File: | build/source/openmp/runtime/src/kmp_runtime.cpp |
Warning: | line 5196, column 12 Dereference of undefined pointer value |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | /* | ||||
2 | * kmp_runtime.cpp -- KPTS runtime support library | ||||
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_atomic.h" | ||||
16 | #include "kmp_environment.h" | ||||
17 | #include "kmp_error.h" | ||||
18 | #include "kmp_i18n.h" | ||||
19 | #include "kmp_io.h" | ||||
20 | #include "kmp_itt.h" | ||||
21 | #include "kmp_settings.h" | ||||
22 | #include "kmp_stats.h" | ||||
23 | #include "kmp_str.h" | ||||
24 | #include "kmp_wait_release.h" | ||||
25 | #include "kmp_wrapper_getpid.h" | ||||
26 | #include "kmp_dispatch.h" | ||||
27 | #if KMP_USE_HIER_SCHED0 | ||||
28 | #include "kmp_dispatch_hier.h" | ||||
29 | #endif | ||||
30 | |||||
31 | #if OMPT_SUPPORT1 | ||||
32 | #include "ompt-specific.h" | ||||
33 | #endif | ||||
34 | #if OMPD_SUPPORT1 | ||||
35 | #include "ompd-specific.h" | ||||
36 | #endif | ||||
37 | |||||
38 | #if OMP_PROFILING_SUPPORT0 | ||||
39 | #include "llvm/Support/TimeProfiler.h" | ||||
40 | static char *ProfileTraceFile = nullptr; | ||||
41 | #endif | ||||
42 | |||||
43 | /* these are temporary issues to be dealt with */ | ||||
44 | #define KMP_USE_PRCTL0 0 | ||||
45 | |||||
46 | #if KMP_OS_WINDOWS0 | ||||
47 | #include <process.h> | ||||
48 | #endif | ||||
49 | |||||
50 | #if KMP_OS_WINDOWS0 | ||||
51 | // windows does not need include files as it doesn't use shared memory | ||||
52 | #else | ||||
53 | #include <sys/mman.h> | ||||
54 | #include <sys/stat.h> | ||||
55 | #include <fcntl.h> | ||||
56 | #define SHM_SIZE1024 1024 | ||||
57 | #endif | ||||
58 | |||||
59 | #if defined(KMP_GOMP_COMPAT) | ||||
60 | char const __kmp_version_alt_comp[] = | ||||
61 | KMP_VERSION_PREFIX"\x00@(#) " "LLVM OMP " "alternative compiler support: yes"; | ||||
62 | #endif /* defined(KMP_GOMP_COMPAT) */ | ||||
63 | |||||
64 | char const __kmp_version_omp_api[] = | ||||
65 | KMP_VERSION_PREFIX"\x00@(#) " "LLVM OMP " "API version: 5.0 (201611)"; | ||||
66 | |||||
67 | #ifdef KMP_DEBUG1 | ||||
68 | char const __kmp_version_lock[] = | ||||
69 | KMP_VERSION_PREFIX"\x00@(#) " "LLVM OMP " "lock type: run time selectable"; | ||||
70 | #endif /* KMP_DEBUG */ | ||||
71 | |||||
72 | #define KMP_MIN(x, y)((x) < (y) ? (x) : (y)) ((x) < (y) ? (x) : (y)) | ||||
73 | |||||
74 | /* ------------------------------------------------------------------------ */ | ||||
75 | |||||
76 | #if KMP_USE_MONITOR | ||||
77 | kmp_info_t __kmp_monitor; | ||||
78 | #endif | ||||
79 | |||||
80 | /* Forward declarations */ | ||||
81 | |||||
82 | void __kmp_cleanup(void); | ||||
83 | |||||
84 | static void __kmp_initialize_info(kmp_info_t *, kmp_team_t *, int tid, | ||||
85 | int gtid); | ||||
86 | static void __kmp_initialize_team(kmp_team_t *team, int new_nproc, | ||||
87 | kmp_internal_control_t *new_icvs, | ||||
88 | ident_t *loc); | ||||
89 | #if KMP_AFFINITY_SUPPORTED1 | ||||
90 | static void __kmp_partition_places(kmp_team_t *team, | ||||
91 | int update_master_only = 0); | ||||
92 | #endif | ||||
93 | static void __kmp_do_serial_initialize(void); | ||||
94 | void __kmp_fork_barrier(int gtid, int tid); | ||||
95 | void __kmp_join_barrier(int gtid); | ||||
96 | void __kmp_setup_icv_copy(kmp_team_t *team, int new_nproc, | ||||
97 | kmp_internal_control_t *new_icvs, ident_t *loc); | ||||
98 | |||||
99 | #ifdef USE_LOAD_BALANCE1 | ||||
100 | static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc); | ||||
101 | #endif | ||||
102 | |||||
103 | static int __kmp_expand_threads(int nNeed); | ||||
104 | #if KMP_OS_WINDOWS0 | ||||
105 | static int __kmp_unregister_root_other_thread(int gtid); | ||||
106 | #endif | ||||
107 | static void __kmp_reap_thread(kmp_info_t *thread, int is_root); | ||||
108 | kmp_info_t *__kmp_thread_pool_insert_pt = NULL__null; | ||||
109 | |||||
110 | void __kmp_resize_dist_barrier(kmp_team_t *team, int old_nthreads, | ||||
111 | int new_nthreads); | ||||
112 | void __kmp_add_threads_to_team(kmp_team_t *team, int new_nthreads); | ||||
113 | |||||
114 | /* Calculate the identifier of the current thread */ | ||||
115 | /* fast (and somewhat portable) way to get unique identifier of executing | ||||
116 | thread. Returns KMP_GTID_DNE if we haven't been assigned a gtid. */ | ||||
117 | int __kmp_get_global_thread_id() { | ||||
118 | int i; | ||||
119 | kmp_info_t **other_threads; | ||||
120 | size_t stack_data; | ||||
121 | char *stack_addr; | ||||
122 | size_t stack_size; | ||||
123 | char *stack_base; | ||||
124 | |||||
125 | KA_TRACE(if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id: entering, nproc=%d all_nproc=%d\n" , __kmp_nth, __kmp_all_nth); } | ||||
126 | 1000,if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id: entering, nproc=%d all_nproc=%d\n" , __kmp_nth, __kmp_all_nth); } | ||||
127 | ("*** __kmp_get_global_thread_id: entering, nproc=%d all_nproc=%d\n",if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id: entering, nproc=%d all_nproc=%d\n" , __kmp_nth, __kmp_all_nth); } | ||||
128 | __kmp_nth, __kmp_all_nth))if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id: entering, nproc=%d all_nproc=%d\n" , __kmp_nth, __kmp_all_nth); }; | ||||
129 | |||||
130 | /* JPH - to handle the case where __kmpc_end(0) is called immediately prior to | ||||
131 | a parallel region, made it return KMP_GTID_DNE to force serial_initialize | ||||
132 | by caller. Had to handle KMP_GTID_DNE at all call-sites, or else guarantee | ||||
133 | __kmp_init_gtid for this to work. */ | ||||
134 | |||||
135 | if (!TCR_4(__kmp_init_gtid)(__kmp_init_gtid)) | ||||
136 | return KMP_GTID_DNE(-2); | ||||
137 | |||||
138 | #ifdef KMP_TDATA_GTID1 | ||||
139 | if (TCR_4(__kmp_gtid_mode)(__kmp_gtid_mode) >= 3) { | ||||
140 | KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using TDATA\n"))if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id: using TDATA\n" ); }; | ||||
141 | return __kmp_gtid; | ||||
142 | } | ||||
143 | #endif | ||||
144 | if (TCR_4(__kmp_gtid_mode)(__kmp_gtid_mode) >= 2) { | ||||
145 | KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using keyed TLS\n"))if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id: using keyed TLS\n" ); }; | ||||
146 | return __kmp_gtid_get_specific(); | ||||
147 | } | ||||
148 | KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using internal alg.\n"))if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id: using internal alg.\n" ); }; | ||||
149 | |||||
150 | stack_addr = (char *)&stack_data; | ||||
151 | other_threads = __kmp_threads; | ||||
152 | |||||
153 | /* ATT: The code below is a source of potential bugs due to unsynchronized | ||||
154 | access to __kmp_threads array. For example: | ||||
155 | 1. Current thread loads other_threads[i] to thr and checks it, it is | ||||
156 | non-NULL. | ||||
157 | 2. Current thread is suspended by OS. | ||||
158 | 3. Another thread unregisters and finishes (debug versions of free() | ||||
159 | may fill memory with something like 0xEF). | ||||
160 | 4. Current thread is resumed. | ||||
161 | 5. Current thread reads junk from *thr. | ||||
162 | TODO: Fix it. --ln */ | ||||
163 | |||||
164 | for (i = 0; i < __kmp_threads_capacity; i++) { | ||||
165 | |||||
166 | kmp_info_t *thr = (kmp_info_t *)TCR_SYNC_PTR(other_threads[i])((void *)(other_threads[i])); | ||||
167 | if (!thr) | ||||
168 | continue; | ||||
169 | |||||
170 | stack_size = (size_t)TCR_PTR(thr->th.th_info.ds.ds_stacksize)((void *)(thr->th.th_info.ds.ds_stacksize)); | ||||
171 | stack_base = (char *)TCR_PTR(thr->th.th_info.ds.ds_stackbase)((void *)(thr->th.th_info.ds.ds_stackbase)); | ||||
172 | |||||
173 | /* stack grows down -- search through all of the active threads */ | ||||
174 | |||||
175 | if (stack_addr <= stack_base) { | ||||
176 | size_t stack_diff = stack_base - stack_addr; | ||||
177 | |||||
178 | if (stack_diff <= stack_size) { | ||||
179 | /* The only way we can be closer than the allocated */ | ||||
180 | /* stack size is if we are running on this thread. */ | ||||
181 | KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == i)if (!(__kmp_gtid_get_specific() == i)) { __kmp_debug_assert("__kmp_gtid_get_specific() == i" , "openmp/runtime/src/kmp_runtime.cpp", 181); }; | ||||
182 | return i; | ||||
183 | } | ||||
184 | } | ||||
185 | } | ||||
186 | |||||
187 | /* get specific to try and determine our gtid */ | ||||
188 | KA_TRACE(1000,if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id: internal alg. failed to find " "thread, using TLS\n"); } | ||||
189 | ("*** __kmp_get_global_thread_id: internal alg. failed to find "if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id: internal alg. failed to find " "thread, using TLS\n"); } | ||||
190 | "thread, using TLS\n"))if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id: internal alg. failed to find " "thread, using TLS\n"); }; | ||||
191 | i = __kmp_gtid_get_specific(); | ||||
192 | |||||
193 | /*fprintf( stderr, "=== %d\n", i ); */ /* GROO */ | ||||
194 | |||||
195 | /* if we havn't been assigned a gtid, then return code */ | ||||
196 | if (i < 0) | ||||
197 | return i; | ||||
198 | |||||
199 | /* dynamically updated stack window for uber threads to avoid get_specific | ||||
200 | call */ | ||||
201 | if (!TCR_4(other_threads[i]->th.th_info.ds.ds_stackgrow)(other_threads[i]->th.th_info.ds.ds_stackgrow)) { | ||||
202 | KMP_FATAL(StackOverflow, i)__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_StackOverflow, i), __kmp_msg_null ); | ||||
203 | } | ||||
204 | |||||
205 | stack_base = (char *)other_threads[i]->th.th_info.ds.ds_stackbase; | ||||
206 | if (stack_addr > stack_base) { | ||||
207 | TCW_PTR(other_threads[i]->th.th_info.ds.ds_stackbase, stack_addr)((other_threads[i]->th.th_info.ds.ds_stackbase)) = ((stack_addr )); | ||||
208 | TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize,((other_threads[i]->th.th_info.ds.ds_stacksize)) = ((other_threads [i]->th.th_info.ds.ds_stacksize + stack_addr - stack_base) ) | ||||
209 | other_threads[i]->th.th_info.ds.ds_stacksize + stack_addr -((other_threads[i]->th.th_info.ds.ds_stacksize)) = ((other_threads [i]->th.th_info.ds.ds_stacksize + stack_addr - stack_base) ) | ||||
210 | stack_base)((other_threads[i]->th.th_info.ds.ds_stacksize)) = ((other_threads [i]->th.th_info.ds.ds_stacksize + stack_addr - stack_base) ); | ||||
211 | } else { | ||||
212 | TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize,((other_threads[i]->th.th_info.ds.ds_stacksize)) = ((stack_base - stack_addr)) | ||||
213 | stack_base - stack_addr)((other_threads[i]->th.th_info.ds.ds_stacksize)) = ((stack_base - stack_addr)); | ||||
214 | } | ||||
215 | |||||
216 | /* Reprint stack bounds for ubermaster since they have been refined */ | ||||
217 | if (__kmp_storage_map) { | ||||
218 | char *stack_end = (char *)other_threads[i]->th.th_info.ds.ds_stackbase; | ||||
219 | char *stack_beg = stack_end - other_threads[i]->th.th_info.ds.ds_stacksize; | ||||
220 | __kmp_print_storage_map_gtid(i, stack_beg, stack_end, | ||||
221 | other_threads[i]->th.th_info.ds.ds_stacksize, | ||||
222 | "th_%d stack (refinement)", i); | ||||
223 | } | ||||
224 | return i; | ||||
225 | } | ||||
226 | |||||
227 | int __kmp_get_global_thread_id_reg() { | ||||
228 | int gtid; | ||||
229 | |||||
230 | if (!__kmp_init_serial) { | ||||
231 | gtid = KMP_GTID_DNE(-2); | ||||
232 | } else | ||||
233 | #ifdef KMP_TDATA_GTID1 | ||||
234 | if (TCR_4(__kmp_gtid_mode)(__kmp_gtid_mode) >= 3) { | ||||
235 | KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using TDATA\n"))if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id_reg: using TDATA\n" ); }; | ||||
236 | gtid = __kmp_gtid; | ||||
237 | } else | ||||
238 | #endif | ||||
239 | if (TCR_4(__kmp_gtid_mode)(__kmp_gtid_mode) >= 2) { | ||||
240 | KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using keyed TLS\n"))if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id_reg: using keyed TLS\n" ); }; | ||||
241 | gtid = __kmp_gtid_get_specific(); | ||||
242 | } else { | ||||
243 | KA_TRACE(1000,if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id_reg: using internal alg.\n" ); } | ||||
244 | ("*** __kmp_get_global_thread_id_reg: using internal alg.\n"))if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id_reg: using internal alg.\n" ); }; | ||||
245 | gtid = __kmp_get_global_thread_id(); | ||||
246 | } | ||||
247 | |||||
248 | /* we must be a new uber master sibling thread */ | ||||
249 | if (gtid == KMP_GTID_DNE(-2)) { | ||||
250 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_get_global_thread_id_reg: Encountered new root thread. " "Registering a new gtid.\n"); } | ||||
251 | ("__kmp_get_global_thread_id_reg: Encountered new root thread. "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_get_global_thread_id_reg: Encountered new root thread. " "Registering a new gtid.\n"); } | ||||
252 | "Registering a new gtid.\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_get_global_thread_id_reg: Encountered new root thread. " "Registering a new gtid.\n"); }; | ||||
253 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); | ||||
254 | if (!__kmp_init_serial) { | ||||
255 | __kmp_do_serial_initialize(); | ||||
256 | gtid = __kmp_gtid_get_specific(); | ||||
257 | } else { | ||||
258 | gtid = __kmp_register_root(FALSE0); | ||||
259 | } | ||||
260 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | ||||
261 | /*__kmp_printf( "+++ %d\n", gtid ); */ /* GROO */ | ||||
262 | } | ||||
263 | |||||
264 | KMP_DEBUG_ASSERT(gtid >= 0)if (!(gtid >= 0)) { __kmp_debug_assert("gtid >= 0", "openmp/runtime/src/kmp_runtime.cpp" , 264); }; | ||||
265 | |||||
266 | return gtid; | ||||
267 | } | ||||
268 | |||||
269 | /* caller must hold forkjoin_lock */ | ||||
270 | void __kmp_check_stack_overlap(kmp_info_t *th) { | ||||
271 | int f; | ||||
272 | char *stack_beg = NULL__null; | ||||
273 | char *stack_end = NULL__null; | ||||
274 | int gtid; | ||||
275 | |||||
276 | KA_TRACE(10, ("__kmp_check_stack_overlap: called\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_check_stack_overlap: called\n" ); }; | ||||
277 | if (__kmp_storage_map) { | ||||
278 | stack_end = (char *)th->th.th_info.ds.ds_stackbase; | ||||
279 | stack_beg = stack_end - th->th.th_info.ds.ds_stacksize; | ||||
280 | |||||
281 | gtid = __kmp_gtid_from_thread(th); | ||||
282 | |||||
283 | if (gtid == KMP_GTID_MONITOR(-4)) { | ||||
284 | __kmp_print_storage_map_gtid( | ||||
285 | gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize, | ||||
286 | "th_%s stack (%s)", "mon", | ||||
287 | (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual"); | ||||
288 | } else { | ||||
289 | __kmp_print_storage_map_gtid( | ||||
290 | gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize, | ||||
291 | "th_%d stack (%s)", gtid, | ||||
292 | (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual"); | ||||
293 | } | ||||
294 | } | ||||
295 | |||||
296 | /* No point in checking ubermaster threads since they use refinement and | ||||
297 | * cannot overlap */ | ||||
298 | gtid = __kmp_gtid_from_thread(th); | ||||
299 | if (__kmp_env_checks == TRUE(!0) && !KMP_UBER_GTID(gtid)) { | ||||
300 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_check_stack_overlap: performing extensive checking\n" ); } | ||||
301 | ("__kmp_check_stack_overlap: performing extensive checking\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_check_stack_overlap: performing extensive checking\n" ); }; | ||||
302 | if (stack_beg == NULL__null) { | ||||
303 | stack_end = (char *)th->th.th_info.ds.ds_stackbase; | ||||
304 | stack_beg = stack_end - th->th.th_info.ds.ds_stacksize; | ||||
305 | } | ||||
306 | |||||
307 | for (f = 0; f < __kmp_threads_capacity; f++) { | ||||
308 | kmp_info_t *f_th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[f])((void *)(__kmp_threads[f])); | ||||
309 | |||||
310 | if (f_th && f_th != th) { | ||||
311 | char *other_stack_end = | ||||
312 | (char *)TCR_PTR(f_th->th.th_info.ds.ds_stackbase)((void *)(f_th->th.th_info.ds.ds_stackbase)); | ||||
313 | char *other_stack_beg = | ||||
314 | other_stack_end - (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize)((void *)(f_th->th.th_info.ds.ds_stacksize)); | ||||
315 | if ((stack_beg > other_stack_beg && stack_beg < other_stack_end) || | ||||
316 | (stack_end > other_stack_beg && stack_end < other_stack_end)) { | ||||
317 | |||||
318 | /* Print the other stack values before the abort */ | ||||
319 | if (__kmp_storage_map) | ||||
320 | __kmp_print_storage_map_gtid( | ||||
321 | -1, other_stack_beg, other_stack_end, | ||||
322 | (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize)((void *)(f_th->th.th_info.ds.ds_stacksize)), | ||||
323 | "th_%d stack (overlapped)", __kmp_gtid_from_thread(f_th)); | ||||
324 | |||||
325 | __kmp_fatal(KMP_MSG(StackOverlap)__kmp_msg_format(kmp_i18n_msg_StackOverlap), KMP_HNT(ChangeStackLimit)__kmp_msg_format(kmp_i18n_hnt_ChangeStackLimit), | ||||
326 | __kmp_msg_null); | ||||
327 | } | ||||
328 | } | ||||
329 | } | ||||
330 | } | ||||
331 | KA_TRACE(10, ("__kmp_check_stack_overlap: returning\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_check_stack_overlap: returning\n" ); }; | ||||
332 | } | ||||
333 | |||||
334 | /* ------------------------------------------------------------------------ */ | ||||
335 | |||||
336 | void __kmp_infinite_loop(void) { | ||||
337 | static int done = FALSE0; | ||||
338 | |||||
339 | while (!done) { | ||||
340 | KMP_YIELD(TRUE){ __kmp_x86_pause(); if (((!0)) && (((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))))))) __kmp_yield (); }; | ||||
341 | } | ||||
342 | } | ||||
343 | |||||
344 | #define MAX_MESSAGE512 512 | ||||
345 | |||||
346 | void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2, size_t size, | ||||
347 | char const *format, ...) { | ||||
348 | char buffer[MAX_MESSAGE512]; | ||||
349 | va_list ap; | ||||
350 | |||||
351 | va_start(ap, format)__builtin_va_start(ap, format); | ||||
352 | KMP_SNPRINTFsnprintf(buffer, sizeof(buffer), "OMP storage map: %p %p%8lu %s\n", p1, | ||||
353 | p2, (unsigned long)size, format); | ||||
354 | __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock); | ||||
355 | __kmp_vprintf(kmp_err, buffer, ap); | ||||
356 | #if KMP_PRINT_DATA_PLACEMENT | ||||
357 | int node; | ||||
358 | if (gtid >= 0) { | ||||
359 | if (p1 <= p2 && (char *)p2 - (char *)p1 == size) { | ||||
360 | if (__kmp_storage_map_verbose) { | ||||
361 | node = __kmp_get_host_node(p1); | ||||
362 | if (node < 0) /* doesn't work, so don't try this next time */ | ||||
363 | __kmp_storage_map_verbose = FALSE0; | ||||
364 | else { | ||||
365 | char *last; | ||||
366 | int lastNode; | ||||
367 | int localProc = __kmp_get_cpu_from_gtid(gtid); | ||||
368 | |||||
369 | const int page_size = KMP_GET_PAGE_SIZE()getpagesize(); | ||||
370 | |||||
371 | p1 = (void *)((size_t)p1 & ~((size_t)page_size - 1)); | ||||
372 | p2 = (void *)(((size_t)p2 - 1) & ~((size_t)page_size - 1)); | ||||
373 | if (localProc >= 0) | ||||
374 | __kmp_printf_no_lock(" GTID %d localNode %d\n", gtid, | ||||
375 | localProc >> 1); | ||||
376 | else | ||||
377 | __kmp_printf_no_lock(" GTID %d\n", gtid); | ||||
378 | #if KMP_USE_PRCTL0 | ||||
379 | /* The more elaborate format is disabled for now because of the prctl | ||||
380 | * hanging bug. */ | ||||
381 | do { | ||||
382 | last = p1; | ||||
383 | lastNode = node; | ||||
384 | /* This loop collates adjacent pages with the same host node. */ | ||||
385 | do { | ||||
386 | (char *)p1 += page_size; | ||||
387 | } while (p1 <= p2 && (node = __kmp_get_host_node(p1)) == lastNode); | ||||
388 | __kmp_printf_no_lock(" %p-%p memNode %d\n", last, (char *)p1 - 1, | ||||
389 | lastNode); | ||||
390 | } while (p1 <= p2); | ||||
391 | #else | ||||
392 | __kmp_printf_no_lock(" %p-%p memNode %d\n", p1, | ||||
393 | (char *)p1 + (page_size - 1), | ||||
394 | __kmp_get_host_node(p1)); | ||||
395 | if (p1 < p2) { | ||||
396 | __kmp_printf_no_lock(" %p-%p memNode %d\n", p2, | ||||
397 | (char *)p2 + (page_size - 1), | ||||
398 | __kmp_get_host_node(p2)); | ||||
399 | } | ||||
400 | #endif | ||||
401 | } | ||||
402 | } | ||||
403 | } else | ||||
404 | __kmp_printf_no_lock(" %s\n", KMP_I18N_STR(StorageMapWarning)__kmp_i18n_catgets(kmp_i18n_str_StorageMapWarning)); | ||||
405 | } | ||||
406 | #endif /* KMP_PRINT_DATA_PLACEMENT */ | ||||
407 | __kmp_release_bootstrap_lock(&__kmp_stdio_lock); | ||||
408 | } | ||||
409 | |||||
410 | void __kmp_warn(char const *format, ...) { | ||||
411 | char buffer[MAX_MESSAGE512]; | ||||
412 | va_list ap; | ||||
413 | |||||
414 | if (__kmp_generate_warnings == kmp_warnings_off) { | ||||
415 | return; | ||||
416 | } | ||||
417 | |||||
418 | va_start(ap, format)__builtin_va_start(ap, format); | ||||
419 | |||||
420 | KMP_SNPRINTFsnprintf(buffer, sizeof(buffer), "OMP warning: %s\n", format); | ||||
421 | __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock); | ||||
422 | __kmp_vprintf(kmp_err, buffer, ap); | ||||
423 | __kmp_release_bootstrap_lock(&__kmp_stdio_lock); | ||||
424 | |||||
425 | va_end(ap)__builtin_va_end(ap); | ||||
426 | } | ||||
427 | |||||
428 | void __kmp_abort_process() { | ||||
429 | // Later threads may stall here, but that's ok because abort() will kill them. | ||||
430 | __kmp_acquire_bootstrap_lock(&__kmp_exit_lock); | ||||
431 | |||||
432 | if (__kmp_debug_buf) { | ||||
433 | __kmp_dump_debug_buffer(); | ||||
434 | } | ||||
435 | |||||
436 | if (KMP_OS_WINDOWS0) { | ||||
437 | // Let other threads know of abnormal termination and prevent deadlock | ||||
438 | // if abort happened during library initialization or shutdown | ||||
439 | __kmp_global.g.g_abort = SIGABRT6; | ||||
440 | |||||
441 | /* On Windows* OS by default abort() causes pop-up error box, which stalls | ||||
442 | nightly testing. Unfortunately, we cannot reliably suppress pop-up error | ||||
443 | boxes. _set_abort_behavior() works well, but this function is not | ||||
444 | available in VS7 (this is not problem for DLL, but it is a problem for | ||||
445 | static OpenMP RTL). SetErrorMode (and so, timelimit utility) does not | ||||
446 | help, at least in some versions of MS C RTL. | ||||
447 | |||||
448 | It seems following sequence is the only way to simulate abort() and | ||||
449 | avoid pop-up error box. */ | ||||
450 | raise(SIGABRT6); | ||||
451 | _exit(3); // Just in case, if signal ignored, exit anyway. | ||||
452 | } else { | ||||
453 | __kmp_unregister_library(); | ||||
454 | abort(); | ||||
455 | } | ||||
456 | |||||
457 | __kmp_infinite_loop(); | ||||
458 | __kmp_release_bootstrap_lock(&__kmp_exit_lock); | ||||
459 | |||||
460 | } // __kmp_abort_process | ||||
461 | |||||
462 | void __kmp_abort_thread(void) { | ||||
463 | // TODO: Eliminate g_abort global variable and this function. | ||||
464 | // In case of abort just call abort(), it will kill all the threads. | ||||
465 | __kmp_infinite_loop(); | ||||
466 | } // __kmp_abort_thread | ||||
467 | |||||
468 | /* Print out the storage map for the major kmp_info_t thread data structures | ||||
469 | that are allocated together. */ | ||||
470 | |||||
471 | static void __kmp_print_thread_storage_map(kmp_info_t *thr, int gtid) { | ||||
472 | __kmp_print_storage_map_gtid(gtid, thr, thr + 1, sizeof(kmp_info_t), "th_%d", | ||||
473 | gtid); | ||||
474 | |||||
475 | __kmp_print_storage_map_gtid(gtid, &thr->th.th_info, &thr->th.th_team, | ||||
476 | sizeof(kmp_desc_t), "th_%d.th_info", gtid); | ||||
477 | |||||
478 | __kmp_print_storage_map_gtid(gtid, &thr->th.th_local, &thr->th.th_pri_head, | ||||
479 | sizeof(kmp_local_t), "th_%d.th_local", gtid); | ||||
480 | |||||
481 | __kmp_print_storage_map_gtid( | ||||
482 | gtid, &thr->th.th_bar[0], &thr->th.th_bar[bs_last_barrier], | ||||
483 | sizeof(kmp_balign_t) * bs_last_barrier, "th_%d.th_bar", gtid); | ||||
484 | |||||
485 | __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_plain_barrier], | ||||
486 | &thr->th.th_bar[bs_plain_barrier + 1], | ||||
487 | sizeof(kmp_balign_t), "th_%d.th_bar[plain]", | ||||
488 | gtid); | ||||
489 | |||||
490 | __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_forkjoin_barrier], | ||||
491 | &thr->th.th_bar[bs_forkjoin_barrier + 1], | ||||
492 | sizeof(kmp_balign_t), "th_%d.th_bar[forkjoin]", | ||||
493 | gtid); | ||||
494 | |||||
495 | #if KMP_FAST_REDUCTION_BARRIER1 | ||||
496 | __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_reduction_barrier], | ||||
497 | &thr->th.th_bar[bs_reduction_barrier + 1], | ||||
498 | sizeof(kmp_balign_t), "th_%d.th_bar[reduction]", | ||||
499 | gtid); | ||||
500 | #endif // KMP_FAST_REDUCTION_BARRIER | ||||
501 | } | ||||
502 | |||||
503 | /* Print out the storage map for the major kmp_team_t team data structures | ||||
504 | that are allocated together. */ | ||||
505 | |||||
506 | static void __kmp_print_team_storage_map(const char *header, kmp_team_t *team, | ||||
507 | int team_id, int num_thr) { | ||||
508 | int num_disp_buff = team->t.t_max_nproc > 1 ? __kmp_dispatch_num_buffers : 2; | ||||
509 | __kmp_print_storage_map_gtid(-1, team, team + 1, sizeof(kmp_team_t), "%s_%d", | ||||
510 | header, team_id); | ||||
511 | |||||
512 | __kmp_print_storage_map_gtid(-1, &team->t.t_bar[0], | ||||
513 | &team->t.t_bar[bs_last_barrier], | ||||
514 | sizeof(kmp_balign_team_t) * bs_last_barrier, | ||||
515 | "%s_%d.t_bar", header, team_id); | ||||
516 | |||||
517 | __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_plain_barrier], | ||||
518 | &team->t.t_bar[bs_plain_barrier + 1], | ||||
519 | sizeof(kmp_balign_team_t), "%s_%d.t_bar[plain]", | ||||
520 | header, team_id); | ||||
521 | |||||
522 | __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_forkjoin_barrier], | ||||
523 | &team->t.t_bar[bs_forkjoin_barrier + 1], | ||||
524 | sizeof(kmp_balign_team_t), | ||||
525 | "%s_%d.t_bar[forkjoin]", header, team_id); | ||||
526 | |||||
527 | #if KMP_FAST_REDUCTION_BARRIER1 | ||||
528 | __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_reduction_barrier], | ||||
529 | &team->t.t_bar[bs_reduction_barrier + 1], | ||||
530 | sizeof(kmp_balign_team_t), | ||||
531 | "%s_%d.t_bar[reduction]", header, team_id); | ||||
532 | #endif // KMP_FAST_REDUCTION_BARRIER | ||||
533 | |||||
534 | __kmp_print_storage_map_gtid( | ||||
535 | -1, &team->t.t_dispatch[0], &team->t.t_dispatch[num_thr], | ||||
536 | sizeof(kmp_disp_t) * num_thr, "%s_%d.t_dispatch", header, team_id); | ||||
537 | |||||
538 | __kmp_print_storage_map_gtid( | ||||
539 | -1, &team->t.t_threads[0], &team->t.t_threads[num_thr], | ||||
540 | sizeof(kmp_info_t *) * num_thr, "%s_%d.t_threads", header, team_id); | ||||
541 | |||||
542 | __kmp_print_storage_map_gtid(-1, &team->t.t_disp_buffer[0], | ||||
543 | &team->t.t_disp_buffer[num_disp_buff], | ||||
544 | sizeof(dispatch_shared_info_t) * num_disp_buff, | ||||
545 | "%s_%d.t_disp_buffer", header, team_id); | ||||
546 | } | ||||
547 | |||||
548 | static void __kmp_init_allocator() { | ||||
549 | __kmp_init_memkind(); | ||||
550 | __kmp_init_target_mem(); | ||||
551 | } | ||||
552 | static void __kmp_fini_allocator() { __kmp_fini_memkind(); } | ||||
553 | |||||
554 | /* ------------------------------------------------------------------------ */ | ||||
555 | |||||
556 | #if ENABLE_LIBOMPTARGET1 | ||||
557 | static void __kmp_init_omptarget() { | ||||
558 | __kmp_init_target_task(); | ||||
559 | } | ||||
560 | #endif | ||||
561 | |||||
562 | /* ------------------------------------------------------------------------ */ | ||||
563 | |||||
564 | #if KMP_DYNAMIC_LIB1 | ||||
565 | #if KMP_OS_WINDOWS0 | ||||
566 | |||||
567 | BOOL WINAPI DllMain(HINSTANCE hInstDLL, DWORD fdwReason, LPVOID lpReserved) { | ||||
568 | //__kmp_acquire_bootstrap_lock( &__kmp_initz_lock ); | ||||
569 | |||||
570 | switch (fdwReason) { | ||||
571 | |||||
572 | case DLL_PROCESS_ATTACH: | ||||
573 | KA_TRACE(10, ("DllMain: PROCESS_ATTACH\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("DllMain: PROCESS_ATTACH\n" ); }; | ||||
574 | |||||
575 | return TRUE(!0); | ||||
576 | |||||
577 | case DLL_PROCESS_DETACH: | ||||
578 | KA_TRACE(10, ("DllMain: PROCESS_DETACH T#%d\n", __kmp_gtid_get_specific()))if (kmp_a_debug >= 10) { __kmp_debug_printf ("DllMain: PROCESS_DETACH T#%d\n" , __kmp_gtid_get_specific()); }; | ||||
579 | |||||
580 | // According to Windows* documentation for DllMain entry point: | ||||
581 | // for DLL_PROCESS_DETACH, lpReserved is used for telling the difference: | ||||
582 | // lpReserved == NULL when FreeLibrary() is called, | ||||
583 | // lpReserved != NULL when the process is terminated. | ||||
584 | // When FreeLibrary() is called, worker threads remain alive. So the | ||||
585 | // runtime's state is consistent and executing proper shutdown is OK. | ||||
586 | // When the process is terminated, worker threads have exited or been | ||||
587 | // forcefully terminated by the OS and only the shutdown thread remains. | ||||
588 | // This can leave the runtime in an inconsistent state. | ||||
589 | // Hence, only attempt proper cleanup when FreeLibrary() is called. | ||||
590 | // Otherwise, rely on OS to reclaim resources. | ||||
591 | if (lpReserved == NULL__null) | ||||
592 | __kmp_internal_end_library(__kmp_gtid_get_specific()); | ||||
593 | |||||
594 | return TRUE(!0); | ||||
595 | |||||
596 | case DLL_THREAD_ATTACH: | ||||
597 | KA_TRACE(10, ("DllMain: THREAD_ATTACH\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("DllMain: THREAD_ATTACH\n" ); }; | ||||
598 | |||||
599 | /* if we want to register new siblings all the time here call | ||||
600 | * __kmp_get_gtid(); */ | ||||
601 | return TRUE(!0); | ||||
602 | |||||
603 | case DLL_THREAD_DETACH: | ||||
604 | KA_TRACE(10, ("DllMain: THREAD_DETACH T#%d\n", __kmp_gtid_get_specific()))if (kmp_a_debug >= 10) { __kmp_debug_printf ("DllMain: THREAD_DETACH T#%d\n" , __kmp_gtid_get_specific()); }; | ||||
605 | |||||
606 | __kmp_internal_end_thread(__kmp_gtid_get_specific()); | ||||
607 | return TRUE(!0); | ||||
608 | } | ||||
609 | |||||
610 | return TRUE(!0); | ||||
611 | } | ||||
612 | |||||
613 | #endif /* KMP_OS_WINDOWS */ | ||||
614 | #endif /* KMP_DYNAMIC_LIB */ | ||||
615 | |||||
616 | /* __kmp_parallel_deo -- Wait until it's our turn. */ | ||||
617 | void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) { | ||||
618 | int gtid = *gtid_ref; | ||||
619 | #ifdef BUILD_PARALLEL_ORDERED1 | ||||
620 | kmp_team_t *team = __kmp_team_from_gtid(gtid); | ||||
621 | #endif /* BUILD_PARALLEL_ORDERED */ | ||||
622 | |||||
623 | if (__kmp_env_consistency_check) { | ||||
624 | if (__kmp_threads[gtid]->th.th_root->r.r_active) | ||||
625 | #if KMP_USE_DYNAMIC_LOCK1 | ||||
626 | __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL__null, 0); | ||||
627 | #else | ||||
628 | __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL__null); | ||||
629 | #endif | ||||
630 | } | ||||
631 | #ifdef BUILD_PARALLEL_ORDERED1 | ||||
632 | if (!team->t.t_serialized) { | ||||
633 | KMP_MB(); | ||||
634 | KMP_WAIT__kmp_wait_4(&team->t.t_ordered.dt.t_value, __kmp_tid_from_gtid(gtid), KMP_EQ__kmp_eq_4, | ||||
635 | NULL__null); | ||||
636 | KMP_MB(); | ||||
637 | } | ||||
638 | #endif /* BUILD_PARALLEL_ORDERED */ | ||||
639 | } | ||||
640 | |||||
641 | /* __kmp_parallel_dxo -- Signal the next task. */ | ||||
642 | void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) { | ||||
643 | int gtid = *gtid_ref; | ||||
644 | #ifdef BUILD_PARALLEL_ORDERED1 | ||||
645 | int tid = __kmp_tid_from_gtid(gtid); | ||||
646 | kmp_team_t *team = __kmp_team_from_gtid(gtid); | ||||
647 | #endif /* BUILD_PARALLEL_ORDERED */ | ||||
648 | |||||
649 | if (__kmp_env_consistency_check) { | ||||
650 | if (__kmp_threads[gtid]->th.th_root->r.r_active) | ||||
651 | __kmp_pop_sync(gtid, ct_ordered_in_parallel, loc_ref); | ||||
652 | } | ||||
653 | #ifdef BUILD_PARALLEL_ORDERED1 | ||||
654 | if (!team->t.t_serialized) { | ||||
655 | KMP_MB(); /* Flush all pending memory write invalidates. */ | ||||
656 | |||||
657 | /* use the tid of the next thread in this team */ | ||||
658 | /* TODO replace with general release procedure */ | ||||
659 | team->t.t_ordered.dt.t_value = ((tid + 1) % team->t.t_nproc); | ||||
660 | |||||
661 | KMP_MB(); /* Flush all pending memory write invalidates. */ | ||||
662 | } | ||||
663 | #endif /* BUILD_PARALLEL_ORDERED */ | ||||
664 | } | ||||
665 | |||||
666 | /* ------------------------------------------------------------------------ */ | ||||
667 | /* The BARRIER for a SINGLE process section is always explicit */ | ||||
668 | |||||
669 | int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws) { | ||||
670 | int status; | ||||
671 | kmp_info_t *th; | ||||
672 | kmp_team_t *team; | ||||
673 | |||||
674 | if (!TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) | ||||
675 | __kmp_parallel_initialize(); | ||||
676 | __kmp_resume_if_soft_paused(); | ||||
677 | |||||
678 | th = __kmp_threads[gtid]; | ||||
679 | team = th->th.th_team; | ||||
680 | status = 0; | ||||
681 | |||||
682 | th->th.th_ident = id_ref; | ||||
683 | |||||
684 | if (team->t.t_serialized) { | ||||
685 | status = 1; | ||||
686 | } else { | ||||
687 | kmp_int32 old_this = th->th.th_local.this_construct; | ||||
688 | |||||
689 | ++th->th.th_local.this_construct; | ||||
690 | /* try to set team count to thread count--success means thread got the | ||||
691 | single block */ | ||||
692 | /* TODO: Should this be acquire or release? */ | ||||
693 | if (team->t.t_construct == old_this) { | ||||
694 | status = __kmp_atomic_compare_store_acq(&team->t.t_construct, old_this, | ||||
695 | th->th.th_local.this_construct); | ||||
696 | } | ||||
697 | #if USE_ITT_BUILD1 | ||||
698 | if (__itt_metadata_add_ptr__kmp_itt_metadata_add_ptr__3_0 && __kmp_forkjoin_frames_mode == 3 && | ||||
699 | KMP_MASTER_GTID(gtid)(0 == __kmp_tid_from_gtid((gtid))) && th->th.th_teams_microtask == NULL__null && | ||||
700 | team->t.t_active_level == 1) { | ||||
701 | // Only report metadata by primary thread of active team at level 1 | ||||
702 | __kmp_itt_metadata_single(id_ref); | ||||
703 | } | ||||
704 | #endif /* USE_ITT_BUILD */ | ||||
705 | } | ||||
706 | |||||
707 | if (__kmp_env_consistency_check) { | ||||
708 | if (status && push_ws) { | ||||
709 | __kmp_push_workshare(gtid, ct_psingle, id_ref); | ||||
710 | } else { | ||||
711 | __kmp_check_workshare(gtid, ct_psingle, id_ref); | ||||
712 | } | ||||
713 | } | ||||
714 | #if USE_ITT_BUILD1 | ||||
715 | if (status) { | ||||
716 | __kmp_itt_single_start(gtid); | ||||
717 | } | ||||
718 | #endif /* USE_ITT_BUILD */ | ||||
719 | return status; | ||||
720 | } | ||||
721 | |||||
722 | void __kmp_exit_single(int gtid) { | ||||
723 | #if USE_ITT_BUILD1 | ||||
724 | __kmp_itt_single_end(gtid); | ||||
725 | #endif /* USE_ITT_BUILD */ | ||||
726 | if (__kmp_env_consistency_check) | ||||
727 | __kmp_pop_workshare(gtid, ct_psingle, NULL__null); | ||||
728 | } | ||||
729 | |||||
730 | /* determine if we can go parallel or must use a serialized parallel region and | ||||
731 | * how many threads we can use | ||||
732 | * set_nproc is the number of threads requested for the team | ||||
733 | * returns 0 if we should serialize or only use one thread, | ||||
734 | * otherwise the number of threads to use | ||||
735 | * The forkjoin lock is held by the caller. */ | ||||
736 | static int __kmp_reserve_threads(kmp_root_t *root, kmp_team_t *parent_team, | ||||
737 | int master_tid, int set_nthreads, | ||||
738 | int enter_teams) { | ||||
739 | int capacity; | ||||
740 | int new_nthreads; | ||||
741 | KMP_DEBUG_ASSERT(__kmp_init_serial)if (!(__kmp_init_serial)) { __kmp_debug_assert("__kmp_init_serial" , "openmp/runtime/src/kmp_runtime.cpp", 741); }; | ||||
742 | KMP_DEBUG_ASSERT(root && parent_team)if (!(root && parent_team)) { __kmp_debug_assert("root && parent_team" , "openmp/runtime/src/kmp_runtime.cpp", 742); }; | ||||
743 | kmp_info_t *this_thr = parent_team->t.t_threads[master_tid]; | ||||
744 | |||||
745 | // If dyn-var is set, dynamically adjust the number of desired threads, | ||||
746 | // according to the method specified by dynamic_mode. | ||||
747 | new_nthreads = set_nthreads; | ||||
748 | if (!get__dynamic_2(parent_team, master_tid)((parent_team)->t.t_threads[(master_tid)]->th.th_current_task ->td_icvs.dynamic)) { | ||||
749 | ; | ||||
750 | } | ||||
751 | #ifdef USE_LOAD_BALANCE1 | ||||
752 | else if (__kmp_global.g.g_dynamic_mode == dynamic_load_balance) { | ||||
753 | new_nthreads = __kmp_load_balance_nproc(root, set_nthreads); | ||||
754 | if (new_nthreads == 1) { | ||||
755 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d load balance reduced " "reservation to 1 thread\n", master_tid); } | ||||
756 | "reservation to 1 thread\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d load balance reduced " "reservation to 1 thread\n", master_tid); } | ||||
757 | master_tid))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d load balance reduced " "reservation to 1 thread\n", master_tid); }; | ||||
758 | return 1; | ||||
759 | } | ||||
760 | if (new_nthreads < set_nthreads) { | ||||
761 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d load balance reduced " "reservation to %d threads\n", master_tid, new_nthreads); } | ||||
762 | "reservation to %d threads\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d load balance reduced " "reservation to %d threads\n", master_tid, new_nthreads); } | ||||
763 | master_tid, new_nthreads))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d load balance reduced " "reservation to %d threads\n", master_tid, new_nthreads); }; | ||||
764 | } | ||||
765 | } | ||||
766 | #endif /* USE_LOAD_BALANCE */ | ||||
767 | else if (__kmp_global.g.g_dynamic_mode == dynamic_thread_limit) { | ||||
768 | new_nthreads = __kmp_avail_proc - __kmp_nth + | ||||
769 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc); | ||||
770 | if (new_nthreads <= 1) { | ||||
771 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d thread limit reduced " "reservation to 1 thread\n", master_tid); } | ||||
772 | "reservation to 1 thread\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d thread limit reduced " "reservation to 1 thread\n", master_tid); } | ||||
773 | master_tid))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d thread limit reduced " "reservation to 1 thread\n", master_tid); }; | ||||
774 | return 1; | ||||
775 | } | ||||
776 | if (new_nthreads < set_nthreads) { | ||||
777 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d thread limit reduced " "reservation to %d threads\n", master_tid, new_nthreads); } | ||||
778 | "reservation to %d threads\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d thread limit reduced " "reservation to %d threads\n", master_tid, new_nthreads); } | ||||
779 | master_tid, new_nthreads))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d thread limit reduced " "reservation to %d threads\n", master_tid, new_nthreads); }; | ||||
780 | } else { | ||||
781 | new_nthreads = set_nthreads; | ||||
782 | } | ||||
783 | } else if (__kmp_global.g.g_dynamic_mode == dynamic_random) { | ||||
784 | if (set_nthreads > 2) { | ||||
785 | new_nthreads = __kmp_get_random(parent_team->t.t_threads[master_tid]); | ||||
786 | new_nthreads = (new_nthreads % set_nthreads) + 1; | ||||
787 | if (new_nthreads == 1) { | ||||
788 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d dynamic random reduced " "reservation to 1 thread\n", master_tid); } | ||||
789 | "reservation to 1 thread\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d dynamic random reduced " "reservation to 1 thread\n", master_tid); } | ||||
790 | master_tid))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d dynamic random reduced " "reservation to 1 thread\n", master_tid); }; | ||||
791 | return 1; | ||||
792 | } | ||||
793 | if (new_nthreads < set_nthreads) { | ||||
794 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d dynamic random reduced " "reservation to %d threads\n", master_tid, new_nthreads); } | ||||
795 | "reservation to %d threads\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d dynamic random reduced " "reservation to %d threads\n", master_tid, new_nthreads); } | ||||
796 | master_tid, new_nthreads))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d dynamic random reduced " "reservation to %d threads\n", master_tid, new_nthreads); }; | ||||
797 | } | ||||
798 | } | ||||
799 | } else { | ||||
800 | KMP_ASSERT(0)if (!(0)) { __kmp_debug_assert("0", "openmp/runtime/src/kmp_runtime.cpp" , 800); }; | ||||
801 | } | ||||
802 | |||||
803 | // Respect KMP_ALL_THREADS/KMP_DEVICE_THREAD_LIMIT. | ||||
804 | if (__kmp_nth + new_nthreads - | ||||
805 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) > | ||||
806 | __kmp_max_nth) { | ||||
807 | int tl_nthreads = __kmp_max_nth - __kmp_nth + | ||||
808 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc); | ||||
809 | if (tl_nthreads <= 0) { | ||||
810 | tl_nthreads = 1; | ||||
811 | } | ||||
812 | |||||
813 | // If dyn-var is false, emit a 1-time warning. | ||||
814 | if (!get__dynamic_2(parent_team, master_tid)((parent_team)->t.t_threads[(master_tid)]->th.th_current_task ->td_icvs.dynamic) && (!__kmp_reserve_warn)) { | ||||
815 | __kmp_reserve_warn = 1; | ||||
816 | __kmp_msg(kmp_ms_warning, | ||||
817 | KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, set_nthreads, tl_nthreads ), | ||||
818 | KMP_HNT(Unset_ALL_THREADS)__kmp_msg_format(kmp_i18n_hnt_Unset_ALL_THREADS), __kmp_msg_null); | ||||
819 | } | ||||
820 | if (tl_nthreads == 1) { | ||||
821 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT " "reduced reservation to 1 thread\n", master_tid); } | ||||
822 | "reduced reservation to 1 thread\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT " "reduced reservation to 1 thread\n", master_tid); } | ||||
823 | master_tid))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT " "reduced reservation to 1 thread\n", master_tid); }; | ||||
824 | return 1; | ||||
825 | } | ||||
826 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT reduced "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT reduced " "reservation to %d threads\n", master_tid, tl_nthreads); } | ||||
827 | "reservation to %d threads\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT reduced " "reservation to %d threads\n", master_tid, tl_nthreads); } | ||||
828 | master_tid, tl_nthreads))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT reduced " "reservation to %d threads\n", master_tid, tl_nthreads); }; | ||||
829 | new_nthreads = tl_nthreads; | ||||
830 | } | ||||
831 | |||||
832 | // Respect OMP_THREAD_LIMIT | ||||
833 | int cg_nthreads = this_thr->th.th_cg_roots->cg_nthreads; | ||||
834 | int max_cg_threads = this_thr->th.th_cg_roots->cg_thread_limit; | ||||
835 | if (cg_nthreads + new_nthreads - | ||||
836 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) > | ||||
837 | max_cg_threads) { | ||||
838 | int tl_nthreads = max_cg_threads - cg_nthreads + | ||||
839 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc); | ||||
840 | if (tl_nthreads <= 0) { | ||||
841 | tl_nthreads = 1; | ||||
842 | } | ||||
843 | |||||
844 | // If dyn-var is false, emit a 1-time warning. | ||||
845 | if (!get__dynamic_2(parent_team, master_tid)((parent_team)->t.t_threads[(master_tid)]->th.th_current_task ->td_icvs.dynamic) && (!__kmp_reserve_warn)) { | ||||
846 | __kmp_reserve_warn = 1; | ||||
847 | __kmp_msg(kmp_ms_warning, | ||||
848 | KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, set_nthreads, tl_nthreads ), | ||||
849 | KMP_HNT(Unset_ALL_THREADS)__kmp_msg_format(kmp_i18n_hnt_Unset_ALL_THREADS), __kmp_msg_null); | ||||
850 | } | ||||
851 | if (tl_nthreads == 1) { | ||||
852 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT " "reduced reservation to 1 thread\n", master_tid); } | ||||
853 | "reduced reservation to 1 thread\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT " "reduced reservation to 1 thread\n", master_tid); } | ||||
854 | master_tid))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT " "reduced reservation to 1 thread\n", master_tid); }; | ||||
855 | return 1; | ||||
856 | } | ||||
857 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT reduced "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT reduced " "reservation to %d threads\n", master_tid, tl_nthreads); } | ||||
858 | "reservation to %d threads\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT reduced " "reservation to %d threads\n", master_tid, tl_nthreads); } | ||||
859 | master_tid, tl_nthreads))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT reduced " "reservation to %d threads\n", master_tid, tl_nthreads); }; | ||||
860 | new_nthreads = tl_nthreads; | ||||
861 | } | ||||
862 | |||||
863 | // Check if the threads array is large enough, or needs expanding. | ||||
864 | // See comment in __kmp_register_root() about the adjustment if | ||||
865 | // __kmp_threads[0] == NULL. | ||||
866 | capacity = __kmp_threads_capacity; | ||||
867 | if (TCR_PTR(__kmp_threads[0])((void *)(__kmp_threads[0])) == NULL__null) { | ||||
868 | --capacity; | ||||
869 | } | ||||
870 | // If it is not for initializing the hidden helper team, we need to take | ||||
871 | // __kmp_hidden_helper_threads_num out of the capacity because it is included | ||||
872 | // in __kmp_threads_capacity. | ||||
873 | if (__kmp_enable_hidden_helper && !TCR_4(__kmp_init_hidden_helper_threads)(__kmp_init_hidden_helper_threads)) { | ||||
874 | capacity -= __kmp_hidden_helper_threads_num; | ||||
875 | } | ||||
876 | if (__kmp_nth + new_nthreads - | ||||
877 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) > | ||||
878 | capacity) { | ||||
879 | // Expand the threads array. | ||||
880 | int slotsRequired = __kmp_nth + new_nthreads - | ||||
881 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) - | ||||
882 | capacity; | ||||
883 | int slotsAdded = __kmp_expand_threads(slotsRequired); | ||||
884 | if (slotsAdded < slotsRequired) { | ||||
885 | // The threads array was not expanded enough. | ||||
886 | new_nthreads -= (slotsRequired - slotsAdded); | ||||
887 | KMP_ASSERT(new_nthreads >= 1)if (!(new_nthreads >= 1)) { __kmp_debug_assert("new_nthreads >= 1" , "openmp/runtime/src/kmp_runtime.cpp", 887); }; | ||||
888 | |||||
889 | // If dyn-var is false, emit a 1-time warning. | ||||
890 | if (!get__dynamic_2(parent_team, master_tid)((parent_team)->t.t_threads[(master_tid)]->th.th_current_task ->td_icvs.dynamic) && (!__kmp_reserve_warn)) { | ||||
891 | __kmp_reserve_warn = 1; | ||||
892 | if (__kmp_tp_cached) { | ||||
893 | __kmp_msg(kmp_ms_warning, | ||||
894 | KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, set_nthreads, new_nthreads ), | ||||
895 | KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity)__kmp_msg_format(kmp_i18n_hnt_Set_ALL_THREADPRIVATE, __kmp_tp_capacity ), | ||||
896 | KMP_HNT(PossibleSystemLimitOnThreads)__kmp_msg_format(kmp_i18n_hnt_PossibleSystemLimitOnThreads), __kmp_msg_null); | ||||
897 | } else { | ||||
898 | __kmp_msg(kmp_ms_warning, | ||||
899 | KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, set_nthreads, new_nthreads ), | ||||
900 | KMP_HNT(SystemLimitOnThreads)__kmp_msg_format(kmp_i18n_hnt_SystemLimitOnThreads), __kmp_msg_null); | ||||
901 | } | ||||
902 | } | ||||
903 | } | ||||
904 | } | ||||
905 | |||||
906 | #ifdef KMP_DEBUG1 | ||||
907 | if (new_nthreads == 1) { | ||||
908 | KC_TRACE(10,if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d serializing team after reclaiming " "dead roots and rechecking; requested %d threads\n", __kmp_get_global_thread_id (), set_nthreads); } | ||||
909 | ("__kmp_reserve_threads: T#%d serializing team after reclaiming "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d serializing team after reclaiming " "dead roots and rechecking; requested %d threads\n", __kmp_get_global_thread_id (), set_nthreads); } | ||||
910 | "dead roots and rechecking; requested %d threads\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d serializing team after reclaiming " "dead roots and rechecking; requested %d threads\n", __kmp_get_global_thread_id (), set_nthreads); } | ||||
911 | __kmp_get_gtid(), set_nthreads))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d serializing team after reclaiming " "dead roots and rechecking; requested %d threads\n", __kmp_get_global_thread_id (), set_nthreads); }; | ||||
912 | } else { | ||||
913 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d allocating %d threads; requested"if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d allocating %d threads; requested" " %d threads\n", __kmp_get_global_thread_id(), new_nthreads, set_nthreads); } | ||||
914 | " %d threads\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d allocating %d threads; requested" " %d threads\n", __kmp_get_global_thread_id(), new_nthreads, set_nthreads); } | ||||
915 | __kmp_get_gtid(), new_nthreads, set_nthreads))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d allocating %d threads; requested" " %d threads\n", __kmp_get_global_thread_id(), new_nthreads, set_nthreads); }; | ||||
916 | } | ||||
917 | #endif // KMP_DEBUG | ||||
918 | return new_nthreads; | ||||
919 | } | ||||
920 | |||||
921 | /* Allocate threads from the thread pool and assign them to the new team. We are | ||||
922 | assured that there are enough threads available, because we checked on that | ||||
923 | earlier within critical section forkjoin */ | ||||
924 | static void __kmp_fork_team_threads(kmp_root_t *root, kmp_team_t *team, | ||||
925 | kmp_info_t *master_th, int master_gtid, | ||||
926 | int fork_teams_workers) { | ||||
927 | int i; | ||||
928 | int use_hot_team; | ||||
929 | |||||
930 | KA_TRACE(10, ("__kmp_fork_team_threads: new_nprocs = %d\n", team->t.t_nproc))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_fork_team_threads: new_nprocs = %d\n" , team->t.t_nproc); }; | ||||
931 | KMP_DEBUG_ASSERT(master_gtid == __kmp_get_gtid())if (!(master_gtid == __kmp_get_global_thread_id())) { __kmp_debug_assert ("master_gtid == __kmp_get_global_thread_id()", "openmp/runtime/src/kmp_runtime.cpp" , 931); }; | ||||
932 | KMP_MB(); | ||||
933 | |||||
934 | /* first, let's setup the primary thread */ | ||||
935 | master_th->th.th_info.ds.ds_tid = 0; | ||||
936 | master_th->th.th_team = team; | ||||
937 | master_th->th.th_team_nproc = team->t.t_nproc; | ||||
938 | master_th->th.th_team_master = master_th; | ||||
939 | master_th->th.th_team_serialized = FALSE0; | ||||
940 | master_th->th.th_dispatch = &team->t.t_dispatch[0]; | ||||
941 | |||||
942 | /* make sure we are not the optimized hot team */ | ||||
943 | #if KMP_NESTED_HOT_TEAMS1 | ||||
944 | use_hot_team = 0; | ||||
945 | kmp_hot_team_ptr_t *hot_teams = master_th->th.th_hot_teams; | ||||
946 | if (hot_teams) { // hot teams array is not allocated if | ||||
947 | // KMP_HOT_TEAMS_MAX_LEVEL=0 | ||||
948 | int level = team->t.t_active_level - 1; // index in array of hot teams | ||||
949 | if (master_th->th.th_teams_microtask) { // are we inside the teams? | ||||
950 | if (master_th->th.th_teams_size.nteams > 1) { | ||||
951 | ++level; // level was not increased in teams construct for | ||||
952 | // team_of_masters | ||||
953 | } | ||||
954 | if (team->t.t_pkfn != (microtask_t)__kmp_teams_master && | ||||
955 | master_th->th.th_teams_level == team->t.t_level) { | ||||
956 | ++level; // level was not increased in teams construct for | ||||
957 | // team_of_workers before the parallel | ||||
958 | } // team->t.t_level will be increased inside parallel | ||||
959 | } | ||||
960 | if (level < __kmp_hot_teams_max_level) { | ||||
961 | if (hot_teams[level].hot_team) { | ||||
962 | // hot team has already been allocated for given level | ||||
963 | KMP_DEBUG_ASSERT(hot_teams[level].hot_team == team)if (!(hot_teams[level].hot_team == team)) { __kmp_debug_assert ("hot_teams[level].hot_team == team", "openmp/runtime/src/kmp_runtime.cpp" , 963); }; | ||||
964 | use_hot_team = 1; // the team is ready to use | ||||
965 | } else { | ||||
966 | use_hot_team = 0; // AC: threads are not allocated yet | ||||
967 | hot_teams[level].hot_team = team; // remember new hot team | ||||
968 | hot_teams[level].hot_team_nth = team->t.t_nproc; | ||||
969 | } | ||||
970 | } else { | ||||
971 | use_hot_team = 0; | ||||
972 | } | ||||
973 | } | ||||
974 | #else | ||||
975 | use_hot_team = team == root->r.r_hot_team; | ||||
976 | #endif | ||||
977 | if (!use_hot_team) { | ||||
978 | |||||
979 | /* install the primary thread */ | ||||
980 | team->t.t_threads[0] = master_th; | ||||
981 | __kmp_initialize_info(master_th, team, 0, master_gtid); | ||||
982 | |||||
983 | /* now, install the worker threads */ | ||||
984 | for (i = 1; i < team->t.t_nproc; i++) { | ||||
985 | |||||
986 | /* fork or reallocate a new thread and install it in team */ | ||||
987 | kmp_info_t *thr = __kmp_allocate_thread(root, team, i); | ||||
988 | team->t.t_threads[i] = thr; | ||||
989 | KMP_DEBUG_ASSERT(thr)if (!(thr)) { __kmp_debug_assert("thr", "openmp/runtime/src/kmp_runtime.cpp" , 989); }; | ||||
990 | KMP_DEBUG_ASSERT(thr->th.th_team == team)if (!(thr->th.th_team == team)) { __kmp_debug_assert("thr->th.th_team == team" , "openmp/runtime/src/kmp_runtime.cpp", 990); }; | ||||
991 | /* align team and thread arrived states */ | ||||
992 | KA_TRACE(20, ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived " "T#%d(%d:%d) join =%llu, plain=%llu\n", __kmp_gtid_from_tid( 0, team), team->t.t_id, 0, __kmp_gtid_from_tid(i, team), team ->t.t_id, i, team->t.t_bar[bs_forkjoin_barrier].b_arrived , team->t.t_bar[bs_plain_barrier].b_arrived); } | ||||
993 | "T#%d(%d:%d) join =%llu, plain=%llu\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived " "T#%d(%d:%d) join =%llu, plain=%llu\n", __kmp_gtid_from_tid( 0, team), team->t.t_id, 0, __kmp_gtid_from_tid(i, team), team ->t.t_id, i, team->t.t_bar[bs_forkjoin_barrier].b_arrived , team->t.t_bar[bs_plain_barrier].b_arrived); } | ||||
994 | __kmp_gtid_from_tid(0, team), team->t.t_id, 0,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived " "T#%d(%d:%d) join =%llu, plain=%llu\n", __kmp_gtid_from_tid( 0, team), team->t.t_id, 0, __kmp_gtid_from_tid(i, team), team ->t.t_id, i, team->t.t_bar[bs_forkjoin_barrier].b_arrived , team->t.t_bar[bs_plain_barrier].b_arrived); } | ||||
995 | __kmp_gtid_from_tid(i, team), team->t.t_id, i,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived " "T#%d(%d:%d) join =%llu, plain=%llu\n", __kmp_gtid_from_tid( 0, team), team->t.t_id, 0, __kmp_gtid_from_tid(i, team), team ->t.t_id, i, team->t.t_bar[bs_forkjoin_barrier].b_arrived , team->t.t_bar[bs_plain_barrier].b_arrived); } | ||||
996 | team->t.t_bar[bs_forkjoin_barrier].b_arrived,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived " "T#%d(%d:%d) join =%llu, plain=%llu\n", __kmp_gtid_from_tid( 0, team), team->t.t_id, 0, __kmp_gtid_from_tid(i, team), team ->t.t_id, i, team->t.t_bar[bs_forkjoin_barrier].b_arrived , team->t.t_bar[bs_plain_barrier].b_arrived); } | ||||
997 | team->t.t_bar[bs_plain_barrier].b_arrived))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived " "T#%d(%d:%d) join =%llu, plain=%llu\n", __kmp_gtid_from_tid( 0, team), team->t.t_id, 0, __kmp_gtid_from_tid(i, team), team ->t.t_id, i, team->t.t_bar[bs_forkjoin_barrier].b_arrived , team->t.t_bar[bs_plain_barrier].b_arrived); }; | ||||
998 | thr->th.th_teams_microtask = master_th->th.th_teams_microtask; | ||||
999 | thr->th.th_teams_level = master_th->th.th_teams_level; | ||||
1000 | thr->th.th_teams_size = master_th->th.th_teams_size; | ||||
1001 | { // Initialize threads' barrier data. | ||||
1002 | int b; | ||||
1003 | kmp_balign_t *balign = team->t.t_threads[i]->th.th_bar; | ||||
1004 | for (b = 0; b < bs_last_barrier; ++b) { | ||||
1005 | balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; | ||||
1006 | KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG)if (!(balign[b].bb.wait_flag != 2)) { __kmp_debug_assert("balign[b].bb.wait_flag != 2" , "openmp/runtime/src/kmp_runtime.cpp", 1006); }; | ||||
1007 | #if USE_DEBUGGER0 | ||||
1008 | balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; | ||||
1009 | #endif | ||||
1010 | } | ||||
1011 | } | ||||
1012 | } | ||||
1013 | |||||
1014 | #if KMP_AFFINITY_SUPPORTED1 | ||||
1015 | // Do not partition the places list for teams construct workers who | ||||
1016 | // haven't actually been forked to do real work yet. This partitioning | ||||
1017 | // will take place in the parallel region nested within the teams construct. | ||||
1018 | if (!fork_teams_workers) { | ||||
1019 | __kmp_partition_places(team); | ||||
1020 | } | ||||
1021 | #endif | ||||
1022 | |||||
1023 | if (team->t.t_nproc > 1 && | ||||
1024 | __kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | ||||
1025 | team->t.b->update_num_threads(team->t.t_nproc); | ||||
1026 | __kmp_add_threads_to_team(team, team->t.t_nproc); | ||||
1027 | } | ||||
1028 | } | ||||
1029 | |||||
1030 | if (__kmp_display_affinity && team->t.t_display_affinity != 1) { | ||||
1031 | for (i = 0; i < team->t.t_nproc; i++) { | ||||
1032 | kmp_info_t *thr = team->t.t_threads[i]; | ||||
1033 | if (thr->th.th_prev_num_threads != team->t.t_nproc || | ||||
1034 | thr->th.th_prev_level != team->t.t_level) { | ||||
1035 | team->t.t_display_affinity = 1; | ||||
1036 | break; | ||||
1037 | } | ||||
1038 | } | ||||
1039 | } | ||||
1040 | |||||
1041 | KMP_MB(); | ||||
1042 | } | ||||
1043 | |||||
1044 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 | ||||
1045 | // Propagate any changes to the floating point control registers out to the team | ||||
1046 | // We try to avoid unnecessary writes to the relevant cache line in the team | ||||
1047 | // structure, so we don't make changes unless they are needed. | ||||
1048 | inline static void propagateFPControl(kmp_team_t *team) { | ||||
1049 | if (__kmp_inherit_fp_control) { | ||||
1050 | kmp_int16 x87_fpu_control_word; | ||||
1051 | kmp_uint32 mxcsr; | ||||
1052 | |||||
1053 | // Get primary thread's values of FPU control flags (both X87 and vector) | ||||
1054 | __kmp_store_x87_fpu_control_word(&x87_fpu_control_word); | ||||
1055 | __kmp_store_mxcsr(&mxcsr); | ||||
1056 | mxcsr &= KMP_X86_MXCSR_MASK0xffffffc0; | ||||
1057 | |||||
1058 | // There is no point looking at t_fp_control_saved here. | ||||
1059 | // If it is TRUE, we still have to update the values if they are different | ||||
1060 | // from those we now have. If it is FALSE we didn't save anything yet, but | ||||
1061 | // our objective is the same. We have to ensure that the values in the team | ||||
1062 | // are the same as those we have. | ||||
1063 | // So, this code achieves what we need whether or not t_fp_control_saved is | ||||
1064 | // true. By checking whether the value needs updating we avoid unnecessary | ||||
1065 | // writes that would put the cache-line into a written state, causing all | ||||
1066 | // threads in the team to have to read it again. | ||||
1067 | KMP_CHECK_UPDATE(team->t.t_x87_fpu_control_word, x87_fpu_control_word)if ((team->t.t_x87_fpu_control_word) != (x87_fpu_control_word )) (team->t.t_x87_fpu_control_word) = (x87_fpu_control_word ); | ||||
1068 | KMP_CHECK_UPDATE(team->t.t_mxcsr, mxcsr)if ((team->t.t_mxcsr) != (mxcsr)) (team->t.t_mxcsr) = ( mxcsr); | ||||
1069 | // Although we don't use this value, other code in the runtime wants to know | ||||
1070 | // whether it should restore them. So we must ensure it is correct. | ||||
1071 | KMP_CHECK_UPDATE(team->t.t_fp_control_saved, TRUE)if ((team->t.t_fp_control_saved) != ((!0))) (team->t.t_fp_control_saved ) = ((!0)); | ||||
1072 | } else { | ||||
1073 | // Similarly here. Don't write to this cache-line in the team structure | ||||
1074 | // unless we have to. | ||||
1075 | KMP_CHECK_UPDATE(team->t.t_fp_control_saved, FALSE)if ((team->t.t_fp_control_saved) != (0)) (team->t.t_fp_control_saved ) = (0); | ||||
1076 | } | ||||
1077 | } | ||||
1078 | |||||
1079 | // Do the opposite, setting the hardware registers to the updated values from | ||||
1080 | // the team. | ||||
1081 | inline static void updateHWFPControl(kmp_team_t *team) { | ||||
1082 | if (__kmp_inherit_fp_control && team->t.t_fp_control_saved) { | ||||
1083 | // Only reset the fp control regs if they have been changed in the team. | ||||
1084 | // the parallel region that we are exiting. | ||||
1085 | kmp_int16 x87_fpu_control_word; | ||||
1086 | kmp_uint32 mxcsr; | ||||
1087 | __kmp_store_x87_fpu_control_word(&x87_fpu_control_word); | ||||
1088 | __kmp_store_mxcsr(&mxcsr); | ||||
1089 | mxcsr &= KMP_X86_MXCSR_MASK0xffffffc0; | ||||
1090 | |||||
1091 | if (team->t.t_x87_fpu_control_word != x87_fpu_control_word) { | ||||
1092 | __kmp_clear_x87_fpu_status_word(); | ||||
1093 | __kmp_load_x87_fpu_control_word(&team->t.t_x87_fpu_control_word); | ||||
1094 | } | ||||
1095 | |||||
1096 | if (team->t.t_mxcsr != mxcsr) { | ||||
1097 | __kmp_load_mxcsr(&team->t.t_mxcsr); | ||||
1098 | } | ||||
1099 | } | ||||
1100 | } | ||||
1101 | #else | ||||
1102 | #define propagateFPControl(x) ((void)0) | ||||
1103 | #define updateHWFPControl(x) ((void)0) | ||||
1104 | #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ | ||||
1105 | |||||
1106 | static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team, | ||||
1107 | int realloc); // forward declaration | ||||
1108 | |||||
1109 | /* Run a parallel region that has been serialized, so runs only in a team of the | ||||
1110 | single primary thread. */ | ||||
1111 | void __kmp_serialized_parallel(ident_t *loc, kmp_int32 global_tid) { | ||||
1112 | kmp_info_t *this_thr; | ||||
1113 | kmp_team_t *serial_team; | ||||
1114 | |||||
1115 | KC_TRACE(10, ("__kmpc_serialized_parallel: called by T#%d\n", global_tid))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: called by T#%d\n" , global_tid); }; | ||||
1116 | |||||
1117 | /* Skip all this code for autopar serialized loops since it results in | ||||
1118 | unacceptable overhead */ | ||||
1119 | if (loc != NULL__null && (loc->flags & KMP_IDENT_AUTOPAR)) | ||||
1120 | return; | ||||
1121 | |||||
1122 | if (!TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) | ||||
1123 | __kmp_parallel_initialize(); | ||||
1124 | __kmp_resume_if_soft_paused(); | ||||
1125 | |||||
1126 | this_thr = __kmp_threads[global_tid]; | ||||
1127 | serial_team = this_thr->th.th_serial_team; | ||||
1128 | |||||
1129 | /* utilize the serialized team held by this thread */ | ||||
1130 | KMP_DEBUG_ASSERT(serial_team)if (!(serial_team)) { __kmp_debug_assert("serial_team", "openmp/runtime/src/kmp_runtime.cpp" , 1130); }; | ||||
1131 | KMP_MB(); | ||||
1132 | |||||
1133 | if (__kmp_tasking_mode != tskm_immediate_exec) { | ||||
1134 | KMP_DEBUG_ASSERT(if (!(this_thr->th.th_task_team == this_thr->th.th_team ->t.t_task_team[this_thr->th.th_task_state])) { __kmp_debug_assert ("this_thr->th.th_task_team == this_thr->th.th_team->t.t_task_team[this_thr->th.th_task_state]" , "openmp/runtime/src/kmp_runtime.cpp", 1136); } | ||||
1135 | this_thr->th.th_task_team ==if (!(this_thr->th.th_task_team == this_thr->th.th_team ->t.t_task_team[this_thr->th.th_task_state])) { __kmp_debug_assert ("this_thr->th.th_task_team == this_thr->th.th_team->t.t_task_team[this_thr->th.th_task_state]" , "openmp/runtime/src/kmp_runtime.cpp", 1136); } | ||||
1136 | this_thr->th.th_team->t.t_task_team[this_thr->th.th_task_state])if (!(this_thr->th.th_task_team == this_thr->th.th_team ->t.t_task_team[this_thr->th.th_task_state])) { __kmp_debug_assert ("this_thr->th.th_task_team == this_thr->th.th_team->t.t_task_team[this_thr->th.th_task_state]" , "openmp/runtime/src/kmp_runtime.cpp", 1136); }; | ||||
1137 | KMP_DEBUG_ASSERT(serial_team->t.t_task_team[this_thr->th.th_task_state] ==if (!(serial_team->t.t_task_team[this_thr->th.th_task_state ] == __null)) { __kmp_debug_assert("serial_team->t.t_task_team[this_thr->th.th_task_state] == __null" , "openmp/runtime/src/kmp_runtime.cpp", 1138); } | ||||
1138 | NULL)if (!(serial_team->t.t_task_team[this_thr->th.th_task_state ] == __null)) { __kmp_debug_assert("serial_team->t.t_task_team[this_thr->th.th_task_state] == __null" , "openmp/runtime/src/kmp_runtime.cpp", 1138); }; | ||||
1139 | KA_TRACE(20, ("__kmpc_serialized_parallel: T#%d pushing task_team %p / "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d pushing task_team %p / " "team %p, new task_team = NULL\n", global_tid, this_thr-> th.th_task_team, this_thr->th.th_team); } | ||||
1140 | "team %p, new task_team = NULL\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d pushing task_team %p / " "team %p, new task_team = NULL\n", global_tid, this_thr-> th.th_task_team, this_thr->th.th_team); } | ||||
1141 | global_tid, this_thr->th.th_task_team, this_thr->th.th_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d pushing task_team %p / " "team %p, new task_team = NULL\n", global_tid, this_thr-> th.th_task_team, this_thr->th.th_team); }; | ||||
1142 | this_thr->th.th_task_team = NULL__null; | ||||
1143 | } | ||||
1144 | |||||
1145 | kmp_proc_bind_t proc_bind = this_thr->th.th_set_proc_bind; | ||||
1146 | if (this_thr->th.th_current_task->td_icvs.proc_bind == proc_bind_false) { | ||||
1147 | proc_bind = proc_bind_false; | ||||
1148 | } else if (proc_bind == proc_bind_default) { | ||||
1149 | // No proc_bind clause was specified, so use the current value | ||||
1150 | // of proc-bind-var for this parallel region. | ||||
1151 | proc_bind = this_thr->th.th_current_task->td_icvs.proc_bind; | ||||
1152 | } | ||||
1153 | // Reset for next parallel region | ||||
1154 | this_thr->th.th_set_proc_bind = proc_bind_default; | ||||
1155 | |||||
1156 | #if OMPT_SUPPORT1 | ||||
1157 | ompt_data_t ompt_parallel_data = ompt_data_none{0}; | ||||
1158 | void *codeptr = OMPT_LOAD_RETURN_ADDRESS(global_tid)__ompt_load_return_address(global_tid); | ||||
1159 | if (ompt_enabled.enabled && | ||||
1160 | this_thr->th.ompt_thread_info.state != ompt_state_overhead) { | ||||
1161 | |||||
1162 | ompt_task_info_t *parent_task_info; | ||||
1163 | parent_task_info = OMPT_CUR_TASK_INFO(this_thr)(&(this_thr->th.th_current_task->ompt_task_info)); | ||||
1164 | |||||
1165 | parent_task_info->frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0)__builtin_frame_address(0); | ||||
1166 | if (ompt_enabled.ompt_callback_parallel_begin) { | ||||
1167 | int team_size = 1; | ||||
1168 | |||||
1169 | ompt_callbacks.ompt_callback(ompt_callback_parallel_begin)ompt_callback_parallel_begin_callback( | ||||
1170 | &(parent_task_info->task_data), &(parent_task_info->frame), | ||||
1171 | &ompt_parallel_data, team_size, | ||||
1172 | ompt_parallel_invoker_program | ompt_parallel_team, codeptr); | ||||
1173 | } | ||||
1174 | } | ||||
1175 | #endif // OMPT_SUPPORT | ||||
1176 | |||||
1177 | if (this_thr->th.th_team != serial_team) { | ||||
1178 | // Nested level will be an index in the nested nthreads array | ||||
1179 | int level = this_thr->th.th_team->t.t_level; | ||||
1180 | |||||
1181 | if (serial_team->t.t_serialized) { | ||||
1182 | /* this serial team was already used | ||||
1183 | TODO increase performance by making this locks more specific */ | ||||
1184 | kmp_team_t *new_team; | ||||
1185 | |||||
1186 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); | ||||
1187 | |||||
1188 | new_team = | ||||
1189 | __kmp_allocate_team(this_thr->th.th_root, 1, 1, | ||||
1190 | #if OMPT_SUPPORT1 | ||||
1191 | ompt_parallel_data, | ||||
1192 | #endif | ||||
1193 | proc_bind, &this_thr->th.th_current_task->td_icvs, | ||||
1194 | 0 USE_NESTED_HOT_ARG(NULL), __null); | ||||
1195 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | ||||
1196 | KMP_ASSERT(new_team)if (!(new_team)) { __kmp_debug_assert("new_team", "openmp/runtime/src/kmp_runtime.cpp" , 1196); }; | ||||
1197 | |||||
1198 | /* setup new serialized team and install it */ | ||||
1199 | new_team->t.t_threads[0] = this_thr; | ||||
1200 | new_team->t.t_parent = this_thr->th.th_team; | ||||
1201 | serial_team = new_team; | ||||
1202 | this_thr->th.th_serial_team = serial_team; | ||||
1203 | |||||
1204 | KF_TRACE(if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d allocated new serial team %p\n" , global_tid, serial_team); } | ||||
1205 | 10,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d allocated new serial team %p\n" , global_tid, serial_team); } | ||||
1206 | ("__kmpc_serialized_parallel: T#%d allocated new serial team %p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d allocated new serial team %p\n" , global_tid, serial_team); } | ||||
1207 | global_tid, serial_team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d allocated new serial team %p\n" , global_tid, serial_team); }; | ||||
1208 | |||||
1209 | /* TODO the above breaks the requirement that if we run out of resources, | ||||
1210 | then we can still guarantee that serialized teams are ok, since we may | ||||
1211 | need to allocate a new one */ | ||||
1212 | } else { | ||||
1213 | KF_TRACE(if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d reusing cached serial team %p\n" , global_tid, serial_team); } | ||||
1214 | 10,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d reusing cached serial team %p\n" , global_tid, serial_team); } | ||||
1215 | ("__kmpc_serialized_parallel: T#%d reusing cached serial team %p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d reusing cached serial team %p\n" , global_tid, serial_team); } | ||||
1216 | global_tid, serial_team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d reusing cached serial team %p\n" , global_tid, serial_team); }; | ||||
1217 | } | ||||
1218 | |||||
1219 | /* we have to initialize this serial team */ | ||||
1220 | KMP_DEBUG_ASSERT(serial_team->t.t_threads)if (!(serial_team->t.t_threads)) { __kmp_debug_assert("serial_team->t.t_threads" , "openmp/runtime/src/kmp_runtime.cpp", 1220); }; | ||||
1221 | KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr)if (!(serial_team->t.t_threads[0] == this_thr)) { __kmp_debug_assert ("serial_team->t.t_threads[0] == this_thr", "openmp/runtime/src/kmp_runtime.cpp" , 1221); }; | ||||
1222 | KMP_DEBUG_ASSERT(this_thr->th.th_team != serial_team)if (!(this_thr->th.th_team != serial_team)) { __kmp_debug_assert ("this_thr->th.th_team != serial_team", "openmp/runtime/src/kmp_runtime.cpp" , 1222); }; | ||||
1223 | serial_team->t.t_ident = loc; | ||||
1224 | serial_team->t.t_serialized = 1; | ||||
1225 | serial_team->t.t_nproc = 1; | ||||
1226 | serial_team->t.t_parent = this_thr->th.th_team; | ||||
1227 | serial_team->t.t_sched.sched = this_thr->th.th_team->t.t_sched.sched; | ||||
1228 | this_thr->th.th_team = serial_team; | ||||
1229 | serial_team->t.t_master_tid = this_thr->th.th_info.ds.ds_tid; | ||||
1230 | |||||
1231 | KF_TRACE(10, ("__kmpc_serialized_parallel: T#%d curtask=%p\n", global_tid,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d curtask=%p\n" , global_tid, this_thr->th.th_current_task); } | ||||
1232 | this_thr->th.th_current_task))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d curtask=%p\n" , global_tid, this_thr->th.th_current_task); }; | ||||
1233 | KMP_ASSERT(this_thr->th.th_current_task->td_flags.executing == 1)if (!(this_thr->th.th_current_task->td_flags.executing == 1)) { __kmp_debug_assert("this_thr->th.th_current_task->td_flags.executing == 1" , "openmp/runtime/src/kmp_runtime.cpp", 1233); }; | ||||
1234 | this_thr->th.th_current_task->td_flags.executing = 0; | ||||
1235 | |||||
1236 | __kmp_push_current_task_to_thread(this_thr, serial_team, 0); | ||||
1237 | |||||
1238 | /* TODO: GEH: do ICVs work for nested serialized teams? Don't we need an | ||||
1239 | implicit task for each serialized task represented by | ||||
1240 | team->t.t_serialized? */ | ||||
1241 | copy_icvs(&this_thr->th.th_current_task->td_icvs, | ||||
1242 | &this_thr->th.th_current_task->td_parent->td_icvs); | ||||
1243 | |||||
1244 | // Thread value exists in the nested nthreads array for the next nested | ||||
1245 | // level | ||||
1246 | if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) { | ||||
1247 | this_thr->th.th_current_task->td_icvs.nproc = | ||||
1248 | __kmp_nested_nth.nth[level + 1]; | ||||
1249 | } | ||||
1250 | |||||
1251 | if (__kmp_nested_proc_bind.used && | ||||
1252 | (level + 1 < __kmp_nested_proc_bind.used)) { | ||||
1253 | this_thr->th.th_current_task->td_icvs.proc_bind = | ||||
1254 | __kmp_nested_proc_bind.bind_types[level + 1]; | ||||
1255 | } | ||||
1256 | |||||
1257 | #if USE_DEBUGGER0 | ||||
1258 | serial_team->t.t_pkfn = (microtask_t)(~0); // For the debugger. | ||||
1259 | #endif | ||||
1260 | this_thr->th.th_info.ds.ds_tid = 0; | ||||
1261 | |||||
1262 | /* set thread cache values */ | ||||
1263 | this_thr->th.th_team_nproc = 1; | ||||
1264 | this_thr->th.th_team_master = this_thr; | ||||
1265 | this_thr->th.th_team_serialized = 1; | ||||
1266 | |||||
1267 | serial_team->t.t_level = serial_team->t.t_parent->t.t_level + 1; | ||||
1268 | serial_team->t.t_active_level = serial_team->t.t_parent->t.t_active_level; | ||||
1269 | serial_team->t.t_def_allocator = this_thr->th.th_def_allocator; // save | ||||
1270 | |||||
1271 | propagateFPControl(serial_team); | ||||
1272 | |||||
1273 | /* check if we need to allocate dispatch buffers stack */ | ||||
1274 | KMP_DEBUG_ASSERT(serial_team->t.t_dispatch)if (!(serial_team->t.t_dispatch)) { __kmp_debug_assert("serial_team->t.t_dispatch" , "openmp/runtime/src/kmp_runtime.cpp", 1274); }; | ||||
1275 | if (!serial_team->t.t_dispatch->th_disp_buffer) { | ||||
1276 | serial_team->t.t_dispatch->th_disp_buffer = | ||||
1277 | (dispatch_private_info_t *)__kmp_allocate(___kmp_allocate((sizeof(dispatch_private_info_t)), "openmp/runtime/src/kmp_runtime.cpp" , 1278) | ||||
1278 | sizeof(dispatch_private_info_t))___kmp_allocate((sizeof(dispatch_private_info_t)), "openmp/runtime/src/kmp_runtime.cpp" , 1278); | ||||
1279 | } | ||||
1280 | this_thr->th.th_dispatch = serial_team->t.t_dispatch; | ||||
1281 | |||||
1282 | KMP_MB(); | ||||
1283 | |||||
1284 | } else { | ||||
1285 | /* this serialized team is already being used, | ||||
1286 | * that's fine, just add another nested level */ | ||||
1287 | KMP_DEBUG_ASSERT(this_thr->th.th_team == serial_team)if (!(this_thr->th.th_team == serial_team)) { __kmp_debug_assert ("this_thr->th.th_team == serial_team", "openmp/runtime/src/kmp_runtime.cpp" , 1287); }; | ||||
1288 | KMP_DEBUG_ASSERT(serial_team->t.t_threads)if (!(serial_team->t.t_threads)) { __kmp_debug_assert("serial_team->t.t_threads" , "openmp/runtime/src/kmp_runtime.cpp", 1288); }; | ||||
1289 | KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr)if (!(serial_team->t.t_threads[0] == this_thr)) { __kmp_debug_assert ("serial_team->t.t_threads[0] == this_thr", "openmp/runtime/src/kmp_runtime.cpp" , 1289); }; | ||||
1290 | ++serial_team->t.t_serialized; | ||||
1291 | this_thr->th.th_team_serialized = serial_team->t.t_serialized; | ||||
1292 | |||||
1293 | // Nested level will be an index in the nested nthreads array | ||||
1294 | int level = this_thr->th.th_team->t.t_level; | ||||
1295 | // Thread value exists in the nested nthreads array for the next nested | ||||
1296 | // level | ||||
1297 | if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) { | ||||
1298 | this_thr->th.th_current_task->td_icvs.nproc = | ||||
1299 | __kmp_nested_nth.nth[level + 1]; | ||||
1300 | } | ||||
1301 | serial_team->t.t_level++; | ||||
1302 | KF_TRACE(10, ("__kmpc_serialized_parallel: T#%d increasing nesting level "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d increasing nesting level " "of serial team %p to %d\n", global_tid, serial_team, serial_team ->t.t_level); } | ||||
1303 | "of serial team %p to %d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d increasing nesting level " "of serial team %p to %d\n", global_tid, serial_team, serial_team ->t.t_level); } | ||||
1304 | global_tid, serial_team, serial_team->t.t_level))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d increasing nesting level " "of serial team %p to %d\n", global_tid, serial_team, serial_team ->t.t_level); }; | ||||
1305 | |||||
1306 | /* allocate/push dispatch buffers stack */ | ||||
1307 | KMP_DEBUG_ASSERT(serial_team->t.t_dispatch)if (!(serial_team->t.t_dispatch)) { __kmp_debug_assert("serial_team->t.t_dispatch" , "openmp/runtime/src/kmp_runtime.cpp", 1307); }; | ||||
1308 | { | ||||
1309 | dispatch_private_info_t *disp_buffer = | ||||
1310 | (dispatch_private_info_t *)__kmp_allocate(___kmp_allocate((sizeof(dispatch_private_info_t)), "openmp/runtime/src/kmp_runtime.cpp" , 1311) | ||||
1311 | sizeof(dispatch_private_info_t))___kmp_allocate((sizeof(dispatch_private_info_t)), "openmp/runtime/src/kmp_runtime.cpp" , 1311); | ||||
1312 | disp_buffer->next = serial_team->t.t_dispatch->th_disp_buffer; | ||||
1313 | serial_team->t.t_dispatch->th_disp_buffer = disp_buffer; | ||||
1314 | } | ||||
1315 | this_thr->th.th_dispatch = serial_team->t.t_dispatch; | ||||
1316 | |||||
1317 | KMP_MB(); | ||||
1318 | } | ||||
1319 | KMP_CHECK_UPDATE(serial_team->t.t_cancel_request, cancel_noreq)if ((serial_team->t.t_cancel_request) != (cancel_noreq)) ( serial_team->t.t_cancel_request) = (cancel_noreq); | ||||
1320 | |||||
1321 | // Perform the display affinity functionality for | ||||
1322 | // serialized parallel regions | ||||
1323 | if (__kmp_display_affinity) { | ||||
1324 | if (this_thr->th.th_prev_level != serial_team->t.t_level || | ||||
1325 | this_thr->th.th_prev_num_threads != 1) { | ||||
1326 | // NULL means use the affinity-format-var ICV | ||||
1327 | __kmp_aux_display_affinity(global_tid, NULL__null); | ||||
1328 | this_thr->th.th_prev_level = serial_team->t.t_level; | ||||
1329 | this_thr->th.th_prev_num_threads = 1; | ||||
1330 | } | ||||
1331 | } | ||||
1332 | |||||
1333 | if (__kmp_env_consistency_check) | ||||
1334 | __kmp_push_parallel(global_tid, NULL__null); | ||||
1335 | #if OMPT_SUPPORT1 | ||||
1336 | serial_team->t.ompt_team_info.master_return_address = codeptr; | ||||
1337 | if (ompt_enabled.enabled && | ||||
1338 | this_thr->th.ompt_thread_info.state != ompt_state_overhead) { | ||||
1339 | OMPT_CUR_TASK_INFO(this_thr)(&(this_thr->th.th_current_task->ompt_task_info))->frame.exit_frame.ptr = | ||||
1340 | OMPT_GET_FRAME_ADDRESS(0)__builtin_frame_address(0); | ||||
1341 | |||||
1342 | ompt_lw_taskteam_t lw_taskteam; | ||||
1343 | __ompt_lw_taskteam_init(&lw_taskteam, this_thr, global_tid, | ||||
1344 | &ompt_parallel_data, codeptr); | ||||
1345 | |||||
1346 | __ompt_lw_taskteam_link(&lw_taskteam, this_thr, 1); | ||||
1347 | // don't use lw_taskteam after linking. content was swaped | ||||
1348 | |||||
1349 | /* OMPT implicit task begin */ | ||||
1350 | if (ompt_enabled.ompt_callback_implicit_task) { | ||||
1351 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | ||||
1352 | ompt_scope_begin, OMPT_CUR_TEAM_DATA(this_thr)(&(this_thr->th.th_team->t.ompt_team_info.parallel_data )), | ||||
1353 | OMPT_CUR_TASK_DATA(this_thr)(&(this_thr->th.th_current_task->ompt_task_info.task_data )), 1, __kmp_tid_from_gtid(global_tid), | ||||
1354 | ompt_task_implicit); // TODO: Can this be ompt_task_initial? | ||||
1355 | OMPT_CUR_TASK_INFO(this_thr)(&(this_thr->th.th_current_task->ompt_task_info))->thread_num = | ||||
1356 | __kmp_tid_from_gtid(global_tid); | ||||
1357 | } | ||||
1358 | |||||
1359 | /* OMPT state */ | ||||
1360 | this_thr->th.ompt_thread_info.state = ompt_state_work_parallel; | ||||
1361 | OMPT_CUR_TASK_INFO(this_thr)(&(this_thr->th.th_current_task->ompt_task_info))->frame.exit_frame.ptr = | ||||
1362 | OMPT_GET_FRAME_ADDRESS(0)__builtin_frame_address(0); | ||||
1363 | } | ||||
1364 | #endif | ||||
1365 | } | ||||
1366 | |||||
1367 | // Test if this fork is for a team closely nested in a teams construct | ||||
1368 | static inline bool __kmp_is_fork_in_teams(kmp_info_t *master_th, | ||||
1369 | microtask_t microtask, int level, | ||||
1370 | int teams_level, kmp_va_list ap) { | ||||
1371 | return (master_th->th.th_teams_microtask && ap && | ||||
1372 | microtask != (microtask_t)__kmp_teams_master && level == teams_level); | ||||
1373 | } | ||||
1374 | |||||
1375 | // Test if this fork is for the teams construct, i.e. to form the outer league | ||||
1376 | // of teams | ||||
1377 | static inline bool __kmp_is_entering_teams(int active_level, int level, | ||||
1378 | int teams_level, kmp_va_list ap) { | ||||
1379 | return ((ap == NULL__null && active_level == 0) || | ||||
1380 | (ap && teams_level > 0 && teams_level == level)); | ||||
1381 | } | ||||
1382 | |||||
1383 | // AC: This is start of parallel that is nested inside teams construct. | ||||
1384 | // The team is actual (hot), all workers are ready at the fork barrier. | ||||
1385 | // No lock needed to initialize the team a bit, then free workers. | ||||
1386 | static inline int | ||||
1387 | __kmp_fork_in_teams(ident_t *loc, int gtid, kmp_team_t *parent_team, | ||||
1388 | kmp_int32 argc, kmp_info_t *master_th, kmp_root_t *root, | ||||
1389 | enum fork_context_e call_context, microtask_t microtask, | ||||
1390 | launch_t invoker, int master_set_numthreads, int level, | ||||
1391 | #if OMPT_SUPPORT1 | ||||
1392 | ompt_data_t ompt_parallel_data, void *return_address, | ||||
1393 | #endif | ||||
1394 | kmp_va_list ap) { | ||||
1395 | void **argv; | ||||
1396 | int i; | ||||
1397 | |||||
1398 | parent_team->t.t_ident = loc; | ||||
1399 | __kmp_alloc_argv_entries(argc, parent_team, TRUE(!0)); | ||||
1400 | parent_team->t.t_argc = argc; | ||||
1401 | argv = (void **)parent_team->t.t_argv; | ||||
1402 | for (i = argc - 1; i >= 0; --i) { | ||||
1403 | *argv++ = va_arg(kmp_va_deref(ap), void *)__builtin_va_arg((*(ap)), void *); | ||||
1404 | } | ||||
1405 | // Increment our nested depth levels, but not increase the serialization | ||||
1406 | if (parent_team == master_th->th.th_serial_team) { | ||||
1407 | // AC: we are in serialized parallel | ||||
1408 | __kmpc_serialized_parallel(loc, gtid); | ||||
1409 | KMP_DEBUG_ASSERT(parent_team->t.t_serialized > 1)if (!(parent_team->t.t_serialized > 1)) { __kmp_debug_assert ("parent_team->t.t_serialized > 1", "openmp/runtime/src/kmp_runtime.cpp" , 1409); }; | ||||
1410 | |||||
1411 | if (call_context == fork_context_gnu) { | ||||
1412 | // AC: need to decrement t_serialized for enquiry functions to work | ||||
1413 | // correctly, will restore at join time | ||||
1414 | parent_team->t.t_serialized--; | ||||
1415 | return TRUE(!0); | ||||
1416 | } | ||||
1417 | |||||
1418 | #if OMPD_SUPPORT1 | ||||
1419 | parent_team->t.t_pkfn = microtask; | ||||
1420 | #endif | ||||
1421 | |||||
1422 | #if OMPT_SUPPORT1 | ||||
1423 | void *dummy; | ||||
1424 | void **exit_frame_p; | ||||
1425 | ompt_data_t *implicit_task_data; | ||||
1426 | ompt_lw_taskteam_t lw_taskteam; | ||||
1427 | |||||
1428 | if (ompt_enabled.enabled) { | ||||
1429 | __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid, | ||||
1430 | &ompt_parallel_data, return_address); | ||||
1431 | exit_frame_p = &(lw_taskteam.ompt_task_info.frame.exit_frame.ptr); | ||||
1432 | |||||
1433 | __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0); | ||||
1434 | // Don't use lw_taskteam after linking. Content was swapped. | ||||
1435 | |||||
1436 | /* OMPT implicit task begin */ | ||||
1437 | implicit_task_data = OMPT_CUR_TASK_DATA(master_th)(&(master_th->th.th_current_task->ompt_task_info.task_data )); | ||||
1438 | if (ompt_enabled.ompt_callback_implicit_task) { | ||||
1439 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num = __kmp_tid_from_gtid(gtid); | ||||
1440 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | ||||
1441 | ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th)(&(master_th->th.th_team->t.ompt_team_info.parallel_data )), implicit_task_data, | ||||
1442 | 1, OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num, ompt_task_implicit); | ||||
1443 | } | ||||
1444 | |||||
1445 | /* OMPT state */ | ||||
1446 | master_th->th.ompt_thread_info.state = ompt_state_work_parallel; | ||||
1447 | } else { | ||||
1448 | exit_frame_p = &dummy; | ||||
1449 | } | ||||
1450 | #endif | ||||
1451 | |||||
1452 | // AC: need to decrement t_serialized for enquiry functions to work | ||||
1453 | // correctly, will restore at join time | ||||
1454 | parent_team->t.t_serialized--; | ||||
1455 | |||||
1456 | { | ||||
1457 | KMP_TIME_PARTITIONED_BLOCK(OMP_parallel)((void)0); | ||||
1458 | KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK)((void)0); | ||||
1459 | __kmp_invoke_microtask(microtask, gtid, 0, argc, parent_team->t.t_argv | ||||
1460 | #if OMPT_SUPPORT1 | ||||
1461 | , | ||||
1462 | exit_frame_p | ||||
1463 | #endif | ||||
1464 | ); | ||||
1465 | } | ||||
1466 | |||||
1467 | #if OMPT_SUPPORT1 | ||||
1468 | if (ompt_enabled.enabled) { | ||||
1469 | *exit_frame_p = NULL__null; | ||||
1470 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->frame.exit_frame = ompt_data_none{0}; | ||||
1471 | if (ompt_enabled.ompt_callback_implicit_task) { | ||||
1472 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | ||||
1473 | ompt_scope_end, NULL__null, implicit_task_data, 1, | ||||
1474 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num, ompt_task_implicit); | ||||
1475 | } | ||||
1476 | ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th)(&(master_th->th.th_team->t.ompt_team_info.parallel_data )); | ||||
1477 | __ompt_lw_taskteam_unlink(master_th); | ||||
1478 | if (ompt_enabled.ompt_callback_parallel_end) { | ||||
1479 | ompt_callbacks.ompt_callback(ompt_callback_parallel_end)ompt_callback_parallel_end_callback( | ||||
1480 | &ompt_parallel_data, OMPT_CUR_TASK_DATA(master_th)(&(master_th->th.th_current_task->ompt_task_info.task_data )), | ||||
1481 | OMPT_INVOKER(call_context)((call_context == fork_context_gnu) ? ompt_parallel_invoker_program : ompt_parallel_invoker_runtime) | ompt_parallel_team, return_address); | ||||
1482 | } | ||||
1483 | master_th->th.ompt_thread_info.state = ompt_state_overhead; | ||||
1484 | } | ||||
1485 | #endif | ||||
1486 | return TRUE(!0); | ||||
1487 | } | ||||
1488 | |||||
1489 | parent_team->t.t_pkfn = microtask; | ||||
1490 | parent_team->t.t_invoke = invoker; | ||||
1491 | KMP_ATOMIC_INC(&root->r.r_in_parallel)(&root->r.r_in_parallel)->fetch_add(1, std::memory_order_acq_rel ); | ||||
1492 | parent_team->t.t_active_level++; | ||||
1493 | parent_team->t.t_level++; | ||||
1494 | parent_team->t.t_def_allocator = master_th->th.th_def_allocator; // save | ||||
1495 | |||||
1496 | // If the threads allocated to the team are less than the thread limit, update | ||||
1497 | // the thread limit here. th_teams_size.nth is specific to this team nested | ||||
1498 | // in a teams construct, the team is fully created, and we're about to do | ||||
1499 | // the actual fork. Best to do this here so that the subsequent uses below | ||||
1500 | // and in the join have the correct value. | ||||
1501 | master_th->th.th_teams_size.nth = parent_team->t.t_nproc; | ||||
1502 | |||||
1503 | #if OMPT_SUPPORT1 | ||||
1504 | if (ompt_enabled.enabled) { | ||||
1505 | ompt_lw_taskteam_t lw_taskteam; | ||||
1506 | __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid, &ompt_parallel_data, | ||||
1507 | return_address); | ||||
1508 | __ompt_lw_taskteam_link(&lw_taskteam, master_th, 1, true); | ||||
1509 | } | ||||
1510 | #endif | ||||
1511 | |||||
1512 | /* Change number of threads in the team if requested */ | ||||
1513 | if (master_set_numthreads) { // The parallel has num_threads clause | ||||
1514 | if (master_set_numthreads <= master_th->th.th_teams_size.nth) { | ||||
1515 | // AC: only can reduce number of threads dynamically, can't increase | ||||
1516 | kmp_info_t **other_threads = parent_team->t.t_threads; | ||||
1517 | // NOTE: if using distributed barrier, we need to run this code block | ||||
1518 | // even when the team size appears not to have changed from the max. | ||||
1519 | int old_proc = master_th->th.th_teams_size.nth; | ||||
1520 | if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | ||||
1521 | __kmp_resize_dist_barrier(parent_team, old_proc, master_set_numthreads); | ||||
1522 | __kmp_add_threads_to_team(parent_team, master_set_numthreads); | ||||
1523 | } | ||||
1524 | parent_team->t.t_nproc = master_set_numthreads; | ||||
1525 | for (i = 0; i < master_set_numthreads; ++i) { | ||||
1526 | other_threads[i]->th.th_team_nproc = master_set_numthreads; | ||||
1527 | } | ||||
1528 | } | ||||
1529 | // Keep extra threads hot in the team for possible next parallels | ||||
1530 | master_th->th.th_set_nproc = 0; | ||||
1531 | } | ||||
1532 | |||||
1533 | #if USE_DEBUGGER0 | ||||
1534 | if (__kmp_debugging) { // Let debugger override number of threads. | ||||
1535 | int nth = __kmp_omp_num_threads(loc); | ||||
1536 | if (nth > 0) { // 0 means debugger doesn't want to change num threads | ||||
1537 | master_set_numthreads = nth; | ||||
1538 | } | ||||
1539 | } | ||||
1540 | #endif | ||||
1541 | |||||
1542 | // Figure out the proc_bind policy for the nested parallel within teams | ||||
1543 | kmp_proc_bind_t proc_bind = master_th->th.th_set_proc_bind; | ||||
1544 | // proc_bind_default means don't update | ||||
1545 | kmp_proc_bind_t proc_bind_icv = proc_bind_default; | ||||
1546 | if (master_th->th.th_current_task->td_icvs.proc_bind == proc_bind_false) { | ||||
1547 | proc_bind = proc_bind_false; | ||||
1548 | } else { | ||||
1549 | // No proc_bind clause specified; use current proc-bind-var | ||||
1550 | if (proc_bind == proc_bind_default) { | ||||
1551 | proc_bind = master_th->th.th_current_task->td_icvs.proc_bind; | ||||
1552 | } | ||||
1553 | /* else: The proc_bind policy was specified explicitly on parallel clause. | ||||
1554 | This overrides proc-bind-var for this parallel region, but does not | ||||
1555 | change proc-bind-var. */ | ||||
1556 | // Figure the value of proc-bind-var for the child threads. | ||||
1557 | if ((level + 1 < __kmp_nested_proc_bind.used) && | ||||
1558 | (__kmp_nested_proc_bind.bind_types[level + 1] != | ||||
1559 | master_th->th.th_current_task->td_icvs.proc_bind)) { | ||||
1560 | proc_bind_icv = __kmp_nested_proc_bind.bind_types[level + 1]; | ||||
1561 | } | ||||
1562 | } | ||||
1563 | KMP_CHECK_UPDATE(parent_team->t.t_proc_bind, proc_bind)if ((parent_team->t.t_proc_bind) != (proc_bind)) (parent_team ->t.t_proc_bind) = (proc_bind); | ||||
1564 | // Need to change the bind-var ICV to correct value for each implicit task | ||||
1565 | if (proc_bind_icv != proc_bind_default && | ||||
1566 | master_th->th.th_current_task->td_icvs.proc_bind != proc_bind_icv) { | ||||
1567 | kmp_info_t **other_threads = parent_team->t.t_threads; | ||||
1568 | for (i = 0; i < master_th->th.th_team_nproc; ++i) { | ||||
1569 | other_threads[i]->th.th_current_task->td_icvs.proc_bind = proc_bind_icv; | ||||
1570 | } | ||||
1571 | } | ||||
1572 | // Reset for next parallel region | ||||
1573 | master_th->th.th_set_proc_bind = proc_bind_default; | ||||
1574 | |||||
1575 | #if USE_ITT_BUILD1 && USE_ITT_NOTIFY1 | ||||
1576 | if (((__itt_frame_submit_v3_ptr__kmp_itt_frame_submit_v3_ptr__3_0 && __itt_get_timestamp_ptr__kmp_itt_get_timestamp_ptr__3_0) || | ||||
1577 | KMP_ITT_DEBUG0) && | ||||
1578 | __kmp_forkjoin_frames_mode == 3 && | ||||
1579 | parent_team->t.t_active_level == 1 // only report frames at level 1 | ||||
1580 | && master_th->th.th_teams_size.nteams == 1) { | ||||
1581 | kmp_uint64 tmp_time = __itt_get_timestamp(!__kmp_itt_get_timestamp_ptr__3_0) ? 0 : __kmp_itt_get_timestamp_ptr__3_0(); | ||||
1582 | master_th->th.th_frame_time = tmp_time; | ||||
1583 | parent_team->t.t_region_time = tmp_time; | ||||
1584 | } | ||||
1585 | if (__itt_stack_caller_create_ptr__kmp_itt_stack_caller_create_ptr__3_0) { | ||||
1586 | KMP_DEBUG_ASSERT(parent_team->t.t_stack_id == NULL)if (!(parent_team->t.t_stack_id == __null)) { __kmp_debug_assert ("parent_team->t.t_stack_id == __null", "openmp/runtime/src/kmp_runtime.cpp" , 1586); }; | ||||
1587 | // create new stack stitching id before entering fork barrier | ||||
1588 | parent_team->t.t_stack_id = __kmp_itt_stack_caller_create(); | ||||
1589 | } | ||||
1590 | #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */ | ||||
1591 | #if KMP_AFFINITY_SUPPORTED1 | ||||
1592 | __kmp_partition_places(parent_team); | ||||
1593 | #endif | ||||
1594 | |||||
1595 | KF_TRACE(10, ("__kmp_fork_in_teams: before internal fork: root=%p, team=%p, "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_in_teams: before internal fork: root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, parent_team, master_th, gtid ); } | ||||
1596 | "master_th=%p, gtid=%d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_in_teams: before internal fork: root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, parent_team, master_th, gtid ); } | ||||
1597 | root, parent_team, master_th, gtid))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_in_teams: before internal fork: root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, parent_team, master_th, gtid ); }; | ||||
1598 | __kmp_internal_fork(loc, gtid, parent_team); | ||||
1599 | KF_TRACE(10, ("__kmp_fork_in_teams: after internal fork: root=%p, team=%p, "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_in_teams: after internal fork: root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, parent_team, master_th, gtid ); } | ||||
1600 | "master_th=%p, gtid=%d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_in_teams: after internal fork: root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, parent_team, master_th, gtid ); } | ||||
1601 | root, parent_team, master_th, gtid))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_in_teams: after internal fork: root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, parent_team, master_th, gtid ); }; | ||||
1602 | |||||
1603 | if (call_context == fork_context_gnu) | ||||
1604 | return TRUE(!0); | ||||
1605 | |||||
1606 | /* Invoke microtask for PRIMARY thread */ | ||||
1607 | KA_TRACE(20, ("__kmp_fork_in_teams: T#%d(%d:0) invoke microtask = %p\n", gtid,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_in_teams: T#%d(%d:0) invoke microtask = %p\n" , gtid, parent_team->t.t_id, parent_team->t.t_pkfn); } | ||||
1608 | parent_team->t.t_id, parent_team->t.t_pkfn))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_in_teams: T#%d(%d:0) invoke microtask = %p\n" , gtid, parent_team->t.t_id, parent_team->t.t_pkfn); }; | ||||
1609 | |||||
1610 | if (!parent_team->t.t_invoke(gtid)) { | ||||
1611 | KMP_ASSERT2(0, "cannot invoke microtask for PRIMARY thread")if (!(0)) { __kmp_debug_assert(("cannot invoke microtask for PRIMARY thread" ), "openmp/runtime/src/kmp_runtime.cpp", 1611); }; | ||||
1612 | } | ||||
1613 | KA_TRACE(20, ("__kmp_fork_in_teams: T#%d(%d:0) done microtask = %p\n", gtid,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_in_teams: T#%d(%d:0) done microtask = %p\n" , gtid, parent_team->t.t_id, parent_team->t.t_pkfn); } | ||||
1614 | parent_team->t.t_id, parent_team->t.t_pkfn))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_in_teams: T#%d(%d:0) done microtask = %p\n" , gtid, parent_team->t.t_id, parent_team->t.t_pkfn); }; | ||||
1615 | KMP_MB(); /* Flush all pending memory write invalidates. */ | ||||
1616 | |||||
1617 | KA_TRACE(20, ("__kmp_fork_in_teams: parallel exit T#%d\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_in_teams: parallel exit T#%d\n" , gtid); }; | ||||
1618 | |||||
1619 | return TRUE(!0); | ||||
1620 | } | ||||
1621 | |||||
1622 | // Create a serialized parallel region | ||||
1623 | static inline int | ||||
1624 | __kmp_serial_fork_call(ident_t *loc, int gtid, enum fork_context_e call_context, | ||||
1625 | kmp_int32 argc, microtask_t microtask, launch_t invoker, | ||||
1626 | kmp_info_t *master_th, kmp_team_t *parent_team, | ||||
1627 | #if OMPT_SUPPORT1 | ||||
1628 | ompt_data_t *ompt_parallel_data, void **return_address, | ||||
1629 | ompt_data_t **parent_task_data, | ||||
1630 | #endif | ||||
1631 | kmp_va_list ap) { | ||||
1632 | kmp_team_t *team; | ||||
1633 | int i; | ||||
1634 | void **argv; | ||||
1635 | |||||
1636 | /* josh todo: hypothetical question: what do we do for OS X*? */ | ||||
1637 | #if KMP_OS_LINUX1 && \ | ||||
1638 | (KMP_ARCH_X860 || KMP_ARCH_X86_641 || KMP_ARCH_ARM || KMP_ARCH_AARCH640) | ||||
1639 | void *args[argc]; | ||||
1640 | #else | ||||
1641 | void **args = (void **)KMP_ALLOCA(argc * sizeof(void *))__builtin_alloca (argc * sizeof(void *)); | ||||
1642 | #endif /* KMP_OS_LINUX && ( KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || \ | ||||
1643 | KMP_ARCH_AARCH64) */ | ||||
1644 | |||||
1645 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_serial_fork_call: T#%d serializing parallel region\n" , gtid); } | ||||
1646 | 20, ("__kmp_serial_fork_call: T#%d serializing parallel region\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_serial_fork_call: T#%d serializing parallel region\n" , gtid); }; | ||||
1647 | |||||
1648 | __kmpc_serialized_parallel(loc, gtid); | ||||
1649 | |||||
1650 | #if OMPD_SUPPORT1 | ||||
1651 | master_th->th.th_serial_team->t.t_pkfn = microtask; | ||||
1652 | #endif | ||||
1653 | |||||
1654 | if (call_context == fork_context_intel) { | ||||
1655 | /* TODO this sucks, use the compiler itself to pass args! :) */ | ||||
1656 | master_th->th.th_serial_team->t.t_ident = loc; | ||||
1657 | if (!ap) { | ||||
1658 | // revert change made in __kmpc_serialized_parallel() | ||||
1659 | master_th->th.th_serial_team->t.t_level--; | ||||
1660 | // Get args from parent team for teams construct | ||||
1661 | |||||
1662 | #if OMPT_SUPPORT1 | ||||
1663 | void *dummy; | ||||
1664 | void **exit_frame_p; | ||||
1665 | ompt_task_info_t *task_info; | ||||
1666 | ompt_lw_taskteam_t lw_taskteam; | ||||
1667 | |||||
1668 | if (ompt_enabled.enabled) { | ||||
1669 | __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid, | ||||
1670 | ompt_parallel_data, *return_address); | ||||
1671 | |||||
1672 | __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0); | ||||
1673 | // don't use lw_taskteam after linking. content was swaped | ||||
1674 | task_info = OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info)); | ||||
1675 | exit_frame_p = &(task_info->frame.exit_frame.ptr); | ||||
1676 | if (ompt_enabled.ompt_callback_implicit_task) { | ||||
1677 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num = __kmp_tid_from_gtid(gtid); | ||||
1678 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | ||||
1679 | ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th)(&(master_th->th.th_team->t.ompt_team_info.parallel_data )), | ||||
1680 | &(task_info->task_data), 1, | ||||
1681 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num, ompt_task_implicit); | ||||
1682 | } | ||||
1683 | |||||
1684 | /* OMPT state */ | ||||
1685 | master_th->th.ompt_thread_info.state = ompt_state_work_parallel; | ||||
1686 | } else { | ||||
1687 | exit_frame_p = &dummy; | ||||
1688 | } | ||||
1689 | #endif | ||||
1690 | |||||
1691 | { | ||||
1692 | KMP_TIME_PARTITIONED_BLOCK(OMP_parallel)((void)0); | ||||
1693 | KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK)((void)0); | ||||
1694 | __kmp_invoke_microtask(microtask, gtid, 0, argc, parent_team->t.t_argv | ||||
1695 | #if OMPT_SUPPORT1 | ||||
1696 | , | ||||
1697 | exit_frame_p | ||||
1698 | #endif | ||||
1699 | ); | ||||
1700 | } | ||||
1701 | |||||
1702 | #if OMPT_SUPPORT1 | ||||
1703 | if (ompt_enabled.enabled) { | ||||
1704 | *exit_frame_p = NULL__null; | ||||
1705 | if (ompt_enabled.ompt_callback_implicit_task) { | ||||
1706 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | ||||
1707 | ompt_scope_end, NULL__null, &(task_info->task_data), 1, | ||||
1708 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num, ompt_task_implicit); | ||||
1709 | } | ||||
1710 | *ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th)(&(master_th->th.th_team->t.ompt_team_info.parallel_data )); | ||||
1711 | __ompt_lw_taskteam_unlink(master_th); | ||||
1712 | if (ompt_enabled.ompt_callback_parallel_end) { | ||||
1713 | ompt_callbacks.ompt_callback(ompt_callback_parallel_end)ompt_callback_parallel_end_callback( | ||||
1714 | ompt_parallel_data, *parent_task_data, | ||||
1715 | OMPT_INVOKER(call_context)((call_context == fork_context_gnu) ? ompt_parallel_invoker_program : ompt_parallel_invoker_runtime) | ompt_parallel_team, *return_address); | ||||
1716 | } | ||||
1717 | master_th->th.ompt_thread_info.state = ompt_state_overhead; | ||||
1718 | } | ||||
1719 | #endif | ||||
1720 | } else if (microtask == (microtask_t)__kmp_teams_master) { | ||||
1721 | KMP_DEBUG_ASSERT(master_th->th.th_team == master_th->th.th_serial_team)if (!(master_th->th.th_team == master_th->th.th_serial_team )) { __kmp_debug_assert("master_th->th.th_team == master_th->th.th_serial_team" , "openmp/runtime/src/kmp_runtime.cpp", 1721); }; | ||||
1722 | team = master_th->th.th_team; | ||||
1723 | // team->t.t_pkfn = microtask; | ||||
1724 | team->t.t_invoke = invoker; | ||||
1725 | __kmp_alloc_argv_entries(argc, team, TRUE(!0)); | ||||
1726 | team->t.t_argc = argc; | ||||
1727 | argv = (void **)team->t.t_argv; | ||||
1728 | if (ap) { | ||||
1729 | for (i = argc - 1; i >= 0; --i) | ||||
1730 | *argv++ = va_arg(kmp_va_deref(ap), void *)__builtin_va_arg((*(ap)), void *); | ||||
1731 | } else { | ||||
1732 | for (i = 0; i < argc; ++i) | ||||
1733 | // Get args from parent team for teams construct | ||||
1734 | argv[i] = parent_team->t.t_argv[i]; | ||||
1735 | } | ||||
1736 | // AC: revert change made in __kmpc_serialized_parallel() | ||||
1737 | // because initial code in teams should have level=0 | ||||
1738 | team->t.t_level--; | ||||
1739 | // AC: call special invoker for outer "parallel" of teams construct | ||||
1740 | invoker(gtid); | ||||
1741 | #if OMPT_SUPPORT1 | ||||
1742 | if (ompt_enabled.enabled) { | ||||
1743 | ompt_task_info_t *task_info = OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info)); | ||||
1744 | if (ompt_enabled.ompt_callback_implicit_task) { | ||||
1745 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | ||||
1746 | ompt_scope_end, NULL__null, &(task_info->task_data), 0, | ||||
1747 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num, ompt_task_initial); | ||||
1748 | } | ||||
1749 | if (ompt_enabled.ompt_callback_parallel_end) { | ||||
1750 | ompt_callbacks.ompt_callback(ompt_callback_parallel_end)ompt_callback_parallel_end_callback( | ||||
1751 | ompt_parallel_data, *parent_task_data, | ||||
1752 | OMPT_INVOKER(call_context)((call_context == fork_context_gnu) ? ompt_parallel_invoker_program : ompt_parallel_invoker_runtime) | ompt_parallel_league, | ||||
1753 | *return_address); | ||||
1754 | } | ||||
1755 | master_th->th.ompt_thread_info.state = ompt_state_overhead; | ||||
1756 | } | ||||
1757 | #endif | ||||
1758 | } else { | ||||
1759 | argv = args; | ||||
1760 | for (i = argc - 1; i >= 0; --i) | ||||
1761 | *argv++ = va_arg(kmp_va_deref(ap), void *)__builtin_va_arg((*(ap)), void *); | ||||
1762 | KMP_MB(); | ||||
1763 | |||||
1764 | #if OMPT_SUPPORT1 | ||||
1765 | void *dummy; | ||||
1766 | void **exit_frame_p; | ||||
1767 | ompt_task_info_t *task_info; | ||||
1768 | ompt_lw_taskteam_t lw_taskteam; | ||||
1769 | ompt_data_t *implicit_task_data; | ||||
1770 | |||||
1771 | if (ompt_enabled.enabled) { | ||||
1772 | __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid, | ||||
1773 | ompt_parallel_data, *return_address); | ||||
1774 | __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0); | ||||
1775 | // don't use lw_taskteam after linking. content was swaped | ||||
1776 | task_info = OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info)); | ||||
1777 | exit_frame_p = &(task_info->frame.exit_frame.ptr); | ||||
1778 | |||||
1779 | /* OMPT implicit task begin */ | ||||
1780 | implicit_task_data = OMPT_CUR_TASK_DATA(master_th)(&(master_th->th.th_current_task->ompt_task_info.task_data )); | ||||
1781 | if (ompt_enabled.ompt_callback_implicit_task) { | ||||
1782 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | ||||
1783 | ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th)(&(master_th->th.th_team->t.ompt_team_info.parallel_data )), | ||||
1784 | implicit_task_data, 1, __kmp_tid_from_gtid(gtid), | ||||
1785 | ompt_task_implicit); | ||||
1786 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num = __kmp_tid_from_gtid(gtid); | ||||
1787 | } | ||||
1788 | |||||
1789 | /* OMPT state */ | ||||
1790 | master_th->th.ompt_thread_info.state = ompt_state_work_parallel; | ||||
1791 | } else { | ||||
1792 | exit_frame_p = &dummy; | ||||
1793 | } | ||||
1794 | #endif | ||||
1795 | |||||
1796 | { | ||||
1797 | KMP_TIME_PARTITIONED_BLOCK(OMP_parallel)((void)0); | ||||
1798 | KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK)((void)0); | ||||
1799 | __kmp_invoke_microtask(microtask, gtid, 0, argc, args | ||||
1800 | #if OMPT_SUPPORT1 | ||||
1801 | , | ||||
1802 | exit_frame_p | ||||
1803 | #endif | ||||
1804 | ); | ||||
1805 | } | ||||
1806 | |||||
1807 | #if OMPT_SUPPORT1 | ||||
1808 | if (ompt_enabled.enabled) { | ||||
1809 | *exit_frame_p = NULL__null; | ||||
1810 | if (ompt_enabled.ompt_callback_implicit_task) { | ||||
1811 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | ||||
1812 | ompt_scope_end, NULL__null, &(task_info->task_data), 1, | ||||
1813 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num, ompt_task_implicit); | ||||
1814 | } | ||||
1815 | |||||
1816 | *ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th)(&(master_th->th.th_team->t.ompt_team_info.parallel_data )); | ||||
1817 | __ompt_lw_taskteam_unlink(master_th); | ||||
1818 | if (ompt_enabled.ompt_callback_parallel_end) { | ||||
1819 | ompt_callbacks.ompt_callback(ompt_callback_parallel_end)ompt_callback_parallel_end_callback( | ||||
1820 | ompt_parallel_data, *parent_task_data, | ||||
1821 | OMPT_INVOKER(call_context)((call_context == fork_context_gnu) ? ompt_parallel_invoker_program : ompt_parallel_invoker_runtime) | ompt_parallel_team, *return_address); | ||||
1822 | } | ||||
1823 | master_th->th.ompt_thread_info.state = ompt_state_overhead; | ||||
1824 | } | ||||
1825 | #endif | ||||
1826 | } | ||||
1827 | } else if (call_context == fork_context_gnu) { | ||||
1828 | #if OMPT_SUPPORT1 | ||||
1829 | if (ompt_enabled.enabled) { | ||||
1830 | ompt_lw_taskteam_t lwt; | ||||
1831 | __ompt_lw_taskteam_init(&lwt, master_th, gtid, ompt_parallel_data, | ||||
1832 | *return_address); | ||||
1833 | |||||
1834 | lwt.ompt_task_info.frame.exit_frame = ompt_data_none{0}; | ||||
1835 | __ompt_lw_taskteam_link(&lwt, master_th, 1); | ||||
1836 | } | ||||
1837 | // don't use lw_taskteam after linking. content was swaped | ||||
1838 | #endif | ||||
1839 | |||||
1840 | // we were called from GNU native code | ||||
1841 | KA_TRACE(20, ("__kmp_serial_fork_call: T#%d serial exit\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_serial_fork_call: T#%d serial exit\n" , gtid); }; | ||||
1842 | return FALSE0; | ||||
1843 | } else { | ||||
1844 | KMP_ASSERT2(call_context < fork_context_last,if (!(call_context < fork_context_last)) { __kmp_debug_assert (("__kmp_serial_fork_call: unknown fork_context parameter"), "openmp/runtime/src/kmp_runtime.cpp" , 1845); } | ||||
1845 | "__kmp_serial_fork_call: unknown fork_context parameter")if (!(call_context < fork_context_last)) { __kmp_debug_assert (("__kmp_serial_fork_call: unknown fork_context parameter"), "openmp/runtime/src/kmp_runtime.cpp" , 1845); }; | ||||
1846 | } | ||||
1847 | |||||
1848 | KA_TRACE(20, ("__kmp_serial_fork_call: T#%d serial exit\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_serial_fork_call: T#%d serial exit\n" , gtid); }; | ||||
1849 | KMP_MB(); | ||||
1850 | return FALSE0; | ||||
1851 | } | ||||
1852 | |||||
1853 | /* most of the work for a fork */ | ||||
1854 | /* return true if we really went parallel, false if serialized */ | ||||
1855 | int __kmp_fork_call(ident_t *loc, int gtid, | ||||
1856 | enum fork_context_e call_context, // Intel, GNU, ... | ||||
1857 | kmp_int32 argc, microtask_t microtask, launch_t invoker, | ||||
1858 | kmp_va_list ap) { | ||||
1859 | void **argv; | ||||
1860 | int i; | ||||
1861 | int master_tid; | ||||
1862 | int master_this_cons; | ||||
1863 | kmp_team_t *team; | ||||
1864 | kmp_team_t *parent_team; | ||||
1865 | kmp_info_t *master_th; | ||||
1866 | kmp_root_t *root; | ||||
1867 | int nthreads; | ||||
1868 | int master_active; | ||||
1869 | int master_set_numthreads; | ||||
1870 | int level; | ||||
1871 | int active_level; | ||||
1872 | int teams_level; | ||||
1873 | #if KMP_NESTED_HOT_TEAMS1 | ||||
1874 | kmp_hot_team_ptr_t **p_hot_teams; | ||||
1875 | #endif | ||||
1876 | { // KMP_TIME_BLOCK | ||||
1877 | KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_fork_call)((void)0); | ||||
1878 | KMP_COUNT_VALUE(OMP_PARALLEL_args, argc)((void)0); | ||||
1879 | |||||
1880 | KA_TRACE(20, ("__kmp_fork_call: enter T#%d\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: enter T#%d\n" , gtid); }; | ||||
1881 | if (__kmp_stkpadding > 0 && __kmp_root[gtid] != NULL__null) { | ||||
1882 | /* Some systems prefer the stack for the root thread(s) to start with */ | ||||
1883 | /* some gap from the parent stack to prevent false sharing. */ | ||||
1884 | void *dummy = KMP_ALLOCA(__kmp_stkpadding)__builtin_alloca (__kmp_stkpadding); | ||||
1885 | /* These 2 lines below are so this does not get optimized out */ | ||||
1886 | if (__kmp_stkpadding > KMP_MAX_STKPADDING(2 * 1024 * 1024)) | ||||
1887 | __kmp_stkpadding += (short)((kmp_int64)dummy); | ||||
1888 | } | ||||
1889 | |||||
1890 | /* initialize if needed */ | ||||
1891 | KMP_DEBUG_ASSERT(if (!(__kmp_init_serial)) { __kmp_debug_assert("__kmp_init_serial" , "openmp/runtime/src/kmp_runtime.cpp", 1892); } | ||||
1892 | __kmp_init_serial)if (!(__kmp_init_serial)) { __kmp_debug_assert("__kmp_init_serial" , "openmp/runtime/src/kmp_runtime.cpp", 1892); }; // AC: potentially unsafe, not in sync with shutdown | ||||
1893 | if (!TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) | ||||
1894 | __kmp_parallel_initialize(); | ||||
1895 | __kmp_resume_if_soft_paused(); | ||||
1896 | |||||
1897 | /* setup current data */ | ||||
1898 | // AC: potentially unsafe, not in sync with library shutdown, | ||||
1899 | // __kmp_threads can be freed | ||||
1900 | master_th = __kmp_threads[gtid]; | ||||
1901 | |||||
1902 | parent_team = master_th->th.th_team; | ||||
1903 | master_tid = master_th->th.th_info.ds.ds_tid; | ||||
1904 | master_this_cons = master_th->th.th_local.this_construct; | ||||
1905 | root = master_th->th.th_root; | ||||
1906 | master_active = root->r.r_active; | ||||
1907 | master_set_numthreads = master_th->th.th_set_nproc; | ||||
1908 | |||||
1909 | #if OMPT_SUPPORT1 | ||||
1910 | ompt_data_t ompt_parallel_data = ompt_data_none{0}; | ||||
1911 | ompt_data_t *parent_task_data; | ||||
1912 | ompt_frame_t *ompt_frame; | ||||
1913 | void *return_address = NULL__null; | ||||
1914 | |||||
1915 | if (ompt_enabled.enabled) { | ||||
1916 | __ompt_get_task_info_internal(0, NULL__null, &parent_task_data, &ompt_frame, | ||||
1917 | NULL__null, NULL__null); | ||||
1918 | return_address = OMPT_LOAD_RETURN_ADDRESS(gtid)__ompt_load_return_address(gtid); | ||||
1919 | } | ||||
1920 | #endif | ||||
1921 | |||||
1922 | // Assign affinity to root thread if it hasn't happened yet | ||||
1923 | __kmp_assign_root_init_mask(); | ||||
1924 | |||||
1925 | // Nested level will be an index in the nested nthreads array | ||||
1926 | level = parent_team->t.t_level; | ||||
1927 | // used to launch non-serial teams even if nested is not allowed | ||||
1928 | active_level = parent_team->t.t_active_level; | ||||
1929 | // needed to check nesting inside the teams | ||||
1930 | teams_level = master_th->th.th_teams_level; | ||||
1931 | #if KMP_NESTED_HOT_TEAMS1 | ||||
1932 | p_hot_teams = &master_th->th.th_hot_teams; | ||||
1933 | if (*p_hot_teams == NULL__null && __kmp_hot_teams_max_level > 0) { | ||||
1934 | *p_hot_teams = (kmp_hot_team_ptr_t *)__kmp_allocate(___kmp_allocate((sizeof(kmp_hot_team_ptr_t) * __kmp_hot_teams_max_level ), "openmp/runtime/src/kmp_runtime.cpp", 1935) | ||||
1935 | sizeof(kmp_hot_team_ptr_t) * __kmp_hot_teams_max_level)___kmp_allocate((sizeof(kmp_hot_team_ptr_t) * __kmp_hot_teams_max_level ), "openmp/runtime/src/kmp_runtime.cpp", 1935); | ||||
1936 | (*p_hot_teams)[0].hot_team = root->r.r_hot_team; | ||||
1937 | // it is either actual or not needed (when active_level > 0) | ||||
1938 | (*p_hot_teams)[0].hot_team_nth = 1; | ||||
1939 | } | ||||
1940 | #endif | ||||
1941 | |||||
1942 | #if OMPT_SUPPORT1 | ||||
1943 | if (ompt_enabled.enabled) { | ||||
1944 | if (ompt_enabled.ompt_callback_parallel_begin) { | ||||
1945 | int team_size = master_set_numthreads | ||||
1946 | ? master_set_numthreads | ||||
1947 | : get__nproc_2(parent_team, master_tid)((parent_team)->t.t_threads[(master_tid)]->th.th_current_task ->td_icvs.nproc); | ||||
1948 | int flags = OMPT_INVOKER(call_context)((call_context == fork_context_gnu) ? ompt_parallel_invoker_program : ompt_parallel_invoker_runtime) | | ||||
1949 | ((microtask == (microtask_t)__kmp_teams_master) | ||||
1950 | ? ompt_parallel_league | ||||
1951 | : ompt_parallel_team); | ||||
1952 | ompt_callbacks.ompt_callback(ompt_callback_parallel_begin)ompt_callback_parallel_begin_callback( | ||||
1953 | parent_task_data, ompt_frame, &ompt_parallel_data, team_size, flags, | ||||
1954 | return_address); | ||||
1955 | } | ||||
1956 | master_th->th.ompt_thread_info.state = ompt_state_overhead; | ||||
1957 | } | ||||
1958 | #endif | ||||
1959 | |||||
1960 | master_th->th.th_ident = loc; | ||||
1961 | |||||
1962 | // Parallel closely nested in teams construct: | ||||
1963 | if (__kmp_is_fork_in_teams(master_th, microtask, level, teams_level, ap)) { | ||||
1964 | return __kmp_fork_in_teams(loc, gtid, parent_team, argc, master_th, root, | ||||
1965 | call_context, microtask, invoker, | ||||
1966 | master_set_numthreads, level, | ||||
1967 | #if OMPT_SUPPORT1 | ||||
1968 | ompt_parallel_data, return_address, | ||||
1969 | #endif | ||||
1970 | ap); | ||||
1971 | } // End parallel closely nested in teams construct | ||||
1972 | |||||
1973 | #if KMP_DEBUG1 | ||||
1974 | if (__kmp_tasking_mode != tskm_immediate_exec) { | ||||
1975 | KMP_DEBUG_ASSERT(master_th->th.th_task_team ==if (!(master_th->th.th_task_team == parent_team->t.t_task_team [master_th->th.th_task_state])) { __kmp_debug_assert("master_th->th.th_task_team == parent_team->t.t_task_team[master_th->th.th_task_state]" , "openmp/runtime/src/kmp_runtime.cpp", 1976); } | ||||
1976 | parent_team->t.t_task_team[master_th->th.th_task_state])if (!(master_th->th.th_task_team == parent_team->t.t_task_team [master_th->th.th_task_state])) { __kmp_debug_assert("master_th->th.th_task_team == parent_team->t.t_task_team[master_th->th.th_task_state]" , "openmp/runtime/src/kmp_runtime.cpp", 1976); }; | ||||
1977 | } | ||||
1978 | #endif | ||||
1979 | |||||
1980 | // Need this to happen before we determine the number of threads, not while | ||||
1981 | // we are allocating the team | ||||
1982 | //__kmp_push_current_task_to_thread(master_th, parent_team, 0); | ||||
1983 | |||||
1984 | // Determine the number of threads | ||||
1985 | int enter_teams = | ||||
1986 | __kmp_is_entering_teams(active_level, level, teams_level, ap); | ||||
1987 | if ((!enter_teams && | ||||
1988 | (parent_team->t.t_active_level >= | ||||
1989 | master_th->th.th_current_task->td_icvs.max_active_levels)) || | ||||
1990 | (__kmp_library == library_serial)) { | ||||
1991 | KC_TRACE(10, ("__kmp_fork_call: T#%d serializing team\n", gtid))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: T#%d serializing team\n" , gtid); }; | ||||
1992 | nthreads = 1; | ||||
1993 | } else { | ||||
1994 | nthreads = master_set_numthreads | ||||
1995 | ? master_set_numthreads | ||||
1996 | // TODO: get nproc directly from current task | ||||
1997 | : get__nproc_2(parent_team, master_tid)((parent_team)->t.t_threads[(master_tid)]->th.th_current_task ->td_icvs.nproc); | ||||
1998 | // Check if we need to take forkjoin lock? (no need for serialized | ||||
1999 | // parallel out of teams construct). | ||||
2000 | if (nthreads > 1) { | ||||
2001 | /* determine how many new threads we can use */ | ||||
2002 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); | ||||
2003 | /* AC: If we execute teams from parallel region (on host), then teams | ||||
2004 | should be created but each can only have 1 thread if nesting is | ||||
2005 | disabled. If teams called from serial region, then teams and their | ||||
2006 | threads should be created regardless of the nesting setting. */ | ||||
2007 | nthreads = __kmp_reserve_threads(root, parent_team, master_tid, | ||||
2008 | nthreads, enter_teams); | ||||
2009 | if (nthreads == 1) { | ||||
2010 | // Free lock for single thread execution here; for multi-thread | ||||
2011 | // execution it will be freed later after team of threads created | ||||
2012 | // and initialized | ||||
2013 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | ||||
2014 | } | ||||
2015 | } | ||||
2016 | } | ||||
2017 | KMP_DEBUG_ASSERT(nthreads > 0)if (!(nthreads > 0)) { __kmp_debug_assert("nthreads > 0" , "openmp/runtime/src/kmp_runtime.cpp", 2017); }; | ||||
2018 | |||||
2019 | // If we temporarily changed the set number of threads then restore it now | ||||
2020 | master_th->th.th_set_nproc = 0; | ||||
2021 | |||||
2022 | if (nthreads == 1) { | ||||
2023 | return __kmp_serial_fork_call(loc, gtid, call_context, argc, microtask, | ||||
2024 | invoker, master_th, parent_team, | ||||
2025 | #if OMPT_SUPPORT1 | ||||
2026 | &ompt_parallel_data, &return_address, | ||||
2027 | &parent_task_data, | ||||
2028 | #endif | ||||
2029 | ap); | ||||
2030 | } // if (nthreads == 1) | ||||
2031 | |||||
2032 | // GEH: only modify the executing flag in the case when not serialized | ||||
2033 | // serialized case is handled in kmpc_serialized_parallel | ||||
2034 | KF_TRACE(10, ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, " "curtask=%p, curtask_max_aclevel=%d\n", parent_team->t.t_active_level , master_th, master_th->th.th_current_task, master_th-> th.th_current_task->td_icvs.max_active_levels); } | ||||
2035 | "curtask=%p, curtask_max_aclevel=%d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, " "curtask=%p, curtask_max_aclevel=%d\n", parent_team->t.t_active_level , master_th, master_th->th.th_current_task, master_th-> th.th_current_task->td_icvs.max_active_levels); } | ||||
2036 | parent_team->t.t_active_level, master_th,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, " "curtask=%p, curtask_max_aclevel=%d\n", parent_team->t.t_active_level , master_th, master_th->th.th_current_task, master_th-> th.th_current_task->td_icvs.max_active_levels); } | ||||
2037 | master_th->th.th_current_task,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, " "curtask=%p, curtask_max_aclevel=%d\n", parent_team->t.t_active_level , master_th, master_th->th.th_current_task, master_th-> th.th_current_task->td_icvs.max_active_levels); } | ||||
2038 | master_th->th.th_current_task->td_icvs.max_active_levels))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, " "curtask=%p, curtask_max_aclevel=%d\n", parent_team->t.t_active_level , master_th, master_th->th.th_current_task, master_th-> th.th_current_task->td_icvs.max_active_levels); }; | ||||
2039 | // TODO: GEH - cannot do this assertion because root thread not set up as | ||||
2040 | // executing | ||||
2041 | // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 1 ); | ||||
2042 | master_th->th.th_current_task->td_flags.executing = 0; | ||||
2043 | |||||
2044 | if (!master_th->th.th_teams_microtask || level > teams_level) { | ||||
2045 | /* Increment our nested depth level */ | ||||
2046 | KMP_ATOMIC_INC(&root->r.r_in_parallel)(&root->r.r_in_parallel)->fetch_add(1, std::memory_order_acq_rel ); | ||||
2047 | } | ||||
2048 | |||||
2049 | // See if we need to make a copy of the ICVs. | ||||
2050 | int nthreads_icv = master_th->th.th_current_task->td_icvs.nproc; | ||||
2051 | if ((level + 1 < __kmp_nested_nth.used) && | ||||
2052 | (__kmp_nested_nth.nth[level + 1] != nthreads_icv)) { | ||||
2053 | nthreads_icv = __kmp_nested_nth.nth[level + 1]; | ||||
2054 | } else { | ||||
2055 | nthreads_icv = 0; // don't update | ||||
2056 | } | ||||
2057 | |||||
2058 | // Figure out the proc_bind_policy for the new team. | ||||
2059 | kmp_proc_bind_t proc_bind = master_th->th.th_set_proc_bind; | ||||
2060 | // proc_bind_default means don't update | ||||
2061 | kmp_proc_bind_t proc_bind_icv = proc_bind_default; | ||||
2062 | if (master_th->th.th_current_task->td_icvs.proc_bind == proc_bind_false) { | ||||
2063 | proc_bind = proc_bind_false; | ||||
2064 | } else { | ||||
2065 | // No proc_bind clause specified; use current proc-bind-var for this | ||||
2066 | // parallel region | ||||
2067 | if (proc_bind == proc_bind_default) { | ||||
2068 | proc_bind = master_th->th.th_current_task->td_icvs.proc_bind; | ||||
2069 | } | ||||
2070 | // Have teams construct take proc_bind value from KMP_TEAMS_PROC_BIND | ||||
2071 | if (master_th->th.th_teams_microtask && | ||||
2072 | microtask == (microtask_t)__kmp_teams_master) { | ||||
2073 | proc_bind = __kmp_teams_proc_bind; | ||||
2074 | } | ||||
2075 | /* else: The proc_bind policy was specified explicitly on parallel clause. | ||||
2076 | This overrides proc-bind-var for this parallel region, but does not | ||||
2077 | change proc-bind-var. */ | ||||
2078 | // Figure the value of proc-bind-var for the child threads. | ||||
2079 | if ((level + 1 < __kmp_nested_proc_bind.used) && | ||||
2080 | (__kmp_nested_proc_bind.bind_types[level + 1] != | ||||
2081 | master_th->th.th_current_task->td_icvs.proc_bind)) { | ||||
2082 | // Do not modify the proc bind icv for the two teams construct forks | ||||
2083 | // They just let the proc bind icv pass through | ||||
2084 | if (!master_th->th.th_teams_microtask || | ||||
2085 | !(microtask == (microtask_t)__kmp_teams_master || ap == NULL__null)) | ||||
2086 | proc_bind_icv = __kmp_nested_proc_bind.bind_types[level + 1]; | ||||
2087 | } | ||||
2088 | } | ||||
2089 | |||||
2090 | // Reset for next parallel region | ||||
2091 | master_th->th.th_set_proc_bind = proc_bind_default; | ||||
2092 | |||||
2093 | if ((nthreads_icv > 0) || (proc_bind_icv != proc_bind_default)) { | ||||
2094 | kmp_internal_control_t new_icvs; | ||||
2095 | copy_icvs(&new_icvs, &master_th->th.th_current_task->td_icvs); | ||||
2096 | new_icvs.next = NULL__null; | ||||
2097 | if (nthreads_icv > 0) { | ||||
2098 | new_icvs.nproc = nthreads_icv; | ||||
2099 | } | ||||
2100 | if (proc_bind_icv != proc_bind_default) { | ||||
2101 | new_icvs.proc_bind = proc_bind_icv; | ||||
2102 | } | ||||
2103 | |||||
2104 | /* allocate a new parallel team */ | ||||
2105 | KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: before __kmp_allocate_team\n" ); }; | ||||
2106 | team = __kmp_allocate_team(root, nthreads, nthreads, | ||||
2107 | #if OMPT_SUPPORT1 | ||||
2108 | ompt_parallel_data, | ||||
2109 | #endif | ||||
2110 | proc_bind, &new_icvs, | ||||
2111 | argc USE_NESTED_HOT_ARG(master_th), master_th); | ||||
2112 | if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) | ||||
2113 | copy_icvs((kmp_internal_control_t *)team->t.b->team_icvs, &new_icvs); | ||||
2114 | } else { | ||||
2115 | /* allocate a new parallel team */ | ||||
2116 | KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: before __kmp_allocate_team\n" ); }; | ||||
2117 | team = __kmp_allocate_team(root, nthreads, nthreads, | ||||
2118 | #if OMPT_SUPPORT1 | ||||
2119 | ompt_parallel_data, | ||||
2120 | #endif | ||||
2121 | proc_bind, | ||||
2122 | &master_th->th.th_current_task->td_icvs, | ||||
2123 | argc USE_NESTED_HOT_ARG(master_th), master_th); | ||||
2124 | if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) | ||||
2125 | copy_icvs((kmp_internal_control_t *)team->t.b->team_icvs, | ||||
2126 | &master_th->th.th_current_task->td_icvs); | ||||
2127 | } | ||||
2128 | KF_TRACE(if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: after __kmp_allocate_team - team = %p\n" , team); } | ||||
2129 | 10, ("__kmp_fork_call: after __kmp_allocate_team - team = %p\n", team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: after __kmp_allocate_team - team = %p\n" , team); }; | ||||
2130 | |||||
2131 | /* setup the new team */ | ||||
2132 | KMP_CHECK_UPDATE(team->t.t_master_tid, master_tid)if ((team->t.t_master_tid) != (master_tid)) (team->t.t_master_tid ) = (master_tid); | ||||
2133 | KMP_CHECK_UPDATE(team->t.t_master_this_cons, master_this_cons)if ((team->t.t_master_this_cons) != (master_this_cons)) (team ->t.t_master_this_cons) = (master_this_cons); | ||||
2134 | KMP_CHECK_UPDATE(team->t.t_ident, loc)if ((team->t.t_ident) != (loc)) (team->t.t_ident) = (loc ); | ||||
2135 | KMP_CHECK_UPDATE(team->t.t_parent, parent_team)if ((team->t.t_parent) != (parent_team)) (team->t.t_parent ) = (parent_team); | ||||
2136 | KMP_CHECK_UPDATE_SYNC(team->t.t_pkfn, microtask)if ((team->t.t_pkfn) != (microtask)) (((team->t.t_pkfn) )) = (((microtask))); | ||||
2137 | #if OMPT_SUPPORT1 | ||||
2138 | KMP_CHECK_UPDATE_SYNC(team->t.ompt_team_info.master_return_address,if ((team->t.ompt_team_info.master_return_address) != (return_address )) (((team->t.ompt_team_info.master_return_address))) = (( (return_address))) | ||||
2139 | return_address)if ((team->t.ompt_team_info.master_return_address) != (return_address )) (((team->t.ompt_team_info.master_return_address))) = (( (return_address))); | ||||
2140 | #endif | ||||
2141 | KMP_CHECK_UPDATE(team->t.t_invoke, invoker)if ((team->t.t_invoke) != (invoker)) (team->t.t_invoke) = (invoker); // TODO move to root, maybe | ||||
2142 | // TODO: parent_team->t.t_level == INT_MAX ??? | ||||
2143 | if (!master_th->th.th_teams_microtask || level > teams_level) { | ||||
2144 | int new_level = parent_team->t.t_level + 1; | ||||
2145 | KMP_CHECK_UPDATE(team->t.t_level, new_level)if ((team->t.t_level) != (new_level)) (team->t.t_level) = (new_level); | ||||
2146 | new_level = parent_team->t.t_active_level + 1; | ||||
2147 | KMP_CHECK_UPDATE(team->t.t_active_level, new_level)if ((team->t.t_active_level) != (new_level)) (team->t.t_active_level ) = (new_level); | ||||
2148 | } else { | ||||
2149 | // AC: Do not increase parallel level at start of the teams construct | ||||
2150 | int new_level = parent_team->t.t_level; | ||||
2151 | KMP_CHECK_UPDATE(team->t.t_level, new_level)if ((team->t.t_level) != (new_level)) (team->t.t_level) = (new_level); | ||||
2152 | new_level = parent_team->t.t_active_level; | ||||
2153 | KMP_CHECK_UPDATE(team->t.t_active_level, new_level)if ((team->t.t_active_level) != (new_level)) (team->t.t_active_level ) = (new_level); | ||||
2154 | } | ||||
2155 | kmp_r_sched_t new_sched = get__sched_2(parent_team, master_tid)((parent_team)->t.t_threads[(master_tid)]->th.th_current_task ->td_icvs.sched); | ||||
2156 | // set primary thread's schedule as new run-time schedule | ||||
2157 | KMP_CHECK_UPDATE(team->t.t_sched.sched, new_sched.sched)if ((team->t.t_sched.sched) != (new_sched.sched)) (team-> t.t_sched.sched) = (new_sched.sched); | ||||
2158 | |||||
2159 | KMP_CHECK_UPDATE(team->t.t_cancel_request, cancel_noreq)if ((team->t.t_cancel_request) != (cancel_noreq)) (team-> t.t_cancel_request) = (cancel_noreq); | ||||
2160 | KMP_CHECK_UPDATE(team->t.t_def_allocator, master_th->th.th_def_allocator)if ((team->t.t_def_allocator) != (master_th->th.th_def_allocator )) (team->t.t_def_allocator) = (master_th->th.th_def_allocator ); | ||||
2161 | |||||
2162 | // Update the floating point rounding in the team if required. | ||||
2163 | propagateFPControl(team); | ||||
2164 | #if OMPD_SUPPORT1 | ||||
2165 | if (ompd_state & OMPD_ENABLE_BP0x1) | ||||
2166 | ompd_bp_parallel_begin(); | ||||
2167 | #endif | ||||
2168 | |||||
2169 | if (__kmp_tasking_mode != tskm_immediate_exec) { | ||||
2170 | // Set primary thread's task team to team's task team. Unless this is hot | ||||
2171 | // team, it should be NULL. | ||||
2172 | KMP_DEBUG_ASSERT(master_th->th.th_task_team ==if (!(master_th->th.th_task_team == parent_team->t.t_task_team [master_th->th.th_task_state])) { __kmp_debug_assert("master_th->th.th_task_team == parent_team->t.t_task_team[master_th->th.th_task_state]" , "openmp/runtime/src/kmp_runtime.cpp", 2173); } | ||||
2173 | parent_team->t.t_task_team[master_th->th.th_task_state])if (!(master_th->th.th_task_team == parent_team->t.t_task_team [master_th->th.th_task_state])) { __kmp_debug_assert("master_th->th.th_task_team == parent_team->t.t_task_team[master_th->th.th_task_state]" , "openmp/runtime/src/kmp_runtime.cpp", 2173); }; | ||||
2174 | KA_TRACE(20, ("__kmp_fork_call: Primary T#%d pushing task_team %p / team "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: Primary T#%d pushing task_team %p / team " "%p, new task_team %p / team %p\n", __kmp_gtid_from_thread(master_th ), master_th->th.th_task_team, parent_team, team->t.t_task_team [master_th->th.th_task_state], team); } | ||||
2175 | "%p, new task_team %p / team %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: Primary T#%d pushing task_team %p / team " "%p, new task_team %p / team %p\n", __kmp_gtid_from_thread(master_th ), master_th->th.th_task_team, parent_team, team->t.t_task_team [master_th->th.th_task_state], team); } | ||||
2176 | __kmp_gtid_from_thread(master_th),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: Primary T#%d pushing task_team %p / team " "%p, new task_team %p / team %p\n", __kmp_gtid_from_thread(master_th ), master_th->th.th_task_team, parent_team, team->t.t_task_team [master_th->th.th_task_state], team); } | ||||
2177 | master_th->th.th_task_team, parent_team,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: Primary T#%d pushing task_team %p / team " "%p, new task_team %p / team %p\n", __kmp_gtid_from_thread(master_th ), master_th->th.th_task_team, parent_team, team->t.t_task_team [master_th->th.th_task_state], team); } | ||||
2178 | team->t.t_task_team[master_th->th.th_task_state], team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: Primary T#%d pushing task_team %p / team " "%p, new task_team %p / team %p\n", __kmp_gtid_from_thread(master_th ), master_th->th.th_task_team, parent_team, team->t.t_task_team [master_th->th.th_task_state], team); }; | ||||
2179 | |||||
2180 | if (active_level || master_th->th.th_task_team) { | ||||
2181 | // Take a memo of primary thread's task_state | ||||
2182 | KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack)if (!(master_th->th.th_task_state_memo_stack)) { __kmp_debug_assert ("master_th->th.th_task_state_memo_stack", "openmp/runtime/src/kmp_runtime.cpp" , 2182); }; | ||||
2183 | if (master_th->th.th_task_state_top >= | ||||
2184 | master_th->th.th_task_state_stack_sz) { // increase size | ||||
2185 | kmp_uint32 new_size = 2 * master_th->th.th_task_state_stack_sz; | ||||
2186 | kmp_uint8 *old_stack, *new_stack; | ||||
2187 | kmp_uint32 i; | ||||
2188 | new_stack = (kmp_uint8 *)__kmp_allocate(new_size)___kmp_allocate((new_size), "openmp/runtime/src/kmp_runtime.cpp" , 2188); | ||||
2189 | for (i = 0; i < master_th->th.th_task_state_stack_sz; ++i) { | ||||
2190 | new_stack[i] = master_th->th.th_task_state_memo_stack[i]; | ||||
2191 | } | ||||
2192 | for (i = master_th->th.th_task_state_stack_sz; i < new_size; | ||||
2193 | ++i) { // zero-init rest of stack | ||||
2194 | new_stack[i] = 0; | ||||
2195 | } | ||||
2196 | old_stack = master_th->th.th_task_state_memo_stack; | ||||
2197 | master_th->th.th_task_state_memo_stack = new_stack; | ||||
2198 | master_th->th.th_task_state_stack_sz = new_size; | ||||
2199 | __kmp_free(old_stack)___kmp_free((old_stack), "openmp/runtime/src/kmp_runtime.cpp" , 2199); | ||||
2200 | } | ||||
2201 | // Store primary thread's task_state on stack | ||||
2202 | master_th->th | ||||
2203 | .th_task_state_memo_stack[master_th->th.th_task_state_top] = | ||||
2204 | master_th->th.th_task_state; | ||||
2205 | master_th->th.th_task_state_top++; | ||||
2206 | #if KMP_NESTED_HOT_TEAMS1 | ||||
2207 | if (master_th->th.th_hot_teams && | ||||
2208 | active_level < __kmp_hot_teams_max_level && | ||||
2209 | team == master_th->th.th_hot_teams[active_level].hot_team) { | ||||
2210 | // Restore primary thread's nested state if nested hot team | ||||
2211 | master_th->th.th_task_state = | ||||
2212 | master_th->th | ||||
2213 | .th_task_state_memo_stack[master_th->th.th_task_state_top]; | ||||
2214 | } else { | ||||
2215 | #endif | ||||
2216 | master_th->th.th_task_state = 0; | ||||
2217 | #if KMP_NESTED_HOT_TEAMS1 | ||||
2218 | } | ||||
2219 | #endif | ||||
2220 | } | ||||
2221 | #if !KMP_NESTED_HOT_TEAMS1 | ||||
2222 | KMP_DEBUG_ASSERT((master_th->th.th_task_team == NULL) ||if (!((master_th->th.th_task_team == __null) || (team == root ->r.r_hot_team))) { __kmp_debug_assert("(master_th->th.th_task_team == __null) || (team == root->r.r_hot_team)" , "openmp/runtime/src/kmp_runtime.cpp", 2223); } | ||||
2223 | (team == root->r.r_hot_team))if (!((master_th->th.th_task_team == __null) || (team == root ->r.r_hot_team))) { __kmp_debug_assert("(master_th->th.th_task_team == __null) || (team == root->r.r_hot_team)" , "openmp/runtime/src/kmp_runtime.cpp", 2223); }; | ||||
2224 | #endif | ||||
2225 | } | ||||
2226 | |||||
2227 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n" , gtid, parent_team->t.t_id, team->t.t_master_tid, team ->t.t_id, team->t.t_nproc); } | ||||
2228 | 20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n" , gtid, parent_team->t.t_id, team->t.t_master_tid, team ->t.t_id, team->t.t_nproc); } | ||||
2229 | ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n" , gtid, parent_team->t.t_id, team->t.t_master_tid, team ->t.t_id, team->t.t_nproc); } | ||||
2230 | gtid, parent_team->t.t_id, team->t.t_master_tid, team->t.t_id,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n" , gtid, parent_team->t.t_id, team->t.t_master_tid, team ->t.t_id, team->t.t_nproc); } | ||||
2231 | team->t.t_nproc))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n" , gtid, parent_team->t.t_id, team->t.t_master_tid, team ->t.t_id, team->t.t_nproc); }; | ||||
2232 | KMP_DEBUG_ASSERT(team != root->r.r_hot_team ||if (!(team != root->r.r_hot_team || (team->t.t_master_tid == 0 && (team->t.t_parent == root->r.r_root_team || team->t.t_parent->t.t_serialized)))) { __kmp_debug_assert ("team != root->r.r_hot_team || (team->t.t_master_tid == 0 && (team->t.t_parent == root->r.r_root_team || team->t.t_parent->t.t_serialized))" , "openmp/runtime/src/kmp_runtime.cpp", 2235); } | ||||
2233 | (team->t.t_master_tid == 0 &&if (!(team != root->r.r_hot_team || (team->t.t_master_tid == 0 && (team->t.t_parent == root->r.r_root_team || team->t.t_parent->t.t_serialized)))) { __kmp_debug_assert ("team != root->r.r_hot_team || (team->t.t_master_tid == 0 && (team->t.t_parent == root->r.r_root_team || team->t.t_parent->t.t_serialized))" , "openmp/runtime/src/kmp_runtime.cpp", 2235); } | ||||
2234 | (team->t.t_parent == root->r.r_root_team ||if (!(team != root->r.r_hot_team || (team->t.t_master_tid == 0 && (team->t.t_parent == root->r.r_root_team || team->t.t_parent->t.t_serialized)))) { __kmp_debug_assert ("team != root->r.r_hot_team || (team->t.t_master_tid == 0 && (team->t.t_parent == root->r.r_root_team || team->t.t_parent->t.t_serialized))" , "openmp/runtime/src/kmp_runtime.cpp", 2235); } | ||||
2235 | team->t.t_parent->t.t_serialized)))if (!(team != root->r.r_hot_team || (team->t.t_master_tid == 0 && (team->t.t_parent == root->r.r_root_team || team->t.t_parent->t.t_serialized)))) { __kmp_debug_assert ("team != root->r.r_hot_team || (team->t.t_master_tid == 0 && (team->t.t_parent == root->r.r_root_team || team->t.t_parent->t.t_serialized))" , "openmp/runtime/src/kmp_runtime.cpp", 2235); }; | ||||
2236 | KMP_MB(); | ||||
2237 | |||||
2238 | /* now, setup the arguments */ | ||||
2239 | argv = (void **)team->t.t_argv; | ||||
2240 | if (ap) { | ||||
2241 | for (i = argc - 1; i >= 0; --i) { | ||||
2242 | void *new_argv = va_arg(kmp_va_deref(ap), void *)__builtin_va_arg((*(ap)), void *); | ||||
2243 | KMP_CHECK_UPDATE(*argv, new_argv)if ((*argv) != (new_argv)) (*argv) = (new_argv); | ||||
2244 | argv++; | ||||
2245 | } | ||||
2246 | } else { | ||||
2247 | for (i = 0; i < argc; ++i) { | ||||
2248 | // Get args from parent team for teams construct | ||||
2249 | KMP_CHECK_UPDATE(argv[i], team->t.t_parent->t.t_argv[i])if ((argv[i]) != (team->t.t_parent->t.t_argv[i])) (argv [i]) = (team->t.t_parent->t.t_argv[i]); | ||||
2250 | } | ||||
2251 | } | ||||
2252 | |||||
2253 | /* now actually fork the threads */ | ||||
2254 | KMP_CHECK_UPDATE(team->t.t_master_active, master_active)if ((team->t.t_master_active) != (master_active)) (team-> t.t_master_active) = (master_active); | ||||
2255 | if (!root->r.r_active) // Only do assignment if it prevents cache ping-pong | ||||
2256 | root->r.r_active = TRUE(!0); | ||||
2257 | |||||
2258 | __kmp_fork_team_threads(root, team, master_th, gtid, !ap); | ||||
2259 | __kmp_setup_icv_copy(team, nthreads, | ||||
2260 | &master_th->th.th_current_task->td_icvs, loc); | ||||
2261 | |||||
2262 | #if OMPT_SUPPORT1 | ||||
2263 | master_th->th.ompt_thread_info.state = ompt_state_work_parallel; | ||||
2264 | #endif | ||||
2265 | |||||
2266 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | ||||
2267 | |||||
2268 | #if USE_ITT_BUILD1 | ||||
2269 | if (team->t.t_active_level == 1 // only report frames at level 1 | ||||
2270 | && !master_th->th.th_teams_microtask) { // not in teams construct | ||||
2271 | #if USE_ITT_NOTIFY1 | ||||
2272 | if ((__itt_frame_submit_v3_ptr__kmp_itt_frame_submit_v3_ptr__3_0 || KMP_ITT_DEBUG0) && | ||||
2273 | (__kmp_forkjoin_frames_mode == 3 || | ||||
2274 | __kmp_forkjoin_frames_mode == 1)) { | ||||
2275 | kmp_uint64 tmp_time = 0; | ||||
2276 | if (__itt_get_timestamp_ptr__kmp_itt_get_timestamp_ptr__3_0) | ||||
2277 | tmp_time = __itt_get_timestamp(!__kmp_itt_get_timestamp_ptr__3_0) ? 0 : __kmp_itt_get_timestamp_ptr__3_0(); | ||||
2278 | // Internal fork - report frame begin | ||||
2279 | master_th->th.th_frame_time = tmp_time; | ||||
2280 | if (__kmp_forkjoin_frames_mode == 3) | ||||
2281 | team->t.t_region_time = tmp_time; | ||||
2282 | } else | ||||
2283 | // only one notification scheme (either "submit" or "forking/joined", not both) | ||||
2284 | #endif /* USE_ITT_NOTIFY */ | ||||
2285 | if ((__itt_frame_begin_v3_ptr__kmp_itt_frame_begin_v3_ptr__3_0 || KMP_ITT_DEBUG0) && | ||||
2286 | __kmp_forkjoin_frames && !__kmp_forkjoin_frames_mode) { | ||||
2287 | // Mark start of "parallel" region for Intel(R) VTune(TM) analyzer. | ||||
2288 | __kmp_itt_region_forking(gtid, team->t.t_nproc, 0); | ||||
2289 | } | ||||
2290 | } | ||||
2291 | #endif /* USE_ITT_BUILD */ | ||||
2292 | |||||
2293 | /* now go on and do the work */ | ||||
2294 | KMP_DEBUG_ASSERT(team == __kmp_threads[gtid]->th.th_team)if (!(team == __kmp_threads[gtid]->th.th_team)) { __kmp_debug_assert ("team == __kmp_threads[gtid]->th.th_team", "openmp/runtime/src/kmp_runtime.cpp" , 2294); }; | ||||
2295 | KMP_MB(); | ||||
2296 | KF_TRACE(10,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_internal_fork : root=%p, team=%p, master_th=%p, gtid=%d\n" , root, team, master_th, gtid); } | ||||
2297 | ("__kmp_internal_fork : root=%p, team=%p, master_th=%p, gtid=%d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_internal_fork : root=%p, team=%p, master_th=%p, gtid=%d\n" , root, team, master_th, gtid); } | ||||
2298 | root, team, master_th, gtid))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_internal_fork : root=%p, team=%p, master_th=%p, gtid=%d\n" , root, team, master_th, gtid); }; | ||||
2299 | |||||
2300 | #if USE_ITT_BUILD1 | ||||
2301 | if (__itt_stack_caller_create_ptr__kmp_itt_stack_caller_create_ptr__3_0) { | ||||
2302 | // create new stack stitching id before entering fork barrier | ||||
2303 | if (!enter_teams) { | ||||
2304 | KMP_DEBUG_ASSERT(team->t.t_stack_id == NULL)if (!(team->t.t_stack_id == __null)) { __kmp_debug_assert( "team->t.t_stack_id == __null", "openmp/runtime/src/kmp_runtime.cpp" , 2304); }; | ||||
2305 | team->t.t_stack_id = __kmp_itt_stack_caller_create(); | ||||
2306 | } else if (parent_team->t.t_serialized) { | ||||
2307 | // keep stack stitching id in the serialized parent_team; | ||||
2308 | // current team will be used for parallel inside the teams; | ||||
2309 | // if parent_team is active, then it already keeps stack stitching id | ||||
2310 | // for the league of teams | ||||
2311 | KMP_DEBUG_ASSERT(parent_team->t.t_stack_id == NULL)if (!(parent_team->t.t_stack_id == __null)) { __kmp_debug_assert ("parent_team->t.t_stack_id == __null", "openmp/runtime/src/kmp_runtime.cpp" , 2311); }; | ||||
2312 | parent_team->t.t_stack_id = __kmp_itt_stack_caller_create(); | ||||
2313 | } | ||||
2314 | } | ||||
2315 | #endif /* USE_ITT_BUILD */ | ||||
2316 | |||||
2317 | // AC: skip __kmp_internal_fork at teams construct, let only primary | ||||
2318 | // threads execute | ||||
2319 | if (ap) { | ||||
2320 | __kmp_internal_fork(loc, gtid, team); | ||||
2321 | KF_TRACE(10, ("__kmp_internal_fork : after : root=%p, team=%p, "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_internal_fork : after : root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, team, master_th, gtid); } | ||||
2322 | "master_th=%p, gtid=%d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_internal_fork : after : root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, team, master_th, gtid); } | ||||
2323 | root, team, master_th, gtid))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_internal_fork : after : root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, team, master_th, gtid); }; | ||||
2324 | } | ||||
2325 | |||||
2326 | if (call_context == fork_context_gnu) { | ||||
2327 | KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: parallel exit T#%d\n" , gtid); }; | ||||
2328 | return TRUE(!0); | ||||
2329 | } | ||||
2330 | |||||
2331 | /* Invoke microtask for PRIMARY thread */ | ||||
2332 | KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n", gtid,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n" , gtid, team->t.t_id, team->t.t_pkfn); } | ||||
2333 | team->t.t_id, team->t.t_pkfn))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n" , gtid, team->t.t_id, team->t.t_pkfn); }; | ||||
2334 | } // END of timer KMP_fork_call block | ||||
2335 | |||||
2336 | #if KMP_STATS_ENABLED0 | ||||
2337 | // If beginning a teams construct, then change thread state | ||||
2338 | stats_state_e previous_state = KMP_GET_THREAD_STATE()((void)0); | ||||
2339 | if (!ap) { | ||||
2340 | KMP_SET_THREAD_STATE(stats_state_e::TEAMS_REGION)((void)0); | ||||
2341 | } | ||||
2342 | #endif | ||||
2343 | |||||
2344 | if (!team->t.t_invoke(gtid)) { | ||||
2345 | KMP_ASSERT2(0, "cannot invoke microtask for PRIMARY thread")if (!(0)) { __kmp_debug_assert(("cannot invoke microtask for PRIMARY thread" ), "openmp/runtime/src/kmp_runtime.cpp", 2345); }; | ||||
2346 | } | ||||
2347 | |||||
2348 | #if KMP_STATS_ENABLED0 | ||||
2349 | // If was beginning of a teams construct, then reset thread state | ||||
2350 | if (!ap) { | ||||
2351 | KMP_SET_THREAD_STATE(previous_state)((void)0); | ||||
2352 | } | ||||
2353 | #endif | ||||
2354 | |||||
2355 | KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n" , gtid, team->t.t_id, team->t.t_pkfn); } | ||||
2356 | team->t.t_id, team->t.t_pkfn))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n" , gtid, team->t.t_id, team->t.t_pkfn); }; | ||||
2357 | KMP_MB(); /* Flush all pending memory write invalidates. */ | ||||
2358 | |||||
2359 | KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: parallel exit T#%d\n" , gtid); }; | ||||
2360 | #if OMPT_SUPPORT1 | ||||
2361 | if (ompt_enabled.enabled) { | ||||
2362 | master_th->th.ompt_thread_info.state = ompt_state_overhead; | ||||
2363 | } | ||||
2364 | #endif | ||||
2365 | |||||
2366 | return TRUE(!0); | ||||
2367 | } | ||||
2368 | |||||
2369 | #if OMPT_SUPPORT1 | ||||
2370 | static inline void __kmp_join_restore_state(kmp_info_t *thread, | ||||
2371 | kmp_team_t *team) { | ||||
2372 | // restore state outside the region | ||||
2373 | thread->th.ompt_thread_info.state = | ||||
2374 | ((team->t.t_serialized) ? ompt_state_work_serial | ||||
2375 | : ompt_state_work_parallel); | ||||
2376 | } | ||||
2377 | |||||
2378 | static inline void __kmp_join_ompt(int gtid, kmp_info_t *thread, | ||||
2379 | kmp_team_t *team, ompt_data_t *parallel_data, | ||||
2380 | int flags, void *codeptr) { | ||||
2381 | ompt_task_info_t *task_info = __ompt_get_task_info_object(0); | ||||
2382 | if (ompt_enabled.ompt_callback_parallel_end) { | ||||
2383 | ompt_callbacks.ompt_callback(ompt_callback_parallel_end)ompt_callback_parallel_end_callback( | ||||
2384 | parallel_data, &(task_info->task_data), flags, codeptr); | ||||
2385 | } | ||||
2386 | |||||
2387 | task_info->frame.enter_frame = ompt_data_none{0}; | ||||
2388 | __kmp_join_restore_state(thread, team); | ||||
2389 | } | ||||
2390 | #endif | ||||
2391 | |||||
2392 | void __kmp_join_call(ident_t *loc, int gtid | ||||
2393 | #if OMPT_SUPPORT1 | ||||
2394 | , | ||||
2395 | enum fork_context_e fork_context | ||||
2396 | #endif | ||||
2397 | , | ||||
2398 | int exit_teams) { | ||||
2399 | KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_join_call)((void)0); | ||||
2400 | kmp_team_t *team; | ||||
2401 | kmp_team_t *parent_team; | ||||
2402 | kmp_info_t *master_th; | ||||
2403 | kmp_root_t *root; | ||||
2404 | int master_active; | ||||
2405 | |||||
2406 | KA_TRACE(20, ("__kmp_join_call: enter T#%d\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: enter T#%d\n" , gtid); }; | ||||
2407 | |||||
2408 | /* setup current data */ | ||||
2409 | master_th = __kmp_threads[gtid]; | ||||
2410 | root = master_th->th.th_root; | ||||
2411 | team = master_th->th.th_team; | ||||
2412 | parent_team = team->t.t_parent; | ||||
2413 | |||||
2414 | master_th->th.th_ident = loc; | ||||
2415 | |||||
2416 | #if OMPT_SUPPORT1 | ||||
2417 | void *team_microtask = (void *)team->t.t_pkfn; | ||||
2418 | // For GOMP interface with serialized parallel, need the | ||||
2419 | // __kmpc_end_serialized_parallel to call hooks for OMPT end-implicit-task | ||||
2420 | // and end-parallel events. | ||||
2421 | if (ompt_enabled.enabled && | ||||
2422 | !(team->t.t_serialized && fork_context == fork_context_gnu)) { | ||||
2423 | master_th->th.ompt_thread_info.state = ompt_state_overhead; | ||||
2424 | } | ||||
2425 | #endif | ||||
2426 | |||||
2427 | #if KMP_DEBUG1 | ||||
2428 | if (__kmp_tasking_mode != tskm_immediate_exec && !exit_teams) { | ||||
2429 | KA_TRACE(20, ("__kmp_join_call: T#%d, old team = %p old task_team = %p, "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: T#%d, old team = %p old task_team = %p, " "th_task_team = %p\n", __kmp_gtid_from_thread(master_th), team , team->t.t_task_team[master_th->th.th_task_state], master_th ->th.th_task_team); } | ||||
2430 | "th_task_team = %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: T#%d, old team = %p old task_team = %p, " "th_task_team = %p\n", __kmp_gtid_from_thread(master_th), team , team->t.t_task_team[master_th->th.th_task_state], master_th ->th.th_task_team); } | ||||
2431 | __kmp_gtid_from_thread(master_th), team,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: T#%d, old team = %p old task_team = %p, " "th_task_team = %p\n", __kmp_gtid_from_thread(master_th), team , team->t.t_task_team[master_th->th.th_task_state], master_th ->th.th_task_team); } | ||||
2432 | team->t.t_task_team[master_th->th.th_task_state],if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: T#%d, old team = %p old task_team = %p, " "th_task_team = %p\n", __kmp_gtid_from_thread(master_th), team , team->t.t_task_team[master_th->th.th_task_state], master_th ->th.th_task_team); } | ||||
2433 | master_th->th.th_task_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: T#%d, old team = %p old task_team = %p, " "th_task_team = %p\n", __kmp_gtid_from_thread(master_th), team , team->t.t_task_team[master_th->th.th_task_state], master_th ->th.th_task_team); }; | ||||
2434 | KMP_DEBUG_ASSERT(master_th->th.th_task_team ==if (!(master_th->th.th_task_team == team->t.t_task_team [master_th->th.th_task_state])) { __kmp_debug_assert("master_th->th.th_task_team == team->t.t_task_team[master_th->th.th_task_state]" , "openmp/runtime/src/kmp_runtime.cpp", 2435); } | ||||
2435 | team->t.t_task_team[master_th->th.th_task_state])if (!(master_th->th.th_task_team == team->t.t_task_team [master_th->th.th_task_state])) { __kmp_debug_assert("master_th->th.th_task_team == team->t.t_task_team[master_th->th.th_task_state]" , "openmp/runtime/src/kmp_runtime.cpp", 2435); }; | ||||
2436 | } | ||||
2437 | #endif | ||||
2438 | |||||
2439 | if (team->t.t_serialized) { | ||||
2440 | if (master_th->th.th_teams_microtask) { | ||||
2441 | // We are in teams construct | ||||
2442 | int level = team->t.t_level; | ||||
2443 | int tlevel = master_th->th.th_teams_level; | ||||
2444 | if (level == tlevel) { | ||||
2445 | // AC: we haven't incremented it earlier at start of teams construct, | ||||
2446 | // so do it here - at the end of teams construct | ||||
2447 | team->t.t_level++; | ||||
2448 | } else if (level == tlevel + 1) { | ||||
2449 | // AC: we are exiting parallel inside teams, need to increment | ||||
2450 | // serialization in order to restore it in the next call to | ||||
2451 | // __kmpc_end_serialized_parallel | ||||
2452 | team->t.t_serialized++; | ||||
2453 | } | ||||
2454 | } | ||||
2455 | __kmpc_end_serialized_parallel(loc, gtid); | ||||
2456 | |||||
2457 | #if OMPT_SUPPORT1 | ||||
2458 | if (ompt_enabled.enabled) { | ||||
2459 | if (fork_context == fork_context_gnu) { | ||||
2460 | __ompt_lw_taskteam_unlink(master_th); | ||||
2461 | } | ||||
2462 | __kmp_join_restore_state(master_th, parent_team); | ||||
2463 | } | ||||
2464 | #endif | ||||
2465 | |||||
2466 | return; | ||||
2467 | } | ||||
2468 | |||||
2469 | master_active = team->t.t_master_active; | ||||
2470 | |||||
2471 | if (!exit_teams) { | ||||
2472 | // AC: No barrier for internal teams at exit from teams construct. | ||||
2473 | // But there is barrier for external team (league). | ||||
2474 | __kmp_internal_join(loc, gtid, team); | ||||
2475 | #if USE_ITT_BUILD1 | ||||
2476 | if (__itt_stack_caller_create_ptr__kmp_itt_stack_caller_create_ptr__3_0) { | ||||
2477 | KMP_DEBUG_ASSERT(team->t.t_stack_id != NULL)if (!(team->t.t_stack_id != __null)) { __kmp_debug_assert( "team->t.t_stack_id != __null", "openmp/runtime/src/kmp_runtime.cpp" , 2477); }; | ||||
2478 | // destroy the stack stitching id after join barrier | ||||
2479 | __kmp_itt_stack_caller_destroy((__itt_caller)team->t.t_stack_id); | ||||
2480 | team->t.t_stack_id = NULL__null; | ||||
2481 | } | ||||
2482 | #endif | ||||
2483 | } else { | ||||
2484 | master_th->th.th_task_state = | ||||
2485 | 0; // AC: no tasking in teams (out of any parallel) | ||||
2486 | #if USE_ITT_BUILD1 | ||||
2487 | if (__itt_stack_caller_create_ptr__kmp_itt_stack_caller_create_ptr__3_0 && parent_team->t.t_serialized) { | ||||
2488 | KMP_DEBUG_ASSERT(parent_team->t.t_stack_id != NULL)if (!(parent_team->t.t_stack_id != __null)) { __kmp_debug_assert ("parent_team->t.t_stack_id != __null", "openmp/runtime/src/kmp_runtime.cpp" , 2488); }; | ||||
2489 | // destroy the stack stitching id on exit from the teams construct | ||||
2490 | // if parent_team is active, then the id will be destroyed later on | ||||
2491 | // by master of the league of teams | ||||
2492 | __kmp_itt_stack_caller_destroy((__itt_caller)parent_team->t.t_stack_id); | ||||
2493 | parent_team->t.t_stack_id = NULL__null; | ||||
2494 | } | ||||
2495 | #endif | ||||
2496 | } | ||||
2497 | |||||
2498 | KMP_MB(); | ||||
2499 | |||||
2500 | #if OMPT_SUPPORT1 | ||||
2501 | ompt_data_t *parallel_data = &(team->t.ompt_team_info.parallel_data); | ||||
2502 | void *codeptr = team->t.ompt_team_info.master_return_address; | ||||
2503 | #endif | ||||
2504 | |||||
2505 | #if USE_ITT_BUILD1 | ||||
2506 | // Mark end of "parallel" region for Intel(R) VTune(TM) analyzer. | ||||
2507 | if (team->t.t_active_level == 1 && | ||||
2508 | (!master_th->th.th_teams_microtask || /* not in teams construct */ | ||||
2509 | master_th->th.th_teams_size.nteams == 1)) { | ||||
2510 | master_th->th.th_ident = loc; | ||||
2511 | // only one notification scheme (either "submit" or "forking/joined", not | ||||
2512 | // both) | ||||
2513 | if ((__itt_frame_submit_v3_ptr__kmp_itt_frame_submit_v3_ptr__3_0 || KMP_ITT_DEBUG0) && | ||||
2514 | __kmp_forkjoin_frames_mode == 3) | ||||
2515 | __kmp_itt_frame_submit(gtid, team->t.t_region_time, | ||||
2516 | master_th->th.th_frame_time, 0, loc, | ||||
2517 | master_th->th.th_team_nproc, 1); | ||||
2518 | else if ((__itt_frame_end_v3_ptr__kmp_itt_frame_end_v3_ptr__3_0 || KMP_ITT_DEBUG0) && | ||||
2519 | !__kmp_forkjoin_frames_mode && __kmp_forkjoin_frames) | ||||
2520 | __kmp_itt_region_joined(gtid); | ||||
2521 | } // active_level == 1 | ||||
2522 | #endif /* USE_ITT_BUILD */ | ||||
2523 | |||||
2524 | #if KMP_AFFINITY_SUPPORTED1 | ||||
2525 | if (!exit_teams) { | ||||
2526 | // Restore master thread's partition. | ||||
2527 | master_th->th.th_first_place = team->t.t_first_place; | ||||
2528 | master_th->th.th_last_place = team->t.t_last_place; | ||||
2529 | } | ||||
2530 | #endif // KMP_AFFINITY_SUPPORTED | ||||
2531 | |||||
2532 | if (master_th->th.th_teams_microtask && !exit_teams && | ||||
2533 | team->t.t_pkfn != (microtask_t)__kmp_teams_master && | ||||
2534 | team->t.t_level == master_th->th.th_teams_level + 1) { | ||||
2535 | // AC: We need to leave the team structure intact at the end of parallel | ||||
2536 | // inside the teams construct, so that at the next parallel same (hot) team | ||||
2537 | // works, only adjust nesting levels | ||||
2538 | #if OMPT_SUPPORT1 | ||||
2539 | ompt_data_t ompt_parallel_data = ompt_data_none{0}; | ||||
2540 | if (ompt_enabled.enabled) { | ||||
2541 | ompt_task_info_t *task_info = __ompt_get_task_info_object(0); | ||||
2542 | if (ompt_enabled.ompt_callback_implicit_task) { | ||||
2543 | int ompt_team_size = team->t.t_nproc; | ||||
2544 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | ||||
2545 | ompt_scope_end, NULL__null, &(task_info->task_data), ompt_team_size, | ||||
2546 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num, ompt_task_implicit); | ||||
2547 | } | ||||
2548 | task_info->frame.exit_frame = ompt_data_none{0}; | ||||
2549 | task_info->task_data = ompt_data_none{0}; | ||||
2550 | ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th)(&(master_th->th.th_team->t.ompt_team_info.parallel_data )); | ||||
2551 | __ompt_lw_taskteam_unlink(master_th); | ||||
2552 | } | ||||
2553 | #endif | ||||
2554 | /* Decrement our nested depth level */ | ||||
2555 | team->t.t_level--; | ||||
2556 | team->t.t_active_level--; | ||||
2557 | KMP_ATOMIC_DEC(&root->r.r_in_parallel)(&root->r.r_in_parallel)->fetch_sub(1, std::memory_order_acq_rel ); | ||||
2558 | |||||
2559 | // Restore number of threads in the team if needed. This code relies on | ||||
2560 | // the proper adjustment of th_teams_size.nth after the fork in | ||||
2561 | // __kmp_teams_master on each teams primary thread in the case that | ||||
2562 | // __kmp_reserve_threads reduced it. | ||||
2563 | if (master_th->th.th_team_nproc < master_th->th.th_teams_size.nth) { | ||||
2564 | int old_num = master_th->th.th_team_nproc; | ||||
2565 | int new_num = master_th->th.th_teams_size.nth; | ||||
2566 | kmp_info_t **other_threads = team->t.t_threads; | ||||
2567 | team->t.t_nproc = new_num; | ||||
2568 | for (int i = 0; i < old_num; ++i) { | ||||
2569 | other_threads[i]->th.th_team_nproc = new_num; | ||||
2570 | } | ||||
2571 | // Adjust states of non-used threads of the team | ||||
2572 | for (int i = old_num; i < new_num; ++i) { | ||||
2573 | // Re-initialize thread's barrier data. | ||||
2574 | KMP_DEBUG_ASSERT(other_threads[i])if (!(other_threads[i])) { __kmp_debug_assert("other_threads[i]" , "openmp/runtime/src/kmp_runtime.cpp", 2574); }; | ||||
2575 | kmp_balign_t *balign = other_threads[i]->th.th_bar; | ||||
2576 | for (int b = 0; b < bs_last_barrier; ++b) { | ||||
2577 | balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; | ||||
2578 | KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG)if (!(balign[b].bb.wait_flag != 2)) { __kmp_debug_assert("balign[b].bb.wait_flag != 2" , "openmp/runtime/src/kmp_runtime.cpp", 2578); }; | ||||
2579 | #if USE_DEBUGGER0 | ||||
2580 | balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; | ||||
2581 | #endif | ||||
2582 | } | ||||
2583 | if (__kmp_tasking_mode != tskm_immediate_exec) { | ||||
2584 | // Synchronize thread's task state | ||||
2585 | other_threads[i]->th.th_task_state = master_th->th.th_task_state; | ||||
2586 | } | ||||
2587 | } | ||||
2588 | } | ||||
2589 | |||||
2590 | #if OMPT_SUPPORT1 | ||||
2591 | if (ompt_enabled.enabled) { | ||||
2592 | __kmp_join_ompt(gtid, master_th, parent_team, &ompt_parallel_data, | ||||
2593 | OMPT_INVOKER(fork_context)((fork_context == fork_context_gnu) ? ompt_parallel_invoker_program : ompt_parallel_invoker_runtime) | ompt_parallel_team, codeptr); | ||||
2594 | } | ||||
2595 | #endif | ||||
2596 | |||||
2597 | return; | ||||
2598 | } | ||||
2599 | |||||
2600 | /* do cleanup and restore the parent team */ | ||||
2601 | master_th->th.th_info.ds.ds_tid = team->t.t_master_tid; | ||||
2602 | master_th->th.th_local.this_construct = team->t.t_master_this_cons; | ||||
2603 | |||||
2604 | master_th->th.th_dispatch = &parent_team->t.t_dispatch[team->t.t_master_tid]; | ||||
2605 | |||||
2606 | /* jc: The following lock has instructions with REL and ACQ semantics, | ||||
2607 | separating the parallel user code called in this parallel region | ||||
2608 | from the serial user code called after this function returns. */ | ||||
2609 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); | ||||
2610 | |||||
2611 | if (!master_th->th.th_teams_microtask || | ||||
2612 | team->t.t_level > master_th->th.th_teams_level) { | ||||
2613 | /* Decrement our nested depth level */ | ||||
2614 | KMP_ATOMIC_DEC(&root->r.r_in_parallel)(&root->r.r_in_parallel)->fetch_sub(1, std::memory_order_acq_rel ); | ||||
2615 | } | ||||
2616 | KMP_DEBUG_ASSERT(root->r.r_in_parallel >= 0)if (!(root->r.r_in_parallel >= 0)) { __kmp_debug_assert ("root->r.r_in_parallel >= 0", "openmp/runtime/src/kmp_runtime.cpp" , 2616); }; | ||||
2617 | |||||
2618 | #if OMPT_SUPPORT1 | ||||
2619 | if (ompt_enabled.enabled) { | ||||
2620 | ompt_task_info_t *task_info = __ompt_get_task_info_object(0); | ||||
2621 | if (ompt_enabled.ompt_callback_implicit_task) { | ||||
2622 | int flags = (team_microtask == (void *)__kmp_teams_master) | ||||
2623 | ? ompt_task_initial | ||||
2624 | : ompt_task_implicit; | ||||
2625 | int ompt_team_size = (flags == ompt_task_initial) ? 0 : team->t.t_nproc; | ||||
2626 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | ||||
2627 | ompt_scope_end, NULL__null, &(task_info->task_data), ompt_team_size, | ||||
2628 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num, flags); | ||||
2629 | } | ||||
2630 | task_info->frame.exit_frame = ompt_data_none{0}; | ||||
2631 | task_info->task_data = ompt_data_none{0}; | ||||
2632 | } | ||||
2633 | #endif | ||||
2634 | |||||
2635 | KF_TRACE(10, ("__kmp_join_call1: T#%d, this_thread=%p team=%p\n", 0,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_join_call1: T#%d, this_thread=%p team=%p\n" , 0, master_th, team); } | ||||
2636 | master_th, team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_join_call1: T#%d, this_thread=%p team=%p\n" , 0, master_th, team); }; | ||||
2637 | __kmp_pop_current_task_from_thread(master_th); | ||||
2638 | |||||
2639 | master_th->th.th_def_allocator = team->t.t_def_allocator; | ||||
2640 | |||||
2641 | #if OMPD_SUPPORT1 | ||||
2642 | if (ompd_state & OMPD_ENABLE_BP0x1) | ||||
2643 | ompd_bp_parallel_end(); | ||||
2644 | #endif | ||||
2645 | updateHWFPControl(team); | ||||
2646 | |||||
2647 | if (root->r.r_active != master_active) | ||||
2648 | root->r.r_active = master_active; | ||||
2649 | |||||
2650 | __kmp_free_team(root, team USE_NESTED_HOT_ARG(, master_th | ||||
2651 | master_th), master_th); // this will free worker threads | ||||
2652 | |||||
2653 | /* this race was fun to find. make sure the following is in the critical | ||||
2654 | region otherwise assertions may fail occasionally since the old team may be | ||||
2655 | reallocated and the hierarchy appears inconsistent. it is actually safe to | ||||
2656 | run and won't cause any bugs, but will cause those assertion failures. it's | ||||
2657 | only one deref&assign so might as well put this in the critical region */ | ||||
2658 | master_th->th.th_team = parent_team; | ||||
2659 | master_th->th.th_team_nproc = parent_team->t.t_nproc; | ||||
2660 | master_th->th.th_team_master = parent_team->t.t_threads[0]; | ||||
2661 | master_th->th.th_team_serialized = parent_team->t.t_serialized; | ||||
2662 | |||||
2663 | /* restore serialized team, if need be */ | ||||
2664 | if (parent_team->t.t_serialized && | ||||
2665 | parent_team != master_th->th.th_serial_team && | ||||
2666 | parent_team != root->r.r_root_team) { | ||||
2667 | __kmp_free_team(root, | ||||
2668 | master_th->th.th_serial_team USE_NESTED_HOT_ARG(NULL), __null); | ||||
2669 | master_th->th.th_serial_team = parent_team; | ||||
2670 | } | ||||
2671 | |||||
2672 | if (__kmp_tasking_mode != tskm_immediate_exec) { | ||||
2673 | if (master_th->th.th_task_state_top > | ||||
2674 | 0) { // Restore task state from memo stack | ||||
2675 | KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack)if (!(master_th->th.th_task_state_memo_stack)) { __kmp_debug_assert ("master_th->th.th_task_state_memo_stack", "openmp/runtime/src/kmp_runtime.cpp" , 2675); }; | ||||
2676 | // Remember primary thread's state if we re-use this nested hot team | ||||
2677 | master_th->th.th_task_state_memo_stack[master_th->th.th_task_state_top] = | ||||
2678 | master_th->th.th_task_state; | ||||
2679 | --master_th->th.th_task_state_top; // pop | ||||
2680 | // Now restore state at this level | ||||
2681 | master_th->th.th_task_state = | ||||
2682 | master_th->th | ||||
2683 | .th_task_state_memo_stack[master_th->th.th_task_state_top]; | ||||
2684 | } else if (team != root->r.r_hot_team) { | ||||
2685 | // Reset the task state of primary thread if we are not hot team because | ||||
2686 | // in this case all the worker threads will be free, and their task state | ||||
2687 | // will be reset. If not reset the primary's, the task state will be | ||||
2688 | // inconsistent. | ||||
2689 | master_th->th.th_task_state = 0; | ||||
2690 | } | ||||
2691 | // Copy the task team from the parent team to the primary thread | ||||
2692 | master_th->th.th_task_team = | ||||
2693 | parent_team->t.t_task_team[master_th->th.th_task_state]; | ||||
2694 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: Primary T#%d restoring task_team %p, team %p\n" , __kmp_gtid_from_thread(master_th), master_th->th.th_task_team , parent_team); } | ||||
2695 | ("__kmp_join_call: Primary T#%d restoring task_team %p, team %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: Primary T#%d restoring task_team %p, team %p\n" , __kmp_gtid_from_thread(master_th), master_th->th.th_task_team , parent_team); } | ||||
2696 | __kmp_gtid_from_thread(master_th), master_th->th.th_task_team,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: Primary T#%d restoring task_team %p, team %p\n" , __kmp_gtid_from_thread(master_th), master_th->th.th_task_team , parent_team); } | ||||
2697 | parent_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: Primary T#%d restoring task_team %p, team %p\n" , __kmp_gtid_from_thread(master_th), master_th->th.th_task_team , parent_team); }; | ||||
2698 | } | ||||
2699 | |||||
2700 | // TODO: GEH - cannot do this assertion because root thread not set up as | ||||
2701 | // executing | ||||
2702 | // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 0 ); | ||||
2703 | master_th->th.th_current_task->td_flags.executing = 1; | ||||
2704 | |||||
2705 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | ||||
2706 | |||||
2707 | #if KMP_AFFINITY_SUPPORTED1 | ||||
2708 | if (master_th->th.th_team->t.t_level == 0 && __kmp_affinity.flags.reset) { | ||||
2709 | __kmp_reset_root_init_mask(gtid); | ||||
2710 | } | ||||
2711 | #endif | ||||
2712 | #if OMPT_SUPPORT1 | ||||
2713 | int flags = | ||||
2714 | OMPT_INVOKER(fork_context)((fork_context == fork_context_gnu) ? ompt_parallel_invoker_program : ompt_parallel_invoker_runtime) | | ||||
2715 | ((team_microtask == (void *)__kmp_teams_master) ? ompt_parallel_league | ||||
2716 | : ompt_parallel_team); | ||||
2717 | if (ompt_enabled.enabled) { | ||||
2718 | __kmp_join_ompt(gtid, master_th, parent_team, parallel_data, flags, | ||||
2719 | codeptr); | ||||
2720 | } | ||||
2721 | #endif | ||||
2722 | |||||
2723 | KMP_MB(); | ||||
2724 | KA_TRACE(20, ("__kmp_join_call: exit T#%d\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: exit T#%d\n" , gtid); }; | ||||
2725 | } | ||||
2726 | |||||
2727 | /* Check whether we should push an internal control record onto the | ||||
2728 | serial team stack. If so, do it. */ | ||||
2729 | void __kmp_save_internal_controls(kmp_info_t *thread) { | ||||
2730 | |||||
2731 | if (thread->th.th_team != thread->th.th_serial_team) { | ||||
2732 | return; | ||||
2733 | } | ||||
2734 | if (thread->th.th_team->t.t_serialized > 1) { | ||||
2735 | int push = 0; | ||||
2736 | |||||
2737 | if (thread->th.th_team->t.t_control_stack_top == NULL__null) { | ||||
2738 | push = 1; | ||||
2739 | } else { | ||||
2740 | if (thread->th.th_team->t.t_control_stack_top->serial_nesting_level != | ||||
2741 | thread->th.th_team->t.t_serialized) { | ||||
2742 | push = 1; | ||||
2743 | } | ||||
2744 | } | ||||
2745 | if (push) { /* push a record on the serial team's stack */ | ||||
2746 | kmp_internal_control_t *control = | ||||
2747 | (kmp_internal_control_t *)__kmp_allocate(___kmp_allocate((sizeof(kmp_internal_control_t)), "openmp/runtime/src/kmp_runtime.cpp" , 2748) | ||||
2748 | sizeof(kmp_internal_control_t))___kmp_allocate((sizeof(kmp_internal_control_t)), "openmp/runtime/src/kmp_runtime.cpp" , 2748); | ||||
2749 | |||||
2750 | copy_icvs(control, &thread->th.th_current_task->td_icvs); | ||||
2751 | |||||
2752 | control->serial_nesting_level = thread->th.th_team->t.t_serialized; | ||||
2753 | |||||
2754 | control->next = thread->th.th_team->t.t_control_stack_top; | ||||
2755 | thread->th.th_team->t.t_control_stack_top = control; | ||||
2756 | } | ||||
2757 | } | ||||
2758 | } | ||||
2759 | |||||
2760 | /* Changes set_nproc */ | ||||
2761 | void __kmp_set_num_threads(int new_nth, int gtid) { | ||||
2762 | kmp_info_t *thread; | ||||
2763 | kmp_root_t *root; | ||||
2764 | |||||
2765 | KF_TRACE(10, ("__kmp_set_num_threads: new __kmp_nth = %d\n", new_nth))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_num_threads: new __kmp_nth = %d\n" , new_nth); }; | ||||
2766 | KMP_DEBUG_ASSERT(__kmp_init_serial)if (!(__kmp_init_serial)) { __kmp_debug_assert("__kmp_init_serial" , "openmp/runtime/src/kmp_runtime.cpp", 2766); }; | ||||
2767 | |||||
2768 | if (new_nth < 1) | ||||
2769 | new_nth = 1; | ||||
2770 | else if (new_nth > __kmp_max_nth) | ||||
2771 | new_nth = __kmp_max_nth; | ||||
2772 | |||||
2773 | KMP_COUNT_VALUE(OMP_set_numthreads, new_nth)((void)0); | ||||
2774 | thread = __kmp_threads[gtid]; | ||||
2775 | if (thread->th.th_current_task->td_icvs.nproc == new_nth) | ||||
2776 | return; // nothing to do | ||||
2777 | |||||
2778 | __kmp_save_internal_controls(thread); | ||||
2779 | |||||
2780 | set__nproc(thread, new_nth)(((thread)->th.th_current_task->td_icvs.nproc) = (new_nth )); | ||||
2781 | |||||
2782 | // If this omp_set_num_threads() call will cause the hot team size to be | ||||
2783 | // reduced (in the absence of a num_threads clause), then reduce it now, | ||||
2784 | // rather than waiting for the next parallel region. | ||||
2785 | root = thread->th.th_root; | ||||
2786 | if (__kmp_init_parallel && (!root->r.r_active) && | ||||
2787 | (root->r.r_hot_team->t.t_nproc > new_nth) | ||||
2788 | #if KMP_NESTED_HOT_TEAMS1 | ||||
2789 | && __kmp_hot_teams_max_level && !__kmp_hot_teams_mode | ||||
2790 | #endif | ||||
2791 | ) { | ||||
2792 | kmp_team_t *hot_team = root->r.r_hot_team; | ||||
2793 | int f; | ||||
2794 | |||||
2795 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); | ||||
2796 | |||||
2797 | if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | ||||
2798 | __kmp_resize_dist_barrier(hot_team, hot_team->t.t_nproc, new_nth); | ||||
2799 | } | ||||
2800 | // Release the extra threads we don't need any more. | ||||
2801 | for (f = new_nth; f < hot_team->t.t_nproc; f++) { | ||||
2802 | KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL)if (!(hot_team->t.t_threads[f] != __null)) { __kmp_debug_assert ("hot_team->t.t_threads[f] != __null", "openmp/runtime/src/kmp_runtime.cpp" , 2802); }; | ||||
2803 | if (__kmp_tasking_mode != tskm_immediate_exec) { | ||||
2804 | // When decreasing team size, threads no longer in the team should unref | ||||
2805 | // task team. | ||||
2806 | hot_team->t.t_threads[f]->th.th_task_team = NULL__null; | ||||
2807 | } | ||||
2808 | __kmp_free_thread(hot_team->t.t_threads[f]); | ||||
2809 | hot_team->t.t_threads[f] = NULL__null; | ||||
2810 | } | ||||
2811 | hot_team->t.t_nproc = new_nth; | ||||
2812 | #if KMP_NESTED_HOT_TEAMS1 | ||||
2813 | if (thread->th.th_hot_teams) { | ||||
2814 | KMP_DEBUG_ASSERT(hot_team == thread->th.th_hot_teams[0].hot_team)if (!(hot_team == thread->th.th_hot_teams[0].hot_team)) { __kmp_debug_assert ("hot_team == thread->th.th_hot_teams[0].hot_team", "openmp/runtime/src/kmp_runtime.cpp" , 2814); }; | ||||
2815 | thread->th.th_hot_teams[0].hot_team_nth = new_nth; | ||||
2816 | } | ||||
2817 | #endif | ||||
2818 | |||||
2819 | if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | ||||
2820 | hot_team->t.b->update_num_threads(new_nth); | ||||
2821 | __kmp_add_threads_to_team(hot_team, new_nth); | ||||
2822 | } | ||||
2823 | |||||
2824 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | ||||
2825 | |||||
2826 | // Update the t_nproc field in the threads that are still active. | ||||
2827 | for (f = 0; f < new_nth; f++) { | ||||
2828 | KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL)if (!(hot_team->t.t_threads[f] != __null)) { __kmp_debug_assert ("hot_team->t.t_threads[f] != __null", "openmp/runtime/src/kmp_runtime.cpp" , 2828); }; | ||||
2829 | hot_team->t.t_threads[f]->th.th_team_nproc = new_nth; | ||||
2830 | } | ||||
2831 | // Special flag in case omp_set_num_threads() call | ||||
2832 | hot_team->t.t_size_changed = -1; | ||||
2833 | } | ||||
2834 | } | ||||
2835 | |||||
2836 | /* Changes max_active_levels */ | ||||
2837 | void __kmp_set_max_active_levels(int gtid, int max_active_levels) { | ||||
2838 | kmp_info_t *thread; | ||||
2839 | |||||
2840 | KF_TRACE(10, ("__kmp_set_max_active_levels: new max_active_levels for thread "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_max_active_levels: new max_active_levels for thread " "%d = (%d)\n", gtid, max_active_levels); } | ||||
2841 | "%d = (%d)\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_max_active_levels: new max_active_levels for thread " "%d = (%d)\n", gtid, max_active_levels); } | ||||
2842 | gtid, max_active_levels))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_max_active_levels: new max_active_levels for thread " "%d = (%d)\n", gtid, max_active_levels); }; | ||||
2843 | KMP_DEBUG_ASSERT(__kmp_init_serial)if (!(__kmp_init_serial)) { __kmp_debug_assert("__kmp_init_serial" , "openmp/runtime/src/kmp_runtime.cpp", 2843); }; | ||||
2844 | |||||
2845 | // validate max_active_levels | ||||
2846 | if (max_active_levels < 0) { | ||||
2847 | KMP_WARNING(ActiveLevelsNegative, max_active_levels)__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_ActiveLevelsNegative , max_active_levels), __kmp_msg_null); | ||||
2848 | // We ignore this call if the user has specified a negative value. | ||||
2849 | // The current setting won't be changed. The last valid setting will be | ||||
2850 | // used. A warning will be issued (if warnings are allowed as controlled by | ||||
2851 | // the KMP_WARNINGS env var). | ||||
2852 | KF_TRACE(10, ("__kmp_set_max_active_levels: the call is ignored: new "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_max_active_levels: the call is ignored: new " "max_active_levels for thread %d = (%d)\n", gtid, max_active_levels ); } | ||||
2853 | "max_active_levels for thread %d = (%d)\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_max_active_levels: the call is ignored: new " "max_active_levels for thread %d = (%d)\n", gtid, max_active_levels ); } | ||||
2854 | gtid, max_active_levels))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_max_active_levels: the call is ignored: new " "max_active_levels for thread %d = (%d)\n", gtid, max_active_levels ); }; | ||||
2855 | return; | ||||
2856 | } | ||||
2857 | if (max_active_levels <= KMP_MAX_ACTIVE_LEVELS_LIMIT2147483647) { | ||||
2858 | // it's OK, the max_active_levels is within the valid range: [ 0; | ||||
2859 | // KMP_MAX_ACTIVE_LEVELS_LIMIT ] | ||||
2860 | // We allow a zero value. (implementation defined behavior) | ||||
2861 | } else { | ||||
2862 | KMP_WARNING(ActiveLevelsExceedLimit, max_active_levels,__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_ActiveLevelsExceedLimit , max_active_levels, 2147483647), __kmp_msg_null) | ||||
2863 | KMP_MAX_ACTIVE_LEVELS_LIMIT)__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_ActiveLevelsExceedLimit , max_active_levels, 2147483647), __kmp_msg_null); | ||||
2864 | max_active_levels = KMP_MAX_ACTIVE_LEVELS_LIMIT2147483647; | ||||
2865 | // Current upper limit is MAX_INT. (implementation defined behavior) | ||||
2866 | // If the input exceeds the upper limit, we correct the input to be the | ||||
2867 | // upper limit. (implementation defined behavior) | ||||
2868 | // Actually, the flow should never get here until we use MAX_INT limit. | ||||
2869 | } | ||||
2870 | KF_TRACE(10, ("__kmp_set_max_active_levels: after validation: new "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_max_active_levels: after validation: new " "max_active_levels for thread %d = (%d)\n", gtid, max_active_levels ); } | ||||
2871 | "max_active_levels for thread %d = (%d)\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_max_active_levels: after validation: new " "max_active_levels for thread %d = (%d)\n", gtid, max_active_levels ); } | ||||
2872 | gtid, max_active_levels))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_max_active_levels: after validation: new " "max_active_levels for thread %d = (%d)\n", gtid, max_active_levels ); }; | ||||
2873 | |||||
2874 | thread = __kmp_threads[gtid]; | ||||
2875 | |||||
2876 | __kmp_save_internal_controls(thread); | ||||
2877 | |||||
2878 | set__max_active_levels(thread, max_active_levels)(((thread)->th.th_current_task->td_icvs.max_active_levels ) = (max_active_levels)); | ||||
2879 | } | ||||
2880 | |||||
2881 | /* Gets max_active_levels */ | ||||
2882 | int __kmp_get_max_active_levels(int gtid) { | ||||
2883 | kmp_info_t *thread; | ||||
2884 | |||||
2885 | KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d\n", gtid))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_get_max_active_levels: thread %d\n" , gtid); }; | ||||
2886 | KMP_DEBUG_ASSERT(__kmp_init_serial)if (!(__kmp_init_serial)) { __kmp_debug_assert("__kmp_init_serial" , "openmp/runtime/src/kmp_runtime.cpp", 2886); }; | ||||
2887 | |||||
2888 | thread = __kmp_threads[gtid]; | ||||
2889 | KMP_DEBUG_ASSERT(thread->th.th_current_task)if (!(thread->th.th_current_task)) { __kmp_debug_assert("thread->th.th_current_task" , "openmp/runtime/src/kmp_runtime.cpp", 2889); }; | ||||
2890 | KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d, curtask=%p, "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_get_max_active_levels: thread %d, curtask=%p, " "curtask_maxaclevel=%d\n", gtid, thread->th.th_current_task , thread->th.th_current_task->td_icvs.max_active_levels ); } | ||||
2891 | "curtask_maxaclevel=%d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_get_max_active_levels: thread %d, curtask=%p, " "curtask_maxaclevel=%d\n", gtid, thread->th.th_current_task , thread->th.th_current_task->td_icvs.max_active_levels ); } | ||||
2892 | gtid, thread->th.th_current_task,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_get_max_active_levels: thread %d, curtask=%p, " "curtask_maxaclevel=%d\n", gtid, thread->th.th_current_task , thread->th.th_current_task->td_icvs.max_active_levels ); } | ||||
2893 | thread->th.th_current_task->td_icvs.max_active_levels))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_get_max_active_levels: thread %d, curtask=%p, " "curtask_maxaclevel=%d\n", gtid, thread->th.th_current_task , thread->th.th_current_task->td_icvs.max_active_levels ); }; | ||||
2894 | return thread->th.th_current_task->td_icvs.max_active_levels; | ||||
2895 | } | ||||
2896 | |||||
2897 | // nteams-var per-device ICV | ||||
2898 | void __kmp_set_num_teams(int num_teams) { | ||||
2899 | if (num_teams > 0) | ||||
2900 | __kmp_nteams = num_teams; | ||||
2901 | } | ||||
2902 | int __kmp_get_max_teams(void) { return __kmp_nteams; } | ||||
2903 | // teams-thread-limit-var per-device ICV | ||||
2904 | void __kmp_set_teams_thread_limit(int limit) { | ||||
2905 | if (limit > 0) | ||||
2906 | __kmp_teams_thread_limit = limit; | ||||
2907 | } | ||||
2908 | int __kmp_get_teams_thread_limit(void) { return __kmp_teams_thread_limit; } | ||||
2909 | |||||
2910 | KMP_BUILD_ASSERT(sizeof(kmp_sched_t) == sizeof(int))static_assert(sizeof(kmp_sched_t) == sizeof(int), "Build condition error" ); | ||||
2911 | KMP_BUILD_ASSERT(sizeof(enum sched_type) == sizeof(int))static_assert(sizeof(enum sched_type) == sizeof(int), "Build condition error" ); | ||||
2912 | |||||
2913 | /* Changes def_sched_var ICV values (run-time schedule kind and chunk) */ | ||||
2914 | void __kmp_set_schedule(int gtid, kmp_sched_t kind, int chunk) { | ||||
2915 | kmp_info_t *thread; | ||||
2916 | kmp_sched_t orig_kind; | ||||
2917 | // kmp_team_t *team; | ||||
2918 | |||||
2919 | KF_TRACE(10, ("__kmp_set_schedule: new schedule for thread %d = (%d, %d)\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_schedule: new schedule for thread %d = (%d, %d)\n" , gtid, (int)kind, chunk); } | ||||
2920 | gtid, (int)kind, chunk))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_schedule: new schedule for thread %d = (%d, %d)\n" , gtid, (int)kind, chunk); }; | ||||
2921 | KMP_DEBUG_ASSERT(__kmp_init_serial)if (!(__kmp_init_serial)) { __kmp_debug_assert("__kmp_init_serial" , "openmp/runtime/src/kmp_runtime.cpp", 2921); }; | ||||
2922 | |||||
2923 | // Check if the kind parameter is valid, correct if needed. | ||||
2924 | // Valid parameters should fit in one of two intervals - standard or extended: | ||||
2925 | // <lower>, <valid>, <upper_std>, <lower_ext>, <valid>, <upper> | ||||
2926 | // 2008-01-25: 0, 1 - 4, 5, 100, 101 - 102, 103 | ||||
2927 | orig_kind = kind; | ||||
2928 | kind = __kmp_sched_without_mods(kind); | ||||
2929 | |||||
2930 | if (kind <= kmp_sched_lower || kind >= kmp_sched_upper || | ||||
2931 | (kind <= kmp_sched_lower_ext && kind >= kmp_sched_upper_std)) { | ||||
2932 | // TODO: Hint needs attention in case we change the default schedule. | ||||
2933 | __kmp_msg(kmp_ms_warning, KMP_MSG(ScheduleKindOutOfRange, kind)__kmp_msg_format(kmp_i18n_msg_ScheduleKindOutOfRange, kind), | ||||
2934 | KMP_HNT(DefaultScheduleKindUsed, "static, no chunk")__kmp_msg_format(kmp_i18n_hnt_DefaultScheduleKindUsed, "static, no chunk" ), | ||||
2935 | __kmp_msg_null); | ||||
2936 | kind = kmp_sched_default; | ||||
2937 | chunk = 0; // ignore chunk value in case of bad kind | ||||
2938 | } | ||||
2939 | |||||
2940 | thread = __kmp_threads[gtid]; | ||||
2941 | |||||
2942 | __kmp_save_internal_controls(thread); | ||||
2943 | |||||
2944 | if (kind < kmp_sched_upper_std) { | ||||
2945 | if (kind == kmp_sched_static && chunk < KMP_DEFAULT_CHUNK1) { | ||||
2946 | // differ static chunked vs. unchunked: chunk should be invalid to | ||||
2947 | // indicate unchunked schedule (which is the default) | ||||
2948 | thread->th.th_current_task->td_icvs.sched.r_sched_type = kmp_sch_static; | ||||
2949 | } else { | ||||
2950 | thread->th.th_current_task->td_icvs.sched.r_sched_type = | ||||
2951 | __kmp_sch_map[kind - kmp_sched_lower - 1]; | ||||
2952 | } | ||||
2953 | } else { | ||||
2954 | // __kmp_sch_map[ kind - kmp_sched_lower_ext + kmp_sched_upper_std - | ||||
2955 | // kmp_sched_lower - 2 ]; | ||||
2956 | thread->th.th_current_task->td_icvs.sched.r_sched_type = | ||||
2957 | __kmp_sch_map[kind - kmp_sched_lower_ext + kmp_sched_upper_std - | ||||
2958 | kmp_sched_lower - 2]; | ||||
2959 | } | ||||
2960 | __kmp_sched_apply_mods_intkind( | ||||
2961 | orig_kind, &(thread->th.th_current_task->td_icvs.sched.r_sched_type)); | ||||
2962 | if (kind == kmp_sched_auto || chunk < 1) { | ||||
2963 | // ignore parameter chunk for schedule auto | ||||
2964 | thread->th.th_current_task->td_icvs.sched.chunk = KMP_DEFAULT_CHUNK1; | ||||
2965 | } else { | ||||
2966 | thread->th.th_current_task->td_icvs.sched.chunk = chunk; | ||||
2967 | } | ||||
2968 | } | ||||
2969 | |||||
2970 | /* Gets def_sched_var ICV values */ | ||||
2971 | void __kmp_get_schedule(int gtid, kmp_sched_t *kind, int *chunk) { | ||||
2972 | kmp_info_t *thread; | ||||
2973 | enum sched_type th_type; | ||||
2974 | |||||
2975 | KF_TRACE(10, ("__kmp_get_schedule: thread %d\n", gtid))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_get_schedule: thread %d\n" , gtid); }; | ||||
2976 | KMP_DEBUG_ASSERT(__kmp_init_serial)if (!(__kmp_init_serial)) { __kmp_debug_assert("__kmp_init_serial" , "openmp/runtime/src/kmp_runtime.cpp", 2976); }; | ||||
2977 | |||||
2978 | thread = __kmp_threads[gtid]; | ||||
2979 | |||||
2980 | th_type = thread->th.th_current_task->td_icvs.sched.r_sched_type; | ||||
2981 | switch (SCHEDULE_WITHOUT_MODIFIERS(th_type)(enum sched_type)( (th_type) & ~(kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic))) { | ||||
2982 | case kmp_sch_static: | ||||
2983 | case kmp_sch_static_greedy: | ||||
2984 | case kmp_sch_static_balanced: | ||||
2985 | *kind = kmp_sched_static; | ||||
2986 | __kmp_sched_apply_mods_stdkind(kind, th_type); | ||||
2987 | *chunk = 0; // chunk was not set, try to show this fact via zero value | ||||
2988 | return; | ||||
2989 | case kmp_sch_static_chunked: | ||||
2990 | *kind = kmp_sched_static; | ||||
2991 | break; | ||||
2992 | case kmp_sch_dynamic_chunked: | ||||
2993 | *kind = kmp_sched_dynamic; | ||||
2994 | break; | ||||
2995 | case kmp_sch_guided_chunked: | ||||
2996 | case kmp_sch_guided_iterative_chunked: | ||||
2997 | case kmp_sch_guided_analytical_chunked: | ||||
2998 | *kind = kmp_sched_guided; | ||||
2999 | break; | ||||
3000 | case kmp_sch_auto: | ||||
3001 | *kind = kmp_sched_auto; | ||||
3002 | break; | ||||
3003 | case kmp_sch_trapezoidal: | ||||
3004 | *kind = kmp_sched_trapezoidal; | ||||
3005 | break; | ||||
3006 | #if KMP_STATIC_STEAL_ENABLED1 | ||||
3007 | case kmp_sch_static_steal: | ||||
3008 | *kind = kmp_sched_static_steal; | ||||
3009 | break; | ||||
3010 | #endif | ||||
3011 | default: | ||||
3012 | KMP_FATAL(UnknownSchedulingType, th_type)__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_UnknownSchedulingType , th_type), __kmp_msg_null); | ||||
3013 | } | ||||
3014 | |||||
3015 | __kmp_sched_apply_mods_stdkind(kind, th_type); | ||||
3016 | *chunk = thread->th.th_current_task->td_icvs.sched.chunk; | ||||
3017 | } | ||||
3018 | |||||
3019 | int __kmp_get_ancestor_thread_num(int gtid, int level) { | ||||
3020 | |||||
3021 | int ii, dd; | ||||
3022 | kmp_team_t *team; | ||||
3023 | kmp_info_t *thr; | ||||
3024 | |||||
3025 | KF_TRACE(10, ("__kmp_get_ancestor_thread_num: thread %d %d\n", gtid, level))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_get_ancestor_thread_num: thread %d %d\n" , gtid, level); }; | ||||
3026 | KMP_DEBUG_ASSERT(__kmp_init_serial)if (!(__kmp_init_serial)) { __kmp_debug_assert("__kmp_init_serial" , "openmp/runtime/src/kmp_runtime.cpp", 3026); }; | ||||
3027 | |||||
3028 | // validate level | ||||
3029 | if (level == 0) | ||||
3030 | return 0; | ||||
3031 | if (level < 0) | ||||
3032 | return -1; | ||||
3033 | thr = __kmp_threads[gtid]; | ||||
3034 | team = thr->th.th_team; | ||||
3035 | ii = team->t.t_level; | ||||
3036 | if (level > ii) | ||||
3037 | return -1; | ||||
3038 | |||||
3039 | if (thr->th.th_teams_microtask) { | ||||
3040 | // AC: we are in teams region where multiple nested teams have same level | ||||
3041 | int tlevel = thr->th.th_teams_level; // the level of the teams construct | ||||
3042 | if (level <= | ||||
3043 | tlevel) { // otherwise usual algorithm works (will not touch the teams) | ||||
3044 | KMP_DEBUG_ASSERT(ii >= tlevel)if (!(ii >= tlevel)) { __kmp_debug_assert("ii >= tlevel" , "openmp/runtime/src/kmp_runtime.cpp", 3044); }; | ||||
3045 | // AC: As we need to pass by the teams league, we need to artificially | ||||
3046 | // increase ii | ||||
3047 | if (ii == tlevel) { | ||||
3048 | ii += 2; // three teams have same level | ||||
3049 | } else { | ||||
3050 | ii++; // two teams have same level | ||||
3051 | } | ||||
3052 | } | ||||
3053 | } | ||||
3054 | |||||
3055 | if (ii == level) | ||||
3056 | return __kmp_tid_from_gtid(gtid); | ||||
3057 | |||||
3058 | dd = team->t.t_serialized; | ||||
3059 | level++; | ||||
3060 | while (ii > level) { | ||||
3061 | for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) { | ||||
3062 | } | ||||
3063 | if ((team->t.t_serialized) && (!dd)) { | ||||
3064 | team = team->t.t_parent; | ||||
3065 | continue; | ||||
3066 | } | ||||
3067 | if (ii > level) { | ||||
3068 | team = team->t.t_parent; | ||||
3069 | dd = team->t.t_serialized; | ||||
3070 | ii--; | ||||
3071 | } | ||||
3072 | } | ||||
3073 | |||||
3074 | return (dd > 1) ? (0) : (team->t.t_master_tid); | ||||
3075 | } | ||||
3076 | |||||
3077 | int __kmp_get_team_size(int gtid, int level) { | ||||
3078 | |||||
3079 | int ii, dd; | ||||
3080 | kmp_team_t *team; | ||||
3081 | kmp_info_t *thr; | ||||
3082 | |||||
3083 | KF_TRACE(10, ("__kmp_get_team_size: thread %d %d\n", gtid, level))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_get_team_size: thread %d %d\n" , gtid, level); }; | ||||
3084 | KMP_DEBUG_ASSERT(__kmp_init_serial)if (!(__kmp_init_serial)) { __kmp_debug_assert("__kmp_init_serial" , "openmp/runtime/src/kmp_runtime.cpp", 3084); }; | ||||
3085 | |||||
3086 | // validate level | ||||
3087 | if (level == 0) | ||||
3088 | return 1; | ||||
3089 | if (level < 0) | ||||
3090 | return -1; | ||||
3091 | thr = __kmp_threads[gtid]; | ||||
3092 | team = thr->th.th_team; | ||||
3093 | ii = team->t.t_level; | ||||
3094 | if (level > ii) | ||||
3095 | return -1; | ||||
3096 | |||||
3097 | if (thr->th.th_teams_microtask) { | ||||
3098 | // AC: we are in teams region where multiple nested teams have same level | ||||
3099 | int tlevel = thr->th.th_teams_level; // the level of the teams construct | ||||
3100 | if (level <= | ||||
3101 | tlevel) { // otherwise usual algorithm works (will not touch the teams) | ||||
3102 | KMP_DEBUG_ASSERT(ii >= tlevel)if (!(ii >= tlevel)) { __kmp_debug_assert("ii >= tlevel" , "openmp/runtime/src/kmp_runtime.cpp", 3102); }; | ||||
3103 | // AC: As we need to pass by the teams league, we need to artificially | ||||
3104 | // increase ii | ||||
3105 | if (ii == tlevel) { | ||||
3106 | ii += 2; // three teams have same level | ||||
3107 | } else { | ||||
3108 | ii++; // two teams have same level | ||||
3109 | } | ||||
3110 | } | ||||
3111 | } | ||||
3112 | |||||
3113 | while (ii > level) { | ||||
3114 | for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) { | ||||
3115 | } | ||||
3116 | if (team->t.t_serialized && (!dd)) { | ||||
3117 | team = team->t.t_parent; | ||||
3118 | continue; | ||||
3119 | } | ||||
3120 | if (ii > level) { | ||||
3121 | team = team->t.t_parent; | ||||
3122 | ii--; | ||||
3123 | } | ||||
3124 | } | ||||
3125 | |||||
3126 | return team->t.t_nproc; | ||||
3127 | } | ||||
3128 | |||||
3129 | kmp_r_sched_t __kmp_get_schedule_global() { | ||||
3130 | // This routine created because pairs (__kmp_sched, __kmp_chunk) and | ||||
3131 | // (__kmp_static, __kmp_guided) may be changed by kmp_set_defaults | ||||
3132 | // independently. So one can get the updated schedule here. | ||||
3133 | |||||
3134 | kmp_r_sched_t r_sched; | ||||
3135 | |||||
3136 | // create schedule from 4 globals: __kmp_sched, __kmp_chunk, __kmp_static, | ||||
3137 | // __kmp_guided. __kmp_sched should keep original value, so that user can set | ||||
3138 | // KMP_SCHEDULE multiple times, and thus have different run-time schedules in | ||||
3139 | // different roots (even in OMP 2.5) | ||||
3140 | enum sched_type s = SCHEDULE_WITHOUT_MODIFIERS(__kmp_sched)(enum sched_type)( (__kmp_sched) & ~(kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)); | ||||
3141 | enum sched_type sched_modifiers = SCHEDULE_GET_MODIFIERS(__kmp_sched)((enum sched_type)( (__kmp_sched) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic))); | ||||
3142 | if (s == kmp_sch_static) { | ||||
3143 | // replace STATIC with more detailed schedule (balanced or greedy) | ||||
3144 | r_sched.r_sched_type = __kmp_static; | ||||
3145 | } else if (s == kmp_sch_guided_chunked) { | ||||
3146 | // replace GUIDED with more detailed schedule (iterative or analytical) | ||||
3147 | r_sched.r_sched_type = __kmp_guided; | ||||
3148 | } else { // (STATIC_CHUNKED), or (DYNAMIC_CHUNKED), or other | ||||
3149 | r_sched.r_sched_type = __kmp_sched; | ||||
3150 | } | ||||
3151 | SCHEDULE_SET_MODIFIERS(r_sched.r_sched_type, sched_modifiers)(r_sched.r_sched_type = (enum sched_type)((kmp_int32)r_sched. r_sched_type | (kmp_int32)sched_modifiers)); | ||||
3152 | |||||
3153 | if (__kmp_chunk < KMP_DEFAULT_CHUNK1) { | ||||
3154 | // __kmp_chunk may be wrong here (if it was not ever set) | ||||
3155 | r_sched.chunk = KMP_DEFAULT_CHUNK1; | ||||
3156 | } else { | ||||
3157 | r_sched.chunk = __kmp_chunk; | ||||
3158 | } | ||||
3159 | |||||
3160 | return r_sched; | ||||
3161 | } | ||||
3162 | |||||
3163 | /* Allocate (realloc == FALSE) * or reallocate (realloc == TRUE) | ||||
3164 | at least argc number of *t_argv entries for the requested team. */ | ||||
3165 | static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team, int realloc) { | ||||
3166 | |||||
3167 | KMP_DEBUG_ASSERT(team)if (!(team)) { __kmp_debug_assert("team", "openmp/runtime/src/kmp_runtime.cpp" , 3167); }; | ||||
3168 | if (!realloc || argc > team->t.t_max_argc) { | ||||
3169 | |||||
3170 | KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: needed entries=%d, "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_alloc_argv_entries: team %d: needed entries=%d, " "current entries=%d\n", team->t.t_id, argc, (realloc) ? team ->t.t_max_argc : 0); } | ||||
3171 | "current entries=%d\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_alloc_argv_entries: team %d: needed entries=%d, " "current entries=%d\n", team->t.t_id, argc, (realloc) ? team ->t.t_max_argc : 0); } | ||||
3172 | team->t.t_id, argc, (realloc) ? team->t.t_max_argc : 0))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_alloc_argv_entries: team %d: needed entries=%d, " "current entries=%d\n", team->t.t_id, argc, (realloc) ? team ->t.t_max_argc : 0); }; | ||||
3173 | /* if previously allocated heap space for args, free them */ | ||||
3174 | if (realloc && team->t.t_argv != &team->t.t_inline_argv[0]) | ||||
3175 | __kmp_free((void *)team->t.t_argv)___kmp_free(((void *)team->t.t_argv), "openmp/runtime/src/kmp_runtime.cpp" , 3175); | ||||
3176 | |||||
3177 | if (argc <= KMP_INLINE_ARGV_ENTRIES(int)((4 * 64 - ((3 * (sizeof(void *)) + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + sizeof(kmp_int16) + sizeof(kmp_uint32)) % 64)) / (sizeof(void *)))) { | ||||
3178 | /* use unused space in the cache line for arguments */ | ||||
3179 | team->t.t_max_argc = KMP_INLINE_ARGV_ENTRIES(int)((4 * 64 - ((3 * (sizeof(void *)) + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + sizeof(kmp_int16) + sizeof(kmp_uint32)) % 64)) / (sizeof(void *))); | ||||
3180 | KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: inline allocate %d "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_alloc_argv_entries: team %d: inline allocate %d " "argv entries\n", team->t.t_id, team->t.t_max_argc); } | ||||
3181 | "argv entries\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_alloc_argv_entries: team %d: inline allocate %d " "argv entries\n", team->t.t_id, team->t.t_max_argc); } | ||||
3182 | team->t.t_id, team->t.t_max_argc))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_alloc_argv_entries: team %d: inline allocate %d " "argv entries\n", team->t.t_id, team->t.t_max_argc); }; | ||||
3183 | team->t.t_argv = &team->t.t_inline_argv[0]; | ||||
3184 | if (__kmp_storage_map) { | ||||
3185 | __kmp_print_storage_map_gtid( | ||||
3186 | -1, &team->t.t_inline_argv[0], | ||||
3187 | &team->t.t_inline_argv[KMP_INLINE_ARGV_ENTRIES(int)((4 * 64 - ((3 * (sizeof(void *)) + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + sizeof(kmp_int16) + sizeof(kmp_uint32)) % 64)) / (sizeof(void *)))], | ||||
3188 | (sizeof(void *) * KMP_INLINE_ARGV_ENTRIES(int)((4 * 64 - ((3 * (sizeof(void *)) + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + sizeof(kmp_int16) + sizeof(kmp_uint32)) % 64)) / (sizeof(void *)))), "team_%d.t_inline_argv", | ||||
3189 | team->t.t_id); | ||||
3190 | } | ||||
3191 | } else { | ||||
3192 | /* allocate space for arguments in the heap */ | ||||
3193 | team->t.t_max_argc = (argc <= (KMP_MIN_MALLOC_ARGV_ENTRIES100 >> 1)) | ||||
3194 | ? KMP_MIN_MALLOC_ARGV_ENTRIES100 | ||||
3195 | : 2 * argc; | ||||
3196 | KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: dynamic allocate %d "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_alloc_argv_entries: team %d: dynamic allocate %d " "argv entries\n", team->t.t_id, team->t.t_max_argc); } | ||||
3197 | "argv entries\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_alloc_argv_entries: team %d: dynamic allocate %d " "argv entries\n", team->t.t_id, team->t.t_max_argc); } | ||||
3198 | team->t.t_id, team->t.t_max_argc))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_alloc_argv_entries: team %d: dynamic allocate %d " "argv entries\n", team->t.t_id, team->t.t_max_argc); }; | ||||
3199 | team->t.t_argv = | ||||
3200 | (void **)__kmp_page_allocate(sizeof(void *) * team->t.t_max_argc)___kmp_page_allocate((sizeof(void *) * team->t.t_max_argc) , "openmp/runtime/src/kmp_runtime.cpp", 3200); | ||||
3201 | if (__kmp_storage_map) { | ||||
3202 | __kmp_print_storage_map_gtid(-1, &team->t.t_argv[0], | ||||
3203 | &team->t.t_argv[team->t.t_max_argc], | ||||
3204 | sizeof(void *) * team->t.t_max_argc, | ||||
3205 | "team_%d.t_argv", team->t.t_id); | ||||
3206 | } | ||||
3207 | } | ||||
3208 | } | ||||
3209 | } | ||||
3210 | |||||
3211 | static void __kmp_allocate_team_arrays(kmp_team_t *team, int max_nth) { | ||||
3212 | int i; | ||||
3213 | int num_disp_buff = max_nth > 1 ? __kmp_dispatch_num_buffers : 2; | ||||
3214 | team->t.t_threads = | ||||
3215 | (kmp_info_t **)__kmp_allocate(sizeof(kmp_info_t *) * max_nth)___kmp_allocate((sizeof(kmp_info_t *) * max_nth), "openmp/runtime/src/kmp_runtime.cpp" , 3215); | ||||
3216 | team->t.t_disp_buffer = (dispatch_shared_info_t *)__kmp_allocate(___kmp_allocate((sizeof(dispatch_shared_info_t) * num_disp_buff ), "openmp/runtime/src/kmp_runtime.cpp", 3217) | ||||
3217 | sizeof(dispatch_shared_info_t) * num_disp_buff)___kmp_allocate((sizeof(dispatch_shared_info_t) * num_disp_buff ), "openmp/runtime/src/kmp_runtime.cpp", 3217); | ||||
3218 | team->t.t_dispatch = | ||||
3219 | (kmp_disp_t *)__kmp_allocate(sizeof(kmp_disp_t) * max_nth)___kmp_allocate((sizeof(kmp_disp_t) * max_nth), "openmp/runtime/src/kmp_runtime.cpp" , 3219); | ||||
3220 | team->t.t_implicit_task_taskdata = | ||||
3221 | (kmp_taskdata_t *)__kmp_allocate(sizeof(kmp_taskdata_t) * max_nth)___kmp_allocate((sizeof(kmp_taskdata_t) * max_nth), "openmp/runtime/src/kmp_runtime.cpp" , 3221); | ||||
3222 | team->t.t_max_nproc = max_nth; | ||||
3223 | |||||
3224 | /* setup dispatch buffers */ | ||||
3225 | for (i = 0; i < num_disp_buff; ++i) { | ||||
3226 | team->t.t_disp_buffer[i].buffer_index = i; | ||||
3227 | team->t.t_disp_buffer[i].doacross_buf_idx = i; | ||||
3228 | } | ||||
3229 | } | ||||
3230 | |||||
3231 | static void __kmp_free_team_arrays(kmp_team_t *team) { | ||||
3232 | /* Note: this does not free the threads in t_threads (__kmp_free_threads) */ | ||||
3233 | int i; | ||||
3234 | for (i = 0; i < team->t.t_max_nproc; ++i) { | ||||
3235 | if (team->t.t_dispatch[i].th_disp_buffer != NULL__null) { | ||||
3236 | __kmp_free(team->t.t_dispatch[i].th_disp_buffer)___kmp_free((team->t.t_dispatch[i].th_disp_buffer), "openmp/runtime/src/kmp_runtime.cpp" , 3236); | ||||
3237 | team->t.t_dispatch[i].th_disp_buffer = NULL__null; | ||||
3238 | } | ||||
3239 | } | ||||
3240 | #if KMP_USE_HIER_SCHED0 | ||||
3241 | __kmp_dispatch_free_hierarchies(team); | ||||
3242 | #endif | ||||
3243 | __kmp_free(team->t.t_threads)___kmp_free((team->t.t_threads), "openmp/runtime/src/kmp_runtime.cpp" , 3243); | ||||
3244 | __kmp_free(team->t.t_disp_buffer)___kmp_free((team->t.t_disp_buffer), "openmp/runtime/src/kmp_runtime.cpp" , 3244); | ||||
3245 | __kmp_free(team->t.t_dispatch)___kmp_free((team->t.t_dispatch), "openmp/runtime/src/kmp_runtime.cpp" , 3245); | ||||
3246 | __kmp_free(team->t.t_implicit_task_taskdata)___kmp_free((team->t.t_implicit_task_taskdata), "openmp/runtime/src/kmp_runtime.cpp" , 3246); | ||||
3247 | team->t.t_threads = NULL__null; | ||||
3248 | team->t.t_disp_buffer = NULL__null; | ||||
3249 | team->t.t_dispatch = NULL__null; | ||||
3250 | team->t.t_implicit_task_taskdata = 0; | ||||
3251 | } | ||||
3252 | |||||
3253 | static void __kmp_reallocate_team_arrays(kmp_team_t *team, int max_nth) { | ||||
3254 | kmp_info_t **oldThreads = team->t.t_threads; | ||||
3255 | |||||
3256 | __kmp_free(team->t.t_disp_buffer)___kmp_free((team->t.t_disp_buffer), "openmp/runtime/src/kmp_runtime.cpp" , 3256); | ||||
3257 | __kmp_free(team->t.t_dispatch)___kmp_free((team->t.t_dispatch), "openmp/runtime/src/kmp_runtime.cpp" , 3257); | ||||
3258 | __kmp_free(team->t.t_implicit_task_taskdata)___kmp_free((team->t.t_implicit_task_taskdata), "openmp/runtime/src/kmp_runtime.cpp" , 3258); | ||||
3259 | __kmp_allocate_team_arrays(team, max_nth); | ||||
3260 | |||||
3261 | KMP_MEMCPYmemcpy(team->t.t_threads, oldThreads, | ||||
3262 | team->t.t_nproc * sizeof(kmp_info_t *)); | ||||
3263 | |||||
3264 | __kmp_free(oldThreads)___kmp_free((oldThreads), "openmp/runtime/src/kmp_runtime.cpp" , 3264); | ||||
3265 | } | ||||
3266 | |||||
3267 | static kmp_internal_control_t __kmp_get_global_icvs(void) { | ||||
3268 | |||||
3269 | kmp_r_sched_t r_sched = | ||||
3270 | __kmp_get_schedule_global(); // get current state of scheduling globals | ||||
3271 | |||||
3272 | KMP_DEBUG_ASSERT(__kmp_nested_proc_bind.used > 0)if (!(__kmp_nested_proc_bind.used > 0)) { __kmp_debug_assert ("__kmp_nested_proc_bind.used > 0", "openmp/runtime/src/kmp_runtime.cpp" , 3272); }; | ||||
3273 | |||||
3274 | kmp_internal_control_t g_icvs = { | ||||
3275 | 0, // int serial_nesting_level; //corresponds to value of th_team_serialized | ||||
3276 | (kmp_int8)__kmp_global.g.g_dynamic, // internal control for dynamic | ||||
3277 | // adjustment of threads (per thread) | ||||
3278 | (kmp_int8)__kmp_env_blocktime, // int bt_set; //internal control for | ||||
3279 | // whether blocktime is explicitly set | ||||
3280 | __kmp_dflt_blocktime, // int blocktime; //internal control for blocktime | ||||
3281 | #if KMP_USE_MONITOR | ||||
3282 | __kmp_bt_intervals, // int bt_intervals; //internal control for blocktime | ||||
3283 | // intervals | ||||
3284 | #endif | ||||
3285 | __kmp_dflt_team_nth, // int nproc; //internal control for # of threads for | ||||
3286 | // next parallel region (per thread) | ||||
3287 | // (use a max ub on value if __kmp_parallel_initialize not called yet) | ||||
3288 | __kmp_cg_max_nth, // int thread_limit; | ||||
3289 | __kmp_dflt_max_active_levels, // int max_active_levels; //internal control | ||||
3290 | // for max_active_levels | ||||
3291 | r_sched, // kmp_r_sched_t sched; //internal control for runtime schedule | ||||
3292 | // {sched,chunk} pair | ||||
3293 | __kmp_nested_proc_bind.bind_types[0], | ||||
3294 | __kmp_default_device, | ||||
3295 | NULL__null // struct kmp_internal_control *next; | ||||
3296 | }; | ||||
3297 | |||||
3298 | return g_icvs; | ||||
3299 | } | ||||
3300 | |||||
3301 | static kmp_internal_control_t __kmp_get_x_global_icvs(const kmp_team_t *team) { | ||||
3302 | |||||
3303 | kmp_internal_control_t gx_icvs; | ||||
3304 | gx_icvs.serial_nesting_level = | ||||
3305 | 0; // probably =team->t.t_serial like in save_inter_controls | ||||
3306 | copy_icvs(&gx_icvs, &team->t.t_threads[0]->th.th_current_task->td_icvs); | ||||
3307 | gx_icvs.next = NULL__null; | ||||
3308 | |||||
3309 | return gx_icvs; | ||||
3310 | } | ||||
3311 | |||||
3312 | static void __kmp_initialize_root(kmp_root_t *root) { | ||||
3313 | int f; | ||||
3314 | kmp_team_t *root_team; | ||||
3315 | kmp_team_t *hot_team; | ||||
3316 | int hot_team_max_nth; | ||||
3317 | kmp_r_sched_t r_sched = | ||||
3318 | __kmp_get_schedule_global(); // get current state of scheduling globals | ||||
3319 | kmp_internal_control_t r_icvs = __kmp_get_global_icvs(); | ||||
3320 | KMP_DEBUG_ASSERT(root)if (!(root)) { __kmp_debug_assert("root", "openmp/runtime/src/kmp_runtime.cpp" , 3320); }; | ||||
3321 | KMP_ASSERT(!root->r.r_begin)if (!(!root->r.r_begin)) { __kmp_debug_assert("!root->r.r_begin" , "openmp/runtime/src/kmp_runtime.cpp", 3321); }; | ||||
3322 | |||||
3323 | /* setup the root state structure */ | ||||
3324 | __kmp_init_lock(&root->r.r_begin_lock); | ||||
3325 | root->r.r_begin = FALSE0; | ||||
3326 | root->r.r_active = FALSE0; | ||||
3327 | root->r.r_in_parallel = 0; | ||||
3328 | root->r.r_blocktime = __kmp_dflt_blocktime; | ||||
3329 | #if KMP_AFFINITY_SUPPORTED1 | ||||
3330 | root->r.r_affinity_assigned = FALSE0; | ||||
3331 | #endif | ||||
3332 | |||||
3333 | /* setup the root team for this task */ | ||||
3334 | /* allocate the root team structure */ | ||||
3335 | KF_TRACE(10, ("__kmp_initialize_root: before root_team\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_root: before root_team\n" ); }; | ||||
3336 | |||||
3337 | root_team = | ||||
3338 | __kmp_allocate_team(root, | ||||
3339 | 1, // new_nproc | ||||
3340 | 1, // max_nproc | ||||
3341 | #if OMPT_SUPPORT1 | ||||
3342 | ompt_data_none{0}, // root parallel id | ||||
3343 | #endif | ||||
3344 | __kmp_nested_proc_bind.bind_types[0], &r_icvs, | ||||
3345 | 0 // argc | ||||
3346 | USE_NESTED_HOT_ARG(NULL), __null // primary thread is unknown | ||||
3347 | ); | ||||
3348 | #if USE_DEBUGGER0 | ||||
3349 | // Non-NULL value should be assigned to make the debugger display the root | ||||
3350 | // team. | ||||
3351 | TCW_SYNC_PTR(root_team->t.t_pkfn, (microtask_t)(~0))((root_team->t.t_pkfn)) = (((microtask_t)(~0))); | ||||
3352 | #endif | ||||
3353 | |||||
3354 | KF_TRACE(10, ("__kmp_initialize_root: after root_team = %p\n", root_team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_root: after root_team = %p\n" , root_team); }; | ||||
3355 | |||||
3356 | root->r.r_root_team = root_team; | ||||
3357 | root_team->t.t_control_stack_top = NULL__null; | ||||
3358 | |||||
3359 | /* initialize root team */ | ||||
3360 | root_team->t.t_threads[0] = NULL__null; | ||||
3361 | root_team->t.t_nproc = 1; | ||||
3362 | root_team->t.t_serialized = 1; | ||||
3363 | // TODO???: root_team->t.t_max_active_levels = __kmp_dflt_max_active_levels; | ||||
3364 | root_team->t.t_sched.sched = r_sched.sched; | ||||
3365 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n" , root_team->t.t_id, 0, 0); } | ||||
3366 | 20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n" , root_team->t.t_id, 0, 0); } | ||||
3367 | ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n" , root_team->t.t_id, 0, 0); } | ||||
3368 | root_team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n" , root_team->t.t_id, 0, 0); }; | ||||
3369 | |||||
3370 | /* setup the hot team for this task */ | ||||
3371 | /* allocate the hot team structure */ | ||||
3372 | KF_TRACE(10, ("__kmp_initialize_root: before hot_team\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_root: before hot_team\n" ); }; | ||||
3373 | |||||
3374 | hot_team = | ||||
3375 | __kmp_allocate_team(root, | ||||
3376 | 1, // new_nproc | ||||
3377 | __kmp_dflt_team_nth_ub * 2, // max_nproc | ||||
3378 | #if OMPT_SUPPORT1 | ||||
3379 | ompt_data_none{0}, // root parallel id | ||||
3380 | #endif | ||||
3381 | __kmp_nested_proc_bind.bind_types[0], &r_icvs, | ||||
3382 | 0 // argc | ||||
3383 | USE_NESTED_HOT_ARG(NULL), __null // primary thread is unknown | ||||
3384 | ); | ||||
3385 | KF_TRACE(10, ("__kmp_initialize_root: after hot_team = %p\n", hot_team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_root: after hot_team = %p\n" , hot_team); }; | ||||
3386 | |||||
3387 | root->r.r_hot_team = hot_team; | ||||
3388 | root_team->t.t_control_stack_top = NULL__null; | ||||
3389 | |||||
3390 | /* first-time initialization */ | ||||
3391 | hot_team->t.t_parent = root_team; | ||||
3392 | |||||
3393 | /* initialize hot team */ | ||||
3394 | hot_team_max_nth = hot_team->t.t_max_nproc; | ||||
3395 | for (f = 0; f < hot_team_max_nth; ++f) { | ||||
3396 | hot_team->t.t_threads[f] = NULL__null; | ||||
3397 | } | ||||
3398 | hot_team->t.t_nproc = 1; | ||||
3399 | // TODO???: hot_team->t.t_max_active_levels = __kmp_dflt_max_active_levels; | ||||
3400 | hot_team->t.t_sched.sched = r_sched.sched; | ||||
3401 | hot_team->t.t_size_changed = 0; | ||||
3402 | } | ||||
3403 | |||||
3404 | #ifdef KMP_DEBUG1 | ||||
3405 | |||||
3406 | typedef struct kmp_team_list_item { | ||||
3407 | kmp_team_p const *entry; | ||||
3408 | struct kmp_team_list_item *next; | ||||
3409 | } kmp_team_list_item_t; | ||||
3410 | typedef kmp_team_list_item_t *kmp_team_list_t; | ||||
3411 | |||||
3412 | static void __kmp_print_structure_team_accum( // Add team to list of teams. | ||||
3413 | kmp_team_list_t list, // List of teams. | ||||
3414 | kmp_team_p const *team // Team to add. | ||||
3415 | ) { | ||||
3416 | |||||
3417 | // List must terminate with item where both entry and next are NULL. | ||||
3418 | // Team is added to the list only once. | ||||
3419 | // List is sorted in ascending order by team id. | ||||
3420 | // Team id is *not* a key. | ||||
3421 | |||||
3422 | kmp_team_list_t l; | ||||
3423 | |||||
3424 | KMP_DEBUG_ASSERT(list != NULL)if (!(list != __null)) { __kmp_debug_assert("list != __null", "openmp/runtime/src/kmp_runtime.cpp", 3424); }; | ||||
3425 | if (team == NULL__null) { | ||||
3426 | return; | ||||
3427 | } | ||||
3428 | |||||
3429 | __kmp_print_structure_team_accum(list, team->t.t_parent); | ||||
3430 | __kmp_print_structure_team_accum(list, team->t.t_next_pool); | ||||
3431 | |||||
3432 | // Search list for the team. | ||||
3433 | l = list; | ||||
3434 | while (l->next != NULL__null && l->entry != team) { | ||||
3435 | l = l->next; | ||||
3436 | } | ||||
3437 | if (l->next != NULL__null) { | ||||
3438 | return; // Team has been added before, exit. | ||||
3439 | } | ||||
3440 | |||||
3441 | // Team is not found. Search list again for insertion point. | ||||
3442 | l = list; | ||||
3443 | while (l->next != NULL__null && l->entry->t.t_id <= team->t.t_id) { | ||||
3444 | l = l->next; | ||||
3445 | } | ||||
3446 | |||||
3447 | // Insert team. | ||||
3448 | { | ||||
3449 | kmp_team_list_item_t *item = (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(malloc(sizeof(kmp_team_list_item_t)) | ||||
3450 | sizeof(kmp_team_list_item_t))malloc(sizeof(kmp_team_list_item_t)); | ||||
3451 | *item = *l; | ||||
3452 | l->entry = team; | ||||
3453 | l->next = item; | ||||
3454 | } | ||||
3455 | } | ||||
3456 | |||||
3457 | static void __kmp_print_structure_team(char const *title, kmp_team_p const *team | ||||
3458 | |||||
3459 | ) { | ||||
3460 | __kmp_printf("%s", title); | ||||
3461 | if (team != NULL__null) { | ||||
3462 | __kmp_printf("%2x %p\n", team->t.t_id, team); | ||||
3463 | } else { | ||||
3464 | __kmp_printf(" - (nil)\n"); | ||||
3465 | } | ||||
3466 | } | ||||
3467 | |||||
3468 | static void __kmp_print_structure_thread(char const *title, | ||||
3469 | kmp_info_p const *thread) { | ||||
3470 | __kmp_printf("%s", title); | ||||
3471 | if (thread != NULL__null) { | ||||
3472 | __kmp_printf("%2d %p\n", thread->th.th_info.ds.ds_gtid, thread); | ||||
3473 | } else { | ||||
3474 | __kmp_printf(" - (nil)\n"); | ||||
3475 | } | ||||
3476 | } | ||||
3477 | |||||
3478 | void __kmp_print_structure(void) { | ||||
3479 | |||||
3480 | kmp_team_list_t list; | ||||
3481 | |||||
3482 | // Initialize list of teams. | ||||
3483 | list = | ||||
3484 | (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(sizeof(kmp_team_list_item_t))malloc(sizeof(kmp_team_list_item_t)); | ||||
3485 | list->entry = NULL__null; | ||||
3486 | list->next = NULL__null; | ||||
3487 | |||||
3488 | __kmp_printf("\n------------------------------\nGlobal Thread " | ||||
3489 | "Table\n------------------------------\n"); | ||||
3490 | { | ||||
3491 | int gtid; | ||||
3492 | for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) { | ||||
3493 | __kmp_printf("%2d", gtid); | ||||
3494 | if (__kmp_threads != NULL__null) { | ||||
3495 | __kmp_printf(" %p", __kmp_threads[gtid]); | ||||
3496 | } | ||||
3497 | if (__kmp_root != NULL__null) { | ||||
3498 | __kmp_printf(" %p", __kmp_root[gtid]); | ||||
3499 | } | ||||
3500 | __kmp_printf("\n"); | ||||
3501 | } | ||||
3502 | } | ||||
3503 | |||||
3504 | // Print out __kmp_threads array. | ||||
3505 | __kmp_printf("\n------------------------------\nThreads\n--------------------" | ||||
3506 | "----------\n"); | ||||
3507 | if (__kmp_threads != NULL__null) { | ||||
3508 | int gtid; | ||||
3509 | for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) { | ||||
3510 | kmp_info_t const *thread = __kmp_threads[gtid]; | ||||
3511 | if (thread != NULL__null) { | ||||
3512 | __kmp_printf("GTID %2d %p:\n", gtid, thread); | ||||
3513 | __kmp_printf(" Our Root: %p\n", thread->th.th_root); | ||||
3514 | __kmp_print_structure_team(" Our Team: ", thread->th.th_team); | ||||
3515 | __kmp_print_structure_team(" Serial Team: ", | ||||
3516 | thread->th.th_serial_team); | ||||
3517 | __kmp_printf(" Threads: %2d\n", thread->th.th_team_nproc); | ||||
3518 | __kmp_print_structure_thread(" Primary: ", | ||||
3519 | thread->th.th_team_master); | ||||
3520 | __kmp_printf(" Serialized?: %2d\n", thread->th.th_team_serialized); | ||||
3521 | __kmp_printf(" Set NProc: %2d\n", thread->th.th_set_nproc); | ||||
3522 | __kmp_printf(" Set Proc Bind: %2d\n", thread->th.th_set_proc_bind); | ||||
3523 | __kmp_print_structure_thread(" Next in pool: ", | ||||
3524 | thread->th.th_next_pool); | ||||
3525 | __kmp_printf("\n"); | ||||
3526 | __kmp_print_structure_team_accum(list, thread->th.th_team); | ||||
3527 | __kmp_print_structure_team_accum(list, thread->th.th_serial_team); | ||||
3528 | } | ||||
3529 | } | ||||
3530 | } else { | ||||
3531 | __kmp_printf("Threads array is not allocated.\n"); | ||||
3532 | } | ||||
3533 | |||||
3534 | // Print out __kmp_root array. | ||||
3535 | __kmp_printf("\n------------------------------\nUbers\n----------------------" | ||||
3536 | "--------\n"); | ||||
3537 | if (__kmp_root != NULL__null) { | ||||
3538 | int gtid; | ||||
3539 | for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) { | ||||
3540 | kmp_root_t const *root = __kmp_root[gtid]; | ||||
3541 | if (root != NULL__null) { | ||||
3542 | __kmp_printf("GTID %2d %p:\n", gtid, root); | ||||
3543 | __kmp_print_structure_team(" Root Team: ", root->r.r_root_team); | ||||
3544 | __kmp_print_structure_team(" Hot Team: ", root->r.r_hot_team); | ||||
3545 | __kmp_print_structure_thread(" Uber Thread: ", | ||||
3546 | root->r.r_uber_thread); | ||||
3547 | __kmp_printf(" Active?: %2d\n", root->r.r_active); | ||||
3548 | __kmp_printf(" In Parallel: %2d\n", | ||||
3549 | KMP_ATOMIC_LD_RLX(&root->r.r_in_parallel)(&root->r.r_in_parallel)->load(std::memory_order_relaxed )); | ||||
3550 | __kmp_printf("\n"); | ||||
3551 | __kmp_print_structure_team_accum(list, root->r.r_root_team); | ||||
3552 | __kmp_print_structure_team_accum(list, root->r.r_hot_team); | ||||
3553 | } | ||||
3554 | } | ||||
3555 | } else { | ||||
3556 | __kmp_printf("Ubers array is not allocated.\n"); | ||||
3557 | } | ||||
3558 | |||||
3559 | __kmp_printf("\n------------------------------\nTeams\n----------------------" | ||||
3560 | "--------\n"); | ||||
3561 | while (list->next != NULL__null) { | ||||
3562 | kmp_team_p const *team = list->entry; | ||||
3563 | int i; | ||||
3564 | __kmp_printf("Team %2x %p:\n", team->t.t_id, team); | ||||
3565 | __kmp_print_structure_team(" Parent Team: ", team->t.t_parent); | ||||
3566 | __kmp_printf(" Primary TID: %2d\n", team->t.t_master_tid); | ||||
3567 | __kmp_printf(" Max threads: %2d\n", team->t.t_max_nproc); | ||||
3568 | __kmp_printf(" Levels of serial: %2d\n", team->t.t_serialized); | ||||
3569 | __kmp_printf(" Number threads: %2d\n", team->t.t_nproc); | ||||
3570 | for (i = 0; i < team->t.t_nproc; ++i) { | ||||
3571 | __kmp_printf(" Thread %2d: ", i); | ||||
3572 | __kmp_print_structure_thread("", team->t.t_threads[i]); | ||||
3573 | } | ||||
3574 | __kmp_print_structure_team(" Next in pool: ", team->t.t_next_pool); | ||||
3575 | __kmp_printf("\n"); | ||||
3576 | list = list->next; | ||||
3577 | } | ||||
3578 | |||||
3579 | // Print out __kmp_thread_pool and __kmp_team_pool. | ||||
3580 | __kmp_printf("\n------------------------------\nPools\n----------------------" | ||||
3581 | "--------\n"); | ||||
3582 | __kmp_print_structure_thread("Thread pool: ", | ||||
3583 | CCAST(kmp_info_t *, __kmp_thread_pool)const_cast<kmp_info_t *>(__kmp_thread_pool)); | ||||
3584 | __kmp_print_structure_team("Team pool: ", | ||||
3585 | CCAST(kmp_team_t *, __kmp_team_pool)const_cast<kmp_team_t *>(__kmp_team_pool)); | ||||
3586 | __kmp_printf("\n"); | ||||
3587 | |||||
3588 | // Free team list. | ||||
3589 | while (list != NULL__null) { | ||||
3590 | kmp_team_list_item_t *item = list; | ||||
3591 | list = list->next; | ||||
3592 | KMP_INTERNAL_FREE(item)free(item); | ||||
3593 | } | ||||
3594 | } | ||||
3595 | |||||
3596 | #endif | ||||
3597 | |||||
3598 | //--------------------------------------------------------------------------- | ||||
3599 | // Stuff for per-thread fast random number generator | ||||
3600 | // Table of primes | ||||
3601 | static const unsigned __kmp_primes[] = { | ||||
3602 | 0x9e3779b1, 0xffe6cc59, 0x2109f6dd, 0x43977ab5, 0xba5703f5, 0xb495a877, | ||||
3603 | 0xe1626741, 0x79695e6b, 0xbc98c09f, 0xd5bee2b3, 0x287488f9, 0x3af18231, | ||||
3604 | 0x9677cd4d, 0xbe3a6929, 0xadc6a877, 0xdcf0674b, 0xbe4d6fe9, 0x5f15e201, | ||||
3605 | 0x99afc3fd, 0xf3f16801, 0xe222cfff, 0x24ba5fdb, 0x0620452d, 0x79f149e3, | ||||
3606 | 0xc8b93f49, 0x972702cd, 0xb07dd827, 0x6c97d5ed, 0x085a3d61, 0x46eb5ea7, | ||||
3607 | 0x3d9910ed, 0x2e687b5b, 0x29609227, 0x6eb081f1, 0x0954c4e1, 0x9d114db9, | ||||
3608 | 0x542acfa9, 0xb3e6bd7b, 0x0742d917, 0xe9f3ffa7, 0x54581edb, 0xf2480f45, | ||||
3609 | 0x0bb9288f, 0xef1affc7, 0x85fa0ca7, 0x3ccc14db, 0xe6baf34b, 0x343377f7, | ||||
3610 | 0x5ca19031, 0xe6d9293b, 0xf0a9f391, 0x5d2e980b, 0xfc411073, 0xc3749363, | ||||
3611 | 0xb892d829, 0x3549366b, 0x629750ad, 0xb98294e5, 0x892d9483, 0xc235baf3, | ||||
3612 | 0x3d2402a3, 0x6bdef3c9, 0xbec333cd, 0x40c9520f}; | ||||
3613 | |||||
3614 | //--------------------------------------------------------------------------- | ||||
3615 | // __kmp_get_random: Get a random number using a linear congruential method. | ||||
3616 | unsigned short __kmp_get_random(kmp_info_t *thread) { | ||||
3617 | unsigned x = thread->th.th_x; | ||||
3618 | unsigned short r = (unsigned short)(x >> 16); | ||||
3619 | |||||
3620 | thread->th.th_x = x * thread->th.th_a + 1; | ||||
3621 | |||||
3622 | KA_TRACE(30, ("__kmp_get_random: THREAD: %d, RETURN: %u\n",if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_get_random: THREAD: %d, RETURN: %u\n" , thread->th.th_info.ds.ds_tid, r); } | ||||
3623 | thread->th.th_info.ds.ds_tid, r))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_get_random: THREAD: %d, RETURN: %u\n" , thread->th.th_info.ds.ds_tid, r); }; | ||||
3624 | |||||
3625 | return r; | ||||
3626 | } | ||||
3627 | //-------------------------------------------------------- | ||||
3628 | // __kmp_init_random: Initialize a random number generator | ||||
3629 | void __kmp_init_random(kmp_info_t *thread) { | ||||
3630 | unsigned seed = thread->th.th_info.ds.ds_tid; | ||||
3631 | |||||
3632 | thread->th.th_a = | ||||
3633 | __kmp_primes[seed % (sizeof(__kmp_primes) / sizeof(__kmp_primes[0]))]; | ||||
3634 | thread->th.th_x = (seed + 1) * thread->th.th_a + 1; | ||||
3635 | KA_TRACE(30,if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_init_random: THREAD: %u; A: %u\n" , seed, thread->th.th_a); } | ||||
3636 | ("__kmp_init_random: THREAD: %u; A: %u\n", seed, thread->th.th_a))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_init_random: THREAD: %u; A: %u\n" , seed, thread->th.th_a); }; | ||||
3637 | } | ||||
3638 | |||||
3639 | #if KMP_OS_WINDOWS0 | ||||
3640 | /* reclaim array entries for root threads that are already dead, returns number | ||||
3641 | * reclaimed */ | ||||
3642 | static int __kmp_reclaim_dead_roots(void) { | ||||
3643 | int i, r = 0; | ||||
3644 | |||||
3645 | for (i = 0; i < __kmp_threads_capacity; ++i) { | ||||
3646 | if (KMP_UBER_GTID(i) && | ||||
3647 | !__kmp_still_running((kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[i])((void *)(__kmp_threads[i]))) && | ||||
3648 | !__kmp_root[i] | ||||
3649 | ->r.r_active) { // AC: reclaim only roots died in non-active state | ||||
3650 | r += __kmp_unregister_root_other_thread(i); | ||||
3651 | } | ||||
3652 | } | ||||
3653 | return r; | ||||
3654 | } | ||||
3655 | #endif | ||||
3656 | |||||
3657 | /* This function attempts to create free entries in __kmp_threads and | ||||
3658 | __kmp_root, and returns the number of free entries generated. | ||||
3659 | |||||
3660 | For Windows* OS static library, the first mechanism used is to reclaim array | ||||
3661 | entries for root threads that are already dead. | ||||
3662 | |||||
3663 | On all platforms, expansion is attempted on the arrays __kmp_threads_ and | ||||
3664 | __kmp_root, with appropriate update to __kmp_threads_capacity. Array | ||||
3665 | capacity is increased by doubling with clipping to __kmp_tp_capacity, if | ||||
3666 | threadprivate cache array has been created. Synchronization with | ||||
3667 | __kmpc_threadprivate_cached is done using __kmp_tp_cached_lock. | ||||
3668 | |||||
3669 | After any dead root reclamation, if the clipping value allows array expansion | ||||
3670 | to result in the generation of a total of nNeed free slots, the function does | ||||
3671 | that expansion. If not, nothing is done beyond the possible initial root | ||||
3672 | thread reclamation. | ||||
3673 | |||||
3674 | If any argument is negative, the behavior is undefined. */ | ||||
3675 | static int __kmp_expand_threads(int nNeed) { | ||||
3676 | int added = 0; | ||||
3677 | int minimumRequiredCapacity; | ||||
3678 | int newCapacity; | ||||
3679 | kmp_info_t **newThreads; | ||||
3680 | kmp_root_t **newRoot; | ||||
3681 | |||||
3682 | // All calls to __kmp_expand_threads should be under __kmp_forkjoin_lock, so | ||||
3683 | // resizing __kmp_threads does not need additional protection if foreign | ||||
3684 | // threads are present | ||||
3685 | |||||
3686 | #if KMP_OS_WINDOWS0 && !KMP_DYNAMIC_LIB1 | ||||
3687 | /* only for Windows static library */ | ||||
3688 | /* reclaim array entries for root threads that are already dead */ | ||||
3689 | added = __kmp_reclaim_dead_roots(); | ||||
3690 | |||||
3691 | if (nNeed) { | ||||
3692 | nNeed -= added; | ||||
3693 | if (nNeed < 0) | ||||
3694 | nNeed = 0; | ||||
3695 | } | ||||
3696 | #endif | ||||
3697 | if (nNeed <= 0) | ||||
3698 | return added; | ||||
3699 | |||||
3700 | // Note that __kmp_threads_capacity is not bounded by __kmp_max_nth. If | ||||
3701 | // __kmp_max_nth is set to some value less than __kmp_sys_max_nth by the | ||||
3702 | // user via KMP_DEVICE_THREAD_LIMIT, then __kmp_threads_capacity may become | ||||
3703 | // > __kmp_max_nth in one of two ways: | ||||
3704 | // | ||||
3705 | // 1) The initialization thread (gtid = 0) exits. __kmp_threads[0] | ||||
3706 | // may not be reused by another thread, so we may need to increase | ||||
3707 | // __kmp_threads_capacity to __kmp_max_nth + 1. | ||||
3708 | // | ||||
3709 | // 2) New foreign root(s) are encountered. We always register new foreign | ||||
3710 | // roots. This may cause a smaller # of threads to be allocated at | ||||
3711 | // subsequent parallel regions, but the worker threads hang around (and | ||||
3712 | // eventually go to sleep) and need slots in the __kmp_threads[] array. | ||||
3713 | // | ||||
3714 | // Anyway, that is the reason for moving the check to see if | ||||
3715 | // __kmp_max_nth was exceeded into __kmp_reserve_threads() | ||||
3716 | // instead of having it performed here. -BB | ||||
3717 | |||||
3718 | KMP_DEBUG_ASSERT(__kmp_sys_max_nth >= __kmp_threads_capacity)if (!(__kmp_sys_max_nth >= __kmp_threads_capacity)) { __kmp_debug_assert ("__kmp_sys_max_nth >= __kmp_threads_capacity", "openmp/runtime/src/kmp_runtime.cpp" , 3718); }; | ||||
3719 | |||||
3720 | /* compute expansion headroom to check if we can expand */ | ||||
3721 | if (__kmp_sys_max_nth - __kmp_threads_capacity < nNeed) { | ||||
3722 | /* possible expansion too small -- give up */ | ||||
3723 | return added; | ||||
3724 | } | ||||
3725 | minimumRequiredCapacity = __kmp_threads_capacity + nNeed; | ||||
3726 | |||||
3727 | newCapacity = __kmp_threads_capacity; | ||||
3728 | do { | ||||
3729 | newCapacity = newCapacity <= (__kmp_sys_max_nth >> 1) ? (newCapacity << 1) | ||||
3730 | : __kmp_sys_max_nth; | ||||
3731 | } while (newCapacity < minimumRequiredCapacity); | ||||
3732 | newThreads = (kmp_info_t **)__kmp_allocate(___kmp_allocate(((sizeof(kmp_info_t *) + sizeof(kmp_root_t *) ) * newCapacity + 64), "openmp/runtime/src/kmp_runtime.cpp", 3733 ) | ||||
3733 | (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * newCapacity + CACHE_LINE)___kmp_allocate(((sizeof(kmp_info_t *) + sizeof(kmp_root_t *) ) * newCapacity + 64), "openmp/runtime/src/kmp_runtime.cpp", 3733 ); | ||||
3734 | newRoot = | ||||
3735 | (kmp_root_t **)((char *)newThreads + sizeof(kmp_info_t *) * newCapacity); | ||||
3736 | KMP_MEMCPYmemcpy(newThreads, __kmp_threads, | ||||
3737 | __kmp_threads_capacity * sizeof(kmp_info_t *)); | ||||
3738 | KMP_MEMCPYmemcpy(newRoot, __kmp_root, | ||||
3739 | __kmp_threads_capacity * sizeof(kmp_root_t *)); | ||||
3740 | // Put old __kmp_threads array on a list. Any ongoing references to the old | ||||
3741 | // list will be valid. This list is cleaned up at library shutdown. | ||||
3742 | kmp_old_threads_list_t *node = | ||||
3743 | (kmp_old_threads_list_t *)__kmp_allocate(sizeof(kmp_old_threads_list_t))___kmp_allocate((sizeof(kmp_old_threads_list_t)), "openmp/runtime/src/kmp_runtime.cpp" , 3743); | ||||
3744 | node->threads = __kmp_threads; | ||||
3745 | node->next = __kmp_old_threads_list; | ||||
3746 | __kmp_old_threads_list = node; | ||||
3747 | |||||
3748 | *(kmp_info_t * *volatile *)&__kmp_threads = newThreads; | ||||
3749 | *(kmp_root_t * *volatile *)&__kmp_root = newRoot; | ||||
3750 | added += newCapacity - __kmp_threads_capacity; | ||||
3751 | *(volatile int *)&__kmp_threads_capacity = newCapacity; | ||||
3752 | |||||
3753 | if (newCapacity > __kmp_tp_capacity) { | ||||
3754 | __kmp_acquire_bootstrap_lock(&__kmp_tp_cached_lock); | ||||
3755 | if (__kmp_tp_cached && newCapacity > __kmp_tp_capacity) { | ||||
3756 | __kmp_threadprivate_resize_cache(newCapacity); | ||||
3757 | } else { // increase __kmp_tp_capacity to correspond with kmp_threads size | ||||
3758 | *(volatile int *)&__kmp_tp_capacity = newCapacity; | ||||
3759 | } | ||||
3760 | __kmp_release_bootstrap_lock(&__kmp_tp_cached_lock); | ||||
3761 | } | ||||
3762 | |||||
3763 | return added; | ||||
3764 | } | ||||
3765 | |||||
3766 | /* Register the current thread as a root thread and obtain our gtid. We must | ||||
3767 | have the __kmp_initz_lock held at this point. Argument TRUE only if are the | ||||
3768 | thread that calls from __kmp_do_serial_initialize() */ | ||||
3769 | int __kmp_register_root(int initial_thread) { | ||||
3770 | kmp_info_t *root_thread; | ||||
3771 | kmp_root_t *root; | ||||
3772 | int gtid; | ||||
3773 | int capacity; | ||||
3774 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); | ||||
3775 | KA_TRACE(20, ("__kmp_register_root: entered\n"))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_register_root: entered\n" ); }; | ||||
3776 | KMP_MB(); | ||||
3777 | |||||
3778 | /* 2007-03-02: | ||||
3779 | If initial thread did not invoke OpenMP RTL yet, and this thread is not an | ||||
3780 | initial one, "__kmp_all_nth >= __kmp_threads_capacity" condition does not | ||||
3781 | work as expected -- it may return false (that means there is at least one | ||||
3782 | empty slot in __kmp_threads array), but it is possible the only free slot | ||||
3783 | is #0, which is reserved for initial thread and so cannot be used for this | ||||
3784 | one. Following code workarounds this bug. | ||||
3785 | |||||
3786 | However, right solution seems to be not reserving slot #0 for initial | ||||
3787 | thread because: | ||||
3788 | (1) there is no magic in slot #0, | ||||
3789 | (2) we cannot detect initial thread reliably (the first thread which does | ||||
3790 | serial initialization may be not a real initial thread). | ||||
3791 | */ | ||||
3792 | capacity = __kmp_threads_capacity; | ||||
3793 | if (!initial_thread && TCR_PTR(__kmp_threads[0])((void *)(__kmp_threads[0])) == NULL__null) { | ||||
3794 | --capacity; | ||||
3795 | } | ||||
3796 | |||||
3797 | // If it is not for initializing the hidden helper team, we need to take | ||||
3798 | // __kmp_hidden_helper_threads_num out of the capacity because it is included | ||||
3799 | // in __kmp_threads_capacity. | ||||
3800 | if (__kmp_enable_hidden_helper && !TCR_4(__kmp_init_hidden_helper_threads)(__kmp_init_hidden_helper_threads)) { | ||||
3801 | capacity -= __kmp_hidden_helper_threads_num; | ||||
3802 | } | ||||
3803 | |||||
3804 | /* see if there are too many threads */ | ||||
3805 | if (__kmp_all_nth >= capacity && !__kmp_expand_threads(1)) { | ||||
3806 | if (__kmp_tp_cached) { | ||||
3807 | __kmp_fatal(KMP_MSG(CantRegisterNewThread)__kmp_msg_format(kmp_i18n_msg_CantRegisterNewThread), | ||||
3808 | KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity)__kmp_msg_format(kmp_i18n_hnt_Set_ALL_THREADPRIVATE, __kmp_tp_capacity ), | ||||
3809 | KMP_HNT(PossibleSystemLimitOnThreads)__kmp_msg_format(kmp_i18n_hnt_PossibleSystemLimitOnThreads), __kmp_msg_null); | ||||
3810 | } else { | ||||
3811 | __kmp_fatal(KMP_MSG(CantRegisterNewThread)__kmp_msg_format(kmp_i18n_msg_CantRegisterNewThread), KMP_HNT(SystemLimitOnThreads)__kmp_msg_format(kmp_i18n_hnt_SystemLimitOnThreads), | ||||
3812 | __kmp_msg_null); | ||||
3813 | } | ||||
3814 | } | ||||
3815 | |||||
3816 | // When hidden helper task is enabled, __kmp_threads is organized as follows: | ||||
3817 | // 0: initial thread, also a regular OpenMP thread. | ||||
3818 | // [1, __kmp_hidden_helper_threads_num]: slots for hidden helper threads. | ||||
3819 | // [__kmp_hidden_helper_threads_num + 1, __kmp_threads_capacity): slots for | ||||
3820 | // regular OpenMP threads. | ||||
3821 | if (TCR_4(__kmp_init_hidden_helper_threads)(__kmp_init_hidden_helper_threads)) { | ||||
3822 | // Find an available thread slot for hidden helper thread. Slots for hidden | ||||
3823 | // helper threads start from 1 to __kmp_hidden_helper_threads_num. | ||||
3824 | for (gtid = 1; TCR_PTR(__kmp_threads[gtid])((void *)(__kmp_threads[gtid])) != NULL__null && | ||||
3825 | gtid <= __kmp_hidden_helper_threads_num; | ||||
3826 | gtid++) | ||||
3827 | ; | ||||
3828 | KMP_ASSERT(gtid <= __kmp_hidden_helper_threads_num)if (!(gtid <= __kmp_hidden_helper_threads_num)) { __kmp_debug_assert ("gtid <= __kmp_hidden_helper_threads_num", "openmp/runtime/src/kmp_runtime.cpp" , 3828); }; | ||||
3829 | KA_TRACE(1, ("__kmp_register_root: found slot in threads array for "if (kmp_a_debug >= 1) { __kmp_debug_printf ("__kmp_register_root: found slot in threads array for " "hidden helper thread: T#%d\n", gtid); } | ||||
3830 | "hidden helper thread: T#%d\n",if (kmp_a_debug >= 1) { __kmp_debug_printf ("__kmp_register_root: found slot in threads array for " "hidden helper thread: T#%d\n", gtid); } | ||||
3831 | gtid))if (kmp_a_debug >= 1) { __kmp_debug_printf ("__kmp_register_root: found slot in threads array for " "hidden helper thread: T#%d\n", gtid); }; | ||||
3832 | } else { | ||||
3833 | /* find an available thread slot */ | ||||
3834 | // Don't reassign the zero slot since we need that to only be used by | ||||
3835 | // initial thread. Slots for hidden helper threads should also be skipped. | ||||
3836 | if (initial_thread && TCR_PTR(__kmp_threads[0])((void *)(__kmp_threads[0])) == NULL__null) { | ||||
3837 | gtid = 0; | ||||
3838 | } else { | ||||
3839 | for (gtid = __kmp_hidden_helper_threads_num + 1; | ||||
3840 | TCR_PTR(__kmp_threads[gtid])((void *)(__kmp_threads[gtid])) != NULL__null; gtid++) | ||||
3841 | ; | ||||
3842 | } | ||||
3843 | KA_TRACE(if (kmp_a_debug >= 1) { __kmp_debug_printf ("__kmp_register_root: found slot in threads array: T#%d\n" , gtid); } | ||||
3844 | 1, ("__kmp_register_root: found slot in threads array: T#%d\n", gtid))if (kmp_a_debug >= 1) { __kmp_debug_printf ("__kmp_register_root: found slot in threads array: T#%d\n" , gtid); }; | ||||
3845 | KMP_ASSERT(gtid < __kmp_threads_capacity)if (!(gtid < __kmp_threads_capacity)) { __kmp_debug_assert ("gtid < __kmp_threads_capacity", "openmp/runtime/src/kmp_runtime.cpp" , 3845); }; | ||||
3846 | } | ||||
3847 | |||||
3848 | /* update global accounting */ | ||||
3849 | __kmp_all_nth++; | ||||
3850 | TCW_4(__kmp_nth, __kmp_nth + 1)(__kmp_nth) = (__kmp_nth + 1); | ||||
3851 | |||||
3852 | // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low | ||||
3853 | // numbers of procs, and method #2 (keyed API call) for higher numbers. | ||||
3854 | if (__kmp_adjust_gtid_mode) { | ||||
3855 | if (__kmp_all_nth >= __kmp_tls_gtid_min) { | ||||
3856 | if (TCR_4(__kmp_gtid_mode)(__kmp_gtid_mode) != 2) { | ||||
3857 | TCW_4(__kmp_gtid_mode, 2)(__kmp_gtid_mode) = (2); | ||||
3858 | } | ||||
3859 | } else { | ||||
3860 | if (TCR_4(__kmp_gtid_mode)(__kmp_gtid_mode) != 1) { | ||||
3861 | TCW_4(__kmp_gtid_mode, 1)(__kmp_gtid_mode) = (1); | ||||
3862 | } | ||||
3863 | } | ||||
3864 | } | ||||
3865 | |||||
3866 | #ifdef KMP_ADJUST_BLOCKTIME1 | ||||
3867 | /* Adjust blocktime to zero if necessary */ | ||||
3868 | /* Middle initialization might not have occurred yet */ | ||||
3869 | if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { | ||||
3870 | if (__kmp_nth > __kmp_avail_proc) { | ||||
3871 | __kmp_zero_bt = TRUE(!0); | ||||
3872 | } | ||||
3873 | } | ||||
3874 | #endif /* KMP_ADJUST_BLOCKTIME */ | ||||
3875 | |||||
3876 | /* setup this new hierarchy */ | ||||
3877 | if (!(root = __kmp_root[gtid])) { | ||||
3878 | root = __kmp_root[gtid] = (kmp_root_t *)__kmp_allocate(sizeof(kmp_root_t))___kmp_allocate((sizeof(kmp_root_t)), "openmp/runtime/src/kmp_runtime.cpp" , 3878); | ||||
3879 | KMP_DEBUG_ASSERT(!root->r.r_root_team)if (!(!root->r.r_root_team)) { __kmp_debug_assert("!root->r.r_root_team" , "openmp/runtime/src/kmp_runtime.cpp", 3879); }; | ||||
3880 | } | ||||
3881 | |||||
3882 | #if KMP_STATS_ENABLED0 | ||||
3883 | // Initialize stats as soon as possible (right after gtid assignment). | ||||
3884 | __kmp_stats_thread_ptr = __kmp_stats_list->push_back(gtid); | ||||
3885 | __kmp_stats_thread_ptr->startLife(); | ||||
3886 | KMP_SET_THREAD_STATE(SERIAL_REGION)((void)0); | ||||
3887 | KMP_INIT_PARTITIONED_TIMERS(OMP_serial)((void)0); | ||||
3888 | #endif | ||||
3889 | __kmp_initialize_root(root); | ||||
3890 | |||||
3891 | /* setup new root thread structure */ | ||||
3892 | if (root->r.r_uber_thread) { | ||||
3893 | root_thread = root->r.r_uber_thread; | ||||
3894 | } else { | ||||
3895 | root_thread = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t))___kmp_allocate((sizeof(kmp_info_t)), "openmp/runtime/src/kmp_runtime.cpp" , 3895); | ||||
3896 | if (__kmp_storage_map) { | ||||
3897 | __kmp_print_thread_storage_map(root_thread, gtid); | ||||
3898 | } | ||||
3899 | root_thread->th.th_info.ds.ds_gtid = gtid; | ||||
3900 | #if OMPT_SUPPORT1 | ||||
3901 | root_thread->th.ompt_thread_info.thread_data = ompt_data_none{0}; | ||||
3902 | #endif | ||||
3903 | root_thread->th.th_root = root; | ||||
3904 | if (__kmp_env_consistency_check) { | ||||
3905 | root_thread->th.th_cons = __kmp_allocate_cons_stack(gtid); | ||||
3906 | } | ||||
3907 | #if USE_FAST_MEMORY3 | ||||
3908 | __kmp_initialize_fast_memory(root_thread); | ||||
3909 | #endif /* USE_FAST_MEMORY */ | ||||
3910 | |||||
3911 | #if KMP_USE_BGET1 | ||||
3912 | KMP_DEBUG_ASSERT(root_thread->th.th_local.bget_data == NULL)if (!(root_thread->th.th_local.bget_data == __null)) { __kmp_debug_assert ("root_thread->th.th_local.bget_data == __null", "openmp/runtime/src/kmp_runtime.cpp" , 3912); }; | ||||
3913 | __kmp_initialize_bget(root_thread); | ||||
3914 | #endif | ||||
3915 | __kmp_init_random(root_thread); // Initialize random number generator | ||||
3916 | } | ||||
3917 | |||||
3918 | /* setup the serial team held in reserve by the root thread */ | ||||
3919 | if (!root_thread->th.th_serial_team) { | ||||
3920 | kmp_internal_control_t r_icvs = __kmp_get_global_icvs(); | ||||
3921 | KF_TRACE(10, ("__kmp_register_root: before serial_team\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_register_root: before serial_team\n" ); }; | ||||
3922 | root_thread->th.th_serial_team = __kmp_allocate_team( | ||||
3923 | root, 1, 1, | ||||
3924 | #if OMPT_SUPPORT1 | ||||
3925 | ompt_data_none{0}, // root parallel id | ||||
3926 | #endif | ||||
3927 | proc_bind_default, &r_icvs, 0 USE_NESTED_HOT_ARG(NULL), __null); | ||||
3928 | } | ||||
3929 | KMP_ASSERT(root_thread->th.th_serial_team)if (!(root_thread->th.th_serial_team)) { __kmp_debug_assert ("root_thread->th.th_serial_team", "openmp/runtime/src/kmp_runtime.cpp" , 3929); }; | ||||
3930 | KF_TRACE(10, ("__kmp_register_root: after serial_team = %p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_register_root: after serial_team = %p\n" , root_thread->th.th_serial_team); } | ||||
3931 | root_thread->th.th_serial_team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_register_root: after serial_team = %p\n" , root_thread->th.th_serial_team); }; | ||||
3932 | |||||
3933 | /* drop root_thread into place */ | ||||
3934 | TCW_SYNC_PTR(__kmp_threads[gtid], root_thread)((__kmp_threads[gtid])) = ((root_thread)); | ||||
3935 | |||||
3936 | root->r.r_root_team->t.t_threads[0] = root_thread; | ||||
3937 | root->r.r_hot_team->t.t_threads[0] = root_thread; | ||||
3938 | root_thread->th.th_serial_team->t.t_threads[0] = root_thread; | ||||
3939 | // AC: the team created in reserve, not for execution (it is unused for now). | ||||
3940 | root_thread->th.th_serial_team->t.t_serialized = 0; | ||||
3941 | root->r.r_uber_thread = root_thread; | ||||
3942 | |||||
3943 | /* initialize the thread, get it ready to go */ | ||||
3944 | __kmp_initialize_info(root_thread, root->r.r_root_team, 0, gtid); | ||||
3945 | TCW_4(__kmp_init_gtid, TRUE)(__kmp_init_gtid) = ((!0)); | ||||
3946 | |||||
3947 | /* prepare the primary thread for get_gtid() */ | ||||
3948 | __kmp_gtid_set_specific(gtid); | ||||
3949 | |||||
3950 | #if USE_ITT_BUILD1 | ||||
3951 | __kmp_itt_thread_name(gtid); | ||||
3952 | #endif /* USE_ITT_BUILD */ | ||||
3953 | |||||
3954 | #ifdef KMP_TDATA_GTID1 | ||||
3955 | __kmp_gtid = gtid; | ||||
3956 | #endif | ||||
3957 | __kmp_create_worker(gtid, root_thread, __kmp_stksize); | ||||
3958 | KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == gtid)if (!(__kmp_gtid_get_specific() == gtid)) { __kmp_debug_assert ("__kmp_gtid_get_specific() == gtid", "openmp/runtime/src/kmp_runtime.cpp" , 3958); }; | ||||
3959 | |||||
3960 | KA_TRACE(20, ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, " "plain=%u\n", gtid, __kmp_gtid_from_tid(0, root->r.r_hot_team ), root->r.r_hot_team->t.t_id, 0, 0, 0); } | ||||
3961 | "plain=%u\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, " "plain=%u\n", gtid, __kmp_gtid_from_tid(0, root->r.r_hot_team ), root->r.r_hot_team->t.t_id, 0, 0, 0); } | ||||
3962 | gtid, __kmp_gtid_from_tid(0, root->r.r_hot_team),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, " "plain=%u\n", gtid, __kmp_gtid_from_tid(0, root->r.r_hot_team ), root->r.r_hot_team->t.t_id, 0, 0, 0); } | ||||
3963 | root->r.r_hot_team->t.t_id, 0, KMP_INIT_BARRIER_STATE,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, " "plain=%u\n", gtid, __kmp_gtid_from_tid(0, root->r.r_hot_team ), root->r.r_hot_team->t.t_id, 0, 0, 0); } | ||||
3964 | KMP_INIT_BARRIER_STATE))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, " "plain=%u\n", gtid, __kmp_gtid_from_tid(0, root->r.r_hot_team ), root->r.r_hot_team->t.t_id, 0, 0, 0); }; | ||||
3965 | { // Initialize barrier data. | ||||
3966 | int b; | ||||
3967 | for (b = 0; b < bs_last_barrier; ++b) { | ||||
3968 | root_thread->th.th_bar[b].bb.b_arrived = KMP_INIT_BARRIER_STATE0; | ||||
3969 | #if USE_DEBUGGER0 | ||||
3970 | root_thread->th.th_bar[b].bb.b_worker_arrived = 0; | ||||
3971 | #endif | ||||
3972 | } | ||||
3973 | } | ||||
3974 | KMP_DEBUG_ASSERT(root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier].b_arrived ==if (!(root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier]. b_arrived == 0)) { __kmp_debug_assert("root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier].b_arrived == 0" , "openmp/runtime/src/kmp_runtime.cpp", 3975); } | ||||
3975 | KMP_INIT_BARRIER_STATE)if (!(root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier]. b_arrived == 0)) { __kmp_debug_assert("root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier].b_arrived == 0" , "openmp/runtime/src/kmp_runtime.cpp", 3975); }; | ||||
3976 | |||||
3977 | #if KMP_AFFINITY_SUPPORTED1 | ||||
3978 | root_thread->th.th_current_place = KMP_PLACE_UNDEFINED(-2); | ||||
3979 | root_thread->th.th_new_place = KMP_PLACE_UNDEFINED(-2); | ||||
3980 | root_thread->th.th_first_place = KMP_PLACE_UNDEFINED(-2); | ||||
3981 | root_thread->th.th_last_place = KMP_PLACE_UNDEFINED(-2); | ||||
3982 | #endif /* KMP_AFFINITY_SUPPORTED */ | ||||
3983 | root_thread->th.th_def_allocator = __kmp_def_allocator; | ||||
3984 | root_thread->th.th_prev_level = 0; | ||||
3985 | root_thread->th.th_prev_num_threads = 1; | ||||
3986 | |||||
3987 | kmp_cg_root_t *tmp = (kmp_cg_root_t *)__kmp_allocate(sizeof(kmp_cg_root_t))___kmp_allocate((sizeof(kmp_cg_root_t)), "openmp/runtime/src/kmp_runtime.cpp" , 3987); | ||||
3988 | tmp->cg_root = root_thread; | ||||
3989 | tmp->cg_thread_limit = __kmp_cg_max_nth; | ||||
3990 | tmp->cg_nthreads = 1; | ||||
3991 | KA_TRACE(100, ("__kmp_register_root: Thread %p created node %p with"if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_register_root: Thread %p created node %p with" " cg_nthreads init to 1\n", root_thread, tmp); } | ||||
3992 | " cg_nthreads init to 1\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_register_root: Thread %p created node %p with" " cg_nthreads init to 1\n", root_thread, tmp); } | ||||
3993 | root_thread, tmp))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_register_root: Thread %p created node %p with" " cg_nthreads init to 1\n", root_thread, tmp); }; | ||||
3994 | tmp->up = NULL__null; | ||||
3995 | root_thread->th.th_cg_roots = tmp; | ||||
3996 | |||||
3997 | __kmp_root_counter++; | ||||
3998 | |||||
3999 | #if OMPT_SUPPORT1 | ||||
4000 | if (!initial_thread && ompt_enabled.enabled) { | ||||
4001 | |||||
4002 | kmp_info_t *root_thread = ompt_get_thread(); | ||||
4003 | |||||
4004 | ompt_set_thread_state(root_thread, ompt_state_overhead); | ||||
4005 | |||||
4006 | if (ompt_enabled.ompt_callback_thread_begin) { | ||||
4007 | ompt_callbacks.ompt_callback(ompt_callback_thread_begin)ompt_callback_thread_begin_callback( | ||||
4008 | ompt_thread_initial, __ompt_get_thread_data_internal()); | ||||
4009 | } | ||||
4010 | ompt_data_t *task_data; | ||||
4011 | ompt_data_t *parallel_data; | ||||
4012 | __ompt_get_task_info_internal(0, NULL__null, &task_data, NULL__null, ¶llel_data, | ||||
4013 | NULL__null); | ||||
4014 | if (ompt_enabled.ompt_callback_implicit_task) { | ||||
4015 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | ||||
4016 | ompt_scope_begin, parallel_data, task_data, 1, 1, ompt_task_initial); | ||||
4017 | } | ||||
4018 | |||||
4019 | ompt_set_thread_state(root_thread, ompt_state_work_serial); | ||||
4020 | } | ||||
4021 | #endif | ||||
4022 | #if OMPD_SUPPORT1 | ||||
4023 | if (ompd_state & OMPD_ENABLE_BP0x1) | ||||
4024 | ompd_bp_thread_begin(); | ||||
4025 | #endif | ||||
4026 | |||||
4027 | KMP_MB(); | ||||
4028 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | ||||
4029 | |||||
4030 | return gtid; | ||||
4031 | } | ||||
4032 | |||||
4033 | #if KMP_NESTED_HOT_TEAMS1 | ||||
4034 | static int __kmp_free_hot_teams(kmp_root_t *root, kmp_info_t *thr, int level, | ||||
4035 | const int max_level) { | ||||
4036 | int i, n, nth; | ||||
4037 | kmp_hot_team_ptr_t *hot_teams = thr->th.th_hot_teams; | ||||
4038 | if (!hot_teams || !hot_teams[level].hot_team) { | ||||
4039 | return 0; | ||||
4040 | } | ||||
4041 | KMP_DEBUG_ASSERT(level < max_level)if (!(level < max_level)) { __kmp_debug_assert("level < max_level" , "openmp/runtime/src/kmp_runtime.cpp", 4041); }; | ||||
4042 | kmp_team_t *team = hot_teams[level].hot_team; | ||||
4043 | nth = hot_teams[level].hot_team_nth; | ||||
4044 | n = nth - 1; // primary thread is not freed | ||||
4045 | if (level < max_level - 1) { | ||||
4046 | for (i = 0; i < nth; ++i) { | ||||
4047 | kmp_info_t *th = team->t.t_threads[i]; | ||||
4048 | n += __kmp_free_hot_teams(root, th, level + 1, max_level); | ||||
4049 | if (i > 0 && th->th.th_hot_teams) { | ||||
4050 | __kmp_free(th->th.th_hot_teams)___kmp_free((th->th.th_hot_teams), "openmp/runtime/src/kmp_runtime.cpp" , 4050); | ||||
4051 | th->th.th_hot_teams = NULL__null; | ||||
4052 | } | ||||
4053 | } | ||||
4054 | } | ||||
4055 | __kmp_free_team(root, team, NULL__null); | ||||
4056 | return n; | ||||
4057 | } | ||||
4058 | #endif | ||||
4059 | |||||
4060 | // Resets a root thread and clear its root and hot teams. | ||||
4061 | // Returns the number of __kmp_threads entries directly and indirectly freed. | ||||
4062 | static int __kmp_reset_root(int gtid, kmp_root_t *root) { | ||||
4063 | kmp_team_t *root_team = root->r.r_root_team; | ||||
4064 | kmp_team_t *hot_team = root->r.r_hot_team; | ||||
4065 | int n = hot_team->t.t_nproc; | ||||
4066 | int i; | ||||
4067 | |||||
4068 | KMP_DEBUG_ASSERT(!root->r.r_active)if (!(!root->r.r_active)) { __kmp_debug_assert("!root->r.r_active" , "openmp/runtime/src/kmp_runtime.cpp", 4068); }; | ||||
4069 | |||||
4070 | root->r.r_root_team = NULL__null; | ||||
4071 | root->r.r_hot_team = NULL__null; | ||||
4072 | // __kmp_free_team() does not free hot teams, so we have to clear r_hot_team | ||||
4073 | // before call to __kmp_free_team(). | ||||
4074 | __kmp_free_team(root, root_team USE_NESTED_HOT_ARG(NULL), __null); | ||||
4075 | #if KMP_NESTED_HOT_TEAMS1 | ||||
4076 | if (__kmp_hot_teams_max_level > | ||||
4077 | 0) { // need to free nested hot teams and their threads if any | ||||
4078 | for (i = 0; i < hot_team->t.t_nproc; ++i) { | ||||
4079 | kmp_info_t *th = hot_team->t.t_threads[i]; | ||||
4080 | if (__kmp_hot_teams_max_level > 1) { | ||||
4081 | n += __kmp_free_hot_teams(root, th, 1, __kmp_hot_teams_max_level); | ||||
4082 | } | ||||
4083 | if (th->th.th_hot_teams) { | ||||
4084 | __kmp_free(th->th.th_hot_teams)___kmp_free((th->th.th_hot_teams), "openmp/runtime/src/kmp_runtime.cpp" , 4084); | ||||
4085 | th->th.th_hot_teams = NULL__null; | ||||
4086 | } | ||||
4087 | } | ||||
4088 | } | ||||
4089 | #endif | ||||
4090 | __kmp_free_team(root, hot_team USE_NESTED_HOT_ARG(NULL), __null); | ||||
4091 | |||||
4092 | // Before we can reap the thread, we need to make certain that all other | ||||
4093 | // threads in the teams that had this root as ancestor have stopped trying to | ||||
4094 | // steal tasks. | ||||
4095 | if (__kmp_tasking_mode != tskm_immediate_exec) { | ||||
4096 | __kmp_wait_to_unref_task_teams(); | ||||
4097 | } | ||||
4098 | |||||
4099 | #if KMP_OS_WINDOWS0 | ||||
4100 | /* Close Handle of root duplicated in __kmp_create_worker (tr #62919) */ | ||||
4101 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_reset_root: free handle, th = %p, handle = %" "lu" "\n", (LPVOID) & (root->r.r_uber_thread->th), root->r.r_uber_thread->th.th_info.ds.ds_thread); } | ||||
4102 | 10, ("__kmp_reset_root: free handle, th = %p, handle = %" KMP_UINTPTR_SPECif (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_reset_root: free handle, th = %p, handle = %" "lu" "\n", (LPVOID) & (root->r.r_uber_thread->th), root->r.r_uber_thread->th.th_info.ds.ds_thread); } | ||||
4103 | "\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_reset_root: free handle, th = %p, handle = %" "lu" "\n", (LPVOID) & (root->r.r_uber_thread->th), root->r.r_uber_thread->th.th_info.ds.ds_thread); } | ||||
4104 | (LPVOID) & (root->r.r_uber_thread->th),if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_reset_root: free handle, th = %p, handle = %" "lu" "\n", (LPVOID) & (root->r.r_uber_thread->th), root->r.r_uber_thread->th.th_info.ds.ds_thread); } | ||||
4105 | root->r.r_uber_thread->th.th_info.ds.ds_thread))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_reset_root: free handle, th = %p, handle = %" "lu" "\n", (LPVOID) & (root->r.r_uber_thread->th), root->r.r_uber_thread->th.th_info.ds.ds_thread); }; | ||||
4106 | __kmp_free_handle(root->r.r_uber_thread->th.th_info.ds.ds_thread); | ||||
4107 | #endif /* KMP_OS_WINDOWS */ | ||||
4108 | |||||
4109 | #if OMPD_SUPPORT1 | ||||
4110 | if (ompd_state & OMPD_ENABLE_BP0x1) | ||||
4111 | ompd_bp_thread_end(); | ||||
4112 | #endif | ||||
4113 | |||||
4114 | #if OMPT_SUPPORT1 | ||||
4115 | ompt_data_t *task_data; | ||||
4116 | ompt_data_t *parallel_data; | ||||
4117 | __ompt_get_task_info_internal(0, NULL__null, &task_data, NULL__null, ¶llel_data, | ||||
4118 | NULL__null); | ||||
4119 | if (ompt_enabled.ompt_callback_implicit_task) { | ||||
4120 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | ||||
4121 | ompt_scope_end, parallel_data, task_data, 0, 1, ompt_task_initial); | ||||
4122 | } | ||||
4123 | if (ompt_enabled.ompt_callback_thread_end) { | ||||
4124 | ompt_callbacks.ompt_callback(ompt_callback_thread_end)ompt_callback_thread_end_callback( | ||||
4125 | &(root->r.r_uber_thread->th.ompt_thread_info.thread_data)); | ||||
4126 | } | ||||
4127 | #endif | ||||
4128 | |||||
4129 | TCW_4(__kmp_nth,(__kmp_nth) = (__kmp_nth - 1) | ||||
4130 | __kmp_nth - 1)(__kmp_nth) = (__kmp_nth - 1); // __kmp_reap_thread will decrement __kmp_all_nth. | ||||
4131 | i = root->r.r_uber_thread->th.th_cg_roots->cg_nthreads--; | ||||
4132 | KA_TRACE(100, ("__kmp_reset_root: Thread %p decrement cg_nthreads on node %p"if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_reset_root: Thread %p decrement cg_nthreads on node %p" " to %d\n", root->r.r_uber_thread, root->r.r_uber_thread ->th.th_cg_roots, root->r.r_uber_thread->th.th_cg_roots ->cg_nthreads); } | ||||
4133 | " to %d\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_reset_root: Thread %p decrement cg_nthreads on node %p" " to %d\n", root->r.r_uber_thread, root->r.r_uber_thread ->th.th_cg_roots, root->r.r_uber_thread->th.th_cg_roots ->cg_nthreads); } | ||||
4134 | root->r.r_uber_thread, root->r.r_uber_thread->th.th_cg_roots,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_reset_root: Thread %p decrement cg_nthreads on node %p" " to %d\n", root->r.r_uber_thread, root->r.r_uber_thread ->th.th_cg_roots, root->r.r_uber_thread->th.th_cg_roots ->cg_nthreads); } | ||||
4135 | root->r.r_uber_thread->th.th_cg_roots->cg_nthreads))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_reset_root: Thread %p decrement cg_nthreads on node %p" " to %d\n", root->r.r_uber_thread, root->r.r_uber_thread ->th.th_cg_roots, root->r.r_uber_thread->th.th_cg_roots ->cg_nthreads); }; | ||||
4136 | if (i == 1) { | ||||
4137 | // need to free contention group structure | ||||
4138 | KMP_DEBUG_ASSERT(root->r.r_uber_thread ==if (!(root->r.r_uber_thread == root->r.r_uber_thread-> th.th_cg_roots->cg_root)) { __kmp_debug_assert("root->r.r_uber_thread == root->r.r_uber_thread->th.th_cg_roots->cg_root" , "openmp/runtime/src/kmp_runtime.cpp", 4139); } | ||||
4139 | root->r.r_uber_thread->th.th_cg_roots->cg_root)if (!(root->r.r_uber_thread == root->r.r_uber_thread-> th.th_cg_roots->cg_root)) { __kmp_debug_assert("root->r.r_uber_thread == root->r.r_uber_thread->th.th_cg_roots->cg_root" , "openmp/runtime/src/kmp_runtime.cpp", 4139); }; | ||||
4140 | KMP_DEBUG_ASSERT(root->r.r_uber_thread->th.th_cg_roots->up == NULL)if (!(root->r.r_uber_thread->th.th_cg_roots->up == __null )) { __kmp_debug_assert("root->r.r_uber_thread->th.th_cg_roots->up == __null" , "openmp/runtime/src/kmp_runtime.cpp", 4140); }; | ||||
4141 | __kmp_free(root->r.r_uber_thread->th.th_cg_roots)___kmp_free((root->r.r_uber_thread->th.th_cg_roots), "openmp/runtime/src/kmp_runtime.cpp" , 4141); | ||||
4142 | root->r.r_uber_thread->th.th_cg_roots = NULL__null; | ||||
4143 | } | ||||
4144 | __kmp_reap_thread(root->r.r_uber_thread, 1); | ||||
4145 | |||||
4146 | // We canot put root thread to __kmp_thread_pool, so we have to reap it | ||||
4147 | // instead of freeing. | ||||
4148 | root->r.r_uber_thread = NULL__null; | ||||
4149 | /* mark root as no longer in use */ | ||||
4150 | root->r.r_begin = FALSE0; | ||||
4151 | |||||
4152 | return n; | ||||
4153 | } | ||||
4154 | |||||
4155 | void __kmp_unregister_root_current_thread(int gtid) { | ||||
4156 | KA_TRACE(1, ("__kmp_unregister_root_current_thread: enter T#%d\n", gtid))if (kmp_a_debug >= 1) { __kmp_debug_printf ("__kmp_unregister_root_current_thread: enter T#%d\n" , gtid); }; | ||||
4157 | /* this lock should be ok, since unregister_root_current_thread is never | ||||
4158 | called during an abort, only during a normal close. furthermore, if you | ||||
4159 | have the forkjoin lock, you should never try to get the initz lock */ | ||||
4160 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); | ||||
4161 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done) || !__kmp_init_serial) { | ||||
4162 | KC_TRACE(10, ("__kmp_unregister_root_current_thread: already finished, "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_unregister_root_current_thread: already finished, " "exiting T#%d\n", gtid); } | ||||
4163 | "exiting T#%d\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_unregister_root_current_thread: already finished, " "exiting T#%d\n", gtid); } | ||||
4164 | gtid))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_unregister_root_current_thread: already finished, " "exiting T#%d\n", gtid); }; | ||||
4165 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | ||||
4166 | return; | ||||
4167 | } | ||||
4168 | kmp_root_t *root = __kmp_root[gtid]; | ||||
4169 | |||||
4170 | KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid])if (!(__kmp_threads && __kmp_threads[gtid])) { __kmp_debug_assert ("__kmp_threads && __kmp_threads[gtid]", "openmp/runtime/src/kmp_runtime.cpp" , 4170); }; | ||||
4171 | KMP_ASSERT(KMP_UBER_GTID(gtid))if (!(KMP_UBER_GTID(gtid))) { __kmp_debug_assert("KMP_UBER_GTID(gtid)" , "openmp/runtime/src/kmp_runtime.cpp", 4171); }; | ||||
4172 | KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root)if (!(root == __kmp_threads[gtid]->th.th_root)) { __kmp_debug_assert ("root == __kmp_threads[gtid]->th.th_root", "openmp/runtime/src/kmp_runtime.cpp" , 4172); }; | ||||
4173 | KMP_ASSERT(root->r.r_active == FALSE)if (!(root->r.r_active == 0)) { __kmp_debug_assert("root->r.r_active == FALSE" , "openmp/runtime/src/kmp_runtime.cpp", 4173); }; | ||||
4174 | |||||
4175 | KMP_MB(); | ||||
4176 | |||||
4177 | kmp_info_t *thread = __kmp_threads[gtid]; | ||||
4178 | kmp_team_t *team = thread->th.th_team; | ||||
4179 | kmp_task_team_t *task_team = thread->th.th_task_team; | ||||
4180 | |||||
4181 | // we need to wait for the proxy tasks before finishing the thread | ||||
4182 | if (task_team != NULL__null && (task_team->tt.tt_found_proxy_tasks || | ||||
4183 | task_team->tt.tt_hidden_helper_task_encountered)) { | ||||
4184 | #if OMPT_SUPPORT1 | ||||
4185 | // the runtime is shutting down so we won't report any events | ||||
4186 | thread->th.ompt_thread_info.state = ompt_state_undefined; | ||||
4187 | #endif | ||||
4188 | __kmp_task_team_wait(thread, team USE_ITT_BUILD_ARG(NULL), __null); | ||||
4189 | } | ||||
4190 | |||||
4191 | __kmp_reset_root(gtid, root); | ||||
4192 | |||||
4193 | KMP_MB(); | ||||
4194 | KC_TRACE(10,if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_unregister_root_current_thread: T#%d unregistered\n" , gtid); } | ||||
4195 | ("__kmp_unregister_root_current_thread: T#%d unregistered\n", gtid))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_unregister_root_current_thread: T#%d unregistered\n" , gtid); }; | ||||
4196 | |||||
4197 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | ||||
4198 | } | ||||
4199 | |||||
4200 | #if KMP_OS_WINDOWS0 | ||||
4201 | /* __kmp_forkjoin_lock must be already held | ||||
4202 | Unregisters a root thread that is not the current thread. Returns the number | ||||
4203 | of __kmp_threads entries freed as a result. */ | ||||
4204 | static int __kmp_unregister_root_other_thread(int gtid) { | ||||
4205 | kmp_root_t *root = __kmp_root[gtid]; | ||||
4206 | int r; | ||||
4207 | |||||
4208 | KA_TRACE(1, ("__kmp_unregister_root_other_thread: enter T#%d\n", gtid))if (kmp_a_debug >= 1) { __kmp_debug_printf ("__kmp_unregister_root_other_thread: enter T#%d\n" , gtid); }; | ||||
4209 | KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid])if (!(__kmp_threads && __kmp_threads[gtid])) { __kmp_debug_assert ("__kmp_threads && __kmp_threads[gtid]", "openmp/runtime/src/kmp_runtime.cpp" , 4209); }; | ||||
4210 | KMP_ASSERT(KMP_UBER_GTID(gtid))if (!(KMP_UBER_GTID(gtid))) { __kmp_debug_assert("KMP_UBER_GTID(gtid)" , "openmp/runtime/src/kmp_runtime.cpp", 4210); }; | ||||
4211 | KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root)if (!(root == __kmp_threads[gtid]->th.th_root)) { __kmp_debug_assert ("root == __kmp_threads[gtid]->th.th_root", "openmp/runtime/src/kmp_runtime.cpp" , 4211); }; | ||||
4212 | KMP_ASSERT(root->r.r_active == FALSE)if (!(root->r.r_active == 0)) { __kmp_debug_assert("root->r.r_active == FALSE" , "openmp/runtime/src/kmp_runtime.cpp", 4212); }; | ||||
4213 | |||||
4214 | r = __kmp_reset_root(gtid, root); | ||||
4215 | KC_TRACE(10,if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_unregister_root_other_thread: T#%d unregistered\n" , gtid); } | ||||
4216 | ("__kmp_unregister_root_other_thread: T#%d unregistered\n", gtid))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_unregister_root_other_thread: T#%d unregistered\n" , gtid); }; | ||||
4217 | return r; | ||||
4218 | } | ||||
4219 | #endif | ||||
4220 | |||||
4221 | #if KMP_DEBUG1 | ||||
4222 | void __kmp_task_info() { | ||||
4223 | |||||
4224 | kmp_int32 gtid = __kmp_entry_gtid()__kmp_get_global_thread_id_reg(); | ||||
4225 | kmp_int32 tid = __kmp_tid_from_gtid(gtid); | ||||
4226 | kmp_info_t *this_thr = __kmp_threads[gtid]; | ||||
4227 | kmp_team_t *steam = this_thr->th.th_serial_team; | ||||
4228 | kmp_team_t *team = this_thr->th.th_team; | ||||
4229 | |||||
4230 | __kmp_printf( | ||||
4231 | "__kmp_task_info: gtid=%d tid=%d t_thread=%p team=%p steam=%p curtask=%p " | ||||
4232 | "ptask=%p\n", | ||||
4233 | gtid, tid, this_thr, team, steam, this_thr->th.th_current_task, | ||||
4234 | team->t.t_implicit_task_taskdata[tid].td_parent); | ||||
4235 | } | ||||
4236 | #endif // KMP_DEBUG | ||||
4237 | |||||
4238 | /* TODO optimize with one big memclr, take out what isn't needed, split | ||||
4239 | responsibility to workers as much as possible, and delay initialization of | ||||
4240 | features as much as possible */ | ||||
4241 | static void __kmp_initialize_info(kmp_info_t *this_thr, kmp_team_t *team, | ||||
4242 | int tid, int gtid) { | ||||
4243 | /* this_thr->th.th_info.ds.ds_gtid is setup in | ||||
4244 | kmp_allocate_thread/create_worker. | ||||
4245 | this_thr->th.th_serial_team is setup in __kmp_allocate_thread */ | ||||
4246 | KMP_DEBUG_ASSERT(this_thr != NULL)if (!(this_thr != __null)) { __kmp_debug_assert("this_thr != __null" , "openmp/runtime/src/kmp_runtime.cpp", 4246); }; | ||||
4247 | KMP_DEBUG_ASSERT(this_thr->th.th_serial_team)if (!(this_thr->th.th_serial_team)) { __kmp_debug_assert("this_thr->th.th_serial_team" , "openmp/runtime/src/kmp_runtime.cpp", 4247); }; | ||||
4248 | KMP_DEBUG_ASSERT(team)if (!(team)) { __kmp_debug_assert("team", "openmp/runtime/src/kmp_runtime.cpp" , 4248); }; | ||||
4249 | KMP_DEBUG_ASSERT(team->t.t_threads)if (!(team->t.t_threads)) { __kmp_debug_assert("team->t.t_threads" , "openmp/runtime/src/kmp_runtime.cpp", 4249); }; | ||||
4250 | KMP_DEBUG_ASSERT(team->t.t_dispatch)if (!(team->t.t_dispatch)) { __kmp_debug_assert("team->t.t_dispatch" , "openmp/runtime/src/kmp_runtime.cpp", 4250); }; | ||||
4251 | kmp_info_t *master = team->t.t_threads[0]; | ||||
4252 | KMP_DEBUG_ASSERT(master)if (!(master)) { __kmp_debug_assert("master", "openmp/runtime/src/kmp_runtime.cpp" , 4252); }; | ||||
4253 | KMP_DEBUG_ASSERT(master->th.th_root)if (!(master->th.th_root)) { __kmp_debug_assert("master->th.th_root" , "openmp/runtime/src/kmp_runtime.cpp", 4253); }; | ||||
4254 | |||||
4255 | KMP_MB(); | ||||
4256 | |||||
4257 | TCW_SYNC_PTR(this_thr->th.th_team, team)((this_thr->th.th_team)) = ((team)); | ||||
4258 | |||||
4259 | this_thr->th.th_info.ds.ds_tid = tid; | ||||
4260 | this_thr->th.th_set_nproc = 0; | ||||
4261 | if (__kmp_tasking_mode != tskm_immediate_exec) | ||||
4262 | // When tasking is possible, threads are not safe to reap until they are | ||||
4263 | // done tasking; this will be set when tasking code is exited in wait | ||||
4264 | this_thr->th.th_reap_state = KMP_NOT_SAFE_TO_REAP0; | ||||
4265 | else // no tasking --> always safe to reap | ||||
4266 | this_thr->th.th_reap_state = KMP_SAFE_TO_REAP1; | ||||
4267 | this_thr->th.th_set_proc_bind = proc_bind_default; | ||||
4268 | #if KMP_AFFINITY_SUPPORTED1 | ||||
4269 | this_thr->th.th_new_place = this_thr->th.th_current_place; | ||||
4270 | #endif | ||||
4271 | this_thr->th.th_root = master->th.th_root; | ||||
4272 | |||||
4273 | /* setup the thread's cache of the team structure */ | ||||
4274 | this_thr->th.th_team_nproc = team->t.t_nproc; | ||||
4275 | this_thr->th.th_team_master = master; | ||||
4276 | this_thr->th.th_team_serialized = team->t.t_serialized; | ||||
4277 | |||||
4278 | KMP_DEBUG_ASSERT(team->t.t_implicit_task_taskdata)if (!(team->t.t_implicit_task_taskdata)) { __kmp_debug_assert ("team->t.t_implicit_task_taskdata", "openmp/runtime/src/kmp_runtime.cpp" , 4278); }; | ||||
4279 | |||||
4280 | KF_TRACE(10, ("__kmp_initialize_info1: T#%d:%d this_thread=%p curtask=%p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_info1: T#%d:%d this_thread=%p curtask=%p\n" , tid, gtid, this_thr, this_thr->th.th_current_task); } | ||||
4281 | tid, gtid, this_thr, this_thr->th.th_current_task))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_info1: T#%d:%d this_thread=%p curtask=%p\n" , tid, gtid, this_thr, this_thr->th.th_current_task); }; | ||||
4282 | |||||
4283 | __kmp_init_implicit_task(this_thr->th.th_team_master->th.th_ident, this_thr, | ||||
4284 | team, tid, TRUE(!0)); | ||||
4285 | |||||
4286 | KF_TRACE(10, ("__kmp_initialize_info2: T#%d:%d this_thread=%p curtask=%p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_info2: T#%d:%d this_thread=%p curtask=%p\n" , tid, gtid, this_thr, this_thr->th.th_current_task); } | ||||
4287 | tid, gtid, this_thr, this_thr->th.th_current_task))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_info2: T#%d:%d this_thread=%p curtask=%p\n" , tid, gtid, this_thr, this_thr->th.th_current_task); }; | ||||
4288 | // TODO: Initialize ICVs from parent; GEH - isn't that already done in | ||||
4289 | // __kmp_initialize_team()? | ||||
4290 | |||||
4291 | /* TODO no worksharing in speculative threads */ | ||||
4292 | this_thr->th.th_dispatch = &team->t.t_dispatch[tid]; | ||||
4293 | |||||
4294 | this_thr->th.th_local.this_construct = 0; | ||||
4295 | |||||
4296 | if (!this_thr->th.th_pri_common) { | ||||
4297 | this_thr->th.th_pri_common = | ||||
4298 | (struct common_table *)__kmp_allocate(sizeof(struct common_table))___kmp_allocate((sizeof(struct common_table)), "openmp/runtime/src/kmp_runtime.cpp" , 4298); | ||||
4299 | if (__kmp_storage_map) { | ||||
4300 | __kmp_print_storage_map_gtid( | ||||
4301 | gtid, this_thr->th.th_pri_common, this_thr->th.th_pri_common + 1, | ||||
4302 | sizeof(struct common_table), "th_%d.th_pri_common\n", gtid); | ||||
4303 | } | ||||
4304 | this_thr->th.th_pri_head = NULL__null; | ||||
4305 | } | ||||
4306 | |||||
4307 | if (this_thr != master && // Primary thread's CG root is initialized elsewhere | ||||
4308 | this_thr->th.th_cg_roots != master->th.th_cg_roots) { // CG root not set | ||||
4309 | // Make new thread's CG root same as primary thread's | ||||
4310 | KMP_DEBUG_ASSERT(master->th.th_cg_roots)if (!(master->th.th_cg_roots)) { __kmp_debug_assert("master->th.th_cg_roots" , "openmp/runtime/src/kmp_runtime.cpp", 4310); }; | ||||
4311 | kmp_cg_root_t *tmp = this_thr->th.th_cg_roots; | ||||
4312 | if (tmp) { | ||||
4313 | // worker changes CG, need to check if old CG should be freed | ||||
4314 | int i = tmp->cg_nthreads--; | ||||
4315 | KA_TRACE(100, ("__kmp_initialize_info: Thread %p decrement cg_nthreads"if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_initialize_info: Thread %p decrement cg_nthreads" " on node %p of thread %p to %d\n", this_thr, tmp, tmp->cg_root , tmp->cg_nthreads); } | ||||
4316 | " on node %p of thread %p to %d\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_initialize_info: Thread %p decrement cg_nthreads" " on node %p of thread %p to %d\n", this_thr, tmp, tmp->cg_root , tmp->cg_nthreads); } | ||||
4317 | this_thr, tmp, tmp->cg_root, tmp->cg_nthreads))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_initialize_info: Thread %p decrement cg_nthreads" " on node %p of thread %p to %d\n", this_thr, tmp, tmp->cg_root , tmp->cg_nthreads); }; | ||||
4318 | if (i == 1) { | ||||
4319 | __kmp_free(tmp)___kmp_free((tmp), "openmp/runtime/src/kmp_runtime.cpp", 4319 ); // last thread left CG --> free it | ||||
4320 | } | ||||
4321 | } | ||||
4322 | this_thr->th.th_cg_roots = master->th.th_cg_roots; | ||||
4323 | // Increment new thread's CG root's counter to add the new thread | ||||
4324 | this_thr->th.th_cg_roots->cg_nthreads++; | ||||
4325 | KA_TRACE(100, ("__kmp_initialize_info: Thread %p increment cg_nthreads on"if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_initialize_info: Thread %p increment cg_nthreads on" " node %p of thread %p to %d\n", this_thr, this_thr->th.th_cg_roots , this_thr->th.th_cg_roots->cg_root, this_thr->th.th_cg_roots ->cg_nthreads); } | ||||
4326 | " node %p of thread %p to %d\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_initialize_info: Thread %p increment cg_nthreads on" " node %p of thread %p to %d\n", this_thr, this_thr->th.th_cg_roots , this_thr->th.th_cg_roots->cg_root, this_thr->th.th_cg_roots ->cg_nthreads); } | ||||
4327 | this_thr, this_thr->th.th_cg_roots,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_initialize_info: Thread %p increment cg_nthreads on" " node %p of thread %p to %d\n", this_thr, this_thr->th.th_cg_roots , this_thr->th.th_cg_roots->cg_root, this_thr->th.th_cg_roots ->cg_nthreads); } | ||||
4328 | this_thr->th.th_cg_roots->cg_root,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_initialize_info: Thread %p increment cg_nthreads on" " node %p of thread %p to %d\n", this_thr, this_thr->th.th_cg_roots , this_thr->th.th_cg_roots->cg_root, this_thr->th.th_cg_roots ->cg_nthreads); } | ||||
4329 | this_thr->th.th_cg_roots->cg_nthreads))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_initialize_info: Thread %p increment cg_nthreads on" " node %p of thread %p to %d\n", this_thr, this_thr->th.th_cg_roots , this_thr->th.th_cg_roots->cg_root, this_thr->th.th_cg_roots ->cg_nthreads); }; | ||||
4330 | this_thr->th.th_current_task->td_icvs.thread_limit = | ||||
4331 | this_thr->th.th_cg_roots->cg_thread_limit; | ||||
4332 | } | ||||
4333 | |||||
4334 | /* Initialize dynamic dispatch */ | ||||
4335 | { | ||||
4336 | volatile kmp_disp_t *dispatch = this_thr->th.th_dispatch; | ||||
4337 | // Use team max_nproc since this will never change for the team. | ||||
4338 | size_t disp_size = | ||||
4339 | sizeof(dispatch_private_info_t) * | ||||
4340 | (team->t.t_max_nproc == 1 ? 1 : __kmp_dispatch_num_buffers); | ||||
4341 | KD_TRACE(10, ("__kmp_initialize_info: T#%d max_nproc: %d\n", gtid,if (kmp_d_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_info: T#%d max_nproc: %d\n" , gtid, team->t.t_max_nproc); } | ||||
4342 | team->t.t_max_nproc))if (kmp_d_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_info: T#%d max_nproc: %d\n" , gtid, team->t.t_max_nproc); }; | ||||
4343 | KMP_ASSERT(dispatch)if (!(dispatch)) { __kmp_debug_assert("dispatch", "openmp/runtime/src/kmp_runtime.cpp" , 4343); }; | ||||
4344 | KMP_DEBUG_ASSERT(team->t.t_dispatch)if (!(team->t.t_dispatch)) { __kmp_debug_assert("team->t.t_dispatch" , "openmp/runtime/src/kmp_runtime.cpp", 4344); }; | ||||
4345 | KMP_DEBUG_ASSERT(dispatch == &team->t.t_dispatch[tid])if (!(dispatch == &team->t.t_dispatch[tid])) { __kmp_debug_assert ("dispatch == &team->t.t_dispatch[tid]", "openmp/runtime/src/kmp_runtime.cpp" , 4345); }; | ||||
4346 | |||||
4347 | dispatch->th_disp_index = 0; | ||||
4348 | dispatch->th_doacross_buf_idx = 0; | ||||
4349 | if (!dispatch->th_disp_buffer) { | ||||
4350 | dispatch->th_disp_buffer = | ||||
4351 | (dispatch_private_info_t *)__kmp_allocate(disp_size)___kmp_allocate((disp_size), "openmp/runtime/src/kmp_runtime.cpp" , 4351); | ||||
4352 | |||||
4353 | if (__kmp_storage_map) { | ||||
4354 | __kmp_print_storage_map_gtid( | ||||
4355 | gtid, &dispatch->th_disp_buffer[0], | ||||
4356 | &dispatch->th_disp_buffer[team->t.t_max_nproc == 1 | ||||
4357 | ? 1 | ||||
4358 | : __kmp_dispatch_num_buffers], | ||||
4359 | disp_size, | ||||
4360 | "th_%d.th_dispatch.th_disp_buffer " | ||||
4361 | "(team_%d.t_dispatch[%d].th_disp_buffer)", | ||||
4362 | gtid, team->t.t_id, gtid); | ||||
4363 | } | ||||
4364 | } else { | ||||
4365 | memset(&dispatch->th_disp_buffer[0], '\0', disp_size); | ||||
4366 | } | ||||
4367 | |||||
4368 | dispatch->th_dispatch_pr_current = 0; | ||||
4369 | dispatch->th_dispatch_sh_current = 0; | ||||
4370 | |||||
4371 | dispatch->th_deo_fcn = 0; /* ORDERED */ | ||||
4372 | dispatch->th_dxo_fcn = 0; /* END ORDERED */ | ||||
4373 | } | ||||
4374 | |||||
4375 | this_thr->th.th_next_pool = NULL__null; | ||||
4376 | |||||
4377 | if (!this_thr->th.th_task_state_memo_stack) { | ||||
4378 | size_t i; | ||||
4379 | this_thr->th.th_task_state_memo_stack = | ||||
4380 | (kmp_uint8 *)__kmp_allocate(4 * sizeof(kmp_uint8))___kmp_allocate((4 * sizeof(kmp_uint8)), "openmp/runtime/src/kmp_runtime.cpp" , 4380); | ||||
4381 | this_thr->th.th_task_state_top = 0; | ||||
4382 | this_thr->th.th_task_state_stack_sz = 4; | ||||
4383 | for (i = 0; i < this_thr->th.th_task_state_stack_sz; | ||||
4384 | ++i) // zero init the stack | ||||
4385 | this_thr->th.th_task_state_memo_stack[i] = 0; | ||||
4386 | } | ||||
4387 | |||||
4388 | KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here)if (!(!this_thr->th.th_spin_here)) { __kmp_debug_assert("!this_thr->th.th_spin_here" , "openmp/runtime/src/kmp_runtime.cpp", 4388); }; | ||||
4389 | KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0)if (!(this_thr->th.th_next_waiting == 0)) { __kmp_debug_assert ("this_thr->th.th_next_waiting == 0", "openmp/runtime/src/kmp_runtime.cpp" , 4389); }; | ||||
4390 | |||||
4391 | KMP_MB(); | ||||
4392 | } | ||||
4393 | |||||
4394 | /* allocate a new thread for the requesting team. this is only called from | ||||
4395 | within a forkjoin critical section. we will first try to get an available | ||||
4396 | thread from the thread pool. if none is available, we will fork a new one | ||||
4397 | assuming we are able to create a new one. this should be assured, as the | ||||
4398 | caller should check on this first. */ | ||||
4399 | kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team, | ||||
4400 | int new_tid) { | ||||
4401 | kmp_team_t *serial_team; | ||||
4402 | kmp_info_t *new_thr; | ||||
4403 | int new_gtid; | ||||
4404 | |||||
4405 | KA_TRACE(20, ("__kmp_allocate_thread: T#%d\n", __kmp_get_gtid()))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_thread: T#%d\n" , __kmp_get_global_thread_id()); }; | ||||
4406 | KMP_DEBUG_ASSERT(root && team)if (!(root && team)) { __kmp_debug_assert("root && team" , "openmp/runtime/src/kmp_runtime.cpp", 4406); }; | ||||
4407 | #if !KMP_NESTED_HOT_TEAMS1 | ||||
4408 | KMP_DEBUG_ASSERT(KMP_MASTER_GTID(__kmp_get_gtid()))if (!((0 == __kmp_tid_from_gtid((__kmp_get_global_thread_id() ))))) { __kmp_debug_assert("(0 == __kmp_tid_from_gtid((__kmp_get_global_thread_id())))" , "openmp/runtime/src/kmp_runtime.cpp", 4408); }; | ||||
4409 | #endif | ||||
4410 | KMP_MB(); | ||||
4411 | |||||
4412 | /* first, try to get one from the thread pool */ | ||||
4413 | if (__kmp_thread_pool) { | ||||
4414 | new_thr = CCAST(kmp_info_t *, __kmp_thread_pool)const_cast<kmp_info_t *>(__kmp_thread_pool); | ||||
4415 | __kmp_thread_pool = (volatile kmp_info_t *)new_thr->th.th_next_pool; | ||||
4416 | if (new_thr == __kmp_thread_pool_insert_pt) { | ||||
4417 | __kmp_thread_pool_insert_pt = NULL__null; | ||||
4418 | } | ||||
4419 | TCW_4(new_thr->th.th_in_pool, FALSE)(new_thr->th.th_in_pool) = (0); | ||||
4420 | __kmp_suspend_initialize_thread(new_thr); | ||||
4421 | __kmp_lock_suspend_mx(new_thr); | ||||
4422 | if (new_thr->th.th_active_in_pool == TRUE(!0)) { | ||||
4423 | KMP_DEBUG_ASSERT(new_thr->th.th_active == TRUE)if (!(new_thr->th.th_active == (!0))) { __kmp_debug_assert ("new_thr->th.th_active == (!0)", "openmp/runtime/src/kmp_runtime.cpp" , 4423); }; | ||||
4424 | KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth)(&__kmp_thread_pool_active_nth)->fetch_sub(1, std::memory_order_acq_rel ); | ||||
4425 | new_thr->th.th_active_in_pool = FALSE0; | ||||
4426 | } | ||||
4427 | __kmp_unlock_suspend_mx(new_thr); | ||||
4428 | |||||
4429 | KA_TRACE(20, ("__kmp_allocate_thread: T#%d using thread T#%d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_thread: T#%d using thread T#%d\n" , __kmp_get_global_thread_id(), new_thr->th.th_info.ds.ds_gtid ); } | ||||
4430 | __kmp_get_gtid(), new_thr->th.th_info.ds.ds_gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_thread: T#%d using thread T#%d\n" , __kmp_get_global_thread_id(), new_thr->th.th_info.ds.ds_gtid ); }; | ||||
4431 | KMP_ASSERT(!new_thr->th.th_team)if (!(!new_thr->th.th_team)) { __kmp_debug_assert("!new_thr->th.th_team" , "openmp/runtime/src/kmp_runtime.cpp", 4431); }; | ||||
4432 | KMP_DEBUG_ASSERT(__kmp_nth < __kmp_threads_capacity)if (!(__kmp_nth < __kmp_threads_capacity)) { __kmp_debug_assert ("__kmp_nth < __kmp_threads_capacity", "openmp/runtime/src/kmp_runtime.cpp" , 4432); }; | ||||
4433 | |||||
4434 | /* setup the thread structure */ | ||||
4435 | __kmp_initialize_info(new_thr, team, new_tid, | ||||
4436 | new_thr->th.th_info.ds.ds_gtid); | ||||
4437 | KMP_DEBUG_ASSERT(new_thr->th.th_serial_team)if (!(new_thr->th.th_serial_team)) { __kmp_debug_assert("new_thr->th.th_serial_team" , "openmp/runtime/src/kmp_runtime.cpp", 4437); }; | ||||
4438 | |||||
4439 | TCW_4(__kmp_nth, __kmp_nth + 1)(__kmp_nth) = (__kmp_nth + 1); | ||||
4440 | |||||
4441 | new_thr->th.th_task_state = 0; | ||||
4442 | new_thr->th.th_task_state_top = 0; | ||||
4443 | new_thr->th.th_task_state_stack_sz = 4; | ||||
4444 | |||||
4445 | if (__kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | ||||
4446 | // Make sure pool thread has transitioned to waiting on own thread struct | ||||
4447 | KMP_DEBUG_ASSERT(new_thr->th.th_used_in_team.load() == 0)if (!(new_thr->th.th_used_in_team.load() == 0)) { __kmp_debug_assert ("new_thr->th.th_used_in_team.load() == 0", "openmp/runtime/src/kmp_runtime.cpp" , 4447); }; | ||||
4448 | // Thread activated in __kmp_allocate_team when increasing team size | ||||
4449 | } | ||||
4450 | |||||
4451 | #ifdef KMP_ADJUST_BLOCKTIME1 | ||||
4452 | /* Adjust blocktime back to zero if necessary */ | ||||
4453 | /* Middle initialization might not have occurred yet */ | ||||
4454 | if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { | ||||
4455 | if (__kmp_nth > __kmp_avail_proc) { | ||||
4456 | __kmp_zero_bt = TRUE(!0); | ||||
4457 | } | ||||
4458 | } | ||||
4459 | #endif /* KMP_ADJUST_BLOCKTIME */ | ||||
4460 | |||||
4461 | #if KMP_DEBUG1 | ||||
4462 | // If thread entered pool via __kmp_free_thread, wait_flag should != | ||||
4463 | // KMP_BARRIER_PARENT_FLAG. | ||||
4464 | int b; | ||||
4465 | kmp_balign_t *balign = new_thr->th.th_bar; | ||||
4466 | for (b = 0; b < bs_last_barrier; ++b) | ||||
4467 | KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG)if (!(balign[b].bb.wait_flag != 2)) { __kmp_debug_assert("balign[b].bb.wait_flag != 2" , "openmp/runtime/src/kmp_runtime.cpp", 4467); }; | ||||
4468 | #endif | ||||
4469 | |||||
4470 | KF_TRACE(10, ("__kmp_allocate_thread: T#%d using thread %p T#%d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: T#%d using thread %p T#%d\n" , __kmp_get_global_thread_id(), new_thr, new_thr->th.th_info .ds.ds_gtid); } | ||||
4471 | __kmp_get_gtid(), new_thr, new_thr->th.th_info.ds.ds_gtid))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: T#%d using thread %p T#%d\n" , __kmp_get_global_thread_id(), new_thr, new_thr->th.th_info .ds.ds_gtid); }; | ||||
4472 | |||||
4473 | KMP_MB(); | ||||
4474 | return new_thr; | ||||
4475 | } | ||||
4476 | |||||
4477 | /* no, well fork a new one */ | ||||
4478 | KMP_ASSERT(__kmp_nth == __kmp_all_nth)if (!(__kmp_nth == __kmp_all_nth)) { __kmp_debug_assert("__kmp_nth == __kmp_all_nth" , "openmp/runtime/src/kmp_runtime.cpp", 4478); }; | ||||
4479 | KMP_ASSERT(__kmp_all_nth < __kmp_threads_capacity)if (!(__kmp_all_nth < __kmp_threads_capacity)) { __kmp_debug_assert ("__kmp_all_nth < __kmp_threads_capacity", "openmp/runtime/src/kmp_runtime.cpp" , 4479); }; | ||||
4480 | |||||
4481 | #if KMP_USE_MONITOR | ||||
4482 | // If this is the first worker thread the RTL is creating, then also | ||||
4483 | // launch the monitor thread. We try to do this as early as possible. | ||||
4484 | if (!TCR_4(__kmp_init_monitor)(__kmp_init_monitor)) { | ||||
4485 | __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock); | ||||
4486 | if (!TCR_4(__kmp_init_monitor)(__kmp_init_monitor)) { | ||||
4487 | KF_TRACE(10, ("before __kmp_create_monitor\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("before __kmp_create_monitor\n" ); }; | ||||
4488 | TCW_4(__kmp_init_monitor, 1)(__kmp_init_monitor) = (1); | ||||
4489 | __kmp_create_monitor(&__kmp_monitor); | ||||
4490 | KF_TRACE(10, ("after __kmp_create_monitor\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("after __kmp_create_monitor\n" ); }; | ||||
4491 | #if KMP_OS_WINDOWS0 | ||||
4492 | // AC: wait until monitor has started. This is a fix for CQ232808. | ||||
4493 | // The reason is that if the library is loaded/unloaded in a loop with | ||||
4494 | // small (parallel) work in between, then there is high probability that | ||||
4495 | // monitor thread started after the library shutdown. At shutdown it is | ||||
4496 | // too late to cope with the problem, because when the primary thread is | ||||
4497 | // in DllMain (process detach) the monitor has no chances to start (it is | ||||
4498 | // blocked), and primary thread has no means to inform the monitor that | ||||
4499 | // the library has gone, because all the memory which the monitor can | ||||
4500 | // access is going to be released/reset. | ||||
4501 | while (TCR_4(__kmp_init_monitor)(__kmp_init_monitor) < 2) { | ||||
4502 | KMP_YIELD(TRUE){ __kmp_x86_pause(); if (((!0)) && (((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))))))) __kmp_yield (); }; | ||||
4503 | } | ||||
4504 | KF_TRACE(10, ("after monitor thread has started\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("after monitor thread has started\n" ); }; | ||||
4505 | #endif | ||||
4506 | } | ||||
4507 | __kmp_release_bootstrap_lock(&__kmp_monitor_lock); | ||||
4508 | } | ||||
4509 | #endif | ||||
4510 | |||||
4511 | KMP_MB(); | ||||
4512 | |||||
4513 | { | ||||
4514 | int new_start_gtid = TCR_4(__kmp_init_hidden_helper_threads)(__kmp_init_hidden_helper_threads) | ||||
4515 | ? 1 | ||||
4516 | : __kmp_hidden_helper_threads_num + 1; | ||||
4517 | |||||
4518 | for (new_gtid = new_start_gtid; TCR_PTR(__kmp_threads[new_gtid])((void *)(__kmp_threads[new_gtid])) != NULL__null; | ||||
4519 | ++new_gtid) { | ||||
4520 | KMP_DEBUG_ASSERT(new_gtid < __kmp_threads_capacity)if (!(new_gtid < __kmp_threads_capacity)) { __kmp_debug_assert ("new_gtid < __kmp_threads_capacity", "openmp/runtime/src/kmp_runtime.cpp" , 4520); }; | ||||
4521 | } | ||||
4522 | |||||
4523 | if (TCR_4(__kmp_init_hidden_helper_threads)(__kmp_init_hidden_helper_threads)) { | ||||
4524 | KMP_DEBUG_ASSERT(new_gtid <= __kmp_hidden_helper_threads_num)if (!(new_gtid <= __kmp_hidden_helper_threads_num)) { __kmp_debug_assert ("new_gtid <= __kmp_hidden_helper_threads_num", "openmp/runtime/src/kmp_runtime.cpp" , 4524); }; | ||||
4525 | } | ||||
4526 | } | ||||
4527 | |||||
4528 | /* allocate space for it. */ | ||||
4529 | new_thr = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t))___kmp_allocate((sizeof(kmp_info_t)), "openmp/runtime/src/kmp_runtime.cpp" , 4529); | ||||
4530 | |||||
4531 | TCW_SYNC_PTR(__kmp_threads[new_gtid], new_thr)((__kmp_threads[new_gtid])) = ((new_thr)); | ||||
4532 | |||||
4533 | #if USE_ITT_BUILD1 && USE_ITT_NOTIFY1 && KMP_DEBUG1 | ||||
4534 | // suppress race conditions detection on synchronization flags in debug mode | ||||
4535 | // this helps to analyze library internals eliminating false positives | ||||
4536 | __itt_suppress_mark_range(!__kmp_itt_suppress_mark_range_ptr__3_0) ? (void)0 : __kmp_itt_suppress_mark_range_ptr__3_0( | ||||
4537 | __itt_suppress_range, __itt_suppress_threading_errors0x000000ff, | ||||
4538 | &new_thr->th.th_sleep_loc, sizeof(new_thr->th.th_sleep_loc)); | ||||
4539 | __itt_suppress_mark_range(!__kmp_itt_suppress_mark_range_ptr__3_0) ? (void)0 : __kmp_itt_suppress_mark_range_ptr__3_0( | ||||
4540 | __itt_suppress_range, __itt_suppress_threading_errors0x000000ff, | ||||
4541 | &new_thr->th.th_reap_state, sizeof(new_thr->th.th_reap_state)); | ||||
4542 | #if KMP_OS_WINDOWS0 | ||||
4543 | __itt_suppress_mark_range(!__kmp_itt_suppress_mark_range_ptr__3_0) ? (void)0 : __kmp_itt_suppress_mark_range_ptr__3_0( | ||||
4544 | __itt_suppress_range, __itt_suppress_threading_errors0x000000ff, | ||||
4545 | &new_thr->th.th_suspend_init, sizeof(new_thr->th.th_suspend_init)); | ||||
4546 | #else | ||||
4547 | __itt_suppress_mark_range(!__kmp_itt_suppress_mark_range_ptr__3_0) ? (void)0 : __kmp_itt_suppress_mark_range_ptr__3_0(__itt_suppress_range, | ||||
4548 | __itt_suppress_threading_errors0x000000ff, | ||||
4549 | &new_thr->th.th_suspend_init_count, | ||||
4550 | sizeof(new_thr->th.th_suspend_init_count)); | ||||
4551 | #endif | ||||
4552 | // TODO: check if we need to also suppress b_arrived flags | ||||
4553 | __itt_suppress_mark_range(!__kmp_itt_suppress_mark_range_ptr__3_0) ? (void)0 : __kmp_itt_suppress_mark_range_ptr__3_0(__itt_suppress_range, | ||||
4554 | __itt_suppress_threading_errors0x000000ff, | ||||
4555 | CCAST(kmp_uint64 *, &new_thr->th.th_bar[0].bb.b_go)const_cast<kmp_uint64 *>(&new_thr->th.th_bar[0]. bb.b_go), | ||||
4556 | sizeof(new_thr->th.th_bar[0].bb.b_go)); | ||||
4557 | __itt_suppress_mark_range(!__kmp_itt_suppress_mark_range_ptr__3_0) ? (void)0 : __kmp_itt_suppress_mark_range_ptr__3_0(__itt_suppress_range, | ||||
4558 | __itt_suppress_threading_errors0x000000ff, | ||||
4559 | CCAST(kmp_uint64 *, &new_thr->th.th_bar[1].bb.b_go)const_cast<kmp_uint64 *>(&new_thr->th.th_bar[1]. bb.b_go), | ||||
4560 | sizeof(new_thr->th.th_bar[1].bb.b_go)); | ||||
4561 | __itt_suppress_mark_range(!__kmp_itt_suppress_mark_range_ptr__3_0) ? (void)0 : __kmp_itt_suppress_mark_range_ptr__3_0(__itt_suppress_range, | ||||
4562 | __itt_suppress_threading_errors0x000000ff, | ||||
4563 | CCAST(kmp_uint64 *, &new_thr->th.th_bar[2].bb.b_go)const_cast<kmp_uint64 *>(&new_thr->th.th_bar[2]. bb.b_go), | ||||
4564 | sizeof(new_thr->th.th_bar[2].bb.b_go)); | ||||
4565 | #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY && KMP_DEBUG */ | ||||
4566 | if (__kmp_storage_map) { | ||||
4567 | __kmp_print_thread_storage_map(new_thr, new_gtid); | ||||
4568 | } | ||||
4569 | |||||
4570 | // add the reserve serialized team, initialized from the team's primary thread | ||||
4571 | { | ||||
4572 | kmp_internal_control_t r_icvs = __kmp_get_x_global_icvs(team); | ||||
4573 | KF_TRACE(10, ("__kmp_allocate_thread: before th_serial/serial_team\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: before th_serial/serial_team\n" ); }; | ||||
4574 | new_thr->th.th_serial_team = serial_team = | ||||
4575 | (kmp_team_t *)__kmp_allocate_team(root, 1, 1, | ||||
4576 | #if OMPT_SUPPORT1 | ||||
4577 | ompt_data_none{0}, // root parallel id | ||||
4578 | #endif | ||||
4579 | proc_bind_default, &r_icvs, | ||||
4580 | 0 USE_NESTED_HOT_ARG(NULL), __null); | ||||
4581 | } | ||||
4582 | KMP_ASSERT(serial_team)if (!(serial_team)) { __kmp_debug_assert("serial_team", "openmp/runtime/src/kmp_runtime.cpp" , 4582); }; | ||||
4583 | serial_team->t.t_serialized = 0; // AC: the team created in reserve, not for | ||||
4584 | // execution (it is unused for now). | ||||
4585 | serial_team->t.t_threads[0] = new_thr; | ||||
4586 | KF_TRACE(10,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n" , new_thr); } | ||||
4587 | ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n" , new_thr); } | ||||
4588 | new_thr))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n" , new_thr); }; | ||||
4589 | |||||
4590 | /* setup the thread structures */ | ||||
4591 | __kmp_initialize_info(new_thr, team, new_tid, new_gtid); | ||||
4592 | |||||
4593 | #if USE_FAST_MEMORY3 | ||||
4594 | __kmp_initialize_fast_memory(new_thr); | ||||
4595 | #endif /* USE_FAST_MEMORY */ | ||||
4596 | |||||
4597 | #if KMP_USE_BGET1 | ||||
4598 | KMP_DEBUG_ASSERT(new_thr->th.th_local.bget_data == NULL)if (!(new_thr->th.th_local.bget_data == __null)) { __kmp_debug_assert ("new_thr->th.th_local.bget_data == __null", "openmp/runtime/src/kmp_runtime.cpp" , 4598); }; | ||||
4599 | __kmp_initialize_bget(new_thr); | ||||
4600 | #endif | ||||
4601 | |||||
4602 | __kmp_init_random(new_thr); // Initialize random number generator | ||||
4603 | |||||
4604 | /* Initialize these only once when thread is grabbed for a team allocation */ | ||||
4605 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_thread: T#%d init go fork=%u, plain=%u\n" , __kmp_get_global_thread_id(), 0, 0); } | ||||
4606 | ("__kmp_allocate_thread: T#%d init go fork=%u, plain=%u\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_thread: T#%d init go fork=%u, plain=%u\n" , __kmp_get_global_thread_id(), 0, 0); } | ||||
4607 | __kmp_get_gtid(), KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_thread: T#%d init go fork=%u, plain=%u\n" , __kmp_get_global_thread_id(), 0, 0); }; | ||||
4608 | |||||
4609 | int b; | ||||
4610 | kmp_balign_t *balign = new_thr->th.th_bar; | ||||
4611 | for (b = 0; b < bs_last_barrier; ++b) { | ||||
4612 | balign[b].bb.b_go = KMP_INIT_BARRIER_STATE0; | ||||
4613 | balign[b].bb.team = NULL__null; | ||||
4614 | balign[b].bb.wait_flag = KMP_BARRIER_NOT_WAITING0; | ||||
4615 | balign[b].bb.use_oncore_barrier = 0; | ||||
4616 | } | ||||
4617 | |||||
4618 | TCW_PTR(new_thr->th.th_sleep_loc, NULL)((new_thr->th.th_sleep_loc)) = ((__null)); | ||||
4619 | new_thr->th.th_sleep_loc_type = flag_unset; | ||||
4620 | |||||
4621 | new_thr->th.th_spin_here = FALSE0; | ||||
4622 | new_thr->th.th_next_waiting = 0; | ||||
4623 | #if KMP_OS_UNIX1 | ||||
4624 | new_thr->th.th_blocking = false; | ||||
4625 | #endif | ||||
4626 | |||||
4627 | #if KMP_AFFINITY_SUPPORTED1 | ||||
4628 | new_thr->th.th_current_place = KMP_PLACE_UNDEFINED(-2); | ||||
4629 | new_thr->th.th_new_place = KMP_PLACE_UNDEFINED(-2); | ||||
4630 | new_thr->th.th_first_place = KMP_PLACE_UNDEFINED(-2); | ||||
4631 | new_thr->th.th_last_place = KMP_PLACE_UNDEFINED(-2); | ||||
4632 | #endif | ||||
4633 | new_thr->th.th_def_allocator = __kmp_def_allocator; | ||||
4634 | new_thr->th.th_prev_level = 0; | ||||
4635 | new_thr->th.th_prev_num_threads = 1; | ||||
4636 | |||||
4637 | TCW_4(new_thr->th.th_in_pool, FALSE)(new_thr->th.th_in_pool) = (0); | ||||
4638 | new_thr->th.th_active_in_pool = FALSE0; | ||||
4639 | TCW_4(new_thr->th.th_active, TRUE)(new_thr->th.th_active) = ((!0)); | ||||
4640 | |||||
4641 | /* adjust the global counters */ | ||||
4642 | __kmp_all_nth++; | ||||
4643 | __kmp_nth++; | ||||
4644 | |||||
4645 | // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low | ||||
4646 | // numbers of procs, and method #2 (keyed API call) for higher numbers. | ||||
4647 | if (__kmp_adjust_gtid_mode) { | ||||
4648 | if (__kmp_all_nth >= __kmp_tls_gtid_min) { | ||||
4649 | if (TCR_4(__kmp_gtid_mode)(__kmp_gtid_mode) != 2) { | ||||
4650 | TCW_4(__kmp_gtid_mode, 2)(__kmp_gtid_mode) = (2); | ||||
4651 | } | ||||
4652 | } else { | ||||
4653 | if (TCR_4(__kmp_gtid_mode)(__kmp_gtid_mode) != 1) { | ||||
4654 | TCW_4(__kmp_gtid_mode, 1)(__kmp_gtid_mode) = (1); | ||||
4655 | } | ||||
4656 | } | ||||
4657 | } | ||||
4658 | |||||
4659 | #ifdef KMP_ADJUST_BLOCKTIME1 | ||||
4660 | /* Adjust blocktime back to zero if necessary */ | ||||
4661 | /* Middle initialization might not have occurred yet */ | ||||
4662 | if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { | ||||
4663 | if (__kmp_nth > __kmp_avail_proc) { | ||||
4664 | __kmp_zero_bt = TRUE(!0); | ||||
4665 | } | ||||
4666 | } | ||||
4667 | #endif /* KMP_ADJUST_BLOCKTIME */ | ||||
4668 | |||||
4669 | /* actually fork it and create the new worker thread */ | ||||
4670 | KF_TRACE(if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: before __kmp_create_worker: %p\n" , new_thr); } | ||||
4671 | 10, ("__kmp_allocate_thread: before __kmp_create_worker: %p\n", new_thr))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: before __kmp_create_worker: %p\n" , new_thr); }; | ||||
4672 | __kmp_create_worker(new_gtid, new_thr, __kmp_stksize); | ||||
4673 | KF_TRACE(10,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: after __kmp_create_worker: %p\n" , new_thr); } | ||||
4674 | ("__kmp_allocate_thread: after __kmp_create_worker: %p\n", new_thr))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: after __kmp_create_worker: %p\n" , new_thr); }; | ||||
4675 | |||||
4676 | KA_TRACE(20, ("__kmp_allocate_thread: T#%d forked T#%d\n", __kmp_get_gtid(),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_thread: T#%d forked T#%d\n" , __kmp_get_global_thread_id(), new_gtid); } | ||||
4677 | new_gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_thread: T#%d forked T#%d\n" , __kmp_get_global_thread_id(), new_gtid); }; | ||||
4678 | KMP_MB(); | ||||
4679 | return new_thr; | ||||
4680 | } | ||||
4681 | |||||
4682 | /* Reinitialize team for reuse. | ||||
4683 | The hot team code calls this case at every fork barrier, so EPCC barrier | ||||
4684 | test are extremely sensitive to changes in it, esp. writes to the team | ||||
4685 | struct, which cause a cache invalidation in all threads. | ||||
4686 | IF YOU TOUCH THIS ROUTINE, RUN EPCC C SYNCBENCH ON A BIG-IRON MACHINE!!! */ | ||||
4687 | static void __kmp_reinitialize_team(kmp_team_t *team, | ||||
4688 | kmp_internal_control_t *new_icvs, | ||||
4689 | ident_t *loc) { | ||||
4690 | KF_TRACE(10, ("__kmp_reinitialize_team: enter this_thread=%p team=%p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_reinitialize_team: enter this_thread=%p team=%p\n" , team->t.t_threads[0], team); } | ||||
4691 | team->t.t_threads[0], team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_reinitialize_team: enter this_thread=%p team=%p\n" , team->t.t_threads[0], team); }; | ||||
4692 | KMP_DEBUG_ASSERT(team && new_icvs)if (!(team && new_icvs)) { __kmp_debug_assert("team && new_icvs" , "openmp/runtime/src/kmp_runtime.cpp", 4692); }; | ||||
4693 | KMP_DEBUG_ASSERT((!TCR_4(__kmp_init_parallel)) || new_icvs->nproc)if (!((!(__kmp_init_parallel)) || new_icvs->nproc)) { __kmp_debug_assert ("(!(__kmp_init_parallel)) || new_icvs->nproc", "openmp/runtime/src/kmp_runtime.cpp" , 4693); }; | ||||
4694 | KMP_CHECK_UPDATE(team->t.t_ident, loc)if ((team->t.t_ident) != (loc)) (team->t.t_ident) = (loc ); | ||||
4695 | |||||
4696 | KMP_CHECK_UPDATE(team->t.t_id, KMP_GEN_TEAM_ID())if ((team->t.t_id) != ((~0))) (team->t.t_id) = ((~0)); | ||||
4697 | // Copy ICVs to the primary thread's implicit taskdata | ||||
4698 | __kmp_init_implicit_task(loc, team->t.t_threads[0], team, 0, FALSE0); | ||||
4699 | copy_icvs(&team->t.t_implicit_task_taskdata[0].td_icvs, new_icvs); | ||||
4700 | |||||
4701 | KF_TRACE(10, ("__kmp_reinitialize_team: exit this_thread=%p team=%p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_reinitialize_team: exit this_thread=%p team=%p\n" , team->t.t_threads[0], team); } | ||||
4702 | team->t.t_threads[0], team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_reinitialize_team: exit this_thread=%p team=%p\n" , team->t.t_threads[0], team); }; | ||||
4703 | } | ||||
4704 | |||||
4705 | /* Initialize the team data structure. | ||||
4706 | This assumes the t_threads and t_max_nproc are already set. | ||||
4707 | Also, we don't touch the arguments */ | ||||
4708 | static void __kmp_initialize_team(kmp_team_t *team, int new_nproc, | ||||
4709 | kmp_internal_control_t *new_icvs, | ||||
4710 | ident_t *loc) { | ||||
4711 | KF_TRACE(10, ("__kmp_initialize_team: enter: team=%p\n", team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_team: enter: team=%p\n" , team); }; | ||||
4712 | |||||
4713 | /* verify */ | ||||
4714 | KMP_DEBUG_ASSERT(team)if (!(team)) { __kmp_debug_assert("team", "openmp/runtime/src/kmp_runtime.cpp" , 4714); }; | ||||
4715 | KMP_DEBUG_ASSERT(new_nproc <= team->t.t_max_nproc)if (!(new_nproc <= team->t.t_max_nproc)) { __kmp_debug_assert ("new_nproc <= team->t.t_max_nproc", "openmp/runtime/src/kmp_runtime.cpp" , 4715); }; | ||||
4716 | KMP_DEBUG_ASSERT(team->t.t_threads)if (!(team->t.t_threads)) { __kmp_debug_assert("team->t.t_threads" , "openmp/runtime/src/kmp_runtime.cpp", 4716); }; | ||||
4717 | KMP_MB(); | ||||
4718 | |||||
4719 | team->t.t_master_tid = 0; /* not needed */ | ||||
4720 | /* team->t.t_master_bar; not needed */ | ||||
4721 | team->t.t_serialized = new_nproc > 1 ? 0 : 1; | ||||
4722 | team->t.t_nproc = new_nproc; | ||||
4723 | |||||
4724 | /* team->t.t_parent = NULL; TODO not needed & would mess up hot team */ | ||||
4725 | team->t.t_next_pool = NULL__null; | ||||
4726 | /* memset( team->t.t_threads, 0, sizeof(kmp_info_t*)*new_nproc ); would mess | ||||
4727 | * up hot team */ | ||||
4728 | |||||
4729 | TCW_SYNC_PTR(team->t.t_pkfn, NULL)((team->t.t_pkfn)) = ((__null)); /* not needed */ | ||||
4730 | team->t.t_invoke = NULL__null; /* not needed */ | ||||
4731 | |||||
4732 | // TODO???: team->t.t_max_active_levels = new_max_active_levels; | ||||
4733 | team->t.t_sched.sched = new_icvs->sched.sched; | ||||
4734 | |||||
4735 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 | ||||
4736 | team->t.t_fp_control_saved = FALSE0; /* not needed */ | ||||
4737 | team->t.t_x87_fpu_control_word = 0; /* not needed */ | ||||
4738 | team->t.t_mxcsr = 0; /* not needed */ | ||||
4739 | #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ | ||||
4740 | |||||
4741 | team->t.t_construct = 0; | ||||
4742 | |||||
4743 | team->t.t_ordered.dt.t_value = 0; | ||||
4744 | team->t.t_master_active = FALSE0; | ||||
4745 | |||||
4746 | #ifdef KMP_DEBUG1 | ||||
4747 | team->t.t_copypriv_data = NULL__null; /* not necessary, but nice for debugging */ | ||||
4748 | #endif | ||||
4749 | #if KMP_OS_WINDOWS0 | ||||
4750 | team->t.t_copyin_counter = 0; /* for barrier-free copyin implementation */ | ||||
4751 | #endif | ||||
4752 | |||||
4753 | team->t.t_control_stack_top = NULL__null; | ||||
4754 | |||||
4755 | __kmp_reinitialize_team(team, new_icvs, loc); | ||||
4756 | |||||
4757 | KMP_MB(); | ||||
4758 | KF_TRACE(10, ("__kmp_initialize_team: exit: team=%p\n", team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_team: exit: team=%p\n" , team); }; | ||||
4759 | } | ||||
4760 | |||||
4761 | #if (KMP_OS_LINUX1 || KMP_OS_FREEBSD0) && KMP_AFFINITY_SUPPORTED1 | ||||
4762 | /* Sets full mask for thread and returns old mask, no changes to structures. */ | ||||
4763 | static void | ||||
4764 | __kmp_set_thread_affinity_mask_full_tmp(kmp_affin_mask_t *old_mask) { | ||||
4765 | if (KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) { | ||||
4766 | int status; | ||||
4767 | if (old_mask != NULL__null) { | ||||
4768 | status = __kmp_get_system_affinity(old_mask, TRUE)(old_mask)->get_system_affinity((!0)); | ||||
4769 | int error = errno(*__errno_location ()); | ||||
4770 | if (status != 0) { | ||||
4771 | __kmp_fatal(KMP_MSG(ChangeThreadAffMaskError)__kmp_msg_format(kmp_i18n_msg_ChangeThreadAffMaskError), KMP_ERR(error)__kmp_msg_error_code(error), | ||||
4772 | __kmp_msg_null); | ||||
4773 | } | ||||
4774 | } | ||||
4775 | __kmp_set_system_affinity(__kmp_affin_fullMask, TRUE)(__kmp_affin_fullMask)->set_system_affinity((!0)); | ||||
4776 | } | ||||
4777 | } | ||||
4778 | #endif | ||||
4779 | |||||
4780 | #if KMP_AFFINITY_SUPPORTED1 | ||||
4781 | |||||
4782 | // __kmp_partition_places() is the heart of the OpenMP 4.0 affinity mechanism. | ||||
4783 | // It calculates the worker + primary thread's partition based upon the parent | ||||
4784 | // thread's partition, and binds each worker to a thread in their partition. | ||||
4785 | // The primary thread's partition should already include its current binding. | ||||
4786 | static void __kmp_partition_places(kmp_team_t *team, int update_master_only) { | ||||
4787 | // Do not partition places for the hidden helper team | ||||
4788 | if (KMP_HIDDEN_HELPER_TEAM(team)(team->t.t_threads[0] == __kmp_hidden_helper_main_thread)) | ||||
4789 | return; | ||||
4790 | // Copy the primary thread's place partition to the team struct | ||||
4791 | kmp_info_t *master_th = team->t.t_threads[0]; | ||||
4792 | KMP_DEBUG_ASSERT(master_th != NULL)if (!(master_th != __null)) { __kmp_debug_assert("master_th != __null" , "openmp/runtime/src/kmp_runtime.cpp", 4792); }; | ||||
4793 | kmp_proc_bind_t proc_bind = team->t.t_proc_bind; | ||||
4794 | int first_place = master_th->th.th_first_place; | ||||
4795 | int last_place = master_th->th.th_last_place; | ||||
4796 | int masters_place = master_th->th.th_current_place; | ||||
4797 | int num_masks = __kmp_affinity.num_masks; | ||||
4798 | team->t.t_first_place = first_place; | ||||
4799 | team->t.t_last_place = last_place; | ||||
4800 | |||||
4801 | KA_TRACE(20, ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) " "bound to place %d partition = [%d,%d]\n", proc_bind, __kmp_gtid_from_thread (team->t.t_threads[0]), team->t.t_id, masters_place, first_place , last_place); } | ||||
4802 | "bound to place %d partition = [%d,%d]\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) " "bound to place %d partition = [%d,%d]\n", proc_bind, __kmp_gtid_from_thread (team->t.t_threads[0]), team->t.t_id, masters_place, first_place , last_place); } | ||||
4803 | proc_bind, __kmp_gtid_from_thread(team->t.t_threads[0]),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) " "bound to place %d partition = [%d,%d]\n", proc_bind, __kmp_gtid_from_thread (team->t.t_threads[0]), team->t.t_id, masters_place, first_place , last_place); } | ||||
4804 | team->t.t_id, masters_place, first_place, last_place))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) " "bound to place %d partition = [%d,%d]\n", proc_bind, __kmp_gtid_from_thread (team->t.t_threads[0]), team->t.t_id, masters_place, first_place , last_place); }; | ||||
4805 | |||||
4806 | switch (proc_bind) { | ||||
4807 | |||||
4808 | case proc_bind_default: | ||||
4809 | // Serial teams might have the proc_bind policy set to proc_bind_default. | ||||
4810 | // Not an issue -- we don't rebind primary thread for any proc_bind policy. | ||||
4811 | KMP_DEBUG_ASSERT(team->t.t_nproc == 1)if (!(team->t.t_nproc == 1)) { __kmp_debug_assert("team->t.t_nproc == 1" , "openmp/runtime/src/kmp_runtime.cpp", 4811); }; | ||||
4812 | break; | ||||
4813 | |||||
4814 | case proc_bind_primary: { | ||||
4815 | int f; | ||||
4816 | int n_th = team->t.t_nproc; | ||||
4817 | for (f = 1; f < n_th; f++) { | ||||
4818 | kmp_info_t *th = team->t.t_threads[f]; | ||||
4819 | KMP_DEBUG_ASSERT(th != NULL)if (!(th != __null)) { __kmp_debug_assert("th != __null", "openmp/runtime/src/kmp_runtime.cpp" , 4819); }; | ||||
4820 | th->th.th_first_place = first_place; | ||||
4821 | th->th.th_last_place = last_place; | ||||
4822 | th->th.th_new_place = masters_place; | ||||
4823 | if (__kmp_display_affinity && masters_place != th->th.th_current_place && | ||||
4824 | team->t.t_display_affinity != 1) { | ||||
4825 | team->t.t_display_affinity = 1; | ||||
4826 | } | ||||
4827 | |||||
4828 | KA_TRACE(100, ("__kmp_partition_places: primary: T#%d(%d:%d) place %d "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: primary: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads [f]), team->t.t_id, f, masters_place, first_place, last_place ); } | ||||
4829 | "partition = [%d,%d]\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: primary: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads [f]), team->t.t_id, f, masters_place, first_place, last_place ); } | ||||
4830 | __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: primary: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads [f]), team->t.t_id, f, masters_place, first_place, last_place ); } | ||||
4831 | f, masters_place, first_place, last_place))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: primary: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads [f]), team->t.t_id, f, masters_place, first_place, last_place ); }; | ||||
4832 | } | ||||
4833 | } break; | ||||
4834 | |||||
4835 | case proc_bind_close: { | ||||
4836 | int f; | ||||
4837 | int n_th = team->t.t_nproc; | ||||
4838 | int n_places; | ||||
4839 | if (first_place <= last_place) { | ||||
4840 | n_places = last_place - first_place + 1; | ||||
4841 | } else { | ||||
4842 | n_places = num_masks - first_place + last_place + 1; | ||||
4843 | } | ||||
4844 | if (n_th <= n_places) { | ||||
4845 | int place = masters_place; | ||||
4846 | for (f = 1; f < n_th; f++) { | ||||
4847 | kmp_info_t *th = team->t.t_threads[f]; | ||||
4848 | KMP_DEBUG_ASSERT(th != NULL)if (!(th != __null)) { __kmp_debug_assert("th != __null", "openmp/runtime/src/kmp_runtime.cpp" , 4848); }; | ||||
4849 | |||||
4850 | if (place == last_place) { | ||||
4851 | place = first_place; | ||||
4852 | } else if (place == (num_masks - 1)) { | ||||
4853 | place = 0; | ||||
4854 | } else { | ||||
4855 | place++; | ||||
4856 | } | ||||
4857 | th->th.th_first_place = first_place; | ||||
4858 | th->th.th_last_place = last_place; | ||||
4859 | th->th.th_new_place = place; | ||||
4860 | if (__kmp_display_affinity && place != th->th.th_current_place && | ||||
4861 | team->t.t_display_affinity != 1) { | ||||
4862 | team->t.t_display_affinity = 1; | ||||
4863 | } | ||||
4864 | |||||
4865 | KA_TRACE(100, ("__kmp_partition_places: close: T#%d(%d:%d) place %d "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: close: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads [f]), team->t.t_id, f, place, first_place, last_place); } | ||||
4866 | "partition = [%d,%d]\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: close: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads [f]), team->t.t_id, f, place, first_place, last_place); } | ||||
4867 | __kmp_gtid_from_thread(team->t.t_threads[f]),if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: close: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads [f]), team->t.t_id, f, place, first_place, last_place); } | ||||
4868 | team->t.t_id, f, place, first_place, last_place))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: close: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads [f]), team->t.t_id, f, place, first_place, last_place); }; | ||||
4869 | } | ||||
4870 | } else { | ||||
4871 | int S, rem, gap, s_count; | ||||
4872 | S = n_th / n_places; | ||||
4873 | s_count = 0; | ||||
4874 | rem = n_th - (S * n_places); | ||||
4875 | gap = rem > 0 ? n_places / rem : n_places; | ||||
4876 | int place = masters_place; | ||||
4877 | int gap_ct = gap; | ||||
4878 | for (f = 0; f < n_th; f++) { | ||||
4879 | kmp_info_t *th = team->t.t_threads[f]; | ||||
4880 | KMP_DEBUG_ASSERT(th != NULL)if (!(th != __null)) { __kmp_debug_assert("th != __null", "openmp/runtime/src/kmp_runtime.cpp" , 4880); }; | ||||
4881 | |||||
4882 | th->th.th_first_place = first_place; | ||||
4883 | th->th.th_last_place = last_place; | ||||
4884 | th->th.th_new_place = place; | ||||
4885 | if (__kmp_display_affinity && place != th->th.th_current_place && | ||||
4886 | team->t.t_display_affinity != 1) { | ||||
4887 | team->t.t_display_affinity = 1; | ||||
4888 | } | ||||
4889 | s_count++; | ||||
4890 | |||||
4891 | if ((s_count == S) && rem && (gap_ct == gap)) { | ||||
4892 | // do nothing, add an extra thread to place on next iteration | ||||
4893 | } else if ((s_count == S + 1) && rem && (gap_ct == gap)) { | ||||
4894 | // we added an extra thread to this place; move to next place | ||||
4895 | if (place == last_place) { | ||||
4896 | place = first_place; | ||||
4897 | } else if (place == (num_masks - 1)) { | ||||
4898 | place = 0; | ||||
4899 | } else { | ||||
4900 | place++; | ||||
4901 | } | ||||
4902 | s_count = 0; | ||||
4903 | gap_ct = 1; | ||||
4904 | rem--; | ||||
4905 | } else if (s_count == S) { // place full; don't add extra | ||||
4906 | if (place == last_place) { | ||||
4907 | place = first_place; | ||||
4908 | } else if (place == (num_masks - 1)) { | ||||
4909 | place = 0; | ||||
4910 | } else { | ||||
4911 | place++; | ||||
4912 | } | ||||
4913 | gap_ct++; | ||||
4914 | s_count = 0; | ||||
4915 | } | ||||
4916 | |||||
4917 | KA_TRACE(100,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: close: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads [f]), team->t.t_id, f, th->th.th_new_place, first_place , last_place); } | ||||
4918 | ("__kmp_partition_places: close: T#%d(%d:%d) place %d "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: close: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads [f]), team->t.t_id, f, th->th.th_new_place, first_place , last_place); } | ||||
4919 | "partition = [%d,%d]\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: close: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads [f]), team->t.t_id, f, th->th.th_new_place, first_place , last_place); } | ||||
4920 | __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id, f,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: close: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads [f]), team->t.t_id, f, th->th.th_new_place, first_place , last_place); } | ||||
4921 | th->th.th_new_place, first_place, last_place))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: close: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads [f]), team->t.t_id, f, th->th.th_new_place, first_place , last_place); }; | ||||
4922 | } | ||||
4923 | KMP_DEBUG_ASSERT(place == masters_place)if (!(place == masters_place)) { __kmp_debug_assert("place == masters_place" , "openmp/runtime/src/kmp_runtime.cpp", 4923); }; | ||||
4924 | } | ||||
4925 | } break; | ||||
4926 | |||||
4927 | case proc_bind_spread: { | ||||
4928 | int f; | ||||
4929 | int n_th = team->t.t_nproc; | ||||
4930 | int n_places; | ||||
4931 | int thidx; | ||||
4932 | if (first_place <= last_place) { | ||||
4933 | n_places = last_place - first_place + 1; | ||||
4934 | } else { | ||||
4935 | n_places = num_masks - first_place + last_place + 1; | ||||
4936 | } | ||||
4937 | if (n_th <= n_places) { | ||||
4938 | int place = -1; | ||||
4939 | |||||
4940 | if (n_places != num_masks) { | ||||
4941 | int S = n_places / n_th; | ||||
4942 | int s_count, rem, gap, gap_ct; | ||||
4943 | |||||
4944 | place = masters_place; | ||||
4945 | rem = n_places - n_th * S; | ||||
4946 | gap = rem ? n_th / rem : 1; | ||||
4947 | gap_ct = gap; | ||||
4948 | thidx = n_th; | ||||
4949 | if (update_master_only == 1) | ||||
4950 | thidx = 1; | ||||
4951 | for (f = 0; f < thidx; f++) { | ||||
4952 | kmp_info_t *th = team->t.t_threads[f]; | ||||
4953 | KMP_DEBUG_ASSERT(th != NULL)if (!(th != __null)) { __kmp_debug_assert("th != __null", "openmp/runtime/src/kmp_runtime.cpp" , 4953); }; | ||||
4954 | |||||
4955 | th->th.th_first_place = place; | ||||
4956 | th->th.th_new_place = place; | ||||
4957 | if (__kmp_display_affinity && place != th->th.th_current_place && | ||||
4958 | team->t.t_display_affinity != 1) { | ||||
4959 | team->t.t_display_affinity = 1; | ||||
4960 | } | ||||
4961 | s_count = 1; | ||||
4962 | while (s_count < S) { | ||||
4963 | if (place == last_place) { | ||||
4964 | place = first_place; | ||||
4965 | } else if (place == (num_masks - 1)) { | ||||
4966 | place = 0; | ||||
4967 | } else { | ||||
4968 | place++; | ||||
4969 | } | ||||
4970 | s_count++; | ||||
4971 | } | ||||
4972 | if (rem && (gap_ct == gap)) { | ||||
4973 | if (place == last_place) { | ||||
4974 | place = first_place; | ||||
4975 | } else if (place == (num_masks - 1)) { | ||||
4976 | place = 0; | ||||
4977 | } else { | ||||
4978 | place++; | ||||
4979 | } | ||||
4980 | rem--; | ||||
4981 | gap_ct = 0; | ||||
4982 | } | ||||
4983 | th->th.th_last_place = place; | ||||
4984 | gap_ct++; | ||||
4985 | |||||
4986 | if (place == last_place) { | ||||
4987 | place = first_place; | ||||
4988 | } else if (place == (num_masks - 1)) { | ||||
4989 | place = 0; | ||||
4990 | } else { | ||||
4991 | place++; | ||||
4992 | } | ||||
4993 | |||||
4994 | KA_TRACE(100,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], num_masks: %u\n", __kmp_gtid_from_thread (team->t.t_threads[f]), team->t.t_id, f, th->th.th_new_place , th->th.th_first_place, th->th.th_last_place, num_masks ); } | ||||
4995 | ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], num_masks: %u\n", __kmp_gtid_from_thread (team->t.t_threads[f]), team->t.t_id, f, th->th.th_new_place , th->th.th_first_place, th->th.th_last_place, num_masks ); } | ||||
4996 | "partition = [%d,%d], num_masks: %u\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], num_masks: %u\n", __kmp_gtid_from_thread (team->t.t_threads[f]), team->t.t_id, f, th->th.th_new_place , th->th.th_first_place, th->th.th_last_place, num_masks ); } | ||||
4997 | __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], num_masks: %u\n", __kmp_gtid_from_thread (team->t.t_threads[f]), team->t.t_id, f, th->th.th_new_place , th->th.th_first_place, th->th.th_last_place, num_masks ); } | ||||
4998 | f, th->th.th_new_place, th->th.th_first_place,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], num_masks: %u\n", __kmp_gtid_from_thread (team->t.t_threads[f]), team->t.t_id, f, th->th.th_new_place , th->th.th_first_place, th->th.th_last_place, num_masks ); } | ||||
4999 | th->th.th_last_place, num_masks))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], num_masks: %u\n", __kmp_gtid_from_thread (team->t.t_threads[f]), team->t.t_id, f, th->th.th_new_place , th->th.th_first_place, th->th.th_last_place, num_masks ); }; | ||||
5000 | } | ||||
5001 | } else { | ||||
5002 | /* Having uniform space of available computation places I can create | ||||
5003 | T partitions of round(P/T) size and put threads into the first | ||||
5004 | place of each partition. */ | ||||
5005 | double current = static_cast<double>(masters_place); | ||||
5006 | double spacing = | ||||
5007 | (static_cast<double>(n_places + 1) / static_cast<double>(n_th)); | ||||
5008 | int first, last; | ||||
5009 | kmp_info_t *th; | ||||
5010 | |||||
5011 | thidx = n_th + 1; | ||||
5012 | if (update_master_only == 1) | ||||
5013 | thidx = 1; | ||||
5014 | for (f = 0; f < thidx; f++) { | ||||
5015 | first = static_cast<int>(current); | ||||
5016 | last = static_cast<int>(current + spacing) - 1; | ||||
5017 | KMP_DEBUG_ASSERT(last >= first)if (!(last >= first)) { __kmp_debug_assert("last >= first" , "openmp/runtime/src/kmp_runtime.cpp", 5017); }; | ||||
5018 | if (first >= n_places) { | ||||
5019 | if (masters_place) { | ||||
5020 | first -= n_places; | ||||
5021 | last -= n_places; | ||||
5022 | if (first == (masters_place + 1)) { | ||||
5023 | KMP_DEBUG_ASSERT(f == n_th)if (!(f == n_th)) { __kmp_debug_assert("f == n_th", "openmp/runtime/src/kmp_runtime.cpp" , 5023); }; | ||||
5024 | first--; | ||||
5025 | } | ||||
5026 | if (last == masters_place) { | ||||
5027 | KMP_DEBUG_ASSERT(f == (n_th - 1))if (!(f == (n_th - 1))) { __kmp_debug_assert("f == (n_th - 1)" , "openmp/runtime/src/kmp_runtime.cpp", 5027); }; | ||||
5028 | last--; | ||||
5029 | } | ||||
5030 | } else { | ||||
5031 | KMP_DEBUG_ASSERT(f == n_th)if (!(f == n_th)) { __kmp_debug_assert("f == n_th", "openmp/runtime/src/kmp_runtime.cpp" , 5031); }; | ||||
5032 | first = 0; | ||||
5033 | last = 0; | ||||
5034 | } | ||||
5035 | } | ||||
5036 | if (last >= n_places) { | ||||
5037 | last = (n_places - 1); | ||||
5038 | } | ||||
5039 | place = first; | ||||
5040 | current += spacing; | ||||
5041 | if (f < n_th) { | ||||
5042 | KMP_DEBUG_ASSERT(0 <= first)if (!(0 <= first)) { __kmp_debug_assert("0 <= first", "openmp/runtime/src/kmp_runtime.cpp" , 5042); }; | ||||
5043 | KMP_DEBUG_ASSERT(n_places > first)if (!(n_places > first)) { __kmp_debug_assert("n_places > first" , "openmp/runtime/src/kmp_runtime.cpp", 5043); }; | ||||
5044 | KMP_DEBUG_ASSERT(0 <= last)if (!(0 <= last)) { __kmp_debug_assert("0 <= last", "openmp/runtime/src/kmp_runtime.cpp" , 5044); }; | ||||
5045 | KMP_DEBUG_ASSERT(n_places > last)if (!(n_places > last)) { __kmp_debug_assert("n_places > last" , "openmp/runtime/src/kmp_runtime.cpp", 5045); }; | ||||
5046 | KMP_DEBUG_ASSERT(last_place >= first_place)if (!(last_place >= first_place)) { __kmp_debug_assert("last_place >= first_place" , "openmp/runtime/src/kmp_runtime.cpp", 5046); }; | ||||
5047 | th = team->t.t_threads[f]; | ||||
5048 | KMP_DEBUG_ASSERT(th)if (!(th)) { __kmp_debug_assert("th", "openmp/runtime/src/kmp_runtime.cpp" , 5048); }; | ||||
5049 | th->th.th_first_place = first; | ||||
5050 | th->th.th_new_place = place; | ||||
5051 | th->th.th_last_place = last; | ||||
5052 | if (__kmp_display_affinity && place != th->th.th_current_place && | ||||
5053 | team->t.t_display_affinity != 1) { | ||||
5054 | team->t.t_display_affinity = 1; | ||||
5055 | } | ||||
5056 | KA_TRACE(100,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], spacing = %.4f\n", __kmp_gtid_from_thread (team->t.t_threads[f]), team->t.t_id, f, th->th.th_new_place , th->th.th_first_place, th->th.th_last_place, spacing) ; } | ||||
5057 | ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], spacing = %.4f\n", __kmp_gtid_from_thread (team->t.t_threads[f]), team->t.t_id, f, th->th.th_new_place , th->th.th_first_place, th->th.th_last_place, spacing) ; } | ||||
5058 | "partition = [%d,%d], spacing = %.4f\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], spacing = %.4f\n", __kmp_gtid_from_thread (team->t.t_threads[f]), team->t.t_id, f, th->th.th_new_place , th->th.th_first_place, th->th.th_last_place, spacing) ; } | ||||
5059 | __kmp_gtid_from_thread(team->t.t_threads[f]),if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], spacing = %.4f\n", __kmp_gtid_from_thread (team->t.t_threads[f]), team->t.t_id, f, th->th.th_new_place , th->th.th_first_place, th->th.th_last_place, spacing) ; } | ||||
5060 | team->t.t_id, f, th->th.th_new_place,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], spacing = %.4f\n", __kmp_gtid_from_thread (team->t.t_threads[f]), team->t.t_id, f, th->th.th_new_place , th->th.th_first_place, th->th.th_last_place, spacing) ; } | ||||
5061 | th->th.th_first_place, th->th.th_last_place, spacing))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], spacing = %.4f\n", __kmp_gtid_from_thread (team->t.t_threads[f]), team->t.t_id, f, th->th.th_new_place , th->th.th_first_place, th->th.th_last_place, spacing) ; }; | ||||
5062 | } | ||||
5063 | } | ||||
5064 | } | ||||
5065 | KMP_DEBUG_ASSERT(update_master_only || place == masters_place)if (!(update_master_only || place == masters_place)) { __kmp_debug_assert ("update_master_only || place == masters_place", "openmp/runtime/src/kmp_runtime.cpp" , 5065); }; | ||||
5066 | } else { | ||||
5067 | int S, rem, gap, s_count; | ||||
5068 | S = n_th / n_places; | ||||
5069 | s_count = 0; | ||||
5070 | rem = n_th - (S * n_places); | ||||
5071 | gap = rem > 0 ? n_places / rem : n_places; | ||||
5072 | int place = masters_place; | ||||
5073 | int gap_ct = gap; | ||||
5074 | thidx = n_th; | ||||
5075 | if (update_master_only == 1) | ||||
5076 | thidx = 1; | ||||
5077 | for (f = 0; f < thidx; f++) { | ||||
5078 | kmp_info_t *th = team->t.t_threads[f]; | ||||
5079 | KMP_DEBUG_ASSERT(th != NULL)if (!(th != __null)) { __kmp_debug_assert("th != __null", "openmp/runtime/src/kmp_runtime.cpp" , 5079); }; | ||||
5080 | |||||
5081 | th->th.th_first_place = place; | ||||
5082 | th->th.th_last_place = place; | ||||
5083 | th->th.th_new_place = place; | ||||
5084 | if (__kmp_display_affinity && place != th->th.th_current_place && | ||||
5085 | team->t.t_display_affinity != 1) { | ||||
5086 | team->t.t_display_affinity = 1; | ||||
5087 | } | ||||
5088 | s_count++; | ||||
5089 | |||||
5090 | if ((s_count == S) && rem && (gap_ct == gap)) { | ||||
5091 | // do nothing, add an extra thread to place on next iteration | ||||
5092 | } else if ((s_count == S + 1) && rem && (gap_ct == gap)) { | ||||
5093 | // we added an extra thread to this place; move on to next place | ||||
5094 | if (place == last_place) { | ||||
5095 | place = first_place; | ||||
5096 | } else if (place == (num_masks - 1)) { | ||||
5097 | place = 0; | ||||
5098 | } else { | ||||
5099 | place++; | ||||
5100 | } | ||||
5101 | s_count = 0; | ||||
5102 | gap_ct = 1; | ||||
5103 | rem--; | ||||
5104 | } else if (s_count == S) { // place is full; don't add extra thread | ||||
5105 | if (place == last_place) { | ||||
5106 | place = first_place; | ||||
5107 | } else if (place == (num_masks - 1)) { | ||||
5108 | place = 0; | ||||
5109 | } else { | ||||
5110 | place++; | ||||
5111 | } | ||||
5112 | gap_ct++; | ||||
5113 | s_count = 0; | ||||
5114 | } | ||||
5115 | |||||
5116 | KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads [f]), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place); } | ||||
5117 | "partition = [%d,%d]\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads [f]), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place); } | ||||
5118 | __kmp_gtid_from_thread(team->t.t_threads[f]),if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads [f]), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place); } | ||||
5119 | team->t.t_id, f, th->th.th_new_place,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads [f]), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place); } | ||||
5120 | th->th.th_first_place, th->th.th_last_place))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads [f]), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place); }; | ||||
5121 | } | ||||
5122 | KMP_DEBUG_ASSERT(update_master_only || place == masters_place)if (!(update_master_only || place == masters_place)) { __kmp_debug_assert ("update_master_only || place == masters_place", "openmp/runtime/src/kmp_runtime.cpp" , 5122); }; | ||||
5123 | } | ||||
5124 | } break; | ||||
5125 | |||||
5126 | default: | ||||
5127 | break; | ||||
5128 | } | ||||
5129 | |||||
5130 | KA_TRACE(20, ("__kmp_partition_places: exit T#%d\n", team->t.t_id))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_partition_places: exit T#%d\n" , team->t.t_id); }; | ||||
5131 | } | ||||
5132 | |||||
5133 | #endif // KMP_AFFINITY_SUPPORTED | ||||
5134 | |||||
5135 | /* allocate a new team data structure to use. take one off of the free pool if | ||||
5136 | available */ | ||||
5137 | kmp_team_t * | ||||
5138 | __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc, | ||||
5139 | #if OMPT_SUPPORT1 | ||||
5140 | ompt_data_t ompt_parallel_data, | ||||
5141 | #endif | ||||
5142 | kmp_proc_bind_t new_proc_bind, | ||||
5143 | kmp_internal_control_t *new_icvs, | ||||
5144 | int argc USE_NESTED_HOT_ARG(kmp_info_t *master), kmp_info_t *master) { | ||||
5145 | KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_allocate_team)((void)0); | ||||
5146 | int f; | ||||
5147 | kmp_team_t *team; | ||||
5148 | int use_hot_team = !root->r.r_active; | ||||
| |||||
5149 | int level = 0; | ||||
5150 | int do_place_partition = 1; | ||||
5151 | |||||
5152 | KA_TRACE(20, ("__kmp_allocate_team: called\n"))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: called\n" ); }; | ||||
5153 | KMP_DEBUG_ASSERT(new_nproc >= 1 && argc >= 0)if (!(new_nproc >= 1 && argc >= 0)) { __kmp_debug_assert ("new_nproc >= 1 && argc >= 0", "openmp/runtime/src/kmp_runtime.cpp" , 5153); }; | ||||
5154 | KMP_DEBUG_ASSERT(max_nproc >= new_nproc)if (!(max_nproc >= new_nproc)) { __kmp_debug_assert("max_nproc >= new_nproc" , "openmp/runtime/src/kmp_runtime.cpp", 5154); }; | ||||
5155 | KMP_MB(); | ||||
5156 | |||||
5157 | #if KMP_NESTED_HOT_TEAMS1 | ||||
5158 | kmp_hot_team_ptr_t *hot_teams; | ||||
5159 | if (master) { | ||||
5160 | team = master->th.th_team; | ||||
5161 | level = team->t.t_active_level; | ||||
5162 | if (master->th.th_teams_microtask) { // in teams construct? | ||||
5163 | if (master->th.th_teams_size.nteams > 1 && | ||||
5164 | ( // #teams > 1 | ||||
5165 | team->t.t_pkfn == | ||||
5166 | (microtask_t)__kmp_teams_master || // inner fork of the teams | ||||
5167 | master->th.th_teams_level < | ||||
5168 | team->t.t_level)) { // or nested parallel inside the teams | ||||
5169 | ++level; // not increment if #teams==1, or for outer fork of the teams; | ||||
5170 | // increment otherwise | ||||
5171 | } | ||||
5172 | // Do not perform the place partition if inner fork of the teams | ||||
5173 | // Wait until nested parallel region encountered inside teams construct | ||||
5174 | if ((master->th.th_teams_size.nteams == 1 && | ||||
5175 | master->th.th_teams_level >= team->t.t_level) || | ||||
5176 | (team->t.t_pkfn == (microtask_t)__kmp_teams_master)) | ||||
5177 | do_place_partition = 0; | ||||
5178 | } | ||||
5179 | hot_teams = master->th.th_hot_teams; | ||||
5180 | if (level < __kmp_hot_teams_max_level && hot_teams && | ||||
5181 | hot_teams[level].hot_team) { | ||||
5182 | // hot team has already been allocated for given level | ||||
5183 | use_hot_team = 1; | ||||
5184 | } else { | ||||
5185 | use_hot_team = 0; | ||||
5186 | } | ||||
5187 | } else { | ||||
5188 | // check we won't access uninitialized hot_teams, just in case | ||||
5189 | KMP_DEBUG_ASSERT(new_nproc == 1)if (!(new_nproc == 1)) { __kmp_debug_assert("new_nproc == 1", "openmp/runtime/src/kmp_runtime.cpp", 5189); }; | ||||
5190 | } | ||||
5191 | #endif | ||||
5192 | // Optimization to use a "hot" team | ||||
5193 | if (use_hot_team
| ||||
5194 | KMP_DEBUG_ASSERT(new_nproc <= max_nproc)if (!(new_nproc <= max_nproc)) { __kmp_debug_assert("new_nproc <= max_nproc" , "openmp/runtime/src/kmp_runtime.cpp", 5194); }; | ||||
5195 | #if KMP_NESTED_HOT_TEAMS1 | ||||
5196 | team = hot_teams[level].hot_team; | ||||
| |||||
5197 | #else | ||||
5198 | team = root->r.r_hot_team; | ||||
5199 | #endif | ||||
5200 | #if KMP_DEBUG1 | ||||
5201 | if (__kmp_tasking_mode != tskm_immediate_exec) { | ||||
5202 | KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: hot team task_team[0] = %p " "task_team[1] = %p before reinit\n", team->t.t_task_team[ 0], team->t.t_task_team[1]); } | ||||
5203 | "task_team[1] = %p before reinit\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: hot team task_team[0] = %p " "task_team[1] = %p before reinit\n", team->t.t_task_team[ 0], team->t.t_task_team[1]); } | ||||
5204 | team->t.t_task_team[0], team->t.t_task_team[1]))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: hot team task_team[0] = %p " "task_team[1] = %p before reinit\n", team->t.t_task_team[ 0], team->t.t_task_team[1]); }; | ||||
5205 | } | ||||
5206 | #endif | ||||
5207 | |||||
5208 | if (team->t.t_nproc != new_nproc && | ||||
5209 | __kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | ||||
5210 | // Distributed barrier may need a resize | ||||
5211 | int old_nthr = team->t.t_nproc; | ||||
5212 | __kmp_resize_dist_barrier(team, old_nthr, new_nproc); | ||||
5213 | } | ||||
5214 | |||||
5215 | // If not doing the place partition, then reset the team's proc bind | ||||
5216 | // to indicate that partitioning of all threads still needs to take place | ||||
5217 | if (do_place_partition == 0) | ||||
5218 | team->t.t_proc_bind = proc_bind_default; | ||||
5219 | // Has the number of threads changed? | ||||
5220 | /* Let's assume the most common case is that the number of threads is | ||||
5221 | unchanged, and put that case first. */ | ||||
5222 | if (team->t.t_nproc == new_nproc) { // Check changes in number of threads | ||||
5223 | KA_TRACE(20, ("__kmp_allocate_team: reusing hot team\n"))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: reusing hot team\n" ); }; | ||||
5224 | // This case can mean that omp_set_num_threads() was called and the hot | ||||
5225 | // team size was already reduced, so we check the special flag | ||||
5226 | if (team->t.t_size_changed == -1) { | ||||
5227 | team->t.t_size_changed = 1; | ||||
5228 | } else { | ||||
5229 | KMP_CHECK_UPDATE(team->t.t_size_changed, 0)if ((team->t.t_size_changed) != (0)) (team->t.t_size_changed ) = (0); | ||||
5230 | } | ||||
5231 | |||||
5232 | // TODO???: team->t.t_max_active_levels = new_max_active_levels; | ||||
5233 | kmp_r_sched_t new_sched = new_icvs->sched; | ||||
5234 | // set primary thread's schedule as new run-time schedule | ||||
5235 | KMP_CHECK_UPDATE(team->t.t_sched.sched, new_sched.sched)if ((team->t.t_sched.sched) != (new_sched.sched)) (team-> t.t_sched.sched) = (new_sched.sched); | ||||
5236 | |||||
5237 | __kmp_reinitialize_team(team, new_icvs, | ||||
5238 | root->r.r_uber_thread->th.th_ident); | ||||
5239 | |||||
5240 | KF_TRACE(10, ("__kmp_allocate_team2: T#%d, this_thread=%p team=%p\n", 0,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_team2: T#%d, this_thread=%p team=%p\n" , 0, team->t.t_threads[0], team); } | ||||
5241 | team->t.t_threads[0], team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_team2: T#%d, this_thread=%p team=%p\n" , 0, team->t.t_threads[0], team); }; | ||||
5242 | __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0); | ||||
5243 | |||||
5244 | #if KMP_AFFINITY_SUPPORTED1 | ||||
5245 | if ((team->t.t_size_changed == 0) && | ||||
5246 | (team->t.t_proc_bind == new_proc_bind)) { | ||||
5247 | if (new_proc_bind == proc_bind_spread) { | ||||
5248 | if (do_place_partition) { | ||||
5249 | // add flag to update only master for spread | ||||
5250 | __kmp_partition_places(team, 1); | ||||
5251 | } | ||||
5252 | } | ||||
5253 | KA_TRACE(200, ("__kmp_allocate_team: reusing hot team #%d bindings: "if (kmp_a_debug >= 200) { __kmp_debug_printf ("__kmp_allocate_team: reusing hot team #%d bindings: " "proc_bind = %d, partition = [%d,%d]\n", team->t.t_id, new_proc_bind , team->t.t_first_place, team->t.t_last_place); } | ||||
5254 | "proc_bind = %d, partition = [%d,%d]\n",if (kmp_a_debug >= 200) { __kmp_debug_printf ("__kmp_allocate_team: reusing hot team #%d bindings: " "proc_bind = %d, partition = [%d,%d]\n", team->t.t_id, new_proc_bind , team->t.t_first_place, team->t.t_last_place); } | ||||
5255 | team->t.t_id, new_proc_bind, team->t.t_first_place,if (kmp_a_debug >= 200) { __kmp_debug_printf ("__kmp_allocate_team: reusing hot team #%d bindings: " "proc_bind = %d, partition = [%d,%d]\n", team->t.t_id, new_proc_bind , team->t.t_first_place, team->t.t_last_place); } | ||||
5256 | team->t.t_last_place))if (kmp_a_debug >= 200) { __kmp_debug_printf ("__kmp_allocate_team: reusing hot team #%d bindings: " "proc_bind = %d, partition = [%d,%d]\n", team->t.t_id, new_proc_bind , team->t.t_first_place, team->t.t_last_place); }; | ||||
5257 | } else { | ||||
5258 | if (do_place_partition) { | ||||
5259 | KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind)if ((team->t.t_proc_bind) != (new_proc_bind)) (team->t. t_proc_bind) = (new_proc_bind); | ||||
5260 | __kmp_partition_places(team); | ||||
5261 | } | ||||
5262 | } | ||||
5263 | #else | ||||
5264 | KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind)if ((team->t.t_proc_bind) != (new_proc_bind)) (team->t. t_proc_bind) = (new_proc_bind); | ||||
5265 | #endif /* KMP_AFFINITY_SUPPORTED */ | ||||
5266 | } else if (team->t.t_nproc > new_nproc) { | ||||
5267 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: decreasing hot team thread count to %d\n" , new_nproc); } | ||||
5268 | ("__kmp_allocate_team: decreasing hot team thread count to %d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: decreasing hot team thread count to %d\n" , new_nproc); } | ||||
5269 | new_nproc))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: decreasing hot team thread count to %d\n" , new_nproc); }; | ||||
5270 | |||||
5271 | team->t.t_size_changed = 1; | ||||
5272 | if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | ||||
5273 | // Barrier size already reduced earlier in this function | ||||
5274 | // Activate team threads via th_used_in_team | ||||
5275 | __kmp_add_threads_to_team(team, new_nproc); | ||||
5276 | } | ||||
5277 | #if KMP_NESTED_HOT_TEAMS1 | ||||
5278 | if (__kmp_hot_teams_mode == 0) { | ||||
5279 | // AC: saved number of threads should correspond to team's value in this | ||||
5280 | // mode, can be bigger in mode 1, when hot team has threads in reserve | ||||
5281 | KMP_DEBUG_ASSERT(hot_teams[level].hot_team_nth == team->t.t_nproc)if (!(hot_teams[level].hot_team_nth == team->t.t_nproc)) { __kmp_debug_assert("hot_teams[level].hot_team_nth == team->t.t_nproc" , "openmp/runtime/src/kmp_runtime.cpp", 5281); }; | ||||
5282 | hot_teams[level].hot_team_nth = new_nproc; | ||||
5283 | #endif // KMP_NESTED_HOT_TEAMS | ||||
5284 | /* release the extra threads we don't need any more */ | ||||
5285 | for (f = new_nproc; f < team->t.t_nproc; f++) { | ||||
5286 | KMP_DEBUG_ASSERT(team->t.t_threads[f])if (!(team->t.t_threads[f])) { __kmp_debug_assert("team->t.t_threads[f]" , "openmp/runtime/src/kmp_runtime.cpp", 5286); }; | ||||
5287 | if (__kmp_tasking_mode != tskm_immediate_exec) { | ||||
5288 | // When decreasing team size, threads no longer in the team should | ||||
5289 | // unref task team. | ||||
5290 | team->t.t_threads[f]->th.th_task_team = NULL__null; | ||||
5291 | } | ||||
5292 | __kmp_free_thread(team->t.t_threads[f]); | ||||
5293 | team->t.t_threads[f] = NULL__null; | ||||
5294 | } | ||||
5295 | #if KMP_NESTED_HOT_TEAMS1 | ||||
5296 | } // (__kmp_hot_teams_mode == 0) | ||||
5297 | else { | ||||
5298 | // When keeping extra threads in team, switch threads to wait on own | ||||
5299 | // b_go flag | ||||
5300 | for (f = new_nproc; f < team->t.t_nproc; ++f) { | ||||
5301 | KMP_DEBUG_ASSERT(team->t.t_threads[f])if (!(team->t.t_threads[f])) { __kmp_debug_assert("team->t.t_threads[f]" , "openmp/runtime/src/kmp_runtime.cpp", 5301); }; | ||||
5302 | kmp_balign_t *balign = team->t.t_threads[f]->th.th_bar; | ||||
5303 | for (int b = 0; b < bs_last_barrier; ++b) { | ||||
5304 | if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG2) { | ||||
5305 | balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG3; | ||||
5306 | } | ||||
5307 | KMP_CHECK_UPDATE(balign[b].bb.leaf_kids, 0)if ((balign[b].bb.leaf_kids) != (0)) (balign[b].bb.leaf_kids) = (0); | ||||
5308 | } | ||||
5309 | } | ||||
5310 | } | ||||
5311 | #endif // KMP_NESTED_HOT_TEAMS | ||||
5312 | team->t.t_nproc = new_nproc; | ||||
5313 | // TODO???: team->t.t_max_active_levels = new_max_active_levels; | ||||
5314 | KMP_CHECK_UPDATE(team->t.t_sched.sched, new_icvs->sched.sched)if ((team->t.t_sched.sched) != (new_icvs->sched.sched)) (team->t.t_sched.sched) = (new_icvs->sched.sched); | ||||
5315 | __kmp_reinitialize_team(team, new_icvs, | ||||
5316 | root->r.r_uber_thread->th.th_ident); | ||||
5317 | |||||
5318 | // Update remaining threads | ||||
5319 | for (f = 0; f < new_nproc; ++f) { | ||||
5320 | team->t.t_threads[f]->th.th_team_nproc = new_nproc; | ||||
5321 | } | ||||
5322 | |||||
5323 | // restore the current task state of the primary thread: should be the | ||||
5324 | // implicit task | ||||
5325 | KF_TRACE(10, ("__kmp_allocate_team: T#%d, this_thread=%p team=%p\n", 0,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_team: T#%d, this_thread=%p team=%p\n" , 0, team->t.t_threads[0], team); } | ||||
5326 | team->t.t_threads[0], team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_team: T#%d, this_thread=%p team=%p\n" , 0, team->t.t_threads[0], team); }; | ||||
5327 | |||||
5328 | __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0); | ||||
5329 | |||||
5330 | #ifdef KMP_DEBUG1 | ||||
5331 | for (f = 0; f < team->t.t_nproc; f++) { | ||||
5332 | KMP_DEBUG_ASSERT(team->t.t_threads[f] &&if (!(team->t.t_threads[f] && team->t.t_threads [f]->th.th_team_nproc == team->t.t_nproc)) { __kmp_debug_assert ("team->t.t_threads[f] && team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc" , "openmp/runtime/src/kmp_runtime.cpp", 5334); } | ||||
5333 | team->t.t_threads[f]->th.th_team_nproc ==if (!(team->t.t_threads[f] && team->t.t_threads [f]->th.th_team_nproc == team->t.t_nproc)) { __kmp_debug_assert ("team->t.t_threads[f] && team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc" , "openmp/runtime/src/kmp_runtime.cpp", 5334); } | ||||
5334 | team->t.t_nproc)if (!(team->t.t_threads[f] && team->t.t_threads [f]->th.th_team_nproc == team->t.t_nproc)) { __kmp_debug_assert ("team->t.t_threads[f] && team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc" , "openmp/runtime/src/kmp_runtime.cpp", 5334); }; | ||||
5335 | } | ||||
5336 | #endif | ||||
5337 | |||||
5338 | if (do_place_partition) { | ||||
5339 | KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind)if ((team->t.t_proc_bind) != (new_proc_bind)) (team->t. t_proc_bind) = (new_proc_bind); | ||||
5340 | #if KMP_AFFINITY_SUPPORTED1 | ||||
5341 | __kmp_partition_places(team); | ||||
5342 | #endif | ||||
5343 | } | ||||
5344 | } else { // team->t.t_nproc < new_nproc | ||||
5345 | #if (KMP_OS_LINUX1 || KMP_OS_FREEBSD0) && KMP_AFFINITY_SUPPORTED1 | ||||
5346 | kmp_affin_mask_t *old_mask; | ||||
5347 | if (KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) { | ||||
5348 | KMP_CPU_ALLOC(old_mask)(old_mask = __kmp_affinity_dispatch->allocate_mask()); | ||||
5349 | } | ||||
5350 | #endif | ||||
5351 | |||||
5352 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: increasing hot team thread count to %d\n" , new_nproc); } | ||||
5353 | ("__kmp_allocate_team: increasing hot team thread count to %d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: increasing hot team thread count to %d\n" , new_nproc); } | ||||
5354 | new_nproc))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: increasing hot team thread count to %d\n" , new_nproc); }; | ||||
5355 | int old_nproc = team->t.t_nproc; // save old value and use to update only | ||||
5356 | team->t.t_size_changed = 1; | ||||
5357 | |||||
5358 | #if KMP_NESTED_HOT_TEAMS1 | ||||
5359 | int avail_threads = hot_teams[level].hot_team_nth; | ||||
5360 | if (new_nproc < avail_threads) | ||||
5361 | avail_threads = new_nproc; | ||||
5362 | kmp_info_t **other_threads = team->t.t_threads; | ||||
5363 | for (f = team->t.t_nproc; f < avail_threads; ++f) { | ||||
5364 | // Adjust barrier data of reserved threads (if any) of the team | ||||
5365 | // Other data will be set in __kmp_initialize_info() below. | ||||
5366 | int b; | ||||
5367 | kmp_balign_t *balign = other_threads[f]->th.th_bar; | ||||
5368 | for (b = 0; b < bs_last_barrier; ++b) { | ||||
5369 | balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; | ||||
5370 | KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG)if (!(balign[b].bb.wait_flag != 2)) { __kmp_debug_assert("balign[b].bb.wait_flag != 2" , "openmp/runtime/src/kmp_runtime.cpp", 5370); }; | ||||
5371 | #if USE_DEBUGGER0 | ||||
5372 | balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; | ||||
5373 | #endif | ||||
5374 | } | ||||
5375 | } | ||||
5376 | if (hot_teams[level].hot_team_nth >= new_nproc) { | ||||
5377 | // we have all needed threads in reserve, no need to allocate any | ||||
5378 | // this only possible in mode 1, cannot have reserved threads in mode 0 | ||||
5379 | KMP_DEBUG_ASSERT(__kmp_hot_teams_mode == 1)if (!(__kmp_hot_teams_mode == 1)) { __kmp_debug_assert("__kmp_hot_teams_mode == 1" , "openmp/runtime/src/kmp_runtime.cpp", 5379); }; | ||||
5380 | team->t.t_nproc = new_nproc; // just get reserved threads involved | ||||
5381 | } else { | ||||
5382 | // We may have some threads in reserve, but not enough; | ||||
5383 | // get reserved threads involved if any. | ||||
5384 | team->t.t_nproc = hot_teams[level].hot_team_nth; | ||||
5385 | hot_teams[level].hot_team_nth = new_nproc; // adjust hot team max size | ||||
5386 | #endif // KMP_NESTED_HOT_TEAMS | ||||
5387 | if (team->t.t_max_nproc < new_nproc) { | ||||
5388 | /* reallocate larger arrays */ | ||||
5389 | __kmp_reallocate_team_arrays(team, new_nproc); | ||||
5390 | __kmp_reinitialize_team(team, new_icvs, NULL__null); | ||||
5391 | } | ||||
5392 | |||||
5393 | #if (KMP_OS_LINUX1 || KMP_OS_FREEBSD0) && KMP_AFFINITY_SUPPORTED1 | ||||
5394 | /* Temporarily set full mask for primary thread before creation of | ||||
5395 | workers. The reason is that workers inherit the affinity from the | ||||
5396 | primary thread, so if a lot of workers are created on the single | ||||
5397 | core quickly, they don't get a chance to set their own affinity for | ||||
5398 | a long time. */ | ||||
5399 | __kmp_set_thread_affinity_mask_full_tmp(old_mask); | ||||
5400 | #endif | ||||
5401 | |||||
5402 | /* allocate new threads for the hot team */ | ||||
5403 | for (f = team->t.t_nproc; f < new_nproc; f++) { | ||||
5404 | kmp_info_t *new_worker = __kmp_allocate_thread(root, team, f); | ||||
5405 | KMP_DEBUG_ASSERT(new_worker)if (!(new_worker)) { __kmp_debug_assert("new_worker", "openmp/runtime/src/kmp_runtime.cpp" , 5405); }; | ||||
5406 | team->t.t_threads[f] = new_worker; | ||||
5407 | |||||
5408 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init T#%d arrived: " "join=%llu, plain=%llu\n", team->t.t_id, __kmp_gtid_from_tid (f, team), team->t.t_id, f, team->t.t_bar[bs_forkjoin_barrier ].b_arrived, team->t.t_bar[bs_plain_barrier].b_arrived); } | ||||
5409 | ("__kmp_allocate_team: team %d init T#%d arrived: "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init T#%d arrived: " "join=%llu, plain=%llu\n", team->t.t_id, __kmp_gtid_from_tid (f, team), team->t.t_id, f, team->t.t_bar[bs_forkjoin_barrier ].b_arrived, team->t.t_bar[bs_plain_barrier].b_arrived); } | ||||
5410 | "join=%llu, plain=%llu\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init T#%d arrived: " "join=%llu, plain=%llu\n", team->t.t_id, __kmp_gtid_from_tid (f, team), team->t.t_id, f, team->t.t_bar[bs_forkjoin_barrier ].b_arrived, team->t.t_bar[bs_plain_barrier].b_arrived); } | ||||
5411 | team->t.t_id, __kmp_gtid_from_tid(f, team), team->t.t_id, f,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init T#%d arrived: " "join=%llu, plain=%llu\n", team->t.t_id, __kmp_gtid_from_tid (f, team), team->t.t_id, f, team->t.t_bar[bs_forkjoin_barrier ].b_arrived, team->t.t_bar[bs_plain_barrier].b_arrived); } | ||||
5412 | team->t.t_bar[bs_forkjoin_barrier].b_arrived,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init T#%d arrived: " "join=%llu, plain=%llu\n", team->t.t_id, __kmp_gtid_from_tid (f, team), team->t.t_id, f, team->t.t_bar[bs_forkjoin_barrier ].b_arrived, team->t.t_bar[bs_plain_barrier].b_arrived); } | ||||
5413 | team->t.t_bar[bs_plain_barrier].b_arrived))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init T#%d arrived: " "join=%llu, plain=%llu\n", team->t.t_id, __kmp_gtid_from_tid (f, team), team->t.t_id, f, team->t.t_bar[bs_forkjoin_barrier ].b_arrived, team->t.t_bar[bs_plain_barrier].b_arrived); }; | ||||
5414 | |||||
5415 | { // Initialize barrier data for new threads. | ||||
5416 | int b; | ||||
5417 | kmp_balign_t *balign = new_worker->th.th_bar; | ||||
5418 | for (b = 0; b < bs_last_barrier; ++b) { | ||||
5419 | balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; | ||||
5420 | KMP_DEBUG_ASSERT(balign[b].bb.wait_flag !=if (!(balign[b].bb.wait_flag != 2)) { __kmp_debug_assert("balign[b].bb.wait_flag != 2" , "openmp/runtime/src/kmp_runtime.cpp", 5421); } | ||||
5421 | KMP_BARRIER_PARENT_FLAG)if (!(balign[b].bb.wait_flag != 2)) { __kmp_debug_assert("balign[b].bb.wait_flag != 2" , "openmp/runtime/src/kmp_runtime.cpp", 5421); }; | ||||
5422 | #if USE_DEBUGGER0 | ||||
5423 | balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; | ||||
5424 | #endif | ||||
5425 | } | ||||
5426 | } | ||||
5427 | } | ||||
5428 | |||||
5429 | #if (KMP_OS_LINUX1 || KMP_OS_FREEBSD0) && KMP_AFFINITY_SUPPORTED1 | ||||
5430 | if (KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) { | ||||
5431 | /* Restore initial primary thread's affinity mask */ | ||||
5432 | __kmp_set_system_affinity(old_mask, TRUE)(old_mask)->set_system_affinity((!0)); | ||||
5433 | KMP_CPU_FREE(old_mask)__kmp_affinity_dispatch->deallocate_mask(old_mask); | ||||
5434 | } | ||||
5435 | #endif | ||||
5436 | #if KMP_NESTED_HOT_TEAMS1 | ||||
5437 | } // end of check of t_nproc vs. new_nproc vs. hot_team_nth | ||||
5438 | #endif // KMP_NESTED_HOT_TEAMS | ||||
5439 | if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | ||||
5440 | // Barrier size already increased earlier in this function | ||||
5441 | // Activate team threads via th_used_in_team | ||||
5442 | __kmp_add_threads_to_team(team, new_nproc); | ||||
5443 | } | ||||
5444 | /* make sure everyone is syncronized */ | ||||
5445 | // new threads below | ||||
5446 | __kmp_initialize_team(team, new_nproc, new_icvs, | ||||
5447 | root->r.r_uber_thread->th.th_ident); | ||||
5448 | |||||
5449 | /* reinitialize the threads */ | ||||
5450 | KMP_DEBUG_ASSERT(team->t.t_nproc == new_nproc)if (!(team->t.t_nproc == new_nproc)) { __kmp_debug_assert( "team->t.t_nproc == new_nproc", "openmp/runtime/src/kmp_runtime.cpp" , 5450); }; | ||||
5451 | for (f = 0; f < team->t.t_nproc; ++f) | ||||
5452 | __kmp_initialize_info(team->t.t_threads[f], team, f, | ||||
5453 | __kmp_gtid_from_tid(f, team)); | ||||
5454 | |||||
5455 | // set th_task_state for new threads in hot team with older thread's state | ||||
5456 | kmp_uint8 old_state = team->t.t_threads[old_nproc - 1]->th.th_task_state; | ||||
5457 | for (f = old_nproc; f < team->t.t_nproc; ++f) | ||||
5458 | team->t.t_threads[f]->th.th_task_state = old_state; | ||||
5459 | |||||
5460 | #ifdef KMP_DEBUG1 | ||||
5461 | for (f = 0; f < team->t.t_nproc; ++f) { | ||||
5462 | KMP_DEBUG_ASSERT(team->t.t_threads[f] &&if (!(team->t.t_threads[f] && team->t.t_threads [f]->th.th_team_nproc == team->t.t_nproc)) { __kmp_debug_assert ("team->t.t_threads[f] && team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc" , "openmp/runtime/src/kmp_runtime.cpp", 5464); } | ||||
5463 | team->t.t_threads[f]->th.th_team_nproc ==if (!(team->t.t_threads[f] && team->t.t_threads [f]->th.th_team_nproc == team->t.t_nproc)) { __kmp_debug_assert ("team->t.t_threads[f] && team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc" , "openmp/runtime/src/kmp_runtime.cpp", 5464); } | ||||
5464 | team->t.t_nproc)if (!(team->t.t_threads[f] && team->t.t_threads [f]->th.th_team_nproc == team->t.t_nproc)) { __kmp_debug_assert ("team->t.t_threads[f] && team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc" , "openmp/runtime/src/kmp_runtime.cpp", 5464); }; | ||||
5465 | } | ||||
5466 | #endif | ||||
5467 | |||||
5468 | if (do_place_partition) { | ||||
5469 | KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind)if ((team->t.t_proc_bind) != (new_proc_bind)) (team->t. t_proc_bind) = (new_proc_bind); | ||||
5470 | #if KMP_AFFINITY_SUPPORTED1 | ||||
5471 | __kmp_partition_places(team); | ||||
5472 | #endif | ||||
5473 | } | ||||
5474 | } // Check changes in number of threads | ||||
5475 | |||||
5476 | kmp_info_t *master = team->t.t_threads[0]; | ||||
5477 | if (master->th.th_teams_microtask) { | ||||
5478 | for (f = 1; f < new_nproc; ++f) { | ||||
5479 | // propagate teams construct specific info to workers | ||||
5480 | kmp_info_t *thr = team->t.t_threads[f]; | ||||
5481 | thr->th.th_teams_microtask = master->th.th_teams_microtask; | ||||
5482 | thr->th.th_teams_level = master->th.th_teams_level; | ||||
5483 | thr->th.th_teams_size = master->th.th_teams_size; | ||||
5484 | } | ||||
5485 | } | ||||
5486 | #if KMP_NESTED_HOT_TEAMS1 | ||||
5487 | if (level) { | ||||
5488 | // Sync barrier state for nested hot teams, not needed for outermost hot | ||||
5489 | // team. | ||||
5490 | for (f = 1; f < new_nproc; ++f) { | ||||
5491 | kmp_info_t *thr = team->t.t_threads[f]; | ||||
5492 | int b; | ||||
5493 | kmp_balign_t *balign = thr->th.th_bar; | ||||
5494 | for (b = 0; b < bs_last_barrier; ++b) { | ||||
5495 | balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; | ||||
5496 | KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG)if (!(balign[b].bb.wait_flag != 2)) { __kmp_debug_assert("balign[b].bb.wait_flag != 2" , "openmp/runtime/src/kmp_runtime.cpp", 5496); }; | ||||
5497 | #if USE_DEBUGGER0 | ||||
5498 | balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; | ||||
5499 | #endif | ||||
5500 | } | ||||
5501 | } | ||||
5502 | } | ||||
5503 | #endif // KMP_NESTED_HOT_TEAMS | ||||
5504 | |||||
5505 | /* reallocate space for arguments if necessary */ | ||||
5506 | __kmp_alloc_argv_entries(argc, team, TRUE(!0)); | ||||
5507 | KMP_CHECK_UPDATE(team->t.t_argc, argc)if ((team->t.t_argc) != (argc)) (team->t.t_argc) = (argc ); | ||||
5508 | // The hot team re-uses the previous task team, | ||||
5509 | // if untouched during the previous release->gather phase. | ||||
5510 | |||||
5511 | KF_TRACE(10, (" hot_team = %p\n", team))if (kmp_f_debug >= 10) { __kmp_debug_printf (" hot_team = %p\n" , team); }; | ||||
5512 | |||||
5513 | #if KMP_DEBUG1 | ||||
5514 | if (__kmp_tasking_mode != tskm_immediate_exec) { | ||||
5515 | KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: hot team task_team[0] = %p " "task_team[1] = %p after reinit\n", team->t.t_task_team[0 ], team->t.t_task_team[1]); } | ||||
5516 | "task_team[1] = %p after reinit\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: hot team task_team[0] = %p " "task_team[1] = %p after reinit\n", team->t.t_task_team[0 ], team->t.t_task_team[1]); } | ||||
5517 | team->t.t_task_team[0], team->t.t_task_team[1]))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: hot team task_team[0] = %p " "task_team[1] = %p after reinit\n", team->t.t_task_team[0 ], team->t.t_task_team[1]); }; | ||||
5518 | } | ||||
5519 | #endif | ||||
5520 | |||||
5521 | #if OMPT_SUPPORT1 | ||||
5522 | __ompt_team_assign_id(team, ompt_parallel_data); | ||||
5523 | #endif | ||||
5524 | |||||
5525 | KMP_MB(); | ||||
5526 | |||||
5527 | return team; | ||||
5528 | } | ||||
5529 | |||||
5530 | /* next, let's try to take one from the team pool */ | ||||
5531 | KMP_MB(); | ||||
5532 | for (team = CCAST(kmp_team_t *, __kmp_team_pool)const_cast<kmp_team_t *>(__kmp_team_pool); (team);) { | ||||
5533 | /* TODO: consider resizing undersized teams instead of reaping them, now | ||||
5534 | that we have a resizing mechanism */ | ||||
5535 | if (team->t.t_max_nproc >= max_nproc) { | ||||
5536 | /* take this team from the team pool */ | ||||
5537 | __kmp_team_pool = team->t.t_next_pool; | ||||
5538 | |||||
5539 | if (max_nproc > 1 && | ||||
5540 | __kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | ||||
5541 | if (!team->t.b) { // Allocate barrier structure | ||||
5542 | team->t.b = distributedBarrier::allocate(__kmp_dflt_team_nth_ub); | ||||
5543 | } | ||||
5544 | } | ||||
5545 | |||||
5546 | /* setup the team for fresh use */ | ||||
5547 | __kmp_initialize_team(team, new_nproc, new_icvs, NULL__null); | ||||
5548 | |||||
5549 | KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: setting task_team[0] %p and " "task_team[1] %p to NULL\n", &team->t.t_task_team[0], &team->t.t_task_team[1]); } | ||||
5550 | "task_team[1] %p to NULL\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: setting task_team[0] %p and " "task_team[1] %p to NULL\n", &team->t.t_task_team[0], &team->t.t_task_team[1]); } | ||||
5551 | &team->t.t_task_team[0], &team->t.t_task_team[1]))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: setting task_team[0] %p and " "task_team[1] %p to NULL\n", &team->t.t_task_team[0], &team->t.t_task_team[1]); }; | ||||
5552 | team->t.t_task_team[0] = NULL__null; | ||||
5553 | team->t.t_task_team[1] = NULL__null; | ||||
5554 | |||||
5555 | /* reallocate space for arguments if necessary */ | ||||
5556 | __kmp_alloc_argv_entries(argc, team, TRUE(!0)); | ||||
5557 | KMP_CHECK_UPDATE(team->t.t_argc, argc)if ((team->t.t_argc) != (argc)) (team->t.t_argc) = (argc ); | ||||
5558 | |||||
5559 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n" , team->t.t_id, 0, 0); } | ||||
5560 | 20, ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n" , team->t.t_id, 0, 0); } | ||||
5561 | team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n" , team->t.t_id, 0, 0); }; | ||||
5562 | { // Initialize barrier data. | ||||
5563 | int b; | ||||
5564 | for (b = 0; b < bs_last_barrier; ++b) { | ||||
5565 | team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE0; | ||||
5566 | #if USE_DEBUGGER0 | ||||
5567 | team->t.t_bar[b].b_master_arrived = 0; | ||||
5568 | team->t.t_bar[b].b_team_arrived = 0; | ||||
5569 | #endif | ||||
5570 | } | ||||
5571 | } | ||||
5572 | |||||
5573 | team->t.t_proc_bind = new_proc_bind; | ||||
5574 | |||||
5575 | KA_TRACE(20, ("__kmp_allocate_team: using team from pool %d.\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: using team from pool %d.\n" , team->t.t_id); } | ||||
5576 | team->t.t_id))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: using team from pool %d.\n" , team->t.t_id); }; | ||||
5577 | |||||
5578 | #if OMPT_SUPPORT1 | ||||
5579 | __ompt_team_assign_id(team, ompt_parallel_data); | ||||
5580 | #endif | ||||
5581 | |||||
5582 | KMP_MB(); | ||||
5583 | |||||
5584 | return team; | ||||
5585 | } | ||||
5586 | |||||
5587 | /* reap team if it is too small, then loop back and check the next one */ | ||||
5588 | // not sure if this is wise, but, will be redone during the hot-teams | ||||
5589 | // rewrite. | ||||
5590 | /* TODO: Use technique to find the right size hot-team, don't reap them */ | ||||
5591 | team = __kmp_reap_team(team); | ||||
5592 | __kmp_team_pool = team; | ||||
5593 | } | ||||
5594 | |||||
5595 | /* nothing available in the pool, no matter, make a new team! */ | ||||
5596 | KMP_MB(); | ||||
5597 | team = (kmp_team_t *)__kmp_allocate(sizeof(kmp_team_t))___kmp_allocate((sizeof(kmp_team_t)), "openmp/runtime/src/kmp_runtime.cpp" , 5597); | ||||
5598 | |||||
5599 | /* and set it up */ | ||||
5600 | team->t.t_max_nproc = max_nproc; | ||||
5601 | if (max_nproc > 1 && | ||||
5602 | __kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | ||||
5603 | // Allocate barrier structure | ||||
5604 | team->t.b = distributedBarrier::allocate(__kmp_dflt_team_nth_ub); | ||||
5605 | } | ||||
5606 | |||||
5607 | /* NOTE well, for some reason allocating one big buffer and dividing it up | ||||
5608 | seems to really hurt performance a lot on the P4, so, let's not use this */ | ||||
5609 | __kmp_allocate_team_arrays(team, max_nproc); | ||||
5610 | |||||
5611 | KA_TRACE(20, ("__kmp_allocate_team: making a new team\n"))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: making a new team\n" ); }; | ||||
5612 | __kmp_initialize_team(team, new_nproc, new_icvs, NULL__null); | ||||
5613 | |||||
5614 | KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and task_team[1] "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: setting task_team[0] %p and task_team[1] " "%p to NULL\n", &team->t.t_task_team[0], &team-> t.t_task_team[1]); } | ||||
5615 | "%p to NULL\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: setting task_team[0] %p and task_team[1] " "%p to NULL\n", &team->t.t_task_team[0], &team-> t.t_task_team[1]); } | ||||
5616 | &team->t.t_task_team[0], &team->t.t_task_team[1]))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: setting task_team[0] %p and task_team[1] " "%p to NULL\n", &team->t.t_task_team[0], &team-> t.t_task_team[1]); }; | ||||
5617 | team->t.t_task_team[0] = NULL__null; // to be removed, as __kmp_allocate zeroes | ||||
5618 | // memory, no need to duplicate | ||||
5619 | team->t.t_task_team[1] = NULL__null; // to be removed, as __kmp_allocate zeroes | ||||
5620 | // memory, no need to duplicate | ||||
5621 | |||||
5622 | if (__kmp_storage_map) { | ||||
5623 | __kmp_print_team_storage_map("team", team, team->t.t_id, new_nproc); | ||||
5624 | } | ||||
5625 | |||||
5626 | /* allocate space for arguments */ | ||||
5627 | __kmp_alloc_argv_entries(argc, team, FALSE0); | ||||
5628 | team->t.t_argc = argc; | ||||
5629 | |||||
5630 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n" , team->t.t_id, 0, 0); } | ||||
5631 | ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n" , team->t.t_id, 0, 0); } | ||||
5632 | team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n" , team->t.t_id, 0, 0); }; | ||||
5633 | { // Initialize barrier data. | ||||
5634 | int b; | ||||
5635 | for (b = 0; b < bs_last_barrier; ++b) { | ||||
5636 | team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE0; | ||||
5637 | #if USE_DEBUGGER0 | ||||
5638 | team->t.t_bar[b].b_master_arrived = 0; | ||||
5639 | team->t.t_bar[b].b_team_arrived = 0; | ||||
5640 | #endif | ||||
5641 | } | ||||
5642 | } | ||||
5643 | |||||
5644 | team->t.t_proc_bind = new_proc_bind; | ||||
5645 | |||||
5646 | #if OMPT_SUPPORT1 | ||||
5647 | __ompt_team_assign_id(team, ompt_parallel_data); | ||||
5648 | team->t.ompt_serialized_team_info = NULL__null; | ||||
5649 | #endif | ||||
5650 | |||||
5651 | KMP_MB(); | ||||
5652 | |||||
5653 | KA_TRACE(20, ("__kmp_allocate_team: done creating a new team %d.\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: done creating a new team %d.\n" , team->t.t_id); } | ||||
5654 | team->t.t_id))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: done creating a new team %d.\n" , team->t.t_id); }; | ||||
5655 | |||||
5656 | return team; | ||||
5657 | } | ||||
5658 | |||||
5659 | /* TODO implement hot-teams at all levels */ | ||||
5660 | /* TODO implement lazy thread release on demand (disband request) */ | ||||
5661 | |||||
5662 | /* free the team. return it to the team pool. release all the threads | ||||
5663 | * associated with it */ | ||||
5664 | void __kmp_free_team(kmp_root_t *root, | ||||
5665 | kmp_team_t *team USE_NESTED_HOT_ARG(kmp_info_t *master), kmp_info_t *master) { | ||||
5666 | int f; | ||||
5667 | KA_TRACE(20, ("__kmp_free_team: T#%d freeing team %d\n", __kmp_get_gtid(),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_team: T#%d freeing team %d\n" , __kmp_get_global_thread_id(), team->t.t_id); } | ||||
5668 | team->t.t_id))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_team: T#%d freeing team %d\n" , __kmp_get_global_thread_id(), team->t.t_id); }; | ||||
5669 | |||||
5670 | /* verify state */ | ||||
5671 | KMP_DEBUG_ASSERT(root)if (!(root)) { __kmp_debug_assert("root", "openmp/runtime/src/kmp_runtime.cpp" , 5671); }; | ||||
5672 | KMP_DEBUG_ASSERT(team)if (!(team)) { __kmp_debug_assert("team", "openmp/runtime/src/kmp_runtime.cpp" , 5672); }; | ||||
5673 | KMP_DEBUG_ASSERT(team->t.t_nproc <= team->t.t_max_nproc)if (!(team->t.t_nproc <= team->t.t_max_nproc)) { __kmp_debug_assert ("team->t.t_nproc <= team->t.t_max_nproc", "openmp/runtime/src/kmp_runtime.cpp" , 5673); }; | ||||
5674 | KMP_DEBUG_ASSERT(team->t.t_threads)if (!(team->t.t_threads)) { __kmp_debug_assert("team->t.t_threads" , "openmp/runtime/src/kmp_runtime.cpp", 5674); }; | ||||
5675 | |||||
5676 | int use_hot_team = team == root->r.r_hot_team; | ||||
5677 | #if KMP_NESTED_HOT_TEAMS1 | ||||
5678 | int level; | ||||
5679 | if (master) { | ||||
5680 | level = team->t.t_active_level - 1; | ||||
5681 | if (master->th.th_teams_microtask) { // in teams construct? | ||||
5682 | if (master->th.th_teams_size.nteams > 1) { | ||||
5683 | ++level; // level was not increased in teams construct for | ||||
5684 | // team_of_masters | ||||
5685 | } | ||||
5686 | if (team->t.t_pkfn != (microtask_t)__kmp_teams_master && | ||||
5687 | master->th.th_teams_level == team->t.t_level) { | ||||
5688 | ++level; // level was not increased in teams construct for | ||||
5689 | // team_of_workers before the parallel | ||||
5690 | } // team->t.t_level will be increased inside parallel | ||||
5691 | } | ||||
5692 | #if KMP_DEBUG1 | ||||
5693 | kmp_hot_team_ptr_t *hot_teams = master->th.th_hot_teams; | ||||
5694 | #endif | ||||
5695 | if (level < __kmp_hot_teams_max_level) { | ||||
5696 | KMP_DEBUG_ASSERT(team == hot_teams[level].hot_team)if (!(team == hot_teams[level].hot_team)) { __kmp_debug_assert ("team == hot_teams[level].hot_team", "openmp/runtime/src/kmp_runtime.cpp" , 5696); }; | ||||
5697 | use_hot_team = 1; | ||||
5698 | } | ||||
5699 | } | ||||
5700 | #endif // KMP_NESTED_HOT_TEAMS | ||||
5701 | |||||
5702 | /* team is done working */ | ||||
5703 | TCW_SYNC_PTR(team->t.t_pkfn,((team->t.t_pkfn)) = ((__null)) | ||||
5704 | NULL)((team->t.t_pkfn)) = ((__null)); // Important for Debugging Support Library. | ||||
5705 | #if KMP_OS_WINDOWS0 | ||||
5706 | team->t.t_copyin_counter = 0; // init counter for possible reuse | ||||
5707 | #endif | ||||
5708 | // Do not reset pointer to parent team to NULL for hot teams. | ||||
5709 | |||||
5710 | /* if we are non-hot team, release our threads */ | ||||
5711 | if (!use_hot_team) { | ||||
5712 | if (__kmp_tasking_mode != tskm_immediate_exec) { | ||||
5713 | // Wait for threads to reach reapable state | ||||
5714 | for (f = 1; f < team->t.t_nproc; ++f) { | ||||
5715 | KMP_DEBUG_ASSERT(team->t.t_threads[f])if (!(team->t.t_threads[f])) { __kmp_debug_assert("team->t.t_threads[f]" , "openmp/runtime/src/kmp_runtime.cpp", 5715); }; | ||||
5716 | kmp_info_t *th = team->t.t_threads[f]; | ||||
5717 | volatile kmp_uint32 *state = &th->th.th_reap_state; | ||||
5718 | while (*state != KMP_SAFE_TO_REAP1) { | ||||
5719 | #if KMP_OS_WINDOWS0 | ||||
5720 | // On Windows a thread can be killed at any time, check this | ||||
5721 | DWORD ecode; | ||||
5722 | if (!__kmp_is_thread_alive(th, &ecode)) { | ||||
5723 | *state = KMP_SAFE_TO_REAP1; // reset the flag for dead thread | ||||
5724 | break; | ||||
5725 | } | ||||
5726 | #endif | ||||
5727 | // first check if thread is sleeping | ||||
5728 | kmp_flag_64<> fl(&th->th.th_bar[bs_forkjoin_barrier].bb.b_go, th); | ||||
5729 | if (fl.is_sleeping()) | ||||
5730 | fl.resume(__kmp_gtid_from_thread(th)); | ||||
5731 | KMP_CPU_PAUSE()__kmp_x86_pause(); | ||||
5732 | } | ||||
5733 | } | ||||
5734 | |||||
5735 | // Delete task teams | ||||
5736 | int tt_idx; | ||||
5737 | for (tt_idx = 0; tt_idx < 2; ++tt_idx) { | ||||
5738 | kmp_task_team_t *task_team = team->t.t_task_team[tt_idx]; | ||||
5739 | if (task_team != NULL__null) { | ||||
5740 | for (f = 0; f < team->t.t_nproc; ++f) { // threads unref task teams | ||||
5741 | KMP_DEBUG_ASSERT(team->t.t_threads[f])if (!(team->t.t_threads[f])) { __kmp_debug_assert("team->t.t_threads[f]" , "openmp/runtime/src/kmp_runtime.cpp", 5741); }; | ||||
5742 | team->t.t_threads[f]->th.th_task_team = NULL__null; | ||||
5743 | } | ||||
5744 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n" , __kmp_get_global_thread_id(), task_team, team->t.t_id); } | ||||
5745 | 20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n" , __kmp_get_global_thread_id(), task_team, team->t.t_id); } | ||||
5746 | ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n" , __kmp_get_global_thread_id(), task_team, team->t.t_id); } | ||||
5747 | __kmp_get_gtid(), task_team, team->t.t_id))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n" , __kmp_get_global_thread_id(), task_team, team->t.t_id); }; | ||||
5748 | #if KMP_NESTED_HOT_TEAMS1 | ||||
5749 | __kmp_free_task_team(master, task_team); | ||||
5750 | #endif | ||||
5751 | team->t.t_task_team[tt_idx] = NULL__null; | ||||
5752 | } | ||||
5753 | } | ||||
5754 | } | ||||
5755 | |||||
5756 | // Reset pointer to parent team only for non-hot teams. | ||||
5757 | team->t.t_parent = NULL__null; | ||||
5758 | team->t.t_level = 0; | ||||
5759 | team->t.t_active_level = 0; | ||||
5760 | |||||
5761 | /* free the worker threads */ | ||||
5762 | for (f = 1; f < team->t.t_nproc; ++f) { | ||||
5763 | KMP_DEBUG_ASSERT(team->t.t_threads[f])if (!(team->t.t_threads[f])) { __kmp_debug_assert("team->t.t_threads[f]" , "openmp/runtime/src/kmp_runtime.cpp", 5763); }; | ||||
5764 | if (__kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | ||||
5765 | KMP_COMPARE_AND_STORE_ACQ32(&(team->t.t_threads[f]->th.th_used_in_team),__sync_bool_compare_and_swap((volatile kmp_uint32 *)(&(team ->t.t_threads[f]->th.th_used_in_team)), (kmp_uint32)(1) , (kmp_uint32)(2)) | ||||
5766 | 1, 2)__sync_bool_compare_and_swap((volatile kmp_uint32 *)(&(team ->t.t_threads[f]->th.th_used_in_team)), (kmp_uint32)(1) , (kmp_uint32)(2)); | ||||
5767 | } | ||||
5768 | __kmp_free_thread(team->t.t_threads[f]); | ||||
5769 | } | ||||
5770 | |||||
5771 | if (__kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | ||||
5772 | if (team->t.b) { | ||||
5773 | // wake up thread at old location | ||||
5774 | team->t.b->go_release(); | ||||
5775 | if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME(2147483647)) { | ||||
5776 | for (f = 1; f < team->t.t_nproc; ++f) { | ||||
5777 | if (team->t.b->sleep[f].sleep) { | ||||
5778 | __kmp_atomic_resume_64( | ||||
5779 | team->t.t_threads[f]->th.th_info.ds.ds_gtid, | ||||
5780 | (kmp_atomic_flag_64<> *)NULL__null); | ||||
5781 | } | ||||
5782 | } | ||||
5783 | } | ||||
5784 | // Wait for threads to be removed from team | ||||
5785 | for (int f = 1; f < team->t.t_nproc; ++f) { | ||||
5786 | while (team->t.t_threads[f]->th.th_used_in_team.load() != 0) | ||||
5787 | KMP_CPU_PAUSE()__kmp_x86_pause(); | ||||
5788 | } | ||||
5789 | } | ||||
5790 | } | ||||
5791 | |||||
5792 | for (f = 1; f < team->t.t_nproc; ++f) { | ||||
5793 | team->t.t_threads[f] = NULL__null; | ||||
5794 | } | ||||
5795 | |||||
5796 | if (team->t.t_max_nproc > 1 && | ||||
5797 | __kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | ||||
5798 | distributedBarrier::deallocate(team->t.b); | ||||
5799 | team->t.b = NULL__null; | ||||
5800 | } | ||||
5801 | /* put the team back in the team pool */ | ||||
5802 | /* TODO limit size of team pool, call reap_team if pool too large */ | ||||
5803 | team->t.t_next_pool = CCAST(kmp_team_t *, __kmp_team_pool)const_cast<kmp_team_t *>(__kmp_team_pool); | ||||
5804 | __kmp_team_pool = (volatile kmp_team_t *)team; | ||||
5805 | } else { // Check if team was created for primary threads in teams construct | ||||
5806 | // See if first worker is a CG root | ||||
5807 | KMP_DEBUG_ASSERT(team->t.t_threads[1] &&if (!(team->t.t_threads[1] && team->t.t_threads [1]->th.th_cg_roots)) { __kmp_debug_assert("team->t.t_threads[1] && team->t.t_threads[1]->th.th_cg_roots" , "openmp/runtime/src/kmp_runtime.cpp", 5808); } | ||||
5808 | team->t.t_threads[1]->th.th_cg_roots)if (!(team->t.t_threads[1] && team->t.t_threads [1]->th.th_cg_roots)) { __kmp_debug_assert("team->t.t_threads[1] && team->t.t_threads[1]->th.th_cg_roots" , "openmp/runtime/src/kmp_runtime.cpp", 5808); }; | ||||
5809 | if (team->t.t_threads[1]->th.th_cg_roots->cg_root == team->t.t_threads[1]) { | ||||
5810 | // Clean up the CG root nodes on workers so that this team can be re-used | ||||
5811 | for (f = 1; f < team->t.t_nproc; ++f) { | ||||
5812 | kmp_info_t *thr = team->t.t_threads[f]; | ||||
5813 | KMP_DEBUG_ASSERT(thr && thr->th.th_cg_roots &&if (!(thr && thr->th.th_cg_roots && thr-> th.th_cg_roots->cg_root == thr)) { __kmp_debug_assert("thr && thr->th.th_cg_roots && thr->th.th_cg_roots->cg_root == thr" , "openmp/runtime/src/kmp_runtime.cpp", 5814); } | ||||
5814 | thr->th.th_cg_roots->cg_root == thr)if (!(thr && thr->th.th_cg_roots && thr-> th.th_cg_roots->cg_root == thr)) { __kmp_debug_assert("thr && thr->th.th_cg_roots && thr->th.th_cg_roots->cg_root == thr" , "openmp/runtime/src/kmp_runtime.cpp", 5814); }; | ||||
5815 | // Pop current CG root off list | ||||
5816 | kmp_cg_root_t *tmp = thr->th.th_cg_roots; | ||||
5817 | thr->th.th_cg_roots = tmp->up; | ||||
5818 | KA_TRACE(100, ("__kmp_free_team: Thread %p popping node %p and moving"if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_free_team: Thread %p popping node %p and moving" " up to node %p. cg_nthreads was %d\n", thr, tmp, thr->th .th_cg_roots, tmp->cg_nthreads); } | ||||
5819 | " up to node %p. cg_nthreads was %d\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_free_team: Thread %p popping node %p and moving" " up to node %p. cg_nthreads was %d\n", thr, tmp, thr->th .th_cg_roots, tmp->cg_nthreads); } | ||||
5820 | thr, tmp, thr->th.th_cg_roots, tmp->cg_nthreads))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_free_team: Thread %p popping node %p and moving" " up to node %p. cg_nthreads was %d\n", thr, tmp, thr->th .th_cg_roots, tmp->cg_nthreads); }; | ||||
5821 | int i = tmp->cg_nthreads--; | ||||
5822 | if (i == 1) { | ||||
5823 | __kmp_free(tmp)___kmp_free((tmp), "openmp/runtime/src/kmp_runtime.cpp", 5823 ); // free CG if we are the last thread in it | ||||
5824 | } | ||||
5825 | // Restore current task's thread_limit from CG root | ||||
5826 | if (thr->th.th_cg_roots) | ||||
5827 | thr->th.th_current_task->td_icvs.thread_limit = | ||||
5828 | thr->th.th_cg_roots->cg_thread_limit; | ||||
5829 | } | ||||
5830 | } | ||||
5831 | } | ||||
5832 | |||||
5833 | KMP_MB(); | ||||
5834 | } | ||||
5835 | |||||
5836 | /* reap the team. destroy it, reclaim all its resources and free its memory */ | ||||
5837 | kmp_team_t *__kmp_reap_team(kmp_team_t *team) { | ||||
5838 | kmp_team_t *next_pool = team->t.t_next_pool; | ||||
5839 | |||||
5840 | KMP_DEBUG_ASSERT(team)if (!(team)) { __kmp_debug_assert("team", "openmp/runtime/src/kmp_runtime.cpp" , 5840); }; | ||||
5841 | KMP_DEBUG_ASSERT(team->t.t_dispatch)if (!(team->t.t_dispatch)) { __kmp_debug_assert("team->t.t_dispatch" , "openmp/runtime/src/kmp_runtime.cpp", 5841); }; | ||||
5842 | KMP_DEBUG_ASSERT(team->t.t_disp_buffer)if (!(team->t.t_disp_buffer)) { __kmp_debug_assert("team->t.t_disp_buffer" , "openmp/runtime/src/kmp_runtime.cpp", 5842); }; | ||||
5843 | KMP_DEBUG_ASSERT(team->t.t_threads)if (!(team->t.t_threads)) { __kmp_debug_assert("team->t.t_threads" , "openmp/runtime/src/kmp_runtime.cpp", 5843); }; | ||||
5844 | KMP_DEBUG_ASSERT(team->t.t_argv)if (!(team->t.t_argv)) { __kmp_debug_assert("team->t.t_argv" , "openmp/runtime/src/kmp_runtime.cpp", 5844); }; | ||||
5845 | |||||
5846 | /* TODO clean the threads that are a part of this? */ | ||||
5847 | |||||
5848 | /* free stuff */ | ||||
5849 | __kmp_free_team_arrays(team); | ||||
5850 | if (team->t.t_argv != &team->t.t_inline_argv[0]) | ||||
5851 | __kmp_free((void *)team->t.t_argv)___kmp_free(((void *)team->t.t_argv), "openmp/runtime/src/kmp_runtime.cpp" , 5851); | ||||
5852 | __kmp_free(team)___kmp_free((team), "openmp/runtime/src/kmp_runtime.cpp", 5852 ); | ||||
5853 | |||||
5854 | KMP_MB(); | ||||
5855 | return next_pool; | ||||
5856 | } | ||||
5857 | |||||
5858 | // Free the thread. Don't reap it, just place it on the pool of available | ||||
5859 | // threads. | ||||
5860 | // | ||||
5861 | // Changes for Quad issue 527845: We need a predictable OMP tid <-> gtid | ||||
5862 | // binding for the affinity mechanism to be useful. | ||||
5863 | // | ||||
5864 | // Now, we always keep the free list (__kmp_thread_pool) sorted by gtid. | ||||
5865 | // However, we want to avoid a potential performance problem by always | ||||
5866 | // scanning through the list to find the correct point at which to insert | ||||
5867 | // the thread (potential N**2 behavior). To do this we keep track of the | ||||
5868 | // last place a thread struct was inserted (__kmp_thread_pool_insert_pt). | ||||
5869 | // With single-level parallelism, threads will always be added to the tail | ||||
5870 | // of the list, kept track of by __kmp_thread_pool_insert_pt. With nested | ||||
5871 | // parallelism, all bets are off and we may need to scan through the entire | ||||
5872 | // free list. | ||||
5873 | // | ||||
5874 | // This change also has a potentially large performance benefit, for some | ||||
5875 | // applications. Previously, as threads were freed from the hot team, they | ||||
5876 | // would be placed back on the free list in inverse order. If the hot team | ||||
5877 | // grew back to it's original size, then the freed thread would be placed | ||||
5878 | // back on the hot team in reverse order. This could cause bad cache | ||||
5879 | // locality problems on programs where the size of the hot team regularly | ||||
5880 | // grew and shrunk. | ||||
5881 | // | ||||
5882 | // Now, for single-level parallelism, the OMP tid is always == gtid. | ||||
5883 | void __kmp_free_thread(kmp_info_t *this_th) { | ||||
5884 | int gtid; | ||||
5885 | kmp_info_t **scan; | ||||
5886 | |||||
5887 | KA_TRACE(20, ("__kmp_free_thread: T#%d putting T#%d back on free pool.\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_thread: T#%d putting T#%d back on free pool.\n" , __kmp_get_global_thread_id(), this_th->th.th_info.ds.ds_gtid ); } | ||||
5888 | __kmp_get_gtid(), this_th->th.th_info.ds.ds_gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_thread: T#%d putting T#%d back on free pool.\n" , __kmp_get_global_thread_id(), this_th->th.th_info.ds.ds_gtid ); }; | ||||
5889 | |||||
5890 | KMP_DEBUG_ASSERT(this_th)if (!(this_th)) { __kmp_debug_assert("this_th", "openmp/runtime/src/kmp_runtime.cpp" , 5890); }; | ||||
5891 | |||||
5892 | // When moving thread to pool, switch thread to wait on own b_go flag, and | ||||
5893 | // uninitialized (NULL team). | ||||
5894 | int b; | ||||
5895 | kmp_balign_t *balign = this_th->th.th_bar; | ||||
5896 | for (b = 0; b < bs_last_barrier; ++b) { | ||||
5897 | if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG2) | ||||
5898 | balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG3; | ||||
5899 | balign[b].bb.team = NULL__null; | ||||
5900 | balign[b].bb.leaf_kids = 0; | ||||
5901 | } | ||||
5902 | this_th->th.th_task_state = 0; | ||||
5903 | this_th->th.th_reap_state = KMP_SAFE_TO_REAP1; | ||||
5904 | |||||
5905 | /* put thread back on the free pool */ | ||||
5906 | TCW_PTR(this_th->th.th_team, NULL)((this_th->th.th_team)) = ((__null)); | ||||
5907 | TCW_PTR(this_th->th.th_root, NULL)((this_th->th.th_root)) = ((__null)); | ||||
5908 | TCW_PTR(this_th->th.th_dispatch, NULL)((this_th->th.th_dispatch)) = ((__null)); /* NOT NEEDED */ | ||||
5909 | |||||
5910 | while (this_th->th.th_cg_roots) { | ||||
5911 | this_th->th.th_cg_roots->cg_nthreads--; | ||||
5912 | KA_TRACE(100, ("__kmp_free_thread: Thread %p decrement cg_nthreads on node"if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_free_thread: Thread %p decrement cg_nthreads on node" " %p of thread %p to %d\n", this_th, this_th->th.th_cg_roots , this_th->th.th_cg_roots->cg_root, this_th->th.th_cg_roots ->cg_nthreads); } | ||||
5913 | " %p of thread %p to %d\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_free_thread: Thread %p decrement cg_nthreads on node" " %p of thread %p to %d\n", this_th, this_th->th.th_cg_roots , this_th->th.th_cg_roots->cg_root, this_th->th.th_cg_roots ->cg_nthreads); } | ||||
5914 | this_th, this_th->th.th_cg_roots,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_free_thread: Thread %p decrement cg_nthreads on node" " %p of thread %p to %d\n", this_th, this_th->th.th_cg_roots , this_th->th.th_cg_roots->cg_root, this_th->th.th_cg_roots ->cg_nthreads); } | ||||
5915 | this_th->th.th_cg_roots->cg_root,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_free_thread: Thread %p decrement cg_nthreads on node" " %p of thread %p to %d\n", this_th, this_th->th.th_cg_roots , this_th->th.th_cg_roots->cg_root, this_th->th.th_cg_roots ->cg_nthreads); } | ||||
5916 | this_th->th.th_cg_roots->cg_nthreads))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_free_thread: Thread %p decrement cg_nthreads on node" " %p of thread %p to %d\n", this_th, this_th->th.th_cg_roots , this_th->th.th_cg_roots->cg_root, this_th->th.th_cg_roots ->cg_nthreads); }; | ||||
5917 | kmp_cg_root_t *tmp = this_th->th.th_cg_roots; | ||||
5918 | if (tmp->cg_root == this_th) { // Thread is a cg_root | ||||
5919 | KMP_DEBUG_ASSERT(tmp->cg_nthreads == 0)if (!(tmp->cg_nthreads == 0)) { __kmp_debug_assert("tmp->cg_nthreads == 0" , "openmp/runtime/src/kmp_runtime.cpp", 5919); }; | ||||
5920 | KA_TRACE(if (kmp_a_debug >= 5) { __kmp_debug_printf ("__kmp_free_thread: Thread %p freeing node %p\n" , this_th, tmp); } | ||||
5921 | 5, ("__kmp_free_thread: Thread %p freeing node %p\n", this_th, tmp))if (kmp_a_debug >= 5) { __kmp_debug_printf ("__kmp_free_thread: Thread %p freeing node %p\n" , this_th, tmp); }; | ||||
5922 | this_th->th.th_cg_roots = tmp->up; | ||||
5923 | __kmp_free(tmp)___kmp_free((tmp), "openmp/runtime/src/kmp_runtime.cpp", 5923 ); | ||||
5924 | } else { // Worker thread | ||||
5925 | if (tmp->cg_nthreads == 0) { // last thread leaves contention group | ||||
5926 | __kmp_free(tmp)___kmp_free((tmp), "openmp/runtime/src/kmp_runtime.cpp", 5926 ); | ||||
5927 | } | ||||
5928 | this_th->th.th_cg_roots = NULL__null; | ||||
5929 | break; | ||||
5930 | } | ||||
5931 | } | ||||
5932 | |||||
5933 | /* If the implicit task assigned to this thread can be used by other threads | ||||
5934 | * -> multiple threads can share the data and try to free the task at | ||||
5935 | * __kmp_reap_thread at exit. This duplicate use of the task data can happen | ||||
5936 | * with higher probability when hot team is disabled but can occurs even when | ||||
5937 | * the hot team is enabled */ | ||||
5938 | __kmp_free_implicit_task(this_th); | ||||
5939 | this_th->th.th_current_task = NULL__null; | ||||
5940 | |||||
5941 | // If the __kmp_thread_pool_insert_pt is already past the new insert | ||||
5942 | // point, then we need to re-scan the entire list. | ||||
5943 | gtid = this_th->th.th_info.ds.ds_gtid; | ||||
5944 | if (__kmp_thread_pool_insert_pt != NULL__null) { | ||||
5945 | KMP_DEBUG_ASSERT(__kmp_thread_pool != NULL)if (!(__kmp_thread_pool != __null)) { __kmp_debug_assert("__kmp_thread_pool != __null" , "openmp/runtime/src/kmp_runtime.cpp", 5945); }; | ||||
5946 | if (__kmp_thread_pool_insert_pt->th.th_info.ds.ds_gtid > gtid) { | ||||
5947 | __kmp_thread_pool_insert_pt = NULL__null; | ||||
5948 | } | ||||
5949 | } | ||||
5950 | |||||
5951 | // Scan down the list to find the place to insert the thread. | ||||
5952 | // scan is the address of a link in the list, possibly the address of | ||||
5953 | // __kmp_thread_pool itself. | ||||
5954 | // | ||||
5955 | // In the absence of nested parallelism, the for loop will have 0 iterations. | ||||
5956 | if (__kmp_thread_pool_insert_pt != NULL__null) { | ||||
5957 | scan = &(__kmp_thread_pool_insert_pt->th.th_next_pool); | ||||
5958 | } else { | ||||
5959 | scan = CCAST(kmp_info_t **, &__kmp_thread_pool)const_cast<kmp_info_t **>(&__kmp_thread_pool); | ||||
5960 | } | ||||
5961 | for (; (*scan != NULL__null) && ((*scan)->th.th_info.ds.ds_gtid < gtid); | ||||
5962 | scan = &((*scan)->th.th_next_pool)) | ||||
5963 | ; | ||||
5964 | |||||
5965 | // Insert the new element on the list, and set __kmp_thread_pool_insert_pt | ||||
5966 | // to its address. | ||||
5967 | TCW_PTR(this_th->th.th_next_pool, *scan)((this_th->th.th_next_pool)) = ((*scan)); | ||||
5968 | __kmp_thread_pool_insert_pt = *scan = this_th; | ||||
5969 | KMP_DEBUG_ASSERT((this_th->th.th_next_pool == NULL) ||if (!((this_th->th.th_next_pool == __null) || (this_th-> th.th_info.ds.ds_gtid < this_th->th.th_next_pool->th .th_info.ds.ds_gtid))) { __kmp_debug_assert("(this_th->th.th_next_pool == __null) || (this_th->th.th_info.ds.ds_gtid < this_th->th.th_next_pool->th.th_info.ds.ds_gtid)" , "openmp/runtime/src/kmp_runtime.cpp", 5971); } | ||||
5970 | (this_th->th.th_info.ds.ds_gtid <if (!((this_th->th.th_next_pool == __null) || (this_th-> th.th_info.ds.ds_gtid < this_th->th.th_next_pool->th .th_info.ds.ds_gtid))) { __kmp_debug_assert("(this_th->th.th_next_pool == __null) || (this_th->th.th_info.ds.ds_gtid < this_th->th.th_next_pool->th.th_info.ds.ds_gtid)" , "openmp/runtime/src/kmp_runtime.cpp", 5971); } | ||||
5971 | this_th->th.th_next_pool->th.th_info.ds.ds_gtid))if (!((this_th->th.th_next_pool == __null) || (this_th-> th.th_info.ds.ds_gtid < this_th->th.th_next_pool->th .th_info.ds.ds_gtid))) { __kmp_debug_assert("(this_th->th.th_next_pool == __null) || (this_th->th.th_info.ds.ds_gtid < this_th->th.th_next_pool->th.th_info.ds.ds_gtid)" , "openmp/runtime/src/kmp_runtime.cpp", 5971); }; | ||||
5972 | TCW_4(this_th->th.th_in_pool, TRUE)(this_th->th.th_in_pool) = ((!0)); | ||||
5973 | __kmp_suspend_initialize_thread(this_th); | ||||
5974 | __kmp_lock_suspend_mx(this_th); | ||||
5975 | if (this_th->th.th_active == TRUE(!0)) { | ||||
5976 | KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth)(&__kmp_thread_pool_active_nth)->fetch_add(1, std::memory_order_acq_rel ); | ||||
5977 | this_th->th.th_active_in_pool = TRUE(!0); | ||||
5978 | } | ||||
5979 | #if KMP_DEBUG1 | ||||
5980 | else { | ||||
5981 | KMP_DEBUG_ASSERT(this_th->th.th_active_in_pool == FALSE)if (!(this_th->th.th_active_in_pool == 0)) { __kmp_debug_assert ("this_th->th.th_active_in_pool == 0", "openmp/runtime/src/kmp_runtime.cpp" , 5981); }; | ||||
5982 | } | ||||
5983 | #endif | ||||
5984 | __kmp_unlock_suspend_mx(this_th); | ||||
5985 | |||||
5986 | TCW_4(__kmp_nth, __kmp_nth - 1)(__kmp_nth) = (__kmp_nth - 1); | ||||
5987 | |||||
5988 | #ifdef KMP_ADJUST_BLOCKTIME1 | ||||
5989 | /* Adjust blocktime back to user setting or default if necessary */ | ||||
5990 | /* Middle initialization might never have occurred */ | ||||
5991 | if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { | ||||
5992 | KMP_DEBUG_ASSERT(__kmp_avail_proc > 0)if (!(__kmp_avail_proc > 0)) { __kmp_debug_assert("__kmp_avail_proc > 0" , "openmp/runtime/src/kmp_runtime.cpp", 5992); }; | ||||
5993 | if (__kmp_nth <= __kmp_avail_proc) { | ||||
5994 | __kmp_zero_bt = FALSE0; | ||||
5995 | } | ||||
5996 | } | ||||
5997 | #endif /* KMP_ADJUST_BLOCKTIME */ | ||||
5998 | |||||
5999 | KMP_MB(); | ||||
6000 | } | ||||
6001 | |||||
6002 | /* ------------------------------------------------------------------------ */ | ||||
6003 | |||||
6004 | void *__kmp_launch_thread(kmp_info_t *this_thr) { | ||||
6005 | #if OMP_PROFILING_SUPPORT0 | ||||
6006 | ProfileTraceFile = getenv("LIBOMPTARGET_PROFILE"); | ||||
6007 | // TODO: add a configuration option for time granularity | ||||
6008 | if (ProfileTraceFile) | ||||
6009 | llvm::timeTraceProfilerInitialize(500 /* us */, "libomptarget"); | ||||
6010 | #endif | ||||
6011 | |||||
6012 | int gtid = this_thr->th.th_info.ds.ds_gtid; | ||||
6013 | /* void *stack_data;*/ | ||||
6014 | kmp_team_t **volatile pteam; | ||||
6015 | |||||
6016 | KMP_MB(); | ||||
6017 | KA_TRACE(10, ("__kmp_launch_thread: T#%d start\n", gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_launch_thread: T#%d start\n" , gtid); }; | ||||
6018 | |||||
6019 | if (__kmp_env_consistency_check) { | ||||
6020 | this_thr->th.th_cons = __kmp_allocate_cons_stack(gtid); // ATT: Memory leak? | ||||
6021 | } | ||||
6022 | |||||
6023 | #if OMPD_SUPPORT1 | ||||
6024 | if (ompd_state & OMPD_ENABLE_BP0x1) | ||||
6025 | ompd_bp_thread_begin(); | ||||
6026 | #endif | ||||
6027 | |||||
6028 | #if OMPT_SUPPORT1 | ||||
6029 | ompt_data_t *thread_data = nullptr; | ||||
6030 | if (ompt_enabled.enabled) { | ||||
6031 | thread_data = &(this_thr->th.ompt_thread_info.thread_data); | ||||
6032 | *thread_data = ompt_data_none{0}; | ||||
6033 | |||||
6034 | this_thr->th.ompt_thread_info.state = ompt_state_overhead; | ||||
6035 | this_thr->th.ompt_thread_info.wait_id = 0; | ||||
6036 | this_thr->th.ompt_thread_info.idle_frame = OMPT_GET_FRAME_ADDRESS(0)__builtin_frame_address(0); | ||||
6037 | this_thr->th.ompt_thread_info.parallel_flags = 0; | ||||
6038 | if (ompt_enabled.ompt_callback_thread_begin) { | ||||
6039 | ompt_callbacks.ompt_callback(ompt_callback_thread_begin)ompt_callback_thread_begin_callback( | ||||
6040 | ompt_thread_worker, thread_data); | ||||
6041 | } | ||||
6042 | this_thr->th.ompt_thread_info.state = ompt_state_idle; | ||||
6043 | } | ||||
6044 | #endif | ||||
6045 | |||||
6046 | /* This is the place where threads wait for work */ | ||||
6047 | while (!TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done)) { | ||||
6048 | KMP_DEBUG_ASSERT(this_thr == __kmp_threads[gtid])if (!(this_thr == __kmp_threads[gtid])) { __kmp_debug_assert( "this_thr == __kmp_threads[gtid]", "openmp/runtime/src/kmp_runtime.cpp" , 6048); }; | ||||
6049 | KMP_MB(); | ||||
6050 | |||||
6051 | /* wait for work to do */ | ||||
6052 | KA_TRACE(20, ("__kmp_launch_thread: T#%d waiting for work\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_launch_thread: T#%d waiting for work\n" , gtid); }; | ||||
6053 | |||||
6054 | /* No tid yet since not part of a team */ | ||||
6055 | __kmp_fork_barrier(gtid, KMP_GTID_DNE(-2)); | ||||
6056 | |||||
6057 | #if OMPT_SUPPORT1 | ||||
6058 | if (ompt_enabled.enabled) { | ||||
6059 | this_thr->th.ompt_thread_info.state = ompt_state_overhead; | ||||
6060 | } | ||||
6061 | #endif | ||||
6062 | |||||
6063 | pteam = &this_thr->th.th_team; | ||||
6064 | |||||
6065 | /* have we been allocated? */ | ||||
6066 | if (TCR_SYNC_PTR(*pteam)((void *)(*pteam)) && !TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done)) { | ||||
6067 | /* we were just woken up, so run our new task */ | ||||
6068 | if (TCR_SYNC_PTR((*pteam)->t.t_pkfn)((void *)((*pteam)->t.t_pkfn)) != NULL__null) { | ||||
6069 | int rc; | ||||
6070 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n" , gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid), (*pteam )->t.t_pkfn); } | ||||
6071 | ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n" , gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid), (*pteam )->t.t_pkfn); } | ||||
6072 | gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n" , gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid), (*pteam )->t.t_pkfn); } | ||||
6073 | (*pteam)->t.t_pkfn))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n" , gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid), (*pteam )->t.t_pkfn); }; | ||||
6074 | |||||
6075 | updateHWFPControl(*pteam); | ||||
6076 | |||||
6077 | #if OMPT_SUPPORT1 | ||||
6078 | if (ompt_enabled.enabled) { | ||||
6079 | this_thr->th.ompt_thread_info.state = ompt_state_work_parallel; | ||||
6080 | } | ||||
6081 | #endif | ||||
6082 | |||||
6083 | rc = (*pteam)->t.t_invoke(gtid); | ||||
6084 | KMP_ASSERT(rc)if (!(rc)) { __kmp_debug_assert("rc", "openmp/runtime/src/kmp_runtime.cpp" , 6084); }; | ||||
6085 | |||||
6086 | KMP_MB(); | ||||
6087 | KA_TRACE(20, ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n" , gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid), (*pteam )->t.t_pkfn); } | ||||
6088 | gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n" , gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid), (*pteam )->t.t_pkfn); } | ||||
6089 | (*pteam)->t.t_pkfn))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n" , gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid), (*pteam )->t.t_pkfn); }; | ||||
6090 | } | ||||
6091 | #if OMPT_SUPPORT1 | ||||
6092 | if (ompt_enabled.enabled) { | ||||
6093 | /* no frame set while outside task */ | ||||
6094 | __ompt_get_task_info_object(0)->frame.exit_frame = ompt_data_none{0}; | ||||
6095 | |||||
6096 | this_thr->th.ompt_thread_info.state = ompt_state_overhead; | ||||
6097 | } | ||||
6098 | #endif | ||||
6099 | /* join barrier after parallel region */ | ||||
6100 | __kmp_join_barrier(gtid); | ||||
6101 | } | ||||
6102 | } | ||||
6103 | TCR_SYNC_PTR((intptr_t)__kmp_global.g.g_done)((void *)((intptr_t)__kmp_global.g.g_done)); | ||||
6104 | |||||
6105 | #if OMPD_SUPPORT1 | ||||
6106 | if (ompd_state & OMPD_ENABLE_BP0x1) | ||||
6107 | ompd_bp_thread_end(); | ||||
6108 | #endif | ||||
6109 | |||||
6110 | #if OMPT_SUPPORT1 | ||||
6111 | if (ompt_enabled.ompt_callback_thread_end) { | ||||
6112 | ompt_callbacks.ompt_callback(ompt_callback_thread_end)ompt_callback_thread_end_callback(thread_data); | ||||
6113 | } | ||||
6114 | #endif | ||||
6115 | |||||
6116 | this_thr->th.th_task_team = NULL__null; | ||||
6117 | /* run the destructors for the threadprivate data for this thread */ | ||||
6118 | __kmp_common_destroy_gtid(gtid); | ||||
6119 | |||||
6120 | KA_TRACE(10, ("__kmp_launch_thread: T#%d done\n", gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_launch_thread: T#%d done\n" , gtid); }; | ||||
6121 | KMP_MB(); | ||||
6122 | |||||
6123 | #if OMP_PROFILING_SUPPORT0 | ||||
6124 | llvm::timeTraceProfilerFinishThread(); | ||||
6125 | #endif | ||||
6126 | return this_thr; | ||||
6127 | } | ||||
6128 | |||||
6129 | /* ------------------------------------------------------------------------ */ | ||||
6130 | |||||
6131 | void __kmp_internal_end_dest(void *specific_gtid) { | ||||
6132 | // Make sure no significant bits are lost | ||||
6133 | int gtid; | ||||
6134 | __kmp_type_convert((kmp_intptr_t)specific_gtid - 1, >id); | ||||
6135 | |||||
6136 | KA_TRACE(30, ("__kmp_internal_end_dest: T#%d\n", gtid))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_internal_end_dest: T#%d\n" , gtid); }; | ||||
6137 | /* NOTE: the gtid is stored as gitd+1 in the thread-local-storage | ||||
6138 | * this is because 0 is reserved for the nothing-stored case */ | ||||
6139 | |||||
6140 | __kmp_internal_end_thread(gtid); | ||||
6141 | } | ||||
6142 | |||||
6143 | #if KMP_OS_UNIX1 && KMP_DYNAMIC_LIB1 | ||||
6144 | |||||
6145 | __attribute__((destructor)) void __kmp_internal_end_dtor(void) { | ||||
6146 | __kmp_internal_end_atexit(); | ||||
6147 | } | ||||
6148 | |||||
6149 | #endif | ||||
6150 | |||||
6151 | /* [Windows] josh: when the atexit handler is called, there may still be more | ||||
6152 | than one thread alive */ | ||||
6153 | void __kmp_internal_end_atexit(void) { | ||||
6154 | KA_TRACE(30, ("__kmp_internal_end_atexit\n"))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_internal_end_atexit\n" ); }; | ||||
6155 | /* [Windows] | ||||
6156 | josh: ideally, we want to completely shutdown the library in this atexit | ||||
6157 | handler, but stat code that depends on thread specific data for gtid fails | ||||
6158 | because that data becomes unavailable at some point during the shutdown, so | ||||
6159 | we call __kmp_internal_end_thread instead. We should eventually remove the | ||||
6160 | dependency on __kmp_get_specific_gtid in the stat code and use | ||||
6161 | __kmp_internal_end_library to cleanly shutdown the library. | ||||
6162 | |||||
6163 | // TODO: Can some of this comment about GVS be removed? | ||||
6164 | I suspect that the offending stat code is executed when the calling thread | ||||
6165 | tries to clean up a dead root thread's data structures, resulting in GVS | ||||
6166 | code trying to close the GVS structures for that thread, but since the stat | ||||
6167 | code uses __kmp_get_specific_gtid to get the gtid with the assumption that | ||||
6168 | the calling thread is cleaning up itself instead of another thread, it get | ||||
6169 | confused. This happens because allowing a thread to unregister and cleanup | ||||
6170 | another thread is a recent modification for addressing an issue. | ||||
6171 | Based on the current design (20050722), a thread may end up | ||||
6172 | trying to unregister another thread only if thread death does not trigger | ||||
6173 | the calling of __kmp_internal_end_thread. For Linux* OS, there is the | ||||
6174 | thread specific data destructor function to detect thread death. For | ||||
6175 | Windows dynamic, there is DllMain(THREAD_DETACH). For Windows static, there | ||||
6176 | is nothing. Thus, the workaround is applicable only for Windows static | ||||
6177 | stat library. */ | ||||
6178 | __kmp_internal_end_library(-1); | ||||
6179 | #if KMP_OS_WINDOWS0 | ||||
6180 | __kmp_close_console(); | ||||
6181 | #endif | ||||
6182 | } | ||||
6183 | |||||
6184 | static void __kmp_reap_thread(kmp_info_t *thread, int is_root) { | ||||
6185 | // It is assumed __kmp_forkjoin_lock is acquired. | ||||
6186 | |||||
6187 | int gtid; | ||||
6188 | |||||
6189 | KMP_DEBUG_ASSERT(thread != NULL)if (!(thread != __null)) { __kmp_debug_assert("thread != __null" , "openmp/runtime/src/kmp_runtime.cpp", 6189); }; | ||||
6190 | |||||
6191 | gtid = thread->th.th_info.ds.ds_gtid; | ||||
6192 | |||||
6193 | if (!is_root) { | ||||
6194 | if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME(2147483647)) { | ||||
6195 | /* Assume the threads are at the fork barrier here */ | ||||
6196 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n" , gtid); } | ||||
6197 | 20, ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n" , gtid); } | ||||
6198 | gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n" , gtid); }; | ||||
6199 | if (__kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | ||||
6200 | while ( | ||||
6201 | !KMP_COMPARE_AND_STORE_ACQ32(&(thread->th.th_used_in_team), 0, 3)__sync_bool_compare_and_swap((volatile kmp_uint32 *)(&(thread ->th.th_used_in_team)), (kmp_uint32)(0), (kmp_uint32)(3))) | ||||
6202 | KMP_CPU_PAUSE()__kmp_x86_pause(); | ||||
6203 | __kmp_resume_32(gtid, (kmp_flag_32<false, false> *)NULL__null); | ||||
6204 | } else { | ||||
6205 | /* Need release fence here to prevent seg faults for tree forkjoin | ||||
6206 | barrier (GEH) */ | ||||
6207 | kmp_flag_64<> flag(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, | ||||
6208 | thread); | ||||
6209 | __kmp_release_64(&flag); | ||||
6210 | } | ||||
6211 | } | ||||
6212 | |||||
6213 | // Terminate OS thread. | ||||
6214 | __kmp_reap_worker(thread); | ||||
6215 | |||||
6216 | // The thread was killed asynchronously. If it was actively | ||||
6217 | // spinning in the thread pool, decrement the global count. | ||||
6218 | // | ||||
6219 | // There is a small timing hole here - if the worker thread was just waking | ||||
6220 | // up after sleeping in the pool, had reset it's th_active_in_pool flag but | ||||
6221 | // not decremented the global counter __kmp_thread_pool_active_nth yet, then | ||||
6222 | // the global counter might not get updated. | ||||
6223 | // | ||||
6224 | // Currently, this can only happen as the library is unloaded, | ||||
6225 | // so there are no harmful side effects. | ||||
6226 | if (thread->th.th_active_in_pool) { | ||||
6227 | thread->th.th_active_in_pool = FALSE0; | ||||
6228 | KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth)(&__kmp_thread_pool_active_nth)->fetch_sub(1, std::memory_order_acq_rel ); | ||||
6229 | KMP_DEBUG_ASSERT(__kmp_thread_pool_active_nth >= 0)if (!(__kmp_thread_pool_active_nth >= 0)) { __kmp_debug_assert ("__kmp_thread_pool_active_nth >= 0", "openmp/runtime/src/kmp_runtime.cpp" , 6229); }; | ||||
6230 | } | ||||
6231 | } | ||||
6232 | |||||
6233 | __kmp_free_implicit_task(thread); | ||||
6234 | |||||
6235 | // Free the fast memory for tasking | ||||
6236 | #if USE_FAST_MEMORY3 | ||||
6237 | __kmp_free_fast_memory(thread); | ||||
6238 | #endif /* USE_FAST_MEMORY */ | ||||
6239 | |||||
6240 | __kmp_suspend_uninitialize_thread(thread); | ||||
6241 | |||||
6242 | KMP_DEBUG_ASSERT(__kmp_threads[gtid] == thread)if (!(__kmp_threads[gtid] == thread)) { __kmp_debug_assert("__kmp_threads[gtid] == thread" , "openmp/runtime/src/kmp_runtime.cpp", 6242); }; | ||||
6243 | TCW_SYNC_PTR(__kmp_threads[gtid], NULL)((__kmp_threads[gtid])) = ((__null)); | ||||
6244 | |||||
6245 | --__kmp_all_nth; | ||||
6246 | // __kmp_nth was decremented when thread is added to the pool. | ||||
6247 | |||||
6248 | #ifdef KMP_ADJUST_BLOCKTIME1 | ||||
6249 | /* Adjust blocktime back to user setting or default if necessary */ | ||||
6250 | /* Middle initialization might never have occurred */ | ||||
6251 | if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { | ||||
6252 | KMP_DEBUG_ASSERT(__kmp_avail_proc > 0)if (!(__kmp_avail_proc > 0)) { __kmp_debug_assert("__kmp_avail_proc > 0" , "openmp/runtime/src/kmp_runtime.cpp", 6252); }; | ||||
6253 | if (__kmp_nth <= __kmp_avail_proc) { | ||||
6254 | __kmp_zero_bt = FALSE0; | ||||
6255 | } | ||||
6256 | } | ||||
6257 | #endif /* KMP_ADJUST_BLOCKTIME */ | ||||
6258 | |||||
6259 | /* free the memory being used */ | ||||
6260 | if (__kmp_env_consistency_check) { | ||||
6261 | if (thread->th.th_cons) { | ||||
6262 | __kmp_free_cons_stack(thread->th.th_cons); | ||||
6263 | thread->th.th_cons = NULL__null; | ||||
6264 | } | ||||
6265 | } | ||||
6266 | |||||
6267 | if (thread->th.th_pri_common != NULL__null) { | ||||
6268 | __kmp_free(thread->th.th_pri_common)___kmp_free((thread->th.th_pri_common), "openmp/runtime/src/kmp_runtime.cpp" , 6268); | ||||
6269 | thread->th.th_pri_common = NULL__null; | ||||
6270 | } | ||||
6271 | |||||
6272 | if (thread->th.th_task_state_memo_stack != NULL__null) { | ||||
6273 | __kmp_free(thread->th.th_task_state_memo_stack)___kmp_free((thread->th.th_task_state_memo_stack), "openmp/runtime/src/kmp_runtime.cpp" , 6273); | ||||
6274 | thread->th.th_task_state_memo_stack = NULL__null; | ||||
6275 | } | ||||
6276 | |||||
6277 | #if KMP_USE_BGET1 | ||||
6278 | if (thread->th.th_local.bget_data != NULL__null) { | ||||
6279 | __kmp_finalize_bget(thread); | ||||
6280 | } | ||||
6281 | #endif | ||||
6282 | |||||
6283 | #if KMP_AFFINITY_SUPPORTED1 | ||||
6284 | if (thread->th.th_affin_mask != NULL__null) { | ||||
6285 | KMP_CPU_FREE(thread->th.th_affin_mask)__kmp_affinity_dispatch->deallocate_mask(thread->th.th_affin_mask ); | ||||
6286 | thread->th.th_affin_mask = NULL__null; | ||||
6287 | } | ||||
6288 | #endif /* KMP_AFFINITY_SUPPORTED */ | ||||
6289 | |||||
6290 | #if KMP_USE_HIER_SCHED0 | ||||
6291 | if (thread->th.th_hier_bar_data != NULL__null) { | ||||
6292 | __kmp_free(thread->th.th_hier_bar_data)___kmp_free((thread->th.th_hier_bar_data), "openmp/runtime/src/kmp_runtime.cpp" , 6292); | ||||
6293 | thread->th.th_hier_bar_data = NULL__null; | ||||
6294 | } | ||||
6295 | #endif | ||||
6296 | |||||
6297 | __kmp_reap_team(thread->th.th_serial_team); | ||||
6298 | thread->th.th_serial_team = NULL__null; | ||||
6299 | __kmp_free(thread)___kmp_free((thread), "openmp/runtime/src/kmp_runtime.cpp", 6299 ); | ||||
6300 | |||||
6301 | KMP_MB(); | ||||
6302 | |||||
6303 | } // __kmp_reap_thread | ||||
6304 | |||||
6305 | static void __kmp_itthash_clean(kmp_info_t *th) { | ||||
6306 | #if USE_ITT_NOTIFY1 | ||||
6307 | if (__kmp_itt_region_domains.count > 0) { | ||||
6308 | for (int i = 0; i < KMP_MAX_FRAME_DOMAINS; ++i) { | ||||
6309 | kmp_itthash_entry_t *bucket = __kmp_itt_region_domains.buckets[i]; | ||||
6310 | while (bucket) { | ||||
6311 | kmp_itthash_entry_t *next = bucket->next_in_bucket; | ||||
6312 | __kmp_thread_free(th, bucket)___kmp_thread_free((th), (bucket), "openmp/runtime/src/kmp_runtime.cpp" , 6312); | ||||
6313 | bucket = next; | ||||
6314 | } | ||||
6315 | } | ||||
6316 | } | ||||
6317 | if (__kmp_itt_barrier_domains.count > 0) { | ||||
6318 | for (int i = 0; i < KMP_MAX_FRAME_DOMAINS; ++i) { | ||||
6319 | kmp_itthash_entry_t *bucket = __kmp_itt_barrier_domains.buckets[i]; | ||||
6320 | while (bucket) { | ||||
6321 | kmp_itthash_entry_t *next = bucket->next_in_bucket; | ||||
6322 | __kmp_thread_free(th, bucket)___kmp_thread_free((th), (bucket), "openmp/runtime/src/kmp_runtime.cpp" , 6322); | ||||
6323 | bucket = next; | ||||
6324 | } | ||||
6325 | } | ||||
6326 | } | ||||
6327 | #endif | ||||
6328 | } | ||||
6329 | |||||
6330 | static void __kmp_internal_end(void) { | ||||
6331 | int i; | ||||
6332 | |||||
6333 | /* First, unregister the library */ | ||||
6334 | __kmp_unregister_library(); | ||||
6335 | |||||
6336 | #if KMP_OS_WINDOWS0 | ||||
6337 | /* In Win static library, we can't tell when a root actually dies, so we | ||||
6338 | reclaim the data structures for any root threads that have died but not | ||||
6339 | unregistered themselves, in order to shut down cleanly. | ||||
6340 | In Win dynamic library we also can't tell when a thread dies. */ | ||||
6341 | __kmp_reclaim_dead_roots(); // AC: moved here to always clean resources of | ||||
6342 | // dead roots | ||||
6343 | #endif | ||||
6344 | |||||
6345 | for (i = 0; i < __kmp_threads_capacity; i++) | ||||
6346 | if (__kmp_root[i]) | ||||
6347 | if (__kmp_root[i]->r.r_active) | ||||
6348 | break; | ||||
6349 | KMP_MB(); /* Flush all pending memory write invalidates. */ | ||||
6350 | TCW_SYNC_4(__kmp_global.g.g_done, TRUE)(__kmp_global.g.g_done) = ((!0)); | ||||
6351 | |||||
6352 | if (i < __kmp_threads_capacity) { | ||||
6353 | #if KMP_USE_MONITOR | ||||
6354 | // 2009-09-08 (lev): Other alive roots found. Why do we kill the monitor?? | ||||
6355 | KMP_MB(); /* Flush all pending memory write invalidates. */ | ||||
6356 | |||||
6357 | // Need to check that monitor was initialized before reaping it. If we are | ||||
6358 | // called form __kmp_atfork_child (which sets __kmp_init_parallel = 0), then | ||||
6359 | // __kmp_monitor will appear to contain valid data, but it is only valid in | ||||
6360 | // the parent process, not the child. | ||||
6361 | // New behavior (201008): instead of keying off of the flag | ||||
6362 | // __kmp_init_parallel, the monitor thread creation is keyed off | ||||
6363 | // of the new flag __kmp_init_monitor. | ||||
6364 | __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock); | ||||
6365 | if (TCR_4(__kmp_init_monitor)(__kmp_init_monitor)) { | ||||
6366 | __kmp_reap_monitor(&__kmp_monitor); | ||||
6367 | TCW_4(__kmp_init_monitor, 0)(__kmp_init_monitor) = (0); | ||||
6368 | } | ||||
6369 | __kmp_release_bootstrap_lock(&__kmp_monitor_lock); | ||||
6370 | KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end: monitor reaped\n" ); }; | ||||
6371 | #endif // KMP_USE_MONITOR | ||||
6372 | } else { | ||||
6373 | /* TODO move this to cleanup code */ | ||||
6374 | #ifdef KMP_DEBUG1 | ||||
6375 | /* make sure that everything has properly ended */ | ||||
6376 | for (i = 0; i < __kmp_threads_capacity; i++) { | ||||
6377 | if (__kmp_root[i]) { | ||||
6378 | // KMP_ASSERT( ! KMP_UBER_GTID( i ) ); // AC: | ||||
6379 | // there can be uber threads alive here | ||||
6380 | KMP_ASSERT(!__kmp_root[i]->r.r_active)if (!(!__kmp_root[i]->r.r_active)) { __kmp_debug_assert("!__kmp_root[i]->r.r_active" , "openmp/runtime/src/kmp_runtime.cpp", 6380); }; // TODO: can they be active? | ||||
6381 | } | ||||
6382 | } | ||||
6383 | #endif | ||||
6384 | |||||
6385 | KMP_MB(); | ||||
6386 | |||||
6387 | // Reap the worker threads. | ||||
6388 | // This is valid for now, but be careful if threads are reaped sooner. | ||||
6389 | while (__kmp_thread_pool != NULL__null) { // Loop thru all the thread in the pool. | ||||
6390 | // Get the next thread from the pool. | ||||
6391 | kmp_info_t *thread = CCAST(kmp_info_t *, __kmp_thread_pool)const_cast<kmp_info_t *>(__kmp_thread_pool); | ||||
6392 | __kmp_thread_pool = thread->th.th_next_pool; | ||||
6393 | // Reap it. | ||||
6394 | KMP_DEBUG_ASSERT(thread->th.th_reap_state == KMP_SAFE_TO_REAP)if (!(thread->th.th_reap_state == 1)) { __kmp_debug_assert ("thread->th.th_reap_state == 1", "openmp/runtime/src/kmp_runtime.cpp" , 6394); }; | ||||
6395 | thread->th.th_next_pool = NULL__null; | ||||
6396 | thread->th.th_in_pool = FALSE0; | ||||
6397 | __kmp_reap_thread(thread, 0); | ||||
6398 | } | ||||
6399 | __kmp_thread_pool_insert_pt = NULL__null; | ||||
6400 | |||||
6401 | // Reap teams. | ||||
6402 | while (__kmp_team_pool != NULL__null) { // Loop thru all the teams in the pool. | ||||
6403 | // Get the next team from the pool. | ||||
6404 | kmp_team_t *team = CCAST(kmp_team_t *, __kmp_team_pool)const_cast<kmp_team_t *>(__kmp_team_pool); | ||||
6405 | __kmp_team_pool = team->t.t_next_pool; | ||||
6406 | // Reap it. | ||||
6407 | team->t.t_next_pool = NULL__null; | ||||
6408 | __kmp_reap_team(team); | ||||
6409 | } | ||||
6410 | |||||
6411 | __kmp_reap_task_teams(); | ||||
6412 | |||||
6413 | #if KMP_OS_UNIX1 | ||||
6414 | // Threads that are not reaped should not access any resources since they | ||||
6415 | // are going to be deallocated soon, so the shutdown sequence should wait | ||||
6416 | // until all threads either exit the final spin-waiting loop or begin | ||||
6417 | // sleeping after the given blocktime. | ||||
6418 | for (i = 0; i < __kmp_threads_capacity; i++) { | ||||
6419 | kmp_info_t *thr = __kmp_threads[i]; | ||||
6420 | while (thr && KMP_ATOMIC_LD_ACQ(&thr->th.th_blocking)(&thr->th.th_blocking)->load(std::memory_order_acquire )) | ||||
6421 | KMP_CPU_PAUSE()__kmp_x86_pause(); | ||||
6422 | } | ||||
6423 | #endif | ||||
6424 | |||||
6425 | for (i = 0; i < __kmp_threads_capacity; ++i) { | ||||
6426 | // TBD: Add some checking... | ||||
6427 | // Something like KMP_DEBUG_ASSERT( __kmp_thread[ i ] == NULL ); | ||||
6428 | } | ||||
6429 | |||||
6430 | /* Make sure all threadprivate destructors get run by joining with all | ||||
6431 | worker threads before resetting this flag */ | ||||
6432 | TCW_SYNC_4(__kmp_init_common, FALSE)(__kmp_init_common) = (0); | ||||
6433 | |||||
6434 | KA_TRACE(10, ("__kmp_internal_end: all workers reaped\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end: all workers reaped\n" ); }; | ||||
6435 | KMP_MB(); | ||||
6436 | |||||
6437 | #if KMP_USE_MONITOR | ||||
6438 | // See note above: One of the possible fixes for CQ138434 / CQ140126 | ||||
6439 | // | ||||
6440 | // FIXME: push both code fragments down and CSE them? | ||||
6441 | // push them into __kmp_cleanup() ? | ||||
6442 | __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock); | ||||
6443 | if (TCR_4(__kmp_init_monitor)(__kmp_init_monitor)) { | ||||
6444 | __kmp_reap_monitor(&__kmp_monitor); | ||||
6445 | TCW_4(__kmp_init_monitor, 0)(__kmp_init_monitor) = (0); | ||||
6446 | } | ||||
6447 | __kmp_release_bootstrap_lock(&__kmp_monitor_lock); | ||||
6448 | KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end: monitor reaped\n" ); }; | ||||
6449 | #endif | ||||
6450 | } /* else !__kmp_global.t_active */ | ||||
6451 | TCW_4(__kmp_init_gtid, FALSE)(__kmp_init_gtid) = (0); | ||||
6452 | KMP_MB(); /* Flush all pending memory write invalidates. */ | ||||
6453 | |||||
6454 | __kmp_cleanup(); | ||||
6455 | #if OMPT_SUPPORT1 | ||||
6456 | ompt_fini(); | ||||
6457 | #endif | ||||
6458 | } | ||||
6459 | |||||
6460 | void __kmp_internal_end_library(int gtid_req) { | ||||
6461 | /* if we have already cleaned up, don't try again, it wouldn't be pretty */ | ||||
6462 | /* this shouldn't be a race condition because __kmp_internal_end() is the | ||||
6463 | only place to clear __kmp_serial_init */ | ||||
6464 | /* we'll check this later too, after we get the lock */ | ||||
6465 | // 2009-09-06: We do not set g_abort without setting g_done. This check looks | ||||
6466 | // redundant, because the next check will work in any case. | ||||
6467 | if (__kmp_global.g.g_abort) { | ||||
6468 | KA_TRACE(11, ("__kmp_internal_end_library: abort, exiting\n"))if (kmp_a_debug >= 11) { __kmp_debug_printf ("__kmp_internal_end_library: abort, exiting\n" ); }; | ||||
6469 | /* TODO abort? */ | ||||
6470 | return; | ||||
6471 | } | ||||
6472 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done) || !__kmp_init_serial) { | ||||
6473 | KA_TRACE(10, ("__kmp_internal_end_library: already finished\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: already finished\n" ); }; | ||||
6474 | return; | ||||
6475 | } | ||||
6476 | |||||
6477 | // If hidden helper team has been initialized, we need to deinit it | ||||
6478 | if (TCR_4(__kmp_init_hidden_helper)(__kmp_init_hidden_helper) && | ||||
6479 | !TCR_4(__kmp_hidden_helper_team_done)(__kmp_hidden_helper_team_done)) { | ||||
6480 | TCW_SYNC_4(__kmp_hidden_helper_team_done, TRUE)(__kmp_hidden_helper_team_done) = ((!0)); | ||||
6481 | // First release the main thread to let it continue its work | ||||
6482 | __kmp_hidden_helper_main_thread_release(); | ||||
6483 | // Wait until the hidden helper team has been destroyed | ||||
6484 | __kmp_hidden_helper_threads_deinitz_wait(); | ||||
6485 | } | ||||
6486 | |||||
6487 | KMP_MB(); /* Flush all pending memory write invalidates. */ | ||||
6488 | /* find out who we are and what we should do */ | ||||
6489 | { | ||||
6490 | int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific(); | ||||
6491 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: enter T#%d (%d)\n" , gtid, gtid_req); } | ||||
6492 | 10, ("__kmp_internal_end_library: enter T#%d (%d)\n", gtid, gtid_req))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: enter T#%d (%d)\n" , gtid, gtid_req); }; | ||||
6493 | if (gtid == KMP_GTID_SHUTDOWN(-3)) { | ||||
6494 | KA_TRACE(10, ("__kmp_internal_end_library: !__kmp_init_runtime, system "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: !__kmp_init_runtime, system " "already shutdown\n"); } | ||||
6495 | "already shutdown\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: !__kmp_init_runtime, system " "already shutdown\n"); }; | ||||
6496 | return; | ||||
6497 | } else if (gtid == KMP_GTID_MONITOR(-4)) { | ||||
6498 | KA_TRACE(10, ("__kmp_internal_end_library: monitor thread, gtid not "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: monitor thread, gtid not " "registered, or system shutdown\n"); } | ||||
6499 | "registered, or system shutdown\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: monitor thread, gtid not " "registered, or system shutdown\n"); }; | ||||
6500 | return; | ||||
6501 | } else if (gtid == KMP_GTID_DNE(-2)) { | ||||
6502 | KA_TRACE(10, ("__kmp_internal_end_library: gtid not registered or system "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: gtid not registered or system " "shutdown\n"); } | ||||
6503 | "shutdown\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: gtid not registered or system " "shutdown\n"); }; | ||||
6504 | /* we don't know who we are, but we may still shutdown the library */ | ||||
6505 | } else if (KMP_UBER_GTID(gtid)) { | ||||
6506 | /* unregister ourselves as an uber thread. gtid is no longer valid */ | ||||
6507 | if (__kmp_root[gtid]->r.r_active) { | ||||
6508 | __kmp_global.g.g_abort = -1; | ||||
6509 | TCW_SYNC_4(__kmp_global.g.g_done, TRUE)(__kmp_global.g.g_done) = ((!0)); | ||||
6510 | __kmp_unregister_library(); | ||||
6511 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: root still active, abort T#%d\n" , gtid); } | ||||
6512 | ("__kmp_internal_end_library: root still active, abort T#%d\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: root still active, abort T#%d\n" , gtid); } | ||||
6513 | gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: root still active, abort T#%d\n" , gtid); }; | ||||
6514 | return; | ||||
6515 | } else { | ||||
6516 | __kmp_itthash_clean(__kmp_threads[gtid]); | ||||
6517 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: unregistering sibling T#%d\n" , gtid); } | ||||
6518 | 10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: unregistering sibling T#%d\n" , gtid); } | ||||
6519 | ("__kmp_internal_end_library: unregistering sibling T#%d\n", gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: unregistering sibling T#%d\n" , gtid); }; | ||||
6520 | __kmp_unregister_root_current_thread(gtid); | ||||
6521 | } | ||||
6522 | } else { | ||||
6523 | /* worker threads may call this function through the atexit handler, if they | ||||
6524 | * call exit() */ | ||||
6525 | /* For now, skip the usual subsequent processing and just dump the debug buffer. | ||||
6526 | TODO: do a thorough shutdown instead */ | ||||
6527 | #ifdef DUMP_DEBUG_ON_EXIT | ||||
6528 | if (__kmp_debug_buf) | ||||
6529 | __kmp_dump_debug_buffer(); | ||||
6530 | #endif | ||||
6531 | // added unregister library call here when we switch to shm linux | ||||
6532 | // if we don't, it will leave lots of files in /dev/shm | ||||
6533 | // cleanup shared memory file before exiting. | ||||
6534 | __kmp_unregister_library(); | ||||
6535 | return; | ||||
6536 | } | ||||
6537 | } | ||||
6538 | /* synchronize the termination process */ | ||||
6539 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); | ||||
6540 | |||||
6541 | /* have we already finished */ | ||||
6542 | if (__kmp_global.g.g_abort) { | ||||
6543 | KA_TRACE(10, ("__kmp_internal_end_library: abort, exiting\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: abort, exiting\n" ); }; | ||||
6544 | /* TODO abort? */ | ||||
6545 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | ||||
6546 | return; | ||||
6547 | } | ||||
6548 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done) || !__kmp_init_serial) { | ||||
6549 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | ||||
6550 | return; | ||||
6551 | } | ||||
6552 | |||||
6553 | /* We need this lock to enforce mutex between this reading of | ||||
6554 | __kmp_threads_capacity and the writing by __kmp_register_root. | ||||
6555 | Alternatively, we can use a counter of roots that is atomically updated by | ||||
6556 | __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and | ||||
6557 | __kmp_internal_end_*. */ | ||||
6558 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); | ||||
6559 | |||||
6560 | /* now we can safely conduct the actual termination */ | ||||
6561 | __kmp_internal_end(); | ||||
6562 | |||||
6563 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | ||||
6564 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | ||||
6565 | |||||
6566 | KA_TRACE(10, ("__kmp_internal_end_library: exit\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: exit\n" ); }; | ||||
6567 | |||||
6568 | #ifdef DUMP_DEBUG_ON_EXIT | ||||
6569 | if (__kmp_debug_buf) | ||||
6570 | __kmp_dump_debug_buffer(); | ||||
6571 | #endif | ||||
6572 | |||||
6573 | #if KMP_OS_WINDOWS0 | ||||
6574 | __kmp_close_console(); | ||||
6575 | #endif | ||||
6576 | |||||
6577 | __kmp_fini_allocator(); | ||||
6578 | |||||
6579 | } // __kmp_internal_end_library | ||||
6580 | |||||
6581 | void __kmp_internal_end_thread(int gtid_req) { | ||||
6582 | int i; | ||||
6583 | |||||
6584 | /* if we have already cleaned up, don't try again, it wouldn't be pretty */ | ||||
6585 | /* this shouldn't be a race condition because __kmp_internal_end() is the | ||||
6586 | * only place to clear __kmp_serial_init */ | ||||
6587 | /* we'll check this later too, after we get the lock */ | ||||
6588 | // 2009-09-06: We do not set g_abort without setting g_done. This check looks | ||||
6589 | // redundant, because the next check will work in any case. | ||||
6590 | if (__kmp_global.g.g_abort) { | ||||
6591 | KA_TRACE(11, ("__kmp_internal_end_thread: abort, exiting\n"))if (kmp_a_debug >= 11) { __kmp_debug_printf ("__kmp_internal_end_thread: abort, exiting\n" ); }; | ||||
6592 | /* TODO abort? */ | ||||
6593 | return; | ||||
6594 | } | ||||
6595 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done) || !__kmp_init_serial) { | ||||
6596 | KA_TRACE(10, ("__kmp_internal_end_thread: already finished\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: already finished\n" ); }; | ||||
6597 | return; | ||||
6598 | } | ||||
6599 | |||||
6600 | // If hidden helper team has been initialized, we need to deinit it | ||||
6601 | if (TCR_4(__kmp_init_hidden_helper)(__kmp_init_hidden_helper) && | ||||
6602 | !TCR_4(__kmp_hidden_helper_team_done)(__kmp_hidden_helper_team_done)) { | ||||
6603 | TCW_SYNC_4(__kmp_hidden_helper_team_done, TRUE)(__kmp_hidden_helper_team_done) = ((!0)); | ||||
6604 | // First release the main thread to let it continue its work | ||||
6605 | __kmp_hidden_helper_main_thread_release(); | ||||
6606 | // Wait until the hidden helper team has been destroyed | ||||
6607 | __kmp_hidden_helper_threads_deinitz_wait(); | ||||
6608 | } | ||||
6609 | |||||
6610 | KMP_MB(); /* Flush all pending memory write invalidates. */ | ||||
6611 | |||||
6612 | /* find out who we are and what we should do */ | ||||
6613 | { | ||||
6614 | int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific(); | ||||
6615 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: enter T#%d (%d)\n" , gtid, gtid_req); } | ||||
6616 | ("__kmp_internal_end_thread: enter T#%d (%d)\n", gtid, gtid_req))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: enter T#%d (%d)\n" , gtid, gtid_req); }; | ||||
6617 | if (gtid == KMP_GTID_SHUTDOWN(-3)) { | ||||
6618 | KA_TRACE(10, ("__kmp_internal_end_thread: !__kmp_init_runtime, system "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: !__kmp_init_runtime, system " "already shutdown\n"); } | ||||
6619 | "already shutdown\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: !__kmp_init_runtime, system " "already shutdown\n"); }; | ||||
6620 | return; | ||||
6621 | } else if (gtid == KMP_GTID_MONITOR(-4)) { | ||||
6622 | KA_TRACE(10, ("__kmp_internal_end_thread: monitor thread, gtid not "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: monitor thread, gtid not " "registered, or system shutdown\n"); } | ||||
6623 | "registered, or system shutdown\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: monitor thread, gtid not " "registered, or system shutdown\n"); }; | ||||
6624 | return; | ||||
6625 | } else if (gtid == KMP_GTID_DNE(-2)) { | ||||
6626 | KA_TRACE(10, ("__kmp_internal_end_thread: gtid not registered or system "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: gtid not registered or system " "shutdown\n"); } | ||||
6627 | "shutdown\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: gtid not registered or system " "shutdown\n"); }; | ||||
6628 | return; | ||||
6629 | /* we don't know who we are */ | ||||
6630 | } else if (KMP_UBER_GTID(gtid)) { | ||||
6631 | /* unregister ourselves as an uber thread. gtid is no longer valid */ | ||||
6632 | if (__kmp_root[gtid]->r.r_active) { | ||||
6633 | __kmp_global.g.g_abort = -1; | ||||
6634 | TCW_SYNC_4(__kmp_global.g.g_done, TRUE)(__kmp_global.g.g_done) = ((!0)); | ||||
6635 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: root still active, abort T#%d\n" , gtid); } | ||||
6636 | ("__kmp_internal_end_thread: root still active, abort T#%d\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: root still active, abort T#%d\n" , gtid); } | ||||
6637 | gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: root still active, abort T#%d\n" , gtid); }; | ||||
6638 | return; | ||||
6639 | } else { | ||||
6640 | KA_TRACE(10, ("__kmp_internal_end_thread: unregistering sibling T#%d\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: unregistering sibling T#%d\n" , gtid); } | ||||
6641 | gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: unregistering sibling T#%d\n" , gtid); }; | ||||
6642 | __kmp_unregister_root_current_thread(gtid); | ||||
6643 | } | ||||
6644 | } else { | ||||
6645 | /* just a worker thread, let's leave */ | ||||
6646 | KA_TRACE(10, ("__kmp_internal_end_thread: worker thread T#%d\n", gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: worker thread T#%d\n" , gtid); }; | ||||
6647 | |||||
6648 | if (gtid >= 0) { | ||||
6649 | __kmp_threads[gtid]->th.th_task_team = NULL__null; | ||||
6650 | } | ||||
6651 | |||||
6652 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n" , gtid); } | ||||
6653 | ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n" , gtid); } | ||||
6654 | gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n" , gtid); }; | ||||
6655 | return; | ||||
6656 | } | ||||
6657 | } | ||||
6658 | #if KMP_DYNAMIC_LIB1 | ||||
6659 | if (__kmp_pause_status != kmp_hard_paused) | ||||
6660 | // AC: lets not shutdown the dynamic library at the exit of uber thread, | ||||
6661 | // because we will better shutdown later in the library destructor. | ||||
6662 | { | ||||
6663 | KA_TRACE(10, ("__kmp_internal_end_thread: exiting T#%d\n", gtid_req))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: exiting T#%d\n" , gtid_req); }; | ||||
6664 | return; | ||||
6665 | } | ||||
6666 | #endif | ||||
6667 | /* synchronize the termination process */ | ||||
6668 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); | ||||
6669 | |||||
6670 | /* have we already finished */ | ||||
6671 | if (__kmp_global.g.g_abort) { | ||||
6672 | KA_TRACE(10, ("__kmp_internal_end_thread: abort, exiting\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: abort, exiting\n" ); }; | ||||
6673 | /* TODO abort? */ | ||||
6674 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | ||||
6675 | return; | ||||
6676 | } | ||||
6677 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done) || !__kmp_init_serial) { | ||||
6678 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | ||||
6679 | return; | ||||
6680 | } | ||||
6681 | |||||
6682 | /* We need this lock to enforce mutex between this reading of | ||||
6683 | __kmp_threads_capacity and the writing by __kmp_register_root. | ||||
6684 | Alternatively, we can use a counter of roots that is atomically updated by | ||||
6685 | __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and | ||||
6686 | __kmp_internal_end_*. */ | ||||
6687 | |||||
6688 | /* should we finish the run-time? are all siblings done? */ | ||||
6689 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); | ||||
6690 | |||||
6691 | for (i = 0; i < __kmp_threads_capacity; ++i) { | ||||
6692 | if (KMP_UBER_GTID(i)) { | ||||
6693 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n" , i); } | ||||
6694 | 10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n" , i); } | ||||
6695 | ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n", i))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n" , i); }; | ||||
6696 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | ||||
6697 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | ||||
6698 | return; | ||||
6699 | } | ||||
6700 | } | ||||
6701 | |||||
6702 | /* now we can safely conduct the actual termination */ | ||||
6703 | |||||
6704 | __kmp_internal_end(); | ||||
6705 | |||||
6706 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | ||||
6707 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | ||||
6708 | |||||
6709 | KA_TRACE(10, ("__kmp_internal_end_thread: exit T#%d\n", gtid_req))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: exit T#%d\n" , gtid_req); }; | ||||
6710 | |||||
6711 | #ifdef DUMP_DEBUG_ON_EXIT | ||||
6712 | if (__kmp_debug_buf) | ||||
6713 | __kmp_dump_debug_buffer(); | ||||
6714 | #endif | ||||
6715 | } // __kmp_internal_end_thread | ||||
6716 | |||||
6717 | // ----------------------------------------------------------------------------- | ||||
6718 | // Library registration stuff. | ||||
6719 | |||||
6720 | static long __kmp_registration_flag = 0; | ||||
6721 | // Random value used to indicate library initialization. | ||||
6722 | static char *__kmp_registration_str = NULL__null; | ||||
6723 | // Value to be saved in env var __KMP_REGISTERED_LIB_<pid>. | ||||
6724 | |||||
6725 | static inline char *__kmp_reg_status_name() { | ||||
6726 | /* On RHEL 3u5 if linked statically, getpid() returns different values in | ||||
6727 | each thread. If registration and unregistration go in different threads | ||||
6728 | (omp_misc_other_root_exit.cpp test case), the name of registered_lib_env | ||||
6729 | env var can not be found, because the name will contain different pid. */ | ||||
6730 | // macOS* complains about name being too long with additional getuid() | ||||
6731 | #if KMP_OS_UNIX1 && !KMP_OS_DARWIN0 && KMP_DYNAMIC_LIB1 | ||||
6732 | return __kmp_str_format("__KMP_REGISTERED_LIB_%d_%d", (int)getpid(), | ||||
6733 | (int)getuid()); | ||||
6734 | #else | ||||
6735 | return __kmp_str_format("__KMP_REGISTERED_LIB_%d", (int)getpid()); | ||||
6736 | #endif | ||||
6737 | } // __kmp_reg_status_get | ||||
6738 | |||||
6739 | #if defined(KMP_USE_SHM) | ||||
6740 | // If /dev/shm is not accessible, we will create a temporary file under /tmp. | ||||
6741 | char *temp_reg_status_file_name = nullptr; | ||||
6742 | #endif | ||||
6743 | |||||
6744 | void __kmp_register_library_startup(void) { | ||||
6745 | |||||
6746 | char *name = __kmp_reg_status_name(); // Name of the environment variable. | ||||
6747 | int done = 0; | ||||
6748 | union { | ||||
6749 | double dtime; | ||||
6750 | long ltime; | ||||
6751 | } time; | ||||
6752 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 | ||||
6753 | __kmp_initialize_system_tick(); | ||||
6754 | #endif | ||||
6755 | __kmp_read_system_time(&time.dtime); | ||||
6756 | __kmp_registration_flag = 0xCAFE0000L | (time.ltime & 0x0000FFFFL); | ||||
6757 | __kmp_registration_str = | ||||
6758 | __kmp_str_format("%p-%lx-%s", &__kmp_registration_flag, | ||||
6759 | __kmp_registration_flag, KMP_LIBRARY_FILE"libomp.so.5"); | ||||
6760 | |||||
6761 | KA_TRACE(50, ("__kmp_register_library_startup: %s=\"%s\"\n", name,if (kmp_a_debug >= 50) { __kmp_debug_printf ("__kmp_register_library_startup: %s=\"%s\"\n" , name, __kmp_registration_str); } | ||||
6762 | __kmp_registration_str))if (kmp_a_debug >= 50) { __kmp_debug_printf ("__kmp_register_library_startup: %s=\"%s\"\n" , name, __kmp_registration_str); }; | ||||
6763 | |||||
6764 | while (!done) { | ||||
6765 | |||||
6766 | char *value = NULL__null; // Actual value of the environment variable. | ||||
6767 | |||||
6768 | #if defined(KMP_USE_SHM) | ||||
6769 | char *shm_name = __kmp_str_format("/%s", name); | ||||
6770 | int shm_preexist = 0; | ||||
6771 | char *data1; | ||||
6772 | int fd1 = shm_open(shm_name, O_CREAT0100 | O_EXCL0200 | O_RDWR02, 0666); | ||||
6773 | if ((fd1 == -1) && (errno(*__errno_location ()) == EEXIST17)) { | ||||
6774 | // file didn't open because it already exists. | ||||
6775 | // try opening existing file | ||||
6776 | fd1 = shm_open(shm_name, O_RDWR02, 0666); | ||||
6777 | if (fd1 == -1) { // file didn't open | ||||
6778 | // error out here | ||||
6779 | __kmp_fatal(KMP_MSG(FunctionError, "Can't open SHM")__kmp_msg_format(kmp_i18n_msg_FunctionError, "Can't open SHM" ), KMP_ERR(0)__kmp_msg_error_code(0), | ||||
6780 | __kmp_msg_null); | ||||
6781 | } else { | ||||
6782 | // able to open existing file | ||||
6783 | shm_preexist = 1; | ||||
6784 | } | ||||
6785 | } else if (fd1 == -1) { | ||||
6786 | // SHM didn't open; it was due to error other than already exists. Try to | ||||
6787 | // create a temp file under /tmp. | ||||
6788 | // TODO: /tmp might not always be the temporary directory. For now we will | ||||
6789 | // not consider TMPDIR. If /tmp is not accessible, we simply error out. | ||||
6790 | char *temp_file_name = __kmp_str_format("/tmp/%sXXXXXX", name); | ||||
6791 | fd1 = mkstemp(temp_file_name); | ||||
6792 | if (fd1 == -1) { | ||||
6793 | // error out here. | ||||
6794 | __kmp_fatal(KMP_MSG(FunctionError, "Can't open TEMP")__kmp_msg_format(kmp_i18n_msg_FunctionError, "Can't open TEMP" ), KMP_ERR(errno)__kmp_msg_error_code((*__errno_location ())), | ||||
6795 | __kmp_msg_null); | ||||
6796 | } | ||||
6797 | temp_reg_status_file_name = temp_file_name; | ||||
6798 | } | ||||
6799 | if (shm_preexist == 0) { | ||||
6800 | // we created SHM now set size | ||||
6801 | if (ftruncate(fd1, SHM_SIZE1024) == -1) { | ||||
6802 | // error occured setting size; | ||||
6803 | __kmp_fatal(KMP_MSG(FunctionError, "Can't set size of SHM")__kmp_msg_format(kmp_i18n_msg_FunctionError, "Can't set size of SHM" ), | ||||
6804 | KMP_ERR(errno)__kmp_msg_error_code((*__errno_location ())), __kmp_msg_null); | ||||
6805 | } | ||||
6806 | } | ||||
6807 | data1 = | ||||
6808 | (char *)mmap(0, SHM_SIZE1024, PROT_READ0x1 | PROT_WRITE0x2, MAP_SHARED0x01, fd1, 0); | ||||
6809 | if (data1 == MAP_FAILED((void *) -1)) { | ||||
6810 | // failed to map shared memory | ||||
6811 | __kmp_fatal(KMP_MSG(FunctionError, "Can't map SHM")__kmp_msg_format(kmp_i18n_msg_FunctionError, "Can't map SHM"), KMP_ERR(errno)__kmp_msg_error_code((*__errno_location ())), | ||||
6812 | __kmp_msg_null); | ||||
6813 | } | ||||
6814 | if (shm_preexist == 0) { // set data to SHM, set value | ||||
6815 | KMP_STRCPY_S(data1, SHM_SIZE, __kmp_registration_str)strcpy(data1, __kmp_registration_str); | ||||
6816 | } | ||||
6817 | // Read value from either what we just wrote or existing file. | ||||
6818 | value = __kmp_str_format("%s", data1); // read value from SHM | ||||
6819 | munmap(data1, SHM_SIZE1024); | ||||
6820 | close(fd1); | ||||
6821 | #else // Windows and unix with static library | ||||
6822 | // Set environment variable, but do not overwrite if it is exist. | ||||
6823 | __kmp_env_set(name, __kmp_registration_str, 0); | ||||
6824 | // read value to see if it got set | ||||
6825 | value = __kmp_env_get(name); | ||||
6826 | #endif | ||||
6827 | |||||
6828 | if (value != NULL__null && strcmp(value, __kmp_registration_str) == 0) { | ||||
6829 | done = 1; // Ok, environment variable set successfully, exit the loop. | ||||
6830 | } else { | ||||
6831 | // Oops. Write failed. Another copy of OpenMP RTL is in memory. | ||||
6832 | // Check whether it alive or dead. | ||||
6833 | int neighbor = 0; // 0 -- unknown status, 1 -- alive, 2 -- dead. | ||||
6834 | char *tail = value; | ||||
6835 | char *flag_addr_str = NULL__null; | ||||
6836 | char *flag_val_str = NULL__null; | ||||
6837 | char const *file_name = NULL__null; | ||||
6838 | __kmp_str_split(tail, '-', &flag_addr_str, &tail); | ||||
6839 | __kmp_str_split(tail, '-', &flag_val_str, &tail); | ||||
6840 | file_name = tail; | ||||
6841 | if (tail != NULL__null) { | ||||
6842 | unsigned long *flag_addr = 0; | ||||
6843 | unsigned long flag_val = 0; | ||||
6844 | KMP_SSCANFsscanf(flag_addr_str, "%p", RCAST(void **, &flag_addr)reinterpret_cast<void **>(&flag_addr)); | ||||
6845 | KMP_SSCANFsscanf(flag_val_str, "%lx", &flag_val); | ||||
6846 | if (flag_addr != 0 && flag_val != 0 && strcmp(file_name, "") != 0) { | ||||
6847 | // First, check whether environment-encoded address is mapped into | ||||
6848 | // addr space. | ||||
6849 | // If so, dereference it to see if it still has the right value. | ||||
6850 | if (__kmp_is_address_mapped(flag_addr) && *flag_addr == flag_val) { | ||||
6851 | neighbor = 1; | ||||
6852 | } else { | ||||
6853 | // If not, then we know the other copy of the library is no longer | ||||
6854 | // running. | ||||
6855 | neighbor = 2; | ||||
6856 | } | ||||
6857 | } | ||||
6858 | } | ||||
6859 | switch (neighbor) { | ||||
6860 | case 0: // Cannot parse environment variable -- neighbor status unknown. | ||||
6861 | // Assume it is the incompatible format of future version of the | ||||
6862 | // library. Assume the other library is alive. | ||||
6863 | // WARN( ... ); // TODO: Issue a warning. | ||||
6864 | file_name = "unknown library"; | ||||
6865 | KMP_FALLTHROUGH()[[fallthrough]]; | ||||
6866 | // Attention! Falling to the next case. That's intentional. | ||||
6867 | case 1: { // Neighbor is alive. | ||||
6868 | // Check it is allowed. | ||||
6869 | char *duplicate_ok = __kmp_env_get("KMP_DUPLICATE_LIB_OK"); | ||||
6870 | if (!__kmp_str_match_true(duplicate_ok)) { | ||||
6871 | // That's not allowed. Issue fatal error. | ||||
6872 | __kmp_fatal(KMP_MSG(DuplicateLibrary, KMP_LIBRARY_FILE, file_name)__kmp_msg_format(kmp_i18n_msg_DuplicateLibrary, "libomp.so.5" , file_name), | ||||
6873 | KMP_HNT(DuplicateLibrary)__kmp_msg_format(kmp_i18n_hnt_DuplicateLibrary), __kmp_msg_null); | ||||
6874 | } | ||||
6875 | KMP_INTERNAL_FREE(duplicate_ok)free(duplicate_ok); | ||||
6876 | __kmp_duplicate_library_ok = 1; | ||||
6877 | done = 1; // Exit the loop. | ||||
6878 | } break; | ||||
6879 | case 2: { // Neighbor is dead. | ||||
6880 | |||||
6881 | #if defined(KMP_USE_SHM) | ||||
6882 | // close shared memory. | ||||
6883 | shm_unlink(shm_name); // this removes file in /dev/shm | ||||
6884 | #else | ||||
6885 | // Clear the variable and try to register library again. | ||||
6886 | __kmp_env_unset(name); | ||||
6887 | #endif | ||||
6888 | } break; | ||||
6889 | default: { | ||||
6890 | KMP_DEBUG_ASSERT(0)if (!(0)) { __kmp_debug_assert("0", "openmp/runtime/src/kmp_runtime.cpp" , 6890); }; | ||||
6891 | } break; | ||||
6892 | } | ||||
6893 | } | ||||
6894 | KMP_INTERNAL_FREE((void *)value)free((void *)value); | ||||
6895 | #if defined(KMP_USE_SHM) | ||||
6896 | KMP_INTERNAL_FREE((void *)shm_name)free((void *)shm_name); | ||||
6897 | #endif | ||||
6898 | } // while | ||||
6899 | KMP_INTERNAL_FREE((void *)name)free((void *)name); | ||||
6900 | |||||
6901 | } // func __kmp_register_library_startup | ||||
6902 | |||||
6903 | void __kmp_unregister_library(void) { | ||||
6904 | |||||
6905 | char *name = __kmp_reg_status_name(); | ||||
6906 | char *value = NULL__null; | ||||
6907 | |||||
6908 | #if defined(KMP_USE_SHM) | ||||
6909 | bool use_shm = true; | ||||
6910 | char *shm_name = __kmp_str_format("/%s", name); | ||||
6911 | int fd1 = shm_open(shm_name, O_RDONLY00, 0666); | ||||
6912 | if (fd1 == -1) { | ||||
6913 | // File did not open. Try the temporary file. | ||||
6914 | use_shm = false; | ||||
6915 | KMP_DEBUG_ASSERT(temp_reg_status_file_name)if (!(temp_reg_status_file_name)) { __kmp_debug_assert("temp_reg_status_file_name" , "openmp/runtime/src/kmp_runtime.cpp", 6915); }; | ||||
6916 | fd1 = open(temp_reg_status_file_name, O_RDONLY00); | ||||
6917 | if (fd1 == -1) { | ||||
6918 | // give it up now. | ||||
6919 | return; | ||||
6920 | } | ||||
6921 | } | ||||
6922 | char *data1 = (char *)mmap(0, SHM_SIZE1024, PROT_READ0x1, MAP_SHARED0x01, fd1, 0); | ||||
6923 | if (data1 != MAP_FAILED((void *) -1)) { | ||||
6924 | value = __kmp_str_format("%s", data1); // read value from SHM | ||||
6925 | munmap(data1, SHM_SIZE1024); | ||||
6926 | } | ||||
6927 | close(fd1); | ||||
6928 | #else | ||||
6929 | value = __kmp_env_get(name); | ||||
6930 | #endif | ||||
6931 | |||||
6932 | KMP_DEBUG_ASSERT(__kmp_registration_flag != 0)if (!(__kmp_registration_flag != 0)) { __kmp_debug_assert("__kmp_registration_flag != 0" , "openmp/runtime/src/kmp_runtime.cpp", 6932); }; | ||||
6933 | KMP_DEBUG_ASSERT(__kmp_registration_str != NULL)if (!(__kmp_registration_str != __null)) { __kmp_debug_assert ("__kmp_registration_str != __null", "openmp/runtime/src/kmp_runtime.cpp" , 6933); }; | ||||
6934 | if (value != NULL__null && strcmp(value, __kmp_registration_str) == 0) { | ||||
6935 | // Ok, this is our variable. Delete it. | ||||
6936 | #if defined(KMP_USE_SHM) | ||||
6937 | if (use_shm) { | ||||
6938 | shm_unlink(shm_name); // this removes file in /dev/shm | ||||
6939 | } else { | ||||
6940 | KMP_DEBUG_ASSERT(temp_reg_status_file_name)if (!(temp_reg_status_file_name)) { __kmp_debug_assert("temp_reg_status_file_name" , "openmp/runtime/src/kmp_runtime.cpp", 6940); }; | ||||
6941 | unlink(temp_reg_status_file_name); // this removes the temp file | ||||
6942 | } | ||||
6943 | #else | ||||
6944 | __kmp_env_unset(name); | ||||
6945 | #endif | ||||
6946 | } | ||||
6947 | |||||
6948 | #if defined(KMP_USE_SHM) | ||||
6949 | KMP_INTERNAL_FREE(shm_name)free(shm_name); | ||||
6950 | if (!use_shm) { | ||||
6951 | KMP_DEBUG_ASSERT(temp_reg_status_file_name)if (!(temp_reg_status_file_name)) { __kmp_debug_assert("temp_reg_status_file_name" , "openmp/runtime/src/kmp_runtime.cpp", 6951); }; | ||||
6952 | KMP_INTERNAL_FREE(temp_reg_status_file_name)free(temp_reg_status_file_name); | ||||
6953 | } | ||||
6954 | #endif | ||||
6955 | |||||
6956 | KMP_INTERNAL_FREE(__kmp_registration_str)free(__kmp_registration_str); | ||||
6957 | KMP_INTERNAL_FREE(value)free(value); | ||||
6958 | KMP_INTERNAL_FREE(name)free(name); | ||||
6959 | |||||
6960 | __kmp_registration_flag = 0; | ||||
6961 | __kmp_registration_str = NULL__null; | ||||
6962 | |||||
6963 | } // __kmp_unregister_library | ||||
6964 | |||||
6965 | // End of Library registration stuff. | ||||
6966 | // ----------------------------------------------------------------------------- | ||||
6967 | |||||
6968 | #if KMP_MIC_SUPPORTED((0 || 1) && (1 || 0)) | ||||
6969 | |||||
6970 | static void __kmp_check_mic_type() { | ||||
6971 | kmp_cpuid_t cpuid_state = {0}; | ||||
6972 | kmp_cpuid_t *cs_p = &cpuid_state; | ||||
6973 | __kmp_x86_cpuid(1, 0, cs_p); | ||||
6974 | // We don't support mic1 at the moment | ||||
6975 | if ((cs_p->eax & 0xff0) == 0xB10) { | ||||
6976 | __kmp_mic_type = mic2; | ||||
6977 | } else if ((cs_p->eax & 0xf0ff0) == 0x50670) { | ||||
6978 | __kmp_mic_type = mic3; | ||||
6979 | } else { | ||||
6980 | __kmp_mic_type = non_mic; | ||||
6981 | } | ||||
6982 | } | ||||
6983 | |||||
6984 | #endif /* KMP_MIC_SUPPORTED */ | ||||
6985 | |||||
6986 | #if KMP_HAVE_UMWAIT((0 || 1) && (1 || 0) && !0) | ||||
6987 | static void __kmp_user_level_mwait_init() { | ||||
6988 | struct kmp_cpuid buf; | ||||
6989 | __kmp_x86_cpuid(7, 0, &buf); | ||||
6990 | __kmp_waitpkg_enabled = ((buf.ecx >> 5) & 1); | ||||
6991 | __kmp_umwait_enabled = __kmp_waitpkg_enabled && __kmp_user_level_mwait; | ||||
6992 | __kmp_tpause_enabled = __kmp_waitpkg_enabled && (__kmp_tpause_state > 0); | ||||
6993 | KF_TRACE(30, ("__kmp_user_level_mwait_init: __kmp_umwait_enabled = %d\n",if (kmp_f_debug >= 30) { __kmp_debug_printf ("__kmp_user_level_mwait_init: __kmp_umwait_enabled = %d\n" , __kmp_umwait_enabled); } | ||||
6994 | __kmp_umwait_enabled))if (kmp_f_debug >= 30) { __kmp_debug_printf ("__kmp_user_level_mwait_init: __kmp_umwait_enabled = %d\n" , __kmp_umwait_enabled); }; | ||||
6995 | } | ||||
6996 | #elif KMP_HAVE_MWAIT((0 || 1) && (1 || 0) && !0) | ||||
6997 | #ifndef AT_INTELPHIUSERMWAIT | ||||
6998 | // Spurious, non-existent value that should always fail to return anything. | ||||
6999 | // Will be replaced with the correct value when we know that. | ||||
7000 | #define AT_INTELPHIUSERMWAIT 10000 | ||||
7001 | #endif | ||||
7002 | // getauxval() function is available in RHEL7 and SLES12. If a system with an | ||||
7003 | // earlier OS is used to build the RTL, we'll use the following internal | ||||
7004 | // function when the entry is not found. | ||||
7005 | unsigned long getauxval(unsigned long) KMP_WEAK_ATTRIBUTE_EXTERNAL; | ||||
7006 | unsigned long getauxval(unsigned long) { return 0; } | ||||
7007 | |||||
7008 | static void __kmp_user_level_mwait_init() { | ||||
7009 | // When getauxval() and correct value of AT_INTELPHIUSERMWAIT are available | ||||
7010 | // use them to find if the user-level mwait is enabled. Otherwise, forcibly | ||||
7011 | // set __kmp_mwait_enabled=TRUE on Intel MIC if the environment variable | ||||
7012 | // KMP_USER_LEVEL_MWAIT was set to TRUE. | ||||
7013 | if (__kmp_mic_type == mic3) { | ||||
7014 | unsigned long res = getauxval(AT_INTELPHIUSERMWAIT); | ||||
7015 | if ((res & 0x1) || __kmp_user_level_mwait) { | ||||
7016 | __kmp_mwait_enabled = TRUE(!0); | ||||
7017 | if (__kmp_user_level_mwait) { | ||||
7018 | KMP_INFORM(EnvMwaitWarn)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_EnvMwaitWarn ), __kmp_msg_null); | ||||
7019 | } | ||||
7020 | } else { | ||||
7021 | __kmp_mwait_enabled = FALSE0; | ||||
7022 | } | ||||
7023 | } | ||||
7024 | KF_TRACE(30, ("__kmp_user_level_mwait_init: __kmp_mic_type = %d, "if (kmp_f_debug >= 30) { __kmp_debug_printf ("__kmp_user_level_mwait_init: __kmp_mic_type = %d, " "__kmp_mwait_enabled = %d\n", __kmp_mic_type, __kmp_mwait_enabled ); } | ||||
7025 | "__kmp_mwait_enabled = %d\n",if (kmp_f_debug >= 30) { __kmp_debug_printf ("__kmp_user_level_mwait_init: __kmp_mic_type = %d, " "__kmp_mwait_enabled = %d\n", __kmp_mic_type, __kmp_mwait_enabled ); } | ||||
7026 | __kmp_mic_type, __kmp_mwait_enabled))if (kmp_f_debug >= 30) { __kmp_debug_printf ("__kmp_user_level_mwait_init: __kmp_mic_type = %d, " "__kmp_mwait_enabled = %d\n", __kmp_mic_type, __kmp_mwait_enabled ); }; | ||||
7027 | } | ||||
7028 | #endif /* KMP_HAVE_UMWAIT */ | ||||
7029 | |||||
7030 | static void __kmp_do_serial_initialize(void) { | ||||
7031 | int i, gtid; | ||||
7032 | size_t size; | ||||
7033 | |||||
7034 | KA_TRACE(10, ("__kmp_do_serial_initialize: enter\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_do_serial_initialize: enter\n" ); }; | ||||
7035 | |||||
7036 | KMP_DEBUG_ASSERT(sizeof(kmp_int32) == 4)if (!(sizeof(kmp_int32) == 4)) { __kmp_debug_assert("sizeof(kmp_int32) == 4" , "openmp/runtime/src/kmp_runtime.cpp", 7036); }; | ||||
7037 | KMP_DEBUG_ASSERT(sizeof(kmp_uint32) == 4)if (!(sizeof(kmp_uint32) == 4)) { __kmp_debug_assert("sizeof(kmp_uint32) == 4" , "openmp/runtime/src/kmp_runtime.cpp", 7037); }; | ||||
7038 | KMP_DEBUG_ASSERT(sizeof(kmp_int64) == 8)if (!(sizeof(kmp_int64) == 8)) { __kmp_debug_assert("sizeof(kmp_int64) == 8" , "openmp/runtime/src/kmp_runtime.cpp", 7038); }; | ||||
7039 | KMP_DEBUG_ASSERT(sizeof(kmp_uint64) == 8)if (!(sizeof(kmp_uint64) == 8)) { __kmp_debug_assert("sizeof(kmp_uint64) == 8" , "openmp/runtime/src/kmp_runtime.cpp", 7039); }; | ||||
7040 | KMP_DEBUG_ASSERT(sizeof(kmp_intptr_t) == sizeof(void *))if (!(sizeof(kmp_intptr_t) == sizeof(void *))) { __kmp_debug_assert ("sizeof(kmp_intptr_t) == sizeof(void *)", "openmp/runtime/src/kmp_runtime.cpp" , 7040); }; | ||||
7041 | |||||
7042 | #if OMPT_SUPPORT1 | ||||
7043 | ompt_pre_init(); | ||||
7044 | #endif | ||||
7045 | #if OMPD_SUPPORT1 | ||||
7046 | __kmp_env_dump(); | ||||
7047 | ompd_init(); | ||||
7048 | #endif | ||||
7049 | |||||
7050 | __kmp_validate_locks(); | ||||
7051 | |||||
7052 | #if ENABLE_LIBOMPTARGET1 | ||||
7053 | /* Initialize functions from libomptarget */ | ||||
7054 | __kmp_init_omptarget(); | ||||
7055 | #endif | ||||
7056 | |||||
7057 | /* Initialize internal memory allocator */ | ||||
7058 | __kmp_init_allocator(); | ||||
7059 | |||||
7060 | /* Register the library startup via an environment variable or via mapped | ||||
7061 | shared memory file and check to see whether another copy of the library is | ||||
7062 | already registered. Since forked child process is often terminated, we | ||||
7063 | postpone the registration till middle initialization in the child */ | ||||
7064 | if (__kmp_need_register_serial) | ||||
7065 | __kmp_register_library_startup(); | ||||
7066 | |||||
7067 | /* TODO reinitialization of library */ | ||||
7068 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done)) { | ||||
7069 | KA_TRACE(10, ("__kmp_do_serial_initialize: reinitialization of library\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_do_serial_initialize: reinitialization of library\n" ); }; | ||||
7070 | } | ||||
7071 | |||||
7072 | __kmp_global.g.g_abort = 0; | ||||
7073 | TCW_SYNC_4(__kmp_global.g.g_done, FALSE)(__kmp_global.g.g_done) = (0); | ||||
7074 | |||||
7075 | /* initialize the locks */ | ||||
7076 | #if KMP_USE_ADAPTIVE_LOCKS(0 || 1) && !0 | ||||
7077 | #if KMP_DEBUG_ADAPTIVE_LOCKS0 | ||||
7078 | __kmp_init_speculative_stats(); | ||||
7079 | #endif | ||||
7080 | #endif | ||||
7081 | #if KMP_STATS_ENABLED0 | ||||
7082 | __kmp_stats_init(); | ||||
7083 | #endif | ||||
7084 | __kmp_init_lock(&__kmp_global_lock); | ||||
7085 | __kmp_init_queuing_lock(&__kmp_dispatch_lock); | ||||
7086 | __kmp_init_lock(&__kmp_debug_lock); | ||||
7087 | __kmp_init_atomic_lock(&__kmp_atomic_lock); | ||||
7088 | __kmp_init_atomic_lock(&__kmp_atomic_lock_1i); | ||||
7089 | __kmp_init_atomic_lock(&__kmp_atomic_lock_2i); | ||||
7090 | __kmp_init_atomic_lock(&__kmp_atomic_lock_4i); | ||||
7091 | __kmp_init_atomic_lock(&__kmp_atomic_lock_4r); | ||||
7092 | __kmp_init_atomic_lock(&__kmp_atomic_lock_8i); | ||||
7093 | __kmp_init_atomic_lock(&__kmp_atomic_lock_8r); | ||||
7094 | __kmp_init_atomic_lock(&__kmp_atomic_lock_8c); | ||||
7095 | __kmp_init_atomic_lock(&__kmp_atomic_lock_10r); | ||||
7096 | __kmp_init_atomic_lock(&__kmp_atomic_lock_16r); | ||||
7097 | __kmp_init_atomic_lock(&__kmp_atomic_lock_16c); | ||||
7098 | __kmp_init_atomic_lock(&__kmp_atomic_lock_20c); | ||||
7099 | __kmp_init_atomic_lock(&__kmp_atomic_lock_32c); | ||||
7100 | __kmp_init_bootstrap_lock(&__kmp_forkjoin_lock); | ||||
7101 | __kmp_init_bootstrap_lock(&__kmp_exit_lock); | ||||
7102 | #if KMP_USE_MONITOR | ||||
7103 | __kmp_init_bootstrap_lock(&__kmp_monitor_lock); | ||||
7104 | #endif | ||||
7105 | __kmp_init_bootstrap_lock(&__kmp_tp_cached_lock); | ||||
7106 | |||||
7107 | /* conduct initialization and initial setup of configuration */ | ||||
7108 | |||||
7109 | __kmp_runtime_initialize(); | ||||
7110 | |||||
7111 | #if KMP_MIC_SUPPORTED((0 || 1) && (1 || 0)) | ||||
7112 | __kmp_check_mic_type(); | ||||
7113 | #endif | ||||
7114 | |||||
7115 | // Some global variable initialization moved here from kmp_env_initialize() | ||||
7116 | #ifdef KMP_DEBUG1 | ||||
7117 | kmp_diag = 0; | ||||
7118 | #endif | ||||
7119 | __kmp_abort_delay = 0; | ||||
7120 | |||||
7121 | // From __kmp_init_dflt_team_nth() | ||||
7122 | /* assume the entire machine will be used */ | ||||
7123 | __kmp_dflt_team_nth_ub = __kmp_xproc; | ||||
7124 | if (__kmp_dflt_team_nth_ub < KMP_MIN_NTH1) { | ||||
7125 | __kmp_dflt_team_nth_ub = KMP_MIN_NTH1; | ||||
7126 | } | ||||
7127 | if (__kmp_dflt_team_nth_ub > __kmp_sys_max_nth) { | ||||
7128 | __kmp_dflt_team_nth_ub = __kmp_sys_max_nth; | ||||
7129 | } | ||||
7130 | __kmp_max_nth = __kmp_sys_max_nth; | ||||
7131 | __kmp_cg_max_nth = __kmp_sys_max_nth; | ||||
7132 | __kmp_teams_max_nth = __kmp_xproc; // set a "reasonable" default | ||||
7133 | if (__kmp_teams_max_nth > __kmp_sys_max_nth) { | ||||
7134 | __kmp_teams_max_nth = __kmp_sys_max_nth; | ||||
7135 | } | ||||
7136 | |||||
7137 | // Three vars below moved here from __kmp_env_initialize() "KMP_BLOCKTIME" | ||||
7138 | // part | ||||
7139 | __kmp_dflt_blocktime = KMP_DEFAULT_BLOCKTIME(__kmp_is_hybrid_cpu() ? (0) : (200)); | ||||
7140 | #if KMP_USE_MONITOR | ||||
7141 | __kmp_monitor_wakeups = | ||||
7142 | KMP_WAKEUPS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups); | ||||
7143 | __kmp_bt_intervals = | ||||
7144 | KMP_INTERVALS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups); | ||||
7145 | #endif | ||||
7146 | // From "KMP_LIBRARY" part of __kmp_env_initialize() | ||||
7147 | __kmp_library = library_throughput; | ||||
7148 | // From KMP_SCHEDULE initialization | ||||
7149 | __kmp_static = kmp_sch_static_balanced; | ||||
7150 | // AC: do not use analytical here, because it is non-monotonous | ||||
7151 | //__kmp_guided = kmp_sch_guided_iterative_chunked; | ||||
7152 | //__kmp_auto = kmp_sch_guided_analytical_chunked; // AC: it is the default, no | ||||
7153 | // need to repeat assignment | ||||
7154 | // Barrier initialization. Moved here from __kmp_env_initialize() Barrier branch | ||||
7155 | // bit control and barrier method control parts | ||||
7156 | #if KMP_FAST_REDUCTION_BARRIER1 | ||||
7157 | #define kmp_reduction_barrier_gather_bb ((int)1) | ||||
7158 | #define kmp_reduction_barrier_release_bb ((int)1) | ||||
7159 | #define kmp_reduction_barrier_gather_pat __kmp_barrier_gather_pat_dflt | ||||
7160 | #define kmp_reduction_barrier_release_pat __kmp_barrier_release_pat_dflt | ||||
7161 | #endif // KMP_FAST_REDUCTION_BARRIER | ||||
7162 | for (i = bs_plain_barrier; i < bs_last_barrier; i++) { | ||||
7163 | __kmp_barrier_gather_branch_bits[i] = __kmp_barrier_gather_bb_dflt; | ||||
7164 | __kmp_barrier_release_branch_bits[i] = __kmp_barrier_release_bb_dflt; | ||||
7165 | __kmp_barrier_gather_pattern[i] = __kmp_barrier_gather_pat_dflt; | ||||
7166 | __kmp_barrier_release_pattern[i] = __kmp_barrier_release_pat_dflt; | ||||
7167 | #if KMP_FAST_REDUCTION_BARRIER1 | ||||
7168 | if (i == bs_reduction_barrier) { // tested and confirmed on ALTIX only ( | ||||
7169 | // lin_64 ): hyper,1 | ||||
7170 | __kmp_barrier_gather_branch_bits[i] = kmp_reduction_barrier_gather_bb; | ||||
7171 | __kmp_barrier_release_branch_bits[i] = kmp_reduction_barrier_release_bb; | ||||
7172 | __kmp_barrier_gather_pattern[i] = kmp_reduction_barrier_gather_pat; | ||||
7173 | __kmp_barrier_release_pattern[i] = kmp_reduction_barrier_release_pat; | ||||
7174 | } | ||||
7175 | #endif // KMP_FAST_REDUCTION_BARRIER | ||||
7176 | } | ||||
7177 | #if KMP_FAST_REDUCTION_BARRIER1 | ||||
7178 | #undef kmp_reduction_barrier_release_pat | ||||
7179 | #undef kmp_reduction_barrier_gather_pat | ||||
7180 | #undef kmp_reduction_barrier_release_bb | ||||
7181 | #undef kmp_reduction_barrier_gather_bb | ||||
7182 | #endif // KMP_FAST_REDUCTION_BARRIER | ||||
7183 | #if KMP_MIC_SUPPORTED((0 || 1) && (1 || 0)) | ||||
7184 | if (__kmp_mic_type == mic2) { // KNC | ||||
7185 | // AC: plane=3,2, forkjoin=2,1 are optimal for 240 threads on KNC | ||||
7186 | __kmp_barrier_gather_branch_bits[bs_plain_barrier] = 3; // plain gather | ||||
7187 | __kmp_barrier_release_branch_bits[bs_forkjoin_barrier] = | ||||
7188 | 1; // forkjoin release | ||||
7189 | __kmp_barrier_gather_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar; | ||||
7190 | __kmp_barrier_release_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar; | ||||
7191 | } | ||||
7192 | #if KMP_FAST_REDUCTION_BARRIER1 | ||||
7193 | if (__kmp_mic_type == mic2) { // KNC | ||||
7194 | __kmp_barrier_gather_pattern[bs_reduction_barrier] = bp_hierarchical_bar; | ||||
7195 | __kmp_barrier_release_pattern[bs_reduction_barrier] = bp_hierarchical_bar; | ||||
7196 | } | ||||
7197 | #endif // KMP_FAST_REDUCTION_BARRIER | ||||
7198 | #endif // KMP_MIC_SUPPORTED | ||||
7199 | |||||
7200 | // From KMP_CHECKS initialization | ||||
7201 | #ifdef KMP_DEBUG1 | ||||
7202 | __kmp_env_checks = TRUE(!0); /* development versions have the extra checks */ | ||||
7203 | #else | ||||
7204 | __kmp_env_checks = FALSE0; /* port versions do not have the extra checks */ | ||||
7205 | #endif | ||||
7206 | |||||
7207 | // From "KMP_FOREIGN_THREADS_THREADPRIVATE" initialization | ||||
7208 | __kmp_foreign_tp = TRUE(!0); | ||||
7209 | |||||
7210 | __kmp_global.g.g_dynamic = FALSE0; | ||||
7211 | __kmp_global.g.g_dynamic_mode = dynamic_default; | ||||
7212 | |||||
7213 | __kmp_init_nesting_mode(); | ||||
7214 | |||||
7215 | __kmp_env_initialize(NULL__null); | ||||
7216 | |||||
7217 | #if KMP_HAVE_MWAIT((0 || 1) && (1 || 0) && !0) || KMP_HAVE_UMWAIT((0 || 1) && (1 || 0) && !0) | ||||
7218 | __kmp_user_level_mwait_init(); | ||||
7219 | #endif | ||||
7220 | // Print all messages in message catalog for testing purposes. | ||||
7221 | #ifdef KMP_DEBUG1 | ||||
7222 | char const *val = __kmp_env_get("KMP_DUMP_CATALOG"); | ||||
7223 | if (__kmp_str_match_true(val)) { | ||||
7224 | kmp_str_buf_t buffer; | ||||
7225 | __kmp_str_buf_init(&buffer){ (&buffer)->str = (&buffer)->bulk; (&buffer )->size = sizeof((&buffer)->bulk); (&buffer)-> used = 0; (&buffer)->bulk[0] = 0; }; | ||||
7226 | __kmp_i18n_dump_catalog(&buffer); | ||||
7227 | __kmp_printf("%s", buffer.str); | ||||
7228 | __kmp_str_buf_free(&buffer); | ||||
7229 | } | ||||
7230 | __kmp_env_free(&val); | ||||
7231 | #endif | ||||
7232 | |||||
7233 | __kmp_threads_capacity = | ||||
7234 | __kmp_initial_threads_capacity(__kmp_dflt_team_nth_ub); | ||||
7235 | // Moved here from __kmp_env_initialize() "KMP_ALL_THREADPRIVATE" part | ||||
7236 | __kmp_tp_capacity = __kmp_default_tp_capacity( | ||||
7237 | __kmp_dflt_team_nth_ub, __kmp_max_nth, __kmp_allThreadsSpecified); | ||||
7238 | |||||
7239 | // If the library is shut down properly, both pools must be NULL. Just in | ||||
7240 | // case, set them to NULL -- some memory may leak, but subsequent code will | ||||
7241 | // work even if pools are not freed. | ||||
7242 | KMP_DEBUG_ASSERT(__kmp_thread_pool == NULL)if (!(__kmp_thread_pool == __null)) { __kmp_debug_assert("__kmp_thread_pool == __null" , "openmp/runtime/src/kmp_runtime.cpp", 7242); }; | ||||
7243 | KMP_DEBUG_ASSERT(__kmp_thread_pool_insert_pt == NULL)if (!(__kmp_thread_pool_insert_pt == __null)) { __kmp_debug_assert ("__kmp_thread_pool_insert_pt == __null", "openmp/runtime/src/kmp_runtime.cpp" , 7243); }; | ||||
7244 | KMP_DEBUG_ASSERT(__kmp_team_pool == NULL)if (!(__kmp_team_pool == __null)) { __kmp_debug_assert("__kmp_team_pool == __null" , "openmp/runtime/src/kmp_runtime.cpp", 7244); }; | ||||
7245 | __kmp_thread_pool = NULL__null; | ||||
7246 | __kmp_thread_pool_insert_pt = NULL__null; | ||||
7247 | __kmp_team_pool = NULL__null; | ||||
7248 | |||||
7249 | /* Allocate all of the variable sized records */ | ||||
7250 | /* NOTE: __kmp_threads_capacity entries are allocated, but the arrays are | ||||
7251 | * expandable */ | ||||
7252 | /* Since allocation is cache-aligned, just add extra padding at the end */ | ||||
7253 | size = | ||||
7254 | (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * __kmp_threads_capacity + | ||||
7255 | CACHE_LINE64; | ||||
7256 | __kmp_threads = (kmp_info_t **)__kmp_allocate(size)___kmp_allocate((size), "openmp/runtime/src/kmp_runtime.cpp", 7256); | ||||
7257 | __kmp_root = (kmp_root_t **)((char *)__kmp_threads + | ||||
7258 | sizeof(kmp_info_t *) * __kmp_threads_capacity); | ||||
7259 | |||||
7260 | /* init thread counts */ | ||||
7261 | KMP_DEBUG_ASSERT(__kmp_all_nth ==if (!(__kmp_all_nth == 0)) { __kmp_debug_assert("__kmp_all_nth == 0" , "openmp/runtime/src/kmp_runtime.cpp", 7262); } | ||||
7262 | 0)if (!(__kmp_all_nth == 0)) { __kmp_debug_assert("__kmp_all_nth == 0" , "openmp/runtime/src/kmp_runtime.cpp", 7262); }; // Asserts fail if the library is reinitializing and | ||||
7263 | KMP_DEBUG_ASSERT(__kmp_nth == 0)if (!(__kmp_nth == 0)) { __kmp_debug_assert("__kmp_nth == 0", "openmp/runtime/src/kmp_runtime.cpp", 7263); }; // something was wrong in termination. | ||||
7264 | __kmp_all_nth = 0; | ||||
7265 | __kmp_nth = 0; | ||||
7266 | |||||
7267 | /* setup the uber master thread and hierarchy */ | ||||
7268 | gtid = __kmp_register_root(TRUE(!0)); | ||||
7269 | KA_TRACE(10, ("__kmp_do_serial_initialize T#%d\n", gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_do_serial_initialize T#%d\n" , gtid); }; | ||||
7270 | KMP_ASSERT(KMP_UBER_GTID(gtid))if (!(KMP_UBER_GTID(gtid))) { __kmp_debug_assert("KMP_UBER_GTID(gtid)" , "openmp/runtime/src/kmp_runtime.cpp", 7270); }; | ||||
7271 | KMP_ASSERT(KMP_INITIAL_GTID(gtid))if (!((0 == (gtid)))) { __kmp_debug_assert("KMP_INITIAL_GTID(gtid)" , "openmp/runtime/src/kmp_runtime.cpp", 7271); }; | ||||
7272 | |||||
7273 | KMP_MB(); /* Flush all pending memory write invalidates. */ | ||||
7274 | |||||
7275 | __kmp_common_initialize(); | ||||
7276 | |||||
7277 | #if KMP_OS_UNIX1 | ||||
7278 | /* invoke the child fork handler */ | ||||
7279 | __kmp_register_atfork(); | ||||
7280 | #endif | ||||
7281 | |||||
7282 | #if !KMP_DYNAMIC_LIB1 || \ | ||||
7283 | ((KMP_COMPILER_ICC0 || KMP_COMPILER_ICX0) && KMP_OS_DARWIN0) | ||||
7284 | { | ||||
7285 | /* Invoke the exit handler when the program finishes, only for static | ||||
7286 | library and macOS* dynamic. For other dynamic libraries, we already | ||||
7287 | have _fini and DllMain. */ | ||||
7288 | int rc = atexit(__kmp_internal_end_atexit); | ||||
7289 | if (rc != 0) { | ||||
7290 | __kmp_fatal(KMP_MSG(FunctionError, "atexit()")__kmp_msg_format(kmp_i18n_msg_FunctionError, "atexit()"), KMP_ERR(rc)__kmp_msg_error_code(rc), | ||||
7291 | __kmp_msg_null); | ||||
7292 | } | ||||
7293 | } | ||||
7294 | #endif | ||||
7295 | |||||
7296 | #if KMP_HANDLE_SIGNALS(1 || 0) | ||||
7297 | #if KMP_OS_UNIX1 | ||||
7298 | /* NOTE: make sure that this is called before the user installs their own | ||||
7299 | signal handlers so that the user handlers are called first. this way they | ||||
7300 | can return false, not call our handler, avoid terminating the library, and | ||||
7301 | continue execution where they left off. */ | ||||
7302 | __kmp_install_signals(FALSE0); | ||||
7303 | #endif /* KMP_OS_UNIX */ | ||||
7304 | #if KMP_OS_WINDOWS0 | ||||
7305 | __kmp_install_signals(TRUE(!0)); | ||||
7306 | #endif /* KMP_OS_WINDOWS */ | ||||
7307 | #endif | ||||
7308 | |||||
7309 | /* we have finished the serial initialization */ | ||||
7310 | __kmp_init_counter++; | ||||
7311 | |||||
7312 | __kmp_init_serial = TRUE(!0); | ||||
7313 | |||||
7314 | if (__kmp_settings) { | ||||
7315 | __kmp_env_print(); | ||||
7316 | } | ||||
7317 | |||||
7318 | if (__kmp_display_env || __kmp_display_env_verbose) { | ||||
7319 | __kmp_env_print_2(); | ||||
7320 | } | ||||
7321 | |||||
7322 | #if OMPT_SUPPORT1 | ||||
7323 | ompt_post_init(); | ||||
7324 | #endif | ||||
7325 | |||||
7326 | KMP_MB(); | ||||
7327 | |||||
7328 | KA_TRACE(10, ("__kmp_do_serial_initialize: exit\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_do_serial_initialize: exit\n" ); }; | ||||
7329 | } | ||||
7330 | |||||
7331 | void __kmp_serial_initialize(void) { | ||||
7332 | if (__kmp_init_serial) { | ||||
7333 | return; | ||||
7334 | } | ||||
7335 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); | ||||
7336 | if (__kmp_init_serial) { | ||||
7337 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | ||||
7338 | return; | ||||
7339 | } | ||||
7340 | __kmp_do_serial_initialize(); | ||||
7341 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | ||||
7342 | } | ||||
7343 | |||||
7344 | static void __kmp_do_middle_initialize(void) { | ||||
7345 | int i, j; | ||||
7346 | int prev_dflt_team_nth; | ||||
7347 | |||||
7348 | if (!__kmp_init_serial) { | ||||
7349 | __kmp_do_serial_initialize(); | ||||
7350 | } | ||||
7351 | |||||
7352 | KA_TRACE(10, ("__kmp_middle_initialize: enter\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_middle_initialize: enter\n" ); }; | ||||
7353 | |||||
7354 | if (UNLIKELY(!__kmp_need_register_serial)__builtin_expect(!!(!__kmp_need_register_serial), 0)) { | ||||
7355 | // We are in a forked child process. The registration was skipped during | ||||
7356 | // serial initialization in __kmp_atfork_child handler. Do it here. | ||||
7357 | __kmp_register_library_startup(); | ||||
7358 | } | ||||
7359 | |||||
7360 | // Save the previous value for the __kmp_dflt_team_nth so that | ||||
7361 | // we can avoid some reinitialization if it hasn't changed. | ||||
7362 | prev_dflt_team_nth = __kmp_dflt_team_nth; | ||||
7363 | |||||
7364 | #if KMP_AFFINITY_SUPPORTED1 | ||||
7365 | // __kmp_affinity_initialize() will try to set __kmp_ncores to the | ||||
7366 | // number of cores on the machine. | ||||
7367 | __kmp_affinity_initialize(__kmp_affinity); | ||||
7368 | |||||
7369 | #endif /* KMP_AFFINITY_SUPPORTED */ | ||||
7370 | |||||
7371 | KMP_ASSERT(__kmp_xproc > 0)if (!(__kmp_xproc > 0)) { __kmp_debug_assert("__kmp_xproc > 0" , "openmp/runtime/src/kmp_runtime.cpp", 7371); }; | ||||
7372 | if (__kmp_avail_proc == 0) { | ||||
7373 | __kmp_avail_proc = __kmp_xproc; | ||||
7374 | } | ||||
7375 | |||||
7376 | // If there were empty places in num_threads list (OMP_NUM_THREADS=,,2,3), | ||||
7377 | // correct them now | ||||
7378 | j = 0; | ||||
7379 | while ((j < __kmp_nested_nth.used) && !__kmp_nested_nth.nth[j]) { | ||||
7380 | __kmp_nested_nth.nth[j] = __kmp_dflt_team_nth = __kmp_dflt_team_nth_ub = | ||||
7381 | __kmp_avail_proc; | ||||
7382 | j++; | ||||
7383 | } | ||||
7384 | |||||
7385 | if (__kmp_dflt_team_nth == 0) { | ||||
7386 | #ifdef KMP_DFLT_NTH_CORES | ||||
7387 | // Default #threads = #cores | ||||
7388 | __kmp_dflt_team_nth = __kmp_ncores; | ||||
7389 | KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = " "__kmp_ncores (%d)\n", __kmp_dflt_team_nth); } | ||||
7390 | "__kmp_ncores (%d)\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = " "__kmp_ncores (%d)\n", __kmp_dflt_team_nth); } | ||||
7391 | __kmp_dflt_team_nth))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = " "__kmp_ncores (%d)\n", __kmp_dflt_team_nth); }; | ||||
7392 | #else | ||||
7393 | // Default #threads = #available OS procs | ||||
7394 | __kmp_dflt_team_nth = __kmp_avail_proc; | ||||
7395 | KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = " "__kmp_avail_proc(%d)\n", __kmp_dflt_team_nth); } | ||||
7396 | "__kmp_avail_proc(%d)\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = " "__kmp_avail_proc(%d)\n", __kmp_dflt_team_nth); } | ||||
7397 | __kmp_dflt_team_nth))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = " "__kmp_avail_proc(%d)\n", __kmp_dflt_team_nth); }; | ||||
7398 | #endif /* KMP_DFLT_NTH_CORES */ | ||||
7399 | } | ||||
7400 | |||||
7401 | if (__kmp_dflt_team_nth < KMP_MIN_NTH1) { | ||||
7402 | __kmp_dflt_team_nth = KMP_MIN_NTH1; | ||||
7403 | } | ||||
7404 | if (__kmp_dflt_team_nth > __kmp_sys_max_nth) { | ||||
7405 | __kmp_dflt_team_nth = __kmp_sys_max_nth; | ||||
7406 | } | ||||
7407 | |||||
7408 | if (__kmp_nesting_mode > 0) | ||||
7409 | __kmp_set_nesting_mode_threads(); | ||||
7410 | |||||
7411 | // There's no harm in continuing if the following check fails, | ||||
7412 | // but it indicates an error in the previous logic. | ||||
7413 | KMP_DEBUG_ASSERT(__kmp_dflt_team_nth <= __kmp_dflt_team_nth_ub)if (!(__kmp_dflt_team_nth <= __kmp_dflt_team_nth_ub)) { __kmp_debug_assert ("__kmp_dflt_team_nth <= __kmp_dflt_team_nth_ub", "openmp/runtime/src/kmp_runtime.cpp" , 7413); }; | ||||
7414 | |||||
7415 | if (__kmp_dflt_team_nth != prev_dflt_team_nth) { | ||||
7416 | // Run through the __kmp_threads array and set the num threads icv for each | ||||
7417 | // root thread that is currently registered with the RTL (which has not | ||||
7418 | // already explicitly set its nthreads-var with a call to | ||||
7419 | // omp_set_num_threads()). | ||||
7420 | for (i = 0; i < __kmp_threads_capacity; i++) { | ||||
7421 | kmp_info_t *thread = __kmp_threads[i]; | ||||
7422 | if (thread == NULL__null) | ||||
7423 | continue; | ||||
7424 | if (thread->th.th_current_task->td_icvs.nproc != 0) | ||||
7425 | continue; | ||||
7426 | |||||
7427 | set__nproc(__kmp_threads[i], __kmp_dflt_team_nth)(((__kmp_threads[i])->th.th_current_task->td_icvs.nproc ) = (__kmp_dflt_team_nth)); | ||||
7428 | } | ||||
7429 | } | ||||
7430 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n" , __kmp_dflt_team_nth); } | ||||
7431 | 20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n" , __kmp_dflt_team_nth); } | ||||
7432 | ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n" , __kmp_dflt_team_nth); } | ||||
7433 | __kmp_dflt_team_nth))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n" , __kmp_dflt_team_nth); }; | ||||
7434 | |||||
7435 | #ifdef KMP_ADJUST_BLOCKTIME1 | ||||
7436 | /* Adjust blocktime to zero if necessary now that __kmp_avail_proc is set */ | ||||
7437 | if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { | ||||
7438 | KMP_DEBUG_ASSERT(__kmp_avail_proc > 0)if (!(__kmp_avail_proc > 0)) { __kmp_debug_assert("__kmp_avail_proc > 0" , "openmp/runtime/src/kmp_runtime.cpp", 7438); }; | ||||
7439 | if (__kmp_nth > __kmp_avail_proc) { | ||||
7440 | __kmp_zero_bt = TRUE(!0); | ||||
7441 | } | ||||
7442 | } | ||||
7443 | #endif /* KMP_ADJUST_BLOCKTIME */ | ||||
7444 | |||||
7445 | /* we have finished middle initialization */ | ||||
7446 | TCW_SYNC_4(__kmp_init_middle, TRUE)(__kmp_init_middle) = ((!0)); | ||||
7447 | |||||
7448 | KA_TRACE(10, ("__kmp_do_middle_initialize: exit\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_do_middle_initialize: exit\n" ); }; | ||||
7449 | } | ||||
7450 | |||||
7451 | void __kmp_middle_initialize(void) { | ||||
7452 | if (__kmp_init_middle) { | ||||
7453 | return; | ||||
7454 | } | ||||
7455 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); | ||||
7456 | if (__kmp_init_middle) { | ||||
7457 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | ||||
7458 | return; | ||||
7459 | } | ||||
7460 | __kmp_do_middle_initialize(); | ||||
7461 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | ||||
7462 | } | ||||
7463 | |||||
7464 | void __kmp_parallel_initialize(void) { | ||||
7465 | int gtid = __kmp_entry_gtid()__kmp_get_global_thread_id_reg(); // this might be a new root | ||||
7466 | |||||
7467 | /* synchronize parallel initialization (for sibling) */ | ||||
7468 | if (TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) | ||||
7469 | return; | ||||
7470 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); | ||||
7471 | if (TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) { | ||||
7472 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | ||||
7473 | return; | ||||
7474 | } | ||||
7475 | |||||
7476 | /* TODO reinitialization after we have already shut down */ | ||||
7477 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done)) { | ||||
7478 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_parallel_initialize: attempt to init while shutting down\n" ); } | ||||
7479 | 10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_parallel_initialize: attempt to init while shutting down\n" ); } | ||||
7480 | ("__kmp_parallel_initialize: attempt to init while shutting down\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_parallel_initialize: attempt to init while shutting down\n" ); }; | ||||
7481 | __kmp_infinite_loop(); | ||||
7482 | } | ||||
7483 | |||||
7484 | /* jc: The lock __kmp_initz_lock is already held, so calling | ||||
7485 | __kmp_serial_initialize would cause a deadlock. So we call | ||||
7486 | __kmp_do_serial_initialize directly. */ | ||||
7487 | if (!__kmp_init_middle) { | ||||
7488 | __kmp_do_middle_initialize(); | ||||
7489 | } | ||||
7490 | __kmp_assign_root_init_mask(); | ||||
7491 | __kmp_resume_if_hard_paused(); | ||||
7492 | |||||
7493 | /* begin initialization */ | ||||
7494 | KA_TRACE(10, ("__kmp_parallel_initialize: enter\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_parallel_initialize: enter\n" ); }; | ||||
7495 | KMP_ASSERT(KMP_UBER_GTID(gtid))if (!(KMP_UBER_GTID(gtid))) { __kmp_debug_assert("KMP_UBER_GTID(gtid)" , "openmp/runtime/src/kmp_runtime.cpp", 7495); }; | ||||
7496 | |||||
7497 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 | ||||
7498 | // Save the FP control regs. | ||||
7499 | // Worker threads will set theirs to these values at thread startup. | ||||
7500 | __kmp_store_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word); | ||||
7501 | __kmp_store_mxcsr(&__kmp_init_mxcsr); | ||||
7502 | __kmp_init_mxcsr &= KMP_X86_MXCSR_MASK0xffffffc0; | ||||
7503 | #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ | ||||
7504 | |||||
7505 | #if KMP_OS_UNIX1 | ||||
7506 | #if KMP_HANDLE_SIGNALS(1 || 0) | ||||
7507 | /* must be after __kmp_serial_initialize */ | ||||
7508 | __kmp_install_signals(TRUE(!0)); | ||||
7509 | #endif | ||||
7510 | #endif | ||||
7511 | |||||
7512 | __kmp_suspend_initialize(); | ||||
7513 | |||||
7514 | #if defined(USE_LOAD_BALANCE1) | ||||
7515 | if (__kmp_global.g.g_dynamic_mode == dynamic_default) { | ||||
7516 | __kmp_global.g.g_dynamic_mode = dynamic_load_balance; | ||||
7517 | } | ||||
7518 | #else | ||||
7519 | if (__kmp_global.g.g_dynamic_mode == dynamic_default) { | ||||
7520 | __kmp_global.g.g_dynamic_mode = dynamic_thread_limit; | ||||
7521 | } | ||||
7522 | #endif | ||||
7523 | |||||
7524 | if (__kmp_version) { | ||||
7525 | __kmp_print_version_2(); | ||||
7526 | } | ||||
7527 | |||||
7528 | /* we have finished parallel initialization */ | ||||
7529 | TCW_SYNC_4(__kmp_init_parallel, TRUE)(__kmp_init_parallel) = ((!0)); | ||||
7530 | |||||
7531 | KMP_MB(); | ||||
7532 | KA_TRACE(10, ("__kmp_parallel_initialize: exit\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_parallel_initialize: exit\n" ); }; | ||||
7533 | |||||
7534 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | ||||
7535 | } | ||||
7536 | |||||
7537 | void __kmp_hidden_helper_initialize() { | ||||
7538 | if (TCR_4(__kmp_init_hidden_helper)(__kmp_init_hidden_helper)) | ||||
7539 | return; | ||||
7540 | |||||
7541 | // __kmp_parallel_initialize is required before we initialize hidden helper | ||||
7542 | if (!TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) | ||||
7543 | __kmp_parallel_initialize(); | ||||
7544 | |||||
7545 | // Double check. Note that this double check should not be placed before | ||||
7546 | // __kmp_parallel_initialize as it will cause dead lock. | ||||
7547 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); | ||||
7548 | if (TCR_4(__kmp_init_hidden_helper)(__kmp_init_hidden_helper)) { | ||||
7549 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | ||||
7550 | return; | ||||
7551 | } | ||||
7552 | |||||
7553 | #if KMP_AFFINITY_SUPPORTED1 | ||||
7554 | // Initialize hidden helper affinity settings. | ||||
7555 | // The above __kmp_parallel_initialize() will initialize | ||||
7556 | // regular affinity (and topology) if not already done. | ||||
7557 | if (!__kmp_hh_affinity.flags.initialized) | ||||
7558 | __kmp_affinity_initialize(__kmp_hh_affinity); | ||||
7559 | #endif | ||||
7560 | |||||
7561 | // Set the count of hidden helper tasks to be executed to zero | ||||
7562 | KMP_ATOMIC_ST_REL(&__kmp_unexecuted_hidden_helper_tasks, 0)(&__kmp_unexecuted_hidden_helper_tasks)->store(0, std:: memory_order_release); | ||||
7563 | |||||
7564 | // Set the global variable indicating that we're initializing hidden helper | ||||
7565 | // team/threads | ||||
7566 | TCW_SYNC_4(__kmp_init_hidden_helper_threads, TRUE)(__kmp_init_hidden_helper_threads) = ((!0)); | ||||
7567 | |||||
7568 | // Platform independent initialization | ||||
7569 | __kmp_do_initialize_hidden_helper_threads(); | ||||
7570 | |||||
7571 | // Wait here for the finish of initialization of hidden helper teams | ||||
7572 | __kmp_hidden_helper_threads_initz_wait(); | ||||
7573 | |||||
7574 | // We have finished hidden helper initialization | ||||
7575 | TCW_SYNC_4(__kmp_init_hidden_helper, TRUE)(__kmp_init_hidden_helper) = ((!0)); | ||||
7576 | |||||
7577 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | ||||
7578 | } | ||||
7579 | |||||
7580 | /* ------------------------------------------------------------------------ */ | ||||
7581 | |||||
7582 | void __kmp_run_before_invoked_task(int gtid, int tid, kmp_info_t *this_thr, | ||||
7583 | kmp_team_t *team) { | ||||
7584 | kmp_disp_t *dispatch; | ||||
7585 | |||||
7586 | KMP_MB(); | ||||
7587 | |||||
7588 | /* none of the threads have encountered any constructs, yet. */ | ||||
7589 | this_thr->th.th_local.this_construct = 0; | ||||
7590 | #if KMP_CACHE_MANAGE | ||||
7591 | KMP_CACHE_PREFETCH(&this_thr->th.th_bar[bs_forkjoin_barrier].bb.b_arrived); | ||||
7592 | #endif /* KMP_CACHE_MANAGE */ | ||||
7593 | dispatch = (kmp_disp_t *)TCR_PTR(this_thr->th.th_dispatch)((void *)(this_thr->th.th_dispatch)); | ||||
7594 | KMP_DEBUG_ASSERT(dispatch)if (!(dispatch)) { __kmp_debug_assert("dispatch", "openmp/runtime/src/kmp_runtime.cpp" , 7594); }; | ||||
7595 | KMP_DEBUG_ASSERT(team->t.t_dispatch)if (!(team->t.t_dispatch)) { __kmp_debug_assert("team->t.t_dispatch" , "openmp/runtime/src/kmp_runtime.cpp", 7595); }; | ||||
7596 | // KMP_DEBUG_ASSERT( this_thr->th.th_dispatch == &team->t.t_dispatch[ | ||||
7597 | // this_thr->th.th_info.ds.ds_tid ] ); | ||||
7598 | |||||
7599 | dispatch->th_disp_index = 0; /* reset the dispatch buffer counter */ | ||||
7600 | dispatch->th_doacross_buf_idx = 0; // reset doacross dispatch buffer counter | ||||
7601 | if (__kmp_env_consistency_check) | ||||
7602 | __kmp_push_parallel(gtid, team->t.t_ident); | ||||
7603 | |||||
7604 | KMP_MB(); /* Flush all pending memory write invalidates. */ | ||||
7605 | } | ||||
7606 | |||||
7607 | void __kmp_run_after_invoked_task(int gtid, int tid, kmp_info_t *this_thr, | ||||
7608 | kmp_team_t *team) { | ||||
7609 | if (__kmp_env_consistency_check) | ||||
7610 | __kmp_pop_parallel(gtid, team->t.t_ident); | ||||
7611 | |||||
7612 | __kmp_finish_implicit_task(this_thr); | ||||
7613 | } | ||||
7614 | |||||
7615 | int __kmp_invoke_task_func(int gtid) { | ||||
7616 | int rc; | ||||
7617 | int tid = __kmp_tid_from_gtid(gtid); | ||||
7618 | kmp_info_t *this_thr = __kmp_threads[gtid]; | ||||
7619 | kmp_team_t *team = this_thr->th.th_team; | ||||
7620 | |||||
7621 | __kmp_run_before_invoked_task(gtid, tid, this_thr, team); | ||||
7622 | #if USE_ITT_BUILD1 | ||||
7623 | if (__itt_stack_caller_create_ptr__kmp_itt_stack_caller_create_ptr__3_0) { | ||||
7624 | // inform ittnotify about entering user's code | ||||
7625 | if (team->t.t_stack_id != NULL__null) { | ||||
7626 | __kmp_itt_stack_callee_enter((__itt_caller)team->t.t_stack_id); | ||||
7627 | } else { | ||||
7628 | KMP_DEBUG_ASSERT(team->t.t_parent->t.t_stack_id != NULL)if (!(team->t.t_parent->t.t_stack_id != __null)) { __kmp_debug_assert ("team->t.t_parent->t.t_stack_id != __null", "openmp/runtime/src/kmp_runtime.cpp" , 7628); }; | ||||
7629 | __kmp_itt_stack_callee_enter( | ||||
7630 | (__itt_caller)team->t.t_parent->t.t_stack_id); | ||||
7631 | } | ||||
7632 | } | ||||
7633 | #endif /* USE_ITT_BUILD */ | ||||
7634 | #if INCLUDE_SSC_MARKS(1 && 1) | ||||
7635 | SSC_MARK_INVOKING()__asm__ __volatile__("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(0xd695) : "%ebx"); | ||||
7636 | #endif | ||||
7637 | |||||
7638 | #if OMPT_SUPPORT1 | ||||
7639 | void *dummy; | ||||
7640 | void **exit_frame_p; | ||||
7641 | ompt_data_t *my_task_data; | ||||
7642 | ompt_data_t *my_parallel_data; | ||||
7643 | int ompt_team_size; | ||||
7644 | |||||
7645 | if (ompt_enabled.enabled) { | ||||
7646 | exit_frame_p = &(team->t.t_implicit_task_taskdata[tid] | ||||
7647 | .ompt_task_info.frame.exit_frame.ptr); | ||||
7648 | } else { | ||||
7649 | exit_frame_p = &dummy; | ||||
7650 | } | ||||
7651 | |||||
7652 | my_task_data = | ||||
7653 | &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data); | ||||
7654 | my_parallel_data = &(team->t.ompt_team_info.parallel_data); | ||||
7655 | if (ompt_enabled.ompt_callback_implicit_task) { | ||||
7656 | ompt_team_size = team->t.t_nproc; | ||||
7657 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | ||||
7658 | ompt_scope_begin, my_parallel_data, my_task_data, ompt_team_size, | ||||
7659 | __kmp_tid_from_gtid(gtid), ompt_task_implicit); | ||||
7660 | OMPT_CUR_TASK_INFO(this_thr)(&(this_thr->th.th_current_task->ompt_task_info))->thread_num = __kmp_tid_from_gtid(gtid); | ||||
7661 | } | ||||
7662 | #endif | ||||
7663 | |||||
7664 | #if KMP_STATS_ENABLED0 | ||||
7665 | stats_state_e previous_state = KMP_GET_THREAD_STATE()((void)0); | ||||
7666 | if (previous_state == stats_state_e::TEAMS_REGION) { | ||||
7667 | KMP_PUSH_PARTITIONED_TIMER(OMP_teams)((void)0); | ||||
7668 | } else { | ||||
7669 | KMP_PUSH_PARTITIONED_TIMER(OMP_parallel)((void)0); | ||||
7670 | } | ||||
7671 | KMP_SET_THREAD_STATE(IMPLICIT_TASK)((void)0); | ||||
7672 | #endif | ||||
7673 | |||||
7674 | rc = __kmp_invoke_microtask((microtask_t)TCR_SYNC_PTR(team->t.t_pkfn)((void *)(team->t.t_pkfn)), gtid, | ||||
7675 | tid, (int)team->t.t_argc, (void **)team->t.t_argv | ||||
7676 | #if OMPT_SUPPORT1 | ||||
7677 | , | ||||
7678 | exit_frame_p | ||||
7679 | #endif | ||||
7680 | ); | ||||
7681 | #if OMPT_SUPPORT1 | ||||
7682 | *exit_frame_p = NULL__null; | ||||
7683 | this_thr->th.ompt_thread_info.parallel_flags |= ompt_parallel_team; | ||||
7684 | #endif | ||||
7685 | |||||
7686 | #if KMP_STATS_ENABLED0 | ||||
7687 | if (previous_state == stats_state_e::TEAMS_REGION) { | ||||
7688 | KMP_SET_THREAD_STATE(previous_state)((void)0); | ||||
7689 | } | ||||
7690 | KMP_POP_PARTITIONED_TIMER()((void)0); | ||||
7691 | #endif | ||||
7692 | |||||
7693 | #if USE_ITT_BUILD1 | ||||
7694 | if (__itt_stack_caller_create_ptr__kmp_itt_stack_caller_create_ptr__3_0) { | ||||
7695 | // inform ittnotify about leaving user's code | ||||
7696 | if (team->t.t_stack_id != NULL__null) { | ||||
7697 | __kmp_itt_stack_callee_leave((__itt_caller)team->t.t_stack_id); | ||||
7698 | } else { | ||||
7699 | KMP_DEBUG_ASSERT(team->t.t_parent->t.t_stack_id != NULL)if (!(team->t.t_parent->t.t_stack_id != __null)) { __kmp_debug_assert ("team->t.t_parent->t.t_stack_id != __null", "openmp/runtime/src/kmp_runtime.cpp" , 7699); }; | ||||
7700 | __kmp_itt_stack_callee_leave( | ||||
7701 | (__itt_caller)team->t.t_parent->t.t_stack_id); | ||||
7702 | } | ||||
7703 | } | ||||
7704 | #endif /* USE_ITT_BUILD */ | ||||
7705 | __kmp_run_after_invoked_task(gtid, tid, this_thr, team); | ||||
7706 | |||||
7707 | return rc; | ||||
7708 | } | ||||
7709 | |||||
7710 | void __kmp_teams_master(int gtid) { | ||||
7711 | // This routine is called by all primary threads in teams construct | ||||
7712 | kmp_info_t *thr = __kmp_threads[gtid]; | ||||
7713 | kmp_team_t *team = thr->th.th_team; | ||||
7714 | ident_t *loc = team->t.t_ident; | ||||
7715 | thr->th.th_set_nproc = thr->th.th_teams_size.nth; | ||||
7716 | KMP_DEBUG_ASSERT(thr->th.th_teams_microtask)if (!(thr->th.th_teams_microtask)) { __kmp_debug_assert("thr->th.th_teams_microtask" , "openmp/runtime/src/kmp_runtime.cpp", 7716); }; | ||||
7717 | KMP_DEBUG_ASSERT(thr->th.th_set_nproc)if (!(thr->th.th_set_nproc)) { __kmp_debug_assert("thr->th.th_set_nproc" , "openmp/runtime/src/kmp_runtime.cpp", 7717); }; | ||||
7718 | KA_TRACE(20, ("__kmp_teams_master: T#%d, Tid %d, microtask %p\n", gtid,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_teams_master: T#%d, Tid %d, microtask %p\n" , gtid, __kmp_tid_from_gtid(gtid), thr->th.th_teams_microtask ); } | ||||
7719 | __kmp_tid_from_gtid(gtid), thr->th.th_teams_microtask))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_teams_master: T#%d, Tid %d, microtask %p\n" , gtid, __kmp_tid_from_gtid(gtid), thr->th.th_teams_microtask ); }; | ||||
7720 | |||||
7721 | // This thread is a new CG root. Set up the proper variables. | ||||
7722 | kmp_cg_root_t *tmp = (kmp_cg_root_t *)__kmp_allocate(sizeof(kmp_cg_root_t))___kmp_allocate((sizeof(kmp_cg_root_t)), "openmp/runtime/src/kmp_runtime.cpp" , 7722); | ||||
7723 | tmp->cg_root = thr; // Make thr the CG root | ||||
7724 | // Init to thread limit stored when league primary threads were forked | ||||
7725 | tmp->cg_thread_limit = thr->th.th_current_task->td_icvs.thread_limit; | ||||
7726 | tmp->cg_nthreads = 1; // Init counter to one active thread, this one | ||||
7727 | KA_TRACE(100, ("__kmp_teams_master: Thread %p created node %p and init"if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_teams_master: Thread %p created node %p and init" " cg_nthreads to 1\n", thr, tmp); } | ||||
7728 | " cg_nthreads to 1\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_teams_master: Thread %p created node %p and init" " cg_nthreads to 1\n", thr, tmp); } | ||||
7729 | thr, tmp))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_teams_master: Thread %p created node %p and init" " cg_nthreads to 1\n", thr, tmp); }; | ||||
7730 | tmp->up = thr->th.th_cg_roots; | ||||
7731 | thr->th.th_cg_roots = tmp; | ||||
7732 | |||||
7733 | // Launch league of teams now, but not let workers execute | ||||
7734 | // (they hang on fork barrier until next parallel) | ||||
7735 | #if INCLUDE_SSC_MARKS(1 && 1) | ||||
7736 | SSC_MARK_FORKING()__asm__ __volatile__("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(0xd693) : "%ebx"); | ||||
7737 | #endif | ||||
7738 | __kmp_fork_call(loc, gtid, fork_context_intel, team->t.t_argc, | ||||
7739 | (microtask_t)thr->th.th_teams_microtask, // "wrapped" task | ||||
7740 | VOLATILE_CAST(launch_t)(launch_t) __kmp_invoke_task_func, NULL__null); | ||||
7741 | #if INCLUDE_SSC_MARKS(1 && 1) | ||||
7742 | SSC_MARK_JOINING()__asm__ __volatile__("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(0xd694) : "%ebx"); | ||||
7743 | #endif | ||||
7744 | // If the team size was reduced from the limit, set it to the new size | ||||
7745 | if (thr->th.th_team_nproc < thr->th.th_teams_size.nth) | ||||
7746 | thr->th.th_teams_size.nth = thr->th.th_team_nproc; | ||||
7747 | // AC: last parameter "1" eliminates join barrier which won't work because | ||||
7748 | // worker threads are in a fork barrier waiting for more parallel regions | ||||
7749 | __kmp_join_call(loc, gtid | ||||
7750 | #if OMPT_SUPPORT1 | ||||
7751 | , | ||||
7752 | fork_context_intel | ||||
7753 | #endif | ||||
7754 | , | ||||
7755 | 1); | ||||
7756 | } | ||||
7757 | |||||
7758 | int __kmp_invoke_teams_master(int gtid) { | ||||
7759 | kmp_info_t *this_thr = __kmp_threads[gtid]; | ||||
7760 | kmp_team_t *team = this_thr->th.th_team; | ||||
7761 | #if KMP_DEBUG1 | ||||
7762 | if (!__kmp_threads[gtid]->th.th_team->t.t_serialized) | ||||
7763 | KMP_DEBUG_ASSERT((void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn ==if (!((void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn == (void *)__kmp_teams_master)) { __kmp_debug_assert("(void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn == (void *)__kmp_teams_master" , "openmp/runtime/src/kmp_runtime.cpp", 7764); } | ||||
7764 | (void *)__kmp_teams_master)if (!((void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn == (void *)__kmp_teams_master)) { __kmp_debug_assert("(void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn == (void *)__kmp_teams_master" , "openmp/runtime/src/kmp_runtime.cpp", 7764); }; | ||||
7765 | #endif | ||||
7766 | __kmp_run_before_invoked_task(gtid, 0, this_thr, team); | ||||
7767 | #if OMPT_SUPPORT1 | ||||
7768 | int tid = __kmp_tid_from_gtid(gtid); | ||||
7769 | ompt_data_t *task_data = | ||||
7770 | &team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data; | ||||
7771 | ompt_data_t *parallel_data = &team->t.ompt_team_info.parallel_data; | ||||
7772 | if (ompt_enabled.ompt_callback_implicit_task) { | ||||
7773 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | ||||
7774 | ompt_scope_begin, parallel_data, task_data, team->t.t_nproc, tid, | ||||
7775 | ompt_task_initial); | ||||
7776 | OMPT_CUR_TASK_INFO(this_thr)(&(this_thr->th.th_current_task->ompt_task_info))->thread_num = tid; | ||||
7777 | } | ||||
7778 | #endif | ||||
7779 | __kmp_teams_master(gtid); | ||||
7780 | #if OMPT_SUPPORT1 | ||||
7781 | this_thr->th.ompt_thread_info.parallel_flags |= ompt_parallel_league; | ||||
7782 | #endif | ||||
7783 | __kmp_run_after_invoked_task(gtid, 0, this_thr, team); | ||||
7784 | return 1; | ||||
7785 | } | ||||
7786 | |||||
7787 | /* this sets the requested number of threads for the next parallel region | ||||
7788 | encountered by this team. since this should be enclosed in the forkjoin | ||||
7789 | critical section it should avoid race conditions with asymmetrical nested | ||||
7790 | parallelism */ | ||||
7791 | |||||
7792 | void __kmp_push_num_threads(ident_t *id, int gtid, int num_threads) { | ||||
7793 | kmp_info_t *thr = __kmp_threads[gtid]; | ||||
7794 | |||||
7795 | if (num_threads > 0) | ||||
7796 | thr->th.th_set_nproc = num_threads; | ||||
7797 | } | ||||
7798 | |||||
7799 | static void __kmp_push_thread_limit(kmp_info_t *thr, int num_teams, | ||||
7800 | int num_threads) { | ||||
7801 | KMP_DEBUG_ASSERT(thr)if (!(thr)) { __kmp_debug_assert("thr", "openmp/runtime/src/kmp_runtime.cpp" , 7801); }; | ||||
7802 | // Remember the number of threads for inner parallel regions | ||||
7803 | if (!TCR_4(__kmp_init_middle)(__kmp_init_middle)) | ||||
7804 | __kmp_middle_initialize(); // get internal globals calculated | ||||
7805 | __kmp_assign_root_init_mask(); | ||||
7806 | KMP_DEBUG_ASSERT(__kmp_avail_proc)if (!(__kmp_avail_proc)) { __kmp_debug_assert("__kmp_avail_proc" , "openmp/runtime/src/kmp_runtime.cpp", 7806); }; | ||||
7807 | KMP_DEBUG_ASSERT(__kmp_dflt_team_nth)if (!(__kmp_dflt_team_nth)) { __kmp_debug_assert("__kmp_dflt_team_nth" , "openmp/runtime/src/kmp_runtime.cpp", 7807); }; | ||||
7808 | |||||
7809 | if (num_threads == 0) { | ||||
7810 | if (__kmp_teams_thread_limit > 0) { | ||||
7811 | num_threads = __kmp_teams_thread_limit; | ||||
7812 | } else { | ||||
7813 | num_threads = __kmp_avail_proc / num_teams; | ||||
7814 | } | ||||
7815 | // adjust num_threads w/o warning as it is not user setting | ||||
7816 | // num_threads = min(num_threads, nthreads-var, thread-limit-var) | ||||
7817 | // no thread_limit clause specified - do not change thread-limit-var ICV | ||||
7818 | if (num_threads > __kmp_dflt_team_nth) { | ||||
7819 | num_threads = __kmp_dflt_team_nth; // honor nthreads-var ICV | ||||
7820 | } | ||||
7821 | if (num_threads > thr->th.th_current_task->td_icvs.thread_limit) { | ||||
7822 | num_threads = thr->th.th_current_task->td_icvs.thread_limit; | ||||
7823 | } // prevent team size to exceed thread-limit-var | ||||
7824 | if (num_teams * num_threads > __kmp_teams_max_nth) { | ||||
7825 | num_threads = __kmp_teams_max_nth / num_teams; | ||||
7826 | } | ||||
7827 | if (num_threads == 0) { | ||||
7828 | num_threads = 1; | ||||
7829 | } | ||||
7830 | } else { | ||||
7831 | if (num_threads < 0) { | ||||
7832 | __kmp_msg(kmp_ms_warning, KMP_MSG(CantFormThrTeam, num_threads, 1)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, num_threads, 1 ), | ||||
7833 | __kmp_msg_null); | ||||
7834 | num_threads = 1; | ||||
7835 | } | ||||
7836 | // This thread will be the primary thread of the league primary threads | ||||
7837 | // Store new thread limit; old limit is saved in th_cg_roots list | ||||
7838 | thr->th.th_current_task->td_icvs.thread_limit = num_threads; | ||||
7839 | // num_threads = min(num_threads, nthreads-var) | ||||
7840 | if (num_threads > __kmp_dflt_team_nth) { | ||||
7841 | num_threads = __kmp_dflt_team_nth; // honor nthreads-var ICV | ||||
7842 | } | ||||
7843 | if (num_teams * num_threads > __kmp_teams_max_nth) { | ||||
7844 | int new_threads = __kmp_teams_max_nth / num_teams; | ||||
7845 | if (new_threads == 0) { | ||||
7846 | new_threads = 1; | ||||
7847 | } | ||||
7848 | if (new_threads != num_threads) { | ||||
7849 | if (!__kmp_reserve_warn) { // user asked for too many threads | ||||
7850 | __kmp_reserve_warn = 1; // conflicts with KMP_TEAMS_THREAD_LIMIT | ||||
7851 | __kmp_msg(kmp_ms_warning, | ||||
7852 | KMP_MSG(CantFormThrTeam, num_threads, new_threads)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, num_threads, new_threads ), | ||||
7853 | KMP_HNT(Unset_ALL_THREADS)__kmp_msg_format(kmp_i18n_hnt_Unset_ALL_THREADS), __kmp_msg_null); | ||||
7854 | } | ||||
7855 | } | ||||
7856 | num_threads = new_threads; | ||||
7857 | } | ||||
7858 | } | ||||
7859 | thr->th.th_teams_size.nth = num_threads; | ||||
7860 | } | ||||
7861 | |||||
7862 | /* this sets the requested number of teams for the teams region and/or | ||||
7863 | the number of threads for the next parallel region encountered */ | ||||
7864 | void __kmp_push_num_teams(ident_t *id, int gtid, int num_teams, | ||||
7865 | int num_threads) { | ||||
7866 | kmp_info_t *thr = __kmp_threads[gtid]; | ||||
7867 | if (num_teams < 0) { | ||||
7868 | // OpenMP specification requires requested values to be positive, | ||||
7869 | // but people can send us any value, so we'd better check | ||||
7870 | __kmp_msg(kmp_ms_warning, KMP_MSG(NumTeamsNotPositive, num_teams, 1)__kmp_msg_format(kmp_i18n_msg_NumTeamsNotPositive, num_teams, 1), | ||||
7871 | __kmp_msg_null); | ||||
7872 | num_teams = 1; | ||||
7873 | } | ||||
7874 | if (num_teams == 0) { | ||||
7875 | if (__kmp_nteams > 0) { | ||||
7876 | num_teams = __kmp_nteams; | ||||
7877 | } else { | ||||
7878 | num_teams = 1; // default number of teams is 1. | ||||
7879 | } | ||||
7880 | } | ||||
7881 | if (num_teams > __kmp_teams_max_nth) { // if too many teams requested? | ||||
7882 | if (!__kmp_reserve_warn) { | ||||
7883 | __kmp_reserve_warn = 1; | ||||
7884 | __kmp_msg(kmp_ms_warning, | ||||
7885 | KMP_MSG(CantFormThrTeam, num_teams, __kmp_teams_max_nth)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, num_teams, __kmp_teams_max_nth ), | ||||
7886 | KMP_HNT(Unset_ALL_THREADS)__kmp_msg_format(kmp_i18n_hnt_Unset_ALL_THREADS), __kmp_msg_null); | ||||
7887 | } | ||||
7888 | num_teams = __kmp_teams_max_nth; | ||||
7889 | } | ||||
7890 | // Set number of teams (number of threads in the outer "parallel" of the | ||||
7891 | // teams) | ||||
7892 | thr->th.th_set_nproc = thr->th.th_teams_size.nteams = num_teams; | ||||
7893 | |||||
7894 | __kmp_push_thread_limit(thr, num_teams, num_threads); | ||||
7895 | } | ||||
7896 | |||||
7897 | /* This sets the requested number of teams for the teams region and/or | ||||
7898 | the number of threads for the next parallel region encountered */ | ||||
7899 | void __kmp_push_num_teams_51(ident_t *id, int gtid, int num_teams_lb, | ||||
7900 | int num_teams_ub, int num_threads) { | ||||
7901 | kmp_info_t *thr = __kmp_threads[gtid]; | ||||
7902 | KMP_DEBUG_ASSERT(num_teams_lb >= 0 && num_teams_ub >= 0)if (!(num_teams_lb >= 0 && num_teams_ub >= 0)) { __kmp_debug_assert("num_teams_lb >= 0 && num_teams_ub >= 0" , "openmp/runtime/src/kmp_runtime.cpp", 7902); }; | ||||
7903 | KMP_DEBUG_ASSERT(num_teams_ub >= num_teams_lb)if (!(num_teams_ub >= num_teams_lb)) { __kmp_debug_assert( "num_teams_ub >= num_teams_lb", "openmp/runtime/src/kmp_runtime.cpp" , 7903); }; | ||||
7904 | KMP_DEBUG_ASSERT(num_threads >= 0)if (!(num_threads >= 0)) { __kmp_debug_assert("num_threads >= 0" , "openmp/runtime/src/kmp_runtime.cpp", 7904); }; | ||||
7905 | |||||
7906 | if (num_teams_lb > num_teams_ub) { | ||||
7907 | __kmp_fatal(KMP_MSG(FailedToCreateTeam, num_teams_lb, num_teams_ub)__kmp_msg_format(kmp_i18n_msg_FailedToCreateTeam, num_teams_lb , num_teams_ub), | ||||
7908 | KMP_HNT(SetNewBound, __kmp_teams_max_nth)__kmp_msg_format(kmp_i18n_hnt_SetNewBound, __kmp_teams_max_nth ), __kmp_msg_null); | ||||
7909 | } | ||||
7910 | |||||
7911 | int num_teams = 1; // defalt number of teams is 1. | ||||
7912 | |||||
7913 | if (num_teams_lb == 0 && num_teams_ub > 0) | ||||
7914 | num_teams_lb = num_teams_ub; | ||||
7915 | |||||
7916 | if (num_teams_lb == 0 && num_teams_ub == 0) { // no num_teams clause | ||||
7917 | num_teams = (__kmp_nteams > 0) ? __kmp_nteams : num_teams; | ||||
7918 | if (num_teams > __kmp_teams_max_nth) { | ||||
7919 | if (!__kmp_reserve_warn) { | ||||
7920 | __kmp_reserve_warn = 1; | ||||
7921 | __kmp_msg(kmp_ms_warning, | ||||
7922 | KMP_MSG(CantFormThrTeam, num_teams, __kmp_teams_max_nth)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, num_teams, __kmp_teams_max_nth ), | ||||
7923 | KMP_HNT(Unset_ALL_THREADS)__kmp_msg_format(kmp_i18n_hnt_Unset_ALL_THREADS), __kmp_msg_null); | ||||
7924 | } | ||||
7925 | num_teams = __kmp_teams_max_nth; | ||||
7926 | } | ||||
7927 | } else if (num_teams_lb == num_teams_ub) { // requires exact number of teams | ||||
7928 | num_teams = num_teams_ub; | ||||
7929 | } else { // num_teams_lb <= num_teams <= num_teams_ub | ||||
7930 | if (num_threads <= 0) { | ||||
7931 | if (num_teams_ub > __kmp_teams_max_nth) { | ||||
7932 | num_teams = num_teams_lb; | ||||
7933 | } else { | ||||
7934 | num_teams = num_teams_ub; | ||||
7935 | } | ||||
7936 | } else { | ||||
7937 | num_teams = (num_threads > __kmp_teams_max_nth) | ||||
7938 | ? num_teams | ||||
7939 | : __kmp_teams_max_nth / num_threads; | ||||
7940 | if (num_teams < num_teams_lb) { | ||||
7941 | num_teams = num_teams_lb; | ||||
7942 | } else if (num_teams > num_teams_ub) { | ||||
7943 | num_teams = num_teams_ub; | ||||
7944 | } | ||||
7945 | } | ||||
7946 | } | ||||
7947 | // Set number of teams (number of threads in the outer "parallel" of the | ||||
7948 | // teams) | ||||
7949 | thr->th.th_set_nproc = thr->th.th_teams_size.nteams = num_teams; | ||||
7950 | |||||
7951 | __kmp_push_thread_limit(thr, num_teams, num_threads); | ||||
7952 | } | ||||
7953 | |||||
7954 | // Set the proc_bind var to use in the following parallel region. | ||||
7955 | void __kmp_push_proc_bind(ident_t *id, int gtid, kmp_proc_bind_t proc_bind) { | ||||
7956 | kmp_info_t *thr = __kmp_threads[gtid]; | ||||
7957 | thr->th.th_set_proc_bind = proc_bind; | ||||
7958 | } | ||||
7959 | |||||
7960 | /* Launch the worker threads into the microtask. */ | ||||
7961 | |||||
7962 | void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team) { | ||||
7963 | kmp_info_t *this_thr = __kmp_threads[gtid]; | ||||
7964 | |||||
7965 | #ifdef KMP_DEBUG1 | ||||
7966 | int f; | ||||
7967 | #endif /* KMP_DEBUG */ | ||||
7968 | |||||
7969 | KMP_DEBUG_ASSERT(team)if (!(team)) { __kmp_debug_assert("team", "openmp/runtime/src/kmp_runtime.cpp" , 7969); }; | ||||
7970 | KMP_DEBUG_ASSERT(this_thr->th.th_team == team)if (!(this_thr->th.th_team == team)) { __kmp_debug_assert( "this_thr->th.th_team == team", "openmp/runtime/src/kmp_runtime.cpp" , 7970); }; | ||||
7971 | KMP_ASSERT(KMP_MASTER_GTID(gtid))if (!((0 == __kmp_tid_from_gtid((gtid))))) { __kmp_debug_assert ("KMP_MASTER_GTID(gtid)", "openmp/runtime/src/kmp_runtime.cpp" , 7971); }; | ||||
7972 | KMP_MB(); /* Flush all pending memory write invalidates. */ | ||||
7973 | |||||
7974 | team->t.t_construct = 0; /* no single directives seen yet */ | ||||
7975 | team->t.t_ordered.dt.t_value = | ||||
7976 | 0; /* thread 0 enters the ordered section first */ | ||||
7977 | |||||
7978 | /* Reset the identifiers on the dispatch buffer */ | ||||
7979 | KMP_DEBUG_ASSERT(team->t.t_disp_buffer)if (!(team->t.t_disp_buffer)) { __kmp_debug_assert("team->t.t_disp_buffer" , "openmp/runtime/src/kmp_runtime.cpp", 7979); }; | ||||
7980 | if (team->t.t_max_nproc > 1) { | ||||
7981 | int i; | ||||
7982 | for (i = 0; i < __kmp_dispatch_num_buffers; ++i) { | ||||
7983 | team->t.t_disp_buffer[i].buffer_index = i; | ||||
7984 | team->t.t_disp_buffer[i].doacross_buf_idx = i; | ||||
7985 | } | ||||
7986 | } else { | ||||
7987 | team->t.t_disp_buffer[0].buffer_index = 0; | ||||
7988 | team->t.t_disp_buffer[0].doacross_buf_idx = 0; | ||||
7989 | } | ||||
7990 | |||||
7991 | KMP_MB(); /* Flush all pending memory write invalidates. */ | ||||
7992 | KMP_ASSERT(this_thr->th.th_team == team)if (!(this_thr->th.th_team == team)) { __kmp_debug_assert( "this_thr->th.th_team == team", "openmp/runtime/src/kmp_runtime.cpp" , 7992); }; | ||||
7993 | |||||
7994 | #ifdef KMP_DEBUG1 | ||||
7995 | for (f = 0; f < team->t.t_nproc; f++) { | ||||
7996 | KMP_DEBUG_ASSERT(team->t.t_threads[f] &&if (!(team->t.t_threads[f] && team->t.t_threads [f]->th.th_team_nproc == team->t.t_nproc)) { __kmp_debug_assert ("team->t.t_threads[f] && team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc" , "openmp/runtime/src/kmp_runtime.cpp", 7997); } | ||||
7997 | team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc)if (!(team->t.t_threads[f] && team->t.t_threads [f]->th.th_team_nproc == team->t.t_nproc)) { __kmp_debug_assert ("team->t.t_threads[f] && team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc" , "openmp/runtime/src/kmp_runtime.cpp", 7997); }; | ||||
7998 | } | ||||
7999 | #endif /* KMP_DEBUG */ | ||||
8000 | |||||
8001 | /* release the worker threads so they may begin working */ | ||||
8002 | __kmp_fork_barrier(gtid, 0); | ||||
8003 | } | ||||
8004 | |||||
8005 | void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team) { | ||||
8006 | kmp_info_t *this_thr = __kmp_threads[gtid]; | ||||
8007 | |||||
8008 | KMP_DEBUG_ASSERT(team)if (!(team)) { __kmp_debug_assert("team", "openmp/runtime/src/kmp_runtime.cpp" , 8008); }; | ||||
8009 | KMP_DEBUG_ASSERT(this_thr->th.th_team == team)if (!(this_thr->th.th_team == team)) { __kmp_debug_assert( "this_thr->th.th_team == team", "openmp/runtime/src/kmp_runtime.cpp" , 8009); }; | ||||
8010 | KMP_ASSERT(KMP_MASTER_GTID(gtid))if (!((0 == __kmp_tid_from_gtid((gtid))))) { __kmp_debug_assert ("KMP_MASTER_GTID(gtid)", "openmp/runtime/src/kmp_runtime.cpp" , 8010); }; | ||||
8011 | KMP_MB(); /* Flush all pending memory write invalidates. */ | ||||
8012 | |||||
8013 | /* Join barrier after fork */ | ||||
8014 | |||||
8015 | #ifdef KMP_DEBUG1 | ||||
8016 | if (__kmp_threads[gtid] && | ||||
8017 | __kmp_threads[gtid]->th.th_team_nproc != team->t.t_nproc) { | ||||
8018 | __kmp_printf("GTID: %d, __kmp_threads[%d]=%p\n", gtid, gtid, | ||||
8019 | __kmp_threads[gtid]); | ||||
8020 | __kmp_printf("__kmp_threads[%d]->th.th_team_nproc=%d, TEAM: %p, " | ||||
8021 | "team->t.t_nproc=%d\n", | ||||
8022 | gtid, __kmp_threads[gtid]->th.th_team_nproc, team, | ||||
8023 | team->t.t_nproc); | ||||
8024 | __kmp_print_structure(); | ||||
8025 | } | ||||
8026 | KMP_DEBUG_ASSERT(__kmp_threads[gtid] &&if (!(__kmp_threads[gtid] && __kmp_threads[gtid]-> th.th_team_nproc == team->t.t_nproc)) { __kmp_debug_assert ("__kmp_threads[gtid] && __kmp_threads[gtid]->th.th_team_nproc == team->t.t_nproc" , "openmp/runtime/src/kmp_runtime.cpp", 8027); } | ||||
8027 | __kmp_threads[gtid]->th.th_team_nproc == team->t.t_nproc)if (!(__kmp_threads[gtid] && __kmp_threads[gtid]-> th.th_team_nproc == team->t.t_nproc)) { __kmp_debug_assert ("__kmp_threads[gtid] && __kmp_threads[gtid]->th.th_team_nproc == team->t.t_nproc" , "openmp/runtime/src/kmp_runtime.cpp", 8027); }; | ||||
8028 | #endif /* KMP_DEBUG */ | ||||
8029 | |||||
8030 | __kmp_join_barrier(gtid); /* wait for everyone */ | ||||
8031 | #if OMPT_SUPPORT1 | ||||
8032 | if (ompt_enabled.enabled && | ||||
8033 | this_thr->th.ompt_thread_info.state == ompt_state_wait_barrier_implicit) { | ||||
8034 | int ds_tid = this_thr->th.th_info.ds.ds_tid; | ||||
8035 | ompt_data_t *task_data = OMPT_CUR_TASK_DATA(this_thr)(&(this_thr->th.th_current_task->ompt_task_info.task_data )); | ||||
8036 | this_thr->th.ompt_thread_info.state = ompt_state_overhead; | ||||
8037 | #if OMPT_OPTIONAL1 | ||||
8038 | void *codeptr = NULL__null; | ||||
8039 | if (KMP_MASTER_TID(ds_tid)(0 == (ds_tid)) && | ||||
8040 | (ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)ompt_callback_sync_region_wait_callback || | ||||
8041 | ompt_callbacks.ompt_callback(ompt_callback_sync_region)ompt_callback_sync_region_callback)) | ||||
8042 | codeptr = OMPT_CUR_TEAM_INFO(this_thr)(&(this_thr->th.th_team->t.ompt_team_info))->master_return_address; | ||||
8043 | |||||
8044 | if (ompt_enabled.ompt_callback_sync_region_wait) { | ||||
8045 | ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)ompt_callback_sync_region_wait_callback( | ||||
8046 | ompt_sync_region_barrier_implicit, ompt_scope_end, NULL__null, task_data, | ||||
8047 | codeptr); | ||||
8048 | } | ||||
8049 | if (ompt_enabled.ompt_callback_sync_region) { | ||||
8050 | ompt_callbacks.ompt_callback(ompt_callback_sync_region)ompt_callback_sync_region_callback( | ||||
8051 | ompt_sync_region_barrier_implicit, ompt_scope_end, NULL__null, task_data, | ||||
8052 | codeptr); | ||||
8053 | } | ||||
8054 | #endif | ||||
8055 | if (!KMP_MASTER_TID(ds_tid)(0 == (ds_tid)) && ompt_enabled.ompt_callback_implicit_task) { | ||||
8056 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | ||||
8057 | ompt_scope_end, NULL__null, task_data, 0, ds_tid, | ||||
8058 | ompt_task_implicit); // TODO: Can this be ompt_task_initial? | ||||
8059 | } | ||||
8060 | } | ||||
8061 | #endif | ||||
8062 | |||||
8063 | KMP_MB(); /* Flush all pending memory write invalidates. */ | ||||
8064 | KMP_ASSERT(this_thr->th.th_team == team)if (!(this_thr->th.th_team == team)) { __kmp_debug_assert( "this_thr->th.th_team == team", "openmp/runtime/src/kmp_runtime.cpp" , 8064); }; | ||||
8065 | } | ||||
8066 | |||||
8067 | /* ------------------------------------------------------------------------ */ | ||||
8068 | |||||
8069 | #ifdef USE_LOAD_BALANCE1 | ||||
8070 | |||||
8071 | // Return the worker threads actively spinning in the hot team, if we | ||||
8072 | // are at the outermost level of parallelism. Otherwise, return 0. | ||||
8073 | static int __kmp_active_hot_team_nproc(kmp_root_t *root) { | ||||
8074 | int i; | ||||
8075 | int retval; | ||||
8076 | kmp_team_t *hot_team; | ||||
8077 | |||||
8078 | if (root->r.r_active) { | ||||
8079 | return 0; | ||||
8080 | } | ||||
8081 | hot_team = root->r.r_hot_team; | ||||
8082 | if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME(2147483647)) { | ||||
8083 | return hot_team->t.t_nproc - 1; // Don't count primary thread | ||||
8084 | } | ||||
8085 | |||||
8086 | // Skip the primary thread - it is accounted for elsewhere. | ||||
8087 | retval = 0; | ||||
8088 | for (i = 1; i < hot_team->t.t_nproc; i++) { | ||||
8089 | if (hot_team->t.t_threads[i]->th.th_active) { | ||||
8090 | retval++; | ||||
8091 | } | ||||
8092 | } | ||||
8093 | return retval; | ||||
8094 | } | ||||
8095 | |||||
8096 | // Perform an automatic adjustment to the number of | ||||
8097 | // threads used by the next parallel region. | ||||
8098 | static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc) { | ||||
8099 | int retval; | ||||
8100 | int pool_active; | ||||
8101 | int hot_team_active; | ||||
8102 | int team_curr_active; | ||||
8103 | int system_active; | ||||
8104 | |||||
8105 | KB_TRACE(20, ("__kmp_load_balance_nproc: called root:%p set_nproc:%d\n", root,if (kmp_b_debug >= 20) { __kmp_debug_printf ("__kmp_load_balance_nproc: called root:%p set_nproc:%d\n" , root, set_nproc); } | ||||
8106 | set_nproc))if (kmp_b_debug >= 20) { __kmp_debug_printf ("__kmp_load_balance_nproc: called root:%p set_nproc:%d\n" , root, set_nproc); }; | ||||
8107 | KMP_DEBUG_ASSERT(root)if (!(root)) { __kmp_debug_assert("root", "openmp/runtime/src/kmp_runtime.cpp" , 8107); }; | ||||
8108 | KMP_DEBUG_ASSERT(root->r.r_root_team->t.t_threads[0]if (!(root->r.r_root_team->t.t_threads[0] ->th.th_current_task ->td_icvs.dynamic == (!0))) { __kmp_debug_assert("root->r.r_root_team->t.t_threads[0] ->th.th_current_task->td_icvs.dynamic == (!0)" , "openmp/runtime/src/kmp_runtime.cpp", 8109); } | ||||
8109 | ->th.th_current_task->td_icvs.dynamic == TRUE)if (!(root->r.r_root_team->t.t_threads[0] ->th.th_current_task ->td_icvs.dynamic == (!0))) { __kmp_debug_assert("root->r.r_root_team->t.t_threads[0] ->th.th_current_task->td_icvs.dynamic == (!0)" , "openmp/runtime/src/kmp_runtime.cpp", 8109); }; | ||||
8110 | KMP_DEBUG_ASSERT(set_nproc > 1)if (!(set_nproc > 1)) { __kmp_debug_assert("set_nproc > 1" , "openmp/runtime/src/kmp_runtime.cpp", 8110); }; | ||||
8111 | |||||
8112 | if (set_nproc == 1) { | ||||
8113 | KB_TRACE(20, ("__kmp_load_balance_nproc: serial execution.\n"))if (kmp_b_debug >= 20) { __kmp_debug_printf ("__kmp_load_balance_nproc: serial execution.\n" ); }; | ||||
8114 | return 1; | ||||
8115 | } | ||||
8116 | |||||
8117 | // Threads that are active in the thread pool, active in the hot team for this | ||||
8118 | // particular root (if we are at the outer par level), and the currently | ||||
8119 | // executing thread (to become the primary thread) are available to add to the | ||||
8120 | // new team, but are currently contributing to the system load, and must be | ||||
8121 | // accounted for. | ||||
8122 | pool_active = __kmp_thread_pool_active_nth; | ||||
8123 | hot_team_active = __kmp_active_hot_team_nproc(root); | ||||
8124 | team_curr_active = pool_active + hot_team_active + 1; | ||||
8125 | |||||
8126 | // Check the system load. | ||||
8127 | system_active = __kmp_get_load_balance(__kmp_avail_proc + team_curr_active); | ||||
8128 | KB_TRACE(30, ("__kmp_load_balance_nproc: system active = %d pool active = %d "if (kmp_b_debug >= 30) { __kmp_debug_printf ("__kmp_load_balance_nproc: system active = %d pool active = %d " "hot team active = %d\n", system_active, pool_active, hot_team_active ); } | ||||
8129 | "hot team active = %d\n",if (kmp_b_debug >= 30) { __kmp_debug_printf ("__kmp_load_balance_nproc: system active = %d pool active = %d " "hot team active = %d\n", system_active, pool_active, hot_team_active ); } | ||||
8130 | system_active, pool_active, hot_team_active))if (kmp_b_debug >= 30) { __kmp_debug_printf ("__kmp_load_balance_nproc: system active = %d pool active = %d " "hot team active = %d\n", system_active, pool_active, hot_team_active ); }; | ||||
8131 | |||||
8132 | if (system_active < 0) { | ||||
8133 | // There was an error reading the necessary info from /proc, so use the | ||||
8134 | // thread limit algorithm instead. Once we set __kmp_global.g.g_dynamic_mode | ||||
8135 | // = dynamic_thread_limit, we shouldn't wind up getting back here. | ||||
8136 | __kmp_global.g.g_dynamic_mode = dynamic_thread_limit; | ||||
8137 | KMP_WARNING(CantLoadBalUsing, "KMP_DYNAMIC_MODE=thread limit")__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_CantLoadBalUsing , "KMP_DYNAMIC_MODE=thread limit"), __kmp_msg_null); | ||||
8138 | |||||
8139 | // Make this call behave like the thread limit algorithm. | ||||
8140 | retval = __kmp_avail_proc - __kmp_nth + | ||||
8141 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc); | ||||
8142 | if (retval > set_nproc) { | ||||
8143 | retval = set_nproc; | ||||
8144 | } | ||||
8145 | if (retval < KMP_MIN_NTH1) { | ||||
8146 | retval = KMP_MIN_NTH1; | ||||
8147 | } | ||||
8148 | |||||
8149 | KB_TRACE(20, ("__kmp_load_balance_nproc: thread limit exit. retval:%d\n",if (kmp_b_debug >= 20) { __kmp_debug_printf ("__kmp_load_balance_nproc: thread limit exit. retval:%d\n" , retval); } | ||||
8150 | retval))if (kmp_b_debug >= 20) { __kmp_debug_printf ("__kmp_load_balance_nproc: thread limit exit. retval:%d\n" , retval); }; | ||||
8151 | return retval; | ||||
8152 | } | ||||
8153 | |||||
8154 | // There is a slight delay in the load balance algorithm in detecting new | ||||
8155 | // running procs. The real system load at this instant should be at least as | ||||
8156 | // large as the #active omp thread that are available to add to the team. | ||||
8157 | if (system_active < team_curr_active) { | ||||
8158 | system_active = team_curr_active; | ||||
8159 | } | ||||
8160 | retval = __kmp_avail_proc - system_active + team_curr_active; | ||||
8161 | if (retval > set_nproc) { | ||||
8162 | retval = set_nproc; | ||||
8163 | } | ||||
8164 | if (retval < KMP_MIN_NTH1) { | ||||
8165 | retval = KMP_MIN_NTH1; | ||||
8166 | } | ||||
8167 | |||||
8168 | KB_TRACE(20, ("__kmp_load_balance_nproc: exit. retval:%d\n", retval))if (kmp_b_debug >= 20) { __kmp_debug_printf ("__kmp_load_balance_nproc: exit. retval:%d\n" , retval); }; | ||||
8169 | return retval; | ||||
8170 | } // __kmp_load_balance_nproc() | ||||
8171 | |||||
8172 | #endif /* USE_LOAD_BALANCE */ | ||||
8173 | |||||
8174 | /* ------------------------------------------------------------------------ */ | ||||
8175 | |||||
8176 | /* NOTE: this is called with the __kmp_init_lock held */ | ||||
8177 | void __kmp_cleanup(void) { | ||||
8178 | int f; | ||||
8179 | |||||
8180 | KA_TRACE(10, ("__kmp_cleanup: enter\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_cleanup: enter\n" ); }; | ||||
8181 | |||||
8182 | if (TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) { | ||||
8183 | #if KMP_HANDLE_SIGNALS(1 || 0) | ||||
8184 | __kmp_remove_signals(); | ||||
8185 | #endif | ||||
8186 | TCW_4(__kmp_init_parallel, FALSE)(__kmp_init_parallel) = (0); | ||||
8187 | } | ||||
8188 | |||||
8189 | if (TCR_4(__kmp_init_middle)(__kmp_init_middle)) { | ||||
8190 | #if KMP_AFFINITY_SUPPORTED1 | ||||
8191 | __kmp_affinity_uninitialize(); | ||||
8192 | #endif /* KMP_AFFINITY_SUPPORTED */ | ||||
8193 | __kmp_cleanup_hierarchy(); | ||||
8194 | TCW_4(__kmp_init_middle, FALSE)(__kmp_init_middle) = (0); | ||||
8195 | } | ||||
8196 | |||||
8197 | KA_TRACE(10, ("__kmp_cleanup: go serial cleanup\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_cleanup: go serial cleanup\n" ); }; | ||||
8198 | |||||
8199 | if (__kmp_init_serial) { | ||||
8200 | __kmp_runtime_destroy(); | ||||
8201 | __kmp_init_serial = FALSE0; | ||||
8202 | } | ||||
8203 | |||||
8204 | __kmp_cleanup_threadprivate_caches(); | ||||
8205 | |||||
8206 | for (f = 0; f < __kmp_threads_capacity; f++) { | ||||
8207 | if (__kmp_root[f] != NULL__null) { | ||||
8208 | __kmp_free(__kmp_root[f])___kmp_free((__kmp_root[f]), "openmp/runtime/src/kmp_runtime.cpp" , 8208); | ||||
8209 | __kmp_root[f] = NULL__null; | ||||
8210 | } | ||||
8211 | } | ||||
8212 | __kmp_free(__kmp_threads)___kmp_free((__kmp_threads), "openmp/runtime/src/kmp_runtime.cpp" , 8212); | ||||
8213 | // __kmp_threads and __kmp_root were allocated at once, as single block, so | ||||
8214 | // there is no need in freeing __kmp_root. | ||||
8215 | __kmp_threads = NULL__null; | ||||
8216 | __kmp_root = NULL__null; | ||||
8217 | __kmp_threads_capacity = 0; | ||||
8218 | |||||
8219 | // Free old __kmp_threads arrays if they exist. | ||||
8220 | kmp_old_threads_list_t *ptr = __kmp_old_threads_list; | ||||
8221 | while (ptr) { | ||||
8222 | kmp_old_threads_list_t *next = ptr->next; | ||||
8223 | __kmp_free(ptr->threads)___kmp_free((ptr->threads), "openmp/runtime/src/kmp_runtime.cpp" , 8223); | ||||
8224 | __kmp_free(ptr)___kmp_free((ptr), "openmp/runtime/src/kmp_runtime.cpp", 8224 ); | ||||
8225 | ptr = next; | ||||
8226 | } | ||||
8227 | |||||
8228 | #if KMP_USE_DYNAMIC_LOCK1 | ||||
8229 | __kmp_cleanup_indirect_user_locks(); | ||||
8230 | #else | ||||
8231 | __kmp_cleanup_user_locks(); | ||||
8232 | #endif | ||||
8233 | #if OMPD_SUPPORT1 | ||||
8234 | if (ompd_state) { | ||||
8235 | __kmp_free(ompd_env_block)___kmp_free((ompd_env_block), "openmp/runtime/src/kmp_runtime.cpp" , 8235); | ||||
8236 | ompd_env_block = NULL__null; | ||||
8237 | ompd_env_block_size = 0; | ||||
8238 | } | ||||
8239 | #endif | ||||
8240 | |||||
8241 | #if KMP_AFFINITY_SUPPORTED1 | ||||
8242 | KMP_INTERNAL_FREE(CCAST(char *, __kmp_cpuinfo_file))free(const_cast<char *>(__kmp_cpuinfo_file)); | ||||
8243 | __kmp_cpuinfo_file = NULL__null; | ||||
8244 | #endif /* KMP_AFFINITY_SUPPORTED */ | ||||
8245 | |||||
8246 | #if KMP_USE_ADAPTIVE_LOCKS(0 || 1) && !0 | ||||
8247 | #if KMP_DEBUG_ADAPTIVE_LOCKS0 | ||||
8248 | __kmp_print_speculative_stats(); | ||||
8249 | #endif | ||||
8250 | #endif | ||||
8251 | KMP_INTERNAL_FREE(__kmp_nested_nth.nth)free(__kmp_nested_nth.nth); | ||||
8252 | __kmp_nested_nth.nth = NULL__null; | ||||
8253 | __kmp_nested_nth.size = 0; | ||||
8254 | __kmp_nested_nth.used = 0; | ||||
8255 | KMP_INTERNAL_FREE(__kmp_nested_proc_bind.bind_types)free(__kmp_nested_proc_bind.bind_types); | ||||
8256 | __kmp_nested_proc_bind.bind_types = NULL__null; | ||||
8257 | __kmp_nested_proc_bind.size = 0; | ||||
8258 | __kmp_nested_proc_bind.used = 0; | ||||
8259 | if (__kmp_affinity_format) { | ||||
8260 | KMP_INTERNAL_FREE(__kmp_affinity_format)free(__kmp_affinity_format); | ||||
8261 | __kmp_affinity_format = NULL__null; | ||||
8262 | } | ||||
8263 | |||||
8264 | __kmp_i18n_catclose(); | ||||
8265 | |||||
8266 | #if KMP_USE_HIER_SCHED0 | ||||
8267 | __kmp_hier_scheds.deallocate(); | ||||
8268 | #endif | ||||
8269 | |||||
8270 | #if KMP_STATS_ENABLED0 | ||||
8271 | __kmp_stats_fini(); | ||||
8272 | #endif | ||||
8273 | |||||
8274 | KA_TRACE(10, ("__kmp_cleanup: exit\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_cleanup: exit\n" ); }; | ||||
8275 | } | ||||
8276 | |||||
8277 | /* ------------------------------------------------------------------------ */ | ||||
8278 | |||||
8279 | int __kmp_ignore_mppbeg(void) { | ||||
8280 | char *env; | ||||
8281 | |||||
8282 | if ((env = getenv("KMP_IGNORE_MPPBEG")) != NULL__null) { | ||||
8283 | if (__kmp_str_match_false(env)) | ||||
8284 | return FALSE0; | ||||
8285 | } | ||||
8286 | // By default __kmpc_begin() is no-op. | ||||
8287 | return TRUE(!0); | ||||
8288 | } | ||||
8289 | |||||
8290 | int __kmp_ignore_mppend(void) { | ||||
8291 | char *env; | ||||
8292 | |||||
8293 | if ((env = getenv("KMP_IGNORE_MPPEND")) != NULL__null) { | ||||
8294 | if (__kmp_str_match_false(env)) | ||||
8295 | return FALSE0; | ||||
8296 | } | ||||
8297 | // By default __kmpc_end() is no-op. | ||||
8298 | return TRUE(!0); | ||||
8299 | } | ||||
8300 | |||||
8301 | void __kmp_internal_begin(void) { | ||||
8302 | int gtid; | ||||
8303 | kmp_root_t *root; | ||||
8304 | |||||
8305 | /* this is a very important step as it will register new sibling threads | ||||
8306 | and assign these new uber threads a new gtid */ | ||||
8307 | gtid = __kmp_entry_gtid()__kmp_get_global_thread_id_reg(); | ||||
8308 | root = __kmp_threads[gtid]->th.th_root; | ||||
8309 | KMP_ASSERT(KMP_UBER_GTID(gtid))if (!(KMP_UBER_GTID(gtid))) { __kmp_debug_assert("KMP_UBER_GTID(gtid)" , "openmp/runtime/src/kmp_runtime.cpp", 8309); }; | ||||
8310 | |||||
8311 | if (root->r.r_begin) | ||||
8312 | return; | ||||
8313 | __kmp_acquire_lock(&root->r.r_begin_lock, gtid); | ||||
8314 | if (root->r.r_begin) { | ||||
8315 | __kmp_release_lock(&root->r.r_begin_lock, gtid); | ||||
8316 | return; | ||||
8317 | } | ||||
8318 | |||||
8319 | root->r.r_begin = TRUE(!0); | ||||
8320 | |||||
8321 | __kmp_release_lock(&root->r.r_begin_lock, gtid); | ||||
8322 | } | ||||
8323 | |||||
8324 | /* ------------------------------------------------------------------------ */ | ||||
8325 | |||||
8326 | void __kmp_user_set_library(enum library_type arg) { | ||||
8327 | int gtid; | ||||
8328 | kmp_root_t *root; | ||||
8329 | kmp_info_t *thread; | ||||
8330 | |||||
8331 | /* first, make sure we are initialized so we can get our gtid */ | ||||
8332 | |||||
8333 | gtid = __kmp_entry_gtid()__kmp_get_global_thread_id_reg(); | ||||
8334 | thread = __kmp_threads[gtid]; | ||||
8335 | |||||
8336 | root = thread->th.th_root; | ||||
8337 | |||||
8338 | KA_TRACE(20, ("__kmp_user_set_library: enter T#%d, arg: %d, %d\n", gtid, arg,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_user_set_library: enter T#%d, arg: %d, %d\n" , gtid, arg, library_serial); } | ||||
8339 | library_serial))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_user_set_library: enter T#%d, arg: %d, %d\n" , gtid, arg, library_serial); }; | ||||
8340 | if (root->r.r_in_parallel) { /* Must be called in serial section of top-level | ||||
8341 | thread */ | ||||
8342 | KMP_WARNING(SetLibraryIncorrectCall)__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_SetLibraryIncorrectCall ), __kmp_msg_null); | ||||
8343 | return; | ||||
8344 | } | ||||
8345 | |||||
8346 | switch (arg) { | ||||
8347 | case library_serial: | ||||
8348 | thread->th.th_set_nproc = 0; | ||||
8349 | set__nproc(thread, 1)(((thread)->th.th_current_task->td_icvs.nproc) = (1)); | ||||
8350 | break; | ||||
8351 | case library_turnaround: | ||||
8352 | thread->th.th_set_nproc = 0; | ||||
8353 | set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth(((thread)->th.th_current_task->td_icvs.nproc) = (__kmp_dflt_team_nth ? __kmp_dflt_team_nth : __kmp_dflt_team_nth_ub)) | ||||
8354 | : __kmp_dflt_team_nth_ub)(((thread)->th.th_current_task->td_icvs.nproc) = (__kmp_dflt_team_nth ? __kmp_dflt_team_nth : __kmp_dflt_team_nth_ub)); | ||||
8355 | break; | ||||
8356 | case library_throughput: | ||||
8357 | thread->th.th_set_nproc = 0; | ||||
8358 | set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth(((thread)->th.th_current_task->td_icvs.nproc) = (__kmp_dflt_team_nth ? __kmp_dflt_team_nth : __kmp_dflt_team_nth_ub)) | ||||
8359 | : __kmp_dflt_team_nth_ub)(((thread)->th.th_current_task->td_icvs.nproc) = (__kmp_dflt_team_nth ? __kmp_dflt_team_nth : __kmp_dflt_team_nth_ub)); | ||||
8360 | break; | ||||
8361 | default: | ||||
8362 | KMP_FATAL(UnknownLibraryType, arg)__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_UnknownLibraryType, arg), __kmp_msg_null); | ||||
8363 | } | ||||
8364 | |||||
8365 | __kmp_aux_set_library(arg); | ||||
8366 | } | ||||
8367 | |||||
8368 | void __kmp_aux_set_stacksize(size_t arg) { | ||||
8369 | if (!__kmp_init_serial) | ||||
8370 | __kmp_serial_initialize(); | ||||
8371 | |||||
8372 | #if KMP_OS_DARWIN0 | ||||
8373 | if (arg & (0x1000 - 1)) { | ||||
8374 | arg &= ~(0x1000 - 1); | ||||
8375 | if (arg + 0x1000) /* check for overflow if we round up */ | ||||
8376 | arg += 0x1000; | ||||
8377 | } | ||||
8378 | #endif | ||||
8379 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); | ||||
8380 | |||||
8381 | /* only change the default stacksize before the first parallel region */ | ||||
8382 | if (!TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) { | ||||
8383 | size_t value = arg; /* argument is in bytes */ | ||||
8384 | |||||
8385 | if (value < __kmp_sys_min_stksize) | ||||
8386 | value = __kmp_sys_min_stksize; | ||||
8387 | else if (value > KMP_MAX_STKSIZE(~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1) ))) | ||||
8388 | value = KMP_MAX_STKSIZE(~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1) )); | ||||
8389 | |||||
8390 | __kmp_stksize = value; | ||||
8391 | |||||
8392 | __kmp_env_stksize = TRUE(!0); /* was KMP_STACKSIZE specified? */ | ||||
8393 | } | ||||
8394 | |||||
8395 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | ||||
8396 | } | ||||
8397 | |||||
8398 | /* set the behaviour of the runtime library */ | ||||
8399 | /* TODO this can cause some odd behaviour with sibling parallelism... */ | ||||
8400 | void __kmp_aux_set_library(enum library_type arg) { | ||||
8401 | __kmp_library = arg; | ||||
8402 | |||||
8403 | switch (__kmp_library) { | ||||
8404 | case library_serial: { | ||||
8405 | KMP_INFORM(LibraryIsSerial)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_LibraryIsSerial ), __kmp_msg_null); | ||||
8406 | } break; | ||||
8407 | case library_turnaround: | ||||
8408 | if (__kmp_use_yield == 1 && !__kmp_use_yield_exp_set) | ||||
8409 | __kmp_use_yield = 2; // only yield when oversubscribed | ||||
8410 | break; | ||||
8411 | case library_throughput: | ||||
8412 | if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME(2147483647)) | ||||
8413 | __kmp_dflt_blocktime = KMP_DEFAULT_BLOCKTIME(__kmp_is_hybrid_cpu() ? (0) : (200)); | ||||
8414 | break; | ||||
8415 | default: | ||||
8416 | KMP_FATAL(UnknownLibraryType, arg)__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_UnknownLibraryType, arg), __kmp_msg_null); | ||||
8417 | } | ||||
8418 | } | ||||
8419 | |||||
8420 | /* Getting team information common for all team API */ | ||||
8421 | // Returns NULL if not in teams construct | ||||
8422 | static kmp_team_t *__kmp_aux_get_team_info(int &teams_serialized) { | ||||
8423 | kmp_info_t *thr = __kmp_entry_thread(); | ||||
8424 | teams_serialized = 0; | ||||
8425 | if (thr->th.th_teams_microtask) { | ||||
8426 | kmp_team_t *team = thr->th.th_team; | ||||
8427 | int tlevel = thr->th.th_teams_level; // the level of the teams construct | ||||
8428 | int ii = team->t.t_level; | ||||
8429 | teams_serialized = team->t.t_serialized; | ||||
8430 | int level = tlevel + 1; | ||||
8431 | KMP_DEBUG_ASSERT(ii >= tlevel)if (!(ii >= tlevel)) { __kmp_debug_assert("ii >= tlevel" , "openmp/runtime/src/kmp_runtime.cpp", 8431); }; | ||||
8432 | while (ii > level) { | ||||
8433 | for (teams_serialized = team->t.t_serialized; | ||||
8434 | (teams_serialized > 0) && (ii > level); teams_serialized--, ii--) { | ||||
8435 | } | ||||
8436 | if (team->t.t_serialized && (!teams_serialized)) { | ||||
8437 | team = team->t.t_parent; | ||||
8438 | continue; | ||||
8439 | } | ||||
8440 | if (ii > level) { | ||||
8441 | team = team->t.t_parent; | ||||
8442 | ii--; | ||||
8443 | } | ||||
8444 | } | ||||
8445 | return team; | ||||
8446 | } | ||||
8447 | return NULL__null; | ||||
8448 | } | ||||
8449 | |||||
8450 | int __kmp_aux_get_team_num() { | ||||
8451 | int serialized; | ||||
8452 | kmp_team_t *team = __kmp_aux_get_team_info(serialized); | ||||
8453 | if (team) { | ||||
8454 | if (serialized > 1) { | ||||
8455 | return 0; // teams region is serialized ( 1 team of 1 thread ). | ||||
8456 | } else { | ||||
8457 | return team->t.t_master_tid; | ||||
8458 | } | ||||
8459 | } | ||||
8460 | return 0; | ||||
8461 | } | ||||
8462 | |||||
8463 | int __kmp_aux_get_num_teams() { | ||||
8464 | int serialized; | ||||
8465 | kmp_team_t *team = __kmp_aux_get_team_info(serialized); | ||||
8466 | if (team) { | ||||
8467 | if (serialized > 1) { | ||||
8468 | return 1; | ||||
8469 | } else { | ||||
8470 | return team->t.t_parent->t.t_nproc; | ||||
8471 | } | ||||
8472 | } | ||||
8473 | return 1; | ||||
8474 | } | ||||
8475 | |||||
8476 | /* ------------------------------------------------------------------------ */ | ||||
8477 | |||||
8478 | /* | ||||
8479 | * Affinity Format Parser | ||||
8480 | * | ||||
8481 | * Field is in form of: %[[[0].]size]type | ||||
8482 | * % and type are required (%% means print a literal '%') | ||||
8483 | * type is either single char or long name surrounded by {}, | ||||
8484 | * e.g., N or {num_threads} | ||||
8485 | * 0 => leading zeros | ||||
8486 | * . => right justified when size is specified | ||||
8487 | * by default output is left justified | ||||
8488 | * size is the *minimum* field length | ||||
8489 | * All other characters are printed as is | ||||
8490 | * | ||||
8491 | * Available field types: | ||||
8492 | * L {thread_level} - omp_get_level() | ||||
8493 | * n {thread_num} - omp_get_thread_num() | ||||
8494 | * h {host} - name of host machine | ||||
8495 | * P {process_id} - process id (integer) | ||||
8496 | * T {thread_identifier} - native thread identifier (integer) | ||||
8497 | * N {num_threads} - omp_get_num_threads() | ||||
8498 | * A {ancestor_tnum} - omp_get_ancestor_thread_num(omp_get_level()-1) | ||||
8499 | * a {thread_affinity} - comma separated list of integers or integer ranges | ||||
8500 | * (values of affinity mask) | ||||
8501 | * | ||||
8502 | * Implementation-specific field types can be added | ||||
8503 | * If a type is unknown, print "undefined" | ||||
8504 | */ | ||||
8505 | |||||
8506 | // Structure holding the short name, long name, and corresponding data type | ||||
8507 | // for snprintf. A table of these will represent the entire valid keyword | ||||
8508 | // field types. | ||||
8509 | typedef struct kmp_affinity_format_field_t { | ||||
8510 | char short_name; // from spec e.g., L -> thread level | ||||
8511 | const char *long_name; // from spec thread_level -> thread level | ||||
8512 | char field_format; // data type for snprintf (typically 'd' or 's' | ||||
8513 | // for integer or string) | ||||
8514 | } kmp_affinity_format_field_t; | ||||
8515 | |||||
8516 | static const kmp_affinity_format_field_t __kmp_affinity_format_table[] = { | ||||
8517 | #if KMP_AFFINITY_SUPPORTED1 | ||||
8518 | {'A', "thread_affinity", 's'}, | ||||
8519 | #endif | ||||
8520 | {'t', "team_num", 'd'}, | ||||
8521 | {'T', "num_teams", 'd'}, | ||||
8522 | {'L', "nesting_level", 'd'}, | ||||
8523 | {'n', "thread_num", 'd'}, | ||||
8524 | {'N', "num_threads", 'd'}, | ||||
8525 | {'a', "ancestor_tnum", 'd'}, | ||||
8526 | {'H', "host", 's'}, | ||||
8527 | {'P', "process_id", 'd'}, | ||||
8528 | {'i', "native_thread_id", 'd'}}; | ||||
8529 | |||||
8530 | // Return the number of characters it takes to hold field | ||||
8531 | static int __kmp_aux_capture_affinity_field(int gtid, const kmp_info_t *th, | ||||
8532 | const char **ptr, | ||||
8533 | kmp_str_buf_t *field_buffer) { | ||||
8534 | int rc, format_index, field_value; | ||||
8535 | const char *width_left, *width_right; | ||||
8536 | bool pad_zeros, right_justify, parse_long_name, found_valid_name; | ||||
8537 | static const int FORMAT_SIZE = 20; | ||||
8538 | char format[FORMAT_SIZE] = {0}; | ||||
8539 | char absolute_short_name = 0; | ||||
8540 | |||||
8541 | KMP_DEBUG_ASSERT(gtid >= 0)if (!(gtid >= 0)) { __kmp_debug_assert("gtid >= 0", "openmp/runtime/src/kmp_runtime.cpp" , 8541); }; | ||||
8542 | KMP_DEBUG_ASSERT(th)if (!(th)) { __kmp_debug_assert("th", "openmp/runtime/src/kmp_runtime.cpp" , 8542); }; | ||||
8543 | KMP_DEBUG_ASSERT(**ptr == '%')if (!(**ptr == '%')) { __kmp_debug_assert("**ptr == '%'", "openmp/runtime/src/kmp_runtime.cpp" , 8543); }; | ||||
8544 | KMP_DEBUG_ASSERT(field_buffer)if (!(field_buffer)) { __kmp_debug_assert("field_buffer", "openmp/runtime/src/kmp_runtime.cpp" , 8544); }; | ||||
8545 | |||||
8546 | __kmp_str_buf_clear(field_buffer); | ||||
8547 | |||||
8548 | // Skip the initial % | ||||
8549 | (*ptr)++; | ||||
8550 | |||||
8551 | // Check for %% first | ||||
8552 | if (**ptr == '%') { | ||||
8553 | __kmp_str_buf_cat(field_buffer, "%", 1); | ||||
8554 | (*ptr)++; // skip over the second % | ||||
8555 | return 1; | ||||
8556 | } | ||||
8557 | |||||
8558 | // Parse field modifiers if they are present | ||||
8559 | pad_zeros = false; | ||||
8560 | if (**ptr == '0') { | ||||
8561 | pad_zeros = true; | ||||
8562 | (*ptr)++; // skip over 0 | ||||
8563 | } | ||||
8564 | right_justify = false; | ||||
8565 | if (**ptr == '.') { | ||||
8566 | right_justify = true; | ||||
8567 | (*ptr)++; // skip over . | ||||
8568 | } | ||||
8569 | // Parse width of field: [width_left, width_right) | ||||
8570 | width_left = width_right = NULL__null; | ||||
8571 | if (**ptr >= '0' && **ptr <= '9') { | ||||
8572 | width_left = *ptr; | ||||
8573 | SKIP_DIGITS(*ptr){ while (*(*ptr) >= '0' && *(*ptr) <= '9') (*ptr )++; }; | ||||
8574 | width_right = *ptr; | ||||
8575 | } | ||||
8576 | |||||
8577 | // Create the format for KMP_SNPRINTF based on flags parsed above | ||||
8578 | format_index = 0; | ||||
8579 | format[format_index++] = '%'; | ||||
8580 | if (!right_justify) | ||||
8581 | format[format_index++] = '-'; | ||||
8582 | if (pad_zeros) | ||||
8583 | format[format_index++] = '0'; | ||||
8584 | if (width_left && width_right) { | ||||
8585 | int i = 0; | ||||
8586 | // Only allow 8 digit number widths. | ||||
8587 | // This also prevents overflowing format variable | ||||
8588 | while (i < 8 && width_left < width_right) { | ||||
8589 | format[format_index++] = *width_left; | ||||
8590 | width_left++; | ||||
8591 | i++; | ||||
8592 | } | ||||
8593 | } | ||||
8594 | |||||
8595 | // Parse a name (long or short) | ||||
8596 | // Canonicalize the name into absolute_short_name | ||||
8597 | found_valid_name = false; | ||||
8598 | parse_long_name = (**ptr == '{'); | ||||
8599 | if (parse_long_name) | ||||
8600 | (*ptr)++; // skip initial left brace | ||||
8601 | for (size_t i = 0; i < sizeof(__kmp_affinity_format_table) / | ||||
8602 | sizeof(__kmp_affinity_format_table[0]); | ||||
8603 | ++i) { | ||||
8604 | char short_name = __kmp_affinity_format_table[i].short_name; | ||||
8605 | const char *long_name = __kmp_affinity_format_table[i].long_name; | ||||
8606 | char field_format = __kmp_affinity_format_table[i].field_format; | ||||
8607 | if (parse_long_name) { | ||||
8608 | size_t length = KMP_STRLENstrlen(long_name); | ||||
8609 | if (strncmp(*ptr, long_name, length) == 0) { | ||||
8610 | found_valid_name = true; | ||||
8611 | (*ptr) += length; // skip the long name | ||||
8612 | } | ||||
8613 | } else if (**ptr == short_name) { | ||||
8614 | found_valid_name = true; | ||||
8615 | (*ptr)++; // skip the short name | ||||
8616 | } | ||||
8617 | if (found_valid_name) { | ||||
8618 | format[format_index++] = field_format; | ||||
8619 | format[format_index++] = '\0'; | ||||
8620 | absolute_short_name = short_name; | ||||
8621 | break; | ||||
8622 | } | ||||
8623 | } | ||||
8624 | if (parse_long_name) { | ||||
8625 | if (**ptr != '}') { | ||||
8626 | absolute_short_name = 0; | ||||
8627 | } else { | ||||
8628 | (*ptr)++; // skip over the right brace | ||||
8629 | } | ||||
8630 | } | ||||
8631 | |||||
8632 | // Attempt to fill the buffer with the requested | ||||
8633 | // value using snprintf within __kmp_str_buf_print() | ||||
8634 | switch (absolute_short_name) { | ||||
8635 | case 't': | ||||
8636 | rc = __kmp_str_buf_print(field_buffer, format, __kmp_aux_get_team_num()); | ||||
8637 | break; | ||||
8638 | case 'T': | ||||
8639 | rc = __kmp_str_buf_print(field_buffer, format, __kmp_aux_get_num_teams()); | ||||
8640 | break; | ||||
8641 | case 'L': | ||||
8642 | rc = __kmp_str_buf_print(field_buffer, format, th->th.th_team->t.t_level); | ||||
8643 | break; | ||||
8644 | case 'n': | ||||
8645 | rc = __kmp_str_buf_print(field_buffer, format, __kmp_tid_from_gtid(gtid)); | ||||
8646 | break; | ||||
8647 | case 'H': { | ||||
8648 | static const int BUFFER_SIZE = 256; | ||||
8649 | char buf[BUFFER_SIZE]; | ||||
8650 | __kmp_expand_host_name(buf, BUFFER_SIZE); | ||||
8651 | rc = __kmp_str_buf_print(field_buffer, format, buf); | ||||
8652 | } break; | ||||
8653 | case 'P': | ||||
8654 | rc = __kmp_str_buf_print(field_buffer, format, getpid()); | ||||
8655 | break; | ||||
8656 | case 'i': | ||||
8657 | rc = __kmp_str_buf_print(field_buffer, format, __kmp_gettid()syscall(186)); | ||||
8658 | break; | ||||
8659 | case 'N': | ||||
8660 | rc = __kmp_str_buf_print(field_buffer, format, th->th.th_team->t.t_nproc); | ||||
8661 | break; | ||||
8662 | case 'a': | ||||
8663 | field_value = | ||||
8664 | __kmp_get_ancestor_thread_num(gtid, th->th.th_team->t.t_level - 1); | ||||
8665 | rc = __kmp_str_buf_print(field_buffer, format, field_value); | ||||
8666 | break; | ||||
8667 | #if KMP_AFFINITY_SUPPORTED1 | ||||
8668 | case 'A': { | ||||
8669 | kmp_str_buf_t buf; | ||||
8670 | __kmp_str_buf_init(&buf){ (&buf)->str = (&buf)->bulk; (&buf)->size = sizeof((&buf)->bulk); (&buf)->used = 0; (& buf)->bulk[0] = 0; }; | ||||
8671 | __kmp_affinity_str_buf_mask(&buf, th->th.th_affin_mask); | ||||
8672 | rc = __kmp_str_buf_print(field_buffer, format, buf.str); | ||||
8673 | __kmp_str_buf_free(&buf); | ||||
8674 | } break; | ||||
8675 | #endif | ||||
8676 | default: | ||||
8677 | // According to spec, If an implementation does not have info for field | ||||
8678 | // type, then "undefined" is printed | ||||
8679 | rc = __kmp_str_buf_print(field_buffer, "%s", "undefined"); | ||||
8680 | // Skip the field | ||||
8681 | if (parse_long_name) { | ||||
8682 | SKIP_TOKEN(*ptr){ while ((*(*ptr) >= '0' && *(*ptr) <= '9') || ( *(*ptr) >= 'a' && *(*ptr) <= 'z') || (*(*ptr) >= 'A' && *(*ptr) <= 'Z') || *(*ptr) == '_') (*ptr)++ ; }; | ||||
8683 | if (**ptr == '}') | ||||
8684 | (*ptr)++; | ||||
8685 | } else { | ||||
8686 | (*ptr)++; | ||||
8687 | } | ||||
8688 | } | ||||
8689 | |||||
8690 | KMP_ASSERT(format_index <= FORMAT_SIZE)if (!(format_index <= FORMAT_SIZE)) { __kmp_debug_assert("format_index <= FORMAT_SIZE" , "openmp/runtime/src/kmp_runtime.cpp", 8690); }; | ||||
8691 | return rc; | ||||
8692 | } | ||||
8693 | |||||
8694 | /* | ||||
8695 | * Return number of characters needed to hold the affinity string | ||||
8696 | * (not including null byte character) | ||||
8697 | * The resultant string is printed to buffer, which the caller can then | ||||
8698 | * handle afterwards | ||||
8699 | */ | ||||
8700 | size_t __kmp_aux_capture_affinity(int gtid, const char *format, | ||||
8701 | kmp_str_buf_t *buffer) { | ||||
8702 | const char *parse_ptr; | ||||
8703 | size_t retval; | ||||
8704 | const kmp_info_t *th; | ||||
8705 | kmp_str_buf_t field; | ||||
8706 | |||||
8707 | KMP_DEBUG_ASSERT(buffer)if (!(buffer)) { __kmp_debug_assert("buffer", "openmp/runtime/src/kmp_runtime.cpp" , 8707); }; | ||||
8708 | KMP_DEBUG_ASSERT(gtid >= 0)if (!(gtid >= 0)) { __kmp_debug_assert("gtid >= 0", "openmp/runtime/src/kmp_runtime.cpp" , 8708); }; | ||||
8709 | |||||
8710 | __kmp_str_buf_init(&field){ (&field)->str = (&field)->bulk; (&field)-> size = sizeof((&field)->bulk); (&field)->used = 0; (&field)->bulk[0] = 0; }; | ||||
8711 | __kmp_str_buf_clear(buffer); | ||||
8712 | |||||
8713 | th = __kmp_threads[gtid]; | ||||
8714 | retval = 0; | ||||
8715 | |||||
8716 | // If format is NULL or zero-length string, then we use | ||||
8717 | // affinity-format-var ICV | ||||
8718 | parse_ptr = format; | ||||
8719 | if (parse_ptr == NULL__null || *parse_ptr == '\0') { | ||||
8720 | parse_ptr = __kmp_affinity_format; | ||||
8721 | } | ||||
8722 | KMP_DEBUG_ASSERT(parse_ptr)if (!(parse_ptr)) { __kmp_debug_assert("parse_ptr", "openmp/runtime/src/kmp_runtime.cpp" , 8722); }; | ||||
8723 | |||||
8724 | while (*parse_ptr != '\0') { | ||||
8725 | // Parse a field | ||||
8726 | if (*parse_ptr == '%') { | ||||
8727 | // Put field in the buffer | ||||
8728 | int rc = __kmp_aux_capture_affinity_field(gtid, th, &parse_ptr, &field); | ||||
8729 | __kmp_str_buf_catbuf(buffer, &field); | ||||
8730 | retval += rc; | ||||
8731 | } else { | ||||
8732 | // Put literal character in buffer | ||||
8733 | __kmp_str_buf_cat(buffer, parse_ptr, 1); | ||||
8734 | retval++; | ||||
8735 | parse_ptr++; | ||||
8736 | } | ||||
8737 | } | ||||
8738 | __kmp_str_buf_free(&field); | ||||
8739 | return retval; | ||||
8740 | } | ||||
8741 | |||||
8742 | // Displays the affinity string to stdout | ||||
8743 | void __kmp_aux_display_affinity(int gtid, const char *format) { | ||||
8744 | kmp_str_buf_t buf; | ||||
8745 | __kmp_str_buf_init(&buf){ (&buf)->str = (&buf)->bulk; (&buf)->size = sizeof((&buf)->bulk); (&buf)->used = 0; (& buf)->bulk[0] = 0; }; | ||||
8746 | __kmp_aux_capture_affinity(gtid, format, &buf); | ||||
8747 | __kmp_fprintf(kmp_out, "%s" KMP_END_OF_LINE"\n", buf.str); | ||||
8748 | __kmp_str_buf_free(&buf); | ||||
8749 | } | ||||
8750 | |||||
8751 | /* ------------------------------------------------------------------------ */ | ||||
8752 | |||||
8753 | void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid) { | ||||
8754 | int blocktime = arg; /* argument is in milliseconds */ | ||||
8755 | #if KMP_USE_MONITOR | ||||
8756 | int bt_intervals; | ||||
8757 | #endif | ||||
8758 | kmp_int8 bt_set; | ||||
8759 | |||||
8760 | __kmp_save_internal_controls(thread); | ||||
8761 | |||||
8762 | /* Normalize and set blocktime for the teams */ | ||||
8763 | if (blocktime < KMP_MIN_BLOCKTIME(0)) | ||||
8764 | blocktime = KMP_MIN_BLOCKTIME(0); | ||||
8765 | else if (blocktime > KMP_MAX_BLOCKTIME(2147483647)) | ||||
8766 | blocktime = KMP_MAX_BLOCKTIME(2147483647); | ||||
8767 | |||||
8768 | set__blocktime_team(thread->th.th_team, tid, blocktime)(((thread->th.th_team)->t.t_threads[(tid)]->th.th_current_task ->td_icvs.blocktime) = (blocktime)); | ||||
8769 | set__blocktime_team(thread->th.th_serial_team, 0, blocktime)(((thread->th.th_serial_team)->t.t_threads[(0)]->th. th_current_task->td_icvs.blocktime) = (blocktime)); | ||||
8770 | |||||
8771 | #if KMP_USE_MONITOR | ||||
8772 | /* Calculate and set blocktime intervals for the teams */ | ||||
8773 | bt_intervals = KMP_INTERVALS_FROM_BLOCKTIME(blocktime, __kmp_monitor_wakeups); | ||||
8774 | |||||
8775 | set__bt_intervals_team(thread->th.th_team, tid, bt_intervals); | ||||
8776 | set__bt_intervals_team(thread->th.th_serial_team, 0, bt_intervals); | ||||
8777 | #endif | ||||
8778 | |||||
8779 | /* Set whether blocktime has been set to "TRUE" */ | ||||
8780 | bt_set = TRUE(!0); | ||||
8781 | |||||
8782 | set__bt_set_team(thread->th.th_team, tid, bt_set)(((thread->th.th_team)->t.t_threads[(tid)]->th.th_current_task ->td_icvs.bt_set) = (bt_set)); | ||||
8783 | set__bt_set_team(thread->th.th_serial_team, 0, bt_set)(((thread->th.th_serial_team)->t.t_threads[(0)]->th. th_current_task->td_icvs.bt_set) = (bt_set)); | ||||
8784 | #if KMP_USE_MONITOR | ||||
8785 | KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, "if (kmp_f_debug >= 10) { __kmp_debug_printf ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, " "bt_intervals=%d, monitor_updates=%d\n", __kmp_gtid_from_tid (tid, thread->th.th_team), thread->th.th_team->t.t_id , tid, blocktime, bt_intervals, __kmp_monitor_wakeups); } | ||||
8786 | "bt_intervals=%d, monitor_updates=%d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, " "bt_intervals=%d, monitor_updates=%d\n", __kmp_gtid_from_tid (tid, thread->th.th_team), thread->th.th_team->t.t_id , tid, blocktime, bt_intervals, __kmp_monitor_wakeups); } | ||||
8787 | __kmp_gtid_from_tid(tid, thread->th.th_team),if (kmp_f_debug >= 10) { __kmp_debug_printf ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, " "bt_intervals=%d, monitor_updates=%d\n", __kmp_gtid_from_tid (tid, thread->th.th_team), thread->th.th_team->t.t_id , tid, blocktime, bt_intervals, __kmp_monitor_wakeups); } | ||||
8788 | thread->th.th_team->t.t_id, tid, blocktime, bt_intervals,if (kmp_f_debug >= 10) { __kmp_debug_printf ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, " "bt_intervals=%d, monitor_updates=%d\n", __kmp_gtid_from_tid (tid, thread->th.th_team), thread->th.th_team->t.t_id , tid, blocktime, bt_intervals, __kmp_monitor_wakeups); } | ||||
8789 | __kmp_monitor_wakeups))if (kmp_f_debug >= 10) { __kmp_debug_printf ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, " "bt_intervals=%d, monitor_updates=%d\n", __kmp_gtid_from_tid (tid, thread->th.th_team), thread->th.th_team->t.t_id , tid, blocktime, bt_intervals, __kmp_monitor_wakeups); }; | ||||
8790 | #else | ||||
8791 | KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d\n" , __kmp_gtid_from_tid(tid, thread->th.th_team), thread-> th.th_team->t.t_id, tid, blocktime); } | ||||
8792 | __kmp_gtid_from_tid(tid, thread->th.th_team),if (kmp_f_debug >= 10) { __kmp_debug_printf ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d\n" , __kmp_gtid_from_tid(tid, thread->th.th_team), thread-> th.th_team->t.t_id, tid, blocktime); } | ||||
8793 | thread->th.th_team->t.t_id, tid, blocktime))if (kmp_f_debug >= 10) { __kmp_debug_printf ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d\n" , __kmp_gtid_from_tid(tid, thread->th.th_team), thread-> th.th_team->t.t_id, tid, blocktime); }; | ||||
8794 | #endif | ||||
8795 | } | ||||
8796 | |||||
8797 | void __kmp_aux_set_defaults(char const *str, size_t len) { | ||||
8798 | if (!__kmp_init_serial) { | ||||
8799 | __kmp_serial_initialize(); | ||||
8800 | } | ||||
8801 | __kmp_env_initialize(str); | ||||
8802 | |||||
8803 | if (__kmp_settings || __kmp_display_env || __kmp_display_env_verbose) { | ||||
8804 | __kmp_env_print(); | ||||
8805 | } | ||||
8806 | } // __kmp_aux_set_defaults | ||||
8807 | |||||
8808 | /* ------------------------------------------------------------------------ */ | ||||
8809 | /* internal fast reduction routines */ | ||||
8810 | |||||
8811 | PACKED_REDUCTION_METHOD_T | ||||
8812 | __kmp_determine_reduction_method( | ||||
8813 | ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, | ||||
8814 | void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data), | ||||
8815 | kmp_critical_name *lck) { | ||||
8816 | |||||
8817 | // Default reduction method: critical construct ( lck != NULL, like in current | ||||
8818 | // PAROPT ) | ||||
8819 | // If ( reduce_data!=NULL && reduce_func!=NULL ): the tree-reduction method | ||||
8820 | // can be selected by RTL | ||||
8821 | // If loc->flags contains KMP_IDENT_ATOMIC_REDUCE, the atomic reduce method | ||||
8822 | // can be selected by RTL | ||||
8823 | // Finally, it's up to OpenMP RTL to make a decision on which method to select | ||||
8824 | // among generated by PAROPT. | ||||
8825 | |||||
8826 | PACKED_REDUCTION_METHOD_T retval; | ||||
8827 | |||||
8828 | int team_size; | ||||
8829 | |||||
8830 | KMP_DEBUG_ASSERT(loc)if (!(loc)) { __kmp_debug_assert("loc", "openmp/runtime/src/kmp_runtime.cpp" , 8830); }; // it would be nice to test ( loc != 0 ) | ||||
8831 | KMP_DEBUG_ASSERT(lck)if (!(lck)) { __kmp_debug_assert("lck", "openmp/runtime/src/kmp_runtime.cpp" , 8831); }; // it would be nice to test ( lck != 0 ) | ||||
8832 | |||||
8833 | #define FAST_REDUCTION_ATOMIC_METHOD_GENERATED \ | ||||
8834 | (loc && \ | ||||
8835 | ((loc->flags & (KMP_IDENT_ATOMIC_REDUCE)) == (KMP_IDENT_ATOMIC_REDUCE))) | ||||
8836 | #define FAST_REDUCTION_TREE_METHOD_GENERATED ((reduce_data) && (reduce_func)) | ||||
8837 | |||||
8838 | retval = critical_reduce_block; | ||||
8839 | |||||
8840 | // another choice of getting a team size (with 1 dynamic deference) is slower | ||||
8841 | team_size = __kmp_get_team_num_threads(global_tid)(__kmp_threads[(global_tid)]->th.th_team->t.t_nproc); | ||||
8842 | if (team_size == 1) { | ||||
8843 | |||||
8844 | retval = empty_reduce_block; | ||||
8845 | |||||
8846 | } else { | ||||
8847 | |||||
8848 | int atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED; | ||||
8849 | |||||
8850 | #if KMP_ARCH_X86_641 || KMP_ARCH_PPC64(0 || 0) || KMP_ARCH_AARCH640 || \ | ||||
8851 | KMP_ARCH_MIPS640 || KMP_ARCH_RISCV640 || KMP_ARCH_LOONGARCH640 | ||||
8852 | |||||
8853 | #if KMP_OS_LINUX1 || KMP_OS_DRAGONFLY0 || KMP_OS_FREEBSD0 || KMP_OS_NETBSD0 || \ | ||||
8854 | KMP_OS_OPENBSD0 || KMP_OS_WINDOWS0 || KMP_OS_DARWIN0 || KMP_OS_HURD0 | ||||
8855 | |||||
8856 | int teamsize_cutoff = 4; | ||||
8857 | |||||
8858 | #if KMP_MIC_SUPPORTED((0 || 1) && (1 || 0)) | ||||
8859 | if (__kmp_mic_type != non_mic) { | ||||
8860 | teamsize_cutoff = 8; | ||||
8861 | } | ||||
8862 | #endif | ||||
8863 | int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED; | ||||
8864 | if (tree_available) { | ||||
8865 | if (team_size <= teamsize_cutoff) { | ||||
8866 | if (atomic_available) { | ||||
8867 | retval = atomic_reduce_block; | ||||
8868 | } | ||||
8869 | } else { | ||||
8870 | retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER(((tree_reduce_block) | (bs_reduction_barrier))); | ||||
8871 | } | ||||
8872 | } else if (atomic_available) { | ||||
8873 | retval = atomic_reduce_block; | ||||
8874 | } | ||||
8875 | #else | ||||
8876 | #error "Unknown or unsupported OS" | ||||
8877 | #endif // KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || | ||||
8878 | // KMP_OS_OPENBSD || KMP_OS_WINDOWS || KMP_OS_DARWIN || KMP_OS_HURD | ||||
8879 | |||||
8880 | #elif KMP_ARCH_X860 || KMP_ARCH_ARM || KMP_ARCH_AARCH || KMP_ARCH_MIPS0 | ||||
8881 | |||||
8882 | #if KMP_OS_LINUX1 || KMP_OS_FREEBSD0 || KMP_OS_WINDOWS0 || KMP_OS_HURD0 | ||||
8883 | |||||
8884 | // basic tuning | ||||
8885 | |||||
8886 | if (atomic_available) { | ||||
8887 | if (num_vars <= 2) { // && ( team_size <= 8 ) due to false-sharing ??? | ||||
8888 | retval = atomic_reduce_block; | ||||
8889 | } | ||||
8890 | } // otherwise: use critical section | ||||
8891 | |||||
8892 | #elif KMP_OS_DARWIN0 | ||||
8893 | |||||
8894 | int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED; | ||||
8895 | if (atomic_available && (num_vars <= 3)) { | ||||
8896 | retval = atomic_reduce_block; | ||||
8897 | } else if (tree_available) { | ||||
8898 | if ((reduce_size > (9 * sizeof(kmp_real64))) && | ||||
8899 | (reduce_size < (2000 * sizeof(kmp_real64)))) { | ||||
8900 | retval = TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER(((tree_reduce_block) | (bs_plain_barrier))); | ||||
8901 | } | ||||
8902 | } // otherwise: use critical section | ||||
8903 | |||||
8904 | #else | ||||
8905 | #error "Unknown or unsupported OS" | ||||
8906 | #endif | ||||
8907 | |||||
8908 | #else | ||||
8909 | #error "Unknown or unsupported architecture" | ||||
8910 | #endif | ||||
8911 | } | ||||
8912 | |||||
8913 | // KMP_FORCE_REDUCTION | ||||
8914 | |||||
8915 | // If the team is serialized (team_size == 1), ignore the forced reduction | ||||
8916 | // method and stay with the unsynchronized method (empty_reduce_block) | ||||
8917 | if (__kmp_force_reduction_method != reduction_method_not_defined && | ||||
8918 | team_size != 1) { | ||||
8919 | |||||
8920 | PACKED_REDUCTION_METHOD_T forced_retval = critical_reduce_block; | ||||
8921 | |||||
8922 | int atomic_available, tree_available; | ||||
8923 | |||||
8924 | switch ((forced_retval = __kmp_force_reduction_method)) { | ||||
8925 | case critical_reduce_block: | ||||
8926 | KMP_ASSERT(lck)if (!(lck)) { __kmp_debug_assert("lck", "openmp/runtime/src/kmp_runtime.cpp" , 8926); }; // lck should be != 0 | ||||
8927 | break; | ||||
8928 | |||||
8929 | case atomic_reduce_block: | ||||
8930 | atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED; | ||||
8931 | if (!atomic_available) { | ||||
8932 | KMP_WARNING(RedMethodNotSupported, "atomic")__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_RedMethodNotSupported , "atomic"), __kmp_msg_null); | ||||
8933 | forced_retval = critical_reduce_block; | ||||
8934 | } | ||||
8935 | break; | ||||
8936 | |||||
8937 | case tree_reduce_block: | ||||
8938 | tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED; | ||||
8939 | if (!tree_available) { | ||||
8940 | KMP_WARNING(RedMethodNotSupported, "tree")__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_RedMethodNotSupported , "tree"), __kmp_msg_null); | ||||
8941 | forced_retval = critical_reduce_block; | ||||
8942 | } else { | ||||
8943 | #if KMP_FAST_REDUCTION_BARRIER1 | ||||
8944 | forced_retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER(((tree_reduce_block) | (bs_reduction_barrier))); | ||||
8945 | #endif | ||||
8946 | } | ||||
8947 | break; | ||||
8948 | |||||
8949 | default: | ||||
8950 | KMP_ASSERT(0)if (!(0)) { __kmp_debug_assert("0", "openmp/runtime/src/kmp_runtime.cpp" , 8950); }; // "unsupported method specified" | ||||
8951 | } | ||||
8952 | |||||
8953 | retval = forced_retval; | ||||
8954 | } | ||||
8955 | |||||
8956 | KA_TRACE(10, ("reduction method selected=%08x\n", retval))if (kmp_a_debug >= 10) { __kmp_debug_printf ("reduction method selected=%08x\n" , retval); }; | ||||
8957 | |||||
8958 | #undef FAST_REDUCTION_TREE_METHOD_GENERATED | ||||
8959 | #undef FAST_REDUCTION_ATOMIC_METHOD_GENERATED | ||||
8960 | |||||
8961 | return (retval); | ||||
8962 | } | ||||
8963 | // this function is for testing set/get/determine reduce method | ||||
8964 | kmp_int32 __kmp_get_reduce_method(void) { | ||||
8965 | return ((__kmp_entry_thread()->th.th_local.packed_reduction_method) >> 8); | ||||
8966 | } | ||||
8967 | |||||
8968 | // Soft pause sets up threads to ignore blocktime and just go to sleep. | ||||
8969 | // Spin-wait code checks __kmp_pause_status and reacts accordingly. | ||||
8970 | void __kmp_soft_pause() { __kmp_pause_status = kmp_soft_paused; } | ||||
8971 | |||||
8972 | // Hard pause shuts down the runtime completely. Resume happens naturally when | ||||
8973 | // OpenMP is used subsequently. | ||||
8974 | void __kmp_hard_pause() { | ||||
8975 | __kmp_pause_status = kmp_hard_paused; | ||||
8976 | __kmp_internal_end_thread(-1); | ||||
8977 | } | ||||
8978 | |||||
8979 | // Soft resume sets __kmp_pause_status, and wakes up all threads. | ||||
8980 | void __kmp_resume_if_soft_paused() { | ||||
8981 | if (__kmp_pause_status == kmp_soft_paused) { | ||||
8982 | __kmp_pause_status = kmp_not_paused; | ||||
8983 | |||||
8984 | for (int gtid = 1; gtid < __kmp_threads_capacity; ++gtid) { | ||||
8985 | kmp_info_t *thread = __kmp_threads[gtid]; | ||||
8986 | if (thread) { // Wake it if sleeping | ||||
8987 | kmp_flag_64<> fl(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, | ||||
8988 | thread); | ||||
8989 | if (fl.is_sleeping()) | ||||
8990 | fl.resume(gtid); | ||||
8991 | else if (__kmp_try_suspend_mx(thread)) { // got suspend lock | ||||
8992 | __kmp_unlock_suspend_mx(thread); // unlock it; it won't sleep | ||||
8993 | } else { // thread holds the lock and may sleep soon | ||||
8994 | do { // until either the thread sleeps, or we can get the lock | ||||
8995 | if (fl.is_sleeping()) { | ||||
8996 | fl.resume(gtid); | ||||
8997 | break; | ||||
8998 | } else if (__kmp_try_suspend_mx(thread)) { | ||||
8999 | __kmp_unlock_suspend_mx(thread); | ||||
9000 | break; | ||||
9001 | } | ||||
9002 | } while (1); | ||||
9003 | } | ||||
9004 | } | ||||
9005 | } | ||||
9006 | } | ||||
9007 | } | ||||
9008 | |||||
9009 | // This function is called via __kmpc_pause_resource. Returns 0 if successful. | ||||
9010 | // TODO: add warning messages | ||||
9011 | int __kmp_pause_resource(kmp_pause_status_t level) { | ||||
9012 | if (level == kmp_not_paused) { // requesting resume | ||||
9013 | if (__kmp_pause_status == kmp_not_paused) { | ||||
9014 | // error message about runtime not being paused, so can't resume | ||||
9015 | return 1; | ||||
9016 | } else { | ||||
9017 | KMP_DEBUG_ASSERT(__kmp_pause_status == kmp_soft_paused ||if (!(__kmp_pause_status == kmp_soft_paused || __kmp_pause_status == kmp_hard_paused)) { __kmp_debug_assert("__kmp_pause_status == kmp_soft_paused || __kmp_pause_status == kmp_hard_paused" , "openmp/runtime/src/kmp_runtime.cpp", 9018); } | ||||
9018 | __kmp_pause_status == kmp_hard_paused)if (!(__kmp_pause_status == kmp_soft_paused || __kmp_pause_status == kmp_hard_paused)) { __kmp_debug_assert("__kmp_pause_status == kmp_soft_paused || __kmp_pause_status == kmp_hard_paused" , "openmp/runtime/src/kmp_runtime.cpp", 9018); }; | ||||
9019 | __kmp_pause_status = kmp_not_paused; | ||||
9020 | return 0; | ||||
9021 | } | ||||
9022 | } else if (level == kmp_soft_paused) { // requesting soft pause | ||||
9023 | if (__kmp_pause_status != kmp_not_paused) { | ||||
9024 | // error message about already being paused | ||||
9025 | return 1; | ||||
9026 | } else { | ||||
9027 | __kmp_soft_pause(); | ||||
9028 | return 0; | ||||
9029 | } | ||||
9030 | } else if (level == kmp_hard_paused) { // requesting hard pause | ||||
9031 | if (__kmp_pause_status != kmp_not_paused) { | ||||
9032 | // error message about already being paused | ||||
9033 | return 1; | ||||
9034 | } else { | ||||
9035 | __kmp_hard_pause(); | ||||
9036 | return 0; | ||||
9037 | } | ||||
9038 | } else { | ||||
9039 | // error message about invalid level | ||||
9040 | return 1; | ||||
9041 | } | ||||
9042 | } | ||||
9043 | |||||
9044 | void __kmp_omp_display_env(int verbose) { | ||||
9045 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); | ||||
9046 | if (__kmp_init_serial == 0) | ||||
9047 | __kmp_do_serial_initialize(); | ||||
9048 | __kmp_display_env_impl(!verbose, verbose); | ||||
9049 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | ||||
9050 | } | ||||
9051 | |||||
9052 | // The team size is changing, so distributed barrier must be modified | ||||
9053 | void __kmp_resize_dist_barrier(kmp_team_t *team, int old_nthreads, | ||||
9054 | int new_nthreads) { | ||||
9055 | KMP_DEBUG_ASSERT(__kmp_barrier_release_pattern[bs_forkjoin_barrier] ==if (!(__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar )) { __kmp_debug_assert("__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar" , "openmp/runtime/src/kmp_runtime.cpp", 9056); } | ||||
9056 | bp_dist_bar)if (!(__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar )) { __kmp_debug_assert("__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar" , "openmp/runtime/src/kmp_runtime.cpp", 9056); }; | ||||
9057 | kmp_info_t **other_threads = team->t.t_threads; | ||||
9058 | |||||
9059 | // We want all the workers to stop waiting on the barrier while we adjust the | ||||
9060 | // size of the team. | ||||
9061 | for (int f = 1; f < old_nthreads; ++f) { | ||||
9062 | KMP_DEBUG_ASSERT(other_threads[f] != NULL)if (!(other_threads[f] != __null)) { __kmp_debug_assert("other_threads[f] != __null" , "openmp/runtime/src/kmp_runtime.cpp", 9062); }; | ||||
9063 | // Ignore threads that are already inactive or not present in the team | ||||
9064 | if (team->t.t_threads[f]->th.th_used_in_team.load() == 0) { | ||||
9065 | // teams construct causes thread_limit to get passed in, and some of | ||||
9066 | // those could be inactive; just ignore them | ||||
9067 | continue; | ||||
9068 | } | ||||
9069 | // If thread is transitioning still to in_use state, wait for it | ||||
9070 | if (team->t.t_threads[f]->th.th_used_in_team.load() == 3) { | ||||
9071 | while (team->t.t_threads[f]->th.th_used_in_team.load() == 3) | ||||
9072 | KMP_CPU_PAUSE()__kmp_x86_pause(); | ||||
9073 | } | ||||
9074 | // The thread should be in_use now | ||||
9075 | KMP_DEBUG_ASSERT(team->t.t_threads[f]->th.th_used_in_team.load() == 1)if (!(team->t.t_threads[f]->th.th_used_in_team.load() == 1)) { __kmp_debug_assert("team->t.t_threads[f]->th.th_used_in_team.load() == 1" , "openmp/runtime/src/kmp_runtime.cpp", 9075); }; | ||||
9076 | // Transition to unused state | ||||
9077 | team->t.t_threads[f]->th.th_used_in_team.store(2); | ||||
9078 | KMP_DEBUG_ASSERT(team->t.t_threads[f]->th.th_used_in_team.load() == 2)if (!(team->t.t_threads[f]->th.th_used_in_team.load() == 2)) { __kmp_debug_assert("team->t.t_threads[f]->th.th_used_in_team.load() == 2" , "openmp/runtime/src/kmp_runtime.cpp", 9078); }; | ||||
9079 | } | ||||
9080 | // Release all the workers | ||||
9081 | team->t.b->go_release(); | ||||
9082 | |||||
9083 | KMP_MFENCE()if (__builtin_expect(!!(!__kmp_cpuinfo.initialized), 0)) { __kmp_query_cpuid (&__kmp_cpuinfo); } if (__kmp_cpuinfo.flags.sse2) { __sync_synchronize (); }; | ||||
9084 | |||||
9085 | // Workers should see transition status 2 and move to 0; but may need to be | ||||
9086 | // woken up first | ||||
9087 | int count = old_nthreads - 1; | ||||
9088 | while (count > 0) { | ||||
9089 | count = old_nthreads - 1; | ||||
9090 | for (int f = 1; f < old_nthreads; ++f) { | ||||
9091 | if (other_threads[f]->th.th_used_in_team.load() != 0) { | ||||
9092 | if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME(2147483647)) { // Wake up the workers | ||||
9093 | kmp_atomic_flag_64<> *flag = (kmp_atomic_flag_64<> *)CCAST(const_cast<void *>(other_threads[f]->th.th_sleep_loc ) | ||||
9094 | void *, other_threads[f]->th.th_sleep_loc)const_cast<void *>(other_threads[f]->th.th_sleep_loc ); | ||||
9095 | __kmp_atomic_resume_64(other_threads[f]->th.th_info.ds.ds_gtid, flag); | ||||
9096 | } | ||||
9097 | } else { | ||||
9098 | KMP_DEBUG_ASSERT(team->t.t_threads[f]->th.th_used_in_team.load() == 0)if (!(team->t.t_threads[f]->th.th_used_in_team.load() == 0)) { __kmp_debug_assert("team->t.t_threads[f]->th.th_used_in_team.load() == 0" , "openmp/runtime/src/kmp_runtime.cpp", 9098); }; | ||||
9099 | count--; | ||||
9100 | } | ||||
9101 | } | ||||
9102 | } | ||||
9103 | // Now update the barrier size | ||||
9104 | team->t.b->update_num_threads(new_nthreads); | ||||
9105 | team->t.b->go_reset(); | ||||
9106 | } | ||||
9107 | |||||
9108 | void __kmp_add_threads_to_team(kmp_team_t *team, int new_nthreads) { | ||||
9109 | // Add the threads back to the team | ||||
9110 | KMP_DEBUG_ASSERT(team)if (!(team)) { __kmp_debug_assert("team", "openmp/runtime/src/kmp_runtime.cpp" , 9110); }; | ||||
9111 | // Threads were paused and pointed at th_used_in_team temporarily during a | ||||
9112 | // resize of the team. We're going to set th_used_in_team to 3 to indicate to | ||||
9113 | // the thread that it should transition itself back into the team. Then, if | ||||
9114 | // blocktime isn't infinite, the thread could be sleeping, so we send a resume | ||||
9115 | // to wake it up. | ||||
9116 | for (int f = 1; f < new_nthreads; ++f) { | ||||
9117 | KMP_DEBUG_ASSERT(team->t.t_threads[f])if (!(team->t.t_threads[f])) { __kmp_debug_assert("team->t.t_threads[f]" , "openmp/runtime/src/kmp_runtime.cpp", 9117); }; | ||||
9118 | KMP_COMPARE_AND_STORE_ACQ32(&(team->t.t_threads[f]->th.th_used_in_team), 0,__sync_bool_compare_and_swap((volatile kmp_uint32 *)(&(team ->t.t_threads[f]->th.th_used_in_team)), (kmp_uint32)(0) , (kmp_uint32)(3)) | ||||
9119 | 3)__sync_bool_compare_and_swap((volatile kmp_uint32 *)(&(team ->t.t_threads[f]->th.th_used_in_team)), (kmp_uint32)(0) , (kmp_uint32)(3)); | ||||
9120 | if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME(2147483647)) { // Wake up sleeping threads | ||||
9121 | __kmp_resume_32(team->t.t_threads[f]->th.th_info.ds.ds_gtid, | ||||
9122 | (kmp_flag_32<false, false> *)NULL__null); | ||||
9123 | } | ||||
9124 | } | ||||
9125 | // The threads should be transitioning to the team; when they are done, they | ||||
9126 | // should have set th_used_in_team to 1. This loop forces master to wait until | ||||
9127 | // all threads have moved into the team and are waiting in the barrier. | ||||
9128 | int count = new_nthreads - 1; | ||||
9129 | while (count > 0) { | ||||
9130 | count = new_nthreads - 1; | ||||
9131 | for (int f = 1; f < new_nthreads; ++f) { | ||||
9132 | if (team->t.t_threads[f]->th.th_used_in_team.load() == 1) { | ||||
9133 | count--; | ||||
9134 | } | ||||
9135 | } | ||||
9136 | } | ||||
9137 | } | ||||
9138 | |||||
9139 | // Globals and functions for hidden helper task | ||||
9140 | kmp_info_t **__kmp_hidden_helper_threads; | ||||
9141 | kmp_info_t *__kmp_hidden_helper_main_thread; | ||||
9142 | std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks; | ||||
9143 | #if KMP_OS_LINUX1 | ||||
9144 | kmp_int32 __kmp_hidden_helper_threads_num = 8; | ||||
9145 | kmp_int32 __kmp_enable_hidden_helper = TRUE(!0); | ||||
9146 | #else | ||||
9147 | kmp_int32 __kmp_hidden_helper_threads_num = 0; | ||||
9148 | kmp_int32 __kmp_enable_hidden_helper = FALSE0; | ||||
9149 | #endif | ||||
9150 | |||||
9151 | namespace { | ||||
9152 | std::atomic<kmp_int32> __kmp_hit_hidden_helper_threads_num; | ||||
9153 | |||||
9154 | void __kmp_hidden_helper_wrapper_fn(int *gtid, int *, ...) { | ||||
9155 | // This is an explicit synchronization on all hidden helper threads in case | ||||
9156 | // that when a regular thread pushes a hidden helper task to one hidden | ||||
9157 | // helper thread, the thread has not been awaken once since they're released | ||||
9158 | // by the main thread after creating the team. | ||||
9159 | KMP_ATOMIC_INC(&__kmp_hit_hidden_helper_threads_num)(&__kmp_hit_hidden_helper_threads_num)->fetch_add(1, std ::memory_order_acq_rel); | ||||
9160 | while (KMP_ATOMIC_LD_ACQ(&__kmp_hit_hidden_helper_threads_num)(&__kmp_hit_hidden_helper_threads_num)->load(std::memory_order_acquire ) != | ||||
9161 | __kmp_hidden_helper_threads_num) | ||||
9162 | ; | ||||
9163 | |||||
9164 | // If main thread, then wait for signal | ||||
9165 | if (__kmpc_master(nullptr, *gtid)) { | ||||
9166 | // First, unset the initial state and release the initial thread | ||||
9167 | TCW_4(__kmp_init_hidden_helper_threads, FALSE)(__kmp_init_hidden_helper_threads) = (0); | ||||
9168 | __kmp_hidden_helper_initz_release(); | ||||
9169 | __kmp_hidden_helper_main_thread_wait(); | ||||
9170 | // Now wake up all worker threads | ||||
9171 | for (int i = 1; i < __kmp_hit_hidden_helper_threads_num; ++i) { | ||||
9172 | __kmp_hidden_helper_worker_thread_signal(); | ||||
9173 | } | ||||
9174 | } | ||||
9175 | } | ||||
9176 | } // namespace | ||||
9177 | |||||
9178 | void __kmp_hidden_helper_threads_initz_routine() { | ||||
9179 | // Create a new root for hidden helper team/threads | ||||
9180 | const int gtid = __kmp_register_root(TRUE(!0)); | ||||
9181 | __kmp_hidden_helper_main_thread = __kmp_threads[gtid]; | ||||
9182 | __kmp_hidden_helper_threads = &__kmp_threads[gtid]; | ||||
9183 | __kmp_hidden_helper_main_thread->th.th_set_nproc = | ||||
9184 | __kmp_hidden_helper_threads_num; | ||||
9185 | |||||
9186 | KMP_ATOMIC_ST_REL(&__kmp_hit_hidden_helper_threads_num, 0)(&__kmp_hit_hidden_helper_threads_num)->store(0, std:: memory_order_release); | ||||
9187 | |||||
9188 | __kmpc_fork_call(nullptr, 0, __kmp_hidden_helper_wrapper_fn); | ||||
9189 | |||||
9190 | // Set the initialization flag to FALSE | ||||
9191 | TCW_SYNC_4(__kmp_init_hidden_helper, FALSE)(__kmp_init_hidden_helper) = (0); | ||||
9192 | |||||
9193 | __kmp_hidden_helper_threads_deinitz_release(); | ||||
9194 | } | ||||
9195 | |||||
9196 | /* Nesting Mode: | ||||
9197 | Set via KMP_NESTING_MODE, which takes an integer. | ||||
9198 | Note: we skip duplicate topology levels, and skip levels with only | ||||
9199 | one entity. | ||||
9200 | KMP_NESTING_MODE=0 is the default, and doesn't use nesting mode. | ||||
9201 | KMP_NESTING_MODE=1 sets as many nesting levels as there are distinct levels | ||||
9202 | in the topology, and initializes the number of threads at each of those | ||||
9203 | levels to the number of entities at each level, respectively, below the | ||||
9204 | entity at the parent level. | ||||
9205 | KMP_NESTING_MODE=N, where N>1, attempts to create up to N nesting levels, | ||||
9206 | but starts with nesting OFF -- max-active-levels-var is 1 -- and requires | ||||
9207 | the user to turn nesting on explicitly. This is an even more experimental | ||||
9208 | option to this experimental feature, and may change or go away in the | ||||
9209 | future. | ||||
9210 | */ | ||||
9211 | |||||
9212 | // Allocate space to store nesting levels | ||||
9213 | void __kmp_init_nesting_mode() { | ||||
9214 | int levels = KMP_HW_LAST; | ||||
9215 | __kmp_nesting_mode_nlevels = levels; | ||||
9216 | __kmp_nesting_nth_level = (int *)KMP_INTERNAL_MALLOC(levels * sizeof(int))malloc(levels * sizeof(int)); | ||||
9217 | for (int i = 0; i < levels; ++i) | ||||
9218 | __kmp_nesting_nth_level[i] = 0; | ||||
9219 | if (__kmp_nested_nth.size < levels) { | ||||
9220 | __kmp_nested_nth.nth = | ||||
9221 | (int *)KMP_INTERNAL_REALLOC(__kmp_nested_nth.nth, levels * sizeof(int))realloc((__kmp_nested_nth.nth), (levels * sizeof(int))); | ||||
9222 | __kmp_nested_nth.size = levels; | ||||
9223 | } | ||||
9224 | } | ||||
9225 | |||||
9226 | // Set # threads for top levels of nesting; must be called after topology set | ||||
9227 | void __kmp_set_nesting_mode_threads() { | ||||
9228 | kmp_info_t *thread = __kmp_threads[__kmp_entry_gtid()__kmp_get_global_thread_id_reg()]; | ||||
9229 | |||||
9230 | if (__kmp_nesting_mode == 1) | ||||
9231 | __kmp_nesting_mode_nlevels = KMP_MAX_ACTIVE_LEVELS_LIMIT2147483647; | ||||
9232 | else if (__kmp_nesting_mode > 1) | ||||
9233 | __kmp_nesting_mode_nlevels = __kmp_nesting_mode; | ||||
9234 | |||||
9235 | if (__kmp_topology) { // use topology info | ||||
9236 | int loc, hw_level; | ||||
9237 | for (loc = 0, hw_level = 0; hw_level < __kmp_topology->get_depth() && | ||||
9238 | loc < __kmp_nesting_mode_nlevels; | ||||
9239 | loc++, hw_level++) { | ||||
9240 | __kmp_nesting_nth_level[loc] = __kmp_topology->get_ratio(hw_level); | ||||
9241 | if (__kmp_nesting_nth_level[loc] == 1) | ||||
9242 | loc--; | ||||
9243 | } | ||||
9244 | // Make sure all cores are used | ||||
9245 | if (__kmp_nesting_mode > 1 && loc > 1) { | ||||
9246 | int core_level = __kmp_topology->get_level(KMP_HW_CORE); | ||||
9247 | int num_cores = __kmp_topology->get_count(core_level); | ||||
9248 | int upper_levels = 1; | ||||
9249 | for (int level = 0; level < loc - 1; ++level) | ||||
9250 | upper_levels *= __kmp_nesting_nth_level[level]; | ||||
9251 | if (upper_levels * __kmp_nesting_nth_level[loc - 1] < num_cores) | ||||
9252 | __kmp_nesting_nth_level[loc - 1] = | ||||
9253 | num_cores / __kmp_nesting_nth_level[loc - 2]; | ||||
9254 | } | ||||
9255 | __kmp_nesting_mode_nlevels = loc; | ||||
9256 | __kmp_nested_nth.used = __kmp_nesting_mode_nlevels; | ||||
9257 | } else { // no topology info available; provide a reasonable guesstimation | ||||
9258 | if (__kmp_avail_proc >= 4) { | ||||
9259 | __kmp_nesting_nth_level[0] = __kmp_avail_proc / 2; | ||||
9260 | __kmp_nesting_nth_level[1] = 2; | ||||
9261 | __kmp_nesting_mode_nlevels = 2; | ||||
9262 | } else { | ||||
9263 | __kmp_nesting_nth_level[0] = __kmp_avail_proc; | ||||
9264 | __kmp_nesting_mode_nlevels = 1; | ||||
9265 | } | ||||
9266 | __kmp_nested_nth.used = __kmp_nesting_mode_nlevels; | ||||
9267 | } | ||||
9268 | for (int i = 0; i < __kmp_nesting_mode_nlevels; ++i) { | ||||
9269 | __kmp_nested_nth.nth[i] = __kmp_nesting_nth_level[i]; | ||||
9270 | } | ||||
9271 | set__nproc(thread, __kmp_nesting_nth_level[0])(((thread)->th.th_current_task->td_icvs.nproc) = (__kmp_nesting_nth_level [0])); | ||||
9272 | if (__kmp_nesting_mode > 1 && __kmp_nesting_mode_nlevels > __kmp_nesting_mode) | ||||
9273 | __kmp_nesting_mode_nlevels = __kmp_nesting_mode; | ||||
9274 | if (get__max_active_levels(thread)((thread)->th.th_current_task->td_icvs.max_active_levels ) > 1) { | ||||
9275 | // if max levels was set, set nesting mode levels to same | ||||
9276 | __kmp_nesting_mode_nlevels = get__max_active_levels(thread)((thread)->th.th_current_task->td_icvs.max_active_levels ); | ||||
9277 | } | ||||
9278 | if (__kmp_nesting_mode == 1) // turn on nesting for this case only | ||||
9279 | set__max_active_levels(thread, __kmp_nesting_mode_nlevels)(((thread)->th.th_current_task->td_icvs.max_active_levels ) = (__kmp_nesting_mode_nlevels)); | ||||
9280 | } | ||||
9281 | |||||
9282 | // Empty symbols to export (see exports_so.txt) when feature is disabled | ||||
9283 | extern "C" { | ||||
9284 | #if !KMP_STATS_ENABLED0 | ||||
9285 | void __kmp_reset_stats() {} | ||||
9286 | #endif | ||||
9287 | #if !USE_DEBUGGER0 | ||||
9288 | int __kmp_omp_debug_struct_info = FALSE0; | ||||
9289 | int __kmp_debugging = FALSE0; | ||||
9290 | #endif | ||||
9291 | #if !USE_ITT_BUILD1 || !USE_ITT_NOTIFY1 | ||||
9292 | void __kmp_itt_fini_ittlib() {} | ||||
9293 | void __kmp_itt_init_ittlib() {} | ||||
9294 | #endif | ||||
9295 | } | ||||
9296 | |||||
9297 | // end of file |