
4430 
4431 #endif /* KMP_GROUP_AFFINITY */
4432 
4433  else if (__kmp_affinity_top_method == affinity_top_method_flat) {
4434  if (__kmp_affinity_verbose) {
4435  KMP_INFORM(AffUsingFlatOS, "KMP_AFFINITY");
4436  }
4437 
4438  depth = __kmp_affinity_create_flat_map(&address2os, &msg_id);
4439  if (depth == 0) {
4440  KMP_EXIT_AFF_NONE;
4441  }
4442  // should not fail
4443  KMP_ASSERT(depth > 0);
4444  KMP_ASSERT(address2os != NULL);
4445  }
4446 
4447 #if KMP_USE_HIER_SCHED
4448  __kmp_dispatch_set_hierarchy_values();
4449 #endif
4450 
4451  if (address2os == NULL) {
4452  if (KMP_AFFINITY_CAPABLE() &&
4453  (__kmp_affinity_verbose ||
4454  (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none)))) {
4455  KMP_WARNING(ErrorInitializeAffinity);
4456  }
4457  __kmp_affinity_type = affinity_none;
4458  __kmp_create_affinity_none_places();
4459  KMP_AFFINITY_DISABLE();
4460  return;
4461  }
4462 
4463  if (__kmp_affinity_gran == affinity_gran_tile
4464 #if KMP_USE_HWLOC
4465  && __kmp_tile_depth == 0
4466 #endif
4467  ) {
4468  // tiles requested but not detected, warn user on this
4469  KMP_WARNING(AffTilesNoTiles, "KMP_AFFINITY");
4470  }
4471 
4472  __kmp_apply_thread_places(&address2os, depth);
4473 
4474  // Create the table of masks, indexed by thread Id.
4475  unsigned maxIndex;
4476  unsigned numUnique;
4477  kmp_affin_mask_t *osId2Mask =
4478  __kmp_create_masks(&maxIndex, &numUnique, address2os, __kmp_avail_proc);
4479  if (__kmp_affinity_gran_levels == 0) {
4480  KMP_DEBUG_ASSERT((int)numUnique == __kmp_avail_proc);
4481  }
4482 
4483  // Set the childNums vector in all Address objects. This must be done before
4484  // we can sort using __kmp_affinity_cmp_Address_child_num(), which takes into
4485  // account the setting of __kmp_affinity_compact.
4486  __kmp_affinity_assign_child_nums(address2os, __kmp_avail_proc);
4487 
4488  switch (__kmp_affinity_type) {
4489 
4490  case affinity_explicit:
4491  KMP_DEBUG_ASSERT(__kmp_affinity_proclist != NULL);
4492  if (__kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) {
4493  __kmp_affinity_process_proclist(
4494  &__kmp_affinity_masks, &__kmp_affinity_num_masks,
4495  __kmp_affinity_proclist, osId2Mask, maxIndex);
4496  } else {
4497  __kmp_affinity_process_placelist(
4498  &__kmp_affinity_masks, &__kmp_affinity_num_masks,
4499  __kmp_affinity_proclist, osId2Mask, maxIndex);
4500  }
4501  if (__kmp_affinity_num_masks == 0) {
4502  if (__kmp_affinity_verbose ||
4503  (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) {
4504  KMP_WARNING(AffNoValidProcID);
4505  }
4506  __kmp_affinity_type = affinity_none;
4507  __kmp_create_affinity_none_places();
4508  return;
4509  }
4510  break;
4511 
4512  // The other affinity types rely on sorting the Addresses according to some
4513  // permutation of the machine topology tree. Set __kmp_affinity_compact and
4514  // __kmp_affinity_offset appropriately, then jump to a common code fragment
4515  // to do the sort and create the array of affinity masks.
4516 
4517  case affinity_logical:
4518  __kmp_affinity_compact = 0;
4519  if (__kmp_affinity_offset) {
4520  __kmp_affinity_offset =
4521  __kmp_nThreadsPerCore * __kmp_affinity_offset % __kmp_avail_proc;
4522  }
4523  goto sortAddresses;
4524 
4525  case affinity_physical:
4526  if (__kmp_nThreadsPerCore > 1) {
4527  __kmp_affinity_compact = 1;
4528  if (__kmp_affinity_compact >= depth) {
4529  __kmp_affinity_compact = 0;
4530  }
4531  } else {
4532  __kmp_affinity_compact = 0;
4533  }
4534  if (__kmp_affinity_offset) {
4535  __kmp_affinity_offset =
4536  __kmp_nThreadsPerCore * __kmp_affinity_offset % __kmp_avail_proc;
4537  }
4538  goto sortAddresses;
4539 
4540  case affinity_scatter:
4541  if (__kmp_affinity_compact >= depth) {
4542  __kmp_affinity_compact = 0;
4543  } else {
4544  __kmp_affinity_compact = depth - 1 - __kmp_affinity_compact;
4545  }
4546  goto sortAddresses;
4547 
4548  case affinity_compact:
4549  if (__kmp_affinity_compact >= depth) {
4550  __kmp_affinity_compact = depth - 1;
4551  }
4552  goto sortAddresses;
4553 
4554  case affinity_balanced:
4555  if (depth <= 1) {
4556  if (__kmp_affinity_verbose || __kmp_affinity_warnings) {
4557  KMP_WARNING(AffBalancedNotAvail, "KMP_AFFINITY");
4558  }
4559  __kmp_affinity_type = affinity_none;
4560  __kmp_create_affinity_none_places();
4561  return;
4562  } else if (!__kmp_affinity_uniform_topology()) {
4563  // Save the depth for further usage
4564  __kmp_aff_depth = depth;
4565 
4566  int core_level = __kmp_affinity_find_core_level(
4567  address2os, __kmp_avail_proc, depth - 1);
4568  int ncores = __kmp_affinity_compute_ncores(address2os, __kmp_avail_proc,
4569  depth - 1, core_level);
4570  int maxprocpercore = __kmp_affinity_max_proc_per_core(
4571  address2os, __kmp_avail_proc, depth - 1, core_level);
4572 
4573  int nproc = ncores * maxprocpercore;
4574  if ((nproc < 2) || (nproc < __kmp_avail_proc)) {
4575  if (__kmp_affinity_verbose || __kmp_affinity_warnings) {
4576  KMP_WARNING(AffBalancedNotAvail, "KMP_AFFINITY");
4577  }
4578  __kmp_affinity_type = affinity_none;
4579  return;
4580  }
4581 
4582  procarr = (int *)__kmp_allocate(sizeof(int) * nproc);
4583  for (int i = 0; i < nproc; i++) {
4584  procarr[i] = -1;
4585  }
4586 
4587  int lastcore = -1;
4588  int inlastcore = 0;
4589  for (int i = 0; i < __kmp_avail_proc; i++) {
4590  int proc = address2os[i].second;
4591  int core =
4592  __kmp_affinity_find_core(address2os, i, depth - 1, core_level);
4593 
4594  if (core == lastcore) {
4595  inlastcore++;
4596  } else {
4597  inlastcore = 0;
4598  }
4599  lastcore = core;
4600 
4601  procarr[core * maxprocpercore + inlastcore] = proc;
4602  }
4603  }
4604  if (__kmp_affinity_compact >= depth) {
4605  __kmp_affinity_compact = depth - 1;
4606  }
4607 
4608  sortAddresses:
4609  // Allocate the gtid->affinity mask table.
4610  if (__kmp_affinity_dups) {
4611  __kmp_affinity_num_masks = __kmp_avail_proc;
4612  } else {
4613  __kmp_affinity_num_masks = numUnique;
4614  }
4615 
4616  if ((__kmp_nested_proc_bind.bind_types[0] != proc_bind_intel) &&
4617  (__kmp_affinity_num_places > 0) &&
4618  ((unsigned)__kmp_affinity_num_places < __kmp_affinity_num_masks)) {
4619  __kmp_affinity_num_masks = __kmp_affinity_num_places;
4620  }
4621 
4622  KMP_CPU_ALLOC_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks);
4623 
4624  // Sort the address2os table according to the current setting of
4625  // __kmp_affinity_compact, then fill out __kmp_affinity_masks.
4626  qsort(address2os, __kmp_avail_proc, sizeof(*address2os),
4627  __kmp_affinity_cmp_Address_child_num);
4628  {
4629  int i;
4630  unsigned j;
4631  for (i = 0, j = 0; i < __kmp_avail_proc; i++) {
4632  if ((!__kmp_affinity_dups) && (!address2os[i].first.leader)) {
4633  continue;
4634  }
4635  unsigned osId = address2os[i].second;
4636  kmp_affin_mask_t *src = KMP_CPU_INDEX(osId2Mask, osId);
4637  kmp_affin_mask_t *dest = KMP_CPU_INDEX(__kmp_affinity_masks, j);
4638  KMP_ASSERT(KMP_CPU_ISSET(osId, src));
4639  KMP_CPU_COPY(dest, src);
4640  if (++j >= __kmp_affinity_num_masks) {
4641  break;
4642  }
4643  }
4644  KMP_DEBUG_ASSERT(j == __kmp_affinity_num_masks);
4645  }
4646  break;
4647 
4648  default:
4649  KMP_ASSERT2(0, "Unexpected affinity setting");
4650  }
4651 
4652  KMP_CPU_FREE_ARRAY(osId2Mask, maxIndex + 1);
4653  machine_hierarchy.init(address2os, __kmp_avail_proc);
4654 }
4655 #undef KMP_EXIT_AFF_NONE
4656 
4657 void __kmp_affinity_initialize(void) {
4658  // Much of the code above was written assuming that if a machine was not
4659  // affinity capable, then __kmp_affinity_type == affinity_none. We now
4660  // explicitly represent this as __kmp_affinity_type == affinity_disabled.
4661  // There are too many checks for __kmp_affinity_type == affinity_none
4662  // in this code. Instead of trying to change them all, check if
4663  // __kmp_affinity_type == affinity_disabled, and if so, slam it with
4664  // affinity_none, call the real initialization routine, then restore
4665  // __kmp_affinity_type to affinity_disabled.
4666  int disabled = (__kmp_affinity_type == affinity_disabled);
4667  if (!KMP_AFFINITY_CAPABLE()) {
4668  KMP_ASSERT(disabled);
4669  }
4670  if (disabled) {
4671  __kmp_affinity_type = affinity_none;
4672  }
4673  __kmp_aux_affinity_initialize();
4674  if (disabled) {
4675  __kmp_affinity_type = affinity_disabled;
4676  }
4677 }
4678 
4679 void __kmp_affinity_uninitialize(void) {
4680  if (__kmp_affinity_masks != NULL) {
4681  KMP_CPU_FREE_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks);
4682  __kmp_affinity_masks = NULL;
4683  }
4684  if (__kmp_affin_fullMask != NULL) {
4685  KMP_CPU_FREE(__kmp_affin_fullMask);
4686  __kmp_affin_fullMask = NULL;
4687  }
4688  __kmp_affinity_num_masks = 0;
4689  __kmp_affinity_type = affinity_default;
4690  __kmp_affinity_num_places = 0;
4691  if (__kmp_affinity_proclist != NULL) {
4692  __kmp_free(__kmp_affinity_proclist);
4693  __kmp_affinity_proclist = NULL;
4694  }
4695  if (address2os != NULL) {
4696  __kmp_free(address2os);
4697  address2os = NULL;
4698  }
4699  if (procarr != NULL) {
4700  __kmp_free(procarr);
4701  procarr = NULL;
4702  }
4703 #if KMP_USE_HWLOC
4704  if (__kmp_hwloc_topology != NULL) {
4705  hwloc_topology_destroy(__kmp_hwloc_topology);
4706  __kmp_hwloc_topology = NULL;
4707  }
4708 #endif
4709  KMPAffinity::destroy_api();
4710 }
4711 
4712 void __kmp_affinity_set_init_mask(int gtid, int isa_root) {
4713  if (!KMP_AFFINITY_CAPABLE()) {
4714  return;
4715  }
4716 
4717  kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
4718  if (th->th.th_affin_mask == NULL) {
4719  KMP_CPU_ALLOC(th->th.th_affin_mask);
4720  } else {
4721  KMP_CPU_ZERO(th->th.th_affin_mask);
4722  }
4723 
4724  // Copy the thread mask to the kmp_info_t structure. If
4725  // __kmp_affinity_type == affinity_none, copy the "full" mask, i.e. one that
4726  // has all of the OS proc ids set, or if __kmp_affinity_respect_mask is set,
4727  // then the full mask is the same as the mask of the initialization thread.
4728  kmp_affin_mask_t *mask;
4729  int i;
4730 
4731  if (KMP_AFFINITY_NON_PROC_BIND) {
4732  if ((__kmp_affinity_type == affinity_none) ||
4733  (__kmp_affinity_type == affinity_balanced)) {
4734 #if KMP_GROUP_AFFINITY
4735  if (__kmp_num_proc_groups > 1) {
4736  return;
4737  }
4738 #endif
4739  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4740  i = 0;
4741  mask = __kmp_affin_fullMask;
4742  } else {
4743  KMP_DEBUG_ASSERT(__kmp_affinity_num_masks > 0);
4744  i = (gtid + __kmp_affinity_offset) % __kmp_affinity_num_masks;
4745  mask = KMP_CPU_INDEX(__kmp_affinity_masks, i);
4746  }
4747  } else {
4748  if ((!isa_root) ||
4749  (__kmp_nested_proc_bind.bind_types[0] == proc_bind_false)) {
4750 #if KMP_GROUP_AFFINITY
4751  if (__kmp_num_proc_groups > 1) {
4752  return;
4753  }
4754 #endif
4755  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4756  i = KMP_PLACE_ALL;
4757  mask = __kmp_affin_fullMask;
4758  } else {
4759  // int i = some hash function or just a counter that doesn't
4760  // always start at 0. Use gtid for now.
4761  KMP_DEBUG_ASSERT(__kmp_affinity_num_masks > 0);
4762  i = (gtid + __kmp_affinity_offset) % __kmp_affinity_num_masks;
4763  mask = KMP_CPU_INDEX(__kmp_affinity_masks, i);
4764  }
4765  }
4766 
4767  th->th.th_current_place = i;
4768  if (isa_root) {
4769  th->th.th_new_place = i;
4770  th->th.th_first_place = 0;
4771  th->th.th_last_place = __kmp_affinity_num_masks - 1;
4772  } else if (KMP_AFFINITY_NON_PROC_BIND) {
4773  // When using a Non-OMP_PROC_BIND affinity method,
4774  // set all threads' place-partition-var to the entire place list
4775  th->th.th_first_place = 0;
4776  th->th.th_last_place = __kmp_affinity_num_masks - 1;
4777  }
4778 
4779  if (i == KMP_PLACE_ALL) {
4780  KA_TRACE(100, ("__kmp_affinity_set_init_mask: binding T#%d to all places\n",
4781  gtid));
4782  } else {
4783  KA_TRACE(100, ("__kmp_affinity_set_init_mask: binding T#%d to place %d\n",
4784  gtid, i));
4785  }
4786 
4787  KMP_CPU_COPY(th->th.th_affin_mask, mask);
4788 
4789  if (__kmp_affinity_verbose
4790  /* to avoid duplicate printing (will be correctly printed on barrier) */
4791  && (__kmp_affinity_type == affinity_none ||
4792  (i != KMP_PLACE_ALL && __kmp_affinity_type != affinity_balanced))) {
4793  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4794  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4795  th->th.th_affin_mask);
4796  KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY", (kmp_int32)getpid(),
4797  __kmp_gettid(), gtid, buf);
4798  }
4799 
4800 #if KMP_OS_WINDOWS
4801  // On Windows* OS, the process affinity mask might have changed. If the user
4802  // didn't request affinity and this call fails, just continue silently.
4803  // See CQ171393.
4804  if (__kmp_affinity_type == affinity_none) {
4805  __kmp_set_system_affinity(th->th.th_affin_mask, FALSE);
4806  } else
4807 #endif
4808  __kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
4809 }
4810 
4811 void __kmp_affinity_set_place(int gtid) {
4812  if (!KMP_AFFINITY_CAPABLE()) {
4813  return;
4814  }
4815 
4816  kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
4817 
4818  KA_TRACE(100, ("__kmp_affinity_set_place: binding T#%d to place %d (current "
4819  "place = %d)\n",
4820  gtid, th->th.th_new_place, th->th.th_current_place));
4821 
4822  // Check that the new place is within this thread's partition.
4823  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
4824  KMP_ASSERT(th->th.th_new_place >= 0);
4825  KMP_ASSERT((unsigned)th->th.th_new_place <= __kmp_affinity_num_masks);
4826  if (th->th.th_first_place <= th->th.th_last_place) {
4827  KMP_ASSERT((th->th.th_new_place >= th->th.th_first_place) &&
4828  (th->th.th_new_place <= th->th.th_last_place));
4829  } else {
4830  KMP_ASSERT((th->th.th_new_place <= th->th.th_first_place) ||
4831  (th->th.th_new_place >= th->th.th_last_place));
4832  }
4833 
4834  // Copy the thread mask to the kmp_info_t structure,
4835  // and set this thread's affinity.
4836  kmp_affin_mask_t *mask =
4837  KMP_CPU_INDEX(__kmp_affinity_masks, th->th.th_new_place);
4838  KMP_CPU_COPY(th->th.th_affin_mask, mask);
4839  th->th.th_current_place = th->th.th_new_place;
4840 
4841  if (__kmp_affinity_verbose) {
4842  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4843  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4844  th->th.th_affin_mask);
4845  KMP_INFORM(BoundToOSProcSet, "OMP_PROC_BIND", (kmp_int32)getpid(),
4846  __kmp_gettid(), gtid, buf);
4847  }
4848  __kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
4849 }
4850 
4851 int __kmp_aux_set_affinity(void **mask) {
4852  int gtid;
4853  kmp_info_t *th;
4854  int retval;
4855 
4856  if (!KMP_AFFINITY_CAPABLE()) {
4857  return -1;
4858  }
4859 
4860  gtid = __kmp_entry_gtid();
4861  KA_TRACE(1000, (""); {
4862  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4863  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4864  (kmp_affin_mask_t *)(*mask));
4865  __kmp_debug_printf(
4866  "kmp_set_affinity: setting affinity mask for thread %d = %s\n", gtid,
4867  buf);
4868  });
4869 
4870  if (__kmp_env_consistency_check) {
4871  if ((mask == NULL) || (*mask == NULL)) {
4872  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4873  } else {
4874  unsigned proc;
4875  int num_procs = 0;
4876 
4877  KMP_CPU_SET_ITERATE(proc, ((kmp_affin_mask_t *)(*mask))) {
4878  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
4879  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4880  }
4881  if (!KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask))) {
4882  continue;
4883  }
4884  num_procs++;
4885  }
4886  if (num_procs == 0) {
4887  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4888  }
4889 
4890 #if KMP_GROUP_AFFINITY
4891  if (__kmp_get_proc_group((kmp_affin_mask_t *)(*mask)) < 0) {
4892  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4893  }
4894 #endif /* KMP_GROUP_AFFINITY */
4895  }
4896  }
4897 
4898  th = __kmp_threads[gtid];
4899  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
4900  retval = __kmp_set_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
4901  if (retval == 0) {
4902  KMP_CPU_COPY(th->th.th_affin_mask, (kmp_affin_mask_t *)(*mask));
4903  }
4904 
4905  th->th.th_current_place = KMP_PLACE_UNDEFINED;
4906  th->th.th_new_place = KMP_PLACE_UNDEFINED;
4907  th->th.th_first_place = 0;
4908  th->th.th_last_place = __kmp_affinity_num_masks - 1;
4909 
4910  // Turn off 4.0 affinity for the current tread at this parallel level.
4911  th->th.th_current_task->td_icvs.proc_bind = proc_bind_false;
4912 
4913  return retval;
4914 }
4915 
4916 int __kmp_aux_get_affinity(void **mask) {
4917  int gtid;
4918  int retval;
4919  kmp_info_t *th;
4920 
4921  if (!KMP_AFFINITY_CAPABLE()) {
4922  return -1;
4923  }
4924 
4925  gtid = __kmp_entry_gtid();
4926  th = __kmp_threads[gtid];
4927  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
4928 
4929  KA_TRACE(1000, (""); {
4930  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4931  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4932  th->th.th_affin_mask);
4933  __kmp_printf("kmp_get_affinity: stored affinity mask for thread %d = %s\n",
4934  gtid, buf);
4935  });
4936 
4937  if (__kmp_env_consistency_check) {
4938  if ((mask == NULL) || (*mask == NULL)) {
4939  KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity");
4940  }
4941  }
4942 
4943 #if !KMP_OS_WINDOWS
4944 
4945  retval = __kmp_get_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
4946  KA_TRACE(1000, (""); {
4947  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4948  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4949  (kmp_affin_mask_t *)(*mask));
4950  __kmp_printf("kmp_get_affinity: system affinity mask for thread %d = %s\n",
4951  gtid, buf);
4952  });
4953  return retval;
4954 
4955 #else
4956 
4957  KMP_CPU_COPY((kmp_affin_mask_t *)(*mask), th->th.th_affin_mask);
4958  return 0;
4959 
4960 #endif /* KMP_OS_WINDOWS */
4961 }
4962 
4963 int __kmp_aux_get_affinity_max_proc() {
4964  if (!KMP_AFFINITY_CAPABLE()) {
4965  return 0;
4966  }
4967 #if KMP_GROUP_AFFINITY
4968  if (__kmp_num_proc_groups > 1) {
4969  return (int)(__kmp_num_proc_groups * sizeof(DWORD_PTR) * CHAR_BIT);
4970  }
4971 #endif
4972  return __kmp_xproc;
4973 }
4974 
4975 int __kmp_aux_set_affinity_mask_proc(int proc, void **mask) {
4976  if (!KMP_AFFINITY_CAPABLE()) {
4977  return -1;
4978  }
4979 
4980  KA_TRACE(1000, (""); {
4981  int gtid = __kmp_entry_gtid();
4982  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4983  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4984  (kmp_affin_mask_t *)(*mask));
4985  __kmp_debug_printf("kmp_set_affinity_mask_proc: setting proc %d in "
4986  "affinity mask for thread %d = %s\n",
4987  proc, gtid, buf);
4988  });
4989 
4990  if (__kmp_env_consistency_check) {
4991  if ((mask == NULL) || (*mask == NULL)) {
4992  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity_mask_proc");
4993  }
4994  }
4995 
4996  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
4997  return -1;
4998  }
4999  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5000  return -2;
5001  }
5002 
5003  KMP_CPU_SET(proc, (kmp_affin_mask_t *)(*mask));
5004  return 0;
5005 }
5006 
5007 int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask) {
5008  if (!KMP_AFFINITY_CAPABLE()) {
5009  return -1;
5010  }
5011 
5012  KA_TRACE(1000, (""); {
5013  int gtid = __kmp_entry_gtid();
5014  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5015  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5016  (kmp_affin_mask_t *)(*mask));
5017  __kmp_debug_printf("kmp_unset_affinity_mask_proc: unsetting proc %d in "
5018  "affinity mask for thread %d = %s\n",
5019  proc, gtid, buf);
5020  });
5021 
5022  if (__kmp_env_consistency_check) {
5023  if ((mask == NULL) || (*mask == NULL)) {
5024  KMP_FATAL(AffinityInvalidMask, "kmp_unset_affinity_mask_proc");
5025  }
5026  }
5027 
5028  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5029  return -1;
5030  }
5031  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5032  return -2;
5033  }
5034 
5035  KMP_CPU_CLR(proc, (kmp_affin_mask_t *)(*mask));
5036  return 0;
5037 }
5038 
5039 int __kmp_aux_get_affinity_mask_proc(int proc, void **mask) {
5040  if (!KMP_AFFINITY_CAPABLE()) {
5041  return -1;
5042  }
5043 
5044  KA_TRACE(1000, (""); {
5045  int gtid = __kmp_entry_gtid();
5046  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5047  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5048  (kmp_affin_mask_t *)(*mask));
5049  __kmp_debug_printf("kmp_get_affinity_mask_proc: getting proc %d in "
5050  "affinity mask for thread %d = %s\n",
5051  proc, gtid, buf);
5052  });
5053 
5054  if (__kmp_env_consistency_check) {
5055  if ((mask == NULL) || (*mask == NULL)) {
5056  KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity_mask_proc");
5057  }
5058  }
5059 
5060  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5061  return -1;
5062  }
5063  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5064  return 0;
5065  }
5066 
5067  return KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask));
5068 }
5069 
5070 // Dynamic affinity settings - Affinity balanced
5071 void __kmp_balanced_affinity(kmp_info_t *th, int nthreads) {
5072  KMP_DEBUG_ASSERT(th);
5073  bool fine_gran = true;
5074  int tid = th->th.th_info.ds.ds_tid;
5075 
5076  switch (__kmp_affinity_gran) {
5077  case affinity_gran_fine:
5078  case affinity_gran_thread:
5079  break;
5080  case affinity_gran_core:
5081  if (__kmp_nThreadsPerCore > 1) {
5082  fine_gran = false;
5083  }
5084  break;
5085  case affinity_gran_package:
5086  if (nCoresPerPkg > 1) {
5087  fine_gran = false;
5088  }
5089  break;
5090  default:
5091  fine_gran = false;
5092  }
5093 
5094  if (__kmp_affinity_uniform_topology()) {
5095  int coreID;
5096  int threadID;
5097  // Number of hyper threads per core in HT machine
5098  int __kmp_nth_per_core = __kmp_avail_proc / __kmp_ncores;
5099  // Number of cores
5100  int ncores = __kmp_ncores;
5101  if ((nPackages > 1) && (__kmp_nth_per_core <= 1)) {
5102  __kmp_nth_per_core = __kmp_avail_proc / nPackages;
5103  ncores = nPackages;
5104  }
5105  // How many threads will be bound to each core
5106  int chunk = nthreads / ncores;
5107  // How many cores will have an additional thread bound to it - "big cores"
5108  int big_cores = nthreads % ncores;
5109  // Number of threads on the big cores
5110  int big_nth = (chunk + 1) * big_cores;
5111  if (tid < big_nth) {
5112  coreID = tid / (chunk + 1);
5113  threadID = (tid % (chunk + 1)) % __kmp_nth_per_core;
5114  } else { // tid >= big_nth
5115  coreID = (tid - big_cores) / chunk;
5116  threadID = ((tid - big_cores) % chunk) % __kmp_nth_per_core;
5117  }
5118 
5119  KMP_DEBUG_ASSERT2(KMP_AFFINITY_CAPABLE(),
5120  "Illegal set affinity operation when not capable");
5121 
5122  kmp_affin_mask_t *mask = th->th.th_affin_mask;
5123  KMP_CPU_ZERO(mask);
5124 
5125  if (fine_gran) {
5126  int osID = address2os[coreID * __kmp_nth_per_core + threadID].second;
5127  KMP_CPU_SET(osID, mask);
5128  } else {
5129  for (int i = 0; i < __kmp_nth_per_core; i++) {
5130  int osID;
5131  osID = address2os[coreID * __kmp_nth_per_core + i].second;
5132  KMP_CPU_SET(osID, mask);
5133  }
5134  }
5135  if (__kmp_affinity_verbose) {
5136  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5137  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
5138  KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY", (kmp_int32)getpid(),
5139  __kmp_gettid(), tid, buf);
5140  }
5141  __kmp_set_system_affinity(mask, TRUE);
5142  } else { // Non-uniform topology
5143 
5144  kmp_affin_mask_t *mask = th->th.th_affin_mask;
5145  KMP_CPU_ZERO(mask);
5146 
5147  int core_level = __kmp_affinity_find_core_level(
5148  address2os, __kmp_avail_proc, __kmp_aff_depth - 1);
5149  int ncores = __kmp_affinity_compute_ncores(address2os, __kmp_avail_proc,
5150  __kmp_aff_depth - 1, core_level);
5151  int nth_per_core = __kmp_affinity_max_proc_per_core(
5152  address2os, __kmp_avail_proc, __kmp_aff_depth - 1, core_level);
5153 
5154  // For performance gain consider the special case nthreads ==
5155  // __kmp_avail_proc
5156  if (nthreads == __kmp_avail_proc) {
5157  if (fine_gran) {
5158  int osID = address2os[tid].second;
5159  KMP_CPU_SET(osID, mask);
5160  } else {
5161  int core = __kmp_affinity_find_core(address2os, tid,
5162  __kmp_aff_depth - 1, core_level);
5163  for (int i = 0; i < __kmp_avail_proc; i++) {
5164  int osID = address2os[i].second;
5165  if (__kmp_affinity_find_core(address2os, i, __kmp_aff_depth - 1,
5166  core_level) == core) {
5167  KMP_CPU_SET(osID, mask);
5168  }
5169  }
5170  }
5171  } else if (nthreads <= ncores) {
5172 
5173  int core = 0;
5174  for (int i = 0; i < ncores; i++) {
5175  // Check if this core from procarr[] is in the mask
5176  int in_mask = 0;
5177  for (int j = 0; j < nth_per_core; j++) {
5178  if (procarr[i * nth_per_core + j] != -1) {
5179  in_mask = 1;
5180  break;
5181  }
5182  }
5183  if (in_mask) {
5184  if (tid == core) {
5185  for (int j = 0; j < nth_per_core; j++) {
5186  int osID = procarr[i * nth_per_core + j];
5187  if (osID != -1) {
5188  KMP_CPU_SET(osID, mask);
5189  // For fine granularity it is enough to set the first available
5190  // osID for this core
5191  if (fine_gran) {
5192  break;
5193  }
5194  }
5195  }
5196  break;
5197  } else {
5198  core++;
5199  }
5200  }
5201  }
5202  } else { // nthreads > ncores
5203  // Array to save the number of processors at each core
5204  int *nproc_at_core = (int *)KMP_ALLOCA(sizeof(int) * ncores);
5205  // Array to save the number of cores with "x" available processors;
5206  int *ncores_with_x_procs =
5207  (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
5208  // Array to save the number of cores with # procs from x to nth_per_core
5209  int *ncores_with_x_to_max_procs =
5210  (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
5211 
5212  for (int i = 0; i <= nth_per_core; i++) {
5213  ncores_with_x_procs[i] = 0;
5214  ncores_with_x_to_max_procs[i] = 0;
5215  }
5216 
5217  for (int i = 0; i < ncores; i++) {
5218  int cnt = 0;
5219  for (int j = 0; j < nth_per_core; j++) {
5220  if (procarr[i * nth_per_core + j] != -1) {
5221  cnt++;
5222  }
5223  }
5224  nproc_at_core[i] = cnt;
5225  ncores_with_x_procs[cnt]++;
5226  }
5227 
5228  for (int i = 0; i <= nth_per_core; i++) {
5229  for (int j = i; j <= nth_per_core; j++) {
5230  ncores_with_x_to_max_procs[i] += ncores_with_x_procs[j];
5231  }
5232  }
5233 
5234  // Max number of processors
5235  int nproc = nth_per_core * ncores;
5236  // An array to keep number of threads per each context
5237  int *newarr = (int *)__kmp_allocate(sizeof(int) * nproc);
5238  for (int i = 0; i < nproc; i++) {
5239  newarr[i] = 0;
5240  }
5241 
5242  int nth = nthreads;
5243  int flag = 0;
5244  while (nth > 0) {
5245  for (int j = 1; j <= nth_per_core; j++) {
5246  int cnt = ncores_with_x_to_max_procs[j];
5247  for (int i = 0; i < ncores; i++) {
5248  // Skip the core with 0 processors
5249  if (nproc_at_core[i] == 0) {
5250  continue;
5251  }
5252  for (int k = 0; k < nth_per_core; k++) {
5253  if (procarr[i * nth_per_core + k] != -1) {
5254  if (newarr[i * nth_per_core + k] == 0) {
5255  newarr[i * nth_per_core + k] = 1;
5256  cnt--;
5257  nth--;
5258  break;
5259  } else {
5260  if (flag != 0) {
5261  newarr[i * nth_per_core + k]++;
5262  cnt--;
5263  nth--;
5264  break;
5265  }
5266  }
5267  }
5268  }
5269  if (cnt == 0 || nth == 0) {
5270  break;
5271  }
5272  }
5273  if (nth == 0) {
5274  break;
5275  }
5276  }
5277  flag = 1;
5278  }
5279  int sum = 0;
5280  for (int i = 0; i < nproc; i++) {
5281  sum += newarr[i];
5282  if (sum > tid) {
5283  if (fine_gran) {
5284  int osID = procarr[i];
5285  KMP_CPU_SET(osID, mask);
5286  } else {
5287  int coreID = i / nth_per_core;
5288  for (int ii = 0; ii < nth_per_core; ii++) {
5289  int osID = procarr[coreID * nth_per_core + ii];
5290  if (osID != -1) {
5291  KMP_CPU_SET(osID, mask);
5292  }
5293  }
5294  }
5295  break;
5296  }
5297  }
5298  __kmp_free(newarr);
5299  }
5300 
5301  if (__kmp_affinity_verbose) {
5302  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5303  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
5304  KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY", (kmp_int32)getpid(),
5305  __kmp_gettid(), tid, buf);
5306  }
5307  __kmp_set_system_affinity(mask, TRUE);
5308  }
5309 }
5310 
5311 #if KMP_OS_LINUX || KMP_OS_FREEBSD
5312 // We don't need this entry for Windows because
5313 // there is GetProcessAffinityMask() api
5314 //
5315 // The intended usage is indicated by these steps:
5316 // 1) The user gets the current affinity mask
5317 // 2) Then sets the affinity by calling this function
5318 // 3) Error check the return value
5319 // 4) Use non-OpenMP parallelization
5320 // 5) Reset the affinity to what was stored in step 1)
5321 #ifdef __cplusplus
5322 extern "C"
5323 #endif
5324  int
5325  kmp_set_thread_affinity_mask_initial()
5326 // the function returns 0 on success,
5327 // -1 if we cannot bind thread
5328 // >0 (errno) if an error happened during binding
5329 {
5330  int gtid = __kmp_get_gtid();
5331  if (gtid < 0) {
5332  // Do not touch non-omp threads
5333  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5334  "non-omp thread, returning\n"));
5335  return -1;
5336  }
5337  if (!KMP_AFFINITY_CAPABLE() || !__kmp_init_middle) {
5338  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5339  "affinity not initialized, returning\n"));
5340  return -1;
5341  }
5342  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5343  "set full mask for thread %d\n",
5344  gtid));
5345  KMP_DEBUG_ASSERT(__kmp_affin_fullMask != NULL);
5346  return __kmp_set_system_affinity(__kmp_affin_fullMask, FALSE);
5347 }
5348 #endif
5349 
5350 #endif // KMP_AFFINITY_SUPPORTED