LLVM OpenMP* Runtime Library
 All Classes Functions Variables Typedefs Enumerations Enumerator Modules Pages
kmp_affinity.cpp
1 /*
2  * kmp_affinity.cpp -- affinity management
3  */
4 
5 //===----------------------------------------------------------------------===//
6 //
7 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
8 // See https://llvm.org/LICENSE.txt for license information.
9 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "kmp.h"
14 #include "kmp_affinity.h"
15 #include "kmp_i18n.h"
16 #include "kmp_io.h"
17 #include "kmp_str.h"
18 #include "kmp_wrapper_getpid.h"
19 #if KMP_USE_HIER_SCHED
20 #include "kmp_dispatch_hier.h"
21 #endif
22 
23 // Store the real or imagined machine hierarchy here
24 static hierarchy_info machine_hierarchy;
25 
26 void __kmp_cleanup_hierarchy() { machine_hierarchy.fini(); }
27 
28 void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar) {
29  kmp_uint32 depth;
30  // The test below is true if affinity is available, but set to "none". Need to
31  // init on first use of hierarchical barrier.
32  if (TCR_1(machine_hierarchy.uninitialized))
33  machine_hierarchy.init(NULL, nproc);
34 
35  // Adjust the hierarchy in case num threads exceeds original
36  if (nproc > machine_hierarchy.base_num_threads)
37  machine_hierarchy.resize(nproc);
38 
39  depth = machine_hierarchy.depth;
40  KMP_DEBUG_ASSERT(depth > 0);
41 
42  thr_bar->depth = depth;
43  thr_bar->base_leaf_kids = (kmp_uint8)machine_hierarchy.numPerLevel[0] - 1;
44  thr_bar->skip_per_level = machine_hierarchy.skipPerLevel;
45 }
46 
47 #if KMP_AFFINITY_SUPPORTED
48 
49 bool KMPAffinity::picked_api = false;
50 
51 void *KMPAffinity::Mask::operator new(size_t n) { return __kmp_allocate(n); }
52 void *KMPAffinity::Mask::operator new[](size_t n) { return __kmp_allocate(n); }
53 void KMPAffinity::Mask::operator delete(void *p) { __kmp_free(p); }
54 void KMPAffinity::Mask::operator delete[](void *p) { __kmp_free(p); }
55 void *KMPAffinity::operator new(size_t n) { return __kmp_allocate(n); }
56 void KMPAffinity::operator delete(void *p) { __kmp_free(p); }
57 
58 void KMPAffinity::pick_api() {
59  KMPAffinity *affinity_dispatch;
60  if (picked_api)
61  return;
62 #if KMP_USE_HWLOC
63  // Only use Hwloc if affinity isn't explicitly disabled and
64  // user requests Hwloc topology method
65  if (__kmp_affinity_top_method == affinity_top_method_hwloc &&
66  __kmp_affinity_type != affinity_disabled) {
67  affinity_dispatch = new KMPHwlocAffinity();
68  } else
69 #endif
70  {
71  affinity_dispatch = new KMPNativeAffinity();
72  }
73  __kmp_affinity_dispatch = affinity_dispatch;
74  picked_api = true;
75 }
76 
77 void KMPAffinity::destroy_api() {
78  if (__kmp_affinity_dispatch != NULL) {
79  delete __kmp_affinity_dispatch;
80  __kmp_affinity_dispatch = NULL;
81  picked_api = false;
82  }
83 }
84 
85 #define KMP_ADVANCE_SCAN(scan) \
86  while (*scan != '\0') { \
87  scan++; \
88  }
89 
90 // Print the affinity mask to the character array in a pretty format.
91 // The format is a comma separated list of non-negative integers or integer
92 // ranges: e.g., 1,2,3-5,7,9-15
93 // The format can also be the string "{<empty>}" if no bits are set in mask
94 char *__kmp_affinity_print_mask(char *buf, int buf_len,
95  kmp_affin_mask_t *mask) {
96  int start = 0, finish = 0, previous = 0;
97  bool first_range;
98  KMP_ASSERT(buf);
99  KMP_ASSERT(buf_len >= 40);
100  KMP_ASSERT(mask);
101  char *scan = buf;
102  char *end = buf + buf_len - 1;
103 
104  // Check for empty set.
105  if (mask->begin() == mask->end()) {
106  KMP_SNPRINTF(scan, end - scan + 1, "{<empty>}");
107  KMP_ADVANCE_SCAN(scan);
108  KMP_ASSERT(scan <= end);
109  return buf;
110  }
111 
112  first_range = true;
113  start = mask->begin();
114  while (1) {
115  // Find next range
116  // [start, previous] is inclusive range of contiguous bits in mask
117  for (finish = mask->next(start), previous = start;
118  finish == previous + 1 && finish != mask->end();
119  finish = mask->next(finish)) {
120  previous = finish;
121  }
122 
123  // The first range does not need a comma printed before it, but the rest
124  // of the ranges do need a comma beforehand
125  if (!first_range) {
126  KMP_SNPRINTF(scan, end - scan + 1, "%s", ",");
127  KMP_ADVANCE_SCAN(scan);
128  } else {
129  first_range = false;
130  }
131  // Range with three or more contiguous bits in the affinity mask
132  if (previous - start > 1) {
133  KMP_SNPRINTF(scan, end - scan + 1, "%d-%d", static_cast<int>(start),
134  static_cast<int>(previous));
135  } else {
136  // Range with one or two contiguous bits in the affinity mask
137  KMP_SNPRINTF(scan, end - scan + 1, "%d", static_cast<int>(start));
138  KMP_ADVANCE_SCAN(scan);
139  if (previous - start > 0) {
140  KMP_SNPRINTF(scan, end - scan + 1, ",%d", static_cast<int>(previous));
141  }
142  }
143  KMP_ADVANCE_SCAN(scan);
144  // Start over with new start point
145  start = finish;
146  if (start == mask->end())
147  break;
148  // Check for overflow
149  if (end - scan < 2)
150  break;
151  }
152 
153  // Check for overflow
154  KMP_ASSERT(scan <= end);
155  return buf;
156 }
157 #undef KMP_ADVANCE_SCAN
158 
159 // Print the affinity mask to the string buffer object in a pretty format
160 // The format is a comma separated list of non-negative integers or integer
161 // ranges: e.g., 1,2,3-5,7,9-15
162 // The format can also be the string "{<empty>}" if no bits are set in mask
163 kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
164  kmp_affin_mask_t *mask) {
165  int start = 0, finish = 0, previous = 0;
166  bool first_range;
167  KMP_ASSERT(buf);
168  KMP_ASSERT(mask);
169 
170  __kmp_str_buf_clear(buf);
171 
172  // Check for empty set.
173  if (mask->begin() == mask->end()) {
174  __kmp_str_buf_print(buf, "%s", "{<empty>}");
175  return buf;
176  }
177 
178  first_range = true;
179  start = mask->begin();
180  while (1) {
181  // Find next range
182  // [start, previous] is inclusive range of contiguous bits in mask
183  for (finish = mask->next(start), previous = start;
184  finish == previous + 1 && finish != mask->end();
185  finish = mask->next(finish)) {
186  previous = finish;
187  }
188 
189  // The first range does not need a comma printed before it, but the rest
190  // of the ranges do need a comma beforehand
191  if (!first_range) {
192  __kmp_str_buf_print(buf, "%s", ",");
193  } else {
194  first_range = false;
195  }
196  // Range with three or more contiguous bits in the affinity mask
197  if (previous - start > 1) {
198  __kmp_str_buf_print(buf, "%d-%d", static_cast<int>(start),
199  static_cast<int>(previous));
200  } else {
201  // Range with one or two contiguous bits in the affinity mask
202  __kmp_str_buf_print(buf, "%d", static_cast<int>(start));
203  if (previous - start > 0) {
204  __kmp_str_buf_print(buf, ",%d", static_cast<int>(previous));
205  }
206  }
207  // Start over with new start point
208  start = finish;
209  if (start == mask->end())
210  break;
211  }
212  return buf;
213 }
214 
215 void __kmp_affinity_entire_machine_mask(kmp_affin_mask_t *mask) {
216  KMP_CPU_ZERO(mask);
217 
218 #if KMP_GROUP_AFFINITY
219 
220  if (__kmp_num_proc_groups > 1) {
221  int group;
222  KMP_DEBUG_ASSERT(__kmp_GetActiveProcessorCount != NULL);
223  for (group = 0; group < __kmp_num_proc_groups; group++) {
224  int i;
225  int num = __kmp_GetActiveProcessorCount(group);
226  for (i = 0; i < num; i++) {
227  KMP_CPU_SET(i + group * (CHAR_BIT * sizeof(DWORD_PTR)), mask);
228  }
229  }
230  } else
231 
232 #endif /* KMP_GROUP_AFFINITY */
233 
234  {
235  int proc;
236  for (proc = 0; proc < __kmp_xproc; proc++) {
237  KMP_CPU_SET(proc, mask);
238  }
239  }
240 }
241 
242 // When sorting by labels, __kmp_affinity_assign_child_nums() must first be
243 // called to renumber the labels from [0..n] and place them into the child_num
244 // vector of the address object. This is done in case the labels used for
245 // the children at one node of the hierarchy differ from those used for
246 // another node at the same level. Example: suppose the machine has 2 nodes
247 // with 2 packages each. The first node contains packages 601 and 602, and
248 // second node contains packages 603 and 604. If we try to sort the table
249 // for "scatter" affinity, the table will still be sorted 601, 602, 603, 604
250 // because we are paying attention to the labels themselves, not the ordinal
251 // child numbers. By using the child numbers in the sort, the result is
252 // {0,0}=601, {0,1}=603, {1,0}=602, {1,1}=604.
253 static void __kmp_affinity_assign_child_nums(AddrUnsPair *address2os,
254  int numAddrs) {
255  KMP_DEBUG_ASSERT(numAddrs > 0);
256  int depth = address2os->first.depth;
257  unsigned *counts = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
258  unsigned *lastLabel = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
259  int labCt;
260  for (labCt = 0; labCt < depth; labCt++) {
261  address2os[0].first.childNums[labCt] = counts[labCt] = 0;
262  lastLabel[labCt] = address2os[0].first.labels[labCt];
263  }
264  int i;
265  for (i = 1; i < numAddrs; i++) {
266  for (labCt = 0; labCt < depth; labCt++) {
267  if (address2os[i].first.labels[labCt] != lastLabel[labCt]) {
268  int labCt2;
269  for (labCt2 = labCt + 1; labCt2 < depth; labCt2++) {
270  counts[labCt2] = 0;
271  lastLabel[labCt2] = address2os[i].first.labels[labCt2];
272  }
273  counts[labCt]++;
274  lastLabel[labCt] = address2os[i].first.labels[labCt];
275  break;
276  }
277  }
278  for (labCt = 0; labCt < depth; labCt++) {
279  address2os[i].first.childNums[labCt] = counts[labCt];
280  }
281  for (; labCt < (int)Address::maxDepth; labCt++) {
282  address2os[i].first.childNums[labCt] = 0;
283  }
284  }
285  __kmp_free(lastLabel);
286  __kmp_free(counts);
287 }
288 
289 // All of the __kmp_affinity_create_*_map() routines should set
290 // __kmp_affinity_masks to a vector of affinity mask objects of length
291 // __kmp_affinity_num_masks, if __kmp_affinity_type != affinity_none, and return
292 // the number of levels in the machine topology tree (zero if
293 // __kmp_affinity_type == affinity_none).
294 //
295 // All of the __kmp_affinity_create_*_map() routines should set
296 // *__kmp_affin_fullMask to the affinity mask for the initialization thread.
297 // They need to save and restore the mask, and it could be needed later, so
298 // saving it is just an optimization to avoid calling kmp_get_system_affinity()
299 // again.
300 kmp_affin_mask_t *__kmp_affin_fullMask = NULL;
301 
302 static int nCoresPerPkg, nPackages;
303 static int __kmp_nThreadsPerCore;
304 #ifndef KMP_DFLT_NTH_CORES
305 static int __kmp_ncores;
306 #endif
307 static int *__kmp_pu_os_idx = NULL;
308 
309 // __kmp_affinity_uniform_topology() doesn't work when called from
310 // places which support arbitrarily many levels in the machine topology
311 // map, i.e. the non-default cases in __kmp_affinity_create_cpuinfo_map()
312 // __kmp_affinity_create_x2apicid_map().
313 inline static bool __kmp_affinity_uniform_topology() {
314  return __kmp_avail_proc == (__kmp_nThreadsPerCore * nCoresPerPkg * nPackages);
315 }
316 
317 // Print out the detailed machine topology map, i.e. the physical locations
318 // of each OS proc.
319 static void __kmp_affinity_print_topology(AddrUnsPair *address2os, int len,
320  int depth, int pkgLevel,
321  int coreLevel, int threadLevel) {
322  int proc;
323 
324  KMP_INFORM(OSProcToPhysicalThreadMap, "KMP_AFFINITY");
325  for (proc = 0; proc < len; proc++) {
326  int level;
327  kmp_str_buf_t buf;
328  __kmp_str_buf_init(&buf);
329  for (level = 0; level < depth; level++) {
330  if (level == threadLevel) {
331  __kmp_str_buf_print(&buf, "%s ", KMP_I18N_STR(Thread));
332  } else if (level == coreLevel) {
333  __kmp_str_buf_print(&buf, "%s ", KMP_I18N_STR(Core));
334  } else if (level == pkgLevel) {
335  __kmp_str_buf_print(&buf, "%s ", KMP_I18N_STR(Package));
336  } else if (level > pkgLevel) {
337  __kmp_str_buf_print(&buf, "%s_%d ", KMP_I18N_STR(Node),
338  level - pkgLevel - 1);
339  } else {
340  __kmp_str_buf_print(&buf, "L%d ", level);
341  }
342  __kmp_str_buf_print(&buf, "%d ", address2os[proc].first.labels[level]);
343  }
344  KMP_INFORM(OSProcMapToPack, "KMP_AFFINITY", address2os[proc].second,
345  buf.str);
346  __kmp_str_buf_free(&buf);
347  }
348 }
349 
350 #if KMP_USE_HWLOC
351 
352 static void __kmp_affinity_print_hwloc_tp(AddrUnsPair *addrP, int len,
353  int depth, int *levels) {
354  int proc;
355  kmp_str_buf_t buf;
356  __kmp_str_buf_init(&buf);
357  KMP_INFORM(OSProcToPhysicalThreadMap, "KMP_AFFINITY");
358  for (proc = 0; proc < len; proc++) {
359  __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Package),
360  addrP[proc].first.labels[0]);
361  if (depth > 1) {
362  int level = 1; // iterate over levels
363  int label = 1; // iterate over labels
364  if (__kmp_numa_detected)
365  // node level follows package
366  if (levels[level++] > 0)
367  __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Node),
368  addrP[proc].first.labels[label++]);
369  if (__kmp_tile_depth > 0)
370  // tile level follows node if any, or package
371  if (levels[level++] > 0)
372  __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Tile),
373  addrP[proc].first.labels[label++]);
374  if (levels[level++] > 0)
375  // core level follows
376  __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Core),
377  addrP[proc].first.labels[label++]);
378  if (levels[level++] > 0)
379  // thread level is the latest
380  __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Thread),
381  addrP[proc].first.labels[label++]);
382  KMP_DEBUG_ASSERT(label == depth);
383  }
384  KMP_INFORM(OSProcMapToPack, "KMP_AFFINITY", addrP[proc].second, buf.str);
385  __kmp_str_buf_clear(&buf);
386  }
387  __kmp_str_buf_free(&buf);
388 }
389 
390 static int nNodePerPkg, nTilePerPkg, nTilePerNode, nCorePerNode, nCorePerTile;
391 
392 // This function removes the topology levels that are radix 1 and don't offer
393 // further information about the topology. The most common example is when you
394 // have one thread context per core, we don't want the extra thread context
395 // level if it offers no unique labels. So they are removed.
396 // return value: the new depth of address2os
397 static int __kmp_affinity_remove_radix_one_levels(AddrUnsPair *addrP, int nTh,
398  int depth, int *levels) {
399  int level;
400  int i;
401  int radix1_detected;
402  int new_depth = depth;
403  for (level = depth - 1; level > 0; --level) {
404  // Detect if this level is radix 1
405  radix1_detected = 1;
406  for (i = 1; i < nTh; ++i) {
407  if (addrP[0].first.labels[level] != addrP[i].first.labels[level]) {
408  // There are differing label values for this level so it stays
409  radix1_detected = 0;
410  break;
411  }
412  }
413  if (!radix1_detected)
414  continue;
415  // Radix 1 was detected
416  --new_depth;
417  levels[level] = -1; // mark level as not present in address2os array
418  if (level == new_depth) {
419  // "turn off" deepest level, just decrement the depth that removes
420  // the level from address2os array
421  for (i = 0; i < nTh; ++i) {
422  addrP[i].first.depth--;
423  }
424  } else {
425  // For other levels, we move labels over and also reduce the depth
426  int j;
427  for (j = level; j < new_depth; ++j) {
428  for (i = 0; i < nTh; ++i) {
429  addrP[i].first.labels[j] = addrP[i].first.labels[j + 1];
430  addrP[i].first.depth--;
431  }
432  levels[j + 1] -= 1;
433  }
434  }
435  }
436  return new_depth;
437 }
438 
439 // Returns the number of objects of type 'type' below 'obj' within the topology
440 // tree structure. e.g., if obj is a HWLOC_OBJ_PACKAGE object, and type is
441 // HWLOC_OBJ_PU, then this will return the number of PU's under the SOCKET
442 // object.
443 static int __kmp_hwloc_get_nobjs_under_obj(hwloc_obj_t obj,
444  hwloc_obj_type_t type) {
445  int retval = 0;
446  hwloc_obj_t first;
447  for (first = hwloc_get_obj_below_by_type(__kmp_hwloc_topology, obj->type,
448  obj->logical_index, type, 0);
449  first != NULL &&
450  hwloc_get_ancestor_obj_by_type(__kmp_hwloc_topology, obj->type, first) ==
451  obj;
452  first = hwloc_get_next_obj_by_type(__kmp_hwloc_topology, first->type,
453  first)) {
454  ++retval;
455  }
456  return retval;
457 }
458 
459 static int __kmp_hwloc_count_children_by_depth(hwloc_topology_t t,
460  hwloc_obj_t o, unsigned depth,
461  hwloc_obj_t *f) {
462  if (o->depth == depth) {
463  if (*f == NULL)
464  *f = o; // output first descendant found
465  return 1;
466  }
467  int sum = 0;
468  for (unsigned i = 0; i < o->arity; i++)
469  sum += __kmp_hwloc_count_children_by_depth(t, o->children[i], depth, f);
470  return sum; // will be 0 if no one found (as PU arity is 0)
471 }
472 
473 static int __kmp_hwloc_count_children_by_type(hwloc_topology_t t, hwloc_obj_t o,
474  hwloc_obj_type_t type,
475  hwloc_obj_t *f) {
476  if (!hwloc_compare_types(o->type, type)) {
477  if (*f == NULL)
478  *f = o; // output first descendant found
479  return 1;
480  }
481  int sum = 0;
482  for (unsigned i = 0; i < o->arity; i++)
483  sum += __kmp_hwloc_count_children_by_type(t, o->children[i], type, f);
484  return sum; // will be 0 if no one found (as PU arity is 0)
485 }
486 
487 static int __kmp_hwloc_process_obj_core_pu(AddrUnsPair *addrPair,
488  int &nActiveThreads,
489  int &num_active_cores,
490  hwloc_obj_t obj, int depth,
491  int *labels) {
492  hwloc_obj_t core = NULL;
493  hwloc_topology_t &tp = __kmp_hwloc_topology;
494  int NC = __kmp_hwloc_count_children_by_type(tp, obj, HWLOC_OBJ_CORE, &core);
495  for (int core_id = 0; core_id < NC; ++core_id, core = core->next_cousin) {
496  hwloc_obj_t pu = NULL;
497  KMP_DEBUG_ASSERT(core != NULL);
498  int num_active_threads = 0;
499  int NT = __kmp_hwloc_count_children_by_type(tp, core, HWLOC_OBJ_PU, &pu);
500  // int NT = core->arity; pu = core->first_child; // faster?
501  for (int pu_id = 0; pu_id < NT; ++pu_id, pu = pu->next_cousin) {
502  KMP_DEBUG_ASSERT(pu != NULL);
503  if (!KMP_CPU_ISSET(pu->os_index, __kmp_affin_fullMask))
504  continue; // skip inactive (inaccessible) unit
505  Address addr(depth + 2);
506  KA_TRACE(20, ("Hwloc inserting %d (%d) %d (%d) %d (%d) into address2os\n",
507  obj->os_index, obj->logical_index, core->os_index,
508  core->logical_index, pu->os_index, pu->logical_index));
509  for (int i = 0; i < depth; ++i)
510  addr.labels[i] = labels[i]; // package, etc.
511  addr.labels[depth] = core_id; // core
512  addr.labels[depth + 1] = pu_id; // pu
513  addrPair[nActiveThreads] = AddrUnsPair(addr, pu->os_index);
514  __kmp_pu_os_idx[nActiveThreads] = pu->os_index;
515  nActiveThreads++;
516  ++num_active_threads; // count active threads per core
517  }
518  if (num_active_threads) { // were there any active threads on the core?
519  ++__kmp_ncores; // count total active cores
520  ++num_active_cores; // count active cores per socket
521  if (num_active_threads > __kmp_nThreadsPerCore)
522  __kmp_nThreadsPerCore = num_active_threads; // calc maximum
523  }
524  }
525  return 0;
526 }
527 
528 // Check if NUMA node detected below the package,
529 // and if tile object is detected and return its depth
530 static int __kmp_hwloc_check_numa() {
531  hwloc_topology_t &tp = __kmp_hwloc_topology;
532  hwloc_obj_t hT, hC, hL, hN, hS; // hwloc objects (pointers to)
533  int depth;
534 
535  // Get some PU
536  hT = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PU, 0);
537  if (hT == NULL) // something has gone wrong
538  return 1;
539 
540  // check NUMA node below PACKAGE
541  hN = hwloc_get_ancestor_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hT);
542  hS = hwloc_get_ancestor_obj_by_type(tp, HWLOC_OBJ_PACKAGE, hT);
543  KMP_DEBUG_ASSERT(hS != NULL);
544  if (hN != NULL && hN->depth > hS->depth) {
545  __kmp_numa_detected = TRUE; // socket includes node(s)
546  if (__kmp_affinity_gran == affinity_gran_node) {
547  __kmp_affinity_gran == affinity_gran_numa;
548  }
549  }
550 
551  // check tile, get object by depth because of multiple caches possible
552  depth = hwloc_get_cache_type_depth(tp, 2, HWLOC_OBJ_CACHE_UNIFIED);
553  hL = hwloc_get_ancestor_obj_by_depth(tp, depth, hT);
554  hC = NULL; // not used, but reset it here just in case
555  if (hL != NULL &&
556  __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE, &hC) > 1)
557  __kmp_tile_depth = depth; // tile consists of multiple cores
558  return 0;
559 }
560 
561 static int __kmp_affinity_create_hwloc_map(AddrUnsPair **address2os,
562  kmp_i18n_id_t *const msg_id) {
563  hwloc_topology_t &tp = __kmp_hwloc_topology; // shortcut of a long name
564  *address2os = NULL;
565  *msg_id = kmp_i18n_null;
566 
567  // Save the affinity mask for the current thread.
568  kmp_affin_mask_t *oldMask;
569  KMP_CPU_ALLOC(oldMask);
570  __kmp_get_system_affinity(oldMask, TRUE);
571  __kmp_hwloc_check_numa();
572 
573  if (!KMP_AFFINITY_CAPABLE()) {
574  // Hack to try and infer the machine topology using only the data
575  // available from cpuid on the current thread, and __kmp_xproc.
576  KMP_ASSERT(__kmp_affinity_type == affinity_none);
577 
578  nCoresPerPkg = __kmp_hwloc_get_nobjs_under_obj(
579  hwloc_get_obj_by_type(tp, HWLOC_OBJ_PACKAGE, 0), HWLOC_OBJ_CORE);
580  __kmp_nThreadsPerCore = __kmp_hwloc_get_nobjs_under_obj(
581  hwloc_get_obj_by_type(tp, HWLOC_OBJ_CORE, 0), HWLOC_OBJ_PU);
582  __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
583  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
584  if (__kmp_affinity_verbose) {
585  KMP_INFORM(AffNotCapableUseLocCpuidL11, "KMP_AFFINITY");
586  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
587  if (__kmp_affinity_uniform_topology()) {
588  KMP_INFORM(Uniform, "KMP_AFFINITY");
589  } else {
590  KMP_INFORM(NonUniform, "KMP_AFFINITY");
591  }
592  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
593  __kmp_nThreadsPerCore, __kmp_ncores);
594  }
595  KMP_CPU_FREE(oldMask);
596  return 0;
597  }
598 
599  int depth = 3;
600  int levels[5] = {0, 1, 2, 3, 4}; // package, [node,] [tile,] core, thread
601  int labels[3] = {0}; // package [,node] [,tile] - head of lables array
602  if (__kmp_numa_detected)
603  ++depth;
604  if (__kmp_tile_depth)
605  ++depth;
606 
607  // Allocate the data structure to be returned.
608  AddrUnsPair *retval =
609  (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * __kmp_avail_proc);
610  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
611  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
612 
613  // When affinity is off, this routine will still be called to set
614  // __kmp_ncores, as well as __kmp_nThreadsPerCore,
615  // nCoresPerPkg, & nPackages. Make sure all these vars are set
616  // correctly, and return if affinity is not enabled.
617 
618  hwloc_obj_t socket, node, tile;
619  int nActiveThreads = 0;
620  int socket_id = 0;
621  // re-calculate globals to count only accessible resources
622  __kmp_ncores = nPackages = nCoresPerPkg = __kmp_nThreadsPerCore = 0;
623  nNodePerPkg = nTilePerPkg = nTilePerNode = nCorePerNode = nCorePerTile = 0;
624  for (socket = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PACKAGE, 0); socket != NULL;
625  socket = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PACKAGE, socket),
626  socket_id++) {
627  labels[0] = socket_id;
628  if (__kmp_numa_detected) {
629  int NN;
630  int n_active_nodes = 0;
631  node = NULL;
632  NN = __kmp_hwloc_count_children_by_type(tp, socket, HWLOC_OBJ_NUMANODE,
633  &node);
634  for (int node_id = 0; node_id < NN; ++node_id, node = node->next_cousin) {
635  labels[1] = node_id;
636  if (__kmp_tile_depth) {
637  // NUMA + tiles
638  int NT;
639  int n_active_tiles = 0;
640  tile = NULL;
641  NT = __kmp_hwloc_count_children_by_depth(tp, node, __kmp_tile_depth,
642  &tile);
643  for (int tl_id = 0; tl_id < NT; ++tl_id, tile = tile->next_cousin) {
644  labels[2] = tl_id;
645  int n_active_cores = 0;
646  __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads,
647  n_active_cores, tile, 3, labels);
648  if (n_active_cores) { // were there any active cores on the socket?
649  ++n_active_tiles; // count active tiles per node
650  if (n_active_cores > nCorePerTile)
651  nCorePerTile = n_active_cores; // calc maximum
652  }
653  }
654  if (n_active_tiles) { // were there any active tiles on the socket?
655  ++n_active_nodes; // count active nodes per package
656  if (n_active_tiles > nTilePerNode)
657  nTilePerNode = n_active_tiles; // calc maximum
658  }
659  } else {
660  // NUMA, no tiles
661  int n_active_cores = 0;
662  __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads,
663  n_active_cores, node, 2, labels);
664  if (n_active_cores) { // were there any active cores on the socket?
665  ++n_active_nodes; // count active nodes per package
666  if (n_active_cores > nCorePerNode)
667  nCorePerNode = n_active_cores; // calc maximum
668  }
669  }
670  }
671  if (n_active_nodes) { // were there any active nodes on the socket?
672  ++nPackages; // count total active packages
673  if (n_active_nodes > nNodePerPkg)
674  nNodePerPkg = n_active_nodes; // calc maximum
675  }
676  } else {
677  if (__kmp_tile_depth) {
678  // no NUMA, tiles
679  int NT;
680  int n_active_tiles = 0;
681  tile = NULL;
682  NT = __kmp_hwloc_count_children_by_depth(tp, socket, __kmp_tile_depth,
683  &tile);
684  for (int tl_id = 0; tl_id < NT; ++tl_id, tile = tile->next_cousin) {
685  labels[1] = tl_id;
686  int n_active_cores = 0;
687  __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads,
688  n_active_cores, tile, 2, labels);
689  if (n_active_cores) { // were there any active cores on the socket?
690  ++n_active_tiles; // count active tiles per package
691  if (n_active_cores > nCorePerTile)
692  nCorePerTile = n_active_cores; // calc maximum
693  }
694  }
695  if (n_active_tiles) { // were there any active tiles on the socket?
696  ++nPackages; // count total active packages
697  if (n_active_tiles > nTilePerPkg)
698  nTilePerPkg = n_active_tiles; // calc maximum
699  }
700  } else {
701  // no NUMA, no tiles
702  int n_active_cores = 0;
703  __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads, n_active_cores,
704  socket, 1, labels);
705  if (n_active_cores) { // were there any active cores on the socket?
706  ++nPackages; // count total active packages
707  if (n_active_cores > nCoresPerPkg)
708  nCoresPerPkg = n_active_cores; // calc maximum
709  }
710  }
711  }
712  }
713 
714  // If there's only one thread context to bind to, return now.
715  KMP_DEBUG_ASSERT(nActiveThreads == __kmp_avail_proc);
716  KMP_ASSERT(nActiveThreads > 0);
717  if (nActiveThreads == 1) {
718  __kmp_ncores = nPackages = 1;
719  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
720  if (__kmp_affinity_verbose) {
721  char buf[KMP_AFFIN_MASK_PRINT_LEN];
722  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
723 
724  KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
725  if (__kmp_affinity_respect_mask) {
726  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
727  } else {
728  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
729  }
730  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
731  KMP_INFORM(Uniform, "KMP_AFFINITY");
732  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
733  __kmp_nThreadsPerCore, __kmp_ncores);
734  }
735 
736  if (__kmp_affinity_type == affinity_none) {
737  __kmp_free(retval);
738  KMP_CPU_FREE(oldMask);
739  return 0;
740  }
741 
742  // Form an Address object which only includes the package level.
743  Address addr(1);
744  addr.labels[0] = retval[0].first.labels[0];
745  retval[0].first = addr;
746 
747  if (__kmp_affinity_gran_levels < 0) {
748  __kmp_affinity_gran_levels = 0;
749  }
750 
751  if (__kmp_affinity_verbose) {
752  __kmp_affinity_print_topology(retval, 1, 1, 0, -1, -1);
753  }
754 
755  *address2os = retval;
756  KMP_CPU_FREE(oldMask);
757  return 1;
758  }
759 
760  // Sort the table by physical Id.
761  qsort(retval, nActiveThreads, sizeof(*retval),
762  __kmp_affinity_cmp_Address_labels);
763 
764  // Check to see if the machine topology is uniform
765  int nPUs = nPackages * __kmp_nThreadsPerCore;
766  if (__kmp_numa_detected) {
767  if (__kmp_tile_depth) { // NUMA + tiles
768  nPUs *= (nNodePerPkg * nTilePerNode * nCorePerTile);
769  } else { // NUMA, no tiles
770  nPUs *= (nNodePerPkg * nCorePerNode);
771  }
772  } else {
773  if (__kmp_tile_depth) { // no NUMA, tiles
774  nPUs *= (nTilePerPkg * nCorePerTile);
775  } else { // no NUMA, no tiles
776  nPUs *= nCoresPerPkg;
777  }
778  }
779  unsigned uniform = (nPUs == nActiveThreads);
780 
781  // Print the machine topology summary.
782  if (__kmp_affinity_verbose) {
783  char mask[KMP_AFFIN_MASK_PRINT_LEN];
784  __kmp_affinity_print_mask(mask, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
785  if (__kmp_affinity_respect_mask) {
786  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", mask);
787  } else {
788  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", mask);
789  }
790  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
791  if (uniform) {
792  KMP_INFORM(Uniform, "KMP_AFFINITY");
793  } else {
794  KMP_INFORM(NonUniform, "KMP_AFFINITY");
795  }
796  if (__kmp_numa_detected) {
797  if (__kmp_tile_depth) { // NUMA + tiles
798  KMP_INFORM(TopologyExtraNoTi, "KMP_AFFINITY", nPackages, nNodePerPkg,
799  nTilePerNode, nCorePerTile, __kmp_nThreadsPerCore,
800  __kmp_ncores);
801  } else { // NUMA, no tiles
802  KMP_INFORM(TopologyExtraNode, "KMP_AFFINITY", nPackages, nNodePerPkg,
803  nCorePerNode, __kmp_nThreadsPerCore, __kmp_ncores);
804  nPUs *= (nNodePerPkg * nCorePerNode);
805  }
806  } else {
807  if (__kmp_tile_depth) { // no NUMA, tiles
808  KMP_INFORM(TopologyExtraTile, "KMP_AFFINITY", nPackages, nTilePerPkg,
809  nCorePerTile, __kmp_nThreadsPerCore, __kmp_ncores);
810  } else { // no NUMA, no tiles
811  kmp_str_buf_t buf;
812  __kmp_str_buf_init(&buf);
813  __kmp_str_buf_print(&buf, "%d", nPackages);
814  KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, nCoresPerPkg,
815  __kmp_nThreadsPerCore, __kmp_ncores);
816  __kmp_str_buf_free(&buf);
817  }
818  }
819  }
820 
821  if (__kmp_affinity_type == affinity_none) {
822  __kmp_free(retval);
823  KMP_CPU_FREE(oldMask);
824  return 0;
825  }
826 
827  int depth_full = depth; // number of levels before compressing
828  // Find any levels with radiix 1, and remove them from the map
829  // (except for the package level).
830  depth = __kmp_affinity_remove_radix_one_levels(retval, nActiveThreads, depth,
831  levels);
832  KMP_DEBUG_ASSERT(__kmp_affinity_gran != affinity_gran_default);
833  if (__kmp_affinity_gran_levels < 0) {
834  // Set the granularity level based on what levels are modeled
835  // in the machine topology map.
836  __kmp_affinity_gran_levels = 0; // lowest level (e.g. fine)
837  if (__kmp_affinity_gran > affinity_gran_thread) {
838  for (int i = 1; i <= depth_full; ++i) {
839  if (__kmp_affinity_gran <= i) // only count deeper levels
840  break;
841  if (levels[depth_full - i] > 0)
842  __kmp_affinity_gran_levels++;
843  }
844  }
845  if (__kmp_affinity_gran > affinity_gran_package)
846  __kmp_affinity_gran_levels++; // e.g. granularity = group
847  }
848 
849  if (__kmp_affinity_verbose)
850  __kmp_affinity_print_hwloc_tp(retval, nActiveThreads, depth, levels);
851 
852  KMP_CPU_FREE(oldMask);
853  *address2os = retval;
854  return depth;
855 }
856 #endif // KMP_USE_HWLOC
857 
858 // If we don't know how to retrieve the machine's processor topology, or
859 // encounter an error in doing so, this routine is called to form a "flat"
860 // mapping of os thread id's <-> processor id's.
861 static int __kmp_affinity_create_flat_map(AddrUnsPair **address2os,
862  kmp_i18n_id_t *const msg_id) {
863  *address2os = NULL;
864  *msg_id = kmp_i18n_null;
865 
866  // Even if __kmp_affinity_type == affinity_none, this routine might still
867  // called to set __kmp_ncores, as well as
868  // __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
869  if (!KMP_AFFINITY_CAPABLE()) {
870  KMP_ASSERT(__kmp_affinity_type == affinity_none);
871  __kmp_ncores = nPackages = __kmp_xproc;
872  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
873  if (__kmp_affinity_verbose) {
874  KMP_INFORM(AffFlatTopology, "KMP_AFFINITY");
875  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
876  KMP_INFORM(Uniform, "KMP_AFFINITY");
877  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
878  __kmp_nThreadsPerCore, __kmp_ncores);
879  }
880  return 0;
881  }
882 
883  // When affinity is off, this routine will still be called to set
884  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
885  // Make sure all these vars are set correctly, and return now if affinity is
886  // not enabled.
887  __kmp_ncores = nPackages = __kmp_avail_proc;
888  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
889  if (__kmp_affinity_verbose) {
890  char buf[KMP_AFFIN_MASK_PRINT_LEN];
891  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
892  __kmp_affin_fullMask);
893 
894  KMP_INFORM(AffCapableUseFlat, "KMP_AFFINITY");
895  if (__kmp_affinity_respect_mask) {
896  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
897  } else {
898  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
899  }
900  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
901  KMP_INFORM(Uniform, "KMP_AFFINITY");
902  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
903  __kmp_nThreadsPerCore, __kmp_ncores);
904  }
905  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
906  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
907  if (__kmp_affinity_type == affinity_none) {
908  int avail_ct = 0;
909  int i;
910  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
911  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask))
912  continue;
913  __kmp_pu_os_idx[avail_ct++] = i; // suppose indices are flat
914  }
915  return 0;
916  }
917 
918  // Contruct the data structure to be returned.
919  *address2os =
920  (AddrUnsPair *)__kmp_allocate(sizeof(**address2os) * __kmp_avail_proc);
921  int avail_ct = 0;
922  int i;
923  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
924  // Skip this proc if it is not included in the machine model.
925  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
926  continue;
927  }
928  __kmp_pu_os_idx[avail_ct] = i; // suppose indices are flat
929  Address addr(1);
930  addr.labels[0] = i;
931  (*address2os)[avail_ct++] = AddrUnsPair(addr, i);
932  }
933  if (__kmp_affinity_verbose) {
934  KMP_INFORM(OSProcToPackage, "KMP_AFFINITY");
935  }
936 
937  if (__kmp_affinity_gran_levels < 0) {
938  // Only the package level is modeled in the machine topology map,
939  // so the #levels of granularity is either 0 or 1.
940  if (__kmp_affinity_gran > affinity_gran_package) {
941  __kmp_affinity_gran_levels = 1;
942  } else {
943  __kmp_affinity_gran_levels = 0;
944  }
945  }
946  return 1;
947 }
948 
949 #if KMP_GROUP_AFFINITY
950 
951 // If multiple Windows* OS processor groups exist, we can create a 2-level
952 // topology map with the groups at level 0 and the individual procs at level 1.
953 // This facilitates letting the threads float among all procs in a group,
954 // if granularity=group (the default when there are multiple groups).
955 static int __kmp_affinity_create_proc_group_map(AddrUnsPair **address2os,
956  kmp_i18n_id_t *const msg_id) {
957  *address2os = NULL;
958  *msg_id = kmp_i18n_null;
959 
960  // If we aren't affinity capable, then return now.
961  // The flat mapping will be used.
962  if (!KMP_AFFINITY_CAPABLE()) {
963  // FIXME set *msg_id
964  return -1;
965  }
966 
967  // Contruct the data structure to be returned.
968  *address2os =
969  (AddrUnsPair *)__kmp_allocate(sizeof(**address2os) * __kmp_avail_proc);
970  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
971  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
972  int avail_ct = 0;
973  int i;
974  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
975  // Skip this proc if it is not included in the machine model.
976  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
977  continue;
978  }
979  __kmp_pu_os_idx[avail_ct] = i; // suppose indices are flat
980  Address addr(2);
981  addr.labels[0] = i / (CHAR_BIT * sizeof(DWORD_PTR));
982  addr.labels[1] = i % (CHAR_BIT * sizeof(DWORD_PTR));
983  (*address2os)[avail_ct++] = AddrUnsPair(addr, i);
984 
985  if (__kmp_affinity_verbose) {
986  KMP_INFORM(AffOSProcToGroup, "KMP_AFFINITY", i, addr.labels[0],
987  addr.labels[1]);
988  }
989  }
990 
991  if (__kmp_affinity_gran_levels < 0) {
992  if (__kmp_affinity_gran == affinity_gran_group) {
993  __kmp_affinity_gran_levels = 1;
994  } else if ((__kmp_affinity_gran == affinity_gran_fine) ||
995  (__kmp_affinity_gran == affinity_gran_thread)) {
996  __kmp_affinity_gran_levels = 0;
997  } else {
998  const char *gran_str = NULL;
999  if (__kmp_affinity_gran == affinity_gran_core) {
1000  gran_str = "core";
1001  } else if (__kmp_affinity_gran == affinity_gran_package) {
1002  gran_str = "package";
1003  } else if (__kmp_affinity_gran == affinity_gran_node) {
1004  gran_str = "node";
1005  } else {
1006  KMP_ASSERT(0);
1007  }
1008 
1009  // Warning: can't use affinity granularity \"gran\" with group topology
1010  // method, using "thread"
1011  __kmp_affinity_gran_levels = 0;
1012  }
1013  }
1014  return 2;
1015 }
1016 
1017 #endif /* KMP_GROUP_AFFINITY */
1018 
1019 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1020 
1021 static int __kmp_cpuid_mask_width(int count) {
1022  int r = 0;
1023 
1024  while ((1 << r) < count)
1025  ++r;
1026  return r;
1027 }
1028 
1029 class apicThreadInfo {
1030 public:
1031  unsigned osId; // param to __kmp_affinity_bind_thread
1032  unsigned apicId; // from cpuid after binding
1033  unsigned maxCoresPerPkg; // ""
1034  unsigned maxThreadsPerPkg; // ""
1035  unsigned pkgId; // inferred from above values
1036  unsigned coreId; // ""
1037  unsigned threadId; // ""
1038 };
1039 
1040 static int __kmp_affinity_cmp_apicThreadInfo_phys_id(const void *a,
1041  const void *b) {
1042  const apicThreadInfo *aa = (const apicThreadInfo *)a;
1043  const apicThreadInfo *bb = (const apicThreadInfo *)b;
1044  if (aa->pkgId < bb->pkgId)
1045  return -1;
1046  if (aa->pkgId > bb->pkgId)
1047  return 1;
1048  if (aa->coreId < bb->coreId)
1049  return -1;
1050  if (aa->coreId > bb->coreId)
1051  return 1;
1052  if (aa->threadId < bb->threadId)
1053  return -1;
1054  if (aa->threadId > bb->threadId)
1055  return 1;
1056  return 0;
1057 }
1058 
1059 // On IA-32 architecture and Intel(R) 64 architecture, we attempt to use
1060 // an algorithm which cycles through the available os threads, setting
1061 // the current thread's affinity mask to that thread, and then retrieves
1062 // the Apic Id for each thread context using the cpuid instruction.
1063 static int __kmp_affinity_create_apicid_map(AddrUnsPair **address2os,
1064  kmp_i18n_id_t *const msg_id) {
1065  kmp_cpuid buf;
1066  *address2os = NULL;
1067  *msg_id = kmp_i18n_null;
1068 
1069  // Check if cpuid leaf 4 is supported.
1070  __kmp_x86_cpuid(0, 0, &buf);
1071  if (buf.eax < 4) {
1072  *msg_id = kmp_i18n_str_NoLeaf4Support;
1073  return -1;
1074  }
1075 
1076  // The algorithm used starts by setting the affinity to each available thread
1077  // and retrieving info from the cpuid instruction, so if we are not capable of
1078  // calling __kmp_get_system_affinity() and _kmp_get_system_affinity(), then we
1079  // need to do something else - use the defaults that we calculated from
1080  // issuing cpuid without binding to each proc.
1081  if (!KMP_AFFINITY_CAPABLE()) {
1082  // Hack to try and infer the machine topology using only the data
1083  // available from cpuid on the current thread, and __kmp_xproc.
1084  KMP_ASSERT(__kmp_affinity_type == affinity_none);
1085 
1086  // Get an upper bound on the number of threads per package using cpuid(1).
1087  // On some OS/chps combinations where HT is supported by the chip but is
1088  // disabled, this value will be 2 on a single core chip. Usually, it will be
1089  // 2 if HT is enabled and 1 if HT is disabled.
1090  __kmp_x86_cpuid(1, 0, &buf);
1091  int maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
1092  if (maxThreadsPerPkg == 0) {
1093  maxThreadsPerPkg = 1;
1094  }
1095 
1096  // The num cores per pkg comes from cpuid(4). 1 must be added to the encoded
1097  // value.
1098  //
1099  // The author of cpu_count.cpp treated this only an upper bound on the
1100  // number of cores, but I haven't seen any cases where it was greater than
1101  // the actual number of cores, so we will treat it as exact in this block of
1102  // code.
1103  //
1104  // First, we need to check if cpuid(4) is supported on this chip. To see if
1105  // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n or
1106  // greater.
1107  __kmp_x86_cpuid(0, 0, &buf);
1108  if (buf.eax >= 4) {
1109  __kmp_x86_cpuid(4, 0, &buf);
1110  nCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
1111  } else {
1112  nCoresPerPkg = 1;
1113  }
1114 
1115  // There is no way to reliably tell if HT is enabled without issuing the
1116  // cpuid instruction from every thread, can correlating the cpuid info, so
1117  // if the machine is not affinity capable, we assume that HT is off. We have
1118  // seen quite a few machines where maxThreadsPerPkg is 2, yet the machine
1119  // does not support HT.
1120  //
1121  // - Older OSes are usually found on machines with older chips, which do not
1122  // support HT.
1123  // - The performance penalty for mistakenly identifying a machine as HT when
1124  // it isn't (which results in blocktime being incorrecly set to 0) is
1125  // greater than the penalty when for mistakenly identifying a machine as
1126  // being 1 thread/core when it is really HT enabled (which results in
1127  // blocktime being incorrectly set to a positive value).
1128  __kmp_ncores = __kmp_xproc;
1129  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
1130  __kmp_nThreadsPerCore = 1;
1131  if (__kmp_affinity_verbose) {
1132  KMP_INFORM(AffNotCapableUseLocCpuid, "KMP_AFFINITY");
1133  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1134  if (__kmp_affinity_uniform_topology()) {
1135  KMP_INFORM(Uniform, "KMP_AFFINITY");
1136  } else {
1137  KMP_INFORM(NonUniform, "KMP_AFFINITY");
1138  }
1139  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
1140  __kmp_nThreadsPerCore, __kmp_ncores);
1141  }
1142  return 0;
1143  }
1144 
1145  // From here on, we can assume that it is safe to call
1146  // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
1147  // __kmp_affinity_type = affinity_none.
1148 
1149  // Save the affinity mask for the current thread.
1150  kmp_affin_mask_t *oldMask;
1151  KMP_CPU_ALLOC(oldMask);
1152  KMP_ASSERT(oldMask != NULL);
1153  __kmp_get_system_affinity(oldMask, TRUE);
1154 
1155  // Run through each of the available contexts, binding the current thread
1156  // to it, and obtaining the pertinent information using the cpuid instr.
1157  //
1158  // The relevant information is:
1159  // - Apic Id: Bits 24:31 of ebx after issuing cpuid(1) - each thread context
1160  // has a uniqie Apic Id, which is of the form pkg# : core# : thread#.
1161  // - Max Threads Per Pkg: Bits 16:23 of ebx after issuing cpuid(1). The value
1162  // of this field determines the width of the core# + thread# fields in the
1163  // Apic Id. It is also an upper bound on the number of threads per
1164  // package, but it has been verified that situations happen were it is not
1165  // exact. In particular, on certain OS/chip combinations where Intel(R)
1166  // Hyper-Threading Technology is supported by the chip but has been
1167  // disabled, the value of this field will be 2 (for a single core chip).
1168  // On other OS/chip combinations supporting Intel(R) Hyper-Threading
1169  // Technology, the value of this field will be 1 when Intel(R)
1170  // Hyper-Threading Technology is disabled and 2 when it is enabled.
1171  // - Max Cores Per Pkg: Bits 26:31 of eax after issuing cpuid(4). The value
1172  // of this field (+1) determines the width of the core# field in the Apic
1173  // Id. The comments in "cpucount.cpp" say that this value is an upper
1174  // bound, but the IA-32 architecture manual says that it is exactly the
1175  // number of cores per package, and I haven't seen any case where it
1176  // wasn't.
1177  //
1178  // From this information, deduce the package Id, core Id, and thread Id,
1179  // and set the corresponding fields in the apicThreadInfo struct.
1180  unsigned i;
1181  apicThreadInfo *threadInfo = (apicThreadInfo *)__kmp_allocate(
1182  __kmp_avail_proc * sizeof(apicThreadInfo));
1183  unsigned nApics = 0;
1184  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
1185  // Skip this proc if it is not included in the machine model.
1186  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
1187  continue;
1188  }
1189  KMP_DEBUG_ASSERT((int)nApics < __kmp_avail_proc);
1190 
1191  __kmp_affinity_dispatch->bind_thread(i);
1192  threadInfo[nApics].osId = i;
1193 
1194  // The apic id and max threads per pkg come from cpuid(1).
1195  __kmp_x86_cpuid(1, 0, &buf);
1196  if (((buf.edx >> 9) & 1) == 0) {
1197  __kmp_set_system_affinity(oldMask, TRUE);
1198  __kmp_free(threadInfo);
1199  KMP_CPU_FREE(oldMask);
1200  *msg_id = kmp_i18n_str_ApicNotPresent;
1201  return -1;
1202  }
1203  threadInfo[nApics].apicId = (buf.ebx >> 24) & 0xff;
1204  threadInfo[nApics].maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
1205  if (threadInfo[nApics].maxThreadsPerPkg == 0) {
1206  threadInfo[nApics].maxThreadsPerPkg = 1;
1207  }
1208 
1209  // Max cores per pkg comes from cpuid(4). 1 must be added to the encoded
1210  // value.
1211  //
1212  // First, we need to check if cpuid(4) is supported on this chip. To see if
1213  // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n
1214  // or greater.
1215  __kmp_x86_cpuid(0, 0, &buf);
1216  if (buf.eax >= 4) {
1217  __kmp_x86_cpuid(4, 0, &buf);
1218  threadInfo[nApics].maxCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
1219  } else {
1220  threadInfo[nApics].maxCoresPerPkg = 1;
1221  }
1222 
1223  // Infer the pkgId / coreId / threadId using only the info obtained locally.
1224  int widthCT = __kmp_cpuid_mask_width(threadInfo[nApics].maxThreadsPerPkg);
1225  threadInfo[nApics].pkgId = threadInfo[nApics].apicId >> widthCT;
1226 
1227  int widthC = __kmp_cpuid_mask_width(threadInfo[nApics].maxCoresPerPkg);
1228  int widthT = widthCT - widthC;
1229  if (widthT < 0) {
1230  // I've never seen this one happen, but I suppose it could, if the cpuid
1231  // instruction on a chip was really screwed up. Make sure to restore the
1232  // affinity mask before the tail call.
1233  __kmp_set_system_affinity(oldMask, TRUE);
1234  __kmp_free(threadInfo);
1235  KMP_CPU_FREE(oldMask);
1236  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1237  return -1;
1238  }
1239 
1240  int maskC = (1 << widthC) - 1;
1241  threadInfo[nApics].coreId = (threadInfo[nApics].apicId >> widthT) & maskC;
1242 
1243  int maskT = (1 << widthT) - 1;
1244  threadInfo[nApics].threadId = threadInfo[nApics].apicId & maskT;
1245 
1246  nApics++;
1247  }
1248 
1249  // We've collected all the info we need.
1250  // Restore the old affinity mask for this thread.
1251  __kmp_set_system_affinity(oldMask, TRUE);
1252 
1253  // If there's only one thread context to bind to, form an Address object
1254  // with depth 1 and return immediately (or, if affinity is off, set
1255  // address2os to NULL and return).
1256  //
1257  // If it is configured to omit the package level when there is only a single
1258  // package, the logic at the end of this routine won't work if there is only
1259  // a single thread - it would try to form an Address object with depth 0.
1260  KMP_ASSERT(nApics > 0);
1261  if (nApics == 1) {
1262  __kmp_ncores = nPackages = 1;
1263  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
1264  if (__kmp_affinity_verbose) {
1265  char buf[KMP_AFFIN_MASK_PRINT_LEN];
1266  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
1267 
1268  KMP_INFORM(AffUseGlobCpuid, "KMP_AFFINITY");
1269  if (__kmp_affinity_respect_mask) {
1270  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
1271  } else {
1272  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
1273  }
1274  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1275  KMP_INFORM(Uniform, "KMP_AFFINITY");
1276  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
1277  __kmp_nThreadsPerCore, __kmp_ncores);
1278  }
1279 
1280  if (__kmp_affinity_type == affinity_none) {
1281  __kmp_free(threadInfo);
1282  KMP_CPU_FREE(oldMask);
1283  return 0;
1284  }
1285 
1286  *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair));
1287  Address addr(1);
1288  addr.labels[0] = threadInfo[0].pkgId;
1289  (*address2os)[0] = AddrUnsPair(addr, threadInfo[0].osId);
1290 
1291  if (__kmp_affinity_gran_levels < 0) {
1292  __kmp_affinity_gran_levels = 0;
1293  }
1294 
1295  if (__kmp_affinity_verbose) {
1296  __kmp_affinity_print_topology(*address2os, 1, 1, 0, -1, -1);
1297  }
1298 
1299  __kmp_free(threadInfo);
1300  KMP_CPU_FREE(oldMask);
1301  return 1;
1302  }
1303 
1304  // Sort the threadInfo table by physical Id.
1305  qsort(threadInfo, nApics, sizeof(*threadInfo),
1306  __kmp_affinity_cmp_apicThreadInfo_phys_id);
1307 
1308  // The table is now sorted by pkgId / coreId / threadId, but we really don't
1309  // know the radix of any of the fields. pkgId's may be sparsely assigned among
1310  // the chips on a system. Although coreId's are usually assigned
1311  // [0 .. coresPerPkg-1] and threadId's are usually assigned
1312  // [0..threadsPerCore-1], we don't want to make any such assumptions.
1313  //
1314  // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
1315  // total # packages) are at this point - we want to determine that now. We
1316  // only have an upper bound on the first two figures.
1317  //
1318  // We also perform a consistency check at this point: the values returned by
1319  // the cpuid instruction for any thread bound to a given package had better
1320  // return the same info for maxThreadsPerPkg and maxCoresPerPkg.
1321  nPackages = 1;
1322  nCoresPerPkg = 1;
1323  __kmp_nThreadsPerCore = 1;
1324  unsigned nCores = 1;
1325 
1326  unsigned pkgCt = 1; // to determine radii
1327  unsigned lastPkgId = threadInfo[0].pkgId;
1328  unsigned coreCt = 1;
1329  unsigned lastCoreId = threadInfo[0].coreId;
1330  unsigned threadCt = 1;
1331  unsigned lastThreadId = threadInfo[0].threadId;
1332 
1333  // intra-pkg consist checks
1334  unsigned prevMaxCoresPerPkg = threadInfo[0].maxCoresPerPkg;
1335  unsigned prevMaxThreadsPerPkg = threadInfo[0].maxThreadsPerPkg;
1336 
1337  for (i = 1; i < nApics; i++) {
1338  if (threadInfo[i].pkgId != lastPkgId) {
1339  nCores++;
1340  pkgCt++;
1341  lastPkgId = threadInfo[i].pkgId;
1342  if ((int)coreCt > nCoresPerPkg)
1343  nCoresPerPkg = coreCt;
1344  coreCt = 1;
1345  lastCoreId = threadInfo[i].coreId;
1346  if ((int)threadCt > __kmp_nThreadsPerCore)
1347  __kmp_nThreadsPerCore = threadCt;
1348  threadCt = 1;
1349  lastThreadId = threadInfo[i].threadId;
1350 
1351  // This is a different package, so go on to the next iteration without
1352  // doing any consistency checks. Reset the consistency check vars, though.
1353  prevMaxCoresPerPkg = threadInfo[i].maxCoresPerPkg;
1354  prevMaxThreadsPerPkg = threadInfo[i].maxThreadsPerPkg;
1355  continue;
1356  }
1357 
1358  if (threadInfo[i].coreId != lastCoreId) {
1359  nCores++;
1360  coreCt++;
1361  lastCoreId = threadInfo[i].coreId;
1362  if ((int)threadCt > __kmp_nThreadsPerCore)
1363  __kmp_nThreadsPerCore = threadCt;
1364  threadCt = 1;
1365  lastThreadId = threadInfo[i].threadId;
1366  } else if (threadInfo[i].threadId != lastThreadId) {
1367  threadCt++;
1368  lastThreadId = threadInfo[i].threadId;
1369  } else {
1370  __kmp_free(threadInfo);
1371  KMP_CPU_FREE(oldMask);
1372  *msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
1373  return -1;
1374  }
1375 
1376  // Check to make certain that the maxCoresPerPkg and maxThreadsPerPkg
1377  // fields agree between all the threads bounds to a given package.
1378  if ((prevMaxCoresPerPkg != threadInfo[i].maxCoresPerPkg) ||
1379  (prevMaxThreadsPerPkg != threadInfo[i].maxThreadsPerPkg)) {
1380  __kmp_free(threadInfo);
1381  KMP_CPU_FREE(oldMask);
1382  *msg_id = kmp_i18n_str_InconsistentCpuidInfo;
1383  return -1;
1384  }
1385  }
1386  nPackages = pkgCt;
1387  if ((int)coreCt > nCoresPerPkg)
1388  nCoresPerPkg = coreCt;
1389  if ((int)threadCt > __kmp_nThreadsPerCore)
1390  __kmp_nThreadsPerCore = threadCt;
1391 
1392  // When affinity is off, this routine will still be called to set
1393  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
1394  // Make sure all these vars are set correctly, and return now if affinity is
1395  // not enabled.
1396  __kmp_ncores = nCores;
1397  if (__kmp_affinity_verbose) {
1398  char buf[KMP_AFFIN_MASK_PRINT_LEN];
1399  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
1400 
1401  KMP_INFORM(AffUseGlobCpuid, "KMP_AFFINITY");
1402  if (__kmp_affinity_respect_mask) {
1403  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
1404  } else {
1405  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
1406  }
1407  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1408  if (__kmp_affinity_uniform_topology()) {
1409  KMP_INFORM(Uniform, "KMP_AFFINITY");
1410  } else {
1411  KMP_INFORM(NonUniform, "KMP_AFFINITY");
1412  }
1413  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
1414  __kmp_nThreadsPerCore, __kmp_ncores);
1415  }
1416  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
1417  KMP_DEBUG_ASSERT(nApics == (unsigned)__kmp_avail_proc);
1418  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
1419  for (i = 0; i < nApics; ++i) {
1420  __kmp_pu_os_idx[i] = threadInfo[i].osId;
1421  }
1422  if (__kmp_affinity_type == affinity_none) {
1423  __kmp_free(threadInfo);
1424  KMP_CPU_FREE(oldMask);
1425  return 0;
1426  }
1427 
1428  // Now that we've determined the number of packages, the number of cores per
1429  // package, and the number of threads per core, we can construct the data
1430  // structure that is to be returned.
1431  int pkgLevel = 0;
1432  int coreLevel = (nCoresPerPkg <= 1) ? -1 : 1;
1433  int threadLevel =
1434  (__kmp_nThreadsPerCore <= 1) ? -1 : ((coreLevel >= 0) ? 2 : 1);
1435  unsigned depth = (pkgLevel >= 0) + (coreLevel >= 0) + (threadLevel >= 0);
1436 
1437  KMP_ASSERT(depth > 0);
1438  *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * nApics);
1439 
1440  for (i = 0; i < nApics; ++i) {
1441  Address addr(depth);
1442  unsigned os = threadInfo[i].osId;
1443  int d = 0;
1444 
1445  if (pkgLevel >= 0) {
1446  addr.labels[d++] = threadInfo[i].pkgId;
1447  }
1448  if (coreLevel >= 0) {
1449  addr.labels[d++] = threadInfo[i].coreId;
1450  }
1451  if (threadLevel >= 0) {
1452  addr.labels[d++] = threadInfo[i].threadId;
1453  }
1454  (*address2os)[i] = AddrUnsPair(addr, os);
1455  }
1456 
1457  if (__kmp_affinity_gran_levels < 0) {
1458  // Set the granularity level based on what levels are modeled in the machine
1459  // topology map.
1460  __kmp_affinity_gran_levels = 0;
1461  if ((threadLevel >= 0) && (__kmp_affinity_gran > affinity_gran_thread)) {
1462  __kmp_affinity_gran_levels++;
1463  }
1464  if ((coreLevel >= 0) && (__kmp_affinity_gran > affinity_gran_core)) {
1465  __kmp_affinity_gran_levels++;
1466  }
1467  if ((pkgLevel >= 0) && (__kmp_affinity_gran > affinity_gran_package)) {
1468  __kmp_affinity_gran_levels++;
1469  }
1470  }
1471 
1472  if (__kmp_affinity_verbose) {
1473  __kmp_affinity_print_topology(*address2os, nApics, depth, pkgLevel,
1474  coreLevel, threadLevel);
1475  }
1476 
1477  __kmp_free(threadInfo);
1478  KMP_CPU_FREE(oldMask);
1479  return depth;
1480 }
1481 
1482 // Intel(R) microarchitecture code name Nehalem, Dunnington and later
1483 // architectures support a newer interface for specifying the x2APIC Ids,
1484 // based on cpuid leaf 11.
1485 static int __kmp_affinity_create_x2apicid_map(AddrUnsPair **address2os,
1486  kmp_i18n_id_t *const msg_id) {
1487  kmp_cpuid buf;
1488  *address2os = NULL;
1489  *msg_id = kmp_i18n_null;
1490 
1491  // Check to see if cpuid leaf 11 is supported.
1492  __kmp_x86_cpuid(0, 0, &buf);
1493  if (buf.eax < 11) {
1494  *msg_id = kmp_i18n_str_NoLeaf11Support;
1495  return -1;
1496  }
1497  __kmp_x86_cpuid(11, 0, &buf);
1498  if (buf.ebx == 0) {
1499  *msg_id = kmp_i18n_str_NoLeaf11Support;
1500  return -1;
1501  }
1502 
1503  // Find the number of levels in the machine topology. While we're at it, get
1504  // the default values for __kmp_nThreadsPerCore & nCoresPerPkg. We will try to
1505  // get more accurate values later by explicitly counting them, but get
1506  // reasonable defaults now, in case we return early.
1507  int level;
1508  int threadLevel = -1;
1509  int coreLevel = -1;
1510  int pkgLevel = -1;
1511  __kmp_nThreadsPerCore = nCoresPerPkg = nPackages = 1;
1512 
1513  for (level = 0;; level++) {
1514  if (level > 31) {
1515  // FIXME: Hack for DPD200163180
1516  //
1517  // If level is big then something went wrong -> exiting
1518  //
1519  // There could actually be 32 valid levels in the machine topology, but so
1520  // far, the only machine we have seen which does not exit this loop before
1521  // iteration 32 has fubar x2APIC settings.
1522  //
1523  // For now, just reject this case based upon loop trip count.
1524  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1525  return -1;
1526  }
1527  __kmp_x86_cpuid(11, level, &buf);
1528  if (buf.ebx == 0) {
1529  if (pkgLevel < 0) {
1530  // Will infer nPackages from __kmp_xproc
1531  pkgLevel = level;
1532  level++;
1533  }
1534  break;
1535  }
1536  int kind = (buf.ecx >> 8) & 0xff;
1537  if (kind == 1) {
1538  // SMT level
1539  threadLevel = level;
1540  coreLevel = -1;
1541  pkgLevel = -1;
1542  __kmp_nThreadsPerCore = buf.ebx & 0xffff;
1543  if (__kmp_nThreadsPerCore == 0) {
1544  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1545  return -1;
1546  }
1547  } else if (kind == 2) {
1548  // core level
1549  coreLevel = level;
1550  pkgLevel = -1;
1551  nCoresPerPkg = buf.ebx & 0xffff;
1552  if (nCoresPerPkg == 0) {
1553  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1554  return -1;
1555  }
1556  } else {
1557  if (level <= 0) {
1558  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1559  return -1;
1560  }
1561  if (pkgLevel >= 0) {
1562  continue;
1563  }
1564  pkgLevel = level;
1565  nPackages = buf.ebx & 0xffff;
1566  if (nPackages == 0) {
1567  *msg_id = kmp_i18n_str_InvalidCpuidInfo;
1568  return -1;
1569  }
1570  }
1571  }
1572  int depth = level;
1573 
1574  // In the above loop, "level" was counted from the finest level (usually
1575  // thread) to the coarsest. The caller expects that we will place the labels
1576  // in (*address2os)[].first.labels[] in the inverse order, so we need to
1577  // invert the vars saying which level means what.
1578  if (threadLevel >= 0) {
1579  threadLevel = depth - threadLevel - 1;
1580  }
1581  if (coreLevel >= 0) {
1582  coreLevel = depth - coreLevel - 1;
1583  }
1584  KMP_DEBUG_ASSERT(pkgLevel >= 0);
1585  pkgLevel = depth - pkgLevel - 1;
1586 
1587  // The algorithm used starts by setting the affinity to each available thread
1588  // and retrieving info from the cpuid instruction, so if we are not capable of
1589  // calling __kmp_get_system_affinity() and _kmp_get_system_affinity(), then we
1590  // need to do something else - use the defaults that we calculated from
1591  // issuing cpuid without binding to each proc.
1592  if (!KMP_AFFINITY_CAPABLE()) {
1593  // Hack to try and infer the machine topology using only the data
1594  // available from cpuid on the current thread, and __kmp_xproc.
1595  KMP_ASSERT(__kmp_affinity_type == affinity_none);
1596 
1597  __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
1598  nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
1599  if (__kmp_affinity_verbose) {
1600  KMP_INFORM(AffNotCapableUseLocCpuidL11, "KMP_AFFINITY");
1601  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1602  if (__kmp_affinity_uniform_topology()) {
1603  KMP_INFORM(Uniform, "KMP_AFFINITY");
1604  } else {
1605  KMP_INFORM(NonUniform, "KMP_AFFINITY");
1606  }
1607  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
1608  __kmp_nThreadsPerCore, __kmp_ncores);
1609  }
1610  return 0;
1611  }
1612 
1613  // From here on, we can assume that it is safe to call
1614  // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
1615  // __kmp_affinity_type = affinity_none.
1616 
1617  // Save the affinity mask for the current thread.
1618  kmp_affin_mask_t *oldMask;
1619  KMP_CPU_ALLOC(oldMask);
1620  __kmp_get_system_affinity(oldMask, TRUE);
1621 
1622  // Allocate the data structure to be returned.
1623  AddrUnsPair *retval =
1624  (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * __kmp_avail_proc);
1625 
1626  // Run through each of the available contexts, binding the current thread
1627  // to it, and obtaining the pertinent information using the cpuid instr.
1628  unsigned int proc;
1629  int nApics = 0;
1630  KMP_CPU_SET_ITERATE(proc, __kmp_affin_fullMask) {
1631  // Skip this proc if it is not included in the machine model.
1632  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
1633  continue;
1634  }
1635  KMP_DEBUG_ASSERT(nApics < __kmp_avail_proc);
1636 
1637  __kmp_affinity_dispatch->bind_thread(proc);
1638 
1639  // Extract labels for each level in the machine topology map from Apic ID.
1640  Address addr(depth);
1641  int prev_shift = 0;
1642 
1643  for (level = 0; level < depth; level++) {
1644  __kmp_x86_cpuid(11, level, &buf);
1645  unsigned apicId = buf.edx;
1646  if (buf.ebx == 0) {
1647  if (level != depth - 1) {
1648  KMP_CPU_FREE(oldMask);
1649  *msg_id = kmp_i18n_str_InconsistentCpuidInfo;
1650  return -1;
1651  }
1652  addr.labels[depth - level - 1] = apicId >> prev_shift;
1653  level++;
1654  break;
1655  }
1656  int shift = buf.eax & 0x1f;
1657  int mask = (1 << shift) - 1;
1658  addr.labels[depth - level - 1] = (apicId & mask) >> prev_shift;
1659  prev_shift = shift;
1660  }
1661  if (level != depth) {
1662  KMP_CPU_FREE(oldMask);
1663  *msg_id = kmp_i18n_str_InconsistentCpuidInfo;
1664  return -1;
1665  }
1666 
1667  retval[nApics] = AddrUnsPair(addr, proc);
1668  nApics++;
1669  }
1670 
1671  // We've collected all the info we need.
1672  // Restore the old affinity mask for this thread.
1673  __kmp_set_system_affinity(oldMask, TRUE);
1674 
1675  // If there's only one thread context to bind to, return now.
1676  KMP_ASSERT(nApics > 0);
1677  if (nApics == 1) {
1678  __kmp_ncores = nPackages = 1;
1679  __kmp_nThreadsPerCore = nCoresPerPkg = 1;
1680  if (__kmp_affinity_verbose) {
1681  char buf[KMP_AFFIN_MASK_PRINT_LEN];
1682  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
1683 
1684  KMP_INFORM(AffUseGlobCpuidL11, "KMP_AFFINITY");
1685  if (__kmp_affinity_respect_mask) {
1686  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
1687  } else {
1688  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
1689  }
1690  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1691  KMP_INFORM(Uniform, "KMP_AFFINITY");
1692  KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
1693  __kmp_nThreadsPerCore, __kmp_ncores);
1694  }
1695 
1696  if (__kmp_affinity_type == affinity_none) {
1697  __kmp_free(retval);
1698  KMP_CPU_FREE(oldMask);
1699  return 0;
1700  }
1701 
1702  // Form an Address object which only includes the package level.
1703  Address addr(1);
1704  addr.labels[0] = retval[0].first.labels[pkgLevel];
1705  retval[0].first = addr;
1706 
1707  if (__kmp_affinity_gran_levels < 0) {
1708  __kmp_affinity_gran_levels = 0;
1709  }
1710 
1711  if (__kmp_affinity_verbose) {
1712  __kmp_affinity_print_topology(retval, 1, 1, 0, -1, -1);
1713  }
1714 
1715  *address2os = retval;
1716  KMP_CPU_FREE(oldMask);
1717  return 1;
1718  }
1719 
1720  // Sort the table by physical Id.
1721  qsort(retval, nApics, sizeof(*retval), __kmp_affinity_cmp_Address_labels);
1722 
1723  // Find the radix at each of the levels.
1724  unsigned *totals = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
1725  unsigned *counts = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
1726  unsigned *maxCt = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
1727  unsigned *last = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
1728  for (level = 0; level < depth; level++) {
1729  totals[level] = 1;
1730  maxCt[level] = 1;
1731  counts[level] = 1;
1732  last[level] = retval[0].first.labels[level];
1733  }
1734 
1735  // From here on, the iteration variable "level" runs from the finest level to
1736  // the coarsest, i.e. we iterate forward through
1737  // (*address2os)[].first.labels[] - in the previous loops, we iterated
1738  // backwards.
1739  for (proc = 1; (int)proc < nApics; proc++) {
1740  int level;
1741  for (level = 0; level < depth; level++) {
1742  if (retval[proc].first.labels[level] != last[level]) {
1743  int j;
1744  for (j = level + 1; j < depth; j++) {
1745  totals[j]++;
1746  counts[j] = 1;
1747  // The line below causes printing incorrect topology information in
1748  // case the max value for some level (maxCt[level]) is encountered
1749  // earlier than some less value while going through the array. For
1750  // example, let pkg0 has 4 cores and pkg1 has 2 cores. Then
1751  // maxCt[1] == 2
1752  // whereas it must be 4.
1753  // TODO!!! Check if it can be commented safely
1754  // maxCt[j] = 1;
1755  last[j] = retval[proc].first.labels[j];
1756  }
1757  totals[level]++;
1758  counts[level]++;
1759  if (counts[level] > maxCt[level]) {
1760  maxCt[level] = counts[level];
1761  }
1762  last[level] = retval[proc].first.labels[level];
1763  break;
1764  } else if (level == depth - 1) {
1765  __kmp_free(last);
1766  __kmp_free(maxCt);
1767  __kmp_free(counts);
1768  __kmp_free(totals);
1769  __kmp_free(retval);
1770  KMP_CPU_FREE(oldMask);
1771  *msg_id = kmp_i18n_str_x2ApicIDsNotUnique;
1772  return -1;
1773  }
1774  }
1775  }
1776 
1777  // When affinity is off, this routine will still be called to set
1778  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
1779  // Make sure all these vars are set correctly, and return if affinity is not
1780  // enabled.
1781  if (threadLevel >= 0) {
1782  __kmp_nThreadsPerCore = maxCt[threadLevel];
1783  } else {
1784  __kmp_nThreadsPerCore = 1;
1785  }
1786  nPackages = totals[pkgLevel];
1787 
1788  if (coreLevel >= 0) {
1789  __kmp_ncores = totals[coreLevel];
1790  nCoresPerPkg = maxCt[coreLevel];
1791  } else {
1792  __kmp_ncores = nPackages;
1793  nCoresPerPkg = 1;
1794  }
1795 
1796  // Check to see if the machine topology is uniform
1797  unsigned prod = maxCt[0];
1798  for (level = 1; level < depth; level++) {
1799  prod *= maxCt[level];
1800  }
1801  bool uniform = (prod == totals[level - 1]);
1802 
1803  // Print the machine topology summary.
1804  if (__kmp_affinity_verbose) {
1805  char mask[KMP_AFFIN_MASK_PRINT_LEN];
1806  __kmp_affinity_print_mask(mask, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
1807 
1808  KMP_INFORM(AffUseGlobCpuidL11, "KMP_AFFINITY");
1809  if (__kmp_affinity_respect_mask) {
1810  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", mask);
1811  } else {
1812  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", mask);
1813  }
1814  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
1815  if (uniform) {
1816  KMP_INFORM(Uniform, "KMP_AFFINITY");
1817  } else {
1818  KMP_INFORM(NonUniform, "KMP_AFFINITY");
1819  }
1820 
1821  kmp_str_buf_t buf;
1822  __kmp_str_buf_init(&buf);
1823 
1824  __kmp_str_buf_print(&buf, "%d", totals[0]);
1825  for (level = 1; level <= pkgLevel; level++) {
1826  __kmp_str_buf_print(&buf, " x %d", maxCt[level]);
1827  }
1828  KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, nCoresPerPkg,
1829  __kmp_nThreadsPerCore, __kmp_ncores);
1830 
1831  __kmp_str_buf_free(&buf);
1832  }
1833  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
1834  KMP_DEBUG_ASSERT(nApics == __kmp_avail_proc);
1835  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
1836  for (proc = 0; (int)proc < nApics; ++proc) {
1837  __kmp_pu_os_idx[proc] = retval[proc].second;
1838  }
1839  if (__kmp_affinity_type == affinity_none) {
1840  __kmp_free(last);
1841  __kmp_free(maxCt);
1842  __kmp_free(counts);
1843  __kmp_free(totals);
1844  __kmp_free(retval);
1845  KMP_CPU_FREE(oldMask);
1846  return 0;
1847  }
1848 
1849  // Find any levels with radiix 1, and remove them from the map
1850  // (except for the package level).
1851  int new_depth = 0;
1852  for (level = 0; level < depth; level++) {
1853  if ((maxCt[level] == 1) && (level != pkgLevel)) {
1854  continue;
1855  }
1856  new_depth++;
1857  }
1858 
1859  // If we are removing any levels, allocate a new vector to return,
1860  // and copy the relevant information to it.
1861  if (new_depth != depth) {
1862  AddrUnsPair *new_retval =
1863  (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * nApics);
1864  for (proc = 0; (int)proc < nApics; proc++) {
1865  Address addr(new_depth);
1866  new_retval[proc] = AddrUnsPair(addr, retval[proc].second);
1867  }
1868  int new_level = 0;
1869  int newPkgLevel = -1;
1870  int newCoreLevel = -1;
1871  int newThreadLevel = -1;
1872  for (level = 0; level < depth; level++) {
1873  if ((maxCt[level] == 1) && (level != pkgLevel)) {
1874  // Remove this level. Never remove the package level
1875  continue;
1876  }
1877  if (level == pkgLevel) {
1878  newPkgLevel = new_level;
1879  }
1880  if (level == coreLevel) {
1881  newCoreLevel = new_level;
1882  }
1883  if (level == threadLevel) {
1884  newThreadLevel = new_level;
1885  }
1886  for (proc = 0; (int)proc < nApics; proc++) {
1887  new_retval[proc].first.labels[new_level] =
1888  retval[proc].first.labels[level];
1889  }
1890  new_level++;
1891  }
1892 
1893  __kmp_free(retval);
1894  retval = new_retval;
1895  depth = new_depth;
1896  pkgLevel = newPkgLevel;
1897  coreLevel = newCoreLevel;
1898  threadLevel = newThreadLevel;
1899  }
1900 
1901  if (__kmp_affinity_gran_levels < 0) {
1902  // Set the granularity level based on what levels are modeled
1903  // in the machine topology map.
1904  __kmp_affinity_gran_levels = 0;
1905  if ((threadLevel >= 0) && (__kmp_affinity_gran > affinity_gran_thread)) {
1906  __kmp_affinity_gran_levels++;
1907  }
1908  if ((coreLevel >= 0) && (__kmp_affinity_gran > affinity_gran_core)) {
1909  __kmp_affinity_gran_levels++;
1910  }
1911  if (__kmp_affinity_gran > affinity_gran_package) {
1912  __kmp_affinity_gran_levels++;
1913  }
1914  }
1915 
1916  if (__kmp_affinity_verbose) {
1917  __kmp_affinity_print_topology(retval, nApics, depth, pkgLevel, coreLevel,
1918  threadLevel);
1919  }
1920 
1921  __kmp_free(last);
1922  __kmp_free(maxCt);
1923  __kmp_free(counts);
1924  __kmp_free(totals);
1925  KMP_CPU_FREE(oldMask);
1926  *address2os = retval;
1927  return depth;
1928 }
1929 
1930 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
1931 
1932 #define osIdIndex 0
1933 #define threadIdIndex 1
1934 #define coreIdIndex 2
1935 #define pkgIdIndex 3
1936 #define nodeIdIndex 4
1937 
1938 typedef unsigned *ProcCpuInfo;
1939 static unsigned maxIndex = pkgIdIndex;
1940 
1941 static int __kmp_affinity_cmp_ProcCpuInfo_phys_id(const void *a,
1942  const void *b) {
1943  unsigned i;
1944  const unsigned *aa = *(unsigned *const *)a;
1945  const unsigned *bb = *(unsigned *const *)b;
1946  for (i = maxIndex;; i--) {
1947  if (aa[i] < bb[i])
1948  return -1;
1949  if (aa[i] > bb[i])
1950  return 1;
1951  if (i == osIdIndex)
1952  break;
1953  }
1954  return 0;
1955 }
1956 
1957 #if KMP_USE_HIER_SCHED
1958 // Set the array sizes for the hierarchy layers
1959 static void __kmp_dispatch_set_hierarchy_values() {
1960  // Set the maximum number of L1's to number of cores
1961  // Set the maximum number of L2's to to either number of cores / 2 for
1962  // Intel(R) Xeon Phi(TM) coprocessor formally codenamed Knights Landing
1963  // Or the number of cores for Intel(R) Xeon(R) processors
1964  // Set the maximum number of NUMA nodes and L3's to number of packages
1965  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1] =
1966  nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
1967  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L1 + 1] = __kmp_ncores;
1968 #if KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
1969  if (__kmp_mic_type >= mic3)
1970  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores / 2;
1971  else
1972 #endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
1973  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores;
1974  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L3 + 1] = nPackages;
1975  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_NUMA + 1] = nPackages;
1976  __kmp_hier_max_units[kmp_hier_layer_e::LAYER_LOOP + 1] = 1;
1977  // Set the number of threads per unit
1978  // Number of hardware threads per L1/L2/L3/NUMA/LOOP
1979  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_THREAD + 1] = 1;
1980  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L1 + 1] =
1981  __kmp_nThreadsPerCore;
1982 #if KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
1983  if (__kmp_mic_type >= mic3)
1984  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
1985  2 * __kmp_nThreadsPerCore;
1986  else
1987 #endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
1988  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
1989  __kmp_nThreadsPerCore;
1990  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L3 + 1] =
1991  nCoresPerPkg * __kmp_nThreadsPerCore;
1992  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_NUMA + 1] =
1993  nCoresPerPkg * __kmp_nThreadsPerCore;
1994  __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_LOOP + 1] =
1995  nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
1996 }
1997 
1998 // Return the index into the hierarchy for this tid and layer type (L1, L2, etc)
1999 // i.e., this thread's L1 or this thread's L2, etc.
2000 int __kmp_dispatch_get_index(int tid, kmp_hier_layer_e type) {
2001  int index = type + 1;
2002  int num_hw_threads = __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1];
2003  KMP_DEBUG_ASSERT(type != kmp_hier_layer_e::LAYER_LAST);
2004  if (type == kmp_hier_layer_e::LAYER_THREAD)
2005  return tid;
2006  else if (type == kmp_hier_layer_e::LAYER_LOOP)
2007  return 0;
2008  KMP_DEBUG_ASSERT(__kmp_hier_max_units[index] != 0);
2009  if (tid >= num_hw_threads)
2010  tid = tid % num_hw_threads;
2011  return (tid / __kmp_hier_threads_per[index]) % __kmp_hier_max_units[index];
2012 }
2013 
2014 // Return the number of t1's per t2
2015 int __kmp_dispatch_get_t1_per_t2(kmp_hier_layer_e t1, kmp_hier_layer_e t2) {
2016  int i1 = t1 + 1;
2017  int i2 = t2 + 1;
2018  KMP_DEBUG_ASSERT(i1 <= i2);
2019  KMP_DEBUG_ASSERT(t1 != kmp_hier_layer_e::LAYER_LAST);
2020  KMP_DEBUG_ASSERT(t2 != kmp_hier_layer_e::LAYER_LAST);
2021  KMP_DEBUG_ASSERT(__kmp_hier_threads_per[i1] != 0);
2022  // (nthreads/t2) / (nthreads/t1) = t1 / t2
2023  return __kmp_hier_threads_per[i2] / __kmp_hier_threads_per[i1];
2024 }
2025 #endif // KMP_USE_HIER_SCHED
2026 
2027 // Parse /proc/cpuinfo (or an alternate file in the same format) to obtain the
2028 // affinity map.
2029 static int __kmp_affinity_create_cpuinfo_map(AddrUnsPair **address2os,
2030  int *line,
2031  kmp_i18n_id_t *const msg_id,
2032  FILE *f) {
2033  *address2os = NULL;
2034  *msg_id = kmp_i18n_null;
2035 
2036  // Scan of the file, and count the number of "processor" (osId) fields,
2037  // and find the highest value of <n> for a node_<n> field.
2038  char buf[256];
2039  unsigned num_records = 0;
2040  while (!feof(f)) {
2041  buf[sizeof(buf) - 1] = 1;
2042  if (!fgets(buf, sizeof(buf), f)) {
2043  // Read errors presumably because of EOF
2044  break;
2045  }
2046 
2047  char s1[] = "processor";
2048  if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
2049  num_records++;
2050  continue;
2051  }
2052 
2053  // FIXME - this will match "node_<n> <garbage>"
2054  unsigned level;
2055  if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
2056  if (nodeIdIndex + level >= maxIndex) {
2057  maxIndex = nodeIdIndex + level;
2058  }
2059  continue;
2060  }
2061  }
2062 
2063  // Check for empty file / no valid processor records, or too many. The number
2064  // of records can't exceed the number of valid bits in the affinity mask.
2065  if (num_records == 0) {
2066  *line = 0;
2067  *msg_id = kmp_i18n_str_NoProcRecords;
2068  return -1;
2069  }
2070  if (num_records > (unsigned)__kmp_xproc) {
2071  *line = 0;
2072  *msg_id = kmp_i18n_str_TooManyProcRecords;
2073  return -1;
2074  }
2075 
2076  // Set the file pointer back to the begginning, so that we can scan the file
2077  // again, this time performing a full parse of the data. Allocate a vector of
2078  // ProcCpuInfo object, where we will place the data. Adding an extra element
2079  // at the end allows us to remove a lot of extra checks for termination
2080  // conditions.
2081  if (fseek(f, 0, SEEK_SET) != 0) {
2082  *line = 0;
2083  *msg_id = kmp_i18n_str_CantRewindCpuinfo;
2084  return -1;
2085  }
2086 
2087  // Allocate the array of records to store the proc info in. The dummy
2088  // element at the end makes the logic in filling them out easier to code.
2089  unsigned **threadInfo =
2090  (unsigned **)__kmp_allocate((num_records + 1) * sizeof(unsigned *));
2091  unsigned i;
2092  for (i = 0; i <= num_records; i++) {
2093  threadInfo[i] =
2094  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2095  }
2096 
2097 #define CLEANUP_THREAD_INFO \
2098  for (i = 0; i <= num_records; i++) { \
2099  __kmp_free(threadInfo[i]); \
2100  } \
2101  __kmp_free(threadInfo);
2102 
2103  // A value of UINT_MAX means that we didn't find the field
2104  unsigned __index;
2105 
2106 #define INIT_PROC_INFO(p) \
2107  for (__index = 0; __index <= maxIndex; __index++) { \
2108  (p)[__index] = UINT_MAX; \
2109  }
2110 
2111  for (i = 0; i <= num_records; i++) {
2112  INIT_PROC_INFO(threadInfo[i]);
2113  }
2114 
2115  unsigned num_avail = 0;
2116  *line = 0;
2117  while (!feof(f)) {
2118  // Create an inner scoping level, so that all the goto targets at the end of
2119  // the loop appear in an outer scoping level. This avoids warnings about
2120  // jumping past an initialization to a target in the same block.
2121  {
2122  buf[sizeof(buf) - 1] = 1;
2123  bool long_line = false;
2124  if (!fgets(buf, sizeof(buf), f)) {
2125  // Read errors presumably because of EOF
2126  // If there is valid data in threadInfo[num_avail], then fake
2127  // a blank line in ensure that the last address gets parsed.
2128  bool valid = false;
2129  for (i = 0; i <= maxIndex; i++) {
2130  if (threadInfo[num_avail][i] != UINT_MAX) {
2131  valid = true;
2132  }
2133  }
2134  if (!valid) {
2135  break;
2136  }
2137  buf[0] = 0;
2138  } else if (!buf[sizeof(buf) - 1]) {
2139  // The line is longer than the buffer. Set a flag and don't
2140  // emit an error if we were going to ignore the line, anyway.
2141  long_line = true;
2142 
2143 #define CHECK_LINE \
2144  if (long_line) { \
2145  CLEANUP_THREAD_INFO; \
2146  *msg_id = kmp_i18n_str_LongLineCpuinfo; \
2147  return -1; \
2148  }
2149  }
2150  (*line)++;
2151 
2152  char s1[] = "processor";
2153  if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
2154  CHECK_LINE;
2155  char *p = strchr(buf + sizeof(s1) - 1, ':');
2156  unsigned val;
2157  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2158  goto no_val;
2159  if (threadInfo[num_avail][osIdIndex] != UINT_MAX)
2160 #if KMP_ARCH_AARCH64
2161  // Handle the old AArch64 /proc/cpuinfo layout differently,
2162  // it contains all of the 'processor' entries listed in a
2163  // single 'Processor' section, therefore the normal looking
2164  // for duplicates in that section will always fail.
2165  num_avail++;
2166 #else
2167  goto dup_field;
2168 #endif
2169  threadInfo[num_avail][osIdIndex] = val;
2170 #if KMP_OS_LINUX && !(KMP_ARCH_X86 || KMP_ARCH_X86_64)
2171  char path[256];
2172  KMP_SNPRINTF(
2173  path, sizeof(path),
2174  "/sys/devices/system/cpu/cpu%u/topology/physical_package_id",
2175  threadInfo[num_avail][osIdIndex]);
2176  __kmp_read_from_file(path, "%u", &threadInfo[num_avail][pkgIdIndex]);
2177 
2178  KMP_SNPRINTF(path, sizeof(path),
2179  "/sys/devices/system/cpu/cpu%u/topology/core_id",
2180  threadInfo[num_avail][osIdIndex]);
2181  __kmp_read_from_file(path, "%u", &threadInfo[num_avail][coreIdIndex]);
2182  continue;
2183 #else
2184  }
2185  char s2[] = "physical id";
2186  if (strncmp(buf, s2, sizeof(s2) - 1) == 0) {
2187  CHECK_LINE;
2188  char *p = strchr(buf + sizeof(s2) - 1, ':');
2189  unsigned val;
2190  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2191  goto no_val;
2192  if (threadInfo[num_avail][pkgIdIndex] != UINT_MAX)
2193  goto dup_field;
2194  threadInfo[num_avail][pkgIdIndex] = val;
2195  continue;
2196  }
2197  char s3[] = "core id";
2198  if (strncmp(buf, s3, sizeof(s3) - 1) == 0) {
2199  CHECK_LINE;
2200  char *p = strchr(buf + sizeof(s3) - 1, ':');
2201  unsigned val;
2202  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2203  goto no_val;
2204  if (threadInfo[num_avail][coreIdIndex] != UINT_MAX)
2205  goto dup_field;
2206  threadInfo[num_avail][coreIdIndex] = val;
2207  continue;
2208 #endif // KMP_OS_LINUX && USE_SYSFS_INFO
2209  }
2210  char s4[] = "thread id";
2211  if (strncmp(buf, s4, sizeof(s4) - 1) == 0) {
2212  CHECK_LINE;
2213  char *p = strchr(buf + sizeof(s4) - 1, ':');
2214  unsigned val;
2215  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2216  goto no_val;
2217  if (threadInfo[num_avail][threadIdIndex] != UINT_MAX)
2218  goto dup_field;
2219  threadInfo[num_avail][threadIdIndex] = val;
2220  continue;
2221  }
2222  unsigned level;
2223  if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
2224  CHECK_LINE;
2225  char *p = strchr(buf + sizeof(s4) - 1, ':');
2226  unsigned val;
2227  if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1))
2228  goto no_val;
2229  KMP_ASSERT(nodeIdIndex + level <= maxIndex);
2230  if (threadInfo[num_avail][nodeIdIndex + level] != UINT_MAX)
2231  goto dup_field;
2232  threadInfo[num_avail][nodeIdIndex + level] = val;
2233  continue;
2234  }
2235 
2236  // We didn't recognize the leading token on the line. There are lots of
2237  // leading tokens that we don't recognize - if the line isn't empty, go on
2238  // to the next line.
2239  if ((*buf != 0) && (*buf != '\n')) {
2240  // If the line is longer than the buffer, read characters
2241  // until we find a newline.
2242  if (long_line) {
2243  int ch;
2244  while (((ch = fgetc(f)) != EOF) && (ch != '\n'))
2245  ;
2246  }
2247  continue;
2248  }
2249 
2250  // A newline has signalled the end of the processor record.
2251  // Check that there aren't too many procs specified.
2252  if ((int)num_avail == __kmp_xproc) {
2253  CLEANUP_THREAD_INFO;
2254  *msg_id = kmp_i18n_str_TooManyEntries;
2255  return -1;
2256  }
2257 
2258  // Check for missing fields. The osId field must be there, and we
2259  // currently require that the physical id field is specified, also.
2260  if (threadInfo[num_avail][osIdIndex] == UINT_MAX) {
2261  CLEANUP_THREAD_INFO;
2262  *msg_id = kmp_i18n_str_MissingProcField;
2263  return -1;
2264  }
2265  if (threadInfo[0][pkgIdIndex] == UINT_MAX) {
2266  CLEANUP_THREAD_INFO;
2267  *msg_id = kmp_i18n_str_MissingPhysicalIDField;
2268  return -1;
2269  }
2270 
2271  // Skip this proc if it is not included in the machine model.
2272  if (!KMP_CPU_ISSET(threadInfo[num_avail][osIdIndex],
2273  __kmp_affin_fullMask)) {
2274  INIT_PROC_INFO(threadInfo[num_avail]);
2275  continue;
2276  }
2277 
2278  // We have a successful parse of this proc's info.
2279  // Increment the counter, and prepare for the next proc.
2280  num_avail++;
2281  KMP_ASSERT(num_avail <= num_records);
2282  INIT_PROC_INFO(threadInfo[num_avail]);
2283  }
2284  continue;
2285 
2286  no_val:
2287  CLEANUP_THREAD_INFO;
2288  *msg_id = kmp_i18n_str_MissingValCpuinfo;
2289  return -1;
2290 
2291  dup_field:
2292  CLEANUP_THREAD_INFO;
2293  *msg_id = kmp_i18n_str_DuplicateFieldCpuinfo;
2294  return -1;
2295  }
2296  *line = 0;
2297 
2298 #if KMP_MIC && REDUCE_TEAM_SIZE
2299  unsigned teamSize = 0;
2300 #endif // KMP_MIC && REDUCE_TEAM_SIZE
2301 
2302  // check for num_records == __kmp_xproc ???
2303 
2304  // If there's only one thread context to bind to, form an Address object with
2305  // depth 1 and return immediately (or, if affinity is off, set address2os to
2306  // NULL and return).
2307  //
2308  // If it is configured to omit the package level when there is only a single
2309  // package, the logic at the end of this routine won't work if there is only a
2310  // single thread - it would try to form an Address object with depth 0.
2311  KMP_ASSERT(num_avail > 0);
2312  KMP_ASSERT(num_avail <= num_records);
2313  if (num_avail == 1) {
2314  __kmp_ncores = 1;
2315  __kmp_nThreadsPerCore = nCoresPerPkg = nPackages = 1;
2316  if (__kmp_affinity_verbose) {
2317  if (!KMP_AFFINITY_CAPABLE()) {
2318  KMP_INFORM(AffNotCapableUseCpuinfo, "KMP_AFFINITY");
2319  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
2320  KMP_INFORM(Uniform, "KMP_AFFINITY");
2321  } else {
2322  char buf[KMP_AFFIN_MASK_PRINT_LEN];
2323  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
2324  __kmp_affin_fullMask);
2325  KMP_INFORM(AffCapableUseCpuinfo, "KMP_AFFINITY");
2326  if (__kmp_affinity_respect_mask) {
2327  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
2328  } else {
2329  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
2330  }
2331  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
2332  KMP_INFORM(Uniform, "KMP_AFFINITY");
2333  }
2334  int index;
2335  kmp_str_buf_t buf;
2336  __kmp_str_buf_init(&buf);
2337  __kmp_str_buf_print(&buf, "1");
2338  for (index = maxIndex - 1; index > pkgIdIndex; index--) {
2339  __kmp_str_buf_print(&buf, " x 1");
2340  }
2341  KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, 1, 1, 1);
2342  __kmp_str_buf_free(&buf);
2343  }
2344 
2345  if (__kmp_affinity_type == affinity_none) {
2346  CLEANUP_THREAD_INFO;
2347  return 0;
2348  }
2349 
2350  *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair));
2351  Address addr(1);
2352  addr.labels[0] = threadInfo[0][pkgIdIndex];
2353  (*address2os)[0] = AddrUnsPair(addr, threadInfo[0][osIdIndex]);
2354 
2355  if (__kmp_affinity_gran_levels < 0) {
2356  __kmp_affinity_gran_levels = 0;
2357  }
2358 
2359  if (__kmp_affinity_verbose) {
2360  __kmp_affinity_print_topology(*address2os, 1, 1, 0, -1, -1);
2361  }
2362 
2363  CLEANUP_THREAD_INFO;
2364  return 1;
2365  }
2366 
2367  // Sort the threadInfo table by physical Id.
2368  qsort(threadInfo, num_avail, sizeof(*threadInfo),
2369  __kmp_affinity_cmp_ProcCpuInfo_phys_id);
2370 
2371  // The table is now sorted by pkgId / coreId / threadId, but we really don't
2372  // know the radix of any of the fields. pkgId's may be sparsely assigned among
2373  // the chips on a system. Although coreId's are usually assigned
2374  // [0 .. coresPerPkg-1] and threadId's are usually assigned
2375  // [0..threadsPerCore-1], we don't want to make any such assumptions.
2376  //
2377  // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
2378  // total # packages) are at this point - we want to determine that now. We
2379  // only have an upper bound on the first two figures.
2380  unsigned *counts =
2381  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2382  unsigned *maxCt =
2383  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2384  unsigned *totals =
2385  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2386  unsigned *lastId =
2387  (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
2388 
2389  bool assign_thread_ids = false;
2390  unsigned threadIdCt;
2391  unsigned index;
2392 
2393 restart_radix_check:
2394  threadIdCt = 0;
2395 
2396  // Initialize the counter arrays with data from threadInfo[0].
2397  if (assign_thread_ids) {
2398  if (threadInfo[0][threadIdIndex] == UINT_MAX) {
2399  threadInfo[0][threadIdIndex] = threadIdCt++;
2400  } else if (threadIdCt <= threadInfo[0][threadIdIndex]) {
2401  threadIdCt = threadInfo[0][threadIdIndex] + 1;
2402  }
2403  }
2404  for (index = 0; index <= maxIndex; index++) {
2405  counts[index] = 1;
2406  maxCt[index] = 1;
2407  totals[index] = 1;
2408  lastId[index] = threadInfo[0][index];
2409  ;
2410  }
2411 
2412  // Run through the rest of the OS procs.
2413  for (i = 1; i < num_avail; i++) {
2414  // Find the most significant index whose id differs from the id for the
2415  // previous OS proc.
2416  for (index = maxIndex; index >= threadIdIndex; index--) {
2417  if (assign_thread_ids && (index == threadIdIndex)) {
2418  // Auto-assign the thread id field if it wasn't specified.
2419  if (threadInfo[i][threadIdIndex] == UINT_MAX) {
2420  threadInfo[i][threadIdIndex] = threadIdCt++;
2421  }
2422  // Apparently the thread id field was specified for some entries and not
2423  // others. Start the thread id counter off at the next higher thread id.
2424  else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
2425  threadIdCt = threadInfo[i][threadIdIndex] + 1;
2426  }
2427  }
2428  if (threadInfo[i][index] != lastId[index]) {
2429  // Run through all indices which are less significant, and reset the
2430  // counts to 1. At all levels up to and including index, we need to
2431  // increment the totals and record the last id.
2432  unsigned index2;
2433  for (index2 = threadIdIndex; index2 < index; index2++) {
2434  totals[index2]++;
2435  if (counts[index2] > maxCt[index2]) {
2436  maxCt[index2] = counts[index2];
2437  }
2438  counts[index2] = 1;
2439  lastId[index2] = threadInfo[i][index2];
2440  }
2441  counts[index]++;
2442  totals[index]++;
2443  lastId[index] = threadInfo[i][index];
2444 
2445  if (assign_thread_ids && (index > threadIdIndex)) {
2446 
2447 #if KMP_MIC && REDUCE_TEAM_SIZE
2448  // The default team size is the total #threads in the machine
2449  // minus 1 thread for every core that has 3 or more threads.
2450  teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
2451 #endif // KMP_MIC && REDUCE_TEAM_SIZE
2452 
2453  // Restart the thread counter, as we are on a new core.
2454  threadIdCt = 0;
2455 
2456  // Auto-assign the thread id field if it wasn't specified.
2457  if (threadInfo[i][threadIdIndex] == UINT_MAX) {
2458  threadInfo[i][threadIdIndex] = threadIdCt++;
2459  }
2460 
2461  // Aparrently the thread id field was specified for some entries and
2462  // not others. Start the thread id counter off at the next higher
2463  // thread id.
2464  else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
2465  threadIdCt = threadInfo[i][threadIdIndex] + 1;
2466  }
2467  }
2468  break;
2469  }
2470  }
2471  if (index < threadIdIndex) {
2472  // If thread ids were specified, it is an error if they are not unique.
2473  // Also, check that we waven't already restarted the loop (to be safe -
2474  // shouldn't need to).
2475  if ((threadInfo[i][threadIdIndex] != UINT_MAX) || assign_thread_ids) {
2476  __kmp_free(lastId);
2477  __kmp_free(totals);
2478  __kmp_free(maxCt);
2479  __kmp_free(counts);
2480  CLEANUP_THREAD_INFO;
2481  *msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
2482  return -1;
2483  }
2484 
2485  // If the thread ids were not specified and we see entries entries that
2486  // are duplicates, start the loop over and assign the thread ids manually.
2487  assign_thread_ids = true;
2488  goto restart_radix_check;
2489  }
2490  }
2491 
2492 #if KMP_MIC && REDUCE_TEAM_SIZE
2493  // The default team size is the total #threads in the machine
2494  // minus 1 thread for every core that has 3 or more threads.
2495  teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
2496 #endif // KMP_MIC && REDUCE_TEAM_SIZE
2497 
2498  for (index = threadIdIndex; index <= maxIndex; index++) {
2499  if (counts[index] > maxCt[index]) {
2500  maxCt[index] = counts[index];
2501  }
2502  }
2503 
2504  __kmp_nThreadsPerCore = maxCt[threadIdIndex];
2505  nCoresPerPkg = maxCt[coreIdIndex];
2506  nPackages = totals[pkgIdIndex];
2507 
2508  // Check to see if the machine topology is uniform
2509  unsigned prod = totals[maxIndex];
2510  for (index = threadIdIndex; index < maxIndex; index++) {
2511  prod *= maxCt[index];
2512  }
2513  bool uniform = (prod == totals[threadIdIndex]);
2514 
2515  // When affinity is off, this routine will still be called to set
2516  // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
2517  // Make sure all these vars are set correctly, and return now if affinity is
2518  // not enabled.
2519  __kmp_ncores = totals[coreIdIndex];
2520 
2521  if (__kmp_affinity_verbose) {
2522  if (!KMP_AFFINITY_CAPABLE()) {
2523  KMP_INFORM(AffNotCapableUseCpuinfo, "KMP_AFFINITY");
2524  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
2525  if (uniform) {
2526  KMP_INFORM(Uniform, "KMP_AFFINITY");
2527  } else {
2528  KMP_INFORM(NonUniform, "KMP_AFFINITY");
2529  }
2530  } else {
2531  char buf[KMP_AFFIN_MASK_PRINT_LEN];
2532  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
2533  __kmp_affin_fullMask);
2534  KMP_INFORM(AffCapableUseCpuinfo, "KMP_AFFINITY");
2535  if (__kmp_affinity_respect_mask) {
2536  KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf);
2537  } else {
2538  KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf);
2539  }
2540  KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
2541  if (uniform) {
2542  KMP_INFORM(Uniform, "KMP_AFFINITY");
2543  } else {
2544  KMP_INFORM(NonUniform, "KMP_AFFINITY");
2545  }
2546  }
2547  kmp_str_buf_t buf;
2548  __kmp_str_buf_init(&buf);
2549 
2550  __kmp_str_buf_print(&buf, "%d", totals[maxIndex]);
2551  for (index = maxIndex - 1; index >= pkgIdIndex; index--) {
2552  __kmp_str_buf_print(&buf, " x %d", maxCt[index]);
2553  }
2554  KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, maxCt[coreIdIndex],
2555  maxCt[threadIdIndex], __kmp_ncores);
2556 
2557  __kmp_str_buf_free(&buf);
2558  }
2559 
2560 #if KMP_MIC && REDUCE_TEAM_SIZE
2561  // Set the default team size.
2562  if ((__kmp_dflt_team_nth == 0) && (teamSize > 0)) {
2563  __kmp_dflt_team_nth = teamSize;
2564  KA_TRACE(20, ("__kmp_affinity_create_cpuinfo_map: setting "
2565  "__kmp_dflt_team_nth = %d\n",
2566  __kmp_dflt_team_nth));
2567  }
2568 #endif // KMP_MIC && REDUCE_TEAM_SIZE
2569 
2570  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
2571  KMP_DEBUG_ASSERT(num_avail == (unsigned)__kmp_avail_proc);
2572  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
2573  for (i = 0; i < num_avail; ++i) { // fill the os indices
2574  __kmp_pu_os_idx[i] = threadInfo[i][osIdIndex];
2575  }
2576 
2577  if (__kmp_affinity_type == affinity_none) {
2578  __kmp_free(lastId);
2579  __kmp_free(totals);
2580  __kmp_free(maxCt);
2581  __kmp_free(counts);
2582  CLEANUP_THREAD_INFO;
2583  return 0;
2584  }
2585 
2586  // Count the number of levels which have more nodes at that level than at the
2587  // parent's level (with there being an implicit root node of the top level).
2588  // This is equivalent to saying that there is at least one node at this level
2589  // which has a sibling. These levels are in the map, and the package level is
2590  // always in the map.
2591  bool *inMap = (bool *)__kmp_allocate((maxIndex + 1) * sizeof(bool));
2592  for (index = threadIdIndex; index < maxIndex; index++) {
2593  KMP_ASSERT(totals[index] >= totals[index + 1]);
2594  inMap[index] = (totals[index] > totals[index + 1]);
2595  }
2596  inMap[maxIndex] = (totals[maxIndex] > 1);
2597  inMap[pkgIdIndex] = true;
2598 
2599  int depth = 0;
2600  for (index = threadIdIndex; index <= maxIndex; index++) {
2601  if (inMap[index]) {
2602  depth++;
2603  }
2604  }
2605  KMP_ASSERT(depth > 0);
2606 
2607  // Construct the data structure that is to be returned.
2608  *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * num_avail);
2609  int pkgLevel = -1;
2610  int coreLevel = -1;
2611  int threadLevel = -1;
2612 
2613  for (i = 0; i < num_avail; ++i) {
2614  Address addr(depth);
2615  unsigned os = threadInfo[i][osIdIndex];
2616  int src_index;
2617  int dst_index = 0;
2618 
2619  for (src_index = maxIndex; src_index >= threadIdIndex; src_index--) {
2620  if (!inMap[src_index]) {
2621  continue;
2622  }
2623  addr.labels[dst_index] = threadInfo[i][src_index];
2624  if (src_index == pkgIdIndex) {
2625  pkgLevel = dst_index;
2626  } else if (src_index == coreIdIndex) {
2627  coreLevel = dst_index;
2628  } else if (src_index == threadIdIndex) {
2629  threadLevel = dst_index;
2630  }
2631  dst_index++;
2632  }
2633  (*address2os)[i] = AddrUnsPair(addr, os);
2634  }
2635 
2636  if (__kmp_affinity_gran_levels < 0) {
2637  // Set the granularity level based on what levels are modeled
2638  // in the machine topology map.
2639  unsigned src_index;
2640  __kmp_affinity_gran_levels = 0;
2641  for (src_index = threadIdIndex; src_index <= maxIndex; src_index++) {
2642  if (!inMap[src_index]) {
2643  continue;
2644  }
2645  switch (src_index) {
2646  case threadIdIndex:
2647  if (__kmp_affinity_gran > affinity_gran_thread) {
2648  __kmp_affinity_gran_levels++;
2649  }
2650 
2651  break;
2652  case coreIdIndex:
2653  if (__kmp_affinity_gran > affinity_gran_core) {
2654  __kmp_affinity_gran_levels++;
2655  }
2656  break;
2657 
2658  case pkgIdIndex:
2659  if (__kmp_affinity_gran > affinity_gran_package) {
2660  __kmp_affinity_gran_levels++;
2661  }
2662  break;
2663  }
2664  }
2665  }
2666 
2667  if (__kmp_affinity_verbose) {
2668  __kmp_affinity_print_topology(*address2os, num_avail, depth, pkgLevel,
2669  coreLevel, threadLevel);
2670  }
2671 
2672  __kmp_free(inMap);
2673  __kmp_free(lastId);
2674  __kmp_free(totals);
2675  __kmp_free(maxCt);
2676  __kmp_free(counts);
2677  CLEANUP_THREAD_INFO;
2678  return depth;
2679 }
2680 
2681 // Create and return a table of affinity masks, indexed by OS thread ID.
2682 // This routine handles OR'ing together all the affinity masks of threads
2683 // that are sufficiently close, if granularity > fine.
2684 static kmp_affin_mask_t *__kmp_create_masks(unsigned *maxIndex,
2685  unsigned *numUnique,
2686  AddrUnsPair *address2os,
2687  unsigned numAddrs) {
2688  // First form a table of affinity masks in order of OS thread id.
2689  unsigned depth;
2690  unsigned maxOsId;
2691  unsigned i;
2692 
2693  KMP_ASSERT(numAddrs > 0);
2694  depth = address2os[0].first.depth;
2695 
2696  maxOsId = 0;
2697  for (i = numAddrs - 1;; --i) {
2698  unsigned osId = address2os[i].second;
2699  if (osId > maxOsId) {
2700  maxOsId = osId;
2701  }
2702  if (i == 0)
2703  break;
2704  }
2705  kmp_affin_mask_t *osId2Mask;
2706  KMP_CPU_ALLOC_ARRAY(osId2Mask, (maxOsId + 1));
2707 
2708  // Sort the address2os table according to physical order. Doing so will put
2709  // all threads on the same core/package/node in consecutive locations.
2710  qsort(address2os, numAddrs, sizeof(*address2os),
2711  __kmp_affinity_cmp_Address_labels);
2712 
2713  KMP_ASSERT(__kmp_affinity_gran_levels >= 0);
2714  if (__kmp_affinity_verbose && (__kmp_affinity_gran_levels > 0)) {
2715  KMP_INFORM(ThreadsMigrate, "KMP_AFFINITY", __kmp_affinity_gran_levels);
2716  }
2717  if (__kmp_affinity_gran_levels >= (int)depth) {
2718  if (__kmp_affinity_verbose ||
2719  (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) {
2720  KMP_WARNING(AffThreadsMayMigrate);
2721  }
2722  }
2723 
2724  // Run through the table, forming the masks for all threads on each core.
2725  // Threads on the same core will have identical "Address" objects, not
2726  // considering the last level, which must be the thread id. All threads on a
2727  // core will appear consecutively.
2728  unsigned unique = 0;
2729  unsigned j = 0; // index of 1st thread on core
2730  unsigned leader = 0;
2731  Address *leaderAddr = &(address2os[0].first);
2732  kmp_affin_mask_t *sum;
2733  KMP_CPU_ALLOC_ON_STACK(sum);
2734  KMP_CPU_ZERO(sum);
2735  KMP_CPU_SET(address2os[0].second, sum);
2736  for (i = 1; i < numAddrs; i++) {
2737  // If this thread is sufficiently close to the leader (within the
2738  // granularity setting), then set the bit for this os thread in the
2739  // affinity mask for this group, and go on to the next thread.
2740  if (leaderAddr->isClose(address2os[i].first, __kmp_affinity_gran_levels)) {
2741  KMP_CPU_SET(address2os[i].second, sum);
2742  continue;
2743  }
2744 
2745  // For every thread in this group, copy the mask to the thread's entry in
2746  // the osId2Mask table. Mark the first address as a leader.
2747  for (; j < i; j++) {
2748  unsigned osId = address2os[j].second;
2749  KMP_DEBUG_ASSERT(osId <= maxOsId);
2750  kmp_affin_mask_t *mask = KMP_CPU_INDEX(osId2Mask, osId);
2751  KMP_CPU_COPY(mask, sum);
2752  address2os[j].first.leader = (j == leader);
2753  }
2754  unique++;
2755 
2756  // Start a new mask.
2757  leader = i;
2758  leaderAddr = &(address2os[i].first);
2759  KMP_CPU_ZERO(sum);
2760  KMP_CPU_SET(address2os[i].second, sum);
2761  }
2762 
2763  // For every thread in last group, copy the mask to the thread's
2764  // entry in the osId2Mask table.
2765  for (; j < i; j++) {
2766  unsigned osId = address2os[j].second;
2767  KMP_DEBUG_ASSERT(osId <= maxOsId);
2768  kmp_affin_mask_t *mask = KMP_CPU_INDEX(osId2Mask, osId);
2769  KMP_CPU_COPY(mask, sum);
2770  address2os[j].first.leader = (j == leader);
2771  }
2772  unique++;
2773  KMP_CPU_FREE_FROM_STACK(sum);
2774 
2775  *maxIndex = maxOsId;
2776  *numUnique = unique;
2777  return osId2Mask;
2778 }
2779 
2780 // Stuff for the affinity proclist parsers. It's easier to declare these vars
2781 // as file-static than to try and pass them through the calling sequence of
2782 // the recursive-descent OMP_PLACES parser.
2783 static kmp_affin_mask_t *newMasks;
2784 static int numNewMasks;
2785 static int nextNewMask;
2786 
2787 #define ADD_MASK(_mask) \
2788  { \
2789  if (nextNewMask >= numNewMasks) { \
2790  int i; \
2791  numNewMasks *= 2; \
2792  kmp_affin_mask_t *temp; \
2793  KMP_CPU_INTERNAL_ALLOC_ARRAY(temp, numNewMasks); \
2794  for (i = 0; i < numNewMasks / 2; i++) { \
2795  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i); \
2796  kmp_affin_mask_t *dest = KMP_CPU_INDEX(temp, i); \
2797  KMP_CPU_COPY(dest, src); \
2798  } \
2799  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks / 2); \
2800  newMasks = temp; \
2801  } \
2802  KMP_CPU_COPY(KMP_CPU_INDEX(newMasks, nextNewMask), (_mask)); \
2803  nextNewMask++; \
2804  }
2805 
2806 #define ADD_MASK_OSID(_osId, _osId2Mask, _maxOsId) \
2807  { \
2808  if (((_osId) > _maxOsId) || \
2809  (!KMP_CPU_ISSET((_osId), KMP_CPU_INDEX((_osId2Mask), (_osId))))) { \
2810  if (__kmp_affinity_verbose || \
2811  (__kmp_affinity_warnings && \
2812  (__kmp_affinity_type != affinity_none))) { \
2813  KMP_WARNING(AffIgnoreInvalidProcID, _osId); \
2814  } \
2815  } else { \
2816  ADD_MASK(KMP_CPU_INDEX(_osId2Mask, (_osId))); \
2817  } \
2818  }
2819 
2820 // Re-parse the proclist (for the explicit affinity type), and form the list
2821 // of affinity newMasks indexed by gtid.
2822 static void __kmp_affinity_process_proclist(kmp_affin_mask_t **out_masks,
2823  unsigned int *out_numMasks,
2824  const char *proclist,
2825  kmp_affin_mask_t *osId2Mask,
2826  int maxOsId) {
2827  int i;
2828  const char *scan = proclist;
2829  const char *next = proclist;
2830 
2831  // We use malloc() for the temporary mask vector, so that we can use
2832  // realloc() to extend it.
2833  numNewMasks = 2;
2834  KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
2835  nextNewMask = 0;
2836  kmp_affin_mask_t *sumMask;
2837  KMP_CPU_ALLOC(sumMask);
2838  int setSize = 0;
2839 
2840  for (;;) {
2841  int start, end, stride;
2842 
2843  SKIP_WS(scan);
2844  next = scan;
2845  if (*next == '\0') {
2846  break;
2847  }
2848 
2849  if (*next == '{') {
2850  int num;
2851  setSize = 0;
2852  next++; // skip '{'
2853  SKIP_WS(next);
2854  scan = next;
2855 
2856  // Read the first integer in the set.
2857  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad proclist");
2858  SKIP_DIGITS(next);
2859  num = __kmp_str_to_int(scan, *next);
2860  KMP_ASSERT2(num >= 0, "bad explicit proc list");
2861 
2862  // Copy the mask for that osId to the sum (union) mask.
2863  if ((num > maxOsId) ||
2864  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
2865  if (__kmp_affinity_verbose ||
2866  (__kmp_affinity_warnings &&
2867  (__kmp_affinity_type != affinity_none))) {
2868  KMP_WARNING(AffIgnoreInvalidProcID, num);
2869  }
2870  KMP_CPU_ZERO(sumMask);
2871  } else {
2872  KMP_CPU_COPY(sumMask, KMP_CPU_INDEX(osId2Mask, num));
2873  setSize = 1;
2874  }
2875 
2876  for (;;) {
2877  // Check for end of set.
2878  SKIP_WS(next);
2879  if (*next == '}') {
2880  next++; // skip '}'
2881  break;
2882  }
2883 
2884  // Skip optional comma.
2885  if (*next == ',') {
2886  next++;
2887  }
2888  SKIP_WS(next);
2889 
2890  // Read the next integer in the set.
2891  scan = next;
2892  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
2893 
2894  SKIP_DIGITS(next);
2895  num = __kmp_str_to_int(scan, *next);
2896  KMP_ASSERT2(num >= 0, "bad explicit proc list");
2897 
2898  // Add the mask for that osId to the sum mask.
2899  if ((num > maxOsId) ||
2900  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
2901  if (__kmp_affinity_verbose ||
2902  (__kmp_affinity_warnings &&
2903  (__kmp_affinity_type != affinity_none))) {
2904  KMP_WARNING(AffIgnoreInvalidProcID, num);
2905  }
2906  } else {
2907  KMP_CPU_UNION(sumMask, KMP_CPU_INDEX(osId2Mask, num));
2908  setSize++;
2909  }
2910  }
2911  if (setSize > 0) {
2912  ADD_MASK(sumMask);
2913  }
2914 
2915  SKIP_WS(next);
2916  if (*next == ',') {
2917  next++;
2918  }
2919  scan = next;
2920  continue;
2921  }
2922 
2923  // Read the first integer.
2924  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
2925  SKIP_DIGITS(next);
2926  start = __kmp_str_to_int(scan, *next);
2927  KMP_ASSERT2(start >= 0, "bad explicit proc list");
2928  SKIP_WS(next);
2929 
2930  // If this isn't a range, then add a mask to the list and go on.
2931  if (*next != '-') {
2932  ADD_MASK_OSID(start, osId2Mask, maxOsId);
2933 
2934  // Skip optional comma.
2935  if (*next == ',') {
2936  next++;
2937  }
2938  scan = next;
2939  continue;
2940  }
2941 
2942  // This is a range. Skip over the '-' and read in the 2nd int.
2943  next++; // skip '-'
2944  SKIP_WS(next);
2945  scan = next;
2946  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
2947  SKIP_DIGITS(next);
2948  end = __kmp_str_to_int(scan, *next);
2949  KMP_ASSERT2(end >= 0, "bad explicit proc list");
2950 
2951  // Check for a stride parameter
2952  stride = 1;
2953  SKIP_WS(next);
2954  if (*next == ':') {
2955  // A stride is specified. Skip over the ':" and read the 3rd int.
2956  int sign = +1;
2957  next++; // skip ':'
2958  SKIP_WS(next);
2959  scan = next;
2960  if (*next == '-') {
2961  sign = -1;
2962  next++;
2963  SKIP_WS(next);
2964  scan = next;
2965  }
2966  KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
2967  SKIP_DIGITS(next);
2968  stride = __kmp_str_to_int(scan, *next);
2969  KMP_ASSERT2(stride >= 0, "bad explicit proc list");
2970  stride *= sign;
2971  }
2972 
2973  // Do some range checks.
2974  KMP_ASSERT2(stride != 0, "bad explicit proc list");
2975  if (stride > 0) {
2976  KMP_ASSERT2(start <= end, "bad explicit proc list");
2977  } else {
2978  KMP_ASSERT2(start >= end, "bad explicit proc list");
2979  }
2980  KMP_ASSERT2((end - start) / stride <= 65536, "bad explicit proc list");
2981 
2982  // Add the mask for each OS proc # to the list.
2983  if (stride > 0) {
2984  do {
2985  ADD_MASK_OSID(start, osId2Mask, maxOsId);
2986  start += stride;
2987  } while (start <= end);
2988  } else {
2989  do {
2990  ADD_MASK_OSID(start, osId2Mask, maxOsId);
2991  start += stride;
2992  } while (start >= end);
2993  }
2994 
2995  // Skip optional comma.
2996  SKIP_WS(next);
2997  if (*next == ',') {
2998  next++;
2999  }
3000  scan = next;
3001  }
3002 
3003  *out_numMasks = nextNewMask;
3004  if (nextNewMask == 0) {
3005  *out_masks = NULL;
3006  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3007  return;
3008  }
3009  KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
3010  for (i = 0; i < nextNewMask; i++) {
3011  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
3012  kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
3013  KMP_CPU_COPY(dest, src);
3014  }
3015  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3016  KMP_CPU_FREE(sumMask);
3017 }
3018 
3019 #if OMP_40_ENABLED
3020 
3021 /*-----------------------------------------------------------------------------
3022 Re-parse the OMP_PLACES proc id list, forming the newMasks for the different
3023 places. Again, Here is the grammar:
3024 
3025 place_list := place
3026 place_list := place , place_list
3027 place := num
3028 place := place : num
3029 place := place : num : signed
3030 place := { subplacelist }
3031 place := ! place // (lowest priority)
3032 subplace_list := subplace
3033 subplace_list := subplace , subplace_list
3034 subplace := num
3035 subplace := num : num
3036 subplace := num : num : signed
3037 signed := num
3038 signed := + signed
3039 signed := - signed
3040 -----------------------------------------------------------------------------*/
3041 
3042 static void __kmp_process_subplace_list(const char **scan,
3043  kmp_affin_mask_t *osId2Mask,
3044  int maxOsId, kmp_affin_mask_t *tempMask,
3045  int *setSize) {
3046  const char *next;
3047 
3048  for (;;) {
3049  int start, count, stride, i;
3050 
3051  // Read in the starting proc id
3052  SKIP_WS(*scan);
3053  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3054  next = *scan;
3055  SKIP_DIGITS(next);
3056  start = __kmp_str_to_int(*scan, *next);
3057  KMP_ASSERT(start >= 0);
3058  *scan = next;
3059 
3060  // valid follow sets are ',' ':' and '}'
3061  SKIP_WS(*scan);
3062  if (**scan == '}' || **scan == ',') {
3063  if ((start > maxOsId) ||
3064  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3065  if (__kmp_affinity_verbose ||
3066  (__kmp_affinity_warnings &&
3067  (__kmp_affinity_type != affinity_none))) {
3068  KMP_WARNING(AffIgnoreInvalidProcID, start);
3069  }
3070  } else {
3071  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3072  (*setSize)++;
3073  }
3074  if (**scan == '}') {
3075  break;
3076  }
3077  (*scan)++; // skip ','
3078  continue;
3079  }
3080  KMP_ASSERT2(**scan == ':', "bad explicit places list");
3081  (*scan)++; // skip ':'
3082 
3083  // Read count parameter
3084  SKIP_WS(*scan);
3085  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3086  next = *scan;
3087  SKIP_DIGITS(next);
3088  count = __kmp_str_to_int(*scan, *next);
3089  KMP_ASSERT(count >= 0);
3090  *scan = next;
3091 
3092  // valid follow sets are ',' ':' and '}'
3093  SKIP_WS(*scan);
3094  if (**scan == '}' || **scan == ',') {
3095  for (i = 0; i < count; i++) {
3096  if ((start > maxOsId) ||
3097  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3098  if (__kmp_affinity_verbose ||
3099  (__kmp_affinity_warnings &&
3100  (__kmp_affinity_type != affinity_none))) {
3101  KMP_WARNING(AffIgnoreInvalidProcID, start);
3102  }
3103  break; // don't proliferate warnings for large count
3104  } else {
3105  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3106  start++;
3107  (*setSize)++;
3108  }
3109  }
3110  if (**scan == '}') {
3111  break;
3112  }
3113  (*scan)++; // skip ','
3114  continue;
3115  }
3116  KMP_ASSERT2(**scan == ':', "bad explicit places list");
3117  (*scan)++; // skip ':'
3118 
3119  // Read stride parameter
3120  int sign = +1;
3121  for (;;) {
3122  SKIP_WS(*scan);
3123  if (**scan == '+') {
3124  (*scan)++; // skip '+'
3125  continue;
3126  }
3127  if (**scan == '-') {
3128  sign *= -1;
3129  (*scan)++; // skip '-'
3130  continue;
3131  }
3132  break;
3133  }
3134  SKIP_WS(*scan);
3135  KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3136  next = *scan;
3137  SKIP_DIGITS(next);
3138  stride = __kmp_str_to_int(*scan, *next);
3139  KMP_ASSERT(stride >= 0);
3140  *scan = next;
3141  stride *= sign;
3142 
3143  // valid follow sets are ',' and '}'
3144  SKIP_WS(*scan);
3145  if (**scan == '}' || **scan == ',') {
3146  for (i = 0; i < count; i++) {
3147  if ((start > maxOsId) ||
3148  (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3149  if (__kmp_affinity_verbose ||
3150  (__kmp_affinity_warnings &&
3151  (__kmp_affinity_type != affinity_none))) {
3152  KMP_WARNING(AffIgnoreInvalidProcID, start);
3153  }
3154  break; // don't proliferate warnings for large count
3155  } else {
3156  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3157  start += stride;
3158  (*setSize)++;
3159  }
3160  }
3161  if (**scan == '}') {
3162  break;
3163  }
3164  (*scan)++; // skip ','
3165  continue;
3166  }
3167 
3168  KMP_ASSERT2(0, "bad explicit places list");
3169  }
3170 }
3171 
3172 static void __kmp_process_place(const char **scan, kmp_affin_mask_t *osId2Mask,
3173  int maxOsId, kmp_affin_mask_t *tempMask,
3174  int *setSize) {
3175  const char *next;
3176 
3177  // valid follow sets are '{' '!' and num
3178  SKIP_WS(*scan);
3179  if (**scan == '{') {
3180  (*scan)++; // skip '{'
3181  __kmp_process_subplace_list(scan, osId2Mask, maxOsId, tempMask, setSize);
3182  KMP_ASSERT2(**scan == '}', "bad explicit places list");
3183  (*scan)++; // skip '}'
3184  } else if (**scan == '!') {
3185  (*scan)++; // skip '!'
3186  __kmp_process_place(scan, osId2Mask, maxOsId, tempMask, setSize);
3187  KMP_CPU_COMPLEMENT(maxOsId, tempMask);
3188  } else if ((**scan >= '0') && (**scan <= '9')) {
3189  next = *scan;
3190  SKIP_DIGITS(next);
3191  int num = __kmp_str_to_int(*scan, *next);
3192  KMP_ASSERT(num >= 0);
3193  if ((num > maxOsId) ||
3194  (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
3195  if (__kmp_affinity_verbose ||
3196  (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) {
3197  KMP_WARNING(AffIgnoreInvalidProcID, num);
3198  }
3199  } else {
3200  KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, num));
3201  (*setSize)++;
3202  }
3203  *scan = next; // skip num
3204  } else {
3205  KMP_ASSERT2(0, "bad explicit places list");
3206  }
3207 }
3208 
3209 // static void
3210 void __kmp_affinity_process_placelist(kmp_affin_mask_t **out_masks,
3211  unsigned int *out_numMasks,
3212  const char *placelist,
3213  kmp_affin_mask_t *osId2Mask,
3214  int maxOsId) {
3215  int i, j, count, stride, sign;
3216  const char *scan = placelist;
3217  const char *next = placelist;
3218 
3219  numNewMasks = 2;
3220  KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
3221  nextNewMask = 0;
3222 
3223  // tempMask is modified based on the previous or initial
3224  // place to form the current place
3225  // previousMask contains the previous place
3226  kmp_affin_mask_t *tempMask;
3227  kmp_affin_mask_t *previousMask;
3228  KMP_CPU_ALLOC(tempMask);
3229  KMP_CPU_ZERO(tempMask);
3230  KMP_CPU_ALLOC(previousMask);
3231  KMP_CPU_ZERO(previousMask);
3232  int setSize = 0;
3233 
3234  for (;;) {
3235  __kmp_process_place(&scan, osId2Mask, maxOsId, tempMask, &setSize);
3236 
3237  // valid follow sets are ',' ':' and EOL
3238  SKIP_WS(scan);
3239  if (*scan == '\0' || *scan == ',') {
3240  if (setSize > 0) {
3241  ADD_MASK(tempMask);
3242  }
3243  KMP_CPU_ZERO(tempMask);
3244  setSize = 0;
3245  if (*scan == '\0') {
3246  break;
3247  }
3248  scan++; // skip ','
3249  continue;
3250  }
3251 
3252  KMP_ASSERT2(*scan == ':', "bad explicit places list");
3253  scan++; // skip ':'
3254 
3255  // Read count parameter
3256  SKIP_WS(scan);
3257  KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
3258  next = scan;
3259  SKIP_DIGITS(next);
3260  count = __kmp_str_to_int(scan, *next);
3261  KMP_ASSERT(count >= 0);
3262  scan = next;
3263 
3264  // valid follow sets are ',' ':' and EOL
3265  SKIP_WS(scan);
3266  if (*scan == '\0' || *scan == ',') {
3267  stride = +1;
3268  } else {
3269  KMP_ASSERT2(*scan == ':', "bad explicit places list");
3270  scan++; // skip ':'
3271 
3272  // Read stride parameter
3273  sign = +1;
3274  for (;;) {
3275  SKIP_WS(scan);
3276  if (*scan == '+') {
3277  scan++; // skip '+'
3278  continue;
3279  }
3280  if (*scan == '-') {
3281  sign *= -1;
3282  scan++; // skip '-'
3283  continue;
3284  }
3285  break;
3286  }
3287  SKIP_WS(scan);
3288  KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
3289  next = scan;
3290  SKIP_DIGITS(next);
3291  stride = __kmp_str_to_int(scan, *next);
3292  KMP_DEBUG_ASSERT(stride >= 0);
3293  scan = next;
3294  stride *= sign;
3295  }
3296 
3297  // Add places determined by initial_place : count : stride
3298  for (i = 0; i < count; i++) {
3299  if (setSize == 0) {
3300  break;
3301  }
3302  // Add the current place, then build the next place (tempMask) from that
3303  KMP_CPU_COPY(previousMask, tempMask);
3304  ADD_MASK(previousMask);
3305  KMP_CPU_ZERO(tempMask);
3306  setSize = 0;
3307  KMP_CPU_SET_ITERATE(j, previousMask) {
3308  if (!KMP_CPU_ISSET(j, previousMask)) {
3309  continue;
3310  }
3311  if ((j + stride > maxOsId) || (j + stride < 0) ||
3312  (!KMP_CPU_ISSET(j, __kmp_affin_fullMask)) ||
3313  (!KMP_CPU_ISSET(j + stride,
3314  KMP_CPU_INDEX(osId2Mask, j + stride)))) {
3315  if ((__kmp_affinity_verbose ||
3316  (__kmp_affinity_warnings &&
3317  (__kmp_affinity_type != affinity_none))) &&
3318  i < count - 1) {
3319  KMP_WARNING(AffIgnoreInvalidProcID, j + stride);
3320  }
3321  continue;
3322  }
3323  KMP_CPU_SET(j + stride, tempMask);
3324  setSize++;
3325  }
3326  }
3327  KMP_CPU_ZERO(tempMask);
3328  setSize = 0;
3329 
3330  // valid follow sets are ',' and EOL
3331  SKIP_WS(scan);
3332  if (*scan == '\0') {
3333  break;
3334  }
3335  if (*scan == ',') {
3336  scan++; // skip ','
3337  continue;
3338  }
3339 
3340  KMP_ASSERT2(0, "bad explicit places list");
3341  }
3342 
3343  *out_numMasks = nextNewMask;
3344  if (nextNewMask == 0) {
3345  *out_masks = NULL;
3346  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3347  return;
3348  }
3349  KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
3350  KMP_CPU_FREE(tempMask);
3351  KMP_CPU_FREE(previousMask);
3352  for (i = 0; i < nextNewMask; i++) {
3353  kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
3354  kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
3355  KMP_CPU_COPY(dest, src);
3356  }
3357  KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3358 }
3359 
3360 #endif /* OMP_40_ENABLED */
3361 
3362 #undef ADD_MASK
3363 #undef ADD_MASK_OSID
3364 
3365 #if KMP_USE_HWLOC
3366 static int __kmp_hwloc_skip_PUs_obj(hwloc_topology_t t, hwloc_obj_t o) {
3367  // skip PUs descendants of the object o
3368  int skipped = 0;
3369  hwloc_obj_t hT = NULL;
3370  int N = __kmp_hwloc_count_children_by_type(t, o, HWLOC_OBJ_PU, &hT);
3371  for (int i = 0; i < N; ++i) {
3372  KMP_DEBUG_ASSERT(hT);
3373  unsigned idx = hT->os_index;
3374  if (KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
3375  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
3376  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
3377  ++skipped;
3378  }
3379  hT = hwloc_get_next_obj_by_type(t, HWLOC_OBJ_PU, hT);
3380  }
3381  return skipped; // count number of skipped units
3382 }
3383 
3384 static int __kmp_hwloc_obj_has_PUs(hwloc_topology_t t, hwloc_obj_t o) {
3385  // check if obj has PUs present in fullMask
3386  hwloc_obj_t hT = NULL;
3387  int N = __kmp_hwloc_count_children_by_type(t, o, HWLOC_OBJ_PU, &hT);
3388  for (int i = 0; i < N; ++i) {
3389  KMP_DEBUG_ASSERT(hT);
3390  unsigned idx = hT->os_index;
3391  if (KMP_CPU_ISSET(idx, __kmp_affin_fullMask))
3392  return 1; // found PU
3393  hT = hwloc_get_next_obj_by_type(t, HWLOC_OBJ_PU, hT);
3394  }
3395  return 0; // no PUs found
3396 }
3397 #endif // KMP_USE_HWLOC
3398 
3399 static void __kmp_apply_thread_places(AddrUnsPair **pAddr, int depth) {
3400  AddrUnsPair *newAddr;
3401  if (__kmp_hws_requested == 0)
3402  goto _exit; // no topology limiting actions requested, exit
3403 #if KMP_USE_HWLOC
3404  if (__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC) {
3405  // Number of subobjects calculated dynamically, this works fine for
3406  // any non-uniform topology.
3407  // L2 cache objects are determined by depth, other objects - by type.
3408  hwloc_topology_t tp = __kmp_hwloc_topology;
3409  int nS = 0, nN = 0, nL = 0, nC = 0,
3410  nT = 0; // logical index including skipped
3411  int nCr = 0, nTr = 0; // number of requested units
3412  int nPkg = 0, nCo = 0, n_new = 0, n_old = 0, nCpP = 0, nTpC = 0; // counters
3413  hwloc_obj_t hT, hC, hL, hN, hS; // hwloc objects (pointers to)
3414  int L2depth, idx;
3415 
3416  // check support of extensions ----------------------------------
3417  int numa_support = 0, tile_support = 0;
3418  if (__kmp_pu_os_idx)
3419  hT = hwloc_get_pu_obj_by_os_index(tp,
3420  __kmp_pu_os_idx[__kmp_avail_proc - 1]);
3421  else
3422  hT = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PU, __kmp_avail_proc - 1);
3423  if (hT == NULL) { // something's gone wrong
3424  KMP_WARNING(AffHWSubsetUnsupported);
3425  goto _exit;
3426  }
3427  // check NUMA node
3428  hN = hwloc_get_ancestor_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hT);
3429  hS = hwloc_get_ancestor_obj_by_type(tp, HWLOC_OBJ_PACKAGE, hT);
3430  if (hN != NULL && hN->depth > hS->depth) {
3431  numa_support = 1; // 1 in case socket includes node(s)
3432  } else if (__kmp_hws_node.num > 0) {
3433  // don't support sockets inside NUMA node (no such HW found for testing)
3434  KMP_WARNING(AffHWSubsetUnsupported);
3435  goto _exit;
3436  }
3437  // check L2 cahce, get object by depth because of multiple caches
3438  L2depth = hwloc_get_cache_type_depth(tp, 2, HWLOC_OBJ_CACHE_UNIFIED);
3439  hL = hwloc_get_ancestor_obj_by_depth(tp, L2depth, hT);
3440  if (hL != NULL &&
3441  __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE, &hC) > 1) {
3442  tile_support = 1; // no sense to count L2 if it includes single core
3443  } else if (__kmp_hws_tile.num > 0) {
3444  if (__kmp_hws_core.num == 0) {
3445  __kmp_hws_core = __kmp_hws_tile; // replace L2 with core
3446  __kmp_hws_tile.num = 0;
3447  } else {
3448  // L2 and core are both requested, but represent same object
3449  KMP_WARNING(AffHWSubsetInvalid);
3450  goto _exit;
3451  }
3452  }
3453  // end of check of extensions -----------------------------------
3454 
3455  // fill in unset items, validate settings -----------------------
3456  if (__kmp_hws_socket.num == 0)
3457  __kmp_hws_socket.num = nPackages; // use all available sockets
3458  if (__kmp_hws_socket.offset >= nPackages) {
3459  KMP_WARNING(AffHWSubsetManySockets);
3460  goto _exit;
3461  }
3462  if (numa_support) {
3463  hN = NULL;
3464  int NN = __kmp_hwloc_count_children_by_type(tp, hS, HWLOC_OBJ_NUMANODE,
3465  &hN); // num nodes in socket
3466  if (__kmp_hws_node.num == 0)
3467  __kmp_hws_node.num = NN; // use all available nodes
3468  if (__kmp_hws_node.offset >= NN) {
3469  KMP_WARNING(AffHWSubsetManyNodes);
3470  goto _exit;
3471  }
3472  if (tile_support) {
3473  // get num tiles in node
3474  int NL = __kmp_hwloc_count_children_by_depth(tp, hN, L2depth, &hL);
3475  if (__kmp_hws_tile.num == 0) {
3476  __kmp_hws_tile.num = NL + 1;
3477  } // use all available tiles, some node may have more tiles, thus +1
3478  if (__kmp_hws_tile.offset >= NL) {
3479  KMP_WARNING(AffHWSubsetManyTiles);
3480  goto _exit;
3481  }
3482  int NC = __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE,
3483  &hC); // num cores in tile
3484  if (__kmp_hws_core.num == 0)
3485  __kmp_hws_core.num = NC; // use all available cores
3486  if (__kmp_hws_core.offset >= NC) {
3487  KMP_WARNING(AffHWSubsetManyCores);
3488  goto _exit;
3489  }
3490  } else { // tile_support
3491  int NC = __kmp_hwloc_count_children_by_type(tp, hN, HWLOC_OBJ_CORE,
3492  &hC); // num cores in node
3493  if (__kmp_hws_core.num == 0)
3494  __kmp_hws_core.num = NC; // use all available cores
3495  if (__kmp_hws_core.offset >= NC) {
3496  KMP_WARNING(AffHWSubsetManyCores);
3497  goto _exit;
3498  }
3499  } // tile_support
3500  } else { // numa_support
3501  if (tile_support) {
3502  // get num tiles in socket
3503  int NL = __kmp_hwloc_count_children_by_depth(tp, hS, L2depth, &hL);
3504  if (__kmp_hws_tile.num == 0)
3505  __kmp_hws_tile.num = NL; // use all available tiles
3506  if (__kmp_hws_tile.offset >= NL) {
3507  KMP_WARNING(AffHWSubsetManyTiles);
3508  goto _exit;
3509  }
3510  int NC = __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE,
3511  &hC); // num cores in tile
3512  if (__kmp_hws_core.num == 0)
3513  __kmp_hws_core.num = NC; // use all available cores
3514  if (__kmp_hws_core.offset >= NC) {
3515  KMP_WARNING(AffHWSubsetManyCores);
3516  goto _exit;
3517  }
3518  } else { // tile_support
3519  int NC = __kmp_hwloc_count_children_by_type(tp, hS, HWLOC_OBJ_CORE,
3520  &hC); // num cores in socket
3521  if (__kmp_hws_core.num == 0)
3522  __kmp_hws_core.num = NC; // use all available cores
3523  if (__kmp_hws_core.offset >= NC) {
3524  KMP_WARNING(AffHWSubsetManyCores);
3525  goto _exit;
3526  }
3527  } // tile_support
3528  }
3529  if (__kmp_hws_proc.num == 0)
3530  __kmp_hws_proc.num = __kmp_nThreadsPerCore; // use all available procs
3531  if (__kmp_hws_proc.offset >= __kmp_nThreadsPerCore) {
3532  KMP_WARNING(AffHWSubsetManyProcs);
3533  goto _exit;
3534  }
3535  // end of validation --------------------------------------------
3536 
3537  if (pAddr) // pAddr is NULL in case of affinity_none
3538  newAddr = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) *
3539  __kmp_avail_proc); // max size
3540  // main loop to form HW subset ----------------------------------
3541  hS = NULL;
3542  int NP = hwloc_get_nbobjs_by_type(tp, HWLOC_OBJ_PACKAGE);
3543  for (int s = 0; s < NP; ++s) {
3544  // Check Socket -----------------------------------------------
3545  hS = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PACKAGE, hS);
3546  if (!__kmp_hwloc_obj_has_PUs(tp, hS))
3547  continue; // skip socket if all PUs are out of fullMask
3548  ++nS; // only count objects those have PUs in affinity mask
3549  if (nS <= __kmp_hws_socket.offset ||
3550  nS > __kmp_hws_socket.num + __kmp_hws_socket.offset) {
3551  n_old += __kmp_hwloc_skip_PUs_obj(tp, hS); // skip socket
3552  continue; // move to next socket
3553  }
3554  nCr = 0; // count number of cores per socket
3555  // socket requested, go down the topology tree
3556  // check 4 cases: (+NUMA+Tile), (+NUMA-Tile), (-NUMA+Tile), (-NUMA-Tile)
3557  if (numa_support) {
3558  nN = 0;
3559  hN = NULL;
3560  // num nodes in current socket
3561  int NN =
3562  __kmp_hwloc_count_children_by_type(tp, hS, HWLOC_OBJ_NUMANODE, &hN);
3563  for (int n = 0; n < NN; ++n) {
3564  // Check NUMA Node ----------------------------------------
3565  if (!__kmp_hwloc_obj_has_PUs(tp, hN)) {
3566  hN = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hN);
3567  continue; // skip node if all PUs are out of fullMask
3568  }
3569  ++nN;
3570  if (nN <= __kmp_hws_node.offset ||
3571  nN > __kmp_hws_node.num + __kmp_hws_node.offset) {
3572  // skip node as not requested
3573  n_old += __kmp_hwloc_skip_PUs_obj(tp, hN); // skip node
3574  hN = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hN);
3575  continue; // move to next node
3576  }
3577  // node requested, go down the topology tree
3578  if (tile_support) {
3579  nL = 0;
3580  hL = NULL;
3581  int NL = __kmp_hwloc_count_children_by_depth(tp, hN, L2depth, &hL);
3582  for (int l = 0; l < NL; ++l) {
3583  // Check L2 (tile) ------------------------------------
3584  if (!__kmp_hwloc_obj_has_PUs(tp, hL)) {
3585  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3586  continue; // skip tile if all PUs are out of fullMask
3587  }
3588  ++nL;
3589  if (nL <= __kmp_hws_tile.offset ||
3590  nL > __kmp_hws_tile.num + __kmp_hws_tile.offset) {
3591  // skip tile as not requested
3592  n_old += __kmp_hwloc_skip_PUs_obj(tp, hL); // skip tile
3593  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3594  continue; // move to next tile
3595  }
3596  // tile requested, go down the topology tree
3597  nC = 0;
3598  hC = NULL;
3599  // num cores in current tile
3600  int NC = __kmp_hwloc_count_children_by_type(tp, hL,
3601  HWLOC_OBJ_CORE, &hC);
3602  for (int c = 0; c < NC; ++c) {
3603  // Check Core ---------------------------------------
3604  if (!__kmp_hwloc_obj_has_PUs(tp, hC)) {
3605  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3606  continue; // skip core if all PUs are out of fullMask
3607  }
3608  ++nC;
3609  if (nC <= __kmp_hws_core.offset ||
3610  nC > __kmp_hws_core.num + __kmp_hws_core.offset) {
3611  // skip node as not requested
3612  n_old += __kmp_hwloc_skip_PUs_obj(tp, hC); // skip core
3613  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3614  continue; // move to next node
3615  }
3616  // core requested, go down to PUs
3617  nT = 0;
3618  nTr = 0;
3619  hT = NULL;
3620  // num procs in current core
3621  int NT = __kmp_hwloc_count_children_by_type(tp, hC,
3622  HWLOC_OBJ_PU, &hT);
3623  for (int t = 0; t < NT; ++t) {
3624  // Check PU ---------------------------------------
3625  idx = hT->os_index;
3626  if (!KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
3627  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3628  continue; // skip PU if not in fullMask
3629  }
3630  ++nT;
3631  if (nT <= __kmp_hws_proc.offset ||
3632  nT > __kmp_hws_proc.num + __kmp_hws_proc.offset) {
3633  // skip PU
3634  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
3635  ++n_old;
3636  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
3637  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3638  continue; // move to next node
3639  }
3640  ++nTr;
3641  if (pAddr) // collect requested thread's data
3642  newAddr[n_new] = (*pAddr)[n_old];
3643  ++n_new;
3644  ++n_old;
3645  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3646  } // threads loop
3647  if (nTr > 0) {
3648  ++nCr; // num cores per socket
3649  ++nCo; // total num cores
3650  if (nTr > nTpC)
3651  nTpC = nTr; // calc max threads per core
3652  }
3653  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3654  } // cores loop
3655  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3656  } // tiles loop
3657  } else { // tile_support
3658  // no tiles, check cores
3659  nC = 0;
3660  hC = NULL;
3661  // num cores in current node
3662  int NC =
3663  __kmp_hwloc_count_children_by_type(tp, hN, HWLOC_OBJ_CORE, &hC);
3664  for (int c = 0; c < NC; ++c) {
3665  // Check Core ---------------------------------------
3666  if (!__kmp_hwloc_obj_has_PUs(tp, hC)) {
3667  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3668  continue; // skip core if all PUs are out of fullMask
3669  }
3670  ++nC;
3671  if (nC <= __kmp_hws_core.offset ||
3672  nC > __kmp_hws_core.num + __kmp_hws_core.offset) {
3673  // skip node as not requested
3674  n_old += __kmp_hwloc_skip_PUs_obj(tp, hC); // skip core
3675  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3676  continue; // move to next node
3677  }
3678  // core requested, go down to PUs
3679  nT = 0;
3680  nTr = 0;
3681  hT = NULL;
3682  int NT =
3683  __kmp_hwloc_count_children_by_type(tp, hC, HWLOC_OBJ_PU, &hT);
3684  for (int t = 0; t < NT; ++t) {
3685  // Check PU ---------------------------------------
3686  idx = hT->os_index;
3687  if (!KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
3688  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3689  continue; // skip PU if not in fullMask
3690  }
3691  ++nT;
3692  if (nT <= __kmp_hws_proc.offset ||
3693  nT > __kmp_hws_proc.num + __kmp_hws_proc.offset) {
3694  // skip PU
3695  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
3696  ++n_old;
3697  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
3698  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3699  continue; // move to next node
3700  }
3701  ++nTr;
3702  if (pAddr) // collect requested thread's data
3703  newAddr[n_new] = (*pAddr)[n_old];
3704  ++n_new;
3705  ++n_old;
3706  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3707  } // threads loop
3708  if (nTr > 0) {
3709  ++nCr; // num cores per socket
3710  ++nCo; // total num cores
3711  if (nTr > nTpC)
3712  nTpC = nTr; // calc max threads per core
3713  }
3714  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3715  } // cores loop
3716  } // tiles support
3717  hN = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hN);
3718  } // nodes loop
3719  } else { // numa_support
3720  // no NUMA support
3721  if (tile_support) {
3722  nL = 0;
3723  hL = NULL;
3724  // num tiles in current socket
3725  int NL = __kmp_hwloc_count_children_by_depth(tp, hS, L2depth, &hL);
3726  for (int l = 0; l < NL; ++l) {
3727  // Check L2 (tile) ------------------------------------
3728  if (!__kmp_hwloc_obj_has_PUs(tp, hL)) {
3729  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3730  continue; // skip tile if all PUs are out of fullMask
3731  }
3732  ++nL;
3733  if (nL <= __kmp_hws_tile.offset ||
3734  nL > __kmp_hws_tile.num + __kmp_hws_tile.offset) {
3735  // skip tile as not requested
3736  n_old += __kmp_hwloc_skip_PUs_obj(tp, hL); // skip tile
3737  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3738  continue; // move to next tile
3739  }
3740  // tile requested, go down the topology tree
3741  nC = 0;
3742  hC = NULL;
3743  // num cores per tile
3744  int NC =
3745  __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE, &hC);
3746  for (int c = 0; c < NC; ++c) {
3747  // Check Core ---------------------------------------
3748  if (!__kmp_hwloc_obj_has_PUs(tp, hC)) {
3749  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3750  continue; // skip core if all PUs are out of fullMask
3751  }
3752  ++nC;
3753  if (nC <= __kmp_hws_core.offset ||
3754  nC > __kmp_hws_core.num + __kmp_hws_core.offset) {
3755  // skip node as not requested
3756  n_old += __kmp_hwloc_skip_PUs_obj(tp, hC); // skip core
3757  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3758  continue; // move to next node
3759  }
3760  // core requested, go down to PUs
3761  nT = 0;
3762  nTr = 0;
3763  hT = NULL;
3764  // num procs per core
3765  int NT =
3766  __kmp_hwloc_count_children_by_type(tp, hC, HWLOC_OBJ_PU, &hT);
3767  for (int t = 0; t < NT; ++t) {
3768  // Check PU ---------------------------------------
3769  idx = hT->os_index;
3770  if (!KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
3771  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3772  continue; // skip PU if not in fullMask
3773  }
3774  ++nT;
3775  if (nT <= __kmp_hws_proc.offset ||
3776  nT > __kmp_hws_proc.num + __kmp_hws_proc.offset) {
3777  // skip PU
3778  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
3779  ++n_old;
3780  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
3781  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3782  continue; // move to next node
3783  }
3784  ++nTr;
3785  if (pAddr) // collect requested thread's data
3786  newAddr[n_new] = (*pAddr)[n_old];
3787  ++n_new;
3788  ++n_old;
3789  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3790  } // threads loop
3791  if (nTr > 0) {
3792  ++nCr; // num cores per socket
3793  ++nCo; // total num cores
3794  if (nTr > nTpC)
3795  nTpC = nTr; // calc max threads per core
3796  }
3797  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3798  } // cores loop
3799  hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
3800  } // tiles loop
3801  } else { // tile_support
3802  // no tiles, check cores
3803  nC = 0;
3804  hC = NULL;
3805  // num cores in socket
3806  int NC =
3807  __kmp_hwloc_count_children_by_type(tp, hS, HWLOC_OBJ_CORE, &hC);
3808  for (int c = 0; c < NC; ++c) {
3809  // Check Core -------------------------------------------
3810  if (!__kmp_hwloc_obj_has_PUs(tp, hC)) {
3811  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3812  continue; // skip core if all PUs are out of fullMask
3813  }
3814  ++nC;
3815  if (nC <= __kmp_hws_core.offset ||
3816  nC > __kmp_hws_core.num + __kmp_hws_core.offset) {
3817  // skip node as not requested
3818  n_old += __kmp_hwloc_skip_PUs_obj(tp, hC); // skip core
3819  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3820  continue; // move to next node
3821  }
3822  // core requested, go down to PUs
3823  nT = 0;
3824  nTr = 0;
3825  hT = NULL;
3826  // num procs per core
3827  int NT =
3828  __kmp_hwloc_count_children_by_type(tp, hC, HWLOC_OBJ_PU, &hT);
3829  for (int t = 0; t < NT; ++t) {
3830  // Check PU ---------------------------------------
3831  idx = hT->os_index;
3832  if (!KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
3833  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3834  continue; // skip PU if not in fullMask
3835  }
3836  ++nT;
3837  if (nT <= __kmp_hws_proc.offset ||
3838  nT > __kmp_hws_proc.num + __kmp_hws_proc.offset) {
3839  // skip PU
3840  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
3841  ++n_old;
3842  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
3843  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3844  continue; // move to next node
3845  }
3846  ++nTr;
3847  if (pAddr) // collect requested thread's data
3848  newAddr[n_new] = (*pAddr)[n_old];
3849  ++n_new;
3850  ++n_old;
3851  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
3852  } // threads loop
3853  if (nTr > 0) {
3854  ++nCr; // num cores per socket
3855  ++nCo; // total num cores
3856  if (nTr > nTpC)
3857  nTpC = nTr; // calc max threads per core
3858  }
3859  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
3860  } // cores loop
3861  } // tiles support
3862  } // numa_support
3863  if (nCr > 0) { // found cores?
3864  ++nPkg; // num sockets
3865  if (nCr > nCpP)
3866  nCpP = nCr; // calc max cores per socket
3867  }
3868  } // sockets loop
3869 
3870  // check the subset is valid
3871  KMP_DEBUG_ASSERT(n_old == __kmp_avail_proc);
3872  KMP_DEBUG_ASSERT(nPkg > 0);
3873  KMP_DEBUG_ASSERT(nCpP > 0);
3874  KMP_DEBUG_ASSERT(nTpC > 0);
3875  KMP_DEBUG_ASSERT(nCo > 0);
3876  KMP_DEBUG_ASSERT(nPkg <= nPackages);
3877  KMP_DEBUG_ASSERT(nCpP <= nCoresPerPkg);
3878  KMP_DEBUG_ASSERT(nTpC <= __kmp_nThreadsPerCore);
3879  KMP_DEBUG_ASSERT(nCo <= __kmp_ncores);
3880 
3881  nPackages = nPkg; // correct num sockets
3882  nCoresPerPkg = nCpP; // correct num cores per socket
3883  __kmp_nThreadsPerCore = nTpC; // correct num threads per core
3884  __kmp_avail_proc = n_new; // correct num procs
3885  __kmp_ncores = nCo; // correct num cores
3886  // hwloc topology method end
3887  } else
3888 #endif // KMP_USE_HWLOC
3889  {
3890  int n_old = 0, n_new = 0, proc_num = 0;
3891  if (__kmp_hws_node.num > 0 || __kmp_hws_tile.num > 0) {
3892  KMP_WARNING(AffHWSubsetNoHWLOC);
3893  goto _exit;
3894  }
3895  if (__kmp_hws_socket.num == 0)
3896  __kmp_hws_socket.num = nPackages; // use all available sockets
3897  if (__kmp_hws_core.num == 0)
3898  __kmp_hws_core.num = nCoresPerPkg; // use all available cores
3899  if (__kmp_hws_proc.num == 0 || __kmp_hws_proc.num > __kmp_nThreadsPerCore)
3900  __kmp_hws_proc.num = __kmp_nThreadsPerCore; // use all HW contexts
3901  if (!__kmp_affinity_uniform_topology()) {
3902  KMP_WARNING(AffHWSubsetNonUniform);
3903  goto _exit; // don't support non-uniform topology
3904  }
3905  if (depth > 3) {
3906  KMP_WARNING(AffHWSubsetNonThreeLevel);
3907  goto _exit; // don't support not-3-level topology
3908  }
3909  if (__kmp_hws_socket.offset + __kmp_hws_socket.num > nPackages) {
3910  KMP_WARNING(AffHWSubsetManySockets);
3911  goto _exit;
3912  }
3913  if (__kmp_hws_core.offset + __kmp_hws_core.num > nCoresPerPkg) {
3914  KMP_WARNING(AffHWSubsetManyCores);
3915  goto _exit;
3916  }
3917  // Form the requested subset
3918  if (pAddr) // pAddr is NULL in case of affinity_none
3919  newAddr = (AddrUnsPair *)__kmp_allocate(
3920  sizeof(AddrUnsPair) * __kmp_hws_socket.num * __kmp_hws_core.num *
3921  __kmp_hws_proc.num);
3922  for (int i = 0; i < nPackages; ++i) {
3923  if (i < __kmp_hws_socket.offset ||
3924  i >= __kmp_hws_socket.offset + __kmp_hws_socket.num) {
3925  // skip not-requested socket
3926  n_old += nCoresPerPkg * __kmp_nThreadsPerCore;
3927  if (__kmp_pu_os_idx != NULL) {
3928  // walk through skipped socket
3929  for (int j = 0; j < nCoresPerPkg; ++j) {
3930  for (int k = 0; k < __kmp_nThreadsPerCore; ++k) {
3931  KMP_CPU_CLR(__kmp_pu_os_idx[proc_num], __kmp_affin_fullMask);
3932  ++proc_num;
3933  }
3934  }
3935  }
3936  } else {
3937  // walk through requested socket
3938  for (int j = 0; j < nCoresPerPkg; ++j) {
3939  if (j < __kmp_hws_core.offset ||
3940  j >= __kmp_hws_core.offset +
3941  __kmp_hws_core.num) { // skip not-requested core
3942  n_old += __kmp_nThreadsPerCore;
3943  if (__kmp_pu_os_idx != NULL) {
3944  for (int k = 0; k < __kmp_nThreadsPerCore; ++k) {
3945  KMP_CPU_CLR(__kmp_pu_os_idx[proc_num], __kmp_affin_fullMask);
3946  ++proc_num;
3947  }
3948  }
3949  } else {
3950  // walk through requested core
3951  for (int k = 0; k < __kmp_nThreadsPerCore; ++k) {
3952  if (k < __kmp_hws_proc.num) {
3953  if (pAddr) // collect requested thread's data
3954  newAddr[n_new] = (*pAddr)[n_old];
3955  n_new++;
3956  } else {
3957  if (__kmp_pu_os_idx != NULL)
3958  KMP_CPU_CLR(__kmp_pu_os_idx[proc_num], __kmp_affin_fullMask);
3959  }
3960  n_old++;
3961  ++proc_num;
3962  }
3963  }
3964  }
3965  }
3966  }
3967  KMP_DEBUG_ASSERT(n_old == nPackages * nCoresPerPkg * __kmp_nThreadsPerCore);
3968  KMP_DEBUG_ASSERT(n_new ==
3969  __kmp_hws_socket.num * __kmp_hws_core.num *
3970  __kmp_hws_proc.num);
3971  nPackages = __kmp_hws_socket.num; // correct nPackages
3972  nCoresPerPkg = __kmp_hws_core.num; // correct nCoresPerPkg
3973  __kmp_nThreadsPerCore = __kmp_hws_proc.num; // correct __kmp_nThreadsPerCore
3974  __kmp_avail_proc = n_new; // correct avail_proc
3975  __kmp_ncores = nPackages * __kmp_hws_core.num; // correct ncores
3976  } // non-hwloc topology method
3977  if (pAddr) {
3978  __kmp_free(*pAddr);
3979  *pAddr = newAddr; // replace old topology with new one
3980  }
3981  if (__kmp_affinity_verbose) {
3982  char m[KMP_AFFIN_MASK_PRINT_LEN];
3983  __kmp_affinity_print_mask(m, KMP_AFFIN_MASK_PRINT_LEN,
3984  __kmp_affin_fullMask);
3985  if (__kmp_affinity_respect_mask) {
3986  KMP_INFORM(InitOSProcSetRespect, "KMP_HW_SUBSET", m);
3987  } else {
3988  KMP_INFORM(InitOSProcSetNotRespect, "KMP_HW_SUBSET", m);
3989  }
3990  KMP_INFORM(AvailableOSProc, "KMP_HW_SUBSET", __kmp_avail_proc);
3991  kmp_str_buf_t buf;
3992  __kmp_str_buf_init(&buf);
3993  __kmp_str_buf_print(&buf, "%d", nPackages);
3994  KMP_INFORM(TopologyExtra, "KMP_HW_SUBSET", buf.str, nCoresPerPkg,
3995  __kmp_nThreadsPerCore, __kmp_ncores);
3996  __kmp_str_buf_free(&buf);
3997  }
3998 _exit:
3999  if (__kmp_pu_os_idx != NULL) {
4000  __kmp_free(__kmp_pu_os_idx);
4001  __kmp_pu_os_idx = NULL;
4002  }
4003 }
4004 
4005 // This function figures out the deepest level at which there is at least one
4006 // cluster/core with more than one processing unit bound to it.
4007 static int __kmp_affinity_find_core_level(const AddrUnsPair *address2os,
4008  int nprocs, int bottom_level) {
4009  int core_level = 0;
4010 
4011  for (int i = 0; i < nprocs; i++) {
4012  for (int j = bottom_level; j > 0; j--) {
4013  if (address2os[i].first.labels[j] > 0) {
4014  if (core_level < (j - 1)) {
4015  core_level = j - 1;
4016  }
4017  }
4018  }
4019  }
4020  return core_level;
4021 }
4022 
4023 // This function counts number of clusters/cores at given level.
4024 static int __kmp_affinity_compute_ncores(const AddrUnsPair *address2os,
4025  int nprocs, int bottom_level,
4026  int core_level) {
4027  int ncores = 0;
4028  int i, j;
4029 
4030  j = bottom_level;
4031  for (i = 0; i < nprocs; i++) {
4032  for (j = bottom_level; j > core_level; j--) {
4033  if ((i + 1) < nprocs) {
4034  if (address2os[i + 1].first.labels[j] > 0) {
4035  break;
4036  }
4037  }
4038  }
4039  if (j == core_level) {
4040  ncores++;
4041  }
4042  }
4043  if (j > core_level) {
4044  // In case of ( nprocs < __kmp_avail_proc ) we may end too deep and miss one
4045  // core. May occur when called from __kmp_affinity_find_core().
4046  ncores++;
4047  }
4048  return ncores;
4049 }
4050 
4051 // This function finds to which cluster/core given processing unit is bound.
4052 static int __kmp_affinity_find_core(const AddrUnsPair *address2os, int proc,
4053  int bottom_level, int core_level) {
4054  return __kmp_affinity_compute_ncores(address2os, proc + 1, bottom_level,
4055  core_level) -
4056  1;
4057 }
4058 
4059 // This function finds maximal number of processing units bound to a
4060 // cluster/core at given level.
4061 static int __kmp_affinity_max_proc_per_core(const AddrUnsPair *address2os,
4062  int nprocs, int bottom_level,
4063  int core_level) {
4064  int maxprocpercore = 0;
4065 
4066  if (core_level < bottom_level) {
4067  for (int i = 0; i < nprocs; i++) {
4068  int percore = address2os[i].first.labels[core_level + 1] + 1;
4069 
4070  if (percore > maxprocpercore) {
4071  maxprocpercore = percore;
4072  }
4073  }
4074  } else {
4075  maxprocpercore = 1;
4076  }
4077  return maxprocpercore;
4078 }
4079 
4080 static AddrUnsPair *address2os = NULL;
4081 static int *procarr = NULL;
4082 static int __kmp_aff_depth = 0;
4083 
4084 #if KMP_USE_HIER_SCHED
4085 #define KMP_EXIT_AFF_NONE \
4086  KMP_ASSERT(__kmp_affinity_type == affinity_none); \
4087  KMP_ASSERT(address2os == NULL); \
4088  __kmp_apply_thread_places(NULL, 0); \
4089  __kmp_create_affinity_none_places(); \
4090  __kmp_dispatch_set_hierarchy_values(); \
4091  return;
4092 #else
4093 #define KMP_EXIT_AFF_NONE \
4094  KMP_ASSERT(__kmp_affinity_type == affinity_none); \
4095  KMP_ASSERT(address2os == NULL); \
4096  __kmp_apply_thread_places(NULL, 0); \
4097  __kmp_create_affinity_none_places(); \
4098  return;
4099 #endif
4100 
4101 // Create a one element mask array (set of places) which only contains the
4102 // initial process's affinity mask
4103 static void __kmp_create_affinity_none_places() {
4104  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4105  KMP_ASSERT(__kmp_affinity_type == affinity_none);
4106  __kmp_affinity_num_masks = 1;
4107  KMP_CPU_ALLOC_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks);
4108  kmp_affin_mask_t *dest = KMP_CPU_INDEX(__kmp_affinity_masks, 0);
4109  KMP_CPU_COPY(dest, __kmp_affin_fullMask);
4110 }
4111 
4112 static int __kmp_affinity_cmp_Address_child_num(const void *a, const void *b) {
4113  const Address *aa = &(((const AddrUnsPair *)a)->first);
4114  const Address *bb = &(((const AddrUnsPair *)b)->first);
4115  unsigned depth = aa->depth;
4116  unsigned i;
4117  KMP_DEBUG_ASSERT(depth == bb->depth);
4118  KMP_DEBUG_ASSERT((unsigned)__kmp_affinity_compact <= depth);
4119  KMP_DEBUG_ASSERT(__kmp_affinity_compact >= 0);
4120  for (i = 0; i < (unsigned)__kmp_affinity_compact; i++) {
4121  int j = depth - i - 1;
4122  if (aa->childNums[j] < bb->childNums[j])
4123  return -1;
4124  if (aa->childNums[j] > bb->childNums[j])
4125  return 1;
4126  }
4127  for (; i < depth; i++) {
4128  int j = i - __kmp_affinity_compact;
4129  if (aa->childNums[j] < bb->childNums[j])
4130  return -1;
4131  if (aa->childNums[j] > bb->childNums[j])
4132  return 1;
4133  }
4134  return 0;
4135 }
4136 
4137 static void __kmp_aux_affinity_initialize(void) {
4138  if (__kmp_affinity_masks != NULL) {
4139  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4140  return;
4141  }
4142 
4143  // Create the "full" mask - this defines all of the processors that we
4144  // consider to be in the machine model. If respect is set, then it is the
4145  // initialization thread's affinity mask. Otherwise, it is all processors that
4146  // we know about on the machine.
4147  if (__kmp_affin_fullMask == NULL) {
4148  KMP_CPU_ALLOC(__kmp_affin_fullMask);
4149  }
4150  if (KMP_AFFINITY_CAPABLE()) {
4151  if (__kmp_affinity_respect_mask) {
4152  __kmp_get_system_affinity(__kmp_affin_fullMask, TRUE);
4153 
4154  // Count the number of available processors.
4155  unsigned i;
4156  __kmp_avail_proc = 0;
4157  KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
4158  if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
4159  continue;
4160  }
4161  __kmp_avail_proc++;
4162  }
4163  if (__kmp_avail_proc > __kmp_xproc) {
4164  if (__kmp_affinity_verbose ||
4165  (__kmp_affinity_warnings &&
4166  (__kmp_affinity_type != affinity_none))) {
4167  KMP_WARNING(ErrorInitializeAffinity);
4168  }
4169  __kmp_affinity_type = affinity_none;
4170  KMP_AFFINITY_DISABLE();
4171  return;
4172  }
4173  } else {
4174  __kmp_affinity_entire_machine_mask(__kmp_affin_fullMask);
4175  __kmp_avail_proc = __kmp_xproc;
4176  }
4177  }
4178 
4179  if (__kmp_affinity_gran == affinity_gran_tile &&
4180  // check if user's request is valid
4181  __kmp_affinity_dispatch->get_api_type() == KMPAffinity::NATIVE_OS) {
4182  KMP_WARNING(AffTilesNoHWLOC, "KMP_AFFINITY");
4183  __kmp_affinity_gran = affinity_gran_package;
4184  }
4185 
4186  int depth = -1;
4187  kmp_i18n_id_t msg_id = kmp_i18n_null;
4188 
4189  // For backward compatibility, setting KMP_CPUINFO_FILE =>
4190  // KMP_TOPOLOGY_METHOD=cpuinfo
4191  if ((__kmp_cpuinfo_file != NULL) &&
4192  (__kmp_affinity_top_method == affinity_top_method_all)) {
4193  __kmp_affinity_top_method = affinity_top_method_cpuinfo;
4194  }
4195 
4196  if (__kmp_affinity_top_method == affinity_top_method_all) {
4197  // In the default code path, errors are not fatal - we just try using
4198  // another method. We only emit a warning message if affinity is on, or the
4199  // verbose flag is set, an the nowarnings flag was not set.
4200  const char *file_name = NULL;
4201  int line = 0;
4202 #if KMP_USE_HWLOC
4203  if (depth < 0 &&
4204  __kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC) {
4205  if (__kmp_affinity_verbose) {
4206  KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
4207  }
4208  if (!__kmp_hwloc_error) {
4209  depth = __kmp_affinity_create_hwloc_map(&address2os, &msg_id);
4210  if (depth == 0) {
4211  KMP_EXIT_AFF_NONE;
4212  } else if (depth < 0 && __kmp_affinity_verbose) {
4213  KMP_INFORM(AffIgnoringHwloc, "KMP_AFFINITY");
4214  }
4215  } else if (__kmp_affinity_verbose) {
4216  KMP_INFORM(AffIgnoringHwloc, "KMP_AFFINITY");
4217  }
4218  }
4219 #endif
4220 
4221 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4222 
4223  if (depth < 0) {
4224  if (__kmp_affinity_verbose) {
4225  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(Decodingx2APIC));
4226  }
4227 
4228  file_name = NULL;
4229  depth = __kmp_affinity_create_x2apicid_map(&address2os, &msg_id);
4230  if (depth == 0) {
4231  KMP_EXIT_AFF_NONE;
4232  }
4233 
4234  if (depth < 0) {
4235  if (__kmp_affinity_verbose) {
4236  if (msg_id != kmp_i18n_null) {
4237  KMP_INFORM(AffInfoStrStr, "KMP_AFFINITY",
4238  __kmp_i18n_catgets(msg_id),
4239  KMP_I18N_STR(DecodingLegacyAPIC));
4240  } else {
4241  KMP_INFORM(AffInfoStr, "KMP_AFFINITY",
4242  KMP_I18N_STR(DecodingLegacyAPIC));
4243  }
4244  }
4245 
4246  file_name = NULL;
4247  depth = __kmp_affinity_create_apicid_map(&address2os, &msg_id);
4248  if (depth == 0) {
4249  KMP_EXIT_AFF_NONE;
4250  }
4251  }
4252  }
4253 
4254 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4255 
4256 #if KMP_OS_LINUX
4257 
4258  if (depth < 0) {
4259  if (__kmp_affinity_verbose) {
4260  if (msg_id != kmp_i18n_null) {
4261  KMP_INFORM(AffStrParseFilename, "KMP_AFFINITY",
4262  __kmp_i18n_catgets(msg_id), "/proc/cpuinfo");
4263  } else {
4264  KMP_INFORM(AffParseFilename, "KMP_AFFINITY", "/proc/cpuinfo");
4265  }
4266  }
4267 
4268  FILE *f = fopen("/proc/cpuinfo", "r");
4269  if (f == NULL) {
4270  msg_id = kmp_i18n_str_CantOpenCpuinfo;
4271  } else {
4272  file_name = "/proc/cpuinfo";
4273  depth =
4274  __kmp_affinity_create_cpuinfo_map(&address2os, &line, &msg_id, f);
4275  fclose(f);
4276  if (depth == 0) {
4277  KMP_EXIT_AFF_NONE;
4278  }
4279  }
4280  }
4281 
4282 #endif /* KMP_OS_LINUX */
4283 
4284 #if KMP_GROUP_AFFINITY
4285 
4286  if ((depth < 0) && (__kmp_num_proc_groups > 1)) {
4287  if (__kmp_affinity_verbose) {
4288  KMP_INFORM(AffWindowsProcGroupMap, "KMP_AFFINITY");
4289  }
4290 
4291  depth = __kmp_affinity_create_proc_group_map(&address2os, &msg_id);
4292  KMP_ASSERT(depth != 0);
4293  }
4294 
4295 #endif /* KMP_GROUP_AFFINITY */
4296 
4297  if (depth < 0) {
4298  if (__kmp_affinity_verbose && (msg_id != kmp_i18n_null)) {
4299  if (file_name == NULL) {
4300  KMP_INFORM(UsingFlatOS, __kmp_i18n_catgets(msg_id));
4301  } else if (line == 0) {
4302  KMP_INFORM(UsingFlatOSFile, file_name, __kmp_i18n_catgets(msg_id));
4303  } else {
4304  KMP_INFORM(UsingFlatOSFileLine, file_name, line,
4305  __kmp_i18n_catgets(msg_id));
4306  }
4307  }
4308  // FIXME - print msg if msg_id = kmp_i18n_null ???
4309 
4310  file_name = "";
4311  depth = __kmp_affinity_create_flat_map(&address2os, &msg_id);
4312  if (depth == 0) {
4313  KMP_EXIT_AFF_NONE;
4314  }
4315  KMP_ASSERT(depth > 0);
4316  KMP_ASSERT(address2os != NULL);
4317  }
4318  }
4319 
4320 #if KMP_USE_HWLOC
4321  else if (__kmp_affinity_top_method == affinity_top_method_hwloc) {
4322  KMP_ASSERT(__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC);
4323  if (__kmp_affinity_verbose) {
4324  KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
4325  }
4326  depth = __kmp_affinity_create_hwloc_map(&address2os, &msg_id);
4327  if (depth == 0) {
4328  KMP_EXIT_AFF_NONE;
4329  }
4330  }
4331 #endif // KMP_USE_HWLOC
4332 
4333 // If the user has specified that a paricular topology discovery method is to be
4334 // used, then we abort if that method fails. The exception is group affinity,
4335 // which might have been implicitly set.
4336 
4337 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4338 
4339  else if (__kmp_affinity_top_method == affinity_top_method_x2apicid) {
4340  if (__kmp_affinity_verbose) {
4341  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(Decodingx2APIC));
4342  }
4343 
4344  depth = __kmp_affinity_create_x2apicid_map(&address2os, &msg_id);
4345  if (depth == 0) {
4346  KMP_EXIT_AFF_NONE;
4347  }
4348  if (depth < 0) {
4349  KMP_ASSERT(msg_id != kmp_i18n_null);
4350  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4351  }
4352  } else if (__kmp_affinity_top_method == affinity_top_method_apicid) {
4353  if (__kmp_affinity_verbose) {
4354  KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(DecodingLegacyAPIC));
4355  }
4356 
4357  depth = __kmp_affinity_create_apicid_map(&address2os, &msg_id);
4358  if (depth == 0) {
4359  KMP_EXIT_AFF_NONE;
4360  }
4361  if (depth < 0) {
4362  KMP_ASSERT(msg_id != kmp_i18n_null);
4363  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4364  }
4365  }
4366 
4367 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4368 
4369  else if (__kmp_affinity_top_method == affinity_top_method_cpuinfo) {
4370  const char *filename;
4371  if (__kmp_cpuinfo_file != NULL) {
4372  filename = __kmp_cpuinfo_file;
4373  } else {
4374  filename = "/proc/cpuinfo";
4375  }
4376 
4377  if (__kmp_affinity_verbose) {
4378  KMP_INFORM(AffParseFilename, "KMP_AFFINITY", filename);
4379  }
4380 
4381  FILE *f = fopen(filename, "r");
4382  if (f == NULL) {
4383  int code = errno;
4384  if (__kmp_cpuinfo_file != NULL) {
4385  __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4386  KMP_HNT(NameComesFrom_CPUINFO_FILE), __kmp_msg_null);
4387  } else {
4388  __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4389  __kmp_msg_null);
4390  }
4391  }
4392  int line = 0;
4393  depth = __kmp_affinity_create_cpuinfo_map(&address2os, &line, &msg_id, f);
4394  fclose(f);
4395  if (depth < 0) {
4396  KMP_ASSERT(msg_id != kmp_i18n_null);
4397  if (line > 0) {
4398  KMP_FATAL(FileLineMsgExiting, filename, line,
4399  __kmp_i18n_catgets(msg_id));
4400  } else {
4401  KMP_FATAL(FileMsgExiting, filename, __kmp_i18n_catgets(msg_id));
4402  }
4403  }
4404  if (__kmp_affinity_type == affinity_none) {
4405  KMP_ASSERT(depth == 0);
4406  KMP_EXIT_AFF_NONE;
4407  }
4408  }
4409 
4410 #if KMP_GROUP_AFFINITY
4411 
4412  else if (__kmp_affinity_top_method == affinity_top_method_group) {
4413  if (__kmp_affinity_verbose) {
4414  KMP_INFORM(AffWindowsProcGroupMap, "KMP_AFFINITY");
4415  }
4416 
4417  depth = __kmp_affinity_create_proc_group_map(&address2os, &msg_id);
4418  KMP_ASSERT(depth != 0);
4419  if (depth < 0) {
4420  KMP_ASSERT(msg_id != kmp_i18n_null);
4421  KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4422  }
4423  }
4424 
4425 #endif /* KMP_GROUP_AFFINITY */
4426 
4427  else if (__kmp_affinity_top_method == affinity_top_method_flat) {
4428  if (__kmp_affinity_verbose) {
4429  KMP_INFORM(AffUsingFlatOS, "KMP_AFFINITY");
4430  }
4431 
4432  depth = __kmp_affinity_create_flat_map(&address2os, &msg_id);
4433  if (depth == 0) {
4434  KMP_EXIT_AFF_NONE;
4435  }
4436  // should not fail
4437  KMP_ASSERT(depth > 0);
4438  KMP_ASSERT(address2os != NULL);
4439  }
4440 
4441 #if KMP_USE_HIER_SCHED
4442  __kmp_dispatch_set_hierarchy_values();
4443 #endif
4444 
4445  if (address2os == NULL) {
4446  if (KMP_AFFINITY_CAPABLE() &&
4447  (__kmp_affinity_verbose ||
4448  (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none)))) {
4449  KMP_WARNING(ErrorInitializeAffinity);
4450  }
4451  __kmp_affinity_type = affinity_none;
4452  __kmp_create_affinity_none_places();
4453  KMP_AFFINITY_DISABLE();
4454  return;
4455  }
4456 
4457  if (__kmp_affinity_gran == affinity_gran_tile
4458 #if KMP_USE_HWLOC
4459  && __kmp_tile_depth == 0
4460 #endif
4461  ) {
4462  // tiles requested but not detected, warn user on this
4463  KMP_WARNING(AffTilesNoTiles, "KMP_AFFINITY");
4464  }
4465 
4466  __kmp_apply_thread_places(&address2os, depth);
4467 
4468  // Create the table of masks, indexed by thread Id.
4469  unsigned maxIndex;
4470  unsigned numUnique;
4471  kmp_affin_mask_t *osId2Mask =
4472  __kmp_create_masks(&maxIndex, &numUnique, address2os, __kmp_avail_proc);
4473  if (__kmp_affinity_gran_levels == 0) {
4474  KMP_DEBUG_ASSERT((int)numUnique == __kmp_avail_proc);
4475  }
4476 
4477  // Set the childNums vector in all Address objects. This must be done before
4478  // we can sort using __kmp_affinity_cmp_Address_child_num(), which takes into
4479  // account the setting of __kmp_affinity_compact.
4480  __kmp_affinity_assign_child_nums(address2os, __kmp_avail_proc);
4481 
4482  switch (__kmp_affinity_type) {
4483 
4484  case affinity_explicit:
4485  KMP_DEBUG_ASSERT(__kmp_affinity_proclist != NULL);
4486 #if OMP_40_ENABLED
4487  if (__kmp_nested_proc_bind.bind_types[0] == proc_bind_intel)
4488 #endif
4489  {
4490  __kmp_affinity_process_proclist(
4491  &__kmp_affinity_masks, &__kmp_affinity_num_masks,
4492  __kmp_affinity_proclist, osId2Mask, maxIndex);
4493  }
4494 #if OMP_40_ENABLED
4495  else {
4496  __kmp_affinity_process_placelist(
4497  &__kmp_affinity_masks, &__kmp_affinity_num_masks,
4498  __kmp_affinity_proclist, osId2Mask, maxIndex);
4499  }
4500 #endif
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 OMP_40_ENABLED
4617  if ((__kmp_nested_proc_bind.bind_types[0] != proc_bind_intel) &&
4618  (__kmp_affinity_num_places > 0) &&
4619  ((unsigned)__kmp_affinity_num_places < __kmp_affinity_num_masks)) {
4620  __kmp_affinity_num_masks = __kmp_affinity_num_places;
4621  }
4622 #endif
4623 
4624  KMP_CPU_ALLOC_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks);
4625 
4626  // Sort the address2os table according to the current setting of
4627  // __kmp_affinity_compact, then fill out __kmp_affinity_masks.
4628  qsort(address2os, __kmp_avail_proc, sizeof(*address2os),
4629  __kmp_affinity_cmp_Address_child_num);
4630  {
4631  int i;
4632  unsigned j;
4633  for (i = 0, j = 0; i < __kmp_avail_proc; i++) {
4634  if ((!__kmp_affinity_dups) && (!address2os[i].first.leader)) {
4635  continue;
4636  }
4637  unsigned osId = address2os[i].second;
4638  kmp_affin_mask_t *src = KMP_CPU_INDEX(osId2Mask, osId);
4639  kmp_affin_mask_t *dest = KMP_CPU_INDEX(__kmp_affinity_masks, j);
4640  KMP_ASSERT(KMP_CPU_ISSET(osId, src));
4641  KMP_CPU_COPY(dest, src);
4642  if (++j >= __kmp_affinity_num_masks) {
4643  break;
4644  }
4645  }
4646  KMP_DEBUG_ASSERT(j == __kmp_affinity_num_masks);
4647  }
4648  break;
4649 
4650  default:
4651  KMP_ASSERT2(0, "Unexpected affinity setting");
4652  }
4653 
4654  KMP_CPU_FREE_ARRAY(osId2Mask, maxIndex + 1);
4655  machine_hierarchy.init(address2os, __kmp_avail_proc);
4656 }
4657 #undef KMP_EXIT_AFF_NONE
4658 
4659 void __kmp_affinity_initialize(void) {
4660  // Much of the code above was written assumming that if a machine was not
4661  // affinity capable, then __kmp_affinity_type == affinity_none. We now
4662  // explicitly represent this as __kmp_affinity_type == affinity_disabled.
4663  // There are too many checks for __kmp_affinity_type == affinity_none
4664  // in this code. Instead of trying to change them all, check if
4665  // __kmp_affinity_type == affinity_disabled, and if so, slam it with
4666  // affinity_none, call the real initialization routine, then restore
4667  // __kmp_affinity_type to affinity_disabled.
4668  int disabled = (__kmp_affinity_type == affinity_disabled);
4669  if (!KMP_AFFINITY_CAPABLE()) {
4670  KMP_ASSERT(disabled);
4671  }
4672  if (disabled) {
4673  __kmp_affinity_type = affinity_none;
4674  }
4675  __kmp_aux_affinity_initialize();
4676  if (disabled) {
4677  __kmp_affinity_type = affinity_disabled;
4678  }
4679 }
4680 
4681 void __kmp_affinity_uninitialize(void) {
4682  if (__kmp_affinity_masks != NULL) {
4683  KMP_CPU_FREE_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks);
4684  __kmp_affinity_masks = NULL;
4685  }
4686  if (__kmp_affin_fullMask != NULL) {
4687  KMP_CPU_FREE(__kmp_affin_fullMask);
4688  __kmp_affin_fullMask = NULL;
4689  }
4690  __kmp_affinity_num_masks = 0;
4691  __kmp_affinity_type = affinity_default;
4692 #if OMP_40_ENABLED
4693  __kmp_affinity_num_places = 0;
4694 #endif
4695  if (__kmp_affinity_proclist != NULL) {
4696  __kmp_free(__kmp_affinity_proclist);
4697  __kmp_affinity_proclist = NULL;
4698  }
4699  if (address2os != NULL) {
4700  __kmp_free(address2os);
4701  address2os = NULL;
4702  }
4703  if (procarr != NULL) {
4704  __kmp_free(procarr);
4705  procarr = NULL;
4706  }
4707 #if KMP_USE_HWLOC
4708  if (__kmp_hwloc_topology != NULL) {
4709  hwloc_topology_destroy(__kmp_hwloc_topology);
4710  __kmp_hwloc_topology = NULL;
4711  }
4712 #endif
4713  KMPAffinity::destroy_api();
4714 }
4715 
4716 void __kmp_affinity_set_init_mask(int gtid, int isa_root) {
4717  if (!KMP_AFFINITY_CAPABLE()) {
4718  return;
4719  }
4720 
4721  kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
4722  if (th->th.th_affin_mask == NULL) {
4723  KMP_CPU_ALLOC(th->th.th_affin_mask);
4724  } else {
4725  KMP_CPU_ZERO(th->th.th_affin_mask);
4726  }
4727 
4728  // Copy the thread mask to the kmp_info_t strucuture. If
4729  // __kmp_affinity_type == affinity_none, copy the "full" mask, i.e. one that
4730  // has all of the OS proc ids set, or if __kmp_affinity_respect_mask is set,
4731  // then the full mask is the same as the mask of the initialization thread.
4732  kmp_affin_mask_t *mask;
4733  int i;
4734 
4735 #if OMP_40_ENABLED
4736  if (KMP_AFFINITY_NON_PROC_BIND)
4737 #endif
4738  {
4739  if ((__kmp_affinity_type == affinity_none) ||
4740  (__kmp_affinity_type == affinity_balanced)) {
4741 #if KMP_GROUP_AFFINITY
4742  if (__kmp_num_proc_groups > 1) {
4743  return;
4744  }
4745 #endif
4746  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4747  i = 0;
4748  mask = __kmp_affin_fullMask;
4749  } else {
4750  KMP_DEBUG_ASSERT(__kmp_affinity_num_masks > 0);
4751  i = (gtid + __kmp_affinity_offset) % __kmp_affinity_num_masks;
4752  mask = KMP_CPU_INDEX(__kmp_affinity_masks, i);
4753  }
4754  }
4755 #if OMP_40_ENABLED
4756  else {
4757  if ((!isa_root) ||
4758  (__kmp_nested_proc_bind.bind_types[0] == proc_bind_false)) {
4759 #if KMP_GROUP_AFFINITY
4760  if (__kmp_num_proc_groups > 1) {
4761  return;
4762  }
4763 #endif
4764  KMP_ASSERT(__kmp_affin_fullMask != NULL);
4765  i = KMP_PLACE_ALL;
4766  mask = __kmp_affin_fullMask;
4767  } else {
4768  // int i = some hash function or just a counter that doesn't
4769  // always start at 0. Use gtid for now.
4770  KMP_DEBUG_ASSERT(__kmp_affinity_num_masks > 0);
4771  i = (gtid + __kmp_affinity_offset) % __kmp_affinity_num_masks;
4772  mask = KMP_CPU_INDEX(__kmp_affinity_masks, i);
4773  }
4774  }
4775 #endif
4776 
4777 #if OMP_40_ENABLED
4778  th->th.th_current_place = i;
4779  if (isa_root) {
4780  th->th.th_new_place = i;
4781  th->th.th_first_place = 0;
4782  th->th.th_last_place = __kmp_affinity_num_masks - 1;
4783  } else if (KMP_AFFINITY_NON_PROC_BIND) {
4784  // When using a Non-OMP_PROC_BIND affinity method,
4785  // set all threads' place-partition-var to the entire place list
4786  th->th.th_first_place = 0;
4787  th->th.th_last_place = __kmp_affinity_num_masks - 1;
4788  }
4789 
4790  if (i == KMP_PLACE_ALL) {
4791  KA_TRACE(100, ("__kmp_affinity_set_init_mask: binding T#%d to all places\n",
4792  gtid));
4793  } else {
4794  KA_TRACE(100, ("__kmp_affinity_set_init_mask: binding T#%d to place %d\n",
4795  gtid, i));
4796  }
4797 #else
4798  if (i == -1) {
4799  KA_TRACE(
4800  100,
4801  ("__kmp_affinity_set_init_mask: binding T#%d to __kmp_affin_fullMask\n",
4802  gtid));
4803  } else {
4804  KA_TRACE(100, ("__kmp_affinity_set_init_mask: binding T#%d to mask %d\n",
4805  gtid, i));
4806  }
4807 #endif /* OMP_40_ENABLED */
4808 
4809  KMP_CPU_COPY(th->th.th_affin_mask, mask);
4810 
4811  if (__kmp_affinity_verbose
4812  /* to avoid duplicate printing (will be correctly printed on barrier) */
4813  && (__kmp_affinity_type == affinity_none ||
4814  (i != KMP_PLACE_ALL && __kmp_affinity_type != affinity_balanced))) {
4815  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4816  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4817  th->th.th_affin_mask);
4818  KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY", (kmp_int32)getpid(),
4819  __kmp_gettid(), gtid, buf);
4820  }
4821 
4822 #if KMP_OS_WINDOWS
4823  // On Windows* OS, the process affinity mask might have changed. If the user
4824  // didn't request affinity and this call fails, just continue silently.
4825  // See CQ171393.
4826  if (__kmp_affinity_type == affinity_none) {
4827  __kmp_set_system_affinity(th->th.th_affin_mask, FALSE);
4828  } else
4829 #endif
4830  __kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
4831 }
4832 
4833 #if OMP_40_ENABLED
4834 
4835 void __kmp_affinity_set_place(int gtid) {
4836  if (!KMP_AFFINITY_CAPABLE()) {
4837  return;
4838  }
4839 
4840  kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
4841 
4842  KA_TRACE(100, ("__kmp_affinity_set_place: binding T#%d to place %d (current "
4843  "place = %d)\n",
4844  gtid, th->th.th_new_place, th->th.th_current_place));
4845 
4846  // Check that the new place is within this thread's partition.
4847  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
4848  KMP_ASSERT(th->th.th_new_place >= 0);
4849  KMP_ASSERT((unsigned)th->th.th_new_place <= __kmp_affinity_num_masks);
4850  if (th->th.th_first_place <= th->th.th_last_place) {
4851  KMP_ASSERT((th->th.th_new_place >= th->th.th_first_place) &&
4852  (th->th.th_new_place <= th->th.th_last_place));
4853  } else {
4854  KMP_ASSERT((th->th.th_new_place <= th->th.th_first_place) ||
4855  (th->th.th_new_place >= th->th.th_last_place));
4856  }
4857 
4858  // Copy the thread mask to the kmp_info_t strucuture,
4859  // and set this thread's affinity.
4860  kmp_affin_mask_t *mask =
4861  KMP_CPU_INDEX(__kmp_affinity_masks, th->th.th_new_place);
4862  KMP_CPU_COPY(th->th.th_affin_mask, mask);
4863  th->th.th_current_place = th->th.th_new_place;
4864 
4865  if (__kmp_affinity_verbose) {
4866  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4867  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4868  th->th.th_affin_mask);
4869  KMP_INFORM(BoundToOSProcSet, "OMP_PROC_BIND", (kmp_int32)getpid(),
4870  __kmp_gettid(), gtid, buf);
4871  }
4872  __kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
4873 }
4874 
4875 #endif /* OMP_40_ENABLED */
4876 
4877 int __kmp_aux_set_affinity(void **mask) {
4878  int gtid;
4879  kmp_info_t *th;
4880  int retval;
4881 
4882  if (!KMP_AFFINITY_CAPABLE()) {
4883  return -1;
4884  }
4885 
4886  gtid = __kmp_entry_gtid();
4887  KA_TRACE(1000, (""); {
4888  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4889  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4890  (kmp_affin_mask_t *)(*mask));
4891  __kmp_debug_printf(
4892  "kmp_set_affinity: setting affinity mask for thread %d = %s\n", gtid,
4893  buf);
4894  });
4895 
4896  if (__kmp_env_consistency_check) {
4897  if ((mask == NULL) || (*mask == NULL)) {
4898  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4899  } else {
4900  unsigned proc;
4901  int num_procs = 0;
4902 
4903  KMP_CPU_SET_ITERATE(proc, ((kmp_affin_mask_t *)(*mask))) {
4904  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
4905  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4906  }
4907  if (!KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask))) {
4908  continue;
4909  }
4910  num_procs++;
4911  }
4912  if (num_procs == 0) {
4913  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4914  }
4915 
4916 #if KMP_GROUP_AFFINITY
4917  if (__kmp_get_proc_group((kmp_affin_mask_t *)(*mask)) < 0) {
4918  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
4919  }
4920 #endif /* KMP_GROUP_AFFINITY */
4921  }
4922  }
4923 
4924  th = __kmp_threads[gtid];
4925  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
4926  retval = __kmp_set_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
4927  if (retval == 0) {
4928  KMP_CPU_COPY(th->th.th_affin_mask, (kmp_affin_mask_t *)(*mask));
4929  }
4930 
4931 #if OMP_40_ENABLED
4932  th->th.th_current_place = KMP_PLACE_UNDEFINED;
4933  th->th.th_new_place = KMP_PLACE_UNDEFINED;
4934  th->th.th_first_place = 0;
4935  th->th.th_last_place = __kmp_affinity_num_masks - 1;
4936 
4937  // Turn off 4.0 affinity for the current tread at this parallel level.
4938  th->th.th_current_task->td_icvs.proc_bind = proc_bind_false;
4939 #endif
4940 
4941  return retval;
4942 }
4943 
4944 int __kmp_aux_get_affinity(void **mask) {
4945  int gtid;
4946  int retval;
4947  kmp_info_t *th;
4948 
4949  if (!KMP_AFFINITY_CAPABLE()) {
4950  return -1;
4951  }
4952 
4953  gtid = __kmp_entry_gtid();
4954  th = __kmp_threads[gtid];
4955  KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
4956 
4957  KA_TRACE(1000, (""); {
4958  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4959  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4960  th->th.th_affin_mask);
4961  __kmp_printf("kmp_get_affinity: stored affinity mask for thread %d = %s\n",
4962  gtid, buf);
4963  });
4964 
4965  if (__kmp_env_consistency_check) {
4966  if ((mask == NULL) || (*mask == NULL)) {
4967  KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity");
4968  }
4969  }
4970 
4971 #if !KMP_OS_WINDOWS
4972 
4973  retval = __kmp_get_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
4974  KA_TRACE(1000, (""); {
4975  char buf[KMP_AFFIN_MASK_PRINT_LEN];
4976  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4977  (kmp_affin_mask_t *)(*mask));
4978  __kmp_printf("kmp_get_affinity: system affinity mask for thread %d = %s\n",
4979  gtid, buf);
4980  });
4981  return retval;
4982 
4983 #else
4984 
4985  KMP_CPU_COPY((kmp_affin_mask_t *)(*mask), th->th.th_affin_mask);
4986  return 0;
4987 
4988 #endif /* KMP_OS_WINDOWS */
4989 }
4990 
4991 int __kmp_aux_get_affinity_max_proc() {
4992  if (!KMP_AFFINITY_CAPABLE()) {
4993  return 0;
4994  }
4995 #if KMP_GROUP_AFFINITY
4996  if (__kmp_num_proc_groups > 1) {
4997  return (int)(__kmp_num_proc_groups * sizeof(DWORD_PTR) * CHAR_BIT);
4998  }
4999 #endif
5000  return __kmp_xproc;
5001 }
5002 
5003 int __kmp_aux_set_affinity_mask_proc(int proc, void **mask) {
5004  if (!KMP_AFFINITY_CAPABLE()) {
5005  return -1;
5006  }
5007 
5008  KA_TRACE(1000, (""); {
5009  int gtid = __kmp_entry_gtid();
5010  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5011  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5012  (kmp_affin_mask_t *)(*mask));
5013  __kmp_debug_printf("kmp_set_affinity_mask_proc: setting proc %d in "
5014  "affinity mask for thread %d = %s\n",
5015  proc, gtid, buf);
5016  });
5017 
5018  if (__kmp_env_consistency_check) {
5019  if ((mask == NULL) || (*mask == NULL)) {
5020  KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity_mask_proc");
5021  }
5022  }
5023 
5024  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5025  return -1;
5026  }
5027  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5028  return -2;
5029  }
5030 
5031  KMP_CPU_SET(proc, (kmp_affin_mask_t *)(*mask));
5032  return 0;
5033 }
5034 
5035 int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask) {
5036  if (!KMP_AFFINITY_CAPABLE()) {
5037  return -1;
5038  }
5039 
5040  KA_TRACE(1000, (""); {
5041  int gtid = __kmp_entry_gtid();
5042  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5043  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5044  (kmp_affin_mask_t *)(*mask));
5045  __kmp_debug_printf("kmp_unset_affinity_mask_proc: unsetting proc %d in "
5046  "affinity mask for thread %d = %s\n",
5047  proc, gtid, buf);
5048  });
5049 
5050  if (__kmp_env_consistency_check) {
5051  if ((mask == NULL) || (*mask == NULL)) {
5052  KMP_FATAL(AffinityInvalidMask, "kmp_unset_affinity_mask_proc");
5053  }
5054  }
5055 
5056  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5057  return -1;
5058  }
5059  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5060  return -2;
5061  }
5062 
5063  KMP_CPU_CLR(proc, (kmp_affin_mask_t *)(*mask));
5064  return 0;
5065 }
5066 
5067 int __kmp_aux_get_affinity_mask_proc(int proc, void **mask) {
5068  if (!KMP_AFFINITY_CAPABLE()) {
5069  return -1;
5070  }
5071 
5072  KA_TRACE(1000, (""); {
5073  int gtid = __kmp_entry_gtid();
5074  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5075  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5076  (kmp_affin_mask_t *)(*mask));
5077  __kmp_debug_printf("kmp_get_affinity_mask_proc: getting proc %d in "
5078  "affinity mask for thread %d = %s\n",
5079  proc, gtid, buf);
5080  });
5081 
5082  if (__kmp_env_consistency_check) {
5083  if ((mask == NULL) || (*mask == NULL)) {
5084  KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity_mask_proc");
5085  }
5086  }
5087 
5088  if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5089  return -1;
5090  }
5091  if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5092  return 0;
5093  }
5094 
5095  return KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask));
5096 }
5097 
5098 // Dynamic affinity settings - Affinity balanced
5099 void __kmp_balanced_affinity(kmp_info_t *th, int nthreads) {
5100  KMP_DEBUG_ASSERT(th);
5101  bool fine_gran = true;
5102  int tid = th->th.th_info.ds.ds_tid;
5103 
5104  switch (__kmp_affinity_gran) {
5105  case affinity_gran_fine:
5106  case affinity_gran_thread:
5107  break;
5108  case affinity_gran_core:
5109  if (__kmp_nThreadsPerCore > 1) {
5110  fine_gran = false;
5111  }
5112  break;
5113  case affinity_gran_package:
5114  if (nCoresPerPkg > 1) {
5115  fine_gran = false;
5116  }
5117  break;
5118  default:
5119  fine_gran = false;
5120  }
5121 
5122  if (__kmp_affinity_uniform_topology()) {
5123  int coreID;
5124  int threadID;
5125  // Number of hyper threads per core in HT machine
5126  int __kmp_nth_per_core = __kmp_avail_proc / __kmp_ncores;
5127  // Number of cores
5128  int ncores = __kmp_ncores;
5129  if ((nPackages > 1) && (__kmp_nth_per_core <= 1)) {
5130  __kmp_nth_per_core = __kmp_avail_proc / nPackages;
5131  ncores = nPackages;
5132  }
5133  // How many threads will be bound to each core
5134  int chunk = nthreads / ncores;
5135  // How many cores will have an additional thread bound to it - "big cores"
5136  int big_cores = nthreads % ncores;
5137  // Number of threads on the big cores
5138  int big_nth = (chunk + 1) * big_cores;
5139  if (tid < big_nth) {
5140  coreID = tid / (chunk + 1);
5141  threadID = (tid % (chunk + 1)) % __kmp_nth_per_core;
5142  } else { // tid >= big_nth
5143  coreID = (tid - big_cores) / chunk;
5144  threadID = ((tid - big_cores) % chunk) % __kmp_nth_per_core;
5145  }
5146 
5147  KMP_DEBUG_ASSERT2(KMP_AFFINITY_CAPABLE(),
5148  "Illegal set affinity operation when not capable");
5149 
5150  kmp_affin_mask_t *mask = th->th.th_affin_mask;
5151  KMP_CPU_ZERO(mask);
5152 
5153  if (fine_gran) {
5154  int osID = address2os[coreID * __kmp_nth_per_core + threadID].second;
5155  KMP_CPU_SET(osID, mask);
5156  } else {
5157  for (int i = 0; i < __kmp_nth_per_core; i++) {
5158  int osID;
5159  osID = address2os[coreID * __kmp_nth_per_core + i].second;
5160  KMP_CPU_SET(osID, mask);
5161  }
5162  }
5163  if (__kmp_affinity_verbose) {
5164  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5165  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
5166  KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY", (kmp_int32)getpid(),
5167  __kmp_gettid(), tid, buf);
5168  }
5169  __kmp_set_system_affinity(mask, TRUE);
5170  } else { // Non-uniform topology
5171 
5172  kmp_affin_mask_t *mask = th->th.th_affin_mask;
5173  KMP_CPU_ZERO(mask);
5174 
5175  int core_level = __kmp_affinity_find_core_level(
5176  address2os, __kmp_avail_proc, __kmp_aff_depth - 1);
5177  int ncores = __kmp_affinity_compute_ncores(address2os, __kmp_avail_proc,
5178  __kmp_aff_depth - 1, core_level);
5179  int nth_per_core = __kmp_affinity_max_proc_per_core(
5180  address2os, __kmp_avail_proc, __kmp_aff_depth - 1, core_level);
5181 
5182  // For performance gain consider the special case nthreads ==
5183  // __kmp_avail_proc
5184  if (nthreads == __kmp_avail_proc) {
5185  if (fine_gran) {
5186  int osID = address2os[tid].second;
5187  KMP_CPU_SET(osID, mask);
5188  } else {
5189  int core = __kmp_affinity_find_core(address2os, tid,
5190  __kmp_aff_depth - 1, core_level);
5191  for (int i = 0; i < __kmp_avail_proc; i++) {
5192  int osID = address2os[i].second;
5193  if (__kmp_affinity_find_core(address2os, i, __kmp_aff_depth - 1,
5194  core_level) == core) {
5195  KMP_CPU_SET(osID, mask);
5196  }
5197  }
5198  }
5199  } else if (nthreads <= ncores) {
5200 
5201  int core = 0;
5202  for (int i = 0; i < ncores; i++) {
5203  // Check if this core from procarr[] is in the mask
5204  int in_mask = 0;
5205  for (int j = 0; j < nth_per_core; j++) {
5206  if (procarr[i * nth_per_core + j] != -1) {
5207  in_mask = 1;
5208  break;
5209  }
5210  }
5211  if (in_mask) {
5212  if (tid == core) {
5213  for (int j = 0; j < nth_per_core; j++) {
5214  int osID = procarr[i * nth_per_core + j];
5215  if (osID != -1) {
5216  KMP_CPU_SET(osID, mask);
5217  // For fine granularity it is enough to set the first available
5218  // osID for this core
5219  if (fine_gran) {
5220  break;
5221  }
5222  }
5223  }
5224  break;
5225  } else {
5226  core++;
5227  }
5228  }
5229  }
5230  } else { // nthreads > ncores
5231  // Array to save the number of processors at each core
5232  int *nproc_at_core = (int *)KMP_ALLOCA(sizeof(int) * ncores);
5233  // Array to save the number of cores with "x" available processors;
5234  int *ncores_with_x_procs =
5235  (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
5236  // Array to save the number of cores with # procs from x to nth_per_core
5237  int *ncores_with_x_to_max_procs =
5238  (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
5239 
5240  for (int i = 0; i <= nth_per_core; i++) {
5241  ncores_with_x_procs[i] = 0;
5242  ncores_with_x_to_max_procs[i] = 0;
5243  }
5244 
5245  for (int i = 0; i < ncores; i++) {
5246  int cnt = 0;
5247  for (int j = 0; j < nth_per_core; j++) {
5248  if (procarr[i * nth_per_core + j] != -1) {
5249  cnt++;
5250  }
5251  }
5252  nproc_at_core[i] = cnt;
5253  ncores_with_x_procs[cnt]++;
5254  }
5255 
5256  for (int i = 0; i <= nth_per_core; i++) {
5257  for (int j = i; j <= nth_per_core; j++) {
5258  ncores_with_x_to_max_procs[i] += ncores_with_x_procs[j];
5259  }
5260  }
5261 
5262  // Max number of processors
5263  int nproc = nth_per_core * ncores;
5264  // An array to keep number of threads per each context
5265  int *newarr = (int *)__kmp_allocate(sizeof(int) * nproc);
5266  for (int i = 0; i < nproc; i++) {
5267  newarr[i] = 0;
5268  }
5269 
5270  int nth = nthreads;
5271  int flag = 0;
5272  while (nth > 0) {
5273  for (int j = 1; j <= nth_per_core; j++) {
5274  int cnt = ncores_with_x_to_max_procs[j];
5275  for (int i = 0; i < ncores; i++) {
5276  // Skip the core with 0 processors
5277  if (nproc_at_core[i] == 0) {
5278  continue;
5279  }
5280  for (int k = 0; k < nth_per_core; k++) {
5281  if (procarr[i * nth_per_core + k] != -1) {
5282  if (newarr[i * nth_per_core + k] == 0) {
5283  newarr[i * nth_per_core + k] = 1;
5284  cnt--;
5285  nth--;
5286  break;
5287  } else {
5288  if (flag != 0) {
5289  newarr[i * nth_per_core + k]++;
5290  cnt--;
5291  nth--;
5292  break;
5293  }
5294  }
5295  }
5296  }
5297  if (cnt == 0 || nth == 0) {
5298  break;
5299  }
5300  }
5301  if (nth == 0) {
5302  break;
5303  }
5304  }
5305  flag = 1;
5306  }
5307  int sum = 0;
5308  for (int i = 0; i < nproc; i++) {
5309  sum += newarr[i];
5310  if (sum > tid) {
5311  if (fine_gran) {
5312  int osID = procarr[i];
5313  KMP_CPU_SET(osID, mask);
5314  } else {
5315  int coreID = i / nth_per_core;
5316  for (int ii = 0; ii < nth_per_core; ii++) {
5317  int osID = procarr[coreID * nth_per_core + ii];
5318  if (osID != -1) {
5319  KMP_CPU_SET(osID, mask);
5320  }
5321  }
5322  }
5323  break;
5324  }
5325  }
5326  __kmp_free(newarr);
5327  }
5328 
5329  if (__kmp_affinity_verbose) {
5330  char buf[KMP_AFFIN_MASK_PRINT_LEN];
5331  __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
5332  KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY", (kmp_int32)getpid(),
5333  __kmp_gettid(), tid, buf);
5334  }
5335  __kmp_set_system_affinity(mask, TRUE);
5336  }
5337 }
5338 
5339 #if KMP_OS_LINUX
5340 // We don't need this entry for Windows because
5341 // there is GetProcessAffinityMask() api
5342 //
5343 // The intended usage is indicated by these steps:
5344 // 1) The user gets the current affinity mask
5345 // 2) Then sets the affinity by calling this function
5346 // 3) Error check the return value
5347 // 4) Use non-OpenMP parallelization
5348 // 5) Reset the affinity to what was stored in step 1)
5349 #ifdef __cplusplus
5350 extern "C"
5351 #endif
5352  int
5353  kmp_set_thread_affinity_mask_initial()
5354 // the function returns 0 on success,
5355 // -1 if we cannot bind thread
5356 // >0 (errno) if an error happened during binding
5357 {
5358  int gtid = __kmp_get_gtid();
5359  if (gtid < 0) {
5360  // Do not touch non-omp threads
5361  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5362  "non-omp thread, returning\n"));
5363  return -1;
5364  }
5365  if (!KMP_AFFINITY_CAPABLE() || !__kmp_init_middle) {
5366  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5367  "affinity not initialized, returning\n"));
5368  return -1;
5369  }
5370  KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5371  "set full mask for thread %d\n",
5372  gtid));
5373  KMP_DEBUG_ASSERT(__kmp_affin_fullMask != NULL);
5374  return __kmp_set_system_affinity(__kmp_affin_fullMask, FALSE);
5375 }
5376 #endif
5377 
5378 #endif // KMP_AFFINITY_SUPPORTED