Bug Summary

File:build/source/openmp/runtime/src/kmp_affinity.cpp
Warning:line 3502, column 15
Value stored to 'next' during its initialization is never read

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -clear-ast-before-backend -disable-llvm-verifier -discard-value-names -main-file-name kmp_affinity.cpp -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mframe-pointer=none -fmath-errno -ffp-contract=on -fno-rounding-math -mconstructor-aliases -funwind-tables=2 -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -ffunction-sections -fdata-sections -fcoverage-compilation-dir=/build/source/build-llvm/tools/clang/stage2-bins -resource-dir /usr/lib/llvm-17/lib/clang/17 -D _DEBUG -D _GLIBCXX_ASSERTIONS -D _GNU_SOURCE -D _LIBCPP_ENABLE_ASSERTIONS -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -D omp_EXPORTS -I projects/openmp/runtime/src -I /build/source/openmp/runtime/src -I include -I /build/source/llvm/include -I /build/source/openmp/runtime/src/i18n -I /build/source/openmp/runtime/src/include -I /build/source/openmp/runtime/src/thirdparty/ittnotify -D _FORTIFY_SOURCE=2 -D NDEBUG -D _GNU_SOURCE -D _REENTRANT -D _FORTIFY_SOURCE=2 -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/x86_64-linux-gnu/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/backward -internal-isystem /usr/lib/llvm-17/lib/clang/17/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../x86_64-linux-gnu/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -fmacro-prefix-map=/build/source/build-llvm/tools/clang/stage2-bins=build-llvm/tools/clang/stage2-bins -fmacro-prefix-map=/build/source/= -fcoverage-prefix-map=/build/source/build-llvm/tools/clang/stage2-bins=build-llvm/tools/clang/stage2-bins -fcoverage-prefix-map=/build/source/= -source-date-epoch 1683717183 -O2 -Wno-unused-command-line-argument -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-class-memaccess -Wno-redundant-move -Wno-pessimizing-move -Wno-noexcept-type -Wno-comment -Wno-misleading-indentation -Wno-extra -Wno-pedantic -Wno-maybe-uninitialized -Wno-class-memaccess -Wno-frame-address -Wno-strict-aliasing -Wno-stringop-truncation -Wno-switch -Wno-uninitialized -Wno-cast-qual -std=c++17 -fdeprecated-macro -fdebug-compilation-dir=/build/source/build-llvm/tools/clang/stage2-bins -fdebug-prefix-map=/build/source/build-llvm/tools/clang/stage2-bins=build-llvm/tools/clang/stage2-bins -fdebug-prefix-map=/build/source/= -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fno-rtti -fgnuc-version=4.2.1 -fcolor-diagnostics -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /tmp/scan-build-2023-05-10-133810-16478-1 -x c++ /build/source/openmp/runtime/src/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_SCHED0
20#include "kmp_dispatch_hier.h"
21#endif
22#if KMP_USE_HWLOC0
23// Copied from hwloc
24#define HWLOC_GROUP_KIND_INTEL_MODULE 102
25#define HWLOC_GROUP_KIND_INTEL_TILE 103
26#define HWLOC_GROUP_KIND_INTEL_DIE 104
27#define HWLOC_GROUP_KIND_WINDOWS_PROCESSOR_GROUP 220
28#endif
29#include <ctype.h>
30
31// The machine topology
32kmp_topology_t *__kmp_topology = nullptr;
33// KMP_HW_SUBSET environment variable
34kmp_hw_subset_t *__kmp_hw_subset = nullptr;
35
36// Store the real or imagined machine hierarchy here
37static hierarchy_info machine_hierarchy;
38
39void __kmp_cleanup_hierarchy() { machine_hierarchy.fini(); }
40
41void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar) {
42 kmp_uint32 depth;
43 // The test below is true if affinity is available, but set to "none". Need to
44 // init on first use of hierarchical barrier.
45 if (TCR_1(machine_hierarchy.uninitialized)(machine_hierarchy.uninitialized))
46 machine_hierarchy.init(nproc);
47
48 // Adjust the hierarchy in case num threads exceeds original
49 if (nproc > machine_hierarchy.base_num_threads)
50 machine_hierarchy.resize(nproc);
51
52 depth = machine_hierarchy.depth;
53 KMP_DEBUG_ASSERT(depth > 0)if (!(depth > 0)) { __kmp_debug_assert("depth > 0", "openmp/runtime/src/kmp_affinity.cpp"
, 53); }
;
54
55 thr_bar->depth = depth;
56 __kmp_type_convert(machine_hierarchy.numPerLevel[0] - 1,
57 &(thr_bar->base_leaf_kids));
58 thr_bar->skip_per_level = machine_hierarchy.skipPerLevel;
59}
60
61static int nCoresPerPkg, nPackages;
62static int __kmp_nThreadsPerCore;
63#ifndef KMP_DFLT_NTH_CORES
64static int __kmp_ncores;
65#endif
66
67const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural) {
68 switch (type) {
69 case KMP_HW_SOCKET:
70 return ((plural) ? KMP_I18N_STR(Sockets)__kmp_i18n_catgets(kmp_i18n_str_Sockets) : KMP_I18N_STR(Socket)__kmp_i18n_catgets(kmp_i18n_str_Socket));
71 case KMP_HW_DIE:
72 return ((plural) ? KMP_I18N_STR(Dice)__kmp_i18n_catgets(kmp_i18n_str_Dice) : KMP_I18N_STR(Die)__kmp_i18n_catgets(kmp_i18n_str_Die));
73 case KMP_HW_MODULE:
74 return ((plural) ? KMP_I18N_STR(Modules)__kmp_i18n_catgets(kmp_i18n_str_Modules) : KMP_I18N_STR(Module)__kmp_i18n_catgets(kmp_i18n_str_Module));
75 case KMP_HW_TILE:
76 return ((plural) ? KMP_I18N_STR(Tiles)__kmp_i18n_catgets(kmp_i18n_str_Tiles) : KMP_I18N_STR(Tile)__kmp_i18n_catgets(kmp_i18n_str_Tile));
77 case KMP_HW_NUMA:
78 return ((plural) ? KMP_I18N_STR(NumaDomains)__kmp_i18n_catgets(kmp_i18n_str_NumaDomains) : KMP_I18N_STR(NumaDomain)__kmp_i18n_catgets(kmp_i18n_str_NumaDomain));
79 case KMP_HW_L3:
80 return ((plural) ? KMP_I18N_STR(L3Caches)__kmp_i18n_catgets(kmp_i18n_str_L3Caches) : KMP_I18N_STR(L3Cache)__kmp_i18n_catgets(kmp_i18n_str_L3Cache));
81 case KMP_HW_L2:
82 return ((plural) ? KMP_I18N_STR(L2Caches)__kmp_i18n_catgets(kmp_i18n_str_L2Caches) : KMP_I18N_STR(L2Cache)__kmp_i18n_catgets(kmp_i18n_str_L2Cache));
83 case KMP_HW_L1:
84 return ((plural) ? KMP_I18N_STR(L1Caches)__kmp_i18n_catgets(kmp_i18n_str_L1Caches) : KMP_I18N_STR(L1Cache)__kmp_i18n_catgets(kmp_i18n_str_L1Cache));
85 case KMP_HW_LLC:
86 return ((plural) ? KMP_I18N_STR(LLCaches)__kmp_i18n_catgets(kmp_i18n_str_LLCaches) : KMP_I18N_STR(LLCache)__kmp_i18n_catgets(kmp_i18n_str_LLCache));
87 case KMP_HW_CORE:
88 return ((plural) ? KMP_I18N_STR(Cores)__kmp_i18n_catgets(kmp_i18n_str_Cores) : KMP_I18N_STR(Core)__kmp_i18n_catgets(kmp_i18n_str_Core));
89 case KMP_HW_THREAD:
90 return ((plural) ? KMP_I18N_STR(Threads)__kmp_i18n_catgets(kmp_i18n_str_Threads) : KMP_I18N_STR(Thread)__kmp_i18n_catgets(kmp_i18n_str_Thread));
91 case KMP_HW_PROC_GROUP:
92 return ((plural) ? KMP_I18N_STR(ProcGroups)__kmp_i18n_catgets(kmp_i18n_str_ProcGroups) : KMP_I18N_STR(ProcGroup)__kmp_i18n_catgets(kmp_i18n_str_ProcGroup));
93 }
94 return KMP_I18N_STR(Unknown)__kmp_i18n_catgets(kmp_i18n_str_Unknown);
95}
96
97const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural) {
98 switch (type) {
99 case KMP_HW_SOCKET:
100 return ((plural) ? "sockets" : "socket");
101 case KMP_HW_DIE:
102 return ((plural) ? "dice" : "die");
103 case KMP_HW_MODULE:
104 return ((plural) ? "modules" : "module");
105 case KMP_HW_TILE:
106 return ((plural) ? "tiles" : "tile");
107 case KMP_HW_NUMA:
108 return ((plural) ? "numa_domains" : "numa_domain");
109 case KMP_HW_L3:
110 return ((plural) ? "l3_caches" : "l3_cache");
111 case KMP_HW_L2:
112 return ((plural) ? "l2_caches" : "l2_cache");
113 case KMP_HW_L1:
114 return ((plural) ? "l1_caches" : "l1_cache");
115 case KMP_HW_LLC:
116 return ((plural) ? "ll_caches" : "ll_cache");
117 case KMP_HW_CORE:
118 return ((plural) ? "cores" : "core");
119 case KMP_HW_THREAD:
120 return ((plural) ? "threads" : "thread");
121 case KMP_HW_PROC_GROUP:
122 return ((plural) ? "proc_groups" : "proc_group");
123 }
124 return ((plural) ? "unknowns" : "unknown");
125}
126
127const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type) {
128 switch (type) {
129 case KMP_HW_CORE_TYPE_UNKNOWN:
130 return "unknown";
131#if KMP_ARCH_X860 || KMP_ARCH_X86_641
132 case KMP_HW_CORE_TYPE_ATOM:
133 return "Intel Atom(R) processor";
134 case KMP_HW_CORE_TYPE_CORE:
135 return "Intel(R) Core(TM) processor";
136#endif
137 }
138 return "unknown";
139}
140
141#if KMP_AFFINITY_SUPPORTED1
142// If affinity is supported, check the affinity
143// verbose and warning flags before printing warning
144#define KMP_AFF_WARNING(s, ...)if (s.flags.verbose || (s.flags.warnings && (s.type !=
affinity_none))) { __kmp_msg(kmp_ms_warning, __kmp_msg_format
(kmp_i18n_msg_...), __kmp_msg_null); }
\
145 if (s.flags.verbose || (s.flags.warnings && (s.type != affinity_none))) { \
146 KMP_WARNING(__VA_ARGS__)__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg___VA_ARGS__
), __kmp_msg_null)
; \
147 }
148#else
149#define KMP_AFF_WARNING(s, ...)if (s.flags.verbose || (s.flags.warnings && (s.type !=
affinity_none))) { __kmp_msg(kmp_ms_warning, __kmp_msg_format
(kmp_i18n_msg_...), __kmp_msg_null); }
KMP_WARNING(__VA_ARGS__)__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg___VA_ARGS__
), __kmp_msg_null)
150#endif
151
152////////////////////////////////////////////////////////////////////////////////
153// kmp_hw_thread_t methods
154int kmp_hw_thread_t::compare_ids(const void *a, const void *b) {
155 const kmp_hw_thread_t *ahwthread = (const kmp_hw_thread_t *)a;
156 const kmp_hw_thread_t *bhwthread = (const kmp_hw_thread_t *)b;
157 int depth = __kmp_topology->get_depth();
158 for (int level = 0; level < depth; ++level) {
159 if (ahwthread->ids[level] < bhwthread->ids[level])
160 return -1;
161 else if (ahwthread->ids[level] > bhwthread->ids[level])
162 return 1;
163 }
164 if (ahwthread->os_id < bhwthread->os_id)
165 return -1;
166 else if (ahwthread->os_id > bhwthread->os_id)
167 return 1;
168 return 0;
169}
170
171#if KMP_AFFINITY_SUPPORTED1
172int kmp_hw_thread_t::compare_compact(const void *a, const void *b) {
173 int i;
174 const kmp_hw_thread_t *aa = (const kmp_hw_thread_t *)a;
175 const kmp_hw_thread_t *bb = (const kmp_hw_thread_t *)b;
176 int depth = __kmp_topology->get_depth();
177 int compact = __kmp_topology->compact;
178 KMP_DEBUG_ASSERT(compact >= 0)if (!(compact >= 0)) { __kmp_debug_assert("compact >= 0"
, "openmp/runtime/src/kmp_affinity.cpp", 178); }
;
179 KMP_DEBUG_ASSERT(compact <= depth)if (!(compact <= depth)) { __kmp_debug_assert("compact <= depth"
, "openmp/runtime/src/kmp_affinity.cpp", 179); }
;
180 for (i = 0; i < compact; i++) {
181 int j = depth - i - 1;
182 if (aa->sub_ids[j] < bb->sub_ids[j])
183 return -1;
184 if (aa->sub_ids[j] > bb->sub_ids[j])
185 return 1;
186 }
187 for (; i < depth; i++) {
188 int j = i - compact;
189 if (aa->sub_ids[j] < bb->sub_ids[j])
190 return -1;
191 if (aa->sub_ids[j] > bb->sub_ids[j])
192 return 1;
193 }
194 return 0;
195}
196#endif
197
198void kmp_hw_thread_t::print() const {
199 int depth = __kmp_topology->get_depth();
200 printf("%4d ", os_id);
201 for (int i = 0; i < depth; ++i) {
202 printf("%4d ", ids[i]);
203 }
204 if (attrs) {
205 if (attrs.is_core_type_valid())
206 printf(" (%s)", __kmp_hw_get_core_type_string(attrs.get_core_type()));
207 if (attrs.is_core_eff_valid())
208 printf(" (eff=%d)", attrs.get_core_eff());
209 }
210 printf("\n");
211}
212
213////////////////////////////////////////////////////////////////////////////////
214// kmp_topology_t methods
215
216// Add a layer to the topology based on the ids. Assume the topology
217// is perfectly nested (i.e., so no object has more than one parent)
218void kmp_topology_t::_insert_layer(kmp_hw_t type, const int *ids) {
219 // Figure out where the layer should go by comparing the ids of the current
220 // layers with the new ids
221 int target_layer;
222 int previous_id = kmp_hw_thread_t::UNKNOWN_ID;
223 int previous_new_id = kmp_hw_thread_t::UNKNOWN_ID;
224
225 // Start from the highest layer and work down to find target layer
226 // If new layer is equal to another layer then put the new layer above
227 for (target_layer = 0; target_layer < depth; ++target_layer) {
228 bool layers_equal = true;
229 bool strictly_above_target_layer = false;
230 for (int i = 0; i < num_hw_threads; ++i) {
231 int id = hw_threads[i].ids[target_layer];
232 int new_id = ids[i];
233 if (id != previous_id && new_id == previous_new_id) {
234 // Found the layer we are strictly above
235 strictly_above_target_layer = true;
236 layers_equal = false;
237 break;
238 } else if (id == previous_id && new_id != previous_new_id) {
239 // Found a layer we are below. Move to next layer and check.
240 layers_equal = false;
241 break;
242 }
243 previous_id = id;
244 previous_new_id = new_id;
245 }
246 if (strictly_above_target_layer || layers_equal)
247 break;
248 }
249
250 // Found the layer we are above. Now move everything to accommodate the new
251 // layer. And put the new ids and type into the topology.
252 for (int i = depth - 1, j = depth; i >= target_layer; --i, --j)
253 types[j] = types[i];
254 types[target_layer] = type;
255 for (int k = 0; k < num_hw_threads; ++k) {
256 for (int i = depth - 1, j = depth; i >= target_layer; --i, --j)
257 hw_threads[k].ids[j] = hw_threads[k].ids[i];
258 hw_threads[k].ids[target_layer] = ids[k];
259 }
260 equivalent[type] = type;
261 depth++;
262}
263
264#if KMP_GROUP_AFFINITY0
265// Insert the Windows Processor Group structure into the topology
266void kmp_topology_t::_insert_windows_proc_groups() {
267 // Do not insert the processor group structure for a single group
268 if (__kmp_num_proc_groups == 1)
269 return;
270 kmp_affin_mask_t *mask;
271 int *ids = (int *)__kmp_allocate(sizeof(int) * num_hw_threads)___kmp_allocate((sizeof(int) * num_hw_threads), "openmp/runtime/src/kmp_affinity.cpp"
, 271)
;
272 KMP_CPU_ALLOC(mask)(mask = __kmp_affinity_dispatch->allocate_mask());
273 for (int i = 0; i < num_hw_threads; ++i) {
274 KMP_CPU_ZERO(mask)(mask)->zero();
275 KMP_CPU_SET(hw_threads[i].os_id, mask)(mask)->set(hw_threads[i].os_id);
276 ids[i] = __kmp_get_proc_group(mask)(mask)->get_proc_group();
277 }
278 KMP_CPU_FREE(mask)__kmp_affinity_dispatch->deallocate_mask(mask);
279 _insert_layer(KMP_HW_PROC_GROUP, ids);
280 __kmp_free(ids)___kmp_free((ids), "openmp/runtime/src/kmp_affinity.cpp", 280
)
;
281}
282#endif
283
284// Remove layers that don't add information to the topology.
285// This is done by having the layer take on the id = UNKNOWN_ID (-1)
286void kmp_topology_t::_remove_radix1_layers() {
287 int preference[KMP_HW_LAST];
288 int top_index1, top_index2;
289 // Set up preference associative array
290 preference[KMP_HW_SOCKET] = 110;
291 preference[KMP_HW_PROC_GROUP] = 100;
292 preference[KMP_HW_CORE] = 95;
293 preference[KMP_HW_THREAD] = 90;
294 preference[KMP_HW_NUMA] = 85;
295 preference[KMP_HW_DIE] = 80;
296 preference[KMP_HW_TILE] = 75;
297 preference[KMP_HW_MODULE] = 73;
298 preference[KMP_HW_L3] = 70;
299 preference[KMP_HW_L2] = 65;
300 preference[KMP_HW_L1] = 60;
301 preference[KMP_HW_LLC] = 5;
302 top_index1 = 0;
303 top_index2 = 1;
304 while (top_index1 < depth - 1 && top_index2 < depth) {
305 kmp_hw_t type1 = types[top_index1];
306 kmp_hw_t type2 = types[top_index2];
307 KMP_ASSERT_VALID_HW_TYPE(type1)if (!(type1 >= (kmp_hw_t)0 && type1 < KMP_HW_LAST
)) { __kmp_debug_assert("type1 >= (kmp_hw_t)0 && type1 < KMP_HW_LAST"
, "openmp/runtime/src/kmp_affinity.cpp", 307); }
;
308 KMP_ASSERT_VALID_HW_TYPE(type2)if (!(type2 >= (kmp_hw_t)0 && type2 < KMP_HW_LAST
)) { __kmp_debug_assert("type2 >= (kmp_hw_t)0 && type2 < KMP_HW_LAST"
, "openmp/runtime/src/kmp_affinity.cpp", 308); }
;
309 // Do not allow the three main topology levels (sockets, cores, threads) to
310 // be compacted down
311 if ((type1 == KMP_HW_THREAD || type1 == KMP_HW_CORE ||
312 type1 == KMP_HW_SOCKET) &&
313 (type2 == KMP_HW_THREAD || type2 == KMP_HW_CORE ||
314 type2 == KMP_HW_SOCKET)) {
315 top_index1 = top_index2++;
316 continue;
317 }
318 bool radix1 = true;
319 bool all_same = true;
320 int id1 = hw_threads[0].ids[top_index1];
321 int id2 = hw_threads[0].ids[top_index2];
322 int pref1 = preference[type1];
323 int pref2 = preference[type2];
324 for (int hwidx = 1; hwidx < num_hw_threads; ++hwidx) {
325 if (hw_threads[hwidx].ids[top_index1] == id1 &&
326 hw_threads[hwidx].ids[top_index2] != id2) {
327 radix1 = false;
328 break;
329 }
330 if (hw_threads[hwidx].ids[top_index2] != id2)
331 all_same = false;
332 id1 = hw_threads[hwidx].ids[top_index1];
333 id2 = hw_threads[hwidx].ids[top_index2];
334 }
335 if (radix1) {
336 // Select the layer to remove based on preference
337 kmp_hw_t remove_type, keep_type;
338 int remove_layer, remove_layer_ids;
339 if (pref1 > pref2) {
340 remove_type = type2;
341 remove_layer = remove_layer_ids = top_index2;
342 keep_type = type1;
343 } else {
344 remove_type = type1;
345 remove_layer = remove_layer_ids = top_index1;
346 keep_type = type2;
347 }
348 // If all the indexes for the second (deeper) layer are the same.
349 // e.g., all are zero, then make sure to keep the first layer's ids
350 if (all_same)
351 remove_layer_ids = top_index2;
352 // Remove radix one type by setting the equivalence, removing the id from
353 // the hw threads and removing the layer from types and depth
354 set_equivalent_type(remove_type, keep_type);
355 for (int idx = 0; idx < num_hw_threads; ++idx) {
356 kmp_hw_thread_t &hw_thread = hw_threads[idx];
357 for (int d = remove_layer_ids; d < depth - 1; ++d)
358 hw_thread.ids[d] = hw_thread.ids[d + 1];
359 }
360 for (int idx = remove_layer; idx < depth - 1; ++idx)
361 types[idx] = types[idx + 1];
362 depth--;
363 } else {
364 top_index1 = top_index2++;
365 }
366 }
367 KMP_ASSERT(depth > 0)if (!(depth > 0)) { __kmp_debug_assert("depth > 0", "openmp/runtime/src/kmp_affinity.cpp"
, 367); }
;
368}
369
370void kmp_topology_t::_set_last_level_cache() {
371 if (get_equivalent_type(KMP_HW_L3) != KMP_HW_UNKNOWN)
372 set_equivalent_type(KMP_HW_LLC, KMP_HW_L3);
373 else if (get_equivalent_type(KMP_HW_L2) != KMP_HW_UNKNOWN)
374 set_equivalent_type(KMP_HW_LLC, KMP_HW_L2);
375#if KMP_MIC_SUPPORTED((0 || 1) && (1 || 0))
376 else if (__kmp_mic_type == mic3) {
377 if (get_equivalent_type(KMP_HW_L2) != KMP_HW_UNKNOWN)
378 set_equivalent_type(KMP_HW_LLC, KMP_HW_L2);
379 else if (get_equivalent_type(KMP_HW_TILE) != KMP_HW_UNKNOWN)
380 set_equivalent_type(KMP_HW_LLC, KMP_HW_TILE);
381 // L2/Tile wasn't detected so just say L1
382 else
383 set_equivalent_type(KMP_HW_LLC, KMP_HW_L1);
384 }
385#endif
386 else if (get_equivalent_type(KMP_HW_L1) != KMP_HW_UNKNOWN)
387 set_equivalent_type(KMP_HW_LLC, KMP_HW_L1);
388 // Fallback is to set last level cache to socket or core
389 if (get_equivalent_type(KMP_HW_LLC) == KMP_HW_UNKNOWN) {
390 if (get_equivalent_type(KMP_HW_SOCKET) != KMP_HW_UNKNOWN)
391 set_equivalent_type(KMP_HW_LLC, KMP_HW_SOCKET);
392 else if (get_equivalent_type(KMP_HW_CORE) != KMP_HW_UNKNOWN)
393 set_equivalent_type(KMP_HW_LLC, KMP_HW_CORE);
394 }
395 KMP_ASSERT(get_equivalent_type(KMP_HW_LLC) != KMP_HW_UNKNOWN)if (!(get_equivalent_type(KMP_HW_LLC) != KMP_HW_UNKNOWN)) { __kmp_debug_assert
("get_equivalent_type(KMP_HW_LLC) != KMP_HW_UNKNOWN", "openmp/runtime/src/kmp_affinity.cpp"
, 395); }
;
396}
397
398// Gather the count of each topology layer and the ratio
399void kmp_topology_t::_gather_enumeration_information() {
400 int previous_id[KMP_HW_LAST];
401 int max[KMP_HW_LAST];
402
403 for (int i = 0; i < depth; ++i) {
404 previous_id[i] = kmp_hw_thread_t::UNKNOWN_ID;
405 max[i] = 0;
406 count[i] = 0;
407 ratio[i] = 0;
408 }
409 int core_level = get_level(KMP_HW_CORE);
410 for (int i = 0; i < num_hw_threads; ++i) {
411 kmp_hw_thread_t &hw_thread = hw_threads[i];
412 for (int layer = 0; layer < depth; ++layer) {
413 int id = hw_thread.ids[layer];
414 if (id != previous_id[layer]) {
415 // Add an additional increment to each count
416 for (int l = layer; l < depth; ++l)
417 count[l]++;
418 // Keep track of topology layer ratio statistics
419 max[layer]++;
420 for (int l = layer + 1; l < depth; ++l) {
421 if (max[l] > ratio[l])
422 ratio[l] = max[l];
423 max[l] = 1;
424 }
425 // Figure out the number of different core types
426 // and efficiencies for hybrid CPUs
427 if (__kmp_is_hybrid_cpu() && core_level >= 0 && layer <= core_level) {
428 if (hw_thread.attrs.is_core_eff_valid() &&
429 hw_thread.attrs.core_eff >= num_core_efficiencies) {
430 // Because efficiencies can range from 0 to max efficiency - 1,
431 // the number of efficiencies is max efficiency + 1
432 num_core_efficiencies = hw_thread.attrs.core_eff + 1;
433 }
434 if (hw_thread.attrs.is_core_type_valid()) {
435 bool found = false;
436 for (int j = 0; j < num_core_types; ++j) {
437 if (hw_thread.attrs.get_core_type() == core_types[j]) {
438 found = true;
439 break;
440 }
441 }
442 if (!found) {
443 KMP_ASSERT(num_core_types < KMP_HW_MAX_NUM_CORE_TYPES)if (!(num_core_types < KMP_HW_MAX_NUM_CORE_TYPES)) { __kmp_debug_assert
("num_core_types < KMP_HW_MAX_NUM_CORE_TYPES", "openmp/runtime/src/kmp_affinity.cpp"
, 443); }
;
444 core_types[num_core_types++] = hw_thread.attrs.get_core_type();
445 }
446 }
447 }
448 break;
449 }
450 }
451 for (int layer = 0; layer < depth; ++layer) {
452 previous_id[layer] = hw_thread.ids[layer];
453 }
454 }
455 for (int layer = 0; layer < depth; ++layer) {
456 if (max[layer] > ratio[layer])
457 ratio[layer] = max[layer];
458 }
459}
460
461int kmp_topology_t::_get_ncores_with_attr(const kmp_hw_attr_t &attr,
462 int above_level,
463 bool find_all) const {
464 int current, current_max;
465 int previous_id[KMP_HW_LAST];
466 for (int i = 0; i < depth; ++i)
467 previous_id[i] = kmp_hw_thread_t::UNKNOWN_ID;
468 int core_level = get_level(KMP_HW_CORE);
469 if (find_all)
470 above_level = -1;
471 KMP_ASSERT(above_level < core_level)if (!(above_level < core_level)) { __kmp_debug_assert("above_level < core_level"
, "openmp/runtime/src/kmp_affinity.cpp", 471); }
;
472 current_max = 0;
473 current = 0;
474 for (int i = 0; i < num_hw_threads; ++i) {
475 kmp_hw_thread_t &hw_thread = hw_threads[i];
476 if (!find_all && hw_thread.ids[above_level] != previous_id[above_level]) {
477 if (current > current_max)
478 current_max = current;
479 current = hw_thread.attrs.contains(attr);
480 } else {
481 for (int level = above_level + 1; level <= core_level; ++level) {
482 if (hw_thread.ids[level] != previous_id[level]) {
483 if (hw_thread.attrs.contains(attr))
484 current++;
485 break;
486 }
487 }
488 }
489 for (int level = 0; level < depth; ++level)
490 previous_id[level] = hw_thread.ids[level];
491 }
492 if (current > current_max)
493 current_max = current;
494 return current_max;
495}
496
497// Find out if the topology is uniform
498void kmp_topology_t::_discover_uniformity() {
499 int num = 1;
500 for (int level = 0; level < depth; ++level)
501 num *= ratio[level];
502 flags.uniform = (num == count[depth - 1]);
503}
504
505// Set all the sub_ids for each hardware thread
506void kmp_topology_t::_set_sub_ids() {
507 int previous_id[KMP_HW_LAST];
508 int sub_id[KMP_HW_LAST];
509
510 for (int i = 0; i < depth; ++i) {
511 previous_id[i] = -1;
512 sub_id[i] = -1;
513 }
514 for (int i = 0; i < num_hw_threads; ++i) {
515 kmp_hw_thread_t &hw_thread = hw_threads[i];
516 // Setup the sub_id
517 for (int j = 0; j < depth; ++j) {
518 if (hw_thread.ids[j] != previous_id[j]) {
519 sub_id[j]++;
520 for (int k = j + 1; k < depth; ++k) {
521 sub_id[k] = 0;
522 }
523 break;
524 }
525 }
526 // Set previous_id
527 for (int j = 0; j < depth; ++j) {
528 previous_id[j] = hw_thread.ids[j];
529 }
530 // Set the sub_ids field
531 for (int j = 0; j < depth; ++j) {
532 hw_thread.sub_ids[j] = sub_id[j];
533 }
534 }
535}
536
537void kmp_topology_t::_set_globals() {
538 // Set nCoresPerPkg, nPackages, __kmp_nThreadsPerCore, __kmp_ncores
539 int core_level, thread_level, package_level;
540 package_level = get_level(KMP_HW_SOCKET);
541#if KMP_GROUP_AFFINITY0
542 if (package_level == -1)
543 package_level = get_level(KMP_HW_PROC_GROUP);
544#endif
545 core_level = get_level(KMP_HW_CORE);
546 thread_level = get_level(KMP_HW_THREAD);
547
548 KMP_ASSERT(core_level != -1)if (!(core_level != -1)) { __kmp_debug_assert("core_level != -1"
, "openmp/runtime/src/kmp_affinity.cpp", 548); }
;
549 KMP_ASSERT(thread_level != -1)if (!(thread_level != -1)) { __kmp_debug_assert("thread_level != -1"
, "openmp/runtime/src/kmp_affinity.cpp", 549); }
;
550
551 __kmp_nThreadsPerCore = calculate_ratio(thread_level, core_level);
552 if (package_level != -1) {
553 nCoresPerPkg = calculate_ratio(core_level, package_level);
554 nPackages = get_count(package_level);
555 } else {
556 // assume one socket
557 nCoresPerPkg = get_count(core_level);
558 nPackages = 1;
559 }
560#ifndef KMP_DFLT_NTH_CORES
561 __kmp_ncores = get_count(core_level);
562#endif
563}
564
565kmp_topology_t *kmp_topology_t::allocate(int nproc, int ndepth,
566 const kmp_hw_t *types) {
567 kmp_topology_t *retval;
568 // Allocate all data in one large allocation
569 size_t size = sizeof(kmp_topology_t) + sizeof(kmp_hw_thread_t) * nproc +
570 sizeof(int) * (size_t)KMP_HW_LAST * 3;
571 char *bytes = (char *)__kmp_allocate(size)___kmp_allocate((size), "openmp/runtime/src/kmp_affinity.cpp"
, 571)
;
572 retval = (kmp_topology_t *)bytes;
573 if (nproc > 0) {
574 retval->hw_threads = (kmp_hw_thread_t *)(bytes + sizeof(kmp_topology_t));
575 } else {
576 retval->hw_threads = nullptr;
577 }
578 retval->num_hw_threads = nproc;
579 retval->depth = ndepth;
580 int *arr =
581 (int *)(bytes + sizeof(kmp_topology_t) + sizeof(kmp_hw_thread_t) * nproc);
582 retval->types = (kmp_hw_t *)arr;
583 retval->ratio = arr + (size_t)KMP_HW_LAST;
584 retval->count = arr + 2 * (size_t)KMP_HW_LAST;
585 retval->num_core_efficiencies = 0;
586 retval->num_core_types = 0;
587 retval->compact = 0;
588 for (int i = 0; i < KMP_HW_MAX_NUM_CORE_TYPES; ++i)
589 retval->core_types[i] = KMP_HW_CORE_TYPE_UNKNOWN;
590 KMP_FOREACH_HW_TYPE(type)for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST; type
= (kmp_hw_t)((int)type + 1))
{ retval->equivalent[type] = KMP_HW_UNKNOWN; }
591 for (int i = 0; i < ndepth; ++i) {
592 retval->types[i] = types[i];
593 retval->equivalent[types[i]] = types[i];
594 }
595 return retval;
596}
597
598void kmp_topology_t::deallocate(kmp_topology_t *topology) {
599 if (topology)
600 __kmp_free(topology)___kmp_free((topology), "openmp/runtime/src/kmp_affinity.cpp"
, 600)
;
601}
602
603bool kmp_topology_t::check_ids() const {
604 // Assume ids have been sorted
605 if (num_hw_threads == 0)
606 return true;
607 for (int i = 1; i < num_hw_threads; ++i) {
608 kmp_hw_thread_t &current_thread = hw_threads[i];
609 kmp_hw_thread_t &previous_thread = hw_threads[i - 1];
610 bool unique = false;
611 for (int j = 0; j < depth; ++j) {
612 if (previous_thread.ids[j] != current_thread.ids[j]) {
613 unique = true;
614 break;
615 }
616 }
617 if (unique)
618 continue;
619 return false;
620 }
621 return true;
622}
623
624void kmp_topology_t::dump() const {
625 printf("***********************\n");
626 printf("*** __kmp_topology: ***\n");
627 printf("***********************\n");
628 printf("* depth: %d\n", depth);
629
630 printf("* types: ");
631 for (int i = 0; i < depth; ++i)
632 printf("%15s ", __kmp_hw_get_keyword(types[i]));
633 printf("\n");
634
635 printf("* ratio: ");
636 for (int i = 0; i < depth; ++i) {
637 printf("%15d ", ratio[i]);
638 }
639 printf("\n");
640
641 printf("* count: ");
642 for (int i = 0; i < depth; ++i) {
643 printf("%15d ", count[i]);
644 }
645 printf("\n");
646
647 printf("* num_core_eff: %d\n", num_core_efficiencies);
648 printf("* num_core_types: %d\n", num_core_types);
649 printf("* core_types: ");
650 for (int i = 0; i < num_core_types; ++i)
651 printf("%3d ", core_types[i]);
652 printf("\n");
653
654 printf("* equivalent map:\n");
655 KMP_FOREACH_HW_TYPE(i)for (kmp_hw_t i = (kmp_hw_t)0; i < KMP_HW_LAST; i = (kmp_hw_t
)((int)i + 1))
{
656 const char *key = __kmp_hw_get_keyword(i);
657 const char *value = __kmp_hw_get_keyword(equivalent[i]);
658 printf("%-15s -> %-15s\n", key, value);
659 }
660
661 printf("* uniform: %s\n", (is_uniform() ? "Yes" : "No"));
662
663 printf("* num_hw_threads: %d\n", num_hw_threads);
664 printf("* hw_threads:\n");
665 for (int i = 0; i < num_hw_threads; ++i) {
666 hw_threads[i].print();
667 }
668 printf("***********************\n");
669}
670
671void kmp_topology_t::print(const char *env_var) const {
672 kmp_str_buf_t buf;
673 int print_types_depth;
674 __kmp_str_buf_init(&buf){ (&buf)->str = (&buf)->bulk; (&buf)->size
= sizeof((&buf)->bulk); (&buf)->used = 0; (&
buf)->bulk[0] = 0; }
;
675 kmp_hw_t print_types[KMP_HW_LAST + 2];
676
677 // Num Available Threads
678 if (num_hw_threads) {
679 KMP_INFORM(AvailableOSProc, env_var, num_hw_threads)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_AvailableOSProc
, env_var, num_hw_threads), __kmp_msg_null)
;
680 } else {
681 KMP_INFORM(AvailableOSProc, env_var, __kmp_xproc)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_AvailableOSProc
, env_var, __kmp_xproc), __kmp_msg_null)
;
682 }
683
684 // Uniform or not
685 if (is_uniform()) {
686 KMP_INFORM(Uniform, env_var)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_Uniform
, env_var), __kmp_msg_null)
;
687 } else {
688 KMP_INFORM(NonUniform, env_var)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_NonUniform
, env_var), __kmp_msg_null)
;
689 }
690
691 // Equivalent types
692 KMP_FOREACH_HW_TYPE(type)for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST; type
= (kmp_hw_t)((int)type + 1))
{
693 kmp_hw_t eq_type = equivalent[type];
694 if (eq_type != KMP_HW_UNKNOWN && eq_type != type) {
695 KMP_INFORM(AffEqualTopologyTypes, env_var,__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_AffEqualTopologyTypes
, env_var, __kmp_hw_get_catalog_string(type), __kmp_hw_get_catalog_string
(eq_type)), __kmp_msg_null)
696 __kmp_hw_get_catalog_string(type),__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_AffEqualTopologyTypes
, env_var, __kmp_hw_get_catalog_string(type), __kmp_hw_get_catalog_string
(eq_type)), __kmp_msg_null)
697 __kmp_hw_get_catalog_string(eq_type))__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_AffEqualTopologyTypes
, env_var, __kmp_hw_get_catalog_string(type), __kmp_hw_get_catalog_string
(eq_type)), __kmp_msg_null)
;
698 }
699 }
700
701 // Quick topology
702 KMP_ASSERT(depth > 0 && depth <= (int)KMP_HW_LAST)if (!(depth > 0 && depth <= (int)KMP_HW_LAST)) {
__kmp_debug_assert("depth > 0 && depth <= (int)KMP_HW_LAST"
, "openmp/runtime/src/kmp_affinity.cpp", 702); }
;
703 // Create a print types array that always guarantees printing
704 // the core and thread level
705 print_types_depth = 0;
706 for (int level = 0; level < depth; ++level)
707 print_types[print_types_depth++] = types[level];
708 if (equivalent[KMP_HW_CORE] != KMP_HW_CORE) {
709 // Force in the core level for quick topology
710 if (print_types[print_types_depth - 1] == KMP_HW_THREAD) {
711 // Force core before thread e.g., 1 socket X 2 threads/socket
712 // becomes 1 socket X 1 core/socket X 2 threads/socket
713 print_types[print_types_depth - 1] = KMP_HW_CORE;
714 print_types[print_types_depth++] = KMP_HW_THREAD;
715 } else {
716 print_types[print_types_depth++] = KMP_HW_CORE;
717 }
718 }
719 // Always put threads at very end of quick topology
720 if (equivalent[KMP_HW_THREAD] != KMP_HW_THREAD)
721 print_types[print_types_depth++] = KMP_HW_THREAD;
722
723 __kmp_str_buf_clear(&buf);
724 kmp_hw_t numerator_type;
725 kmp_hw_t denominator_type = KMP_HW_UNKNOWN;
726 int core_level = get_level(KMP_HW_CORE);
727 int ncores = get_count(core_level);
728
729 for (int plevel = 0, level = 0; plevel < print_types_depth; ++plevel) {
730 int c;
731 bool plural;
732 numerator_type = print_types[plevel];
733 KMP_ASSERT_VALID_HW_TYPE(numerator_type)if (!(numerator_type >= (kmp_hw_t)0 && numerator_type
< KMP_HW_LAST)) { __kmp_debug_assert("numerator_type >= (kmp_hw_t)0 && numerator_type < KMP_HW_LAST"
, "openmp/runtime/src/kmp_affinity.cpp", 733); }
;
734 if (equivalent[numerator_type] != numerator_type)
735 c = 1;
736 else
737 c = get_ratio(level++);
738 plural = (c > 1);
739 if (plevel == 0) {
740 __kmp_str_buf_print(&buf, "%d %s", c,
741 __kmp_hw_get_catalog_string(numerator_type, plural));
742 } else {
743 __kmp_str_buf_print(&buf, " x %d %s/%s", c,
744 __kmp_hw_get_catalog_string(numerator_type, plural),
745 __kmp_hw_get_catalog_string(denominator_type));
746 }
747 denominator_type = numerator_type;
748 }
749 KMP_INFORM(TopologyGeneric, env_var, buf.str, ncores)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_TopologyGeneric
, env_var, buf.str, ncores), __kmp_msg_null)
;
750
751 // Hybrid topology information
752 if (__kmp_is_hybrid_cpu()) {
753 for (int i = 0; i < num_core_types; ++i) {
754 kmp_hw_core_type_t core_type = core_types[i];
755 kmp_hw_attr_t attr;
756 attr.clear();
757 attr.set_core_type(core_type);
758 int ncores = get_ncores_with_attr(attr);
759 if (ncores > 0) {
760 KMP_INFORM(TopologyHybrid, env_var, ncores,__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_TopologyHybrid
, env_var, ncores, __kmp_hw_get_core_type_string(core_type)),
__kmp_msg_null)
761 __kmp_hw_get_core_type_string(core_type))__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_TopologyHybrid
, env_var, ncores, __kmp_hw_get_core_type_string(core_type)),
__kmp_msg_null)
;
762 KMP_ASSERT(num_core_efficiencies <= KMP_HW_MAX_NUM_CORE_EFFS)if (!(num_core_efficiencies <= 8)) { __kmp_debug_assert("num_core_efficiencies <= KMP_HW_MAX_NUM_CORE_EFFS"
, "openmp/runtime/src/kmp_affinity.cpp", 762); }
763 for (int eff = 0; eff < num_core_efficiencies; ++eff) {
764 attr.set_core_eff(eff);
765 int ncores_with_eff = get_ncores_with_attr(attr);
766 if (ncores_with_eff > 0) {
767 KMP_INFORM(TopologyHybridCoreEff, env_var, ncores_with_eff, eff)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_TopologyHybridCoreEff
, env_var, ncores_with_eff, eff), __kmp_msg_null)
;
768 }
769 }
770 }
771 }
772 }
773
774 if (num_hw_threads <= 0) {
775 __kmp_str_buf_free(&buf);
776 return;
777 }
778
779 // Full OS proc to hardware thread map
780 KMP_INFORM(OSProcToPhysicalThreadMap, env_var)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_OSProcToPhysicalThreadMap
, env_var), __kmp_msg_null)
;
781 for (int i = 0; i < num_hw_threads; i++) {
782 __kmp_str_buf_clear(&buf);
783 for (int level = 0; level < depth; ++level) {
784 kmp_hw_t type = types[level];
785 __kmp_str_buf_print(&buf, "%s ", __kmp_hw_get_catalog_string(type));
786 __kmp_str_buf_print(&buf, "%d ", hw_threads[i].ids[level]);
787 }
788 if (__kmp_is_hybrid_cpu())
789 __kmp_str_buf_print(
790 &buf, "(%s)",
791 __kmp_hw_get_core_type_string(hw_threads[i].attrs.get_core_type()));
792 KMP_INFORM(OSProcMapToPack, env_var, hw_threads[i].os_id, buf.str)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_OSProcMapToPack
, env_var, hw_threads[i].os_id, buf.str), __kmp_msg_null)
;
793 }
794
795 __kmp_str_buf_free(&buf);
796}
797
798#if KMP_AFFINITY_SUPPORTED1
799void kmp_topology_t::set_granularity(kmp_affinity_t &affinity) const {
800 const char *env_var = affinity.env_var;
801 // Set the number of affinity granularity levels
802 if (affinity.gran_levels < 0) {
803 kmp_hw_t gran_type = get_equivalent_type(affinity.gran);
804 // Check if user's granularity request is valid
805 if (gran_type == KMP_HW_UNKNOWN) {
806 // First try core, then thread, then package
807 kmp_hw_t gran_types[3] = {KMP_HW_CORE, KMP_HW_THREAD, KMP_HW_SOCKET};
808 for (auto g : gran_types) {
809 if (get_equivalent_type(g) != KMP_HW_UNKNOWN) {
810 gran_type = g;
811 break;
812 }
813 }
814 KMP_ASSERT(gran_type != KMP_HW_UNKNOWN)if (!(gran_type != KMP_HW_UNKNOWN)) { __kmp_debug_assert("gran_type != KMP_HW_UNKNOWN"
, "openmp/runtime/src/kmp_affinity.cpp", 814); }
;
815 // Warn user what granularity setting will be used instead
816 KMP_AFF_WARNING(affinity, AffGranularityBad, env_var,if (affinity.flags.verbose || (affinity.flags.warnings &&
(affinity.type != affinity_none))) { __kmp_msg(kmp_ms_warning
, __kmp_msg_format(kmp_i18n_msg_AffGranularityBad, env_var, __kmp_hw_get_catalog_string
(affinity.gran), __kmp_hw_get_catalog_string(gran_type)), __kmp_msg_null
); }
817 __kmp_hw_get_catalog_string(affinity.gran),if (affinity.flags.verbose || (affinity.flags.warnings &&
(affinity.type != affinity_none))) { __kmp_msg(kmp_ms_warning
, __kmp_msg_format(kmp_i18n_msg_AffGranularityBad, env_var, __kmp_hw_get_catalog_string
(affinity.gran), __kmp_hw_get_catalog_string(gran_type)), __kmp_msg_null
); }
818 __kmp_hw_get_catalog_string(gran_type))if (affinity.flags.verbose || (affinity.flags.warnings &&
(affinity.type != affinity_none))) { __kmp_msg(kmp_ms_warning
, __kmp_msg_format(kmp_i18n_msg_AffGranularityBad, env_var, __kmp_hw_get_catalog_string
(affinity.gran), __kmp_hw_get_catalog_string(gran_type)), __kmp_msg_null
); }
;
819 affinity.gran = gran_type;
820 }
821#if KMP_GROUP_AFFINITY0
822 // If more than one processor group exists, and the level of
823 // granularity specified by the user is too coarse, then the
824 // granularity must be adjusted "down" to processor group affinity
825 // because threads can only exist within one processor group.
826 // For example, if a user sets granularity=socket and there are two
827 // processor groups that cover a socket, then the runtime must
828 // restrict the granularity down to the processor group level.
829 if (__kmp_num_proc_groups > 1) {
830 int gran_depth = get_level(gran_type);
831 int proc_group_depth = get_level(KMP_HW_PROC_GROUP);
832 if (gran_depth >= 0 && proc_group_depth >= 0 &&
833 gran_depth < proc_group_depth) {
834 KMP_AFF_WARNING(affinity, AffGranTooCoarseProcGroup, env_var,if (affinity.flags.verbose || (affinity.flags.warnings &&
(affinity.type != affinity_none))) { __kmp_msg(kmp_ms_warning
, __kmp_msg_format(kmp_i18n_msg_AffGranTooCoarseProcGroup, env_var
, __kmp_hw_get_catalog_string(affinity.gran)), __kmp_msg_null
); }
835 __kmp_hw_get_catalog_string(affinity.gran))if (affinity.flags.verbose || (affinity.flags.warnings &&
(affinity.type != affinity_none))) { __kmp_msg(kmp_ms_warning
, __kmp_msg_format(kmp_i18n_msg_AffGranTooCoarseProcGroup, env_var
, __kmp_hw_get_catalog_string(affinity.gran)), __kmp_msg_null
); }
;
836 affinity.gran = gran_type = KMP_HW_PROC_GROUP;
837 }
838 }
839#endif
840 affinity.gran_levels = 0;
841 for (int i = depth - 1; i >= 0 && get_type(i) != gran_type; --i)
842 affinity.gran_levels++;
843 }
844}
845#endif
846
847void kmp_topology_t::canonicalize() {
848#if KMP_GROUP_AFFINITY0
849 _insert_windows_proc_groups();
850#endif
851 _remove_radix1_layers();
852 _gather_enumeration_information();
853 _discover_uniformity();
854 _set_sub_ids();
855 _set_globals();
856 _set_last_level_cache();
857
858#if KMP_MIC_SUPPORTED((0 || 1) && (1 || 0))
859 // Manually Add L2 = Tile equivalence
860 if (__kmp_mic_type == mic3) {
861 if (get_level(KMP_HW_L2) != -1)
862 set_equivalent_type(KMP_HW_TILE, KMP_HW_L2);
863 else if (get_level(KMP_HW_TILE) != -1)
864 set_equivalent_type(KMP_HW_L2, KMP_HW_TILE);
865 }
866#endif
867
868 // Perform post canonicalization checking
869 KMP_ASSERT(depth > 0)if (!(depth > 0)) { __kmp_debug_assert("depth > 0", "openmp/runtime/src/kmp_affinity.cpp"
, 869); }
;
870 for (int level = 0; level < depth; ++level) {
871 // All counts, ratios, and types must be valid
872 KMP_ASSERT(count[level] > 0 && ratio[level] > 0)if (!(count[level] > 0 && ratio[level] > 0)) { __kmp_debug_assert
("count[level] > 0 && ratio[level] > 0", "openmp/runtime/src/kmp_affinity.cpp"
, 872); }
;
873 KMP_ASSERT_VALID_HW_TYPE(types[level])if (!(types[level] >= (kmp_hw_t)0 && types[level] <
KMP_HW_LAST)) { __kmp_debug_assert("types[level] >= (kmp_hw_t)0 && types[level] < KMP_HW_LAST"
, "openmp/runtime/src/kmp_affinity.cpp", 873); }
;
874 // Detected types must point to themselves
875 KMP_ASSERT(equivalent[types[level]] == types[level])if (!(equivalent[types[level]] == types[level])) { __kmp_debug_assert
("equivalent[types[level]] == types[level]", "openmp/runtime/src/kmp_affinity.cpp"
, 875); }
;
876 }
877}
878
879// Canonicalize an explicit packages X cores/pkg X threads/core topology
880void kmp_topology_t::canonicalize(int npackages, int ncores_per_pkg,
881 int nthreads_per_core, int ncores) {
882 int ndepth = 3;
883 depth = ndepth;
884 KMP_FOREACH_HW_TYPE(i)for (kmp_hw_t i = (kmp_hw_t)0; i < KMP_HW_LAST; i = (kmp_hw_t
)((int)i + 1))
{ equivalent[i] = KMP_HW_UNKNOWN; }
885 for (int level = 0; level < depth; ++level) {
886 count[level] = 0;
887 ratio[level] = 0;
888 }
889 count[0] = npackages;
890 count[1] = ncores;
891 count[2] = __kmp_xproc;
892 ratio[0] = npackages;
893 ratio[1] = ncores_per_pkg;
894 ratio[2] = nthreads_per_core;
895 equivalent[KMP_HW_SOCKET] = KMP_HW_SOCKET;
896 equivalent[KMP_HW_CORE] = KMP_HW_CORE;
897 equivalent[KMP_HW_THREAD] = KMP_HW_THREAD;
898 types[0] = KMP_HW_SOCKET;
899 types[1] = KMP_HW_CORE;
900 types[2] = KMP_HW_THREAD;
901 //__kmp_avail_proc = __kmp_xproc;
902 _discover_uniformity();
903}
904
905// Represents running sub IDs for a single core attribute where
906// attribute values have SIZE possibilities.
907template <size_t SIZE, typename IndexFunc> struct kmp_sub_ids_t {
908 int last_level; // last level in topology to consider for sub_ids
909 int sub_id[SIZE]; // The sub ID for a given attribute value
910 int prev_sub_id[KMP_HW_LAST];
911 IndexFunc indexer;
912
913public:
914 kmp_sub_ids_t(int last_level) : last_level(last_level) {
915 KMP_ASSERT(last_level < KMP_HW_LAST)if (!(last_level < KMP_HW_LAST)) { __kmp_debug_assert("last_level < KMP_HW_LAST"
, "openmp/runtime/src/kmp_affinity.cpp", 915); }
;
916 for (size_t i = 0; i < SIZE; ++i)
917 sub_id[i] = -1;
918 for (size_t i = 0; i < KMP_HW_LAST; ++i)
919 prev_sub_id[i] = -1;
920 }
921 void update(const kmp_hw_thread_t &hw_thread) {
922 int idx = indexer(hw_thread);
923 KMP_ASSERT(idx < (int)SIZE)if (!(idx < (int)SIZE)) { __kmp_debug_assert("idx < (int)SIZE"
, "openmp/runtime/src/kmp_affinity.cpp", 923); }
;
924 for (int level = 0; level <= last_level; ++level) {
925 if (hw_thread.sub_ids[level] != prev_sub_id[level]) {
926 if (level < last_level)
927 sub_id[idx] = -1;
928 sub_id[idx]++;
929 break;
930 }
931 }
932 for (int level = 0; level <= last_level; ++level)
933 prev_sub_id[level] = hw_thread.sub_ids[level];
934 }
935 int get_sub_id(const kmp_hw_thread_t &hw_thread) const {
936 return sub_id[indexer(hw_thread)];
937 }
938};
939
940static kmp_str_buf_t *
941__kmp_hw_get_catalog_core_string(const kmp_hw_attr_t &attr, kmp_str_buf_t *buf,
942 bool plural) {
943 __kmp_str_buf_init(buf){ (buf)->str = (buf)->bulk; (buf)->size = sizeof((buf
)->bulk); (buf)->used = 0; (buf)->bulk[0] = 0; }
;
944 if (attr.is_core_type_valid())
945 __kmp_str_buf_print(buf, "%s %s",
946 __kmp_hw_get_core_type_string(attr.get_core_type()),
947 __kmp_hw_get_catalog_string(KMP_HW_CORE, plural));
948 else
949 __kmp_str_buf_print(buf, "%s eff=%d",
950 __kmp_hw_get_catalog_string(KMP_HW_CORE, plural),
951 attr.get_core_eff());
952 return buf;
953}
954
955// Apply the KMP_HW_SUBSET envirable to the topology
956// Returns true if KMP_HW_SUBSET filtered any processors
957// otherwise, returns false
958bool kmp_topology_t::filter_hw_subset() {
959 // If KMP_HW_SUBSET wasn't requested, then do nothing.
960 if (!__kmp_hw_subset)
961 return false;
962
963 // First, sort the KMP_HW_SUBSET items by the machine topology
964 __kmp_hw_subset->sort();
965
966 // Check to see if KMP_HW_SUBSET is a valid subset of the detected topology
967 bool using_core_types = false;
968 bool using_core_effs = false;
969 int hw_subset_depth = __kmp_hw_subset->get_depth();
970 kmp_hw_t specified[KMP_HW_LAST];
971 int *topology_levels = (int *)KMP_ALLOCA(sizeof(int) * hw_subset_depth)__builtin_alloca (sizeof(int) * hw_subset_depth);
972 KMP_ASSERT(hw_subset_depth > 0)if (!(hw_subset_depth > 0)) { __kmp_debug_assert("hw_subset_depth > 0"
, "openmp/runtime/src/kmp_affinity.cpp", 972); }
;
973 KMP_FOREACH_HW_TYPE(i)for (kmp_hw_t i = (kmp_hw_t)0; i < KMP_HW_LAST; i = (kmp_hw_t
)((int)i + 1))
{ specified[i] = KMP_HW_UNKNOWN; }
974 int core_level = get_level(KMP_HW_CORE);
975 for (int i = 0; i < hw_subset_depth; ++i) {
976 int max_count;
977 const kmp_hw_subset_t::item_t &item = __kmp_hw_subset->at(i);
978 int num = item.num[0];
979 int offset = item.offset[0];
980 kmp_hw_t type = item.type;
981 kmp_hw_t equivalent_type = equivalent[type];
982 int level = get_level(type);
983 topology_levels[i] = level;
984
985 // Check to see if current layer is in detected machine topology
986 if (equivalent_type != KMP_HW_UNKNOWN) {
987 __kmp_hw_subset->at(i).type = equivalent_type;
988 } else {
989 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetNotExistGeneric,if (__kmp_affinity.flags.verbose || (__kmp_affinity.flags.warnings
&& (__kmp_affinity.type != affinity_none))) { __kmp_msg
(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_AffHWSubsetNotExistGeneric
, __kmp_hw_get_catalog_string(type)), __kmp_msg_null); }
990 __kmp_hw_get_catalog_string(type))if (__kmp_affinity.flags.verbose || (__kmp_affinity.flags.warnings
&& (__kmp_affinity.type != affinity_none))) { __kmp_msg
(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_AffHWSubsetNotExistGeneric
, __kmp_hw_get_catalog_string(type)), __kmp_msg_null); }
;
991 return false;
992 }
993
994 // Check to see if current layer has already been
995 // specified either directly or through an equivalent type
996 if (specified[equivalent_type] != KMP_HW_UNKNOWN) {
997 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetEqvLayers,if (__kmp_affinity.flags.verbose || (__kmp_affinity.flags.warnings
&& (__kmp_affinity.type != affinity_none))) { __kmp_msg
(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_AffHWSubsetEqvLayers
, __kmp_hw_get_catalog_string(type), __kmp_hw_get_catalog_string
(specified[equivalent_type])), __kmp_msg_null); }
998 __kmp_hw_get_catalog_string(type),if (__kmp_affinity.flags.verbose || (__kmp_affinity.flags.warnings
&& (__kmp_affinity.type != affinity_none))) { __kmp_msg
(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_AffHWSubsetEqvLayers
, __kmp_hw_get_catalog_string(type), __kmp_hw_get_catalog_string
(specified[equivalent_type])), __kmp_msg_null); }
999 __kmp_hw_get_catalog_string(specified[equivalent_type]))if (__kmp_affinity.flags.verbose || (__kmp_affinity.flags.warnings
&& (__kmp_affinity.type != affinity_none))) { __kmp_msg
(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_AffHWSubsetEqvLayers
, __kmp_hw_get_catalog_string(type), __kmp_hw_get_catalog_string
(specified[equivalent_type])), __kmp_msg_null); }
;
1000 return false;
1001 }
1002 specified[equivalent_type] = type;
1003
1004 // Check to see if each layer's num & offset parameters are valid
1005 max_count = get_ratio(level);
1006 if (max_count < 0 ||
1007 (num != kmp_hw_subset_t::USE_ALL && num + offset > max_count)) {
1008 bool plural = (num > 1);
1009 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetManyGeneric,if (__kmp_affinity.flags.verbose || (__kmp_affinity.flags.warnings
&& (__kmp_affinity.type != affinity_none))) { __kmp_msg
(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_AffHWSubsetManyGeneric
, __kmp_hw_get_catalog_string(type, plural)), __kmp_msg_null)
; }
1010 __kmp_hw_get_catalog_string(type, plural))if (__kmp_affinity.flags.verbose || (__kmp_affinity.flags.warnings
&& (__kmp_affinity.type != affinity_none))) { __kmp_msg
(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_AffHWSubsetManyGeneric
, __kmp_hw_get_catalog_string(type, plural)), __kmp_msg_null)
; }
;
1011 return false;
1012 }
1013
1014 // Check to see if core attributes are consistent
1015 if (core_level == level) {
1016 // Determine which core attributes are specified
1017 for (int j = 0; j < item.num_attrs; ++j) {
1018 if (item.attr[j].is_core_type_valid())
1019 using_core_types = true;
1020 if (item.attr[j].is_core_eff_valid())
1021 using_core_effs = true;
1022 }
1023
1024 // Check if using a single core attribute on non-hybrid arch.
1025 // Do not ignore all of KMP_HW_SUBSET, just ignore the attribute.
1026 //
1027 // Check if using multiple core attributes on non-hyrbid arch.
1028 // Ignore all of KMP_HW_SUBSET if this is the case.
1029 if ((using_core_effs || using_core_types) && !__kmp_is_hybrid_cpu()) {
1030 if (item.num_attrs == 1) {
1031 if (using_core_effs) {
1032 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetIgnoringAttr,if (__kmp_affinity.flags.verbose || (__kmp_affinity.flags.warnings
&& (__kmp_affinity.type != affinity_none))) { __kmp_msg
(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_AffHWSubsetIgnoringAttr
, "efficiency"), __kmp_msg_null); }
1033 "efficiency")if (__kmp_affinity.flags.verbose || (__kmp_affinity.flags.warnings
&& (__kmp_affinity.type != affinity_none))) { __kmp_msg
(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_AffHWSubsetIgnoringAttr
, "efficiency"), __kmp_msg_null); }
;
1034 } else {
1035 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetIgnoringAttr,if (__kmp_affinity.flags.verbose || (__kmp_affinity.flags.warnings
&& (__kmp_affinity.type != affinity_none))) { __kmp_msg
(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_AffHWSubsetIgnoringAttr
, "core_type"), __kmp_msg_null); }
1036 "core_type")if (__kmp_affinity.flags.verbose || (__kmp_affinity.flags.warnings
&& (__kmp_affinity.type != affinity_none))) { __kmp_msg
(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_AffHWSubsetIgnoringAttr
, "core_type"), __kmp_msg_null); }
;
1037 }
1038 using_core_effs = false;
1039 using_core_types = false;
1040 } else {
1041 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetAttrsNonHybrid)if (__kmp_affinity.flags.verbose || (__kmp_affinity.flags.warnings
&& (__kmp_affinity.type != affinity_none))) { __kmp_msg
(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_AffHWSubsetAttrsNonHybrid
), __kmp_msg_null); }
;
1042 return false;
1043 }
1044 }
1045
1046 // Check if using both core types and core efficiencies together
1047 if (using_core_types && using_core_effs) {
1048 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetIncompat, "core_type",if (__kmp_affinity.flags.verbose || (__kmp_affinity.flags.warnings
&& (__kmp_affinity.type != affinity_none))) { __kmp_msg
(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_AffHWSubsetIncompat
, "core_type", "efficiency"), __kmp_msg_null); }
1049 "efficiency")if (__kmp_affinity.flags.verbose || (__kmp_affinity.flags.warnings
&& (__kmp_affinity.type != affinity_none))) { __kmp_msg
(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_AffHWSubsetIncompat
, "core_type", "efficiency"), __kmp_msg_null); }
;
1050 return false;
1051 }
1052
1053 // Check that core efficiency values are valid
1054 if (using_core_effs) {
1055 for (int j = 0; j < item.num_attrs; ++j) {
1056 if (item.attr[j].is_core_eff_valid()) {
1057 int core_eff = item.attr[j].get_core_eff();
1058 if (core_eff < 0 || core_eff >= num_core_efficiencies) {
1059 kmp_str_buf_t buf;
1060 __kmp_str_buf_init(&buf){ (&buf)->str = (&buf)->bulk; (&buf)->size
= sizeof((&buf)->bulk); (&buf)->used = 0; (&
buf)->bulk[0] = 0; }
;
1061 __kmp_str_buf_print(&buf, "%d", item.attr[j].get_core_eff());
1062 __kmp_msg(kmp_ms_warning,
1063 KMP_MSG(AffHWSubsetAttrInvalid, "efficiency", buf.str)__kmp_msg_format(kmp_i18n_msg_AffHWSubsetAttrInvalid, "efficiency"
, buf.str)
,
1064 KMP_HNT(ValidValuesRange, 0, num_core_efficiencies - 1)__kmp_msg_format(kmp_i18n_hnt_ValidValuesRange, 0, num_core_efficiencies
- 1)
,
1065 __kmp_msg_null);
1066 __kmp_str_buf_free(&buf);
1067 return false;
1068 }
1069 }
1070 }
1071 }
1072
1073 // Check that the number of requested cores with attributes is valid
1074 if (using_core_types || using_core_effs) {
1075 for (int j = 0; j < item.num_attrs; ++j) {
1076 int num = item.num[j];
1077 int offset = item.offset[j];
1078 int level_above = core_level - 1;
1079 if (level_above >= 0) {
1080 max_count = get_ncores_with_attr_per(item.attr[j], level_above);
1081 if (max_count <= 0 ||
1082 (num != kmp_hw_subset_t::USE_ALL && num + offset > max_count)) {
1083 kmp_str_buf_t buf;
1084 __kmp_hw_get_catalog_core_string(item.attr[j], &buf, num > 0);
1085 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetManyGeneric, buf.str)if (__kmp_affinity.flags.verbose || (__kmp_affinity.flags.warnings
&& (__kmp_affinity.type != affinity_none))) { __kmp_msg
(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_AffHWSubsetManyGeneric
, buf.str), __kmp_msg_null); }
;
1086 __kmp_str_buf_free(&buf);
1087 return false;
1088 }
1089 }
1090 }
1091 }
1092
1093 if ((using_core_types || using_core_effs) && item.num_attrs > 1) {
1094 for (int j = 0; j < item.num_attrs; ++j) {
1095 // Ambiguous use of specific core attribute + generic core
1096 // e.g., 4c & 3c:intel_core or 4c & 3c:eff1
1097 if (!item.attr[j]) {
1098 kmp_hw_attr_t other_attr;
1099 for (int k = 0; k < item.num_attrs; ++k) {
1100 if (item.attr[k] != item.attr[j]) {
1101 other_attr = item.attr[k];
1102 break;
1103 }
1104 }
1105 kmp_str_buf_t buf;
1106 __kmp_hw_get_catalog_core_string(other_attr, &buf, item.num[j] > 0);
1107 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetIncompat,if (__kmp_affinity.flags.verbose || (__kmp_affinity.flags.warnings
&& (__kmp_affinity.type != affinity_none))) { __kmp_msg
(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_AffHWSubsetIncompat
, __kmp_hw_get_catalog_string(KMP_HW_CORE), buf.str), __kmp_msg_null
); }
1108 __kmp_hw_get_catalog_string(KMP_HW_CORE), buf.str)if (__kmp_affinity.flags.verbose || (__kmp_affinity.flags.warnings
&& (__kmp_affinity.type != affinity_none))) { __kmp_msg
(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_AffHWSubsetIncompat
, __kmp_hw_get_catalog_string(KMP_HW_CORE), buf.str), __kmp_msg_null
); }
;
1109 __kmp_str_buf_free(&buf);
1110 return false;
1111 }
1112 // Allow specifying a specific core type or core eff exactly once
1113 for (int k = 0; k < j; ++k) {
1114 if (!item.attr[j] || !item.attr[k])
1115 continue;
1116 if (item.attr[k] == item.attr[j]) {
1117 kmp_str_buf_t buf;
1118 __kmp_hw_get_catalog_core_string(item.attr[j], &buf,
1119 item.num[j] > 0);
1120 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetAttrRepeat, buf.str)if (__kmp_affinity.flags.verbose || (__kmp_affinity.flags.warnings
&& (__kmp_affinity.type != affinity_none))) { __kmp_msg
(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_AffHWSubsetAttrRepeat
, buf.str), __kmp_msg_null); }
;
1121 __kmp_str_buf_free(&buf);
1122 return false;
1123 }
1124 }
1125 }
1126 }
1127 }
1128 }
1129
1130 struct core_type_indexer {
1131 int operator()(const kmp_hw_thread_t &t) const {
1132 switch (t.attrs.get_core_type()) {
1133#if KMP_ARCH_X860 || KMP_ARCH_X86_641
1134 case KMP_HW_CORE_TYPE_ATOM:
1135 return 1;
1136 case KMP_HW_CORE_TYPE_CORE:
1137 return 2;
1138#endif
1139 case KMP_HW_CORE_TYPE_UNKNOWN:
1140 return 0;
1141 }
1142 KMP_ASSERT(0)if (!(0)) { __kmp_debug_assert("0", "openmp/runtime/src/kmp_affinity.cpp"
, 1142); }
;
1143 return 0;
1144 }
1145 };
1146 struct core_eff_indexer {
1147 int operator()(const kmp_hw_thread_t &t) const {
1148 return t.attrs.get_core_eff();
1149 }
1150 };
1151
1152 kmp_sub_ids_t<KMP_HW_MAX_NUM_CORE_TYPES, core_type_indexer> core_type_sub_ids(
1153 core_level);
1154 kmp_sub_ids_t<KMP_HW_MAX_NUM_CORE_EFFS8, core_eff_indexer> core_eff_sub_ids(
1155 core_level);
1156
1157 // Determine which hardware threads should be filtered.
1158 int num_filtered = 0;
1159 bool *filtered = (bool *)__kmp_allocate(sizeof(bool) * num_hw_threads)___kmp_allocate((sizeof(bool) * num_hw_threads), "openmp/runtime/src/kmp_affinity.cpp"
, 1159)
;
1160 for (int i = 0; i < num_hw_threads; ++i) {
1161 kmp_hw_thread_t &hw_thread = hw_threads[i];
1162 // Update type_sub_id
1163 if (using_core_types)
1164 core_type_sub_ids.update(hw_thread);
1165 if (using_core_effs)
1166 core_eff_sub_ids.update(hw_thread);
1167
1168 // Check to see if this hardware thread should be filtered
1169 bool should_be_filtered = false;
1170 for (int hw_subset_index = 0; hw_subset_index < hw_subset_depth;
1171 ++hw_subset_index) {
1172 const auto &hw_subset_item = __kmp_hw_subset->at(hw_subset_index);
1173 int level = topology_levels[hw_subset_index];
1174 if (level == -1)
1175 continue;
1176 if ((using_core_effs || using_core_types) && level == core_level) {
1177 // Look for the core attribute in KMP_HW_SUBSET which corresponds
1178 // to this hardware thread's core attribute. Use this num,offset plus
1179 // the running sub_id for the particular core attribute of this hardware
1180 // thread to determine if the hardware thread should be filtered or not.
1181 int attr_idx;
1182 kmp_hw_core_type_t core_type = hw_thread.attrs.get_core_type();
1183 int core_eff = hw_thread.attrs.get_core_eff();
1184 for (attr_idx = 0; attr_idx < hw_subset_item.num_attrs; ++attr_idx) {
1185 if (using_core_types &&
1186 hw_subset_item.attr[attr_idx].get_core_type() == core_type)
1187 break;
1188 if (using_core_effs &&
1189 hw_subset_item.attr[attr_idx].get_core_eff() == core_eff)
1190 break;
1191 }
1192 // This core attribute isn't in the KMP_HW_SUBSET so always filter it.
1193 if (attr_idx == hw_subset_item.num_attrs) {
1194 should_be_filtered = true;
1195 break;
1196 }
1197 int sub_id;
1198 int num = hw_subset_item.num[attr_idx];
1199 int offset = hw_subset_item.offset[attr_idx];
1200 if (using_core_types)
1201 sub_id = core_type_sub_ids.get_sub_id(hw_thread);
1202 else
1203 sub_id = core_eff_sub_ids.get_sub_id(hw_thread);
1204 if (sub_id < offset ||
1205 (num != kmp_hw_subset_t::USE_ALL && sub_id >= offset + num)) {
1206 should_be_filtered = true;
1207 break;
1208 }
1209 } else {
1210 int num = hw_subset_item.num[0];
1211 int offset = hw_subset_item.offset[0];
1212 if (hw_thread.sub_ids[level] < offset ||
1213 (num != kmp_hw_subset_t::USE_ALL &&
1214 hw_thread.sub_ids[level] >= offset + num)) {
1215 should_be_filtered = true;
1216 break;
1217 }
1218 }
1219 }
1220 // Collect filtering information
1221 filtered[i] = should_be_filtered;
1222 if (should_be_filtered)
1223 num_filtered++;
1224 }
1225
1226 // One last check that we shouldn't allow filtering entire machine
1227 if (num_filtered == num_hw_threads) {
1228 KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetAllFiltered)if (__kmp_affinity.flags.verbose || (__kmp_affinity.flags.warnings
&& (__kmp_affinity.type != affinity_none))) { __kmp_msg
(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_AffHWSubsetAllFiltered
), __kmp_msg_null); }
;
1229 __kmp_free(filtered)___kmp_free((filtered), "openmp/runtime/src/kmp_affinity.cpp"
, 1229)
;
1230 return false;
1231 }
1232
1233 // Apply the filter
1234 int new_index = 0;
1235 for (int i = 0; i < num_hw_threads; ++i) {
1236 if (!filtered[i]) {
1237 if (i != new_index)
1238 hw_threads[new_index] = hw_threads[i];
1239 new_index++;
1240 } else {
1241#if KMP_AFFINITY_SUPPORTED1
1242 KMP_CPU_CLR(hw_threads[i].os_id, __kmp_affin_fullMask)(__kmp_affin_fullMask)->clear(hw_threads[i].os_id);
1243#endif
1244 __kmp_avail_proc--;
1245 }
1246 }
1247
1248 KMP_DEBUG_ASSERT(new_index <= num_hw_threads)if (!(new_index <= num_hw_threads)) { __kmp_debug_assert("new_index <= num_hw_threads"
, "openmp/runtime/src/kmp_affinity.cpp", 1248); }
;
1249 num_hw_threads = new_index;
1250
1251 // Post hardware subset canonicalization
1252 _gather_enumeration_information();
1253 _discover_uniformity();
1254 _set_globals();
1255 _set_last_level_cache();
1256 __kmp_free(filtered)___kmp_free((filtered), "openmp/runtime/src/kmp_affinity.cpp"
, 1256)
;
1257 return true;
1258}
1259
1260bool kmp_topology_t::is_close(int hwt1, int hwt2, int hw_level) const {
1261 if (hw_level >= depth)
1262 return true;
1263 bool retval = true;
1264 const kmp_hw_thread_t &t1 = hw_threads[hwt1];
1265 const kmp_hw_thread_t &t2 = hw_threads[hwt2];
1266 for (int i = 0; i < (depth - hw_level); ++i) {
1267 if (t1.ids[i] != t2.ids[i])
1268 return false;
1269 }
1270 return retval;
1271}
1272
1273////////////////////////////////////////////////////////////////////////////////
1274
1275#if KMP_AFFINITY_SUPPORTED1
1276class kmp_affinity_raii_t {
1277 kmp_affin_mask_t *mask;
1278 bool restored;
1279
1280public:
1281 kmp_affinity_raii_t() : restored(false) {
1282 KMP_CPU_ALLOC(mask)(mask = __kmp_affinity_dispatch->allocate_mask());
1283 KMP_ASSERT(mask != NULL)if (!(mask != __null)) { __kmp_debug_assert("mask != NULL", "openmp/runtime/src/kmp_affinity.cpp"
, 1283); }
;
1284 __kmp_get_system_affinity(mask, TRUE)(mask)->get_system_affinity((!0));
1285 }
1286 void restore() {
1287 __kmp_set_system_affinity(mask, TRUE)(mask)->set_system_affinity((!0));
1288 KMP_CPU_FREE(mask)__kmp_affinity_dispatch->deallocate_mask(mask);
1289 restored = true;
1290 }
1291 ~kmp_affinity_raii_t() {
1292 if (!restored) {
1293 __kmp_set_system_affinity(mask, TRUE)(mask)->set_system_affinity((!0));
1294 KMP_CPU_FREE(mask)__kmp_affinity_dispatch->deallocate_mask(mask);
1295 }
1296 }
1297};
1298
1299bool KMPAffinity::picked_api = false;
1300
1301void *KMPAffinity::Mask::operator new(size_t n) { return __kmp_allocate(n)___kmp_allocate((n), "openmp/runtime/src/kmp_affinity.cpp", 1301
)
; }
1302void *KMPAffinity::Mask::operator new[](size_t n) { return __kmp_allocate(n)___kmp_allocate((n), "openmp/runtime/src/kmp_affinity.cpp", 1302
)
; }
1303void KMPAffinity::Mask::operator delete(void *p) { __kmp_free(p)___kmp_free((p), "openmp/runtime/src/kmp_affinity.cpp", 1303); }
1304void KMPAffinity::Mask::operator delete[](void *p) { __kmp_free(p)___kmp_free((p), "openmp/runtime/src/kmp_affinity.cpp", 1304); }
1305void *KMPAffinity::operator new(size_t n) { return __kmp_allocate(n)___kmp_allocate((n), "openmp/runtime/src/kmp_affinity.cpp", 1305
)
; }
1306void KMPAffinity::operator delete(void *p) { __kmp_free(p)___kmp_free((p), "openmp/runtime/src/kmp_affinity.cpp", 1306); }
1307
1308void KMPAffinity::pick_api() {
1309 KMPAffinity *affinity_dispatch;
1310 if (picked_api)
1311 return;
1312#if KMP_USE_HWLOC0
1313 // Only use Hwloc if affinity isn't explicitly disabled and
1314 // user requests Hwloc topology method
1315 if (__kmp_affinity_top_method == affinity_top_method_hwloc &&
1316 __kmp_affinity.type != affinity_disabled) {
1317 affinity_dispatch = new KMPHwlocAffinity();
1318 } else
1319#endif
1320 {
1321 affinity_dispatch = new KMPNativeAffinity();
1322 }
1323 __kmp_affinity_dispatch = affinity_dispatch;
1324 picked_api = true;
1325}
1326
1327void KMPAffinity::destroy_api() {
1328 if (__kmp_affinity_dispatch != NULL__null) {
1329 delete __kmp_affinity_dispatch;
1330 __kmp_affinity_dispatch = NULL__null;
1331 picked_api = false;
1332 }
1333}
1334
1335#define KMP_ADVANCE_SCAN(scan) \
1336 while (*scan != '\0') { \
1337 scan++; \
1338 }
1339
1340// Print the affinity mask to the character array in a pretty format.
1341// The format is a comma separated list of non-negative integers or integer
1342// ranges: e.g., 1,2,3-5,7,9-15
1343// The format can also be the string "{<empty>}" if no bits are set in mask
1344char *__kmp_affinity_print_mask(char *buf, int buf_len,
1345 kmp_affin_mask_t *mask) {
1346 int start = 0, finish = 0, previous = 0;
1347 bool first_range;
1348 KMP_ASSERT(buf)if (!(buf)) { __kmp_debug_assert("buf", "openmp/runtime/src/kmp_affinity.cpp"
, 1348); }
;
1349 KMP_ASSERT(buf_len >= 40)if (!(buf_len >= 40)) { __kmp_debug_assert("buf_len >= 40"
, "openmp/runtime/src/kmp_affinity.cpp", 1349); }
;
1350 KMP_ASSERT(mask)if (!(mask)) { __kmp_debug_assert("mask", "openmp/runtime/src/kmp_affinity.cpp"
, 1350); }
;
1351 char *scan = buf;
1352 char *end = buf + buf_len - 1;
1353
1354 // Check for empty set.
1355 if (mask->begin() == mask->end()) {
1356 KMP_SNPRINTFsnprintf(scan, end - scan + 1, "{<empty>}");
1357 KMP_ADVANCE_SCAN(scan);
1358 KMP_ASSERT(scan <= end)if (!(scan <= end)) { __kmp_debug_assert("scan <= end",
"openmp/runtime/src/kmp_affinity.cpp", 1358); }
;
1359 return buf;
1360 }
1361
1362 first_range = true;
1363 start = mask->begin();
1364 while (1) {
1365 // Find next range
1366 // [start, previous] is inclusive range of contiguous bits in mask
1367 for (finish = mask->next(start), previous = start;
1368 finish == previous + 1 && finish != mask->end();
1369 finish = mask->next(finish)) {
1370 previous = finish;
1371 }
1372
1373 // The first range does not need a comma printed before it, but the rest
1374 // of the ranges do need a comma beforehand
1375 if (!first_range) {
1376 KMP_SNPRINTFsnprintf(scan, end - scan + 1, "%s", ",");
1377 KMP_ADVANCE_SCAN(scan);
1378 } else {
1379 first_range = false;
1380 }
1381 // Range with three or more contiguous bits in the affinity mask
1382 if (previous - start > 1) {
1383 KMP_SNPRINTFsnprintf(scan, end - scan + 1, "%u-%u", start, previous);
1384 } else {
1385 // Range with one or two contiguous bits in the affinity mask
1386 KMP_SNPRINTFsnprintf(scan, end - scan + 1, "%u", start);
1387 KMP_ADVANCE_SCAN(scan);
1388 if (previous - start > 0) {
1389 KMP_SNPRINTFsnprintf(scan, end - scan + 1, ",%u", previous);
1390 }
1391 }
1392 KMP_ADVANCE_SCAN(scan);
1393 // Start over with new start point
1394 start = finish;
1395 if (start == mask->end())
1396 break;
1397 // Check for overflow
1398 if (end - scan < 2)
1399 break;
1400 }
1401
1402 // Check for overflow
1403 KMP_ASSERT(scan <= end)if (!(scan <= end)) { __kmp_debug_assert("scan <= end",
"openmp/runtime/src/kmp_affinity.cpp", 1403); }
;
1404 return buf;
1405}
1406#undef KMP_ADVANCE_SCAN
1407
1408// Print the affinity mask to the string buffer object in a pretty format
1409// The format is a comma separated list of non-negative integers or integer
1410// ranges: e.g., 1,2,3-5,7,9-15
1411// The format can also be the string "{<empty>}" if no bits are set in mask
1412kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
1413 kmp_affin_mask_t *mask) {
1414 int start = 0, finish = 0, previous = 0;
1415 bool first_range;
1416 KMP_ASSERT(buf)if (!(buf)) { __kmp_debug_assert("buf", "openmp/runtime/src/kmp_affinity.cpp"
, 1416); }
;
1417 KMP_ASSERT(mask)if (!(mask)) { __kmp_debug_assert("mask", "openmp/runtime/src/kmp_affinity.cpp"
, 1417); }
;
1418
1419 __kmp_str_buf_clear(buf);
1420
1421 // Check for empty set.
1422 if (mask->begin() == mask->end()) {
1423 __kmp_str_buf_print(buf, "%s", "{<empty>}");
1424 return buf;
1425 }
1426
1427 first_range = true;
1428 start = mask->begin();
1429 while (1) {
1430 // Find next range
1431 // [start, previous] is inclusive range of contiguous bits in mask
1432 for (finish = mask->next(start), previous = start;
1433 finish == previous + 1 && finish != mask->end();
1434 finish = mask->next(finish)) {
1435 previous = finish;
1436 }
1437
1438 // The first range does not need a comma printed before it, but the rest
1439 // of the ranges do need a comma beforehand
1440 if (!first_range) {
1441 __kmp_str_buf_print(buf, "%s", ",");
1442 } else {
1443 first_range = false;
1444 }
1445 // Range with three or more contiguous bits in the affinity mask
1446 if (previous - start > 1) {
1447 __kmp_str_buf_print(buf, "%u-%u", start, previous);
1448 } else {
1449 // Range with one or two contiguous bits in the affinity mask
1450 __kmp_str_buf_print(buf, "%u", start);
1451 if (previous - start > 0) {
1452 __kmp_str_buf_print(buf, ",%u", previous);
1453 }
1454 }
1455 // Start over with new start point
1456 start = finish;
1457 if (start == mask->end())
1458 break;
1459 }
1460 return buf;
1461}
1462
1463// Return (possibly empty) affinity mask representing the offline CPUs
1464// Caller must free the mask
1465kmp_affin_mask_t *__kmp_affinity_get_offline_cpus() {
1466 kmp_affin_mask_t *offline;
1467 KMP_CPU_ALLOC(offline)(offline = __kmp_affinity_dispatch->allocate_mask());
1468 KMP_CPU_ZERO(offline)(offline)->zero();
1469#if KMP_OS_LINUX1
1470 int n, begin_cpu, end_cpu;
1471 kmp_safe_raii_file_t offline_file;
1472 auto skip_ws = [](FILE *f) {
1473 int c;
1474 do {
1475 c = fgetc(f);
1476 } while (isspace(c));
1477 if (c != EOF(-1))
1478 ungetc(c, f);
1479 };
1480 // File contains CSV of integer ranges representing the offline CPUs
1481 // e.g., 1,2,4-7,9,11-15
1482 int status = offline_file.try_open("/sys/devices/system/cpu/offline", "r");
1483 if (status != 0)
1484 return offline;
1485 while (!feof(offline_file)) {
1486 skip_ws(offline_file);
1487 n = fscanf(offline_file, "%d", &begin_cpu);
1488 if (n != 1)
1489 break;
1490 skip_ws(offline_file);
1491 int c = fgetc(offline_file);
1492 if (c == EOF(-1) || c == ',') {
1493 // Just single CPU
1494 end_cpu = begin_cpu;
1495 } else if (c == '-') {
1496 // Range of CPUs
1497 skip_ws(offline_file);
1498 n = fscanf(offline_file, "%d", &end_cpu);
1499 if (n != 1)
1500 break;
1501 skip_ws(offline_file);
1502 c = fgetc(offline_file); // skip ','
1503 } else {
1504 // Syntax problem
1505 break;
1506 }
1507 // Ensure a valid range of CPUs
1508 if (begin_cpu < 0 || begin_cpu >= __kmp_xproc || end_cpu < 0 ||
1509 end_cpu >= __kmp_xproc || begin_cpu > end_cpu) {
1510 continue;
1511 }
1512 // Insert [begin_cpu, end_cpu] into offline mask
1513 for (int cpu = begin_cpu; cpu <= end_cpu; ++cpu) {
1514 KMP_CPU_SET(cpu, offline)(offline)->set(cpu);
1515 }
1516 }
1517#endif
1518 return offline;
1519}
1520
1521// Return the number of available procs
1522int __kmp_affinity_entire_machine_mask(kmp_affin_mask_t *mask) {
1523 int avail_proc = 0;
1524 KMP_CPU_ZERO(mask)(mask)->zero();
1525
1526#if KMP_GROUP_AFFINITY0
1527
1528 if (__kmp_num_proc_groups > 1) {
1529 int group;
1530 KMP_DEBUG_ASSERT(__kmp_GetActiveProcessorCount != NULL)if (!(__kmp_GetActiveProcessorCount != __null)) { __kmp_debug_assert
("__kmp_GetActiveProcessorCount != __null", "openmp/runtime/src/kmp_affinity.cpp"
, 1530); }
;
1531 for (group = 0; group < __kmp_num_proc_groups; group++) {
1532 int i;
1533 int num = __kmp_GetActiveProcessorCount(group);
1534 for (i = 0; i < num; i++) {
1535 KMP_CPU_SET(i + group * (CHAR_BIT * sizeof(DWORD_PTR)), mask)(mask)->set(i + group * (8 * sizeof(DWORD_PTR)));
1536 avail_proc++;
1537 }
1538 }
1539 } else
1540
1541#endif /* KMP_GROUP_AFFINITY */
1542
1543 {
1544 int proc;
1545 kmp_affin_mask_t *offline_cpus = __kmp_affinity_get_offline_cpus();
1546 for (proc = 0; proc < __kmp_xproc; proc++) {
1547 // Skip offline CPUs
1548 if (KMP_CPU_ISSET(proc, offline_cpus)(offline_cpus)->is_set(proc))
1549 continue;
1550 KMP_CPU_SET(proc, mask)(mask)->set(proc);
1551 avail_proc++;
1552 }
1553 KMP_CPU_FREE(offline_cpus)__kmp_affinity_dispatch->deallocate_mask(offline_cpus);
1554 }
1555
1556 return avail_proc;
1557}
1558
1559// All of the __kmp_affinity_create_*_map() routines should allocate the
1560// internal topology object and set the layer ids for it. Each routine
1561// returns a boolean on whether it was successful at doing so.
1562kmp_affin_mask_t *__kmp_affin_fullMask = NULL__null;
1563// Original mask is a subset of full mask in multiple processor groups topology
1564kmp_affin_mask_t *__kmp_affin_origMask = NULL__null;
1565
1566#if KMP_USE_HWLOC0
1567static inline bool __kmp_hwloc_is_cache_type(hwloc_obj_t obj) {
1568#if HWLOC_API_VERSION >= 0x00020000
1569 return hwloc_obj_type_is_cache(obj->type);
1570#else
1571 return obj->type == HWLOC_OBJ_CACHE;
1572#endif
1573}
1574
1575// Returns KMP_HW_* type derived from HWLOC_* type
1576static inline kmp_hw_t __kmp_hwloc_type_2_topology_type(hwloc_obj_t obj) {
1577
1578 if (__kmp_hwloc_is_cache_type(obj)) {
1579 if (obj->attr->cache.type == HWLOC_OBJ_CACHE_INSTRUCTION)
1580 return KMP_HW_UNKNOWN;
1581 switch (obj->attr->cache.depth) {
1582 case 1:
1583 return KMP_HW_L1;
1584 case 2:
1585#if KMP_MIC_SUPPORTED((0 || 1) && (1 || 0))
1586 if (__kmp_mic_type == mic3) {
1587 return KMP_HW_TILE;
1588 }
1589#endif
1590 return KMP_HW_L2;
1591 case 3:
1592 return KMP_HW_L3;
1593 }
1594 return KMP_HW_UNKNOWN;
1595 }
1596
1597 switch (obj->type) {
1598 case HWLOC_OBJ_PACKAGE:
1599 return KMP_HW_SOCKET;
1600 case HWLOC_OBJ_NUMANODE:
1601 return KMP_HW_NUMA;
1602 case HWLOC_OBJ_CORE:
1603 return KMP_HW_CORE;
1604 case HWLOC_OBJ_PU:
1605 return KMP_HW_THREAD;
1606 case HWLOC_OBJ_GROUP:
1607#if HWLOC_API_VERSION >= 0x00020000
1608 if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_DIE)
1609 return KMP_HW_DIE;
1610 else if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_TILE)
1611 return KMP_HW_TILE;
1612 else if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_MODULE)
1613 return KMP_HW_MODULE;
1614 else if (obj->attr->group.kind == HWLOC_GROUP_KIND_WINDOWS_PROCESSOR_GROUP)
1615 return KMP_HW_PROC_GROUP;
1616#endif
1617 return KMP_HW_UNKNOWN;
1618#if HWLOC_API_VERSION >= 0x00020100
1619 case HWLOC_OBJ_DIE:
1620 return KMP_HW_DIE;
1621#endif
1622 }
1623 return KMP_HW_UNKNOWN;
1624}
1625
1626// Returns the number of objects of type 'type' below 'obj' within the topology
1627// tree structure. e.g., if obj is a HWLOC_OBJ_PACKAGE object, and type is
1628// HWLOC_OBJ_PU, then this will return the number of PU's under the SOCKET
1629// object.
1630static int __kmp_hwloc_get_nobjs_under_obj(hwloc_obj_t obj,
1631 hwloc_obj_type_t type) {
1632 int retval = 0;
1633 hwloc_obj_t first;
1634 for (first = hwloc_get_obj_below_by_type(__kmp_hwloc_topology, obj->type,
1635 obj->logical_index, type, 0);
1636 first != NULL__null && hwloc_get_ancestor_obj_by_type(__kmp_hwloc_topology,
1637 obj->type, first) == obj;
1638 first = hwloc_get_next_obj_by_type(__kmp_hwloc_topology, first->type,
1639 first)) {
1640 ++retval;
1641 }
1642 return retval;
1643}
1644
1645// This gets the sub_id for a lower object under a higher object in the
1646// topology tree
1647static int __kmp_hwloc_get_sub_id(hwloc_topology_t t, hwloc_obj_t higher,
1648 hwloc_obj_t lower) {
1649 hwloc_obj_t obj;
1650 hwloc_obj_type_t ltype = lower->type;
1651 int lindex = lower->logical_index - 1;
1652 int sub_id = 0;
1653 // Get the previous lower object
1654 obj = hwloc_get_obj_by_type(t, ltype, lindex);
1655 while (obj && lindex >= 0 &&
1656 hwloc_bitmap_isincluded(obj->cpuset, higher->cpuset)) {
1657 if (obj->userdata) {
1658 sub_id = (int)(RCAST(kmp_intptr_t, obj->userdata)reinterpret_cast<kmp_intptr_t>(obj->userdata));
1659 break;
1660 }
1661 sub_id++;
1662 lindex--;
1663 obj = hwloc_get_obj_by_type(t, ltype, lindex);
1664 }
1665 // store sub_id + 1 so that 0 is differed from NULL
1666 lower->userdata = RCAST(void *, sub_id + 1)reinterpret_cast<void *>(sub_id + 1);
1667 return sub_id;
1668}
1669
1670static bool __kmp_affinity_create_hwloc_map(kmp_i18n_id_t *const msg_id) {
1671 kmp_hw_t type;
1672 int hw_thread_index, sub_id;
1673 int depth;
1674 hwloc_obj_t pu, obj, root, prev;
1675 kmp_hw_t types[KMP_HW_LAST];
1676 hwloc_obj_type_t hwloc_types[KMP_HW_LAST];
1677
1678 hwloc_topology_t tp = __kmp_hwloc_topology;
1679 *msg_id = kmp_i18n_null;
1680 if (__kmp_affinity.flags.verbose) {
1681 KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY")__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_AffUsingHwloc
, "KMP_AFFINITY"), __kmp_msg_null)
;
1682 }
1683
1684 if (!KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) {
1685 // Hack to try and infer the machine topology using only the data
1686 // available from hwloc on the current thread, and __kmp_xproc.
1687 KMP_ASSERT(__kmp_affinity.type == affinity_none)if (!(__kmp_affinity.type == affinity_none)) { __kmp_debug_assert
("__kmp_affinity.type == affinity_none", "openmp/runtime/src/kmp_affinity.cpp"
, 1687); }
;
1688 // hwloc only guarantees existance of PU object, so check PACKAGE and CORE
1689 hwloc_obj_t o = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PACKAGE, 0);
1690 if (o != NULL__null)
1691 nCoresPerPkg = __kmp_hwloc_get_nobjs_under_obj(o, HWLOC_OBJ_CORE);
1692 else
1693 nCoresPerPkg = 1; // no PACKAGE found
1694 o = hwloc_get_obj_by_type(tp, HWLOC_OBJ_CORE, 0);
1695 if (o != NULL__null)
1696 __kmp_nThreadsPerCore = __kmp_hwloc_get_nobjs_under_obj(o, HWLOC_OBJ_PU);
1697 else
1698 __kmp_nThreadsPerCore = 1; // no CORE found
1699 __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
1700 if (nCoresPerPkg == 0)
1701 nCoresPerPkg = 1; // to prevent possible division by 0
1702 nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
1703 return true;
1704 }
1705
1706#if HWLOC_API_VERSION >= 0x00020400
1707 // Handle multiple types of cores if they exist on the system
1708 int nr_cpu_kinds = hwloc_cpukinds_get_nr(tp, 0);
1709
1710 typedef struct kmp_hwloc_cpukinds_info_t {
1711 int efficiency;
1712 kmp_hw_core_type_t core_type;
1713 hwloc_bitmap_t mask;
1714 } kmp_hwloc_cpukinds_info_t;
1715 kmp_hwloc_cpukinds_info_t *cpukinds = nullptr;
1716
1717 if (nr_cpu_kinds > 0) {
1718 unsigned nr_infos;
1719 struct hwloc_info_s *infos;
1720 cpukinds = (kmp_hwloc_cpukinds_info_t *)__kmp_allocate(___kmp_allocate((sizeof(kmp_hwloc_cpukinds_info_t) * nr_cpu_kinds
), "openmp/runtime/src/kmp_affinity.cpp", 1721)
1721 sizeof(kmp_hwloc_cpukinds_info_t) * nr_cpu_kinds)___kmp_allocate((sizeof(kmp_hwloc_cpukinds_info_t) * nr_cpu_kinds
), "openmp/runtime/src/kmp_affinity.cpp", 1721)
;
1722 for (unsigned idx = 0; idx < (unsigned)nr_cpu_kinds; ++idx) {
1723 cpukinds[idx].efficiency = -1;
1724 cpukinds[idx].core_type = KMP_HW_CORE_TYPE_UNKNOWN;
1725 cpukinds[idx].mask = hwloc_bitmap_alloc();
1726 if (hwloc_cpukinds_get_info(tp, idx, cpukinds[idx].mask,
1727 &cpukinds[idx].efficiency, &nr_infos, &infos,
1728 0) == 0) {
1729 for (unsigned i = 0; i < nr_infos; ++i) {
1730 if (__kmp_str_match("CoreType", 8, infos[i].name)) {
1731#if KMP_ARCH_X860 || KMP_ARCH_X86_641
1732 if (__kmp_str_match("IntelAtom", 9, infos[i].value)) {
1733 cpukinds[idx].core_type = KMP_HW_CORE_TYPE_ATOM;
1734 break;
1735 } else if (__kmp_str_match("IntelCore", 9, infos[i].value)) {
1736 cpukinds[idx].core_type = KMP_HW_CORE_TYPE_CORE;
1737 break;
1738 }
1739#endif
1740 }
1741 }
1742 }
1743 }
1744 }
1745#endif
1746
1747 root = hwloc_get_root_obj(tp);
1748
1749 // Figure out the depth and types in the topology
1750 depth = 0;
1751 pu = hwloc_get_pu_obj_by_os_index(tp, __kmp_affin_fullMask->begin());
1752 KMP_ASSERT(pu)if (!(pu)) { __kmp_debug_assert("pu", "openmp/runtime/src/kmp_affinity.cpp"
, 1752); }
;
1753 obj = pu;
1754 types[depth] = KMP_HW_THREAD;
1755 hwloc_types[depth] = obj->type;
1756 depth++;
1757 while (obj != root && obj != NULL__null) {
1758 obj = obj->parent;
1759#if HWLOC_API_VERSION >= 0x00020000
1760 if (obj->memory_arity) {
1761 hwloc_obj_t memory;
1762 for (memory = obj->memory_first_child; memory;
1763 memory = hwloc_get_next_child(tp, obj, memory)) {
1764 if (memory->type == HWLOC_OBJ_NUMANODE)
1765 break;
1766 }
1767 if (memory && memory->type == HWLOC_OBJ_NUMANODE) {
1768 types[depth] = KMP_HW_NUMA;
1769 hwloc_types[depth] = memory->type;
1770 depth++;
1771 }
1772 }
1773#endif
1774 type = __kmp_hwloc_type_2_topology_type(obj);
1775 if (type != KMP_HW_UNKNOWN) {
1776 types[depth] = type;
1777 hwloc_types[depth] = obj->type;
1778 depth++;
1779 }
1780 }
1781 KMP_ASSERT(depth > 0)if (!(depth > 0)) { __kmp_debug_assert("depth > 0", "openmp/runtime/src/kmp_affinity.cpp"
, 1781); }
;
1782
1783 // Get the order for the types correct
1784 for (int i = 0, j = depth - 1; i < j; ++i, --j) {
1785 hwloc_obj_type_t hwloc_temp = hwloc_types[i];
1786 kmp_hw_t temp = types[i];
1787 types[i] = types[j];
1788 types[j] = temp;
1789 hwloc_types[i] = hwloc_types[j];
1790 hwloc_types[j] = hwloc_temp;
1791 }
1792
1793 // Allocate the data structure to be returned.
1794 __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
1795
1796 hw_thread_index = 0;
1797 pu = NULL__null;
1798 while ((pu = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, pu))) {
1799 int index = depth - 1;
1800 bool included = KMP_CPU_ISSET(pu->os_index, __kmp_affin_fullMask)(__kmp_affin_fullMask)->is_set(pu->os_index);
1801 kmp_hw_thread_t &hw_thread = __kmp_topology->at(hw_thread_index);
1802 if (included) {
1803 hw_thread.clear();
1804 hw_thread.ids[index] = pu->logical_index;
1805 hw_thread.os_id = pu->os_index;
1806 // If multiple core types, then set that attribute for the hardware thread
1807#if HWLOC_API_VERSION >= 0x00020400
1808 if (cpukinds) {
1809 int cpukind_index = -1;
1810 for (int i = 0; i < nr_cpu_kinds; ++i) {
1811 if (hwloc_bitmap_isset(cpukinds[i].mask, hw_thread.os_id)) {
1812 cpukind_index = i;
1813 break;
1814 }
1815 }
1816 if (cpukind_index >= 0) {
1817 hw_thread.attrs.set_core_type(cpukinds[cpukind_index].core_type);
1818 hw_thread.attrs.set_core_eff(cpukinds[cpukind_index].efficiency);
1819 }
1820 }
1821#endif
1822 index--;
1823 }
1824 obj = pu;
1825 prev = obj;
1826 while (obj != root && obj != NULL__null) {
1827 obj = obj->parent;
1828#if HWLOC_API_VERSION >= 0x00020000
1829 // NUMA Nodes are handled differently since they are not within the
1830 // parent/child structure anymore. They are separate children
1831 // of obj (memory_first_child points to first memory child)
1832 if (obj->memory_arity) {
1833 hwloc_obj_t memory;
1834 for (memory = obj->memory_first_child; memory;
1835 memory = hwloc_get_next_child(tp, obj, memory)) {
1836 if (memory->type == HWLOC_OBJ_NUMANODE)
1837 break;
1838 }
1839 if (memory && memory->type == HWLOC_OBJ_NUMANODE) {
1840 sub_id = __kmp_hwloc_get_sub_id(tp, memory, prev);
1841 if (included) {
1842 hw_thread.ids[index] = memory->logical_index;
1843 hw_thread.ids[index + 1] = sub_id;
1844 index--;
1845 }
1846 prev = memory;
1847 }
1848 prev = obj;
1849 }
1850#endif
1851 type = __kmp_hwloc_type_2_topology_type(obj);
1852 if (type != KMP_HW_UNKNOWN) {
1853 sub_id = __kmp_hwloc_get_sub_id(tp, obj, prev);
1854 if (included) {
1855 hw_thread.ids[index] = obj->logical_index;
1856 hw_thread.ids[index + 1] = sub_id;
1857 index--;
1858 }
1859 prev = obj;
1860 }
1861 }
1862 if (included)
1863 hw_thread_index++;
1864 }
1865
1866#if HWLOC_API_VERSION >= 0x00020400
1867 // Free the core types information
1868 if (cpukinds) {
1869 for (int idx = 0; idx < nr_cpu_kinds; ++idx)
1870 hwloc_bitmap_free(cpukinds[idx].mask);
1871 __kmp_free(cpukinds)___kmp_free((cpukinds), "openmp/runtime/src/kmp_affinity.cpp"
, 1871)
;
1872 }
1873#endif
1874 __kmp_topology->sort_ids();
1875 return true;
1876}
1877#endif // KMP_USE_HWLOC
1878
1879// If we don't know how to retrieve the machine's processor topology, or
1880// encounter an error in doing so, this routine is called to form a "flat"
1881// mapping of os thread id's <-> processor id's.
1882static bool __kmp_affinity_create_flat_map(kmp_i18n_id_t *const msg_id) {
1883 *msg_id = kmp_i18n_null;
1884 int depth = 3;
1885 kmp_hw_t types[] = {KMP_HW_SOCKET, KMP_HW_CORE, KMP_HW_THREAD};
1886
1887 if (__kmp_affinity.flags.verbose) {
1888 KMP_INFORM(UsingFlatOS, "KMP_AFFINITY")__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_UsingFlatOS
, "KMP_AFFINITY"), __kmp_msg_null)
;
1889 }
1890
1891 // Even if __kmp_affinity.type == affinity_none, this routine might still
1892 // be called to set __kmp_ncores, as well as
1893 // __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
1894 if (!KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) {
1895 KMP_ASSERT(__kmp_affinity.type == affinity_none)if (!(__kmp_affinity.type == affinity_none)) { __kmp_debug_assert
("__kmp_affinity.type == affinity_none", "openmp/runtime/src/kmp_affinity.cpp"
, 1895); }
;
1896 __kmp_ncores = nPackages = __kmp_xproc;
1897 __kmp_nThreadsPerCore = nCoresPerPkg = 1;
1898 return true;
1899 }
1900
1901 // When affinity is off, this routine will still be called to set
1902 // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
1903 // Make sure all these vars are set correctly, and return now if affinity is
1904 // not enabled.
1905 __kmp_ncores = nPackages = __kmp_avail_proc;
1906 __kmp_nThreadsPerCore = nCoresPerPkg = 1;
1907
1908 // Construct the data structure to be returned.
1909 __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
1910 int avail_ct = 0;
1911 int i;
1912 KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask)for (i = (__kmp_affin_fullMask)->begin(); (int)i != (__kmp_affin_fullMask
)->end(); i = (__kmp_affin_fullMask)->next(i))
{
1913 // Skip this proc if it is not included in the machine model.
1914 if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)(__kmp_affin_fullMask)->is_set(i)) {
1915 continue;
1916 }
1917 kmp_hw_thread_t &hw_thread = __kmp_topology->at(avail_ct);
1918 hw_thread.clear();
1919 hw_thread.os_id = i;
1920 hw_thread.ids[0] = i;
1921 hw_thread.ids[1] = 0;
1922 hw_thread.ids[2] = 0;
1923 avail_ct++;
1924 }
1925 if (__kmp_affinity.flags.verbose) {
1926 KMP_INFORM(OSProcToPackage, "KMP_AFFINITY")__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_OSProcToPackage
, "KMP_AFFINITY"), __kmp_msg_null)
;
1927 }
1928 return true;
1929}
1930
1931#if KMP_GROUP_AFFINITY0
1932// If multiple Windows* OS processor groups exist, we can create a 2-level
1933// topology map with the groups at level 0 and the individual procs at level 1.
1934// This facilitates letting the threads float among all procs in a group,
1935// if granularity=group (the default when there are multiple groups).
1936static bool __kmp_affinity_create_proc_group_map(kmp_i18n_id_t *const msg_id) {
1937 *msg_id = kmp_i18n_null;
1938 int depth = 3;
1939 kmp_hw_t types[] = {KMP_HW_PROC_GROUP, KMP_HW_CORE, KMP_HW_THREAD};
1940 const static size_t BITS_PER_GROUP = CHAR_BIT8 * sizeof(DWORD_PTR);
1941
1942 if (__kmp_affinity.flags.verbose) {
1943 KMP_INFORM(AffWindowsProcGroupMap, "KMP_AFFINITY")__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_AffWindowsProcGroupMap
, "KMP_AFFINITY"), __kmp_msg_null)
;
1944 }
1945
1946 // If we aren't affinity capable, then use flat topology
1947 if (!KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) {
1948 KMP_ASSERT(__kmp_affinity.type == affinity_none)if (!(__kmp_affinity.type == affinity_none)) { __kmp_debug_assert
("__kmp_affinity.type == affinity_none", "openmp/runtime/src/kmp_affinity.cpp"
, 1948); }
;
1949 nPackages = __kmp_num_proc_groups;
1950 __kmp_nThreadsPerCore = 1;
1951 __kmp_ncores = __kmp_xproc;
1952 nCoresPerPkg = nPackages / __kmp_ncores;
1953 return true;
1954 }
1955
1956 // Construct the data structure to be returned.
1957 __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
1958 int avail_ct = 0;
1959 int i;
1960 KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask)for (i = (__kmp_affin_fullMask)->begin(); (int)i != (__kmp_affin_fullMask
)->end(); i = (__kmp_affin_fullMask)->next(i))
{
1961 // Skip this proc if it is not included in the machine model.
1962 if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)(__kmp_affin_fullMask)->is_set(i)) {
1963 continue;
1964 }
1965 kmp_hw_thread_t &hw_thread = __kmp_topology->at(avail_ct++);
1966 hw_thread.clear();
1967 hw_thread.os_id = i;
1968 hw_thread.ids[0] = i / BITS_PER_GROUP;
1969 hw_thread.ids[1] = hw_thread.ids[2] = i % BITS_PER_GROUP;
1970 }
1971 return true;
1972}
1973#endif /* KMP_GROUP_AFFINITY */
1974
1975#if KMP_ARCH_X860 || KMP_ARCH_X86_641
1976
1977template <kmp_uint32 LSB, kmp_uint32 MSB>
1978static inline unsigned __kmp_extract_bits(kmp_uint32 v) {
1979 const kmp_uint32 SHIFT_LEFT = sizeof(kmp_uint32) * 8 - 1 - MSB;
1980 const kmp_uint32 SHIFT_RIGHT = LSB;
1981 kmp_uint32 retval = v;
1982 retval <<= SHIFT_LEFT;
1983 retval >>= (SHIFT_LEFT + SHIFT_RIGHT);
1984 return retval;
1985}
1986
1987static int __kmp_cpuid_mask_width(int count) {
1988 int r = 0;
1989
1990 while ((1 << r) < count)
1991 ++r;
1992 return r;
1993}
1994
1995class apicThreadInfo {
1996public:
1997 unsigned osId; // param to __kmp_affinity_bind_thread
1998 unsigned apicId; // from cpuid after binding
1999 unsigned maxCoresPerPkg; // ""
2000 unsigned maxThreadsPerPkg; // ""
2001 unsigned pkgId; // inferred from above values
2002 unsigned coreId; // ""
2003 unsigned threadId; // ""
2004};
2005
2006static int __kmp_affinity_cmp_apicThreadInfo_phys_id(const void *a,
2007 const void *b) {
2008 const apicThreadInfo *aa = (const apicThreadInfo *)a;
2009 const apicThreadInfo *bb = (const apicThreadInfo *)b;
2010 if (aa->pkgId < bb->pkgId)
2011 return -1;
2012 if (aa->pkgId > bb->pkgId)
2013 return 1;
2014 if (aa->coreId < bb->coreId)
2015 return -1;
2016 if (aa->coreId > bb->coreId)
2017 return 1;
2018 if (aa->threadId < bb->threadId)
2019 return -1;
2020 if (aa->threadId > bb->threadId)
2021 return 1;
2022 return 0;
2023}
2024
2025class kmp_cache_info_t {
2026public:
2027 struct info_t {
2028 unsigned level, mask;
2029 };
2030 kmp_cache_info_t() : depth(0) { get_leaf4_levels(); }
2031 size_t get_depth() const { return depth; }
2032 info_t &operator[](size_t index) { return table[index]; }
2033 const info_t &operator[](size_t index) const { return table[index]; }
2034
2035 static kmp_hw_t get_topology_type(unsigned level) {
2036 KMP_DEBUG_ASSERT(level >= 1 && level <= MAX_CACHE_LEVEL)if (!(level >= 1 && level <= MAX_CACHE_LEVEL)) {
__kmp_debug_assert("level >= 1 && level <= MAX_CACHE_LEVEL"
, "openmp/runtime/src/kmp_affinity.cpp", 2036); }
;
2037 switch (level) {
2038 case 1:
2039 return KMP_HW_L1;
2040 case 2:
2041 return KMP_HW_L2;
2042 case 3:
2043 return KMP_HW_L3;
2044 }
2045 return KMP_HW_UNKNOWN;
2046 }
2047
2048private:
2049 static const int MAX_CACHE_LEVEL = 3;
2050
2051 size_t depth;
2052 info_t table[MAX_CACHE_LEVEL];
2053
2054 void get_leaf4_levels() {
2055 unsigned level = 0;
2056 while (depth < MAX_CACHE_LEVEL) {
2057 unsigned cache_type, max_threads_sharing;
2058 unsigned cache_level, cache_mask_width;
2059 kmp_cpuid buf2;
2060 __kmp_x86_cpuid(4, level, &buf2);
2061 cache_type = __kmp_extract_bits<0, 4>(buf2.eax);
2062 if (!cache_type)
2063 break;
2064 // Skip instruction caches
2065 if (cache_type == 2) {
2066 level++;
2067 continue;
2068 }
2069 max_threads_sharing = __kmp_extract_bits<14, 25>(buf2.eax) + 1;
2070 cache_mask_width = __kmp_cpuid_mask_width(max_threads_sharing);
2071 cache_level = __kmp_extract_bits<5, 7>(buf2.eax);
2072 table[depth].level = cache_level;
2073 table[depth].mask = ((-1) << cache_mask_width);
2074 depth++;
2075 level++;
2076 }
2077 }
2078};
2079
2080// On IA-32 architecture and Intel(R) 64 architecture, we attempt to use
2081// an algorithm which cycles through the available os threads, setting
2082// the current thread's affinity mask to that thread, and then retrieves
2083// the Apic Id for each thread context using the cpuid instruction.
2084static bool __kmp_affinity_create_apicid_map(kmp_i18n_id_t *const msg_id) {
2085 kmp_cpuid buf;
2086 *msg_id = kmp_i18n_null;
2087
2088 if (__kmp_affinity.flags.verbose) {
2089 KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(DecodingLegacyAPIC))__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_AffInfoStr
, "KMP_AFFINITY", __kmp_i18n_catgets(kmp_i18n_str_DecodingLegacyAPIC
)), __kmp_msg_null)
;
2090 }
2091
2092 // Check if cpuid leaf 4 is supported.
2093 __kmp_x86_cpuid(0, 0, &buf);
2094 if (buf.eax < 4) {
2095 *msg_id = kmp_i18n_str_NoLeaf4Support;
2096 return false;
2097 }
2098
2099 // The algorithm used starts by setting the affinity to each available thread
2100 // and retrieving info from the cpuid instruction, so if we are not capable of
2101 // calling __kmp_get_system_affinity() and _kmp_get_system_affinity(), then we
2102 // need to do something else - use the defaults that we calculated from
2103 // issuing cpuid without binding to each proc.
2104 if (!KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) {
2105 // Hack to try and infer the machine topology using only the data
2106 // available from cpuid on the current thread, and __kmp_xproc.
2107 KMP_ASSERT(__kmp_affinity.type == affinity_none)if (!(__kmp_affinity.type == affinity_none)) { __kmp_debug_assert
("__kmp_affinity.type == affinity_none", "openmp/runtime/src/kmp_affinity.cpp"
, 2107); }
;
2108
2109 // Get an upper bound on the number of threads per package using cpuid(1).
2110 // On some OS/chps combinations where HT is supported by the chip but is
2111 // disabled, this value will be 2 on a single core chip. Usually, it will be
2112 // 2 if HT is enabled and 1 if HT is disabled.
2113 __kmp_x86_cpuid(1, 0, &buf);
2114 int maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
2115 if (maxThreadsPerPkg == 0) {
2116 maxThreadsPerPkg = 1;
2117 }
2118
2119 // The num cores per pkg comes from cpuid(4). 1 must be added to the encoded
2120 // value.
2121 //
2122 // The author of cpu_count.cpp treated this only an upper bound on the
2123 // number of cores, but I haven't seen any cases where it was greater than
2124 // the actual number of cores, so we will treat it as exact in this block of
2125 // code.
2126 //
2127 // First, we need to check if cpuid(4) is supported on this chip. To see if
2128 // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n or
2129 // greater.
2130 __kmp_x86_cpuid(0, 0, &buf);
2131 if (buf.eax >= 4) {
2132 __kmp_x86_cpuid(4, 0, &buf);
2133 nCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
2134 } else {
2135 nCoresPerPkg = 1;
2136 }
2137
2138 // There is no way to reliably tell if HT is enabled without issuing the
2139 // cpuid instruction from every thread, can correlating the cpuid info, so
2140 // if the machine is not affinity capable, we assume that HT is off. We have
2141 // seen quite a few machines where maxThreadsPerPkg is 2, yet the machine
2142 // does not support HT.
2143 //
2144 // - Older OSes are usually found on machines with older chips, which do not
2145 // support HT.
2146 // - The performance penalty for mistakenly identifying a machine as HT when
2147 // it isn't (which results in blocktime being incorrectly set to 0) is
2148 // greater than the penalty when for mistakenly identifying a machine as
2149 // being 1 thread/core when it is really HT enabled (which results in
2150 // blocktime being incorrectly set to a positive value).
2151 __kmp_ncores = __kmp_xproc;
2152 nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
2153 __kmp_nThreadsPerCore = 1;
2154 return true;
2155 }
2156
2157 // From here on, we can assume that it is safe to call
2158 // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
2159 // __kmp_affinity.type = affinity_none.
2160
2161 // Save the affinity mask for the current thread.
2162 kmp_affinity_raii_t previous_affinity;
2163
2164 // Run through each of the available contexts, binding the current thread
2165 // to it, and obtaining the pertinent information using the cpuid instr.
2166 //
2167 // The relevant information is:
2168 // - Apic Id: Bits 24:31 of ebx after issuing cpuid(1) - each thread context
2169 // has a uniqie Apic Id, which is of the form pkg# : core# : thread#.
2170 // - Max Threads Per Pkg: Bits 16:23 of ebx after issuing cpuid(1). The value
2171 // of this field determines the width of the core# + thread# fields in the
2172 // Apic Id. It is also an upper bound on the number of threads per
2173 // package, but it has been verified that situations happen were it is not
2174 // exact. In particular, on certain OS/chip combinations where Intel(R)
2175 // Hyper-Threading Technology is supported by the chip but has been
2176 // disabled, the value of this field will be 2 (for a single core chip).
2177 // On other OS/chip combinations supporting Intel(R) Hyper-Threading
2178 // Technology, the value of this field will be 1 when Intel(R)
2179 // Hyper-Threading Technology is disabled and 2 when it is enabled.
2180 // - Max Cores Per Pkg: Bits 26:31 of eax after issuing cpuid(4). The value
2181 // of this field (+1) determines the width of the core# field in the Apic
2182 // Id. The comments in "cpucount.cpp" say that this value is an upper
2183 // bound, but the IA-32 architecture manual says that it is exactly the
2184 // number of cores per package, and I haven't seen any case where it
2185 // wasn't.
2186 //
2187 // From this information, deduce the package Id, core Id, and thread Id,
2188 // and set the corresponding fields in the apicThreadInfo struct.
2189 unsigned i;
2190 apicThreadInfo *threadInfo = (apicThreadInfo *)__kmp_allocate(___kmp_allocate((__kmp_avail_proc * sizeof(apicThreadInfo)), "openmp/runtime/src/kmp_affinity.cpp"
, 2191)
2191 __kmp_avail_proc * sizeof(apicThreadInfo))___kmp_allocate((__kmp_avail_proc * sizeof(apicThreadInfo)), "openmp/runtime/src/kmp_affinity.cpp"
, 2191)
;
2192 unsigned nApics = 0;
2193 KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask)for (i = (__kmp_affin_fullMask)->begin(); (int)i != (__kmp_affin_fullMask
)->end(); i = (__kmp_affin_fullMask)->next(i))
{
2194 // Skip this proc if it is not included in the machine model.
2195 if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)(__kmp_affin_fullMask)->is_set(i)) {
2196 continue;
2197 }
2198 KMP_DEBUG_ASSERT((int)nApics < __kmp_avail_proc)if (!((int)nApics < __kmp_avail_proc)) { __kmp_debug_assert
("(int)nApics < __kmp_avail_proc", "openmp/runtime/src/kmp_affinity.cpp"
, 2198); }
;
2199
2200 __kmp_affinity_dispatch->bind_thread(i);
2201 threadInfo[nApics].osId = i;
2202
2203 // The apic id and max threads per pkg come from cpuid(1).
2204 __kmp_x86_cpuid(1, 0, &buf);
2205 if (((buf.edx >> 9) & 1) == 0) {
2206 __kmp_free(threadInfo)___kmp_free((threadInfo), "openmp/runtime/src/kmp_affinity.cpp"
, 2206)
;
2207 *msg_id = kmp_i18n_str_ApicNotPresent;
2208 return false;
2209 }
2210 threadInfo[nApics].apicId = (buf.ebx >> 24) & 0xff;
2211 threadInfo[nApics].maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
2212 if (threadInfo[nApics].maxThreadsPerPkg == 0) {
2213 threadInfo[nApics].maxThreadsPerPkg = 1;
2214 }
2215
2216 // Max cores per pkg comes from cpuid(4). 1 must be added to the encoded
2217 // value.
2218 //
2219 // First, we need to check if cpuid(4) is supported on this chip. To see if
2220 // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n
2221 // or greater.
2222 __kmp_x86_cpuid(0, 0, &buf);
2223 if (buf.eax >= 4) {
2224 __kmp_x86_cpuid(4, 0, &buf);
2225 threadInfo[nApics].maxCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
2226 } else {
2227 threadInfo[nApics].maxCoresPerPkg = 1;
2228 }
2229
2230 // Infer the pkgId / coreId / threadId using only the info obtained locally.
2231 int widthCT = __kmp_cpuid_mask_width(threadInfo[nApics].maxThreadsPerPkg);
2232 threadInfo[nApics].pkgId = threadInfo[nApics].apicId >> widthCT;
2233
2234 int widthC = __kmp_cpuid_mask_width(threadInfo[nApics].maxCoresPerPkg);
2235 int widthT = widthCT - widthC;
2236 if (widthT < 0) {
2237 // I've never seen this one happen, but I suppose it could, if the cpuid
2238 // instruction on a chip was really screwed up. Make sure to restore the
2239 // affinity mask before the tail call.
2240 __kmp_free(threadInfo)___kmp_free((threadInfo), "openmp/runtime/src/kmp_affinity.cpp"
, 2240)
;
2241 *msg_id = kmp_i18n_str_InvalidCpuidInfo;
2242 return false;
2243 }
2244
2245 int maskC = (1 << widthC) - 1;
2246 threadInfo[nApics].coreId = (threadInfo[nApics].apicId >> widthT) & maskC;
2247
2248 int maskT = (1 << widthT) - 1;
2249 threadInfo[nApics].threadId = threadInfo[nApics].apicId & maskT;
2250
2251 nApics++;
2252 }
2253
2254 // We've collected all the info we need.
2255 // Restore the old affinity mask for this thread.
2256 previous_affinity.restore();
2257
2258 // Sort the threadInfo table by physical Id.
2259 qsort(threadInfo, nApics, sizeof(*threadInfo),
2260 __kmp_affinity_cmp_apicThreadInfo_phys_id);
2261
2262 // The table is now sorted by pkgId / coreId / threadId, but we really don't
2263 // know the radix of any of the fields. pkgId's may be sparsely assigned among
2264 // the chips on a system. Although coreId's are usually assigned
2265 // [0 .. coresPerPkg-1] and threadId's are usually assigned
2266 // [0..threadsPerCore-1], we don't want to make any such assumptions.
2267 //
2268 // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
2269 // total # packages) are at this point - we want to determine that now. We
2270 // only have an upper bound on the first two figures.
2271 //
2272 // We also perform a consistency check at this point: the values returned by
2273 // the cpuid instruction for any thread bound to a given package had better
2274 // return the same info for maxThreadsPerPkg and maxCoresPerPkg.
2275 nPackages = 1;
2276 nCoresPerPkg = 1;
2277 __kmp_nThreadsPerCore = 1;
2278 unsigned nCores = 1;
2279
2280 unsigned pkgCt = 1; // to determine radii
2281 unsigned lastPkgId = threadInfo[0].pkgId;
2282 unsigned coreCt = 1;
2283 unsigned lastCoreId = threadInfo[0].coreId;
2284 unsigned threadCt = 1;
2285 unsigned lastThreadId = threadInfo[0].threadId;
2286
2287 // intra-pkg consist checks
2288 unsigned prevMaxCoresPerPkg = threadInfo[0].maxCoresPerPkg;
2289 unsigned prevMaxThreadsPerPkg = threadInfo[0].maxThreadsPerPkg;
2290
2291 for (i = 1; i < nApics; i++) {
2292 if (threadInfo[i].pkgId != lastPkgId) {
2293 nCores++;
2294 pkgCt++;
2295 lastPkgId = threadInfo[i].pkgId;
2296 if ((int)coreCt > nCoresPerPkg)
2297 nCoresPerPkg = coreCt;
2298 coreCt = 1;
2299 lastCoreId = threadInfo[i].coreId;
2300 if ((int)threadCt > __kmp_nThreadsPerCore)
2301 __kmp_nThreadsPerCore = threadCt;
2302 threadCt = 1;
2303 lastThreadId = threadInfo[i].threadId;
2304
2305 // This is a different package, so go on to the next iteration without
2306 // doing any consistency checks. Reset the consistency check vars, though.
2307 prevMaxCoresPerPkg = threadInfo[i].maxCoresPerPkg;
2308 prevMaxThreadsPerPkg = threadInfo[i].maxThreadsPerPkg;
2309 continue;
2310 }
2311
2312 if (threadInfo[i].coreId != lastCoreId) {
2313 nCores++;
2314 coreCt++;
2315 lastCoreId = threadInfo[i].coreId;
2316 if ((int)threadCt > __kmp_nThreadsPerCore)
2317 __kmp_nThreadsPerCore = threadCt;
2318 threadCt = 1;
2319 lastThreadId = threadInfo[i].threadId;
2320 } else if (threadInfo[i].threadId != lastThreadId) {
2321 threadCt++;
2322 lastThreadId = threadInfo[i].threadId;
2323 } else {
2324 __kmp_free(threadInfo)___kmp_free((threadInfo), "openmp/runtime/src/kmp_affinity.cpp"
, 2324)
;
2325 *msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
2326 return false;
2327 }
2328
2329 // Check to make certain that the maxCoresPerPkg and maxThreadsPerPkg
2330 // fields agree between all the threads bounds to a given package.
2331 if ((prevMaxCoresPerPkg != threadInfo[i].maxCoresPerPkg) ||
2332 (prevMaxThreadsPerPkg != threadInfo[i].maxThreadsPerPkg)) {
2333 __kmp_free(threadInfo)___kmp_free((threadInfo), "openmp/runtime/src/kmp_affinity.cpp"
, 2333)
;
2334 *msg_id = kmp_i18n_str_InconsistentCpuidInfo;
2335 return false;
2336 }
2337 }
2338 // When affinity is off, this routine will still be called to set
2339 // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
2340 // Make sure all these vars are set correctly
2341 nPackages = pkgCt;
2342 if ((int)coreCt > nCoresPerPkg)
2343 nCoresPerPkg = coreCt;
2344 if ((int)threadCt > __kmp_nThreadsPerCore)
2345 __kmp_nThreadsPerCore = threadCt;
2346 __kmp_ncores = nCores;
2347 KMP_DEBUG_ASSERT(nApics == (unsigned)__kmp_avail_proc)if (!(nApics == (unsigned)__kmp_avail_proc)) { __kmp_debug_assert
("nApics == (unsigned)__kmp_avail_proc", "openmp/runtime/src/kmp_affinity.cpp"
, 2347); }
;
2348
2349 // Now that we've determined the number of packages, the number of cores per
2350 // package, and the number of threads per core, we can construct the data
2351 // structure that is to be returned.
2352 int idx = 0;
2353 int pkgLevel = 0;
2354 int coreLevel = 1;
2355 int threadLevel = 2;
2356 //(__kmp_nThreadsPerCore <= 1) ? -1 : ((coreLevel >= 0) ? 2 : 1);
2357 int depth = (pkgLevel >= 0) + (coreLevel >= 0) + (threadLevel >= 0);
2358 kmp_hw_t types[3];
2359 if (pkgLevel >= 0)
2360 types[idx++] = KMP_HW_SOCKET;
2361 if (coreLevel >= 0)
2362 types[idx++] = KMP_HW_CORE;
2363 if (threadLevel >= 0)
2364 types[idx++] = KMP_HW_THREAD;
2365
2366 KMP_ASSERT(depth > 0)if (!(depth > 0)) { __kmp_debug_assert("depth > 0", "openmp/runtime/src/kmp_affinity.cpp"
, 2366); }
;
2367 __kmp_topology = kmp_topology_t::allocate(nApics, depth, types);
2368
2369 for (i = 0; i < nApics; ++i) {
2370 idx = 0;
2371 unsigned os = threadInfo[i].osId;
2372 kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
2373 hw_thread.clear();
2374
2375 if (pkgLevel >= 0) {
2376 hw_thread.ids[idx++] = threadInfo[i].pkgId;
2377 }
2378 if (coreLevel >= 0) {
2379 hw_thread.ids[idx++] = threadInfo[i].coreId;
2380 }
2381 if (threadLevel >= 0) {
2382 hw_thread.ids[idx++] = threadInfo[i].threadId;
2383 }
2384 hw_thread.os_id = os;
2385 }
2386
2387 __kmp_free(threadInfo)___kmp_free((threadInfo), "openmp/runtime/src/kmp_affinity.cpp"
, 2387)
;
2388 __kmp_topology->sort_ids();
2389 if (!__kmp_topology->check_ids()) {
2390 kmp_topology_t::deallocate(__kmp_topology);
2391 __kmp_topology = nullptr;
2392 *msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
2393 return false;
2394 }
2395 return true;
2396}
2397
2398// Hybrid cpu detection using CPUID.1A
2399// Thread should be pinned to processor already
2400static void __kmp_get_hybrid_info(kmp_hw_core_type_t *type, int *efficiency,
2401 unsigned *native_model_id) {
2402 kmp_cpuid buf;
2403 __kmp_x86_cpuid(0x1a, 0, &buf);
2404 *type = (kmp_hw_core_type_t)__kmp_extract_bits<24, 31>(buf.eax);
2405 switch (*type) {
2406 case KMP_HW_CORE_TYPE_ATOM:
2407 *efficiency = 0;
2408 break;
2409 case KMP_HW_CORE_TYPE_CORE:
2410 *efficiency = 1;
2411 break;
2412 default:
2413 *efficiency = 0;
2414 }
2415 *native_model_id = __kmp_extract_bits<0, 23>(buf.eax);
2416}
2417
2418// Intel(R) microarchitecture code name Nehalem, Dunnington and later
2419// architectures support a newer interface for specifying the x2APIC Ids,
2420// based on CPUID.B or CPUID.1F
2421/*
2422 * CPUID.B or 1F, Input ECX (sub leaf # aka level number)
2423 Bits Bits Bits Bits
2424 31-16 15-8 7-4 4-0
2425---+-----------+--------------+-------------+-----------------+
2426EAX| reserved | reserved | reserved | Bits to Shift |
2427---+-----------|--------------+-------------+-----------------|
2428EBX| reserved | Num logical processors at level (16 bits) |
2429---+-----------|--------------+-------------------------------|
2430ECX| reserved | Level Type | Level Number (8 bits) |
2431---+-----------+--------------+-------------------------------|
2432EDX| X2APIC ID (32 bits) |
2433---+----------------------------------------------------------+
2434*/
2435
2436enum {
2437 INTEL_LEVEL_TYPE_INVALID = 0, // Package level
2438 INTEL_LEVEL_TYPE_SMT = 1,
2439 INTEL_LEVEL_TYPE_CORE = 2,
2440 INTEL_LEVEL_TYPE_MODULE = 3,
2441 INTEL_LEVEL_TYPE_TILE = 4,
2442 INTEL_LEVEL_TYPE_DIE = 5,
2443 INTEL_LEVEL_TYPE_LAST = 6,
2444};
2445
2446struct cpuid_level_info_t {
2447 unsigned level_type, mask, mask_width, nitems, cache_mask;
2448};
2449
2450static kmp_hw_t __kmp_intel_type_2_topology_type(int intel_type) {
2451 switch (intel_type) {
2452 case INTEL_LEVEL_TYPE_INVALID:
2453 return KMP_HW_SOCKET;
2454 case INTEL_LEVEL_TYPE_SMT:
2455 return KMP_HW_THREAD;
2456 case INTEL_LEVEL_TYPE_CORE:
2457 return KMP_HW_CORE;
2458 case INTEL_LEVEL_TYPE_TILE:
2459 return KMP_HW_TILE;
2460 case INTEL_LEVEL_TYPE_MODULE:
2461 return KMP_HW_MODULE;
2462 case INTEL_LEVEL_TYPE_DIE:
2463 return KMP_HW_DIE;
2464 }
2465 return KMP_HW_UNKNOWN;
2466}
2467
2468// This function takes the topology leaf, a levels array to store the levels
2469// detected and a bitmap of the known levels.
2470// Returns the number of levels in the topology
2471static unsigned
2472__kmp_x2apicid_get_levels(int leaf,
2473 cpuid_level_info_t levels[INTEL_LEVEL_TYPE_LAST],
2474 kmp_uint64 known_levels) {
2475 unsigned level, levels_index;
2476 unsigned level_type, mask_width, nitems;
2477 kmp_cpuid buf;
2478
2479 // New algorithm has known topology layers act as highest unknown topology
2480 // layers when unknown topology layers exist.
2481 // e.g., Suppose layers were SMT <X> CORE <Y> <Z> PACKAGE, where <X> <Y> <Z>
2482 // are unknown topology layers, Then SMT will take the characteristics of
2483 // (SMT x <X>) and CORE will take the characteristics of (CORE x <Y> x <Z>).
2484 // This eliminates unknown portions of the topology while still keeping the
2485 // correct structure.
2486 level = levels_index = 0;
2487 do {
2488 __kmp_x86_cpuid(leaf, level, &buf);
2489 level_type = __kmp_extract_bits<8, 15>(buf.ecx);
2490 mask_width = __kmp_extract_bits<0, 4>(buf.eax);
2491 nitems = __kmp_extract_bits<0, 15>(buf.ebx);
2492 if (level_type != INTEL_LEVEL_TYPE_INVALID && nitems == 0)
2493 return 0;
2494
2495 if (known_levels & (1ull << level_type)) {
2496 // Add a new level to the topology
2497 KMP_ASSERT(levels_index < INTEL_LEVEL_TYPE_LAST)if (!(levels_index < INTEL_LEVEL_TYPE_LAST)) { __kmp_debug_assert
("levels_index < INTEL_LEVEL_TYPE_LAST", "openmp/runtime/src/kmp_affinity.cpp"
, 2497); }
;
2498 levels[levels_index].level_type = level_type;
2499 levels[levels_index].mask_width = mask_width;
2500 levels[levels_index].nitems = nitems;
2501 levels_index++;
2502 } else {
2503 // If it is an unknown level, then logically move the previous layer up
2504 if (levels_index > 0) {
2505 levels[levels_index - 1].mask_width = mask_width;
2506 levels[levels_index - 1].nitems = nitems;
2507 }
2508 }
2509 level++;
2510 } while (level_type != INTEL_LEVEL_TYPE_INVALID);
2511
2512 // Set the masks to & with apicid
2513 for (unsigned i = 0; i < levels_index; ++i) {
2514 if (levels[i].level_type != INTEL_LEVEL_TYPE_INVALID) {
2515 levels[i].mask = ~((-1) << levels[i].mask_width);
2516 levels[i].cache_mask = (-1) << levels[i].mask_width;
2517 for (unsigned j = 0; j < i; ++j)
2518 levels[i].mask ^= levels[j].mask;
2519 } else {
2520 KMP_DEBUG_ASSERT(levels_index > 0)if (!(levels_index > 0)) { __kmp_debug_assert("levels_index > 0"
, "openmp/runtime/src/kmp_affinity.cpp", 2520); }
;
2521 levels[i].mask = (-1) << levels[i - 1].mask_width;
2522 levels[i].cache_mask = 0;
2523 }
2524 }
2525 return levels_index;
2526}
2527
2528static bool __kmp_affinity_create_x2apicid_map(kmp_i18n_id_t *const msg_id) {
2529
2530 cpuid_level_info_t levels[INTEL_LEVEL_TYPE_LAST];
2531 kmp_hw_t types[INTEL_LEVEL_TYPE_LAST];
2532 unsigned levels_index;
2533 kmp_cpuid buf;
2534 kmp_uint64 known_levels;
2535 int topology_leaf, highest_leaf, apic_id;
2536 int num_leaves;
2537 static int leaves[] = {0, 0};
2538
2539 kmp_i18n_id_t leaf_message_id;
2540
2541 KMP_BUILD_ASSERT(sizeof(known_levels) * CHAR_BIT > KMP_HW_LAST)static_assert(sizeof(known_levels) * 8 > KMP_HW_LAST, "Build condition error"
)
;
2542
2543 *msg_id = kmp_i18n_null;
2544 if (__kmp_affinity.flags.verbose) {
2545 KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(Decodingx2APIC))__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_AffInfoStr
, "KMP_AFFINITY", __kmp_i18n_catgets(kmp_i18n_str_Decodingx2APIC
)), __kmp_msg_null)
;
2546 }
2547
2548 // Figure out the known topology levels
2549 known_levels = 0ull;
2550 for (int i = 0; i < INTEL_LEVEL_TYPE_LAST; ++i) {
2551 if (__kmp_intel_type_2_topology_type(i) != KMP_HW_UNKNOWN) {
2552 known_levels |= (1ull << i);
2553 }
2554 }
2555
2556 // Get the highest cpuid leaf supported
2557 __kmp_x86_cpuid(0, 0, &buf);
2558 highest_leaf = buf.eax;
2559
2560 // If a specific topology method was requested, only allow that specific leaf
2561 // otherwise, try both leaves 31 and 11 in that order
2562 num_leaves = 0;
2563 if (__kmp_affinity_top_method == affinity_top_method_x2apicid) {
2564 num_leaves = 1;
2565 leaves[0] = 11;
2566 leaf_message_id = kmp_i18n_str_NoLeaf11Support;
2567 } else if (__kmp_affinity_top_method == affinity_top_method_x2apicid_1f) {
2568 num_leaves = 1;
2569 leaves[0] = 31;
2570 leaf_message_id = kmp_i18n_str_NoLeaf31Support;
2571 } else {
2572 num_leaves = 2;
2573 leaves[0] = 31;
2574 leaves[1] = 11;
2575 leaf_message_id = kmp_i18n_str_NoLeaf11Support;
2576 }
2577
2578 // Check to see if cpuid leaf 31 or 11 is supported.
2579 __kmp_nThreadsPerCore = nCoresPerPkg = nPackages = 1;
2580 topology_leaf = -1;
2581 for (int i = 0; i < num_leaves; ++i) {
2582 int leaf = leaves[i];
2583 if (highest_leaf < leaf)
2584 continue;
2585 __kmp_x86_cpuid(leaf, 0, &buf);
2586 if (buf.ebx == 0)
2587 continue;
2588 topology_leaf = leaf;
2589 levels_index = __kmp_x2apicid_get_levels(leaf, levels, known_levels);
2590 if (levels_index == 0)
2591 continue;
2592 break;
2593 }
2594 if (topology_leaf == -1 || levels_index == 0) {
2595 *msg_id = leaf_message_id;
2596 return false;
2597 }
2598 KMP_ASSERT(levels_index <= INTEL_LEVEL_TYPE_LAST)if (!(levels_index <= INTEL_LEVEL_TYPE_LAST)) { __kmp_debug_assert
("levels_index <= INTEL_LEVEL_TYPE_LAST", "openmp/runtime/src/kmp_affinity.cpp"
, 2598); }
;
2599
2600 // The algorithm used starts by setting the affinity to each available thread
2601 // and retrieving info from the cpuid instruction, so if we are not capable of
2602 // calling __kmp_get_system_affinity() and __kmp_get_system_affinity(), then
2603 // we need to do something else - use the defaults that we calculated from
2604 // issuing cpuid without binding to each proc.
2605 if (!KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) {
2606 // Hack to try and infer the machine topology using only the data
2607 // available from cpuid on the current thread, and __kmp_xproc.
2608 KMP_ASSERT(__kmp_affinity.type == affinity_none)if (!(__kmp_affinity.type == affinity_none)) { __kmp_debug_assert
("__kmp_affinity.type == affinity_none", "openmp/runtime/src/kmp_affinity.cpp"
, 2608); }
;
2609 for (unsigned i = 0; i < levels_index; ++i) {
2610 if (levels[i].level_type == INTEL_LEVEL_TYPE_SMT) {
2611 __kmp_nThreadsPerCore = levels[i].nitems;
2612 } else if (levels[i].level_type == INTEL_LEVEL_TYPE_CORE) {
2613 nCoresPerPkg = levels[i].nitems;
2614 }
2615 }
2616 __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
2617 nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
2618 return true;
2619 }
2620
2621 // Allocate the data structure to be returned.
2622 int depth = levels_index;
2623 for (int i = depth - 1, j = 0; i >= 0; --i, ++j)
2624 types[j] = __kmp_intel_type_2_topology_type(levels[i].level_type);
2625 __kmp_topology =
2626 kmp_topology_t::allocate(__kmp_avail_proc, levels_index, types);
2627
2628 // Insert equivalent cache types if they exist
2629 kmp_cache_info_t cache_info;
2630 for (size_t i = 0; i < cache_info.get_depth(); ++i) {
2631 const kmp_cache_info_t::info_t &info = cache_info[i];
2632 unsigned cache_mask = info.mask;
2633 unsigned cache_level = info.level;
2634 for (unsigned j = 0; j < levels_index; ++j) {
2635 unsigned hw_cache_mask = levels[j].cache_mask;
2636 kmp_hw_t cache_type = kmp_cache_info_t::get_topology_type(cache_level);
2637 if (hw_cache_mask == cache_mask && j < levels_index - 1) {
2638 kmp_hw_t type =
2639 __kmp_intel_type_2_topology_type(levels[j + 1].level_type);
2640 __kmp_topology->set_equivalent_type(cache_type, type);
2641 }
2642 }
2643 }
2644
2645 // From here on, we can assume that it is safe to call
2646 // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
2647 // __kmp_affinity.type = affinity_none.
2648
2649 // Save the affinity mask for the current thread.
2650 kmp_affinity_raii_t previous_affinity;
2651
2652 // Run through each of the available contexts, binding the current thread
2653 // to it, and obtaining the pertinent information using the cpuid instr.
2654 unsigned int proc;
2655 int hw_thread_index = 0;
2656 KMP_CPU_SET_ITERATE(proc, __kmp_affin_fullMask)for (proc = (__kmp_affin_fullMask)->begin(); (int)proc != (
__kmp_affin_fullMask)->end(); proc = (__kmp_affin_fullMask
)->next(proc))
{
2657 cpuid_level_info_t my_levels[INTEL_LEVEL_TYPE_LAST];
2658 unsigned my_levels_index;
2659
2660 // Skip this proc if it is not included in the machine model.
2661 if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)(__kmp_affin_fullMask)->is_set(proc)) {
2662 continue;
2663 }
2664 KMP_DEBUG_ASSERT(hw_thread_index < __kmp_avail_proc)if (!(hw_thread_index < __kmp_avail_proc)) { __kmp_debug_assert
("hw_thread_index < __kmp_avail_proc", "openmp/runtime/src/kmp_affinity.cpp"
, 2664); }
;
2665
2666 __kmp_affinity_dispatch->bind_thread(proc);
2667
2668 // New algorithm
2669 __kmp_x86_cpuid(topology_leaf, 0, &buf);
2670 apic_id = buf.edx;
2671 kmp_hw_thread_t &hw_thread = __kmp_topology->at(hw_thread_index);
2672 my_levels_index =
2673 __kmp_x2apicid_get_levels(topology_leaf, my_levels, known_levels);
2674 if (my_levels_index == 0 || my_levels_index != levels_index) {
2675 *msg_id = kmp_i18n_str_InvalidCpuidInfo;
2676 return false;
2677 }
2678 hw_thread.clear();
2679 hw_thread.os_id = proc;
2680 // Put in topology information
2681 for (unsigned j = 0, idx = depth - 1; j < my_levels_index; ++j, --idx) {
2682 hw_thread.ids[idx] = apic_id & my_levels[j].mask;
2683 if (j > 0) {
2684 hw_thread.ids[idx] >>= my_levels[j - 1].mask_width;
2685 }
2686 }
2687 // Hybrid information
2688 if (__kmp_is_hybrid_cpu() && highest_leaf >= 0x1a) {
2689 kmp_hw_core_type_t type;
2690 unsigned native_model_id;
2691 int efficiency;
2692 __kmp_get_hybrid_info(&type, &efficiency, &native_model_id);
2693 hw_thread.attrs.set_core_type(type);
2694 hw_thread.attrs.set_core_eff(efficiency);
2695 }
2696 hw_thread_index++;
2697 }
2698 KMP_ASSERT(hw_thread_index > 0)if (!(hw_thread_index > 0)) { __kmp_debug_assert("hw_thread_index > 0"
, "openmp/runtime/src/kmp_affinity.cpp", 2698); }
;
2699 __kmp_topology->sort_ids();
2700 if (!__kmp_topology->check_ids()) {
2701 kmp_topology_t::deallocate(__kmp_topology);
2702 __kmp_topology = nullptr;
2703 *msg_id = kmp_i18n_str_x2ApicIDsNotUnique;
2704 return false;
2705 }
2706 return true;
2707}
2708#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2709
2710#define osIdIndex0 0
2711#define threadIdIndex1 1
2712#define coreIdIndex2 2
2713#define pkgIdIndex3 3
2714#define nodeIdIndex4 4
2715
2716typedef unsigned *ProcCpuInfo;
2717static unsigned maxIndex = pkgIdIndex3;
2718
2719static int __kmp_affinity_cmp_ProcCpuInfo_phys_id(const void *a,
2720 const void *b) {
2721 unsigned i;
2722 const unsigned *aa = *(unsigned *const *)a;
2723 const unsigned *bb = *(unsigned *const *)b;
2724 for (i = maxIndex;; i--) {
2725 if (aa[i] < bb[i])
2726 return -1;
2727 if (aa[i] > bb[i])
2728 return 1;
2729 if (i == osIdIndex0)
2730 break;
2731 }
2732 return 0;
2733}
2734
2735#if KMP_USE_HIER_SCHED0
2736// Set the array sizes for the hierarchy layers
2737static void __kmp_dispatch_set_hierarchy_values() {
2738 // Set the maximum number of L1's to number of cores
2739 // Set the maximum number of L2's to to either number of cores / 2 for
2740 // Intel(R) Xeon Phi(TM) coprocessor formally codenamed Knights Landing
2741 // Or the number of cores for Intel(R) Xeon(R) processors
2742 // Set the maximum number of NUMA nodes and L3's to number of packages
2743 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1] =
2744 nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
2745 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L1 + 1] = __kmp_ncores;
2746#if KMP_ARCH_X86_641 && (KMP_OS_LINUX1 || KMP_OS_FREEBSD0 || KMP_OS_WINDOWS0) && \
2747 KMP_MIC_SUPPORTED((0 || 1) && (1 || 0))
2748 if (__kmp_mic_type >= mic3)
2749 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores / 2;
2750 else
2751#endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
2752 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores;
2753 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L3 + 1] = nPackages;
2754 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_NUMA + 1] = nPackages;
2755 __kmp_hier_max_units[kmp_hier_layer_e::LAYER_LOOP + 1] = 1;
2756 // Set the number of threads per unit
2757 // Number of hardware threads per L1/L2/L3/NUMA/LOOP
2758 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_THREAD + 1] = 1;
2759 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L1 + 1] =
2760 __kmp_nThreadsPerCore;
2761#if KMP_ARCH_X86_641 && (KMP_OS_LINUX1 || KMP_OS_FREEBSD0 || KMP_OS_WINDOWS0) && \
2762 KMP_MIC_SUPPORTED((0 || 1) && (1 || 0))
2763 if (__kmp_mic_type >= mic3)
2764 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
2765 2 * __kmp_nThreadsPerCore;
2766 else
2767#endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
2768 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
2769 __kmp_nThreadsPerCore;
2770 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L3 + 1] =
2771 nCoresPerPkg * __kmp_nThreadsPerCore;
2772 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_NUMA + 1] =
2773 nCoresPerPkg * __kmp_nThreadsPerCore;
2774 __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_LOOP + 1] =
2775 nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
2776}
2777
2778// Return the index into the hierarchy for this tid and layer type (L1, L2, etc)
2779// i.e., this thread's L1 or this thread's L2, etc.
2780int __kmp_dispatch_get_index(int tid, kmp_hier_layer_e type) {
2781 int index = type + 1;
2782 int num_hw_threads = __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1];
2783 KMP_DEBUG_ASSERT(type != kmp_hier_layer_e::LAYER_LAST)if (!(type != kmp_hier_layer_e::LAYER_LAST)) { __kmp_debug_assert
("type != kmp_hier_layer_e::LAYER_LAST", "openmp/runtime/src/kmp_affinity.cpp"
, 2783); }
;
2784 if (type == kmp_hier_layer_e::LAYER_THREAD)
2785 return tid;
2786 else if (type == kmp_hier_layer_e::LAYER_LOOP)
2787 return 0;
2788 KMP_DEBUG_ASSERT(__kmp_hier_max_units[index] != 0)if (!(__kmp_hier_max_units[index] != 0)) { __kmp_debug_assert
("__kmp_hier_max_units[index] != 0", "openmp/runtime/src/kmp_affinity.cpp"
, 2788); }
;
2789 if (tid >= num_hw_threads)
2790 tid = tid % num_hw_threads;
2791 return (tid / __kmp_hier_threads_per[index]) % __kmp_hier_max_units[index];
2792}
2793
2794// Return the number of t1's per t2
2795int __kmp_dispatch_get_t1_per_t2(kmp_hier_layer_e t1, kmp_hier_layer_e t2) {
2796 int i1 = t1 + 1;
2797 int i2 = t2 + 1;
2798 KMP_DEBUG_ASSERT(i1 <= i2)if (!(i1 <= i2)) { __kmp_debug_assert("i1 <= i2", "openmp/runtime/src/kmp_affinity.cpp"
, 2798); }
;
2799 KMP_DEBUG_ASSERT(t1 != kmp_hier_layer_e::LAYER_LAST)if (!(t1 != kmp_hier_layer_e::LAYER_LAST)) { __kmp_debug_assert
("t1 != kmp_hier_layer_e::LAYER_LAST", "openmp/runtime/src/kmp_affinity.cpp"
, 2799); }
;
2800 KMP_DEBUG_ASSERT(t2 != kmp_hier_layer_e::LAYER_LAST)if (!(t2 != kmp_hier_layer_e::LAYER_LAST)) { __kmp_debug_assert
("t2 != kmp_hier_layer_e::LAYER_LAST", "openmp/runtime/src/kmp_affinity.cpp"
, 2800); }
;
2801 KMP_DEBUG_ASSERT(__kmp_hier_threads_per[i1] != 0)if (!(__kmp_hier_threads_per[i1] != 0)) { __kmp_debug_assert(
"__kmp_hier_threads_per[i1] != 0", "openmp/runtime/src/kmp_affinity.cpp"
, 2801); }
;
2802 // (nthreads/t2) / (nthreads/t1) = t1 / t2
2803 return __kmp_hier_threads_per[i2] / __kmp_hier_threads_per[i1];
2804}
2805#endif // KMP_USE_HIER_SCHED
2806
2807static inline const char *__kmp_cpuinfo_get_filename() {
2808 const char *filename;
2809 if (__kmp_cpuinfo_file != nullptr)
2810 filename = __kmp_cpuinfo_file;
2811 else
2812 filename = "/proc/cpuinfo";
2813 return filename;
2814}
2815
2816static inline const char *__kmp_cpuinfo_get_envvar() {
2817 const char *envvar = nullptr;
2818 if (__kmp_cpuinfo_file != nullptr)
2819 envvar = "KMP_CPUINFO_FILE";
2820 return envvar;
2821}
2822
2823// Parse /proc/cpuinfo (or an alternate file in the same format) to obtain the
2824// affinity map.
2825static bool __kmp_affinity_create_cpuinfo_map(int *line,
2826 kmp_i18n_id_t *const msg_id) {
2827 const char *filename = __kmp_cpuinfo_get_filename();
2828 const char *envvar = __kmp_cpuinfo_get_envvar();
2829 *msg_id = kmp_i18n_null;
2830
2831 if (__kmp_affinity.flags.verbose) {
2832 KMP_INFORM(AffParseFilename, "KMP_AFFINITY", filename)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_AffParseFilename
, "KMP_AFFINITY", filename), __kmp_msg_null)
;
2833 }
2834
2835 kmp_safe_raii_file_t f(filename, "r", envvar);
2836
2837 // Scan of the file, and count the number of "processor" (osId) fields,
2838 // and find the highest value of <n> for a node_<n> field.
2839 char buf[256];
2840 unsigned num_records = 0;
2841 while (!feof(f)) {
2842 buf[sizeof(buf) - 1] = 1;
2843 if (!fgets(buf, sizeof(buf), f)) {
2844 // Read errors presumably because of EOF
2845 break;
2846 }
2847
2848 char s1[] = "processor";
2849 if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
2850 num_records++;
2851 continue;
2852 }
2853
2854 // FIXME - this will match "node_<n> <garbage>"
2855 unsigned level;
2856 if (KMP_SSCANFsscanf(buf, "node_%u id", &level) == 1) {
2857 // validate the input fisrt:
2858 if (level > (unsigned)__kmp_xproc) { // level is too big
2859 level = __kmp_xproc;
2860 }
2861 if (nodeIdIndex4 + level >= maxIndex) {
2862 maxIndex = nodeIdIndex4 + level;
2863 }
2864 continue;
2865 }
2866 }
2867
2868 // Check for empty file / no valid processor records, or too many. The number
2869 // of records can't exceed the number of valid bits in the affinity mask.
2870 if (num_records == 0) {
2871 *msg_id = kmp_i18n_str_NoProcRecords;
2872 return false;
2873 }
2874 if (num_records > (unsigned)__kmp_xproc) {
2875 *msg_id = kmp_i18n_str_TooManyProcRecords;
2876 return false;
2877 }
2878
2879 // Set the file pointer back to the beginning, so that we can scan the file
2880 // again, this time performing a full parse of the data. Allocate a vector of
2881 // ProcCpuInfo object, where we will place the data. Adding an extra element
2882 // at the end allows us to remove a lot of extra checks for termination
2883 // conditions.
2884 if (fseek(f, 0, SEEK_SET0) != 0) {
2885 *msg_id = kmp_i18n_str_CantRewindCpuinfo;
2886 return false;
2887 }
2888
2889 // Allocate the array of records to store the proc info in. The dummy
2890 // element at the end makes the logic in filling them out easier to code.
2891 unsigned **threadInfo =
2892 (unsigned **)__kmp_allocate((num_records + 1) * sizeof(unsigned *))___kmp_allocate(((num_records + 1) * sizeof(unsigned *)), "openmp/runtime/src/kmp_affinity.cpp"
, 2892)
;
2893 unsigned i;
2894 for (i = 0; i <= num_records; i++) {
2895 threadInfo[i] =
2896 (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned))___kmp_allocate(((maxIndex + 1) * sizeof(unsigned)), "openmp/runtime/src/kmp_affinity.cpp"
, 2896)
;
2897 }
2898
2899#define CLEANUP_THREAD_INFOfor (i = 0; i <= num_records; i++) { ___kmp_free((threadInfo
[i]), "openmp/runtime/src/kmp_affinity.cpp", 2899); } ___kmp_free
((threadInfo), "openmp/runtime/src/kmp_affinity.cpp", 2899);
\
2900 for (i = 0; i <= num_records; i++) { \
2901 __kmp_free(threadInfo[i])___kmp_free((threadInfo[i]), "openmp/runtime/src/kmp_affinity.cpp"
, 2901)
; \
2902 } \
2903 __kmp_free(threadInfo)___kmp_free((threadInfo), "openmp/runtime/src/kmp_affinity.cpp"
, 2903)
;
2904
2905 // A value of UINT_MAX means that we didn't find the field
2906 unsigned __index;
2907
2908#define INIT_PROC_INFO(p)for (__index = 0; __index <= maxIndex; __index++) { (p)[__index
] = (2147483647 *2U +1U); }
\
2909 for (__index = 0; __index <= maxIndex; __index++) { \
2910 (p)[__index] = UINT_MAX(2147483647 *2U +1U); \
2911 }
2912
2913 for (i = 0; i <= num_records; i++) {
2914 INIT_PROC_INFO(threadInfo[i])for (__index = 0; __index <= maxIndex; __index++) { (threadInfo
[i])[__index] = (2147483647 *2U +1U); }
;
2915 }
2916
2917 unsigned num_avail = 0;
2918 *line = 0;
2919 while (!feof(f)) {
2920 // Create an inner scoping level, so that all the goto targets at the end of
2921 // the loop appear in an outer scoping level. This avoids warnings about
2922 // jumping past an initialization to a target in the same block.
2923 {
2924 buf[sizeof(buf) - 1] = 1;
2925 bool long_line = false;
2926 if (!fgets(buf, sizeof(buf), f)) {
2927 // Read errors presumably because of EOF
2928 // If there is valid data in threadInfo[num_avail], then fake
2929 // a blank line in ensure that the last address gets parsed.
2930 bool valid = false;
2931 for (i = 0; i <= maxIndex; i++) {
2932 if (threadInfo[num_avail][i] != UINT_MAX(2147483647 *2U +1U)) {
2933 valid = true;
2934 }
2935 }
2936 if (!valid) {
2937 break;
2938 }
2939 buf[0] = 0;
2940 } else if (!buf[sizeof(buf) - 1]) {
2941 // The line is longer than the buffer. Set a flag and don't
2942 // emit an error if we were going to ignore the line, anyway.
2943 long_line = true;
2944
2945#define CHECK_LINEif (long_line) { for (i = 0; i <= num_records; i++) { ___kmp_free
((threadInfo[i]), "openmp/runtime/src/kmp_affinity.cpp", 2945
); } ___kmp_free((threadInfo), "openmp/runtime/src/kmp_affinity.cpp"
, 2945);; *msg_id = kmp_i18n_str_LongLineCpuinfo; return false
; }
\
2946 if (long_line) { \
2947 CLEANUP_THREAD_INFOfor (i = 0; i <= num_records; i++) { ___kmp_free((threadInfo
[i]), "openmp/runtime/src/kmp_affinity.cpp", 2947); } ___kmp_free
((threadInfo), "openmp/runtime/src/kmp_affinity.cpp", 2947);
; \
2948 *msg_id = kmp_i18n_str_LongLineCpuinfo; \
2949 return false; \
2950 }
2951 }
2952 (*line)++;
2953
2954#if KMP_ARCH_LOONGARCH640
2955 // The parsing logic of /proc/cpuinfo in this function highly depends on
2956 // the blank lines between each processor info block. But on LoongArch a
2957 // blank line exists before the first processor info block (i.e. after the
2958 // "system type" line). This blank line was added because the "system
2959 // type" line is unrelated to any of the CPUs. We must skip this line so
2960 // that the original logic works on LoongArch.
2961 if (*buf == '\n' && *line == 2)
2962 continue;
2963#endif
2964
2965 char s1[] = "processor";
2966 if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
2967 CHECK_LINEif (long_line) { for (i = 0; i <= num_records; i++) { ___kmp_free
((threadInfo[i]), "openmp/runtime/src/kmp_affinity.cpp", 2967
); } ___kmp_free((threadInfo), "openmp/runtime/src/kmp_affinity.cpp"
, 2967);; *msg_id = kmp_i18n_str_LongLineCpuinfo; return false
; }
;
2968 char *p = strchr(buf + sizeof(s1) - 1, ':');
2969 unsigned val;
2970 if ((p == NULL__null) || (KMP_SSCANFsscanf(p + 1, "%u\n", &val) != 1))
2971 goto no_val;
2972 if (threadInfo[num_avail][osIdIndex0] != UINT_MAX(2147483647 *2U +1U))
2973#if KMP_ARCH_AARCH640
2974 // Handle the old AArch64 /proc/cpuinfo layout differently,
2975 // it contains all of the 'processor' entries listed in a
2976 // single 'Processor' section, therefore the normal looking
2977 // for duplicates in that section will always fail.
2978 num_avail++;
2979#else
2980 goto dup_field;
2981#endif
2982 threadInfo[num_avail][osIdIndex0] = val;
2983#if KMP_OS_LINUX1 && !(KMP_ARCH_X860 || KMP_ARCH_X86_641)
2984 char path[256];
2985 KMP_SNPRINTFsnprintf(
2986 path, sizeof(path),
2987 "/sys/devices/system/cpu/cpu%u/topology/physical_package_id",
2988 threadInfo[num_avail][osIdIndex0]);
2989 __kmp_read_from_file(path, "%u", &threadInfo[num_avail][pkgIdIndex3]);
2990
2991 KMP_SNPRINTFsnprintf(path, sizeof(path),
2992 "/sys/devices/system/cpu/cpu%u/topology/core_id",
2993 threadInfo[num_avail][osIdIndex0]);
2994 __kmp_read_from_file(path, "%u", &threadInfo[num_avail][coreIdIndex2]);
2995 continue;
2996#else
2997 }
2998 char s2[] = "physical id";
2999 if (strncmp(buf, s2, sizeof(s2) - 1) == 0) {
3000 CHECK_LINEif (long_line) { for (i = 0; i <= num_records; i++) { ___kmp_free
((threadInfo[i]), "openmp/runtime/src/kmp_affinity.cpp", 3000
); } ___kmp_free((threadInfo), "openmp/runtime/src/kmp_affinity.cpp"
, 3000);; *msg_id = kmp_i18n_str_LongLineCpuinfo; return false
; }
;
3001 char *p = strchr(buf + sizeof(s2) - 1, ':');
3002 unsigned val;
3003 if ((p == NULL__null) || (KMP_SSCANFsscanf(p + 1, "%u\n", &val) != 1))
3004 goto no_val;
3005 if (threadInfo[num_avail][pkgIdIndex3] != UINT_MAX(2147483647 *2U +1U))
3006 goto dup_field;
3007 threadInfo[num_avail][pkgIdIndex3] = val;
3008 continue;
3009 }
3010 char s3[] = "core id";
3011 if (strncmp(buf, s3, sizeof(s3) - 1) == 0) {
3012 CHECK_LINEif (long_line) { for (i = 0; i <= num_records; i++) { ___kmp_free
((threadInfo[i]), "openmp/runtime/src/kmp_affinity.cpp", 3012
); } ___kmp_free((threadInfo), "openmp/runtime/src/kmp_affinity.cpp"
, 3012);; *msg_id = kmp_i18n_str_LongLineCpuinfo; return false
; }
;
3013 char *p = strchr(buf + sizeof(s3) - 1, ':');
3014 unsigned val;
3015 if ((p == NULL__null) || (KMP_SSCANFsscanf(p + 1, "%u\n", &val) != 1))
3016 goto no_val;
3017 if (threadInfo[num_avail][coreIdIndex2] != UINT_MAX(2147483647 *2U +1U))
3018 goto dup_field;
3019 threadInfo[num_avail][coreIdIndex2] = val;
3020 continue;
3021#endif // KMP_OS_LINUX && USE_SYSFS_INFO
3022 }
3023 char s4[] = "thread id";
3024 if (strncmp(buf, s4, sizeof(s4) - 1) == 0) {
3025 CHECK_LINEif (long_line) { for (i = 0; i <= num_records; i++) { ___kmp_free
((threadInfo[i]), "openmp/runtime/src/kmp_affinity.cpp", 3025
); } ___kmp_free((threadInfo), "openmp/runtime/src/kmp_affinity.cpp"
, 3025);; *msg_id = kmp_i18n_str_LongLineCpuinfo; return false
; }
;
3026 char *p = strchr(buf + sizeof(s4) - 1, ':');
3027 unsigned val;
3028 if ((p == NULL__null) || (KMP_SSCANFsscanf(p + 1, "%u\n", &val) != 1))
3029 goto no_val;
3030 if (threadInfo[num_avail][threadIdIndex1] != UINT_MAX(2147483647 *2U +1U))
3031 goto dup_field;
3032 threadInfo[num_avail][threadIdIndex1] = val;
3033 continue;
3034 }
3035 unsigned level;
3036 if (KMP_SSCANFsscanf(buf, "node_%u id", &level) == 1) {
3037 CHECK_LINEif (long_line) { for (i = 0; i <= num_records; i++) { ___kmp_free
((threadInfo[i]), "openmp/runtime/src/kmp_affinity.cpp", 3037
); } ___kmp_free((threadInfo), "openmp/runtime/src/kmp_affinity.cpp"
, 3037);; *msg_id = kmp_i18n_str_LongLineCpuinfo; return false
; }
;
3038 char *p = strchr(buf + sizeof(s4) - 1, ':');
3039 unsigned val;
3040 if ((p == NULL__null) || (KMP_SSCANFsscanf(p + 1, "%u\n", &val) != 1))
3041 goto no_val;
3042 // validate the input before using level:
3043 if (level > (unsigned)__kmp_xproc) { // level is too big
3044 level = __kmp_xproc;
3045 }
3046 if (threadInfo[num_avail][nodeIdIndex4 + level] != UINT_MAX(2147483647 *2U +1U))
3047 goto dup_field;
3048 threadInfo[num_avail][nodeIdIndex4 + level] = val;
3049 continue;
3050 }
3051
3052 // We didn't recognize the leading token on the line. There are lots of
3053 // leading tokens that we don't recognize - if the line isn't empty, go on
3054 // to the next line.
3055 if ((*buf != 0) && (*buf != '\n')) {
3056 // If the line is longer than the buffer, read characters
3057 // until we find a newline.
3058 if (long_line) {
3059 int ch;
3060 while (((ch = fgetc(f)) != EOF(-1)) && (ch != '\n'))
3061 ;
3062 }
3063 continue;
3064 }
3065
3066 // A newline has signalled the end of the processor record.
3067 // Check that there aren't too many procs specified.
3068 if ((int)num_avail == __kmp_xproc) {
3069 CLEANUP_THREAD_INFOfor (i = 0; i <= num_records; i++) { ___kmp_free((threadInfo
[i]), "openmp/runtime/src/kmp_affinity.cpp", 3069); } ___kmp_free
((threadInfo), "openmp/runtime/src/kmp_affinity.cpp", 3069);
;
3070 *msg_id = kmp_i18n_str_TooManyEntries;
3071 return false;
3072 }
3073
3074 // Check for missing fields. The osId field must be there, and we
3075 // currently require that the physical id field is specified, also.
3076 if (threadInfo[num_avail][osIdIndex0] == UINT_MAX(2147483647 *2U +1U)) {
3077 CLEANUP_THREAD_INFOfor (i = 0; i <= num_records; i++) { ___kmp_free((threadInfo
[i]), "openmp/runtime/src/kmp_affinity.cpp", 3077); } ___kmp_free
((threadInfo), "openmp/runtime/src/kmp_affinity.cpp", 3077);
;
3078 *msg_id = kmp_i18n_str_MissingProcField;
3079 return false;
3080 }
3081 if (threadInfo[0][pkgIdIndex3] == UINT_MAX(2147483647 *2U +1U)) {
3082 CLEANUP_THREAD_INFOfor (i = 0; i <= num_records; i++) { ___kmp_free((threadInfo
[i]), "openmp/runtime/src/kmp_affinity.cpp", 3082); } ___kmp_free
((threadInfo), "openmp/runtime/src/kmp_affinity.cpp", 3082);
;
3083 *msg_id = kmp_i18n_str_MissingPhysicalIDField;
3084 return false;
3085 }
3086
3087 // Skip this proc if it is not included in the machine model.
3088 if (KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0) &&
3089 !KMP_CPU_ISSET(threadInfo[num_avail][osIdIndex],(__kmp_affin_fullMask)->is_set(threadInfo[num_avail][0])
3090 __kmp_affin_fullMask)(__kmp_affin_fullMask)->is_set(threadInfo[num_avail][0])) {
3091 INIT_PROC_INFO(threadInfo[num_avail])for (__index = 0; __index <= maxIndex; __index++) { (threadInfo
[num_avail])[__index] = (2147483647 *2U +1U); }
;
3092 continue;
3093 }
3094
3095 // We have a successful parse of this proc's info.
3096 // Increment the counter, and prepare for the next proc.
3097 num_avail++;
3098 KMP_ASSERT(num_avail <= num_records)if (!(num_avail <= num_records)) { __kmp_debug_assert("num_avail <= num_records"
, "openmp/runtime/src/kmp_affinity.cpp", 3098); }
;
3099 INIT_PROC_INFO(threadInfo[num_avail])for (__index = 0; __index <= maxIndex; __index++) { (threadInfo
[num_avail])[__index] = (2147483647 *2U +1U); }
;
3100 }
3101 continue;
3102
3103 no_val:
3104 CLEANUP_THREAD_INFOfor (i = 0; i <= num_records; i++) { ___kmp_free((threadInfo
[i]), "openmp/runtime/src/kmp_affinity.cpp", 3104); } ___kmp_free
((threadInfo), "openmp/runtime/src/kmp_affinity.cpp", 3104);
;
3105 *msg_id = kmp_i18n_str_MissingValCpuinfo;
3106 return false;
3107
3108 dup_field:
3109 CLEANUP_THREAD_INFOfor (i = 0; i <= num_records; i++) { ___kmp_free((threadInfo
[i]), "openmp/runtime/src/kmp_affinity.cpp", 3109); } ___kmp_free
((threadInfo), "openmp/runtime/src/kmp_affinity.cpp", 3109);
;
3110 *msg_id = kmp_i18n_str_DuplicateFieldCpuinfo;
3111 return false;
3112 }
3113 *line = 0;
3114
3115#if KMP_MIC0 && REDUCE_TEAM_SIZE
3116 unsigned teamSize = 0;
3117#endif // KMP_MIC && REDUCE_TEAM_SIZE
3118
3119 // check for num_records == __kmp_xproc ???
3120
3121 // If it is configured to omit the package level when there is only a single
3122 // package, the logic at the end of this routine won't work if there is only a
3123 // single thread
3124 KMP_ASSERT(num_avail > 0)if (!(num_avail > 0)) { __kmp_debug_assert("num_avail > 0"
, "openmp/runtime/src/kmp_affinity.cpp", 3124); }
;
3125 KMP_ASSERT(num_avail <= num_records)if (!(num_avail <= num_records)) { __kmp_debug_assert("num_avail <= num_records"
, "openmp/runtime/src/kmp_affinity.cpp", 3125); }
;
3126
3127 // Sort the threadInfo table by physical Id.
3128 qsort(threadInfo, num_avail, sizeof(*threadInfo),
3129 __kmp_affinity_cmp_ProcCpuInfo_phys_id);
3130
3131 // The table is now sorted by pkgId / coreId / threadId, but we really don't
3132 // know the radix of any of the fields. pkgId's may be sparsely assigned among
3133 // the chips on a system. Although coreId's are usually assigned
3134 // [0 .. coresPerPkg-1] and threadId's are usually assigned
3135 // [0..threadsPerCore-1], we don't want to make any such assumptions.
3136 //
3137 // For that matter, we don't know what coresPerPkg and threadsPerCore (or the
3138 // total # packages) are at this point - we want to determine that now. We
3139 // only have an upper bound on the first two figures.
3140 unsigned *counts =
3141 (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned))___kmp_allocate(((maxIndex + 1) * sizeof(unsigned)), "openmp/runtime/src/kmp_affinity.cpp"
, 3141)
;
3142 unsigned *maxCt =
3143 (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned))___kmp_allocate(((maxIndex + 1) * sizeof(unsigned)), "openmp/runtime/src/kmp_affinity.cpp"
, 3143)
;
3144 unsigned *totals =
3145 (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned))___kmp_allocate(((maxIndex + 1) * sizeof(unsigned)), "openmp/runtime/src/kmp_affinity.cpp"
, 3145)
;
3146 unsigned *lastId =
3147 (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned))___kmp_allocate(((maxIndex + 1) * sizeof(unsigned)), "openmp/runtime/src/kmp_affinity.cpp"
, 3147)
;
3148
3149 bool assign_thread_ids = false;
3150 unsigned threadIdCt;
3151 unsigned index;
3152
3153restart_radix_check:
3154 threadIdCt = 0;
3155
3156 // Initialize the counter arrays with data from threadInfo[0].
3157 if (assign_thread_ids) {
3158 if (threadInfo[0][threadIdIndex1] == UINT_MAX(2147483647 *2U +1U)) {
3159 threadInfo[0][threadIdIndex1] = threadIdCt++;
3160 } else if (threadIdCt <= threadInfo[0][threadIdIndex1]) {
3161 threadIdCt = threadInfo[0][threadIdIndex1] + 1;
3162 }
3163 }
3164 for (index = 0; index <= maxIndex; index++) {
3165 counts[index] = 1;
3166 maxCt[index] = 1;
3167 totals[index] = 1;
3168 lastId[index] = threadInfo[0][index];
3169 ;
3170 }
3171
3172 // Run through the rest of the OS procs.
3173 for (i = 1; i < num_avail; i++) {
3174 // Find the most significant index whose id differs from the id for the
3175 // previous OS proc.
3176 for (index = maxIndex; index >= threadIdIndex1; index--) {
3177 if (assign_thread_ids && (index == threadIdIndex1)) {
3178 // Auto-assign the thread id field if it wasn't specified.
3179 if (threadInfo[i][threadIdIndex1] == UINT_MAX(2147483647 *2U +1U)) {
3180 threadInfo[i][threadIdIndex1] = threadIdCt++;
3181 }
3182 // Apparently the thread id field was specified for some entries and not
3183 // others. Start the thread id counter off at the next higher thread id.
3184 else if (threadIdCt <= threadInfo[i][threadIdIndex1]) {
3185 threadIdCt = threadInfo[i][threadIdIndex1] + 1;
3186 }
3187 }
3188 if (threadInfo[i][index] != lastId[index]) {
3189 // Run through all indices which are less significant, and reset the
3190 // counts to 1. At all levels up to and including index, we need to
3191 // increment the totals and record the last id.
3192 unsigned index2;
3193 for (index2 = threadIdIndex1; index2 < index; index2++) {
3194 totals[index2]++;
3195 if (counts[index2] > maxCt[index2]) {
3196 maxCt[index2] = counts[index2];
3197 }
3198 counts[index2] = 1;
3199 lastId[index2] = threadInfo[i][index2];
3200 }
3201 counts[index]++;
3202 totals[index]++;
3203 lastId[index] = threadInfo[i][index];
3204
3205 if (assign_thread_ids && (index > threadIdIndex1)) {
3206
3207#if KMP_MIC0 && REDUCE_TEAM_SIZE
3208 // The default team size is the total #threads in the machine
3209 // minus 1 thread for every core that has 3 or more threads.
3210 teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
3211#endif // KMP_MIC && REDUCE_TEAM_SIZE
3212
3213 // Restart the thread counter, as we are on a new core.
3214 threadIdCt = 0;
3215
3216 // Auto-assign the thread id field if it wasn't specified.
3217 if (threadInfo[i][threadIdIndex1] == UINT_MAX(2147483647 *2U +1U)) {
3218 threadInfo[i][threadIdIndex1] = threadIdCt++;
3219 }
3220
3221 // Apparently the thread id field was specified for some entries and
3222 // not others. Start the thread id counter off at the next higher
3223 // thread id.
3224 else if (threadIdCt <= threadInfo[i][threadIdIndex1]) {
3225 threadIdCt = threadInfo[i][threadIdIndex1] + 1;
3226 }
3227 }
3228 break;
3229 }
3230 }
3231 if (index < threadIdIndex1) {
3232 // If thread ids were specified, it is an error if they are not unique.
3233 // Also, check that we waven't already restarted the loop (to be safe -
3234 // shouldn't need to).
3235 if ((threadInfo[i][threadIdIndex1] != UINT_MAX(2147483647 *2U +1U)) || assign_thread_ids) {
3236 __kmp_free(lastId)___kmp_free((lastId), "openmp/runtime/src/kmp_affinity.cpp", 3236
)
;
3237 __kmp_free(totals)___kmp_free((totals), "openmp/runtime/src/kmp_affinity.cpp", 3237
)
;
3238 __kmp_free(maxCt)___kmp_free((maxCt), "openmp/runtime/src/kmp_affinity.cpp", 3238
)
;
3239 __kmp_free(counts)___kmp_free((counts), "openmp/runtime/src/kmp_affinity.cpp", 3239
)
;
3240 CLEANUP_THREAD_INFOfor (i = 0; i <= num_records; i++) { ___kmp_free((threadInfo
[i]), "openmp/runtime/src/kmp_affinity.cpp", 3240); } ___kmp_free
((threadInfo), "openmp/runtime/src/kmp_affinity.cpp", 3240);
;
3241 *msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
3242 return false;
3243 }
3244
3245 // If the thread ids were not specified and we see entries entries that
3246 // are duplicates, start the loop over and assign the thread ids manually.
3247 assign_thread_ids = true;
3248 goto restart_radix_check;
3249 }
3250 }
3251
3252#if KMP_MIC0 && REDUCE_TEAM_SIZE
3253 // The default team size is the total #threads in the machine
3254 // minus 1 thread for every core that has 3 or more threads.
3255 teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
3256#endif // KMP_MIC && REDUCE_TEAM_SIZE
3257
3258 for (index = threadIdIndex1; index <= maxIndex; index++) {
3259 if (counts[index] > maxCt[index]) {
3260 maxCt[index] = counts[index];
3261 }
3262 }
3263
3264 __kmp_nThreadsPerCore = maxCt[threadIdIndex1];
3265 nCoresPerPkg = maxCt[coreIdIndex2];
3266 nPackages = totals[pkgIdIndex3];
3267
3268 // When affinity is off, this routine will still be called to set
3269 // __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
3270 // Make sure all these vars are set correctly, and return now if affinity is
3271 // not enabled.
3272 __kmp_ncores = totals[coreIdIndex2];
3273 if (!KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) {
3274 KMP_ASSERT(__kmp_affinity.type == affinity_none)if (!(__kmp_affinity.type == affinity_none)) { __kmp_debug_assert
("__kmp_affinity.type == affinity_none", "openmp/runtime/src/kmp_affinity.cpp"
, 3274); }
;
3275 return true;
3276 }
3277
3278#if KMP_MIC0 && REDUCE_TEAM_SIZE
3279 // Set the default team size.
3280 if ((__kmp_dflt_team_nth == 0) && (teamSize > 0)) {
3281 __kmp_dflt_team_nth = teamSize;
3282 KA_TRACE(20, ("__kmp_affinity_create_cpuinfo_map: setting "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_affinity_create_cpuinfo_map: setting "
"__kmp_dflt_team_nth = %d\n", __kmp_dflt_team_nth); }
3283 "__kmp_dflt_team_nth = %d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_affinity_create_cpuinfo_map: setting "
"__kmp_dflt_team_nth = %d\n", __kmp_dflt_team_nth); }
3284 __kmp_dflt_team_nth))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_affinity_create_cpuinfo_map: setting "
"__kmp_dflt_team_nth = %d\n", __kmp_dflt_team_nth); }
;
3285 }
3286#endif // KMP_MIC && REDUCE_TEAM_SIZE
3287
3288 KMP_DEBUG_ASSERT(num_avail == (unsigned)__kmp_avail_proc)if (!(num_avail == (unsigned)__kmp_avail_proc)) { __kmp_debug_assert
("num_avail == (unsigned)__kmp_avail_proc", "openmp/runtime/src/kmp_affinity.cpp"
, 3288); }
;
3289
3290 // Count the number of levels which have more nodes at that level than at the
3291 // parent's level (with there being an implicit root node of the top level).
3292 // This is equivalent to saying that there is at least one node at this level
3293 // which has a sibling. These levels are in the map, and the package level is
3294 // always in the map.
3295 bool *inMap = (bool *)__kmp_allocate((maxIndex + 1) * sizeof(bool))___kmp_allocate(((maxIndex + 1) * sizeof(bool)), "openmp/runtime/src/kmp_affinity.cpp"
, 3295)
;
3296 for (index = threadIdIndex1; index < maxIndex; index++) {
3297 KMP_ASSERT(totals[index] >= totals[index + 1])if (!(totals[index] >= totals[index + 1])) { __kmp_debug_assert
("totals[index] >= totals[index + 1]", "openmp/runtime/src/kmp_affinity.cpp"
, 3297); }
;
3298 inMap[index] = (totals[index] > totals[index + 1]);
3299 }
3300 inMap[maxIndex] = (totals[maxIndex] > 1);
3301 inMap[pkgIdIndex3] = true;
3302 inMap[coreIdIndex2] = true;
3303 inMap[threadIdIndex1] = true;
3304
3305 int depth = 0;
3306 int idx = 0;
3307 kmp_hw_t types[KMP_HW_LAST];
3308 int pkgLevel = -1;
3309 int coreLevel = -1;
3310 int threadLevel = -1;
3311 for (index = threadIdIndex1; index <= maxIndex; index++) {
3312 if (inMap[index]) {
3313 depth++;
3314 }
3315 }
3316 if (inMap[pkgIdIndex3]) {
3317 pkgLevel = idx;
3318 types[idx++] = KMP_HW_SOCKET;
3319 }
3320 if (inMap[coreIdIndex2]) {
3321 coreLevel = idx;
3322 types[idx++] = KMP_HW_CORE;
3323 }
3324 if (inMap[threadIdIndex1]) {
3325 threadLevel = idx;
3326 types[idx++] = KMP_HW_THREAD;
3327 }
3328 KMP_ASSERT(depth > 0)if (!(depth > 0)) { __kmp_debug_assert("depth > 0", "openmp/runtime/src/kmp_affinity.cpp"
, 3328); }
;
3329
3330 // Construct the data structure that is to be returned.
3331 __kmp_topology = kmp_topology_t::allocate(num_avail, depth, types);
3332
3333 for (i = 0; i < num_avail; ++i) {
3334 unsigned os = threadInfo[i][osIdIndex0];
3335 int src_index;
3336 kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
3337 hw_thread.clear();
3338 hw_thread.os_id = os;
3339
3340 idx = 0;
3341 for (src_index = maxIndex; src_index >= threadIdIndex1; src_index--) {
3342 if (!inMap[src_index]) {
3343 continue;
3344 }
3345 if (src_index == pkgIdIndex3) {
3346 hw_thread.ids[pkgLevel] = threadInfo[i][src_index];
3347 } else if (src_index == coreIdIndex2) {
3348 hw_thread.ids[coreLevel] = threadInfo[i][src_index];
3349 } else if (src_index == threadIdIndex1) {
3350 hw_thread.ids[threadLevel] = threadInfo[i][src_index];
3351 }
3352 }
3353 }
3354
3355 __kmp_free(inMap)___kmp_free((inMap), "openmp/runtime/src/kmp_affinity.cpp", 3355
)
;
3356 __kmp_free(lastId)___kmp_free((lastId), "openmp/runtime/src/kmp_affinity.cpp", 3356
)
;
3357 __kmp_free(totals)___kmp_free((totals), "openmp/runtime/src/kmp_affinity.cpp", 3357
)
;
3358 __kmp_free(maxCt)___kmp_free((maxCt), "openmp/runtime/src/kmp_affinity.cpp", 3358
)
;
3359 __kmp_free(counts)___kmp_free((counts), "openmp/runtime/src/kmp_affinity.cpp", 3359
)
;
3360 CLEANUP_THREAD_INFOfor (i = 0; i <= num_records; i++) { ___kmp_free((threadInfo
[i]), "openmp/runtime/src/kmp_affinity.cpp", 3360); } ___kmp_free
((threadInfo), "openmp/runtime/src/kmp_affinity.cpp", 3360);
;
3361 __kmp_topology->sort_ids();
3362 if (!__kmp_topology->check_ids()) {
3363 kmp_topology_t::deallocate(__kmp_topology);
3364 __kmp_topology = nullptr;
3365 *msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
3366 return false;
3367 }
3368 return true;
3369}
3370
3371// Create and return a table of affinity masks, indexed by OS thread ID.
3372// This routine handles OR'ing together all the affinity masks of threads
3373// that are sufficiently close, if granularity > fine.
3374static void __kmp_create_os_id_masks(unsigned *numUnique,
3375 kmp_affinity_t &affinity) {
3376 // First form a table of affinity masks in order of OS thread id.
3377 int maxOsId;
3378 int i;
3379 int numAddrs = __kmp_topology->get_num_hw_threads();
3380 int depth = __kmp_topology->get_depth();
3381 const char *env_var = affinity.env_var;
3382 KMP_ASSERT(numAddrs)if (!(numAddrs)) { __kmp_debug_assert("numAddrs", "openmp/runtime/src/kmp_affinity.cpp"
, 3382); }
;
3383 KMP_ASSERT(depth)if (!(depth)) { __kmp_debug_assert("depth", "openmp/runtime/src/kmp_affinity.cpp"
, 3383); }
;
3384
3385 maxOsId = 0;
3386 for (i = numAddrs - 1;; --i) {
3387 int osId = __kmp_topology->at(i).os_id;
3388 if (osId > maxOsId) {
3389 maxOsId = osId;
3390 }
3391 if (i == 0)
3392 break;
3393 }
3394 affinity.num_os_id_masks = maxOsId + 1;
3395 KMP_CPU_ALLOC_ARRAY(affinity.os_id_masks, affinity.num_os_id_masks)(affinity.os_id_masks = __kmp_affinity_dispatch->allocate_mask_array
(affinity.num_os_id_masks))
;
3396 KMP_ASSERT(affinity.gran_levels >= 0)if (!(affinity.gran_levels >= 0)) { __kmp_debug_assert("affinity.gran_levels >= 0"
, "openmp/runtime/src/kmp_affinity.cpp", 3396); }
;
3397 if (affinity.flags.verbose && (affinity.gran_levels > 0)) {
3398 KMP_INFORM(ThreadsMigrate, env_var, affinity.gran_levels)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_ThreadsMigrate
, env_var, affinity.gran_levels), __kmp_msg_null)
;
3399 }
3400 if (affinity.gran_levels >= (int)depth) {
3401 KMP_AFF_WARNING(affinity, AffThreadsMayMigrate)if (affinity.flags.verbose || (affinity.flags.warnings &&
(affinity.type != affinity_none))) { __kmp_msg(kmp_ms_warning
, __kmp_msg_format(kmp_i18n_msg_AffThreadsMayMigrate), __kmp_msg_null
); }
;
3402 }
3403
3404 // Run through the table, forming the masks for all threads on each core.
3405 // Threads on the same core will have identical kmp_hw_thread_t objects, not
3406 // considering the last level, which must be the thread id. All threads on a
3407 // core will appear consecutively.
3408 int unique = 0;
3409 int j = 0; // index of 1st thread on core
3410 int leader = 0;
3411 kmp_affin_mask_t *sum;
3412 KMP_CPU_ALLOC_ON_STACK(sum)(sum = __kmp_affinity_dispatch->allocate_mask());
3413 KMP_CPU_ZERO(sum)(sum)->zero();
3414 KMP_CPU_SET(__kmp_topology->at(0).os_id, sum)(sum)->set(__kmp_topology->at(0).os_id);
3415 for (i = 1; i < numAddrs; i++) {
3416 // If this thread is sufficiently close to the leader (within the
3417 // granularity setting), then set the bit for this os thread in the
3418 // affinity mask for this group, and go on to the next thread.
3419 if (__kmp_topology->is_close(leader, i, affinity.gran_levels)) {
3420 KMP_CPU_SET(__kmp_topology->at(i).os_id, sum)(sum)->set(__kmp_topology->at(i).os_id);
3421 continue;
3422 }
3423
3424 // For every thread in this group, copy the mask to the thread's entry in
3425 // the OS Id mask table. Mark the first address as a leader.
3426 for (; j < i; j++) {
3427 int osId = __kmp_topology->at(j).os_id;
3428 KMP_DEBUG_ASSERT(osId <= maxOsId)if (!(osId <= maxOsId)) { __kmp_debug_assert("osId <= maxOsId"
, "openmp/runtime/src/kmp_affinity.cpp", 3428); }
;
3429 kmp_affin_mask_t *mask = KMP_CPU_INDEX(affinity.os_id_masks, osId)__kmp_affinity_dispatch->index_mask_array(affinity.os_id_masks
, osId)
;
3430 KMP_CPU_COPY(mask, sum)(mask)->copy(sum);
3431 __kmp_topology->at(j).leader = (j == leader);
3432 }
3433 unique++;
3434
3435 // Start a new mask.
3436 leader = i;
3437 KMP_CPU_ZERO(sum)(sum)->zero();
3438 KMP_CPU_SET(__kmp_topology->at(i).os_id, sum)(sum)->set(__kmp_topology->at(i).os_id);
3439 }
3440
3441 // For every thread in last group, copy the mask to the thread's
3442 // entry in the OS Id mask table.
3443 for (; j < i; j++) {
3444 int osId = __kmp_topology->at(j).os_id;
3445 KMP_DEBUG_ASSERT(osId <= maxOsId)if (!(osId <= maxOsId)) { __kmp_debug_assert("osId <= maxOsId"
, "openmp/runtime/src/kmp_affinity.cpp", 3445); }
;
3446 kmp_affin_mask_t *mask = KMP_CPU_INDEX(affinity.os_id_masks, osId)__kmp_affinity_dispatch->index_mask_array(affinity.os_id_masks
, osId)
;
3447 KMP_CPU_COPY(mask, sum)(mask)->copy(sum);
3448 __kmp_topology->at(j).leader = (j == leader);
3449 }
3450 unique++;
3451 KMP_CPU_FREE_FROM_STACK(sum)__kmp_affinity_dispatch->deallocate_mask(sum);
3452
3453 *numUnique = unique;
3454}
3455
3456// Stuff for the affinity proclist parsers. It's easier to declare these vars
3457// as file-static than to try and pass them through the calling sequence of
3458// the recursive-descent OMP_PLACES parser.
3459static kmp_affin_mask_t *newMasks;
3460static int numNewMasks;
3461static int nextNewMask;
3462
3463#define ADD_MASK(_mask) \
3464 { \
3465 if (nextNewMask >= numNewMasks) { \
3466 int i; \
3467 numNewMasks *= 2; \
3468 kmp_affin_mask_t *temp; \
3469 KMP_CPU_INTERNAL_ALLOC_ARRAY(temp, numNewMasks)(temp = __kmp_affinity_dispatch->allocate_mask_array(numNewMasks
))
; \
3470 for (i = 0; i < numNewMasks / 2; i++) { \
3471 kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i)__kmp_affinity_dispatch->index_mask_array(newMasks, i); \
3472 kmp_affin_mask_t *dest = KMP_CPU_INDEX(temp, i)__kmp_affinity_dispatch->index_mask_array(temp, i); \
3473 KMP_CPU_COPY(dest, src)(dest)->copy(src); \
3474 } \
3475 KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks / 2)__kmp_affinity_dispatch->deallocate_mask_array(newMasks); \
3476 newMasks = temp; \
3477 } \
3478 KMP_CPU_COPY(KMP_CPU_INDEX(newMasks, nextNewMask), (_mask))(__kmp_affinity_dispatch->index_mask_array(newMasks, nextNewMask
))->copy((_mask))
; \
3479 nextNewMask++; \
3480 }
3481
3482#define ADD_MASK_OSID(_osId, _osId2Mask, _maxOsId) \
3483 { \
3484 if (((_osId) > _maxOsId) || \
3485 (!KMP_CPU_ISSET((_osId), KMP_CPU_INDEX((_osId2Mask), (_osId)))(__kmp_affinity_dispatch->index_mask_array((_osId2Mask), (
_osId)))->is_set((_osId))
)) { \
3486 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, _osId)if (affinity.flags.verbose || (affinity.flags.warnings &&
(affinity.type != affinity_none))) { __kmp_msg(kmp_ms_warning
, __kmp_msg_format(kmp_i18n_msg_AffIgnoreInvalidProcID, _osId
), __kmp_msg_null); }
; \
3487 } else { \
3488 ADD_MASK(KMP_CPU_INDEX(_osId2Mask, (_osId))__kmp_affinity_dispatch->index_mask_array(_osId2Mask, (_osId
))
); \
3489 } \
3490 }
3491
3492// Re-parse the proclist (for the explicit affinity type), and form the list
3493// of affinity newMasks indexed by gtid.
3494static void __kmp_affinity_process_proclist(kmp_affinity_t &affinity) {
3495 int i;
3496 kmp_affin_mask_t **out_masks = &affinity.masks;
3497 unsigned *out_numMasks = &affinity.num_masks;
3498 const char *proclist = affinity.proclist;
3499 kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
3500 int maxOsId = affinity.num_os_id_masks - 1;
3501 const char *scan = proclist;
3502 const char *next = proclist;
Value stored to 'next' during its initialization is never read
3503
3504 // We use malloc() for the temporary mask vector, so that we can use
3505 // realloc() to extend it.
3506 numNewMasks = 2;
3507 KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks)(newMasks = __kmp_affinity_dispatch->allocate_mask_array(numNewMasks
))
;
3508 nextNewMask = 0;
3509 kmp_affin_mask_t *sumMask;
3510 KMP_CPU_ALLOC(sumMask)(sumMask = __kmp_affinity_dispatch->allocate_mask());
3511 int setSize = 0;
3512
3513 for (;;) {
3514 int start, end, stride;
3515
3516 SKIP_WS(scan){ while (*(scan) == ' ' || *(scan) == '\t') (scan)++; };
3517 next = scan;
3518 if (*next == '\0') {
3519 break;
3520 }
3521
3522 if (*next == '{') {
3523 int num;
3524 setSize = 0;
3525 next++; // skip '{'
3526 SKIP_WS(next){ while (*(next) == ' ' || *(next) == '\t') (next)++; };
3527 scan = next;
3528
3529 // Read the first integer in the set.
3530 KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad proclist")if (!((*next >= '0') && (*next <= '9'))) { __kmp_debug_assert
(("bad proclist"), "openmp/runtime/src/kmp_affinity.cpp", 3530
); }
;
3531 SKIP_DIGITS(next){ while (*(next) >= '0' && *(next) <= '9') (next
)++; }
;
3532 num = __kmp_str_to_int(scan, *next);
3533 KMP_ASSERT2(num >= 0, "bad explicit proc list")if (!(num >= 0)) { __kmp_debug_assert(("bad explicit proc list"
), "openmp/runtime/src/kmp_affinity.cpp", 3533); }
;
3534
3535 // Copy the mask for that osId to the sum (union) mask.
3536 if ((num > maxOsId) ||
3537 (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num))(__kmp_affinity_dispatch->index_mask_array(osId2Mask, num)
)->is_set(num)
)) {
3538 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, num)if (affinity.flags.verbose || (affinity.flags.warnings &&
(affinity.type != affinity_none))) { __kmp_msg(kmp_ms_warning
, __kmp_msg_format(kmp_i18n_msg_AffIgnoreInvalidProcID, num),
__kmp_msg_null); }
;
3539 KMP_CPU_ZERO(sumMask)(sumMask)->zero();
3540 } else {
3541 KMP_CPU_COPY(sumMask, KMP_CPU_INDEX(osId2Mask, num))(sumMask)->copy(__kmp_affinity_dispatch->index_mask_array
(osId2Mask, num))
;
3542 setSize = 1;
3543 }
3544
3545 for (;;) {
3546 // Check for end of set.
3547 SKIP_WS(next){ while (*(next) == ' ' || *(next) == '\t') (next)++; };
3548 if (*next == '}') {
3549 next++; // skip '}'
3550 break;
3551 }
3552
3553 // Skip optional comma.
3554 if (*next == ',') {
3555 next++;
3556 }
3557 SKIP_WS(next){ while (*(next) == ' ' || *(next) == '\t') (next)++; };
3558
3559 // Read the next integer in the set.
3560 scan = next;
3561 KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list")if (!((*next >= '0') && (*next <= '9'))) { __kmp_debug_assert
(("bad explicit proc list"), "openmp/runtime/src/kmp_affinity.cpp"
, 3561); }
;
3562
3563 SKIP_DIGITS(next){ while (*(next) >= '0' && *(next) <= '9') (next
)++; }
;
3564 num = __kmp_str_to_int(scan, *next);
3565 KMP_ASSERT2(num >= 0, "bad explicit proc list")if (!(num >= 0)) { __kmp_debug_assert(("bad explicit proc list"
), "openmp/runtime/src/kmp_affinity.cpp", 3565); }
;
3566
3567 // Add the mask for that osId to the sum mask.
3568 if ((num > maxOsId) ||
3569 (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num))(__kmp_affinity_dispatch->index_mask_array(osId2Mask, num)
)->is_set(num)
)) {
3570 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, num)if (affinity.flags.verbose || (affinity.flags.warnings &&
(affinity.type != affinity_none))) { __kmp_msg(kmp_ms_warning
, __kmp_msg_format(kmp_i18n_msg_AffIgnoreInvalidProcID, num),
__kmp_msg_null); }
;
3571 } else {
3572 KMP_CPU_UNION(sumMask, KMP_CPU_INDEX(osId2Mask, num))(sumMask)->bitwise_or(__kmp_affinity_dispatch->index_mask_array
(osId2Mask, num))
;
3573 setSize++;
3574 }
3575 }
3576 if (setSize > 0) {
3577 ADD_MASK(sumMask);
3578 }
3579
3580 SKIP_WS(next){ while (*(next) == ' ' || *(next) == '\t') (next)++; };
3581 if (*next == ',') {
3582 next++;
3583 }
3584 scan = next;
3585 continue;
3586 }
3587
3588 // Read the first integer.
3589 KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list")if (!((*next >= '0') && (*next <= '9'))) { __kmp_debug_assert
(("bad explicit proc list"), "openmp/runtime/src/kmp_affinity.cpp"
, 3589); }
;
3590 SKIP_DIGITS(next){ while (*(next) >= '0' && *(next) <= '9') (next
)++; }
;
3591 start = __kmp_str_to_int(scan, *next);
3592 KMP_ASSERT2(start >= 0, "bad explicit proc list")if (!(start >= 0)) { __kmp_debug_assert(("bad explicit proc list"
), "openmp/runtime/src/kmp_affinity.cpp", 3592); }
;
3593 SKIP_WS(next){ while (*(next) == ' ' || *(next) == '\t') (next)++; };
3594
3595 // If this isn't a range, then add a mask to the list and go on.
3596 if (*next != '-') {
3597 ADD_MASK_OSID(start, osId2Mask, maxOsId);
3598
3599 // Skip optional comma.
3600 if (*next == ',') {
3601 next++;
3602 }
3603 scan = next;
3604 continue;
3605 }
3606
3607 // This is a range. Skip over the '-' and read in the 2nd int.
3608 next++; // skip '-'
3609 SKIP_WS(next){ while (*(next) == ' ' || *(next) == '\t') (next)++; };
3610 scan = next;
3611 KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list")if (!((*next >= '0') && (*next <= '9'))) { __kmp_debug_assert
(("bad explicit proc list"), "openmp/runtime/src/kmp_affinity.cpp"
, 3611); }
;
3612 SKIP_DIGITS(next){ while (*(next) >= '0' && *(next) <= '9') (next
)++; }
;
3613 end = __kmp_str_to_int(scan, *next);
3614 KMP_ASSERT2(end >= 0, "bad explicit proc list")if (!(end >= 0)) { __kmp_debug_assert(("bad explicit proc list"
), "openmp/runtime/src/kmp_affinity.cpp", 3614); }
;
3615
3616 // Check for a stride parameter
3617 stride = 1;
3618 SKIP_WS(next){ while (*(next) == ' ' || *(next) == '\t') (next)++; };
3619 if (*next == ':') {
3620 // A stride is specified. Skip over the ':" and read the 3rd int.
3621 int sign = +1;
3622 next++; // skip ':'
3623 SKIP_WS(next){ while (*(next) == ' ' || *(next) == '\t') (next)++; };
3624 scan = next;
3625 if (*next == '-') {
3626 sign = -1;
3627 next++;
3628 SKIP_WS(next){ while (*(next) == ' ' || *(next) == '\t') (next)++; };
3629 scan = next;
3630 }
3631 KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list")if (!((*next >= '0') && (*next <= '9'))) { __kmp_debug_assert
(("bad explicit proc list"), "openmp/runtime/src/kmp_affinity.cpp"
, 3631); }
;
3632 SKIP_DIGITS(next){ while (*(next) >= '0' && *(next) <= '9') (next
)++; }
;
3633 stride = __kmp_str_to_int(scan, *next);
3634 KMP_ASSERT2(stride >= 0, "bad explicit proc list")if (!(stride >= 0)) { __kmp_debug_assert(("bad explicit proc list"
), "openmp/runtime/src/kmp_affinity.cpp", 3634); }
;
3635 stride *= sign;
3636 }
3637
3638 // Do some range checks.
3639 KMP_ASSERT2(stride != 0, "bad explicit proc list")if (!(stride != 0)) { __kmp_debug_assert(("bad explicit proc list"
), "openmp/runtime/src/kmp_affinity.cpp", 3639); }
;
3640 if (stride > 0) {
3641 KMP_ASSERT2(start <= end, "bad explicit proc list")if (!(start <= end)) { __kmp_debug_assert(("bad explicit proc list"
), "openmp/runtime/src/kmp_affinity.cpp", 3641); }
;
3642 } else {
3643 KMP_ASSERT2(start >= end, "bad explicit proc list")if (!(start >= end)) { __kmp_debug_assert(("bad explicit proc list"
), "openmp/runtime/src/kmp_affinity.cpp", 3643); }
;
3644 }
3645 KMP_ASSERT2((end - start) / stride <= 65536, "bad explicit proc list")if (!((end - start) / stride <= 65536)) { __kmp_debug_assert
(("bad explicit proc list"), "openmp/runtime/src/kmp_affinity.cpp"
, 3645); }
;
3646
3647 // Add the mask for each OS proc # to the list.
3648 if (stride > 0) {
3649 do {
3650 ADD_MASK_OSID(start, osId2Mask, maxOsId);
3651 start += stride;
3652 } while (start <= end);
3653 } else {
3654 do {
3655 ADD_MASK_OSID(start, osId2Mask, maxOsId);
3656 start += stride;
3657 } while (start >= end);
3658 }
3659
3660 // Skip optional comma.
3661 SKIP_WS(next){ while (*(next) == ' ' || *(next) == '\t') (next)++; };
3662 if (*next == ',') {
3663 next++;
3664 }
3665 scan = next;
3666 }
3667
3668 *out_numMasks = nextNewMask;
3669 if (nextNewMask == 0) {
3670 *out_masks = NULL__null;
3671 KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks)__kmp_affinity_dispatch->deallocate_mask_array(newMasks);
3672 return;
3673 }
3674 KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask)((*out_masks) = __kmp_affinity_dispatch->allocate_mask_array
(nextNewMask))
;
3675 for (i = 0; i < nextNewMask; i++) {
3676 kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i)__kmp_affinity_dispatch->index_mask_array(newMasks, i);
3677 kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i)__kmp_affinity_dispatch->index_mask_array((*out_masks), i);
3678 KMP_CPU_COPY(dest, src)(dest)->copy(src);
3679 }
3680 KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks)__kmp_affinity_dispatch->deallocate_mask_array(newMasks);
3681 KMP_CPU_FREE(sumMask)__kmp_affinity_dispatch->deallocate_mask(sumMask);
3682}
3683
3684/*-----------------------------------------------------------------------------
3685Re-parse the OMP_PLACES proc id list, forming the newMasks for the different
3686places. Again, Here is the grammar:
3687
3688place_list := place
3689place_list := place , place_list
3690place := num
3691place := place : num
3692place := place : num : signed
3693place := { subplacelist }
3694place := ! place // (lowest priority)
3695subplace_list := subplace
3696subplace_list := subplace , subplace_list
3697subplace := num
3698subplace := num : num
3699subplace := num : num : signed
3700signed := num
3701signed := + signed
3702signed := - signed
3703-----------------------------------------------------------------------------*/
3704static void __kmp_process_subplace_list(const char **scan,
3705 kmp_affinity_t &affinity, int maxOsId,
3706 kmp_affin_mask_t *tempMask,
3707 int *setSize) {
3708 const char *next;
3709 kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
3710
3711 for (;;) {
3712 int start, count, stride, i;
3713
3714 // Read in the starting proc id
3715 SKIP_WS(*scan){ while (*(*scan) == ' ' || *(*scan) == '\t') (*scan)++; };
3716 KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list")if (!((**scan >= '0') && (**scan <= '9'))) { __kmp_debug_assert
(("bad explicit places list"), "openmp/runtime/src/kmp_affinity.cpp"
, 3716); }
;
3717 next = *scan;
3718 SKIP_DIGITS(next){ while (*(next) >= '0' && *(next) <= '9') (next
)++; }
;
3719 start = __kmp_str_to_int(*scan, *next);
3720 KMP_ASSERT(start >= 0)if (!(start >= 0)) { __kmp_debug_assert("start >= 0", "openmp/runtime/src/kmp_affinity.cpp"
, 3720); }
;
3721 *scan = next;
3722
3723 // valid follow sets are ',' ':' and '}'
3724 SKIP_WS(*scan){ while (*(*scan) == ' ' || *(*scan) == '\t') (*scan)++; };
3725 if (**scan == '}' || **scan == ',') {
3726 if ((start > maxOsId) ||
3727 (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start))(__kmp_affinity_dispatch->index_mask_array(osId2Mask, start
))->is_set(start)
)) {
3728 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, start)if (affinity.flags.verbose || (affinity.flags.warnings &&
(affinity.type != affinity_none))) { __kmp_msg(kmp_ms_warning
, __kmp_msg_format(kmp_i18n_msg_AffIgnoreInvalidProcID, start
), __kmp_msg_null); }
;
3729 } else {
3730 KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start))(tempMask)->bitwise_or(__kmp_affinity_dispatch->index_mask_array
(osId2Mask, start))
;
3731 (*setSize)++;
3732 }
3733 if (**scan == '}') {
3734 break;
3735 }
3736 (*scan)++; // skip ','
3737 continue;
3738 }
3739 KMP_ASSERT2(**scan == ':', "bad explicit places list")if (!(**scan == ':')) { __kmp_debug_assert(("bad explicit places list"
), "openmp/runtime/src/kmp_affinity.cpp", 3739); }
;
3740 (*scan)++; // skip ':'
3741
3742 // Read count parameter
3743 SKIP_WS(*scan){ while (*(*scan) == ' ' || *(*scan) == '\t') (*scan)++; };
3744 KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list")if (!((**scan >= '0') && (**scan <= '9'))) { __kmp_debug_assert
(("bad explicit places list"), "openmp/runtime/src/kmp_affinity.cpp"
, 3744); }
;
3745 next = *scan;
3746 SKIP_DIGITS(next){ while (*(next) >= '0' && *(next) <= '9') (next
)++; }
;
3747 count = __kmp_str_to_int(*scan, *next);
3748 KMP_ASSERT(count >= 0)if (!(count >= 0)) { __kmp_debug_assert("count >= 0", "openmp/runtime/src/kmp_affinity.cpp"
, 3748); }
;
3749 *scan = next;
3750
3751 // valid follow sets are ',' ':' and '}'
3752 SKIP_WS(*scan){ while (*(*scan) == ' ' || *(*scan) == '\t') (*scan)++; };
3753 if (**scan == '}' || **scan == ',') {
3754 for (i = 0; i < count; i++) {
3755 if ((start > maxOsId) ||
3756 (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start))(__kmp_affinity_dispatch->index_mask_array(osId2Mask, start
))->is_set(start)
)) {
3757 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, start)if (affinity.flags.verbose || (affinity.flags.warnings &&
(affinity.type != affinity_none))) { __kmp_msg(kmp_ms_warning
, __kmp_msg_format(kmp_i18n_msg_AffIgnoreInvalidProcID, start
), __kmp_msg_null); }
;
3758 break; // don't proliferate warnings for large count
3759 } else {
3760 KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start))(tempMask)->bitwise_or(__kmp_affinity_dispatch->index_mask_array
(osId2Mask, start))
;
3761 start++;
3762 (*setSize)++;
3763 }
3764 }
3765 if (**scan == '}') {
3766 break;
3767 }
3768 (*scan)++; // skip ','
3769 continue;
3770 }
3771 KMP_ASSERT2(**scan == ':', "bad explicit places list")if (!(**scan == ':')) { __kmp_debug_assert(("bad explicit places list"
), "openmp/runtime/src/kmp_affinity.cpp", 3771); }
;
3772 (*scan)++; // skip ':'
3773
3774 // Read stride parameter
3775 int sign = +1;
3776 for (;;) {
3777 SKIP_WS(*scan){ while (*(*scan) == ' ' || *(*scan) == '\t') (*scan)++; };
3778 if (**scan == '+') {
3779 (*scan)++; // skip '+'
3780 continue;
3781 }
3782 if (**scan == '-') {
3783 sign *= -1;
3784 (*scan)++; // skip '-'
3785 continue;
3786 }
3787 break;
3788 }
3789 SKIP_WS(*scan){ while (*(*scan) == ' ' || *(*scan) == '\t') (*scan)++; };
3790 KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list")if (!((**scan >= '0') && (**scan <= '9'))) { __kmp_debug_assert
(("bad explicit places list"), "openmp/runtime/src/kmp_affinity.cpp"
, 3790); }
;
3791 next = *scan;
3792 SKIP_DIGITS(next){ while (*(next) >= '0' && *(next) <= '9') (next
)++; }
;
3793 stride = __kmp_str_to_int(*scan, *next);
3794 KMP_ASSERT(stride >= 0)if (!(stride >= 0)) { __kmp_debug_assert("stride >= 0",
"openmp/runtime/src/kmp_affinity.cpp", 3794); }
;
3795 *scan = next;
3796 stride *= sign;
3797
3798 // valid follow sets are ',' and '}'
3799 SKIP_WS(*scan){ while (*(*scan) == ' ' || *(*scan) == '\t') (*scan)++; };
3800 if (**scan == '}' || **scan == ',') {
3801 for (i = 0; i < count; i++) {
3802 if ((start > maxOsId) ||
3803 (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start))(__kmp_affinity_dispatch->index_mask_array(osId2Mask, start
))->is_set(start)
)) {
3804 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, start)if (affinity.flags.verbose || (affinity.flags.warnings &&
(affinity.type != affinity_none))) { __kmp_msg(kmp_ms_warning
, __kmp_msg_format(kmp_i18n_msg_AffIgnoreInvalidProcID, start
), __kmp_msg_null); }
;
3805 break; // don't proliferate warnings for large count
3806 } else {
3807 KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start))(tempMask)->bitwise_or(__kmp_affinity_dispatch->index_mask_array
(osId2Mask, start))
;
3808 start += stride;
3809 (*setSize)++;
3810 }
3811 }
3812 if (**scan == '}') {
3813 break;
3814 }
3815 (*scan)++; // skip ','
3816 continue;
3817 }
3818
3819 KMP_ASSERT2(0, "bad explicit places list")if (!(0)) { __kmp_debug_assert(("bad explicit places list"), "openmp/runtime/src/kmp_affinity.cpp"
, 3819); }
;
3820 }
3821}
3822
3823static void __kmp_process_place(const char **scan, kmp_affinity_t &affinity,
3824 int maxOsId, kmp_affin_mask_t *tempMask,
3825 int *setSize) {
3826 const char *next;
3827 kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
3828
3829 // valid follow sets are '{' '!' and num
3830 SKIP_WS(*scan){ while (*(*scan) == ' ' || *(*scan) == '\t') (*scan)++; };
3831 if (**scan == '{') {
3832 (*scan)++; // skip '{'
3833 __kmp_process_subplace_list(scan, affinity, maxOsId, tempMask, setSize);
3834 KMP_ASSERT2(**scan == '}', "bad explicit places list")if (!(**scan == '}')) { __kmp_debug_assert(("bad explicit places list"
), "openmp/runtime/src/kmp_affinity.cpp", 3834); }
;
3835 (*scan)++; // skip '}'
3836 } else if (**scan == '!') {
3837 (*scan)++; // skip '!'
3838 __kmp_process_place(scan, affinity, maxOsId, tempMask, setSize);
3839 KMP_CPU_COMPLEMENT(maxOsId, tempMask)(tempMask)->bitwise_not();
3840 } else if ((**scan >= '0') && (**scan <= '9')) {
3841 next = *scan;
3842 SKIP_DIGITS(next){ while (*(next) >= '0' && *(next) <= '9') (next
)++; }
;
3843 int num = __kmp_str_to_int(*scan, *next);
3844 KMP_ASSERT(num >= 0)if (!(num >= 0)) { __kmp_debug_assert("num >= 0", "openmp/runtime/src/kmp_affinity.cpp"
, 3844); }
;
3845 if ((num > maxOsId) ||
3846 (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num))(__kmp_affinity_dispatch->index_mask_array(osId2Mask, num)
)->is_set(num)
)) {
3847 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, num)if (affinity.flags.verbose || (affinity.flags.warnings &&
(affinity.type != affinity_none))) { __kmp_msg(kmp_ms_warning
, __kmp_msg_format(kmp_i18n_msg_AffIgnoreInvalidProcID, num),
__kmp_msg_null); }
;
3848 } else {
3849 KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, num))(tempMask)->bitwise_or(__kmp_affinity_dispatch->index_mask_array
(osId2Mask, num))
;
3850 (*setSize)++;
3851 }
3852 *scan = next; // skip num
3853 } else {
3854 KMP_ASSERT2(0, "bad explicit places list")if (!(0)) { __kmp_debug_assert(("bad explicit places list"), "openmp/runtime/src/kmp_affinity.cpp"
, 3854); }
;
3855 }
3856}
3857
3858// static void
3859void __kmp_affinity_process_placelist(kmp_affinity_t &affinity) {
3860 int i, j, count, stride, sign;
3861 kmp_affin_mask_t **out_masks = &affinity.masks;
3862 unsigned *out_numMasks = &affinity.num_masks;
3863 const char *placelist = affinity.proclist;
3864 kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
3865 int maxOsId = affinity.num_os_id_masks - 1;
3866 const char *scan = placelist;
3867 const char *next = placelist;
3868
3869 numNewMasks = 2;
3870 KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks)(newMasks = __kmp_affinity_dispatch->allocate_mask_array(numNewMasks
))
;
3871 nextNewMask = 0;
3872
3873 // tempMask is modified based on the previous or initial
3874 // place to form the current place
3875 // previousMask contains the previous place
3876 kmp_affin_mask_t *tempMask;
3877 kmp_affin_mask_t *previousMask;
3878 KMP_CPU_ALLOC(tempMask)(tempMask = __kmp_affinity_dispatch->allocate_mask());
3879 KMP_CPU_ZERO(tempMask)(tempMask)->zero();
3880 KMP_CPU_ALLOC(previousMask)(previousMask = __kmp_affinity_dispatch->allocate_mask());
3881 KMP_CPU_ZERO(previousMask)(previousMask)->zero();
3882 int setSize = 0;
3883
3884 for (;;) {
3885 __kmp_process_place(&scan, affinity, maxOsId, tempMask, &setSize);
3886
3887 // valid follow sets are ',' ':' and EOL
3888 SKIP_WS(scan){ while (*(scan) == ' ' || *(scan) == '\t') (scan)++; };
3889 if (*scan == '\0' || *scan == ',') {
3890 if (setSize > 0) {
3891 ADD_MASK(tempMask);
3892 }
3893 KMP_CPU_ZERO(tempMask)(tempMask)->zero();
3894 setSize = 0;
3895 if (*scan == '\0') {
3896 break;
3897 }
3898 scan++; // skip ','
3899 continue;
3900 }
3901
3902 KMP_ASSERT2(*scan == ':', "bad explicit places list")if (!(*scan == ':')) { __kmp_debug_assert(("bad explicit places list"
), "openmp/runtime/src/kmp_affinity.cpp", 3902); }
;
3903 scan++; // skip ':'
3904
3905 // Read count parameter
3906 SKIP_WS(scan){ while (*(scan) == ' ' || *(scan) == '\t') (scan)++; };
3907 KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list")if (!((*scan >= '0') && (*scan <= '9'))) { __kmp_debug_assert
(("bad explicit places list"), "openmp/runtime/src/kmp_affinity.cpp"
, 3907); }
;
3908 next = scan;
3909 SKIP_DIGITS(next){ while (*(next) >= '0' && *(next) <= '9') (next
)++; }
;
3910 count = __kmp_str_to_int(scan, *next);
3911 KMP_ASSERT(count >= 0)if (!(count >= 0)) { __kmp_debug_assert("count >= 0", "openmp/runtime/src/kmp_affinity.cpp"
, 3911); }
;
3912 scan = next;
3913
3914 // valid follow sets are ',' ':' and EOL
3915 SKIP_WS(scan){ while (*(scan) == ' ' || *(scan) == '\t') (scan)++; };
3916 if (*scan == '\0' || *scan == ',') {
3917 stride = +1;
3918 } else {
3919 KMP_ASSERT2(*scan == ':', "bad explicit places list")if (!(*scan == ':')) { __kmp_debug_assert(("bad explicit places list"
), "openmp/runtime/src/kmp_affinity.cpp", 3919); }
;
3920 scan++; // skip ':'
3921
3922 // Read stride parameter
3923 sign = +1;
3924 for (;;) {
3925 SKIP_WS(scan){ while (*(scan) == ' ' || *(scan) == '\t') (scan)++; };
3926 if (*scan == '+') {
3927 scan++; // skip '+'
3928 continue;
3929 }
3930 if (*scan == '-') {
3931 sign *= -1;
3932 scan++; // skip '-'
3933 continue;
3934 }
3935 break;
3936 }
3937 SKIP_WS(scan){ while (*(scan) == ' ' || *(scan) == '\t') (scan)++; };
3938 KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list")if (!((*scan >= '0') && (*scan <= '9'))) { __kmp_debug_assert
(("bad explicit places list"), "openmp/runtime/src/kmp_affinity.cpp"
, 3938); }
;
3939 next = scan;
3940 SKIP_DIGITS(next){ while (*(next) >= '0' && *(next) <= '9') (next
)++; }
;
3941 stride = __kmp_str_to_int(scan, *next);
3942 KMP_DEBUG_ASSERT(stride >= 0)if (!(stride >= 0)) { __kmp_debug_assert("stride >= 0",
"openmp/runtime/src/kmp_affinity.cpp", 3942); }
;
3943 scan = next;
3944 stride *= sign;
3945 }
3946
3947 // Add places determined by initial_place : count : stride
3948 for (i = 0; i < count; i++) {
3949 if (setSize == 0) {
3950 break;
3951 }
3952 // Add the current place, then build the next place (tempMask) from that
3953 KMP_CPU_COPY(previousMask, tempMask)(previousMask)->copy(tempMask);
3954 ADD_MASK(previousMask);
3955 KMP_CPU_ZERO(tempMask)(tempMask)->zero();
3956 setSize = 0;
3957 KMP_CPU_SET_ITERATE(j, previousMask)for (j = (previousMask)->begin(); (int)j != (previousMask)
->end(); j = (previousMask)->next(j))
{
3958 if (!KMP_CPU_ISSET(j, previousMask)(previousMask)->is_set(j)) {
3959 continue;
3960 }
3961 if ((j + stride > maxOsId) || (j + stride < 0) ||
3962 (!KMP_CPU_ISSET(j, __kmp_affin_fullMask)(__kmp_affin_fullMask)->is_set(j)) ||
3963 (!KMP_CPU_ISSET(j + stride,(__kmp_affinity_dispatch->index_mask_array(osId2Mask, j + stride
))->is_set(j + stride)
3964 KMP_CPU_INDEX(osId2Mask, j + stride))(__kmp_affinity_dispatch->index_mask_array(osId2Mask, j + stride
))->is_set(j + stride)
)) {
3965 if (i < count - 1) {
3966 KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, j + stride)if (affinity.flags.verbose || (affinity.flags.warnings &&
(affinity.type != affinity_none))) { __kmp_msg(kmp_ms_warning
, __kmp_msg_format(kmp_i18n_msg_AffIgnoreInvalidProcID, j + stride
), __kmp_msg_null); }
;
3967 }
3968 continue;
3969 }
3970 KMP_CPU_SET(j + stride, tempMask)(tempMask)->set(j + stride);
3971 setSize++;
3972 }
3973 }
3974 KMP_CPU_ZERO(tempMask)(tempMask)->zero();
3975 setSize = 0;
3976
3977 // valid follow sets are ',' and EOL
3978 SKIP_WS(scan){ while (*(scan) == ' ' || *(scan) == '\t') (scan)++; };
3979 if (*scan == '\0') {
3980 break;
3981 }
3982 if (*scan == ',') {
3983 scan++; // skip ','
3984 continue;
3985 }
3986
3987 KMP_ASSERT2(0, "bad explicit places list")if (!(0)) { __kmp_debug_assert(("bad explicit places list"), "openmp/runtime/src/kmp_affinity.cpp"
, 3987); }
;
3988 }
3989
3990 *out_numMasks = nextNewMask;
3991 if (nextNewMask == 0) {
3992 *out_masks = NULL__null;
3993 KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks)__kmp_affinity_dispatch->deallocate_mask_array(newMasks);
3994 return;
3995 }
3996 KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask)((*out_masks) = __kmp_affinity_dispatch->allocate_mask_array
(nextNewMask))
;
3997 KMP_CPU_FREE(tempMask)__kmp_affinity_dispatch->deallocate_mask(tempMask);
3998 KMP_CPU_FREE(previousMask)__kmp_affinity_dispatch->deallocate_mask(previousMask);
3999 for (i = 0; i < nextNewMask; i++) {
4000 kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i)__kmp_affinity_dispatch->index_mask_array(newMasks, i);
4001 kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i)__kmp_affinity_dispatch->index_mask_array((*out_masks), i);
4002 KMP_CPU_COPY(dest, src)(dest)->copy(src);
4003 }
4004 KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks)__kmp_affinity_dispatch->deallocate_mask_array(newMasks);
4005}
4006
4007#undef ADD_MASK
4008#undef ADD_MASK_OSID
4009
4010// This function figures out the deepest level at which there is at least one
4011// cluster/core with more than one processing unit bound to it.
4012static int __kmp_affinity_find_core_level(int nprocs, int bottom_level) {
4013 int core_level = 0;
4014
4015 for (int i = 0; i < nprocs; i++) {
4016 const kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
4017 for (int j = bottom_level; j > 0; j--) {
4018 if (hw_thread.ids[j] > 0) {
4019 if (core_level < (j - 1)) {
4020 core_level = j - 1;
4021 }
4022 }
4023 }
4024 }
4025 return core_level;
4026}
4027
4028// This function counts number of clusters/cores at given level.
4029static int __kmp_affinity_compute_ncores(int nprocs, int bottom_level,
4030 int core_level) {
4031 return __kmp_topology->get_count(core_level);
4032}
4033// This function finds to which cluster/core given processing unit is bound.
4034static int __kmp_affinity_find_core(int proc, int bottom_level,
4035 int core_level) {
4036 int core = 0;
4037 KMP_DEBUG_ASSERT(proc >= 0 && proc < __kmp_topology->get_num_hw_threads())if (!(proc >= 0 && proc < __kmp_topology->get_num_hw_threads
())) { __kmp_debug_assert("proc >= 0 && proc < __kmp_topology->get_num_hw_threads()"
, "openmp/runtime/src/kmp_affinity.cpp", 4037); }
;
4038 for (int i = 0; i <= proc; ++i) {
4039 if (i + 1 <= proc) {
4040 for (int j = 0; j <= core_level; ++j) {
4041 if (__kmp_topology->at(i + 1).sub_ids[j] !=
4042 __kmp_topology->at(i).sub_ids[j]) {
4043 core++;
4044 break;
4045 }
4046 }
4047 }
4048 }
4049 return core;
4050}
4051
4052// This function finds maximal number of processing units bound to a
4053// cluster/core at given level.
4054static int __kmp_affinity_max_proc_per_core(int nprocs, int bottom_level,
4055 int core_level) {
4056 if (core_level >= bottom_level)
4057 return 1;
4058 int thread_level = __kmp_topology->get_level(KMP_HW_THREAD);
4059 return __kmp_topology->calculate_ratio(thread_level, core_level);
4060}
4061
4062static int *procarr = NULL__null;
4063static int __kmp_aff_depth = 0;
4064static int *__kmp_osid_to_hwthread_map = NULL__null;
4065
4066static void __kmp_affinity_get_mask_topology_info(const kmp_affin_mask_t *mask,
4067 kmp_affinity_ids_t &ids,
4068 kmp_affinity_attrs_t &attrs) {
4069 if (!KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0))
4070 return;
4071
4072 // Initiailze ids and attrs thread data
4073 for (int i = 0; i < KMP_HW_LAST; ++i)
4074 ids[i] = kmp_hw_thread_t::UNKNOWN_ID;
4075 attrs = KMP_AFFINITY_ATTRS_UNKNOWN{ KMP_HW_CORE_TYPE_UNKNOWN, kmp_hw_attr_t::UNKNOWN_CORE_EFF, 0
, 0 }
;
4076
4077 // Iterate through each os id within the mask and determine
4078 // the topology id and attribute information
4079 int cpu;
4080 int depth = __kmp_topology->get_depth();
4081 KMP_CPU_SET_ITERATE(cpu, mask)for (cpu = (mask)->begin(); (int)cpu != (mask)->end(); cpu
= (mask)->next(cpu))
{
4082 int osid_idx = __kmp_osid_to_hwthread_map[cpu];
4083 const kmp_hw_thread_t &hw_thread = __kmp_topology->at(osid_idx);
4084 for (int level = 0; level < depth; ++level) {
4085 kmp_hw_t type = __kmp_topology->get_type(level);
4086 int id = hw_thread.sub_ids[level];
4087 if (ids[type] == kmp_hw_thread_t::UNKNOWN_ID || ids[type] == id) {
4088 ids[type] = id;
4089 } else {
4090 // This mask spans across multiple topology units, set it as such
4091 // and mark every level below as such as well.
4092 ids[type] = kmp_hw_thread_t::MULTIPLE_ID;
4093 for (; level < depth; ++level) {
4094 kmp_hw_t type = __kmp_topology->get_type(level);
4095 ids[type] = kmp_hw_thread_t::MULTIPLE_ID;
4096 }
4097 }
4098 }
4099 if (!attrs.valid) {
4100 attrs.core_type = hw_thread.attrs.get_core_type();
4101 attrs.core_eff = hw_thread.attrs.get_core_eff();
4102 attrs.valid = 1;
4103 } else {
4104 // This mask spans across multiple attributes, set it as such
4105 if (attrs.core_type != hw_thread.attrs.get_core_type())
4106 attrs.core_type = KMP_HW_CORE_TYPE_UNKNOWN;
4107 if (attrs.core_eff != hw_thread.attrs.get_core_eff())
4108 attrs.core_eff = kmp_hw_attr_t::UNKNOWN_CORE_EFF;
4109 }
4110 }
4111}
4112
4113static void __kmp_affinity_get_thread_topology_info(kmp_info_t *th) {
4114 if (!KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0))
4115 return;
4116 const kmp_affin_mask_t *mask = th->th.th_affin_mask;
4117 kmp_affinity_ids_t &ids = th->th.th_topology_ids;
4118 kmp_affinity_attrs_t &attrs = th->th.th_topology_attrs;
4119 __kmp_affinity_get_mask_topology_info(mask, ids, attrs);
4120}
4121
4122// Assign the topology information to each place in the place list
4123// A thread can then grab not only its affinity mask, but the topology
4124// information associated with that mask. e.g., Which socket is a thread on
4125static void __kmp_affinity_get_topology_info(kmp_affinity_t &affinity) {
4126 if (!KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0))
4127 return;
4128 if (affinity.type != affinity_none) {
4129 KMP_ASSERT(affinity.num_os_id_masks)if (!(affinity.num_os_id_masks)) { __kmp_debug_assert("affinity.num_os_id_masks"
, "openmp/runtime/src/kmp_affinity.cpp", 4129); }
;
4130 KMP_ASSERT(affinity.os_id_masks)if (!(affinity.os_id_masks)) { __kmp_debug_assert("affinity.os_id_masks"
, "openmp/runtime/src/kmp_affinity.cpp", 4130); }
;
4131 }
4132 KMP_ASSERT(affinity.num_masks)if (!(affinity.num_masks)) { __kmp_debug_assert("affinity.num_masks"
, "openmp/runtime/src/kmp_affinity.cpp", 4132); }
;
4133 KMP_ASSERT(affinity.masks)if (!(affinity.masks)) { __kmp_debug_assert("affinity.masks",
"openmp/runtime/src/kmp_affinity.cpp", 4133); }
;
4134 KMP_ASSERT(__kmp_affin_fullMask)if (!(__kmp_affin_fullMask)) { __kmp_debug_assert("__kmp_affin_fullMask"
, "openmp/runtime/src/kmp_affinity.cpp", 4134); }
;
4135
4136 int max_cpu = __kmp_affin_fullMask->get_max_cpu();
4137 int num_hw_threads = __kmp_topology->get_num_hw_threads();
4138
4139 // Allocate thread topology information
4140 if (!affinity.ids) {
4141 affinity.ids = (kmp_affinity_ids_t *)__kmp_allocate(___kmp_allocate((sizeof(kmp_affinity_ids_t) * affinity.num_masks
), "openmp/runtime/src/kmp_affinity.cpp", 4142)
4142 sizeof(kmp_affinity_ids_t) * affinity.num_masks)___kmp_allocate((sizeof(kmp_affinity_ids_t) * affinity.num_masks
), "openmp/runtime/src/kmp_affinity.cpp", 4142)
;
4143 }
4144 if (!affinity.attrs) {
4145 affinity.attrs = (kmp_affinity_attrs_t *)__kmp_allocate(___kmp_allocate((sizeof(kmp_affinity_attrs_t) * affinity.num_masks
), "openmp/runtime/src/kmp_affinity.cpp", 4146)
4146 sizeof(kmp_affinity_attrs_t) * affinity.num_masks)___kmp_allocate((sizeof(kmp_affinity_attrs_t) * affinity.num_masks
), "openmp/runtime/src/kmp_affinity.cpp", 4146)
;
4147 }
4148 if (!__kmp_osid_to_hwthread_map) {
4149 // Want the +1 because max_cpu should be valid index into map
4150 __kmp_osid_to_hwthread_map =
4151 (int *)__kmp_allocate(sizeof(int) * (max_cpu + 1))___kmp_allocate((sizeof(int) * (max_cpu + 1)), "openmp/runtime/src/kmp_affinity.cpp"
, 4151)
;
4152 }
4153
4154 // Create the OS proc to hardware thread map
4155 for (int hw_thread = 0; hw_thread < num_hw_threads; ++hw_thread)
4156 __kmp_osid_to_hwthread_map[__kmp_topology->at(hw_thread).os_id] = hw_thread;
4157
4158 for (unsigned i = 0; i < affinity.num_masks; ++i) {
4159 kmp_affinity_ids_t &ids = affinity.ids[i];
4160 kmp_affinity_attrs_t &attrs = affinity.attrs[i];
4161 kmp_affin_mask_t *mask = KMP_CPU_INDEX(affinity.masks, i)__kmp_affinity_dispatch->index_mask_array(affinity.masks, i
)
;
4162 __kmp_affinity_get_mask_topology_info(mask, ids, attrs);
4163 }
4164}
4165
4166// Create a one element mask array (set of places) which only contains the
4167// initial process's affinity mask
4168static void __kmp_create_affinity_none_places(kmp_affinity_t &affinity) {
4169 KMP_ASSERT(__kmp_affin_fullMask != NULL)if (!(__kmp_affin_fullMask != __null)) { __kmp_debug_assert("__kmp_affin_fullMask != NULL"
, "openmp/runtime/src/kmp_affinity.cpp", 4169); }
;
4170 KMP_ASSERT(affinity.type == affinity_none)if (!(affinity.type == affinity_none)) { __kmp_debug_assert("affinity.type == affinity_none"
, "openmp/runtime/src/kmp_affinity.cpp", 4170); }
;
4171 affinity.num_masks = 1;
4172 KMP_CPU_ALLOC_ARRAY(affinity.masks, affinity.num_masks)(affinity.masks = __kmp_affinity_dispatch->allocate_mask_array
(affinity.num_masks))
;
4173 kmp_affin_mask_t *dest = KMP_CPU_INDEX(affinity.masks, 0)__kmp_affinity_dispatch->index_mask_array(affinity.masks, 0
)
;
4174 KMP_CPU_COPY(dest, __kmp_affin_fullMask)(dest)->copy(__kmp_affin_fullMask);
4175 __kmp_affinity_get_topology_info(affinity);
4176}
4177
4178static void __kmp_aux_affinity_initialize_masks(kmp_affinity_t &affinity) {
4179 // Create the "full" mask - this defines all of the processors that we
4180 // consider to be in the machine model. If respect is set, then it is the
4181 // initialization thread's affinity mask. Otherwise, it is all processors that
4182 // we know about on the machine.
4183 int verbose = affinity.flags.verbose;
4184 const char *env_var = affinity.env_var;
4185
4186 // Already initialized
4187 if (__kmp_affin_fullMask && __kmp_affin_origMask)
4188 return;
4189
4190 if (__kmp_affin_fullMask == NULL__null) {
4191 KMP_CPU_ALLOC(__kmp_affin_fullMask)(__kmp_affin_fullMask = __kmp_affinity_dispatch->allocate_mask
())
;
4192 }
4193 if (__kmp_affin_origMask == NULL__null) {
4194 KMP_CPU_ALLOC(__kmp_affin_origMask)(__kmp_affin_origMask = __kmp_affinity_dispatch->allocate_mask
())
;
4195 }
4196 if (KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) {
4197 __kmp_get_system_affinity(__kmp_affin_fullMask, TRUE)(__kmp_affin_fullMask)->get_system_affinity((!0));
4198 // Make a copy before possible expanding to the entire machine mask
4199 __kmp_affin_origMask->copy(__kmp_affin_fullMask);
4200 if (affinity.flags.respect) {
4201 // Count the number of available processors.
4202 unsigned i;
4203 __kmp_avail_proc = 0;
4204 KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask)for (i = (__kmp_affin_fullMask)->begin(); (int)i != (__kmp_affin_fullMask
)->end(); i = (__kmp_affin_fullMask)->next(i))
{
4205 if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)(__kmp_affin_fullMask)->is_set(i)) {
4206 continue;
4207 }
4208 __kmp_avail_proc++;
4209 }
4210 if (__kmp_avail_proc > __kmp_xproc) {
4211 KMP_AFF_WARNING(affinity, ErrorInitializeAffinity)if (affinity.flags.verbose || (affinity.flags.warnings &&
(affinity.type != affinity_none))) { __kmp_msg(kmp_ms_warning
, __kmp_msg_format(kmp_i18n_msg_ErrorInitializeAffinity), __kmp_msg_null
); }
;
4212 affinity.type = affinity_none;
4213 KMP_AFFINITY_DISABLE()(__kmp_affin_mask_size = 0);
4214 return;
4215 }
4216
4217 if (verbose) {
4218 char buf[KMP_AFFIN_MASK_PRINT_LEN1024];
4219 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN1024,
4220 __kmp_affin_fullMask);
4221 KMP_INFORM(InitOSProcSetRespect, env_var, buf)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_InitOSProcSetRespect
, env_var, buf), __kmp_msg_null)
;
4222 }
4223 } else {
4224 if (verbose) {
4225 char buf[KMP_AFFIN_MASK_PRINT_LEN1024];
4226 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN1024,
4227 __kmp_affin_fullMask);
4228 KMP_INFORM(InitOSProcSetNotRespect, env_var, buf)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_InitOSProcSetNotRespect
, env_var, buf), __kmp_msg_null)
;
4229 }
4230 __kmp_avail_proc =
4231 __kmp_affinity_entire_machine_mask(__kmp_affin_fullMask);
4232#if KMP_OS_WINDOWS0
4233 if (__kmp_num_proc_groups <= 1) {
4234 // Copy expanded full mask if topology has single processor group
4235 __kmp_affin_origMask->copy(__kmp_affin_fullMask);
4236 }
4237 // Set the process affinity mask since threads' affinity
4238 // masks must be subset of process mask in Windows* OS
4239 __kmp_affin_fullMask->set_process_affinity(true);
4240#endif
4241 }
4242 }
4243}
4244
4245static bool __kmp_aux_affinity_initialize_topology(kmp_affinity_t &affinity) {
4246 bool success = false;
4247 const char *env_var = affinity.env_var;
4248 kmp_i18n_id_t msg_id = kmp_i18n_null;
4249 int verbose = affinity.flags.verbose;
4250
4251 // For backward compatibility, setting KMP_CPUINFO_FILE =>
4252 // KMP_TOPOLOGY_METHOD=cpuinfo
4253 if ((__kmp_cpuinfo_file != NULL__null) &&
4254 (__kmp_affinity_top_method == affinity_top_method_all)) {
4255 __kmp_affinity_top_method = affinity_top_method_cpuinfo;
4256 }
4257
4258 if (__kmp_affinity_top_method == affinity_top_method_all) {
4259// In the default code path, errors are not fatal - we just try using
4260// another method. We only emit a warning message if affinity is on, or the
4261// verbose flag is set, an the nowarnings flag was not set.
4262#if KMP_USE_HWLOC0
4263 if (!success &&
4264 __kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC) {
4265 if (!__kmp_hwloc_error) {
4266 success = __kmp_affinity_create_hwloc_map(&msg_id);
4267 if (!success && verbose) {
4268 KMP_INFORM(AffIgnoringHwloc, env_var)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_AffIgnoringHwloc
, env_var), __kmp_msg_null)
;
4269 }
4270 } else if (verbose) {
4271 KMP_INFORM(AffIgnoringHwloc, env_var)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_AffIgnoringHwloc
, env_var), __kmp_msg_null)
;
4272 }
4273 }
4274#endif
4275
4276#if KMP_ARCH_X860 || KMP_ARCH_X86_641
4277 if (!success) {
4278 success = __kmp_affinity_create_x2apicid_map(&msg_id);
4279 if (!success && verbose && msg_id != kmp_i18n_null) {
4280 KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id))__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_AffInfoStr
, env_var, __kmp_i18n_catgets(msg_id)), __kmp_msg_null)
;
4281 }
4282 }
4283 if (!success) {
4284 success = __kmp_affinity_create_apicid_map(&msg_id);
4285 if (!success && verbose && msg_id != kmp_i18n_null) {
4286 KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id))__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_AffInfoStr
, env_var, __kmp_i18n_catgets(msg_id)), __kmp_msg_null)
;
4287 }
4288 }
4289#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4290
4291#if KMP_OS_LINUX1
4292 if (!success) {
4293 int line = 0;
4294 success = __kmp_affinity_create_cpuinfo_map(&line, &msg_id);
4295 if (!success && verbose && msg_id != kmp_i18n_null) {
4296 KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id))__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_AffInfoStr
, env_var, __kmp_i18n_catgets(msg_id)), __kmp_msg_null)
;
4297 }
4298 }
4299#endif /* KMP_OS_LINUX */
4300
4301#if KMP_GROUP_AFFINITY0
4302 if (!success && (__kmp_num_proc_groups > 1)) {
4303 success = __kmp_affinity_create_proc_group_map(&msg_id);
4304 if (!success && verbose && msg_id != kmp_i18n_null) {
4305 KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id))__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_AffInfoStr
, env_var, __kmp_i18n_catgets(msg_id)), __kmp_msg_null)
;
4306 }
4307 }
4308#endif /* KMP_GROUP_AFFINITY */
4309
4310 if (!success) {
4311 success = __kmp_affinity_create_flat_map(&msg_id);
4312 if (!success && verbose && msg_id != kmp_i18n_null) {
4313 KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id))__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_AffInfoStr
, env_var, __kmp_i18n_catgets(msg_id)), __kmp_msg_null)
;
4314 }
4315 KMP_ASSERT(success)if (!(success)) { __kmp_debug_assert("success", "openmp/runtime/src/kmp_affinity.cpp"
, 4315); }
;
4316 }
4317 }
4318
4319// If the user has specified that a paricular topology discovery method is to be
4320// used, then we abort if that method fails. The exception is group affinity,
4321// which might have been implicitly set.
4322#if KMP_USE_HWLOC0
4323 else if (__kmp_affinity_top_method == affinity_top_method_hwloc) {
4324 KMP_ASSERT(__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC)if (!(__kmp_affinity_dispatch->get_api_type() == KMPAffinity
::HWLOC)) { __kmp_debug_assert("__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC"
, "openmp/runtime/src/kmp_affinity.cpp", 4324); }
;
4325 success = __kmp_affinity_create_hwloc_map(&msg_id);
4326 if (!success) {
4327 KMP_ASSERT(msg_id != kmp_i18n_null)if (!(msg_id != kmp_i18n_null)) { __kmp_debug_assert("msg_id != kmp_i18n_null"
, "openmp/runtime/src/kmp_affinity.cpp", 4327); }
;
4328 KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id))__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_MsgExiting, __kmp_i18n_catgets
(msg_id)), __kmp_msg_null)
;
4329 }
4330 }
4331#endif // KMP_USE_HWLOC
4332
4333#if KMP_ARCH_X860 || KMP_ARCH_X86_641
4334 else if (__kmp_affinity_top_method == affinity_top_method_x2apicid ||
4335 __kmp_affinity_top_method == affinity_top_method_x2apicid_1f) {
4336 success = __kmp_affinity_create_x2apicid_map(&msg_id);
4337 if (!success) {
4338 KMP_ASSERT(msg_id != kmp_i18n_null)if (!(msg_id != kmp_i18n_null)) { __kmp_debug_assert("msg_id != kmp_i18n_null"
, "openmp/runtime/src/kmp_affinity.cpp", 4338); }
;
4339 KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id))__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_MsgExiting, __kmp_i18n_catgets
(msg_id)), __kmp_msg_null)
;
4340 }
4341 } else if (__kmp_affinity_top_method == affinity_top_method_apicid) {
4342 success = __kmp_affinity_create_apicid_map(&msg_id);
4343 if (!success) {
4344 KMP_ASSERT(msg_id != kmp_i18n_null)if (!(msg_id != kmp_i18n_null)) { __kmp_debug_assert("msg_id != kmp_i18n_null"
, "openmp/runtime/src/kmp_affinity.cpp", 4344); }
;
4345 KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id))__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_MsgExiting, __kmp_i18n_catgets
(msg_id)), __kmp_msg_null)
;
4346 }
4347 }
4348#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4349
4350 else if (__kmp_affinity_top_method == affinity_top_method_cpuinfo) {
4351 int line = 0;
4352 success = __kmp_affinity_create_cpuinfo_map(&line, &msg_id);
4353 if (!success) {
4354 KMP_ASSERT(msg_id != kmp_i18n_null)if (!(msg_id != kmp_i18n_null)) { __kmp_debug_assert("msg_id != kmp_i18n_null"
, "openmp/runtime/src/kmp_affinity.cpp", 4354); }
;
4355 const char *filename = __kmp_cpuinfo_get_filename();
4356 if (line > 0) {
4357 KMP_FATAL(FileLineMsgExiting, filename, line,__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_FileLineMsgExiting,
filename, line, __kmp_i18n_catgets(msg_id)), __kmp_msg_null)
4358 __kmp_i18n_catgets(msg_id))__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_FileLineMsgExiting,
filename, line, __kmp_i18n_catgets(msg_id)), __kmp_msg_null)
;
4359 } else {
4360 KMP_FATAL(FileMsgExiting, filename, __kmp_i18n_catgets(msg_id))__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_FileMsgExiting, filename
, __kmp_i18n_catgets(msg_id)), __kmp_msg_null)
;
4361 }
4362 }
4363 }
4364
4365#if KMP_GROUP_AFFINITY0
4366 else if (__kmp_affinity_top_method == affinity_top_method_group) {
4367 success = __kmp_affinity_create_proc_group_map(&msg_id);
4368 KMP_ASSERT(success)if (!(success)) { __kmp_debug_assert("success", "openmp/runtime/src/kmp_affinity.cpp"
, 4368); }
;
4369 if (!success) {
4370 KMP_ASSERT(msg_id != kmp_i18n_null)if (!(msg_id != kmp_i18n_null)) { __kmp_debug_assert("msg_id != kmp_i18n_null"
, "openmp/runtime/src/kmp_affinity.cpp", 4370); }
;
4371 KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id))__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_MsgExiting, __kmp_i18n_catgets
(msg_id)), __kmp_msg_null)
;
4372 }
4373 }
4374#endif /* KMP_GROUP_AFFINITY */
4375
4376 else if (__kmp_affinity_top_method == affinity_top_method_flat) {
4377 success = __kmp_affinity_create_flat_map(&msg_id);
4378 // should not fail
4379 KMP_ASSERT(success)if (!(success)) { __kmp_debug_assert("success", "openmp/runtime/src/kmp_affinity.cpp"
, 4379); }
;
4380 }
4381
4382 // Early exit if topology could not be created
4383 if (!__kmp_topology) {
4384 if (KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) {
4385 KMP_AFF_WARNING(affinity, ErrorInitializeAffinity)if (affinity.flags.verbose || (affinity.flags.warnings &&
(affinity.type != affinity_none))) { __kmp_msg(kmp_ms_warning
, __kmp_msg_format(kmp_i18n_msg_ErrorInitializeAffinity), __kmp_msg_null
); }
;
4386 }
4387 if (nPackages > 0 && nCoresPerPkg > 0 && __kmp_nThreadsPerCore > 0 &&
4388 __kmp_ncores > 0) {
4389 __kmp_topology = kmp_topology_t::allocate(0, 0, NULL__null);
4390 __kmp_topology->canonicalize(nPackages, nCoresPerPkg,
4391 __kmp_nThreadsPerCore, __kmp_ncores);
4392 if (verbose) {
4393 __kmp_topology->print(env_var);
4394 }
4395 }
4396 return false;
4397 }
4398
4399 // Canonicalize, print (if requested), apply KMP_HW_SUBSET
4400 __kmp_topology->canonicalize();
4401 if (verbose)
4402 __kmp_topology->print(env_var);
4403 bool filtered = __kmp_topology->filter_hw_subset();
4404 if (filtered) {
4405#if KMP_OS_WINDOWS0
4406 // Copy filtered full mask if topology has single processor group
4407 if (__kmp_num_proc_groups <= 1)
4408#endif
4409 __kmp_affin_origMask->copy(__kmp_affin_fullMask);
4410 }
4411 if (filtered && verbose)
4412 __kmp_topology->print("KMP_HW_SUBSET");
4413 return success;
4414}
4415
4416static void __kmp_aux_affinity_initialize(kmp_affinity_t &affinity) {
4417 bool is_regular_affinity = (&affinity == &__kmp_affinity);
4418 bool is_hidden_helper_affinity = (&affinity == &__kmp_hh_affinity);
4419 const char *env_var = affinity.env_var;
4420
4421 if (affinity.flags.initialized) {
4422 KMP_ASSERT(__kmp_affin_fullMask != NULL)if (!(__kmp_affin_fullMask != __null)) { __kmp_debug_assert("__kmp_affin_fullMask != NULL"
, "openmp/runtime/src/kmp_affinity.cpp", 4422); }
;
4423 return;
4424 }
4425
4426 if (is_regular_affinity && (!__kmp_affin_fullMask || !__kmp_affin_origMask))
4427 __kmp_aux_affinity_initialize_masks(affinity);
4428
4429 if (is_regular_affinity && !__kmp_topology) {
4430 bool success = __kmp_aux_affinity_initialize_topology(affinity);
4431 if (success) {
4432 // Initialize other data structures which depend on the topology
4433 machine_hierarchy.init(__kmp_topology->get_num_hw_threads());
4434 KMP_ASSERT(__kmp_avail_proc == __kmp_topology->get_num_hw_threads())if (!(__kmp_avail_proc == __kmp_topology->get_num_hw_threads
())) { __kmp_debug_assert("__kmp_avail_proc == __kmp_topology->get_num_hw_threads()"
, "openmp/runtime/src/kmp_affinity.cpp", 4434); }
;
4435 } else {
4436 affinity.type = affinity_none;
4437 KMP_AFFINITY_DISABLE()(__kmp_affin_mask_size = 0);
4438 }
4439 }
4440
4441 // If KMP_AFFINITY=none, then only create the single "none" place
4442 // which is the process's initial affinity mask or the number of
4443 // hardware threads depending on respect,norespect
4444 if (affinity.type == affinity_none) {
4445 __kmp_create_affinity_none_places(affinity);
4446#if KMP_USE_HIER_SCHED0
4447 __kmp_dispatch_set_hierarchy_values();
4448#endif
4449 affinity.flags.initialized = TRUE(!0);
4450 return;
4451 }
4452
4453 __kmp_topology->set_granularity(affinity);
4454 int depth = __kmp_topology->get_depth();
4455
4456 // Create the table of masks, indexed by thread Id.
4457 unsigned numUnique;
4458 __kmp_create_os_id_masks(&numUnique, affinity);
4459 if (affinity.gran_levels == 0) {
4460 KMP_DEBUG_ASSERT((int)numUnique == __kmp_avail_proc)if (!((int)numUnique == __kmp_avail_proc)) { __kmp_debug_assert
("(int)numUnique == __kmp_avail_proc", "openmp/runtime/src/kmp_affinity.cpp"
, 4460); }
;
4461 }
4462
4463 switch (affinity.type) {
4464
4465 case affinity_explicit:
4466 KMP_DEBUG_ASSERT(affinity.proclist != NULL)if (!(affinity.proclist != __null)) { __kmp_debug_assert("affinity.proclist != __null"
, "openmp/runtime/src/kmp_affinity.cpp", 4466); }
;
4467 if (is_hidden_helper_affinity ||
4468 __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) {
4469 __kmp_affinity_process_proclist(affinity);
4470 } else {
4471 __kmp_affinity_process_placelist(affinity);
4472 }
4473 if (affinity.num_masks == 0) {
4474 KMP_AFF_WARNING(affinity, AffNoValidProcID)if (affinity.flags.verbose || (affinity.flags.warnings &&
(affinity.type != affinity_none))) { __kmp_msg(kmp_ms_warning
, __kmp_msg_format(kmp_i18n_msg_AffNoValidProcID), __kmp_msg_null
); }
;
4475 affinity.type = affinity_none;
4476 __kmp_create_affinity_none_places(affinity);
4477 affinity.flags.initialized = TRUE(!0);
4478 return;
4479 }
4480 break;
4481
4482 // The other affinity types rely on sorting the hardware threads according to
4483 // some permutation of the machine topology tree. Set affinity.compact
4484 // and affinity.offset appropriately, then jump to a common code
4485 // fragment to do the sort and create the array of affinity masks.
4486 case affinity_logical:
4487 affinity.compact = 0;
4488 if (affinity.offset) {
4489 affinity.offset =
4490 __kmp_nThreadsPerCore * affinity.offset % __kmp_avail_proc;
4491 }
4492 goto sortTopology;
4493
4494 case affinity_physical:
4495 if (__kmp_nThreadsPerCore > 1) {
4496 affinity.compact = 1;
4497 if (affinity.compact >= depth) {
4498 affinity.compact = 0;
4499 }
4500 } else {
4501 affinity.compact = 0;
4502 }
4503 if (affinity.offset) {
4504 affinity.offset =
4505 __kmp_nThreadsPerCore * affinity.offset % __kmp_avail_proc;
4506 }
4507 goto sortTopology;
4508
4509 case affinity_scatter:
4510 if (affinity.compact >= depth) {
4511 affinity.compact = 0;
4512 } else {
4513 affinity.compact = depth - 1 - affinity.compact;
4514 }
4515 goto sortTopology;
4516
4517 case affinity_compact:
4518 if (affinity.compact >= depth) {
4519 affinity.compact = depth - 1;
4520 }
4521 goto sortTopology;
4522
4523 case affinity_balanced:
4524 if (depth <= 1 || is_hidden_helper_affinity) {
4525 KMP_AFF_WARNING(affinity, AffBalancedNotAvail, env_var)if (affinity.flags.verbose || (affinity.flags.warnings &&
(affinity.type != affinity_none))) { __kmp_msg(kmp_ms_warning
, __kmp_msg_format(kmp_i18n_msg_AffBalancedNotAvail, env_var)
, __kmp_msg_null); }
;
4526 affinity.type = affinity_none;
4527 __kmp_create_affinity_none_places(affinity);
4528 affinity.flags.initialized = TRUE(!0);
4529 return;
4530 } else if (!__kmp_topology->is_uniform()) {
4531 // Save the depth for further usage
4532 __kmp_aff_depth = depth;
4533
4534 int core_level =
4535 __kmp_affinity_find_core_level(__kmp_avail_proc, depth - 1);
4536 int ncores = __kmp_affinity_compute_ncores(__kmp_avail_proc, depth - 1,
4537 core_level);
4538 int maxprocpercore = __kmp_affinity_max_proc_per_core(
4539 __kmp_avail_proc, depth - 1, core_level);
4540
4541 int nproc = ncores * maxprocpercore;
4542 if ((nproc < 2) || (nproc < __kmp_avail_proc)) {
4543 KMP_AFF_WARNING(affinity, AffBalancedNotAvail, env_var)if (affinity.flags.verbose || (affinity.flags.warnings &&
(affinity.type != affinity_none))) { __kmp_msg(kmp_ms_warning
, __kmp_msg_format(kmp_i18n_msg_AffBalancedNotAvail, env_var)
, __kmp_msg_null); }
;
4544 affinity.type = affinity_none;
4545 __kmp_create_affinity_none_places(affinity);
4546 affinity.flags.initialized = TRUE(!0);
4547 return;
4548 }
4549
4550 procarr = (int *)__kmp_allocate(sizeof(int) * nproc)___kmp_allocate((sizeof(int) * nproc), "openmp/runtime/src/kmp_affinity.cpp"
, 4550)
;
4551 for (int i = 0; i < nproc; i++) {
4552 procarr[i] = -1;
4553 }
4554
4555 int lastcore = -1;
4556 int inlastcore = 0;
4557 for (int i = 0; i < __kmp_avail_proc; i++) {
4558 int proc = __kmp_topology->at(i).os_id;
4559 int core = __kmp_affinity_find_core(i, depth - 1, core_level);
4560
4561 if (core == lastcore) {
4562 inlastcore++;
4563 } else {
4564 inlastcore = 0;
4565 }
4566 lastcore = core;
4567
4568 procarr[core * maxprocpercore + inlastcore] = proc;
4569 }
4570 }
4571 if (affinity.compact >= depth) {
4572 affinity.compact = depth - 1;
4573 }
4574
4575 sortTopology:
4576 // Allocate the gtid->affinity mask table.
4577 if (affinity.flags.dups) {
4578 affinity.num_masks = __kmp_avail_proc;
4579 } else {
4580 affinity.num_masks = numUnique;
4581 }
4582
4583 if ((__kmp_nested_proc_bind.bind_types[0] != proc_bind_intel) &&
4584 (__kmp_affinity_num_places > 0) &&
4585 ((unsigned)__kmp_affinity_num_places < affinity.num_masks) &&
4586 !is_hidden_helper_affinity) {
4587 affinity.num_masks = __kmp_affinity_num_places;
4588 }
4589
4590 KMP_CPU_ALLOC_ARRAY(affinity.masks, affinity.num_masks)(affinity.masks = __kmp_affinity_dispatch->allocate_mask_array
(affinity.num_masks))
;
4591
4592 // Sort the topology table according to the current setting of
4593 // affinity.compact, then fill out affinity.masks.
4594 __kmp_topology->sort_compact(affinity);
4595 {
4596 int i;
4597 unsigned j;
4598 int num_hw_threads = __kmp_topology->get_num_hw_threads();
4599 for (i = 0, j = 0; i < num_hw_threads; i++) {
4600 if ((!affinity.flags.dups) && (!__kmp_topology->at(i).leader)) {
4601 continue;
4602 }
4603 int osId = __kmp_topology->at(i).os_id;
4604
4605 kmp_affin_mask_t *src = KMP_CPU_INDEX(affinity.os_id_masks, osId)__kmp_affinity_dispatch->index_mask_array(affinity.os_id_masks
, osId)
;
4606 kmp_affin_mask_t *dest = KMP_CPU_INDEX(affinity.masks, j)__kmp_affinity_dispatch->index_mask_array(affinity.masks, j
)
;
4607 KMP_ASSERT(KMP_CPU_ISSET(osId, src))if (!((src)->is_set(osId))) { __kmp_debug_assert("KMP_CPU_ISSET(osId, src)"
, "openmp/runtime/src/kmp_affinity.cpp", 4607); }
;
4608 KMP_CPU_COPY(dest, src)(dest)->copy(src);
4609 if (++j >= affinity.num_masks) {
4610 break;
4611 }
4612 }
4613 KMP_DEBUG_ASSERT(j == affinity.num_masks)if (!(j == affinity.num_masks)) { __kmp_debug_assert("j == affinity.num_masks"
, "openmp/runtime/src/kmp_affinity.cpp", 4613); }
;
4614 }
4615 // Sort the topology back using ids
4616 __kmp_topology->sort_ids();
4617 break;
4618
4619 default:
4620 KMP_ASSERT2(0, "Unexpected affinity setting")if (!(0)) { __kmp_debug_assert(("Unexpected affinity setting"
), "openmp/runtime/src/kmp_affinity.cpp", 4620); }
;
4621 }
4622 __kmp_affinity_get_topology_info(affinity);
4623 affinity.flags.initialized = TRUE(!0);
4624}
4625
4626void __kmp_affinity_initialize(kmp_affinity_t &affinity) {
4627 // Much of the code above was written assuming that if a machine was not
4628 // affinity capable, then affinity type == affinity_none.
4629 // We now explicitly represent this as affinity type == affinity_disabled.
4630 // There are too many checks for affinity type == affinity_none in this code.
4631 // Instead of trying to change them all, check if
4632 // affinity type == affinity_disabled, and if so, slam it with affinity_none,
4633 // call the real initialization routine, then restore affinity type to
4634 // affinity_disabled.
4635 int disabled = (affinity.type == affinity_disabled);
4636 if (!KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0))
4637 KMP_ASSERT(disabled)if (!(disabled)) { __kmp_debug_assert("disabled", "openmp/runtime/src/kmp_affinity.cpp"
, 4637); }
;
4638 if (disabled)
4639 affinity.type = affinity_none;
4640 __kmp_aux_affinity_initialize(affinity);
4641 if (disabled)
4642 affinity.type = affinity_disabled;
4643}
4644
4645void __kmp_affinity_uninitialize(void) {
4646 for (kmp_affinity_t *affinity : __kmp_affinities) {
4647 if (affinity->masks != NULL__null)
4648 KMP_CPU_FREE_ARRAY(affinity->masks, affinity->num_masks)__kmp_affinity_dispatch->deallocate_mask_array(affinity->
masks)
;
4649 if (affinity->os_id_masks != NULL__null)
4650 KMP_CPU_FREE_ARRAY(affinity->os_id_masks, affinity->num_os_id_masks)__kmp_affinity_dispatch->deallocate_mask_array(affinity->
os_id_masks)
;
4651 if (affinity->proclist != NULL__null)
4652 __kmp_free(affinity->proclist)___kmp_free((affinity->proclist), "openmp/runtime/src/kmp_affinity.cpp"
, 4652)
;
4653 if (affinity->ids != NULL__null)
4654 __kmp_free(affinity->ids)___kmp_free((affinity->ids), "openmp/runtime/src/kmp_affinity.cpp"
, 4654)
;
4655 if (affinity->attrs != NULL__null)
4656 __kmp_free(affinity->attrs)___kmp_free((affinity->attrs), "openmp/runtime/src/kmp_affinity.cpp"
, 4656)
;
4657 *affinity = KMP_AFFINITY_INIT(affinity->env_var){ nullptr, affinity_default, KMP_HW_UNKNOWN, -1, 0, 0, {(!0),
0, (!0), (2), 0, 0}, 0, nullptr, nullptr, nullptr, 0, nullptr
, affinity->env_var }
;
4658 }
4659 if (__kmp_affin_origMask != NULL__null) {
4660 if (KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) {
4661 __kmp_set_system_affinity(__kmp_affin_origMask, FALSE)(__kmp_affin_origMask)->set_system_affinity(0);
4662 }
4663 KMP_CPU_FREE(__kmp_affin_origMask)__kmp_affinity_dispatch->deallocate_mask(__kmp_affin_origMask
)
;
4664 __kmp_affin_origMask = NULL__null;
4665 }
4666 __kmp_affinity_num_places = 0;
4667 if (procarr != NULL__null) {
4668 __kmp_free(procarr)___kmp_free((procarr), "openmp/runtime/src/kmp_affinity.cpp",
4668)
;
4669 procarr = NULL__null;
4670 }
4671 if (__kmp_osid_to_hwthread_map) {
4672 __kmp_free(__kmp_osid_to_hwthread_map)___kmp_free((__kmp_osid_to_hwthread_map), "openmp/runtime/src/kmp_affinity.cpp"
, 4672)
;
4673 __kmp_osid_to_hwthread_map = NULL__null;
4674 }
4675#if KMP_USE_HWLOC0
4676 if (__kmp_hwloc_topology != NULL__null) {
4677 hwloc_topology_destroy(__kmp_hwloc_topology);
4678 __kmp_hwloc_topology = NULL__null;
4679 }
4680#endif
4681 if (__kmp_hw_subset) {
4682 kmp_hw_subset_t::deallocate(__kmp_hw_subset);
4683 __kmp_hw_subset = nullptr;
4684 }
4685 if (__kmp_topology) {
4686 kmp_topology_t::deallocate(__kmp_topology);
4687 __kmp_topology = nullptr;
4688 }
4689 KMPAffinity::destroy_api();
4690}
4691
4692static void __kmp_select_mask_by_gtid(int gtid, const kmp_affinity_t *affinity,
4693 int *place, kmp_affin_mask_t **mask) {
4694 int mask_idx;
4695 bool is_hidden_helper = KMP_HIDDEN_HELPER_THREAD(gtid)((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num
)
;
4696 if (is_hidden_helper)
4697 // The first gtid is the regular primary thread, the second gtid is the main
4698 // thread of hidden team which does not participate in task execution.
4699 mask_idx = gtid - 2;
4700 else
4701 mask_idx = __kmp_adjust_gtid_for_hidden_helpers(gtid);
4702 KMP_DEBUG_ASSERT(affinity->num_masks > 0)if (!(affinity->num_masks > 0)) { __kmp_debug_assert("affinity->num_masks > 0"
, "openmp/runtime/src/kmp_affinity.cpp", 4702); }
;
4703 *place = (mask_idx + affinity->offset) % affinity->num_masks;
4704 *mask = KMP_CPU_INDEX(affinity->masks, *place)__kmp_affinity_dispatch->index_mask_array(affinity->masks
, *place)
;
4705}
4706
4707// This function initializes the per-thread data concerning affinity including
4708// the mask and topology information
4709void __kmp_affinity_set_init_mask(int gtid, int isa_root) {
4710
4711 kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid])((void *)(__kmp_threads[gtid]));
4712
4713 // Set the thread topology information to default of unknown
4714 for (int id = 0; id < KMP_HW_LAST; ++id)
4715 th->th.th_topology_ids[id] = kmp_hw_thread_t::UNKNOWN_ID;
4716 th->th.th_topology_attrs = KMP_AFFINITY_ATTRS_UNKNOWN{ KMP_HW_CORE_TYPE_UNKNOWN, kmp_hw_attr_t::UNKNOWN_CORE_EFF, 0
, 0 }
;
4717
4718 if (!KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) {
4719 return;
4720 }
4721
4722 if (th->th.th_affin_mask == NULL__null) {
4723 KMP_CPU_ALLOC(th->th.th_affin_mask)(th->th.th_affin_mask = __kmp_affinity_dispatch->allocate_mask
())
;
4724 } else {
4725 KMP_CPU_ZERO(th->th.th_affin_mask)(th->th.th_affin_mask)->zero();
4726 }
4727
4728 // Copy the thread mask to the kmp_info_t structure. If
4729 // __kmp_affinity.type == affinity_none, copy the "full" mask, i.e.
4730 // one that has all of the OS proc ids set, or if
4731 // __kmp_affinity.flags.respect is set, then the full mask is the
4732 // same as the mask of the initialization thread.
4733 kmp_affin_mask_t *mask;
4734 int i;
4735 const kmp_affinity_t *affinity;
4736 const char *env_var;
4737 bool is_hidden_helper = KMP_HIDDEN_HELPER_THREAD(gtid)((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num
)
;
4738
4739 if (is_hidden_helper)
4740 affinity = &__kmp_hh_affinity;
4741 else
4742 affinity = &__kmp_affinity;
4743 env_var = affinity->env_var;
4744
4745 if (KMP_AFFINITY_NON_PROC_BIND((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || __kmp_nested_proc_bind
.bind_types[0] == proc_bind_intel) && (__kmp_affinity
.num_masks > 0 || __kmp_affinity.type == affinity_balanced
))
|| is_hidden_helper) {
4746 if ((affinity->type == affinity_none) ||
4747 (affinity->type == affinity_balanced) ||
4748 KMP_HIDDEN_HELPER_MAIN_THREAD(gtid)((gtid) == 1 && (gtid) <= __kmp_hidden_helper_threads_num
)
) {
4749#if KMP_GROUP_AFFINITY0
4750 if (__kmp_num_proc_groups > 1) {
4751 return;
4752 }
4753#endif
4754 KMP_ASSERT(__kmp_affin_fullMask != NULL)if (!(__kmp_affin_fullMask != __null)) { __kmp_debug_assert("__kmp_affin_fullMask != NULL"
, "openmp/runtime/src/kmp_affinity.cpp", 4754); }
;
4755 i = 0;
4756 mask = __kmp_affin_fullMask;
4757 } else {
4758 __kmp_select_mask_by_gtid(gtid, affinity, &i, &mask);
4759 }
4760 } else {
4761 if (!isa_root || __kmp_nested_proc_bind.bind_types[0] == proc_bind_false) {
4762#if KMP_GROUP_AFFINITY0
4763 if (__kmp_num_proc_groups > 1) {
4764 return;
4765 }
4766#endif
4767 KMP_ASSERT(__kmp_affin_fullMask != NULL)if (!(__kmp_affin_fullMask != __null)) { __kmp_debug_assert("__kmp_affin_fullMask != NULL"
, "openmp/runtime/src/kmp_affinity.cpp", 4767); }
;
4768 i = KMP_PLACE_ALL(-1);
4769 mask = __kmp_affin_fullMask;
4770 } else {
4771 __kmp_select_mask_by_gtid(gtid, affinity, &i, &mask);
4772 }
4773 }
4774
4775 th->th.th_current_place = i;
4776 if (isa_root && !is_hidden_helper) {
4777 th->th.th_new_place = i;
4778 th->th.th_first_place = 0;
4779 th->th.th_last_place = affinity->num_masks - 1;
4780 } else if (KMP_AFFINITY_NON_PROC_BIND((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || __kmp_nested_proc_bind
.bind_types[0] == proc_bind_intel) && (__kmp_affinity
.num_masks > 0 || __kmp_affinity.type == affinity_balanced
))
) {
4781 // When using a Non-OMP_PROC_BIND affinity method,
4782 // set all threads' place-partition-var to the entire place list
4783 th->th.th_first_place = 0;
4784 th->th.th_last_place = affinity->num_masks - 1;
4785 }
4786 // Copy topology information associated with the place
4787 if (i >= 0) {
4788 th->th.th_topology_ids = __kmp_affinity.ids[i];
4789 th->th.th_topology_attrs = __kmp_affinity.attrs[i];
4790 }
4791
4792 if (i == KMP_PLACE_ALL(-1)) {
4793 KA_TRACE(100, ("__kmp_affinity_set_init_mask: binding T#%d to all places\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_affinity_set_init_mask: binding T#%d to all places\n"
, gtid); }
4794 gtid))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_affinity_set_init_mask: binding T#%d to all places\n"
, gtid); }
;
4795 } else {
4796 KA_TRACE(100, ("__kmp_affinity_set_init_mask: binding T#%d to place %d\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_affinity_set_init_mask: binding T#%d to place %d\n"
, gtid, i); }
4797 gtid, i))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_affinity_set_init_mask: binding T#%d to place %d\n"
, gtid, i); }
;
4798 }
4799
4800 KMP_CPU_COPY(th->th.th_affin_mask, mask)(th->th.th_affin_mask)->copy(mask);
4801
4802 /* to avoid duplicate printing (will be correctly printed on barrier) */
4803 if (affinity->flags.verbose &&
4804 (affinity->type == affinity_none ||
4805 (i != KMP_PLACE_ALL(-1) && affinity->type != affinity_balanced)) &&
4806 !KMP_HIDDEN_HELPER_MAIN_THREAD(gtid)((gtid) == 1 && (gtid) <= __kmp_hidden_helper_threads_num
)
) {
4807 char buf[KMP_AFFIN_MASK_PRINT_LEN1024];
4808 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN1024,
4809 th->th.th_affin_mask);
4810 KMP_INFORM(BoundToOSProcSet, env_var, (kmp_int32)getpid(), __kmp_gettid(),__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_BoundToOSProcSet
, env_var, (kmp_int32)getpid(), syscall(186), gtid, buf), __kmp_msg_null
)
4811 gtid, buf)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_BoundToOSProcSet
, env_var, (kmp_int32)getpid(), syscall(186), gtid, buf), __kmp_msg_null
)
;
4812 }
4813
4814#if KMP_OS_WINDOWS0
4815 // On Windows* OS, the process affinity mask might have changed. If the user
4816 // didn't request affinity and this call fails, just continue silently.
4817 // See CQ171393.
4818 if (affinity->type == affinity_none) {
4819 __kmp_set_system_affinity(th->th.th_affin_mask, FALSE)(th->th.th_affin_mask)->set_system_affinity(0);
4820 } else
4821#endif
4822 __kmp_set_system_affinity(th->th.th_affin_mask, TRUE)(th->th.th_affin_mask)->set_system_affinity((!0));
4823}
4824
4825void __kmp_affinity_set_place(int gtid) {
4826 // Hidden helper threads should not be affected by OMP_PLACES/OMP_PROC_BIND
4827 if (!KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0) || KMP_HIDDEN_HELPER_THREAD(gtid)((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num
)
) {
4828 return;
4829 }
4830
4831 kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid])((void *)(__kmp_threads[gtid]));
4832
4833 KA_TRACE(100, ("__kmp_affinity_set_place: binding T#%d to place %d (current "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_affinity_set_place: binding T#%d to place %d (current "
"place = %d)\n", gtid, th->th.th_new_place, th->th.th_current_place
); }
4834 "place = %d)\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_affinity_set_place: binding T#%d to place %d (current "
"place = %d)\n", gtid, th->th.th_new_place, th->th.th_current_place
); }
4835 gtid, th->th.th_new_place, th->th.th_current_place))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_affinity_set_place: binding T#%d to place %d (current "
"place = %d)\n", gtid, th->th.th_new_place, th->th.th_current_place
); }
;
4836
4837 // Check that the new place is within this thread's partition.
4838 KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL)if (!(th->th.th_affin_mask != __null)) { __kmp_debug_assert
("th->th.th_affin_mask != __null", "openmp/runtime/src/kmp_affinity.cpp"
, 4838); }
;
4839 KMP_ASSERT(th->th.th_new_place >= 0)if (!(th->th.th_new_place >= 0)) { __kmp_debug_assert("th->th.th_new_place >= 0"
, "openmp/runtime/src/kmp_affinity.cpp", 4839); }
;
4840 KMP_ASSERT((unsigned)th->th.th_new_place <= __kmp_affinity.num_masks)if (!((unsigned)th->th.th_new_place <= __kmp_affinity.num_masks
)) { __kmp_debug_assert("(unsigned)th->th.th_new_place <= __kmp_affinity.num_masks"
, "openmp/runtime/src/kmp_affinity.cpp", 4840); }
;
4841 if (th->th.th_first_place <= th->th.th_last_place) {
4842 KMP_ASSERT((th->th.th_new_place >= th->th.th_first_place) &&if (!((th->th.th_new_place >= th->th.th_first_place)
&& (th->th.th_new_place <= th->th.th_last_place
))) { __kmp_debug_assert("(th->th.th_new_place >= th->th.th_first_place) && (th->th.th_new_place <= th->th.th_last_place)"
, "openmp/runtime/src/kmp_affinity.cpp", 4843); }
4843 (th->th.th_new_place <= th->th.th_last_place))if (!((th->th.th_new_place >= th->th.th_first_place)
&& (th->th.th_new_place <= th->th.th_last_place
))) { __kmp_debug_assert("(th->th.th_new_place >= th->th.th_first_place) && (th->th.th_new_place <= th->th.th_last_place)"
, "openmp/runtime/src/kmp_affinity.cpp", 4843); }
;
4844 } else {
4845 KMP_ASSERT((th->th.th_new_place <= th->th.th_first_place) ||if (!((th->th.th_new_place <= th->th.th_first_place)
|| (th->th.th_new_place >= th->th.th_last_place))) {
__kmp_debug_assert("(th->th.th_new_place <= th->th.th_first_place) || (th->th.th_new_place >= th->th.th_last_place)"
, "openmp/runtime/src/kmp_affinity.cpp", 4846); }
4846 (th->th.th_new_place >= th->th.th_last_place))if (!((th->th.th_new_place <= th->th.th_first_place)
|| (th->th.th_new_place >= th->th.th_last_place))) {
__kmp_debug_assert("(th->th.th_new_place <= th->th.th_first_place) || (th->th.th_new_place >= th->th.th_last_place)"
, "openmp/runtime/src/kmp_affinity.cpp", 4846); }
;
4847 }
4848
4849 // Copy the thread mask to the kmp_info_t structure,
4850 // and set this thread's affinity.
4851 kmp_affin_mask_t *mask =
4852 KMP_CPU_INDEX(__kmp_affinity.masks, th->th.th_new_place)__kmp_affinity_dispatch->index_mask_array(__kmp_affinity.masks
, th->th.th_new_place)
;
4853 KMP_CPU_COPY(th->th.th_affin_mask, mask)(th->th.th_affin_mask)->copy(mask);
4854 th->th.th_current_place = th->th.th_new_place;
4855 // Copy topology information associated with the place
4856 th->th.th_topology_ids = __kmp_affinity.ids[th->th.th_new_place];
4857 th->th.th_topology_attrs = __kmp_affinity.attrs[th->th.th_new_place];
4858
4859 if (__kmp_affinity.flags.verbose) {
4860 char buf[KMP_AFFIN_MASK_PRINT_LEN1024];
4861 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN1024,
4862 th->th.th_affin_mask);
4863 KMP_INFORM(BoundToOSProcSet, "OMP_PROC_BIND", (kmp_int32)getpid(),__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_BoundToOSProcSet
, "OMP_PROC_BIND", (kmp_int32)getpid(), syscall(186), gtid, buf
), __kmp_msg_null)
4864 __kmp_gettid(), gtid, buf)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_BoundToOSProcSet
, "OMP_PROC_BIND", (kmp_int32)getpid(), syscall(186), gtid, buf
), __kmp_msg_null)
;
4865 }
4866 __kmp_set_system_affinity(th->th.th_affin_mask, TRUE)(th->th.th_affin_mask)->set_system_affinity((!0));
4867}
4868
4869int __kmp_aux_set_affinity(void **mask) {
4870 int gtid;
4871 kmp_info_t *th;
4872 int retval;
4873
4874 if (!KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) {
4875 return -1;
4876 }
4877
4878 gtid = __kmp_entry_gtid()__kmp_get_global_thread_id_reg();
4879 KA_TRACE(if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, (kmp_affin_mask_t
*)(*mask)); __kmp_debug_printf( "kmp_set_affinity: setting affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4880 1000, (""); {if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, (kmp_affin_mask_t
*)(*mask)); __kmp_debug_printf( "kmp_set_affinity: setting affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4881 char buf[KMP_AFFIN_MASK_PRINT_LEN];if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, (kmp_affin_mask_t
*)(*mask)); __kmp_debug_printf( "kmp_set_affinity: setting affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4882 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, (kmp_affin_mask_t
*)(*mask)); __kmp_debug_printf( "kmp_set_affinity: setting affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4883 (kmp_affin_mask_t *)(*mask));if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, (kmp_affin_mask_t
*)(*mask)); __kmp_debug_printf( "kmp_set_affinity: setting affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4884 __kmp_debug_printf(if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, (kmp_affin_mask_t
*)(*mask)); __kmp_debug_printf( "kmp_set_affinity: setting affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4885 "kmp_set_affinity: setting affinity mask for thread %d = %s\n",if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, (kmp_affin_mask_t
*)(*mask)); __kmp_debug_printf( "kmp_set_affinity: setting affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4886 gtid, buf);if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, (kmp_affin_mask_t
*)(*mask)); __kmp_debug_printf( "kmp_set_affinity: setting affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4887 })if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, (kmp_affin_mask_t
*)(*mask)); __kmp_debug_printf( "kmp_set_affinity: setting affinity mask for thread %d = %s\n"
, gtid, buf); }; }
;
4888
4889 if (__kmp_env_consistency_check) {
4890 if ((mask == NULL__null) || (*mask == NULL__null)) {
4891 KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity")__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_AffinityInvalidMask
, "kmp_set_affinity"), __kmp_msg_null)
;
4892 } else {
4893 unsigned proc;
4894 int num_procs = 0;
4895
4896 KMP_CPU_SET_ITERATE(proc, ((kmp_affin_mask_t *)(*mask)))for (proc = (((kmp_affin_mask_t *)(*mask)))->begin(); (int
)proc != (((kmp_affin_mask_t *)(*mask)))->end(); proc = ((
(kmp_affin_mask_t *)(*mask)))->next(proc))
{
4897 if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)(__kmp_affin_fullMask)->is_set(proc)) {
4898 KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity")__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_AffinityInvalidMask
, "kmp_set_affinity"), __kmp_msg_null)
;
4899 }
4900 if (!KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask))((kmp_affin_mask_t *)(*mask))->is_set(proc)) {
4901 continue;
4902 }
4903 num_procs++;
4904 }
4905 if (num_procs == 0) {
4906 KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity")__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_AffinityInvalidMask
, "kmp_set_affinity"), __kmp_msg_null)
;
4907 }
4908
4909#if KMP_GROUP_AFFINITY0
4910 if (__kmp_get_proc_group((kmp_affin_mask_t *)(*mask))((kmp_affin_mask_t *)(*mask))->get_proc_group() < 0) {
4911 KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity")__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_AffinityInvalidMask
, "kmp_set_affinity"), __kmp_msg_null)
;
4912 }
4913#endif /* KMP_GROUP_AFFINITY */
4914 }
4915 }
4916
4917 th = __kmp_threads[gtid];
4918 KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL)if (!(th->th.th_affin_mask != __null)) { __kmp_debug_assert
("th->th.th_affin_mask != __null", "openmp/runtime/src/kmp_affinity.cpp"
, 4918); }
;
4919 retval = __kmp_set_system_affinity((kmp_affin_mask_t *)(*mask), FALSE)((kmp_affin_mask_t *)(*mask))->set_system_affinity(0);
4920 if (retval == 0) {
4921 KMP_CPU_COPY(th->th.th_affin_mask, (kmp_affin_mask_t *)(*mask))(th->th.th_affin_mask)->copy((kmp_affin_mask_t *)(*mask
))
;
4922 }
4923
4924 th->th.th_current_place = KMP_PLACE_UNDEFINED(-2);
4925 th->th.th_new_place = KMP_PLACE_UNDEFINED(-2);
4926 th->th.th_first_place = 0;
4927 th->th.th_last_place = __kmp_affinity.num_masks - 1;
4928
4929 // Turn off 4.0 affinity for the current tread at this parallel level.
4930 th->th.th_current_task->td_icvs.proc_bind = proc_bind_false;
4931
4932 return retval;
4933}
4934
4935int __kmp_aux_get_affinity(void **mask) {
4936 int gtid;
4937 int retval;
4938#if KMP_OS_WINDOWS0 || KMP_DEBUG1
4939 kmp_info_t *th;
4940#endif
4941 if (!KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) {
4942 return -1;
4943 }
4944
4945 gtid = __kmp_entry_gtid()__kmp_get_global_thread_id_reg();
4946#if KMP_OS_WINDOWS0 || KMP_DEBUG1
4947 th = __kmp_threads[gtid];
4948#else
4949 (void)gtid; // unused variable
4950#endif
4951 KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL)if (!(th->th.th_affin_mask != __null)) { __kmp_debug_assert
("th->th.th_affin_mask != __null", "openmp/runtime/src/kmp_affinity.cpp"
, 4951); }
;
4952
4953 KA_TRACE(if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, th->th.th_affin_mask
); __kmp_printf( "kmp_get_affinity: stored affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4954 1000, (""); {if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, th->th.th_affin_mask
); __kmp_printf( "kmp_get_affinity: stored affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4955 char buf[KMP_AFFIN_MASK_PRINT_LEN];if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, th->th.th_affin_mask
); __kmp_printf( "kmp_get_affinity: stored affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4956 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, th->th.th_affin_mask
); __kmp_printf( "kmp_get_affinity: stored affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4957 th->th.th_affin_mask);if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, th->th.th_affin_mask
); __kmp_printf( "kmp_get_affinity: stored affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4958 __kmp_printf(if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, th->th.th_affin_mask
); __kmp_printf( "kmp_get_affinity: stored affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4959 "kmp_get_affinity: stored affinity mask for thread %d = %s\n", gtid,if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, th->th.th_affin_mask
); __kmp_printf( "kmp_get_affinity: stored affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4960 buf);if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, th->th.th_affin_mask
); __kmp_printf( "kmp_get_affinity: stored affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4961 })if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, th->th.th_affin_mask
); __kmp_printf( "kmp_get_affinity: stored affinity mask for thread %d = %s\n"
, gtid, buf); }; }
;
4962
4963 if (__kmp_env_consistency_check) {
4964 if ((mask == NULL__null) || (*mask == NULL__null)) {
4965 KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity")__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_AffinityInvalidMask
, "kmp_get_affinity"), __kmp_msg_null)
;
4966 }
4967 }
4968
4969#if !KMP_OS_WINDOWS0
4970
4971 retval = __kmp_get_system_affinity((kmp_affin_mask_t *)(*mask), FALSE)((kmp_affin_mask_t *)(*mask))->get_system_affinity(0);
4972 KA_TRACE(if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, (kmp_affin_mask_t
*)(*mask)); __kmp_printf( "kmp_get_affinity: system affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4973 1000, (""); {if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, (kmp_affin_mask_t
*)(*mask)); __kmp_printf( "kmp_get_affinity: system affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4974 char buf[KMP_AFFIN_MASK_PRINT_LEN];if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, (kmp_affin_mask_t
*)(*mask)); __kmp_printf( "kmp_get_affinity: system affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4975 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, (kmp_affin_mask_t
*)(*mask)); __kmp_printf( "kmp_get_affinity: system affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4976 (kmp_affin_mask_t *)(*mask));if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, (kmp_affin_mask_t
*)(*mask)); __kmp_printf( "kmp_get_affinity: system affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4977 __kmp_printf(if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, (kmp_affin_mask_t
*)(*mask)); __kmp_printf( "kmp_get_affinity: system affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4978 "kmp_get_affinity: system affinity mask for thread %d = %s\n", gtid,if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, (kmp_affin_mask_t
*)(*mask)); __kmp_printf( "kmp_get_affinity: system affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4979 buf);if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, (kmp_affin_mask_t
*)(*mask)); __kmp_printf( "kmp_get_affinity: system affinity mask for thread %d = %s\n"
, gtid, buf); }; }
4980 })if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { char
buf[1024]; __kmp_affinity_print_mask(buf, 1024, (kmp_affin_mask_t
*)(*mask)); __kmp_printf( "kmp_get_affinity: system affinity mask for thread %d = %s\n"
, gtid, buf); }; }
;
4981 return retval;
4982
4983#else
4984 (void)retval;
4985
4986 KMP_CPU_COPY((kmp_affin_mask_t *)(*mask), th->th.th_affin_mask)((kmp_affin_mask_t *)(*mask))->copy(th->th.th_affin_mask
)
;
4987 return 0;
4988
4989#endif /* KMP_OS_WINDOWS */
4990}
4991
4992int __kmp_aux_get_affinity_max_proc() {
4993 if (!KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) {
4994 return 0;
4995 }
4996#if KMP_GROUP_AFFINITY0
4997 if (__kmp_num_proc_groups > 1) {
4998 return (int)(__kmp_num_proc_groups * sizeof(DWORD_PTR) * CHAR_BIT8);
4999 }
5000#endif
5001 return __kmp_xproc;
5002}
5003
5004int __kmp_aux_set_affinity_mask_proc(int proc, void **mask) {
5005 if (!KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) {
5006 return -1;
5007 }
5008
5009 KA_TRACE(if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_set_affinity_mask_proc: setting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5010 1000, (""); {if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_set_affinity_mask_proc: setting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5011 int gtid = __kmp_entry_gtid();if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_set_affinity_mask_proc: setting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5012 char buf[KMP_AFFIN_MASK_PRINT_LEN];if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_set_affinity_mask_proc: setting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5013 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_set_affinity_mask_proc: setting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5014 (kmp_affin_mask_t *)(*mask));if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_set_affinity_mask_proc: setting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5015 __kmp_debug_printf("kmp_set_affinity_mask_proc: setting proc %d in "if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_set_affinity_mask_proc: setting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5016 "affinity mask for thread %d = %s\n",if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_set_affinity_mask_proc: setting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5017 proc, gtid, buf);if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_set_affinity_mask_proc: setting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5018 })if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_set_affinity_mask_proc: setting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
;
5019
5020 if (__kmp_env_consistency_check) {
5021 if ((mask == NULL__null) || (*mask == NULL__null)) {
5022 KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity_mask_proc")__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_AffinityInvalidMask
, "kmp_set_affinity_mask_proc"), __kmp_msg_null)
;
5023 }
5024 }
5025
5026 if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5027 return -1;
5028 }
5029 if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)(__kmp_affin_fullMask)->is_set(proc)) {
5030 return -2;
5031 }
5032
5033 KMP_CPU_SET(proc, (kmp_affin_mask_t *)(*mask))((kmp_affin_mask_t *)(*mask))->set(proc);
5034 return 0;
5035}
5036
5037int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask) {
5038 if (!KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) {
5039 return -1;
5040 }
5041
5042 KA_TRACE(if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_unset_affinity_mask_proc: unsetting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5043 1000, (""); {if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_unset_affinity_mask_proc: unsetting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5044 int gtid = __kmp_entry_gtid();if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_unset_affinity_mask_proc: unsetting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5045 char buf[KMP_AFFIN_MASK_PRINT_LEN];if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_unset_affinity_mask_proc: unsetting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5046 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_unset_affinity_mask_proc: unsetting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5047 (kmp_affin_mask_t *)(*mask));if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_unset_affinity_mask_proc: unsetting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5048 __kmp_debug_printf("kmp_unset_affinity_mask_proc: unsetting proc %d in "if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_unset_affinity_mask_proc: unsetting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5049 "affinity mask for thread %d = %s\n",if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_unset_affinity_mask_proc: unsetting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5050 proc, gtid, buf);if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_unset_affinity_mask_proc: unsetting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5051 })if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_unset_affinity_mask_proc: unsetting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
;
5052
5053 if (__kmp_env_consistency_check) {
5054 if ((mask == NULL__null) || (*mask == NULL__null)) {
5055 KMP_FATAL(AffinityInvalidMask, "kmp_unset_affinity_mask_proc")__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_AffinityInvalidMask
, "kmp_unset_affinity_mask_proc"), __kmp_msg_null)
;
5056 }
5057 }
5058
5059 if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5060 return -1;
5061 }
5062 if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)(__kmp_affin_fullMask)->is_set(proc)) {
5063 return -2;
5064 }
5065
5066 KMP_CPU_CLR(proc, (kmp_affin_mask_t *)(*mask))((kmp_affin_mask_t *)(*mask))->clear(proc);
5067 return 0;
5068}
5069
5070int __kmp_aux_get_affinity_mask_proc(int proc, void **mask) {
5071 if (!KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) {
5072 return -1;
5073 }
5074
5075 KA_TRACE(if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_get_affinity_mask_proc: getting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5076 1000, (""); {if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_get_affinity_mask_proc: getting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5077 int gtid = __kmp_entry_gtid();if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_get_affinity_mask_proc: getting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5078 char buf[KMP_AFFIN_MASK_PRINT_LEN];if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_get_affinity_mask_proc: getting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5079 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_get_affinity_mask_proc: getting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5080 (kmp_affin_mask_t *)(*mask));if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_get_affinity_mask_proc: getting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5081 __kmp_debug_printf("kmp_get_affinity_mask_proc: getting proc %d in "if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_get_affinity_mask_proc: getting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5082 "affinity mask for thread %d = %s\n",if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_get_affinity_mask_proc: getting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5083 proc, gtid, buf);if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_get_affinity_mask_proc: getting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
5084 })if (kmp_a_debug >= 1000) { __kmp_debug_printf (""); { int gtid
= __kmp_get_global_thread_id_reg(); char buf[1024]; __kmp_affinity_print_mask
(buf, 1024, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf(
"kmp_get_affinity_mask_proc: getting proc %d in " "affinity mask for thread %d = %s\n"
, proc, gtid, buf); }; }
;
5085
5086 if (__kmp_env_consistency_check) {
5087 if ((mask == NULL__null) || (*mask == NULL__null)) {
5088 KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity_mask_proc")__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_AffinityInvalidMask
, "kmp_get_affinity_mask_proc"), __kmp_msg_null)
;
5089 }
5090 }
5091
5092 if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5093 return -1;
5094 }
5095 if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)(__kmp_affin_fullMask)->is_set(proc)) {
5096 return 0;
5097 }
5098
5099 return KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask))((kmp_affin_mask_t *)(*mask))->is_set(proc);
5100}
5101
5102// Dynamic affinity settings - Affinity balanced
5103void __kmp_balanced_affinity(kmp_info_t *th, int nthreads) {
5104 KMP_DEBUG_ASSERT(th)if (!(th)) { __kmp_debug_assert("th", "openmp/runtime/src/kmp_affinity.cpp"
, 5104); }
;
5105 bool fine_gran = true;
5106 int tid = th->th.th_info.ds.ds_tid;
5107 const char *env_var = "KMP_AFFINITY";
5108
5109 // Do not perform balanced affinity for the hidden helper threads
5110 if (KMP_HIDDEN_HELPER_THREAD(__kmp_gtid_from_thread(th))((__kmp_gtid_from_thread(th)) >= 1 && (__kmp_gtid_from_thread
(th)) <= __kmp_hidden_helper_threads_num)
)
5111 return;
5112
5113 switch (__kmp_affinity.gran) {
5114 case KMP_HW_THREAD:
5115 break;
5116 case KMP_HW_CORE:
5117 if (__kmp_nThreadsPerCore > 1) {
5118 fine_gran = false;
5119 }
5120 break;
5121 case KMP_HW_SOCKET:
5122 if (nCoresPerPkg > 1) {
5123 fine_gran = false;
5124 }
5125 break;
5126 default:
5127 fine_gran = false;
5128 }
5129
5130 if (__kmp_topology->is_uniform()) {
5131 int coreID;
5132 int threadID;
5133 // Number of hyper threads per core in HT machine
5134 int __kmp_nth_per_core = __kmp_avail_proc / __kmp_ncores;
5135 // Number of cores
5136 int ncores = __kmp_ncores;
5137 if ((nPackages > 1) && (__kmp_nth_per_core <= 1)) {
5138 __kmp_nth_per_core = __kmp_avail_proc / nPackages;
5139 ncores = nPackages;
5140 }
5141 // How many threads will be bound to each core
5142 int chunk = nthreads / ncores;
5143 // How many cores will have an additional thread bound to it - "big cores"
5144 int big_cores = nthreads % ncores;
5145 // Number of threads on the big cores
5146 int big_nth = (chunk + 1) * big_cores;
5147 if (tid < big_nth) {
5148 coreID = tid / (chunk + 1);
5149 threadID = (tid % (chunk + 1)) % __kmp_nth_per_core;
5150 } else { // tid >= big_nth
5151 coreID = (tid - big_cores) / chunk;
5152 threadID = ((tid - big_cores) % chunk) % __kmp_nth_per_core;
5153 }
5154 KMP_DEBUG_ASSERT2(KMP_AFFINITY_CAPABLE(),if (!((__kmp_affin_mask_size > 0))) { __kmp_debug_assert((
"Illegal set affinity operation when not capable"), "openmp/runtime/src/kmp_affinity.cpp"
, 5155); }
5155 "Illegal set affinity operation when not capable")if (!((__kmp_affin_mask_size > 0))) { __kmp_debug_assert((
"Illegal set affinity operation when not capable"), "openmp/runtime/src/kmp_affinity.cpp"
, 5155); }
;
5156
5157 kmp_affin_mask_t *mask = th->th.th_affin_mask;
5158 KMP_CPU_ZERO(mask)(mask)->zero();
5159
5160 if (fine_gran) {
5161 int osID =
5162 __kmp_topology->at(coreID * __kmp_nth_per_core + threadID).os_id;
5163 KMP_CPU_SET(osID, mask)(mask)->set(osID);
5164 } else {
5165 for (int i = 0; i < __kmp_nth_per_core; i++) {
5166 int osID;
5167 osID = __kmp_topology->at(coreID * __kmp_nth_per_core + i).os_id;
5168 KMP_CPU_SET(osID, mask)(mask)->set(osID);
5169 }
5170 }
5171 if (__kmp_affinity.flags.verbose) {
5172 char buf[KMP_AFFIN_MASK_PRINT_LEN1024];
5173 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN1024, mask);
5174 KMP_INFORM(BoundToOSProcSet, env_var, (kmp_int32)getpid(), __kmp_gettid(),__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_BoundToOSProcSet
, env_var, (kmp_int32)getpid(), syscall(186), tid, buf), __kmp_msg_null
)
5175 tid, buf)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_BoundToOSProcSet
, env_var, (kmp_int32)getpid(), syscall(186), tid, buf), __kmp_msg_null
)
;
5176 }
5177 __kmp_affinity_get_thread_topology_info(th);
5178 __kmp_set_system_affinity(mask, TRUE)(mask)->set_system_affinity((!0));
5179 } else { // Non-uniform topology
5180
5181 kmp_affin_mask_t *mask = th->th.th_affin_mask;
5182 KMP_CPU_ZERO(mask)(mask)->zero();
5183
5184 int core_level =
5185 __kmp_affinity_find_core_level(__kmp_avail_proc, __kmp_aff_depth - 1);
5186 int ncores = __kmp_affinity_compute_ncores(__kmp_avail_proc,
5187 __kmp_aff_depth - 1, core_level);
5188 int nth_per_core = __kmp_affinity_max_proc_per_core(
5189 __kmp_avail_proc, __kmp_aff_depth - 1, core_level);
5190
5191 // For performance gain consider the special case nthreads ==
5192 // __kmp_avail_proc
5193 if (nthreads == __kmp_avail_proc) {
5194 if (fine_gran) {
5195 int osID = __kmp_topology->at(tid).os_id;
5196 KMP_CPU_SET(osID, mask)(mask)->set(osID);
5197 } else {
5198 int core =
5199 __kmp_affinity_find_core(tid, __kmp_aff_depth - 1, core_level);
5200 for (int i = 0; i < __kmp_avail_proc; i++) {
5201 int osID = __kmp_topology->at(i).os_id;
5202 if (__kmp_affinity_find_core(i, __kmp_aff_depth - 1, core_level) ==
5203 core) {
5204 KMP_CPU_SET(osID, mask)(mask)->set(osID);
5205 }
5206 }
5207 }
5208 } else if (nthreads <= ncores) {
5209
5210 int core = 0;
5211 for (int i = 0; i < ncores; i++) {
5212 // Check if this core from procarr[] is in the mask
5213 int in_mask = 0;
5214 for (int j = 0; j < nth_per_core; j++) {
5215 if (procarr[i * nth_per_core + j] != -1) {
5216 in_mask = 1;
5217 break;
5218 }
5219 }
5220 if (in_mask) {
5221 if (tid == core) {
5222 for (int j = 0; j < nth_per_core; j++) {
5223 int osID = procarr[i * nth_per_core + j];
5224 if (osID != -1) {
5225 KMP_CPU_SET(osID, mask)(mask)->set(osID);
5226 // For fine granularity it is enough to set the first available
5227 // osID for this core
5228 if (fine_gran) {
5229 break;
5230 }
5231 }
5232 }
5233 break;
5234 } else {
5235 core++;
5236 }
5237 }
5238 }
5239 } else { // nthreads > ncores
5240 // Array to save the number of processors at each core
5241 int *nproc_at_core = (int *)KMP_ALLOCA(sizeof(int) * ncores)__builtin_alloca (sizeof(int) * ncores);
5242 // Array to save the number of cores with "x" available processors;
5243 int *ncores_with_x_procs =
5244 (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1))__builtin_alloca (sizeof(int) * (nth_per_core + 1));
5245 // Array to save the number of cores with # procs from x to nth_per_core
5246 int *ncores_with_x_to_max_procs =
5247 (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1))__builtin_alloca (sizeof(int) * (nth_per_core + 1));
5248
5249 for (int i = 0; i <= nth_per_core; i++) {
5250 ncores_with_x_procs[i] = 0;
5251 ncores_with_x_to_max_procs[i] = 0;
5252 }
5253
5254 for (int i = 0; i < ncores; i++) {
5255 int cnt = 0;
5256 for (int j = 0; j < nth_per_core; j++) {
5257 if (procarr[i * nth_per_core + j] != -1) {
5258 cnt++;
5259 }
5260 }
5261 nproc_at_core[i] = cnt;
5262 ncores_with_x_procs[cnt]++;
5263 }
5264
5265 for (int i = 0; i <= nth_per_core; i++) {
5266 for (int j = i; j <= nth_per_core; j++) {
5267 ncores_with_x_to_max_procs[i] += ncores_with_x_procs[j];
5268 }
5269 }
5270
5271 // Max number of processors
5272 int nproc = nth_per_core * ncores;
5273 // An array to keep number of threads per each context
5274 int *newarr = (int *)__kmp_allocate(sizeof(int) * nproc)___kmp_allocate((sizeof(int) * nproc), "openmp/runtime/src/kmp_affinity.cpp"
, 5274)
;
5275 for (int i = 0; i < nproc; i++) {
5276 newarr[i] = 0;
5277 }
5278
5279 int nth = nthreads;
5280 int flag = 0;
5281 while (nth > 0) {
5282 for (int j = 1; j <= nth_per_core; j++) {
5283 int cnt = ncores_with_x_to_max_procs[j];
5284 for (int i = 0; i < ncores; i++) {
5285 // Skip the core with 0 processors
5286 if (nproc_at_core[i] == 0) {
5287 continue;
5288 }
5289 for (int k = 0; k < nth_per_core; k++) {
5290 if (procarr[i * nth_per_core + k] != -1) {
5291 if (newarr[i * nth_per_core + k] == 0) {
5292 newarr[i * nth_per_core + k] = 1;
5293 cnt--;
5294 nth--;
5295 break;
5296 } else {
5297 if (flag != 0) {
5298 newarr[i * nth_per_core + k]++;
5299 cnt--;
5300 nth--;
5301 break;
5302 }
5303 }
5304 }
5305 }
5306 if (cnt == 0 || nth == 0) {
5307 break;
5308 }
5309 }
5310 if (nth == 0) {
5311 break;
5312 }
5313 }
5314 flag = 1;
5315 }
5316 int sum = 0;
5317 for (int i = 0; i < nproc; i++) {
5318 sum += newarr[i];
5319 if (sum > tid) {
5320 if (fine_gran) {
5321 int osID = procarr[i];
5322 KMP_CPU_SET(osID, mask)(mask)->set(osID);
5323 } else {
5324 int coreID = i / nth_per_core;
5325 for (int ii = 0; ii < nth_per_core; ii++) {
5326 int osID = procarr[coreID * nth_per_core + ii];
5327 if (osID != -1) {
5328 KMP_CPU_SET(osID, mask)(mask)->set(osID);
5329 }
5330 }
5331 }
5332 break;
5333 }
5334 }
5335 __kmp_free(newarr)___kmp_free((newarr), "openmp/runtime/src/kmp_affinity.cpp", 5335
)
;
5336 }
5337
5338 if (__kmp_affinity.flags.verbose) {
5339 char buf[KMP_AFFIN_MASK_PRINT_LEN1024];
5340 __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN1024, mask);
5341 KMP_INFORM(BoundToOSProcSet, env_var, (kmp_int32)getpid(), __kmp_gettid(),__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_BoundToOSProcSet
, env_var, (kmp_int32)getpid(), syscall(186), tid, buf), __kmp_msg_null
)
5342 tid, buf)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_BoundToOSProcSet
, env_var, (kmp_int32)getpid(), syscall(186), tid, buf), __kmp_msg_null
)
;
5343 }
5344 __kmp_affinity_get_thread_topology_info(th);
5345 __kmp_set_system_affinity(mask, TRUE)(mask)->set_system_affinity((!0));
5346 }
5347}
5348
5349#if KMP_OS_LINUX1 || KMP_OS_FREEBSD0
5350// We don't need this entry for Windows because
5351// there is GetProcessAffinityMask() api
5352//
5353// The intended usage is indicated by these steps:
5354// 1) The user gets the current affinity mask
5355// 2) Then sets the affinity by calling this function
5356// 3) Error check the return value
5357// 4) Use non-OpenMP parallelization
5358// 5) Reset the affinity to what was stored in step 1)
5359#ifdef __cplusplus201703L
5360extern "C"
5361#endif
5362 int
5363 kmp_set_thread_affinity_mask_initial()
5364// the function returns 0 on success,
5365// -1 if we cannot bind thread
5366// >0 (errno) if an error happened during binding
5367{
5368 int gtid = __kmp_get_gtid()__kmp_get_global_thread_id();
5369 if (gtid < 0) {
5370 // Do not touch non-omp threads
5371 KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "if (kmp_a_debug >= 30) { __kmp_debug_printf ("kmp_set_thread_affinity_mask_initial: "
"non-omp thread, returning\n"); }
5372 "non-omp thread, returning\n"))if (kmp_a_debug >= 30) { __kmp_debug_printf ("kmp_set_thread_affinity_mask_initial: "
"non-omp thread, returning\n"); }
;
5373 return -1;
5374 }
5375 if (!KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0) || !__kmp_init_middle) {
5376 KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "if (kmp_a_debug >= 30) { __kmp_debug_printf ("kmp_set_thread_affinity_mask_initial: "
"affinity not initialized, returning\n"); }
5377 "affinity not initialized, returning\n"))if (kmp_a_debug >= 30) { __kmp_debug_printf ("kmp_set_thread_affinity_mask_initial: "
"affinity not initialized, returning\n"); }
;
5378 return -1;
5379 }
5380 KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "if (kmp_a_debug >= 30) { __kmp_debug_printf ("kmp_set_thread_affinity_mask_initial: "
"set full mask for thread %d\n", gtid); }
5381 "set full mask for thread %d\n",if (kmp_a_debug >= 30) { __kmp_debug_printf ("kmp_set_thread_affinity_mask_initial: "
"set full mask for thread %d\n", gtid); }
5382 gtid))if (kmp_a_debug >= 30) { __kmp_debug_printf ("kmp_set_thread_affinity_mask_initial: "
"set full mask for thread %d\n", gtid); }
;
5383 KMP_DEBUG_ASSERT(__kmp_affin_fullMask != NULL)if (!(__kmp_affin_fullMask != __null)) { __kmp_debug_assert("__kmp_affin_fullMask != __null"
, "openmp/runtime/src/kmp_affinity.cpp", 5383); }
;
5384 return __kmp_set_system_affinity(__kmp_affin_fullMask, FALSE)(__kmp_affin_fullMask)->set_system_affinity(0);
5385}
5386#endif
5387
5388#endif // KMP_AFFINITY_SUPPORTED