File: | build/source/openmp/runtime/src/kmp_runtime.cpp |
Warning: | line 1451, column 9 1st function call argument is an uninitialized 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 KMP_DYNAMIC_LIB1 | |||
557 | #if KMP_OS_WINDOWS0 | |||
558 | ||||
559 | BOOL WINAPI DllMain(HINSTANCE hInstDLL, DWORD fdwReason, LPVOID lpReserved) { | |||
560 | //__kmp_acquire_bootstrap_lock( &__kmp_initz_lock ); | |||
561 | ||||
562 | switch (fdwReason) { | |||
563 | ||||
564 | case DLL_PROCESS_ATTACH: | |||
565 | KA_TRACE(10, ("DllMain: PROCESS_ATTACH\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("DllMain: PROCESS_ATTACH\n" ); }; | |||
566 | ||||
567 | return TRUE(!0); | |||
568 | ||||
569 | case DLL_PROCESS_DETACH: | |||
570 | 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()); }; | |||
571 | ||||
572 | // According to Windows* documentation for DllMain entry point: | |||
573 | // for DLL_PROCESS_DETACH, lpReserved is used for telling the difference: | |||
574 | // lpReserved == NULL when FreeLibrary() is called, | |||
575 | // lpReserved != NULL when the process is terminated. | |||
576 | // When FreeLibrary() is called, worker threads remain alive. So the | |||
577 | // runtime's state is consistent and executing proper shutdown is OK. | |||
578 | // When the process is terminated, worker threads have exited or been | |||
579 | // forcefully terminated by the OS and only the shutdown thread remains. | |||
580 | // This can leave the runtime in an inconsistent state. | |||
581 | // Hence, only attempt proper cleanup when FreeLibrary() is called. | |||
582 | // Otherwise, rely on OS to reclaim resources. | |||
583 | if (lpReserved == NULL__null) | |||
584 | __kmp_internal_end_library(__kmp_gtid_get_specific()); | |||
585 | ||||
586 | return TRUE(!0); | |||
587 | ||||
588 | case DLL_THREAD_ATTACH: | |||
589 | KA_TRACE(10, ("DllMain: THREAD_ATTACH\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("DllMain: THREAD_ATTACH\n" ); }; | |||
590 | ||||
591 | /* if we want to register new siblings all the time here call | |||
592 | * __kmp_get_gtid(); */ | |||
593 | return TRUE(!0); | |||
594 | ||||
595 | case DLL_THREAD_DETACH: | |||
596 | 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()); }; | |||
597 | ||||
598 | __kmp_internal_end_thread(__kmp_gtid_get_specific()); | |||
599 | return TRUE(!0); | |||
600 | } | |||
601 | ||||
602 | return TRUE(!0); | |||
603 | } | |||
604 | ||||
605 | #endif /* KMP_OS_WINDOWS */ | |||
606 | #endif /* KMP_DYNAMIC_LIB */ | |||
607 | ||||
608 | /* __kmp_parallel_deo -- Wait until it's our turn. */ | |||
609 | void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) { | |||
610 | int gtid = *gtid_ref; | |||
611 | #ifdef BUILD_PARALLEL_ORDERED1 | |||
612 | kmp_team_t *team = __kmp_team_from_gtid(gtid); | |||
613 | #endif /* BUILD_PARALLEL_ORDERED */ | |||
614 | ||||
615 | if (__kmp_env_consistency_check) { | |||
616 | if (__kmp_threads[gtid]->th.th_root->r.r_active) | |||
617 | #if KMP_USE_DYNAMIC_LOCK1 | |||
618 | __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL__null, 0); | |||
619 | #else | |||
620 | __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL__null); | |||
621 | #endif | |||
622 | } | |||
623 | #ifdef BUILD_PARALLEL_ORDERED1 | |||
624 | if (!team->t.t_serialized) { | |||
625 | KMP_MB(); | |||
626 | KMP_WAIT__kmp_wait_4(&team->t.t_ordered.dt.t_value, __kmp_tid_from_gtid(gtid), KMP_EQ__kmp_eq_4, | |||
627 | NULL__null); | |||
628 | KMP_MB(); | |||
629 | } | |||
630 | #endif /* BUILD_PARALLEL_ORDERED */ | |||
631 | } | |||
632 | ||||
633 | /* __kmp_parallel_dxo -- Signal the next task. */ | |||
634 | void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) { | |||
635 | int gtid = *gtid_ref; | |||
636 | #ifdef BUILD_PARALLEL_ORDERED1 | |||
637 | int tid = __kmp_tid_from_gtid(gtid); | |||
638 | kmp_team_t *team = __kmp_team_from_gtid(gtid); | |||
639 | #endif /* BUILD_PARALLEL_ORDERED */ | |||
640 | ||||
641 | if (__kmp_env_consistency_check) { | |||
642 | if (__kmp_threads[gtid]->th.th_root->r.r_active) | |||
643 | __kmp_pop_sync(gtid, ct_ordered_in_parallel, loc_ref); | |||
644 | } | |||
645 | #ifdef BUILD_PARALLEL_ORDERED1 | |||
646 | if (!team->t.t_serialized) { | |||
647 | KMP_MB(); /* Flush all pending memory write invalidates. */ | |||
648 | ||||
649 | /* use the tid of the next thread in this team */ | |||
650 | /* TODO replace with general release procedure */ | |||
651 | team->t.t_ordered.dt.t_value = ((tid + 1) % team->t.t_nproc); | |||
652 | ||||
653 | KMP_MB(); /* Flush all pending memory write invalidates. */ | |||
654 | } | |||
655 | #endif /* BUILD_PARALLEL_ORDERED */ | |||
656 | } | |||
657 | ||||
658 | /* ------------------------------------------------------------------------ */ | |||
659 | /* The BARRIER for a SINGLE process section is always explicit */ | |||
660 | ||||
661 | int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws) { | |||
662 | int status; | |||
663 | kmp_info_t *th; | |||
664 | kmp_team_t *team; | |||
665 | ||||
666 | if (!TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) | |||
667 | __kmp_parallel_initialize(); | |||
668 | __kmp_resume_if_soft_paused(); | |||
669 | ||||
670 | th = __kmp_threads[gtid]; | |||
671 | team = th->th.th_team; | |||
672 | status = 0; | |||
673 | ||||
674 | th->th.th_ident = id_ref; | |||
675 | ||||
676 | if (team->t.t_serialized) { | |||
677 | status = 1; | |||
678 | } else { | |||
679 | kmp_int32 old_this = th->th.th_local.this_construct; | |||
680 | ||||
681 | ++th->th.th_local.this_construct; | |||
682 | /* try to set team count to thread count--success means thread got the | |||
683 | single block */ | |||
684 | /* TODO: Should this be acquire or release? */ | |||
685 | if (team->t.t_construct == old_this) { | |||
686 | status = __kmp_atomic_compare_store_acq(&team->t.t_construct, old_this, | |||
687 | th->th.th_local.this_construct); | |||
688 | } | |||
689 | #if USE_ITT_BUILD1 | |||
690 | if (__itt_metadata_add_ptr__kmp_itt_metadata_add_ptr__3_0 && __kmp_forkjoin_frames_mode == 3 && | |||
691 | KMP_MASTER_GTID(gtid)(0 == __kmp_tid_from_gtid((gtid))) && th->th.th_teams_microtask == NULL__null && | |||
692 | team->t.t_active_level == 1) { | |||
693 | // Only report metadata by primary thread of active team at level 1 | |||
694 | __kmp_itt_metadata_single(id_ref); | |||
695 | } | |||
696 | #endif /* USE_ITT_BUILD */ | |||
697 | } | |||
698 | ||||
699 | if (__kmp_env_consistency_check) { | |||
700 | if (status && push_ws) { | |||
701 | __kmp_push_workshare(gtid, ct_psingle, id_ref); | |||
702 | } else { | |||
703 | __kmp_check_workshare(gtid, ct_psingle, id_ref); | |||
704 | } | |||
705 | } | |||
706 | #if USE_ITT_BUILD1 | |||
707 | if (status) { | |||
708 | __kmp_itt_single_start(gtid); | |||
709 | } | |||
710 | #endif /* USE_ITT_BUILD */ | |||
711 | return status; | |||
712 | } | |||
713 | ||||
714 | void __kmp_exit_single(int gtid) { | |||
715 | #if USE_ITT_BUILD1 | |||
716 | __kmp_itt_single_end(gtid); | |||
717 | #endif /* USE_ITT_BUILD */ | |||
718 | if (__kmp_env_consistency_check) | |||
719 | __kmp_pop_workshare(gtid, ct_psingle, NULL__null); | |||
720 | } | |||
721 | ||||
722 | /* determine if we can go parallel or must use a serialized parallel region and | |||
723 | * how many threads we can use | |||
724 | * set_nproc is the number of threads requested for the team | |||
725 | * returns 0 if we should serialize or only use one thread, | |||
726 | * otherwise the number of threads to use | |||
727 | * The forkjoin lock is held by the caller. */ | |||
728 | static int __kmp_reserve_threads(kmp_root_t *root, kmp_team_t *parent_team, | |||
729 | int master_tid, int set_nthreads, | |||
730 | int enter_teams) { | |||
731 | int capacity; | |||
732 | int new_nthreads; | |||
733 | KMP_DEBUG_ASSERT(__kmp_init_serial)if (!(__kmp_init_serial)) { __kmp_debug_assert("__kmp_init_serial" , "openmp/runtime/src/kmp_runtime.cpp", 733); }; | |||
734 | KMP_DEBUG_ASSERT(root && parent_team)if (!(root && parent_team)) { __kmp_debug_assert("root && parent_team" , "openmp/runtime/src/kmp_runtime.cpp", 734); }; | |||
735 | kmp_info_t *this_thr = parent_team->t.t_threads[master_tid]; | |||
736 | ||||
737 | // If dyn-var is set, dynamically adjust the number of desired threads, | |||
738 | // according to the method specified by dynamic_mode. | |||
739 | new_nthreads = set_nthreads; | |||
740 | if (!get__dynamic_2(parent_team, master_tid)((parent_team)->t.t_threads[(master_tid)]->th.th_current_task ->td_icvs.dynamic)) { | |||
741 | ; | |||
742 | } | |||
743 | #ifdef USE_LOAD_BALANCE1 | |||
744 | else if (__kmp_global.g.g_dynamic_mode == dynamic_load_balance) { | |||
745 | new_nthreads = __kmp_load_balance_nproc(root, set_nthreads); | |||
746 | if (new_nthreads == 1) { | |||
747 | 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); } | |||
748 | "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); } | |||
749 | 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); }; | |||
750 | return 1; | |||
751 | } | |||
752 | if (new_nthreads < set_nthreads) { | |||
753 | 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); } | |||
754 | "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); } | |||
755 | 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); }; | |||
756 | } | |||
757 | } | |||
758 | #endif /* USE_LOAD_BALANCE */ | |||
759 | else if (__kmp_global.g.g_dynamic_mode == dynamic_thread_limit) { | |||
760 | new_nthreads = __kmp_avail_proc - __kmp_nth + | |||
761 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc); | |||
762 | if (new_nthreads <= 1) { | |||
763 | 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); } | |||
764 | "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); } | |||
765 | 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); }; | |||
766 | return 1; | |||
767 | } | |||
768 | if (new_nthreads < set_nthreads) { | |||
769 | 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); } | |||
770 | "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); } | |||
771 | 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); }; | |||
772 | } else { | |||
773 | new_nthreads = set_nthreads; | |||
774 | } | |||
775 | } else if (__kmp_global.g.g_dynamic_mode == dynamic_random) { | |||
776 | if (set_nthreads > 2) { | |||
777 | new_nthreads = __kmp_get_random(parent_team->t.t_threads[master_tid]); | |||
778 | new_nthreads = (new_nthreads % set_nthreads) + 1; | |||
779 | if (new_nthreads == 1) { | |||
780 | 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); } | |||
781 | "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); } | |||
782 | 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); }; | |||
783 | return 1; | |||
784 | } | |||
785 | if (new_nthreads < set_nthreads) { | |||
786 | 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); } | |||
787 | "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); } | |||
788 | 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); }; | |||
789 | } | |||
790 | } | |||
791 | } else { | |||
792 | KMP_ASSERT(0)if (!(0)) { __kmp_debug_assert("0", "openmp/runtime/src/kmp_runtime.cpp" , 792); }; | |||
793 | } | |||
794 | ||||
795 | // Respect KMP_ALL_THREADS/KMP_DEVICE_THREAD_LIMIT. | |||
796 | if (__kmp_nth + new_nthreads - | |||
797 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) > | |||
798 | __kmp_max_nth) { | |||
799 | int tl_nthreads = __kmp_max_nth - __kmp_nth + | |||
800 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc); | |||
801 | if (tl_nthreads <= 0) { | |||
802 | tl_nthreads = 1; | |||
803 | } | |||
804 | ||||
805 | // If dyn-var is false, emit a 1-time warning. | |||
806 | 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)) { | |||
807 | __kmp_reserve_warn = 1; | |||
808 | __kmp_msg(kmp_ms_warning, | |||
809 | KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, set_nthreads, tl_nthreads ), | |||
810 | KMP_HNT(Unset_ALL_THREADS)__kmp_msg_format(kmp_i18n_hnt_Unset_ALL_THREADS), __kmp_msg_null); | |||
811 | } | |||
812 | if (tl_nthreads == 1) { | |||
813 | 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); } | |||
814 | "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); } | |||
815 | 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); }; | |||
816 | return 1; | |||
817 | } | |||
818 | 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); } | |||
819 | "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); } | |||
820 | 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); }; | |||
821 | new_nthreads = tl_nthreads; | |||
822 | } | |||
823 | ||||
824 | // Respect OMP_THREAD_LIMIT | |||
825 | int cg_nthreads = this_thr->th.th_cg_roots->cg_nthreads; | |||
826 | int max_cg_threads = this_thr->th.th_cg_roots->cg_thread_limit; | |||
827 | if (cg_nthreads + new_nthreads - | |||
828 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) > | |||
829 | max_cg_threads) { | |||
830 | int tl_nthreads = max_cg_threads - cg_nthreads + | |||
831 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc); | |||
832 | if (tl_nthreads <= 0) { | |||
833 | tl_nthreads = 1; | |||
834 | } | |||
835 | ||||
836 | // If dyn-var is false, emit a 1-time warning. | |||
837 | 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)) { | |||
838 | __kmp_reserve_warn = 1; | |||
839 | __kmp_msg(kmp_ms_warning, | |||
840 | KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, set_nthreads, tl_nthreads ), | |||
841 | KMP_HNT(Unset_ALL_THREADS)__kmp_msg_format(kmp_i18n_hnt_Unset_ALL_THREADS), __kmp_msg_null); | |||
842 | } | |||
843 | if (tl_nthreads == 1) { | |||
844 | 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); } | |||
845 | "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); } | |||
846 | 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); }; | |||
847 | return 1; | |||
848 | } | |||
849 | 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); } | |||
850 | "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); } | |||
851 | 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); }; | |||
852 | new_nthreads = tl_nthreads; | |||
853 | } | |||
854 | ||||
855 | // Check if the threads array is large enough, or needs expanding. | |||
856 | // See comment in __kmp_register_root() about the adjustment if | |||
857 | // __kmp_threads[0] == NULL. | |||
858 | capacity = __kmp_threads_capacity; | |||
859 | if (TCR_PTR(__kmp_threads[0])((void *)(__kmp_threads[0])) == NULL__null) { | |||
860 | --capacity; | |||
861 | } | |||
862 | // If it is not for initializing the hidden helper team, we need to take | |||
863 | // __kmp_hidden_helper_threads_num out of the capacity because it is included | |||
864 | // in __kmp_threads_capacity. | |||
865 | if (__kmp_enable_hidden_helper && !TCR_4(__kmp_init_hidden_helper_threads)(__kmp_init_hidden_helper_threads)) { | |||
866 | capacity -= __kmp_hidden_helper_threads_num; | |||
867 | } | |||
868 | if (__kmp_nth + new_nthreads - | |||
869 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) > | |||
870 | capacity) { | |||
871 | // Expand the threads array. | |||
872 | int slotsRequired = __kmp_nth + new_nthreads - | |||
873 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) - | |||
874 | capacity; | |||
875 | int slotsAdded = __kmp_expand_threads(slotsRequired); | |||
876 | if (slotsAdded < slotsRequired) { | |||
877 | // The threads array was not expanded enough. | |||
878 | new_nthreads -= (slotsRequired - slotsAdded); | |||
879 | KMP_ASSERT(new_nthreads >= 1)if (!(new_nthreads >= 1)) { __kmp_debug_assert("new_nthreads >= 1" , "openmp/runtime/src/kmp_runtime.cpp", 879); }; | |||
880 | ||||
881 | // If dyn-var is false, emit a 1-time warning. | |||
882 | 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)) { | |||
883 | __kmp_reserve_warn = 1; | |||
884 | if (__kmp_tp_cached) { | |||
885 | __kmp_msg(kmp_ms_warning, | |||
886 | KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, set_nthreads, new_nthreads ), | |||
887 | KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity)__kmp_msg_format(kmp_i18n_hnt_Set_ALL_THREADPRIVATE, __kmp_tp_capacity ), | |||
888 | KMP_HNT(PossibleSystemLimitOnThreads)__kmp_msg_format(kmp_i18n_hnt_PossibleSystemLimitOnThreads), __kmp_msg_null); | |||
889 | } else { | |||
890 | __kmp_msg(kmp_ms_warning, | |||
891 | KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, set_nthreads, new_nthreads ), | |||
892 | KMP_HNT(SystemLimitOnThreads)__kmp_msg_format(kmp_i18n_hnt_SystemLimitOnThreads), __kmp_msg_null); | |||
893 | } | |||
894 | } | |||
895 | } | |||
896 | } | |||
897 | ||||
898 | #ifdef KMP_DEBUG1 | |||
899 | if (new_nthreads == 1) { | |||
900 | 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); } | |||
901 | ("__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); } | |||
902 | "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); } | |||
903 | __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); }; | |||
904 | } else { | |||
905 | 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); } | |||
906 | " %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); } | |||
907 | __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); }; | |||
908 | } | |||
909 | #endif // KMP_DEBUG | |||
910 | return new_nthreads; | |||
911 | } | |||
912 | ||||
913 | /* Allocate threads from the thread pool and assign them to the new team. We are | |||
914 | assured that there are enough threads available, because we checked on that | |||
915 | earlier within critical section forkjoin */ | |||
916 | static void __kmp_fork_team_threads(kmp_root_t *root, kmp_team_t *team, | |||
917 | kmp_info_t *master_th, int master_gtid, | |||
918 | int fork_teams_workers) { | |||
919 | int i; | |||
920 | int use_hot_team; | |||
921 | ||||
922 | 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); }; | |||
923 | 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" , 923); }; | |||
924 | KMP_MB(); | |||
925 | ||||
926 | /* first, let's setup the primary thread */ | |||
927 | master_th->th.th_info.ds.ds_tid = 0; | |||
928 | master_th->th.th_team = team; | |||
929 | master_th->th.th_team_nproc = team->t.t_nproc; | |||
930 | master_th->th.th_team_master = master_th; | |||
931 | master_th->th.th_team_serialized = FALSE0; | |||
932 | master_th->th.th_dispatch = &team->t.t_dispatch[0]; | |||
933 | ||||
934 | /* make sure we are not the optimized hot team */ | |||
935 | #if KMP_NESTED_HOT_TEAMS1 | |||
936 | use_hot_team = 0; | |||
937 | kmp_hot_team_ptr_t *hot_teams = master_th->th.th_hot_teams; | |||
938 | if (hot_teams) { // hot teams array is not allocated if | |||
939 | // KMP_HOT_TEAMS_MAX_LEVEL=0 | |||
940 | int level = team->t.t_active_level - 1; // index in array of hot teams | |||
941 | if (master_th->th.th_teams_microtask) { // are we inside the teams? | |||
942 | if (master_th->th.th_teams_size.nteams > 1) { | |||
943 | ++level; // level was not increased in teams construct for | |||
944 | // team_of_masters | |||
945 | } | |||
946 | if (team->t.t_pkfn != (microtask_t)__kmp_teams_master && | |||
947 | master_th->th.th_teams_level == team->t.t_level) { | |||
948 | ++level; // level was not increased in teams construct for | |||
949 | // team_of_workers before the parallel | |||
950 | } // team->t.t_level will be increased inside parallel | |||
951 | } | |||
952 | if (level < __kmp_hot_teams_max_level) { | |||
953 | if (hot_teams[level].hot_team) { | |||
954 | // hot team has already been allocated for given level | |||
955 | 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" , 955); }; | |||
956 | use_hot_team = 1; // the team is ready to use | |||
957 | } else { | |||
958 | use_hot_team = 0; // AC: threads are not allocated yet | |||
959 | hot_teams[level].hot_team = team; // remember new hot team | |||
960 | hot_teams[level].hot_team_nth = team->t.t_nproc; | |||
961 | } | |||
962 | } else { | |||
963 | use_hot_team = 0; | |||
964 | } | |||
965 | } | |||
966 | #else | |||
967 | use_hot_team = team == root->r.r_hot_team; | |||
968 | #endif | |||
969 | if (!use_hot_team) { | |||
970 | ||||
971 | /* install the primary thread */ | |||
972 | team->t.t_threads[0] = master_th; | |||
973 | __kmp_initialize_info(master_th, team, 0, master_gtid); | |||
974 | ||||
975 | /* now, install the worker threads */ | |||
976 | for (i = 1; i < team->t.t_nproc; i++) { | |||
977 | ||||
978 | /* fork or reallocate a new thread and install it in team */ | |||
979 | kmp_info_t *thr = __kmp_allocate_thread(root, team, i); | |||
980 | team->t.t_threads[i] = thr; | |||
981 | KMP_DEBUG_ASSERT(thr)if (!(thr)) { __kmp_debug_assert("thr", "openmp/runtime/src/kmp_runtime.cpp" , 981); }; | |||
982 | 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", 982); }; | |||
983 | /* align team and thread arrived states */ | |||
984 | 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); } | |||
985 | "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); } | |||
986 | __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); } | |||
987 | __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); } | |||
988 | 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); } | |||
989 | 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); }; | |||
990 | thr->th.th_teams_microtask = master_th->th.th_teams_microtask; | |||
991 | thr->th.th_teams_level = master_th->th.th_teams_level; | |||
992 | thr->th.th_teams_size = master_th->th.th_teams_size; | |||
993 | { // Initialize threads' barrier data. | |||
994 | int b; | |||
995 | kmp_balign_t *balign = team->t.t_threads[i]->th.th_bar; | |||
996 | for (b = 0; b < bs_last_barrier; ++b) { | |||
997 | balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; | |||
998 | 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", 998); }; | |||
999 | #if USE_DEBUGGER0 | |||
1000 | balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; | |||
1001 | #endif | |||
1002 | } | |||
1003 | } | |||
1004 | } | |||
1005 | ||||
1006 | #if KMP_AFFINITY_SUPPORTED1 | |||
1007 | // Do not partition the places list for teams construct workers who | |||
1008 | // haven't actually been forked to do real work yet. This partitioning | |||
1009 | // will take place in the parallel region nested within the teams construct. | |||
1010 | if (!fork_teams_workers) { | |||
1011 | __kmp_partition_places(team); | |||
1012 | } | |||
1013 | #endif | |||
1014 | } | |||
1015 | ||||
1016 | if (__kmp_display_affinity && team->t.t_display_affinity != 1) { | |||
1017 | for (i = 0; i < team->t.t_nproc; i++) { | |||
1018 | kmp_info_t *thr = team->t.t_threads[i]; | |||
1019 | if (thr->th.th_prev_num_threads != team->t.t_nproc || | |||
1020 | thr->th.th_prev_level != team->t.t_level) { | |||
1021 | team->t.t_display_affinity = 1; | |||
1022 | break; | |||
1023 | } | |||
1024 | } | |||
1025 | } | |||
1026 | ||||
1027 | KMP_MB(); | |||
1028 | } | |||
1029 | ||||
1030 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 | |||
1031 | // Propagate any changes to the floating point control registers out to the team | |||
1032 | // We try to avoid unnecessary writes to the relevant cache line in the team | |||
1033 | // structure, so we don't make changes unless they are needed. | |||
1034 | inline static void propagateFPControl(kmp_team_t *team) { | |||
1035 | if (__kmp_inherit_fp_control) { | |||
1036 | kmp_int16 x87_fpu_control_word; | |||
1037 | kmp_uint32 mxcsr; | |||
1038 | ||||
1039 | // Get primary thread's values of FPU control flags (both X87 and vector) | |||
1040 | __kmp_store_x87_fpu_control_word(&x87_fpu_control_word); | |||
1041 | __kmp_store_mxcsr(&mxcsr); | |||
1042 | mxcsr &= KMP_X86_MXCSR_MASK0xffffffc0; | |||
1043 | ||||
1044 | // There is no point looking at t_fp_control_saved here. | |||
1045 | // If it is TRUE, we still have to update the values if they are different | |||
1046 | // from those we now have. If it is FALSE we didn't save anything yet, but | |||
1047 | // our objective is the same. We have to ensure that the values in the team | |||
1048 | // are the same as those we have. | |||
1049 | // So, this code achieves what we need whether or not t_fp_control_saved is | |||
1050 | // true. By checking whether the value needs updating we avoid unnecessary | |||
1051 | // writes that would put the cache-line into a written state, causing all | |||
1052 | // threads in the team to have to read it again. | |||
1053 | 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 ); | |||
1054 | KMP_CHECK_UPDATE(team->t.t_mxcsr, mxcsr)if ((team->t.t_mxcsr) != (mxcsr)) (team->t.t_mxcsr) = ( mxcsr); | |||
1055 | // Although we don't use this value, other code in the runtime wants to know | |||
1056 | // whether it should restore them. So we must ensure it is correct. | |||
1057 | 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)); | |||
1058 | } else { | |||
1059 | // Similarly here. Don't write to this cache-line in the team structure | |||
1060 | // unless we have to. | |||
1061 | 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); | |||
1062 | } | |||
1063 | } | |||
1064 | ||||
1065 | // Do the opposite, setting the hardware registers to the updated values from | |||
1066 | // the team. | |||
1067 | inline static void updateHWFPControl(kmp_team_t *team) { | |||
1068 | if (__kmp_inherit_fp_control && team->t.t_fp_control_saved) { | |||
1069 | // Only reset the fp control regs if they have been changed in the team. | |||
1070 | // the parallel region that we are exiting. | |||
1071 | kmp_int16 x87_fpu_control_word; | |||
1072 | kmp_uint32 mxcsr; | |||
1073 | __kmp_store_x87_fpu_control_word(&x87_fpu_control_word); | |||
1074 | __kmp_store_mxcsr(&mxcsr); | |||
1075 | mxcsr &= KMP_X86_MXCSR_MASK0xffffffc0; | |||
1076 | ||||
1077 | if (team->t.t_x87_fpu_control_word != x87_fpu_control_word) { | |||
1078 | __kmp_clear_x87_fpu_status_word(); | |||
1079 | __kmp_load_x87_fpu_control_word(&team->t.t_x87_fpu_control_word); | |||
1080 | } | |||
1081 | ||||
1082 | if (team->t.t_mxcsr != mxcsr) { | |||
1083 | __kmp_load_mxcsr(&team->t.t_mxcsr); | |||
1084 | } | |||
1085 | } | |||
1086 | } | |||
1087 | #else | |||
1088 | #define propagateFPControl(x) ((void)0) | |||
1089 | #define updateHWFPControl(x) ((void)0) | |||
1090 | #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ | |||
1091 | ||||
1092 | static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team, | |||
1093 | int realloc); // forward declaration | |||
1094 | ||||
1095 | /* Run a parallel region that has been serialized, so runs only in a team of the | |||
1096 | single primary thread. */ | |||
1097 | void __kmp_serialized_parallel(ident_t *loc, kmp_int32 global_tid) { | |||
1098 | kmp_info_t *this_thr; | |||
1099 | kmp_team_t *serial_team; | |||
1100 | ||||
1101 | 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); }; | |||
1102 | ||||
1103 | /* Skip all this code for autopar serialized loops since it results in | |||
1104 | unacceptable overhead */ | |||
1105 | if (loc != NULL__null && (loc->flags & KMP_IDENT_AUTOPAR)) | |||
1106 | return; | |||
1107 | ||||
1108 | if (!TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) | |||
1109 | __kmp_parallel_initialize(); | |||
1110 | __kmp_resume_if_soft_paused(); | |||
1111 | ||||
1112 | this_thr = __kmp_threads[global_tid]; | |||
1113 | serial_team = this_thr->th.th_serial_team; | |||
1114 | ||||
1115 | /* utilize the serialized team held by this thread */ | |||
1116 | KMP_DEBUG_ASSERT(serial_team)if (!(serial_team)) { __kmp_debug_assert("serial_team", "openmp/runtime/src/kmp_runtime.cpp" , 1116); }; | |||
1117 | KMP_MB(); | |||
1118 | ||||
1119 | if (__kmp_tasking_mode != tskm_immediate_exec) { | |||
1120 | 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", 1122); } | |||
1121 | 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", 1122); } | |||
1122 | 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", 1122); }; | |||
1123 | 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", 1124); } | |||
1124 | 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", 1124); }; | |||
1125 | 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); } | |||
1126 | "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); } | |||
1127 | 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); }; | |||
1128 | this_thr->th.th_task_team = NULL__null; | |||
1129 | } | |||
1130 | ||||
1131 | kmp_proc_bind_t proc_bind = this_thr->th.th_set_proc_bind; | |||
1132 | if (this_thr->th.th_current_task->td_icvs.proc_bind == proc_bind_false) { | |||
1133 | proc_bind = proc_bind_false; | |||
1134 | } else if (proc_bind == proc_bind_default) { | |||
1135 | // No proc_bind clause was specified, so use the current value | |||
1136 | // of proc-bind-var for this parallel region. | |||
1137 | proc_bind = this_thr->th.th_current_task->td_icvs.proc_bind; | |||
1138 | } | |||
1139 | // Reset for next parallel region | |||
1140 | this_thr->th.th_set_proc_bind = proc_bind_default; | |||
1141 | ||||
1142 | #if OMPT_SUPPORT1 | |||
1143 | ompt_data_t ompt_parallel_data = ompt_data_none{0}; | |||
1144 | void *codeptr = OMPT_LOAD_RETURN_ADDRESS(global_tid)__ompt_load_return_address(global_tid); | |||
1145 | if (ompt_enabled.enabled && | |||
1146 | this_thr->th.ompt_thread_info.state != ompt_state_overhead) { | |||
1147 | ||||
1148 | ompt_task_info_t *parent_task_info; | |||
1149 | parent_task_info = OMPT_CUR_TASK_INFO(this_thr)(&(this_thr->th.th_current_task->ompt_task_info)); | |||
1150 | ||||
1151 | parent_task_info->frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0)__builtin_frame_address(0); | |||
1152 | if (ompt_enabled.ompt_callback_parallel_begin) { | |||
1153 | int team_size = 1; | |||
1154 | ||||
1155 | ompt_callbacks.ompt_callback(ompt_callback_parallel_begin)ompt_callback_parallel_begin_callback( | |||
1156 | &(parent_task_info->task_data), &(parent_task_info->frame), | |||
1157 | &ompt_parallel_data, team_size, | |||
1158 | ompt_parallel_invoker_program | ompt_parallel_team, codeptr); | |||
1159 | } | |||
1160 | } | |||
1161 | #endif // OMPT_SUPPORT | |||
1162 | ||||
1163 | if (this_thr->th.th_team != serial_team) { | |||
1164 | // Nested level will be an index in the nested nthreads array | |||
1165 | int level = this_thr->th.th_team->t.t_level; | |||
1166 | ||||
1167 | if (serial_team->t.t_serialized) { | |||
1168 | /* this serial team was already used | |||
1169 | TODO increase performance by making this locks more specific */ | |||
1170 | kmp_team_t *new_team; | |||
1171 | ||||
1172 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); | |||
1173 | ||||
1174 | new_team = | |||
1175 | __kmp_allocate_team(this_thr->th.th_root, 1, 1, | |||
1176 | #if OMPT_SUPPORT1 | |||
1177 | ompt_parallel_data, | |||
1178 | #endif | |||
1179 | proc_bind, &this_thr->th.th_current_task->td_icvs, | |||
1180 | 0 USE_NESTED_HOT_ARG(NULL), __null); | |||
1181 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | |||
1182 | KMP_ASSERT(new_team)if (!(new_team)) { __kmp_debug_assert("new_team", "openmp/runtime/src/kmp_runtime.cpp" , 1182); }; | |||
1183 | ||||
1184 | /* setup new serialized team and install it */ | |||
1185 | new_team->t.t_threads[0] = this_thr; | |||
1186 | new_team->t.t_parent = this_thr->th.th_team; | |||
1187 | serial_team = new_team; | |||
1188 | this_thr->th.th_serial_team = serial_team; | |||
1189 | ||||
1190 | 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); } | |||
1191 | 10,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d allocated new serial team %p\n" , global_tid, serial_team); } | |||
1192 | ("__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); } | |||
1193 | 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); }; | |||
1194 | ||||
1195 | /* TODO the above breaks the requirement that if we run out of resources, | |||
1196 | then we can still guarantee that serialized teams are ok, since we may | |||
1197 | need to allocate a new one */ | |||
1198 | } else { | |||
1199 | 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); } | |||
1200 | 10,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d reusing cached serial team %p\n" , global_tid, serial_team); } | |||
1201 | ("__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); } | |||
1202 | 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); }; | |||
1203 | } | |||
1204 | ||||
1205 | /* we have to initialize this serial team */ | |||
1206 | 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", 1206); }; | |||
1207 | 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" , 1207); }; | |||
1208 | 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" , 1208); }; | |||
1209 | serial_team->t.t_ident = loc; | |||
1210 | serial_team->t.t_serialized = 1; | |||
1211 | serial_team->t.t_nproc = 1; | |||
1212 | serial_team->t.t_parent = this_thr->th.th_team; | |||
1213 | serial_team->t.t_sched.sched = this_thr->th.th_team->t.t_sched.sched; | |||
1214 | this_thr->th.th_team = serial_team; | |||
1215 | serial_team->t.t_master_tid = this_thr->th.th_info.ds.ds_tid; | |||
1216 | ||||
1217 | 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); } | |||
1218 | 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); }; | |||
1219 | 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", 1219); }; | |||
1220 | this_thr->th.th_current_task->td_flags.executing = 0; | |||
1221 | ||||
1222 | __kmp_push_current_task_to_thread(this_thr, serial_team, 0); | |||
1223 | ||||
1224 | /* TODO: GEH: do ICVs work for nested serialized teams? Don't we need an | |||
1225 | implicit task for each serialized task represented by | |||
1226 | team->t.t_serialized? */ | |||
1227 | copy_icvs(&this_thr->th.th_current_task->td_icvs, | |||
1228 | &this_thr->th.th_current_task->td_parent->td_icvs); | |||
1229 | ||||
1230 | // Thread value exists in the nested nthreads array for the next nested | |||
1231 | // level | |||
1232 | if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) { | |||
1233 | this_thr->th.th_current_task->td_icvs.nproc = | |||
1234 | __kmp_nested_nth.nth[level + 1]; | |||
1235 | } | |||
1236 | ||||
1237 | if (__kmp_nested_proc_bind.used && | |||
1238 | (level + 1 < __kmp_nested_proc_bind.used)) { | |||
1239 | this_thr->th.th_current_task->td_icvs.proc_bind = | |||
1240 | __kmp_nested_proc_bind.bind_types[level + 1]; | |||
1241 | } | |||
1242 | ||||
1243 | #if USE_DEBUGGER0 | |||
1244 | serial_team->t.t_pkfn = (microtask_t)(~0); // For the debugger. | |||
1245 | #endif | |||
1246 | this_thr->th.th_info.ds.ds_tid = 0; | |||
1247 | ||||
1248 | /* set thread cache values */ | |||
1249 | this_thr->th.th_team_nproc = 1; | |||
1250 | this_thr->th.th_team_master = this_thr; | |||
1251 | this_thr->th.th_team_serialized = 1; | |||
1252 | ||||
1253 | serial_team->t.t_level = serial_team->t.t_parent->t.t_level + 1; | |||
1254 | serial_team->t.t_active_level = serial_team->t.t_parent->t.t_active_level; | |||
1255 | serial_team->t.t_def_allocator = this_thr->th.th_def_allocator; // save | |||
1256 | ||||
1257 | propagateFPControl(serial_team); | |||
1258 | ||||
1259 | /* check if we need to allocate dispatch buffers stack */ | |||
1260 | 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", 1260); }; | |||
1261 | if (!serial_team->t.t_dispatch->th_disp_buffer) { | |||
1262 | serial_team->t.t_dispatch->th_disp_buffer = | |||
1263 | (dispatch_private_info_t *)__kmp_allocate(___kmp_allocate((sizeof(dispatch_private_info_t)), "openmp/runtime/src/kmp_runtime.cpp" , 1264) | |||
1264 | sizeof(dispatch_private_info_t))___kmp_allocate((sizeof(dispatch_private_info_t)), "openmp/runtime/src/kmp_runtime.cpp" , 1264); | |||
1265 | } | |||
1266 | this_thr->th.th_dispatch = serial_team->t.t_dispatch; | |||
1267 | ||||
1268 | KMP_MB(); | |||
1269 | ||||
1270 | } else { | |||
1271 | /* this serialized team is already being used, | |||
1272 | * that's fine, just add another nested level */ | |||
1273 | 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" , 1273); }; | |||
1274 | 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", 1274); }; | |||
1275 | 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" , 1275); }; | |||
1276 | ++serial_team->t.t_serialized; | |||
1277 | this_thr->th.th_team_serialized = serial_team->t.t_serialized; | |||
1278 | ||||
1279 | // Nested level will be an index in the nested nthreads array | |||
1280 | int level = this_thr->th.th_team->t.t_level; | |||
1281 | // Thread value exists in the nested nthreads array for the next nested | |||
1282 | // level | |||
1283 | if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) { | |||
1284 | this_thr->th.th_current_task->td_icvs.nproc = | |||
1285 | __kmp_nested_nth.nth[level + 1]; | |||
1286 | } | |||
1287 | serial_team->t.t_level++; | |||
1288 | 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); } | |||
1289 | "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); } | |||
1290 | 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); }; | |||
1291 | ||||
1292 | /* allocate/push dispatch buffers stack */ | |||
1293 | 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", 1293); }; | |||
1294 | { | |||
1295 | dispatch_private_info_t *disp_buffer = | |||
1296 | (dispatch_private_info_t *)__kmp_allocate(___kmp_allocate((sizeof(dispatch_private_info_t)), "openmp/runtime/src/kmp_runtime.cpp" , 1297) | |||
1297 | sizeof(dispatch_private_info_t))___kmp_allocate((sizeof(dispatch_private_info_t)), "openmp/runtime/src/kmp_runtime.cpp" , 1297); | |||
1298 | disp_buffer->next = serial_team->t.t_dispatch->th_disp_buffer; | |||
1299 | serial_team->t.t_dispatch->th_disp_buffer = disp_buffer; | |||
1300 | } | |||
1301 | this_thr->th.th_dispatch = serial_team->t.t_dispatch; | |||
1302 | ||||
1303 | KMP_MB(); | |||
1304 | } | |||
1305 | 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); | |||
1306 | ||||
1307 | // Perform the display affinity functionality for | |||
1308 | // serialized parallel regions | |||
1309 | if (__kmp_display_affinity) { | |||
1310 | if (this_thr->th.th_prev_level != serial_team->t.t_level || | |||
1311 | this_thr->th.th_prev_num_threads != 1) { | |||
1312 | // NULL means use the affinity-format-var ICV | |||
1313 | __kmp_aux_display_affinity(global_tid, NULL__null); | |||
1314 | this_thr->th.th_prev_level = serial_team->t.t_level; | |||
1315 | this_thr->th.th_prev_num_threads = 1; | |||
1316 | } | |||
1317 | } | |||
1318 | ||||
1319 | if (__kmp_env_consistency_check) | |||
1320 | __kmp_push_parallel(global_tid, NULL__null); | |||
1321 | #if OMPT_SUPPORT1 | |||
1322 | serial_team->t.ompt_team_info.master_return_address = codeptr; | |||
1323 | if (ompt_enabled.enabled && | |||
1324 | this_thr->th.ompt_thread_info.state != ompt_state_overhead) { | |||
1325 | OMPT_CUR_TASK_INFO(this_thr)(&(this_thr->th.th_current_task->ompt_task_info))->frame.exit_frame.ptr = | |||
1326 | OMPT_GET_FRAME_ADDRESS(0)__builtin_frame_address(0); | |||
1327 | ||||
1328 | ompt_lw_taskteam_t lw_taskteam; | |||
1329 | __ompt_lw_taskteam_init(&lw_taskteam, this_thr, global_tid, | |||
1330 | &ompt_parallel_data, codeptr); | |||
1331 | ||||
1332 | __ompt_lw_taskteam_link(&lw_taskteam, this_thr, 1); | |||
1333 | // don't use lw_taskteam after linking. content was swaped | |||
1334 | ||||
1335 | /* OMPT implicit task begin */ | |||
1336 | if (ompt_enabled.ompt_callback_implicit_task) { | |||
1337 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | |||
1338 | ompt_scope_begin, OMPT_CUR_TEAM_DATA(this_thr)(&(this_thr->th.th_team->t.ompt_team_info.parallel_data )), | |||
1339 | OMPT_CUR_TASK_DATA(this_thr)(&(this_thr->th.th_current_task->ompt_task_info.task_data )), 1, __kmp_tid_from_gtid(global_tid), | |||
1340 | ompt_task_implicit); // TODO: Can this be ompt_task_initial? | |||
1341 | OMPT_CUR_TASK_INFO(this_thr)(&(this_thr->th.th_current_task->ompt_task_info))->thread_num = | |||
1342 | __kmp_tid_from_gtid(global_tid); | |||
1343 | } | |||
1344 | ||||
1345 | /* OMPT state */ | |||
1346 | this_thr->th.ompt_thread_info.state = ompt_state_work_parallel; | |||
1347 | OMPT_CUR_TASK_INFO(this_thr)(&(this_thr->th.th_current_task->ompt_task_info))->frame.exit_frame.ptr = | |||
1348 | OMPT_GET_FRAME_ADDRESS(0)__builtin_frame_address(0); | |||
1349 | } | |||
1350 | #endif | |||
1351 | } | |||
1352 | ||||
1353 | /* most of the work for a fork */ | |||
1354 | /* return true if we really went parallel, false if serialized */ | |||
1355 | int __kmp_fork_call(ident_t *loc, int gtid, | |||
1356 | enum fork_context_e call_context, // Intel, GNU, ... | |||
1357 | kmp_int32 argc, microtask_t microtask, launch_t invoker, | |||
1358 | kmp_va_list ap) { | |||
1359 | void **argv; | |||
1360 | int i; | |||
1361 | int master_tid; | |||
1362 | int master_this_cons; | |||
1363 | kmp_team_t *team; | |||
1364 | kmp_team_t *parent_team; | |||
1365 | kmp_info_t *master_th; | |||
1366 | kmp_root_t *root; | |||
1367 | int nthreads; | |||
1368 | int master_active; | |||
1369 | int master_set_numthreads; | |||
1370 | int level; | |||
1371 | int active_level; | |||
1372 | int teams_level; | |||
1373 | #if KMP_NESTED_HOT_TEAMS1 | |||
1374 | kmp_hot_team_ptr_t **p_hot_teams; | |||
1375 | #endif | |||
1376 | { // KMP_TIME_BLOCK | |||
1377 | KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_fork_call)((void)0); | |||
1378 | KMP_COUNT_VALUE(OMP_PARALLEL_args, argc)((void)0); | |||
1379 | ||||
1380 | 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); }; | |||
| ||||
1381 | if (__kmp_stkpadding > 0 && __kmp_root[gtid] != NULL__null) { | |||
1382 | /* Some systems prefer the stack for the root thread(s) to start with */ | |||
1383 | /* some gap from the parent stack to prevent false sharing. */ | |||
1384 | void *dummy = KMP_ALLOCA(__kmp_stkpadding)__builtin_alloca (__kmp_stkpadding); | |||
1385 | /* These 2 lines below are so this does not get optimized out */ | |||
1386 | if (__kmp_stkpadding > KMP_MAX_STKPADDING(2 * 1024 * 1024)) | |||
1387 | __kmp_stkpadding += (short)((kmp_int64)dummy); | |||
1388 | } | |||
1389 | ||||
1390 | /* initialize if needed */ | |||
1391 | KMP_DEBUG_ASSERT(if (!(__kmp_init_serial)) { __kmp_debug_assert("__kmp_init_serial" , "openmp/runtime/src/kmp_runtime.cpp", 1392); } | |||
1392 | __kmp_init_serial)if (!(__kmp_init_serial)) { __kmp_debug_assert("__kmp_init_serial" , "openmp/runtime/src/kmp_runtime.cpp", 1392); }; // AC: potentially unsafe, not in sync with shutdown | |||
1393 | if (!TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) | |||
1394 | __kmp_parallel_initialize(); | |||
1395 | __kmp_resume_if_soft_paused(); | |||
1396 | ||||
1397 | /* setup current data */ | |||
1398 | master_th = __kmp_threads[gtid]; // AC: potentially unsafe, not in sync with | |||
1399 | // shutdown | |||
1400 | parent_team = master_th->th.th_team; | |||
1401 | master_tid = master_th->th.th_info.ds.ds_tid; | |||
1402 | master_this_cons = master_th->th.th_local.this_construct; | |||
1403 | root = master_th->th.th_root; | |||
1404 | master_active = root->r.r_active; | |||
1405 | master_set_numthreads = master_th->th.th_set_nproc; | |||
1406 | ||||
1407 | #if OMPT_SUPPORT1 | |||
1408 | ompt_data_t ompt_parallel_data = ompt_data_none{0}; | |||
1409 | ompt_data_t *parent_task_data; | |||
1410 | ompt_frame_t *ompt_frame; | |||
1411 | ompt_data_t *implicit_task_data; | |||
1412 | void *return_address = NULL__null; | |||
1413 | ||||
1414 | if (ompt_enabled.enabled) { | |||
1415 | __ompt_get_task_info_internal(0, NULL__null, &parent_task_data, &ompt_frame, | |||
1416 | NULL__null, NULL__null); | |||
1417 | return_address = OMPT_LOAD_RETURN_ADDRESS(gtid)__ompt_load_return_address(gtid); | |||
1418 | } | |||
1419 | #endif | |||
1420 | ||||
1421 | // Assign affinity to root thread if it hasn't happened yet | |||
1422 | __kmp_assign_root_init_mask(); | |||
1423 | ||||
1424 | // Nested level will be an index in the nested nthreads array | |||
1425 | level = parent_team->t.t_level; | |||
1426 | // used to launch non-serial teams even if nested is not allowed | |||
1427 | active_level = parent_team->t.t_active_level; | |||
1428 | // needed to check nesting inside the teams | |||
1429 | teams_level = master_th->th.th_teams_level; | |||
1430 | #if KMP_NESTED_HOT_TEAMS1 | |||
1431 | p_hot_teams = &master_th->th.th_hot_teams; | |||
1432 | if (*p_hot_teams == NULL__null && __kmp_hot_teams_max_level > 0) { | |||
1433 | *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", 1434) | |||
1434 | 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", 1434); | |||
1435 | (*p_hot_teams)[0].hot_team = root->r.r_hot_team; | |||
1436 | // it is either actual or not needed (when active_level > 0) | |||
1437 | (*p_hot_teams)[0].hot_team_nth = 1; | |||
1438 | } | |||
1439 | #endif | |||
1440 | ||||
1441 | #if OMPT_SUPPORT1 | |||
1442 | if (ompt_enabled.enabled) { | |||
1443 | if (ompt_enabled.ompt_callback_parallel_begin) { | |||
1444 | int team_size = master_set_numthreads | |||
1445 | ? master_set_numthreads | |||
1446 | : get__nproc_2(parent_team, master_tid)((parent_team)->t.t_threads[(master_tid)]->th.th_current_task ->td_icvs.nproc); | |||
1447 | int flags = OMPT_INVOKER(call_context)((call_context == fork_context_gnu) ? ompt_parallel_invoker_program : ompt_parallel_invoker_runtime) | | |||
1448 | ((microtask == (microtask_t)__kmp_teams_master) | |||
1449 | ? ompt_parallel_league | |||
1450 | : ompt_parallel_team); | |||
1451 | ompt_callbacks.ompt_callback(ompt_callback_parallel_begin)ompt_callback_parallel_begin_callback( | |||
| ||||
1452 | parent_task_data, ompt_frame, &ompt_parallel_data, team_size, flags, | |||
1453 | return_address); | |||
1454 | } | |||
1455 | master_th->th.ompt_thread_info.state = ompt_state_overhead; | |||
1456 | } | |||
1457 | #endif | |||
1458 | ||||
1459 | master_th->th.th_ident = loc; | |||
1460 | ||||
1461 | if (master_th->th.th_teams_microtask && ap && | |||
1462 | microtask != (microtask_t)__kmp_teams_master && level == teams_level) { | |||
1463 | // AC: This is start of parallel that is nested inside teams construct. | |||
1464 | // The team is actual (hot), all workers are ready at the fork barrier. | |||
1465 | // No lock needed to initialize the team a bit, then free workers. | |||
1466 | parent_team->t.t_ident = loc; | |||
1467 | __kmp_alloc_argv_entries(argc, parent_team, TRUE(!0)); | |||
1468 | parent_team->t.t_argc = argc; | |||
1469 | argv = (void **)parent_team->t.t_argv; | |||
1470 | for (i = argc - 1; i >= 0; --i) | |||
1471 | *argv++ = va_arg(kmp_va_deref(ap), void *)__builtin_va_arg((*(ap)), void *); | |||
1472 | // Increment our nested depth levels, but not increase the serialization | |||
1473 | if (parent_team == master_th->th.th_serial_team) { | |||
1474 | // AC: we are in serialized parallel | |||
1475 | __kmpc_serialized_parallel(loc, gtid); | |||
1476 | 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" , 1476); }; | |||
1477 | ||||
1478 | if (call_context == fork_context_gnu) { | |||
1479 | // AC: need to decrement t_serialized for enquiry functions to work | |||
1480 | // correctly, will restore at join time | |||
1481 | parent_team->t.t_serialized--; | |||
1482 | return TRUE(!0); | |||
1483 | } | |||
1484 | ||||
1485 | #if OMPD_SUPPORT1 | |||
1486 | parent_team->t.t_pkfn = microtask; | |||
1487 | #endif | |||
1488 | ||||
1489 | #if OMPT_SUPPORT1 | |||
1490 | void *dummy; | |||
1491 | void **exit_frame_p; | |||
1492 | ||||
1493 | ompt_lw_taskteam_t lw_taskteam; | |||
1494 | ||||
1495 | if (ompt_enabled.enabled) { | |||
1496 | __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid, | |||
1497 | &ompt_parallel_data, return_address); | |||
1498 | exit_frame_p = &(lw_taskteam.ompt_task_info.frame.exit_frame.ptr); | |||
1499 | ||||
1500 | __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0); | |||
1501 | // don't use lw_taskteam after linking. content was swaped | |||
1502 | ||||
1503 | /* OMPT implicit task begin */ | |||
1504 | implicit_task_data = OMPT_CUR_TASK_DATA(master_th)(&(master_th->th.th_current_task->ompt_task_info.task_data )); | |||
1505 | if (ompt_enabled.ompt_callback_implicit_task) { | |||
1506 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num = | |||
1507 | __kmp_tid_from_gtid(gtid); | |||
1508 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | |||
1509 | ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th)(&(master_th->th.th_team->t.ompt_team_info.parallel_data )), | |||
1510 | implicit_task_data, 1, | |||
1511 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num, ompt_task_implicit); | |||
1512 | } | |||
1513 | ||||
1514 | /* OMPT state */ | |||
1515 | master_th->th.ompt_thread_info.state = ompt_state_work_parallel; | |||
1516 | } else { | |||
1517 | exit_frame_p = &dummy; | |||
1518 | } | |||
1519 | #endif | |||
1520 | // AC: need to decrement t_serialized for enquiry functions to work | |||
1521 | // correctly, will restore at join time | |||
1522 | parent_team->t.t_serialized--; | |||
1523 | ||||
1524 | { | |||
1525 | KMP_TIME_PARTITIONED_BLOCK(OMP_parallel)((void)0); | |||
1526 | KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK)((void)0); | |||
1527 | __kmp_invoke_microtask(microtask, gtid, 0, argc, parent_team->t.t_argv | |||
1528 | #if OMPT_SUPPORT1 | |||
1529 | , | |||
1530 | exit_frame_p | |||
1531 | #endif | |||
1532 | ); | |||
1533 | } | |||
1534 | ||||
1535 | #if OMPT_SUPPORT1 | |||
1536 | if (ompt_enabled.enabled) { | |||
1537 | *exit_frame_p = NULL__null; | |||
1538 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->frame.exit_frame = ompt_data_none{0}; | |||
1539 | if (ompt_enabled.ompt_callback_implicit_task) { | |||
1540 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | |||
1541 | ompt_scope_end, NULL__null, implicit_task_data, 1, | |||
1542 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num, ompt_task_implicit); | |||
1543 | } | |||
1544 | ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th)(&(master_th->th.th_team->t.ompt_team_info.parallel_data )); | |||
1545 | __ompt_lw_taskteam_unlink(master_th); | |||
1546 | if (ompt_enabled.ompt_callback_parallel_end) { | |||
1547 | ompt_callbacks.ompt_callback(ompt_callback_parallel_end)ompt_callback_parallel_end_callback( | |||
1548 | &ompt_parallel_data, OMPT_CUR_TASK_DATA(master_th)(&(master_th->th.th_current_task->ompt_task_info.task_data )), | |||
1549 | OMPT_INVOKER(call_context)((call_context == fork_context_gnu) ? ompt_parallel_invoker_program : ompt_parallel_invoker_runtime) | ompt_parallel_team, | |||
1550 | return_address); | |||
1551 | } | |||
1552 | master_th->th.ompt_thread_info.state = ompt_state_overhead; | |||
1553 | } | |||
1554 | #endif | |||
1555 | return TRUE(!0); | |||
1556 | } | |||
1557 | ||||
1558 | parent_team->t.t_pkfn = microtask; | |||
1559 | parent_team->t.t_invoke = invoker; | |||
1560 | KMP_ATOMIC_INC(&root->r.r_in_parallel)(&root->r.r_in_parallel)->fetch_add(1, std::memory_order_acq_rel ); | |||
1561 | parent_team->t.t_active_level++; | |||
1562 | parent_team->t.t_level++; | |||
1563 | parent_team->t.t_def_allocator = master_th->th.th_def_allocator; // save | |||
1564 | ||||
1565 | #if OMPT_SUPPORT1 | |||
1566 | if (ompt_enabled.enabled) { | |||
1567 | ompt_lw_taskteam_t lw_taskteam; | |||
1568 | __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid, | |||
1569 | &ompt_parallel_data, return_address); | |||
1570 | __ompt_lw_taskteam_link(&lw_taskteam, master_th, 1, true); | |||
1571 | } | |||
1572 | #endif | |||
1573 | ||||
1574 | /* Change number of threads in the team if requested */ | |||
1575 | if (master_set_numthreads) { // The parallel has num_threads clause | |||
1576 | if (master_set_numthreads <= master_th->th.th_teams_size.nth) { | |||
1577 | // AC: only can reduce number of threads dynamically, can't increase | |||
1578 | kmp_info_t **other_threads = parent_team->t.t_threads; | |||
1579 | // NOTE: if using distributed barrier, we need to run this code block | |||
1580 | // even when the team size appears not to have changed from the max. | |||
1581 | int old_proc = master_th->th.th_teams_size.nth; | |||
1582 | if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == | |||
1583 | bp_dist_bar) { | |||
1584 | __kmp_resize_dist_barrier(parent_team, old_proc, | |||
1585 | master_set_numthreads); | |||
1586 | __kmp_add_threads_to_team(parent_team, master_set_numthreads); | |||
1587 | } | |||
1588 | parent_team->t.t_nproc = master_set_numthreads; | |||
1589 | for (i = 0; i < master_set_numthreads; ++i) { | |||
1590 | other_threads[i]->th.th_team_nproc = master_set_numthreads; | |||
1591 | } | |||
1592 | } | |||
1593 | // Keep extra threads hot in the team for possible next parallels | |||
1594 | master_th->th.th_set_nproc = 0; | |||
1595 | } | |||
1596 | ||||
1597 | #if USE_DEBUGGER0 | |||
1598 | if (__kmp_debugging) { // Let debugger override number of threads. | |||
1599 | int nth = __kmp_omp_num_threads(loc); | |||
1600 | if (nth > 0) { // 0 means debugger doesn't want to change num threads | |||
1601 | master_set_numthreads = nth; | |||
1602 | } | |||
1603 | } | |||
1604 | #endif | |||
1605 | ||||
1606 | // Figure out the proc_bind policy for the nested parallel within teams | |||
1607 | kmp_proc_bind_t proc_bind = master_th->th.th_set_proc_bind; | |||
1608 | // proc_bind_default means don't update | |||
1609 | kmp_proc_bind_t proc_bind_icv = proc_bind_default; | |||
1610 | if (master_th->th.th_current_task->td_icvs.proc_bind == proc_bind_false) { | |||
1611 | proc_bind = proc_bind_false; | |||
1612 | } else { | |||
1613 | // No proc_bind clause specified; use current proc-bind-var | |||
1614 | if (proc_bind == proc_bind_default) { | |||
1615 | proc_bind = master_th->th.th_current_task->td_icvs.proc_bind; | |||
1616 | } | |||
1617 | /* else: The proc_bind policy was specified explicitly on parallel | |||
1618 | clause. | |||
1619 | This overrides proc-bind-var for this parallel region, but does not | |||
1620 | change proc-bind-var. */ | |||
1621 | // Figure the value of proc-bind-var for the child threads. | |||
1622 | if ((level + 1 < __kmp_nested_proc_bind.used) && | |||
1623 | (__kmp_nested_proc_bind.bind_types[level + 1] != | |||
1624 | master_th->th.th_current_task->td_icvs.proc_bind)) { | |||
1625 | proc_bind_icv = __kmp_nested_proc_bind.bind_types[level + 1]; | |||
1626 | } | |||
1627 | } | |||
1628 | 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); | |||
1629 | // Need to change the bind-var ICV to correct value for each implicit task | |||
1630 | if (proc_bind_icv != proc_bind_default && | |||
1631 | master_th->th.th_current_task->td_icvs.proc_bind != proc_bind_icv) { | |||
1632 | kmp_info_t **other_threads = parent_team->t.t_threads; | |||
1633 | for (i = 0; i < master_th->th.th_team_nproc; ++i) { | |||
1634 | other_threads[i]->th.th_current_task->td_icvs.proc_bind = | |||
1635 | proc_bind_icv; | |||
1636 | } | |||
1637 | } | |||
1638 | // Reset for next parallel region | |||
1639 | master_th->th.th_set_proc_bind = proc_bind_default; | |||
1640 | ||||
1641 | #if USE_ITT_BUILD1 && USE_ITT_NOTIFY1 | |||
1642 | 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) || | |||
1643 | KMP_ITT_DEBUG0) && | |||
1644 | __kmp_forkjoin_frames_mode == 3 && | |||
1645 | parent_team->t.t_active_level == 1 // only report frames at level 1 | |||
1646 | && master_th->th.th_teams_size.nteams == 1) { | |||
1647 | kmp_uint64 tmp_time = __itt_get_timestamp(!__kmp_itt_get_timestamp_ptr__3_0) ? 0 : __kmp_itt_get_timestamp_ptr__3_0(); | |||
1648 | master_th->th.th_frame_time = tmp_time; | |||
1649 | parent_team->t.t_region_time = tmp_time; | |||
1650 | } | |||
1651 | if (__itt_stack_caller_create_ptr__kmp_itt_stack_caller_create_ptr__3_0) { | |||
1652 | 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" , 1652); }; | |||
1653 | // create new stack stitching id before entering fork barrier | |||
1654 | parent_team->t.t_stack_id = __kmp_itt_stack_caller_create(); | |||
1655 | } | |||
1656 | #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */ | |||
1657 | #if KMP_AFFINITY_SUPPORTED1 | |||
1658 | __kmp_partition_places(parent_team); | |||
1659 | #endif | |||
1660 | ||||
1661 | KF_TRACE(10, ("__kmp_fork_call: before internal fork: root=%p, team=%p, "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: before internal fork: root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, parent_team, master_th, gtid ); } | |||
1662 | "master_th=%p, gtid=%d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: before internal fork: root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, parent_team, master_th, gtid ); } | |||
1663 | root, parent_team, master_th, gtid))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: before internal fork: root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, parent_team, master_th, gtid ); }; | |||
1664 | __kmp_internal_fork(loc, gtid, parent_team); | |||
1665 | KF_TRACE(10, ("__kmp_fork_call: after internal fork: root=%p, team=%p, "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: after internal fork: root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, parent_team, master_th, gtid ); } | |||
1666 | "master_th=%p, gtid=%d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: after internal fork: root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, parent_team, master_th, gtid ); } | |||
1667 | root, parent_team, master_th, gtid))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: after internal fork: root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, parent_team, master_th, gtid ); }; | |||
1668 | ||||
1669 | if (call_context == fork_context_gnu) | |||
1670 | return TRUE(!0); | |||
1671 | ||||
1672 | /* Invoke microtask for PRIMARY thread */ | |||
1673 | 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, parent_team->t.t_id, parent_team->t.t_pkfn); } | |||
1674 | parent_team->t.t_id, parent_team->t.t_pkfn))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n" , gtid, parent_team->t.t_id, parent_team->t.t_pkfn); }; | |||
1675 | ||||
1676 | if (!parent_team->t.t_invoke(gtid)) { | |||
1677 | 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", 1677); }; | |||
1678 | } | |||
1679 | 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, parent_team->t.t_id, parent_team->t.t_pkfn); } | |||
1680 | parent_team->t.t_id, parent_team->t.t_pkfn))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n" , gtid, parent_team->t.t_id, parent_team->t.t_pkfn); }; | |||
1681 | KMP_MB(); /* Flush all pending memory write invalidates. */ | |||
1682 | ||||
1683 | 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); }; | |||
1684 | ||||
1685 | return TRUE(!0); | |||
1686 | } // Parallel closely nested in teams construct | |||
1687 | ||||
1688 | #if KMP_DEBUG1 | |||
1689 | if (__kmp_tasking_mode != tskm_immediate_exec) { | |||
1690 | 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", 1691); } | |||
1691 | 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", 1691); }; | |||
1692 | } | |||
1693 | #endif | |||
1694 | ||||
1695 | // Need this to happen before we determine the number of threads, not while | |||
1696 | // we are allocating the team | |||
1697 | //__kmp_push_current_task_to_thread(master_th, parent_team, 0); | |||
1698 | int enter_teams = 0; | |||
1699 | if (parent_team->t.t_active_level >= | |||
1700 | master_th->th.th_current_task->td_icvs.max_active_levels) { | |||
1701 | nthreads = 1; | |||
1702 | } else { | |||
1703 | enter_teams = ((ap == NULL__null && active_level == 0) || | |||
1704 | (ap && teams_level > 0 && teams_level == level)); | |||
1705 | nthreads = master_set_numthreads | |||
1706 | ? master_set_numthreads | |||
1707 | // TODO: get nproc directly from current task | |||
1708 | : get__nproc_2(parent_team, master_tid)((parent_team)->t.t_threads[(master_tid)]->th.th_current_task ->td_icvs.nproc); | |||
1709 | // Check if we need to take forkjoin lock? (no need for serialized | |||
1710 | // parallel out of teams construct). This code moved here from | |||
1711 | // __kmp_reserve_threads() to speedup nested serialized parallels. | |||
1712 | if (nthreads > 1) { | |||
1713 | if ((get__max_active_levels(master_th)((master_th)->th.th_current_task->td_icvs.max_active_levels ) == 1 && | |||
1714 | (root->r.r_in_parallel && !enter_teams)) || | |||
1715 | (__kmp_library == library_serial)) { | |||
1716 | KC_TRACE(10, ("__kmp_fork_call: T#%d serializing team; requested %d"if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: T#%d serializing team; requested %d" " threads\n", gtid, nthreads); } | |||
1717 | " threads\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: T#%d serializing team; requested %d" " threads\n", gtid, nthreads); } | |||
1718 | gtid, nthreads))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: T#%d serializing team; requested %d" " threads\n", gtid, nthreads); }; | |||
1719 | nthreads = 1; | |||
1720 | } | |||
1721 | } | |||
1722 | if (nthreads > 1) { | |||
1723 | /* determine how many new threads we can use */ | |||
1724 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); | |||
1725 | /* AC: If we execute teams from parallel region (on host), then teams | |||
1726 | should be created but each can only have 1 thread if nesting is | |||
1727 | disabled. If teams called from serial region, then teams and their | |||
1728 | threads should be created regardless of the nesting setting. */ | |||
1729 | nthreads = __kmp_reserve_threads(root, parent_team, master_tid, | |||
1730 | nthreads, enter_teams); | |||
1731 | if (nthreads == 1) { | |||
1732 | // Free lock for single thread execution here; for multi-thread | |||
1733 | // execution it will be freed later after team of threads created | |||
1734 | // and initialized | |||
1735 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | |||
1736 | } | |||
1737 | } | |||
1738 | } | |||
1739 | KMP_DEBUG_ASSERT(nthreads > 0)if (!(nthreads > 0)) { __kmp_debug_assert("nthreads > 0" , "openmp/runtime/src/kmp_runtime.cpp", 1739); }; | |||
1740 | ||||
1741 | // If we temporarily changed the set number of threads then restore it now | |||
1742 | master_th->th.th_set_nproc = 0; | |||
1743 | ||||
1744 | /* create a serialized parallel region? */ | |||
1745 | if (nthreads == 1) { | |||
1746 | /* josh todo: hypothetical question: what do we do for OS X*? */ | |||
1747 | #if KMP_OS_LINUX1 && \ | |||
1748 | (KMP_ARCH_X860 || KMP_ARCH_X86_641 || KMP_ARCH_ARM || KMP_ARCH_AARCH640) | |||
1749 | void *args[argc]; | |||
1750 | #else | |||
1751 | void **args = (void **)KMP_ALLOCA(argc * sizeof(void *))__builtin_alloca (argc * sizeof(void *)); | |||
1752 | #endif /* KMP_OS_LINUX && ( KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || \ | |||
1753 | KMP_ARCH_AARCH64) */ | |||
1754 | ||||
1755 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d serializing parallel region\n" , gtid); } | |||
1756 | ("__kmp_fork_call: T#%d serializing parallel region\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d serializing parallel region\n" , gtid); }; | |||
1757 | ||||
1758 | __kmpc_serialized_parallel(loc, gtid); | |||
1759 | ||||
1760 | #if OMPD_SUPPORT1 | |||
1761 | master_th->th.th_serial_team->t.t_pkfn = microtask; | |||
1762 | #endif | |||
1763 | ||||
1764 | if (call_context == fork_context_intel) { | |||
1765 | /* TODO this sucks, use the compiler itself to pass args! :) */ | |||
1766 | master_th->th.th_serial_team->t.t_ident = loc; | |||
1767 | if (!ap) { | |||
1768 | // revert change made in __kmpc_serialized_parallel() | |||
1769 | master_th->th.th_serial_team->t.t_level--; | |||
1770 | // Get args from parent team for teams construct | |||
1771 | ||||
1772 | #if OMPT_SUPPORT1 | |||
1773 | void *dummy; | |||
1774 | void **exit_frame_p; | |||
1775 | ompt_task_info_t *task_info; | |||
1776 | ||||
1777 | ompt_lw_taskteam_t lw_taskteam; | |||
1778 | ||||
1779 | if (ompt_enabled.enabled) { | |||
1780 | __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid, | |||
1781 | &ompt_parallel_data, return_address); | |||
1782 | ||||
1783 | __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0); | |||
1784 | // don't use lw_taskteam after linking. content was swaped | |||
1785 | ||||
1786 | task_info = OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info)); | |||
1787 | exit_frame_p = &(task_info->frame.exit_frame.ptr); | |||
1788 | if (ompt_enabled.ompt_callback_implicit_task) { | |||
1789 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num = | |||
1790 | __kmp_tid_from_gtid(gtid); | |||
1791 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | |||
1792 | ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th)(&(master_th->th.th_team->t.ompt_team_info.parallel_data )), | |||
1793 | &(task_info->task_data), 1, | |||
1794 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num, | |||
1795 | ompt_task_implicit); | |||
1796 | } | |||
1797 | ||||
1798 | /* OMPT state */ | |||
1799 | master_th->th.ompt_thread_info.state = ompt_state_work_parallel; | |||
1800 | } else { | |||
1801 | exit_frame_p = &dummy; | |||
1802 | } | |||
1803 | #endif | |||
1804 | ||||
1805 | { | |||
1806 | KMP_TIME_PARTITIONED_BLOCK(OMP_parallel)((void)0); | |||
1807 | KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK)((void)0); | |||
1808 | __kmp_invoke_microtask(microtask, gtid, 0, argc, | |||
1809 | parent_team->t.t_argv | |||
1810 | #if OMPT_SUPPORT1 | |||
1811 | , | |||
1812 | exit_frame_p | |||
1813 | #endif | |||
1814 | ); | |||
1815 | } | |||
1816 | ||||
1817 | #if OMPT_SUPPORT1 | |||
1818 | if (ompt_enabled.enabled) { | |||
1819 | *exit_frame_p = NULL__null; | |||
1820 | if (ompt_enabled.ompt_callback_implicit_task) { | |||
1821 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | |||
1822 | ompt_scope_end, NULL__null, &(task_info->task_data), 1, | |||
1823 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num, | |||
1824 | ompt_task_implicit); | |||
1825 | } | |||
1826 | ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th)(&(master_th->th.th_team->t.ompt_team_info.parallel_data )); | |||
1827 | __ompt_lw_taskteam_unlink(master_th); | |||
1828 | if (ompt_enabled.ompt_callback_parallel_end) { | |||
1829 | ompt_callbacks.ompt_callback(ompt_callback_parallel_end)ompt_callback_parallel_end_callback( | |||
1830 | &ompt_parallel_data, parent_task_data, | |||
1831 | OMPT_INVOKER(call_context)((call_context == fork_context_gnu) ? ompt_parallel_invoker_program : ompt_parallel_invoker_runtime) | ompt_parallel_team, | |||
1832 | return_address); | |||
1833 | } | |||
1834 | master_th->th.ompt_thread_info.state = ompt_state_overhead; | |||
1835 | } | |||
1836 | #endif | |||
1837 | } else if (microtask == (microtask_t)__kmp_teams_master) { | |||
1838 | KMP_DEBUG_ASSERT(master_th->th.th_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", 1839); } | |||
1839 | 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", 1839); }; | |||
1840 | team = master_th->th.th_team; | |||
1841 | // team->t.t_pkfn = microtask; | |||
1842 | team->t.t_invoke = invoker; | |||
1843 | __kmp_alloc_argv_entries(argc, team, TRUE(!0)); | |||
1844 | team->t.t_argc = argc; | |||
1845 | argv = (void **)team->t.t_argv; | |||
1846 | if (ap) { | |||
1847 | for (i = argc - 1; i >= 0; --i) | |||
1848 | *argv++ = va_arg(kmp_va_deref(ap), void *)__builtin_va_arg((*(ap)), void *); | |||
1849 | } else { | |||
1850 | for (i = 0; i < argc; ++i) | |||
1851 | // Get args from parent team for teams construct | |||
1852 | argv[i] = parent_team->t.t_argv[i]; | |||
1853 | } | |||
1854 | // AC: revert change made in __kmpc_serialized_parallel() | |||
1855 | // because initial code in teams should have level=0 | |||
1856 | team->t.t_level--; | |||
1857 | // AC: call special invoker for outer "parallel" of teams construct | |||
1858 | invoker(gtid); | |||
1859 | #if OMPT_SUPPORT1 | |||
1860 | if (ompt_enabled.enabled) { | |||
1861 | ompt_task_info_t *task_info = OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info)); | |||
1862 | if (ompt_enabled.ompt_callback_implicit_task) { | |||
1863 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | |||
1864 | ompt_scope_end, NULL__null, &(task_info->task_data), 0, | |||
1865 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num, ompt_task_initial); | |||
1866 | } | |||
1867 | if (ompt_enabled.ompt_callback_parallel_end) { | |||
1868 | ompt_callbacks.ompt_callback(ompt_callback_parallel_end)ompt_callback_parallel_end_callback( | |||
1869 | &ompt_parallel_data, parent_task_data, | |||
1870 | OMPT_INVOKER(call_context)((call_context == fork_context_gnu) ? ompt_parallel_invoker_program : ompt_parallel_invoker_runtime) | ompt_parallel_league, | |||
1871 | return_address); | |||
1872 | } | |||
1873 | master_th->th.ompt_thread_info.state = ompt_state_overhead; | |||
1874 | } | |||
1875 | #endif | |||
1876 | } else { | |||
1877 | argv = args; | |||
1878 | for (i = argc - 1; i >= 0; --i) | |||
1879 | *argv++ = va_arg(kmp_va_deref(ap), void *)__builtin_va_arg((*(ap)), void *); | |||
1880 | KMP_MB(); | |||
1881 | ||||
1882 | #if OMPT_SUPPORT1 | |||
1883 | void *dummy; | |||
1884 | void **exit_frame_p; | |||
1885 | ompt_task_info_t *task_info; | |||
1886 | ||||
1887 | ompt_lw_taskteam_t lw_taskteam; | |||
1888 | ||||
1889 | if (ompt_enabled.enabled) { | |||
1890 | __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid, | |||
1891 | &ompt_parallel_data, return_address); | |||
1892 | __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0); | |||
1893 | // don't use lw_taskteam after linking. content was swaped | |||
1894 | task_info = OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info)); | |||
1895 | exit_frame_p = &(task_info->frame.exit_frame.ptr); | |||
1896 | ||||
1897 | /* OMPT implicit task begin */ | |||
1898 | implicit_task_data = OMPT_CUR_TASK_DATA(master_th)(&(master_th->th.th_current_task->ompt_task_info.task_data )); | |||
1899 | if (ompt_enabled.ompt_callback_implicit_task) { | |||
1900 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | |||
1901 | ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th)(&(master_th->th.th_team->t.ompt_team_info.parallel_data )), | |||
1902 | implicit_task_data, 1, __kmp_tid_from_gtid(gtid), | |||
1903 | ompt_task_implicit); | |||
1904 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num = | |||
1905 | __kmp_tid_from_gtid(gtid); | |||
1906 | } | |||
1907 | ||||
1908 | /* OMPT state */ | |||
1909 | master_th->th.ompt_thread_info.state = ompt_state_work_parallel; | |||
1910 | } else { | |||
1911 | exit_frame_p = &dummy; | |||
1912 | } | |||
1913 | #endif | |||
1914 | ||||
1915 | { | |||
1916 | KMP_TIME_PARTITIONED_BLOCK(OMP_parallel)((void)0); | |||
1917 | KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK)((void)0); | |||
1918 | __kmp_invoke_microtask(microtask, gtid, 0, argc, args | |||
1919 | #if OMPT_SUPPORT1 | |||
1920 | , | |||
1921 | exit_frame_p | |||
1922 | #endif | |||
1923 | ); | |||
1924 | } | |||
1925 | ||||
1926 | #if OMPT_SUPPORT1 | |||
1927 | if (ompt_enabled.enabled) { | |||
1928 | *exit_frame_p = NULL__null; | |||
1929 | if (ompt_enabled.ompt_callback_implicit_task) { | |||
1930 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | |||
1931 | ompt_scope_end, NULL__null, &(task_info->task_data), 1, | |||
1932 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num, | |||
1933 | ompt_task_implicit); | |||
1934 | } | |||
1935 | ||||
1936 | ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th)(&(master_th->th.th_team->t.ompt_team_info.parallel_data )); | |||
1937 | __ompt_lw_taskteam_unlink(master_th); | |||
1938 | if (ompt_enabled.ompt_callback_parallel_end) { | |||
1939 | ompt_callbacks.ompt_callback(ompt_callback_parallel_end)ompt_callback_parallel_end_callback( | |||
1940 | &ompt_parallel_data, parent_task_data, | |||
1941 | OMPT_INVOKER(call_context)((call_context == fork_context_gnu) ? ompt_parallel_invoker_program : ompt_parallel_invoker_runtime) | ompt_parallel_team, | |||
1942 | return_address); | |||
1943 | } | |||
1944 | master_th->th.ompt_thread_info.state = ompt_state_overhead; | |||
1945 | } | |||
1946 | #endif | |||
1947 | } | |||
1948 | } else if (call_context == fork_context_gnu) { | |||
1949 | #if OMPT_SUPPORT1 | |||
1950 | if (ompt_enabled.enabled) { | |||
1951 | ompt_lw_taskteam_t lwt; | |||
1952 | __ompt_lw_taskteam_init(&lwt, master_th, gtid, &ompt_parallel_data, | |||
1953 | return_address); | |||
1954 | ||||
1955 | lwt.ompt_task_info.frame.exit_frame = ompt_data_none{0}; | |||
1956 | __ompt_lw_taskteam_link(&lwt, master_th, 1); | |||
1957 | } | |||
1958 | // don't use lw_taskteam after linking. content was swaped | |||
1959 | #endif | |||
1960 | ||||
1961 | // we were called from GNU native code | |||
1962 | KA_TRACE(20, ("__kmp_fork_call: T#%d serial exit\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d serial exit\n" , gtid); }; | |||
1963 | return FALSE0; | |||
1964 | } else { | |||
1965 | KMP_ASSERT2(call_context < fork_context_last,if (!(call_context < fork_context_last)) { __kmp_debug_assert (("__kmp_fork_call: unknown fork_context parameter"), "openmp/runtime/src/kmp_runtime.cpp" , 1966); } | |||
1966 | "__kmp_fork_call: unknown fork_context parameter")if (!(call_context < fork_context_last)) { __kmp_debug_assert (("__kmp_fork_call: unknown fork_context parameter"), "openmp/runtime/src/kmp_runtime.cpp" , 1966); }; | |||
1967 | } | |||
1968 | ||||
1969 | KA_TRACE(20, ("__kmp_fork_call: T#%d serial exit\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d serial exit\n" , gtid); }; | |||
1970 | KMP_MB(); | |||
1971 | return FALSE0; | |||
1972 | } // if (nthreads == 1) | |||
1973 | ||||
1974 | // GEH: only modify the executing flag in the case when not serialized | |||
1975 | // serialized case is handled in kmpc_serialized_parallel | |||
1976 | 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); } | |||
1977 | "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); } | |||
1978 | 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); } | |||
1979 | 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); } | |||
1980 | 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); }; | |||
1981 | // TODO: GEH - cannot do this assertion because root thread not set up as | |||
1982 | // executing | |||
1983 | // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 1 ); | |||
1984 | master_th->th.th_current_task->td_flags.executing = 0; | |||
1985 | ||||
1986 | if (!master_th->th.th_teams_microtask || level > teams_level) { | |||
1987 | /* Increment our nested depth level */ | |||
1988 | KMP_ATOMIC_INC(&root->r.r_in_parallel)(&root->r.r_in_parallel)->fetch_add(1, std::memory_order_acq_rel ); | |||
1989 | } | |||
1990 | ||||
1991 | // See if we need to make a copy of the ICVs. | |||
1992 | int nthreads_icv = master_th->th.th_current_task->td_icvs.nproc; | |||
1993 | if ((level + 1 < __kmp_nested_nth.used) && | |||
1994 | (__kmp_nested_nth.nth[level + 1] != nthreads_icv)) { | |||
1995 | nthreads_icv = __kmp_nested_nth.nth[level + 1]; | |||
1996 | } else { | |||
1997 | nthreads_icv = 0; // don't update | |||
1998 | } | |||
1999 | ||||
2000 | // Figure out the proc_bind_policy for the new team. | |||
2001 | kmp_proc_bind_t proc_bind = master_th->th.th_set_proc_bind; | |||
2002 | // proc_bind_default means don't update | |||
2003 | kmp_proc_bind_t proc_bind_icv = proc_bind_default; | |||
2004 | if (master_th->th.th_current_task->td_icvs.proc_bind == proc_bind_false) { | |||
2005 | proc_bind = proc_bind_false; | |||
2006 | } else { | |||
2007 | // No proc_bind clause specified; use current proc-bind-var for this | |||
2008 | // parallel region | |||
2009 | if (proc_bind == proc_bind_default) { | |||
2010 | proc_bind = master_th->th.th_current_task->td_icvs.proc_bind; | |||
2011 | } | |||
2012 | // Have teams construct take proc_bind value from KMP_TEAMS_PROC_BIND | |||
2013 | if (master_th->th.th_teams_microtask && | |||
2014 | microtask == (microtask_t)__kmp_teams_master) { | |||
2015 | proc_bind = __kmp_teams_proc_bind; | |||
2016 | } | |||
2017 | /* else: The proc_bind policy was specified explicitly on parallel clause. | |||
2018 | This overrides proc-bind-var for this parallel region, but does not | |||
2019 | change proc-bind-var. */ | |||
2020 | // Figure the value of proc-bind-var for the child threads. | |||
2021 | if ((level + 1 < __kmp_nested_proc_bind.used) && | |||
2022 | (__kmp_nested_proc_bind.bind_types[level + 1] != | |||
2023 | master_th->th.th_current_task->td_icvs.proc_bind)) { | |||
2024 | // Do not modify the proc bind icv for the two teams construct forks | |||
2025 | // They just let the proc bind icv pass through | |||
2026 | if (!master_th->th.th_teams_microtask || | |||
2027 | !(microtask == (microtask_t)__kmp_teams_master || ap == NULL__null)) | |||
2028 | proc_bind_icv = __kmp_nested_proc_bind.bind_types[level + 1]; | |||
2029 | } | |||
2030 | } | |||
2031 | ||||
2032 | // Reset for next parallel region | |||
2033 | master_th->th.th_set_proc_bind = proc_bind_default; | |||
2034 | ||||
2035 | if ((nthreads_icv > 0) || (proc_bind_icv != proc_bind_default)) { | |||
2036 | kmp_internal_control_t new_icvs; | |||
2037 | copy_icvs(&new_icvs, &master_th->th.th_current_task->td_icvs); | |||
2038 | new_icvs.next = NULL__null; | |||
2039 | if (nthreads_icv > 0) { | |||
2040 | new_icvs.nproc = nthreads_icv; | |||
2041 | } | |||
2042 | if (proc_bind_icv != proc_bind_default) { | |||
2043 | new_icvs.proc_bind = proc_bind_icv; | |||
2044 | } | |||
2045 | ||||
2046 | /* allocate a new parallel team */ | |||
2047 | 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" ); }; | |||
2048 | team = __kmp_allocate_team(root, nthreads, nthreads, | |||
2049 | #if OMPT_SUPPORT1 | |||
2050 | ompt_parallel_data, | |||
2051 | #endif | |||
2052 | proc_bind, &new_icvs, | |||
2053 | argc USE_NESTED_HOT_ARG(master_th), master_th); | |||
2054 | if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) | |||
2055 | copy_icvs((kmp_internal_control_t *)team->t.b->team_icvs, &new_icvs); | |||
2056 | } else { | |||
2057 | /* allocate a new parallel team */ | |||
2058 | 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" ); }; | |||
2059 | team = __kmp_allocate_team(root, nthreads, nthreads, | |||
2060 | #if OMPT_SUPPORT1 | |||
2061 | ompt_parallel_data, | |||
2062 | #endif | |||
2063 | proc_bind, | |||
2064 | &master_th->th.th_current_task->td_icvs, | |||
2065 | argc USE_NESTED_HOT_ARG(master_th), master_th); | |||
2066 | if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) | |||
2067 | copy_icvs((kmp_internal_control_t *)team->t.b->team_icvs, | |||
2068 | &master_th->th.th_current_task->td_icvs); | |||
2069 | } | |||
2070 | KF_TRACE(if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: after __kmp_allocate_team - team = %p\n" , team); } | |||
2071 | 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); }; | |||
2072 | ||||
2073 | /* setup the new team */ | |||
2074 | 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); | |||
2075 | 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); | |||
2076 | KMP_CHECK_UPDATE(team->t.t_ident, loc)if ((team->t.t_ident) != (loc)) (team->t.t_ident) = (loc ); | |||
2077 | KMP_CHECK_UPDATE(team->t.t_parent, parent_team)if ((team->t.t_parent) != (parent_team)) (team->t.t_parent ) = (parent_team); | |||
2078 | KMP_CHECK_UPDATE_SYNC(team->t.t_pkfn, microtask)if ((team->t.t_pkfn) != (microtask)) (((team->t.t_pkfn) )) = (((microtask))); | |||
2079 | #if OMPT_SUPPORT1 | |||
2080 | 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))) | |||
2081 | return_address)if ((team->t.ompt_team_info.master_return_address) != (return_address )) (((team->t.ompt_team_info.master_return_address))) = (( (return_address))); | |||
2082 | #endif | |||
2083 | KMP_CHECK_UPDATE(team->t.t_invoke, invoker)if ((team->t.t_invoke) != (invoker)) (team->t.t_invoke) = (invoker); // TODO move to root, maybe | |||
2084 | // TODO: parent_team->t.t_level == INT_MAX ??? | |||
2085 | if (!master_th->th.th_teams_microtask || level > teams_level) { | |||
2086 | int new_level = parent_team->t.t_level + 1; | |||
2087 | KMP_CHECK_UPDATE(team->t.t_level, new_level)if ((team->t.t_level) != (new_level)) (team->t.t_level) = (new_level); | |||
2088 | new_level = parent_team->t.t_active_level + 1; | |||
2089 | 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); | |||
2090 | } else { | |||
2091 | // AC: Do not increase parallel level at start of the teams construct | |||
2092 | int new_level = parent_team->t.t_level; | |||
2093 | KMP_CHECK_UPDATE(team->t.t_level, new_level)if ((team->t.t_level) != (new_level)) (team->t.t_level) = (new_level); | |||
2094 | new_level = parent_team->t.t_active_level; | |||
2095 | 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); | |||
2096 | } | |||
2097 | 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); | |||
2098 | // set primary thread's schedule as new run-time schedule | |||
2099 | 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); | |||
2100 | ||||
2101 | 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); | |||
2102 | 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 ); | |||
2103 | ||||
2104 | // Update the floating point rounding in the team if required. | |||
2105 | propagateFPControl(team); | |||
2106 | #if OMPD_SUPPORT1 | |||
2107 | if (ompd_state & OMPD_ENABLE_BP0x1) | |||
2108 | ompd_bp_parallel_begin(); | |||
2109 | #endif | |||
2110 | ||||
2111 | if (__kmp_tasking_mode != tskm_immediate_exec) { | |||
2112 | // Set primary thread's task team to team's task team. Unless this is hot | |||
2113 | // team, it should be NULL. | |||
2114 | 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", 2115); } | |||
2115 | 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", 2115); }; | |||
2116 | 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); } | |||
2117 | "%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); } | |||
2118 | __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); } | |||
2119 | 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); } | |||
2120 | 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); }; | |||
2121 | ||||
2122 | if (active_level || master_th->th.th_task_team) { | |||
2123 | // Take a memo of primary thread's task_state | |||
2124 | 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" , 2124); }; | |||
2125 | if (master_th->th.th_task_state_top >= | |||
2126 | master_th->th.th_task_state_stack_sz) { // increase size | |||
2127 | kmp_uint32 new_size = 2 * master_th->th.th_task_state_stack_sz; | |||
2128 | kmp_uint8 *old_stack, *new_stack; | |||
2129 | kmp_uint32 i; | |||
2130 | new_stack = (kmp_uint8 *)__kmp_allocate(new_size)___kmp_allocate((new_size), "openmp/runtime/src/kmp_runtime.cpp" , 2130); | |||
2131 | for (i = 0; i < master_th->th.th_task_state_stack_sz; ++i) { | |||
2132 | new_stack[i] = master_th->th.th_task_state_memo_stack[i]; | |||
2133 | } | |||
2134 | for (i = master_th->th.th_task_state_stack_sz; i < new_size; | |||
2135 | ++i) { // zero-init rest of stack | |||
2136 | new_stack[i] = 0; | |||
2137 | } | |||
2138 | old_stack = master_th->th.th_task_state_memo_stack; | |||
2139 | master_th->th.th_task_state_memo_stack = new_stack; | |||
2140 | master_th->th.th_task_state_stack_sz = new_size; | |||
2141 | __kmp_free(old_stack)___kmp_free((old_stack), "openmp/runtime/src/kmp_runtime.cpp" , 2141); | |||
2142 | } | |||
2143 | // Store primary thread's task_state on stack | |||
2144 | master_th->th | |||
2145 | .th_task_state_memo_stack[master_th->th.th_task_state_top] = | |||
2146 | master_th->th.th_task_state; | |||
2147 | master_th->th.th_task_state_top++; | |||
2148 | #if KMP_NESTED_HOT_TEAMS1 | |||
2149 | if (master_th->th.th_hot_teams && | |||
2150 | active_level < __kmp_hot_teams_max_level && | |||
2151 | team == master_th->th.th_hot_teams[active_level].hot_team) { | |||
2152 | // Restore primary thread's nested state if nested hot team | |||
2153 | master_th->th.th_task_state = | |||
2154 | master_th->th | |||
2155 | .th_task_state_memo_stack[master_th->th.th_task_state_top]; | |||
2156 | } else { | |||
2157 | #endif | |||
2158 | master_th->th.th_task_state = 0; | |||
2159 | #if KMP_NESTED_HOT_TEAMS1 | |||
2160 | } | |||
2161 | #endif | |||
2162 | } | |||
2163 | #if !KMP_NESTED_HOT_TEAMS1 | |||
2164 | 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", 2165); } | |||
2165 | (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", 2165); }; | |||
2166 | #endif | |||
2167 | } | |||
2168 | ||||
2169 | 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); } | |||
2170 | 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); } | |||
2171 | ("__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); } | |||
2172 | 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); } | |||
2173 | 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); }; | |||
2174 | 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", 2177); } | |||
2175 | (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", 2177); } | |||
2176 | (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", 2177); } | |||
2177 | 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", 2177); }; | |||
2178 | KMP_MB(); | |||
2179 | ||||
2180 | /* now, setup the arguments */ | |||
2181 | argv = (void **)team->t.t_argv; | |||
2182 | if (ap) { | |||
2183 | for (i = argc - 1; i >= 0; --i) { | |||
2184 | void *new_argv = va_arg(kmp_va_deref(ap), void *)__builtin_va_arg((*(ap)), void *); | |||
2185 | KMP_CHECK_UPDATE(*argv, new_argv)if ((*argv) != (new_argv)) (*argv) = (new_argv); | |||
2186 | argv++; | |||
2187 | } | |||
2188 | } else { | |||
2189 | for (i = 0; i < argc; ++i) { | |||
2190 | // Get args from parent team for teams construct | |||
2191 | 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]); | |||
2192 | } | |||
2193 | } | |||
2194 | ||||
2195 | /* now actually fork the threads */ | |||
2196 | 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); | |||
2197 | if (!root->r.r_active) // Only do assignment if it prevents cache ping-pong | |||
2198 | root->r.r_active = TRUE(!0); | |||
2199 | ||||
2200 | __kmp_fork_team_threads(root, team, master_th, gtid, !ap); | |||
2201 | __kmp_setup_icv_copy(team, nthreads, | |||
2202 | &master_th->th.th_current_task->td_icvs, loc); | |||
2203 | ||||
2204 | #if OMPT_SUPPORT1 | |||
2205 | master_th->th.ompt_thread_info.state = ompt_state_work_parallel; | |||
2206 | #endif | |||
2207 | ||||
2208 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | |||
2209 | ||||
2210 | #if USE_ITT_BUILD1 | |||
2211 | if (team->t.t_active_level == 1 // only report frames at level 1 | |||
2212 | && !master_th->th.th_teams_microtask) { // not in teams construct | |||
2213 | #if USE_ITT_NOTIFY1 | |||
2214 | if ((__itt_frame_submit_v3_ptr__kmp_itt_frame_submit_v3_ptr__3_0 || KMP_ITT_DEBUG0) && | |||
2215 | (__kmp_forkjoin_frames_mode == 3 || | |||
2216 | __kmp_forkjoin_frames_mode == 1)) { | |||
2217 | kmp_uint64 tmp_time = 0; | |||
2218 | if (__itt_get_timestamp_ptr__kmp_itt_get_timestamp_ptr__3_0) | |||
2219 | tmp_time = __itt_get_timestamp(!__kmp_itt_get_timestamp_ptr__3_0) ? 0 : __kmp_itt_get_timestamp_ptr__3_0(); | |||
2220 | // Internal fork - report frame begin | |||
2221 | master_th->th.th_frame_time = tmp_time; | |||
2222 | if (__kmp_forkjoin_frames_mode == 3) | |||
2223 | team->t.t_region_time = tmp_time; | |||
2224 | } else | |||
2225 | // only one notification scheme (either "submit" or "forking/joined", not both) | |||
2226 | #endif /* USE_ITT_NOTIFY */ | |||
2227 | if ((__itt_frame_begin_v3_ptr__kmp_itt_frame_begin_v3_ptr__3_0 || KMP_ITT_DEBUG0) && | |||
2228 | __kmp_forkjoin_frames && !__kmp_forkjoin_frames_mode) { | |||
2229 | // Mark start of "parallel" region for Intel(R) VTune(TM) analyzer. | |||
2230 | __kmp_itt_region_forking(gtid, team->t.t_nproc, 0); | |||
2231 | } | |||
2232 | } | |||
2233 | #endif /* USE_ITT_BUILD */ | |||
2234 | ||||
2235 | /* now go on and do the work */ | |||
2236 | 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" , 2236); }; | |||
2237 | KMP_MB(); | |||
2238 | 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); } | |||
2239 | ("__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); } | |||
2240 | 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); }; | |||
2241 | ||||
2242 | #if USE_ITT_BUILD1 | |||
2243 | if (__itt_stack_caller_create_ptr__kmp_itt_stack_caller_create_ptr__3_0) { | |||
2244 | // create new stack stitching id before entering fork barrier | |||
2245 | if (!enter_teams) { | |||
2246 | 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" , 2246); }; | |||
2247 | team->t.t_stack_id = __kmp_itt_stack_caller_create(); | |||
2248 | } else if (parent_team->t.t_serialized) { | |||
2249 | // keep stack stitching id in the serialized parent_team; | |||
2250 | // current team will be used for parallel inside the teams; | |||
2251 | // if parent_team is active, then it already keeps stack stitching id | |||
2252 | // for the league of teams | |||
2253 | 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" , 2253); }; | |||
2254 | parent_team->t.t_stack_id = __kmp_itt_stack_caller_create(); | |||
2255 | } | |||
2256 | } | |||
2257 | #endif /* USE_ITT_BUILD */ | |||
2258 | ||||
2259 | // AC: skip __kmp_internal_fork at teams construct, let only primary | |||
2260 | // threads execute | |||
2261 | if (ap) { | |||
2262 | __kmp_internal_fork(loc, gtid, team); | |||
2263 | 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); } | |||
2264 | "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); } | |||
2265 | 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); }; | |||
2266 | } | |||
2267 | ||||
2268 | if (call_context == fork_context_gnu) { | |||
2269 | 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); }; | |||
2270 | return TRUE(!0); | |||
2271 | } | |||
2272 | ||||
2273 | /* Invoke microtask for PRIMARY thread */ | |||
2274 | 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); } | |||
2275 | 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); }; | |||
2276 | } // END of timer KMP_fork_call block | |||
2277 | ||||
2278 | #if KMP_STATS_ENABLED0 | |||
2279 | // If beginning a teams construct, then change thread state | |||
2280 | stats_state_e previous_state = KMP_GET_THREAD_STATE()((void)0); | |||
2281 | if (!ap) { | |||
2282 | KMP_SET_THREAD_STATE(stats_state_e::TEAMS_REGION)((void)0); | |||
2283 | } | |||
2284 | #endif | |||
2285 | ||||
2286 | if (!team->t.t_invoke(gtid)) { | |||
2287 | 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", 2287); }; | |||
2288 | } | |||
2289 | ||||
2290 | #if KMP_STATS_ENABLED0 | |||
2291 | // If was beginning of a teams construct, then reset thread state | |||
2292 | if (!ap) { | |||
2293 | KMP_SET_THREAD_STATE(previous_state)((void)0); | |||
2294 | } | |||
2295 | #endif | |||
2296 | ||||
2297 | 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); } | |||
2298 | 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); }; | |||
2299 | KMP_MB(); /* Flush all pending memory write invalidates. */ | |||
2300 | ||||
2301 | 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); }; | |||
2302 | #if OMPT_SUPPORT1 | |||
2303 | if (ompt_enabled.enabled) { | |||
2304 | master_th->th.ompt_thread_info.state = ompt_state_overhead; | |||
2305 | } | |||
2306 | #endif | |||
2307 | ||||
2308 | return TRUE(!0); | |||
2309 | } | |||
2310 | ||||
2311 | #if OMPT_SUPPORT1 | |||
2312 | static inline void __kmp_join_restore_state(kmp_info_t *thread, | |||
2313 | kmp_team_t *team) { | |||
2314 | // restore state outside the region | |||
2315 | thread->th.ompt_thread_info.state = | |||
2316 | ((team->t.t_serialized) ? ompt_state_work_serial | |||
2317 | : ompt_state_work_parallel); | |||
2318 | } | |||
2319 | ||||
2320 | static inline void __kmp_join_ompt(int gtid, kmp_info_t *thread, | |||
2321 | kmp_team_t *team, ompt_data_t *parallel_data, | |||
2322 | int flags, void *codeptr) { | |||
2323 | ompt_task_info_t *task_info = __ompt_get_task_info_object(0); | |||
2324 | if (ompt_enabled.ompt_callback_parallel_end) { | |||
2325 | ompt_callbacks.ompt_callback(ompt_callback_parallel_end)ompt_callback_parallel_end_callback( | |||
2326 | parallel_data, &(task_info->task_data), flags, codeptr); | |||
2327 | } | |||
2328 | ||||
2329 | task_info->frame.enter_frame = ompt_data_none{0}; | |||
2330 | __kmp_join_restore_state(thread, team); | |||
2331 | } | |||
2332 | #endif | |||
2333 | ||||
2334 | void __kmp_join_call(ident_t *loc, int gtid | |||
2335 | #if OMPT_SUPPORT1 | |||
2336 | , | |||
2337 | enum fork_context_e fork_context | |||
2338 | #endif | |||
2339 | , | |||
2340 | int exit_teams) { | |||
2341 | KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_join_call)((void)0); | |||
2342 | kmp_team_t *team; | |||
2343 | kmp_team_t *parent_team; | |||
2344 | kmp_info_t *master_th; | |||
2345 | kmp_root_t *root; | |||
2346 | int master_active; | |||
2347 | ||||
2348 | 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); }; | |||
2349 | ||||
2350 | /* setup current data */ | |||
2351 | master_th = __kmp_threads[gtid]; | |||
2352 | root = master_th->th.th_root; | |||
2353 | team = master_th->th.th_team; | |||
2354 | parent_team = team->t.t_parent; | |||
2355 | ||||
2356 | master_th->th.th_ident = loc; | |||
2357 | ||||
2358 | #if OMPT_SUPPORT1 | |||
2359 | void *team_microtask = (void *)team->t.t_pkfn; | |||
2360 | // For GOMP interface with serialized parallel, need the | |||
2361 | // __kmpc_end_serialized_parallel to call hooks for OMPT end-implicit-task | |||
2362 | // and end-parallel events. | |||
2363 | if (ompt_enabled.enabled && | |||
2364 | !(team->t.t_serialized && fork_context == fork_context_gnu)) { | |||
2365 | master_th->th.ompt_thread_info.state = ompt_state_overhead; | |||
2366 | } | |||
2367 | #endif | |||
2368 | ||||
2369 | #if KMP_DEBUG1 | |||
2370 | if (__kmp_tasking_mode != tskm_immediate_exec && !exit_teams) { | |||
2371 | 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); } | |||
2372 | "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); } | |||
2373 | __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); } | |||
2374 | 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); } | |||
2375 | 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); }; | |||
2376 | 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", 2377); } | |||
2377 | 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", 2377); }; | |||
2378 | } | |||
2379 | #endif | |||
2380 | ||||
2381 | if (team->t.t_serialized) { | |||
2382 | if (master_th->th.th_teams_microtask) { | |||
2383 | // We are in teams construct | |||
2384 | int level = team->t.t_level; | |||
2385 | int tlevel = master_th->th.th_teams_level; | |||
2386 | if (level == tlevel) { | |||
2387 | // AC: we haven't incremented it earlier at start of teams construct, | |||
2388 | // so do it here - at the end of teams construct | |||
2389 | team->t.t_level++; | |||
2390 | } else if (level == tlevel + 1) { | |||
2391 | // AC: we are exiting parallel inside teams, need to increment | |||
2392 | // serialization in order to restore it in the next call to | |||
2393 | // __kmpc_end_serialized_parallel | |||
2394 | team->t.t_serialized++; | |||
2395 | } | |||
2396 | } | |||
2397 | __kmpc_end_serialized_parallel(loc, gtid); | |||
2398 | ||||
2399 | #if OMPT_SUPPORT1 | |||
2400 | if (ompt_enabled.enabled) { | |||
2401 | if (fork_context == fork_context_gnu) { | |||
2402 | __ompt_lw_taskteam_unlink(master_th); | |||
2403 | } | |||
2404 | __kmp_join_restore_state(master_th, parent_team); | |||
2405 | } | |||
2406 | #endif | |||
2407 | ||||
2408 | return; | |||
2409 | } | |||
2410 | ||||
2411 | master_active = team->t.t_master_active; | |||
2412 | ||||
2413 | if (!exit_teams) { | |||
2414 | // AC: No barrier for internal teams at exit from teams construct. | |||
2415 | // But there is barrier for external team (league). | |||
2416 | __kmp_internal_join(loc, gtid, team); | |||
2417 | #if USE_ITT_BUILD1 | |||
2418 | if (__itt_stack_caller_create_ptr__kmp_itt_stack_caller_create_ptr__3_0) { | |||
2419 | 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" , 2419); }; | |||
2420 | // destroy the stack stitching id after join barrier | |||
2421 | __kmp_itt_stack_caller_destroy((__itt_caller)team->t.t_stack_id); | |||
2422 | team->t.t_stack_id = NULL__null; | |||
2423 | } | |||
2424 | #endif | |||
2425 | } else { | |||
2426 | master_th->th.th_task_state = | |||
2427 | 0; // AC: no tasking in teams (out of any parallel) | |||
2428 | #if USE_ITT_BUILD1 | |||
2429 | if (__itt_stack_caller_create_ptr__kmp_itt_stack_caller_create_ptr__3_0 && parent_team->t.t_serialized) { | |||
2430 | 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" , 2430); }; | |||
2431 | // destroy the stack stitching id on exit from the teams construct | |||
2432 | // if parent_team is active, then the id will be destroyed later on | |||
2433 | // by master of the league of teams | |||
2434 | __kmp_itt_stack_caller_destroy((__itt_caller)parent_team->t.t_stack_id); | |||
2435 | parent_team->t.t_stack_id = NULL__null; | |||
2436 | } | |||
2437 | #endif | |||
2438 | ||||
2439 | if (team->t.t_nproc > 1 && | |||
2440 | __kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | |||
2441 | team->t.b->update_num_threads(team->t.t_nproc); | |||
2442 | __kmp_add_threads_to_team(team, team->t.t_nproc); | |||
2443 | } | |||
2444 | } | |||
2445 | ||||
2446 | KMP_MB(); | |||
2447 | ||||
2448 | #if OMPT_SUPPORT1 | |||
2449 | ompt_data_t *parallel_data = &(team->t.ompt_team_info.parallel_data); | |||
2450 | void *codeptr = team->t.ompt_team_info.master_return_address; | |||
2451 | #endif | |||
2452 | ||||
2453 | #if USE_ITT_BUILD1 | |||
2454 | // Mark end of "parallel" region for Intel(R) VTune(TM) analyzer. | |||
2455 | if (team->t.t_active_level == 1 && | |||
2456 | (!master_th->th.th_teams_microtask || /* not in teams construct */ | |||
2457 | master_th->th.th_teams_size.nteams == 1)) { | |||
2458 | master_th->th.th_ident = loc; | |||
2459 | // only one notification scheme (either "submit" or "forking/joined", not | |||
2460 | // both) | |||
2461 | if ((__itt_frame_submit_v3_ptr__kmp_itt_frame_submit_v3_ptr__3_0 || KMP_ITT_DEBUG0) && | |||
2462 | __kmp_forkjoin_frames_mode == 3) | |||
2463 | __kmp_itt_frame_submit(gtid, team->t.t_region_time, | |||
2464 | master_th->th.th_frame_time, 0, loc, | |||
2465 | master_th->th.th_team_nproc, 1); | |||
2466 | else if ((__itt_frame_end_v3_ptr__kmp_itt_frame_end_v3_ptr__3_0 || KMP_ITT_DEBUG0) && | |||
2467 | !__kmp_forkjoin_frames_mode && __kmp_forkjoin_frames) | |||
2468 | __kmp_itt_region_joined(gtid); | |||
2469 | } // active_level == 1 | |||
2470 | #endif /* USE_ITT_BUILD */ | |||
2471 | ||||
2472 | #if KMP_AFFINITY_SUPPORTED1 | |||
2473 | if (!exit_teams) { | |||
2474 | // Restore master thread's partition. | |||
2475 | master_th->th.th_first_place = team->t.t_first_place; | |||
2476 | master_th->th.th_last_place = team->t.t_last_place; | |||
2477 | } | |||
2478 | #endif // KMP_AFFINITY_SUPPORTED | |||
2479 | ||||
2480 | if (master_th->th.th_teams_microtask && !exit_teams && | |||
2481 | team->t.t_pkfn != (microtask_t)__kmp_teams_master && | |||
2482 | team->t.t_level == master_th->th.th_teams_level + 1) { | |||
2483 | // AC: We need to leave the team structure intact at the end of parallel | |||
2484 | // inside the teams construct, so that at the next parallel same (hot) team | |||
2485 | // works, only adjust nesting levels | |||
2486 | #if OMPT_SUPPORT1 | |||
2487 | ompt_data_t ompt_parallel_data = ompt_data_none{0}; | |||
2488 | if (ompt_enabled.enabled) { | |||
2489 | ompt_task_info_t *task_info = __ompt_get_task_info_object(0); | |||
2490 | if (ompt_enabled.ompt_callback_implicit_task) { | |||
2491 | int ompt_team_size = team->t.t_nproc; | |||
2492 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | |||
2493 | ompt_scope_end, NULL__null, &(task_info->task_data), ompt_team_size, | |||
2494 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num, ompt_task_implicit); | |||
2495 | } | |||
2496 | task_info->frame.exit_frame = ompt_data_none{0}; | |||
2497 | task_info->task_data = ompt_data_none{0}; | |||
2498 | ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th)(&(master_th->th.th_team->t.ompt_team_info.parallel_data )); | |||
2499 | __ompt_lw_taskteam_unlink(master_th); | |||
2500 | } | |||
2501 | #endif | |||
2502 | /* Decrement our nested depth level */ | |||
2503 | team->t.t_level--; | |||
2504 | team->t.t_active_level--; | |||
2505 | KMP_ATOMIC_DEC(&root->r.r_in_parallel)(&root->r.r_in_parallel)->fetch_sub(1, std::memory_order_acq_rel ); | |||
2506 | ||||
2507 | // Restore number of threads in the team if needed. This code relies on | |||
2508 | // the proper adjustment of th_teams_size.nth after the fork in | |||
2509 | // __kmp_teams_master on each teams primary thread in the case that | |||
2510 | // __kmp_reserve_threads reduced it. | |||
2511 | if (master_th->th.th_team_nproc < master_th->th.th_teams_size.nth) { | |||
2512 | int old_num = master_th->th.th_team_nproc; | |||
2513 | int new_num = master_th->th.th_teams_size.nth; | |||
2514 | kmp_info_t **other_threads = team->t.t_threads; | |||
2515 | team->t.t_nproc = new_num; | |||
2516 | for (int i = 0; i < old_num; ++i) { | |||
2517 | other_threads[i]->th.th_team_nproc = new_num; | |||
2518 | } | |||
2519 | // Adjust states of non-used threads of the team | |||
2520 | for (int i = old_num; i < new_num; ++i) { | |||
2521 | // Re-initialize thread's barrier data. | |||
2522 | KMP_DEBUG_ASSERT(other_threads[i])if (!(other_threads[i])) { __kmp_debug_assert("other_threads[i]" , "openmp/runtime/src/kmp_runtime.cpp", 2522); }; | |||
2523 | kmp_balign_t *balign = other_threads[i]->th.th_bar; | |||
2524 | for (int b = 0; b < bs_last_barrier; ++b) { | |||
2525 | balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; | |||
2526 | 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", 2526); }; | |||
2527 | #if USE_DEBUGGER0 | |||
2528 | balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; | |||
2529 | #endif | |||
2530 | } | |||
2531 | if (__kmp_tasking_mode != tskm_immediate_exec) { | |||
2532 | // Synchronize thread's task state | |||
2533 | other_threads[i]->th.th_task_state = master_th->th.th_task_state; | |||
2534 | } | |||
2535 | } | |||
2536 | } | |||
2537 | ||||
2538 | #if OMPT_SUPPORT1 | |||
2539 | if (ompt_enabled.enabled) { | |||
2540 | __kmp_join_ompt(gtid, master_th, parent_team, &ompt_parallel_data, | |||
2541 | OMPT_INVOKER(fork_context)((fork_context == fork_context_gnu) ? ompt_parallel_invoker_program : ompt_parallel_invoker_runtime) | ompt_parallel_team, codeptr); | |||
2542 | } | |||
2543 | #endif | |||
2544 | ||||
2545 | return; | |||
2546 | } | |||
2547 | ||||
2548 | /* do cleanup and restore the parent team */ | |||
2549 | master_th->th.th_info.ds.ds_tid = team->t.t_master_tid; | |||
2550 | master_th->th.th_local.this_construct = team->t.t_master_this_cons; | |||
2551 | ||||
2552 | master_th->th.th_dispatch = &parent_team->t.t_dispatch[team->t.t_master_tid]; | |||
2553 | ||||
2554 | /* jc: The following lock has instructions with REL and ACQ semantics, | |||
2555 | separating the parallel user code called in this parallel region | |||
2556 | from the serial user code called after this function returns. */ | |||
2557 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); | |||
2558 | ||||
2559 | if (!master_th->th.th_teams_microtask || | |||
2560 | team->t.t_level > master_th->th.th_teams_level) { | |||
2561 | /* Decrement our nested depth level */ | |||
2562 | KMP_ATOMIC_DEC(&root->r.r_in_parallel)(&root->r.r_in_parallel)->fetch_sub(1, std::memory_order_acq_rel ); | |||
2563 | } | |||
2564 | 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" , 2564); }; | |||
2565 | ||||
2566 | #if OMPT_SUPPORT1 | |||
2567 | if (ompt_enabled.enabled) { | |||
2568 | ompt_task_info_t *task_info = __ompt_get_task_info_object(0); | |||
2569 | if (ompt_enabled.ompt_callback_implicit_task) { | |||
2570 | int flags = (team_microtask == (void *)__kmp_teams_master) | |||
2571 | ? ompt_task_initial | |||
2572 | : ompt_task_implicit; | |||
2573 | int ompt_team_size = (flags == ompt_task_initial) ? 0 : team->t.t_nproc; | |||
2574 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | |||
2575 | ompt_scope_end, NULL__null, &(task_info->task_data), ompt_team_size, | |||
2576 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num, flags); | |||
2577 | } | |||
2578 | task_info->frame.exit_frame = ompt_data_none{0}; | |||
2579 | task_info->task_data = ompt_data_none{0}; | |||
2580 | } | |||
2581 | #endif | |||
2582 | ||||
2583 | 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); } | |||
2584 | 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); }; | |||
2585 | __kmp_pop_current_task_from_thread(master_th); | |||
2586 | ||||
2587 | master_th->th.th_def_allocator = team->t.t_def_allocator; | |||
2588 | ||||
2589 | #if OMPD_SUPPORT1 | |||
2590 | if (ompd_state & OMPD_ENABLE_BP0x1) | |||
2591 | ompd_bp_parallel_end(); | |||
2592 | #endif | |||
2593 | updateHWFPControl(team); | |||
2594 | ||||
2595 | if (root->r.r_active != master_active) | |||
2596 | root->r.r_active = master_active; | |||
2597 | ||||
2598 | __kmp_free_team(root, team USE_NESTED_HOT_ARG(, master_th | |||
2599 | master_th), master_th); // this will free worker threads | |||
2600 | ||||
2601 | /* this race was fun to find. make sure the following is in the critical | |||
2602 | region otherwise assertions may fail occasionally since the old team may be | |||
2603 | reallocated and the hierarchy appears inconsistent. it is actually safe to | |||
2604 | run and won't cause any bugs, but will cause those assertion failures. it's | |||
2605 | only one deref&assign so might as well put this in the critical region */ | |||
2606 | master_th->th.th_team = parent_team; | |||
2607 | master_th->th.th_team_nproc = parent_team->t.t_nproc; | |||
2608 | master_th->th.th_team_master = parent_team->t.t_threads[0]; | |||
2609 | master_th->th.th_team_serialized = parent_team->t.t_serialized; | |||
2610 | ||||
2611 | /* restore serialized team, if need be */ | |||
2612 | if (parent_team->t.t_serialized && | |||
2613 | parent_team != master_th->th.th_serial_team && | |||
2614 | parent_team != root->r.r_root_team) { | |||
2615 | __kmp_free_team(root, | |||
2616 | master_th->th.th_serial_team USE_NESTED_HOT_ARG(NULL), __null); | |||
2617 | master_th->th.th_serial_team = parent_team; | |||
2618 | } | |||
2619 | ||||
2620 | if (__kmp_tasking_mode != tskm_immediate_exec) { | |||
2621 | if (master_th->th.th_task_state_top > | |||
2622 | 0) { // Restore task state from memo stack | |||
2623 | 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" , 2623); }; | |||
2624 | // Remember primary thread's state if we re-use this nested hot team | |||
2625 | master_th->th.th_task_state_memo_stack[master_th->th.th_task_state_top] = | |||
2626 | master_th->th.th_task_state; | |||
2627 | --master_th->th.th_task_state_top; // pop | |||
2628 | // Now restore state at this level | |||
2629 | master_th->th.th_task_state = | |||
2630 | master_th->th | |||
2631 | .th_task_state_memo_stack[master_th->th.th_task_state_top]; | |||
2632 | } | |||
2633 | // Copy the task team from the parent team to the primary thread | |||
2634 | master_th->th.th_task_team = | |||
2635 | parent_team->t.t_task_team[master_th->th.th_task_state]; | |||
2636 | 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); } | |||
2637 | ("__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); } | |||
2638 | __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); } | |||
2639 | 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); }; | |||
2640 | } | |||
2641 | ||||
2642 | // TODO: GEH - cannot do this assertion because root thread not set up as | |||
2643 | // executing | |||
2644 | // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 0 ); | |||
2645 | master_th->th.th_current_task->td_flags.executing = 1; | |||
2646 | ||||
2647 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | |||
2648 | ||||
2649 | #if KMP_AFFINITY_SUPPORTED1 | |||
2650 | if (master_th->th.th_team->t.t_level == 0 && __kmp_affinity.flags.reset) { | |||
2651 | __kmp_reset_root_init_mask(gtid); | |||
2652 | } | |||
2653 | #endif | |||
2654 | #if OMPT_SUPPORT1 | |||
2655 | int flags = | |||
2656 | OMPT_INVOKER(fork_context)((fork_context == fork_context_gnu) ? ompt_parallel_invoker_program : ompt_parallel_invoker_runtime) | | |||
2657 | ((team_microtask == (void *)__kmp_teams_master) ? ompt_parallel_league | |||
2658 | : ompt_parallel_team); | |||
2659 | if (ompt_enabled.enabled) { | |||
2660 | __kmp_join_ompt(gtid, master_th, parent_team, parallel_data, flags, | |||
2661 | codeptr); | |||
2662 | } | |||
2663 | #endif | |||
2664 | ||||
2665 | KMP_MB(); | |||
2666 | 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); }; | |||
2667 | } | |||
2668 | ||||
2669 | /* Check whether we should push an internal control record onto the | |||
2670 | serial team stack. If so, do it. */ | |||
2671 | void __kmp_save_internal_controls(kmp_info_t *thread) { | |||
2672 | ||||
2673 | if (thread->th.th_team != thread->th.th_serial_team) { | |||
2674 | return; | |||
2675 | } | |||
2676 | if (thread->th.th_team->t.t_serialized > 1) { | |||
2677 | int push = 0; | |||
2678 | ||||
2679 | if (thread->th.th_team->t.t_control_stack_top == NULL__null) { | |||
2680 | push = 1; | |||
2681 | } else { | |||
2682 | if (thread->th.th_team->t.t_control_stack_top->serial_nesting_level != | |||
2683 | thread->th.th_team->t.t_serialized) { | |||
2684 | push = 1; | |||
2685 | } | |||
2686 | } | |||
2687 | if (push) { /* push a record on the serial team's stack */ | |||
2688 | kmp_internal_control_t *control = | |||
2689 | (kmp_internal_control_t *)__kmp_allocate(___kmp_allocate((sizeof(kmp_internal_control_t)), "openmp/runtime/src/kmp_runtime.cpp" , 2690) | |||
2690 | sizeof(kmp_internal_control_t))___kmp_allocate((sizeof(kmp_internal_control_t)), "openmp/runtime/src/kmp_runtime.cpp" , 2690); | |||
2691 | ||||
2692 | copy_icvs(control, &thread->th.th_current_task->td_icvs); | |||
2693 | ||||
2694 | control->serial_nesting_level = thread->th.th_team->t.t_serialized; | |||
2695 | ||||
2696 | control->next = thread->th.th_team->t.t_control_stack_top; | |||
2697 | thread->th.th_team->t.t_control_stack_top = control; | |||
2698 | } | |||
2699 | } | |||
2700 | } | |||
2701 | ||||
2702 | /* Changes set_nproc */ | |||
2703 | void __kmp_set_num_threads(int new_nth, int gtid) { | |||
2704 | kmp_info_t *thread; | |||
2705 | kmp_root_t *root; | |||
2706 | ||||
2707 | 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); }; | |||
2708 | KMP_DEBUG_ASSERT(__kmp_init_serial)if (!(__kmp_init_serial)) { __kmp_debug_assert("__kmp_init_serial" , "openmp/runtime/src/kmp_runtime.cpp", 2708); }; | |||
2709 | ||||
2710 | if (new_nth < 1) | |||
2711 | new_nth = 1; | |||
2712 | else if (new_nth > __kmp_max_nth) | |||
2713 | new_nth = __kmp_max_nth; | |||
2714 | ||||
2715 | KMP_COUNT_VALUE(OMP_set_numthreads, new_nth)((void)0); | |||
2716 | thread = __kmp_threads[gtid]; | |||
2717 | if (thread->th.th_current_task->td_icvs.nproc == new_nth) | |||
2718 | return; // nothing to do | |||
2719 | ||||
2720 | __kmp_save_internal_controls(thread); | |||
2721 | ||||
2722 | set__nproc(thread, new_nth)(((thread)->th.th_current_task->td_icvs.nproc) = (new_nth )); | |||
2723 | ||||
2724 | // If this omp_set_num_threads() call will cause the hot team size to be | |||
2725 | // reduced (in the absence of a num_threads clause), then reduce it now, | |||
2726 | // rather than waiting for the next parallel region. | |||
2727 | root = thread->th.th_root; | |||
2728 | if (__kmp_init_parallel && (!root->r.r_active) && | |||
2729 | (root->r.r_hot_team->t.t_nproc > new_nth) | |||
2730 | #if KMP_NESTED_HOT_TEAMS1 | |||
2731 | && __kmp_hot_teams_max_level && !__kmp_hot_teams_mode | |||
2732 | #endif | |||
2733 | ) { | |||
2734 | kmp_team_t *hot_team = root->r.r_hot_team; | |||
2735 | int f; | |||
2736 | ||||
2737 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); | |||
2738 | ||||
2739 | if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | |||
2740 | __kmp_resize_dist_barrier(hot_team, hot_team->t.t_nproc, new_nth); | |||
2741 | } | |||
2742 | // Release the extra threads we don't need any more. | |||
2743 | for (f = new_nth; f < hot_team->t.t_nproc; f++) { | |||
2744 | 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" , 2744); }; | |||
2745 | if (__kmp_tasking_mode != tskm_immediate_exec) { | |||
2746 | // When decreasing team size, threads no longer in the team should unref | |||
2747 | // task team. | |||
2748 | hot_team->t.t_threads[f]->th.th_task_team = NULL__null; | |||
2749 | } | |||
2750 | __kmp_free_thread(hot_team->t.t_threads[f]); | |||
2751 | hot_team->t.t_threads[f] = NULL__null; | |||
2752 | } | |||
2753 | hot_team->t.t_nproc = new_nth; | |||
2754 | #if KMP_NESTED_HOT_TEAMS1 | |||
2755 | if (thread->th.th_hot_teams) { | |||
2756 | 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" , 2756); }; | |||
2757 | thread->th.th_hot_teams[0].hot_team_nth = new_nth; | |||
2758 | } | |||
2759 | #endif | |||
2760 | ||||
2761 | if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | |||
2762 | hot_team->t.b->update_num_threads(new_nth); | |||
2763 | __kmp_add_threads_to_team(hot_team, new_nth); | |||
2764 | } | |||
2765 | ||||
2766 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | |||
2767 | ||||
2768 | // Update the t_nproc field in the threads that are still active. | |||
2769 | for (f = 0; f < new_nth; f++) { | |||
2770 | 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" , 2770); }; | |||
2771 | hot_team->t.t_threads[f]->th.th_team_nproc = new_nth; | |||
2772 | } | |||
2773 | // Special flag in case omp_set_num_threads() call | |||
2774 | hot_team->t.t_size_changed = -1; | |||
2775 | } | |||
2776 | } | |||
2777 | ||||
2778 | /* Changes max_active_levels */ | |||
2779 | void __kmp_set_max_active_levels(int gtid, int max_active_levels) { | |||
2780 | kmp_info_t *thread; | |||
2781 | ||||
2782 | 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); } | |||
2783 | "%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); } | |||
2784 | 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); }; | |||
2785 | KMP_DEBUG_ASSERT(__kmp_init_serial)if (!(__kmp_init_serial)) { __kmp_debug_assert("__kmp_init_serial" , "openmp/runtime/src/kmp_runtime.cpp", 2785); }; | |||
2786 | ||||
2787 | // validate max_active_levels | |||
2788 | if (max_active_levels < 0) { | |||
2789 | KMP_WARNING(ActiveLevelsNegative, max_active_levels)__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_ActiveLevelsNegative , max_active_levels), __kmp_msg_null); | |||
2790 | // We ignore this call if the user has specified a negative value. | |||
2791 | // The current setting won't be changed. The last valid setting will be | |||
2792 | // used. A warning will be issued (if warnings are allowed as controlled by | |||
2793 | // the KMP_WARNINGS env var). | |||
2794 | 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 ); } | |||
2795 | "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 ); } | |||
2796 | 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 ); }; | |||
2797 | return; | |||
2798 | } | |||
2799 | if (max_active_levels <= KMP_MAX_ACTIVE_LEVELS_LIMIT2147483647) { | |||
2800 | // it's OK, the max_active_levels is within the valid range: [ 0; | |||
2801 | // KMP_MAX_ACTIVE_LEVELS_LIMIT ] | |||
2802 | // We allow a zero value. (implementation defined behavior) | |||
2803 | } else { | |||
2804 | 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) | |||
2805 | KMP_MAX_ACTIVE_LEVELS_LIMIT)__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_ActiveLevelsExceedLimit , max_active_levels, 2147483647), __kmp_msg_null); | |||
2806 | max_active_levels = KMP_MAX_ACTIVE_LEVELS_LIMIT2147483647; | |||
2807 | // Current upper limit is MAX_INT. (implementation defined behavior) | |||
2808 | // If the input exceeds the upper limit, we correct the input to be the | |||
2809 | // upper limit. (implementation defined behavior) | |||
2810 | // Actually, the flow should never get here until we use MAX_INT limit. | |||
2811 | } | |||
2812 | 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 ); } | |||
2813 | "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 ); } | |||
2814 | 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 ); }; | |||
2815 | ||||
2816 | thread = __kmp_threads[gtid]; | |||
2817 | ||||
2818 | __kmp_save_internal_controls(thread); | |||
2819 | ||||
2820 | set__max_active_levels(thread, max_active_levels)(((thread)->th.th_current_task->td_icvs.max_active_levels ) = (max_active_levels)); | |||
2821 | } | |||
2822 | ||||
2823 | /* Gets max_active_levels */ | |||
2824 | int __kmp_get_max_active_levels(int gtid) { | |||
2825 | kmp_info_t *thread; | |||
2826 | ||||
2827 | 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); }; | |||
2828 | KMP_DEBUG_ASSERT(__kmp_init_serial)if (!(__kmp_init_serial)) { __kmp_debug_assert("__kmp_init_serial" , "openmp/runtime/src/kmp_runtime.cpp", 2828); }; | |||
2829 | ||||
2830 | thread = __kmp_threads[gtid]; | |||
2831 | 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", 2831); }; | |||
2832 | 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 ); } | |||
2833 | "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 ); } | |||
2834 | 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 ); } | |||
2835 | 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 ); }; | |||
2836 | return thread->th.th_current_task->td_icvs.max_active_levels; | |||
2837 | } | |||
2838 | ||||
2839 | // nteams-var per-device ICV | |||
2840 | void __kmp_set_num_teams(int num_teams) { | |||
2841 | if (num_teams > 0) | |||
2842 | __kmp_nteams = num_teams; | |||
2843 | } | |||
2844 | int __kmp_get_max_teams(void) { return __kmp_nteams; } | |||
2845 | // teams-thread-limit-var per-device ICV | |||
2846 | void __kmp_set_teams_thread_limit(int limit) { | |||
2847 | if (limit > 0) | |||
2848 | __kmp_teams_thread_limit = limit; | |||
2849 | } | |||
2850 | int __kmp_get_teams_thread_limit(void) { return __kmp_teams_thread_limit; } | |||
2851 | ||||
2852 | KMP_BUILD_ASSERT(sizeof(kmp_sched_t) == sizeof(int))static_assert(sizeof(kmp_sched_t) == sizeof(int), "Build condition error" ); | |||
2853 | KMP_BUILD_ASSERT(sizeof(enum sched_type) == sizeof(int))static_assert(sizeof(enum sched_type) == sizeof(int), "Build condition error" ); | |||
2854 | ||||
2855 | /* Changes def_sched_var ICV values (run-time schedule kind and chunk) */ | |||
2856 | void __kmp_set_schedule(int gtid, kmp_sched_t kind, int chunk) { | |||
2857 | kmp_info_t *thread; | |||
2858 | kmp_sched_t orig_kind; | |||
2859 | // kmp_team_t *team; | |||
2860 | ||||
2861 | 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); } | |||
2862 | 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); }; | |||
2863 | KMP_DEBUG_ASSERT(__kmp_init_serial)if (!(__kmp_init_serial)) { __kmp_debug_assert("__kmp_init_serial" , "openmp/runtime/src/kmp_runtime.cpp", 2863); }; | |||
2864 | ||||
2865 | // Check if the kind parameter is valid, correct if needed. | |||
2866 | // Valid parameters should fit in one of two intervals - standard or extended: | |||
2867 | // <lower>, <valid>, <upper_std>, <lower_ext>, <valid>, <upper> | |||
2868 | // 2008-01-25: 0, 1 - 4, 5, 100, 101 - 102, 103 | |||
2869 | orig_kind = kind; | |||
2870 | kind = __kmp_sched_without_mods(kind); | |||
2871 | ||||
2872 | if (kind <= kmp_sched_lower || kind >= kmp_sched_upper || | |||
2873 | (kind <= kmp_sched_lower_ext && kind >= kmp_sched_upper_std)) { | |||
2874 | // TODO: Hint needs attention in case we change the default schedule. | |||
2875 | __kmp_msg(kmp_ms_warning, KMP_MSG(ScheduleKindOutOfRange, kind)__kmp_msg_format(kmp_i18n_msg_ScheduleKindOutOfRange, kind), | |||
2876 | KMP_HNT(DefaultScheduleKindUsed, "static, no chunk")__kmp_msg_format(kmp_i18n_hnt_DefaultScheduleKindUsed, "static, no chunk" ), | |||
2877 | __kmp_msg_null); | |||
2878 | kind = kmp_sched_default; | |||
2879 | chunk = 0; // ignore chunk value in case of bad kind | |||
2880 | } | |||
2881 | ||||
2882 | thread = __kmp_threads[gtid]; | |||
2883 | ||||
2884 | __kmp_save_internal_controls(thread); | |||
2885 | ||||
2886 | if (kind < kmp_sched_upper_std) { | |||
2887 | if (kind == kmp_sched_static && chunk < KMP_DEFAULT_CHUNK1) { | |||
2888 | // differ static chunked vs. unchunked: chunk should be invalid to | |||
2889 | // indicate unchunked schedule (which is the default) | |||
2890 | thread->th.th_current_task->td_icvs.sched.r_sched_type = kmp_sch_static; | |||
2891 | } else { | |||
2892 | thread->th.th_current_task->td_icvs.sched.r_sched_type = | |||
2893 | __kmp_sch_map[kind - kmp_sched_lower - 1]; | |||
2894 | } | |||
2895 | } else { | |||
2896 | // __kmp_sch_map[ kind - kmp_sched_lower_ext + kmp_sched_upper_std - | |||
2897 | // kmp_sched_lower - 2 ]; | |||
2898 | thread->th.th_current_task->td_icvs.sched.r_sched_type = | |||
2899 | __kmp_sch_map[kind - kmp_sched_lower_ext + kmp_sched_upper_std - | |||
2900 | kmp_sched_lower - 2]; | |||
2901 | } | |||
2902 | __kmp_sched_apply_mods_intkind( | |||
2903 | orig_kind, &(thread->th.th_current_task->td_icvs.sched.r_sched_type)); | |||
2904 | if (kind == kmp_sched_auto || chunk < 1) { | |||
2905 | // ignore parameter chunk for schedule auto | |||
2906 | thread->th.th_current_task->td_icvs.sched.chunk = KMP_DEFAULT_CHUNK1; | |||
2907 | } else { | |||
2908 | thread->th.th_current_task->td_icvs.sched.chunk = chunk; | |||
2909 | } | |||
2910 | } | |||
2911 | ||||
2912 | /* Gets def_sched_var ICV values */ | |||
2913 | void __kmp_get_schedule(int gtid, kmp_sched_t *kind, int *chunk) { | |||
2914 | kmp_info_t *thread; | |||
2915 | enum sched_type th_type; | |||
2916 | ||||
2917 | 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); }; | |||
2918 | KMP_DEBUG_ASSERT(__kmp_init_serial)if (!(__kmp_init_serial)) { __kmp_debug_assert("__kmp_init_serial" , "openmp/runtime/src/kmp_runtime.cpp", 2918); }; | |||
2919 | ||||
2920 | thread = __kmp_threads[gtid]; | |||
2921 | ||||
2922 | th_type = thread->th.th_current_task->td_icvs.sched.r_sched_type; | |||
2923 | switch (SCHEDULE_WITHOUT_MODIFIERS(th_type)(enum sched_type)( (th_type) & ~(kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic))) { | |||
2924 | case kmp_sch_static: | |||
2925 | case kmp_sch_static_greedy: | |||
2926 | case kmp_sch_static_balanced: | |||
2927 | *kind = kmp_sched_static; | |||
2928 | __kmp_sched_apply_mods_stdkind(kind, th_type); | |||
2929 | *chunk = 0; // chunk was not set, try to show this fact via zero value | |||
2930 | return; | |||
2931 | case kmp_sch_static_chunked: | |||
2932 | *kind = kmp_sched_static; | |||
2933 | break; | |||
2934 | case kmp_sch_dynamic_chunked: | |||
2935 | *kind = kmp_sched_dynamic; | |||
2936 | break; | |||
2937 | case kmp_sch_guided_chunked: | |||
2938 | case kmp_sch_guided_iterative_chunked: | |||
2939 | case kmp_sch_guided_analytical_chunked: | |||
2940 | *kind = kmp_sched_guided; | |||
2941 | break; | |||
2942 | case kmp_sch_auto: | |||
2943 | *kind = kmp_sched_auto; | |||
2944 | break; | |||
2945 | case kmp_sch_trapezoidal: | |||
2946 | *kind = kmp_sched_trapezoidal; | |||
2947 | break; | |||
2948 | #if KMP_STATIC_STEAL_ENABLED1 | |||
2949 | case kmp_sch_static_steal: | |||
2950 | *kind = kmp_sched_static_steal; | |||
2951 | break; | |||
2952 | #endif | |||
2953 | default: | |||
2954 | KMP_FATAL(UnknownSchedulingType, th_type)__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_UnknownSchedulingType , th_type), __kmp_msg_null); | |||
2955 | } | |||
2956 | ||||
2957 | __kmp_sched_apply_mods_stdkind(kind, th_type); | |||
2958 | *chunk = thread->th.th_current_task->td_icvs.sched.chunk; | |||
2959 | } | |||
2960 | ||||
2961 | int __kmp_get_ancestor_thread_num(int gtid, int level) { | |||
2962 | ||||
2963 | int ii, dd; | |||
2964 | kmp_team_t *team; | |||
2965 | kmp_info_t *thr; | |||
2966 | ||||
2967 | 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); }; | |||
2968 | KMP_DEBUG_ASSERT(__kmp_init_serial)if (!(__kmp_init_serial)) { __kmp_debug_assert("__kmp_init_serial" , "openmp/runtime/src/kmp_runtime.cpp", 2968); }; | |||
2969 | ||||
2970 | // validate level | |||
2971 | if (level == 0) | |||
2972 | return 0; | |||
2973 | if (level < 0) | |||
2974 | return -1; | |||
2975 | thr = __kmp_threads[gtid]; | |||
2976 | team = thr->th.th_team; | |||
2977 | ii = team->t.t_level; | |||
2978 | if (level > ii) | |||
2979 | return -1; | |||
2980 | ||||
2981 | if (thr->th.th_teams_microtask) { | |||
2982 | // AC: we are in teams region where multiple nested teams have same level | |||
2983 | int tlevel = thr->th.th_teams_level; // the level of the teams construct | |||
2984 | if (level <= | |||
2985 | tlevel) { // otherwise usual algorithm works (will not touch the teams) | |||
2986 | KMP_DEBUG_ASSERT(ii >= tlevel)if (!(ii >= tlevel)) { __kmp_debug_assert("ii >= tlevel" , "openmp/runtime/src/kmp_runtime.cpp", 2986); }; | |||
2987 | // AC: As we need to pass by the teams league, we need to artificially | |||
2988 | // increase ii | |||
2989 | if (ii == tlevel) { | |||
2990 | ii += 2; // three teams have same level | |||
2991 | } else { | |||
2992 | ii++; // two teams have same level | |||
2993 | } | |||
2994 | } | |||
2995 | } | |||
2996 | ||||
2997 | if (ii == level) | |||
2998 | return __kmp_tid_from_gtid(gtid); | |||
2999 | ||||
3000 | dd = team->t.t_serialized; | |||
3001 | level++; | |||
3002 | while (ii > level) { | |||
3003 | for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) { | |||
3004 | } | |||
3005 | if ((team->t.t_serialized) && (!dd)) { | |||
3006 | team = team->t.t_parent; | |||
3007 | continue; | |||
3008 | } | |||
3009 | if (ii > level) { | |||
3010 | team = team->t.t_parent; | |||
3011 | dd = team->t.t_serialized; | |||
3012 | ii--; | |||
3013 | } | |||
3014 | } | |||
3015 | ||||
3016 | return (dd > 1) ? (0) : (team->t.t_master_tid); | |||
3017 | } | |||
3018 | ||||
3019 | int __kmp_get_team_size(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_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); }; | |||
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 1; | |||
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 | while (ii > level) { | |||
3056 | for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) { | |||
3057 | } | |||
3058 | if (team->t.t_serialized && (!dd)) { | |||
3059 | team = team->t.t_parent; | |||
3060 | continue; | |||
3061 | } | |||
3062 | if (ii > level) { | |||
3063 | team = team->t.t_parent; | |||
3064 | ii--; | |||
3065 | } | |||
3066 | } | |||
3067 | ||||
3068 | return team->t.t_nproc; | |||
3069 | } | |||
3070 | ||||
3071 | kmp_r_sched_t __kmp_get_schedule_global() { | |||
3072 | // This routine created because pairs (__kmp_sched, __kmp_chunk) and | |||
3073 | // (__kmp_static, __kmp_guided) may be changed by kmp_set_defaults | |||
3074 | // independently. So one can get the updated schedule here. | |||
3075 | ||||
3076 | kmp_r_sched_t r_sched; | |||
3077 | ||||
3078 | // create schedule from 4 globals: __kmp_sched, __kmp_chunk, __kmp_static, | |||
3079 | // __kmp_guided. __kmp_sched should keep original value, so that user can set | |||
3080 | // KMP_SCHEDULE multiple times, and thus have different run-time schedules in | |||
3081 | // different roots (even in OMP 2.5) | |||
3082 | enum sched_type s = SCHEDULE_WITHOUT_MODIFIERS(__kmp_sched)(enum sched_type)( (__kmp_sched) & ~(kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)); | |||
3083 | enum sched_type sched_modifiers = SCHEDULE_GET_MODIFIERS(__kmp_sched)((enum sched_type)( (__kmp_sched) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic))); | |||
3084 | if (s == kmp_sch_static) { | |||
3085 | // replace STATIC with more detailed schedule (balanced or greedy) | |||
3086 | r_sched.r_sched_type = __kmp_static; | |||
3087 | } else if (s == kmp_sch_guided_chunked) { | |||
3088 | // replace GUIDED with more detailed schedule (iterative or analytical) | |||
3089 | r_sched.r_sched_type = __kmp_guided; | |||
3090 | } else { // (STATIC_CHUNKED), or (DYNAMIC_CHUNKED), or other | |||
3091 | r_sched.r_sched_type = __kmp_sched; | |||
3092 | } | |||
3093 | 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)); | |||
3094 | ||||
3095 | if (__kmp_chunk < KMP_DEFAULT_CHUNK1) { | |||
3096 | // __kmp_chunk may be wrong here (if it was not ever set) | |||
3097 | r_sched.chunk = KMP_DEFAULT_CHUNK1; | |||
3098 | } else { | |||
3099 | r_sched.chunk = __kmp_chunk; | |||
3100 | } | |||
3101 | ||||
3102 | return r_sched; | |||
3103 | } | |||
3104 | ||||
3105 | /* Allocate (realloc == FALSE) * or reallocate (realloc == TRUE) | |||
3106 | at least argc number of *t_argv entries for the requested team. */ | |||
3107 | static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team, int realloc) { | |||
3108 | ||||
3109 | KMP_DEBUG_ASSERT(team)if (!(team)) { __kmp_debug_assert("team", "openmp/runtime/src/kmp_runtime.cpp" , 3109); }; | |||
3110 | if (!realloc || argc > team->t.t_max_argc) { | |||
3111 | ||||
3112 | 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); } | |||
3113 | "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); } | |||
3114 | 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); }; | |||
3115 | /* if previously allocated heap space for args, free them */ | |||
3116 | if (realloc && team->t.t_argv != &team->t.t_inline_argv[0]) | |||
3117 | __kmp_free((void *)team->t.t_argv)___kmp_free(((void *)team->t.t_argv), "openmp/runtime/src/kmp_runtime.cpp" , 3117); | |||
3118 | ||||
3119 | 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 *)))) { | |||
3120 | /* use unused space in the cache line for arguments */ | |||
3121 | 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 *))); | |||
3122 | 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); } | |||
3123 | "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); } | |||
3124 | 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); }; | |||
3125 | team->t.t_argv = &team->t.t_inline_argv[0]; | |||
3126 | if (__kmp_storage_map) { | |||
3127 | __kmp_print_storage_map_gtid( | |||
3128 | -1, &team->t.t_inline_argv[0], | |||
3129 | &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 *)))], | |||
3130 | (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", | |||
3131 | team->t.t_id); | |||
3132 | } | |||
3133 | } else { | |||
3134 | /* allocate space for arguments in the heap */ | |||
3135 | team->t.t_max_argc = (argc <= (KMP_MIN_MALLOC_ARGV_ENTRIES100 >> 1)) | |||
3136 | ? KMP_MIN_MALLOC_ARGV_ENTRIES100 | |||
3137 | : 2 * argc; | |||
3138 | 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); } | |||
3139 | "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); } | |||
3140 | 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); }; | |||
3141 | team->t.t_argv = | |||
3142 | (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", 3142); | |||
3143 | if (__kmp_storage_map) { | |||
3144 | __kmp_print_storage_map_gtid(-1, &team->t.t_argv[0], | |||
3145 | &team->t.t_argv[team->t.t_max_argc], | |||
3146 | sizeof(void *) * team->t.t_max_argc, | |||
3147 | "team_%d.t_argv", team->t.t_id); | |||
3148 | } | |||
3149 | } | |||
3150 | } | |||
3151 | } | |||
3152 | ||||
3153 | static void __kmp_allocate_team_arrays(kmp_team_t *team, int max_nth) { | |||
3154 | int i; | |||
3155 | int num_disp_buff = max_nth > 1 ? __kmp_dispatch_num_buffers : 2; | |||
3156 | team->t.t_threads = | |||
3157 | (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" , 3157); | |||
3158 | 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", 3159) | |||
3159 | sizeof(dispatch_shared_info_t) * num_disp_buff)___kmp_allocate((sizeof(dispatch_shared_info_t) * num_disp_buff ), "openmp/runtime/src/kmp_runtime.cpp", 3159); | |||
3160 | team->t.t_dispatch = | |||
3161 | (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" , 3161); | |||
3162 | team->t.t_implicit_task_taskdata = | |||
3163 | (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" , 3163); | |||
3164 | team->t.t_max_nproc = max_nth; | |||
3165 | ||||
3166 | /* setup dispatch buffers */ | |||
3167 | for (i = 0; i < num_disp_buff; ++i) { | |||
3168 | team->t.t_disp_buffer[i].buffer_index = i; | |||
3169 | team->t.t_disp_buffer[i].doacross_buf_idx = i; | |||
3170 | } | |||
3171 | } | |||
3172 | ||||
3173 | static void __kmp_free_team_arrays(kmp_team_t *team) { | |||
3174 | /* Note: this does not free the threads in t_threads (__kmp_free_threads) */ | |||
3175 | int i; | |||
3176 | for (i = 0; i < team->t.t_max_nproc; ++i) { | |||
3177 | if (team->t.t_dispatch[i].th_disp_buffer != NULL__null) { | |||
3178 | __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" , 3178); | |||
3179 | team->t.t_dispatch[i].th_disp_buffer = NULL__null; | |||
3180 | } | |||
3181 | } | |||
3182 | #if KMP_USE_HIER_SCHED0 | |||
3183 | __kmp_dispatch_free_hierarchies(team); | |||
3184 | #endif | |||
3185 | __kmp_free(team->t.t_threads)___kmp_free((team->t.t_threads), "openmp/runtime/src/kmp_runtime.cpp" , 3185); | |||
3186 | __kmp_free(team->t.t_disp_buffer)___kmp_free((team->t.t_disp_buffer), "openmp/runtime/src/kmp_runtime.cpp" , 3186); | |||
3187 | __kmp_free(team->t.t_dispatch)___kmp_free((team->t.t_dispatch), "openmp/runtime/src/kmp_runtime.cpp" , 3187); | |||
3188 | __kmp_free(team->t.t_implicit_task_taskdata)___kmp_free((team->t.t_implicit_task_taskdata), "openmp/runtime/src/kmp_runtime.cpp" , 3188); | |||
3189 | team->t.t_threads = NULL__null; | |||
3190 | team->t.t_disp_buffer = NULL__null; | |||
3191 | team->t.t_dispatch = NULL__null; | |||
3192 | team->t.t_implicit_task_taskdata = 0; | |||
3193 | } | |||
3194 | ||||
3195 | static void __kmp_reallocate_team_arrays(kmp_team_t *team, int max_nth) { | |||
3196 | kmp_info_t **oldThreads = team->t.t_threads; | |||
3197 | ||||
3198 | __kmp_free(team->t.t_disp_buffer)___kmp_free((team->t.t_disp_buffer), "openmp/runtime/src/kmp_runtime.cpp" , 3198); | |||
3199 | __kmp_free(team->t.t_dispatch)___kmp_free((team->t.t_dispatch), "openmp/runtime/src/kmp_runtime.cpp" , 3199); | |||
3200 | __kmp_free(team->t.t_implicit_task_taskdata)___kmp_free((team->t.t_implicit_task_taskdata), "openmp/runtime/src/kmp_runtime.cpp" , 3200); | |||
3201 | __kmp_allocate_team_arrays(team, max_nth); | |||
3202 | ||||
3203 | KMP_MEMCPYmemcpy(team->t.t_threads, oldThreads, | |||
3204 | team->t.t_nproc * sizeof(kmp_info_t *)); | |||
3205 | ||||
3206 | __kmp_free(oldThreads)___kmp_free((oldThreads), "openmp/runtime/src/kmp_runtime.cpp" , 3206); | |||
3207 | } | |||
3208 | ||||
3209 | static kmp_internal_control_t __kmp_get_global_icvs(void) { | |||
3210 | ||||
3211 | kmp_r_sched_t r_sched = | |||
3212 | __kmp_get_schedule_global(); // get current state of scheduling globals | |||
3213 | ||||
3214 | 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" , 3214); }; | |||
3215 | ||||
3216 | kmp_internal_control_t g_icvs = { | |||
3217 | 0, // int serial_nesting_level; //corresponds to value of th_team_serialized | |||
3218 | (kmp_int8)__kmp_global.g.g_dynamic, // internal control for dynamic | |||
3219 | // adjustment of threads (per thread) | |||
3220 | (kmp_int8)__kmp_env_blocktime, // int bt_set; //internal control for | |||
3221 | // whether blocktime is explicitly set | |||
3222 | __kmp_dflt_blocktime, // int blocktime; //internal control for blocktime | |||
3223 | #if KMP_USE_MONITOR | |||
3224 | __kmp_bt_intervals, // int bt_intervals; //internal control for blocktime | |||
3225 | // intervals | |||
3226 | #endif | |||
3227 | __kmp_dflt_team_nth, // int nproc; //internal control for # of threads for | |||
3228 | // next parallel region (per thread) | |||
3229 | // (use a max ub on value if __kmp_parallel_initialize not called yet) | |||
3230 | __kmp_cg_max_nth, // int thread_limit; | |||
3231 | __kmp_dflt_max_active_levels, // int max_active_levels; //internal control | |||
3232 | // for max_active_levels | |||
3233 | r_sched, // kmp_r_sched_t sched; //internal control for runtime schedule | |||
3234 | // {sched,chunk} pair | |||
3235 | __kmp_nested_proc_bind.bind_types[0], | |||
3236 | __kmp_default_device, | |||
3237 | NULL__null // struct kmp_internal_control *next; | |||
3238 | }; | |||
3239 | ||||
3240 | return g_icvs; | |||
3241 | } | |||
3242 | ||||
3243 | static kmp_internal_control_t __kmp_get_x_global_icvs(const kmp_team_t *team) { | |||
3244 | ||||
3245 | kmp_internal_control_t gx_icvs; | |||
3246 | gx_icvs.serial_nesting_level = | |||
3247 | 0; // probably =team->t.t_serial like in save_inter_controls | |||
3248 | copy_icvs(&gx_icvs, &team->t.t_threads[0]->th.th_current_task->td_icvs); | |||
3249 | gx_icvs.next = NULL__null; | |||
3250 | ||||
3251 | return gx_icvs; | |||
3252 | } | |||
3253 | ||||
3254 | static void __kmp_initialize_root(kmp_root_t *root) { | |||
3255 | int f; | |||
3256 | kmp_team_t *root_team; | |||
3257 | kmp_team_t *hot_team; | |||
3258 | int hot_team_max_nth; | |||
3259 | kmp_r_sched_t r_sched = | |||
3260 | __kmp_get_schedule_global(); // get current state of scheduling globals | |||
3261 | kmp_internal_control_t r_icvs = __kmp_get_global_icvs(); | |||
3262 | KMP_DEBUG_ASSERT(root)if (!(root)) { __kmp_debug_assert("root", "openmp/runtime/src/kmp_runtime.cpp" , 3262); }; | |||
3263 | KMP_ASSERT(!root->r.r_begin)if (!(!root->r.r_begin)) { __kmp_debug_assert("!root->r.r_begin" , "openmp/runtime/src/kmp_runtime.cpp", 3263); }; | |||
3264 | ||||
3265 | /* setup the root state structure */ | |||
3266 | __kmp_init_lock(&root->r.r_begin_lock); | |||
3267 | root->r.r_begin = FALSE0; | |||
3268 | root->r.r_active = FALSE0; | |||
3269 | root->r.r_in_parallel = 0; | |||
3270 | root->r.r_blocktime = __kmp_dflt_blocktime; | |||
3271 | #if KMP_AFFINITY_SUPPORTED1 | |||
3272 | root->r.r_affinity_assigned = FALSE0; | |||
3273 | #endif | |||
3274 | ||||
3275 | /* setup the root team for this task */ | |||
3276 | /* allocate the root team structure */ | |||
3277 | 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" ); }; | |||
3278 | ||||
3279 | root_team = | |||
3280 | __kmp_allocate_team(root, | |||
3281 | 1, // new_nproc | |||
3282 | 1, // max_nproc | |||
3283 | #if OMPT_SUPPORT1 | |||
3284 | ompt_data_none{0}, // root parallel id | |||
3285 | #endif | |||
3286 | __kmp_nested_proc_bind.bind_types[0], &r_icvs, | |||
3287 | 0 // argc | |||
3288 | USE_NESTED_HOT_ARG(NULL), __null // primary thread is unknown | |||
3289 | ); | |||
3290 | #if USE_DEBUGGER0 | |||
3291 | // Non-NULL value should be assigned to make the debugger display the root | |||
3292 | // team. | |||
3293 | TCW_SYNC_PTR(root_team->t.t_pkfn, (microtask_t)(~0))((root_team->t.t_pkfn)) = (((microtask_t)(~0))); | |||
3294 | #endif | |||
3295 | ||||
3296 | 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); }; | |||
3297 | ||||
3298 | root->r.r_root_team = root_team; | |||
3299 | root_team->t.t_control_stack_top = NULL__null; | |||
3300 | ||||
3301 | /* initialize root team */ | |||
3302 | root_team->t.t_threads[0] = NULL__null; | |||
3303 | root_team->t.t_nproc = 1; | |||
3304 | root_team->t.t_serialized = 1; | |||
3305 | // TODO???: root_team->t.t_max_active_levels = __kmp_dflt_max_active_levels; | |||
3306 | root_team->t.t_sched.sched = r_sched.sched; | |||
3307 | 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); } | |||
3308 | 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); } | |||
3309 | ("__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); } | |||
3310 | 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); }; | |||
3311 | ||||
3312 | /* setup the hot team for this task */ | |||
3313 | /* allocate the hot team structure */ | |||
3314 | 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" ); }; | |||
3315 | ||||
3316 | hot_team = | |||
3317 | __kmp_allocate_team(root, | |||
3318 | 1, // new_nproc | |||
3319 | __kmp_dflt_team_nth_ub * 2, // max_nproc | |||
3320 | #if OMPT_SUPPORT1 | |||
3321 | ompt_data_none{0}, // root parallel id | |||
3322 | #endif | |||
3323 | __kmp_nested_proc_bind.bind_types[0], &r_icvs, | |||
3324 | 0 // argc | |||
3325 | USE_NESTED_HOT_ARG(NULL), __null // primary thread is unknown | |||
3326 | ); | |||
3327 | 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); }; | |||
3328 | ||||
3329 | root->r.r_hot_team = hot_team; | |||
3330 | root_team->t.t_control_stack_top = NULL__null; | |||
3331 | ||||
3332 | /* first-time initialization */ | |||
3333 | hot_team->t.t_parent = root_team; | |||
3334 | ||||
3335 | /* initialize hot team */ | |||
3336 | hot_team_max_nth = hot_team->t.t_max_nproc; | |||
3337 | for (f = 0; f < hot_team_max_nth; ++f) { | |||
3338 | hot_team->t.t_threads[f] = NULL__null; | |||
3339 | } | |||
3340 | hot_team->t.t_nproc = 1; | |||
3341 | // TODO???: hot_team->t.t_max_active_levels = __kmp_dflt_max_active_levels; | |||
3342 | hot_team->t.t_sched.sched = r_sched.sched; | |||
3343 | hot_team->t.t_size_changed = 0; | |||
3344 | } | |||
3345 | ||||
3346 | #ifdef KMP_DEBUG1 | |||
3347 | ||||
3348 | typedef struct kmp_team_list_item { | |||
3349 | kmp_team_p const *entry; | |||
3350 | struct kmp_team_list_item *next; | |||
3351 | } kmp_team_list_item_t; | |||
3352 | typedef kmp_team_list_item_t *kmp_team_list_t; | |||
3353 | ||||
3354 | static void __kmp_print_structure_team_accum( // Add team to list of teams. | |||
3355 | kmp_team_list_t list, // List of teams. | |||
3356 | kmp_team_p const *team // Team to add. | |||
3357 | ) { | |||
3358 | ||||
3359 | // List must terminate with item where both entry and next are NULL. | |||
3360 | // Team is added to the list only once. | |||
3361 | // List is sorted in ascending order by team id. | |||
3362 | // Team id is *not* a key. | |||
3363 | ||||
3364 | kmp_team_list_t l; | |||
3365 | ||||
3366 | KMP_DEBUG_ASSERT(list != NULL)if (!(list != __null)) { __kmp_debug_assert("list != __null", "openmp/runtime/src/kmp_runtime.cpp", 3366); }; | |||
3367 | if (team == NULL__null) { | |||
3368 | return; | |||
3369 | } | |||
3370 | ||||
3371 | __kmp_print_structure_team_accum(list, team->t.t_parent); | |||
3372 | __kmp_print_structure_team_accum(list, team->t.t_next_pool); | |||
3373 | ||||
3374 | // Search list for the team. | |||
3375 | l = list; | |||
3376 | while (l->next != NULL__null && l->entry != team) { | |||
3377 | l = l->next; | |||
3378 | } | |||
3379 | if (l->next != NULL__null) { | |||
3380 | return; // Team has been added before, exit. | |||
3381 | } | |||
3382 | ||||
3383 | // Team is not found. Search list again for insertion point. | |||
3384 | l = list; | |||
3385 | while (l->next != NULL__null && l->entry->t.t_id <= team->t.t_id) { | |||
3386 | l = l->next; | |||
3387 | } | |||
3388 | ||||
3389 | // Insert team. | |||
3390 | { | |||
3391 | kmp_team_list_item_t *item = (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(malloc(sizeof(kmp_team_list_item_t)) | |||
3392 | sizeof(kmp_team_list_item_t))malloc(sizeof(kmp_team_list_item_t)); | |||
3393 | *item = *l; | |||
3394 | l->entry = team; | |||
3395 | l->next = item; | |||
3396 | } | |||
3397 | } | |||
3398 | ||||
3399 | static void __kmp_print_structure_team(char const *title, kmp_team_p const *team | |||
3400 | ||||
3401 | ) { | |||
3402 | __kmp_printf("%s", title); | |||
3403 | if (team != NULL__null) { | |||
3404 | __kmp_printf("%2x %p\n", team->t.t_id, team); | |||
3405 | } else { | |||
3406 | __kmp_printf(" - (nil)\n"); | |||
3407 | } | |||
3408 | } | |||
3409 | ||||
3410 | static void __kmp_print_structure_thread(char const *title, | |||
3411 | kmp_info_p const *thread) { | |||
3412 | __kmp_printf("%s", title); | |||
3413 | if (thread != NULL__null) { | |||
3414 | __kmp_printf("%2d %p\n", thread->th.th_info.ds.ds_gtid, thread); | |||
3415 | } else { | |||
3416 | __kmp_printf(" - (nil)\n"); | |||
3417 | } | |||
3418 | } | |||
3419 | ||||
3420 | void __kmp_print_structure(void) { | |||
3421 | ||||
3422 | kmp_team_list_t list; | |||
3423 | ||||
3424 | // Initialize list of teams. | |||
3425 | list = | |||
3426 | (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(sizeof(kmp_team_list_item_t))malloc(sizeof(kmp_team_list_item_t)); | |||
3427 | list->entry = NULL__null; | |||
3428 | list->next = NULL__null; | |||
3429 | ||||
3430 | __kmp_printf("\n------------------------------\nGlobal Thread " | |||
3431 | "Table\n------------------------------\n"); | |||
3432 | { | |||
3433 | int gtid; | |||
3434 | for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) { | |||
3435 | __kmp_printf("%2d", gtid); | |||
3436 | if (__kmp_threads != NULL__null) { | |||
3437 | __kmp_printf(" %p", __kmp_threads[gtid]); | |||
3438 | } | |||
3439 | if (__kmp_root != NULL__null) { | |||
3440 | __kmp_printf(" %p", __kmp_root[gtid]); | |||
3441 | } | |||
3442 | __kmp_printf("\n"); | |||
3443 | } | |||
3444 | } | |||
3445 | ||||
3446 | // Print out __kmp_threads array. | |||
3447 | __kmp_printf("\n------------------------------\nThreads\n--------------------" | |||
3448 | "----------\n"); | |||
3449 | if (__kmp_threads != NULL__null) { | |||
3450 | int gtid; | |||
3451 | for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) { | |||
3452 | kmp_info_t const *thread = __kmp_threads[gtid]; | |||
3453 | if (thread != NULL__null) { | |||
3454 | __kmp_printf("GTID %2d %p:\n", gtid, thread); | |||
3455 | __kmp_printf(" Our Root: %p\n", thread->th.th_root); | |||
3456 | __kmp_print_structure_team(" Our Team: ", thread->th.th_team); | |||
3457 | __kmp_print_structure_team(" Serial Team: ", | |||
3458 | thread->th.th_serial_team); | |||
3459 | __kmp_printf(" Threads: %2d\n", thread->th.th_team_nproc); | |||
3460 | __kmp_print_structure_thread(" Primary: ", | |||
3461 | thread->th.th_team_master); | |||
3462 | __kmp_printf(" Serialized?: %2d\n", thread->th.th_team_serialized); | |||
3463 | __kmp_printf(" Set NProc: %2d\n", thread->th.th_set_nproc); | |||
3464 | __kmp_printf(" Set Proc Bind: %2d\n", thread->th.th_set_proc_bind); | |||
3465 | __kmp_print_structure_thread(" Next in pool: ", | |||
3466 | thread->th.th_next_pool); | |||
3467 | __kmp_printf("\n"); | |||
3468 | __kmp_print_structure_team_accum(list, thread->th.th_team); | |||
3469 | __kmp_print_structure_team_accum(list, thread->th.th_serial_team); | |||
3470 | } | |||
3471 | } | |||
3472 | } else { | |||
3473 | __kmp_printf("Threads array is not allocated.\n"); | |||
3474 | } | |||
3475 | ||||
3476 | // Print out __kmp_root array. | |||
3477 | __kmp_printf("\n------------------------------\nUbers\n----------------------" | |||
3478 | "--------\n"); | |||
3479 | if (__kmp_root != NULL__null) { | |||
3480 | int gtid; | |||
3481 | for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) { | |||
3482 | kmp_root_t const *root = __kmp_root[gtid]; | |||
3483 | if (root != NULL__null) { | |||
3484 | __kmp_printf("GTID %2d %p:\n", gtid, root); | |||
3485 | __kmp_print_structure_team(" Root Team: ", root->r.r_root_team); | |||
3486 | __kmp_print_structure_team(" Hot Team: ", root->r.r_hot_team); | |||
3487 | __kmp_print_structure_thread(" Uber Thread: ", | |||
3488 | root->r.r_uber_thread); | |||
3489 | __kmp_printf(" Active?: %2d\n", root->r.r_active); | |||
3490 | __kmp_printf(" In Parallel: %2d\n", | |||
3491 | KMP_ATOMIC_LD_RLX(&root->r.r_in_parallel)(&root->r.r_in_parallel)->load(std::memory_order_relaxed )); | |||
3492 | __kmp_printf("\n"); | |||
3493 | __kmp_print_structure_team_accum(list, root->r.r_root_team); | |||
3494 | __kmp_print_structure_team_accum(list, root->r.r_hot_team); | |||
3495 | } | |||
3496 | } | |||
3497 | } else { | |||
3498 | __kmp_printf("Ubers array is not allocated.\n"); | |||
3499 | } | |||
3500 | ||||
3501 | __kmp_printf("\n------------------------------\nTeams\n----------------------" | |||
3502 | "--------\n"); | |||
3503 | while (list->next != NULL__null) { | |||
3504 | kmp_team_p const *team = list->entry; | |||
3505 | int i; | |||
3506 | __kmp_printf("Team %2x %p:\n", team->t.t_id, team); | |||
3507 | __kmp_print_structure_team(" Parent Team: ", team->t.t_parent); | |||
3508 | __kmp_printf(" Primary TID: %2d\n", team->t.t_master_tid); | |||
3509 | __kmp_printf(" Max threads: %2d\n", team->t.t_max_nproc); | |||
3510 | __kmp_printf(" Levels of serial: %2d\n", team->t.t_serialized); | |||
3511 | __kmp_printf(" Number threads: %2d\n", team->t.t_nproc); | |||
3512 | for (i = 0; i < team->t.t_nproc; ++i) { | |||
3513 | __kmp_printf(" Thread %2d: ", i); | |||
3514 | __kmp_print_structure_thread("", team->t.t_threads[i]); | |||
3515 | } | |||
3516 | __kmp_print_structure_team(" Next in pool: ", team->t.t_next_pool); | |||
3517 | __kmp_printf("\n"); | |||
3518 | list = list->next; | |||
3519 | } | |||
3520 | ||||
3521 | // Print out __kmp_thread_pool and __kmp_team_pool. | |||
3522 | __kmp_printf("\n------------------------------\nPools\n----------------------" | |||
3523 | "--------\n"); | |||
3524 | __kmp_print_structure_thread("Thread pool: ", | |||
3525 | CCAST(kmp_info_t *, __kmp_thread_pool)const_cast<kmp_info_t *>(__kmp_thread_pool)); | |||
3526 | __kmp_print_structure_team("Team pool: ", | |||
3527 | CCAST(kmp_team_t *, __kmp_team_pool)const_cast<kmp_team_t *>(__kmp_team_pool)); | |||
3528 | __kmp_printf("\n"); | |||
3529 | ||||
3530 | // Free team list. | |||
3531 | while (list != NULL__null) { | |||
3532 | kmp_team_list_item_t *item = list; | |||
3533 | list = list->next; | |||
3534 | KMP_INTERNAL_FREE(item)free(item); | |||
3535 | } | |||
3536 | } | |||
3537 | ||||
3538 | #endif | |||
3539 | ||||
3540 | //--------------------------------------------------------------------------- | |||
3541 | // Stuff for per-thread fast random number generator | |||
3542 | // Table of primes | |||
3543 | static const unsigned __kmp_primes[] = { | |||
3544 | 0x9e3779b1, 0xffe6cc59, 0x2109f6dd, 0x43977ab5, 0xba5703f5, 0xb495a877, | |||
3545 | 0xe1626741, 0x79695e6b, 0xbc98c09f, 0xd5bee2b3, 0x287488f9, 0x3af18231, | |||
3546 | 0x9677cd4d, 0xbe3a6929, 0xadc6a877, 0xdcf0674b, 0xbe4d6fe9, 0x5f15e201, | |||
3547 | 0x99afc3fd, 0xf3f16801, 0xe222cfff, 0x24ba5fdb, 0x0620452d, 0x79f149e3, | |||
3548 | 0xc8b93f49, 0x972702cd, 0xb07dd827, 0x6c97d5ed, 0x085a3d61, 0x46eb5ea7, | |||
3549 | 0x3d9910ed, 0x2e687b5b, 0x29609227, 0x6eb081f1, 0x0954c4e1, 0x9d114db9, | |||
3550 | 0x542acfa9, 0xb3e6bd7b, 0x0742d917, 0xe9f3ffa7, 0x54581edb, 0xf2480f45, | |||
3551 | 0x0bb9288f, 0xef1affc7, 0x85fa0ca7, 0x3ccc14db, 0xe6baf34b, 0x343377f7, | |||
3552 | 0x5ca19031, 0xe6d9293b, 0xf0a9f391, 0x5d2e980b, 0xfc411073, 0xc3749363, | |||
3553 | 0xb892d829, 0x3549366b, 0x629750ad, 0xb98294e5, 0x892d9483, 0xc235baf3, | |||
3554 | 0x3d2402a3, 0x6bdef3c9, 0xbec333cd, 0x40c9520f}; | |||
3555 | ||||
3556 | //--------------------------------------------------------------------------- | |||
3557 | // __kmp_get_random: Get a random number using a linear congruential method. | |||
3558 | unsigned short __kmp_get_random(kmp_info_t *thread) { | |||
3559 | unsigned x = thread->th.th_x; | |||
3560 | unsigned short r = (unsigned short)(x >> 16); | |||
3561 | ||||
3562 | thread->th.th_x = x * thread->th.th_a + 1; | |||
3563 | ||||
3564 | 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); } | |||
3565 | 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); }; | |||
3566 | ||||
3567 | return r; | |||
3568 | } | |||
3569 | //-------------------------------------------------------- | |||
3570 | // __kmp_init_random: Initialize a random number generator | |||
3571 | void __kmp_init_random(kmp_info_t *thread) { | |||
3572 | unsigned seed = thread->th.th_info.ds.ds_tid; | |||
3573 | ||||
3574 | thread->th.th_a = | |||
3575 | __kmp_primes[seed % (sizeof(__kmp_primes) / sizeof(__kmp_primes[0]))]; | |||
3576 | thread->th.th_x = (seed + 1) * thread->th.th_a + 1; | |||
3577 | KA_TRACE(30,if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_init_random: THREAD: %u; A: %u\n" , seed, thread->th.th_a); } | |||
3578 | ("__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); }; | |||
3579 | } | |||
3580 | ||||
3581 | #if KMP_OS_WINDOWS0 | |||
3582 | /* reclaim array entries for root threads that are already dead, returns number | |||
3583 | * reclaimed */ | |||
3584 | static int __kmp_reclaim_dead_roots(void) { | |||
3585 | int i, r = 0; | |||
3586 | ||||
3587 | for (i = 0; i < __kmp_threads_capacity; ++i) { | |||
3588 | if (KMP_UBER_GTID(i) && | |||
3589 | !__kmp_still_running((kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[i])((void *)(__kmp_threads[i]))) && | |||
3590 | !__kmp_root[i] | |||
3591 | ->r.r_active) { // AC: reclaim only roots died in non-active state | |||
3592 | r += __kmp_unregister_root_other_thread(i); | |||
3593 | } | |||
3594 | } | |||
3595 | return r; | |||
3596 | } | |||
3597 | #endif | |||
3598 | ||||
3599 | /* This function attempts to create free entries in __kmp_threads and | |||
3600 | __kmp_root, and returns the number of free entries generated. | |||
3601 | ||||
3602 | For Windows* OS static library, the first mechanism used is to reclaim array | |||
3603 | entries for root threads that are already dead. | |||
3604 | ||||
3605 | On all platforms, expansion is attempted on the arrays __kmp_threads_ and | |||
3606 | __kmp_root, with appropriate update to __kmp_threads_capacity. Array | |||
3607 | capacity is increased by doubling with clipping to __kmp_tp_capacity, if | |||
3608 | threadprivate cache array has been created. Synchronization with | |||
3609 | __kmpc_threadprivate_cached is done using __kmp_tp_cached_lock. | |||
3610 | ||||
3611 | After any dead root reclamation, if the clipping value allows array expansion | |||
3612 | to result in the generation of a total of nNeed free slots, the function does | |||
3613 | that expansion. If not, nothing is done beyond the possible initial root | |||
3614 | thread reclamation. | |||
3615 | ||||
3616 | If any argument is negative, the behavior is undefined. */ | |||
3617 | static int __kmp_expand_threads(int nNeed) { | |||
3618 | int added = 0; | |||
3619 | int minimumRequiredCapacity; | |||
3620 | int newCapacity; | |||
3621 | kmp_info_t **newThreads; | |||
3622 | kmp_root_t **newRoot; | |||
3623 | ||||
3624 | // All calls to __kmp_expand_threads should be under __kmp_forkjoin_lock, so | |||
3625 | // resizing __kmp_threads does not need additional protection if foreign | |||
3626 | // threads are present | |||
3627 | ||||
3628 | #if KMP_OS_WINDOWS0 && !KMP_DYNAMIC_LIB1 | |||
3629 | /* only for Windows static library */ | |||
3630 | /* reclaim array entries for root threads that are already dead */ | |||
3631 | added = __kmp_reclaim_dead_roots(); | |||
3632 | ||||
3633 | if (nNeed) { | |||
3634 | nNeed -= added; | |||
3635 | if (nNeed < 0) | |||
3636 | nNeed = 0; | |||
3637 | } | |||
3638 | #endif | |||
3639 | if (nNeed <= 0) | |||
3640 | return added; | |||
3641 | ||||
3642 | // Note that __kmp_threads_capacity is not bounded by __kmp_max_nth. If | |||
3643 | // __kmp_max_nth is set to some value less than __kmp_sys_max_nth by the | |||
3644 | // user via KMP_DEVICE_THREAD_LIMIT, then __kmp_threads_capacity may become | |||
3645 | // > __kmp_max_nth in one of two ways: | |||
3646 | // | |||
3647 | // 1) The initialization thread (gtid = 0) exits. __kmp_threads[0] | |||
3648 | // may not be reused by another thread, so we may need to increase | |||
3649 | // __kmp_threads_capacity to __kmp_max_nth + 1. | |||
3650 | // | |||
3651 | // 2) New foreign root(s) are encountered. We always register new foreign | |||
3652 | // roots. This may cause a smaller # of threads to be allocated at | |||
3653 | // subsequent parallel regions, but the worker threads hang around (and | |||
3654 | // eventually go to sleep) and need slots in the __kmp_threads[] array. | |||
3655 | // | |||
3656 | // Anyway, that is the reason for moving the check to see if | |||
3657 | // __kmp_max_nth was exceeded into __kmp_reserve_threads() | |||
3658 | // instead of having it performed here. -BB | |||
3659 | ||||
3660 | 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" , 3660); }; | |||
3661 | ||||
3662 | /* compute expansion headroom to check if we can expand */ | |||
3663 | if (__kmp_sys_max_nth - __kmp_threads_capacity < nNeed) { | |||
3664 | /* possible expansion too small -- give up */ | |||
3665 | return added; | |||
3666 | } | |||
3667 | minimumRequiredCapacity = __kmp_threads_capacity + nNeed; | |||
3668 | ||||
3669 | newCapacity = __kmp_threads_capacity; | |||
3670 | do { | |||
3671 | newCapacity = newCapacity <= (__kmp_sys_max_nth >> 1) ? (newCapacity << 1) | |||
3672 | : __kmp_sys_max_nth; | |||
3673 | } while (newCapacity < minimumRequiredCapacity); | |||
3674 | newThreads = (kmp_info_t **)__kmp_allocate(___kmp_allocate(((sizeof(kmp_info_t *) + sizeof(kmp_root_t *) ) * newCapacity + 64), "openmp/runtime/src/kmp_runtime.cpp", 3675 ) | |||
3675 | (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", 3675 ); | |||
3676 | newRoot = | |||
3677 | (kmp_root_t **)((char *)newThreads + sizeof(kmp_info_t *) * newCapacity); | |||
3678 | KMP_MEMCPYmemcpy(newThreads, __kmp_threads, | |||
3679 | __kmp_threads_capacity * sizeof(kmp_info_t *)); | |||
3680 | KMP_MEMCPYmemcpy(newRoot, __kmp_root, | |||
3681 | __kmp_threads_capacity * sizeof(kmp_root_t *)); | |||
3682 | // Put old __kmp_threads array on a list. Any ongoing references to the old | |||
3683 | // list will be valid. This list is cleaned up at library shutdown. | |||
3684 | kmp_old_threads_list_t *node = | |||
3685 | (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" , 3685); | |||
3686 | node->threads = __kmp_threads; | |||
3687 | node->next = __kmp_old_threads_list; | |||
3688 | __kmp_old_threads_list = node; | |||
3689 | ||||
3690 | *(kmp_info_t * *volatile *)&__kmp_threads = newThreads; | |||
3691 | *(kmp_root_t * *volatile *)&__kmp_root = newRoot; | |||
3692 | added += newCapacity - __kmp_threads_capacity; | |||
3693 | *(volatile int *)&__kmp_threads_capacity = newCapacity; | |||
3694 | ||||
3695 | if (newCapacity > __kmp_tp_capacity) { | |||
3696 | __kmp_acquire_bootstrap_lock(&__kmp_tp_cached_lock); | |||
3697 | if (__kmp_tp_cached && newCapacity > __kmp_tp_capacity) { | |||
3698 | __kmp_threadprivate_resize_cache(newCapacity); | |||
3699 | } else { // increase __kmp_tp_capacity to correspond with kmp_threads size | |||
3700 | *(volatile int *)&__kmp_tp_capacity = newCapacity; | |||
3701 | } | |||
3702 | __kmp_release_bootstrap_lock(&__kmp_tp_cached_lock); | |||
3703 | } | |||
3704 | ||||
3705 | return added; | |||
3706 | } | |||
3707 | ||||
3708 | /* Register the current thread as a root thread and obtain our gtid. We must | |||
3709 | have the __kmp_initz_lock held at this point. Argument TRUE only if are the | |||
3710 | thread that calls from __kmp_do_serial_initialize() */ | |||
3711 | int __kmp_register_root(int initial_thread) { | |||
3712 | kmp_info_t *root_thread; | |||
3713 | kmp_root_t *root; | |||
3714 | int gtid; | |||
3715 | int capacity; | |||
3716 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); | |||
3717 | KA_TRACE(20, ("__kmp_register_root: entered\n"))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_register_root: entered\n" ); }; | |||
3718 | KMP_MB(); | |||
3719 | ||||
3720 | /* 2007-03-02: | |||
3721 | If initial thread did not invoke OpenMP RTL yet, and this thread is not an | |||
3722 | initial one, "__kmp_all_nth >= __kmp_threads_capacity" condition does not | |||
3723 | work as expected -- it may return false (that means there is at least one | |||
3724 | empty slot in __kmp_threads array), but it is possible the only free slot | |||
3725 | is #0, which is reserved for initial thread and so cannot be used for this | |||
3726 | one. Following code workarounds this bug. | |||
3727 | ||||
3728 | However, right solution seems to be not reserving slot #0 for initial | |||
3729 | thread because: | |||
3730 | (1) there is no magic in slot #0, | |||
3731 | (2) we cannot detect initial thread reliably (the first thread which does | |||
3732 | serial initialization may be not a real initial thread). | |||
3733 | */ | |||
3734 | capacity = __kmp_threads_capacity; | |||
3735 | if (!initial_thread && TCR_PTR(__kmp_threads[0])((void *)(__kmp_threads[0])) == NULL__null) { | |||
3736 | --capacity; | |||
3737 | } | |||
3738 | ||||
3739 | // If it is not for initializing the hidden helper team, we need to take | |||
3740 | // __kmp_hidden_helper_threads_num out of the capacity because it is included | |||
3741 | // in __kmp_threads_capacity. | |||
3742 | if (__kmp_enable_hidden_helper && !TCR_4(__kmp_init_hidden_helper_threads)(__kmp_init_hidden_helper_threads)) { | |||
3743 | capacity -= __kmp_hidden_helper_threads_num; | |||
3744 | } | |||
3745 | ||||
3746 | /* see if there are too many threads */ | |||
3747 | if (__kmp_all_nth >= capacity && !__kmp_expand_threads(1)) { | |||
3748 | if (__kmp_tp_cached) { | |||
3749 | __kmp_fatal(KMP_MSG(CantRegisterNewThread)__kmp_msg_format(kmp_i18n_msg_CantRegisterNewThread), | |||
3750 | KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity)__kmp_msg_format(kmp_i18n_hnt_Set_ALL_THREADPRIVATE, __kmp_tp_capacity ), | |||
3751 | KMP_HNT(PossibleSystemLimitOnThreads)__kmp_msg_format(kmp_i18n_hnt_PossibleSystemLimitOnThreads), __kmp_msg_null); | |||
3752 | } else { | |||
3753 | __kmp_fatal(KMP_MSG(CantRegisterNewThread)__kmp_msg_format(kmp_i18n_msg_CantRegisterNewThread), KMP_HNT(SystemLimitOnThreads)__kmp_msg_format(kmp_i18n_hnt_SystemLimitOnThreads), | |||
3754 | __kmp_msg_null); | |||
3755 | } | |||
3756 | } | |||
3757 | ||||
3758 | // When hidden helper task is enabled, __kmp_threads is organized as follows: | |||
3759 | // 0: initial thread, also a regular OpenMP thread. | |||
3760 | // [1, __kmp_hidden_helper_threads_num]: slots for hidden helper threads. | |||
3761 | // [__kmp_hidden_helper_threads_num + 1, __kmp_threads_capacity): slots for | |||
3762 | // regular OpenMP threads. | |||
3763 | if (TCR_4(__kmp_init_hidden_helper_threads)(__kmp_init_hidden_helper_threads)) { | |||
3764 | // Find an available thread slot for hidden helper thread. Slots for hidden | |||
3765 | // helper threads start from 1 to __kmp_hidden_helper_threads_num. | |||
3766 | for (gtid = 1; TCR_PTR(__kmp_threads[gtid])((void *)(__kmp_threads[gtid])) != NULL__null && | |||
3767 | gtid <= __kmp_hidden_helper_threads_num; | |||
3768 | gtid++) | |||
3769 | ; | |||
3770 | 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" , 3770); }; | |||
3771 | 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); } | |||
3772 | "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); } | |||
3773 | 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); }; | |||
3774 | } else { | |||
3775 | /* find an available thread slot */ | |||
3776 | // Don't reassign the zero slot since we need that to only be used by | |||
3777 | // initial thread. Slots for hidden helper threads should also be skipped. | |||
3778 | if (initial_thread && TCR_PTR(__kmp_threads[0])((void *)(__kmp_threads[0])) == NULL__null) { | |||
3779 | gtid = 0; | |||
3780 | } else { | |||
3781 | for (gtid = __kmp_hidden_helper_threads_num + 1; | |||
3782 | TCR_PTR(__kmp_threads[gtid])((void *)(__kmp_threads[gtid])) != NULL__null; gtid++) | |||
3783 | ; | |||
3784 | } | |||
3785 | KA_TRACE(if (kmp_a_debug >= 1) { __kmp_debug_printf ("__kmp_register_root: found slot in threads array: T#%d\n" , gtid); } | |||
3786 | 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); }; | |||
3787 | KMP_ASSERT(gtid < __kmp_threads_capacity)if (!(gtid < __kmp_threads_capacity)) { __kmp_debug_assert ("gtid < __kmp_threads_capacity", "openmp/runtime/src/kmp_runtime.cpp" , 3787); }; | |||
3788 | } | |||
3789 | ||||
3790 | /* update global accounting */ | |||
3791 | __kmp_all_nth++; | |||
3792 | TCW_4(__kmp_nth, __kmp_nth + 1)(__kmp_nth) = (__kmp_nth + 1); | |||
3793 | ||||
3794 | // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low | |||
3795 | // numbers of procs, and method #2 (keyed API call) for higher numbers. | |||
3796 | if (__kmp_adjust_gtid_mode) { | |||
3797 | if (__kmp_all_nth >= __kmp_tls_gtid_min) { | |||
3798 | if (TCR_4(__kmp_gtid_mode)(__kmp_gtid_mode) != 2) { | |||
3799 | TCW_4(__kmp_gtid_mode, 2)(__kmp_gtid_mode) = (2); | |||
3800 | } | |||
3801 | } else { | |||
3802 | if (TCR_4(__kmp_gtid_mode)(__kmp_gtid_mode) != 1) { | |||
3803 | TCW_4(__kmp_gtid_mode, 1)(__kmp_gtid_mode) = (1); | |||
3804 | } | |||
3805 | } | |||
3806 | } | |||
3807 | ||||
3808 | #ifdef KMP_ADJUST_BLOCKTIME1 | |||
3809 | /* Adjust blocktime to zero if necessary */ | |||
3810 | /* Middle initialization might not have occurred yet */ | |||
3811 | if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { | |||
3812 | if (__kmp_nth > __kmp_avail_proc) { | |||
3813 | __kmp_zero_bt = TRUE(!0); | |||
3814 | } | |||
3815 | } | |||
3816 | #endif /* KMP_ADJUST_BLOCKTIME */ | |||
3817 | ||||
3818 | /* setup this new hierarchy */ | |||
3819 | if (!(root = __kmp_root[gtid])) { | |||
3820 | 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" , 3820); | |||
3821 | 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", 3821); }; | |||
3822 | } | |||
3823 | ||||
3824 | #if KMP_STATS_ENABLED0 | |||
3825 | // Initialize stats as soon as possible (right after gtid assignment). | |||
3826 | __kmp_stats_thread_ptr = __kmp_stats_list->push_back(gtid); | |||
3827 | __kmp_stats_thread_ptr->startLife(); | |||
3828 | KMP_SET_THREAD_STATE(SERIAL_REGION)((void)0); | |||
3829 | KMP_INIT_PARTITIONED_TIMERS(OMP_serial)((void)0); | |||
3830 | #endif | |||
3831 | __kmp_initialize_root(root); | |||
3832 | ||||
3833 | /* setup new root thread structure */ | |||
3834 | if (root->r.r_uber_thread) { | |||
3835 | root_thread = root->r.r_uber_thread; | |||
3836 | } else { | |||
3837 | root_thread = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t))___kmp_allocate((sizeof(kmp_info_t)), "openmp/runtime/src/kmp_runtime.cpp" , 3837); | |||
3838 | if (__kmp_storage_map) { | |||
3839 | __kmp_print_thread_storage_map(root_thread, gtid); | |||
3840 | } | |||
3841 | root_thread->th.th_info.ds.ds_gtid = gtid; | |||
3842 | #if OMPT_SUPPORT1 | |||
3843 | root_thread->th.ompt_thread_info.thread_data = ompt_data_none{0}; | |||
3844 | #endif | |||
3845 | root_thread->th.th_root = root; | |||
3846 | if (__kmp_env_consistency_check) { | |||
3847 | root_thread->th.th_cons = __kmp_allocate_cons_stack(gtid); | |||
3848 | } | |||
3849 | #if USE_FAST_MEMORY3 | |||
3850 | __kmp_initialize_fast_memory(root_thread); | |||
3851 | #endif /* USE_FAST_MEMORY */ | |||
3852 | ||||
3853 | #if KMP_USE_BGET1 | |||
3854 | 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" , 3854); }; | |||
3855 | __kmp_initialize_bget(root_thread); | |||
3856 | #endif | |||
3857 | __kmp_init_random(root_thread); // Initialize random number generator | |||
3858 | } | |||
3859 | ||||
3860 | /* setup the serial team held in reserve by the root thread */ | |||
3861 | if (!root_thread->th.th_serial_team) { | |||
3862 | kmp_internal_control_t r_icvs = __kmp_get_global_icvs(); | |||
3863 | 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" ); }; | |||
3864 | root_thread->th.th_serial_team = __kmp_allocate_team( | |||
3865 | root, 1, 1, | |||
3866 | #if OMPT_SUPPORT1 | |||
3867 | ompt_data_none{0}, // root parallel id | |||
3868 | #endif | |||
3869 | proc_bind_default, &r_icvs, 0 USE_NESTED_HOT_ARG(NULL), __null); | |||
3870 | } | |||
3871 | 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" , 3871); }; | |||
3872 | 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); } | |||
3873 | 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); }; | |||
3874 | ||||
3875 | /* drop root_thread into place */ | |||
3876 | TCW_SYNC_PTR(__kmp_threads[gtid], root_thread)((__kmp_threads[gtid])) = ((root_thread)); | |||
3877 | ||||
3878 | root->r.r_root_team->t.t_threads[0] = root_thread; | |||
3879 | root->r.r_hot_team->t.t_threads[0] = root_thread; | |||
3880 | root_thread->th.th_serial_team->t.t_threads[0] = root_thread; | |||
3881 | // AC: the team created in reserve, not for execution (it is unused for now). | |||
3882 | root_thread->th.th_serial_team->t.t_serialized = 0; | |||
3883 | root->r.r_uber_thread = root_thread; | |||
3884 | ||||
3885 | /* initialize the thread, get it ready to go */ | |||
3886 | __kmp_initialize_info(root_thread, root->r.r_root_team, 0, gtid); | |||
3887 | TCW_4(__kmp_init_gtid, TRUE)(__kmp_init_gtid) = ((!0)); | |||
3888 | ||||
3889 | /* prepare the primary thread for get_gtid() */ | |||
3890 | __kmp_gtid_set_specific(gtid); | |||
3891 | ||||
3892 | #if USE_ITT_BUILD1 | |||
3893 | __kmp_itt_thread_name(gtid); | |||
3894 | #endif /* USE_ITT_BUILD */ | |||
3895 | ||||
3896 | #ifdef KMP_TDATA_GTID1 | |||
3897 | __kmp_gtid = gtid; | |||
3898 | #endif | |||
3899 | __kmp_create_worker(gtid, root_thread, __kmp_stksize); | |||
3900 | 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" , 3900); }; | |||
3901 | ||||
3902 | 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); } | |||
3903 | "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); } | |||
3904 | 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); } | |||
3905 | 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); } | |||
3906 | 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); }; | |||
3907 | { // Initialize barrier data. | |||
3908 | int b; | |||
3909 | for (b = 0; b < bs_last_barrier; ++b) { | |||
3910 | root_thread->th.th_bar[b].bb.b_arrived = KMP_INIT_BARRIER_STATE0; | |||
3911 | #if USE_DEBUGGER0 | |||
3912 | root_thread->th.th_bar[b].bb.b_worker_arrived = 0; | |||
3913 | #endif | |||
3914 | } | |||
3915 | } | |||
3916 | 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", 3917); } | |||
3917 | 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", 3917); }; | |||
3918 | ||||
3919 | #if KMP_AFFINITY_SUPPORTED1 | |||
3920 | root_thread->th.th_current_place = KMP_PLACE_UNDEFINED(-2); | |||
3921 | root_thread->th.th_new_place = KMP_PLACE_UNDEFINED(-2); | |||
3922 | root_thread->th.th_first_place = KMP_PLACE_UNDEFINED(-2); | |||
3923 | root_thread->th.th_last_place = KMP_PLACE_UNDEFINED(-2); | |||
3924 | #endif /* KMP_AFFINITY_SUPPORTED */ | |||
3925 | root_thread->th.th_def_allocator = __kmp_def_allocator; | |||
3926 | root_thread->th.th_prev_level = 0; | |||
3927 | root_thread->th.th_prev_num_threads = 1; | |||
3928 | ||||
3929 | 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" , 3929); | |||
3930 | tmp->cg_root = root_thread; | |||
3931 | tmp->cg_thread_limit = __kmp_cg_max_nth; | |||
3932 | tmp->cg_nthreads = 1; | |||
3933 | 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); } | |||
3934 | " 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); } | |||
3935 | 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); }; | |||
3936 | tmp->up = NULL__null; | |||
3937 | root_thread->th.th_cg_roots = tmp; | |||
3938 | ||||
3939 | __kmp_root_counter++; | |||
3940 | ||||
3941 | #if OMPT_SUPPORT1 | |||
3942 | if (!initial_thread && ompt_enabled.enabled) { | |||
3943 | ||||
3944 | kmp_info_t *root_thread = ompt_get_thread(); | |||
3945 | ||||
3946 | ompt_set_thread_state(root_thread, ompt_state_overhead); | |||
3947 | ||||
3948 | if (ompt_enabled.ompt_callback_thread_begin) { | |||
3949 | ompt_callbacks.ompt_callback(ompt_callback_thread_begin)ompt_callback_thread_begin_callback( | |||
3950 | ompt_thread_initial, __ompt_get_thread_data_internal()); | |||
3951 | } | |||
3952 | ompt_data_t *task_data; | |||
3953 | ompt_data_t *parallel_data; | |||
3954 | __ompt_get_task_info_internal(0, NULL__null, &task_data, NULL__null, ¶llel_data, | |||
3955 | NULL__null); | |||
3956 | if (ompt_enabled.ompt_callback_implicit_task) { | |||
3957 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | |||
3958 | ompt_scope_begin, parallel_data, task_data, 1, 1, ompt_task_initial); | |||
3959 | } | |||
3960 | ||||
3961 | ompt_set_thread_state(root_thread, ompt_state_work_serial); | |||
3962 | } | |||
3963 | #endif | |||
3964 | #if OMPD_SUPPORT1 | |||
3965 | if (ompd_state & OMPD_ENABLE_BP0x1) | |||
3966 | ompd_bp_thread_begin(); | |||
3967 | #endif | |||
3968 | ||||
3969 | KMP_MB(); | |||
3970 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | |||
3971 | ||||
3972 | return gtid; | |||
3973 | } | |||
3974 | ||||
3975 | #if KMP_NESTED_HOT_TEAMS1 | |||
3976 | static int __kmp_free_hot_teams(kmp_root_t *root, kmp_info_t *thr, int level, | |||
3977 | const int max_level) { | |||
3978 | int i, n, nth; | |||
3979 | kmp_hot_team_ptr_t *hot_teams = thr->th.th_hot_teams; | |||
3980 | if (!hot_teams || !hot_teams[level].hot_team) { | |||
3981 | return 0; | |||
3982 | } | |||
3983 | KMP_DEBUG_ASSERT(level < max_level)if (!(level < max_level)) { __kmp_debug_assert("level < max_level" , "openmp/runtime/src/kmp_runtime.cpp", 3983); }; | |||
3984 | kmp_team_t *team = hot_teams[level].hot_team; | |||
3985 | nth = hot_teams[level].hot_team_nth; | |||
3986 | n = nth - 1; // primary thread is not freed | |||
3987 | if (level < max_level - 1) { | |||
3988 | for (i = 0; i < nth; ++i) { | |||
3989 | kmp_info_t *th = team->t.t_threads[i]; | |||
3990 | n += __kmp_free_hot_teams(root, th, level + 1, max_level); | |||
3991 | if (i > 0 && th->th.th_hot_teams) { | |||
3992 | __kmp_free(th->th.th_hot_teams)___kmp_free((th->th.th_hot_teams), "openmp/runtime/src/kmp_runtime.cpp" , 3992); | |||
3993 | th->th.th_hot_teams = NULL__null; | |||
3994 | } | |||
3995 | } | |||
3996 | } | |||
3997 | __kmp_free_team(root, team, NULL__null); | |||
3998 | return n; | |||
3999 | } | |||
4000 | #endif | |||
4001 | ||||
4002 | // Resets a root thread and clear its root and hot teams. | |||
4003 | // Returns the number of __kmp_threads entries directly and indirectly freed. | |||
4004 | static int __kmp_reset_root(int gtid, kmp_root_t *root) { | |||
4005 | kmp_team_t *root_team = root->r.r_root_team; | |||
4006 | kmp_team_t *hot_team = root->r.r_hot_team; | |||
4007 | int n = hot_team->t.t_nproc; | |||
4008 | int i; | |||
4009 | ||||
4010 | 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", 4010); }; | |||
4011 | ||||
4012 | root->r.r_root_team = NULL__null; | |||
4013 | root->r.r_hot_team = NULL__null; | |||
4014 | // __kmp_free_team() does not free hot teams, so we have to clear r_hot_team | |||
4015 | // before call to __kmp_free_team(). | |||
4016 | __kmp_free_team(root, root_team USE_NESTED_HOT_ARG(NULL), __null); | |||
4017 | #if KMP_NESTED_HOT_TEAMS1 | |||
4018 | if (__kmp_hot_teams_max_level > | |||
4019 | 0) { // need to free nested hot teams and their threads if any | |||
4020 | for (i = 0; i < hot_team->t.t_nproc; ++i) { | |||
4021 | kmp_info_t *th = hot_team->t.t_threads[i]; | |||
4022 | if (__kmp_hot_teams_max_level > 1) { | |||
4023 | n += __kmp_free_hot_teams(root, th, 1, __kmp_hot_teams_max_level); | |||
4024 | } | |||
4025 | if (th->th.th_hot_teams) { | |||
4026 | __kmp_free(th->th.th_hot_teams)___kmp_free((th->th.th_hot_teams), "openmp/runtime/src/kmp_runtime.cpp" , 4026); | |||
4027 | th->th.th_hot_teams = NULL__null; | |||
4028 | } | |||
4029 | } | |||
4030 | } | |||
4031 | #endif | |||
4032 | __kmp_free_team(root, hot_team USE_NESTED_HOT_ARG(NULL), __null); | |||
4033 | ||||
4034 | // Before we can reap the thread, we need to make certain that all other | |||
4035 | // threads in the teams that had this root as ancestor have stopped trying to | |||
4036 | // steal tasks. | |||
4037 | if (__kmp_tasking_mode != tskm_immediate_exec) { | |||
4038 | __kmp_wait_to_unref_task_teams(); | |||
4039 | } | |||
4040 | ||||
4041 | #if KMP_OS_WINDOWS0 | |||
4042 | /* Close Handle of root duplicated in __kmp_create_worker (tr #62919) */ | |||
4043 | 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); } | |||
4044 | 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); } | |||
4045 | "\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); } | |||
4046 | (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); } | |||
4047 | 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); }; | |||
4048 | __kmp_free_handle(root->r.r_uber_thread->th.th_info.ds.ds_thread); | |||
4049 | #endif /* KMP_OS_WINDOWS */ | |||
4050 | ||||
4051 | #if OMPD_SUPPORT1 | |||
4052 | if (ompd_state & OMPD_ENABLE_BP0x1) | |||
4053 | ompd_bp_thread_end(); | |||
4054 | #endif | |||
4055 | ||||
4056 | #if OMPT_SUPPORT1 | |||
4057 | ompt_data_t *task_data; | |||
4058 | ompt_data_t *parallel_data; | |||
4059 | __ompt_get_task_info_internal(0, NULL__null, &task_data, NULL__null, ¶llel_data, | |||
4060 | NULL__null); | |||
4061 | if (ompt_enabled.ompt_callback_implicit_task) { | |||
4062 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | |||
4063 | ompt_scope_end, parallel_data, task_data, 0, 1, ompt_task_initial); | |||
4064 | } | |||
4065 | if (ompt_enabled.ompt_callback_thread_end) { | |||
4066 | ompt_callbacks.ompt_callback(ompt_callback_thread_end)ompt_callback_thread_end_callback( | |||
4067 | &(root->r.r_uber_thread->th.ompt_thread_info.thread_data)); | |||
4068 | } | |||
4069 | #endif | |||
4070 | ||||
4071 | TCW_4(__kmp_nth,(__kmp_nth) = (__kmp_nth - 1) | |||
4072 | __kmp_nth - 1)(__kmp_nth) = (__kmp_nth - 1); // __kmp_reap_thread will decrement __kmp_all_nth. | |||
4073 | i = root->r.r_uber_thread->th.th_cg_roots->cg_nthreads--; | |||
4074 | 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); } | |||
4075 | " 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); } | |||
4076 | 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); } | |||
4077 | 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); }; | |||
4078 | if (i == 1) { | |||
4079 | // need to free contention group structure | |||
4080 | 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", 4081); } | |||
4081 | 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", 4081); }; | |||
4082 | 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", 4082); }; | |||
4083 | __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" , 4083); | |||
4084 | root->r.r_uber_thread->th.th_cg_roots = NULL__null; | |||
4085 | } | |||
4086 | __kmp_reap_thread(root->r.r_uber_thread, 1); | |||
4087 | ||||
4088 | // We canot put root thread to __kmp_thread_pool, so we have to reap it | |||
4089 | // instead of freeing. | |||
4090 | root->r.r_uber_thread = NULL__null; | |||
4091 | /* mark root as no longer in use */ | |||
4092 | root->r.r_begin = FALSE0; | |||
4093 | ||||
4094 | return n; | |||
4095 | } | |||
4096 | ||||
4097 | void __kmp_unregister_root_current_thread(int gtid) { | |||
4098 | 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); }; | |||
4099 | /* this lock should be ok, since unregister_root_current_thread is never | |||
4100 | called during an abort, only during a normal close. furthermore, if you | |||
4101 | have the forkjoin lock, you should never try to get the initz lock */ | |||
4102 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); | |||
4103 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done) || !__kmp_init_serial) { | |||
4104 | 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); } | |||
4105 | "exiting T#%d\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_unregister_root_current_thread: already finished, " "exiting T#%d\n", gtid); } | |||
4106 | gtid))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_unregister_root_current_thread: already finished, " "exiting T#%d\n", gtid); }; | |||
4107 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | |||
4108 | return; | |||
4109 | } | |||
4110 | kmp_root_t *root = __kmp_root[gtid]; | |||
4111 | ||||
4112 | 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" , 4112); }; | |||
4113 | KMP_ASSERT(KMP_UBER_GTID(gtid))if (!(KMP_UBER_GTID(gtid))) { __kmp_debug_assert("KMP_UBER_GTID(gtid)" , "openmp/runtime/src/kmp_runtime.cpp", 4113); }; | |||
4114 | 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" , 4114); }; | |||
4115 | 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", 4115); }; | |||
4116 | ||||
4117 | KMP_MB(); | |||
4118 | ||||
4119 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
4120 | kmp_team_t *team = thread->th.th_team; | |||
4121 | kmp_task_team_t *task_team = thread->th.th_task_team; | |||
4122 | ||||
4123 | // we need to wait for the proxy tasks before finishing the thread | |||
4124 | if (task_team != NULL__null && (task_team->tt.tt_found_proxy_tasks || | |||
4125 | task_team->tt.tt_hidden_helper_task_encountered)) { | |||
4126 | #if OMPT_SUPPORT1 | |||
4127 | // the runtime is shutting down so we won't report any events | |||
4128 | thread->th.ompt_thread_info.state = ompt_state_undefined; | |||
4129 | #endif | |||
4130 | __kmp_task_team_wait(thread, team USE_ITT_BUILD_ARG(NULL), __null); | |||
4131 | } | |||
4132 | ||||
4133 | __kmp_reset_root(gtid, root); | |||
4134 | ||||
4135 | KMP_MB(); | |||
4136 | KC_TRACE(10,if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_unregister_root_current_thread: T#%d unregistered\n" , gtid); } | |||
4137 | ("__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); }; | |||
4138 | ||||
4139 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | |||
4140 | } | |||
4141 | ||||
4142 | #if KMP_OS_WINDOWS0 | |||
4143 | /* __kmp_forkjoin_lock must be already held | |||
4144 | Unregisters a root thread that is not the current thread. Returns the number | |||
4145 | of __kmp_threads entries freed as a result. */ | |||
4146 | static int __kmp_unregister_root_other_thread(int gtid) { | |||
4147 | kmp_root_t *root = __kmp_root[gtid]; | |||
4148 | int r; | |||
4149 | ||||
4150 | 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); }; | |||
4151 | 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" , 4151); }; | |||
4152 | KMP_ASSERT(KMP_UBER_GTID(gtid))if (!(KMP_UBER_GTID(gtid))) { __kmp_debug_assert("KMP_UBER_GTID(gtid)" , "openmp/runtime/src/kmp_runtime.cpp", 4152); }; | |||
4153 | 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" , 4153); }; | |||
4154 | 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", 4154); }; | |||
4155 | ||||
4156 | r = __kmp_reset_root(gtid, root); | |||
4157 | KC_TRACE(10,if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_unregister_root_other_thread: T#%d unregistered\n" , gtid); } | |||
4158 | ("__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); }; | |||
4159 | return r; | |||
4160 | } | |||
4161 | #endif | |||
4162 | ||||
4163 | #if KMP_DEBUG1 | |||
4164 | void __kmp_task_info() { | |||
4165 | ||||
4166 | kmp_int32 gtid = __kmp_entry_gtid()__kmp_get_global_thread_id_reg(); | |||
4167 | kmp_int32 tid = __kmp_tid_from_gtid(gtid); | |||
4168 | kmp_info_t *this_thr = __kmp_threads[gtid]; | |||
4169 | kmp_team_t *steam = this_thr->th.th_serial_team; | |||
4170 | kmp_team_t *team = this_thr->th.th_team; | |||
4171 | ||||
4172 | __kmp_printf( | |||
4173 | "__kmp_task_info: gtid=%d tid=%d t_thread=%p team=%p steam=%p curtask=%p " | |||
4174 | "ptask=%p\n", | |||
4175 | gtid, tid, this_thr, team, steam, this_thr->th.th_current_task, | |||
4176 | team->t.t_implicit_task_taskdata[tid].td_parent); | |||
4177 | } | |||
4178 | #endif // KMP_DEBUG | |||
4179 | ||||
4180 | /* TODO optimize with one big memclr, take out what isn't needed, split | |||
4181 | responsibility to workers as much as possible, and delay initialization of | |||
4182 | features as much as possible */ | |||
4183 | static void __kmp_initialize_info(kmp_info_t *this_thr, kmp_team_t *team, | |||
4184 | int tid, int gtid) { | |||
4185 | /* this_thr->th.th_info.ds.ds_gtid is setup in | |||
4186 | kmp_allocate_thread/create_worker. | |||
4187 | this_thr->th.th_serial_team is setup in __kmp_allocate_thread */ | |||
4188 | KMP_DEBUG_ASSERT(this_thr != NULL)if (!(this_thr != __null)) { __kmp_debug_assert("this_thr != __null" , "openmp/runtime/src/kmp_runtime.cpp", 4188); }; | |||
4189 | 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", 4189); }; | |||
4190 | KMP_DEBUG_ASSERT(team)if (!(team)) { __kmp_debug_assert("team", "openmp/runtime/src/kmp_runtime.cpp" , 4190); }; | |||
4191 | 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", 4191); }; | |||
4192 | 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", 4192); }; | |||
4193 | kmp_info_t *master = team->t.t_threads[0]; | |||
4194 | KMP_DEBUG_ASSERT(master)if (!(master)) { __kmp_debug_assert("master", "openmp/runtime/src/kmp_runtime.cpp" , 4194); }; | |||
4195 | 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", 4195); }; | |||
4196 | ||||
4197 | KMP_MB(); | |||
4198 | ||||
4199 | TCW_SYNC_PTR(this_thr->th.th_team, team)((this_thr->th.th_team)) = ((team)); | |||
4200 | ||||
4201 | this_thr->th.th_info.ds.ds_tid = tid; | |||
4202 | this_thr->th.th_set_nproc = 0; | |||
4203 | if (__kmp_tasking_mode != tskm_immediate_exec) | |||
4204 | // When tasking is possible, threads are not safe to reap until they are | |||
4205 | // done tasking; this will be set when tasking code is exited in wait | |||
4206 | this_thr->th.th_reap_state = KMP_NOT_SAFE_TO_REAP0; | |||
4207 | else // no tasking --> always safe to reap | |||
4208 | this_thr->th.th_reap_state = KMP_SAFE_TO_REAP1; | |||
4209 | this_thr->th.th_set_proc_bind = proc_bind_default; | |||
4210 | #if KMP_AFFINITY_SUPPORTED1 | |||
4211 | this_thr->th.th_new_place = this_thr->th.th_current_place; | |||
4212 | #endif | |||
4213 | this_thr->th.th_root = master->th.th_root; | |||
4214 | ||||
4215 | /* setup the thread's cache of the team structure */ | |||
4216 | this_thr->th.th_team_nproc = team->t.t_nproc; | |||
4217 | this_thr->th.th_team_master = master; | |||
4218 | this_thr->th.th_team_serialized = team->t.t_serialized; | |||
4219 | ||||
4220 | 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" , 4220); }; | |||
4221 | ||||
4222 | 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); } | |||
4223 | 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); }; | |||
4224 | ||||
4225 | __kmp_init_implicit_task(this_thr->th.th_team_master->th.th_ident, this_thr, | |||
4226 | team, tid, TRUE(!0)); | |||
4227 | ||||
4228 | 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); } | |||
4229 | 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); }; | |||
4230 | // TODO: Initialize ICVs from parent; GEH - isn't that already done in | |||
4231 | // __kmp_initialize_team()? | |||
4232 | ||||
4233 | /* TODO no worksharing in speculative threads */ | |||
4234 | this_thr->th.th_dispatch = &team->t.t_dispatch[tid]; | |||
4235 | ||||
4236 | this_thr->th.th_local.this_construct = 0; | |||
4237 | ||||
4238 | if (!this_thr->th.th_pri_common) { | |||
4239 | this_thr->th.th_pri_common = | |||
4240 | (struct common_table *)__kmp_allocate(sizeof(struct common_table))___kmp_allocate((sizeof(struct common_table)), "openmp/runtime/src/kmp_runtime.cpp" , 4240); | |||
4241 | if (__kmp_storage_map) { | |||
4242 | __kmp_print_storage_map_gtid( | |||
4243 | gtid, this_thr->th.th_pri_common, this_thr->th.th_pri_common + 1, | |||
4244 | sizeof(struct common_table), "th_%d.th_pri_common\n", gtid); | |||
4245 | } | |||
4246 | this_thr->th.th_pri_head = NULL__null; | |||
4247 | } | |||
4248 | ||||
4249 | if (this_thr != master && // Primary thread's CG root is initialized elsewhere | |||
4250 | this_thr->th.th_cg_roots != master->th.th_cg_roots) { // CG root not set | |||
4251 | // Make new thread's CG root same as primary thread's | |||
4252 | 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", 4252); }; | |||
4253 | kmp_cg_root_t *tmp = this_thr->th.th_cg_roots; | |||
4254 | if (tmp) { | |||
4255 | // worker changes CG, need to check if old CG should be freed | |||
4256 | int i = tmp->cg_nthreads--; | |||
4257 | 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); } | |||
4258 | " 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); } | |||
4259 | 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); }; | |||
4260 | if (i == 1) { | |||
4261 | __kmp_free(tmp)___kmp_free((tmp), "openmp/runtime/src/kmp_runtime.cpp", 4261 ); // last thread left CG --> free it | |||
4262 | } | |||
4263 | } | |||
4264 | this_thr->th.th_cg_roots = master->th.th_cg_roots; | |||
4265 | // Increment new thread's CG root's counter to add the new thread | |||
4266 | this_thr->th.th_cg_roots->cg_nthreads++; | |||
4267 | 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); } | |||
4268 | " 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); } | |||
4269 | 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); } | |||
4270 | 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); } | |||
4271 | 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); }; | |||
4272 | this_thr->th.th_current_task->td_icvs.thread_limit = | |||
4273 | this_thr->th.th_cg_roots->cg_thread_limit; | |||
4274 | } | |||
4275 | ||||
4276 | /* Initialize dynamic dispatch */ | |||
4277 | { | |||
4278 | volatile kmp_disp_t *dispatch = this_thr->th.th_dispatch; | |||
4279 | // Use team max_nproc since this will never change for the team. | |||
4280 | size_t disp_size = | |||
4281 | sizeof(dispatch_private_info_t) * | |||
4282 | (team->t.t_max_nproc == 1 ? 1 : __kmp_dispatch_num_buffers); | |||
4283 | 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); } | |||
4284 | 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); }; | |||
4285 | KMP_ASSERT(dispatch)if (!(dispatch)) { __kmp_debug_assert("dispatch", "openmp/runtime/src/kmp_runtime.cpp" , 4285); }; | |||
4286 | 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", 4286); }; | |||
4287 | 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" , 4287); }; | |||
4288 | ||||
4289 | dispatch->th_disp_index = 0; | |||
4290 | dispatch->th_doacross_buf_idx = 0; | |||
4291 | if (!dispatch->th_disp_buffer) { | |||
4292 | dispatch->th_disp_buffer = | |||
4293 | (dispatch_private_info_t *)__kmp_allocate(disp_size)___kmp_allocate((disp_size), "openmp/runtime/src/kmp_runtime.cpp" , 4293); | |||
4294 | ||||
4295 | if (__kmp_storage_map) { | |||
4296 | __kmp_print_storage_map_gtid( | |||
4297 | gtid, &dispatch->th_disp_buffer[0], | |||
4298 | &dispatch->th_disp_buffer[team->t.t_max_nproc == 1 | |||
4299 | ? 1 | |||
4300 | : __kmp_dispatch_num_buffers], | |||
4301 | disp_size, | |||
4302 | "th_%d.th_dispatch.th_disp_buffer " | |||
4303 | "(team_%d.t_dispatch[%d].th_disp_buffer)", | |||
4304 | gtid, team->t.t_id, gtid); | |||
4305 | } | |||
4306 | } else { | |||
4307 | memset(&dispatch->th_disp_buffer[0], '\0', disp_size); | |||
4308 | } | |||
4309 | ||||
4310 | dispatch->th_dispatch_pr_current = 0; | |||
4311 | dispatch->th_dispatch_sh_current = 0; | |||
4312 | ||||
4313 | dispatch->th_deo_fcn = 0; /* ORDERED */ | |||
4314 | dispatch->th_dxo_fcn = 0; /* END ORDERED */ | |||
4315 | } | |||
4316 | ||||
4317 | this_thr->th.th_next_pool = NULL__null; | |||
4318 | ||||
4319 | if (!this_thr->th.th_task_state_memo_stack) { | |||
4320 | size_t i; | |||
4321 | this_thr->th.th_task_state_memo_stack = | |||
4322 | (kmp_uint8 *)__kmp_allocate(4 * sizeof(kmp_uint8))___kmp_allocate((4 * sizeof(kmp_uint8)), "openmp/runtime/src/kmp_runtime.cpp" , 4322); | |||
4323 | this_thr->th.th_task_state_top = 0; | |||
4324 | this_thr->th.th_task_state_stack_sz = 4; | |||
4325 | for (i = 0; i < this_thr->th.th_task_state_stack_sz; | |||
4326 | ++i) // zero init the stack | |||
4327 | this_thr->th.th_task_state_memo_stack[i] = 0; | |||
4328 | } | |||
4329 | ||||
4330 | 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", 4330); }; | |||
4331 | 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" , 4331); }; | |||
4332 | ||||
4333 | KMP_MB(); | |||
4334 | } | |||
4335 | ||||
4336 | /* allocate a new thread for the requesting team. this is only called from | |||
4337 | within a forkjoin critical section. we will first try to get an available | |||
4338 | thread from the thread pool. if none is available, we will fork a new one | |||
4339 | assuming we are able to create a new one. this should be assured, as the | |||
4340 | caller should check on this first. */ | |||
4341 | kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team, | |||
4342 | int new_tid) { | |||
4343 | kmp_team_t *serial_team; | |||
4344 | kmp_info_t *new_thr; | |||
4345 | int new_gtid; | |||
4346 | ||||
4347 | 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()); }; | |||
4348 | KMP_DEBUG_ASSERT(root && team)if (!(root && team)) { __kmp_debug_assert("root && team" , "openmp/runtime/src/kmp_runtime.cpp", 4348); }; | |||
4349 | #if !KMP_NESTED_HOT_TEAMS1 | |||
4350 | 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", 4350); }; | |||
4351 | #endif | |||
4352 | KMP_MB(); | |||
4353 | ||||
4354 | /* first, try to get one from the thread pool */ | |||
4355 | if (__kmp_thread_pool) { | |||
4356 | new_thr = CCAST(kmp_info_t *, __kmp_thread_pool)const_cast<kmp_info_t *>(__kmp_thread_pool); | |||
4357 | __kmp_thread_pool = (volatile kmp_info_t *)new_thr->th.th_next_pool; | |||
4358 | if (new_thr == __kmp_thread_pool_insert_pt) { | |||
4359 | __kmp_thread_pool_insert_pt = NULL__null; | |||
4360 | } | |||
4361 | TCW_4(new_thr->th.th_in_pool, FALSE)(new_thr->th.th_in_pool) = (0); | |||
4362 | __kmp_suspend_initialize_thread(new_thr); | |||
4363 | __kmp_lock_suspend_mx(new_thr); | |||
4364 | if (new_thr->th.th_active_in_pool == TRUE(!0)) { | |||
4365 | 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" , 4365); }; | |||
4366 | KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth)(&__kmp_thread_pool_active_nth)->fetch_sub(1, std::memory_order_acq_rel ); | |||
4367 | new_thr->th.th_active_in_pool = FALSE0; | |||
4368 | } | |||
4369 | __kmp_unlock_suspend_mx(new_thr); | |||
4370 | ||||
4371 | 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 ); } | |||
4372 | __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 ); }; | |||
4373 | 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", 4373); }; | |||
4374 | 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" , 4374); }; | |||
4375 | ||||
4376 | /* setup the thread structure */ | |||
4377 | __kmp_initialize_info(new_thr, team, new_tid, | |||
4378 | new_thr->th.th_info.ds.ds_gtid); | |||
4379 | 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", 4379); }; | |||
4380 | ||||
4381 | TCW_4(__kmp_nth, __kmp_nth + 1)(__kmp_nth) = (__kmp_nth + 1); | |||
4382 | ||||
4383 | new_thr->th.th_task_state = 0; | |||
4384 | new_thr->th.th_task_state_top = 0; | |||
4385 | new_thr->th.th_task_state_stack_sz = 4; | |||
4386 | ||||
4387 | if (__kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | |||
4388 | // Make sure pool thread has transitioned to waiting on own thread struct | |||
4389 | 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" , 4389); }; | |||
4390 | // Thread activated in __kmp_allocate_team when increasing team size | |||
4391 | } | |||
4392 | ||||
4393 | #ifdef KMP_ADJUST_BLOCKTIME1 | |||
4394 | /* Adjust blocktime back to zero if necessary */ | |||
4395 | /* Middle initialization might not have occurred yet */ | |||
4396 | if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { | |||
4397 | if (__kmp_nth > __kmp_avail_proc) { | |||
4398 | __kmp_zero_bt = TRUE(!0); | |||
4399 | } | |||
4400 | } | |||
4401 | #endif /* KMP_ADJUST_BLOCKTIME */ | |||
4402 | ||||
4403 | #if KMP_DEBUG1 | |||
4404 | // If thread entered pool via __kmp_free_thread, wait_flag should != | |||
4405 | // KMP_BARRIER_PARENT_FLAG. | |||
4406 | int b; | |||
4407 | kmp_balign_t *balign = new_thr->th.th_bar; | |||
4408 | for (b = 0; b < bs_last_barrier; ++b) | |||
4409 | 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", 4409); }; | |||
4410 | #endif | |||
4411 | ||||
4412 | 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); } | |||
4413 | __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); }; | |||
4414 | ||||
4415 | KMP_MB(); | |||
4416 | return new_thr; | |||
4417 | } | |||
4418 | ||||
4419 | /* no, well fork a new one */ | |||
4420 | 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", 4420); }; | |||
4421 | 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" , 4421); }; | |||
4422 | ||||
4423 | #if KMP_USE_MONITOR | |||
4424 | // If this is the first worker thread the RTL is creating, then also | |||
4425 | // launch the monitor thread. We try to do this as early as possible. | |||
4426 | if (!TCR_4(__kmp_init_monitor)(__kmp_init_monitor)) { | |||
4427 | __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock); | |||
4428 | if (!TCR_4(__kmp_init_monitor)(__kmp_init_monitor)) { | |||
4429 | KF_TRACE(10, ("before __kmp_create_monitor\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("before __kmp_create_monitor\n" ); }; | |||
4430 | TCW_4(__kmp_init_monitor, 1)(__kmp_init_monitor) = (1); | |||
4431 | __kmp_create_monitor(&__kmp_monitor); | |||
4432 | KF_TRACE(10, ("after __kmp_create_monitor\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("after __kmp_create_monitor\n" ); }; | |||
4433 | #if KMP_OS_WINDOWS0 | |||
4434 | // AC: wait until monitor has started. This is a fix for CQ232808. | |||
4435 | // The reason is that if the library is loaded/unloaded in a loop with | |||
4436 | // small (parallel) work in between, then there is high probability that | |||
4437 | // monitor thread started after the library shutdown. At shutdown it is | |||
4438 | // too late to cope with the problem, because when the primary thread is | |||
4439 | // in DllMain (process detach) the monitor has no chances to start (it is | |||
4440 | // blocked), and primary thread has no means to inform the monitor that | |||
4441 | // the library has gone, because all the memory which the monitor can | |||
4442 | // access is going to be released/reset. | |||
4443 | while (TCR_4(__kmp_init_monitor)(__kmp_init_monitor) < 2) { | |||
4444 | 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 (); }; | |||
4445 | } | |||
4446 | KF_TRACE(10, ("after monitor thread has started\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("after monitor thread has started\n" ); }; | |||
4447 | #endif | |||
4448 | } | |||
4449 | __kmp_release_bootstrap_lock(&__kmp_monitor_lock); | |||
4450 | } | |||
4451 | #endif | |||
4452 | ||||
4453 | KMP_MB(); | |||
4454 | ||||
4455 | { | |||
4456 | int new_start_gtid = TCR_4(__kmp_init_hidden_helper_threads)(__kmp_init_hidden_helper_threads) | |||
4457 | ? 1 | |||
4458 | : __kmp_hidden_helper_threads_num + 1; | |||
4459 | ||||
4460 | for (new_gtid = new_start_gtid; TCR_PTR(__kmp_threads[new_gtid])((void *)(__kmp_threads[new_gtid])) != NULL__null; | |||
4461 | ++new_gtid) { | |||
4462 | 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" , 4462); }; | |||
4463 | } | |||
4464 | ||||
4465 | if (TCR_4(__kmp_init_hidden_helper_threads)(__kmp_init_hidden_helper_threads)) { | |||
4466 | 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" , 4466); }; | |||
4467 | } | |||
4468 | } | |||
4469 | ||||
4470 | /* allocate space for it. */ | |||
4471 | new_thr = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t))___kmp_allocate((sizeof(kmp_info_t)), "openmp/runtime/src/kmp_runtime.cpp" , 4471); | |||
4472 | ||||
4473 | TCW_SYNC_PTR(__kmp_threads[new_gtid], new_thr)((__kmp_threads[new_gtid])) = ((new_thr)); | |||
4474 | ||||
4475 | #if USE_ITT_BUILD1 && USE_ITT_NOTIFY1 && KMP_DEBUG1 | |||
4476 | // suppress race conditions detection on synchronization flags in debug mode | |||
4477 | // this helps to analyze library internals eliminating false positives | |||
4478 | __itt_suppress_mark_range(!__kmp_itt_suppress_mark_range_ptr__3_0) ? (void)0 : __kmp_itt_suppress_mark_range_ptr__3_0( | |||
4479 | __itt_suppress_range, __itt_suppress_threading_errors0x000000ff, | |||
4480 | &new_thr->th.th_sleep_loc, sizeof(new_thr->th.th_sleep_loc)); | |||
4481 | __itt_suppress_mark_range(!__kmp_itt_suppress_mark_range_ptr__3_0) ? (void)0 : __kmp_itt_suppress_mark_range_ptr__3_0( | |||
4482 | __itt_suppress_range, __itt_suppress_threading_errors0x000000ff, | |||
4483 | &new_thr->th.th_reap_state, sizeof(new_thr->th.th_reap_state)); | |||
4484 | #if KMP_OS_WINDOWS0 | |||
4485 | __itt_suppress_mark_range(!__kmp_itt_suppress_mark_range_ptr__3_0) ? (void)0 : __kmp_itt_suppress_mark_range_ptr__3_0( | |||
4486 | __itt_suppress_range, __itt_suppress_threading_errors0x000000ff, | |||
4487 | &new_thr->th.th_suspend_init, sizeof(new_thr->th.th_suspend_init)); | |||
4488 | #else | |||
4489 | __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, | |||
4490 | __itt_suppress_threading_errors0x000000ff, | |||
4491 | &new_thr->th.th_suspend_init_count, | |||
4492 | sizeof(new_thr->th.th_suspend_init_count)); | |||
4493 | #endif | |||
4494 | // TODO: check if we need to also suppress b_arrived flags | |||
4495 | __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, | |||
4496 | __itt_suppress_threading_errors0x000000ff, | |||
4497 | 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), | |||
4498 | sizeof(new_thr->th.th_bar[0].bb.b_go)); | |||
4499 | __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, | |||
4500 | __itt_suppress_threading_errors0x000000ff, | |||
4501 | 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), | |||
4502 | sizeof(new_thr->th.th_bar[1].bb.b_go)); | |||
4503 | __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, | |||
4504 | __itt_suppress_threading_errors0x000000ff, | |||
4505 | 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), | |||
4506 | sizeof(new_thr->th.th_bar[2].bb.b_go)); | |||
4507 | #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY && KMP_DEBUG */ | |||
4508 | if (__kmp_storage_map) { | |||
4509 | __kmp_print_thread_storage_map(new_thr, new_gtid); | |||
4510 | } | |||
4511 | ||||
4512 | // add the reserve serialized team, initialized from the team's primary thread | |||
4513 | { | |||
4514 | kmp_internal_control_t r_icvs = __kmp_get_x_global_icvs(team); | |||
4515 | 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" ); }; | |||
4516 | new_thr->th.th_serial_team = serial_team = | |||
4517 | (kmp_team_t *)__kmp_allocate_team(root, 1, 1, | |||
4518 | #if OMPT_SUPPORT1 | |||
4519 | ompt_data_none{0}, // root parallel id | |||
4520 | #endif | |||
4521 | proc_bind_default, &r_icvs, | |||
4522 | 0 USE_NESTED_HOT_ARG(NULL), __null); | |||
4523 | } | |||
4524 | KMP_ASSERT(serial_team)if (!(serial_team)) { __kmp_debug_assert("serial_team", "openmp/runtime/src/kmp_runtime.cpp" , 4524); }; | |||
4525 | serial_team->t.t_serialized = 0; // AC: the team created in reserve, not for | |||
4526 | // execution (it is unused for now). | |||
4527 | serial_team->t.t_threads[0] = new_thr; | |||
4528 | 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); } | |||
4529 | ("__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); } | |||
4530 | new_thr))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n" , new_thr); }; | |||
4531 | ||||
4532 | /* setup the thread structures */ | |||
4533 | __kmp_initialize_info(new_thr, team, new_tid, new_gtid); | |||
4534 | ||||
4535 | #if USE_FAST_MEMORY3 | |||
4536 | __kmp_initialize_fast_memory(new_thr); | |||
4537 | #endif /* USE_FAST_MEMORY */ | |||
4538 | ||||
4539 | #if KMP_USE_BGET1 | |||
4540 | 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" , 4540); }; | |||
4541 | __kmp_initialize_bget(new_thr); | |||
4542 | #endif | |||
4543 | ||||
4544 | __kmp_init_random(new_thr); // Initialize random number generator | |||
4545 | ||||
4546 | /* Initialize these only once when thread is grabbed for a team allocation */ | |||
4547 | 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); } | |||
4548 | ("__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); } | |||
4549 | __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); }; | |||
4550 | ||||
4551 | int b; | |||
4552 | kmp_balign_t *balign = new_thr->th.th_bar; | |||
4553 | for (b = 0; b < bs_last_barrier; ++b) { | |||
4554 | balign[b].bb.b_go = KMP_INIT_BARRIER_STATE0; | |||
4555 | balign[b].bb.team = NULL__null; | |||
4556 | balign[b].bb.wait_flag = KMP_BARRIER_NOT_WAITING0; | |||
4557 | balign[b].bb.use_oncore_barrier = 0; | |||
4558 | } | |||
4559 | ||||
4560 | TCW_PTR(new_thr->th.th_sleep_loc, NULL)((new_thr->th.th_sleep_loc)) = ((__null)); | |||
4561 | new_thr->th.th_sleep_loc_type = flag_unset; | |||
4562 | ||||
4563 | new_thr->th.th_spin_here = FALSE0; | |||
4564 | new_thr->th.th_next_waiting = 0; | |||
4565 | #if KMP_OS_UNIX1 | |||
4566 | new_thr->th.th_blocking = false; | |||
4567 | #endif | |||
4568 | ||||
4569 | #if KMP_AFFINITY_SUPPORTED1 | |||
4570 | new_thr->th.th_current_place = KMP_PLACE_UNDEFINED(-2); | |||
4571 | new_thr->th.th_new_place = KMP_PLACE_UNDEFINED(-2); | |||
4572 | new_thr->th.th_first_place = KMP_PLACE_UNDEFINED(-2); | |||
4573 | new_thr->th.th_last_place = KMP_PLACE_UNDEFINED(-2); | |||
4574 | #endif | |||
4575 | new_thr->th.th_def_allocator = __kmp_def_allocator; | |||
4576 | new_thr->th.th_prev_level = 0; | |||
4577 | new_thr->th.th_prev_num_threads = 1; | |||
4578 | ||||
4579 | TCW_4(new_thr->th.th_in_pool, FALSE)(new_thr->th.th_in_pool) = (0); | |||
4580 | new_thr->th.th_active_in_pool = FALSE0; | |||
4581 | TCW_4(new_thr->th.th_active, TRUE)(new_thr->th.th_active) = ((!0)); | |||
4582 | ||||
4583 | /* adjust the global counters */ | |||
4584 | __kmp_all_nth++; | |||
4585 | __kmp_nth++; | |||
4586 | ||||
4587 | // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low | |||
4588 | // numbers of procs, and method #2 (keyed API call) for higher numbers. | |||
4589 | if (__kmp_adjust_gtid_mode) { | |||
4590 | if (__kmp_all_nth >= __kmp_tls_gtid_min) { | |||
4591 | if (TCR_4(__kmp_gtid_mode)(__kmp_gtid_mode) != 2) { | |||
4592 | TCW_4(__kmp_gtid_mode, 2)(__kmp_gtid_mode) = (2); | |||
4593 | } | |||
4594 | } else { | |||
4595 | if (TCR_4(__kmp_gtid_mode)(__kmp_gtid_mode) != 1) { | |||
4596 | TCW_4(__kmp_gtid_mode, 1)(__kmp_gtid_mode) = (1); | |||
4597 | } | |||
4598 | } | |||
4599 | } | |||
4600 | ||||
4601 | #ifdef KMP_ADJUST_BLOCKTIME1 | |||
4602 | /* Adjust blocktime back to zero if necessary */ | |||
4603 | /* Middle initialization might not have occurred yet */ | |||
4604 | if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { | |||
4605 | if (__kmp_nth > __kmp_avail_proc) { | |||
4606 | __kmp_zero_bt = TRUE(!0); | |||
4607 | } | |||
4608 | } | |||
4609 | #endif /* KMP_ADJUST_BLOCKTIME */ | |||
4610 | ||||
4611 | /* actually fork it and create the new worker thread */ | |||
4612 | KF_TRACE(if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: before __kmp_create_worker: %p\n" , new_thr); } | |||
4613 | 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); }; | |||
4614 | __kmp_create_worker(new_gtid, new_thr, __kmp_stksize); | |||
4615 | KF_TRACE(10,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: after __kmp_create_worker: %p\n" , new_thr); } | |||
4616 | ("__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); }; | |||
4617 | ||||
4618 | 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); } | |||
4619 | 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); }; | |||
4620 | KMP_MB(); | |||
4621 | return new_thr; | |||
4622 | } | |||
4623 | ||||
4624 | /* Reinitialize team for reuse. | |||
4625 | The hot team code calls this case at every fork barrier, so EPCC barrier | |||
4626 | test are extremely sensitive to changes in it, esp. writes to the team | |||
4627 | struct, which cause a cache invalidation in all threads. | |||
4628 | IF YOU TOUCH THIS ROUTINE, RUN EPCC C SYNCBENCH ON A BIG-IRON MACHINE!!! */ | |||
4629 | static void __kmp_reinitialize_team(kmp_team_t *team, | |||
4630 | kmp_internal_control_t *new_icvs, | |||
4631 | ident_t *loc) { | |||
4632 | 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); } | |||
4633 | 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); }; | |||
4634 | KMP_DEBUG_ASSERT(team && new_icvs)if (!(team && new_icvs)) { __kmp_debug_assert("team && new_icvs" , "openmp/runtime/src/kmp_runtime.cpp", 4634); }; | |||
4635 | 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" , 4635); }; | |||
4636 | KMP_CHECK_UPDATE(team->t.t_ident, loc)if ((team->t.t_ident) != (loc)) (team->t.t_ident) = (loc ); | |||
4637 | ||||
4638 | KMP_CHECK_UPDATE(team->t.t_id, KMP_GEN_TEAM_ID())if ((team->t.t_id) != ((~0))) (team->t.t_id) = ((~0)); | |||
4639 | // Copy ICVs to the primary thread's implicit taskdata | |||
4640 | __kmp_init_implicit_task(loc, team->t.t_threads[0], team, 0, FALSE0); | |||
4641 | copy_icvs(&team->t.t_implicit_task_taskdata[0].td_icvs, new_icvs); | |||
4642 | ||||
4643 | 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); } | |||
4644 | 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); }; | |||
4645 | } | |||
4646 | ||||
4647 | /* Initialize the team data structure. | |||
4648 | This assumes the t_threads and t_max_nproc are already set. | |||
4649 | Also, we don't touch the arguments */ | |||
4650 | static void __kmp_initialize_team(kmp_team_t *team, int new_nproc, | |||
4651 | kmp_internal_control_t *new_icvs, | |||
4652 | ident_t *loc) { | |||
4653 | 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); }; | |||
4654 | ||||
4655 | /* verify */ | |||
4656 | KMP_DEBUG_ASSERT(team)if (!(team)) { __kmp_debug_assert("team", "openmp/runtime/src/kmp_runtime.cpp" , 4656); }; | |||
4657 | 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" , 4657); }; | |||
4658 | 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", 4658); }; | |||
4659 | KMP_MB(); | |||
4660 | ||||
4661 | team->t.t_master_tid = 0; /* not needed */ | |||
4662 | /* team->t.t_master_bar; not needed */ | |||
4663 | team->t.t_serialized = new_nproc > 1 ? 0 : 1; | |||
4664 | team->t.t_nproc = new_nproc; | |||
4665 | ||||
4666 | /* team->t.t_parent = NULL; TODO not needed & would mess up hot team */ | |||
4667 | team->t.t_next_pool = NULL__null; | |||
4668 | /* memset( team->t.t_threads, 0, sizeof(kmp_info_t*)*new_nproc ); would mess | |||
4669 | * up hot team */ | |||
4670 | ||||
4671 | TCW_SYNC_PTR(team->t.t_pkfn, NULL)((team->t.t_pkfn)) = ((__null)); /* not needed */ | |||
4672 | team->t.t_invoke = NULL__null; /* not needed */ | |||
4673 | ||||
4674 | // TODO???: team->t.t_max_active_levels = new_max_active_levels; | |||
4675 | team->t.t_sched.sched = new_icvs->sched.sched; | |||
4676 | ||||
4677 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 | |||
4678 | team->t.t_fp_control_saved = FALSE0; /* not needed */ | |||
4679 | team->t.t_x87_fpu_control_word = 0; /* not needed */ | |||
4680 | team->t.t_mxcsr = 0; /* not needed */ | |||
4681 | #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ | |||
4682 | ||||
4683 | team->t.t_construct = 0; | |||
4684 | ||||
4685 | team->t.t_ordered.dt.t_value = 0; | |||
4686 | team->t.t_master_active = FALSE0; | |||
4687 | ||||
4688 | #ifdef KMP_DEBUG1 | |||
4689 | team->t.t_copypriv_data = NULL__null; /* not necessary, but nice for debugging */ | |||
4690 | #endif | |||
4691 | #if KMP_OS_WINDOWS0 | |||
4692 | team->t.t_copyin_counter = 0; /* for barrier-free copyin implementation */ | |||
4693 | #endif | |||
4694 | ||||
4695 | team->t.t_control_stack_top = NULL__null; | |||
4696 | ||||
4697 | __kmp_reinitialize_team(team, new_icvs, loc); | |||
4698 | ||||
4699 | KMP_MB(); | |||
4700 | 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); }; | |||
4701 | } | |||
4702 | ||||
4703 | #if (KMP_OS_LINUX1 || KMP_OS_FREEBSD0) && KMP_AFFINITY_SUPPORTED1 | |||
4704 | /* Sets full mask for thread and returns old mask, no changes to structures. */ | |||
4705 | static void | |||
4706 | __kmp_set_thread_affinity_mask_full_tmp(kmp_affin_mask_t *old_mask) { | |||
4707 | if (KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) { | |||
4708 | int status; | |||
4709 | if (old_mask != NULL__null) { | |||
4710 | status = __kmp_get_system_affinity(old_mask, TRUE)(old_mask)->get_system_affinity((!0)); | |||
4711 | int error = errno(*__errno_location ()); | |||
4712 | if (status != 0) { | |||
4713 | __kmp_fatal(KMP_MSG(ChangeThreadAffMaskError)__kmp_msg_format(kmp_i18n_msg_ChangeThreadAffMaskError), KMP_ERR(error)__kmp_msg_error_code(error), | |||
4714 | __kmp_msg_null); | |||
4715 | } | |||
4716 | } | |||
4717 | __kmp_set_system_affinity(__kmp_affin_fullMask, TRUE)(__kmp_affin_fullMask)->set_system_affinity((!0)); | |||
4718 | } | |||
4719 | } | |||
4720 | #endif | |||
4721 | ||||
4722 | #if KMP_AFFINITY_SUPPORTED1 | |||
4723 | ||||
4724 | // __kmp_partition_places() is the heart of the OpenMP 4.0 affinity mechanism. | |||
4725 | // It calculates the worker + primary thread's partition based upon the parent | |||
4726 | // thread's partition, and binds each worker to a thread in their partition. | |||
4727 | // The primary thread's partition should already include its current binding. | |||
4728 | static void __kmp_partition_places(kmp_team_t *team, int update_master_only) { | |||
4729 | // Do not partition places for the hidden helper team | |||
4730 | if (KMP_HIDDEN_HELPER_TEAM(team)(team->t.t_threads[0] == __kmp_hidden_helper_main_thread)) | |||
4731 | return; | |||
4732 | // Copy the primary thread's place partition to the team struct | |||
4733 | kmp_info_t *master_th = team->t.t_threads[0]; | |||
4734 | KMP_DEBUG_ASSERT(master_th != NULL)if (!(master_th != __null)) { __kmp_debug_assert("master_th != __null" , "openmp/runtime/src/kmp_runtime.cpp", 4734); }; | |||
4735 | kmp_proc_bind_t proc_bind = team->t.t_proc_bind; | |||
4736 | int first_place = master_th->th.th_first_place; | |||
4737 | int last_place = master_th->th.th_last_place; | |||
4738 | int masters_place = master_th->th.th_current_place; | |||
4739 | int num_masks = __kmp_affinity.num_masks; | |||
4740 | team->t.t_first_place = first_place; | |||
4741 | team->t.t_last_place = last_place; | |||
4742 | ||||
4743 | 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); } | |||
4744 | "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); } | |||
4745 | 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); } | |||
4746 | 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); }; | |||
4747 | ||||
4748 | switch (proc_bind) { | |||
4749 | ||||
4750 | case proc_bind_default: | |||
4751 | // Serial teams might have the proc_bind policy set to proc_bind_default. | |||
4752 | // Not an issue -- we don't rebind primary thread for any proc_bind policy. | |||
4753 | 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", 4753); }; | |||
4754 | break; | |||
4755 | ||||
4756 | case proc_bind_primary: { | |||
4757 | int f; | |||
4758 | int n_th = team->t.t_nproc; | |||
4759 | for (f = 1; f < n_th; f++) { | |||
4760 | kmp_info_t *th = team->t.t_threads[f]; | |||
4761 | KMP_DEBUG_ASSERT(th != NULL)if (!(th != __null)) { __kmp_debug_assert("th != __null", "openmp/runtime/src/kmp_runtime.cpp" , 4761); }; | |||
4762 | th->th.th_first_place = first_place; | |||
4763 | th->th.th_last_place = last_place; | |||
4764 | th->th.th_new_place = masters_place; | |||
4765 | if (__kmp_display_affinity && masters_place != th->th.th_current_place && | |||
4766 | team->t.t_display_affinity != 1) { | |||
4767 | team->t.t_display_affinity = 1; | |||
4768 | } | |||
4769 | ||||
4770 | 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 ); } | |||
4771 | "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 ); } | |||
4772 | __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 ); } | |||
4773 | 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 ); }; | |||
4774 | } | |||
4775 | } break; | |||
4776 | ||||
4777 | case proc_bind_close: { | |||
4778 | int f; | |||
4779 | int n_th = team->t.t_nproc; | |||
4780 | int n_places; | |||
4781 | if (first_place <= last_place) { | |||
4782 | n_places = last_place - first_place + 1; | |||
4783 | } else { | |||
4784 | n_places = num_masks - first_place + last_place + 1; | |||
4785 | } | |||
4786 | if (n_th <= n_places) { | |||
4787 | int place = masters_place; | |||
4788 | for (f = 1; f < n_th; f++) { | |||
4789 | kmp_info_t *th = team->t.t_threads[f]; | |||
4790 | KMP_DEBUG_ASSERT(th != NULL)if (!(th != __null)) { __kmp_debug_assert("th != __null", "openmp/runtime/src/kmp_runtime.cpp" , 4790); }; | |||
4791 | ||||
4792 | if (place == last_place) { | |||
4793 | place = first_place; | |||
4794 | } else if (place == (num_masks - 1)) { | |||
4795 | place = 0; | |||
4796 | } else { | |||
4797 | place++; | |||
4798 | } | |||
4799 | th->th.th_first_place = first_place; | |||
4800 | th->th.th_last_place = last_place; | |||
4801 | th->th.th_new_place = place; | |||
4802 | if (__kmp_display_affinity && place != th->th.th_current_place && | |||
4803 | team->t.t_display_affinity != 1) { | |||
4804 | team->t.t_display_affinity = 1; | |||
4805 | } | |||
4806 | ||||
4807 | 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); } | |||
4808 | "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); } | |||
4809 | __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); } | |||
4810 | 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); }; | |||
4811 | } | |||
4812 | } else { | |||
4813 | int S, rem, gap, s_count; | |||
4814 | S = n_th / n_places; | |||
4815 | s_count = 0; | |||
4816 | rem = n_th - (S * n_places); | |||
4817 | gap = rem > 0 ? n_places / rem : n_places; | |||
4818 | int place = masters_place; | |||
4819 | int gap_ct = gap; | |||
4820 | for (f = 0; f < n_th; f++) { | |||
4821 | kmp_info_t *th = team->t.t_threads[f]; | |||
4822 | KMP_DEBUG_ASSERT(th != NULL)if (!(th != __null)) { __kmp_debug_assert("th != __null", "openmp/runtime/src/kmp_runtime.cpp" , 4822); }; | |||
4823 | ||||
4824 | th->th.th_first_place = first_place; | |||
4825 | th->th.th_last_place = last_place; | |||
4826 | th->th.th_new_place = place; | |||
4827 | if (__kmp_display_affinity && place != th->th.th_current_place && | |||
4828 | team->t.t_display_affinity != 1) { | |||
4829 | team->t.t_display_affinity = 1; | |||
4830 | } | |||
4831 | s_count++; | |||
4832 | ||||
4833 | if ((s_count == S) && rem && (gap_ct == gap)) { | |||
4834 | // do nothing, add an extra thread to place on next iteration | |||
4835 | } else if ((s_count == S + 1) && rem && (gap_ct == gap)) { | |||
4836 | // we added an extra thread to this place; move to next place | |||
4837 | if (place == last_place) { | |||
4838 | place = first_place; | |||
4839 | } else if (place == (num_masks - 1)) { | |||
4840 | place = 0; | |||
4841 | } else { | |||
4842 | place++; | |||
4843 | } | |||
4844 | s_count = 0; | |||
4845 | gap_ct = 1; | |||
4846 | rem--; | |||
4847 | } else if (s_count == S) { // place full; don't add extra | |||
4848 | if (place == last_place) { | |||
4849 | place = first_place; | |||
4850 | } else if (place == (num_masks - 1)) { | |||
4851 | place = 0; | |||
4852 | } else { | |||
4853 | place++; | |||
4854 | } | |||
4855 | gap_ct++; | |||
4856 | s_count = 0; | |||
4857 | } | |||
4858 | ||||
4859 | 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); } | |||
4860 | ("__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); } | |||
4861 | "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); } | |||
4862 | __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); } | |||
4863 | 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); }; | |||
4864 | } | |||
4865 | KMP_DEBUG_ASSERT(place == masters_place)if (!(place == masters_place)) { __kmp_debug_assert("place == masters_place" , "openmp/runtime/src/kmp_runtime.cpp", 4865); }; | |||
4866 | } | |||
4867 | } break; | |||
4868 | ||||
4869 | case proc_bind_spread: { | |||
4870 | int f; | |||
4871 | int n_th = team->t.t_nproc; | |||
4872 | int n_places; | |||
4873 | int thidx; | |||
4874 | if (first_place <= last_place) { | |||
4875 | n_places = last_place - first_place + 1; | |||
4876 | } else { | |||
4877 | n_places = num_masks - first_place + last_place + 1; | |||
4878 | } | |||
4879 | if (n_th <= n_places) { | |||
4880 | int place = -1; | |||
4881 | ||||
4882 | if (n_places != num_masks) { | |||
4883 | int S = n_places / n_th; | |||
4884 | int s_count, rem, gap, gap_ct; | |||
4885 | ||||
4886 | place = masters_place; | |||
4887 | rem = n_places - n_th * S; | |||
4888 | gap = rem ? n_th / rem : 1; | |||
4889 | gap_ct = gap; | |||
4890 | thidx = n_th; | |||
4891 | if (update_master_only == 1) | |||
4892 | thidx = 1; | |||
4893 | for (f = 0; f < thidx; f++) { | |||
4894 | kmp_info_t *th = team->t.t_threads[f]; | |||
4895 | KMP_DEBUG_ASSERT(th != NULL)if (!(th != __null)) { __kmp_debug_assert("th != __null", "openmp/runtime/src/kmp_runtime.cpp" , 4895); }; | |||
4896 | ||||
4897 | th->th.th_first_place = place; | |||
4898 | th->th.th_new_place = place; | |||
4899 | if (__kmp_display_affinity && place != th->th.th_current_place && | |||
4900 | team->t.t_display_affinity != 1) { | |||
4901 | team->t.t_display_affinity = 1; | |||
4902 | } | |||
4903 | s_count = 1; | |||
4904 | while (s_count < S) { | |||
4905 | if (place == last_place) { | |||
4906 | place = first_place; | |||
4907 | } else if (place == (num_masks - 1)) { | |||
4908 | place = 0; | |||
4909 | } else { | |||
4910 | place++; | |||
4911 | } | |||
4912 | s_count++; | |||
4913 | } | |||
4914 | if (rem && (gap_ct == gap)) { | |||
4915 | if (place == last_place) { | |||
4916 | place = first_place; | |||
4917 | } else if (place == (num_masks - 1)) { | |||
4918 | place = 0; | |||
4919 | } else { | |||
4920 | place++; | |||
4921 | } | |||
4922 | rem--; | |||
4923 | gap_ct = 0; | |||
4924 | } | |||
4925 | th->th.th_last_place = place; | |||
4926 | gap_ct++; | |||
4927 | ||||
4928 | if (place == last_place) { | |||
4929 | place = first_place; | |||
4930 | } else if (place == (num_masks - 1)) { | |||
4931 | place = 0; | |||
4932 | } else { | |||
4933 | place++; | |||
4934 | } | |||
4935 | ||||
4936 | 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 ); } | |||
4937 | ("__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 ); } | |||
4938 | "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 ); } | |||
4939 | __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 ); } | |||
4940 | 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 ); } | |||
4941 | 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 ); }; | |||
4942 | } | |||
4943 | } else { | |||
4944 | /* Having uniform space of available computation places I can create | |||
4945 | T partitions of round(P/T) size and put threads into the first | |||
4946 | place of each partition. */ | |||
4947 | double current = static_cast<double>(masters_place); | |||
4948 | double spacing = | |||
4949 | (static_cast<double>(n_places + 1) / static_cast<double>(n_th)); | |||
4950 | int first, last; | |||
4951 | kmp_info_t *th; | |||
4952 | ||||
4953 | thidx = n_th + 1; | |||
4954 | if (update_master_only == 1) | |||
4955 | thidx = 1; | |||
4956 | for (f = 0; f < thidx; f++) { | |||
4957 | first = static_cast<int>(current); | |||
4958 | last = static_cast<int>(current + spacing) - 1; | |||
4959 | KMP_DEBUG_ASSERT(last >= first)if (!(last >= first)) { __kmp_debug_assert("last >= first" , "openmp/runtime/src/kmp_runtime.cpp", 4959); }; | |||
4960 | if (first >= n_places) { | |||
4961 | if (masters_place) { | |||
4962 | first -= n_places; | |||
4963 | last -= n_places; | |||
4964 | if (first == (masters_place + 1)) { | |||
4965 | KMP_DEBUG_ASSERT(f == n_th)if (!(f == n_th)) { __kmp_debug_assert("f == n_th", "openmp/runtime/src/kmp_runtime.cpp" , 4965); }; | |||
4966 | first--; | |||
4967 | } | |||
4968 | if (last == masters_place) { | |||
4969 | 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", 4969); }; | |||
4970 | last--; | |||
4971 | } | |||
4972 | } else { | |||
4973 | KMP_DEBUG_ASSERT(f == n_th)if (!(f == n_th)) { __kmp_debug_assert("f == n_th", "openmp/runtime/src/kmp_runtime.cpp" , 4973); }; | |||
4974 | first = 0; | |||
4975 | last = 0; | |||
4976 | } | |||
4977 | } | |||
4978 | if (last >= n_places) { | |||
4979 | last = (n_places - 1); | |||
4980 | } | |||
4981 | place = first; | |||
4982 | current += spacing; | |||
4983 | if (f < n_th) { | |||
4984 | KMP_DEBUG_ASSERT(0 <= first)if (!(0 <= first)) { __kmp_debug_assert("0 <= first", "openmp/runtime/src/kmp_runtime.cpp" , 4984); }; | |||
4985 | KMP_DEBUG_ASSERT(n_places > first)if (!(n_places > first)) { __kmp_debug_assert("n_places > first" , "openmp/runtime/src/kmp_runtime.cpp", 4985); }; | |||
4986 | KMP_DEBUG_ASSERT(0 <= last)if (!(0 <= last)) { __kmp_debug_assert("0 <= last", "openmp/runtime/src/kmp_runtime.cpp" , 4986); }; | |||
4987 | KMP_DEBUG_ASSERT(n_places > last)if (!(n_places > last)) { __kmp_debug_assert("n_places > last" , "openmp/runtime/src/kmp_runtime.cpp", 4987); }; | |||
4988 | 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", 4988); }; | |||
4989 | th = team->t.t_threads[f]; | |||
4990 | KMP_DEBUG_ASSERT(th)if (!(th)) { __kmp_debug_assert("th", "openmp/runtime/src/kmp_runtime.cpp" , 4990); }; | |||
4991 | th->th.th_first_place = first; | |||
4992 | th->th.th_new_place = place; | |||
4993 | th->th.th_last_place = last; | |||
4994 | if (__kmp_display_affinity && place != th->th.th_current_place && | |||
4995 | team->t.t_display_affinity != 1) { | |||
4996 | team->t.t_display_affinity = 1; | |||
4997 | } | |||
4998 | 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) ; } | |||
4999 | ("__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) ; } | |||
5000 | "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) ; } | |||
5001 | __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) ; } | |||
5002 | 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) ; } | |||
5003 | 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) ; }; | |||
5004 | } | |||
5005 | } | |||
5006 | } | |||
5007 | 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" , 5007); }; | |||
5008 | } else { | |||
5009 | int S, rem, gap, s_count; | |||
5010 | S = n_th / n_places; | |||
5011 | s_count = 0; | |||
5012 | rem = n_th - (S * n_places); | |||
5013 | gap = rem > 0 ? n_places / rem : n_places; | |||
5014 | int place = masters_place; | |||
5015 | int gap_ct = gap; | |||
5016 | thidx = n_th; | |||
5017 | if (update_master_only == 1) | |||
5018 | thidx = 1; | |||
5019 | for (f = 0; f < thidx; f++) { | |||
5020 | kmp_info_t *th = team->t.t_threads[f]; | |||
5021 | KMP_DEBUG_ASSERT(th != NULL)if (!(th != __null)) { __kmp_debug_assert("th != __null", "openmp/runtime/src/kmp_runtime.cpp" , 5021); }; | |||
5022 | ||||
5023 | th->th.th_first_place = place; | |||
5024 | th->th.th_last_place = place; | |||
5025 | th->th.th_new_place = place; | |||
5026 | if (__kmp_display_affinity && place != th->th.th_current_place && | |||
5027 | team->t.t_display_affinity != 1) { | |||
5028 | team->t.t_display_affinity = 1; | |||
5029 | } | |||
5030 | s_count++; | |||
5031 | ||||
5032 | if ((s_count == S) && rem && (gap_ct == gap)) { | |||
5033 | // do nothing, add an extra thread to place on next iteration | |||
5034 | } else if ((s_count == S + 1) && rem && (gap_ct == gap)) { | |||
5035 | // we added an extra thread to this place; move on to next place | |||
5036 | if (place == last_place) { | |||
5037 | place = first_place; | |||
5038 | } else if (place == (num_masks - 1)) { | |||
5039 | place = 0; | |||
5040 | } else { | |||
5041 | place++; | |||
5042 | } | |||
5043 | s_count = 0; | |||
5044 | gap_ct = 1; | |||
5045 | rem--; | |||
5046 | } else if (s_count == S) { // place is full; don't add extra thread | |||
5047 | if (place == last_place) { | |||
5048 | place = first_place; | |||
5049 | } else if (place == (num_masks - 1)) { | |||
5050 | place = 0; | |||
5051 | } else { | |||
5052 | place++; | |||
5053 | } | |||
5054 | gap_ct++; | |||
5055 | s_count = 0; | |||
5056 | } | |||
5057 | ||||
5058 | 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); } | |||
5059 | "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); } | |||
5060 | __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); } | |||
5061 | 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); } | |||
5062 | 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); }; | |||
5063 | } | |||
5064 | 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" , 5064); }; | |||
5065 | } | |||
5066 | } break; | |||
5067 | ||||
5068 | default: | |||
5069 | break; | |||
5070 | } | |||
5071 | ||||
5072 | 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); }; | |||
5073 | } | |||
5074 | ||||
5075 | #endif // KMP_AFFINITY_SUPPORTED | |||
5076 | ||||
5077 | /* allocate a new team data structure to use. take one off of the free pool if | |||
5078 | available */ | |||
5079 | kmp_team_t * | |||
5080 | __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc, | |||
5081 | #if OMPT_SUPPORT1 | |||
5082 | ompt_data_t ompt_parallel_data, | |||
5083 | #endif | |||
5084 | kmp_proc_bind_t new_proc_bind, | |||
5085 | kmp_internal_control_t *new_icvs, | |||
5086 | int argc USE_NESTED_HOT_ARG(kmp_info_t *master), kmp_info_t *master) { | |||
5087 | KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_allocate_team)((void)0); | |||
5088 | int f; | |||
5089 | kmp_team_t *team; | |||
5090 | int use_hot_team = !root->r.r_active; | |||
5091 | int level = 0; | |||
5092 | int do_place_partition = 1; | |||
5093 | ||||
5094 | KA_TRACE(20, ("__kmp_allocate_team: called\n"))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: called\n" ); }; | |||
5095 | 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" , 5095); }; | |||
5096 | 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", 5096); }; | |||
5097 | KMP_MB(); | |||
5098 | ||||
5099 | #if KMP_NESTED_HOT_TEAMS1 | |||
5100 | kmp_hot_team_ptr_t *hot_teams; | |||
5101 | if (master) { | |||
5102 | team = master->th.th_team; | |||
5103 | level = team->t.t_active_level; | |||
5104 | if (master->th.th_teams_microtask) { // in teams construct? | |||
5105 | if (master->th.th_teams_size.nteams > 1 && | |||
5106 | ( // #teams > 1 | |||
5107 | team->t.t_pkfn == | |||
5108 | (microtask_t)__kmp_teams_master || // inner fork of the teams | |||
5109 | master->th.th_teams_level < | |||
5110 | team->t.t_level)) { // or nested parallel inside the teams | |||
5111 | ++level; // not increment if #teams==1, or for outer fork of the teams; | |||
5112 | // increment otherwise | |||
5113 | } | |||
5114 | // Do not perform the place partition if inner fork of the teams | |||
5115 | // Wait until nested parallel region encountered inside teams construct | |||
5116 | if ((master->th.th_teams_size.nteams == 1 && | |||
5117 | master->th.th_teams_level >= team->t.t_level) || | |||
5118 | (team->t.t_pkfn == (microtask_t)__kmp_teams_master)) | |||
5119 | do_place_partition = 0; | |||
5120 | } | |||
5121 | hot_teams = master->th.th_hot_teams; | |||
5122 | if (level < __kmp_hot_teams_max_level && hot_teams && | |||
5123 | hot_teams[level].hot_team) { | |||
5124 | // hot team has already been allocated for given level | |||
5125 | use_hot_team = 1; | |||
5126 | } else { | |||
5127 | use_hot_team = 0; | |||
5128 | } | |||
5129 | } else { | |||
5130 | // check we won't access uninitialized hot_teams, just in case | |||
5131 | KMP_DEBUG_ASSERT(new_nproc == 1)if (!(new_nproc == 1)) { __kmp_debug_assert("new_nproc == 1", "openmp/runtime/src/kmp_runtime.cpp", 5131); }; | |||
5132 | } | |||
5133 | #endif | |||
5134 | // Optimization to use a "hot" team | |||
5135 | if (use_hot_team && new_nproc > 1) { | |||
5136 | 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", 5136); }; | |||
5137 | #if KMP_NESTED_HOT_TEAMS1 | |||
5138 | team = hot_teams[level].hot_team; | |||
5139 | #else | |||
5140 | team = root->r.r_hot_team; | |||
5141 | #endif | |||
5142 | #if KMP_DEBUG1 | |||
5143 | if (__kmp_tasking_mode != tskm_immediate_exec) { | |||
5144 | 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]); } | |||
5145 | "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]); } | |||
5146 | 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]); }; | |||
5147 | } | |||
5148 | #endif | |||
5149 | ||||
5150 | if (team->t.t_nproc != new_nproc && | |||
5151 | __kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | |||
5152 | // Distributed barrier may need a resize | |||
5153 | int old_nthr = team->t.t_nproc; | |||
5154 | __kmp_resize_dist_barrier(team, old_nthr, new_nproc); | |||
5155 | } | |||
5156 | ||||
5157 | // If not doing the place partition, then reset the team's proc bind | |||
5158 | // to indicate that partitioning of all threads still needs to take place | |||
5159 | if (do_place_partition == 0) | |||
5160 | team->t.t_proc_bind = proc_bind_default; | |||
5161 | // Has the number of threads changed? | |||
5162 | /* Let's assume the most common case is that the number of threads is | |||
5163 | unchanged, and put that case first. */ | |||
5164 | if (team->t.t_nproc == new_nproc) { // Check changes in number of threads | |||
5165 | 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" ); }; | |||
5166 | // This case can mean that omp_set_num_threads() was called and the hot | |||
5167 | // team size was already reduced, so we check the special flag | |||
5168 | if (team->t.t_size_changed == -1) { | |||
5169 | team->t.t_size_changed = 1; | |||
5170 | } else { | |||
5171 | KMP_CHECK_UPDATE(team->t.t_size_changed, 0)if ((team->t.t_size_changed) != (0)) (team->t.t_size_changed ) = (0); | |||
5172 | } | |||
5173 | ||||
5174 | // TODO???: team->t.t_max_active_levels = new_max_active_levels; | |||
5175 | kmp_r_sched_t new_sched = new_icvs->sched; | |||
5176 | // set primary thread's schedule as new run-time schedule | |||
5177 | 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); | |||
5178 | ||||
5179 | __kmp_reinitialize_team(team, new_icvs, | |||
5180 | root->r.r_uber_thread->th.th_ident); | |||
5181 | ||||
5182 | 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); } | |||
5183 | 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); }; | |||
5184 | __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0); | |||
5185 | ||||
5186 | #if KMP_AFFINITY_SUPPORTED1 | |||
5187 | if ((team->t.t_size_changed == 0) && | |||
5188 | (team->t.t_proc_bind == new_proc_bind)) { | |||
5189 | if (new_proc_bind == proc_bind_spread) { | |||
5190 | if (do_place_partition) { | |||
5191 | // add flag to update only master for spread | |||
5192 | __kmp_partition_places(team, 1); | |||
5193 | } | |||
5194 | } | |||
5195 | 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); } | |||
5196 | "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); } | |||
5197 | 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); } | |||
5198 | 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); }; | |||
5199 | } else { | |||
5200 | if (do_place_partition) { | |||
5201 | 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); | |||
5202 | __kmp_partition_places(team); | |||
5203 | } | |||
5204 | } | |||
5205 | #else | |||
5206 | 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); | |||
5207 | #endif /* KMP_AFFINITY_SUPPORTED */ | |||
5208 | } else if (team->t.t_nproc > new_nproc) { | |||
5209 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: decreasing hot team thread count to %d\n" , new_nproc); } | |||
5210 | ("__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); } | |||
5211 | new_nproc))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: decreasing hot team thread count to %d\n" , new_nproc); }; | |||
5212 | ||||
5213 | team->t.t_size_changed = 1; | |||
5214 | if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | |||
5215 | // Barrier size already reduced earlier in this function | |||
5216 | // Activate team threads via th_used_in_team | |||
5217 | __kmp_add_threads_to_team(team, new_nproc); | |||
5218 | } | |||
5219 | #if KMP_NESTED_HOT_TEAMS1 | |||
5220 | if (__kmp_hot_teams_mode == 0) { | |||
5221 | // AC: saved number of threads should correspond to team's value in this | |||
5222 | // mode, can be bigger in mode 1, when hot team has threads in reserve | |||
5223 | 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", 5223); }; | |||
5224 | hot_teams[level].hot_team_nth = new_nproc; | |||
5225 | #endif // KMP_NESTED_HOT_TEAMS | |||
5226 | /* release the extra threads we don't need any more */ | |||
5227 | for (f = new_nproc; f < team->t.t_nproc; f++) { | |||
5228 | 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", 5228); }; | |||
5229 | if (__kmp_tasking_mode != tskm_immediate_exec) { | |||
5230 | // When decreasing team size, threads no longer in the team should | |||
5231 | // unref task team. | |||
5232 | team->t.t_threads[f]->th.th_task_team = NULL__null; | |||
5233 | } | |||
5234 | __kmp_free_thread(team->t.t_threads[f]); | |||
5235 | team->t.t_threads[f] = NULL__null; | |||
5236 | } | |||
5237 | #if KMP_NESTED_HOT_TEAMS1 | |||
5238 | } // (__kmp_hot_teams_mode == 0) | |||
5239 | else { | |||
5240 | // When keeping extra threads in team, switch threads to wait on own | |||
5241 | // b_go flag | |||
5242 | for (f = new_nproc; f < team->t.t_nproc; ++f) { | |||
5243 | 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", 5243); }; | |||
5244 | kmp_balign_t *balign = team->t.t_threads[f]->th.th_bar; | |||
5245 | for (int b = 0; b < bs_last_barrier; ++b) { | |||
5246 | if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG2) { | |||
5247 | balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG3; | |||
5248 | } | |||
5249 | KMP_CHECK_UPDATE(balign[b].bb.leaf_kids, 0)if ((balign[b].bb.leaf_kids) != (0)) (balign[b].bb.leaf_kids) = (0); | |||
5250 | } | |||
5251 | } | |||
5252 | } | |||
5253 | #endif // KMP_NESTED_HOT_TEAMS | |||
5254 | team->t.t_nproc = new_nproc; | |||
5255 | // TODO???: team->t.t_max_active_levels = new_max_active_levels; | |||
5256 | 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); | |||
5257 | __kmp_reinitialize_team(team, new_icvs, | |||
5258 | root->r.r_uber_thread->th.th_ident); | |||
5259 | ||||
5260 | // Update remaining threads | |||
5261 | for (f = 0; f < new_nproc; ++f) { | |||
5262 | team->t.t_threads[f]->th.th_team_nproc = new_nproc; | |||
5263 | } | |||
5264 | ||||
5265 | // restore the current task state of the primary thread: should be the | |||
5266 | // implicit task | |||
5267 | 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); } | |||
5268 | 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); }; | |||
5269 | ||||
5270 | __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0); | |||
5271 | ||||
5272 | #ifdef KMP_DEBUG1 | |||
5273 | for (f = 0; f < team->t.t_nproc; f++) { | |||
5274 | 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", 5276); } | |||
5275 | 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", 5276); } | |||
5276 | 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", 5276); }; | |||
5277 | } | |||
5278 | #endif | |||
5279 | ||||
5280 | if (do_place_partition) { | |||
5281 | 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); | |||
5282 | #if KMP_AFFINITY_SUPPORTED1 | |||
5283 | __kmp_partition_places(team); | |||
5284 | #endif | |||
5285 | } | |||
5286 | } else { // team->t.t_nproc < new_nproc | |||
5287 | #if (KMP_OS_LINUX1 || KMP_OS_FREEBSD0) && KMP_AFFINITY_SUPPORTED1 | |||
5288 | kmp_affin_mask_t *old_mask; | |||
5289 | if (KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) { | |||
5290 | KMP_CPU_ALLOC(old_mask)(old_mask = __kmp_affinity_dispatch->allocate_mask()); | |||
5291 | } | |||
5292 | #endif | |||
5293 | ||||
5294 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: increasing hot team thread count to %d\n" , new_nproc); } | |||
5295 | ("__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); } | |||
5296 | new_nproc))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: increasing hot team thread count to %d\n" , new_nproc); }; | |||
5297 | int old_nproc = team->t.t_nproc; // save old value and use to update only | |||
5298 | team->t.t_size_changed = 1; | |||
5299 | ||||
5300 | #if KMP_NESTED_HOT_TEAMS1 | |||
5301 | int avail_threads = hot_teams[level].hot_team_nth; | |||
5302 | if (new_nproc < avail_threads) | |||
5303 | avail_threads = new_nproc; | |||
5304 | kmp_info_t **other_threads = team->t.t_threads; | |||
5305 | for (f = team->t.t_nproc; f < avail_threads; ++f) { | |||
5306 | // Adjust barrier data of reserved threads (if any) of the team | |||
5307 | // Other data will be set in __kmp_initialize_info() below. | |||
5308 | int b; | |||
5309 | kmp_balign_t *balign = other_threads[f]->th.th_bar; | |||
5310 | for (b = 0; b < bs_last_barrier; ++b) { | |||
5311 | balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; | |||
5312 | 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", 5312); }; | |||
5313 | #if USE_DEBUGGER0 | |||
5314 | balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; | |||
5315 | #endif | |||
5316 | } | |||
5317 | } | |||
5318 | if (hot_teams[level].hot_team_nth >= new_nproc) { | |||
5319 | // we have all needed threads in reserve, no need to allocate any | |||
5320 | // this only possible in mode 1, cannot have reserved threads in mode 0 | |||
5321 | 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", 5321); }; | |||
5322 | team->t.t_nproc = new_nproc; // just get reserved threads involved | |||
5323 | } else { | |||
5324 | // We may have some threads in reserve, but not enough; | |||
5325 | // get reserved threads involved if any. | |||
5326 | team->t.t_nproc = hot_teams[level].hot_team_nth; | |||
5327 | hot_teams[level].hot_team_nth = new_nproc; // adjust hot team max size | |||
5328 | #endif // KMP_NESTED_HOT_TEAMS | |||
5329 | if (team->t.t_max_nproc < new_nproc) { | |||
5330 | /* reallocate larger arrays */ | |||
5331 | __kmp_reallocate_team_arrays(team, new_nproc); | |||
5332 | __kmp_reinitialize_team(team, new_icvs, NULL__null); | |||
5333 | } | |||
5334 | ||||
5335 | #if (KMP_OS_LINUX1 || KMP_OS_FREEBSD0) && KMP_AFFINITY_SUPPORTED1 | |||
5336 | /* Temporarily set full mask for primary thread before creation of | |||
5337 | workers. The reason is that workers inherit the affinity from the | |||
5338 | primary thread, so if a lot of workers are created on the single | |||
5339 | core quickly, they don't get a chance to set their own affinity for | |||
5340 | a long time. */ | |||
5341 | __kmp_set_thread_affinity_mask_full_tmp(old_mask); | |||
5342 | #endif | |||
5343 | ||||
5344 | /* allocate new threads for the hot team */ | |||
5345 | for (f = team->t.t_nproc; f < new_nproc; f++) { | |||
5346 | kmp_info_t *new_worker = __kmp_allocate_thread(root, team, f); | |||
5347 | KMP_DEBUG_ASSERT(new_worker)if (!(new_worker)) { __kmp_debug_assert("new_worker", "openmp/runtime/src/kmp_runtime.cpp" , 5347); }; | |||
5348 | team->t.t_threads[f] = new_worker; | |||
5349 | ||||
5350 | 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); } | |||
5351 | ("__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); } | |||
5352 | "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); } | |||
5353 | 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); } | |||
5354 | 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); } | |||
5355 | 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); }; | |||
5356 | ||||
5357 | { // Initialize barrier data for new threads. | |||
5358 | int b; | |||
5359 | kmp_balign_t *balign = new_worker->th.th_bar; | |||
5360 | for (b = 0; b < bs_last_barrier; ++b) { | |||
5361 | balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; | |||
5362 | 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", 5363); } | |||
5363 | 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", 5363); }; | |||
5364 | #if USE_DEBUGGER0 | |||
5365 | balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; | |||
5366 | #endif | |||
5367 | } | |||
5368 | } | |||
5369 | } | |||
5370 | ||||
5371 | #if (KMP_OS_LINUX1 || KMP_OS_FREEBSD0) && KMP_AFFINITY_SUPPORTED1 | |||
5372 | if (KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) { | |||
5373 | /* Restore initial primary thread's affinity mask */ | |||
5374 | __kmp_set_system_affinity(old_mask, TRUE)(old_mask)->set_system_affinity((!0)); | |||
5375 | KMP_CPU_FREE(old_mask)__kmp_affinity_dispatch->deallocate_mask(old_mask); | |||
5376 | } | |||
5377 | #endif | |||
5378 | #if KMP_NESTED_HOT_TEAMS1 | |||
5379 | } // end of check of t_nproc vs. new_nproc vs. hot_team_nth | |||
5380 | #endif // KMP_NESTED_HOT_TEAMS | |||
5381 | if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | |||
5382 | // Barrier size already increased earlier in this function | |||
5383 | // Activate team threads via th_used_in_team | |||
5384 | __kmp_add_threads_to_team(team, new_nproc); | |||
5385 | } | |||
5386 | /* make sure everyone is syncronized */ | |||
5387 | // new threads below | |||
5388 | __kmp_initialize_team(team, new_nproc, new_icvs, | |||
5389 | root->r.r_uber_thread->th.th_ident); | |||
5390 | ||||
5391 | /* reinitialize the threads */ | |||
5392 | 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" , 5392); }; | |||
5393 | for (f = 0; f < team->t.t_nproc; ++f) | |||
5394 | __kmp_initialize_info(team->t.t_threads[f], team, f, | |||
5395 | __kmp_gtid_from_tid(f, team)); | |||
5396 | ||||
5397 | if (level) { // set th_task_state for new threads in nested hot team | |||
5398 | // __kmp_initialize_info() no longer zeroes th_task_state, so we should | |||
5399 | // only need to set the th_task_state for the new threads. th_task_state | |||
5400 | // for primary thread will not be accurate until after this in | |||
5401 | // __kmp_fork_call(), so we look to the primary thread's memo_stack to | |||
5402 | // get the correct value. | |||
5403 | for (f = old_nproc; f < team->t.t_nproc; ++f) | |||
5404 | team->t.t_threads[f]->th.th_task_state = | |||
5405 | team->t.t_threads[0]->th.th_task_state_memo_stack[level]; | |||
5406 | } else { // set th_task_state for new threads in non-nested hot team | |||
5407 | // copy primary thread's state | |||
5408 | kmp_uint8 old_state = team->t.t_threads[0]->th.th_task_state; | |||
5409 | for (f = old_nproc; f < team->t.t_nproc; ++f) | |||
5410 | team->t.t_threads[f]->th.th_task_state = old_state; | |||
5411 | } | |||
5412 | ||||
5413 | #ifdef KMP_DEBUG1 | |||
5414 | for (f = 0; f < team->t.t_nproc; ++f) { | |||
5415 | 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", 5417); } | |||
5416 | 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", 5417); } | |||
5417 | 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", 5417); }; | |||
5418 | } | |||
5419 | #endif | |||
5420 | ||||
5421 | if (do_place_partition) { | |||
5422 | 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); | |||
5423 | #if KMP_AFFINITY_SUPPORTED1 | |||
5424 | __kmp_partition_places(team); | |||
5425 | #endif | |||
5426 | } | |||
5427 | } // Check changes in number of threads | |||
5428 | ||||
5429 | kmp_info_t *master = team->t.t_threads[0]; | |||
5430 | if (master->th.th_teams_microtask) { | |||
5431 | for (f = 1; f < new_nproc; ++f) { | |||
5432 | // propagate teams construct specific info to workers | |||
5433 | kmp_info_t *thr = team->t.t_threads[f]; | |||
5434 | thr->th.th_teams_microtask = master->th.th_teams_microtask; | |||
5435 | thr->th.th_teams_level = master->th.th_teams_level; | |||
5436 | thr->th.th_teams_size = master->th.th_teams_size; | |||
5437 | } | |||
5438 | } | |||
5439 | #if KMP_NESTED_HOT_TEAMS1 | |||
5440 | if (level) { | |||
5441 | // Sync barrier state for nested hot teams, not needed for outermost hot | |||
5442 | // team. | |||
5443 | for (f = 1; f < new_nproc; ++f) { | |||
5444 | kmp_info_t *thr = team->t.t_threads[f]; | |||
5445 | int b; | |||
5446 | kmp_balign_t *balign = thr->th.th_bar; | |||
5447 | for (b = 0; b < bs_last_barrier; ++b) { | |||
5448 | balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; | |||
5449 | 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", 5449); }; | |||
5450 | #if USE_DEBUGGER0 | |||
5451 | balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; | |||
5452 | #endif | |||
5453 | } | |||
5454 | } | |||
5455 | } | |||
5456 | #endif // KMP_NESTED_HOT_TEAMS | |||
5457 | ||||
5458 | /* reallocate space for arguments if necessary */ | |||
5459 | __kmp_alloc_argv_entries(argc, team, TRUE(!0)); | |||
5460 | KMP_CHECK_UPDATE(team->t.t_argc, argc)if ((team->t.t_argc) != (argc)) (team->t.t_argc) = (argc ); | |||
5461 | // The hot team re-uses the previous task team, | |||
5462 | // if untouched during the previous release->gather phase. | |||
5463 | ||||
5464 | KF_TRACE(10, (" hot_team = %p\n", team))if (kmp_f_debug >= 10) { __kmp_debug_printf (" hot_team = %p\n" , team); }; | |||
5465 | ||||
5466 | #if KMP_DEBUG1 | |||
5467 | if (__kmp_tasking_mode != tskm_immediate_exec) { | |||
5468 | 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]); } | |||
5469 | "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]); } | |||
5470 | 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]); }; | |||
5471 | } | |||
5472 | #endif | |||
5473 | ||||
5474 | #if OMPT_SUPPORT1 | |||
5475 | __ompt_team_assign_id(team, ompt_parallel_data); | |||
5476 | #endif | |||
5477 | ||||
5478 | KMP_MB(); | |||
5479 | ||||
5480 | return team; | |||
5481 | } | |||
5482 | ||||
5483 | /* next, let's try to take one from the team pool */ | |||
5484 | KMP_MB(); | |||
5485 | for (team = CCAST(kmp_team_t *, __kmp_team_pool)const_cast<kmp_team_t *>(__kmp_team_pool); (team);) { | |||
5486 | /* TODO: consider resizing undersized teams instead of reaping them, now | |||
5487 | that we have a resizing mechanism */ | |||
5488 | if (team->t.t_max_nproc >= max_nproc) { | |||
5489 | /* take this team from the team pool */ | |||
5490 | __kmp_team_pool = team->t.t_next_pool; | |||
5491 | ||||
5492 | if (max_nproc > 1 && | |||
5493 | __kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | |||
5494 | if (!team->t.b) { // Allocate barrier structure | |||
5495 | team->t.b = distributedBarrier::allocate(__kmp_dflt_team_nth_ub); | |||
5496 | } | |||
5497 | } | |||
5498 | ||||
5499 | /* setup the team for fresh use */ | |||
5500 | __kmp_initialize_team(team, new_nproc, new_icvs, NULL__null); | |||
5501 | ||||
5502 | 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]); } | |||
5503 | "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]); } | |||
5504 | &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]); }; | |||
5505 | team->t.t_task_team[0] = NULL__null; | |||
5506 | team->t.t_task_team[1] = NULL__null; | |||
5507 | ||||
5508 | /* reallocate space for arguments if necessary */ | |||
5509 | __kmp_alloc_argv_entries(argc, team, TRUE(!0)); | |||
5510 | KMP_CHECK_UPDATE(team->t.t_argc, argc)if ((team->t.t_argc) != (argc)) (team->t.t_argc) = (argc ); | |||
5511 | ||||
5512 | 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); } | |||
5513 | 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); } | |||
5514 | 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); }; | |||
5515 | { // Initialize barrier data. | |||
5516 | int b; | |||
5517 | for (b = 0; b < bs_last_barrier; ++b) { | |||
5518 | team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE0; | |||
5519 | #if USE_DEBUGGER0 | |||
5520 | team->t.t_bar[b].b_master_arrived = 0; | |||
5521 | team->t.t_bar[b].b_team_arrived = 0; | |||
5522 | #endif | |||
5523 | } | |||
5524 | } | |||
5525 | ||||
5526 | team->t.t_proc_bind = new_proc_bind; | |||
5527 | ||||
5528 | 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); } | |||
5529 | 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); }; | |||
5530 | ||||
5531 | #if OMPT_SUPPORT1 | |||
5532 | __ompt_team_assign_id(team, ompt_parallel_data); | |||
5533 | #endif | |||
5534 | ||||
5535 | KMP_MB(); | |||
5536 | ||||
5537 | return team; | |||
5538 | } | |||
5539 | ||||
5540 | /* reap team if it is too small, then loop back and check the next one */ | |||
5541 | // not sure if this is wise, but, will be redone during the hot-teams | |||
5542 | // rewrite. | |||
5543 | /* TODO: Use technique to find the right size hot-team, don't reap them */ | |||
5544 | team = __kmp_reap_team(team); | |||
5545 | __kmp_team_pool = team; | |||
5546 | } | |||
5547 | ||||
5548 | /* nothing available in the pool, no matter, make a new team! */ | |||
5549 | KMP_MB(); | |||
5550 | team = (kmp_team_t *)__kmp_allocate(sizeof(kmp_team_t))___kmp_allocate((sizeof(kmp_team_t)), "openmp/runtime/src/kmp_runtime.cpp" , 5550); | |||
5551 | ||||
5552 | /* and set it up */ | |||
5553 | team->t.t_max_nproc = max_nproc; | |||
5554 | if (max_nproc > 1 && | |||
5555 | __kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | |||
5556 | // Allocate barrier structure | |||
5557 | team->t.b = distributedBarrier::allocate(__kmp_dflt_team_nth_ub); | |||
5558 | } | |||
5559 | ||||
5560 | /* NOTE well, for some reason allocating one big buffer and dividing it up | |||
5561 | seems to really hurt performance a lot on the P4, so, let's not use this */ | |||
5562 | __kmp_allocate_team_arrays(team, max_nproc); | |||
5563 | ||||
5564 | 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" ); }; | |||
5565 | __kmp_initialize_team(team, new_nproc, new_icvs, NULL__null); | |||
5566 | ||||
5567 | 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]); } | |||
5568 | "%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]); } | |||
5569 | &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]); }; | |||
5570 | team->t.t_task_team[0] = NULL__null; // to be removed, as __kmp_allocate zeroes | |||
5571 | // memory, no need to duplicate | |||
5572 | team->t.t_task_team[1] = NULL__null; // to be removed, as __kmp_allocate zeroes | |||
5573 | // memory, no need to duplicate | |||
5574 | ||||
5575 | if (__kmp_storage_map) { | |||
5576 | __kmp_print_team_storage_map("team", team, team->t.t_id, new_nproc); | |||
5577 | } | |||
5578 | ||||
5579 | /* allocate space for arguments */ | |||
5580 | __kmp_alloc_argv_entries(argc, team, FALSE0); | |||
5581 | team->t.t_argc = argc; | |||
5582 | ||||
5583 | 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); } | |||
5584 | ("__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); } | |||
5585 | 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); }; | |||
5586 | { // Initialize barrier data. | |||
5587 | int b; | |||
5588 | for (b = 0; b < bs_last_barrier; ++b) { | |||
5589 | team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE0; | |||
5590 | #if USE_DEBUGGER0 | |||
5591 | team->t.t_bar[b].b_master_arrived = 0; | |||
5592 | team->t.t_bar[b].b_team_arrived = 0; | |||
5593 | #endif | |||
5594 | } | |||
5595 | } | |||
5596 | ||||
5597 | team->t.t_proc_bind = new_proc_bind; | |||
5598 | ||||
5599 | #if OMPT_SUPPORT1 | |||
5600 | __ompt_team_assign_id(team, ompt_parallel_data); | |||
5601 | team->t.ompt_serialized_team_info = NULL__null; | |||
5602 | #endif | |||
5603 | ||||
5604 | KMP_MB(); | |||
5605 | ||||
5606 | 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); } | |||
5607 | 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); }; | |||
5608 | ||||
5609 | return team; | |||
5610 | } | |||
5611 | ||||
5612 | /* TODO implement hot-teams at all levels */ | |||
5613 | /* TODO implement lazy thread release on demand (disband request) */ | |||
5614 | ||||
5615 | /* free the team. return it to the team pool. release all the threads | |||
5616 | * associated with it */ | |||
5617 | void __kmp_free_team(kmp_root_t *root, | |||
5618 | kmp_team_t *team USE_NESTED_HOT_ARG(kmp_info_t *master), kmp_info_t *master) { | |||
5619 | int f; | |||
5620 | 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); } | |||
5621 | 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); }; | |||
5622 | ||||
5623 | /* verify state */ | |||
5624 | KMP_DEBUG_ASSERT(root)if (!(root)) { __kmp_debug_assert("root", "openmp/runtime/src/kmp_runtime.cpp" , 5624); }; | |||
5625 | KMP_DEBUG_ASSERT(team)if (!(team)) { __kmp_debug_assert("team", "openmp/runtime/src/kmp_runtime.cpp" , 5625); }; | |||
5626 | 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" , 5626); }; | |||
5627 | 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", 5627); }; | |||
5628 | ||||
5629 | int use_hot_team = team == root->r.r_hot_team; | |||
5630 | #if KMP_NESTED_HOT_TEAMS1 | |||
5631 | int level; | |||
5632 | if (master) { | |||
5633 | level = team->t.t_active_level - 1; | |||
5634 | if (master->th.th_teams_microtask) { // in teams construct? | |||
5635 | if (master->th.th_teams_size.nteams > 1) { | |||
5636 | ++level; // level was not increased in teams construct for | |||
5637 | // team_of_masters | |||
5638 | } | |||
5639 | if (team->t.t_pkfn != (microtask_t)__kmp_teams_master && | |||
5640 | master->th.th_teams_level == team->t.t_level) { | |||
5641 | ++level; // level was not increased in teams construct for | |||
5642 | // team_of_workers before the parallel | |||
5643 | } // team->t.t_level will be increased inside parallel | |||
5644 | } | |||
5645 | #if KMP_DEBUG1 | |||
5646 | kmp_hot_team_ptr_t *hot_teams = master->th.th_hot_teams; | |||
5647 | #endif | |||
5648 | if (level < __kmp_hot_teams_max_level) { | |||
5649 | 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" , 5649); }; | |||
5650 | use_hot_team = 1; | |||
5651 | } | |||
5652 | } | |||
5653 | #endif // KMP_NESTED_HOT_TEAMS | |||
5654 | ||||
5655 | /* team is done working */ | |||
5656 | TCW_SYNC_PTR(team->t.t_pkfn,((team->t.t_pkfn)) = ((__null)) | |||
5657 | NULL)((team->t.t_pkfn)) = ((__null)); // Important for Debugging Support Library. | |||
5658 | #if KMP_OS_WINDOWS0 | |||
5659 | team->t.t_copyin_counter = 0; // init counter for possible reuse | |||
5660 | #endif | |||
5661 | // Do not reset pointer to parent team to NULL for hot teams. | |||
5662 | ||||
5663 | /* if we are non-hot team, release our threads */ | |||
5664 | if (!use_hot_team) { | |||
5665 | if (__kmp_tasking_mode != tskm_immediate_exec) { | |||
5666 | // Wait for threads to reach reapable state | |||
5667 | for (f = 1; f < team->t.t_nproc; ++f) { | |||
5668 | 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", 5668); }; | |||
5669 | kmp_info_t *th = team->t.t_threads[f]; | |||
5670 | volatile kmp_uint32 *state = &th->th.th_reap_state; | |||
5671 | while (*state != KMP_SAFE_TO_REAP1) { | |||
5672 | #if KMP_OS_WINDOWS0 | |||
5673 | // On Windows a thread can be killed at any time, check this | |||
5674 | DWORD ecode; | |||
5675 | if (!__kmp_is_thread_alive(th, &ecode)) { | |||
5676 | *state = KMP_SAFE_TO_REAP1; // reset the flag for dead thread | |||
5677 | break; | |||
5678 | } | |||
5679 | #endif | |||
5680 | // first check if thread is sleeping | |||
5681 | kmp_flag_64<> fl(&th->th.th_bar[bs_forkjoin_barrier].bb.b_go, th); | |||
5682 | if (fl.is_sleeping()) | |||
5683 | fl.resume(__kmp_gtid_from_thread(th)); | |||
5684 | KMP_CPU_PAUSE()__kmp_x86_pause(); | |||
5685 | } | |||
5686 | } | |||
5687 | ||||
5688 | // Delete task teams | |||
5689 | int tt_idx; | |||
5690 | for (tt_idx = 0; tt_idx < 2; ++tt_idx) { | |||
5691 | kmp_task_team_t *task_team = team->t.t_task_team[tt_idx]; | |||
5692 | if (task_team != NULL__null) { | |||
5693 | for (f = 0; f < team->t.t_nproc; ++f) { // threads unref task teams | |||
5694 | 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", 5694); }; | |||
5695 | team->t.t_threads[f]->th.th_task_team = NULL__null; | |||
5696 | } | |||
5697 | 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); } | |||
5698 | 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); } | |||
5699 | ("__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); } | |||
5700 | __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); }; | |||
5701 | #if KMP_NESTED_HOT_TEAMS1 | |||
5702 | __kmp_free_task_team(master, task_team); | |||
5703 | #endif | |||
5704 | team->t.t_task_team[tt_idx] = NULL__null; | |||
5705 | } | |||
5706 | } | |||
5707 | } | |||
5708 | ||||
5709 | // Reset pointer to parent team only for non-hot teams. | |||
5710 | team->t.t_parent = NULL__null; | |||
5711 | team->t.t_level = 0; | |||
5712 | team->t.t_active_level = 0; | |||
5713 | ||||
5714 | /* free the worker threads */ | |||
5715 | for (f = 1; f < team->t.t_nproc; ++f) { | |||
5716 | 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", 5716); }; | |||
5717 | if (__kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | |||
5718 | 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)) | |||
5719 | 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)); | |||
5720 | } | |||
5721 | __kmp_free_thread(team->t.t_threads[f]); | |||
5722 | } | |||
5723 | ||||
5724 | if (__kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | |||
5725 | if (team->t.b) { | |||
5726 | // wake up thread at old location | |||
5727 | team->t.b->go_release(); | |||
5728 | if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME(2147483647)) { | |||
5729 | for (f = 1; f < team->t.t_nproc; ++f) { | |||
5730 | if (team->t.b->sleep[f].sleep) { | |||
5731 | __kmp_atomic_resume_64( | |||
5732 | team->t.t_threads[f]->th.th_info.ds.ds_gtid, | |||
5733 | (kmp_atomic_flag_64<> *)NULL__null); | |||
5734 | } | |||
5735 | } | |||
5736 | } | |||
5737 | // Wait for threads to be removed from team | |||
5738 | for (int f = 1; f < team->t.t_nproc; ++f) { | |||
5739 | while (team->t.t_threads[f]->th.th_used_in_team.load() != 0) | |||
5740 | KMP_CPU_PAUSE()__kmp_x86_pause(); | |||
5741 | } | |||
5742 | } | |||
5743 | } | |||
5744 | ||||
5745 | for (f = 1; f < team->t.t_nproc; ++f) { | |||
5746 | team->t.t_threads[f] = NULL__null; | |||
5747 | } | |||
5748 | ||||
5749 | if (team->t.t_max_nproc > 1 && | |||
5750 | __kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | |||
5751 | distributedBarrier::deallocate(team->t.b); | |||
5752 | team->t.b = NULL__null; | |||
5753 | } | |||
5754 | /* put the team back in the team pool */ | |||
5755 | /* TODO limit size of team pool, call reap_team if pool too large */ | |||
5756 | team->t.t_next_pool = CCAST(kmp_team_t *, __kmp_team_pool)const_cast<kmp_team_t *>(__kmp_team_pool); | |||
5757 | __kmp_team_pool = (volatile kmp_team_t *)team; | |||
5758 | } else { // Check if team was created for primary threads in teams construct | |||
5759 | // See if first worker is a CG root | |||
5760 | 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", 5761); } | |||
5761 | 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", 5761); }; | |||
5762 | if (team->t.t_threads[1]->th.th_cg_roots->cg_root == team->t.t_threads[1]) { | |||
5763 | // Clean up the CG root nodes on workers so that this team can be re-used | |||
5764 | for (f = 1; f < team->t.t_nproc; ++f) { | |||
5765 | kmp_info_t *thr = team->t.t_threads[f]; | |||
5766 | 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", 5767); } | |||
5767 | 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", 5767); }; | |||
5768 | // Pop current CG root off list | |||
5769 | kmp_cg_root_t *tmp = thr->th.th_cg_roots; | |||
5770 | thr->th.th_cg_roots = tmp->up; | |||
5771 | 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); } | |||
5772 | " 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); } | |||
5773 | 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); }; | |||
5774 | int i = tmp->cg_nthreads--; | |||
5775 | if (i == 1) { | |||
5776 | __kmp_free(tmp)___kmp_free((tmp), "openmp/runtime/src/kmp_runtime.cpp", 5776 ); // free CG if we are the last thread in it | |||
5777 | } | |||
5778 | // Restore current task's thread_limit from CG root | |||
5779 | if (thr->th.th_cg_roots) | |||
5780 | thr->th.th_current_task->td_icvs.thread_limit = | |||
5781 | thr->th.th_cg_roots->cg_thread_limit; | |||
5782 | } | |||
5783 | } | |||
5784 | } | |||
5785 | ||||
5786 | KMP_MB(); | |||
5787 | } | |||
5788 | ||||
5789 | /* reap the team. destroy it, reclaim all its resources and free its memory */ | |||
5790 | kmp_team_t *__kmp_reap_team(kmp_team_t *team) { | |||
5791 | kmp_team_t *next_pool = team->t.t_next_pool; | |||
5792 | ||||
5793 | KMP_DEBUG_ASSERT(team)if (!(team)) { __kmp_debug_assert("team", "openmp/runtime/src/kmp_runtime.cpp" , 5793); }; | |||
5794 | 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", 5794); }; | |||
5795 | 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", 5795); }; | |||
5796 | 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", 5796); }; | |||
5797 | 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", 5797); }; | |||
5798 | ||||
5799 | /* TODO clean the threads that are a part of this? */ | |||
5800 | ||||
5801 | /* free stuff */ | |||
5802 | __kmp_free_team_arrays(team); | |||
5803 | if (team->t.t_argv != &team->t.t_inline_argv[0]) | |||
5804 | __kmp_free((void *)team->t.t_argv)___kmp_free(((void *)team->t.t_argv), "openmp/runtime/src/kmp_runtime.cpp" , 5804); | |||
5805 | __kmp_free(team)___kmp_free((team), "openmp/runtime/src/kmp_runtime.cpp", 5805 ); | |||
5806 | ||||
5807 | KMP_MB(); | |||
5808 | return next_pool; | |||
5809 | } | |||
5810 | ||||
5811 | // Free the thread. Don't reap it, just place it on the pool of available | |||
5812 | // threads. | |||
5813 | // | |||
5814 | // Changes for Quad issue 527845: We need a predictable OMP tid <-> gtid | |||
5815 | // binding for the affinity mechanism to be useful. | |||
5816 | // | |||
5817 | // Now, we always keep the free list (__kmp_thread_pool) sorted by gtid. | |||
5818 | // However, we want to avoid a potential performance problem by always | |||
5819 | // scanning through the list to find the correct point at which to insert | |||
5820 | // the thread (potential N**2 behavior). To do this we keep track of the | |||
5821 | // last place a thread struct was inserted (__kmp_thread_pool_insert_pt). | |||
5822 | // With single-level parallelism, threads will always be added to the tail | |||
5823 | // of the list, kept track of by __kmp_thread_pool_insert_pt. With nested | |||
5824 | // parallelism, all bets are off and we may need to scan through the entire | |||
5825 | // free list. | |||
5826 | // | |||
5827 | // This change also has a potentially large performance benefit, for some | |||
5828 | // applications. Previously, as threads were freed from the hot team, they | |||
5829 | // would be placed back on the free list in inverse order. If the hot team | |||
5830 | // grew back to it's original size, then the freed thread would be placed | |||
5831 | // back on the hot team in reverse order. This could cause bad cache | |||
5832 | // locality problems on programs where the size of the hot team regularly | |||
5833 | // grew and shrunk. | |||
5834 | // | |||
5835 | // Now, for single-level parallelism, the OMP tid is always == gtid. | |||
5836 | void __kmp_free_thread(kmp_info_t *this_th) { | |||
5837 | int gtid; | |||
5838 | kmp_info_t **scan; | |||
5839 | ||||
5840 | 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 ); } | |||
5841 | __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 ); }; | |||
5842 | ||||
5843 | KMP_DEBUG_ASSERT(this_th)if (!(this_th)) { __kmp_debug_assert("this_th", "openmp/runtime/src/kmp_runtime.cpp" , 5843); }; | |||
5844 | ||||
5845 | // When moving thread to pool, switch thread to wait on own b_go flag, and | |||
5846 | // uninitialized (NULL team). | |||
5847 | int b; | |||
5848 | kmp_balign_t *balign = this_th->th.th_bar; | |||
5849 | for (b = 0; b < bs_last_barrier; ++b) { | |||
5850 | if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG2) | |||
5851 | balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG3; | |||
5852 | balign[b].bb.team = NULL__null; | |||
5853 | balign[b].bb.leaf_kids = 0; | |||
5854 | } | |||
5855 | this_th->th.th_task_state = 0; | |||
5856 | this_th->th.th_reap_state = KMP_SAFE_TO_REAP1; | |||
5857 | ||||
5858 | /* put thread back on the free pool */ | |||
5859 | TCW_PTR(this_th->th.th_team, NULL)((this_th->th.th_team)) = ((__null)); | |||
5860 | TCW_PTR(this_th->th.th_root, NULL)((this_th->th.th_root)) = ((__null)); | |||
5861 | TCW_PTR(this_th->th.th_dispatch, NULL)((this_th->th.th_dispatch)) = ((__null)); /* NOT NEEDED */ | |||
5862 | ||||
5863 | while (this_th->th.th_cg_roots) { | |||
5864 | this_th->th.th_cg_roots->cg_nthreads--; | |||
5865 | 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); } | |||
5866 | " %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); } | |||
5867 | 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); } | |||
5868 | 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); } | |||
5869 | 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); }; | |||
5870 | kmp_cg_root_t *tmp = this_th->th.th_cg_roots; | |||
5871 | if (tmp->cg_root == this_th) { // Thread is a cg_root | |||
5872 | 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", 5872); }; | |||
5873 | KA_TRACE(if (kmp_a_debug >= 5) { __kmp_debug_printf ("__kmp_free_thread: Thread %p freeing node %p\n" , this_th, tmp); } | |||
5874 | 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); }; | |||
5875 | this_th->th.th_cg_roots = tmp->up; | |||
5876 | __kmp_free(tmp)___kmp_free((tmp), "openmp/runtime/src/kmp_runtime.cpp", 5876 ); | |||
5877 | } else { // Worker thread | |||
5878 | if (tmp->cg_nthreads == 0) { // last thread leaves contention group | |||
5879 | __kmp_free(tmp)___kmp_free((tmp), "openmp/runtime/src/kmp_runtime.cpp", 5879 ); | |||
5880 | } | |||
5881 | this_th->th.th_cg_roots = NULL__null; | |||
5882 | break; | |||
5883 | } | |||
5884 | } | |||
5885 | ||||
5886 | /* If the implicit task assigned to this thread can be used by other threads | |||
5887 | * -> multiple threads can share the data and try to free the task at | |||
5888 | * __kmp_reap_thread at exit. This duplicate use of the task data can happen | |||
5889 | * with higher probability when hot team is disabled but can occurs even when | |||
5890 | * the hot team is enabled */ | |||
5891 | __kmp_free_implicit_task(this_th); | |||
5892 | this_th->th.th_current_task = NULL__null; | |||
5893 | ||||
5894 | // If the __kmp_thread_pool_insert_pt is already past the new insert | |||
5895 | // point, then we need to re-scan the entire list. | |||
5896 | gtid = this_th->th.th_info.ds.ds_gtid; | |||
5897 | if (__kmp_thread_pool_insert_pt != NULL__null) { | |||
5898 | 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", 5898); }; | |||
5899 | if (__kmp_thread_pool_insert_pt->th.th_info.ds.ds_gtid > gtid) { | |||
5900 | __kmp_thread_pool_insert_pt = NULL__null; | |||
5901 | } | |||
5902 | } | |||
5903 | ||||
5904 | // Scan down the list to find the place to insert the thread. | |||
5905 | // scan is the address of a link in the list, possibly the address of | |||
5906 | // __kmp_thread_pool itself. | |||
5907 | // | |||
5908 | // In the absence of nested parallelism, the for loop will have 0 iterations. | |||
5909 | if (__kmp_thread_pool_insert_pt != NULL__null) { | |||
5910 | scan = &(__kmp_thread_pool_insert_pt->th.th_next_pool); | |||
5911 | } else { | |||
5912 | scan = CCAST(kmp_info_t **, &__kmp_thread_pool)const_cast<kmp_info_t **>(&__kmp_thread_pool); | |||
5913 | } | |||
5914 | for (; (*scan != NULL__null) && ((*scan)->th.th_info.ds.ds_gtid < gtid); | |||
5915 | scan = &((*scan)->th.th_next_pool)) | |||
5916 | ; | |||
5917 | ||||
5918 | // Insert the new element on the list, and set __kmp_thread_pool_insert_pt | |||
5919 | // to its address. | |||
5920 | TCW_PTR(this_th->th.th_next_pool, *scan)((this_th->th.th_next_pool)) = ((*scan)); | |||
5921 | __kmp_thread_pool_insert_pt = *scan = this_th; | |||
5922 | 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", 5924); } | |||
5923 | (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", 5924); } | |||
5924 | 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", 5924); }; | |||
5925 | TCW_4(this_th->th.th_in_pool, TRUE)(this_th->th.th_in_pool) = ((!0)); | |||
5926 | __kmp_suspend_initialize_thread(this_th); | |||
5927 | __kmp_lock_suspend_mx(this_th); | |||
5928 | if (this_th->th.th_active == TRUE(!0)) { | |||
5929 | KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth)(&__kmp_thread_pool_active_nth)->fetch_add(1, std::memory_order_acq_rel ); | |||
5930 | this_th->th.th_active_in_pool = TRUE(!0); | |||
5931 | } | |||
5932 | #if KMP_DEBUG1 | |||
5933 | else { | |||
5934 | 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" , 5934); }; | |||
5935 | } | |||
5936 | #endif | |||
5937 | __kmp_unlock_suspend_mx(this_th); | |||
5938 | ||||
5939 | TCW_4(__kmp_nth, __kmp_nth - 1)(__kmp_nth) = (__kmp_nth - 1); | |||
5940 | ||||
5941 | #ifdef KMP_ADJUST_BLOCKTIME1 | |||
5942 | /* Adjust blocktime back to user setting or default if necessary */ | |||
5943 | /* Middle initialization might never have occurred */ | |||
5944 | if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { | |||
5945 | 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", 5945); }; | |||
5946 | if (__kmp_nth <= __kmp_avail_proc) { | |||
5947 | __kmp_zero_bt = FALSE0; | |||
5948 | } | |||
5949 | } | |||
5950 | #endif /* KMP_ADJUST_BLOCKTIME */ | |||
5951 | ||||
5952 | KMP_MB(); | |||
5953 | } | |||
5954 | ||||
5955 | /* ------------------------------------------------------------------------ */ | |||
5956 | ||||
5957 | void *__kmp_launch_thread(kmp_info_t *this_thr) { | |||
5958 | #if OMP_PROFILING_SUPPORT0 | |||
5959 | ProfileTraceFile = getenv("LIBOMPTARGET_PROFILE"); | |||
5960 | // TODO: add a configuration option for time granularity | |||
5961 | if (ProfileTraceFile) | |||
5962 | llvm::timeTraceProfilerInitialize(500 /* us */, "libomptarget"); | |||
5963 | #endif | |||
5964 | ||||
5965 | int gtid = this_thr->th.th_info.ds.ds_gtid; | |||
5966 | /* void *stack_data;*/ | |||
5967 | kmp_team_t **volatile pteam; | |||
5968 | ||||
5969 | KMP_MB(); | |||
5970 | 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); }; | |||
5971 | ||||
5972 | if (__kmp_env_consistency_check) { | |||
5973 | this_thr->th.th_cons = __kmp_allocate_cons_stack(gtid); // ATT: Memory leak? | |||
5974 | } | |||
5975 | ||||
5976 | #if OMPD_SUPPORT1 | |||
5977 | if (ompd_state & OMPD_ENABLE_BP0x1) | |||
5978 | ompd_bp_thread_begin(); | |||
5979 | #endif | |||
5980 | ||||
5981 | #if OMPT_SUPPORT1 | |||
5982 | ompt_data_t *thread_data = nullptr; | |||
5983 | if (ompt_enabled.enabled) { | |||
5984 | thread_data = &(this_thr->th.ompt_thread_info.thread_data); | |||
5985 | *thread_data = ompt_data_none{0}; | |||
5986 | ||||
5987 | this_thr->th.ompt_thread_info.state = ompt_state_overhead; | |||
5988 | this_thr->th.ompt_thread_info.wait_id = 0; | |||
5989 | this_thr->th.ompt_thread_info.idle_frame = OMPT_GET_FRAME_ADDRESS(0)__builtin_frame_address(0); | |||
5990 | this_thr->th.ompt_thread_info.parallel_flags = 0; | |||
5991 | if (ompt_enabled.ompt_callback_thread_begin) { | |||
5992 | ompt_callbacks.ompt_callback(ompt_callback_thread_begin)ompt_callback_thread_begin_callback( | |||
5993 | ompt_thread_worker, thread_data); | |||
5994 | } | |||
5995 | this_thr->th.ompt_thread_info.state = ompt_state_idle; | |||
5996 | } | |||
5997 | #endif | |||
5998 | ||||
5999 | /* This is the place where threads wait for work */ | |||
6000 | while (!TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done)) { | |||
6001 | 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" , 6001); }; | |||
6002 | KMP_MB(); | |||
6003 | ||||
6004 | /* wait for work to do */ | |||
6005 | 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); }; | |||
6006 | ||||
6007 | /* No tid yet since not part of a team */ | |||
6008 | __kmp_fork_barrier(gtid, KMP_GTID_DNE(-2)); | |||
6009 | ||||
6010 | #if OMPT_SUPPORT1 | |||
6011 | if (ompt_enabled.enabled) { | |||
6012 | this_thr->th.ompt_thread_info.state = ompt_state_overhead; | |||
6013 | } | |||
6014 | #endif | |||
6015 | ||||
6016 | pteam = &this_thr->th.th_team; | |||
6017 | ||||
6018 | /* have we been allocated? */ | |||
6019 | if (TCR_SYNC_PTR(*pteam)((void *)(*pteam)) && !TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done)) { | |||
6020 | /* we were just woken up, so run our new task */ | |||
6021 | if (TCR_SYNC_PTR((*pteam)->t.t_pkfn)((void *)((*pteam)->t.t_pkfn)) != NULL__null) { | |||
6022 | int rc; | |||
6023 | 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); } | |||
6024 | ("__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); } | |||
6025 | 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); } | |||
6026 | (*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); }; | |||
6027 | ||||
6028 | updateHWFPControl(*pteam); | |||
6029 | ||||
6030 | #if OMPT_SUPPORT1 | |||
6031 | if (ompt_enabled.enabled) { | |||
6032 | this_thr->th.ompt_thread_info.state = ompt_state_work_parallel; | |||
6033 | } | |||
6034 | #endif | |||
6035 | ||||
6036 | rc = (*pteam)->t.t_invoke(gtid); | |||
6037 | KMP_ASSERT(rc)if (!(rc)) { __kmp_debug_assert("rc", "openmp/runtime/src/kmp_runtime.cpp" , 6037); }; | |||
6038 | ||||
6039 | KMP_MB(); | |||
6040 | 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); } | |||
6041 | 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); } | |||
6042 | (*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); }; | |||
6043 | } | |||
6044 | #if OMPT_SUPPORT1 | |||
6045 | if (ompt_enabled.enabled) { | |||
6046 | /* no frame set while outside task */ | |||
6047 | __ompt_get_task_info_object(0)->frame.exit_frame = ompt_data_none{0}; | |||
6048 | ||||
6049 | this_thr->th.ompt_thread_info.state = ompt_state_overhead; | |||
6050 | } | |||
6051 | #endif | |||
6052 | /* join barrier after parallel region */ | |||
6053 | __kmp_join_barrier(gtid); | |||
6054 | } | |||
6055 | } | |||
6056 | TCR_SYNC_PTR((intptr_t)__kmp_global.g.g_done)((void *)((intptr_t)__kmp_global.g.g_done)); | |||
6057 | ||||
6058 | #if OMPD_SUPPORT1 | |||
6059 | if (ompd_state & OMPD_ENABLE_BP0x1) | |||
6060 | ompd_bp_thread_end(); | |||
6061 | #endif | |||
6062 | ||||
6063 | #if OMPT_SUPPORT1 | |||
6064 | if (ompt_enabled.ompt_callback_thread_end) { | |||
6065 | ompt_callbacks.ompt_callback(ompt_callback_thread_end)ompt_callback_thread_end_callback(thread_data); | |||
6066 | } | |||
6067 | #endif | |||
6068 | ||||
6069 | this_thr->th.th_task_team = NULL__null; | |||
6070 | /* run the destructors for the threadprivate data for this thread */ | |||
6071 | __kmp_common_destroy_gtid(gtid); | |||
6072 | ||||
6073 | 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); }; | |||
6074 | KMP_MB(); | |||
6075 | ||||
6076 | #if OMP_PROFILING_SUPPORT0 | |||
6077 | llvm::timeTraceProfilerFinishThread(); | |||
6078 | #endif | |||
6079 | return this_thr; | |||
6080 | } | |||
6081 | ||||
6082 | /* ------------------------------------------------------------------------ */ | |||
6083 | ||||
6084 | void __kmp_internal_end_dest(void *specific_gtid) { | |||
6085 | // Make sure no significant bits are lost | |||
6086 | int gtid; | |||
6087 | __kmp_type_convert((kmp_intptr_t)specific_gtid - 1, >id); | |||
6088 | ||||
6089 | 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); }; | |||
6090 | /* NOTE: the gtid is stored as gitd+1 in the thread-local-storage | |||
6091 | * this is because 0 is reserved for the nothing-stored case */ | |||
6092 | ||||
6093 | __kmp_internal_end_thread(gtid); | |||
6094 | } | |||
6095 | ||||
6096 | #if KMP_OS_UNIX1 && KMP_DYNAMIC_LIB1 | |||
6097 | ||||
6098 | __attribute__((destructor)) void __kmp_internal_end_dtor(void) { | |||
6099 | __kmp_internal_end_atexit(); | |||
6100 | } | |||
6101 | ||||
6102 | #endif | |||
6103 | ||||
6104 | /* [Windows] josh: when the atexit handler is called, there may still be more | |||
6105 | than one thread alive */ | |||
6106 | void __kmp_internal_end_atexit(void) { | |||
6107 | KA_TRACE(30, ("__kmp_internal_end_atexit\n"))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_internal_end_atexit\n" ); }; | |||
6108 | /* [Windows] | |||
6109 | josh: ideally, we want to completely shutdown the library in this atexit | |||
6110 | handler, but stat code that depends on thread specific data for gtid fails | |||
6111 | because that data becomes unavailable at some point during the shutdown, so | |||
6112 | we call __kmp_internal_end_thread instead. We should eventually remove the | |||
6113 | dependency on __kmp_get_specific_gtid in the stat code and use | |||
6114 | __kmp_internal_end_library to cleanly shutdown the library. | |||
6115 | ||||
6116 | // TODO: Can some of this comment about GVS be removed? | |||
6117 | I suspect that the offending stat code is executed when the calling thread | |||
6118 | tries to clean up a dead root thread's data structures, resulting in GVS | |||
6119 | code trying to close the GVS structures for that thread, but since the stat | |||
6120 | code uses __kmp_get_specific_gtid to get the gtid with the assumption that | |||
6121 | the calling thread is cleaning up itself instead of another thread, it get | |||
6122 | confused. This happens because allowing a thread to unregister and cleanup | |||
6123 | another thread is a recent modification for addressing an issue. | |||
6124 | Based on the current design (20050722), a thread may end up | |||
6125 | trying to unregister another thread only if thread death does not trigger | |||
6126 | the calling of __kmp_internal_end_thread. For Linux* OS, there is the | |||
6127 | thread specific data destructor function to detect thread death. For | |||
6128 | Windows dynamic, there is DllMain(THREAD_DETACH). For Windows static, there | |||
6129 | is nothing. Thus, the workaround is applicable only for Windows static | |||
6130 | stat library. */ | |||
6131 | __kmp_internal_end_library(-1); | |||
6132 | #if KMP_OS_WINDOWS0 | |||
6133 | __kmp_close_console(); | |||
6134 | #endif | |||
6135 | } | |||
6136 | ||||
6137 | static void __kmp_reap_thread(kmp_info_t *thread, int is_root) { | |||
6138 | // It is assumed __kmp_forkjoin_lock is acquired. | |||
6139 | ||||
6140 | int gtid; | |||
6141 | ||||
6142 | KMP_DEBUG_ASSERT(thread != NULL)if (!(thread != __null)) { __kmp_debug_assert("thread != __null" , "openmp/runtime/src/kmp_runtime.cpp", 6142); }; | |||
6143 | ||||
6144 | gtid = thread->th.th_info.ds.ds_gtid; | |||
6145 | ||||
6146 | if (!is_root) { | |||
6147 | if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME(2147483647)) { | |||
6148 | /* Assume the threads are at the fork barrier here */ | |||
6149 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n" , gtid); } | |||
6150 | 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); } | |||
6151 | gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n" , gtid); }; | |||
6152 | if (__kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { | |||
6153 | while ( | |||
6154 | !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))) | |||
6155 | KMP_CPU_PAUSE()__kmp_x86_pause(); | |||
6156 | __kmp_resume_32(gtid, (kmp_flag_32<false, false> *)NULL__null); | |||
6157 | } else { | |||
6158 | /* Need release fence here to prevent seg faults for tree forkjoin | |||
6159 | barrier (GEH) */ | |||
6160 | kmp_flag_64<> flag(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, | |||
6161 | thread); | |||
6162 | __kmp_release_64(&flag); | |||
6163 | } | |||
6164 | } | |||
6165 | ||||
6166 | // Terminate OS thread. | |||
6167 | __kmp_reap_worker(thread); | |||
6168 | ||||
6169 | // The thread was killed asynchronously. If it was actively | |||
6170 | // spinning in the thread pool, decrement the global count. | |||
6171 | // | |||
6172 | // There is a small timing hole here - if the worker thread was just waking | |||
6173 | // up after sleeping in the pool, had reset it's th_active_in_pool flag but | |||
6174 | // not decremented the global counter __kmp_thread_pool_active_nth yet, then | |||
6175 | // the global counter might not get updated. | |||
6176 | // | |||
6177 | // Currently, this can only happen as the library is unloaded, | |||
6178 | // so there are no harmful side effects. | |||
6179 | if (thread->th.th_active_in_pool) { | |||
6180 | thread->th.th_active_in_pool = FALSE0; | |||
6181 | KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth)(&__kmp_thread_pool_active_nth)->fetch_sub(1, std::memory_order_acq_rel ); | |||
6182 | 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" , 6182); }; | |||
6183 | } | |||
6184 | } | |||
6185 | ||||
6186 | __kmp_free_implicit_task(thread); | |||
6187 | ||||
6188 | // Free the fast memory for tasking | |||
6189 | #if USE_FAST_MEMORY3 | |||
6190 | __kmp_free_fast_memory(thread); | |||
6191 | #endif /* USE_FAST_MEMORY */ | |||
6192 | ||||
6193 | __kmp_suspend_uninitialize_thread(thread); | |||
6194 | ||||
6195 | 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", 6195); }; | |||
6196 | TCW_SYNC_PTR(__kmp_threads[gtid], NULL)((__kmp_threads[gtid])) = ((__null)); | |||
6197 | ||||
6198 | --__kmp_all_nth; | |||
6199 | // __kmp_nth was decremented when thread is added to the pool. | |||
6200 | ||||
6201 | #ifdef KMP_ADJUST_BLOCKTIME1 | |||
6202 | /* Adjust blocktime back to user setting or default if necessary */ | |||
6203 | /* Middle initialization might never have occurred */ | |||
6204 | if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { | |||
6205 | 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", 6205); }; | |||
6206 | if (__kmp_nth <= __kmp_avail_proc) { | |||
6207 | __kmp_zero_bt = FALSE0; | |||
6208 | } | |||
6209 | } | |||
6210 | #endif /* KMP_ADJUST_BLOCKTIME */ | |||
6211 | ||||
6212 | /* free the memory being used */ | |||
6213 | if (__kmp_env_consistency_check) { | |||
6214 | if (thread->th.th_cons) { | |||
6215 | __kmp_free_cons_stack(thread->th.th_cons); | |||
6216 | thread->th.th_cons = NULL__null; | |||
6217 | } | |||
6218 | } | |||
6219 | ||||
6220 | if (thread->th.th_pri_common != NULL__null) { | |||
6221 | __kmp_free(thread->th.th_pri_common)___kmp_free((thread->th.th_pri_common), "openmp/runtime/src/kmp_runtime.cpp" , 6221); | |||
6222 | thread->th.th_pri_common = NULL__null; | |||
6223 | } | |||
6224 | ||||
6225 | if (thread->th.th_task_state_memo_stack != NULL__null) { | |||
6226 | __kmp_free(thread->th.th_task_state_memo_stack)___kmp_free((thread->th.th_task_state_memo_stack), "openmp/runtime/src/kmp_runtime.cpp" , 6226); | |||
6227 | thread->th.th_task_state_memo_stack = NULL__null; | |||
6228 | } | |||
6229 | ||||
6230 | #if KMP_USE_BGET1 | |||
6231 | if (thread->th.th_local.bget_data != NULL__null) { | |||
6232 | __kmp_finalize_bget(thread); | |||
6233 | } | |||
6234 | #endif | |||
6235 | ||||
6236 | #if KMP_AFFINITY_SUPPORTED1 | |||
6237 | if (thread->th.th_affin_mask != NULL__null) { | |||
6238 | KMP_CPU_FREE(thread->th.th_affin_mask)__kmp_affinity_dispatch->deallocate_mask(thread->th.th_affin_mask ); | |||
6239 | thread->th.th_affin_mask = NULL__null; | |||
6240 | } | |||
6241 | #endif /* KMP_AFFINITY_SUPPORTED */ | |||
6242 | ||||
6243 | #if KMP_USE_HIER_SCHED0 | |||
6244 | if (thread->th.th_hier_bar_data != NULL__null) { | |||
6245 | __kmp_free(thread->th.th_hier_bar_data)___kmp_free((thread->th.th_hier_bar_data), "openmp/runtime/src/kmp_runtime.cpp" , 6245); | |||
6246 | thread->th.th_hier_bar_data = NULL__null; | |||
6247 | } | |||
6248 | #endif | |||
6249 | ||||
6250 | __kmp_reap_team(thread->th.th_serial_team); | |||
6251 | thread->th.th_serial_team = NULL__null; | |||
6252 | __kmp_free(thread)___kmp_free((thread), "openmp/runtime/src/kmp_runtime.cpp", 6252 ); | |||
6253 | ||||
6254 | KMP_MB(); | |||
6255 | ||||
6256 | } // __kmp_reap_thread | |||
6257 | ||||
6258 | static void __kmp_itthash_clean(kmp_info_t *th) { | |||
6259 | #if USE_ITT_NOTIFY1 | |||
6260 | if (__kmp_itt_region_domains.count > 0) { | |||
6261 | for (int i = 0; i < KMP_MAX_FRAME_DOMAINS; ++i) { | |||
6262 | kmp_itthash_entry_t *bucket = __kmp_itt_region_domains.buckets[i]; | |||
6263 | while (bucket) { | |||
6264 | kmp_itthash_entry_t *next = bucket->next_in_bucket; | |||
6265 | __kmp_thread_free(th, bucket)___kmp_thread_free((th), (bucket), "openmp/runtime/src/kmp_runtime.cpp" , 6265); | |||
6266 | bucket = next; | |||
6267 | } | |||
6268 | } | |||
6269 | } | |||
6270 | if (__kmp_itt_barrier_domains.count > 0) { | |||
6271 | for (int i = 0; i < KMP_MAX_FRAME_DOMAINS; ++i) { | |||
6272 | kmp_itthash_entry_t *bucket = __kmp_itt_barrier_domains.buckets[i]; | |||
6273 | while (bucket) { | |||
6274 | kmp_itthash_entry_t *next = bucket->next_in_bucket; | |||
6275 | __kmp_thread_free(th, bucket)___kmp_thread_free((th), (bucket), "openmp/runtime/src/kmp_runtime.cpp" , 6275); | |||
6276 | bucket = next; | |||
6277 | } | |||
6278 | } | |||
6279 | } | |||
6280 | #endif | |||
6281 | } | |||
6282 | ||||
6283 | static void __kmp_internal_end(void) { | |||
6284 | int i; | |||
6285 | ||||
6286 | /* First, unregister the library */ | |||
6287 | __kmp_unregister_library(); | |||
6288 | ||||
6289 | #if KMP_OS_WINDOWS0 | |||
6290 | /* In Win static library, we can't tell when a root actually dies, so we | |||
6291 | reclaim the data structures for any root threads that have died but not | |||
6292 | unregistered themselves, in order to shut down cleanly. | |||
6293 | In Win dynamic library we also can't tell when a thread dies. */ | |||
6294 | __kmp_reclaim_dead_roots(); // AC: moved here to always clean resources of | |||
6295 | // dead roots | |||
6296 | #endif | |||
6297 | ||||
6298 | for (i = 0; i < __kmp_threads_capacity; i++) | |||
6299 | if (__kmp_root[i]) | |||
6300 | if (__kmp_root[i]->r.r_active) | |||
6301 | break; | |||
6302 | KMP_MB(); /* Flush all pending memory write invalidates. */ | |||
6303 | TCW_SYNC_4(__kmp_global.g.g_done, TRUE)(__kmp_global.g.g_done) = ((!0)); | |||
6304 | ||||
6305 | if (i < __kmp_threads_capacity) { | |||
6306 | #if KMP_USE_MONITOR | |||
6307 | // 2009-09-08 (lev): Other alive roots found. Why do we kill the monitor?? | |||
6308 | KMP_MB(); /* Flush all pending memory write invalidates. */ | |||
6309 | ||||
6310 | // Need to check that monitor was initialized before reaping it. If we are | |||
6311 | // called form __kmp_atfork_child (which sets __kmp_init_parallel = 0), then | |||
6312 | // __kmp_monitor will appear to contain valid data, but it is only valid in | |||
6313 | // the parent process, not the child. | |||
6314 | // New behavior (201008): instead of keying off of the flag | |||
6315 | // __kmp_init_parallel, the monitor thread creation is keyed off | |||
6316 | // of the new flag __kmp_init_monitor. | |||
6317 | __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock); | |||
6318 | if (TCR_4(__kmp_init_monitor)(__kmp_init_monitor)) { | |||
6319 | __kmp_reap_monitor(&__kmp_monitor); | |||
6320 | TCW_4(__kmp_init_monitor, 0)(__kmp_init_monitor) = (0); | |||
6321 | } | |||
6322 | __kmp_release_bootstrap_lock(&__kmp_monitor_lock); | |||
6323 | KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end: monitor reaped\n" ); }; | |||
6324 | #endif // KMP_USE_MONITOR | |||
6325 | } else { | |||
6326 | /* TODO move this to cleanup code */ | |||
6327 | #ifdef KMP_DEBUG1 | |||
6328 | /* make sure that everything has properly ended */ | |||
6329 | for (i = 0; i < __kmp_threads_capacity; i++) { | |||
6330 | if (__kmp_root[i]) { | |||
6331 | // KMP_ASSERT( ! KMP_UBER_GTID( i ) ); // AC: | |||
6332 | // there can be uber threads alive here | |||
6333 | 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", 6333); }; // TODO: can they be active? | |||
6334 | } | |||
6335 | } | |||
6336 | #endif | |||
6337 | ||||
6338 | KMP_MB(); | |||
6339 | ||||
6340 | // Reap the worker threads. | |||
6341 | // This is valid for now, but be careful if threads are reaped sooner. | |||
6342 | while (__kmp_thread_pool != NULL__null) { // Loop thru all the thread in the pool. | |||
6343 | // Get the next thread from the pool. | |||
6344 | kmp_info_t *thread = CCAST(kmp_info_t *, __kmp_thread_pool)const_cast<kmp_info_t *>(__kmp_thread_pool); | |||
6345 | __kmp_thread_pool = thread->th.th_next_pool; | |||
6346 | // Reap it. | |||
6347 | 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" , 6347); }; | |||
6348 | thread->th.th_next_pool = NULL__null; | |||
6349 | thread->th.th_in_pool = FALSE0; | |||
6350 | __kmp_reap_thread(thread, 0); | |||
6351 | } | |||
6352 | __kmp_thread_pool_insert_pt = NULL__null; | |||
6353 | ||||
6354 | // Reap teams. | |||
6355 | while (__kmp_team_pool != NULL__null) { // Loop thru all the teams in the pool. | |||
6356 | // Get the next team from the pool. | |||
6357 | kmp_team_t *team = CCAST(kmp_team_t *, __kmp_team_pool)const_cast<kmp_team_t *>(__kmp_team_pool); | |||
6358 | __kmp_team_pool = team->t.t_next_pool; | |||
6359 | // Reap it. | |||
6360 | team->t.t_next_pool = NULL__null; | |||
6361 | __kmp_reap_team(team); | |||
6362 | } | |||
6363 | ||||
6364 | __kmp_reap_task_teams(); | |||
6365 | ||||
6366 | #if KMP_OS_UNIX1 | |||
6367 | // Threads that are not reaped should not access any resources since they | |||
6368 | // are going to be deallocated soon, so the shutdown sequence should wait | |||
6369 | // until all threads either exit the final spin-waiting loop or begin | |||
6370 | // sleeping after the given blocktime. | |||
6371 | for (i = 0; i < __kmp_threads_capacity; i++) { | |||
6372 | kmp_info_t *thr = __kmp_threads[i]; | |||
6373 | while (thr && KMP_ATOMIC_LD_ACQ(&thr->th.th_blocking)(&thr->th.th_blocking)->load(std::memory_order_acquire )) | |||
6374 | KMP_CPU_PAUSE()__kmp_x86_pause(); | |||
6375 | } | |||
6376 | #endif | |||
6377 | ||||
6378 | for (i = 0; i < __kmp_threads_capacity; ++i) { | |||
6379 | // TBD: Add some checking... | |||
6380 | // Something like KMP_DEBUG_ASSERT( __kmp_thread[ i ] == NULL ); | |||
6381 | } | |||
6382 | ||||
6383 | /* Make sure all threadprivate destructors get run by joining with all | |||
6384 | worker threads before resetting this flag */ | |||
6385 | TCW_SYNC_4(__kmp_init_common, FALSE)(__kmp_init_common) = (0); | |||
6386 | ||||
6387 | 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" ); }; | |||
6388 | KMP_MB(); | |||
6389 | ||||
6390 | #if KMP_USE_MONITOR | |||
6391 | // See note above: One of the possible fixes for CQ138434 / CQ140126 | |||
6392 | // | |||
6393 | // FIXME: push both code fragments down and CSE them? | |||
6394 | // push them into __kmp_cleanup() ? | |||
6395 | __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock); | |||
6396 | if (TCR_4(__kmp_init_monitor)(__kmp_init_monitor)) { | |||
6397 | __kmp_reap_monitor(&__kmp_monitor); | |||
6398 | TCW_4(__kmp_init_monitor, 0)(__kmp_init_monitor) = (0); | |||
6399 | } | |||
6400 | __kmp_release_bootstrap_lock(&__kmp_monitor_lock); | |||
6401 | KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end: monitor reaped\n" ); }; | |||
6402 | #endif | |||
6403 | } /* else !__kmp_global.t_active */ | |||
6404 | TCW_4(__kmp_init_gtid, FALSE)(__kmp_init_gtid) = (0); | |||
6405 | KMP_MB(); /* Flush all pending memory write invalidates. */ | |||
6406 | ||||
6407 | __kmp_cleanup(); | |||
6408 | #if OMPT_SUPPORT1 | |||
6409 | ompt_fini(); | |||
6410 | #endif | |||
6411 | } | |||
6412 | ||||
6413 | void __kmp_internal_end_library(int gtid_req) { | |||
6414 | /* if we have already cleaned up, don't try again, it wouldn't be pretty */ | |||
6415 | /* this shouldn't be a race condition because __kmp_internal_end() is the | |||
6416 | only place to clear __kmp_serial_init */ | |||
6417 | /* we'll check this later too, after we get the lock */ | |||
6418 | // 2009-09-06: We do not set g_abort without setting g_done. This check looks | |||
6419 | // redundant, because the next check will work in any case. | |||
6420 | if (__kmp_global.g.g_abort) { | |||
6421 | 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" ); }; | |||
6422 | /* TODO abort? */ | |||
6423 | return; | |||
6424 | } | |||
6425 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done) || !__kmp_init_serial) { | |||
6426 | 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" ); }; | |||
6427 | return; | |||
6428 | } | |||
6429 | ||||
6430 | // If hidden helper team has been initialized, we need to deinit it | |||
6431 | if (TCR_4(__kmp_init_hidden_helper)(__kmp_init_hidden_helper) && | |||
6432 | !TCR_4(__kmp_hidden_helper_team_done)(__kmp_hidden_helper_team_done)) { | |||
6433 | TCW_SYNC_4(__kmp_hidden_helper_team_done, TRUE)(__kmp_hidden_helper_team_done) = ((!0)); | |||
6434 | // First release the main thread to let it continue its work | |||
6435 | __kmp_hidden_helper_main_thread_release(); | |||
6436 | // Wait until the hidden helper team has been destroyed | |||
6437 | __kmp_hidden_helper_threads_deinitz_wait(); | |||
6438 | } | |||
6439 | ||||
6440 | KMP_MB(); /* Flush all pending memory write invalidates. */ | |||
6441 | /* find out who we are and what we should do */ | |||
6442 | { | |||
6443 | int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific(); | |||
6444 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: enter T#%d (%d)\n" , gtid, gtid_req); } | |||
6445 | 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); }; | |||
6446 | if (gtid == KMP_GTID_SHUTDOWN(-3)) { | |||
6447 | 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"); } | |||
6448 | "already shutdown\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: !__kmp_init_runtime, system " "already shutdown\n"); }; | |||
6449 | return; | |||
6450 | } else if (gtid == KMP_GTID_MONITOR(-4)) { | |||
6451 | 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"); } | |||
6452 | "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"); }; | |||
6453 | return; | |||
6454 | } else if (gtid == KMP_GTID_DNE(-2)) { | |||
6455 | 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"); } | |||
6456 | "shutdown\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: gtid not registered or system " "shutdown\n"); }; | |||
6457 | /* we don't know who we are, but we may still shutdown the library */ | |||
6458 | } else if (KMP_UBER_GTID(gtid)) { | |||
6459 | /* unregister ourselves as an uber thread. gtid is no longer valid */ | |||
6460 | if (__kmp_root[gtid]->r.r_active) { | |||
6461 | __kmp_global.g.g_abort = -1; | |||
6462 | TCW_SYNC_4(__kmp_global.g.g_done, TRUE)(__kmp_global.g.g_done) = ((!0)); | |||
6463 | __kmp_unregister_library(); | |||
6464 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: root still active, abort T#%d\n" , gtid); } | |||
6465 | ("__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); } | |||
6466 | gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: root still active, abort T#%d\n" , gtid); }; | |||
6467 | return; | |||
6468 | } else { | |||
6469 | __kmp_itthash_clean(__kmp_threads[gtid]); | |||
6470 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: unregistering sibling T#%d\n" , gtid); } | |||
6471 | 10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: unregistering sibling T#%d\n" , gtid); } | |||
6472 | ("__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); }; | |||
6473 | __kmp_unregister_root_current_thread(gtid); | |||
6474 | } | |||
6475 | } else { | |||
6476 | /* worker threads may call this function through the atexit handler, if they | |||
6477 | * call exit() */ | |||
6478 | /* For now, skip the usual subsequent processing and just dump the debug buffer. | |||
6479 | TODO: do a thorough shutdown instead */ | |||
6480 | #ifdef DUMP_DEBUG_ON_EXIT | |||
6481 | if (__kmp_debug_buf) | |||
6482 | __kmp_dump_debug_buffer(); | |||
6483 | #endif | |||
6484 | // added unregister library call here when we switch to shm linux | |||
6485 | // if we don't, it will leave lots of files in /dev/shm | |||
6486 | // cleanup shared memory file before exiting. | |||
6487 | __kmp_unregister_library(); | |||
6488 | return; | |||
6489 | } | |||
6490 | } | |||
6491 | /* synchronize the termination process */ | |||
6492 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); | |||
6493 | ||||
6494 | /* have we already finished */ | |||
6495 | if (__kmp_global.g.g_abort) { | |||
6496 | 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" ); }; | |||
6497 | /* TODO abort? */ | |||
6498 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | |||
6499 | return; | |||
6500 | } | |||
6501 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done) || !__kmp_init_serial) { | |||
6502 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | |||
6503 | return; | |||
6504 | } | |||
6505 | ||||
6506 | /* We need this lock to enforce mutex between this reading of | |||
6507 | __kmp_threads_capacity and the writing by __kmp_register_root. | |||
6508 | Alternatively, we can use a counter of roots that is atomically updated by | |||
6509 | __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and | |||
6510 | __kmp_internal_end_*. */ | |||
6511 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); | |||
6512 | ||||
6513 | /* now we can safely conduct the actual termination */ | |||
6514 | __kmp_internal_end(); | |||
6515 | ||||
6516 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | |||
6517 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | |||
6518 | ||||
6519 | KA_TRACE(10, ("__kmp_internal_end_library: exit\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: exit\n" ); }; | |||
6520 | ||||
6521 | #ifdef DUMP_DEBUG_ON_EXIT | |||
6522 | if (__kmp_debug_buf) | |||
6523 | __kmp_dump_debug_buffer(); | |||
6524 | #endif | |||
6525 | ||||
6526 | #if KMP_OS_WINDOWS0 | |||
6527 | __kmp_close_console(); | |||
6528 | #endif | |||
6529 | ||||
6530 | __kmp_fini_allocator(); | |||
6531 | ||||
6532 | } // __kmp_internal_end_library | |||
6533 | ||||
6534 | void __kmp_internal_end_thread(int gtid_req) { | |||
6535 | int i; | |||
6536 | ||||
6537 | /* if we have already cleaned up, don't try again, it wouldn't be pretty */ | |||
6538 | /* this shouldn't be a race condition because __kmp_internal_end() is the | |||
6539 | * only place to clear __kmp_serial_init */ | |||
6540 | /* we'll check this later too, after we get the lock */ | |||
6541 | // 2009-09-06: We do not set g_abort without setting g_done. This check looks | |||
6542 | // redundant, because the next check will work in any case. | |||
6543 | if (__kmp_global.g.g_abort) { | |||
6544 | 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" ); }; | |||
6545 | /* TODO abort? */ | |||
6546 | return; | |||
6547 | } | |||
6548 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done) || !__kmp_init_serial) { | |||
6549 | 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" ); }; | |||
6550 | return; | |||
6551 | } | |||
6552 | ||||
6553 | // If hidden helper team has been initialized, we need to deinit it | |||
6554 | if (TCR_4(__kmp_init_hidden_helper)(__kmp_init_hidden_helper) && | |||
6555 | !TCR_4(__kmp_hidden_helper_team_done)(__kmp_hidden_helper_team_done)) { | |||
6556 | TCW_SYNC_4(__kmp_hidden_helper_team_done, TRUE)(__kmp_hidden_helper_team_done) = ((!0)); | |||
6557 | // First release the main thread to let it continue its work | |||
6558 | __kmp_hidden_helper_main_thread_release(); | |||
6559 | // Wait until the hidden helper team has been destroyed | |||
6560 | __kmp_hidden_helper_threads_deinitz_wait(); | |||
6561 | } | |||
6562 | ||||
6563 | KMP_MB(); /* Flush all pending memory write invalidates. */ | |||
6564 | ||||
6565 | /* find out who we are and what we should do */ | |||
6566 | { | |||
6567 | int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific(); | |||
6568 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: enter T#%d (%d)\n" , gtid, gtid_req); } | |||
6569 | ("__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); }; | |||
6570 | if (gtid == KMP_GTID_SHUTDOWN(-3)) { | |||
6571 | 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"); } | |||
6572 | "already shutdown\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: !__kmp_init_runtime, system " "already shutdown\n"); }; | |||
6573 | return; | |||
6574 | } else if (gtid == KMP_GTID_MONITOR(-4)) { | |||
6575 | 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"); } | |||
6576 | "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"); }; | |||
6577 | return; | |||
6578 | } else if (gtid == KMP_GTID_DNE(-2)) { | |||
6579 | 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"); } | |||
6580 | "shutdown\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: gtid not registered or system " "shutdown\n"); }; | |||
6581 | return; | |||
6582 | /* we don't know who we are */ | |||
6583 | } else if (KMP_UBER_GTID(gtid)) { | |||
6584 | /* unregister ourselves as an uber thread. gtid is no longer valid */ | |||
6585 | if (__kmp_root[gtid]->r.r_active) { | |||
6586 | __kmp_global.g.g_abort = -1; | |||
6587 | TCW_SYNC_4(__kmp_global.g.g_done, TRUE)(__kmp_global.g.g_done) = ((!0)); | |||
6588 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: root still active, abort T#%d\n" , gtid); } | |||
6589 | ("__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); } | |||
6590 | gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: root still active, abort T#%d\n" , gtid); }; | |||
6591 | return; | |||
6592 | } else { | |||
6593 | 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); } | |||
6594 | gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: unregistering sibling T#%d\n" , gtid); }; | |||
6595 | __kmp_unregister_root_current_thread(gtid); | |||
6596 | } | |||
6597 | } else { | |||
6598 | /* just a worker thread, let's leave */ | |||
6599 | 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); }; | |||
6600 | ||||
6601 | if (gtid >= 0) { | |||
6602 | __kmp_threads[gtid]->th.th_task_team = NULL__null; | |||
6603 | } | |||
6604 | ||||
6605 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n" , gtid); } | |||
6606 | ("__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); } | |||
6607 | gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n" , gtid); }; | |||
6608 | return; | |||
6609 | } | |||
6610 | } | |||
6611 | #if KMP_DYNAMIC_LIB1 | |||
6612 | if (__kmp_pause_status != kmp_hard_paused) | |||
6613 | // AC: lets not shutdown the dynamic library at the exit of uber thread, | |||
6614 | // because we will better shutdown later in the library destructor. | |||
6615 | { | |||
6616 | 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); }; | |||
6617 | return; | |||
6618 | } | |||
6619 | #endif | |||
6620 | /* synchronize the termination process */ | |||
6621 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); | |||
6622 | ||||
6623 | /* have we already finished */ | |||
6624 | if (__kmp_global.g.g_abort) { | |||
6625 | 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" ); }; | |||
6626 | /* TODO abort? */ | |||
6627 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | |||
6628 | return; | |||
6629 | } | |||
6630 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done) || !__kmp_init_serial) { | |||
6631 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | |||
6632 | return; | |||
6633 | } | |||
6634 | ||||
6635 | /* We need this lock to enforce mutex between this reading of | |||
6636 | __kmp_threads_capacity and the writing by __kmp_register_root. | |||
6637 | Alternatively, we can use a counter of roots that is atomically updated by | |||
6638 | __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and | |||
6639 | __kmp_internal_end_*. */ | |||
6640 | ||||
6641 | /* should we finish the run-time? are all siblings done? */ | |||
6642 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); | |||
6643 | ||||
6644 | for (i = 0; i < __kmp_threads_capacity; ++i) { | |||
6645 | if (KMP_UBER_GTID(i)) { | |||
6646 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n" , i); } | |||
6647 | 10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n" , i); } | |||
6648 | ("__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); }; | |||
6649 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | |||
6650 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | |||
6651 | return; | |||
6652 | } | |||
6653 | } | |||
6654 | ||||
6655 | /* now we can safely conduct the actual termination */ | |||
6656 | ||||
6657 | __kmp_internal_end(); | |||
6658 | ||||
6659 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); | |||
6660 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | |||
6661 | ||||
6662 | 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); }; | |||
6663 | ||||
6664 | #ifdef DUMP_DEBUG_ON_EXIT | |||
6665 | if (__kmp_debug_buf) | |||
6666 | __kmp_dump_debug_buffer(); | |||
6667 | #endif | |||
6668 | } // __kmp_internal_end_thread | |||
6669 | ||||
6670 | // ----------------------------------------------------------------------------- | |||
6671 | // Library registration stuff. | |||
6672 | ||||
6673 | static long __kmp_registration_flag = 0; | |||
6674 | // Random value used to indicate library initialization. | |||
6675 | static char *__kmp_registration_str = NULL__null; | |||
6676 | // Value to be saved in env var __KMP_REGISTERED_LIB_<pid>. | |||
6677 | ||||
6678 | static inline char *__kmp_reg_status_name() { | |||
6679 | /* On RHEL 3u5 if linked statically, getpid() returns different values in | |||
6680 | each thread. If registration and unregistration go in different threads | |||
6681 | (omp_misc_other_root_exit.cpp test case), the name of registered_lib_env | |||
6682 | env var can not be found, because the name will contain different pid. */ | |||
6683 | // macOS* complains about name being too long with additional getuid() | |||
6684 | #if KMP_OS_UNIX1 && !KMP_OS_DARWIN0 && KMP_DYNAMIC_LIB1 | |||
6685 | return __kmp_str_format("__KMP_REGISTERED_LIB_%d_%d", (int)getpid(), | |||
6686 | (int)getuid()); | |||
6687 | #else | |||
6688 | return __kmp_str_format("__KMP_REGISTERED_LIB_%d", (int)getpid()); | |||
6689 | #endif | |||
6690 | } // __kmp_reg_status_get | |||
6691 | ||||
6692 | void __kmp_register_library_startup(void) { | |||
6693 | ||||
6694 | char *name = __kmp_reg_status_name(); // Name of the environment variable. | |||
6695 | int done = 0; | |||
6696 | union { | |||
6697 | double dtime; | |||
6698 | long ltime; | |||
6699 | } time; | |||
6700 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 | |||
6701 | __kmp_initialize_system_tick(); | |||
6702 | #endif | |||
6703 | __kmp_read_system_time(&time.dtime); | |||
6704 | __kmp_registration_flag = 0xCAFE0000L | (time.ltime & 0x0000FFFFL); | |||
6705 | __kmp_registration_str = | |||
6706 | __kmp_str_format("%p-%lx-%s", &__kmp_registration_flag, | |||
6707 | __kmp_registration_flag, KMP_LIBRARY_FILE"libomp.so.5"); | |||
6708 | ||||
6709 | 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); } | |||
6710 | __kmp_registration_str))if (kmp_a_debug >= 50) { __kmp_debug_printf ("__kmp_register_library_startup: %s=\"%s\"\n" , name, __kmp_registration_str); }; | |||
6711 | ||||
6712 | while (!done) { | |||
6713 | ||||
6714 | char *value = NULL__null; // Actual value of the environment variable. | |||
6715 | ||||
6716 | #if defined(KMP_USE_SHM) | |||
6717 | char *shm_name = __kmp_str_format("/%s", name); | |||
6718 | int shm_preexist = 0; | |||
6719 | char *data1; | |||
6720 | int fd1 = shm_open(shm_name, O_CREAT0100 | O_EXCL0200 | O_RDWR02, 0666); | |||
6721 | if ((fd1 == -1) && (errno(*__errno_location ()) == EEXIST17)) { | |||
6722 | // file didn't open because it already exists. | |||
6723 | // try opening existing file | |||
6724 | fd1 = shm_open(shm_name, O_RDWR02, 0666); | |||
6725 | if (fd1 == -1) { // file didn't open | |||
6726 | // error out here | |||
6727 | __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), | |||
6728 | __kmp_msg_null); | |||
6729 | } else { | |||
6730 | // able to open existing file | |||
6731 | shm_preexist = 1; | |||
6732 | } | |||
6733 | } else if (fd1 == -1) { // SHM didn't open; it was due to error other than | |||
6734 | // already exists. | |||
6735 | // error out here. | |||
6736 | __kmp_fatal(KMP_MSG(FunctionError, "Can't open SHM2")__kmp_msg_format(kmp_i18n_msg_FunctionError, "Can't open SHM2" ), KMP_ERR(errno)__kmp_msg_error_code((*__errno_location ())), | |||
6737 | __kmp_msg_null); | |||
6738 | } | |||
6739 | if (shm_preexist == 0) { | |||
6740 | // we created SHM now set size | |||
6741 | if (ftruncate(fd1, SHM_SIZE1024) == -1) { | |||
6742 | // error occured setting size; | |||
6743 | __kmp_fatal(KMP_MSG(FunctionError, "Can't set size of SHM")__kmp_msg_format(kmp_i18n_msg_FunctionError, "Can't set size of SHM" ), | |||
6744 | KMP_ERR(errno)__kmp_msg_error_code((*__errno_location ())), __kmp_msg_null); | |||
6745 | } | |||
6746 | } | |||
6747 | data1 = | |||
6748 | (char *)mmap(0, SHM_SIZE1024, PROT_READ0x1 | PROT_WRITE0x2, MAP_SHARED0x01, fd1, 0); | |||
6749 | if (data1 == MAP_FAILED((void *) -1)) { | |||
6750 | // failed to map shared memory | |||
6751 | __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 ())), | |||
6752 | __kmp_msg_null); | |||
6753 | } | |||
6754 | if (shm_preexist == 0) { // set data to SHM, set value | |||
6755 | KMP_STRCPY_S(data1, SHM_SIZE, __kmp_registration_str)strcpy(data1, __kmp_registration_str); | |||
6756 | } | |||
6757 | // Read value from either what we just wrote or existing file. | |||
6758 | value = __kmp_str_format("%s", data1); // read value from SHM | |||
6759 | munmap(data1, SHM_SIZE1024); | |||
6760 | close(fd1); | |||
6761 | #else // Windows and unix with static library | |||
6762 | // Set environment variable, but do not overwrite if it is exist. | |||
6763 | __kmp_env_set(name, __kmp_registration_str, 0); | |||
6764 | // read value to see if it got set | |||
6765 | value = __kmp_env_get(name); | |||
6766 | #endif | |||
6767 | ||||
6768 | if (value != NULL__null && strcmp(value, __kmp_registration_str) == 0) { | |||
6769 | done = 1; // Ok, environment variable set successfully, exit the loop. | |||
6770 | } else { | |||
6771 | // Oops. Write failed. Another copy of OpenMP RTL is in memory. | |||
6772 | // Check whether it alive or dead. | |||
6773 | int neighbor = 0; // 0 -- unknown status, 1 -- alive, 2 -- dead. | |||
6774 | char *tail = value; | |||
6775 | char *flag_addr_str = NULL__null; | |||
6776 | char *flag_val_str = NULL__null; | |||
6777 | char const *file_name = NULL__null; | |||
6778 | __kmp_str_split(tail, '-', &flag_addr_str, &tail); | |||
6779 | __kmp_str_split(tail, '-', &flag_val_str, &tail); | |||
6780 | file_name = tail; | |||
6781 | if (tail != NULL__null) { | |||
6782 | unsigned long *flag_addr = 0; | |||
6783 | unsigned long flag_val = 0; | |||
6784 | KMP_SSCANFsscanf(flag_addr_str, "%p", RCAST(void **, &flag_addr)reinterpret_cast<void **>(&flag_addr)); | |||
6785 | KMP_SSCANFsscanf(flag_val_str, "%lx", &flag_val); | |||
6786 | if (flag_addr != 0 && flag_val != 0 && strcmp(file_name, "") != 0) { | |||
6787 | // First, check whether environment-encoded address is mapped into | |||
6788 | // addr space. | |||
6789 | // If so, dereference it to see if it still has the right value. | |||
6790 | if (__kmp_is_address_mapped(flag_addr) && *flag_addr == flag_val) { | |||
6791 | neighbor = 1; | |||
6792 | } else { | |||
6793 | // If not, then we know the other copy of the library is no longer | |||
6794 | // running. | |||
6795 | neighbor = 2; | |||
6796 | } | |||
6797 | } | |||
6798 | } | |||
6799 | switch (neighbor) { | |||
6800 | case 0: // Cannot parse environment variable -- neighbor status unknown. | |||
6801 | // Assume it is the incompatible format of future version of the | |||
6802 | // library. Assume the other library is alive. | |||
6803 | // WARN( ... ); // TODO: Issue a warning. | |||
6804 | file_name = "unknown library"; | |||
6805 | KMP_FALLTHROUGH()[[fallthrough]]; | |||
6806 | // Attention! Falling to the next case. That's intentional. | |||
6807 | case 1: { // Neighbor is alive. | |||
6808 | // Check it is allowed. | |||
6809 | char *duplicate_ok = __kmp_env_get("KMP_DUPLICATE_LIB_OK"); | |||
6810 | if (!__kmp_str_match_true(duplicate_ok)) { | |||
6811 | // That's not allowed. Issue fatal error. | |||
6812 | __kmp_fatal(KMP_MSG(DuplicateLibrary, KMP_LIBRARY_FILE, file_name)__kmp_msg_format(kmp_i18n_msg_DuplicateLibrary, "libomp.so.5" , file_name), | |||
6813 | KMP_HNT(DuplicateLibrary)__kmp_msg_format(kmp_i18n_hnt_DuplicateLibrary), __kmp_msg_null); | |||
6814 | } | |||
6815 | KMP_INTERNAL_FREE(duplicate_ok)free(duplicate_ok); | |||
6816 | __kmp_duplicate_library_ok = 1; | |||
6817 | done = 1; // Exit the loop. | |||
6818 | } break; | |||
6819 | case 2: { // Neighbor is dead. | |||
6820 | ||||
6821 | #if defined(KMP_USE_SHM) | |||
6822 | // close shared memory. | |||
6823 | shm_unlink(shm_name); // this removes file in /dev/shm | |||
6824 | #else | |||
6825 | // Clear the variable and try to register library again. | |||
6826 | __kmp_env_unset(name); | |||
6827 | #endif | |||
6828 | } break; | |||
6829 | default: { | |||
6830 | KMP_DEBUG_ASSERT(0)if (!(0)) { __kmp_debug_assert("0", "openmp/runtime/src/kmp_runtime.cpp" , 6830); }; | |||
6831 | } break; | |||
6832 | } | |||
6833 | } | |||
6834 | KMP_INTERNAL_FREE((void *)value)free((void *)value); | |||
6835 | #if defined(KMP_USE_SHM) | |||
6836 | KMP_INTERNAL_FREE((void *)shm_name)free((void *)shm_name); | |||
6837 | #endif | |||
6838 | } // while | |||
6839 | KMP_INTERNAL_FREE((void *)name)free((void *)name); | |||
6840 | ||||
6841 | } // func __kmp_register_library_startup | |||
6842 | ||||
6843 | void __kmp_unregister_library(void) { | |||
6844 | ||||
6845 | char *name = __kmp_reg_status_name(); | |||
6846 | char *value = NULL__null; | |||
6847 | ||||
6848 | #if defined(KMP_USE_SHM) | |||
6849 | char *shm_name = __kmp_str_format("/%s", name); | |||
6850 | int fd1 = shm_open(shm_name, O_RDONLY00, 0666); | |||
6851 | if (fd1 == -1) { | |||
6852 | // file did not open. return. | |||
6853 | return; | |||
6854 | } | |||
6855 | char *data1 = (char *)mmap(0, SHM_SIZE1024, PROT_READ0x1, MAP_SHARED0x01, fd1, 0); | |||
6856 | if (data1 != MAP_FAILED((void *) -1)) { | |||
6857 | value = __kmp_str_format("%s", data1); // read value from SHM | |||
6858 | munmap(data1, SHM_SIZE1024); | |||
6859 | } | |||
6860 | close(fd1); | |||
6861 | #else | |||
6862 | value = __kmp_env_get(name); | |||
6863 | #endif | |||
6864 | ||||
6865 | 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", 6865); }; | |||
6866 | 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" , 6866); }; | |||
6867 | if (value != NULL__null && strcmp(value, __kmp_registration_str) == 0) { | |||
6868 | // Ok, this is our variable. Delete it. | |||
6869 | #if defined(KMP_USE_SHM) | |||
6870 | shm_unlink(shm_name); // this removes file in /dev/shm | |||
6871 | #else | |||
6872 | __kmp_env_unset(name); | |||
6873 | #endif | |||
6874 | } | |||
6875 | ||||
6876 | #if defined(KMP_USE_SHM) | |||
6877 | KMP_INTERNAL_FREE(shm_name)free(shm_name); | |||
6878 | #endif | |||
6879 | ||||
6880 | KMP_INTERNAL_FREE(__kmp_registration_str)free(__kmp_registration_str); | |||
6881 | KMP_INTERNAL_FREE(value)free(value); | |||
6882 | KMP_INTERNAL_FREE(name)free(name); | |||
6883 | ||||
6884 | __kmp_registration_flag = 0; | |||
6885 | __kmp_registration_str = NULL__null; | |||
6886 | ||||
6887 | } // __kmp_unregister_library | |||
6888 | ||||
6889 | // End of Library registration stuff. | |||
6890 | // ----------------------------------------------------------------------------- | |||
6891 | ||||
6892 | #if KMP_MIC_SUPPORTED((0 || 1) && (1 || 0)) | |||
6893 | ||||
6894 | static void __kmp_check_mic_type() { | |||
6895 | kmp_cpuid_t cpuid_state = {0}; | |||
6896 | kmp_cpuid_t *cs_p = &cpuid_state; | |||
6897 | __kmp_x86_cpuid(1, 0, cs_p); | |||
6898 | // We don't support mic1 at the moment | |||
6899 | if ((cs_p->eax & 0xff0) == 0xB10) { | |||
6900 | __kmp_mic_type = mic2; | |||
6901 | } else if ((cs_p->eax & 0xf0ff0) == 0x50670) { | |||
6902 | __kmp_mic_type = mic3; | |||
6903 | } else { | |||
6904 | __kmp_mic_type = non_mic; | |||
6905 | } | |||
6906 | } | |||
6907 | ||||
6908 | #endif /* KMP_MIC_SUPPORTED */ | |||
6909 | ||||
6910 | #if KMP_HAVE_UMWAIT((0 || 1) && (1 || 0) && !0) | |||
6911 | static void __kmp_user_level_mwait_init() { | |||
6912 | struct kmp_cpuid buf; | |||
6913 | __kmp_x86_cpuid(7, 0, &buf); | |||
6914 | __kmp_waitpkg_enabled = ((buf.ecx >> 5) & 1); | |||
6915 | __kmp_umwait_enabled = __kmp_waitpkg_enabled && __kmp_user_level_mwait; | |||
6916 | __kmp_tpause_enabled = __kmp_waitpkg_enabled && (__kmp_tpause_state > 0); | |||
6917 | 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); } | |||
6918 | __kmp_umwait_enabled))if (kmp_f_debug >= 30) { __kmp_debug_printf ("__kmp_user_level_mwait_init: __kmp_umwait_enabled = %d\n" , __kmp_umwait_enabled); }; | |||
6919 | } | |||
6920 | #elif KMP_HAVE_MWAIT((0 || 1) && (1 || 0) && !0) | |||
6921 | #ifndef AT_INTELPHIUSERMWAIT | |||
6922 | // Spurious, non-existent value that should always fail to return anything. | |||
6923 | // Will be replaced with the correct value when we know that. | |||
6924 | #define AT_INTELPHIUSERMWAIT 10000 | |||
6925 | #endif | |||
6926 | // getauxval() function is available in RHEL7 and SLES12. If a system with an | |||
6927 | // earlier OS is used to build the RTL, we'll use the following internal | |||
6928 | // function when the entry is not found. | |||
6929 | unsigned long getauxval(unsigned long) KMP_WEAK_ATTRIBUTE_EXTERNAL; | |||
6930 | unsigned long getauxval(unsigned long) { return 0; } | |||
6931 | ||||
6932 | static void __kmp_user_level_mwait_init() { | |||
6933 | // When getauxval() and correct value of AT_INTELPHIUSERMWAIT are available | |||
6934 | // use them to find if the user-level mwait is enabled. Otherwise, forcibly | |||
6935 | // set __kmp_mwait_enabled=TRUE on Intel MIC if the environment variable | |||
6936 | // KMP_USER_LEVEL_MWAIT was set to TRUE. | |||
6937 | if (__kmp_mic_type == mic3) { | |||
6938 | unsigned long res = getauxval(AT_INTELPHIUSERMWAIT); | |||
6939 | if ((res & 0x1) || __kmp_user_level_mwait) { | |||
6940 | __kmp_mwait_enabled = TRUE(!0); | |||
6941 | if (__kmp_user_level_mwait) { | |||
6942 | KMP_INFORM(EnvMwaitWarn)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_EnvMwaitWarn ), __kmp_msg_null); | |||
6943 | } | |||
6944 | } else { | |||
6945 | __kmp_mwait_enabled = FALSE0; | |||
6946 | } | |||
6947 | } | |||
6948 | 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 ); } | |||
6949 | "__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 ); } | |||
6950 | __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 ); }; | |||
6951 | } | |||
6952 | #endif /* KMP_HAVE_UMWAIT */ | |||
6953 | ||||
6954 | static void __kmp_do_serial_initialize(void) { | |||
6955 | int i, gtid; | |||
6956 | size_t size; | |||
6957 | ||||
6958 | KA_TRACE(10, ("__kmp_do_serial_initialize: enter\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_do_serial_initialize: enter\n" ); }; | |||
6959 | ||||
6960 | 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", 6960); }; | |||
6961 | 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", 6961); }; | |||
6962 | 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", 6962); }; | |||
6963 | 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", 6963); }; | |||
6964 | 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" , 6964); }; | |||
6965 | ||||
6966 | #if OMPT_SUPPORT1 | |||
6967 | ompt_pre_init(); | |||
6968 | #endif | |||
6969 | #if OMPD_SUPPORT1 | |||
6970 | __kmp_env_dump(); | |||
6971 | ompd_init(); | |||
6972 | #endif | |||
6973 | ||||
6974 | __kmp_validate_locks(); | |||
6975 | ||||
6976 | /* Initialize internal memory allocator */ | |||
6977 | __kmp_init_allocator(); | |||
6978 | ||||
6979 | /* Register the library startup via an environment variable or via mapped | |||
6980 | shared memory file and check to see whether another copy of the library is | |||
6981 | already registered. Since forked child process is often terminated, we | |||
6982 | postpone the registration till middle initialization in the child */ | |||
6983 | if (__kmp_need_register_serial) | |||
6984 | __kmp_register_library_startup(); | |||
6985 | ||||
6986 | /* TODO reinitialization of library */ | |||
6987 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done)) { | |||
6988 | 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" ); }; | |||
6989 | } | |||
6990 | ||||
6991 | __kmp_global.g.g_abort = 0; | |||
6992 | TCW_SYNC_4(__kmp_global.g.g_done, FALSE)(__kmp_global.g.g_done) = (0); | |||
6993 | ||||
6994 | /* initialize the locks */ | |||
6995 | #if KMP_USE_ADAPTIVE_LOCKS(0 || 1) && !0 | |||
6996 | #if KMP_DEBUG_ADAPTIVE_LOCKS0 | |||
6997 | __kmp_init_speculative_stats(); | |||
6998 | #endif | |||
6999 | #endif | |||
7000 | #if KMP_STATS_ENABLED0 | |||
7001 | __kmp_stats_init(); | |||
7002 | #endif | |||
7003 | __kmp_init_lock(&__kmp_global_lock); | |||
7004 | __kmp_init_queuing_lock(&__kmp_dispatch_lock); | |||
7005 | __kmp_init_lock(&__kmp_debug_lock); | |||
7006 | __kmp_init_atomic_lock(&__kmp_atomic_lock); | |||
7007 | __kmp_init_atomic_lock(&__kmp_atomic_lock_1i); | |||
7008 | __kmp_init_atomic_lock(&__kmp_atomic_lock_2i); | |||
7009 | __kmp_init_atomic_lock(&__kmp_atomic_lock_4i); | |||
7010 | __kmp_init_atomic_lock(&__kmp_atomic_lock_4r); | |||
7011 | __kmp_init_atomic_lock(&__kmp_atomic_lock_8i); | |||
7012 | __kmp_init_atomic_lock(&__kmp_atomic_lock_8r); | |||
7013 | __kmp_init_atomic_lock(&__kmp_atomic_lock_8c); | |||
7014 | __kmp_init_atomic_lock(&__kmp_atomic_lock_10r); | |||
7015 | __kmp_init_atomic_lock(&__kmp_atomic_lock_16r); | |||
7016 | __kmp_init_atomic_lock(&__kmp_atomic_lock_16c); | |||
7017 | __kmp_init_atomic_lock(&__kmp_atomic_lock_20c); | |||
7018 | __kmp_init_atomic_lock(&__kmp_atomic_lock_32c); | |||
7019 | __kmp_init_bootstrap_lock(&__kmp_forkjoin_lock); | |||
7020 | __kmp_init_bootstrap_lock(&__kmp_exit_lock); | |||
7021 | #if KMP_USE_MONITOR | |||
7022 | __kmp_init_bootstrap_lock(&__kmp_monitor_lock); | |||
7023 | #endif | |||
7024 | __kmp_init_bootstrap_lock(&__kmp_tp_cached_lock); | |||
7025 | ||||
7026 | /* conduct initialization and initial setup of configuration */ | |||
7027 | ||||
7028 | __kmp_runtime_initialize(); | |||
7029 | ||||
7030 | #if KMP_MIC_SUPPORTED((0 || 1) && (1 || 0)) | |||
7031 | __kmp_check_mic_type(); | |||
7032 | #endif | |||
7033 | ||||
7034 | // Some global variable initialization moved here from kmp_env_initialize() | |||
7035 | #ifdef KMP_DEBUG1 | |||
7036 | kmp_diag = 0; | |||
7037 | #endif | |||
7038 | __kmp_abort_delay = 0; | |||
7039 | ||||
7040 | // From __kmp_init_dflt_team_nth() | |||
7041 | /* assume the entire machine will be used */ | |||
7042 | __kmp_dflt_team_nth_ub = __kmp_xproc; | |||
7043 | if (__kmp_dflt_team_nth_ub < KMP_MIN_NTH1) { | |||
7044 | __kmp_dflt_team_nth_ub = KMP_MIN_NTH1; | |||
7045 | } | |||
7046 | if (__kmp_dflt_team_nth_ub > __kmp_sys_max_nth) { | |||
7047 | __kmp_dflt_team_nth_ub = __kmp_sys_max_nth; | |||
7048 | } | |||
7049 | __kmp_max_nth = __kmp_sys_max_nth; | |||
7050 | __kmp_cg_max_nth = __kmp_sys_max_nth; | |||
7051 | __kmp_teams_max_nth = __kmp_xproc; // set a "reasonable" default | |||
7052 | if (__kmp_teams_max_nth > __kmp_sys_max_nth) { | |||
7053 | __kmp_teams_max_nth = __kmp_sys_max_nth; | |||
7054 | } | |||
7055 | ||||
7056 | // Three vars below moved here from __kmp_env_initialize() "KMP_BLOCKTIME" | |||
7057 | // part | |||
7058 | __kmp_dflt_blocktime = KMP_DEFAULT_BLOCKTIME(__kmp_is_hybrid_cpu() ? (0) : (200)); | |||
7059 | #if KMP_USE_MONITOR | |||
7060 | __kmp_monitor_wakeups = | |||
7061 | KMP_WAKEUPS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups); | |||
7062 | __kmp_bt_intervals = | |||
7063 | KMP_INTERVALS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups); | |||
7064 | #endif | |||
7065 | // From "KMP_LIBRARY" part of __kmp_env_initialize() | |||
7066 | __kmp_library = library_throughput; | |||
7067 | // From KMP_SCHEDULE initialization | |||
7068 | __kmp_static = kmp_sch_static_balanced; | |||
7069 | // AC: do not use analytical here, because it is non-monotonous | |||
7070 | //__kmp_guided = kmp_sch_guided_iterative_chunked; | |||
7071 | //__kmp_auto = kmp_sch_guided_analytical_chunked; // AC: it is the default, no | |||
7072 | // need to repeat assignment | |||
7073 | // Barrier initialization. Moved here from __kmp_env_initialize() Barrier branch | |||
7074 | // bit control and barrier method control parts | |||
7075 | #if KMP_FAST_REDUCTION_BARRIER1 | |||
7076 | #define kmp_reduction_barrier_gather_bb ((int)1) | |||
7077 | #define kmp_reduction_barrier_release_bb ((int)1) | |||
7078 | #define kmp_reduction_barrier_gather_pat __kmp_barrier_gather_pat_dflt | |||
7079 | #define kmp_reduction_barrier_release_pat __kmp_barrier_release_pat_dflt | |||
7080 | #endif // KMP_FAST_REDUCTION_BARRIER | |||
7081 | for (i = bs_plain_barrier; i < bs_last_barrier; i++) { | |||
7082 | __kmp_barrier_gather_branch_bits[i] = __kmp_barrier_gather_bb_dflt; | |||
7083 | __kmp_barrier_release_branch_bits[i] = __kmp_barrier_release_bb_dflt; | |||
7084 | __kmp_barrier_gather_pattern[i] = __kmp_barrier_gather_pat_dflt; | |||
7085 | __kmp_barrier_release_pattern[i] = __kmp_barrier_release_pat_dflt; | |||
7086 | #if KMP_FAST_REDUCTION_BARRIER1 | |||
7087 | if (i == bs_reduction_barrier) { // tested and confirmed on ALTIX only ( | |||
7088 | // lin_64 ): hyper,1 | |||
7089 | __kmp_barrier_gather_branch_bits[i] = kmp_reduction_barrier_gather_bb; | |||
7090 | __kmp_barrier_release_branch_bits[i] = kmp_reduction_barrier_release_bb; | |||
7091 | __kmp_barrier_gather_pattern[i] = kmp_reduction_barrier_gather_pat; | |||
7092 | __kmp_barrier_release_pattern[i] = kmp_reduction_barrier_release_pat; | |||
7093 | } | |||
7094 | #endif // KMP_FAST_REDUCTION_BARRIER | |||
7095 | } | |||
7096 | #if KMP_FAST_REDUCTION_BARRIER1 | |||
7097 | #undef kmp_reduction_barrier_release_pat | |||
7098 | #undef kmp_reduction_barrier_gather_pat | |||
7099 | #undef kmp_reduction_barrier_release_bb | |||
7100 | #undef kmp_reduction_barrier_gather_bb | |||
7101 | #endif // KMP_FAST_REDUCTION_BARRIER | |||
7102 | #if KMP_MIC_SUPPORTED((0 || 1) && (1 || 0)) | |||
7103 | if (__kmp_mic_type == mic2) { // KNC | |||
7104 | // AC: plane=3,2, forkjoin=2,1 are optimal for 240 threads on KNC | |||
7105 | __kmp_barrier_gather_branch_bits[bs_plain_barrier] = 3; // plain gather | |||
7106 | __kmp_barrier_release_branch_bits[bs_forkjoin_barrier] = | |||
7107 | 1; // forkjoin release | |||
7108 | __kmp_barrier_gather_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar; | |||
7109 | __kmp_barrier_release_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar; | |||
7110 | } | |||
7111 | #if KMP_FAST_REDUCTION_BARRIER1 | |||
7112 | if (__kmp_mic_type == mic2) { // KNC | |||
7113 | __kmp_barrier_gather_pattern[bs_reduction_barrier] = bp_hierarchical_bar; | |||
7114 | __kmp_barrier_release_pattern[bs_reduction_barrier] = bp_hierarchical_bar; | |||
7115 | } | |||
7116 | #endif // KMP_FAST_REDUCTION_BARRIER | |||
7117 | #endif // KMP_MIC_SUPPORTED | |||
7118 | ||||
7119 | // From KMP_CHECKS initialization | |||
7120 | #ifdef KMP_DEBUG1 | |||
7121 | __kmp_env_checks = TRUE(!0); /* development versions have the extra checks */ | |||
7122 | #else | |||
7123 | __kmp_env_checks = FALSE0; /* port versions do not have the extra checks */ | |||
7124 | #endif | |||
7125 | ||||
7126 | // From "KMP_FOREIGN_THREADS_THREADPRIVATE" initialization | |||
7127 | __kmp_foreign_tp = TRUE(!0); | |||
7128 | ||||
7129 | __kmp_global.g.g_dynamic = FALSE0; | |||
7130 | __kmp_global.g.g_dynamic_mode = dynamic_default; | |||
7131 | ||||
7132 | __kmp_init_nesting_mode(); | |||
7133 | ||||
7134 | __kmp_env_initialize(NULL__null); | |||
7135 | ||||
7136 | #if KMP_HAVE_MWAIT((0 || 1) && (1 || 0) && !0) || KMP_HAVE_UMWAIT((0 || 1) && (1 || 0) && !0) | |||
7137 | __kmp_user_level_mwait_init(); | |||
7138 | #endif | |||
7139 | // Print all messages in message catalog for testing purposes. | |||
7140 | #ifdef KMP_DEBUG1 | |||
7141 | char const *val = __kmp_env_get("KMP_DUMP_CATALOG"); | |||
7142 | if (__kmp_str_match_true(val)) { | |||
7143 | kmp_str_buf_t buffer; | |||
7144 | __kmp_str_buf_init(&buffer){ (&buffer)->str = (&buffer)->bulk; (&buffer )->size = sizeof((&buffer)->bulk); (&buffer)-> used = 0; (&buffer)->bulk[0] = 0; }; | |||
7145 | __kmp_i18n_dump_catalog(&buffer); | |||
7146 | __kmp_printf("%s", buffer.str); | |||
7147 | __kmp_str_buf_free(&buffer); | |||
7148 | } | |||
7149 | __kmp_env_free(&val); | |||
7150 | #endif | |||
7151 | ||||
7152 | __kmp_threads_capacity = | |||
7153 | __kmp_initial_threads_capacity(__kmp_dflt_team_nth_ub); | |||
7154 | // Moved here from __kmp_env_initialize() "KMP_ALL_THREADPRIVATE" part | |||
7155 | __kmp_tp_capacity = __kmp_default_tp_capacity( | |||
7156 | __kmp_dflt_team_nth_ub, __kmp_max_nth, __kmp_allThreadsSpecified); | |||
7157 | ||||
7158 | // If the library is shut down properly, both pools must be NULL. Just in | |||
7159 | // case, set them to NULL -- some memory may leak, but subsequent code will | |||
7160 | // work even if pools are not freed. | |||
7161 | 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", 7161); }; | |||
7162 | 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" , 7162); }; | |||
7163 | 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", 7163); }; | |||
7164 | __kmp_thread_pool = NULL__null; | |||
7165 | __kmp_thread_pool_insert_pt = NULL__null; | |||
7166 | __kmp_team_pool = NULL__null; | |||
7167 | ||||
7168 | /* Allocate all of the variable sized records */ | |||
7169 | /* NOTE: __kmp_threads_capacity entries are allocated, but the arrays are | |||
7170 | * expandable */ | |||
7171 | /* Since allocation is cache-aligned, just add extra padding at the end */ | |||
7172 | size = | |||
7173 | (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * __kmp_threads_capacity + | |||
7174 | CACHE_LINE64; | |||
7175 | __kmp_threads = (kmp_info_t **)__kmp_allocate(size)___kmp_allocate((size), "openmp/runtime/src/kmp_runtime.cpp", 7175); | |||
7176 | __kmp_root = (kmp_root_t **)((char *)__kmp_threads + | |||
7177 | sizeof(kmp_info_t *) * __kmp_threads_capacity); | |||
7178 | ||||
7179 | /* init thread counts */ | |||
7180 | KMP_DEBUG_ASSERT(__kmp_all_nth ==if (!(__kmp_all_nth == 0)) { __kmp_debug_assert("__kmp_all_nth == 0" , "openmp/runtime/src/kmp_runtime.cpp", 7181); } | |||
7181 | 0)if (!(__kmp_all_nth == 0)) { __kmp_debug_assert("__kmp_all_nth == 0" , "openmp/runtime/src/kmp_runtime.cpp", 7181); }; // Asserts fail if the library is reinitializing and | |||
7182 | KMP_DEBUG_ASSERT(__kmp_nth == 0)if (!(__kmp_nth == 0)) { __kmp_debug_assert("__kmp_nth == 0", "openmp/runtime/src/kmp_runtime.cpp", 7182); }; // something was wrong in termination. | |||
7183 | __kmp_all_nth = 0; | |||
7184 | __kmp_nth = 0; | |||
7185 | ||||
7186 | /* setup the uber master thread and hierarchy */ | |||
7187 | gtid = __kmp_register_root(TRUE(!0)); | |||
7188 | 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); }; | |||
7189 | KMP_ASSERT(KMP_UBER_GTID(gtid))if (!(KMP_UBER_GTID(gtid))) { __kmp_debug_assert("KMP_UBER_GTID(gtid)" , "openmp/runtime/src/kmp_runtime.cpp", 7189); }; | |||
7190 | KMP_ASSERT(KMP_INITIAL_GTID(gtid))if (!((0 == (gtid)))) { __kmp_debug_assert("KMP_INITIAL_GTID(gtid)" , "openmp/runtime/src/kmp_runtime.cpp", 7190); }; | |||
7191 | ||||
7192 | KMP_MB(); /* Flush all pending memory write invalidates. */ | |||
7193 | ||||
7194 | __kmp_common_initialize(); | |||
7195 | ||||
7196 | #if KMP_OS_UNIX1 | |||
7197 | /* invoke the child fork handler */ | |||
7198 | __kmp_register_atfork(); | |||
7199 | #endif | |||
7200 | ||||
7201 | #if !KMP_DYNAMIC_LIB1 | |||
7202 | { | |||
7203 | /* Invoke the exit handler when the program finishes, only for static | |||
7204 | library. For dynamic library, we already have _fini and DllMain. */ | |||
7205 | int rc = atexit(__kmp_internal_end_atexit); | |||
7206 | if (rc != 0) { | |||
7207 | __kmp_fatal(KMP_MSG(FunctionError, "atexit()")__kmp_msg_format(kmp_i18n_msg_FunctionError, "atexit()"), KMP_ERR(rc)__kmp_msg_error_code(rc), | |||
7208 | __kmp_msg_null); | |||
7209 | } | |||
7210 | } | |||
7211 | #endif | |||
7212 | ||||
7213 | #if KMP_HANDLE_SIGNALS(1 || 0) | |||
7214 | #if KMP_OS_UNIX1 | |||
7215 | /* NOTE: make sure that this is called before the user installs their own | |||
7216 | signal handlers so that the user handlers are called first. this way they | |||
7217 | can return false, not call our handler, avoid terminating the library, and | |||
7218 | continue execution where they left off. */ | |||
7219 | __kmp_install_signals(FALSE0); | |||
7220 | #endif /* KMP_OS_UNIX */ | |||
7221 | #if KMP_OS_WINDOWS0 | |||
7222 | __kmp_install_signals(TRUE(!0)); | |||
7223 | #endif /* KMP_OS_WINDOWS */ | |||
7224 | #endif | |||
7225 | ||||
7226 | /* we have finished the serial initialization */ | |||
7227 | __kmp_init_counter++; | |||
7228 | ||||
7229 | __kmp_init_serial = TRUE(!0); | |||
7230 | ||||
7231 | if (__kmp_settings) { | |||
7232 | __kmp_env_print(); | |||
7233 | } | |||
7234 | ||||
7235 | if (__kmp_display_env || __kmp_display_env_verbose) { | |||
7236 | __kmp_env_print_2(); | |||
7237 | } | |||
7238 | ||||
7239 | #if OMPT_SUPPORT1 | |||
7240 | ompt_post_init(); | |||
7241 | #endif | |||
7242 | ||||
7243 | KMP_MB(); | |||
7244 | ||||
7245 | KA_TRACE(10, ("__kmp_do_serial_initialize: exit\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_do_serial_initialize: exit\n" ); }; | |||
7246 | } | |||
7247 | ||||
7248 | void __kmp_serial_initialize(void) { | |||
7249 | if (__kmp_init_serial) { | |||
7250 | return; | |||
7251 | } | |||
7252 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); | |||
7253 | if (__kmp_init_serial) { | |||
7254 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | |||
7255 | return; | |||
7256 | } | |||
7257 | __kmp_do_serial_initialize(); | |||
7258 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | |||
7259 | } | |||
7260 | ||||
7261 | static void __kmp_do_middle_initialize(void) { | |||
7262 | int i, j; | |||
7263 | int prev_dflt_team_nth; | |||
7264 | ||||
7265 | if (!__kmp_init_serial) { | |||
7266 | __kmp_do_serial_initialize(); | |||
7267 | } | |||
7268 | ||||
7269 | KA_TRACE(10, ("__kmp_middle_initialize: enter\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_middle_initialize: enter\n" ); }; | |||
7270 | ||||
7271 | if (UNLIKELY(!__kmp_need_register_serial)__builtin_expect(!!(!__kmp_need_register_serial), 0)) { | |||
7272 | // We are in a forked child process. The registration was skipped during | |||
7273 | // serial initialization in __kmp_atfork_child handler. Do it here. | |||
7274 | __kmp_register_library_startup(); | |||
7275 | } | |||
7276 | ||||
7277 | // Save the previous value for the __kmp_dflt_team_nth so that | |||
7278 | // we can avoid some reinitialization if it hasn't changed. | |||
7279 | prev_dflt_team_nth = __kmp_dflt_team_nth; | |||
7280 | ||||
7281 | #if KMP_AFFINITY_SUPPORTED1 | |||
7282 | // __kmp_affinity_initialize() will try to set __kmp_ncores to the | |||
7283 | // number of cores on the machine. | |||
7284 | __kmp_affinity_initialize(__kmp_affinity); | |||
7285 | ||||
7286 | #endif /* KMP_AFFINITY_SUPPORTED */ | |||
7287 | ||||
7288 | KMP_ASSERT(__kmp_xproc > 0)if (!(__kmp_xproc > 0)) { __kmp_debug_assert("__kmp_xproc > 0" , "openmp/runtime/src/kmp_runtime.cpp", 7288); }; | |||
7289 | if (__kmp_avail_proc == 0) { | |||
7290 | __kmp_avail_proc = __kmp_xproc; | |||
7291 | } | |||
7292 | ||||
7293 | // If there were empty places in num_threads list (OMP_NUM_THREADS=,,2,3), | |||
7294 | // correct them now | |||
7295 | j = 0; | |||
7296 | while ((j < __kmp_nested_nth.used) && !__kmp_nested_nth.nth[j]) { | |||
7297 | __kmp_nested_nth.nth[j] = __kmp_dflt_team_nth = __kmp_dflt_team_nth_ub = | |||
7298 | __kmp_avail_proc; | |||
7299 | j++; | |||
7300 | } | |||
7301 | ||||
7302 | if (__kmp_dflt_team_nth == 0) { | |||
7303 | #ifdef KMP_DFLT_NTH_CORES | |||
7304 | // Default #threads = #cores | |||
7305 | __kmp_dflt_team_nth = __kmp_ncores; | |||
7306 | 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); } | |||
7307 | "__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); } | |||
7308 | __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); }; | |||
7309 | #else | |||
7310 | // Default #threads = #available OS procs | |||
7311 | __kmp_dflt_team_nth = __kmp_avail_proc; | |||
7312 | 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); } | |||
7313 | "__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); } | |||
7314 | __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); }; | |||
7315 | #endif /* KMP_DFLT_NTH_CORES */ | |||
7316 | } | |||
7317 | ||||
7318 | if (__kmp_dflt_team_nth < KMP_MIN_NTH1) { | |||
7319 | __kmp_dflt_team_nth = KMP_MIN_NTH1; | |||
7320 | } | |||
7321 | if (__kmp_dflt_team_nth > __kmp_sys_max_nth) { | |||
7322 | __kmp_dflt_team_nth = __kmp_sys_max_nth; | |||
7323 | } | |||
7324 | ||||
7325 | if (__kmp_nesting_mode > 0) | |||
7326 | __kmp_set_nesting_mode_threads(); | |||
7327 | ||||
7328 | // There's no harm in continuing if the following check fails, | |||
7329 | // but it indicates an error in the previous logic. | |||
7330 | 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" , 7330); }; | |||
7331 | ||||
7332 | if (__kmp_dflt_team_nth != prev_dflt_team_nth) { | |||
7333 | // Run through the __kmp_threads array and set the num threads icv for each | |||
7334 | // root thread that is currently registered with the RTL (which has not | |||
7335 | // already explicitly set its nthreads-var with a call to | |||
7336 | // omp_set_num_threads()). | |||
7337 | for (i = 0; i < __kmp_threads_capacity; i++) { | |||
7338 | kmp_info_t *thread = __kmp_threads[i]; | |||
7339 | if (thread == NULL__null) | |||
7340 | continue; | |||
7341 | if (thread->th.th_current_task->td_icvs.nproc != 0) | |||
7342 | continue; | |||
7343 | ||||
7344 | set__nproc(__kmp_threads[i], __kmp_dflt_team_nth)(((__kmp_threads[i])->th.th_current_task->td_icvs.nproc ) = (__kmp_dflt_team_nth)); | |||
7345 | } | |||
7346 | } | |||
7347 | 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); } | |||
7348 | 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); } | |||
7349 | ("__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); } | |||
7350 | __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); }; | |||
7351 | ||||
7352 | #ifdef KMP_ADJUST_BLOCKTIME1 | |||
7353 | /* Adjust blocktime to zero if necessary now that __kmp_avail_proc is set */ | |||
7354 | if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { | |||
7355 | 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", 7355); }; | |||
7356 | if (__kmp_nth > __kmp_avail_proc) { | |||
7357 | __kmp_zero_bt = TRUE(!0); | |||
7358 | } | |||
7359 | } | |||
7360 | #endif /* KMP_ADJUST_BLOCKTIME */ | |||
7361 | ||||
7362 | /* we have finished middle initialization */ | |||
7363 | TCW_SYNC_4(__kmp_init_middle, TRUE)(__kmp_init_middle) = ((!0)); | |||
7364 | ||||
7365 | KA_TRACE(10, ("__kmp_do_middle_initialize: exit\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_do_middle_initialize: exit\n" ); }; | |||
7366 | } | |||
7367 | ||||
7368 | void __kmp_middle_initialize(void) { | |||
7369 | if (__kmp_init_middle) { | |||
7370 | return; | |||
7371 | } | |||
7372 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); | |||
7373 | if (__kmp_init_middle) { | |||
7374 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | |||
7375 | return; | |||
7376 | } | |||
7377 | __kmp_do_middle_initialize(); | |||
7378 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | |||
7379 | } | |||
7380 | ||||
7381 | void __kmp_parallel_initialize(void) { | |||
7382 | int gtid = __kmp_entry_gtid()__kmp_get_global_thread_id_reg(); // this might be a new root | |||
7383 | ||||
7384 | /* synchronize parallel initialization (for sibling) */ | |||
7385 | if (TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) | |||
7386 | return; | |||
7387 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); | |||
7388 | if (TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) { | |||
7389 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | |||
7390 | return; | |||
7391 | } | |||
7392 | ||||
7393 | /* TODO reinitialization after we have already shut down */ | |||
7394 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done)) { | |||
7395 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_parallel_initialize: attempt to init while shutting down\n" ); } | |||
7396 | 10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_parallel_initialize: attempt to init while shutting down\n" ); } | |||
7397 | ("__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" ); }; | |||
7398 | __kmp_infinite_loop(); | |||
7399 | } | |||
7400 | ||||
7401 | /* jc: The lock __kmp_initz_lock is already held, so calling | |||
7402 | __kmp_serial_initialize would cause a deadlock. So we call | |||
7403 | __kmp_do_serial_initialize directly. */ | |||
7404 | if (!__kmp_init_middle) { | |||
7405 | __kmp_do_middle_initialize(); | |||
7406 | } | |||
7407 | __kmp_assign_root_init_mask(); | |||
7408 | __kmp_resume_if_hard_paused(); | |||
7409 | ||||
7410 | /* begin initialization */ | |||
7411 | KA_TRACE(10, ("__kmp_parallel_initialize: enter\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_parallel_initialize: enter\n" ); }; | |||
7412 | KMP_ASSERT(KMP_UBER_GTID(gtid))if (!(KMP_UBER_GTID(gtid))) { __kmp_debug_assert("KMP_UBER_GTID(gtid)" , "openmp/runtime/src/kmp_runtime.cpp", 7412); }; | |||
7413 | ||||
7414 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 | |||
7415 | // Save the FP control regs. | |||
7416 | // Worker threads will set theirs to these values at thread startup. | |||
7417 | __kmp_store_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word); | |||
7418 | __kmp_store_mxcsr(&__kmp_init_mxcsr); | |||
7419 | __kmp_init_mxcsr &= KMP_X86_MXCSR_MASK0xffffffc0; | |||
7420 | #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ | |||
7421 | ||||
7422 | #if KMP_OS_UNIX1 | |||
7423 | #if KMP_HANDLE_SIGNALS(1 || 0) | |||
7424 | /* must be after __kmp_serial_initialize */ | |||
7425 | __kmp_install_signals(TRUE(!0)); | |||
7426 | #endif | |||
7427 | #endif | |||
7428 | ||||
7429 | __kmp_suspend_initialize(); | |||
7430 | ||||
7431 | #if defined(USE_LOAD_BALANCE1) | |||
7432 | if (__kmp_global.g.g_dynamic_mode == dynamic_default) { | |||
7433 | __kmp_global.g.g_dynamic_mode = dynamic_load_balance; | |||
7434 | } | |||
7435 | #else | |||
7436 | if (__kmp_global.g.g_dynamic_mode == dynamic_default) { | |||
7437 | __kmp_global.g.g_dynamic_mode = dynamic_thread_limit; | |||
7438 | } | |||
7439 | #endif | |||
7440 | ||||
7441 | if (__kmp_version) { | |||
7442 | __kmp_print_version_2(); | |||
7443 | } | |||
7444 | ||||
7445 | /* we have finished parallel initialization */ | |||
7446 | TCW_SYNC_4(__kmp_init_parallel, TRUE)(__kmp_init_parallel) = ((!0)); | |||
7447 | ||||
7448 | KMP_MB(); | |||
7449 | KA_TRACE(10, ("__kmp_parallel_initialize: exit\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_parallel_initialize: exit\n" ); }; | |||
7450 | ||||
7451 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | |||
7452 | } | |||
7453 | ||||
7454 | void __kmp_hidden_helper_initialize() { | |||
7455 | if (TCR_4(__kmp_init_hidden_helper)(__kmp_init_hidden_helper)) | |||
7456 | return; | |||
7457 | ||||
7458 | // __kmp_parallel_initialize is required before we initialize hidden helper | |||
7459 | if (!TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) | |||
7460 | __kmp_parallel_initialize(); | |||
7461 | ||||
7462 | // Double check. Note that this double check should not be placed before | |||
7463 | // __kmp_parallel_initialize as it will cause dead lock. | |||
7464 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); | |||
7465 | if (TCR_4(__kmp_init_hidden_helper)(__kmp_init_hidden_helper)) { | |||
7466 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | |||
7467 | return; | |||
7468 | } | |||
7469 | ||||
7470 | #if KMP_AFFINITY_SUPPORTED1 | |||
7471 | // Initialize hidden helper affinity settings. | |||
7472 | // The above __kmp_parallel_initialize() will initialize | |||
7473 | // regular affinity (and topology) if not already done. | |||
7474 | if (!__kmp_hh_affinity.flags.initialized) | |||
7475 | __kmp_affinity_initialize(__kmp_hh_affinity); | |||
7476 | #endif | |||
7477 | ||||
7478 | // Set the count of hidden helper tasks to be executed to zero | |||
7479 | KMP_ATOMIC_ST_REL(&__kmp_unexecuted_hidden_helper_tasks, 0)(&__kmp_unexecuted_hidden_helper_tasks)->store(0, std:: memory_order_release); | |||
7480 | ||||
7481 | // Set the global variable indicating that we're initializing hidden helper | |||
7482 | // team/threads | |||
7483 | TCW_SYNC_4(__kmp_init_hidden_helper_threads, TRUE)(__kmp_init_hidden_helper_threads) = ((!0)); | |||
7484 | ||||
7485 | // Platform independent initialization | |||
7486 | __kmp_do_initialize_hidden_helper_threads(); | |||
7487 | ||||
7488 | // Wait here for the finish of initialization of hidden helper teams | |||
7489 | __kmp_hidden_helper_threads_initz_wait(); | |||
7490 | ||||
7491 | // We have finished hidden helper initialization | |||
7492 | TCW_SYNC_4(__kmp_init_hidden_helper, TRUE)(__kmp_init_hidden_helper) = ((!0)); | |||
7493 | ||||
7494 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | |||
7495 | } | |||
7496 | ||||
7497 | /* ------------------------------------------------------------------------ */ | |||
7498 | ||||
7499 | void __kmp_run_before_invoked_task(int gtid, int tid, kmp_info_t *this_thr, | |||
7500 | kmp_team_t *team) { | |||
7501 | kmp_disp_t *dispatch; | |||
7502 | ||||
7503 | KMP_MB(); | |||
7504 | ||||
7505 | /* none of the threads have encountered any constructs, yet. */ | |||
7506 | this_thr->th.th_local.this_construct = 0; | |||
7507 | #if KMP_CACHE_MANAGE | |||
7508 | KMP_CACHE_PREFETCH(&this_thr->th.th_bar[bs_forkjoin_barrier].bb.b_arrived); | |||
7509 | #endif /* KMP_CACHE_MANAGE */ | |||
7510 | dispatch = (kmp_disp_t *)TCR_PTR(this_thr->th.th_dispatch)((void *)(this_thr->th.th_dispatch)); | |||
7511 | KMP_DEBUG_ASSERT(dispatch)if (!(dispatch)) { __kmp_debug_assert("dispatch", "openmp/runtime/src/kmp_runtime.cpp" , 7511); }; | |||
7512 | 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", 7512); }; | |||
7513 | // KMP_DEBUG_ASSERT( this_thr->th.th_dispatch == &team->t.t_dispatch[ | |||
7514 | // this_thr->th.th_info.ds.ds_tid ] ); | |||
7515 | ||||
7516 | dispatch->th_disp_index = 0; /* reset the dispatch buffer counter */ | |||
7517 | dispatch->th_doacross_buf_idx = 0; // reset doacross dispatch buffer counter | |||
7518 | if (__kmp_env_consistency_check) | |||
7519 | __kmp_push_parallel(gtid, team->t.t_ident); | |||
7520 | ||||
7521 | KMP_MB(); /* Flush all pending memory write invalidates. */ | |||
7522 | } | |||
7523 | ||||
7524 | void __kmp_run_after_invoked_task(int gtid, int tid, kmp_info_t *this_thr, | |||
7525 | kmp_team_t *team) { | |||
7526 | if (__kmp_env_consistency_check) | |||
7527 | __kmp_pop_parallel(gtid, team->t.t_ident); | |||
7528 | ||||
7529 | __kmp_finish_implicit_task(this_thr); | |||
7530 | } | |||
7531 | ||||
7532 | int __kmp_invoke_task_func(int gtid) { | |||
7533 | int rc; | |||
7534 | int tid = __kmp_tid_from_gtid(gtid); | |||
7535 | kmp_info_t *this_thr = __kmp_threads[gtid]; | |||
7536 | kmp_team_t *team = this_thr->th.th_team; | |||
7537 | ||||
7538 | __kmp_run_before_invoked_task(gtid, tid, this_thr, team); | |||
7539 | #if USE_ITT_BUILD1 | |||
7540 | if (__itt_stack_caller_create_ptr__kmp_itt_stack_caller_create_ptr__3_0) { | |||
7541 | // inform ittnotify about entering user's code | |||
7542 | if (team->t.t_stack_id != NULL__null) { | |||
7543 | __kmp_itt_stack_callee_enter((__itt_caller)team->t.t_stack_id); | |||
7544 | } else { | |||
7545 | 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" , 7545); }; | |||
7546 | __kmp_itt_stack_callee_enter( | |||
7547 | (__itt_caller)team->t.t_parent->t.t_stack_id); | |||
7548 | } | |||
7549 | } | |||
7550 | #endif /* USE_ITT_BUILD */ | |||
7551 | #if INCLUDE_SSC_MARKS(1 && 1) | |||
7552 | SSC_MARK_INVOKING()__asm__ __volatile__("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(0xd695) : "%ebx"); | |||
7553 | #endif | |||
7554 | ||||
7555 | #if OMPT_SUPPORT1 | |||
7556 | void *dummy; | |||
7557 | void **exit_frame_p; | |||
7558 | ompt_data_t *my_task_data; | |||
7559 | ompt_data_t *my_parallel_data; | |||
7560 | int ompt_team_size; | |||
7561 | ||||
7562 | if (ompt_enabled.enabled) { | |||
7563 | exit_frame_p = &(team->t.t_implicit_task_taskdata[tid] | |||
7564 | .ompt_task_info.frame.exit_frame.ptr); | |||
7565 | } else { | |||
7566 | exit_frame_p = &dummy; | |||
7567 | } | |||
7568 | ||||
7569 | my_task_data = | |||
7570 | &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data); | |||
7571 | my_parallel_data = &(team->t.ompt_team_info.parallel_data); | |||
7572 | if (ompt_enabled.ompt_callback_implicit_task) { | |||
7573 | ompt_team_size = team->t.t_nproc; | |||
7574 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | |||
7575 | ompt_scope_begin, my_parallel_data, my_task_data, ompt_team_size, | |||
7576 | __kmp_tid_from_gtid(gtid), ompt_task_implicit); | |||
7577 | OMPT_CUR_TASK_INFO(this_thr)(&(this_thr->th.th_current_task->ompt_task_info))->thread_num = __kmp_tid_from_gtid(gtid); | |||
7578 | } | |||
7579 | #endif | |||
7580 | ||||
7581 | #if KMP_STATS_ENABLED0 | |||
7582 | stats_state_e previous_state = KMP_GET_THREAD_STATE()((void)0); | |||
7583 | if (previous_state == stats_state_e::TEAMS_REGION) { | |||
7584 | KMP_PUSH_PARTITIONED_TIMER(OMP_teams)((void)0); | |||
7585 | } else { | |||
7586 | KMP_PUSH_PARTITIONED_TIMER(OMP_parallel)((void)0); | |||
7587 | } | |||
7588 | KMP_SET_THREAD_STATE(IMPLICIT_TASK)((void)0); | |||
7589 | #endif | |||
7590 | ||||
7591 | rc = __kmp_invoke_microtask((microtask_t)TCR_SYNC_PTR(team->t.t_pkfn)((void *)(team->t.t_pkfn)), gtid, | |||
7592 | tid, (int)team->t.t_argc, (void **)team->t.t_argv | |||
7593 | #if OMPT_SUPPORT1 | |||
7594 | , | |||
7595 | exit_frame_p | |||
7596 | #endif | |||
7597 | ); | |||
7598 | #if OMPT_SUPPORT1 | |||
7599 | *exit_frame_p = NULL__null; | |||
7600 | this_thr->th.ompt_thread_info.parallel_flags |= ompt_parallel_team; | |||
7601 | #endif | |||
7602 | ||||
7603 | #if KMP_STATS_ENABLED0 | |||
7604 | if (previous_state == stats_state_e::TEAMS_REGION) { | |||
7605 | KMP_SET_THREAD_STATE(previous_state)((void)0); | |||
7606 | } | |||
7607 | KMP_POP_PARTITIONED_TIMER()((void)0); | |||
7608 | #endif | |||
7609 | ||||
7610 | #if USE_ITT_BUILD1 | |||
7611 | if (__itt_stack_caller_create_ptr__kmp_itt_stack_caller_create_ptr__3_0) { | |||
7612 | // inform ittnotify about leaving user's code | |||
7613 | if (team->t.t_stack_id != NULL__null) { | |||
7614 | __kmp_itt_stack_callee_leave((__itt_caller)team->t.t_stack_id); | |||
7615 | } else { | |||
7616 | 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" , 7616); }; | |||
7617 | __kmp_itt_stack_callee_leave( | |||
7618 | (__itt_caller)team->t.t_parent->t.t_stack_id); | |||
7619 | } | |||
7620 | } | |||
7621 | #endif /* USE_ITT_BUILD */ | |||
7622 | __kmp_run_after_invoked_task(gtid, tid, this_thr, team); | |||
7623 | ||||
7624 | return rc; | |||
7625 | } | |||
7626 | ||||
7627 | void __kmp_teams_master(int gtid) { | |||
7628 | // This routine is called by all primary threads in teams construct | |||
7629 | kmp_info_t *thr = __kmp_threads[gtid]; | |||
7630 | kmp_team_t *team = thr->th.th_team; | |||
7631 | ident_t *loc = team->t.t_ident; | |||
7632 | thr->th.th_set_nproc = thr->th.th_teams_size.nth; | |||
7633 | 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", 7633); }; | |||
7634 | 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", 7634); }; | |||
7635 | 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 ); } | |||
7636 | __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 ); }; | |||
7637 | ||||
7638 | // This thread is a new CG root. Set up the proper variables. | |||
7639 | 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" , 7639); | |||
7640 | tmp->cg_root = thr; // Make thr the CG root | |||
7641 | // Init to thread limit stored when league primary threads were forked | |||
7642 | tmp->cg_thread_limit = thr->th.th_current_task->td_icvs.thread_limit; | |||
7643 | tmp->cg_nthreads = 1; // Init counter to one active thread, this one | |||
7644 | 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); } | |||
7645 | " 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); } | |||
7646 | 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); }; | |||
7647 | tmp->up = thr->th.th_cg_roots; | |||
7648 | thr->th.th_cg_roots = tmp; | |||
7649 | ||||
7650 | // Launch league of teams now, but not let workers execute | |||
7651 | // (they hang on fork barrier until next parallel) | |||
7652 | #if INCLUDE_SSC_MARKS(1 && 1) | |||
7653 | SSC_MARK_FORKING()__asm__ __volatile__("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(0xd693) : "%ebx"); | |||
7654 | #endif | |||
7655 | __kmp_fork_call(loc, gtid, fork_context_intel, team->t.t_argc, | |||
7656 | (microtask_t)thr->th.th_teams_microtask, // "wrapped" task | |||
7657 | VOLATILE_CAST(launch_t)(launch_t) __kmp_invoke_task_func, NULL__null); | |||
7658 | #if INCLUDE_SSC_MARKS(1 && 1) | |||
7659 | SSC_MARK_JOINING()__asm__ __volatile__("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(0xd694) : "%ebx"); | |||
7660 | #endif | |||
7661 | // If the team size was reduced from the limit, set it to the new size | |||
7662 | if (thr->th.th_team_nproc < thr->th.th_teams_size.nth) | |||
7663 | thr->th.th_teams_size.nth = thr->th.th_team_nproc; | |||
7664 | // AC: last parameter "1" eliminates join barrier which won't work because | |||
7665 | // worker threads are in a fork barrier waiting for more parallel regions | |||
7666 | __kmp_join_call(loc, gtid | |||
7667 | #if OMPT_SUPPORT1 | |||
7668 | , | |||
7669 | fork_context_intel | |||
7670 | #endif | |||
7671 | , | |||
7672 | 1); | |||
7673 | } | |||
7674 | ||||
7675 | int __kmp_invoke_teams_master(int gtid) { | |||
7676 | kmp_info_t *this_thr = __kmp_threads[gtid]; | |||
7677 | kmp_team_t *team = this_thr->th.th_team; | |||
7678 | #if KMP_DEBUG1 | |||
7679 | if (!__kmp_threads[gtid]->th.th_team->t.t_serialized) | |||
7680 | 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", 7681); } | |||
7681 | (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", 7681); }; | |||
7682 | #endif | |||
7683 | __kmp_run_before_invoked_task(gtid, 0, this_thr, team); | |||
7684 | #if OMPT_SUPPORT1 | |||
7685 | int tid = __kmp_tid_from_gtid(gtid); | |||
7686 | ompt_data_t *task_data = | |||
7687 | &team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data; | |||
7688 | ompt_data_t *parallel_data = &team->t.ompt_team_info.parallel_data; | |||
7689 | if (ompt_enabled.ompt_callback_implicit_task) { | |||
7690 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | |||
7691 | ompt_scope_begin, parallel_data, task_data, team->t.t_nproc, tid, | |||
7692 | ompt_task_initial); | |||
7693 | OMPT_CUR_TASK_INFO(this_thr)(&(this_thr->th.th_current_task->ompt_task_info))->thread_num = tid; | |||
7694 | } | |||
7695 | #endif | |||
7696 | __kmp_teams_master(gtid); | |||
7697 | #if OMPT_SUPPORT1 | |||
7698 | this_thr->th.ompt_thread_info.parallel_flags |= ompt_parallel_league; | |||
7699 | #endif | |||
7700 | __kmp_run_after_invoked_task(gtid, 0, this_thr, team); | |||
7701 | return 1; | |||
7702 | } | |||
7703 | ||||
7704 | /* this sets the requested number of threads for the next parallel region | |||
7705 | encountered by this team. since this should be enclosed in the forkjoin | |||
7706 | critical section it should avoid race conditions with asymmetrical nested | |||
7707 | parallelism */ | |||
7708 | ||||
7709 | void __kmp_push_num_threads(ident_t *id, int gtid, int num_threads) { | |||
7710 | kmp_info_t *thr = __kmp_threads[gtid]; | |||
7711 | ||||
7712 | if (num_threads > 0) | |||
7713 | thr->th.th_set_nproc = num_threads; | |||
7714 | } | |||
7715 | ||||
7716 | static void __kmp_push_thread_limit(kmp_info_t *thr, int num_teams, | |||
7717 | int num_threads) { | |||
7718 | KMP_DEBUG_ASSERT(thr)if (!(thr)) { __kmp_debug_assert("thr", "openmp/runtime/src/kmp_runtime.cpp" , 7718); }; | |||
7719 | // Remember the number of threads for inner parallel regions | |||
7720 | if (!TCR_4(__kmp_init_middle)(__kmp_init_middle)) | |||
7721 | __kmp_middle_initialize(); // get internal globals calculated | |||
7722 | __kmp_assign_root_init_mask(); | |||
7723 | KMP_DEBUG_ASSERT(__kmp_avail_proc)if (!(__kmp_avail_proc)) { __kmp_debug_assert("__kmp_avail_proc" , "openmp/runtime/src/kmp_runtime.cpp", 7723); }; | |||
7724 | 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", 7724); }; | |||
7725 | ||||
7726 | if (num_threads == 0) { | |||
7727 | if (__kmp_teams_thread_limit > 0) { | |||
7728 | num_threads = __kmp_teams_thread_limit; | |||
7729 | } else { | |||
7730 | num_threads = __kmp_avail_proc / num_teams; | |||
7731 | } | |||
7732 | // adjust num_threads w/o warning as it is not user setting | |||
7733 | // num_threads = min(num_threads, nthreads-var, thread-limit-var) | |||
7734 | // no thread_limit clause specified - do not change thread-limit-var ICV | |||
7735 | if (num_threads > __kmp_dflt_team_nth) { | |||
7736 | num_threads = __kmp_dflt_team_nth; // honor nthreads-var ICV | |||
7737 | } | |||
7738 | if (num_threads > thr->th.th_current_task->td_icvs.thread_limit) { | |||
7739 | num_threads = thr->th.th_current_task->td_icvs.thread_limit; | |||
7740 | } // prevent team size to exceed thread-limit-var | |||
7741 | if (num_teams * num_threads > __kmp_teams_max_nth) { | |||
7742 | num_threads = __kmp_teams_max_nth / num_teams; | |||
7743 | } | |||
7744 | if (num_threads == 0) { | |||
7745 | num_threads = 1; | |||
7746 | } | |||
7747 | } else { | |||
7748 | if (num_threads < 0) { | |||
7749 | __kmp_msg(kmp_ms_warning, KMP_MSG(CantFormThrTeam, num_threads, 1)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, num_threads, 1 ), | |||
7750 | __kmp_msg_null); | |||
7751 | num_threads = 1; | |||
7752 | } | |||
7753 | // This thread will be the primary thread of the league primary threads | |||
7754 | // Store new thread limit; old limit is saved in th_cg_roots list | |||
7755 | thr->th.th_current_task->td_icvs.thread_limit = num_threads; | |||
7756 | // num_threads = min(num_threads, nthreads-var) | |||
7757 | if (num_threads > __kmp_dflt_team_nth) { | |||
7758 | num_threads = __kmp_dflt_team_nth; // honor nthreads-var ICV | |||
7759 | } | |||
7760 | if (num_teams * num_threads > __kmp_teams_max_nth) { | |||
7761 | int new_threads = __kmp_teams_max_nth / num_teams; | |||
7762 | if (new_threads == 0) { | |||
7763 | new_threads = 1; | |||
7764 | } | |||
7765 | if (new_threads != num_threads) { | |||
7766 | if (!__kmp_reserve_warn) { // user asked for too many threads | |||
7767 | __kmp_reserve_warn = 1; // conflicts with KMP_TEAMS_THREAD_LIMIT | |||
7768 | __kmp_msg(kmp_ms_warning, | |||
7769 | KMP_MSG(CantFormThrTeam, num_threads, new_threads)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, num_threads, new_threads ), | |||
7770 | KMP_HNT(Unset_ALL_THREADS)__kmp_msg_format(kmp_i18n_hnt_Unset_ALL_THREADS), __kmp_msg_null); | |||
7771 | } | |||
7772 | } | |||
7773 | num_threads = new_threads; | |||
7774 | } | |||
7775 | } | |||
7776 | thr->th.th_teams_size.nth = num_threads; | |||
7777 | } | |||
7778 | ||||
7779 | /* this sets the requested number of teams for the teams region and/or | |||
7780 | the number of threads for the next parallel region encountered */ | |||
7781 | void __kmp_push_num_teams(ident_t *id, int gtid, int num_teams, | |||
7782 | int num_threads) { | |||
7783 | kmp_info_t *thr = __kmp_threads[gtid]; | |||
7784 | if (num_teams < 0) { | |||
7785 | // OpenMP specification requires requested values to be positive, | |||
7786 | // but people can send us any value, so we'd better check | |||
7787 | __kmp_msg(kmp_ms_warning, KMP_MSG(NumTeamsNotPositive, num_teams, 1)__kmp_msg_format(kmp_i18n_msg_NumTeamsNotPositive, num_teams, 1), | |||
7788 | __kmp_msg_null); | |||
7789 | num_teams = 1; | |||
7790 | } | |||
7791 | if (num_teams == 0) { | |||
7792 | if (__kmp_nteams > 0) { | |||
7793 | num_teams = __kmp_nteams; | |||
7794 | } else { | |||
7795 | num_teams = 1; // default number of teams is 1. | |||
7796 | } | |||
7797 | } | |||
7798 | if (num_teams > __kmp_teams_max_nth) { // if too many teams requested? | |||
7799 | if (!__kmp_reserve_warn) { | |||
7800 | __kmp_reserve_warn = 1; | |||
7801 | __kmp_msg(kmp_ms_warning, | |||
7802 | KMP_MSG(CantFormThrTeam, num_teams, __kmp_teams_max_nth)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, num_teams, __kmp_teams_max_nth ), | |||
7803 | KMP_HNT(Unset_ALL_THREADS)__kmp_msg_format(kmp_i18n_hnt_Unset_ALL_THREADS), __kmp_msg_null); | |||
7804 | } | |||
7805 | num_teams = __kmp_teams_max_nth; | |||
7806 | } | |||
7807 | // Set number of teams (number of threads in the outer "parallel" of the | |||
7808 | // teams) | |||
7809 | thr->th.th_set_nproc = thr->th.th_teams_size.nteams = num_teams; | |||
7810 | ||||
7811 | __kmp_push_thread_limit(thr, num_teams, num_threads); | |||
7812 | } | |||
7813 | ||||
7814 | /* This sets the requested number of teams for the teams region and/or | |||
7815 | the number of threads for the next parallel region encountered */ | |||
7816 | void __kmp_push_num_teams_51(ident_t *id, int gtid, int num_teams_lb, | |||
7817 | int num_teams_ub, int num_threads) { | |||
7818 | kmp_info_t *thr = __kmp_threads[gtid]; | |||
7819 | 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", 7819); }; | |||
7820 | 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" , 7820); }; | |||
7821 | KMP_DEBUG_ASSERT(num_threads >= 0)if (!(num_threads >= 0)) { __kmp_debug_assert("num_threads >= 0" , "openmp/runtime/src/kmp_runtime.cpp", 7821); }; | |||
7822 | ||||
7823 | if (num_teams_lb > num_teams_ub) { | |||
7824 | __kmp_fatal(KMP_MSG(FailedToCreateTeam, num_teams_lb, num_teams_ub)__kmp_msg_format(kmp_i18n_msg_FailedToCreateTeam, num_teams_lb , num_teams_ub), | |||
7825 | KMP_HNT(SetNewBound, __kmp_teams_max_nth)__kmp_msg_format(kmp_i18n_hnt_SetNewBound, __kmp_teams_max_nth ), __kmp_msg_null); | |||
7826 | } | |||
7827 | ||||
7828 | int num_teams = 1; // defalt number of teams is 1. | |||
7829 | ||||
7830 | if (num_teams_lb == 0 && num_teams_ub > 0) | |||
7831 | num_teams_lb = num_teams_ub; | |||
7832 | ||||
7833 | if (num_teams_lb == 0 && num_teams_ub == 0) { // no num_teams clause | |||
7834 | num_teams = (__kmp_nteams > 0) ? __kmp_nteams : num_teams; | |||
7835 | if (num_teams > __kmp_teams_max_nth) { | |||
7836 | if (!__kmp_reserve_warn) { | |||
7837 | __kmp_reserve_warn = 1; | |||
7838 | __kmp_msg(kmp_ms_warning, | |||
7839 | KMP_MSG(CantFormThrTeam, num_teams, __kmp_teams_max_nth)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, num_teams, __kmp_teams_max_nth ), | |||
7840 | KMP_HNT(Unset_ALL_THREADS)__kmp_msg_format(kmp_i18n_hnt_Unset_ALL_THREADS), __kmp_msg_null); | |||
7841 | } | |||
7842 | num_teams = __kmp_teams_max_nth; | |||
7843 | } | |||
7844 | } else if (num_teams_lb == num_teams_ub) { // requires exact number of teams | |||
7845 | num_teams = num_teams_ub; | |||
7846 | } else { // num_teams_lb <= num_teams <= num_teams_ub | |||
7847 | if (num_threads <= 0) { | |||
7848 | if (num_teams_ub > __kmp_teams_max_nth) { | |||
7849 | num_teams = num_teams_lb; | |||
7850 | } else { | |||
7851 | num_teams = num_teams_ub; | |||
7852 | } | |||
7853 | } else { | |||
7854 | num_teams = (num_threads > __kmp_teams_max_nth) | |||
7855 | ? num_teams | |||
7856 | : __kmp_teams_max_nth / num_threads; | |||
7857 | if (num_teams < num_teams_lb) { | |||
7858 | num_teams = num_teams_lb; | |||
7859 | } else if (num_teams > num_teams_ub) { | |||
7860 | num_teams = num_teams_ub; | |||
7861 | } | |||
7862 | } | |||
7863 | } | |||
7864 | // Set number of teams (number of threads in the outer "parallel" of the | |||
7865 | // teams) | |||
7866 | thr->th.th_set_nproc = thr->th.th_teams_size.nteams = num_teams; | |||
7867 | ||||
7868 | __kmp_push_thread_limit(thr, num_teams, num_threads); | |||
7869 | } | |||
7870 | ||||
7871 | // Set the proc_bind var to use in the following parallel region. | |||
7872 | void __kmp_push_proc_bind(ident_t *id, int gtid, kmp_proc_bind_t proc_bind) { | |||
7873 | kmp_info_t *thr = __kmp_threads[gtid]; | |||
7874 | thr->th.th_set_proc_bind = proc_bind; | |||
7875 | } | |||
7876 | ||||
7877 | /* Launch the worker threads into the microtask. */ | |||
7878 | ||||
7879 | void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team) { | |||
7880 | kmp_info_t *this_thr = __kmp_threads[gtid]; | |||
7881 | ||||
7882 | #ifdef KMP_DEBUG1 | |||
7883 | int f; | |||
7884 | #endif /* KMP_DEBUG */ | |||
7885 | ||||
7886 | KMP_DEBUG_ASSERT(team)if (!(team)) { __kmp_debug_assert("team", "openmp/runtime/src/kmp_runtime.cpp" , 7886); }; | |||
7887 | 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" , 7887); }; | |||
7888 | 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" , 7888); }; | |||
7889 | KMP_MB(); /* Flush all pending memory write invalidates. */ | |||
7890 | ||||
7891 | team->t.t_construct = 0; /* no single directives seen yet */ | |||
7892 | team->t.t_ordered.dt.t_value = | |||
7893 | 0; /* thread 0 enters the ordered section first */ | |||
7894 | ||||
7895 | /* Reset the identifiers on the dispatch buffer */ | |||
7896 | 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", 7896); }; | |||
7897 | if (team->t.t_max_nproc > 1) { | |||
7898 | int i; | |||
7899 | for (i = 0; i < __kmp_dispatch_num_buffers; ++i) { | |||
7900 | team->t.t_disp_buffer[i].buffer_index = i; | |||
7901 | team->t.t_disp_buffer[i].doacross_buf_idx = i; | |||
7902 | } | |||
7903 | } else { | |||
7904 | team->t.t_disp_buffer[0].buffer_index = 0; | |||
7905 | team->t.t_disp_buffer[0].doacross_buf_idx = 0; | |||
7906 | } | |||
7907 | ||||
7908 | KMP_MB(); /* Flush all pending memory write invalidates. */ | |||
7909 | 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" , 7909); }; | |||
7910 | ||||
7911 | #ifdef KMP_DEBUG1 | |||
7912 | for (f = 0; f < team->t.t_nproc; f++) { | |||
7913 | 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", 7914); } | |||
7914 | 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", 7914); }; | |||
7915 | } | |||
7916 | #endif /* KMP_DEBUG */ | |||
7917 | ||||
7918 | /* release the worker threads so they may begin working */ | |||
7919 | __kmp_fork_barrier(gtid, 0); | |||
7920 | } | |||
7921 | ||||
7922 | void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team) { | |||
7923 | kmp_info_t *this_thr = __kmp_threads[gtid]; | |||
7924 | ||||
7925 | KMP_DEBUG_ASSERT(team)if (!(team)) { __kmp_debug_assert("team", "openmp/runtime/src/kmp_runtime.cpp" , 7925); }; | |||
7926 | 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" , 7926); }; | |||
7927 | 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" , 7927); }; | |||
7928 | KMP_MB(); /* Flush all pending memory write invalidates. */ | |||
7929 | ||||
7930 | /* Join barrier after fork */ | |||
7931 | ||||
7932 | #ifdef KMP_DEBUG1 | |||
7933 | if (__kmp_threads[gtid] && | |||
7934 | __kmp_threads[gtid]->th.th_team_nproc != team->t.t_nproc) { | |||
7935 | __kmp_printf("GTID: %d, __kmp_threads[%d]=%p\n", gtid, gtid, | |||
7936 | __kmp_threads[gtid]); | |||
7937 | __kmp_printf("__kmp_threads[%d]->th.th_team_nproc=%d, TEAM: %p, " | |||
7938 | "team->t.t_nproc=%d\n", | |||
7939 | gtid, __kmp_threads[gtid]->th.th_team_nproc, team, | |||
7940 | team->t.t_nproc); | |||
7941 | __kmp_print_structure(); | |||
7942 | } | |||
7943 | 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", 7944); } | |||
7944 | __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", 7944); }; | |||
7945 | #endif /* KMP_DEBUG */ | |||
7946 | ||||
7947 | __kmp_join_barrier(gtid); /* wait for everyone */ | |||
7948 | #if OMPT_SUPPORT1 | |||
7949 | if (ompt_enabled.enabled && | |||
7950 | this_thr->th.ompt_thread_info.state == ompt_state_wait_barrier_implicit) { | |||
7951 | int ds_tid = this_thr->th.th_info.ds.ds_tid; | |||
7952 | ompt_data_t *task_data = OMPT_CUR_TASK_DATA(this_thr)(&(this_thr->th.th_current_task->ompt_task_info.task_data )); | |||
7953 | this_thr->th.ompt_thread_info.state = ompt_state_overhead; | |||
7954 | #if OMPT_OPTIONAL1 | |||
7955 | void *codeptr = NULL__null; | |||
7956 | if (KMP_MASTER_TID(ds_tid)(0 == (ds_tid)) && | |||
7957 | (ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)ompt_callback_sync_region_wait_callback || | |||
7958 | ompt_callbacks.ompt_callback(ompt_callback_sync_region)ompt_callback_sync_region_callback)) | |||
7959 | codeptr = OMPT_CUR_TEAM_INFO(this_thr)(&(this_thr->th.th_team->t.ompt_team_info))->master_return_address; | |||
7960 | ||||
7961 | if (ompt_enabled.ompt_callback_sync_region_wait) { | |||
7962 | ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)ompt_callback_sync_region_wait_callback( | |||
7963 | ompt_sync_region_barrier_implicit, ompt_scope_end, NULL__null, task_data, | |||
7964 | codeptr); | |||
7965 | } | |||
7966 | if (ompt_enabled.ompt_callback_sync_region) { | |||
7967 | ompt_callbacks.ompt_callback(ompt_callback_sync_region)ompt_callback_sync_region_callback( | |||
7968 | ompt_sync_region_barrier_implicit, ompt_scope_end, NULL__null, task_data, | |||
7969 | codeptr); | |||
7970 | } | |||
7971 | #endif | |||
7972 | if (!KMP_MASTER_TID(ds_tid)(0 == (ds_tid)) && ompt_enabled.ompt_callback_implicit_task) { | |||
7973 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( | |||
7974 | ompt_scope_end, NULL__null, task_data, 0, ds_tid, | |||
7975 | ompt_task_implicit); // TODO: Can this be ompt_task_initial? | |||
7976 | } | |||
7977 | } | |||
7978 | #endif | |||
7979 | ||||
7980 | KMP_MB(); /* Flush all pending memory write invalidates. */ | |||
7981 | 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" , 7981); }; | |||
7982 | } | |||
7983 | ||||
7984 | /* ------------------------------------------------------------------------ */ | |||
7985 | ||||
7986 | #ifdef USE_LOAD_BALANCE1 | |||
7987 | ||||
7988 | // Return the worker threads actively spinning in the hot team, if we | |||
7989 | // are at the outermost level of parallelism. Otherwise, return 0. | |||
7990 | static int __kmp_active_hot_team_nproc(kmp_root_t *root) { | |||
7991 | int i; | |||
7992 | int retval; | |||
7993 | kmp_team_t *hot_team; | |||
7994 | ||||
7995 | if (root->r.r_active) { | |||
7996 | return 0; | |||
7997 | } | |||
7998 | hot_team = root->r.r_hot_team; | |||
7999 | if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME(2147483647)) { | |||
8000 | return hot_team->t.t_nproc - 1; // Don't count primary thread | |||
8001 | } | |||
8002 | ||||
8003 | // Skip the primary thread - it is accounted for elsewhere. | |||
8004 | retval = 0; | |||
8005 | for (i = 1; i < hot_team->t.t_nproc; i++) { | |||
8006 | if (hot_team->t.t_threads[i]->th.th_active) { | |||
8007 | retval++; | |||
8008 | } | |||
8009 | } | |||
8010 | return retval; | |||
8011 | } | |||
8012 | ||||
8013 | // Perform an automatic adjustment to the number of | |||
8014 | // threads used by the next parallel region. | |||
8015 | static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc) { | |||
8016 | int retval; | |||
8017 | int pool_active; | |||
8018 | int hot_team_active; | |||
8019 | int team_curr_active; | |||
8020 | int system_active; | |||
8021 | ||||
8022 | 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); } | |||
8023 | set_nproc))if (kmp_b_debug >= 20) { __kmp_debug_printf ("__kmp_load_balance_nproc: called root:%p set_nproc:%d\n" , root, set_nproc); }; | |||
8024 | KMP_DEBUG_ASSERT(root)if (!(root)) { __kmp_debug_assert("root", "openmp/runtime/src/kmp_runtime.cpp" , 8024); }; | |||
8025 | 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", 8026); } | |||
8026 | ->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", 8026); }; | |||
8027 | KMP_DEBUG_ASSERT(set_nproc > 1)if (!(set_nproc > 1)) { __kmp_debug_assert("set_nproc > 1" , "openmp/runtime/src/kmp_runtime.cpp", 8027); }; | |||
8028 | ||||
8029 | if (set_nproc == 1) { | |||
8030 | 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" ); }; | |||
8031 | return 1; | |||
8032 | } | |||
8033 | ||||
8034 | // Threads that are active in the thread pool, active in the hot team for this | |||
8035 | // particular root (if we are at the outer par level), and the currently | |||
8036 | // executing thread (to become the primary thread) are available to add to the | |||
8037 | // new team, but are currently contributing to the system load, and must be | |||
8038 | // accounted for. | |||
8039 | pool_active = __kmp_thread_pool_active_nth; | |||
8040 | hot_team_active = __kmp_active_hot_team_nproc(root); | |||
8041 | team_curr_active = pool_active + hot_team_active + 1; | |||
8042 | ||||
8043 | // Check the system load. | |||
8044 | system_active = __kmp_get_load_balance(__kmp_avail_proc + team_curr_active); | |||
8045 | 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 ); } | |||
8046 | "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 ); } | |||
8047 | 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 ); }; | |||
8048 | ||||
8049 | if (system_active < 0) { | |||
8050 | // There was an error reading the necessary info from /proc, so use the | |||
8051 | // thread limit algorithm instead. Once we set __kmp_global.g.g_dynamic_mode | |||
8052 | // = dynamic_thread_limit, we shouldn't wind up getting back here. | |||
8053 | __kmp_global.g.g_dynamic_mode = dynamic_thread_limit; | |||
8054 | 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); | |||
8055 | ||||
8056 | // Make this call behave like the thread limit algorithm. | |||
8057 | retval = __kmp_avail_proc - __kmp_nth + | |||
8058 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc); | |||
8059 | if (retval > set_nproc) { | |||
8060 | retval = set_nproc; | |||
8061 | } | |||
8062 | if (retval < KMP_MIN_NTH1) { | |||
8063 | retval = KMP_MIN_NTH1; | |||
8064 | } | |||
8065 | ||||
8066 | 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); } | |||
8067 | retval))if (kmp_b_debug >= 20) { __kmp_debug_printf ("__kmp_load_balance_nproc: thread limit exit. retval:%d\n" , retval); }; | |||
8068 | return retval; | |||
8069 | } | |||
8070 | ||||
8071 | // There is a slight delay in the load balance algorithm in detecting new | |||
8072 | // running procs. The real system load at this instant should be at least as | |||
8073 | // large as the #active omp thread that are available to add to the team. | |||
8074 | if (system_active < team_curr_active) { | |||
8075 | system_active = team_curr_active; | |||
8076 | } | |||
8077 | retval = __kmp_avail_proc - system_active + team_curr_active; | |||
8078 | if (retval > set_nproc) { | |||
8079 | retval = set_nproc; | |||
8080 | } | |||
8081 | if (retval < KMP_MIN_NTH1) { | |||
8082 | retval = KMP_MIN_NTH1; | |||
8083 | } | |||
8084 | ||||
8085 | 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); }; | |||
8086 | return retval; | |||
8087 | } // __kmp_load_balance_nproc() | |||
8088 | ||||
8089 | #endif /* USE_LOAD_BALANCE */ | |||
8090 | ||||
8091 | /* ------------------------------------------------------------------------ */ | |||
8092 | ||||
8093 | /* NOTE: this is called with the __kmp_init_lock held */ | |||
8094 | void __kmp_cleanup(void) { | |||
8095 | int f; | |||
8096 | ||||
8097 | KA_TRACE(10, ("__kmp_cleanup: enter\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_cleanup: enter\n" ); }; | |||
8098 | ||||
8099 | if (TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) { | |||
8100 | #if KMP_HANDLE_SIGNALS(1 || 0) | |||
8101 | __kmp_remove_signals(); | |||
8102 | #endif | |||
8103 | TCW_4(__kmp_init_parallel, FALSE)(__kmp_init_parallel) = (0); | |||
8104 | } | |||
8105 | ||||
8106 | if (TCR_4(__kmp_init_middle)(__kmp_init_middle)) { | |||
8107 | #if KMP_AFFINITY_SUPPORTED1 | |||
8108 | __kmp_affinity_uninitialize(); | |||
8109 | #endif /* KMP_AFFINITY_SUPPORTED */ | |||
8110 | __kmp_cleanup_hierarchy(); | |||
8111 | TCW_4(__kmp_init_middle, FALSE)(__kmp_init_middle) = (0); | |||
8112 | } | |||
8113 | ||||
8114 | KA_TRACE(10, ("__kmp_cleanup: go serial cleanup\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_cleanup: go serial cleanup\n" ); }; | |||
8115 | ||||
8116 | if (__kmp_init_serial) { | |||
8117 | __kmp_runtime_destroy(); | |||
8118 | __kmp_init_serial = FALSE0; | |||
8119 | } | |||
8120 | ||||
8121 | __kmp_cleanup_threadprivate_caches(); | |||
8122 | ||||
8123 | for (f = 0; f < __kmp_threads_capacity; f++) { | |||
8124 | if (__kmp_root[f] != NULL__null) { | |||
8125 | __kmp_free(__kmp_root[f])___kmp_free((__kmp_root[f]), "openmp/runtime/src/kmp_runtime.cpp" , 8125); | |||
8126 | __kmp_root[f] = NULL__null; | |||
8127 | } | |||
8128 | } | |||
8129 | __kmp_free(__kmp_threads)___kmp_free((__kmp_threads), "openmp/runtime/src/kmp_runtime.cpp" , 8129); | |||
8130 | // __kmp_threads and __kmp_root were allocated at once, as single block, so | |||
8131 | // there is no need in freeing __kmp_root. | |||
8132 | __kmp_threads = NULL__null; | |||
8133 | __kmp_root = NULL__null; | |||
8134 | __kmp_threads_capacity = 0; | |||
8135 | ||||
8136 | // Free old __kmp_threads arrays if they exist. | |||
8137 | kmp_old_threads_list_t *ptr = __kmp_old_threads_list; | |||
8138 | while (ptr) { | |||
8139 | kmp_old_threads_list_t *next = ptr->next; | |||
8140 | __kmp_free(ptr->threads)___kmp_free((ptr->threads), "openmp/runtime/src/kmp_runtime.cpp" , 8140); | |||
8141 | __kmp_free(ptr)___kmp_free((ptr), "openmp/runtime/src/kmp_runtime.cpp", 8141 ); | |||
8142 | ptr = next; | |||
8143 | } | |||
8144 | ||||
8145 | #if KMP_USE_DYNAMIC_LOCK1 | |||
8146 | __kmp_cleanup_indirect_user_locks(); | |||
8147 | #else | |||
8148 | __kmp_cleanup_user_locks(); | |||
8149 | #endif | |||
8150 | #if OMPD_SUPPORT1 | |||
8151 | if (ompd_state) { | |||
8152 | __kmp_free(ompd_env_block)___kmp_free((ompd_env_block), "openmp/runtime/src/kmp_runtime.cpp" , 8152); | |||
8153 | ompd_env_block = NULL__null; | |||
8154 | ompd_env_block_size = 0; | |||
8155 | } | |||
8156 | #endif | |||
8157 | ||||
8158 | #if KMP_AFFINITY_SUPPORTED1 | |||
8159 | KMP_INTERNAL_FREE(CCAST(char *, __kmp_cpuinfo_file))free(const_cast<char *>(__kmp_cpuinfo_file)); | |||
8160 | __kmp_cpuinfo_file = NULL__null; | |||
8161 | #endif /* KMP_AFFINITY_SUPPORTED */ | |||
8162 | ||||
8163 | #if KMP_USE_ADAPTIVE_LOCKS(0 || 1) && !0 | |||
8164 | #if KMP_DEBUG_ADAPTIVE_LOCKS0 | |||
8165 | __kmp_print_speculative_stats(); | |||
8166 | #endif | |||
8167 | #endif | |||
8168 | KMP_INTERNAL_FREE(__kmp_nested_nth.nth)free(__kmp_nested_nth.nth); | |||
8169 | __kmp_nested_nth.nth = NULL__null; | |||
8170 | __kmp_nested_nth.size = 0; | |||
8171 | __kmp_nested_nth.used = 0; | |||
8172 | KMP_INTERNAL_FREE(__kmp_nested_proc_bind.bind_types)free(__kmp_nested_proc_bind.bind_types); | |||
8173 | __kmp_nested_proc_bind.bind_types = NULL__null; | |||
8174 | __kmp_nested_proc_bind.size = 0; | |||
8175 | __kmp_nested_proc_bind.used = 0; | |||
8176 | if (__kmp_affinity_format) { | |||
8177 | KMP_INTERNAL_FREE(__kmp_affinity_format)free(__kmp_affinity_format); | |||
8178 | __kmp_affinity_format = NULL__null; | |||
8179 | } | |||
8180 | ||||
8181 | __kmp_i18n_catclose(); | |||
8182 | ||||
8183 | #if KMP_USE_HIER_SCHED0 | |||
8184 | __kmp_hier_scheds.deallocate(); | |||
8185 | #endif | |||
8186 | ||||
8187 | #if KMP_STATS_ENABLED0 | |||
8188 | __kmp_stats_fini(); | |||
8189 | #endif | |||
8190 | ||||
8191 | KA_TRACE(10, ("__kmp_cleanup: exit\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_cleanup: exit\n" ); }; | |||
8192 | } | |||
8193 | ||||
8194 | /* ------------------------------------------------------------------------ */ | |||
8195 | ||||
8196 | int __kmp_ignore_mppbeg(void) { | |||
8197 | char *env; | |||
8198 | ||||
8199 | if ((env = getenv("KMP_IGNORE_MPPBEG")) != NULL__null) { | |||
8200 | if (__kmp_str_match_false(env)) | |||
8201 | return FALSE0; | |||
8202 | } | |||
8203 | // By default __kmpc_begin() is no-op. | |||
8204 | return TRUE(!0); | |||
8205 | } | |||
8206 | ||||
8207 | int __kmp_ignore_mppend(void) { | |||
8208 | char *env; | |||
8209 | ||||
8210 | if ((env = getenv("KMP_IGNORE_MPPEND")) != NULL__null) { | |||
8211 | if (__kmp_str_match_false(env)) | |||
8212 | return FALSE0; | |||
8213 | } | |||
8214 | // By default __kmpc_end() is no-op. | |||
8215 | return TRUE(!0); | |||
8216 | } | |||
8217 | ||||
8218 | void __kmp_internal_begin(void) { | |||
8219 | int gtid; | |||
8220 | kmp_root_t *root; | |||
8221 | ||||
8222 | /* this is a very important step as it will register new sibling threads | |||
8223 | and assign these new uber threads a new gtid */ | |||
8224 | gtid = __kmp_entry_gtid()__kmp_get_global_thread_id_reg(); | |||
8225 | root = __kmp_threads[gtid]->th.th_root; | |||
8226 | KMP_ASSERT(KMP_UBER_GTID(gtid))if (!(KMP_UBER_GTID(gtid))) { __kmp_debug_assert("KMP_UBER_GTID(gtid)" , "openmp/runtime/src/kmp_runtime.cpp", 8226); }; | |||
8227 | ||||
8228 | if (root->r.r_begin) | |||
8229 | return; | |||
8230 | __kmp_acquire_lock(&root->r.r_begin_lock, gtid); | |||
8231 | if (root->r.r_begin) { | |||
8232 | __kmp_release_lock(&root->r.r_begin_lock, gtid); | |||
8233 | return; | |||
8234 | } | |||
8235 | ||||
8236 | root->r.r_begin = TRUE(!0); | |||
8237 | ||||
8238 | __kmp_release_lock(&root->r.r_begin_lock, gtid); | |||
8239 | } | |||
8240 | ||||
8241 | /* ------------------------------------------------------------------------ */ | |||
8242 | ||||
8243 | void __kmp_user_set_library(enum library_type arg) { | |||
8244 | int gtid; | |||
8245 | kmp_root_t *root; | |||
8246 | kmp_info_t *thread; | |||
8247 | ||||
8248 | /* first, make sure we are initialized so we can get our gtid */ | |||
8249 | ||||
8250 | gtid = __kmp_entry_gtid()__kmp_get_global_thread_id_reg(); | |||
8251 | thread = __kmp_threads[gtid]; | |||
8252 | ||||
8253 | root = thread->th.th_root; | |||
8254 | ||||
8255 | 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); } | |||
8256 | 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); }; | |||
8257 | if (root->r.r_in_parallel) { /* Must be called in serial section of top-level | |||
8258 | thread */ | |||
8259 | KMP_WARNING(SetLibraryIncorrectCall)__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_SetLibraryIncorrectCall ), __kmp_msg_null); | |||
8260 | return; | |||
8261 | } | |||
8262 | ||||
8263 | switch (arg) { | |||
8264 | case library_serial: | |||
8265 | thread->th.th_set_nproc = 0; | |||
8266 | set__nproc(thread, 1)(((thread)->th.th_current_task->td_icvs.nproc) = (1)); | |||
8267 | break; | |||
8268 | case library_turnaround: | |||
8269 | thread->th.th_set_nproc = 0; | |||
8270 | 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)) | |||
8271 | : __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)); | |||
8272 | break; | |||
8273 | case library_throughput: | |||
8274 | thread->th.th_set_nproc = 0; | |||
8275 | 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)) | |||
8276 | : __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)); | |||
8277 | break; | |||
8278 | default: | |||
8279 | KMP_FATAL(UnknownLibraryType, arg)__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_UnknownLibraryType, arg), __kmp_msg_null); | |||
8280 | } | |||
8281 | ||||
8282 | __kmp_aux_set_library(arg); | |||
8283 | } | |||
8284 | ||||
8285 | void __kmp_aux_set_stacksize(size_t arg) { | |||
8286 | if (!__kmp_init_serial) | |||
8287 | __kmp_serial_initialize(); | |||
8288 | ||||
8289 | #if KMP_OS_DARWIN0 | |||
8290 | if (arg & (0x1000 - 1)) { | |||
8291 | arg &= ~(0x1000 - 1); | |||
8292 | if (arg + 0x1000) /* check for overflow if we round up */ | |||
8293 | arg += 0x1000; | |||
8294 | } | |||
8295 | #endif | |||
8296 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); | |||
8297 | ||||
8298 | /* only change the default stacksize before the first parallel region */ | |||
8299 | if (!TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) { | |||
8300 | size_t value = arg; /* argument is in bytes */ | |||
8301 | ||||
8302 | if (value < __kmp_sys_min_stksize) | |||
8303 | value = __kmp_sys_min_stksize; | |||
8304 | else if (value > KMP_MAX_STKSIZE(~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1) ))) | |||
8305 | value = KMP_MAX_STKSIZE(~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1) )); | |||
8306 | ||||
8307 | __kmp_stksize = value; | |||
8308 | ||||
8309 | __kmp_env_stksize = TRUE(!0); /* was KMP_STACKSIZE specified? */ | |||
8310 | } | |||
8311 | ||||
8312 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | |||
8313 | } | |||
8314 | ||||
8315 | /* set the behaviour of the runtime library */ | |||
8316 | /* TODO this can cause some odd behaviour with sibling parallelism... */ | |||
8317 | void __kmp_aux_set_library(enum library_type arg) { | |||
8318 | __kmp_library = arg; | |||
8319 | ||||
8320 | switch (__kmp_library) { | |||
8321 | case library_serial: { | |||
8322 | KMP_INFORM(LibraryIsSerial)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_LibraryIsSerial ), __kmp_msg_null); | |||
8323 | } break; | |||
8324 | case library_turnaround: | |||
8325 | if (__kmp_use_yield == 1 && !__kmp_use_yield_exp_set) | |||
8326 | __kmp_use_yield = 2; // only yield when oversubscribed | |||
8327 | break; | |||
8328 | case library_throughput: | |||
8329 | if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME(2147483647)) | |||
8330 | __kmp_dflt_blocktime = KMP_DEFAULT_BLOCKTIME(__kmp_is_hybrid_cpu() ? (0) : (200)); | |||
8331 | break; | |||
8332 | default: | |||
8333 | KMP_FATAL(UnknownLibraryType, arg)__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_UnknownLibraryType, arg), __kmp_msg_null); | |||
8334 | } | |||
8335 | } | |||
8336 | ||||
8337 | /* Getting team information common for all team API */ | |||
8338 | // Returns NULL if not in teams construct | |||
8339 | static kmp_team_t *__kmp_aux_get_team_info(int &teams_serialized) { | |||
8340 | kmp_info_t *thr = __kmp_entry_thread(); | |||
8341 | teams_serialized = 0; | |||
8342 | if (thr->th.th_teams_microtask) { | |||
8343 | kmp_team_t *team = thr->th.th_team; | |||
8344 | int tlevel = thr->th.th_teams_level; // the level of the teams construct | |||
8345 | int ii = team->t.t_level; | |||
8346 | teams_serialized = team->t.t_serialized; | |||
8347 | int level = tlevel + 1; | |||
8348 | KMP_DEBUG_ASSERT(ii >= tlevel)if (!(ii >= tlevel)) { __kmp_debug_assert("ii >= tlevel" , "openmp/runtime/src/kmp_runtime.cpp", 8348); }; | |||
8349 | while (ii > level) { | |||
8350 | for (teams_serialized = team->t.t_serialized; | |||
8351 | (teams_serialized > 0) && (ii > level); teams_serialized--, ii--) { | |||
8352 | } | |||
8353 | if (team->t.t_serialized && (!teams_serialized)) { | |||
8354 | team = team->t.t_parent; | |||
8355 | continue; | |||
8356 | } | |||
8357 | if (ii > level) { | |||
8358 | team = team->t.t_parent; | |||
8359 | ii--; | |||
8360 | } | |||
8361 | } | |||
8362 | return team; | |||
8363 | } | |||
8364 | return NULL__null; | |||
8365 | } | |||
8366 | ||||
8367 | int __kmp_aux_get_team_num() { | |||
8368 | int serialized; | |||
8369 | kmp_team_t *team = __kmp_aux_get_team_info(serialized); | |||
8370 | if (team) { | |||
8371 | if (serialized > 1) { | |||
8372 | return 0; // teams region is serialized ( 1 team of 1 thread ). | |||
8373 | } else { | |||
8374 | return team->t.t_master_tid; | |||
8375 | } | |||
8376 | } | |||
8377 | return 0; | |||
8378 | } | |||
8379 | ||||
8380 | int __kmp_aux_get_num_teams() { | |||
8381 | int serialized; | |||
8382 | kmp_team_t *team = __kmp_aux_get_team_info(serialized); | |||
8383 | if (team) { | |||
8384 | if (serialized > 1) { | |||
8385 | return 1; | |||
8386 | } else { | |||
8387 | return team->t.t_parent->t.t_nproc; | |||
8388 | } | |||
8389 | } | |||
8390 | return 1; | |||
8391 | } | |||
8392 | ||||
8393 | /* ------------------------------------------------------------------------ */ | |||
8394 | ||||
8395 | /* | |||
8396 | * Affinity Format Parser | |||
8397 | * | |||
8398 | * Field is in form of: %[[[0].]size]type | |||
8399 | * % and type are required (%% means print a literal '%') | |||
8400 | * type is either single char or long name surrounded by {}, | |||
8401 | * e.g., N or {num_threads} | |||
8402 | * 0 => leading zeros | |||
8403 | * . => right justified when size is specified | |||
8404 | * by default output is left justified | |||
8405 | * size is the *minimum* field length | |||
8406 | * All other characters are printed as is | |||
8407 | * | |||
8408 | * Available field types: | |||
8409 | * L {thread_level} - omp_get_level() | |||
8410 | * n {thread_num} - omp_get_thread_num() | |||
8411 | * h {host} - name of host machine | |||
8412 | * P {process_id} - process id (integer) | |||
8413 | * T {thread_identifier} - native thread identifier (integer) | |||
8414 | * N {num_threads} - omp_get_num_threads() | |||
8415 | * A {ancestor_tnum} - omp_get_ancestor_thread_num(omp_get_level()-1) | |||
8416 | * a {thread_affinity} - comma separated list of integers or integer ranges | |||
8417 | * (values of affinity mask) | |||
8418 | * | |||
8419 | * Implementation-specific field types can be added | |||
8420 | * If a type is unknown, print "undefined" | |||
8421 | */ | |||
8422 | ||||
8423 | // Structure holding the short name, long name, and corresponding data type | |||
8424 | // for snprintf. A table of these will represent the entire valid keyword | |||
8425 | // field types. | |||
8426 | typedef struct kmp_affinity_format_field_t { | |||
8427 | char short_name; // from spec e.g., L -> thread level | |||
8428 | const char *long_name; // from spec thread_level -> thread level | |||
8429 | char field_format; // data type for snprintf (typically 'd' or 's' | |||
8430 | // for integer or string) | |||
8431 | } kmp_affinity_format_field_t; | |||
8432 | ||||
8433 | static const kmp_affinity_format_field_t __kmp_affinity_format_table[] = { | |||
8434 | #if KMP_AFFINITY_SUPPORTED1 | |||
8435 | {'A', "thread_affinity", 's'}, | |||
8436 | #endif | |||
8437 | {'t', "team_num", 'd'}, | |||
8438 | {'T', "num_teams", 'd'}, | |||
8439 | {'L', "nesting_level", 'd'}, | |||
8440 | {'n', "thread_num", 'd'}, | |||
8441 | {'N', "num_threads", 'd'}, | |||
8442 | {'a', "ancestor_tnum", 'd'}, | |||
8443 | {'H', "host", 's'}, | |||
8444 | {'P', "process_id", 'd'}, | |||
8445 | {'i', "native_thread_id", 'd'}}; | |||
8446 | ||||
8447 | // Return the number of characters it takes to hold field | |||
8448 | static int __kmp_aux_capture_affinity_field(int gtid, const kmp_info_t *th, | |||
8449 | const char **ptr, | |||
8450 | kmp_str_buf_t *field_buffer) { | |||
8451 | int rc, format_index, field_value; | |||
8452 | const char *width_left, *width_right; | |||
8453 | bool pad_zeros, right_justify, parse_long_name, found_valid_name; | |||
8454 | static const int FORMAT_SIZE = 20; | |||
8455 | char format[FORMAT_SIZE] = {0}; | |||
8456 | char absolute_short_name = 0; | |||
8457 | ||||
8458 | KMP_DEBUG_ASSERT(gtid >= 0)if (!(gtid >= 0)) { __kmp_debug_assert("gtid >= 0", "openmp/runtime/src/kmp_runtime.cpp" , 8458); }; | |||
8459 | KMP_DEBUG_ASSERT(th)if (!(th)) { __kmp_debug_assert("th", "openmp/runtime/src/kmp_runtime.cpp" , 8459); }; | |||
8460 | KMP_DEBUG_ASSERT(**ptr == '%')if (!(**ptr == '%')) { __kmp_debug_assert("**ptr == '%'", "openmp/runtime/src/kmp_runtime.cpp" , 8460); }; | |||
8461 | KMP_DEBUG_ASSERT(field_buffer)if (!(field_buffer)) { __kmp_debug_assert("field_buffer", "openmp/runtime/src/kmp_runtime.cpp" , 8461); }; | |||
8462 | ||||
8463 | __kmp_str_buf_clear(field_buffer); | |||
8464 | ||||
8465 | // Skip the initial % | |||
8466 | (*ptr)++; | |||
8467 | ||||
8468 | // Check for %% first | |||
8469 | if (**ptr == '%') { | |||
8470 | __kmp_str_buf_cat(field_buffer, "%", 1); | |||
8471 | (*ptr)++; // skip over the second % | |||
8472 | return 1; | |||
8473 | } | |||
8474 | ||||
8475 | // Parse field modifiers if they are present | |||
8476 | pad_zeros = false; | |||
8477 | if (**ptr == '0') { | |||
8478 | pad_zeros = true; | |||
8479 | (*ptr)++; // skip over 0 | |||
8480 | } | |||
8481 | right_justify = false; | |||
8482 | if (**ptr == '.') { | |||
8483 | right_justify = true; | |||
8484 | (*ptr)++; // skip over . | |||
8485 | } | |||
8486 | // Parse width of field: [width_left, width_right) | |||
8487 | width_left = width_right = NULL__null; | |||
8488 | if (**ptr >= '0' && **ptr <= '9') { | |||
8489 | width_left = *ptr; | |||
8490 | SKIP_DIGITS(*ptr){ while (*(*ptr) >= '0' && *(*ptr) <= '9') (*ptr )++; }; | |||
8491 | width_right = *ptr; | |||
8492 | } | |||
8493 | ||||
8494 | // Create the format for KMP_SNPRINTF based on flags parsed above | |||
8495 | format_index = 0; | |||
8496 | format[format_index++] = '%'; | |||
8497 | if (!right_justify) | |||
8498 | format[format_index++] = '-'; | |||
8499 | if (pad_zeros) | |||
8500 | format[format_index++] = '0'; | |||
8501 | if (width_left && width_right) { | |||
8502 | int i = 0; | |||
8503 | // Only allow 8 digit number widths. | |||
8504 | // This also prevents overflowing format variable | |||
8505 | while (i < 8 && width_left < width_right) { | |||
8506 | format[format_index++] = *width_left; | |||
8507 | width_left++; | |||
8508 | i++; | |||
8509 | } | |||
8510 | } | |||
8511 | ||||
8512 | // Parse a name (long or short) | |||
8513 | // Canonicalize the name into absolute_short_name | |||
8514 | found_valid_name = false; | |||
8515 | parse_long_name = (**ptr == '{'); | |||
8516 | if (parse_long_name) | |||
8517 | (*ptr)++; // skip initial left brace | |||
8518 | for (size_t i = 0; i < sizeof(__kmp_affinity_format_table) / | |||
8519 | sizeof(__kmp_affinity_format_table[0]); | |||
8520 | ++i) { | |||
8521 | char short_name = __kmp_affinity_format_table[i].short_name; | |||
8522 | const char *long_name = __kmp_affinity_format_table[i].long_name; | |||
8523 | char field_format = __kmp_affinity_format_table[i].field_format; | |||
8524 | if (parse_long_name) { | |||
8525 | size_t length = KMP_STRLENstrlen(long_name); | |||
8526 | if (strncmp(*ptr, long_name, length) == 0) { | |||
8527 | found_valid_name = true; | |||
8528 | (*ptr) += length; // skip the long name | |||
8529 | } | |||
8530 | } else if (**ptr == short_name) { | |||
8531 | found_valid_name = true; | |||
8532 | (*ptr)++; // skip the short name | |||
8533 | } | |||
8534 | if (found_valid_name) { | |||
8535 | format[format_index++] = field_format; | |||
8536 | format[format_index++] = '\0'; | |||
8537 | absolute_short_name = short_name; | |||
8538 | break; | |||
8539 | } | |||
8540 | } | |||
8541 | if (parse_long_name) { | |||
8542 | if (**ptr != '}') { | |||
8543 | absolute_short_name = 0; | |||
8544 | } else { | |||
8545 | (*ptr)++; // skip over the right brace | |||
8546 | } | |||
8547 | } | |||
8548 | ||||
8549 | // Attempt to fill the buffer with the requested | |||
8550 | // value using snprintf within __kmp_str_buf_print() | |||
8551 | switch (absolute_short_name) { | |||
8552 | case 't': | |||
8553 | rc = __kmp_str_buf_print(field_buffer, format, __kmp_aux_get_team_num()); | |||
8554 | break; | |||
8555 | case 'T': | |||
8556 | rc = __kmp_str_buf_print(field_buffer, format, __kmp_aux_get_num_teams()); | |||
8557 | break; | |||
8558 | case 'L': | |||
8559 | rc = __kmp_str_buf_print(field_buffer, format, th->th.th_team->t.t_level); | |||
8560 | break; | |||
8561 | case 'n': | |||
8562 | rc = __kmp_str_buf_print(field_buffer, format, __kmp_tid_from_gtid(gtid)); | |||
8563 | break; | |||
8564 | case 'H': { | |||
8565 | static const int BUFFER_SIZE = 256; | |||
8566 | char buf[BUFFER_SIZE]; | |||
8567 | __kmp_expand_host_name(buf, BUFFER_SIZE); | |||
8568 | rc = __kmp_str_buf_print(field_buffer, format, buf); | |||
8569 | } break; | |||
8570 | case 'P': | |||
8571 | rc = __kmp_str_buf_print(field_buffer, format, getpid()); | |||
8572 | break; | |||
8573 | case 'i': | |||
8574 | rc = __kmp_str_buf_print(field_buffer, format, __kmp_gettid()syscall(186)); | |||
8575 | break; | |||
8576 | case 'N': | |||
8577 | rc = __kmp_str_buf_print(field_buffer, format, th->th.th_team->t.t_nproc); | |||
8578 | break; | |||
8579 | case 'a': | |||
8580 | field_value = | |||
8581 | __kmp_get_ancestor_thread_num(gtid, th->th.th_team->t.t_level - 1); | |||
8582 | rc = __kmp_str_buf_print(field_buffer, format, field_value); | |||
8583 | break; | |||
8584 | #if KMP_AFFINITY_SUPPORTED1 | |||
8585 | case 'A': { | |||
8586 | kmp_str_buf_t buf; | |||
8587 | __kmp_str_buf_init(&buf){ (&buf)->str = (&buf)->bulk; (&buf)->size = sizeof((&buf)->bulk); (&buf)->used = 0; (& buf)->bulk[0] = 0; }; | |||
8588 | __kmp_affinity_str_buf_mask(&buf, th->th.th_affin_mask); | |||
8589 | rc = __kmp_str_buf_print(field_buffer, format, buf.str); | |||
8590 | __kmp_str_buf_free(&buf); | |||
8591 | } break; | |||
8592 | #endif | |||
8593 | default: | |||
8594 | // According to spec, If an implementation does not have info for field | |||
8595 | // type, then "undefined" is printed | |||
8596 | rc = __kmp_str_buf_print(field_buffer, "%s", "undefined"); | |||
8597 | // Skip the field | |||
8598 | if (parse_long_name) { | |||
8599 | SKIP_TOKEN(*ptr){ while ((*(*ptr) >= '0' && *(*ptr) <= '9') || ( *(*ptr) >= 'a' && *(*ptr) <= 'z') || (*(*ptr) >= 'A' && *(*ptr) <= 'Z') || *(*ptr) == '_') (*ptr)++ ; }; | |||
8600 | if (**ptr == '}') | |||
8601 | (*ptr)++; | |||
8602 | } else { | |||
8603 | (*ptr)++; | |||
8604 | } | |||
8605 | } | |||
8606 | ||||
8607 | KMP_ASSERT(format_index <= FORMAT_SIZE)if (!(format_index <= FORMAT_SIZE)) { __kmp_debug_assert("format_index <= FORMAT_SIZE" , "openmp/runtime/src/kmp_runtime.cpp", 8607); }; | |||
8608 | return rc; | |||
8609 | } | |||
8610 | ||||
8611 | /* | |||
8612 | * Return number of characters needed to hold the affinity string | |||
8613 | * (not including null byte character) | |||
8614 | * The resultant string is printed to buffer, which the caller can then | |||
8615 | * handle afterwards | |||
8616 | */ | |||
8617 | size_t __kmp_aux_capture_affinity(int gtid, const char *format, | |||
8618 | kmp_str_buf_t *buffer) { | |||
8619 | const char *parse_ptr; | |||
8620 | size_t retval; | |||
8621 | const kmp_info_t *th; | |||
8622 | kmp_str_buf_t field; | |||
8623 | ||||
8624 | KMP_DEBUG_ASSERT(buffer)if (!(buffer)) { __kmp_debug_assert("buffer", "openmp/runtime/src/kmp_runtime.cpp" , 8624); }; | |||
8625 | KMP_DEBUG_ASSERT(gtid >= 0)if (!(gtid >= 0)) { __kmp_debug_assert("gtid >= 0", "openmp/runtime/src/kmp_runtime.cpp" , 8625); }; | |||
8626 | ||||
8627 | __kmp_str_buf_init(&field){ (&field)->str = (&field)->bulk; (&field)-> size = sizeof((&field)->bulk); (&field)->used = 0; (&field)->bulk[0] = 0; }; | |||
8628 | __kmp_str_buf_clear(buffer); | |||
8629 | ||||
8630 | th = __kmp_threads[gtid]; | |||
8631 | retval = 0; | |||
8632 | ||||
8633 | // If format is NULL or zero-length string, then we use | |||
8634 | // affinity-format-var ICV | |||
8635 | parse_ptr = format; | |||
8636 | if (parse_ptr == NULL__null || *parse_ptr == '\0') { | |||
8637 | parse_ptr = __kmp_affinity_format; | |||
8638 | } | |||
8639 | KMP_DEBUG_ASSERT(parse_ptr)if (!(parse_ptr)) { __kmp_debug_assert("parse_ptr", "openmp/runtime/src/kmp_runtime.cpp" , 8639); }; | |||
8640 | ||||
8641 | while (*parse_ptr != '\0') { | |||
8642 | // Parse a field | |||
8643 | if (*parse_ptr == '%') { | |||
8644 | // Put field in the buffer | |||
8645 | int rc = __kmp_aux_capture_affinity_field(gtid, th, &parse_ptr, &field); | |||
8646 | __kmp_str_buf_catbuf(buffer, &field); | |||
8647 | retval += rc; | |||
8648 | } else { | |||
8649 | // Put literal character in buffer | |||
8650 | __kmp_str_buf_cat(buffer, parse_ptr, 1); | |||
8651 | retval++; | |||
8652 | parse_ptr++; | |||
8653 | } | |||
8654 | } | |||
8655 | __kmp_str_buf_free(&field); | |||
8656 | return retval; | |||
8657 | } | |||
8658 | ||||
8659 | // Displays the affinity string to stdout | |||
8660 | void __kmp_aux_display_affinity(int gtid, const char *format) { | |||
8661 | kmp_str_buf_t buf; | |||
8662 | __kmp_str_buf_init(&buf){ (&buf)->str = (&buf)->bulk; (&buf)->size = sizeof((&buf)->bulk); (&buf)->used = 0; (& buf)->bulk[0] = 0; }; | |||
8663 | __kmp_aux_capture_affinity(gtid, format, &buf); | |||
8664 | __kmp_fprintf(kmp_out, "%s" KMP_END_OF_LINE"\n", buf.str); | |||
8665 | __kmp_str_buf_free(&buf); | |||
8666 | } | |||
8667 | ||||
8668 | /* ------------------------------------------------------------------------ */ | |||
8669 | ||||
8670 | void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid) { | |||
8671 | int blocktime = arg; /* argument is in milliseconds */ | |||
8672 | #if KMP_USE_MONITOR | |||
8673 | int bt_intervals; | |||
8674 | #endif | |||
8675 | kmp_int8 bt_set; | |||
8676 | ||||
8677 | __kmp_save_internal_controls(thread); | |||
8678 | ||||
8679 | /* Normalize and set blocktime for the teams */ | |||
8680 | if (blocktime < KMP_MIN_BLOCKTIME(0)) | |||
8681 | blocktime = KMP_MIN_BLOCKTIME(0); | |||
8682 | else if (blocktime > KMP_MAX_BLOCKTIME(2147483647)) | |||
8683 | blocktime = KMP_MAX_BLOCKTIME(2147483647); | |||
8684 | ||||
8685 | 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)); | |||
8686 | 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)); | |||
8687 | ||||
8688 | #if KMP_USE_MONITOR | |||
8689 | /* Calculate and set blocktime intervals for the teams */ | |||
8690 | bt_intervals = KMP_INTERVALS_FROM_BLOCKTIME(blocktime, __kmp_monitor_wakeups); | |||
8691 | ||||
8692 | set__bt_intervals_team(thread->th.th_team, tid, bt_intervals); | |||
8693 | set__bt_intervals_team(thread->th.th_serial_team, 0, bt_intervals); | |||
8694 | #endif | |||
8695 | ||||
8696 | /* Set whether blocktime has been set to "TRUE" */ | |||
8697 | bt_set = TRUE(!0); | |||
8698 | ||||
8699 | 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)); | |||
8700 | 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)); | |||
8701 | #if KMP_USE_MONITOR | |||
8702 | 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); } | |||
8703 | "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); } | |||
8704 | __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); } | |||
8705 | 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); } | |||
8706 | __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); }; | |||
8707 | #else | |||
8708 | 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); } | |||
8709 | __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); } | |||
8710 | 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); }; | |||
8711 | #endif | |||
8712 | } | |||
8713 | ||||
8714 | void __kmp_aux_set_defaults(char const *str, size_t len) { | |||
8715 | if (!__kmp_init_serial) { | |||
8716 | __kmp_serial_initialize(); | |||
8717 | } | |||
8718 | __kmp_env_initialize(str); | |||
8719 | ||||
8720 | if (__kmp_settings || __kmp_display_env || __kmp_display_env_verbose) { | |||
8721 | __kmp_env_print(); | |||
8722 | } | |||
8723 | } // __kmp_aux_set_defaults | |||
8724 | ||||
8725 | /* ------------------------------------------------------------------------ */ | |||
8726 | /* internal fast reduction routines */ | |||
8727 | ||||
8728 | PACKED_REDUCTION_METHOD_T | |||
8729 | __kmp_determine_reduction_method( | |||
8730 | ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, | |||
8731 | void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data), | |||
8732 | kmp_critical_name *lck) { | |||
8733 | ||||
8734 | // Default reduction method: critical construct ( lck != NULL, like in current | |||
8735 | // PAROPT ) | |||
8736 | // If ( reduce_data!=NULL && reduce_func!=NULL ): the tree-reduction method | |||
8737 | // can be selected by RTL | |||
8738 | // If loc->flags contains KMP_IDENT_ATOMIC_REDUCE, the atomic reduce method | |||
8739 | // can be selected by RTL | |||
8740 | // Finally, it's up to OpenMP RTL to make a decision on which method to select | |||
8741 | // among generated by PAROPT. | |||
8742 | ||||
8743 | PACKED_REDUCTION_METHOD_T retval; | |||
8744 | ||||
8745 | int team_size; | |||
8746 | ||||
8747 | KMP_DEBUG_ASSERT(loc)if (!(loc)) { __kmp_debug_assert("loc", "openmp/runtime/src/kmp_runtime.cpp" , 8747); }; // it would be nice to test ( loc != 0 ) | |||
8748 | KMP_DEBUG_ASSERT(lck)if (!(lck)) { __kmp_debug_assert("lck", "openmp/runtime/src/kmp_runtime.cpp" , 8748); }; // it would be nice to test ( lck != 0 ) | |||
8749 | ||||
8750 | #define FAST_REDUCTION_ATOMIC_METHOD_GENERATED \ | |||
8751 | (loc && \ | |||
8752 | ((loc->flags & (KMP_IDENT_ATOMIC_REDUCE)) == (KMP_IDENT_ATOMIC_REDUCE))) | |||
8753 | #define FAST_REDUCTION_TREE_METHOD_GENERATED ((reduce_data) && (reduce_func)) | |||
8754 | ||||
8755 | retval = critical_reduce_block; | |||
8756 | ||||
8757 | // another choice of getting a team size (with 1 dynamic deference) is slower | |||
8758 | team_size = __kmp_get_team_num_threads(global_tid)(__kmp_threads[(global_tid)]->th.th_team->t.t_nproc); | |||
8759 | if (team_size == 1) { | |||
8760 | ||||
8761 | retval = empty_reduce_block; | |||
8762 | ||||
8763 | } else { | |||
8764 | ||||
8765 | int atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED; | |||
8766 | ||||
8767 | #if KMP_ARCH_X86_641 || KMP_ARCH_PPC64(0 || 0) || KMP_ARCH_AARCH640 || \ | |||
8768 | KMP_ARCH_MIPS640 || KMP_ARCH_RISCV640 || KMP_ARCH_LOONGARCH640 | |||
8769 | ||||
8770 | #if KMP_OS_LINUX1 || KMP_OS_DRAGONFLY0 || KMP_OS_FREEBSD0 || KMP_OS_NETBSD0 || \ | |||
8771 | KMP_OS_OPENBSD0 || KMP_OS_WINDOWS0 || KMP_OS_DARWIN0 || KMP_OS_HURD0 | |||
8772 | ||||
8773 | int teamsize_cutoff = 4; | |||
8774 | ||||
8775 | #if KMP_MIC_SUPPORTED((0 || 1) && (1 || 0)) | |||
8776 | if (__kmp_mic_type != non_mic) { | |||
8777 | teamsize_cutoff = 8; | |||
8778 | } | |||
8779 | #endif | |||
8780 | int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED; | |||
8781 | if (tree_available) { | |||
8782 | if (team_size <= teamsize_cutoff) { | |||
8783 | if (atomic_available) { | |||
8784 | retval = atomic_reduce_block; | |||
8785 | } | |||
8786 | } else { | |||
8787 | retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER(((tree_reduce_block) | (bs_reduction_barrier))); | |||
8788 | } | |||
8789 | } else if (atomic_available) { | |||
8790 | retval = atomic_reduce_block; | |||
8791 | } | |||
8792 | #else | |||
8793 | #error "Unknown or unsupported OS" | |||
8794 | #endif // KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || | |||
8795 | // KMP_OS_OPENBSD || KMP_OS_WINDOWS || KMP_OS_DARWIN || KMP_OS_HURD | |||
8796 | ||||
8797 | #elif KMP_ARCH_X860 || KMP_ARCH_ARM || KMP_ARCH_AARCH || KMP_ARCH_MIPS0 | |||
8798 | ||||
8799 | #if KMP_OS_LINUX1 || KMP_OS_FREEBSD0 || KMP_OS_WINDOWS0 || KMP_OS_HURD0 | |||
8800 | ||||
8801 | // basic tuning | |||
8802 | ||||
8803 | if (atomic_available) { | |||
8804 | if (num_vars <= 2) { // && ( team_size <= 8 ) due to false-sharing ??? | |||
8805 | retval = atomic_reduce_block; | |||
8806 | } | |||
8807 | } // otherwise: use critical section | |||
8808 | ||||
8809 | #elif KMP_OS_DARWIN0 | |||
8810 | ||||
8811 | int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED; | |||
8812 | if (atomic_available && (num_vars <= 3)) { | |||
8813 | retval = atomic_reduce_block; | |||
8814 | } else if (tree_available) { | |||
8815 | if ((reduce_size > (9 * sizeof(kmp_real64))) && | |||
8816 | (reduce_size < (2000 * sizeof(kmp_real64)))) { | |||
8817 | retval = TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER(((tree_reduce_block) | (bs_plain_barrier))); | |||
8818 | } | |||
8819 | } // otherwise: use critical section | |||
8820 | ||||
8821 | #else | |||
8822 | #error "Unknown or unsupported OS" | |||
8823 | #endif | |||
8824 | ||||
8825 | #else | |||
8826 | #error "Unknown or unsupported architecture" | |||
8827 | #endif | |||
8828 | } | |||
8829 | ||||
8830 | // KMP_FORCE_REDUCTION | |||
8831 | ||||
8832 | // If the team is serialized (team_size == 1), ignore the forced reduction | |||
8833 | // method and stay with the unsynchronized method (empty_reduce_block) | |||
8834 | if (__kmp_force_reduction_method != reduction_method_not_defined && | |||
8835 | team_size != 1) { | |||
8836 | ||||
8837 | PACKED_REDUCTION_METHOD_T forced_retval = critical_reduce_block; | |||
8838 | ||||
8839 | int atomic_available, tree_available; | |||
8840 | ||||
8841 | switch ((forced_retval = __kmp_force_reduction_method)) { | |||
8842 | case critical_reduce_block: | |||
8843 | KMP_ASSERT(lck)if (!(lck)) { __kmp_debug_assert("lck", "openmp/runtime/src/kmp_runtime.cpp" , 8843); }; // lck should be != 0 | |||
8844 | break; | |||
8845 | ||||
8846 | case atomic_reduce_block: | |||
8847 | atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED; | |||
8848 | if (!atomic_available) { | |||
8849 | KMP_WARNING(RedMethodNotSupported, "atomic")__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_RedMethodNotSupported , "atomic"), __kmp_msg_null); | |||
8850 | forced_retval = critical_reduce_block; | |||
8851 | } | |||
8852 | break; | |||
8853 | ||||
8854 | case tree_reduce_block: | |||
8855 | tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED; | |||
8856 | if (!tree_available) { | |||
8857 | KMP_WARNING(RedMethodNotSupported, "tree")__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_RedMethodNotSupported , "tree"), __kmp_msg_null); | |||
8858 | forced_retval = critical_reduce_block; | |||
8859 | } else { | |||
8860 | #if KMP_FAST_REDUCTION_BARRIER1 | |||
8861 | forced_retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER(((tree_reduce_block) | (bs_reduction_barrier))); | |||
8862 | #endif | |||
8863 | } | |||
8864 | break; | |||
8865 | ||||
8866 | default: | |||
8867 | KMP_ASSERT(0)if (!(0)) { __kmp_debug_assert("0", "openmp/runtime/src/kmp_runtime.cpp" , 8867); }; // "unsupported method specified" | |||
8868 | } | |||
8869 | ||||
8870 | retval = forced_retval; | |||
8871 | } | |||
8872 | ||||
8873 | KA_TRACE(10, ("reduction method selected=%08x\n", retval))if (kmp_a_debug >= 10) { __kmp_debug_printf ("reduction method selected=%08x\n" , retval); }; | |||
8874 | ||||
8875 | #undef FAST_REDUCTION_TREE_METHOD_GENERATED | |||
8876 | #undef FAST_REDUCTION_ATOMIC_METHOD_GENERATED | |||
8877 | ||||
8878 | return (retval); | |||
8879 | } | |||
8880 | // this function is for testing set/get/determine reduce method | |||
8881 | kmp_int32 __kmp_get_reduce_method(void) { | |||
8882 | return ((__kmp_entry_thread()->th.th_local.packed_reduction_method) >> 8); | |||
8883 | } | |||
8884 | ||||
8885 | // Soft pause sets up threads to ignore blocktime and just go to sleep. | |||
8886 | // Spin-wait code checks __kmp_pause_status and reacts accordingly. | |||
8887 | void __kmp_soft_pause() { __kmp_pause_status = kmp_soft_paused; } | |||
8888 | ||||
8889 | // Hard pause shuts down the runtime completely. Resume happens naturally when | |||
8890 | // OpenMP is used subsequently. | |||
8891 | void __kmp_hard_pause() { | |||
8892 | __kmp_pause_status = kmp_hard_paused; | |||
8893 | __kmp_internal_end_thread(-1); | |||
8894 | } | |||
8895 | ||||
8896 | // Soft resume sets __kmp_pause_status, and wakes up all threads. | |||
8897 | void __kmp_resume_if_soft_paused() { | |||
8898 | if (__kmp_pause_status == kmp_soft_paused) { | |||
8899 | __kmp_pause_status = kmp_not_paused; | |||
8900 | ||||
8901 | for (int gtid = 1; gtid < __kmp_threads_capacity; ++gtid) { | |||
8902 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
8903 | if (thread) { // Wake it if sleeping | |||
8904 | kmp_flag_64<> fl(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, | |||
8905 | thread); | |||
8906 | if (fl.is_sleeping()) | |||
8907 | fl.resume(gtid); | |||
8908 | else if (__kmp_try_suspend_mx(thread)) { // got suspend lock | |||
8909 | __kmp_unlock_suspend_mx(thread); // unlock it; it won't sleep | |||
8910 | } else { // thread holds the lock and may sleep soon | |||
8911 | do { // until either the thread sleeps, or we can get the lock | |||
8912 | if (fl.is_sleeping()) { | |||
8913 | fl.resume(gtid); | |||
8914 | break; | |||
8915 | } else if (__kmp_try_suspend_mx(thread)) { | |||
8916 | __kmp_unlock_suspend_mx(thread); | |||
8917 | break; | |||
8918 | } | |||
8919 | } while (1); | |||
8920 | } | |||
8921 | } | |||
8922 | } | |||
8923 | } | |||
8924 | } | |||
8925 | ||||
8926 | // This function is called via __kmpc_pause_resource. Returns 0 if successful. | |||
8927 | // TODO: add warning messages | |||
8928 | int __kmp_pause_resource(kmp_pause_status_t level) { | |||
8929 | if (level == kmp_not_paused) { // requesting resume | |||
8930 | if (__kmp_pause_status == kmp_not_paused) { | |||
8931 | // error message about runtime not being paused, so can't resume | |||
8932 | return 1; | |||
8933 | } else { | |||
8934 | 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", 8935); } | |||
8935 | __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", 8935); }; | |||
8936 | __kmp_pause_status = kmp_not_paused; | |||
8937 | return 0; | |||
8938 | } | |||
8939 | } else if (level == kmp_soft_paused) { // requesting soft pause | |||
8940 | if (__kmp_pause_status != kmp_not_paused) { | |||
8941 | // error message about already being paused | |||
8942 | return 1; | |||
8943 | } else { | |||
8944 | __kmp_soft_pause(); | |||
8945 | return 0; | |||
8946 | } | |||
8947 | } else if (level == kmp_hard_paused) { // requesting hard pause | |||
8948 | if (__kmp_pause_status != kmp_not_paused) { | |||
8949 | // error message about already being paused | |||
8950 | return 1; | |||
8951 | } else { | |||
8952 | __kmp_hard_pause(); | |||
8953 | return 0; | |||
8954 | } | |||
8955 | } else { | |||
8956 | // error message about invalid level | |||
8957 | return 1; | |||
8958 | } | |||
8959 | } | |||
8960 | ||||
8961 | void __kmp_omp_display_env(int verbose) { | |||
8962 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); | |||
8963 | if (__kmp_init_serial == 0) | |||
8964 | __kmp_do_serial_initialize(); | |||
8965 | __kmp_display_env_impl(!verbose, verbose); | |||
8966 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); | |||
8967 | } | |||
8968 | ||||
8969 | // The team size is changing, so distributed barrier must be modified | |||
8970 | void __kmp_resize_dist_barrier(kmp_team_t *team, int old_nthreads, | |||
8971 | int new_nthreads) { | |||
8972 | 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", 8973); } | |||
8973 | 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", 8973); }; | |||
8974 | kmp_info_t **other_threads = team->t.t_threads; | |||
8975 | ||||
8976 | // We want all the workers to stop waiting on the barrier while we adjust the | |||
8977 | // size of the team. | |||
8978 | for (int f = 1; f < old_nthreads; ++f) { | |||
8979 | 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", 8979); }; | |||
8980 | // Ignore threads that are already inactive or not present in the team | |||
8981 | if (team->t.t_threads[f]->th.th_used_in_team.load() == 0) { | |||
8982 | // teams construct causes thread_limit to get passed in, and some of | |||
8983 | // those could be inactive; just ignore them | |||
8984 | continue; | |||
8985 | } | |||
8986 | // If thread is transitioning still to in_use state, wait for it | |||
8987 | if (team->t.t_threads[f]->th.th_used_in_team.load() == 3) { | |||
8988 | while (team->t.t_threads[f]->th.th_used_in_team.load() == 3) | |||
8989 | KMP_CPU_PAUSE()__kmp_x86_pause(); | |||
8990 | } | |||
8991 | // The thread should be in_use now | |||
8992 | 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", 8992); }; | |||
8993 | // Transition to unused state | |||
8994 | team->t.t_threads[f]->th.th_used_in_team.store(2); | |||
8995 | 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", 8995); }; | |||
8996 | } | |||
8997 | // Release all the workers | |||
8998 | team->t.b->go_release(); | |||
8999 | ||||
9000 | KMP_MFENCE()if (__builtin_expect(!!(!__kmp_cpuinfo.initialized), 0)) { __kmp_query_cpuid (&__kmp_cpuinfo); } if (__kmp_cpuinfo.flags.sse2) { __sync_synchronize (); }; | |||
9001 | ||||
9002 | // Workers should see transition status 2 and move to 0; but may need to be | |||
9003 | // woken up first | |||
9004 | int count = old_nthreads - 1; | |||
9005 | while (count > 0) { | |||
9006 | count = old_nthreads - 1; | |||
9007 | for (int f = 1; f < old_nthreads; ++f) { | |||
9008 | if (other_threads[f]->th.th_used_in_team.load() != 0) { | |||
9009 | if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME(2147483647)) { // Wake up the workers | |||
9010 | kmp_atomic_flag_64<> *flag = (kmp_atomic_flag_64<> *)CCAST(const_cast<void *>(other_threads[f]->th.th_sleep_loc ) | |||
9011 | void *, other_threads[f]->th.th_sleep_loc)const_cast<void *>(other_threads[f]->th.th_sleep_loc ); | |||
9012 | __kmp_atomic_resume_64(other_threads[f]->th.th_info.ds.ds_gtid, flag); | |||
9013 | } | |||
9014 | } else { | |||
9015 | 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", 9015); }; | |||
9016 | count--; | |||
9017 | } | |||
9018 | } | |||
9019 | } | |||
9020 | // Now update the barrier size | |||
9021 | team->t.b->update_num_threads(new_nthreads); | |||
9022 | team->t.b->go_reset(); | |||
9023 | } | |||
9024 | ||||
9025 | void __kmp_add_threads_to_team(kmp_team_t *team, int new_nthreads) { | |||
9026 | // Add the threads back to the team | |||
9027 | KMP_DEBUG_ASSERT(team)if (!(team)) { __kmp_debug_assert("team", "openmp/runtime/src/kmp_runtime.cpp" , 9027); }; | |||
9028 | // Threads were paused and pointed at th_used_in_team temporarily during a | |||
9029 | // resize of the team. We're going to set th_used_in_team to 3 to indicate to | |||
9030 | // the thread that it should transition itself back into the team. Then, if | |||
9031 | // blocktime isn't infinite, the thread could be sleeping, so we send a resume | |||
9032 | // to wake it up. | |||
9033 | for (int f = 1; f < new_nthreads; ++f) { | |||
9034 | 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", 9034); }; | |||
9035 | 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)) | |||
9036 | 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)); | |||
9037 | if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME(2147483647)) { // Wake up sleeping threads | |||
9038 | __kmp_resume_32(team->t.t_threads[f]->th.th_info.ds.ds_gtid, | |||
9039 | (kmp_flag_32<false, false> *)NULL__null); | |||
9040 | } | |||
9041 | } | |||
9042 | // The threads should be transitioning to the team; when they are done, they | |||
9043 | // should have set th_used_in_team to 1. This loop forces master to wait until | |||
9044 | // all threads have moved into the team and are waiting in the barrier. | |||
9045 | int count = new_nthreads - 1; | |||
9046 | while (count > 0) { | |||
9047 | count = new_nthreads - 1; | |||
9048 | for (int f = 1; f < new_nthreads; ++f) { | |||
9049 | if (team->t.t_threads[f]->th.th_used_in_team.load() == 1) { | |||
9050 | count--; | |||
9051 | } | |||
9052 | } | |||
9053 | } | |||
9054 | } | |||
9055 | ||||
9056 | // Globals and functions for hidden helper task | |||
9057 | kmp_info_t **__kmp_hidden_helper_threads; | |||
9058 | kmp_info_t *__kmp_hidden_helper_main_thread; | |||
9059 | std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks; | |||
9060 | #if KMP_OS_LINUX1 | |||
9061 | kmp_int32 __kmp_hidden_helper_threads_num = 8; | |||
9062 | kmp_int32 __kmp_enable_hidden_helper = TRUE(!0); | |||
9063 | #else | |||
9064 | kmp_int32 __kmp_hidden_helper_threads_num = 0; | |||
9065 | kmp_int32 __kmp_enable_hidden_helper = FALSE0; | |||
9066 | #endif | |||
9067 | ||||
9068 | namespace { | |||
9069 | std::atomic<kmp_int32> __kmp_hit_hidden_helper_threads_num; | |||
9070 | ||||
9071 | void __kmp_hidden_helper_wrapper_fn(int *gtid, int *, ...) { | |||
9072 | // This is an explicit synchronization on all hidden helper threads in case | |||
9073 | // that when a regular thread pushes a hidden helper task to one hidden | |||
9074 | // helper thread, the thread has not been awaken once since they're released | |||
9075 | // by the main thread after creating the team. | |||
9076 | KMP_ATOMIC_INC(&__kmp_hit_hidden_helper_threads_num)(&__kmp_hit_hidden_helper_threads_num)->fetch_add(1, std ::memory_order_acq_rel); | |||
9077 | while (KMP_ATOMIC_LD_ACQ(&__kmp_hit_hidden_helper_threads_num)(&__kmp_hit_hidden_helper_threads_num)->load(std::memory_order_acquire ) != | |||
9078 | __kmp_hidden_helper_threads_num) | |||
9079 | ; | |||
9080 | ||||
9081 | // If main thread, then wait for signal | |||
9082 | if (__kmpc_master(nullptr, *gtid)) { | |||
9083 | // First, unset the initial state and release the initial thread | |||
9084 | TCW_4(__kmp_init_hidden_helper_threads, FALSE)(__kmp_init_hidden_helper_threads) = (0); | |||
9085 | __kmp_hidden_helper_initz_release(); | |||
9086 | __kmp_hidden_helper_main_thread_wait(); | |||
9087 | // Now wake up all worker threads | |||
9088 | for (int i = 1; i < __kmp_hit_hidden_helper_threads_num; ++i) { | |||
9089 | __kmp_hidden_helper_worker_thread_signal(); | |||
9090 | } | |||
9091 | } | |||
9092 | } | |||
9093 | } // namespace | |||
9094 | ||||
9095 | void __kmp_hidden_helper_threads_initz_routine() { | |||
9096 | // Create a new root for hidden helper team/threads | |||
9097 | const int gtid = __kmp_register_root(TRUE(!0)); | |||
9098 | __kmp_hidden_helper_main_thread = __kmp_threads[gtid]; | |||
9099 | __kmp_hidden_helper_threads = &__kmp_threads[gtid]; | |||
9100 | __kmp_hidden_helper_main_thread->th.th_set_nproc = | |||
9101 | __kmp_hidden_helper_threads_num; | |||
9102 | ||||
9103 | KMP_ATOMIC_ST_REL(&__kmp_hit_hidden_helper_threads_num, 0)(&__kmp_hit_hidden_helper_threads_num)->store(0, std:: memory_order_release); | |||
9104 | ||||
9105 | __kmpc_fork_call(nullptr, 0, __kmp_hidden_helper_wrapper_fn); | |||
9106 | ||||
9107 | // Set the initialization flag to FALSE | |||
9108 | TCW_SYNC_4(__kmp_init_hidden_helper, FALSE)(__kmp_init_hidden_helper) = (0); | |||
9109 | ||||
9110 | __kmp_hidden_helper_threads_deinitz_release(); | |||
9111 | } | |||
9112 | ||||
9113 | /* Nesting Mode: | |||
9114 | Set via KMP_NESTING_MODE, which takes an integer. | |||
9115 | Note: we skip duplicate topology levels, and skip levels with only | |||
9116 | one entity. | |||
9117 | KMP_NESTING_MODE=0 is the default, and doesn't use nesting mode. | |||
9118 | KMP_NESTING_MODE=1 sets as many nesting levels as there are distinct levels | |||
9119 | in the topology, and initializes the number of threads at each of those | |||
9120 | levels to the number of entities at each level, respectively, below the | |||
9121 | entity at the parent level. | |||
9122 | KMP_NESTING_MODE=N, where N>1, attempts to create up to N nesting levels, | |||
9123 | but starts with nesting OFF -- max-active-levels-var is 1 -- and requires | |||
9124 | the user to turn nesting on explicitly. This is an even more experimental | |||
9125 | option to this experimental feature, and may change or go away in the | |||
9126 | future. | |||
9127 | */ | |||
9128 | ||||
9129 | // Allocate space to store nesting levels | |||
9130 | void __kmp_init_nesting_mode() { | |||
9131 | int levels = KMP_HW_LAST; | |||
9132 | __kmp_nesting_mode_nlevels = levels; | |||
9133 | __kmp_nesting_nth_level = (int *)KMP_INTERNAL_MALLOC(levels * sizeof(int))malloc(levels * sizeof(int)); | |||
9134 | for (int i = 0; i < levels; ++i) | |||
9135 | __kmp_nesting_nth_level[i] = 0; | |||
9136 | if (__kmp_nested_nth.size < levels) { | |||
9137 | __kmp_nested_nth.nth = | |||
9138 | (int *)KMP_INTERNAL_REALLOC(__kmp_nested_nth.nth, levels * sizeof(int))realloc((__kmp_nested_nth.nth), (levels * sizeof(int))); | |||
9139 | __kmp_nested_nth.size = levels; | |||
9140 | } | |||
9141 | } | |||
9142 | ||||
9143 | // Set # threads for top levels of nesting; must be called after topology set | |||
9144 | void __kmp_set_nesting_mode_threads() { | |||
9145 | kmp_info_t *thread = __kmp_threads[__kmp_entry_gtid()__kmp_get_global_thread_id_reg()]; | |||
9146 | ||||
9147 | if (__kmp_nesting_mode == 1) | |||
9148 | __kmp_nesting_mode_nlevels = KMP_MAX_ACTIVE_LEVELS_LIMIT2147483647; | |||
9149 | else if (__kmp_nesting_mode > 1) | |||
9150 | __kmp_nesting_mode_nlevels = __kmp_nesting_mode; | |||
9151 | ||||
9152 | if (__kmp_topology) { // use topology info | |||
9153 | int loc, hw_level; | |||
9154 | for (loc = 0, hw_level = 0; hw_level < __kmp_topology->get_depth() && | |||
9155 | loc < __kmp_nesting_mode_nlevels; | |||
9156 | loc++, hw_level++) { | |||
9157 | __kmp_nesting_nth_level[loc] = __kmp_topology->get_ratio(hw_level); | |||
9158 | if (__kmp_nesting_nth_level[loc] == 1) | |||
9159 | loc--; | |||
9160 | } | |||
9161 | // Make sure all cores are used | |||
9162 | if (__kmp_nesting_mode > 1 && loc > 1) { | |||
9163 | int core_level = __kmp_topology->get_level(KMP_HW_CORE); | |||
9164 | int num_cores = __kmp_topology->get_count(core_level); | |||
9165 | int upper_levels = 1; | |||
9166 | for (int level = 0; level < loc - 1; ++level) | |||
9167 | upper_levels *= __kmp_nesting_nth_level[level]; | |||
9168 | if (upper_levels * __kmp_nesting_nth_level[loc - 1] < num_cores) | |||
9169 | __kmp_nesting_nth_level[loc - 1] = | |||
9170 | num_cores / __kmp_nesting_nth_level[loc - 2]; | |||
9171 | } | |||
9172 | __kmp_nesting_mode_nlevels = loc; | |||
9173 | __kmp_nested_nth.used = __kmp_nesting_mode_nlevels; | |||
9174 | } else { // no topology info available; provide a reasonable guesstimation | |||
9175 | if (__kmp_avail_proc >= 4) { | |||
9176 | __kmp_nesting_nth_level[0] = __kmp_avail_proc / 2; | |||
9177 | __kmp_nesting_nth_level[1] = 2; | |||
9178 | __kmp_nesting_mode_nlevels = 2; | |||
9179 | } else { | |||
9180 | __kmp_nesting_nth_level[0] = __kmp_avail_proc; | |||
9181 | __kmp_nesting_mode_nlevels = 1; | |||
9182 | } | |||
9183 | __kmp_nested_nth.used = __kmp_nesting_mode_nlevels; | |||
9184 | } | |||
9185 | for (int i = 0; i < __kmp_nesting_mode_nlevels; ++i) { | |||
9186 | __kmp_nested_nth.nth[i] = __kmp_nesting_nth_level[i]; | |||
9187 | } | |||
9188 | set__nproc(thread, __kmp_nesting_nth_level[0])(((thread)->th.th_current_task->td_icvs.nproc) = (__kmp_nesting_nth_level [0])); | |||
9189 | if (__kmp_nesting_mode > 1 && __kmp_nesting_mode_nlevels > __kmp_nesting_mode) | |||
9190 | __kmp_nesting_mode_nlevels = __kmp_nesting_mode; | |||
9191 | if (get__max_active_levels(thread)((thread)->th.th_current_task->td_icvs.max_active_levels ) > 1) { | |||
9192 | // if max levels was set, set nesting mode levels to same | |||
9193 | __kmp_nesting_mode_nlevels = get__max_active_levels(thread)((thread)->th.th_current_task->td_icvs.max_active_levels ); | |||
9194 | } | |||
9195 | if (__kmp_nesting_mode == 1) // turn on nesting for this case only | |||
9196 | set__max_active_levels(thread, __kmp_nesting_mode_nlevels)(((thread)->th.th_current_task->td_icvs.max_active_levels ) = (__kmp_nesting_mode_nlevels)); | |||
9197 | } | |||
9198 | ||||
9199 | // Empty symbols to export (see exports_so.txt) when feature is disabled | |||
9200 | extern "C" { | |||
9201 | #if !KMP_STATS_ENABLED0 | |||
9202 | void __kmp_reset_stats() {} | |||
9203 | #endif | |||
9204 | #if !USE_DEBUGGER0 | |||
9205 | int __kmp_omp_debug_struct_info = FALSE0; | |||
9206 | int __kmp_debugging = FALSE0; | |||
9207 | #endif | |||
9208 | #if !USE_ITT_BUILD1 || !USE_ITT_NOTIFY1 | |||
9209 | void __kmp_itt_fini_ittlib() {} | |||
9210 | void __kmp_itt_init_ittlib() {} | |||
9211 | #endif | |||
9212 | } | |||
9213 | ||||
9214 | // end of file |
1 | /*! \file */ |
2 | /* |
3 | * kmp.h -- KPTS runtime header file. |
4 | */ |
5 | |
6 | //===----------------------------------------------------------------------===// |
7 | // |
8 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
9 | // See https://llvm.org/LICENSE.txt for license information. |
10 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #ifndef KMP_H |
15 | #define KMP_H |
16 | |
17 | #include "kmp_config.h" |
18 | |
19 | /* #define BUILD_PARALLEL_ORDERED 1 */ |
20 | |
21 | /* This fix replaces gettimeofday with clock_gettime for better scalability on |
22 | the Altix. Requires user code to be linked with -lrt. */ |
23 | //#define FIX_SGI_CLOCK |
24 | |
25 | /* Defines for OpenMP 3.0 tasking and auto scheduling */ |
26 | |
27 | #ifndef KMP_STATIC_STEAL_ENABLED1 |
28 | #define KMP_STATIC_STEAL_ENABLED1 1 |
29 | #endif |
30 | |
31 | #define TASK_CURRENT_NOT_QUEUED0 0 |
32 | #define TASK_CURRENT_QUEUED1 1 |
33 | |
34 | #ifdef BUILD_TIED_TASK_STACK |
35 | #define TASK_STACK_EMPTY 0 // entries when the stack is empty |
36 | #define TASK_STACK_BLOCK_BITS 5 // Used in TASK_STACK_SIZE and TASK_STACK_MASK |
37 | // Number of entries in each task stack array |
38 | #define TASK_STACK_BLOCK_SIZE (1 << TASK_STACK_BLOCK_BITS) |
39 | // Mask for determining index into stack block |
40 | #define TASK_STACK_INDEX_MASK (TASK_STACK_BLOCK_SIZE - 1) |
41 | #endif // BUILD_TIED_TASK_STACK |
42 | |
43 | #define TASK_NOT_PUSHED1 1 |
44 | #define TASK_SUCCESSFULLY_PUSHED0 0 |
45 | #define TASK_TIED1 1 |
46 | #define TASK_UNTIED0 0 |
47 | #define TASK_EXPLICIT1 1 |
48 | #define TASK_IMPLICIT0 0 |
49 | #define TASK_PROXY1 1 |
50 | #define TASK_FULL0 0 |
51 | #define TASK_DETACHABLE1 1 |
52 | #define TASK_UNDETACHABLE0 0 |
53 | |
54 | #define KMP_CANCEL_THREADS |
55 | #define KMP_THREAD_ATTR |
56 | |
57 | // Android does not have pthread_cancel. Undefine KMP_CANCEL_THREADS if being |
58 | // built on Android |
59 | #if defined(__ANDROID__) |
60 | #undef KMP_CANCEL_THREADS |
61 | #endif |
62 | |
63 | #include <signal.h> |
64 | #include <stdarg.h> |
65 | #include <stddef.h> |
66 | #include <stdio.h> |
67 | #include <stdlib.h> |
68 | #include <string.h> |
69 | #include <limits> |
70 | #include <type_traits> |
71 | /* include <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad |
72 | Microsoft library. Some macros provided below to replace these functions */ |
73 | #ifndef __ABSOFT_WIN |
74 | #include <sys/types.h> |
75 | #endif |
76 | #include <limits.h> |
77 | #include <time.h> |
78 | |
79 | #include <errno(*__errno_location ()).h> |
80 | |
81 | #include "kmp_os.h" |
82 | |
83 | #include "kmp_safe_c_api.h" |
84 | |
85 | #if KMP_STATS_ENABLED0 |
86 | class kmp_stats_list; |
87 | #endif |
88 | |
89 | #if KMP_USE_HIER_SCHED0 |
90 | // Only include hierarchical scheduling if affinity is supported |
91 | #undef KMP_USE_HIER_SCHED0 |
92 | #define KMP_USE_HIER_SCHED0 KMP_AFFINITY_SUPPORTED1 |
93 | #endif |
94 | |
95 | #if KMP_USE_HWLOC0 && KMP_AFFINITY_SUPPORTED1 |
96 | #include "hwloc.h" |
97 | #ifndef HWLOC_OBJ_NUMANODE |
98 | #define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE |
99 | #endif |
100 | #ifndef HWLOC_OBJ_PACKAGE |
101 | #define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET |
102 | #endif |
103 | #endif |
104 | |
105 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 |
106 | #include <xmmintrin.h> |
107 | #endif |
108 | |
109 | // The below has to be defined before including "kmp_barrier.h". |
110 | #define KMP_INTERNAL_MALLOC(sz)malloc(sz) malloc(sz) |
111 | #define KMP_INTERNAL_FREE(p)free(p) free(p) |
112 | #define KMP_INTERNAL_REALLOC(p, sz)realloc((p), (sz)) realloc((p), (sz)) |
113 | #define KMP_INTERNAL_CALLOC(n, sz)calloc((n), (sz)) calloc((n), (sz)) |
114 | |
115 | #include "kmp_debug.h" |
116 | #include "kmp_lock.h" |
117 | #include "kmp_version.h" |
118 | #include "kmp_barrier.h" |
119 | #if USE_DEBUGGER0 |
120 | #include "kmp_debugger.h" |
121 | #endif |
122 | #include "kmp_i18n.h" |
123 | |
124 | #define KMP_HANDLE_SIGNALS(1 || 0) (KMP_OS_UNIX1 || KMP_OS_WINDOWS0) |
125 | |
126 | #include "kmp_wrapper_malloc.h" |
127 | #if KMP_OS_UNIX1 |
128 | #include <unistd.h> |
129 | #if !defined NSIG(64 + 1) && defined _NSIG(64 + 1) |
130 | #define NSIG(64 + 1) _NSIG(64 + 1) |
131 | #endif |
132 | #endif |
133 | |
134 | #if KMP_OS_LINUX1 |
135 | #pragma weak clock_gettime |
136 | #endif |
137 | |
138 | #if OMPT_SUPPORT1 |
139 | #include "ompt-internal.h" |
140 | #endif |
141 | |
142 | #if OMPD_SUPPORT1 |
143 | #include "ompd-specific.h" |
144 | #endif |
145 | |
146 | #ifndef UNLIKELY |
147 | #define UNLIKELY(x)__builtin_expect(!!(x), 0) (x) |
148 | #endif |
149 | |
150 | // Affinity format function |
151 | #include "kmp_str.h" |
152 | |
153 | // 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64. |
154 | // 3 - fast allocation using sync, non-sync free lists of any size, non-self |
155 | // free lists of limited size. |
156 | #ifndef USE_FAST_MEMORY3 |
157 | #define USE_FAST_MEMORY3 3 |
158 | #endif |
159 | |
160 | #ifndef KMP_NESTED_HOT_TEAMS1 |
161 | #define KMP_NESTED_HOT_TEAMS1 0 |
162 | #define USE_NESTED_HOT_ARG(x), x |
163 | #else |
164 | #if KMP_NESTED_HOT_TEAMS1 |
165 | #define USE_NESTED_HOT_ARG(x), x , x |
166 | #else |
167 | #define USE_NESTED_HOT_ARG(x), x |
168 | #endif |
169 | #endif |
170 | |
171 | // Assume using BGET compare_exchange instruction instead of lock by default. |
172 | #ifndef USE_CMP_XCHG_FOR_BGET1 |
173 | #define USE_CMP_XCHG_FOR_BGET1 1 |
174 | #endif |
175 | |
176 | // Test to see if queuing lock is better than bootstrap lock for bget |
177 | // #ifndef USE_QUEUING_LOCK_FOR_BGET |
178 | // #define USE_QUEUING_LOCK_FOR_BGET |
179 | // #endif |
180 | |
181 | #define KMP_NSEC_PER_SEC1000000000L 1000000000L |
182 | #define KMP_USEC_PER_SEC1000000L 1000000L |
183 | |
184 | /*! |
185 | @ingroup BASIC_TYPES |
186 | @{ |
187 | */ |
188 | |
189 | /*! |
190 | Values for bit flags used in the ident_t to describe the fields. |
191 | */ |
192 | enum { |
193 | /*! Use trampoline for internal microtasks */ |
194 | KMP_IDENT_IMB = 0x01, |
195 | /*! Use c-style ident structure */ |
196 | KMP_IDENT_KMPC = 0x02, |
197 | /* 0x04 is no longer used */ |
198 | /*! Entry point generated by auto-parallelization */ |
199 | KMP_IDENT_AUTOPAR = 0x08, |
200 | /*! Compiler generates atomic reduction option for kmpc_reduce* */ |
201 | KMP_IDENT_ATOMIC_REDUCE = 0x10, |
202 | /*! To mark a 'barrier' directive in user code */ |
203 | KMP_IDENT_BARRIER_EXPL = 0x20, |
204 | /*! To Mark implicit barriers. */ |
205 | KMP_IDENT_BARRIER_IMPL = 0x0040, |
206 | KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0, |
207 | KMP_IDENT_BARRIER_IMPL_FOR = 0x0040, |
208 | KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0, |
209 | |
210 | KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140, |
211 | KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0, |
212 | |
213 | /*! To mark a static loop in OMPT callbacks */ |
214 | KMP_IDENT_WORK_LOOP = 0x200, |
215 | /*! To mark a sections directive in OMPT callbacks */ |
216 | KMP_IDENT_WORK_SECTIONS = 0x400, |
217 | /*! To mark a distribute construct in OMPT callbacks */ |
218 | KMP_IDENT_WORK_DISTRIBUTE = 0x800, |
219 | /*! Atomic hint; bottom four bits as omp_sync_hint_t. Top four reserved and |
220 | not currently used. If one day we need more bits, then we can use |
221 | an invalid combination of hints to mean that another, larger field |
222 | should be used in a different flag. */ |
223 | KMP_IDENT_ATOMIC_HINT_MASK = 0xFF0000, |
224 | KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000, |
225 | KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000, |
226 | KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000, |
227 | KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000, |
228 | KMP_IDENT_OPENMP_SPEC_VERSION_MASK = 0xFF000000 |
229 | }; |
230 | |
231 | /*! |
232 | * The ident structure that describes a source location. |
233 | */ |
234 | typedef struct ident { |
235 | kmp_int32 reserved_1; /**< might be used in Fortran; see above */ |
236 | kmp_int32 flags; /**< also f.flags; KMP_IDENT_xxx flags; KMP_IDENT_KMPC |
237 | identifies this union member */ |
238 | kmp_int32 reserved_2; /**< not really used in Fortran any more; see above */ |
239 | #if USE_ITT_BUILD1 |
240 | /* but currently used for storing region-specific ITT */ |
241 | /* contextual information. */ |
242 | #endif /* USE_ITT_BUILD */ |
243 | kmp_int32 reserved_3; /**< source[4] in Fortran, do not use for C++ */ |
244 | char const *psource; /**< String describing the source location. |
245 | The string is composed of semi-colon separated fields |
246 | which describe the source file, the function and a pair |
247 | of line numbers that delimit the construct. */ |
248 | // Returns the OpenMP version in form major*10+minor (e.g., 50 for 5.0) |
249 | kmp_int32 get_openmp_version() { |
250 | return (((flags & KMP_IDENT_OPENMP_SPEC_VERSION_MASK) >> 24) & 0xFF); |
251 | } |
252 | } ident_t; |
253 | /*! |
254 | @} |
255 | */ |
256 | |
257 | // Some forward declarations. |
258 | typedef union kmp_team kmp_team_t; |
259 | typedef struct kmp_taskdata kmp_taskdata_t; |
260 | typedef union kmp_task_team kmp_task_team_t; |
261 | typedef union kmp_team kmp_team_p; |
262 | typedef union kmp_info kmp_info_p; |
263 | typedef union kmp_root kmp_root_p; |
264 | |
265 | template <bool C = false, bool S = true> class kmp_flag_32; |
266 | template <bool C = false, bool S = true> class kmp_flag_64; |
267 | template <bool C = false, bool S = true> class kmp_atomic_flag_64; |
268 | class kmp_flag_oncore; |
269 | |
270 | #ifdef __cplusplus201703L |
271 | extern "C" { |
272 | #endif |
273 | |
274 | /* ------------------------------------------------------------------------ */ |
275 | |
276 | /* Pack two 32-bit signed integers into a 64-bit signed integer */ |
277 | /* ToDo: Fix word ordering for big-endian machines. */ |
278 | #define KMP_PACK_64(HIGH_32, LOW_32)((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64 )(LOW_32))) \ |
279 | ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32))) |
280 | |
281 | // Generic string manipulation macros. Assume that _x is of type char * |
282 | #define SKIP_WS(_x){ while (*(_x) == ' ' || *(_x) == '\t') (_x)++; } \ |
283 | { \ |
284 | while (*(_x) == ' ' || *(_x) == '\t') \ |
285 | (_x)++; \ |
286 | } |
287 | #define SKIP_DIGITS(_x){ while (*(_x) >= '0' && *(_x) <= '9') (_x)++; } \ |
288 | { \ |
289 | while (*(_x) >= '0' && *(_x) <= '9') \ |
290 | (_x)++; \ |
291 | } |
292 | #define SKIP_TOKEN(_x){ while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x ) >= 'a' && *(_x) <= 'z') || (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') (_x)++; } \ |
293 | { \ |
294 | while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \ |
295 | (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \ |
296 | (_x)++; \ |
297 | } |
298 | #define SKIP_TO(_x, _c){ while (*(_x) != '\0' && *(_x) != (_c)) (_x)++; } \ |
299 | { \ |
300 | while (*(_x) != '\0' && *(_x) != (_c)) \ |
301 | (_x)++; \ |
302 | } |
303 | |
304 | /* ------------------------------------------------------------------------ */ |
305 | |
306 | #define KMP_MAX(x, y)((x) > (y) ? (x) : (y)) ((x) > (y) ? (x) : (y)) |
307 | #define KMP_MIN(x, y)((x) < (y) ? (x) : (y)) ((x) < (y) ? (x) : (y)) |
308 | |
309 | /* ------------------------------------------------------------------------ */ |
310 | /* Enumeration types */ |
311 | |
312 | enum kmp_state_timer { |
313 | ts_stop, |
314 | ts_start, |
315 | ts_pause, |
316 | |
317 | ts_last_state |
318 | }; |
319 | |
320 | enum dynamic_mode { |
321 | dynamic_default, |
322 | #ifdef USE_LOAD_BALANCE1 |
323 | dynamic_load_balance, |
324 | #endif /* USE_LOAD_BALANCE */ |
325 | dynamic_random, |
326 | dynamic_thread_limit, |
327 | dynamic_max |
328 | }; |
329 | |
330 | /* external schedule constants, duplicate enum omp_sched in omp.h in order to |
331 | * not include it here */ |
332 | #ifndef KMP_SCHED_TYPE_DEFINED |
333 | #define KMP_SCHED_TYPE_DEFINED |
334 | typedef enum kmp_sched { |
335 | kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check |
336 | // Note: need to adjust __kmp_sch_map global array in case enum is changed |
337 | kmp_sched_static = 1, // mapped to kmp_sch_static_chunked (33) |
338 | kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked (35) |
339 | kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked (36) |
340 | kmp_sched_auto = 4, // mapped to kmp_sch_auto (38) |
341 | kmp_sched_upper_std = 5, // upper bound for standard schedules |
342 | kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules |
343 | kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39) |
344 | #if KMP_STATIC_STEAL_ENABLED1 |
345 | kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44) |
346 | #endif |
347 | kmp_sched_upper, |
348 | kmp_sched_default = kmp_sched_static, // default scheduling |
349 | kmp_sched_monotonic = 0x80000000 |
350 | } kmp_sched_t; |
351 | #endif |
352 | |
353 | /*! |
354 | @ingroup WORK_SHARING |
355 | * Describes the loop schedule to be used for a parallel for loop. |
356 | */ |
357 | enum sched_type : kmp_int32 { |
358 | kmp_sch_lower = 32, /**< lower bound for unordered values */ |
359 | kmp_sch_static_chunked = 33, |
360 | kmp_sch_static = 34, /**< static unspecialized */ |
361 | kmp_sch_dynamic_chunked = 35, |
362 | kmp_sch_guided_chunked = 36, /**< guided unspecialized */ |
363 | kmp_sch_runtime = 37, |
364 | kmp_sch_auto = 38, /**< auto */ |
365 | kmp_sch_trapezoidal = 39, |
366 | |
367 | /* accessible only through KMP_SCHEDULE environment variable */ |
368 | kmp_sch_static_greedy = 40, |
369 | kmp_sch_static_balanced = 41, |
370 | /* accessible only through KMP_SCHEDULE environment variable */ |
371 | kmp_sch_guided_iterative_chunked = 42, |
372 | kmp_sch_guided_analytical_chunked = 43, |
373 | /* accessible only through KMP_SCHEDULE environment variable */ |
374 | kmp_sch_static_steal = 44, |
375 | |
376 | /* static with chunk adjustment (e.g., simd) */ |
377 | kmp_sch_static_balanced_chunked = 45, |
378 | kmp_sch_guided_simd = 46, /**< guided with chunk adjustment */ |
379 | kmp_sch_runtime_simd = 47, /**< runtime with chunk adjustment */ |
380 | |
381 | /* accessible only through KMP_SCHEDULE environment variable */ |
382 | kmp_sch_upper, /**< upper bound for unordered values */ |
383 | |
384 | kmp_ord_lower = 64, /**< lower bound for ordered values, must be power of 2 */ |
385 | kmp_ord_static_chunked = 65, |
386 | kmp_ord_static = 66, /**< ordered static unspecialized */ |
387 | kmp_ord_dynamic_chunked = 67, |
388 | kmp_ord_guided_chunked = 68, |
389 | kmp_ord_runtime = 69, |
390 | kmp_ord_auto = 70, /**< ordered auto */ |
391 | kmp_ord_trapezoidal = 71, |
392 | kmp_ord_upper, /**< upper bound for ordered values */ |
393 | |
394 | /* Schedules for Distribute construct */ |
395 | kmp_distribute_static_chunked = 91, /**< distribute static chunked */ |
396 | kmp_distribute_static = 92, /**< distribute static unspecialized */ |
397 | |
398 | /* For the "nomerge" versions, kmp_dispatch_next*() will always return a |
399 | single iteration/chunk, even if the loop is serialized. For the schedule |
400 | types listed above, the entire iteration vector is returned if the loop is |
401 | serialized. This doesn't work for gcc/gcomp sections. */ |
402 | kmp_nm_lower = 160, /**< lower bound for nomerge values */ |
403 | |
404 | kmp_nm_static_chunked = |
405 | (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower), |
406 | kmp_nm_static = 162, /**< static unspecialized */ |
407 | kmp_nm_dynamic_chunked = 163, |
408 | kmp_nm_guided_chunked = 164, /**< guided unspecialized */ |
409 | kmp_nm_runtime = 165, |
410 | kmp_nm_auto = 166, /**< auto */ |
411 | kmp_nm_trapezoidal = 167, |
412 | |
413 | /* accessible only through KMP_SCHEDULE environment variable */ |
414 | kmp_nm_static_greedy = 168, |
415 | kmp_nm_static_balanced = 169, |
416 | /* accessible only through KMP_SCHEDULE environment variable */ |
417 | kmp_nm_guided_iterative_chunked = 170, |
418 | kmp_nm_guided_analytical_chunked = 171, |
419 | kmp_nm_static_steal = |
420 | 172, /* accessible only through OMP_SCHEDULE environment variable */ |
421 | |
422 | kmp_nm_ord_static_chunked = 193, |
423 | kmp_nm_ord_static = 194, /**< ordered static unspecialized */ |
424 | kmp_nm_ord_dynamic_chunked = 195, |
425 | kmp_nm_ord_guided_chunked = 196, |
426 | kmp_nm_ord_runtime = 197, |
427 | kmp_nm_ord_auto = 198, /**< auto */ |
428 | kmp_nm_ord_trapezoidal = 199, |
429 | kmp_nm_upper, /**< upper bound for nomerge values */ |
430 | |
431 | /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since |
432 | we need to distinguish the three possible cases (no modifier, monotonic |
433 | modifier, nonmonotonic modifier), we need separate bits for each modifier. |
434 | The absence of monotonic does not imply nonmonotonic, especially since 4.5 |
435 | says that the behaviour of the "no modifier" case is implementation defined |
436 | in 4.5, but will become "nonmonotonic" in 5.0. |
437 | |
438 | Since we're passing a full 32 bit value, we can use a couple of high bits |
439 | for these flags; out of paranoia we avoid the sign bit. |
440 | |
441 | These modifiers can be or-ed into non-static schedules by the compiler to |
442 | pass the additional information. They will be stripped early in the |
443 | processing in __kmp_dispatch_init when setting up schedules, so most of the |
444 | code won't ever see schedules with these bits set. */ |
445 | kmp_sch_modifier_monotonic = |
446 | (1 << 29), /**< Set if the monotonic schedule modifier was present */ |
447 | kmp_sch_modifier_nonmonotonic = |
448 | (1 << 30), /**< Set if the nonmonotonic schedule modifier was present */ |
449 | |
450 | #define SCHEDULE_WITHOUT_MODIFIERS(s)(enum sched_type)( (s) & ~(kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) \ |
451 | (enum sched_type)( \ |
452 | (s) & ~(kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) |
453 | #define SCHEDULE_HAS_MONOTONIC(s)(((s)&kmp_sch_modifier_monotonic) != 0) (((s)&kmp_sch_modifier_monotonic) != 0) |
454 | #define SCHEDULE_HAS_NONMONOTONIC(s)(((s)&kmp_sch_modifier_nonmonotonic) != 0) (((s)&kmp_sch_modifier_nonmonotonic) != 0) |
455 | #define SCHEDULE_HAS_NO_MODIFIERS(s)(((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic )) == 0) \ |
456 | (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0) |
457 | #define SCHEDULE_GET_MODIFIERS(s)((enum sched_type)( (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic))) \ |
458 | ((enum sched_type)( \ |
459 | (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic))) |
460 | #define SCHEDULE_SET_MODIFIERS(s, m)(s = (enum sched_type)((kmp_int32)s | (kmp_int32)m)) \ |
461 | (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m)) |
462 | #define SCHEDULE_NONMONOTONIC0 0 |
463 | #define SCHEDULE_MONOTONIC1 1 |
464 | |
465 | kmp_sch_default = kmp_sch_static /**< default scheduling algorithm */ |
466 | }; |
467 | |
468 | // Apply modifiers on internal kind to standard kind |
469 | static inline void |
470 | __kmp_sched_apply_mods_stdkind(kmp_sched_t *kind, |
471 | enum sched_type internal_kind) { |
472 | if (SCHEDULE_HAS_MONOTONIC(internal_kind)(((internal_kind)&kmp_sch_modifier_monotonic) != 0)) { |
473 | *kind = (kmp_sched_t)((int)*kind | (int)kmp_sched_monotonic); |
474 | } |
475 | } |
476 | |
477 | // Apply modifiers on standard kind to internal kind |
478 | static inline void |
479 | __kmp_sched_apply_mods_intkind(kmp_sched_t kind, |
480 | enum sched_type *internal_kind) { |
481 | if ((int)kind & (int)kmp_sched_monotonic) { |
482 | *internal_kind = (enum sched_type)((int)*internal_kind | |
483 | (int)kmp_sch_modifier_monotonic); |
484 | } |
485 | } |
486 | |
487 | // Get standard schedule without modifiers |
488 | static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) { |
489 | return (kmp_sched_t)((int)kind & ~((int)kmp_sched_monotonic)); |
490 | } |
491 | |
492 | /* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */ |
493 | typedef union kmp_r_sched { |
494 | struct { |
495 | enum sched_type r_sched_type; |
496 | int chunk; |
497 | }; |
498 | kmp_int64 sched; |
499 | } kmp_r_sched_t; |
500 | |
501 | extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our |
502 | // internal schedule types |
503 | |
504 | enum library_type { |
505 | library_none, |
506 | library_serial, |
507 | library_turnaround, |
508 | library_throughput |
509 | }; |
510 | |
511 | #if KMP_OS_LINUX1 |
512 | enum clock_function_type { |
513 | clock_function_gettimeofday, |
514 | clock_function_clock_gettime |
515 | }; |
516 | #endif /* KMP_OS_LINUX */ |
517 | |
518 | #if KMP_MIC_SUPPORTED((0 || 1) && (1 || 0)) |
519 | enum mic_type { non_mic, mic1, mic2, mic3, dummy }; |
520 | #endif |
521 | |
522 | /* -- fast reduction stuff ------------------------------------------------ */ |
523 | |
524 | #undef KMP_FAST_REDUCTION_BARRIER1 |
525 | #define KMP_FAST_REDUCTION_BARRIER1 1 |
526 | |
527 | #undef KMP_FAST_REDUCTION_CORE_DUO1 |
528 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 |
529 | #define KMP_FAST_REDUCTION_CORE_DUO1 1 |
530 | #endif |
531 | |
532 | enum _reduction_method { |
533 | reduction_method_not_defined = 0, |
534 | critical_reduce_block = (1 << 8), |
535 | atomic_reduce_block = (2 << 8), |
536 | tree_reduce_block = (3 << 8), |
537 | empty_reduce_block = (4 << 8) |
538 | }; |
539 | |
540 | // Description of the packed_reduction_method variable: |
541 | // The packed_reduction_method variable consists of two enum types variables |
542 | // that are packed together into 0-th byte and 1-st byte: |
543 | // 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of |
544 | // barrier that will be used in fast reduction: bs_plain_barrier or |
545 | // bs_reduction_barrier |
546 | // 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will |
547 | // be used in fast reduction; |
548 | // Reduction method is of 'enum _reduction_method' type and it's defined the way |
549 | // so that the bits of 0-th byte are empty, so no need to execute a shift |
550 | // instruction while packing/unpacking |
551 | |
552 | #if KMP_FAST_REDUCTION_BARRIER1 |
553 | #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type)((reduction_method) | (barrier_type)) \ |
554 | ((reduction_method) | (barrier_type)) |
555 | |
556 | #define UNPACK_REDUCTION_METHOD(packed_reduction_method)((enum _reduction_method)((packed_reduction_method) & (0x0000FF00 ))) \ |
557 | ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00))) |
558 | |
559 | #define UNPACK_REDUCTION_BARRIER(packed_reduction_method)((enum barrier_type)((packed_reduction_method) & (0x000000FF ))) \ |
560 | ((enum barrier_type)((packed_reduction_method) & (0x000000FF))) |
561 | #else |
562 | #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type)((reduction_method) | (barrier_type)) \ |
563 | (reduction_method) |
564 | |
565 | #define UNPACK_REDUCTION_METHOD(packed_reduction_method)((enum _reduction_method)((packed_reduction_method) & (0x0000FF00 ))) \ |
566 | (packed_reduction_method) |
567 | |
568 | #define UNPACK_REDUCTION_BARRIER(packed_reduction_method)((enum barrier_type)((packed_reduction_method) & (0x000000FF ))) (bs_plain_barrier) |
569 | #endif |
570 | |
571 | #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block)((((enum _reduction_method)((packed_reduction_method) & ( 0x0000FF00)))) == (which_reduction_block)) \ |
572 | ((UNPACK_REDUCTION_METHOD(packed_reduction_method)((enum _reduction_method)((packed_reduction_method) & (0x0000FF00 )))) == \ |
573 | (which_reduction_block)) |
574 | |
575 | #if KMP_FAST_REDUCTION_BARRIER1 |
576 | #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER(((tree_reduce_block) | (bs_reduction_barrier))) \ |
577 | (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier)((tree_reduce_block) | (bs_reduction_barrier))) |
578 | |
579 | #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER(((tree_reduce_block) | (bs_plain_barrier))) \ |
580 | (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier)((tree_reduce_block) | (bs_plain_barrier))) |
581 | #endif |
582 | |
583 | typedef int PACKED_REDUCTION_METHOD_T; |
584 | |
585 | /* -- end of fast reduction stuff ----------------------------------------- */ |
586 | |
587 | #if KMP_OS_WINDOWS0 |
588 | #define USE_CBLKDATA |
589 | #if KMP_MSVC_COMPAT0 |
590 | #pragma warning(push) |
591 | #pragma warning(disable : 271 310) |
592 | #endif |
593 | #include <windows.h> |
594 | #if KMP_MSVC_COMPAT0 |
595 | #pragma warning(pop) |
596 | #endif |
597 | #endif |
598 | |
599 | #if KMP_OS_UNIX1 |
600 | #include <dlfcn.h> |
601 | #include <pthread.h> |
602 | #endif |
603 | |
604 | enum kmp_hw_t : int { |
605 | KMP_HW_UNKNOWN = -1, |
606 | KMP_HW_SOCKET = 0, |
607 | KMP_HW_PROC_GROUP, |
608 | KMP_HW_NUMA, |
609 | KMP_HW_DIE, |
610 | KMP_HW_LLC, |
611 | KMP_HW_L3, |
612 | KMP_HW_TILE, |
613 | KMP_HW_MODULE, |
614 | KMP_HW_L2, |
615 | KMP_HW_L1, |
616 | KMP_HW_CORE, |
617 | KMP_HW_THREAD, |
618 | KMP_HW_LAST |
619 | }; |
620 | |
621 | typedef enum kmp_hw_core_type_t { |
622 | KMP_HW_CORE_TYPE_UNKNOWN = 0x0, |
623 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 |
624 | KMP_HW_CORE_TYPE_ATOM = 0x20, |
625 | KMP_HW_CORE_TYPE_CORE = 0x40, |
626 | KMP_HW_MAX_NUM_CORE_TYPES = 3, |
627 | #else |
628 | KMP_HW_MAX_NUM_CORE_TYPES = 1, |
629 | #endif |
630 | } kmp_hw_core_type_t; |
631 | |
632 | #define KMP_HW_MAX_NUM_CORE_EFFS8 8 |
633 | |
634 | #define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type)if (!(type >= (kmp_hw_t)0 && type < KMP_HW_LAST )) { __kmp_debug_assert("type >= (kmp_hw_t)0 && type < KMP_HW_LAST" , "openmp/runtime/src/kmp.h", 634); } \ |
635 | KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)if (!(type >= (kmp_hw_t)0 && type < KMP_HW_LAST )) { __kmp_debug_assert("type >= (kmp_hw_t)0 && type < KMP_HW_LAST" , "openmp/runtime/src/kmp.h", 635); } |
636 | #define KMP_ASSERT_VALID_HW_TYPE(type)if (!(type >= (kmp_hw_t)0 && type < KMP_HW_LAST )) { __kmp_debug_assert("type >= (kmp_hw_t)0 && type < KMP_HW_LAST" , "openmp/runtime/src/kmp.h", 636); } \ |
637 | KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)if (!(type >= (kmp_hw_t)0 && type < KMP_HW_LAST )) { __kmp_debug_assert("type >= (kmp_hw_t)0 && type < KMP_HW_LAST" , "openmp/runtime/src/kmp.h", 637); } |
638 | |
639 | #define KMP_FOREACH_HW_TYPE(type)for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST; type = (kmp_hw_t)((int)type + 1)) \ |
640 | for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST; \ |
641 | type = (kmp_hw_t)((int)type + 1)) |
642 | |
643 | const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural = false); |
644 | const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural = false); |
645 | const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type); |
646 | |
647 | /* Only Linux* OS and Windows* OS support thread affinity. */ |
648 | #if KMP_AFFINITY_SUPPORTED1 |
649 | |
650 | // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later). |
651 | #if KMP_OS_WINDOWS0 |
652 | #if _MSC_VER < 1600 && KMP_MSVC_COMPAT0 |
653 | typedef struct GROUP_AFFINITY { |
654 | KAFFINITY Mask; |
655 | WORD Group; |
656 | WORD Reserved[3]; |
657 | } GROUP_AFFINITY; |
658 | #endif /* _MSC_VER < 1600 */ |
659 | #if KMP_GROUP_AFFINITY0 |
660 | extern int __kmp_num_proc_groups; |
661 | #else |
662 | static const int __kmp_num_proc_groups = 1; |
663 | #endif /* KMP_GROUP_AFFINITY */ |
664 | typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD); |
665 | extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount; |
666 | |
667 | typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void); |
668 | extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount; |
669 | |
670 | typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *); |
671 | extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity; |
672 | |
673 | typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *, |
674 | GROUP_AFFINITY *); |
675 | extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity; |
676 | #endif /* KMP_OS_WINDOWS */ |
677 | |
678 | #if KMP_USE_HWLOC0 |
679 | extern hwloc_topology_t __kmp_hwloc_topology; |
680 | extern int __kmp_hwloc_error; |
681 | #endif |
682 | |
683 | extern size_t __kmp_affin_mask_size; |
684 | #define KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0) (__kmp_affin_mask_size > 0) |
685 | #define KMP_AFFINITY_DISABLE()(__kmp_affin_mask_size = 0) (__kmp_affin_mask_size = 0) |
686 | #define KMP_AFFINITY_ENABLE(mask_size)(__kmp_affin_mask_size = mask_size) (__kmp_affin_mask_size = mask_size) |
687 | #define KMP_CPU_SET_ITERATE(i, mask)for (i = (mask)->begin(); (int)i != (mask)->end(); i = ( mask)->next(i)) \ |
688 | for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i)) |
689 | #define KMP_CPU_SET(i, mask)(mask)->set(i) (mask)->set(i) |
690 | #define KMP_CPU_ISSET(i, mask)(mask)->is_set(i) (mask)->is_set(i) |
691 | #define KMP_CPU_CLR(i, mask)(mask)->clear(i) (mask)->clear(i) |
692 | #define KMP_CPU_ZERO(mask)(mask)->zero() (mask)->zero() |
693 | #define KMP_CPU_COPY(dest, src)(dest)->copy(src) (dest)->copy(src) |
694 | #define KMP_CPU_AND(dest, src)(dest)->bitwise_and(src) (dest)->bitwise_and(src) |
695 | #define KMP_CPU_COMPLEMENT(max_bit_number, mask)(mask)->bitwise_not() (mask)->bitwise_not() |
696 | #define KMP_CPU_UNION(dest, src)(dest)->bitwise_or(src) (dest)->bitwise_or(src) |
697 | #define KMP_CPU_ALLOC(ptr)(ptr = __kmp_affinity_dispatch->allocate_mask()) (ptr = __kmp_affinity_dispatch->allocate_mask()) |
698 | #define KMP_CPU_FREE(ptr)__kmp_affinity_dispatch->deallocate_mask(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr) |
699 | #define KMP_CPU_ALLOC_ON_STACK(ptr)(ptr = __kmp_affinity_dispatch->allocate_mask()) KMP_CPU_ALLOC(ptr)(ptr = __kmp_affinity_dispatch->allocate_mask()) |
700 | #define KMP_CPU_FREE_FROM_STACK(ptr)__kmp_affinity_dispatch->deallocate_mask(ptr) KMP_CPU_FREE(ptr)__kmp_affinity_dispatch->deallocate_mask(ptr) |
701 | #define KMP_CPU_INTERNAL_ALLOC(ptr)(ptr = __kmp_affinity_dispatch->allocate_mask()) KMP_CPU_ALLOC(ptr)(ptr = __kmp_affinity_dispatch->allocate_mask()) |
702 | #define KMP_CPU_INTERNAL_FREE(ptr)__kmp_affinity_dispatch->deallocate_mask(ptr) KMP_CPU_FREE(ptr)__kmp_affinity_dispatch->deallocate_mask(ptr) |
703 | #define KMP_CPU_INDEX(arr, i)__kmp_affinity_dispatch->index_mask_array(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i) |
704 | #define KMP_CPU_ALLOC_ARRAY(arr, n)(arr = __kmp_affinity_dispatch->allocate_mask_array(n)) \ |
705 | (arr = __kmp_affinity_dispatch->allocate_mask_array(n)) |
706 | #define KMP_CPU_FREE_ARRAY(arr, n)__kmp_affinity_dispatch->deallocate_mask_array(arr) \ |
707 | __kmp_affinity_dispatch->deallocate_mask_array(arr) |
708 | #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n)(arr = __kmp_affinity_dispatch->allocate_mask_array(n)) KMP_CPU_ALLOC_ARRAY(arr, n)(arr = __kmp_affinity_dispatch->allocate_mask_array(n)) |
709 | #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n)__kmp_affinity_dispatch->deallocate_mask_array(arr) KMP_CPU_FREE_ARRAY(arr, n)__kmp_affinity_dispatch->deallocate_mask_array(arr) |
710 | #define __kmp_get_system_affinity(mask, abort_bool)(mask)->get_system_affinity(abort_bool) \ |
711 | (mask)->get_system_affinity(abort_bool) |
712 | #define __kmp_set_system_affinity(mask, abort_bool)(mask)->set_system_affinity(abort_bool) \ |
713 | (mask)->set_system_affinity(abort_bool) |
714 | #define __kmp_get_proc_group(mask)(mask)->get_proc_group() (mask)->get_proc_group() |
715 | |
716 | class KMPAffinity { |
717 | public: |
718 | class Mask { |
719 | public: |
720 | void *operator new(size_t n); |
721 | void operator delete(void *p); |
722 | void *operator new[](size_t n); |
723 | void operator delete[](void *p); |
724 | virtual ~Mask() {} |
725 | // Set bit i to 1 |
726 | virtual void set(int i) {} |
727 | // Return bit i |
728 | virtual bool is_set(int i) const { return false; } |
729 | // Set bit i to 0 |
730 | virtual void clear(int i) {} |
731 | // Zero out entire mask |
732 | virtual void zero() {} |
733 | // Copy src into this mask |
734 | virtual void copy(const Mask *src) {} |
735 | // this &= rhs |
736 | virtual void bitwise_and(const Mask *rhs) {} |
737 | // this |= rhs |
738 | virtual void bitwise_or(const Mask *rhs) {} |
739 | // this = ~this |
740 | virtual void bitwise_not() {} |
741 | // API for iterating over an affinity mask |
742 | // for (int i = mask->begin(); i != mask->end(); i = mask->next(i)) |
743 | virtual int begin() const { return 0; } |
744 | virtual int end() const { return 0; } |
745 | virtual int next(int previous) const { return 0; } |
746 | #if KMP_OS_WINDOWS0 |
747 | virtual int set_process_affinity(bool abort_on_error) const { return -1; } |
748 | #endif |
749 | // Set the system's affinity to this affinity mask's value |
750 | virtual int set_system_affinity(bool abort_on_error) const { return -1; } |
751 | // Set this affinity mask to the current system affinity |
752 | virtual int get_system_affinity(bool abort_on_error) { return -1; } |
753 | // Only 1 DWORD in the mask should have any procs set. |
754 | // Return the appropriate index, or -1 for an invalid mask. |
755 | virtual int get_proc_group() const { return -1; } |
756 | }; |
757 | void *operator new(size_t n); |
758 | void operator delete(void *p); |
759 | // Need virtual destructor |
760 | virtual ~KMPAffinity() = default; |
761 | // Determine if affinity is capable |
762 | virtual void determine_capable(const char *env_var) {} |
763 | // Bind the current thread to os proc |
764 | virtual void bind_thread(int proc) {} |
765 | // Factory functions to allocate/deallocate a mask |
766 | virtual Mask *allocate_mask() { return nullptr; } |
767 | virtual void deallocate_mask(Mask *m) {} |
768 | virtual Mask *allocate_mask_array(int num) { return nullptr; } |
769 | virtual void deallocate_mask_array(Mask *m) {} |
770 | virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; } |
771 | static void pick_api(); |
772 | static void destroy_api(); |
773 | enum api_type { |
774 | NATIVE_OS |
775 | #if KMP_USE_HWLOC0 |
776 | , |
777 | HWLOC |
778 | #endif |
779 | }; |
780 | virtual api_type get_api_type() const { |
781 | KMP_ASSERT(0)if (!(0)) { __kmp_debug_assert("0", "openmp/runtime/src/kmp.h" , 781); }; |
782 | return NATIVE_OS; |
783 | } |
784 | |
785 | private: |
786 | static bool picked_api; |
787 | }; |
788 | |
789 | typedef KMPAffinity::Mask kmp_affin_mask_t; |
790 | extern KMPAffinity *__kmp_affinity_dispatch; |
791 | |
792 | // Declare local char buffers with this size for printing debug and info |
793 | // messages, using __kmp_affinity_print_mask(). |
794 | #define KMP_AFFIN_MASK_PRINT_LEN1024 1024 |
795 | |
796 | enum affinity_type { |
797 | affinity_none = 0, |
798 | affinity_physical, |
799 | affinity_logical, |
800 | affinity_compact, |
801 | affinity_scatter, |
802 | affinity_explicit, |
803 | affinity_balanced, |
804 | affinity_disabled, // not used outsize the env var parser |
805 | affinity_default |
806 | }; |
807 | |
808 | enum affinity_top_method { |
809 | affinity_top_method_all = 0, // try all (supported) methods, in order |
810 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 |
811 | affinity_top_method_apicid, |
812 | affinity_top_method_x2apicid, |
813 | affinity_top_method_x2apicid_1f, |
814 | #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ |
815 | affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too |
816 | #if KMP_GROUP_AFFINITY0 |
817 | affinity_top_method_group, |
818 | #endif /* KMP_GROUP_AFFINITY */ |
819 | affinity_top_method_flat, |
820 | #if KMP_USE_HWLOC0 |
821 | affinity_top_method_hwloc, |
822 | #endif |
823 | affinity_top_method_default |
824 | }; |
825 | |
826 | #define affinity_respect_mask_default(2) (2) |
827 | |
828 | typedef struct kmp_affinity_flags_t { |
829 | unsigned dups : 1; |
830 | unsigned verbose : 1; |
831 | unsigned warnings : 1; |
832 | unsigned respect : 2; |
833 | unsigned reset : 1; |
834 | unsigned initialized : 1; |
835 | unsigned reserved : 25; |
836 | } kmp_affinity_flags_t; |
837 | KMP_BUILD_ASSERT(sizeof(kmp_affinity_flags_t) == 4)static_assert(sizeof(kmp_affinity_flags_t) == 4, "Build condition error" ); |
838 | |
839 | typedef struct kmp_affinity_t { |
840 | char *proclist; |
841 | enum affinity_type type; |
842 | kmp_hw_t gran; |
843 | int gran_levels; |
844 | int compact; |
845 | int offset; |
846 | kmp_affinity_flags_t flags; |
847 | unsigned num_masks; |
848 | kmp_affin_mask_t *masks; |
849 | unsigned num_os_id_masks; |
850 | kmp_affin_mask_t *os_id_masks; |
851 | const char *env_var; |
852 | } kmp_affinity_t; |
853 | |
854 | #define KMP_AFFINITY_INIT(env){ nullptr, affinity_default, KMP_HW_UNKNOWN, -1, 0, 0, {(!0), 0, (!0), (2), 0, 0}, 0, nullptr, 0, nullptr, env } \ |
855 | { \ |
856 | nullptr, affinity_default, KMP_HW_UNKNOWN, -1, 0, 0, \ |
857 | {TRUE(!0), FALSE0, TRUE(!0), affinity_respect_mask_default(2), FALSE0, FALSE0}, 0, \ |
858 | nullptr, 0, nullptr, env \ |
859 | } |
860 | |
861 | extern enum affinity_top_method __kmp_affinity_top_method; |
862 | extern kmp_affinity_t __kmp_affinity; |
863 | extern kmp_affinity_t __kmp_hh_affinity; |
864 | extern kmp_affinity_t *__kmp_affinities[2]; |
865 | |
866 | extern void __kmp_affinity_bind_thread(int which); |
867 | |
868 | extern kmp_affin_mask_t *__kmp_affin_fullMask; |
869 | extern kmp_affin_mask_t *__kmp_affin_origMask; |
870 | extern char *__kmp_cpuinfo_file; |
871 | |
872 | #endif /* KMP_AFFINITY_SUPPORTED */ |
873 | |
874 | // This needs to be kept in sync with the values in omp.h !!! |
875 | typedef enum kmp_proc_bind_t { |
876 | proc_bind_false = 0, |
877 | proc_bind_true, |
878 | proc_bind_primary, |
879 | proc_bind_close, |
880 | proc_bind_spread, |
881 | proc_bind_intel, // use KMP_AFFINITY interface |
882 | proc_bind_default |
883 | } kmp_proc_bind_t; |
884 | |
885 | typedef struct kmp_nested_proc_bind_t { |
886 | kmp_proc_bind_t *bind_types; |
887 | int size; |
888 | int used; |
889 | } kmp_nested_proc_bind_t; |
890 | |
891 | extern kmp_nested_proc_bind_t __kmp_nested_proc_bind; |
892 | extern kmp_proc_bind_t __kmp_teams_proc_bind; |
893 | |
894 | extern int __kmp_display_affinity; |
895 | extern char *__kmp_affinity_format; |
896 | static const size_t KMP_AFFINITY_FORMAT_SIZE = 512; |
897 | #if OMPT_SUPPORT1 |
898 | extern int __kmp_tool; |
899 | extern char *__kmp_tool_libraries; |
900 | #endif // OMPT_SUPPORT |
901 | |
902 | #if KMP_AFFINITY_SUPPORTED1 |
903 | #define KMP_PLACE_ALL(-1) (-1) |
904 | #define KMP_PLACE_UNDEFINED(-2) (-2) |
905 | // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES? |
906 | #define KMP_AFFINITY_NON_PROC_BIND((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || __kmp_nested_proc_bind .bind_types[0] == proc_bind_intel) && (__kmp_affinity .num_masks > 0 || __kmp_affinity.type == affinity_balanced )) \ |
907 | ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \ |
908 | __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \ |
909 | (__kmp_affinity.num_masks > 0 || __kmp_affinity.type == affinity_balanced)) |
910 | #endif /* KMP_AFFINITY_SUPPORTED */ |
911 | |
912 | extern int __kmp_affinity_num_places; |
913 | |
914 | typedef enum kmp_cancel_kind_t { |
915 | cancel_noreq = 0, |
916 | cancel_parallel = 1, |
917 | cancel_loop = 2, |
918 | cancel_sections = 3, |
919 | cancel_taskgroup = 4 |
920 | } kmp_cancel_kind_t; |
921 | |
922 | // KMP_HW_SUBSET support: |
923 | typedef struct kmp_hws_item { |
924 | int num; |
925 | int offset; |
926 | } kmp_hws_item_t; |
927 | |
928 | extern kmp_hws_item_t __kmp_hws_socket; |
929 | extern kmp_hws_item_t __kmp_hws_die; |
930 | extern kmp_hws_item_t __kmp_hws_node; |
931 | extern kmp_hws_item_t __kmp_hws_tile; |
932 | extern kmp_hws_item_t __kmp_hws_core; |
933 | extern kmp_hws_item_t __kmp_hws_proc; |
934 | extern int __kmp_hws_requested; |
935 | extern int __kmp_hws_abs_flag; // absolute or per-item number requested |
936 | |
937 | /* ------------------------------------------------------------------------ */ |
938 | |
939 | #define KMP_PAD(type, sz)(sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1)) \ |
940 | (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1)) |
941 | |
942 | // We need to avoid using -1 as a GTID as +1 is added to the gtid |
943 | // when storing it in a lock, and the value 0 is reserved. |
944 | #define KMP_GTID_DNE(-2) (-2) /* Does not exist */ |
945 | #define KMP_GTID_SHUTDOWN(-3) (-3) /* Library is shutting down */ |
946 | #define KMP_GTID_MONITOR(-4) (-4) /* Monitor thread ID */ |
947 | #define KMP_GTID_UNKNOWN(-5) (-5) /* Is not known */ |
948 | #define KMP_GTID_MIN(-6) (-6) /* Minimal gtid for low bound check in DEBUG */ |
949 | |
950 | /* OpenMP 5.0 Memory Management support */ |
951 | |
952 | #ifndef __OMP_H |
953 | // Duplicate type definitions from omp.h |
954 | typedef uintptr_t omp_uintptr_t; |
955 | |
956 | typedef enum { |
957 | omp_atk_sync_hint = 1, |
958 | omp_atk_alignment = 2, |
959 | omp_atk_access = 3, |
960 | omp_atk_pool_size = 4, |
961 | omp_atk_fallback = 5, |
962 | omp_atk_fb_data = 6, |
963 | omp_atk_pinned = 7, |
964 | omp_atk_partition = 8 |
965 | } omp_alloctrait_key_t; |
966 | |
967 | typedef enum { |
968 | omp_atv_false = 0, |
969 | omp_atv_true = 1, |
970 | omp_atv_contended = 3, |
971 | omp_atv_uncontended = 4, |
972 | omp_atv_serialized = 5, |
973 | omp_atv_sequential = omp_atv_serialized, // (deprecated) |
974 | omp_atv_private = 6, |
975 | omp_atv_all = 7, |
976 | omp_atv_thread = 8, |
977 | omp_atv_pteam = 9, |
978 | omp_atv_cgroup = 10, |
979 | omp_atv_default_mem_fb = 11, |
980 | omp_atv_null_fb = 12, |
981 | omp_atv_abort_fb = 13, |
982 | omp_atv_allocator_fb = 14, |
983 | omp_atv_environment = 15, |
984 | omp_atv_nearest = 16, |
985 | omp_atv_blocked = 17, |
986 | omp_atv_interleaved = 18 |
987 | } omp_alloctrait_value_t; |
988 | #define omp_atv_default((omp_uintptr_t)-1) ((omp_uintptr_t)-1) |
989 | |
990 | typedef void *omp_memspace_handle_t; |
991 | extern omp_memspace_handle_t const omp_default_mem_space; |
992 | extern omp_memspace_handle_t const omp_large_cap_mem_space; |
993 | extern omp_memspace_handle_t const omp_const_mem_space; |
994 | extern omp_memspace_handle_t const omp_high_bw_mem_space; |
995 | extern omp_memspace_handle_t const omp_low_lat_mem_space; |
996 | extern omp_memspace_handle_t const llvm_omp_target_host_mem_space; |
997 | extern omp_memspace_handle_t const llvm_omp_target_shared_mem_space; |
998 | extern omp_memspace_handle_t const llvm_omp_target_device_mem_space; |
999 | |
1000 | typedef struct { |
1001 | omp_alloctrait_key_t key; |
1002 | omp_uintptr_t value; |
1003 | } omp_alloctrait_t; |
1004 | |
1005 | typedef void *omp_allocator_handle_t; |
1006 | extern omp_allocator_handle_t const omp_null_allocator; |
1007 | extern omp_allocator_handle_t const omp_default_mem_alloc; |
1008 | extern omp_allocator_handle_t const omp_large_cap_mem_alloc; |
1009 | extern omp_allocator_handle_t const omp_const_mem_alloc; |
1010 | extern omp_allocator_handle_t const omp_high_bw_mem_alloc; |
1011 | extern omp_allocator_handle_t const omp_low_lat_mem_alloc; |
1012 | extern omp_allocator_handle_t const omp_cgroup_mem_alloc; |
1013 | extern omp_allocator_handle_t const omp_pteam_mem_alloc; |
1014 | extern omp_allocator_handle_t const omp_thread_mem_alloc; |
1015 | extern omp_allocator_handle_t const llvm_omp_target_host_mem_alloc; |
1016 | extern omp_allocator_handle_t const llvm_omp_target_shared_mem_alloc; |
1017 | extern omp_allocator_handle_t const llvm_omp_target_device_mem_alloc; |
1018 | extern omp_allocator_handle_t const kmp_max_mem_alloc; |
1019 | extern omp_allocator_handle_t __kmp_def_allocator; |
1020 | |
1021 | // end of duplicate type definitions from omp.h |
1022 | #endif |
1023 | |
1024 | extern int __kmp_memkind_available; |
1025 | |
1026 | typedef omp_memspace_handle_t kmp_memspace_t; // placeholder |
1027 | |
1028 | typedef struct kmp_allocator_t { |
1029 | omp_memspace_handle_t memspace; |
1030 | void **memkind; // pointer to memkind |
1031 | size_t alignment; |
1032 | omp_alloctrait_value_t fb; |
1033 | kmp_allocator_t *fb_data; |
1034 | kmp_uint64 pool_size; |
1035 | kmp_uint64 pool_used; |
1036 | } kmp_allocator_t; |
1037 | |
1038 | extern omp_allocator_handle_t __kmpc_init_allocator(int gtid, |
1039 | omp_memspace_handle_t, |
1040 | int ntraits, |
1041 | omp_alloctrait_t traits[]); |
1042 | extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al); |
1043 | extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al); |
1044 | extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid); |
1045 | // external interfaces, may be used by compiler |
1046 | extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al); |
1047 | extern void *__kmpc_aligned_alloc(int gtid, size_t align, size_t sz, |
1048 | omp_allocator_handle_t al); |
1049 | extern void *__kmpc_calloc(int gtid, size_t nmemb, size_t sz, |
1050 | omp_allocator_handle_t al); |
1051 | extern void *__kmpc_realloc(int gtid, void *ptr, size_t sz, |
1052 | omp_allocator_handle_t al, |
1053 | omp_allocator_handle_t free_al); |
1054 | extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al); |
1055 | // internal interfaces, contain real implementation |
1056 | extern void *__kmp_alloc(int gtid, size_t align, size_t sz, |
1057 | omp_allocator_handle_t al); |
1058 | extern void *__kmp_calloc(int gtid, size_t align, size_t nmemb, size_t sz, |
1059 | omp_allocator_handle_t al); |
1060 | extern void *__kmp_realloc(int gtid, void *ptr, size_t sz, |
1061 | omp_allocator_handle_t al, |
1062 | omp_allocator_handle_t free_al); |
1063 | extern void ___kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al); |
1064 | |
1065 | extern void __kmp_init_memkind(); |
1066 | extern void __kmp_fini_memkind(); |
1067 | extern void __kmp_init_target_mem(); |
1068 | |
1069 | /* ------------------------------------------------------------------------ */ |
1070 | |
1071 | #define KMP_UINT64_MAX(~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3 )) - 1))) \ |
1072 | (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1))) |
1073 | |
1074 | #define KMP_MIN_NTH1 1 |
1075 | |
1076 | #ifndef KMP_MAX_NTH2147483647 |
1077 | #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX2147483647 |
1078 | #define KMP_MAX_NTH2147483647 PTHREAD_THREADS_MAX |
1079 | #else |
1080 | #define KMP_MAX_NTH2147483647 INT_MAX2147483647 |
1081 | #endif |
1082 | #endif /* KMP_MAX_NTH */ |
1083 | |
1084 | #ifdef PTHREAD_STACK_MIN16384 |
1085 | #define KMP_MIN_STKSIZE16384 PTHREAD_STACK_MIN16384 |
1086 | #else |
1087 | #define KMP_MIN_STKSIZE16384 ((size_t)(32 * 1024)) |
1088 | #endif |
1089 | |
1090 | #define KMP_MAX_STKSIZE(~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1) )) (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1))) |
1091 | |
1092 | #if KMP_ARCH_X860 |
1093 | #define KMP_DEFAULT_STKSIZE((size_t)(4 * 1024 * 1024)) ((size_t)(2 * 1024 * 1024)) |
1094 | #elif KMP_ARCH_X86_641 |
1095 | #define KMP_DEFAULT_STKSIZE((size_t)(4 * 1024 * 1024)) ((size_t)(4 * 1024 * 1024)) |
1096 | #define KMP_BACKUP_STKSIZE((size_t)(2 * 1024 * 1024)) ((size_t)(2 * 1024 * 1024)) |
1097 | #else |
1098 | #define KMP_DEFAULT_STKSIZE((size_t)(4 * 1024 * 1024)) ((size_t)(1024 * 1024)) |
1099 | #endif |
1100 | |
1101 | #define KMP_DEFAULT_MALLOC_POOL_INCR((size_t)(1024 * 1024)) ((size_t)(1024 * 1024)) |
1102 | #define KMP_MIN_MALLOC_POOL_INCR((size_t)(4 * 1024)) ((size_t)(4 * 1024)) |
1103 | #define KMP_MAX_MALLOC_POOL_INCR(~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1) )) \ |
1104 | (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1))) |
1105 | |
1106 | #define KMP_MIN_STKOFFSET(0) (0) |
1107 | #define KMP_MAX_STKOFFSET(~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1) )) KMP_MAX_STKSIZE(~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1) )) |
1108 | #if KMP_OS_DARWIN0 |
1109 | #define KMP_DEFAULT_STKOFFSET64 KMP_MIN_STKOFFSET(0) |
1110 | #else |
1111 | #define KMP_DEFAULT_STKOFFSET64 CACHE_LINE64 |
1112 | #endif |
1113 | |
1114 | #define KMP_MIN_STKPADDING(0) (0) |
1115 | #define KMP_MAX_STKPADDING(2 * 1024 * 1024) (2 * 1024 * 1024) |
1116 | |
1117 | #define KMP_BLOCKTIME_MULTIPLIER(1000) \ |
1118 | (1000) /* number of blocktime units per second */ |
1119 | #define KMP_MIN_BLOCKTIME(0) (0) |
1120 | #define KMP_MAX_BLOCKTIME(2147483647) \ |
1121 | (INT_MAX2147483647) /* Must be this for "infinite" setting the work */ |
1122 | |
1123 | /* __kmp_blocktime is in milliseconds */ |
1124 | #define KMP_DEFAULT_BLOCKTIME(__kmp_is_hybrid_cpu() ? (0) : (200)) (__kmp_is_hybrid_cpu() ? (0) : (200)) |
1125 | |
1126 | #if KMP_USE_MONITOR |
1127 | #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024)) |
1128 | #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second |
1129 | #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec |
1130 | |
1131 | /* Calculate new number of monitor wakeups for a specific block time based on |
1132 | previous monitor_wakeups. Only allow increasing number of wakeups */ |
1133 | #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \ |
1134 | (((blocktime) == KMP_MAX_BLOCKTIME(2147483647)) ? (monitor_wakeups) \ |
1135 | : ((blocktime) == KMP_MIN_BLOCKTIME(0)) ? KMP_MAX_MONITOR_WAKEUPS \ |
1136 | : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER(1000) / (blocktime))) \ |
1137 | ? (monitor_wakeups) \ |
1138 | : (KMP_BLOCKTIME_MULTIPLIER(1000)) / (blocktime)) |
1139 | |
1140 | /* Calculate number of intervals for a specific block time based on |
1141 | monitor_wakeups */ |
1142 | #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \ |
1143 | (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER(1000) / (monitor_wakeups)) - 1) / \ |
1144 | (KMP_BLOCKTIME_MULTIPLIER(1000) / (monitor_wakeups))) |
1145 | #else |
1146 | #define KMP_BLOCKTIME(team, tid)(((team)->t.t_threads[(tid)]->th.th_current_task->td_icvs .bt_set) ? ((team)->t.t_threads[(tid)]->th.th_current_task ->td_icvs.blocktime) : __kmp_dflt_blocktime) \ |
1147 | (get__bt_set(team, tid)((team)->t.t_threads[(tid)]->th.th_current_task->td_icvs .bt_set) ? get__blocktime(team, tid)((team)->t.t_threads[(tid)]->th.th_current_task->td_icvs .blocktime) : __kmp_dflt_blocktime) |
1148 | #if KMP_OS_UNIX1 && (KMP_ARCH_X860 || KMP_ARCH_X86_641) |
1149 | // HW TSC is used to reduce overhead (clock tick instead of nanosecond). |
1150 | extern kmp_uint64 __kmp_ticks_per_msec; |
1151 | #if KMP_COMPILER_ICC0 || KMP_COMPILER_ICX0 |
1152 | #define KMP_NOW()__kmp_hardware_timestamp() ((kmp_uint64)_rdtsc()) |
1153 | #else |
1154 | #define KMP_NOW()__kmp_hardware_timestamp() __kmp_hardware_timestamp() |
1155 | #endif |
1156 | #define KMP_NOW_MSEC()(__kmp_hardware_timestamp() / __kmp_ticks_per_msec) (KMP_NOW()__kmp_hardware_timestamp() / __kmp_ticks_per_msec) |
1157 | #define KMP_BLOCKTIME_INTERVAL(team, tid)((((team)->t.t_threads[(tid)]->th.th_current_task->td_icvs .bt_set) ? ((team)->t.t_threads[(tid)]->th.th_current_task ->td_icvs.blocktime) : __kmp_dflt_blocktime) * __kmp_ticks_per_msec ) \ |
1158 | (KMP_BLOCKTIME(team, tid)(((team)->t.t_threads[(tid)]->th.th_current_task->td_icvs .bt_set) ? ((team)->t.t_threads[(tid)]->th.th_current_task ->td_icvs.blocktime) : __kmp_dflt_blocktime) * __kmp_ticks_per_msec) |
1159 | #define KMP_BLOCKING(goal, count)((goal) > __kmp_hardware_timestamp()) ((goal) > KMP_NOW()__kmp_hardware_timestamp()) |
1160 | #else |
1161 | // System time is retrieved sporadically while blocking. |
1162 | extern kmp_uint64 __kmp_now_nsec(); |
1163 | #define KMP_NOW()__kmp_hardware_timestamp() __kmp_now_nsec() |
1164 | #define KMP_NOW_MSEC()(__kmp_hardware_timestamp() / __kmp_ticks_per_msec) (KMP_NOW()__kmp_hardware_timestamp() / KMP_USEC_PER_SEC1000000L) |
1165 | #define KMP_BLOCKTIME_INTERVAL(team, tid)((((team)->t.t_threads[(tid)]->th.th_current_task->td_icvs .bt_set) ? ((team)->t.t_threads[(tid)]->th.th_current_task ->td_icvs.blocktime) : __kmp_dflt_blocktime) * __kmp_ticks_per_msec ) \ |
1166 | (KMP_BLOCKTIME(team, tid)(((team)->t.t_threads[(tid)]->th.th_current_task->td_icvs .bt_set) ? ((team)->t.t_threads[(tid)]->th.th_current_task ->td_icvs.blocktime) : __kmp_dflt_blocktime) * KMP_USEC_PER_SEC1000000L) |
1167 | #define KMP_BLOCKING(goal, count)((goal) > __kmp_hardware_timestamp()) ((count) % 1000 != 0 || (goal) > KMP_NOW()__kmp_hardware_timestamp()) |
1168 | #endif |
1169 | #endif // KMP_USE_MONITOR |
1170 | |
1171 | #define KMP_MIN_STATSCOLS40 40 |
1172 | #define KMP_MAX_STATSCOLS4096 4096 |
1173 | #define KMP_DEFAULT_STATSCOLS80 80 |
1174 | |
1175 | #define KMP_MIN_INTERVAL0 0 |
1176 | #define KMP_MAX_INTERVAL(2147483647 - 1) (INT_MAX2147483647 - 1) |
1177 | #define KMP_DEFAULT_INTERVAL0 0 |
1178 | |
1179 | #define KMP_MIN_CHUNK1 1 |
1180 | #define KMP_MAX_CHUNK(2147483647 - 1) (INT_MAX2147483647 - 1) |
1181 | #define KMP_DEFAULT_CHUNK1 1 |
1182 | |
1183 | #define KMP_MIN_DISP_NUM_BUFF1 1 |
1184 | #define KMP_DFLT_DISP_NUM_BUFF7 7 |
1185 | #define KMP_MAX_DISP_NUM_BUFF4096 4096 |
1186 | |
1187 | #define KMP_MAX_ORDERED8 8 |
1188 | |
1189 | #define KMP_MAX_FIELDS32 32 |
1190 | |
1191 | #define KMP_MAX_BRANCH_BITS31 31 |
1192 | |
1193 | #define KMP_MAX_ACTIVE_LEVELS_LIMIT2147483647 INT_MAX2147483647 |
1194 | |
1195 | #define KMP_MAX_DEFAULT_DEVICE_LIMIT2147483647 INT_MAX2147483647 |
1196 | |
1197 | #define KMP_MAX_TASK_PRIORITY_LIMIT2147483647 INT_MAX2147483647 |
1198 | |
1199 | /* Minimum number of threads before switch to TLS gtid (experimentally |
1200 | determined) */ |
1201 | /* josh TODO: what about OS X* tuning? */ |
1202 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 |
1203 | #define KMP_TLS_GTID_MIN5 5 |
1204 | #else |
1205 | #define KMP_TLS_GTID_MIN5 INT_MAX2147483647 |
1206 | #endif |
1207 | |
1208 | #define KMP_MASTER_TID(tid)(0 == (tid)) (0 == (tid)) |
1209 | #define KMP_WORKER_TID(tid)(0 != (tid)) (0 != (tid)) |
1210 | |
1211 | #define KMP_MASTER_GTID(gtid)(0 == __kmp_tid_from_gtid((gtid))) (0 == __kmp_tid_from_gtid((gtid))) |
1212 | #define KMP_WORKER_GTID(gtid)(0 != __kmp_tid_from_gtid((gtid))) (0 != __kmp_tid_from_gtid((gtid))) |
1213 | #define KMP_INITIAL_GTID(gtid)(0 == (gtid)) (0 == (gtid)) |
1214 | |
1215 | #ifndef TRUE(!0) |
1216 | #define FALSE0 0 |
1217 | #define TRUE(!0) (!FALSE0) |
1218 | #endif |
1219 | |
1220 | /* NOTE: all of the following constants must be even */ |
1221 | |
1222 | #if KMP_OS_WINDOWS0 |
1223 | #define KMP_INIT_WAIT1024U 64U /* initial number of spin-tests */ |
1224 | #define KMP_NEXT_WAIT512U 32U /* susequent number of spin-tests */ |
1225 | #elif KMP_OS_LINUX1 |
1226 | #define KMP_INIT_WAIT1024U 1024U /* initial number of spin-tests */ |
1227 | #define KMP_NEXT_WAIT512U 512U /* susequent number of spin-tests */ |
1228 | #elif KMP_OS_DARWIN0 |
1229 | /* TODO: tune for KMP_OS_DARWIN */ |
1230 | #define KMP_INIT_WAIT1024U 1024U /* initial number of spin-tests */ |
1231 | #define KMP_NEXT_WAIT512U 512U /* susequent number of spin-tests */ |
1232 | #elif KMP_OS_DRAGONFLY0 |
1233 | /* TODO: tune for KMP_OS_DRAGONFLY */ |
1234 | #define KMP_INIT_WAIT1024U 1024U /* initial number of spin-tests */ |
1235 | #define KMP_NEXT_WAIT512U 512U /* susequent number of spin-tests */ |
1236 | #elif KMP_OS_FREEBSD0 |
1237 | /* TODO: tune for KMP_OS_FREEBSD */ |
1238 | #define KMP_INIT_WAIT1024U 1024U /* initial number of spin-tests */ |
1239 | #define KMP_NEXT_WAIT512U 512U /* susequent number of spin-tests */ |
1240 | #elif KMP_OS_NETBSD0 |
1241 | /* TODO: tune for KMP_OS_NETBSD */ |
1242 | #define KMP_INIT_WAIT1024U 1024U /* initial number of spin-tests */ |
1243 | #define KMP_NEXT_WAIT512U 512U /* susequent number of spin-tests */ |
1244 | #elif KMP_OS_HURD0 |
1245 | /* TODO: tune for KMP_OS_HURD */ |
1246 | #define KMP_INIT_WAIT1024U 1024U /* initial number of spin-tests */ |
1247 | #define KMP_NEXT_WAIT512U 512U /* susequent number of spin-tests */ |
1248 | #elif KMP_OS_OPENBSD0 |
1249 | /* TODO: tune for KMP_OS_OPENBSD */ |
1250 | #define KMP_INIT_WAIT1024U 1024U /* initial number of spin-tests */ |
1251 | #define KMP_NEXT_WAIT512U 512U /* susequent number of spin-tests */ |
1252 | #endif |
1253 | |
1254 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 |
1255 | typedef struct kmp_cpuid { |
1256 | kmp_uint32 eax; |
1257 | kmp_uint32 ebx; |
1258 | kmp_uint32 ecx; |
1259 | kmp_uint32 edx; |
1260 | } kmp_cpuid_t; |
1261 | |
1262 | typedef struct kmp_cpuinfo_flags_t { |
1263 | unsigned sse2 : 1; // 0 if SSE2 instructions are not supported, 1 otherwise. |
1264 | unsigned rtm : 1; // 0 if RTM instructions are not supported, 1 otherwise. |
1265 | unsigned hybrid : 1; |
1266 | unsigned reserved : 29; // Ensure size of 32 bits |
1267 | } kmp_cpuinfo_flags_t; |
1268 | |
1269 | typedef struct kmp_cpuinfo { |
1270 | int initialized; // If 0, other fields are not initialized. |
1271 | int signature; // CPUID(1).EAX |
1272 | int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family) |
1273 | int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended |
1274 | // Model << 4 ) + Model) |
1275 | int stepping; // CPUID(1).EAX[3:0] ( Stepping ) |
1276 | kmp_cpuinfo_flags_t flags; |
1277 | int apic_id; |
1278 | int physical_id; |
1279 | int logical_id; |
1280 | kmp_uint64 frequency; // Nominal CPU frequency in Hz. |
1281 | char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004) |
1282 | } kmp_cpuinfo_t; |
1283 | |
1284 | extern void __kmp_query_cpuid(kmp_cpuinfo_t *p); |
1285 | |
1286 | #if KMP_OS_UNIX1 |
1287 | // subleaf is only needed for cache and topology discovery and can be set to |
1288 | // zero in most cases |
1289 | static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) { |
1290 | __asm__ __volatile__("cpuid" |
1291 | : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx) |
1292 | : "a"(leaf), "c"(subleaf)); |
1293 | } |
1294 | // Load p into FPU control word |
1295 | static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) { |
1296 | __asm__ __volatile__("fldcw %0" : : "m"(*p)); |
1297 | } |
1298 | // Store FPU control word into p |
1299 | static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) { |
1300 | __asm__ __volatile__("fstcw %0" : "=m"(*p)); |
1301 | } |
1302 | static inline void __kmp_clear_x87_fpu_status_word() { |
1303 | #if KMP_MIC0 |
1304 | // 32-bit protected mode x87 FPU state |
1305 | struct x87_fpu_state { |
1306 | unsigned cw; |
1307 | unsigned sw; |
1308 | unsigned tw; |
1309 | unsigned fip; |
1310 | unsigned fips; |
1311 | unsigned fdp; |
1312 | unsigned fds; |
1313 | }; |
1314 | struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0}; |
1315 | __asm__ __volatile__("fstenv %0\n\t" // store FP env |
1316 | "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW |
1317 | "fldenv %0\n\t" // load FP env back |
1318 | : "+m"(fpu_state), "+m"(fpu_state.sw)); |
1319 | #else |
1320 | __asm__ __volatile__("fnclex"); |
1321 | #endif // KMP_MIC |
1322 | } |
1323 | #if __SSE__1 |
1324 | static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); } |
1325 | static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); } |
1326 | #else |
1327 | static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {} |
1328 | static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; } |
1329 | #endif |
1330 | #else |
1331 | // Windows still has these as external functions in assembly file |
1332 | extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p); |
1333 | extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p); |
1334 | extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p); |
1335 | extern void __kmp_clear_x87_fpu_status_word(); |
1336 | static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); } |
1337 | static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); } |
1338 | #endif // KMP_OS_UNIX |
1339 | |
1340 | #define KMP_X86_MXCSR_MASK0xffffffc0 0xffffffc0 /* ignore status flags (6 lsb) */ |
1341 | |
1342 | // User-level Monitor/Mwait |
1343 | #if KMP_HAVE_UMWAIT((0 || 1) && (1 || 0) && !0) |
1344 | // We always try for UMWAIT first |
1345 | #if KMP_HAVE_WAITPKG_INTRINSICS1 |
1346 | #if KMP_HAVE_IMMINTRIN_H1 |
1347 | #include <immintrin.h> |
1348 | #elif KMP_HAVE_INTRIN_H0 |
1349 | #include <intrin.h> |
1350 | #endif |
1351 | #endif // KMP_HAVE_WAITPKG_INTRINSICS |
1352 | |
1353 | KMP_ATTRIBUTE_TARGET_WAITPKG__attribute__((target("waitpkg"))) |
1354 | static inline int __kmp_tpause(uint32_t hint, uint64_t counter) { |
1355 | #if !KMP_HAVE_WAITPKG_INTRINSICS1 |
1356 | uint32_t timeHi = uint32_t(counter >> 32); |
1357 | uint32_t timeLo = uint32_t(counter & 0xffffffff); |
1358 | char flag; |
1359 | __asm__ volatile("#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n" |
1360 | "setb %0" |
1361 | // The "=q" restraint means any register accessible as rl |
1362 | // in 32-bit mode: a, b, c, and d; |
1363 | // in 64-bit mode: any integer register |
1364 | : "=q"(flag) |
1365 | : "a"(timeLo), "d"(timeHi), "c"(hint) |
1366 | :); |
1367 | return flag; |
1368 | #else |
1369 | return _tpause(hint, counter); |
1370 | #endif |
1371 | } |
1372 | KMP_ATTRIBUTE_TARGET_WAITPKG__attribute__((target("waitpkg"))) |
1373 | static inline void __kmp_umonitor(void *cacheline) { |
1374 | #if !KMP_HAVE_WAITPKG_INTRINSICS1 |
1375 | __asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 " |
1376 | : |
1377 | : "a"(cacheline) |
1378 | :); |
1379 | #else |
1380 | _umonitor(cacheline); |
1381 | #endif |
1382 | } |
1383 | KMP_ATTRIBUTE_TARGET_WAITPKG__attribute__((target("waitpkg"))) |
1384 | static inline int __kmp_umwait(uint32_t hint, uint64_t counter) { |
1385 | #if !KMP_HAVE_WAITPKG_INTRINSICS1 |
1386 | uint32_t timeHi = uint32_t(counter >> 32); |
1387 | uint32_t timeLo = uint32_t(counter & 0xffffffff); |
1388 | char flag; |
1389 | __asm__ volatile("#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n" |
1390 | "setb %0" |
1391 | // The "=q" restraint means any register accessible as rl |
1392 | // in 32-bit mode: a, b, c, and d; |
1393 | // in 64-bit mode: any integer register |
1394 | : "=q"(flag) |
1395 | : "a"(timeLo), "d"(timeHi), "c"(hint) |
1396 | :); |
1397 | return flag; |
1398 | #else |
1399 | return _umwait(hint, counter); |
1400 | #endif |
1401 | } |
1402 | #elif KMP_HAVE_MWAIT((0 || 1) && (1 || 0) && !0) |
1403 | #if KMP_OS_UNIX1 |
1404 | #include <pmmintrin.h> |
1405 | #else |
1406 | #include <intrin.h> |
1407 | #endif |
1408 | #if KMP_OS_UNIX1 |
1409 | __attribute__((target("sse3"))) |
1410 | #endif |
1411 | static inline void |
1412 | __kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) { |
1413 | _mm_monitor(cacheline, extensions, hints); |
1414 | } |
1415 | #if KMP_OS_UNIX1 |
1416 | __attribute__((target("sse3"))) |
1417 | #endif |
1418 | static inline void |
1419 | __kmp_mm_mwait(unsigned extensions, unsigned hints) { |
1420 | _mm_mwait(extensions, hints); |
1421 | } |
1422 | #endif // KMP_HAVE_UMWAIT |
1423 | |
1424 | #if KMP_ARCH_X860 |
1425 | extern void __kmp_x86_pause(void); |
1426 | #elif KMP_MIC0 |
1427 | // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed |
1428 | // regression after removal of extra PAUSE from spin loops. Changing |
1429 | // the delay from 100 to 300 showed even better performance than double PAUSE |
1430 | // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC. |
1431 | static inline void __kmp_x86_pause(void) { _mm_delay_32(300); } |
1432 | #else |
1433 | static inline void __kmp_x86_pause(void) { _mm_pause(); } |
1434 | #endif |
1435 | #define KMP_CPU_PAUSE()__kmp_x86_pause() __kmp_x86_pause() |
1436 | #elif KMP_ARCH_PPC64(0 || 0) |
1437 | #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1") |
1438 | #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2") |
1439 | #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory") |
1440 | #define KMP_CPU_PAUSE()__kmp_x86_pause() \ |
1441 | do { \ |
1442 | KMP_PPC64_PRI_LOW(); \ |
1443 | KMP_PPC64_PRI_MED(); \ |
1444 | KMP_PPC64_PRI_LOC_MB(); \ |
1445 | } while (0) |
1446 | #else |
1447 | #define KMP_CPU_PAUSE()__kmp_x86_pause() /* nothing to do */ |
1448 | #endif |
1449 | |
1450 | #define KMP_INIT_YIELD(count){ (count) = __kmp_yield_init; } \ |
1451 | { (count) = __kmp_yield_init; } |
1452 | |
1453 | #define KMP_INIT_BACKOFF(time){ (time) = __kmp_pause_init; } \ |
1454 | { (time) = __kmp_pause_init; } |
1455 | |
1456 | #define KMP_OVERSUBSCRIBED((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc )) \ |
1457 | (TCR_4(__kmp_nth)(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)) |
1458 | |
1459 | #define KMP_TRY_YIELD((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && ( ((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc ))))) \ |
1460 | ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc ))))) |
1461 | |
1462 | #define KMP_TRY_YIELD_OVERSUB((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (( (__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc )))) \ |
1463 | ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc )))) |
1464 | |
1465 | #define KMP_YIELD(cond){ __kmp_x86_pause(); if ((cond) && (((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))))))) __kmp_yield (); } \ |
1466 | { \ |
1467 | KMP_CPU_PAUSE()__kmp_x86_pause(); \ |
1468 | if ((cond) && (KMP_TRY_YIELD((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && ( ((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc ))))))) \ |
1469 | __kmp_yield(); \ |
1470 | } |
1471 | |
1472 | #define KMP_YIELD_OVERSUB(){ __kmp_x86_pause(); if ((((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)))))) __kmp_yield(); } \ |
1473 | { \ |
1474 | KMP_CPU_PAUSE()__kmp_x86_pause(); \ |
1475 | if ((KMP_TRY_YIELD_OVERSUB((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (( (__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc )))))) \ |
1476 | __kmp_yield(); \ |
1477 | } |
1478 | |
1479 | // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround, |
1480 | // there should be no yielding since initial value from KMP_INIT_YIELD() is odd. |
1481 | #define KMP_YIELD_SPIN(count){ __kmp_x86_pause(); if (((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)))))) { (count) -= 2; if (!(count)) { __kmp_yield (); (count) = __kmp_yield_next; } } } \ |
1482 | { \ |
1483 | KMP_CPU_PAUSE()__kmp_x86_pause(); \ |
1484 | if (KMP_TRY_YIELD((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && ( ((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc )))))) { \ |
1485 | (count) -= 2; \ |
1486 | if (!(count)) { \ |
1487 | __kmp_yield(); \ |
1488 | (count) = __kmp_yield_next; \ |
1489 | } \ |
1490 | } \ |
1491 | } |
1492 | |
1493 | // If TPAUSE is available & enabled, use it. If oversubscribed, use the slower |
1494 | // (C0.2) state, which improves performance of other SMT threads on the same |
1495 | // core, otherwise, use the fast (C0.1) default state, or whatever the user has |
1496 | // requested. Uses a timed TPAUSE, and exponential backoff. If TPAUSE isn't |
1497 | // available, fall back to the regular CPU pause and yield combination. |
1498 | #if KMP_HAVE_UMWAIT((0 || 1) && (1 || 0) && !0) |
1499 | #define KMP_TPAUSE_MAX_MASK((kmp_uint64)0xFFFF) ((kmp_uint64)0xFFFF) |
1500 | #define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time){ if (__kmp_tpause_enabled) { if (((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))) { __kmp_tpause(0, (time) ); } else { __kmp_tpause(__kmp_tpause_hint, (time)); } (time) = (time << 1 | 1) & ((kmp_uint64)0xFFFF); } else { __kmp_x86_pause(); if ((((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)))))) { __kmp_yield(); } else if (__kmp_use_yield == 1) { (count) -= 2; if (!(count)) { __kmp_yield(); (count) = __kmp_yield_next; } } } } \ |
1501 | { \ |
1502 | if (__kmp_tpause_enabled) { \ |
1503 | if (KMP_OVERSUBSCRIBED((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc ))) { \ |
1504 | __kmp_tpause(0, (time)); \ |
1505 | } else { \ |
1506 | __kmp_tpause(__kmp_tpause_hint, (time)); \ |
1507 | } \ |
1508 | (time) = (time << 1 | 1) & KMP_TPAUSE_MAX_MASK((kmp_uint64)0xFFFF); \ |
1509 | } else { \ |
1510 | KMP_CPU_PAUSE()__kmp_x86_pause(); \ |
1511 | if ((KMP_TRY_YIELD_OVERSUB((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (( (__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc )))))) { \ |
1512 | __kmp_yield(); \ |
1513 | } else if (__kmp_use_yield == 1) { \ |
1514 | (count) -= 2; \ |
1515 | if (!(count)) { \ |
1516 | __kmp_yield(); \ |
1517 | (count) = __kmp_yield_next; \ |
1518 | } \ |
1519 | } \ |
1520 | } \ |
1521 | } |
1522 | #else |
1523 | #define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time){ if (__kmp_tpause_enabled) { if (((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))) { __kmp_tpause(0, (time) ); } else { __kmp_tpause(__kmp_tpause_hint, (time)); } (time) = (time << 1 | 1) & ((kmp_uint64)0xFFFF); } else { __kmp_x86_pause(); if ((((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)))))) { __kmp_yield(); } else if (__kmp_use_yield == 1) { (count) -= 2; if (!(count)) { __kmp_yield(); (count) = __kmp_yield_next; } } } } \ |
1524 | { \ |
1525 | KMP_CPU_PAUSE()__kmp_x86_pause(); \ |
1526 | if ((KMP_TRY_YIELD_OVERSUB((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (( (__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc )))))) \ |
1527 | __kmp_yield(); \ |
1528 | else if (__kmp_use_yield == 1) { \ |
1529 | (count) -= 2; \ |
1530 | if (!(count)) { \ |
1531 | __kmp_yield(); \ |
1532 | (count) = __kmp_yield_next; \ |
1533 | } \ |
1534 | } \ |
1535 | } |
1536 | #endif // KMP_HAVE_UMWAIT |
1537 | |
1538 | /* ------------------------------------------------------------------------ */ |
1539 | /* Support datatypes for the orphaned construct nesting checks. */ |
1540 | /* ------------------------------------------------------------------------ */ |
1541 | |
1542 | /* When adding to this enum, add its corresponding string in cons_text_c[] |
1543 | * array in kmp_error.cpp */ |
1544 | enum cons_type { |
1545 | ct_none, |
1546 | ct_parallel, |
1547 | ct_pdo, |
1548 | ct_pdo_ordered, |
1549 | ct_psections, |
1550 | ct_psingle, |
1551 | ct_critical, |
1552 | ct_ordered_in_parallel, |
1553 | ct_ordered_in_pdo, |
1554 | ct_master, |
1555 | ct_reduce, |
1556 | ct_barrier, |
1557 | ct_masked |
1558 | }; |
1559 | |
1560 | #define IS_CONS_TYPE_ORDERED(ct)((ct) == ct_pdo_ordered) ((ct) == ct_pdo_ordered) |
1561 | |
1562 | struct cons_data { |
1563 | ident_t const *ident; |
1564 | enum cons_type type; |
1565 | int prev; |
1566 | kmp_user_lock_p |
1567 | name; /* address exclusively for critical section name comparison */ |
1568 | }; |
1569 | |
1570 | struct cons_header { |
1571 | int p_top, w_top, s_top; |
1572 | int stack_size, stack_top; |
1573 | struct cons_data *stack_data; |
1574 | }; |
1575 | |
1576 | struct kmp_region_info { |
1577 | char *text; |
1578 | int offset[KMP_MAX_FIELDS32]; |
1579 | int length[KMP_MAX_FIELDS32]; |
1580 | }; |
1581 | |
1582 | /* ---------------------------------------------------------------------- */ |
1583 | /* ---------------------------------------------------------------------- */ |
1584 | |
1585 | #if KMP_OS_WINDOWS0 |
1586 | typedef HANDLE kmp_thread_t; |
1587 | typedef DWORD kmp_key_t; |
1588 | #endif /* KMP_OS_WINDOWS */ |
1589 | |
1590 | #if KMP_OS_UNIX1 |
1591 | typedef pthread_t kmp_thread_t; |
1592 | typedef pthread_key_t kmp_key_t; |
1593 | #endif |
1594 | |
1595 | extern kmp_key_t __kmp_gtid_threadprivate_key; |
1596 | |
1597 | typedef struct kmp_sys_info { |
1598 | long maxrss; /* the maximum resident set size utilized (in kilobytes) */ |
1599 | long minflt; /* the number of page faults serviced without any I/O */ |
1600 | long majflt; /* the number of page faults serviced that required I/O */ |
1601 | long nswap; /* the number of times a process was "swapped" out of memory */ |
1602 | long inblock; /* the number of times the file system had to perform input */ |
1603 | long oublock; /* the number of times the file system had to perform output */ |
1604 | long nvcsw; /* the number of times a context switch was voluntarily */ |
1605 | long nivcsw; /* the number of times a context switch was forced */ |
1606 | } kmp_sys_info_t; |
1607 | |
1608 | #if USE_ITT_BUILD1 |
1609 | // We cannot include "kmp_itt.h" due to circular dependency. Declare the only |
1610 | // required type here. Later we will check the type meets requirements. |
1611 | typedef int kmp_itt_mark_t; |
1612 | #define KMP_ITT_DEBUG0 0 |
1613 | #endif /* USE_ITT_BUILD */ |
1614 | |
1615 | typedef kmp_int32 kmp_critical_name[8]; |
1616 | |
1617 | /*! |
1618 | @ingroup PARALLEL |
1619 | The type for a microtask which gets passed to @ref __kmpc_fork_call(). |
1620 | The arguments to the outlined function are |
1621 | @param global_tid the global thread identity of the thread executing the |
1622 | function. |
1623 | @param bound_tid the local identity of the thread executing the function |
1624 | @param ... pointers to shared variables accessed by the function. |
1625 | */ |
1626 | typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...); |
1627 | typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth, |
1628 | ...); |
1629 | |
1630 | /*! |
1631 | @ingroup THREADPRIVATE |
1632 | @{ |
1633 | */ |
1634 | /* --------------------------------------------------------------------------- |
1635 | */ |
1636 | /* Threadprivate initialization/finalization function declarations */ |
1637 | |
1638 | /* for non-array objects: __kmpc_threadprivate_register() */ |
1639 | |
1640 | /*! |
1641 | Pointer to the constructor function. |
1642 | The first argument is the <tt>this</tt> pointer |
1643 | */ |
1644 | typedef void *(*kmpc_ctor)(void *); |
1645 | |
1646 | /*! |
1647 | Pointer to the destructor function. |
1648 | The first argument is the <tt>this</tt> pointer |
1649 | */ |
1650 | typedef void (*kmpc_dtor)( |
1651 | void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel |
1652 | compiler */ |
1653 | /*! |
1654 | Pointer to an alternate constructor. |
1655 | The first argument is the <tt>this</tt> pointer. |
1656 | */ |
1657 | typedef void *(*kmpc_cctor)(void *, void *); |
1658 | |
1659 | /* for array objects: __kmpc_threadprivate_register_vec() */ |
1660 | /* First arg: "this" pointer */ |
1661 | /* Last arg: number of array elements */ |
1662 | /*! |
1663 | Array constructor. |
1664 | First argument is the <tt>this</tt> pointer |
1665 | Second argument the number of array elements. |
1666 | */ |
1667 | typedef void *(*kmpc_ctor_vec)(void *, size_t); |
1668 | /*! |
1669 | Pointer to the array destructor function. |
1670 | The first argument is the <tt>this</tt> pointer |
1671 | Second argument the number of array elements. |
1672 | */ |
1673 | typedef void (*kmpc_dtor_vec)(void *, size_t); |
1674 | /*! |
1675 | Array constructor. |
1676 | First argument is the <tt>this</tt> pointer |
1677 | Third argument the number of array elements. |
1678 | */ |
1679 | typedef void *(*kmpc_cctor_vec)(void *, void *, |
1680 | size_t); /* function unused by compiler */ |
1681 | |
1682 | /*! |
1683 | @} |
1684 | */ |
1685 | |
1686 | /* keeps tracked of threadprivate cache allocations for cleanup later */ |
1687 | typedef struct kmp_cached_addr { |
1688 | void **addr; /* address of allocated cache */ |
1689 | void ***compiler_cache; /* pointer to compiler's cache */ |
1690 | void *data; /* pointer to global data */ |
1691 | struct kmp_cached_addr *next; /* pointer to next cached address */ |
1692 | } kmp_cached_addr_t; |
1693 | |
1694 | struct private_data { |
1695 | struct private_data *next; /* The next descriptor in the list */ |
1696 | void *data; /* The data buffer for this descriptor */ |
1697 | int more; /* The repeat count for this descriptor */ |
1698 | size_t size; /* The data size for this descriptor */ |
1699 | }; |
1700 | |
1701 | struct private_common { |
1702 | struct private_common *next; |
1703 | struct private_common *link; |
1704 | void *gbl_addr; |
1705 | void *par_addr; /* par_addr == gbl_addr for PRIMARY thread */ |
1706 | size_t cmn_size; |
1707 | }; |
1708 | |
1709 | struct shared_common { |
1710 | struct shared_common *next; |
1711 | struct private_data *pod_init; |
1712 | void *obj_init; |
1713 | void *gbl_addr; |
1714 | union { |
1715 | kmpc_ctor ctor; |
1716 | kmpc_ctor_vec ctorv; |
1717 | } ct; |
1718 | union { |
1719 | kmpc_cctor cctor; |
1720 | kmpc_cctor_vec cctorv; |
1721 | } cct; |
1722 | union { |
1723 | kmpc_dtor dtor; |
1724 | kmpc_dtor_vec dtorv; |
1725 | } dt; |
1726 | size_t vec_len; |
1727 | int is_vec; |
1728 | size_t cmn_size; |
1729 | }; |
1730 | |
1731 | #define KMP_HASH_TABLE_LOG29 9 /* log2 of the hash table size */ |
1732 | #define KMP_HASH_TABLE_SIZE(1 << 9) \ |
1733 | (1 << KMP_HASH_TABLE_LOG29) /* size of the hash table */ |
1734 | #define KMP_HASH_SHIFT3 3 /* throw away this many low bits from the address */ |
1735 | #define KMP_HASH(x)((((kmp_uintptr_t)x) >> 3) & ((1 << 9) - 1)) \ |
1736 | ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT3) & (KMP_HASH_TABLE_SIZE(1 << 9) - 1)) |
1737 | |
1738 | struct common_table { |
1739 | struct private_common *data[KMP_HASH_TABLE_SIZE(1 << 9)]; |
1740 | }; |
1741 | |
1742 | struct shared_table { |
1743 | struct shared_common *data[KMP_HASH_TABLE_SIZE(1 << 9)]; |
1744 | }; |
1745 | |
1746 | /* ------------------------------------------------------------------------ */ |
1747 | |
1748 | #if KMP_USE_HIER_SCHED0 |
1749 | // Shared barrier data that exists inside a single unit of the scheduling |
1750 | // hierarchy |
1751 | typedef struct kmp_hier_private_bdata_t { |
1752 | kmp_int32 num_active; |
1753 | kmp_uint64 index; |
1754 | kmp_uint64 wait_val[2]; |
1755 | } kmp_hier_private_bdata_t; |
1756 | #endif |
1757 | |
1758 | typedef struct kmp_sched_flags { |
1759 | unsigned ordered : 1; |
1760 | unsigned nomerge : 1; |
1761 | unsigned contains_last : 1; |
1762 | #if KMP_USE_HIER_SCHED0 |
1763 | unsigned use_hier : 1; |
1764 | unsigned unused : 28; |
1765 | #else |
1766 | unsigned unused : 29; |
1767 | #endif |
1768 | } kmp_sched_flags_t; |
1769 | |
1770 | KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4)static_assert(sizeof(kmp_sched_flags_t) == 4, "Build condition error" ); |
1771 | |
1772 | #if KMP_STATIC_STEAL_ENABLED1 |
1773 | typedef struct KMP_ALIGN_CACHE__attribute__((aligned(64))) dispatch_private_info32 { |
1774 | kmp_int32 count; |
1775 | kmp_int32 ub; |
1776 | /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */ |
1777 | kmp_int32 lb; |
1778 | kmp_int32 st; |
1779 | kmp_int32 tc; |
1780 | kmp_lock_t *steal_lock; // lock used for chunk stealing |
1781 | // KMP_ALIGN(32) ensures (if the KMP_ALIGN macro is turned on) |
1782 | // a) parm3 is properly aligned and |
1783 | // b) all parm1-4 are on the same cache line. |
1784 | // Because of parm1-4 are used together, performance seems to be better |
1785 | // if they are on the same cache line (not measured though). |
1786 | |
1787 | struct KMP_ALIGN(32)__attribute__((aligned(32))) { // AC: changed 16 to 32 in order to simplify template |
1788 | kmp_int32 parm1; // structures in kmp_dispatch.cpp. This should |
1789 | kmp_int32 parm2; // make no real change at least while padding is off. |
1790 | kmp_int32 parm3; |
1791 | kmp_int32 parm4; |
1792 | }; |
1793 | |
1794 | kmp_uint32 ordered_lower; |
1795 | kmp_uint32 ordered_upper; |
1796 | #if KMP_OS_WINDOWS0 |
1797 | kmp_int32 last_upper; |
1798 | #endif /* KMP_OS_WINDOWS */ |
1799 | } dispatch_private_info32_t; |
1800 | |
1801 | typedef struct KMP_ALIGN_CACHE__attribute__((aligned(64))) dispatch_private_info64 { |
1802 | kmp_int64 count; // current chunk number for static & static-steal scheduling |
1803 | kmp_int64 ub; /* upper-bound */ |
1804 | /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */ |
1805 | kmp_int64 lb; /* lower-bound */ |
1806 | kmp_int64 st; /* stride */ |
1807 | kmp_int64 tc; /* trip count (number of iterations) */ |
1808 | kmp_lock_t *steal_lock; // lock used for chunk stealing |
1809 | /* parm[1-4] are used in different ways by different scheduling algorithms */ |
1810 | |
1811 | // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on ) |
1812 | // a) parm3 is properly aligned and |
1813 | // b) all parm1-4 are in the same cache line. |
1814 | // Because of parm1-4 are used together, performance seems to be better |
1815 | // if they are in the same line (not measured though). |
1816 | |
1817 | struct KMP_ALIGN(32)__attribute__((aligned(32))) { |
1818 | kmp_int64 parm1; |
1819 | kmp_int64 parm2; |
1820 | kmp_int64 parm3; |
1821 | kmp_int64 parm4; |
1822 | }; |
1823 | |
1824 | kmp_uint64 ordered_lower; |
1825 | kmp_uint64 ordered_upper; |
1826 | #if KMP_OS_WINDOWS0 |
1827 | kmp_int64 last_upper; |
1828 | #endif /* KMP_OS_WINDOWS */ |
1829 | } dispatch_private_info64_t; |
1830 | #else /* KMP_STATIC_STEAL_ENABLED */ |
1831 | typedef struct KMP_ALIGN_CACHE__attribute__((aligned(64))) dispatch_private_info32 { |
1832 | kmp_int32 lb; |
1833 | kmp_int32 ub; |
1834 | kmp_int32 st; |
1835 | kmp_int32 tc; |
1836 | |
1837 | kmp_int32 parm1; |
1838 | kmp_int32 parm2; |
1839 | kmp_int32 parm3; |
1840 | kmp_int32 parm4; |
1841 | |
1842 | kmp_int32 count; |
1843 | |
1844 | kmp_uint32 ordered_lower; |
1845 | kmp_uint32 ordered_upper; |
1846 | #if KMP_OS_WINDOWS0 |
1847 | kmp_int32 last_upper; |
1848 | #endif /* KMP_OS_WINDOWS */ |
1849 | } dispatch_private_info32_t; |
1850 | |
1851 | typedef struct KMP_ALIGN_CACHE__attribute__((aligned(64))) dispatch_private_info64 { |
1852 | kmp_int64 lb; /* lower-bound */ |
1853 | kmp_int64 ub; /* upper-bound */ |
1854 | kmp_int64 st; /* stride */ |
1855 | kmp_int64 tc; /* trip count (number of iterations) */ |
1856 | |
1857 | /* parm[1-4] are used in different ways by different scheduling algorithms */ |
1858 | kmp_int64 parm1; |
1859 | kmp_int64 parm2; |
1860 | kmp_int64 parm3; |
1861 | kmp_int64 parm4; |
1862 | |
1863 | kmp_int64 count; /* current chunk number for static scheduling */ |
1864 | |
1865 | kmp_uint64 ordered_lower; |
1866 | kmp_uint64 ordered_upper; |
1867 | #if KMP_OS_WINDOWS0 |
1868 | kmp_int64 last_upper; |
1869 | #endif /* KMP_OS_WINDOWS */ |
1870 | } dispatch_private_info64_t; |
1871 | #endif /* KMP_STATIC_STEAL_ENABLED */ |
1872 | |
1873 | typedef struct KMP_ALIGN_CACHE__attribute__((aligned(64))) dispatch_private_info { |
1874 | union private_info { |
1875 | dispatch_private_info32_t p32; |
1876 | dispatch_private_info64_t p64; |
1877 | } u; |
1878 | enum sched_type schedule; /* scheduling algorithm */ |
1879 | kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */ |
1880 | std::atomic<kmp_uint32> steal_flag; // static_steal only, state of a buffer |
1881 | kmp_int32 ordered_bumped; |
1882 | // Stack of buffers for nest of serial regions |
1883 | struct dispatch_private_info *next; |
1884 | kmp_int32 type_size; /* the size of types in private_info */ |
1885 | #if KMP_USE_HIER_SCHED0 |
1886 | kmp_int32 hier_id; |
1887 | void *parent; /* hierarchical scheduling parent pointer */ |
1888 | #endif |
1889 | enum cons_type pushed_ws; |
1890 | } dispatch_private_info_t; |
1891 | |
1892 | typedef struct dispatch_shared_info32 { |
1893 | /* chunk index under dynamic, number of idle threads under static-steal; |
1894 | iteration index otherwise */ |
1895 | volatile kmp_uint32 iteration; |
1896 | volatile kmp_int32 num_done; |
1897 | volatile kmp_uint32 ordered_iteration; |
1898 | // Dummy to retain the structure size after making ordered_iteration scalar |
1899 | kmp_int32 ordered_dummy[KMP_MAX_ORDERED8 - 1]; |
1900 | } dispatch_shared_info32_t; |
1901 | |
1902 | typedef struct dispatch_shared_info64 { |
1903 | /* chunk index under dynamic, number of idle threads under static-steal; |
1904 | iteration index otherwise */ |
1905 | volatile kmp_uint64 iteration; |
1906 | volatile kmp_int64 num_done; |
1907 | volatile kmp_uint64 ordered_iteration; |
1908 | // Dummy to retain the structure size after making ordered_iteration scalar |
1909 | kmp_int64 ordered_dummy[KMP_MAX_ORDERED8 - 3]; |
1910 | } dispatch_shared_info64_t; |
1911 | |
1912 | typedef struct dispatch_shared_info { |
1913 | union shared_info { |
1914 | dispatch_shared_info32_t s32; |
1915 | dispatch_shared_info64_t s64; |
1916 | } u; |
1917 | volatile kmp_uint32 buffer_index; |
1918 | volatile kmp_int32 doacross_buf_idx; // teamwise index |
1919 | volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1) |
1920 | kmp_int32 doacross_num_done; // count finished threads |
1921 | #if KMP_USE_HIER_SCHED0 |
1922 | void *hier; |
1923 | #endif |
1924 | #if KMP_USE_HWLOC0 |
1925 | // When linking with libhwloc, the ORDERED EPCC test slows down on big |
1926 | // machines (> 48 cores). Performance analysis showed that a cache thrash |
1927 | // was occurring and this padding helps alleviate the problem. |
1928 | char padding[64]; |
1929 | #endif |
1930 | } dispatch_shared_info_t; |
1931 | |
1932 | typedef struct kmp_disp { |
1933 | /* Vector for ORDERED SECTION */ |
1934 | void (*th_deo_fcn)(int *gtid, int *cid, ident_t *); |
1935 | /* Vector for END ORDERED SECTION */ |
1936 | void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *); |
1937 | |
1938 | dispatch_shared_info_t *th_dispatch_sh_current; |
1939 | dispatch_private_info_t *th_dispatch_pr_current; |
1940 | |
1941 | dispatch_private_info_t *th_disp_buffer; |
1942 | kmp_uint32 th_disp_index; |
1943 | kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index |
1944 | volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags |
1945 | kmp_int64 *th_doacross_info; // info on loop bounds |
1946 | #if KMP_USE_INTERNODE_ALIGNMENT0 |
1947 | char more_padding[INTERNODE_CACHE_LINE4096]; |
1948 | #endif |
1949 | } kmp_disp_t; |
1950 | |
1951 | /* ------------------------------------------------------------------------ */ |
1952 | /* Barrier stuff */ |
1953 | |
1954 | /* constants for barrier state update */ |
1955 | #define KMP_INIT_BARRIER_STATE0 0 /* should probably start from zero */ |
1956 | #define KMP_BARRIER_SLEEP_BIT0 0 /* bit used for suspend/sleep part of state */ |
1957 | #define KMP_BARRIER_UNUSED_BIT1 1 // bit that must never be set for valid state |
1958 | #define KMP_BARRIER_BUMP_BIT2 2 /* lsb used for bump of go/arrived state */ |
1959 | |
1960 | #define KMP_BARRIER_SLEEP_STATE(1 << 0) (1 << KMP_BARRIER_SLEEP_BIT0) |
1961 | #define KMP_BARRIER_UNUSED_STATE(1 << 1) (1 << KMP_BARRIER_UNUSED_BIT1) |
1962 | #define KMP_BARRIER_STATE_BUMP(1 << 2) (1 << KMP_BARRIER_BUMP_BIT2) |
1963 | |
1964 | #if (KMP_BARRIER_SLEEP_BIT0 >= KMP_BARRIER_BUMP_BIT2) |
1965 | #error "Barrier sleep bit must be smaller than barrier bump bit" |
1966 | #endif |
1967 | #if (KMP_BARRIER_UNUSED_BIT1 >= KMP_BARRIER_BUMP_BIT2) |
1968 | #error "Barrier unused bit must be smaller than barrier bump bit" |
1969 | #endif |
1970 | |
1971 | // Constants for release barrier wait state: currently, hierarchical only |
1972 | #define KMP_BARRIER_NOT_WAITING0 0 // Normal state; worker not in wait_sleep |
1973 | #define KMP_BARRIER_OWN_FLAG1 \ |
1974 | 1 // Normal state; worker waiting on own b_go flag in release |
1975 | #define KMP_BARRIER_PARENT_FLAG2 \ |
1976 | 2 // Special state; worker waiting on parent's b_go flag in release |
1977 | #define KMP_BARRIER_SWITCH_TO_OWN_FLAG3 \ |
1978 | 3 // Special state; tells worker to shift from parent to own b_go |
1979 | #define KMP_BARRIER_SWITCHING4 \ |
1980 | 4 // Special state; worker resets appropriate flag on wake-up |
1981 | |
1982 | #define KMP_NOT_SAFE_TO_REAP0 \ |
1983 | 0 // Thread th_reap_state: not safe to reap (tasking) |
1984 | #define KMP_SAFE_TO_REAP1 1 // Thread th_reap_state: safe to reap (not tasking) |
1985 | |
1986 | // The flag_type describes the storage used for the flag. |
1987 | enum flag_type { |
1988 | flag32, /**< atomic 32 bit flags */ |
1989 | flag64, /**< 64 bit flags */ |
1990 | atomic_flag64, /**< atomic 64 bit flags */ |
1991 | flag_oncore, /**< special 64-bit flag for on-core barrier (hierarchical) */ |
1992 | flag_unset |
1993 | }; |
1994 | |
1995 | enum barrier_type { |
1996 | bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction |
1997 | barriers if enabled) */ |
1998 | bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */ |
1999 | #if KMP_FAST_REDUCTION_BARRIER1 |
2000 | bs_reduction_barrier, /* 2, All barriers that are used in reduction */ |
2001 | #endif // KMP_FAST_REDUCTION_BARRIER |
2002 | bs_last_barrier /* Just a placeholder to mark the end */ |
2003 | }; |
2004 | |
2005 | // to work with reduction barriers just like with plain barriers |
2006 | #if !KMP_FAST_REDUCTION_BARRIER1 |
2007 | #define bs_reduction_barrier bs_plain_barrier |
2008 | #endif // KMP_FAST_REDUCTION_BARRIER |
2009 | |
2010 | typedef enum kmp_bar_pat { /* Barrier communication patterns */ |
2011 | bp_linear_bar = |
2012 | 0, /* Single level (degenerate) tree */ |
2013 | bp_tree_bar = |
2014 | 1, /* Balanced tree with branching factor 2^n */ |
2015 | bp_hyper_bar = 2, /* Hypercube-embedded tree with min |
2016 | branching factor 2^n */ |
2017 | bp_hierarchical_bar = 3, /* Machine hierarchy tree */ |
2018 | bp_dist_bar = 4, /* Distributed barrier */ |
2019 | bp_last_bar /* Placeholder to mark the end */ |
2020 | } kmp_bar_pat_e; |
2021 | |
2022 | #define KMP_BARRIER_ICV_PUSH1 1 |
2023 | |
2024 | /* Record for holding the values of the internal controls stack records */ |
2025 | typedef struct kmp_internal_control { |
2026 | int serial_nesting_level; /* corresponds to the value of the |
2027 | th_team_serialized field */ |
2028 | kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per |
2029 | thread) */ |
2030 | kmp_int8 |
2031 | bt_set; /* internal control for whether blocktime is explicitly set */ |
2032 | int blocktime; /* internal control for blocktime */ |
2033 | #if KMP_USE_MONITOR |
2034 | int bt_intervals; /* internal control for blocktime intervals */ |
2035 | #endif |
2036 | int nproc; /* internal control for #threads for next parallel region (per |
2037 | thread) */ |
2038 | int thread_limit; /* internal control for thread-limit-var */ |
2039 | int max_active_levels; /* internal control for max_active_levels */ |
2040 | kmp_r_sched_t |
2041 | sched; /* internal control for runtime schedule {sched,chunk} pair */ |
2042 | kmp_proc_bind_t proc_bind; /* internal control for affinity */ |
2043 | kmp_int32 default_device; /* internal control for default device */ |
2044 | struct kmp_internal_control *next; |
2045 | } kmp_internal_control_t; |
2046 | |
2047 | static inline void copy_icvs(kmp_internal_control_t *dst, |
2048 | kmp_internal_control_t *src) { |
2049 | *dst = *src; |
2050 | } |
2051 | |
2052 | /* Thread barrier needs volatile barrier fields */ |
2053 | typedef struct KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_bstate { |
2054 | // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all |
2055 | // uses of it). It is not explicitly aligned below, because we *don't* want |
2056 | // it to be padded -- instead, we fit b_go into the same cache line with |
2057 | // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier. |
2058 | kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread |
2059 | // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with |
2060 | // same NGO store |
2061 | volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical) |
2062 | KMP_ALIGN_CACHE__attribute__((aligned(64))) volatile kmp_uint64 |
2063 | b_arrived; // STATE => task reached synch point. |
2064 | kmp_uint32 *skip_per_level; |
2065 | kmp_uint32 my_level; |
2066 | kmp_int32 parent_tid; |
2067 | kmp_int32 old_tid; |
2068 | kmp_uint32 depth; |
2069 | struct kmp_bstate *parent_bar; |
2070 | kmp_team_t *team; |
2071 | kmp_uint64 leaf_state; |
2072 | kmp_uint32 nproc; |
2073 | kmp_uint8 base_leaf_kids; |
2074 | kmp_uint8 leaf_kids; |
2075 | kmp_uint8 offset; |
2076 | kmp_uint8 wait_flag; |
2077 | kmp_uint8 use_oncore_barrier; |
2078 | #if USE_DEBUGGER0 |
2079 | // The following field is intended for the debugger solely. Only the worker |
2080 | // thread itself accesses this field: the worker increases it by 1 when it |
2081 | // arrives to a barrier. |
2082 | KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_uint b_worker_arrived; |
2083 | #endif /* USE_DEBUGGER */ |
2084 | } kmp_bstate_t; |
2085 | |
2086 | union KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_barrier_union { |
2087 | double b_align; /* use worst case alignment */ |
2088 | char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)(sizeof(kmp_bstate_t) + (64 - ((sizeof(kmp_bstate_t) - 1) % ( 64)) - 1))]; |
2089 | kmp_bstate_t bb; |
2090 | }; |
2091 | |
2092 | typedef union kmp_barrier_union kmp_balign_t; |
2093 | |
2094 | /* Team barrier needs only non-volatile arrived counter */ |
2095 | union KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_barrier_team_union { |
2096 | double b_align; /* use worst case alignment */ |
2097 | char b_pad[CACHE_LINE64]; |
2098 | struct { |
2099 | kmp_uint64 b_arrived; /* STATE => task reached synch point. */ |
2100 | #if USE_DEBUGGER0 |
2101 | // The following two fields are indended for the debugger solely. Only |
2102 | // primary thread of the team accesses these fields: the first one is |
2103 | // increased by 1 when the primary thread arrives to a barrier, the second |
2104 | // one is increased by one when all the threads arrived. |
2105 | kmp_uint b_master_arrived; |
2106 | kmp_uint b_team_arrived; |
2107 | #endif |
2108 | }; |
2109 | }; |
2110 | |
2111 | typedef union kmp_barrier_team_union kmp_balign_team_t; |
2112 | |
2113 | /* Padding for Linux* OS pthreads condition variables and mutexes used to signal |
2114 | threads when a condition changes. This is to workaround an NPTL bug where |
2115 | padding was added to pthread_cond_t which caused the initialization routine |
2116 | to write outside of the structure if compiled on pre-NPTL threads. */ |
2117 | #if KMP_OS_WINDOWS0 |
2118 | typedef struct kmp_win32_mutex { |
2119 | /* The Lock */ |
2120 | CRITICAL_SECTION cs; |
2121 | } kmp_win32_mutex_t; |
2122 | |
2123 | typedef struct kmp_win32_cond { |
2124 | /* Count of the number of waiters. */ |
2125 | int waiters_count_; |
2126 | |
2127 | /* Serialize access to <waiters_count_> */ |
2128 | kmp_win32_mutex_t waiters_count_lock_; |
2129 | |
2130 | /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */ |
2131 | int release_count_; |
2132 | |
2133 | /* Keeps track of the current "generation" so that we don't allow */ |
2134 | /* one thread to steal all the "releases" from the broadcast. */ |
2135 | int wait_generation_count_; |
2136 | |
2137 | /* A manual-reset event that's used to block and release waiting threads. */ |
2138 | HANDLE event_; |
2139 | } kmp_win32_cond_t; |
2140 | #endif |
2141 | |
2142 | #if KMP_OS_UNIX1 |
2143 | |
2144 | union KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_cond_union { |
2145 | double c_align; |
2146 | char c_pad[CACHE_LINE64]; |
2147 | pthread_cond_t c_cond; |
2148 | }; |
2149 | |
2150 | typedef union kmp_cond_union kmp_cond_align_t; |
2151 | |
2152 | union KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_mutex_union { |
2153 | double m_align; |
2154 | char m_pad[CACHE_LINE64]; |
2155 | pthread_mutex_t m_mutex; |
2156 | }; |
2157 | |
2158 | typedef union kmp_mutex_union kmp_mutex_align_t; |
2159 | |
2160 | #endif /* KMP_OS_UNIX */ |
2161 | |
2162 | typedef struct kmp_desc_base { |
2163 | void *ds_stackbase; |
2164 | size_t ds_stacksize; |
2165 | int ds_stackgrow; |
2166 | kmp_thread_t ds_thread; |
2167 | volatile int ds_tid; |
2168 | int ds_gtid; |
2169 | #if KMP_OS_WINDOWS0 |
2170 | volatile int ds_alive; |
2171 | DWORD ds_thread_id; |
2172 | /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes. |
2173 | However, debugger support (libomp_db) cannot work with handles, because they |
2174 | uncomparable. For example, debugger requests info about thread with handle h. |
2175 | h is valid within debugger process, and meaningless within debugee process. |
2176 | Even if h is duped by call to DuplicateHandle(), so the result h' is valid |
2177 | within debugee process, but it is a *new* handle which does *not* equal to |
2178 | any other handle in debugee... The only way to compare handles is convert |
2179 | them to system-wide ids. GetThreadId() function is available only in |
2180 | Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available |
2181 | on all Windows* OS flavours (including Windows* 95). Thus, we have to get |
2182 | thread id by call to GetCurrentThreadId() from within the thread and save it |
2183 | to let libomp_db identify threads. */ |
2184 | #endif /* KMP_OS_WINDOWS */ |
2185 | } kmp_desc_base_t; |
2186 | |
2187 | typedef union KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_desc { |
2188 | double ds_align; /* use worst case alignment */ |
2189 | char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)(sizeof(kmp_desc_base_t) + (64 - ((sizeof(kmp_desc_base_t) - 1 ) % (64)) - 1))]; |
2190 | kmp_desc_base_t ds; |
2191 | } kmp_desc_t; |
2192 | |
2193 | typedef struct kmp_local { |
2194 | volatile int this_construct; /* count of single's encountered by thread */ |
2195 | void *reduce_data; |
2196 | #if KMP_USE_BGET1 |
2197 | void *bget_data; |
2198 | void *bget_list; |
2199 | #if !USE_CMP_XCHG_FOR_BGET1 |
2200 | #ifdef USE_QUEUING_LOCK_FOR_BGET |
2201 | kmp_lock_t bget_lock; /* Lock for accessing bget free list */ |
2202 | #else |
2203 | kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be |
2204 | // bootstrap lock so we can use it at library |
2205 | // shutdown. |
2206 | #endif /* USE_LOCK_FOR_BGET */ |
2207 | #endif /* ! USE_CMP_XCHG_FOR_BGET */ |
2208 | #endif /* KMP_USE_BGET */ |
2209 | |
2210 | PACKED_REDUCTION_METHOD_T |
2211 | packed_reduction_method; /* stored by __kmpc_reduce*(), used by |
2212 | __kmpc_end_reduce*() */ |
2213 | |
2214 | } kmp_local_t; |
2215 | |
2216 | #define KMP_CHECK_UPDATE(a, b)if ((a) != (b)) (a) = (b) \ |
2217 | if ((a) != (b)) \ |
2218 | (a) = (b) |
2219 | #define KMP_CHECK_UPDATE_SYNC(a, b)if ((a) != (b)) (((a))) = (((b))) \ |
2220 | if ((a) != (b)) \ |
2221 | TCW_SYNC_PTR((a), (b))(((a))) = (((b))) |
2222 | |
2223 | #define get__blocktime(xteam, xtid)((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs .blocktime) \ |
2224 | ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) |
2225 | #define get__bt_set(xteam, xtid)((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs .bt_set) \ |
2226 | ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) |
2227 | #if KMP_USE_MONITOR |
2228 | #define get__bt_intervals(xteam, xtid) \ |
2229 | ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) |
2230 | #endif |
2231 | |
2232 | #define get__dynamic_2(xteam, xtid)((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs .dynamic) \ |
2233 | ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic) |
2234 | #define get__nproc_2(xteam, xtid)((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs .nproc) \ |
2235 | ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc) |
2236 | #define get__sched_2(xteam, xtid)((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs .sched) \ |
2237 | ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched) |
2238 | |
2239 | #define set__blocktime_team(xteam, xtid, xval)(((xteam)->t.t_threads[(xtid)]->th.th_current_task-> td_icvs.blocktime) = (xval)) \ |
2240 | (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \ |
2241 | (xval)) |
2242 | |
2243 | #if KMP_USE_MONITOR |
2244 | #define set__bt_intervals_team(xteam, xtid, xval) \ |
2245 | (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \ |
2246 | (xval)) |
2247 | #endif |
2248 | |
2249 | #define set__bt_set_team(xteam, xtid, xval)(((xteam)->t.t_threads[(xtid)]->th.th_current_task-> td_icvs.bt_set) = (xval)) \ |
2250 | (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval)) |
2251 | |
2252 | #define set__dynamic(xthread, xval)(((xthread)->th.th_current_task->td_icvs.dynamic) = (xval )) \ |
2253 | (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval)) |
2254 | #define get__dynamic(xthread)(((xthread)->th.th_current_task->td_icvs.dynamic) ? ((! 0)) : (0)) \ |
2255 | (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE(!0)) : (FTN_FALSE0)) |
2256 | |
2257 | #define set__nproc(xthread, xval)(((xthread)->th.th_current_task->td_icvs.nproc) = (xval )) \ |
2258 | (((xthread)->th.th_current_task->td_icvs.nproc) = (xval)) |
2259 | |
2260 | #define set__thread_limit(xthread, xval)(((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval)) \ |
2261 | (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval)) |
2262 | |
2263 | #define set__max_active_levels(xthread, xval)(((xthread)->th.th_current_task->td_icvs.max_active_levels ) = (xval)) \ |
2264 | (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval)) |
2265 | |
2266 | #define get__max_active_levels(xthread)((xthread)->th.th_current_task->td_icvs.max_active_levels ) \ |
2267 | ((xthread)->th.th_current_task->td_icvs.max_active_levels) |
2268 | |
2269 | #define set__sched(xthread, xval)(((xthread)->th.th_current_task->td_icvs.sched) = (xval )) \ |
2270 | (((xthread)->th.th_current_task->td_icvs.sched) = (xval)) |
2271 | |
2272 | #define set__proc_bind(xthread, xval)(((xthread)->th.th_current_task->td_icvs.proc_bind) = ( xval)) \ |
2273 | (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval)) |
2274 | #define get__proc_bind(xthread)((xthread)->th.th_current_task->td_icvs.proc_bind) \ |
2275 | ((xthread)->th.th_current_task->td_icvs.proc_bind) |
2276 | |
2277 | // OpenMP tasking data structures |
2278 | |
2279 | typedef enum kmp_tasking_mode { |
2280 | tskm_immediate_exec = 0, |
2281 | tskm_extra_barrier = 1, |
2282 | tskm_task_teams = 2, |
2283 | tskm_max = 2 |
2284 | } kmp_tasking_mode_t; |
2285 | |
2286 | extern kmp_tasking_mode_t |
2287 | __kmp_tasking_mode; /* determines how/when to execute tasks */ |
2288 | extern int __kmp_task_stealing_constraint; |
2289 | extern int __kmp_enable_task_throttling; |
2290 | extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if |
2291 | // specified, defaults to 0 otherwise |
2292 | // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise |
2293 | extern kmp_int32 __kmp_max_task_priority; |
2294 | // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise |
2295 | extern kmp_uint64 __kmp_taskloop_min_tasks; |
2296 | |
2297 | /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with |
2298 | taskdata first */ |
2299 | #define KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1) (((kmp_taskdata_t *)task) - 1) |
2300 | #define KMP_TASKDATA_TO_TASK(taskdata)(kmp_task_t *)(taskdata + 1) (kmp_task_t *)(taskdata + 1) |
2301 | |
2302 | // The tt_found_tasks flag is a signal to all threads in the team that tasks |
2303 | // were spawned and queued since the previous barrier release. |
2304 | #define KMP_TASKING_ENABLED(task_team)((!0) == ((task_team)->tt.tt_found_tasks)) \ |
2305 | (TRUE(!0) == TCR_SYNC_4((task_team)->tt.tt_found_tasks)((task_team)->tt.tt_found_tasks)) |
2306 | /*! |
2307 | @ingroup BASIC_TYPES |
2308 | @{ |
2309 | */ |
2310 | |
2311 | /*! |
2312 | */ |
2313 | typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *); |
2314 | |
2315 | typedef union kmp_cmplrdata { |
2316 | kmp_int32 priority; /**< priority specified by user for the task */ |
2317 | kmp_routine_entry_t |
2318 | destructors; /* pointer to function to invoke deconstructors of |
2319 | firstprivate C++ objects */ |
2320 | /* future data */ |
2321 | } kmp_cmplrdata_t; |
2322 | |
2323 | /* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */ |
2324 | /*! |
2325 | */ |
2326 | typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */ |
2327 | void *shareds; /**< pointer to block of pointers to shared vars */ |
2328 | kmp_routine_entry_t |
2329 | routine; /**< pointer to routine to call for executing task */ |
2330 | kmp_int32 part_id; /**< part id for the task */ |
2331 | kmp_cmplrdata_t |
2332 | data1; /* Two known optional additions: destructors and priority */ |
2333 | kmp_cmplrdata_t data2; /* Process destructors first, priority second */ |
2334 | /* future data */ |
2335 | /* private vars */ |
2336 | } kmp_task_t; |
2337 | |
2338 | /*! |
2339 | @} |
2340 | */ |
2341 | |
2342 | typedef struct kmp_taskgroup { |
2343 | std::atomic<kmp_int32> count; // number of allocated and incomplete tasks |
2344 | std::atomic<kmp_int32> |
2345 | cancel_request; // request for cancellation of this taskgroup |
2346 | struct kmp_taskgroup *parent; // parent taskgroup |
2347 | // Block of data to perform task reduction |
2348 | void *reduce_data; // reduction related info |
2349 | kmp_int32 reduce_num_data; // number of data items to reduce |
2350 | uintptr_t *gomp_data; // gomp reduction data |
2351 | } kmp_taskgroup_t; |
2352 | |
2353 | // forward declarations |
2354 | typedef union kmp_depnode kmp_depnode_t; |
2355 | typedef struct kmp_depnode_list kmp_depnode_list_t; |
2356 | typedef struct kmp_dephash_entry kmp_dephash_entry_t; |
2357 | |
2358 | // macros for checking dep flag as an integer |
2359 | #define KMP_DEP_IN0x1 0x1 |
2360 | #define KMP_DEP_OUT0x2 0x2 |
2361 | #define KMP_DEP_INOUT0x3 0x3 |
2362 | #define KMP_DEP_MTX0x4 0x4 |
2363 | #define KMP_DEP_SET0x8 0x8 |
2364 | #define KMP_DEP_ALL0x80 0x80 |
2365 | // Compiler sends us this info: |
2366 | typedef struct kmp_depend_info { |
2367 | kmp_intptr_t base_addr; |
2368 | size_t len; |
2369 | union { |
2370 | kmp_uint8 flag; // flag as an unsigned char |
2371 | struct { // flag as a set of 8 bits |
2372 | unsigned in : 1; |
2373 | unsigned out : 1; |
2374 | unsigned mtx : 1; |
2375 | unsigned set : 1; |
2376 | unsigned unused : 3; |
2377 | unsigned all : 1; |
2378 | } flags; |
2379 | }; |
2380 | } kmp_depend_info_t; |
2381 | |
2382 | // Internal structures to work with task dependencies: |
2383 | struct kmp_depnode_list { |
2384 | kmp_depnode_t *node; |
2385 | kmp_depnode_list_t *next; |
2386 | }; |
2387 | |
2388 | // Max number of mutexinoutset dependencies per node |
2389 | #define MAX_MTX_DEPS4 4 |
2390 | |
2391 | typedef struct kmp_base_depnode { |
2392 | kmp_depnode_list_t *successors; /* used under lock */ |
2393 | kmp_task_t *task; /* non-NULL if depnode is active, used under lock */ |
2394 | kmp_lock_t *mtx_locks[MAX_MTX_DEPS4]; /* lock mutexinoutset dependent tasks */ |
2395 | kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */ |
2396 | kmp_lock_t lock; /* guards shared fields: task, successors */ |
2397 | #if KMP_SUPPORT_GRAPH_OUTPUT |
2398 | kmp_uint32 id; |
2399 | #endif |
2400 | std::atomic<kmp_int32> npredecessors; |
2401 | std::atomic<kmp_int32> nrefs; |
2402 | } kmp_base_depnode_t; |
2403 | |
2404 | union KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_depnode { |
2405 | double dn_align; /* use worst case alignment */ |
2406 | char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)(sizeof(kmp_base_depnode_t) + (64 - ((sizeof(kmp_base_depnode_t ) - 1) % (64)) - 1))]; |
2407 | kmp_base_depnode_t dn; |
2408 | }; |
2409 | |
2410 | struct kmp_dephash_entry { |
2411 | kmp_intptr_t addr; |
2412 | kmp_depnode_t *last_out; |
2413 | kmp_depnode_list_t *last_set; |
2414 | kmp_depnode_list_t *prev_set; |
2415 | kmp_uint8 last_flag; |
2416 | kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */ |
2417 | kmp_dephash_entry_t *next_in_bucket; |
2418 | }; |
2419 | |
2420 | typedef struct kmp_dephash { |
2421 | kmp_dephash_entry_t **buckets; |
2422 | size_t size; |
2423 | kmp_depnode_t *last_all; |
2424 | size_t generation; |
2425 | kmp_uint32 nelements; |
2426 | kmp_uint32 nconflicts; |
2427 | } kmp_dephash_t; |
2428 | |
2429 | typedef struct kmp_task_affinity_info { |
2430 | kmp_intptr_t base_addr; |
2431 | size_t len; |
2432 | struct { |
2433 | bool flag1 : 1; |
2434 | bool flag2 : 1; |
2435 | kmp_int32 reserved : 30; |
2436 | } flags; |
2437 | } kmp_task_affinity_info_t; |
2438 | |
2439 | typedef enum kmp_event_type_t { |
2440 | KMP_EVENT_UNINITIALIZED = 0, |
2441 | KMP_EVENT_ALLOW_COMPLETION = 1 |
2442 | } kmp_event_type_t; |
2443 | |
2444 | typedef struct { |
2445 | kmp_event_type_t type; |
2446 | kmp_tas_lock_t lock; |
2447 | union { |
2448 | kmp_task_t *task; |
2449 | } ed; |
2450 | } kmp_event_t; |
2451 | |
2452 | #ifdef BUILD_TIED_TASK_STACK |
2453 | |
2454 | /* Tied Task stack definitions */ |
2455 | typedef struct kmp_stack_block { |
2456 | kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE]; |
2457 | struct kmp_stack_block *sb_next; |
2458 | struct kmp_stack_block *sb_prev; |
2459 | } kmp_stack_block_t; |
2460 | |
2461 | typedef struct kmp_task_stack { |
2462 | kmp_stack_block_t ts_first_block; // first block of stack entries |
2463 | kmp_taskdata_t **ts_top; // pointer to the top of stack |
2464 | kmp_int32 ts_entries; // number of entries on the stack |
2465 | } kmp_task_stack_t; |
2466 | |
2467 | #endif // BUILD_TIED_TASK_STACK |
2468 | |
2469 | typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */ |
2470 | /* Compiler flags */ /* Total compiler flags must be 16 bits */ |
2471 | unsigned tiedness : 1; /* task is either tied (1) or untied (0) */ |
2472 | unsigned final : 1; /* task is final(1) so execute immediately */ |
2473 | unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0 |
2474 | code path */ |
2475 | unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to |
2476 | invoke destructors from the runtime */ |
2477 | unsigned proxy : 1; /* task is a proxy task (it will be executed outside the |
2478 | context of the RTL) */ |
2479 | unsigned priority_specified : 1; /* set if the compiler provides priority |
2480 | setting for the task */ |
2481 | unsigned detachable : 1; /* 1 == can detach */ |
2482 | unsigned hidden_helper : 1; /* 1 == hidden helper task */ |
2483 | unsigned reserved : 8; /* reserved for compiler use */ |
2484 | |
2485 | /* Library flags */ /* Total library flags must be 16 bits */ |
2486 | unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */ |
2487 | unsigned task_serial : 1; // task is executed immediately (1) or deferred (0) |
2488 | unsigned tasking_ser : 1; // all tasks in team are either executed immediately |
2489 | // (1) or may be deferred (0) |
2490 | unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel |
2491 | // (0) [>= 2 threads] |
2492 | /* If either team_serial or tasking_ser is set, task team may be NULL */ |
2493 | /* Task State Flags: */ |
2494 | unsigned started : 1; /* 1==started, 0==not started */ |
2495 | unsigned executing : 1; /* 1==executing, 0==not executing */ |
2496 | unsigned complete : 1; /* 1==complete, 0==not complete */ |
2497 | unsigned freed : 1; /* 1==freed, 0==allocated */ |
2498 | unsigned native : 1; /* 1==gcc-compiled task, 0==intel */ |
2499 | unsigned reserved31 : 7; /* reserved for library use */ |
2500 | |
2501 | } kmp_tasking_flags_t; |
2502 | |
2503 | struct kmp_taskdata { /* aligned during dynamic allocation */ |
2504 | kmp_int32 td_task_id; /* id, assigned by debugger */ |
2505 | kmp_tasking_flags_t td_flags; /* task flags */ |
2506 | kmp_team_t *td_team; /* team for this task */ |
2507 | kmp_info_p *td_alloc_thread; /* thread that allocated data structures */ |
2508 | /* Currently not used except for perhaps IDB */ |
2509 | kmp_taskdata_t *td_parent; /* parent task */ |
2510 | kmp_int32 td_level; /* task nesting level */ |
2511 | std::atomic<kmp_int32> td_untied_count; // untied task active parts counter |
2512 | ident_t *td_ident; /* task identifier */ |
2513 | // Taskwait data. |
2514 | ident_t *td_taskwait_ident; |
2515 | kmp_uint32 td_taskwait_counter; |
2516 | kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */ |
2517 | KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_internal_control_t |
2518 | td_icvs; /* Internal control variables for the task */ |
2519 | KMP_ALIGN_CACHE__attribute__((aligned(64))) std::atomic<kmp_int32> |
2520 | td_allocated_child_tasks; /* Child tasks (+ current task) not yet |
2521 | deallocated */ |
2522 | std::atomic<kmp_int32> |
2523 | td_incomplete_child_tasks; /* Child tasks not yet complete */ |
2524 | kmp_taskgroup_t |
2525 | *td_taskgroup; // Each task keeps pointer to its current taskgroup |
2526 | kmp_dephash_t |
2527 | *td_dephash; // Dependencies for children tasks are tracked from here |
2528 | kmp_depnode_t |
2529 | *td_depnode; // Pointer to graph node if this task has dependencies |
2530 | kmp_task_team_t *td_task_team; |
2531 | size_t td_size_alloc; // Size of task structure, including shareds etc. |
2532 | #if defined(KMP_GOMP_COMPAT) |
2533 | // 4 or 8 byte integers for the loop bounds in GOMP_taskloop |
2534 | kmp_int32 td_size_loop_bounds; |
2535 | #endif |
2536 | kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint |
2537 | #if defined(KMP_GOMP_COMPAT) |
2538 | // GOMP sends in a copy function for copy constructors |
2539 | void (*td_copy_func)(void *, void *); |
2540 | #endif |
2541 | kmp_event_t td_allow_completion_event; |
2542 | #if OMPT_SUPPORT1 |
2543 | ompt_task_info_t ompt_task_info; |
2544 | #endif |
2545 | }; // struct kmp_taskdata |
2546 | |
2547 | // Make sure padding above worked |
2548 | KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0)static_assert(sizeof(kmp_taskdata_t) % sizeof(void *) == 0, "Build condition error" ); |
2549 | |
2550 | // Data for task team but per thread |
2551 | typedef struct kmp_base_thread_data { |
2552 | kmp_info_p *td_thr; // Pointer back to thread info |
2553 | // Used only in __kmp_execute_tasks_template, maybe not avail until task is |
2554 | // queued? |
2555 | kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque |
2556 | kmp_taskdata_t * |
2557 | *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated |
2558 | kmp_int32 td_deque_size; // Size of deck |
2559 | kmp_uint32 td_deque_head; // Head of deque (will wrap) |
2560 | kmp_uint32 td_deque_tail; // Tail of deque (will wrap) |
2561 | kmp_int32 td_deque_ntasks; // Number of tasks in deque |
2562 | // GEH: shouldn't this be volatile since used in while-spin? |
2563 | kmp_int32 td_deque_last_stolen; // Thread number of last successful steal |
2564 | #ifdef BUILD_TIED_TASK_STACK |
2565 | kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task |
2566 | // scheduling constraint |
2567 | #endif // BUILD_TIED_TASK_STACK |
2568 | } kmp_base_thread_data_t; |
2569 | |
2570 | #define TASK_DEQUE_BITS8 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE |
2571 | #define INITIAL_TASK_DEQUE_SIZE(1 << 8) (1 << TASK_DEQUE_BITS8) |
2572 | |
2573 | #define TASK_DEQUE_SIZE(td)((td).td_deque_size) ((td).td_deque_size) |
2574 | #define TASK_DEQUE_MASK(td)((td).td_deque_size - 1) ((td).td_deque_size - 1) |
2575 | |
2576 | typedef union KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_thread_data { |
2577 | kmp_base_thread_data_t td; |
2578 | double td_align; /* use worst case alignment */ |
2579 | char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)(sizeof(kmp_base_thread_data_t) + (64 - ((sizeof(kmp_base_thread_data_t ) - 1) % (64)) - 1))]; |
2580 | } kmp_thread_data_t; |
2581 | |
2582 | typedef struct kmp_task_pri { |
2583 | kmp_thread_data_t td; |
2584 | kmp_int32 priority; |
2585 | kmp_task_pri *next; |
2586 | } kmp_task_pri_t; |
2587 | |
2588 | // Data for task teams which are used when tasking is enabled for the team |
2589 | typedef struct kmp_base_task_team { |
2590 | kmp_bootstrap_lock_t |
2591 | tt_threads_lock; /* Lock used to allocate per-thread part of task team */ |
2592 | /* must be bootstrap lock since used at library shutdown*/ |
2593 | |
2594 | // TODO: check performance vs kmp_tas_lock_t |
2595 | kmp_bootstrap_lock_t tt_task_pri_lock; /* Lock to access priority tasks */ |
2596 | kmp_task_pri_t *tt_task_pri_list; |
2597 | |
2598 | kmp_task_team_t *tt_next; /* For linking the task team free list */ |
2599 | kmp_thread_data_t |
2600 | *tt_threads_data; /* Array of per-thread structures for task team */ |
2601 | /* Data survives task team deallocation */ |
2602 | kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while |
2603 | executing this team? */ |
2604 | /* TRUE means tt_threads_data is set up and initialized */ |
2605 | kmp_int32 tt_nproc; /* #threads in team */ |
2606 | kmp_int32 tt_max_threads; // # entries allocated for threads_data array |
2607 | kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier |
2608 | kmp_int32 tt_untied_task_encountered; |
2609 | std::atomic<kmp_int32> tt_num_task_pri; // number of priority tasks enqueued |
2610 | // There is hidden helper thread encountered in this task team so that we must |
2611 | // wait when waiting on task team |
2612 | kmp_int32 tt_hidden_helper_task_encountered; |
2613 | |
2614 | KMP_ALIGN_CACHE__attribute__((aligned(64))) |
2615 | std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */ |
2616 | |
2617 | KMP_ALIGN_CACHE__attribute__((aligned(64))) |
2618 | volatile kmp_uint32 |
2619 | tt_active; /* is the team still actively executing tasks */ |
2620 | } kmp_base_task_team_t; |
2621 | |
2622 | union KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_task_team { |
2623 | kmp_base_task_team_t tt; |
2624 | double tt_align; /* use worst case alignment */ |
2625 | char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)(sizeof(kmp_base_task_team_t) + (64 - ((sizeof(kmp_base_task_team_t ) - 1) % (64)) - 1))]; |
2626 | }; |
2627 | |
2628 | #if (USE_FAST_MEMORY3 == 3) || (USE_FAST_MEMORY3 == 5) |
2629 | // Free lists keep same-size free memory slots for fast memory allocation |
2630 | // routines |
2631 | typedef struct kmp_free_list { |
2632 | void *th_free_list_self; // Self-allocated tasks free list |
2633 | void *th_free_list_sync; // Self-allocated tasks stolen/returned by other |
2634 | // threads |
2635 | void *th_free_list_other; // Non-self free list (to be returned to owner's |
2636 | // sync list) |
2637 | } kmp_free_list_t; |
2638 | #endif |
2639 | #if KMP_NESTED_HOT_TEAMS1 |
2640 | // Hot teams array keeps hot teams and their sizes for given thread. Hot teams |
2641 | // are not put in teams pool, and they don't put threads in threads pool. |
2642 | typedef struct kmp_hot_team_ptr { |
2643 | kmp_team_p *hot_team; // pointer to hot_team of given nesting level |
2644 | kmp_int32 hot_team_nth; // number of threads allocated for the hot_team |
2645 | } kmp_hot_team_ptr_t; |
2646 | #endif |
2647 | typedef struct kmp_teams_size { |
2648 | kmp_int32 nteams; // number of teams in a league |
2649 | kmp_int32 nth; // number of threads in each team of the league |
2650 | } kmp_teams_size_t; |
2651 | |
2652 | // This struct stores a thread that acts as a "root" for a contention |
2653 | // group. Contention groups are rooted at kmp_root threads, but also at |
2654 | // each primary thread of each team created in the teams construct. |
2655 | // This struct therefore also stores a thread_limit associated with |
2656 | // that contention group, and a counter to track the number of threads |
2657 | // active in that contention group. Each thread has a list of these: CG |
2658 | // root threads have an entry in their list in which cg_root refers to |
2659 | // the thread itself, whereas other workers in the CG will have a |
2660 | // single entry where cg_root is same as the entry containing their CG |
2661 | // root. When a thread encounters a teams construct, it will add a new |
2662 | // entry to the front of its list, because it now roots a new CG. |
2663 | typedef struct kmp_cg_root { |
2664 | kmp_info_p *cg_root; // "root" thread for a contention group |
2665 | // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or |
2666 | // thread_limit clause for teams primary threads |
2667 | kmp_int32 cg_thread_limit; |
2668 | kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root |
2669 | struct kmp_cg_root *up; // pointer to higher level CG root in list |
2670 | } kmp_cg_root_t; |
2671 | |
2672 | // OpenMP thread data structures |
2673 | |
2674 | typedef struct KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_base_info { |
2675 | /* Start with the readonly data which is cache aligned and padded. This is |
2676 | written before the thread starts working by the primary thread. Uber |
2677 | masters may update themselves later. Usage does not consider serialized |
2678 | regions. */ |
2679 | kmp_desc_t th_info; |
2680 | kmp_team_p *th_team; /* team we belong to */ |
2681 | kmp_root_p *th_root; /* pointer to root of task hierarchy */ |
2682 | kmp_info_p *th_next_pool; /* next available thread in the pool */ |
2683 | kmp_disp_t *th_dispatch; /* thread's dispatch data */ |
2684 | int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */ |
2685 | |
2686 | /* The following are cached from the team info structure */ |
2687 | /* TODO use these in more places as determined to be needed via profiling */ |
2688 | int th_team_nproc; /* number of threads in a team */ |
2689 | kmp_info_p *th_team_master; /* the team's primary thread */ |
2690 | int th_team_serialized; /* team is serialized */ |
2691 | microtask_t th_teams_microtask; /* save entry address for teams construct */ |
2692 | int th_teams_level; /* save initial level of teams construct */ |
2693 | /* it is 0 on device but may be any on host */ |
2694 | |
2695 | /* The blocktime info is copied from the team struct to the thread struct */ |
2696 | /* at the start of a barrier, and the values stored in the team are used */ |
2697 | /* at points in the code where the team struct is no longer guaranteed */ |
2698 | /* to exist (from the POV of worker threads). */ |
2699 | #if KMP_USE_MONITOR |
2700 | int th_team_bt_intervals; |
2701 | int th_team_bt_set; |
2702 | #else |
2703 | kmp_uint64 th_team_bt_intervals; |
2704 | #endif |
2705 | |
2706 | #if KMP_AFFINITY_SUPPORTED1 |
2707 | kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */ |
2708 | #endif |
2709 | omp_allocator_handle_t th_def_allocator; /* default allocator */ |
2710 | /* The data set by the primary thread at reinit, then R/W by the worker */ |
2711 | KMP_ALIGN_CACHE__attribute__((aligned(64))) int |
2712 | th_set_nproc; /* if > 0, then only use this request for the next fork */ |
2713 | #if KMP_NESTED_HOT_TEAMS1 |
2714 | kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */ |
2715 | #endif |
2716 | kmp_proc_bind_t |
2717 | th_set_proc_bind; /* if != proc_bind_default, use request for next fork */ |
2718 | kmp_teams_size_t |
2719 | th_teams_size; /* number of teams/threads in teams construct */ |
2720 | #if KMP_AFFINITY_SUPPORTED1 |
2721 | int th_current_place; /* place currently bound to */ |
2722 | int th_new_place; /* place to bind to in par reg */ |
2723 | int th_first_place; /* first place in partition */ |
2724 | int th_last_place; /* last place in partition */ |
2725 | #endif |
2726 | int th_prev_level; /* previous level for affinity format */ |
2727 | int th_prev_num_threads; /* previous num_threads for affinity format */ |
2728 | #if USE_ITT_BUILD1 |
2729 | kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */ |
2730 | kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */ |
2731 | kmp_uint64 th_frame_time; /* frame timestamp */ |
2732 | #endif /* USE_ITT_BUILD */ |
2733 | kmp_local_t th_local; |
2734 | struct private_common *th_pri_head; |
2735 | |
2736 | /* Now the data only used by the worker (after initial allocation) */ |
2737 | /* TODO the first serial team should actually be stored in the info_t |
2738 | structure. this will help reduce initial allocation overhead */ |
2739 | KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_team_p |
2740 | *th_serial_team; /*serialized team held in reserve*/ |
2741 | |
2742 | #if OMPT_SUPPORT1 |
2743 | ompt_thread_info_t ompt_thread_info; |
2744 | #endif |
2745 | |
2746 | /* The following are also read by the primary thread during reinit */ |
2747 | struct common_table *th_pri_common; |
2748 | |
2749 | volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */ |
2750 | /* while awaiting queuing lock acquire */ |
2751 | |
2752 | volatile void *th_sleep_loc; // this points at a kmp_flag<T> |
2753 | flag_type th_sleep_loc_type; // enum type of flag stored in th_sleep_loc |
2754 | |
2755 | ident_t *th_ident; |
2756 | unsigned th_x; // Random number generator data |
2757 | unsigned th_a; // Random number generator data |
2758 | |
2759 | /* Tasking-related data for the thread */ |
2760 | kmp_task_team_t *th_task_team; // Task team struct |
2761 | kmp_taskdata_t *th_current_task; // Innermost Task being executed |
2762 | kmp_uint8 th_task_state; // alternating 0/1 for task team identification |
2763 | kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state |
2764 | // at nested levels |
2765 | kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack |
2766 | kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack |
2767 | kmp_uint32 th_reap_state; // Non-zero indicates thread is not |
2768 | // tasking, thus safe to reap |
2769 | |
2770 | /* More stuff for keeping track of active/sleeping threads (this part is |
2771 | written by the worker thread) */ |
2772 | kmp_uint8 th_active_in_pool; // included in count of #active threads in pool |
2773 | int th_active; // ! sleeping; 32 bits for TCR/TCW |
2774 | std::atomic<kmp_uint32> th_used_in_team; // Flag indicating use in team |
2775 | // 0 = not used in team; 1 = used in team; |
2776 | // 2 = transitioning to not used in team; 3 = transitioning to used in team |
2777 | struct cons_header *th_cons; // used for consistency check |
2778 | #if KMP_USE_HIER_SCHED0 |
2779 | // used for hierarchical scheduling |
2780 | kmp_hier_private_bdata_t *th_hier_bar_data; |
2781 | #endif |
2782 | |
2783 | /* Add the syncronizing data which is cache aligned and padded. */ |
2784 | KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_balign_t th_bar[bs_last_barrier]; |
2785 | |
2786 | KMP_ALIGN_CACHE__attribute__((aligned(64))) volatile kmp_int32 |
2787 | th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */ |
2788 | |
2789 | #if (USE_FAST_MEMORY3 == 3) || (USE_FAST_MEMORY3 == 5) |
2790 | #define NUM_LISTS4 4 |
2791 | kmp_free_list_t th_free_lists[NUM_LISTS4]; // Free lists for fast memory |
2792 | // allocation routines |
2793 | #endif |
2794 | |
2795 | #if KMP_OS_WINDOWS0 |
2796 | kmp_win32_cond_t th_suspend_cv; |
2797 | kmp_win32_mutex_t th_suspend_mx; |
2798 | std::atomic<int> th_suspend_init; |
2799 | #endif |
2800 | #if KMP_OS_UNIX1 |
2801 | kmp_cond_align_t th_suspend_cv; |
2802 | kmp_mutex_align_t th_suspend_mx; |
2803 | std::atomic<int> th_suspend_init_count; |
2804 | #endif |
2805 | |
2806 | #if USE_ITT_BUILD1 |
2807 | kmp_itt_mark_t th_itt_mark_single; |
2808 | // alignment ??? |
2809 | #endif /* USE_ITT_BUILD */ |
2810 | #if KMP_STATS_ENABLED0 |
2811 | kmp_stats_list *th_stats; |
2812 | #endif |
2813 | #if KMP_OS_UNIX1 |
2814 | std::atomic<bool> th_blocking; |
2815 | #endif |
2816 | kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread |
2817 | } kmp_base_info_t; |
2818 | |
2819 | typedef union KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_info { |
2820 | double th_align; /* use worst case alignment */ |
2821 | char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)(sizeof(kmp_base_info_t) + (64 - ((sizeof(kmp_base_info_t) - 1 ) % (64)) - 1))]; |
2822 | kmp_base_info_t th; |
2823 | } kmp_info_t; |
2824 | |
2825 | // OpenMP thread team data structures |
2826 | |
2827 | typedef struct kmp_base_data { |
2828 | volatile kmp_uint32 t_value; |
2829 | } kmp_base_data_t; |
2830 | |
2831 | typedef union KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_sleep_team { |
2832 | double dt_align; /* use worst case alignment */ |
2833 | char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)(sizeof(kmp_base_data_t) + (64 - ((sizeof(kmp_base_data_t) - 1 ) % (64)) - 1))]; |
2834 | kmp_base_data_t dt; |
2835 | } kmp_sleep_team_t; |
2836 | |
2837 | typedef union KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_ordered_team { |
2838 | double dt_align; /* use worst case alignment */ |
2839 | char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)(sizeof(kmp_base_data_t) + (64 - ((sizeof(kmp_base_data_t) - 1 ) % (64)) - 1))]; |
2840 | kmp_base_data_t dt; |
2841 | } kmp_ordered_team_t; |
2842 | |
2843 | typedef int (*launch_t)(int gtid); |
2844 | |
2845 | /* Minimum number of ARGV entries to malloc if necessary */ |
2846 | #define KMP_MIN_MALLOC_ARGV_ENTRIES100 100 |
2847 | |
2848 | // Set up how many argv pointers will fit in cache lines containing |
2849 | // t_inline_argv. Historically, we have supported at least 96 bytes. Using a |
2850 | // larger value for more space between the primary write/worker read section and |
2851 | // read/write by all section seems to buy more performance on EPCC PARALLEL. |
2852 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 |
2853 | #define KMP_INLINE_ARGV_BYTES(4 * 64 - ((3 * (sizeof(void *)) + 2 * sizeof(int) + 2 * sizeof (kmp_int8) + sizeof(kmp_int16) + sizeof(kmp_uint32)) % 64)) \ |
2854 | (4 * CACHE_LINE64 - \ |
2855 | ((3 * KMP_PTR_SKIP(sizeof(void *)) + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \ |
2856 | sizeof(kmp_int16) + sizeof(kmp_uint32)) % \ |
2857 | CACHE_LINE64)) |
2858 | #else |
2859 | #define KMP_INLINE_ARGV_BYTES(4 * 64 - ((3 * (sizeof(void *)) + 2 * sizeof(int) + 2 * sizeof (kmp_int8) + sizeof(kmp_int16) + sizeof(kmp_uint32)) % 64)) \ |
2860 | (2 * CACHE_LINE64 - ((3 * KMP_PTR_SKIP(sizeof(void *)) + 2 * sizeof(int)) % CACHE_LINE64)) |
2861 | #endif |
2862 | #define 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 *))) (int)(KMP_INLINE_ARGV_BYTES(4 * 64 - ((3 * (sizeof(void *)) + 2 * sizeof(int) + 2 * sizeof (kmp_int8) + sizeof(kmp_int16) + sizeof(kmp_uint32)) % 64)) / KMP_PTR_SKIP(sizeof(void *))) |
2863 | |
2864 | typedef struct KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_base_team { |
2865 | // Synchronization Data |
2866 | // --------------------------------------------------------------------------- |
2867 | KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_ordered_team_t t_ordered; |
2868 | kmp_balign_team_t t_bar[bs_last_barrier]; |
2869 | std::atomic<int> t_construct; // count of single directive encountered by team |
2870 | char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron |
2871 | |
2872 | // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups |
2873 | std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier |
2874 | std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions |
2875 | |
2876 | // Primary thread only |
2877 | // --------------------------------------------------------------------------- |
2878 | KMP_ALIGN_CACHE__attribute__((aligned(64))) int t_master_tid; // tid of primary thread in parent team |
2879 | int t_master_this_cons; // "this_construct" single counter of primary thread |
2880 | // in parent team |
2881 | ident_t *t_ident; // if volatile, have to change too much other crud to |
2882 | // volatile too |
2883 | kmp_team_p *t_parent; // parent team |
2884 | kmp_team_p *t_next_pool; // next free team in the team pool |
2885 | kmp_disp_t *t_dispatch; // thread's dispatch data |
2886 | kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2 |
2887 | kmp_proc_bind_t t_proc_bind; // bind type for par region |
2888 | #if USE_ITT_BUILD1 |
2889 | kmp_uint64 t_region_time; // region begin timestamp |
2890 | #endif /* USE_ITT_BUILD */ |
2891 | |
2892 | // Primary thread write, workers read |
2893 | // -------------------------------------------------------------------------- |
2894 | KMP_ALIGN_CACHE__attribute__((aligned(64))) void **t_argv; |
2895 | int t_argc; |
2896 | int t_nproc; // number of threads in team |
2897 | microtask_t t_pkfn; |
2898 | launch_t t_invoke; // procedure to launch the microtask |
2899 | |
2900 | #if OMPT_SUPPORT1 |
2901 | ompt_team_info_t ompt_team_info; |
2902 | ompt_lw_taskteam_t *ompt_serialized_team_info; |
2903 | #endif |
2904 | |
2905 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 |
2906 | kmp_int8 t_fp_control_saved; |
2907 | kmp_int8 t_pad2b; |
2908 | kmp_int16 t_x87_fpu_control_word; // FP control regs |
2909 | kmp_uint32 t_mxcsr; |
2910 | #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ |
2911 | |
2912 | void *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 *)))]; |
2913 | |
2914 | KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_info_t **t_threads; |
2915 | kmp_taskdata_t |
2916 | *t_implicit_task_taskdata; // Taskdata for the thread's implicit task |
2917 | int t_level; // nested parallel level |
2918 | |
2919 | KMP_ALIGN_CACHE__attribute__((aligned(64))) int t_max_argc; |
2920 | int t_max_nproc; // max threads this team can handle (dynamically expandable) |
2921 | int t_serialized; // levels deep of serialized teams |
2922 | dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system |
2923 | int t_id; // team's id, assigned by debugger. |
2924 | int t_active_level; // nested active parallel level |
2925 | kmp_r_sched_t t_sched; // run-time schedule for the team |
2926 | #if KMP_AFFINITY_SUPPORTED1 |
2927 | int t_first_place; // first & last place in parent thread's partition. |
2928 | int t_last_place; // Restore these values to primary thread after par region. |
2929 | #endif // KMP_AFFINITY_SUPPORTED |
2930 | int t_display_affinity; |
2931 | int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via |
2932 | // omp_set_num_threads() call |
2933 | omp_allocator_handle_t t_def_allocator; /* default allocator */ |
2934 | |
2935 | // Read/write by workers as well |
2936 | #if (KMP_ARCH_X860 || KMP_ARCH_X86_641) |
2937 | // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf |
2938 | // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra |
2939 | // padding serves to fix the performance of epcc 'parallel' and 'barrier' when |
2940 | // CACHE_LINE=64. TODO: investigate more and get rid if this padding. |
2941 | char dummy_padding[1024]; |
2942 | #endif |
2943 | // Internal control stack for additional nested teams. |
2944 | KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_internal_control_t *t_control_stack_top; |
2945 | // for SERIALIZED teams nested 2 or more levels deep |
2946 | // typed flag to store request state of cancellation |
2947 | std::atomic<kmp_int32> t_cancel_request; |
2948 | int t_master_active; // save on fork, restore on join |
2949 | void *t_copypriv_data; // team specific pointer to copyprivate data array |
2950 | #if KMP_OS_WINDOWS0 |
2951 | std::atomic<kmp_uint32> t_copyin_counter; |
2952 | #endif |
2953 | #if USE_ITT_BUILD1 |
2954 | void *t_stack_id; // team specific stack stitching id (for ittnotify) |
2955 | #endif /* USE_ITT_BUILD */ |
2956 | distributedBarrier *b; // Distributed barrier data associated with team |
2957 | } kmp_base_team_t; |
2958 | |
2959 | union KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_team { |
2960 | kmp_base_team_t t; |
2961 | double t_align; /* use worst case alignment */ |
2962 | char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)(sizeof(kmp_base_team_t) + (64 - ((sizeof(kmp_base_team_t) - 1 ) % (64)) - 1))]; |
2963 | }; |
2964 | |
2965 | typedef union KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_time_global { |
2966 | double dt_align; /* use worst case alignment */ |
2967 | char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)(sizeof(kmp_base_data_t) + (64 - ((sizeof(kmp_base_data_t) - 1 ) % (64)) - 1))]; |
2968 | kmp_base_data_t dt; |
2969 | } kmp_time_global_t; |
2970 | |
2971 | typedef struct kmp_base_global { |
2972 | /* cache-aligned */ |
2973 | kmp_time_global_t g_time; |
2974 | |
2975 | /* non cache-aligned */ |
2976 | volatile int g_abort; |
2977 | volatile int g_done; |
2978 | |
2979 | int g_dynamic; |
2980 | enum dynamic_mode g_dynamic_mode; |
2981 | } kmp_base_global_t; |
2982 | |
2983 | typedef union KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_global { |
2984 | kmp_base_global_t g; |
2985 | double g_align; /* use worst case alignment */ |
2986 | char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)(sizeof(kmp_base_global_t) + (64 - ((sizeof(kmp_base_global_t ) - 1) % (64)) - 1))]; |
2987 | } kmp_global_t; |
2988 | |
2989 | typedef struct kmp_base_root { |
2990 | // TODO: GEH - combine r_active with r_in_parallel then r_active == |
2991 | // (r_in_parallel>= 0) |
2992 | // TODO: GEH - then replace r_active with t_active_levels if we can to reduce |
2993 | // the synch overhead or keeping r_active |
2994 | volatile int r_active; /* TRUE if some region in a nest has > 1 thread */ |
2995 | // keeps a count of active parallel regions per root |
2996 | std::atomic<int> r_in_parallel; |
2997 | // GEH: This is misnamed, should be r_active_levels |
2998 | kmp_team_t *r_root_team; |
2999 | kmp_team_t *r_hot_team; |
3000 | kmp_info_t *r_uber_thread; |
3001 | kmp_lock_t r_begin_lock; |
3002 | volatile int r_begin; |
3003 | int r_blocktime; /* blocktime for this root and descendants */ |
3004 | #if KMP_AFFINITY_SUPPORTED1 |
3005 | int r_affinity_assigned; |
3006 | #endif // KMP_AFFINITY_SUPPORTED |
3007 | } kmp_base_root_t; |
3008 | |
3009 | typedef union KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_root { |
3010 | kmp_base_root_t r; |
3011 | double r_align; /* use worst case alignment */ |
3012 | char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)(sizeof(kmp_base_root_t) + (64 - ((sizeof(kmp_base_root_t) - 1 ) % (64)) - 1))]; |
3013 | } kmp_root_t; |
3014 | |
3015 | struct fortran_inx_info { |
3016 | kmp_int32 data; |
3017 | }; |
3018 | |
3019 | // This list type exists to hold old __kmp_threads arrays so that |
3020 | // old references to them may complete while reallocation takes place when |
3021 | // expanding the array. The items in this list are kept alive until library |
3022 | // shutdown. |
3023 | typedef struct kmp_old_threads_list_t { |
3024 | kmp_info_t **threads; |
3025 | struct kmp_old_threads_list_t *next; |
3026 | } kmp_old_threads_list_t; |
3027 | |
3028 | /* ------------------------------------------------------------------------ */ |
3029 | |
3030 | extern int __kmp_settings; |
3031 | extern int __kmp_duplicate_library_ok; |
3032 | #if USE_ITT_BUILD1 |
3033 | extern int __kmp_forkjoin_frames; |
3034 | extern int __kmp_forkjoin_frames_mode; |
3035 | #endif |
3036 | extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method; |
3037 | extern int __kmp_determ_red; |
3038 | |
3039 | #ifdef KMP_DEBUG1 |
3040 | extern int kmp_a_debug; |
3041 | extern int kmp_b_debug; |
3042 | extern int kmp_c_debug; |
3043 | extern int kmp_d_debug; |
3044 | extern int kmp_e_debug; |
3045 | extern int kmp_f_debug; |
3046 | #endif /* KMP_DEBUG */ |
3047 | |
3048 | /* For debug information logging using rotating buffer */ |
3049 | #define KMP_DEBUG_BUF_LINES_INIT512 512 |
3050 | #define KMP_DEBUG_BUF_LINES_MIN1 1 |
3051 | |
3052 | #define KMP_DEBUG_BUF_CHARS_INIT128 128 |
3053 | #define KMP_DEBUG_BUF_CHARS_MIN2 2 |
3054 | |
3055 | extern int |
3056 | __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */ |
3057 | extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */ |
3058 | extern int |
3059 | __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */ |
3060 | extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer |
3061 | entry pointer */ |
3062 | |
3063 | extern char *__kmp_debug_buffer; /* Debug buffer itself */ |
3064 | extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines |
3065 | printed in buffer so far */ |
3066 | extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase |
3067 | recommended in warnings */ |
3068 | /* end rotating debug buffer */ |
3069 | |
3070 | #ifdef KMP_DEBUG1 |
3071 | extern int __kmp_par_range; /* +1 => only go par for constructs in range */ |
3072 | |
3073 | #define KMP_PAR_RANGE_ROUTINE_LEN1024 1024 |
3074 | extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN1024]; |
3075 | #define KMP_PAR_RANGE_FILENAME_LEN1024 1024 |
3076 | extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN1024]; |
3077 | extern int __kmp_par_range_lb; |
3078 | extern int __kmp_par_range_ub; |
3079 | #endif |
3080 | |
3081 | /* For printing out dynamic storage map for threads and teams */ |
3082 | extern int |
3083 | __kmp_storage_map; /* True means print storage map for threads and teams */ |
3084 | extern int __kmp_storage_map_verbose; /* True means storage map includes |
3085 | placement info */ |
3086 | extern int __kmp_storage_map_verbose_specified; |
3087 | |
3088 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 |
3089 | extern kmp_cpuinfo_t __kmp_cpuinfo; |
3090 | static inline bool __kmp_is_hybrid_cpu() { return __kmp_cpuinfo.flags.hybrid; } |
3091 | #elif KMP_OS_DARWIN0 && KMP_ARCH_AARCH640 |
3092 | static inline bool __kmp_is_hybrid_cpu() { return true; } |
3093 | #else |
3094 | static inline bool __kmp_is_hybrid_cpu() { return false; } |
3095 | #endif |
3096 | |
3097 | extern volatile int __kmp_init_serial; |
3098 | extern volatile int __kmp_init_gtid; |
3099 | extern volatile int __kmp_init_common; |
3100 | extern volatile int __kmp_need_register_serial; |
3101 | extern volatile int __kmp_init_middle; |
3102 | extern volatile int __kmp_init_parallel; |
3103 | #if KMP_USE_MONITOR |
3104 | extern volatile int __kmp_init_monitor; |
3105 | #endif |
3106 | extern volatile int __kmp_init_user_locks; |
3107 | extern volatile int __kmp_init_hidden_helper_threads; |
3108 | extern int __kmp_init_counter; |
3109 | extern int __kmp_root_counter; |
3110 | extern int __kmp_version; |
3111 | |
3112 | /* list of address of allocated caches for commons */ |
3113 | extern kmp_cached_addr_t *__kmp_threadpriv_cache_list; |
3114 | |
3115 | /* Barrier algorithm types and options */ |
3116 | extern kmp_uint32 __kmp_barrier_gather_bb_dflt; |
3117 | extern kmp_uint32 __kmp_barrier_release_bb_dflt; |
3118 | extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt; |
3119 | extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt; |
3120 | extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier]; |
3121 | extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier]; |
3122 | extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier]; |
3123 | extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier]; |
3124 | extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier]; |
3125 | extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier]; |
3126 | extern char const *__kmp_barrier_type_name[bs_last_barrier]; |
3127 | extern char const *__kmp_barrier_pattern_name[bp_last_bar]; |
3128 | |
3129 | /* Global Locks */ |
3130 | extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */ |
3131 | extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */ |
3132 | extern kmp_bootstrap_lock_t __kmp_task_team_lock; |
3133 | extern kmp_bootstrap_lock_t |
3134 | __kmp_exit_lock; /* exit() is not always thread-safe */ |
3135 | #if KMP_USE_MONITOR |
3136 | extern kmp_bootstrap_lock_t |
3137 | __kmp_monitor_lock; /* control monitor thread creation */ |
3138 | #endif |
3139 | extern kmp_bootstrap_lock_t |
3140 | __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and |
3141 | __kmp_threads expansion to co-exist */ |
3142 | |
3143 | extern kmp_lock_t __kmp_global_lock; /* control OS/global access */ |
3144 | extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access */ |
3145 | extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */ |
3146 | |
3147 | extern enum library_type __kmp_library; |
3148 | |
3149 | extern enum sched_type __kmp_sched; /* default runtime scheduling */ |
3150 | extern enum sched_type __kmp_static; /* default static scheduling method */ |
3151 | extern enum sched_type __kmp_guided; /* default guided scheduling method */ |
3152 | extern enum sched_type __kmp_auto; /* default auto scheduling method */ |
3153 | extern int __kmp_chunk; /* default runtime chunk size */ |
3154 | extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */ |
3155 | |
3156 | extern size_t __kmp_stksize; /* stack size per thread */ |
3157 | #if KMP_USE_MONITOR |
3158 | extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */ |
3159 | #endif |
3160 | extern size_t __kmp_stkoffset; /* stack offset per thread */ |
3161 | extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */ |
3162 | |
3163 | extern size_t |
3164 | __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */ |
3165 | extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */ |
3166 | extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */ |
3167 | extern int __kmp_env_checks; /* was KMP_CHECKS specified? */ |
3168 | extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified? |
3169 | extern int __kmp_generate_warnings; /* should we issue warnings? */ |
3170 | extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */ |
3171 | |
3172 | #ifdef DEBUG_SUSPEND |
3173 | extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */ |
3174 | #endif |
3175 | |
3176 | extern kmp_int32 __kmp_use_yield; |
3177 | extern kmp_int32 __kmp_use_yield_exp_set; |
3178 | extern kmp_uint32 __kmp_yield_init; |
3179 | extern kmp_uint32 __kmp_yield_next; |
3180 | extern kmp_uint64 __kmp_pause_init; |
3181 | |
3182 | /* ------------------------------------------------------------------------- */ |
3183 | extern int __kmp_allThreadsSpecified; |
3184 | |
3185 | extern size_t __kmp_align_alloc; |
3186 | /* following data protected by initialization routines */ |
3187 | extern int __kmp_xproc; /* number of processors in the system */ |
3188 | extern int __kmp_avail_proc; /* number of processors available to the process */ |
3189 | extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */ |
3190 | extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */ |
3191 | // maximum total number of concurrently-existing threads on device |
3192 | extern int __kmp_max_nth; |
3193 | // maximum total number of concurrently-existing threads in a contention group |
3194 | extern int __kmp_cg_max_nth; |
3195 | extern int __kmp_teams_max_nth; // max threads used in a teams construct |
3196 | extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and |
3197 | __kmp_root */ |
3198 | extern int __kmp_dflt_team_nth; /* default number of threads in a parallel |
3199 | region a la OMP_NUM_THREADS */ |
3200 | extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial |
3201 | initialization */ |
3202 | extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is |
3203 | used (fixed) */ |
3204 | extern int __kmp_tp_cached; /* whether threadprivate cache has been created |
3205 | (__kmpc_threadprivate_cached()) */ |
3206 | extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before |
3207 | blocking (env setting) */ |
3208 | extern bool __kmp_wpolicy_passive; /* explicitly set passive wait policy */ |
3209 | #if KMP_USE_MONITOR |
3210 | extern int |
3211 | __kmp_monitor_wakeups; /* number of times monitor wakes up per second */ |
3212 | extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before |
3213 | blocking */ |
3214 | #endif |
3215 | #ifdef KMP_ADJUST_BLOCKTIME1 |
3216 | extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */ |
3217 | #endif /* KMP_ADJUST_BLOCKTIME */ |
3218 | #ifdef KMP_DFLT_NTH_CORES |
3219 | extern int __kmp_ncores; /* Total number of cores for threads placement */ |
3220 | #endif |
3221 | /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */ |
3222 | extern int __kmp_abort_delay; |
3223 | |
3224 | extern int __kmp_need_register_atfork_specified; |
3225 | extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork |
3226 | to install fork handler */ |
3227 | extern int __kmp_gtid_mode; /* Method of getting gtid, values: |
3228 | 0 - not set, will be set at runtime |
3229 | 1 - using stack search |
3230 | 2 - dynamic TLS (pthread_getspecific(Linux* OS/OS |
3231 | X*) or TlsGetValue(Windows* OS)) |
3232 | 3 - static TLS (__declspec(thread) __kmp_gtid), |
3233 | Linux* OS .so only. */ |
3234 | extern int |
3235 | __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */ |
3236 | #ifdef KMP_TDATA_GTID1 |
3237 | extern KMP_THREAD_LOCAL__thread int __kmp_gtid; |
3238 | #endif |
3239 | extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */ |
3240 | extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread |
3241 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 |
3242 | extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork |
3243 | extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg |
3244 | extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */ |
3245 | #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ |
3246 | |
3247 | // max_active_levels for nested parallelism enabled by default via |
3248 | // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND |
3249 | extern int __kmp_dflt_max_active_levels; |
3250 | // Indicates whether value of __kmp_dflt_max_active_levels was already |
3251 | // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false |
3252 | extern bool __kmp_dflt_max_active_levels_set; |
3253 | extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in |
3254 | concurrent execution per team */ |
3255 | #if KMP_NESTED_HOT_TEAMS1 |
3256 | extern int __kmp_hot_teams_mode; |
3257 | extern int __kmp_hot_teams_max_level; |
3258 | #endif |
3259 | |
3260 | #if KMP_OS_LINUX1 |
3261 | extern enum clock_function_type __kmp_clock_function; |
3262 | extern int __kmp_clock_function_param; |
3263 | #endif /* KMP_OS_LINUX */ |
3264 | |
3265 | #if KMP_MIC_SUPPORTED((0 || 1) && (1 || 0)) |
3266 | extern enum mic_type __kmp_mic_type; |
3267 | #endif |
3268 | |
3269 | #ifdef USE_LOAD_BALANCE1 |
3270 | extern double __kmp_load_balance_interval; // load balance algorithm interval |
3271 | #endif /* USE_LOAD_BALANCE */ |
3272 | |
3273 | // OpenMP 3.1 - Nested num threads array |
3274 | typedef struct kmp_nested_nthreads_t { |
3275 | int *nth; |
3276 | int size; |
3277 | int used; |
3278 | } kmp_nested_nthreads_t; |
3279 | |
3280 | extern kmp_nested_nthreads_t __kmp_nested_nth; |
3281 | |
3282 | #if KMP_USE_ADAPTIVE_LOCKS(0 || 1) && !0 |
3283 | |
3284 | // Parameters for the speculative lock backoff system. |
3285 | struct kmp_adaptive_backoff_params_t { |
3286 | // Number of soft retries before it counts as a hard retry. |
3287 | kmp_uint32 max_soft_retries; |
3288 | // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to |
3289 | // the right |
3290 | kmp_uint32 max_badness; |
3291 | }; |
3292 | |
3293 | extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params; |
3294 | |
3295 | #if KMP_DEBUG_ADAPTIVE_LOCKS0 |
3296 | extern const char *__kmp_speculative_statsfile; |
3297 | #endif |
3298 | |
3299 | #endif // KMP_USE_ADAPTIVE_LOCKS |
3300 | |
3301 | extern int __kmp_display_env; /* TRUE or FALSE */ |
3302 | extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */ |
3303 | extern int __kmp_omp_cancellation; /* TRUE or FALSE */ |
3304 | extern int __kmp_nteams; |
3305 | extern int __kmp_teams_thread_limit; |
3306 | |
3307 | /* ------------------------------------------------------------------------- */ |
3308 | |
3309 | /* the following are protected by the fork/join lock */ |
3310 | /* write: lock read: anytime */ |
3311 | extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */ |
3312 | /* Holds old arrays of __kmp_threads until library shutdown */ |
3313 | extern kmp_old_threads_list_t *__kmp_old_threads_list; |
3314 | /* read/write: lock */ |
3315 | extern volatile kmp_team_t *__kmp_team_pool; |
3316 | extern volatile kmp_info_t *__kmp_thread_pool; |
3317 | extern kmp_info_t *__kmp_thread_pool_insert_pt; |
3318 | |
3319 | // total num threads reachable from some root thread including all root threads |
3320 | extern volatile int __kmp_nth; |
3321 | /* total number of threads reachable from some root thread including all root |
3322 | threads, and those in the thread pool */ |
3323 | extern volatile int __kmp_all_nth; |
3324 | extern std::atomic<int> __kmp_thread_pool_active_nth; |
3325 | |
3326 | extern kmp_root_t **__kmp_root; /* root of thread hierarchy */ |
3327 | /* end data protected by fork/join lock */ |
3328 | /* ------------------------------------------------------------------------- */ |
3329 | |
3330 | #define __kmp_get_gtid()__kmp_get_global_thread_id() __kmp_get_global_thread_id() |
3331 | #define __kmp_entry_gtid()__kmp_get_global_thread_id_reg() __kmp_get_global_thread_id_reg() |
3332 | #define __kmp_get_tid()(__kmp_tid_from_gtid(__kmp_get_global_thread_id())) (__kmp_tid_from_gtid(__kmp_get_gtid()__kmp_get_global_thread_id())) |
3333 | #define __kmp_get_team()(__kmp_threads[(__kmp_get_global_thread_id())]->th.th_team ) (__kmp_threads[(__kmp_get_gtid()__kmp_get_global_thread_id())]->th.th_team) |
3334 | #define __kmp_get_thread()(__kmp_thread_from_gtid(__kmp_get_global_thread_id())) (__kmp_thread_from_gtid(__kmp_get_gtid()__kmp_get_global_thread_id())) |
3335 | |
3336 | // AT: Which way is correct? |
3337 | // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc; |
3338 | // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc; |
3339 | #define __kmp_get_team_num_threads(gtid)(__kmp_threads[(gtid)]->th.th_team->t.t_nproc) \ |
3340 | (__kmp_threads[(gtid)]->th.th_team->t.t_nproc) |
3341 | |
3342 | static inline bool KMP_UBER_GTID(int gtid) { |
3343 | KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN)if (!(gtid >= (-6))) { __kmp_debug_assert("gtid >= (-6)" , "openmp/runtime/src/kmp.h", 3343); }; |
3344 | KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity)if (!(gtid < __kmp_threads_capacity)) { __kmp_debug_assert ("gtid < __kmp_threads_capacity", "openmp/runtime/src/kmp.h" , 3344); }; |
3345 | return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] && |
3346 | __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread); |
3347 | } |
3348 | |
3349 | static inline int __kmp_tid_from_gtid(int gtid) { |
3350 | KMP_DEBUG_ASSERT(gtid >= 0)if (!(gtid >= 0)) { __kmp_debug_assert("gtid >= 0", "openmp/runtime/src/kmp.h" , 3350); }; |
3351 | return __kmp_threads[gtid]->th.th_info.ds.ds_tid; |
3352 | } |
3353 | |
3354 | static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) { |
3355 | KMP_DEBUG_ASSERT(tid >= 0 && team)if (!(tid >= 0 && team)) { __kmp_debug_assert("tid >= 0 && team" , "openmp/runtime/src/kmp.h", 3355); }; |
3356 | return team->t.t_threads[tid]->th.th_info.ds.ds_gtid; |
3357 | } |
3358 | |
3359 | static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) { |
3360 | KMP_DEBUG_ASSERT(thr)if (!(thr)) { __kmp_debug_assert("thr", "openmp/runtime/src/kmp.h" , 3360); }; |
3361 | return thr->th.th_info.ds.ds_gtid; |
3362 | } |
3363 | |
3364 | static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) { |
3365 | KMP_DEBUG_ASSERT(gtid >= 0)if (!(gtid >= 0)) { __kmp_debug_assert("gtid >= 0", "openmp/runtime/src/kmp.h" , 3365); }; |
3366 | return __kmp_threads[gtid]; |
3367 | } |
3368 | |
3369 | static inline kmp_team_t *__kmp_team_from_gtid(int gtid) { |
3370 | KMP_DEBUG_ASSERT(gtid >= 0)if (!(gtid >= 0)) { __kmp_debug_assert("gtid >= 0", "openmp/runtime/src/kmp.h" , 3370); }; |
3371 | return __kmp_threads[gtid]->th.th_team; |
3372 | } |
3373 | |
3374 | static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) { |
3375 | if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity)__builtin_expect(!!(gtid < 0 || gtid >= __kmp_threads_capacity ), 0)) |
3376 | KMP_FATAL(ThreadIdentInvalid)__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_ThreadIdentInvalid) , __kmp_msg_null); |
3377 | } |
3378 | |
3379 | #if KMP_HAVE_MWAIT((0 || 1) && (1 || 0) && !0) || KMP_HAVE_UMWAIT((0 || 1) && (1 || 0) && !0) |
3380 | extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT |
3381 | extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled |
3382 | extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled |
3383 | extern int __kmp_mwait_hints; // Hints to pass in to mwait |
3384 | #endif |
3385 | |
3386 | #if KMP_HAVE_UMWAIT((0 || 1) && (1 || 0) && !0) |
3387 | extern int __kmp_waitpkg_enabled; // Runtime check if waitpkg exists |
3388 | extern int __kmp_tpause_state; // 0 (default), 1=C0.1, 2=C0.2; from KMP_TPAUSE |
3389 | extern int __kmp_tpause_hint; // 1=C0.1 (default), 0=C0.2; from KMP_TPAUSE |
3390 | extern int __kmp_tpause_enabled; // 0 (default), 1 (KMP_TPAUSE is non-zero) |
3391 | #endif |
3392 | |
3393 | /* ------------------------------------------------------------------------- */ |
3394 | |
3395 | extern kmp_global_t __kmp_global; /* global status */ |
3396 | |
3397 | extern kmp_info_t __kmp_monitor; |
3398 | // For Debugging Support Library |
3399 | extern std::atomic<kmp_int32> __kmp_team_counter; |
3400 | // For Debugging Support Library |
3401 | extern std::atomic<kmp_int32> __kmp_task_counter; |
3402 | |
3403 | #if USE_DEBUGGER0 |
3404 | #define _KMP_GEN_ID(counter)(~0) \ |
3405 | (__kmp_debugging ? KMP_ATOMIC_INC(&counter)(&counter)->fetch_add(1, std::memory_order_acq_rel) + 1 : ~0) |
3406 | #else |
3407 | #define _KMP_GEN_ID(counter)(~0) (~0) |
3408 | #endif /* USE_DEBUGGER */ |
3409 | |
3410 | #define KMP_GEN_TASK_ID()(~0) _KMP_GEN_ID(__kmp_task_counter)(~0) |
3411 | #define KMP_GEN_TEAM_ID()(~0) _KMP_GEN_ID(__kmp_team_counter)(~0) |
3412 | |
3413 | /* ------------------------------------------------------------------------ */ |
3414 | |
3415 | extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2, |
3416 | size_t size, char const *format, ...); |
3417 | |
3418 | extern void __kmp_serial_initialize(void); |
3419 | extern void __kmp_middle_initialize(void); |
3420 | extern void __kmp_parallel_initialize(void); |
3421 | |
3422 | extern void __kmp_internal_begin(void); |
3423 | extern void __kmp_internal_end_library(int gtid); |
3424 | extern void __kmp_internal_end_thread(int gtid); |
3425 | extern void __kmp_internal_end_atexit(void); |
3426 | extern void __kmp_internal_end_dtor(void); |
3427 | extern void __kmp_internal_end_dest(void *); |
3428 | |
3429 | extern int __kmp_register_root(int initial_thread); |
3430 | extern void __kmp_unregister_root(int gtid); |
3431 | extern void __kmp_unregister_library(void); // called by __kmp_internal_end() |
3432 | |
3433 | extern int __kmp_ignore_mppbeg(void); |
3434 | extern int __kmp_ignore_mppend(void); |
3435 | |
3436 | extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws); |
3437 | extern void __kmp_exit_single(int gtid); |
3438 | |
3439 | extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref); |
3440 | extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref); |
3441 | |
3442 | #ifdef USE_LOAD_BALANCE1 |
3443 | extern int __kmp_get_load_balance(int); |
3444 | #endif |
3445 | |
3446 | extern int __kmp_get_global_thread_id(void); |
3447 | extern int __kmp_get_global_thread_id_reg(void); |
3448 | extern void __kmp_exit_thread(int exit_status); |
3449 | extern void __kmp_abort(char const *format, ...); |
3450 | extern void __kmp_abort_thread(void); |
3451 | KMP_NORETURN[[noreturn]] extern void __kmp_abort_process(void); |
3452 | extern void __kmp_warn(char const *format, ...); |
3453 | |
3454 | extern void __kmp_set_num_threads(int new_nth, int gtid); |
3455 | |
3456 | // Returns current thread (pointer to kmp_info_t). Current thread *must* be |
3457 | // registered. |
3458 | static inline kmp_info_t *__kmp_entry_thread() { |
3459 | int gtid = __kmp_entry_gtid()__kmp_get_global_thread_id_reg(); |
3460 | |
3461 | return __kmp_threads[gtid]; |
3462 | } |
3463 | |
3464 | extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels); |
3465 | extern int __kmp_get_max_active_levels(int gtid); |
3466 | extern int __kmp_get_ancestor_thread_num(int gtid, int level); |
3467 | extern int __kmp_get_team_size(int gtid, int level); |
3468 | extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk); |
3469 | extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk); |
3470 | |
3471 | extern unsigned short __kmp_get_random(kmp_info_t *thread); |
3472 | extern void __kmp_init_random(kmp_info_t *thread); |
3473 | |
3474 | extern kmp_r_sched_t __kmp_get_schedule_global(void); |
3475 | extern void __kmp_adjust_num_threads(int new_nproc); |
3476 | extern void __kmp_check_stksize(size_t *val); |
3477 | |
3478 | extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL, char const *_file_, int _line_); |
3479 | extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL, char const *_file_, int _line_); |
3480 | extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL, char const *_file_, int _line_); |
3481 | #define __kmp_allocate(size)___kmp_allocate((size), "openmp/runtime/src/kmp.h", 3481) ___kmp_allocate((size)KMP_SRC_LOC_CURR, "openmp/runtime/src/kmp.h", 3481) |
3482 | #define __kmp_page_allocate(size)___kmp_page_allocate((size), "openmp/runtime/src/kmp.h", 3482 ) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR, "openmp/runtime/src/kmp.h", 3482) |
3483 | #define __kmp_free(ptr)___kmp_free((ptr), "openmp/runtime/src/kmp.h", 3483) ___kmp_free((ptr)KMP_SRC_LOC_CURR, "openmp/runtime/src/kmp.h", 3483) |
3484 | |
3485 | #if USE_FAST_MEMORY3 |
3486 | extern void *___kmp_fast_allocate(kmp_info_t *this_thr, |
3487 | size_t size KMP_SRC_LOC_DECL, char const *_file_, int _line_); |
3488 | extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL, char const *_file_, int _line_); |
3489 | extern void __kmp_free_fast_memory(kmp_info_t *this_thr); |
3490 | extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr); |
3491 | #define __kmp_fast_allocate(this_thr, size)___kmp_fast_allocate((this_thr), (size), "openmp/runtime/src/kmp.h" , 3491) \ |
3492 | ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR, "openmp/runtime/src/kmp.h", 3492) |
3493 | #define __kmp_fast_free(this_thr, ptr)___kmp_fast_free((this_thr), (ptr), "openmp/runtime/src/kmp.h" , 3493) \ |
3494 | ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR, "openmp/runtime/src/kmp.h", 3494) |
3495 | #endif |
3496 | |
3497 | extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL, char const *_file_, int _line_); |
3498 | extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem, |
3499 | size_t elsize KMP_SRC_LOC_DECL, char const *_file_, int _line_); |
3500 | extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr, |
3501 | size_t size KMP_SRC_LOC_DECL, char const *_file_, int _line_); |
3502 | extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL, char const *_file_, int _line_); |
3503 | #define __kmp_thread_malloc(th, size)___kmp_thread_malloc((th), (size), "openmp/runtime/src/kmp.h" , 3503) \ |
3504 | ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR, "openmp/runtime/src/kmp.h", 3504) |
3505 | #define __kmp_thread_calloc(th, nelem, elsize)___kmp_thread_calloc((th), (nelem), (elsize), "openmp/runtime/src/kmp.h" , 3505) \ |
3506 | ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR, "openmp/runtime/src/kmp.h", 3506) |
3507 | #define __kmp_thread_realloc(th, ptr, size)___kmp_thread_realloc((th), (ptr), (size), "openmp/runtime/src/kmp.h" , 3507) \ |
3508 | ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR, "openmp/runtime/src/kmp.h", 3508) |
3509 | #define __kmp_thread_free(th, ptr)___kmp_thread_free((th), (ptr), "openmp/runtime/src/kmp.h", 3509 ) \ |
3510 | ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR, "openmp/runtime/src/kmp.h", 3510) |
3511 | |
3512 | extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads); |
3513 | |
3514 | extern void __kmp_push_proc_bind(ident_t *loc, int gtid, |
3515 | kmp_proc_bind_t proc_bind); |
3516 | extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams, |
3517 | int num_threads); |
3518 | extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb, |
3519 | int num_teams_ub, int num_threads); |
3520 | |
3521 | extern void __kmp_yield(); |
3522 | |
3523 | extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid, |
3524 | enum sched_type schedule, kmp_int32 lb, |
3525 | kmp_int32 ub, kmp_int32 st, kmp_int32 chunk); |
3526 | extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid, |
3527 | enum sched_type schedule, kmp_uint32 lb, |
3528 | kmp_uint32 ub, kmp_int32 st, |
3529 | kmp_int32 chunk); |
3530 | extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid, |
3531 | enum sched_type schedule, kmp_int64 lb, |
3532 | kmp_int64 ub, kmp_int64 st, kmp_int64 chunk); |
3533 | extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid, |
3534 | enum sched_type schedule, kmp_uint64 lb, |
3535 | kmp_uint64 ub, kmp_int64 st, |
3536 | kmp_int64 chunk); |
3537 | |
3538 | extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid, |
3539 | kmp_int32 *p_last, kmp_int32 *p_lb, |
3540 | kmp_int32 *p_ub, kmp_int32 *p_st); |
3541 | extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid, |
3542 | kmp_int32 *p_last, kmp_uint32 *p_lb, |
3543 | kmp_uint32 *p_ub, kmp_int32 *p_st); |
3544 | extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid, |
3545 | kmp_int32 *p_last, kmp_int64 *p_lb, |
3546 | kmp_int64 *p_ub, kmp_int64 *p_st); |
3547 | extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid, |
3548 | kmp_int32 *p_last, kmp_uint64 *p_lb, |
3549 | kmp_uint64 *p_ub, kmp_int64 *p_st); |
3550 | |
3551 | extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid); |
3552 | extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid); |
3553 | extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid); |
3554 | extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid); |
3555 | |
3556 | #ifdef KMP_GOMP_COMPAT |
3557 | |
3558 | extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid, |
3559 | enum sched_type schedule, kmp_int32 lb, |
3560 | kmp_int32 ub, kmp_int32 st, |
3561 | kmp_int32 chunk, int push_ws); |
3562 | extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid, |
3563 | enum sched_type schedule, kmp_uint32 lb, |
3564 | kmp_uint32 ub, kmp_int32 st, |
3565 | kmp_int32 chunk, int push_ws); |
3566 | extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid, |
3567 | enum sched_type schedule, kmp_int64 lb, |
3568 | kmp_int64 ub, kmp_int64 st, |
3569 | kmp_int64 chunk, int push_ws); |
3570 | extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid, |
3571 | enum sched_type schedule, kmp_uint64 lb, |
3572 | kmp_uint64 ub, kmp_int64 st, |
3573 | kmp_int64 chunk, int push_ws); |
3574 | extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid); |
3575 | extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid); |
3576 | extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid); |
3577 | extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid); |
3578 | |
3579 | #endif /* KMP_GOMP_COMPAT */ |
3580 | |
3581 | extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker); |
3582 | extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker); |
3583 | extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker); |
3584 | extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker); |
3585 | extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker); |
3586 | extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker, |
3587 | kmp_uint32 (*pred)(kmp_uint32, kmp_uint32), |
3588 | void *obj); |
3589 | extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker, |
3590 | kmp_uint32 (*pred)(void *, kmp_uint32), void *obj); |
3591 | |
3592 | extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag, |
3593 | int final_spin |
3594 | #if USE_ITT_BUILD1 |
3595 | , |
3596 | void *itt_sync_obj |
3597 | #endif |
3598 | ); |
3599 | extern void __kmp_release_64(kmp_flag_64<> *flag); |
3600 | |
3601 | extern void __kmp_infinite_loop(void); |
3602 | |
3603 | extern void __kmp_cleanup(void); |
3604 | |
3605 | #if KMP_HANDLE_SIGNALS(1 || 0) |
3606 | extern int __kmp_handle_signals; |
3607 | extern void __kmp_install_signals(int parallel_init); |
3608 | extern void __kmp_remove_signals(void); |
3609 | #endif |
3610 | |
3611 | extern void __kmp_clear_system_time(void); |
3612 | extern void __kmp_read_system_time(double *delta); |
3613 | |
3614 | extern void __kmp_check_stack_overlap(kmp_info_t *thr); |
3615 | |
3616 | extern void __kmp_expand_host_name(char *buffer, size_t size); |
3617 | extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern); |
3618 | |
3619 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 || (KMP_OS_WINDOWS0 && (KMP_ARCH_AARCH640 || KMP_ARCH_ARM)) |
3620 | extern void |
3621 | __kmp_initialize_system_tick(void); /* Initialize timer tick value */ |
3622 | #endif |
3623 | |
3624 | extern void |
3625 | __kmp_runtime_initialize(void); /* machine specific initialization */ |
3626 | extern void __kmp_runtime_destroy(void); |
3627 | |
3628 | #if KMP_AFFINITY_SUPPORTED1 |
3629 | extern char *__kmp_affinity_print_mask(char *buf, int buf_len, |
3630 | kmp_affin_mask_t *mask); |
3631 | extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf, |
3632 | kmp_affin_mask_t *mask); |
3633 | extern void __kmp_affinity_initialize(kmp_affinity_t &affinity); |
3634 | extern void __kmp_affinity_uninitialize(void); |
3635 | extern void __kmp_affinity_set_init_mask( |
3636 | int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */ |
3637 | extern void __kmp_affinity_set_place(int gtid); |
3638 | extern void __kmp_affinity_determine_capable(const char *env_var); |
3639 | extern int __kmp_aux_set_affinity(void **mask); |
3640 | extern int __kmp_aux_get_affinity(void **mask); |
3641 | extern int __kmp_aux_get_affinity_max_proc(); |
3642 | extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask); |
3643 | extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask); |
3644 | extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask); |
3645 | extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size); |
3646 | #if KMP_OS_LINUX1 || KMP_OS_FREEBSD0 |
3647 | extern int kmp_set_thread_affinity_mask_initial(void); |
3648 | #endif |
3649 | static inline void __kmp_assign_root_init_mask() { |
3650 | int gtid = __kmp_entry_gtid()__kmp_get_global_thread_id_reg(); |
3651 | kmp_root_t *r = __kmp_threads[gtid]->th.th_root; |
3652 | if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) { |
3653 | __kmp_affinity_set_init_mask(gtid, TRUE(!0)); |
3654 | r->r.r_affinity_assigned = TRUE(!0); |
3655 | } |
3656 | } |
3657 | static inline void __kmp_reset_root_init_mask(int gtid) { |
3658 | if (!KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) |
3659 | return; |
3660 | kmp_info_t *th = __kmp_threads[gtid]; |
3661 | kmp_root_t *r = th->th.th_root; |
3662 | if (r->r.r_uber_thread == th && r->r.r_affinity_assigned) { |
3663 | __kmp_set_system_affinity(__kmp_affin_origMask, FALSE)(__kmp_affin_origMask)->set_system_affinity(0); |
3664 | KMP_CPU_COPY(th->th.th_affin_mask, __kmp_affin_origMask)(th->th.th_affin_mask)->copy(__kmp_affin_origMask); |
3665 | r->r.r_affinity_assigned = FALSE0; |
3666 | } |
3667 | } |
3668 | #else /* KMP_AFFINITY_SUPPORTED */ |
3669 | #define __kmp_assign_root_init_mask() /* Nothing */ |
3670 | static inline void __kmp_reset_root_init_mask(int gtid) {} |
3671 | #endif /* KMP_AFFINITY_SUPPORTED */ |
3672 | // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the |
3673 | // format string is for affinity, so platforms that do not support |
3674 | // affinity can still use the other fields, e.g., %n for num_threads |
3675 | extern size_t __kmp_aux_capture_affinity(int gtid, const char *format, |
3676 | kmp_str_buf_t *buffer); |
3677 | extern void __kmp_aux_display_affinity(int gtid, const char *format); |
3678 | |
3679 | extern void __kmp_cleanup_hierarchy(); |
3680 | extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar); |
3681 | |
3682 | #if KMP_USE_FUTEX(1 && (0 || 1 || KMP_ARCH_ARM || 0)) |
3683 | |
3684 | extern int __kmp_futex_determine_capable(void); |
3685 | |
3686 | #endif // KMP_USE_FUTEX |
3687 | |
3688 | extern void __kmp_gtid_set_specific(int gtid); |
3689 | extern int __kmp_gtid_get_specific(void); |
3690 | |
3691 | extern double __kmp_read_cpu_time(void); |
3692 | |
3693 | extern int __kmp_read_system_info(struct kmp_sys_info *info); |
3694 | |
3695 | #if KMP_USE_MONITOR |
3696 | extern void __kmp_create_monitor(kmp_info_t *th); |
3697 | #endif |
3698 | |
3699 | extern void *__kmp_launch_thread(kmp_info_t *thr); |
3700 | |
3701 | extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size); |
3702 | |
3703 | #if KMP_OS_WINDOWS0 |
3704 | extern int __kmp_still_running(kmp_info_t *th); |
3705 | extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val); |
3706 | extern void __kmp_free_handle(kmp_thread_t tHandle); |
3707 | #endif |
3708 | |
3709 | #if KMP_USE_MONITOR |
3710 | extern void __kmp_reap_monitor(kmp_info_t *th); |
3711 | #endif |
3712 | extern void __kmp_reap_worker(kmp_info_t *th); |
3713 | extern void __kmp_terminate_thread(int gtid); |
3714 | |
3715 | extern int __kmp_try_suspend_mx(kmp_info_t *th); |
3716 | extern void __kmp_lock_suspend_mx(kmp_info_t *th); |
3717 | extern void __kmp_unlock_suspend_mx(kmp_info_t *th); |
3718 | |
3719 | extern void __kmp_elapsed(double *); |
3720 | extern void __kmp_elapsed_tick(double *); |
3721 | |
3722 | extern void __kmp_enable(int old_state); |
3723 | extern void __kmp_disable(int *old_state); |
3724 | |
3725 | extern void __kmp_thread_sleep(int millis); |
3726 | |
3727 | extern void __kmp_common_initialize(void); |
3728 | extern void __kmp_common_destroy(void); |
3729 | extern void __kmp_common_destroy_gtid(int gtid); |
3730 | |
3731 | #if KMP_OS_UNIX1 |
3732 | extern void __kmp_register_atfork(void); |
3733 | #endif |
3734 | extern void __kmp_suspend_initialize(void); |
3735 | extern void __kmp_suspend_initialize_thread(kmp_info_t *th); |
3736 | extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th); |
3737 | |
3738 | extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team, |
3739 | int tid); |
3740 | extern kmp_team_t * |
3741 | __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc, |
3742 | #if OMPT_SUPPORT1 |
3743 | ompt_data_t ompt_parallel_data, |
3744 | #endif |
3745 | kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs, |
3746 | int argc USE_NESTED_HOT_ARG(kmp_info_t *thr), kmp_info_t *thr); |
3747 | extern void __kmp_free_thread(kmp_info_t *); |
3748 | extern void __kmp_free_team(kmp_root_t *, |
3749 | kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *), kmp_info_t *); |
3750 | extern kmp_team_t *__kmp_reap_team(kmp_team_t *); |
3751 | |
3752 | /* ------------------------------------------------------------------------ */ |
3753 | |
3754 | extern void __kmp_initialize_bget(kmp_info_t *th); |
3755 | extern void __kmp_finalize_bget(kmp_info_t *th); |
3756 | |
3757 | KMP_EXPORTextern void *kmpc_malloc(size_t size); |
3758 | KMP_EXPORTextern void *kmpc_aligned_malloc(size_t size, size_t alignment); |
3759 | KMP_EXPORTextern void *kmpc_calloc(size_t nelem, size_t elsize); |
3760 | KMP_EXPORTextern void *kmpc_realloc(void *ptr, size_t size); |
3761 | KMP_EXPORTextern void kmpc_free(void *ptr); |
3762 | |
3763 | /* declarations for internal use */ |
3764 | |
3765 | extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split, |
3766 | size_t reduce_size, void *reduce_data, |
3767 | void (*reduce)(void *, void *)); |
3768 | extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid); |
3769 | extern int __kmp_barrier_gomp_cancel(int gtid); |
3770 | |
3771 | /*! |
3772 | * Tell the fork call which compiler generated the fork call, and therefore how |
3773 | * to deal with the call. |
3774 | */ |
3775 | enum fork_context_e { |
3776 | fork_context_gnu, /**< Called from GNU generated code, so must not invoke the |
3777 | microtask internally. */ |
3778 | fork_context_intel, /**< Called from Intel generated code. */ |
3779 | fork_context_last |
3780 | }; |
3781 | extern int __kmp_fork_call(ident_t *loc, int gtid, |
3782 | enum fork_context_e fork_context, kmp_int32 argc, |
3783 | microtask_t microtask, launch_t invoker, |
3784 | kmp_va_list ap); |
3785 | |
3786 | extern void __kmp_join_call(ident_t *loc, int gtid |
3787 | #if OMPT_SUPPORT1 |
3788 | , |
3789 | enum fork_context_e fork_context |
3790 | #endif |
3791 | , |
3792 | int exit_teams = 0); |
3793 | |
3794 | extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid); |
3795 | extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team); |
3796 | extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team); |
3797 | extern int __kmp_invoke_task_func(int gtid); |
3798 | extern void __kmp_run_before_invoked_task(int gtid, int tid, |
3799 | kmp_info_t *this_thr, |
3800 | kmp_team_t *team); |
3801 | extern void __kmp_run_after_invoked_task(int gtid, int tid, |
3802 | kmp_info_t *this_thr, |
3803 | kmp_team_t *team); |
3804 | |
3805 | // should never have been exported |
3806 | KMP_EXPORTextern int __kmpc_invoke_task_func(int gtid); |
3807 | extern int __kmp_invoke_teams_master(int gtid); |
3808 | extern void __kmp_teams_master(int gtid); |
3809 | extern int __kmp_aux_get_team_num(); |
3810 | extern int __kmp_aux_get_num_teams(); |
3811 | extern void __kmp_save_internal_controls(kmp_info_t *thread); |
3812 | extern void __kmp_user_set_library(enum library_type arg); |
3813 | extern void __kmp_aux_set_library(enum library_type arg); |
3814 | extern void __kmp_aux_set_stacksize(size_t arg); |
3815 | extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid); |
3816 | extern void __kmp_aux_set_defaults(char const *str, size_t len); |
3817 | |
3818 | /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */ |
3819 | void kmpc_set_blocktime(int arg); |
3820 | void ompc_set_nested(int flag); |
3821 | void ompc_set_dynamic(int flag); |
3822 | void ompc_set_num_threads(int arg); |
3823 | |
3824 | extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr, |
3825 | kmp_team_t *team, int tid); |
3826 | extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr); |
3827 | extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid, |
3828 | kmp_tasking_flags_t *flags, |
3829 | size_t sizeof_kmp_task_t, |
3830 | size_t sizeof_shareds, |
3831 | kmp_routine_entry_t task_entry); |
3832 | extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr, |
3833 | kmp_team_t *team, int tid, |
3834 | int set_curr_task); |
3835 | extern void __kmp_finish_implicit_task(kmp_info_t *this_thr); |
3836 | extern void __kmp_free_implicit_task(kmp_info_t *this_thr); |
3837 | |
3838 | extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref, |
3839 | int gtid, |
3840 | kmp_task_t *task); |
3841 | extern void __kmp_fulfill_event(kmp_event_t *event); |
3842 | |
3843 | extern void __kmp_free_task_team(kmp_info_t *thread, |
3844 | kmp_task_team_t *task_team); |
3845 | extern void __kmp_reap_task_teams(void); |
3846 | extern void __kmp_wait_to_unref_task_teams(void); |
3847 | extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team, |
3848 | int always); |
3849 | extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team); |
3850 | extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team |
3851 | #if USE_ITT_BUILD1 |
3852 | , |
3853 | void *itt_sync_obj |
3854 | #endif /* USE_ITT_BUILD */ |
3855 | , |
3856 | int wait = 1); |
3857 | extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread, |
3858 | int gtid); |
3859 | |
3860 | extern int __kmp_is_address_mapped(void *addr); |
3861 | extern kmp_uint64 __kmp_hardware_timestamp(void); |
3862 | |
3863 | #if KMP_OS_UNIX1 |
3864 | extern int __kmp_read_from_file(char const *path, char const *format, ...); |
3865 | #endif |
3866 | |
3867 | /* ------------------------------------------------------------------------ */ |
3868 | // |
3869 | // Assembly routines that have no compiler intrinsic replacement |
3870 | // |
3871 | |
3872 | extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc, |
3873 | void *argv[] |
3874 | #if OMPT_SUPPORT1 |
3875 | , |
3876 | void **exit_frame_ptr |
3877 | #endif |
3878 | ); |
3879 | |
3880 | /* ------------------------------------------------------------------------ */ |
3881 | |
3882 | KMP_EXPORTextern void __kmpc_begin(ident_t *, kmp_int32 flags); |
3883 | KMP_EXPORTextern void __kmpc_end(ident_t *); |
3884 | |
3885 | KMP_EXPORTextern void __kmpc_threadprivate_register_vec(ident_t *, void *data, |
3886 | kmpc_ctor_vec ctor, |
3887 | kmpc_cctor_vec cctor, |
3888 | kmpc_dtor_vec dtor, |
3889 | size_t vector_length); |
3890 | KMP_EXPORTextern void __kmpc_threadprivate_register(ident_t *, void *data, |
3891 | kmpc_ctor ctor, kmpc_cctor cctor, |
3892 | kmpc_dtor dtor); |
3893 | KMP_EXPORTextern void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid, |
3894 | void *data, size_t size); |
3895 | |
3896 | KMP_EXPORTextern kmp_int32 __kmpc_global_thread_num(ident_t *); |
3897 | KMP_EXPORTextern kmp_int32 __kmpc_global_num_threads(ident_t *); |
3898 | KMP_EXPORTextern kmp_int32 __kmpc_bound_thread_num(ident_t *); |
3899 | KMP_EXPORTextern kmp_int32 __kmpc_bound_num_threads(ident_t *); |
3900 | |
3901 | KMP_EXPORTextern kmp_int32 __kmpc_ok_to_fork(ident_t *); |
3902 | KMP_EXPORTextern void __kmpc_fork_call(ident_t *, kmp_int32 nargs, |
3903 | kmpc_micro microtask, ...); |
3904 | |
3905 | KMP_EXPORTextern void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid); |
3906 | KMP_EXPORTextern void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid); |
3907 | |
3908 | KMP_EXPORTextern void __kmpc_flush(ident_t *); |
3909 | KMP_EXPORTextern void __kmpc_barrier(ident_t *, kmp_int32 global_tid); |
3910 | KMP_EXPORTextern kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid); |
3911 | KMP_EXPORTextern void __kmpc_end_master(ident_t *, kmp_int32 global_tid); |
3912 | KMP_EXPORTextern kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid, |
3913 | kmp_int32 filter); |
3914 | KMP_EXPORTextern void __kmpc_end_masked(ident_t *, kmp_int32 global_tid); |
3915 | KMP_EXPORTextern void __kmpc_ordered(ident_t *, kmp_int32 global_tid); |
3916 | KMP_EXPORTextern void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid); |
3917 | KMP_EXPORTextern void __kmpc_critical(ident_t *, kmp_int32 global_tid, |
3918 | kmp_critical_name *); |
3919 | KMP_EXPORTextern void __kmpc_end_critical(ident_t *, kmp_int32 global_tid, |
3920 | kmp_critical_name *); |
3921 | KMP_EXPORTextern void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid, |
3922 | kmp_critical_name *, uint32_t hint); |
3923 | |
3924 | KMP_EXPORTextern kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid); |
3925 | KMP_EXPORTextern void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid); |
3926 | |
3927 | KMP_EXPORTextern kmp_int32 __kmpc_barrier_master_nowait(ident_t *, |
3928 | kmp_int32 global_tid); |
3929 | |
3930 | KMP_EXPORTextern kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid); |
3931 | KMP_EXPORTextern void __kmpc_end_single(ident_t *, kmp_int32 global_tid); |
3932 | |
3933 | KMP_EXPORTextern kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid); |
3934 | KMP_EXPORTextern kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid, |
3935 | kmp_int32 numberOfSections); |
3936 | KMP_EXPORTextern void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid); |
3937 | |
3938 | KMP_EXPORTextern void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid, |
3939 | kmp_int32 schedtype, kmp_int32 *plastiter, |
3940 | kmp_int *plower, kmp_int *pupper, |
3941 | kmp_int *pstride, kmp_int incr, |
3942 | kmp_int chunk); |
3943 | |
3944 | KMP_EXPORTextern void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid); |
3945 | |
3946 | KMP_EXPORTextern void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid, |
3947 | size_t cpy_size, void *cpy_data, |
3948 | void (*cpy_func)(void *, void *), |
3949 | kmp_int32 didit); |
3950 | |
3951 | KMP_EXPORTextern void *__kmpc_copyprivate_light(ident_t *loc, kmp_int32 gtid, |
3952 | void *cpy_data); |
3953 | |
3954 | extern void KMPC_SET_NUM_THREADS(int arg); |
3955 | extern void KMPC_SET_DYNAMIC(int flag); |
3956 | extern void KMPC_SET_NESTED(int flag); |
3957 | |
3958 | /* OMP 3.0 tasking interface routines */ |
3959 | KMP_EXPORTextern kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid, |
3960 | kmp_task_t *new_task); |
3961 | KMP_EXPORTextern kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid, |
3962 | kmp_int32 flags, |
3963 | size_t sizeof_kmp_task_t, |
3964 | size_t sizeof_shareds, |
3965 | kmp_routine_entry_t task_entry); |
3966 | KMP_EXPORTextern kmp_task_t *__kmpc_omp_target_task_alloc( |
3967 | ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, |
3968 | size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id); |
3969 | KMP_EXPORTextern void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid, |
3970 | kmp_task_t *task); |
3971 | KMP_EXPORTextern void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid, |
3972 | kmp_task_t *task); |
3973 | KMP_EXPORTextern kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid, |
3974 | kmp_task_t *new_task); |
3975 | KMP_EXPORTextern kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid); |
3976 | /* __kmpc_omp_taskwait_51 : Function for OpenMP 5.1 nowait clause. |
3977 | * Placeholder for taskwait with nowait clause.*/ |
3978 | KMP_EXPORTextern kmp_int32 __kmpc_omp_taskwait_51(ident_t *loc_ref, kmp_int32 gtid, |
3979 | kmp_int32 has_no_wait); |
3980 | |
3981 | KMP_EXPORTextern kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid, |
3982 | int end_part); |
3983 | |
3984 | #if TASK_UNUSED |
3985 | void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task); |
3986 | void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid, |
3987 | kmp_task_t *task); |
3988 | #endif // TASK_UNUSED |
3989 | |
3990 | /* ------------------------------------------------------------------------ */ |
3991 | |
3992 | KMP_EXPORTextern void __kmpc_taskgroup(ident_t *loc, int gtid); |
3993 | KMP_EXPORTextern void __kmpc_end_taskgroup(ident_t *loc, int gtid); |
3994 | |
3995 | KMP_EXPORTextern kmp_int32 __kmpc_omp_task_with_deps( |
3996 | ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps, |
3997 | kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, |
3998 | kmp_depend_info_t *noalias_dep_list); |
3999 | KMP_EXPORTextern void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid, |
4000 | kmp_int32 ndeps, |
4001 | kmp_depend_info_t *dep_list, |
4002 | kmp_int32 ndeps_noalias, |
4003 | kmp_depend_info_t *noalias_dep_list); |
4004 | /* __kmpc_omp_taskwait_deps_51 : Function for OpenMP 5.1 nowait clause. |
4005 | * Placeholder for taskwait with nowait clause.*/ |
4006 | KMP_EXPORTextern void __kmpc_omp_taskwait_deps_51(ident_t *loc_ref, kmp_int32 gtid, |
4007 | kmp_int32 ndeps, |
4008 | kmp_depend_info_t *dep_list, |
4009 | kmp_int32 ndeps_noalias, |
4010 | kmp_depend_info_t *noalias_dep_list, |
4011 | kmp_int32 has_no_wait); |
4012 | |
4013 | extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task, |
4014 | bool serialize_immediate); |
4015 | |
4016 | KMP_EXPORTextern kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid, |
4017 | kmp_int32 cncl_kind); |
4018 | KMP_EXPORTextern kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid, |
4019 | kmp_int32 cncl_kind); |
4020 | KMP_EXPORTextern kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid); |
4021 | KMP_EXPORTextern int __kmp_get_cancellation_status(int cancel_kind); |
4022 | |
4023 | KMP_EXPORTextern void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask); |
4024 | KMP_EXPORTextern void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask); |
4025 | KMP_EXPORTextern void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task, |
4026 | kmp_int32 if_val, kmp_uint64 *lb, |
4027 | kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup, |
4028 | kmp_int32 sched, kmp_uint64 grainsize, |
4029 | void *task_dup); |
4030 | KMP_EXPORTextern void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid, |
4031 | kmp_task_t *task, kmp_int32 if_val, |
4032 | kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, |
4033 | kmp_int32 nogroup, kmp_int32 sched, |
4034 | kmp_uint64 grainsize, kmp_int32 modifier, |
4035 | void *task_dup); |
4036 | KMP_EXPORTextern void *__kmpc_task_reduction_init(int gtid, int num_data, void *data); |
4037 | KMP_EXPORTextern void *__kmpc_taskred_init(int gtid, int num_data, void *data); |
4038 | KMP_EXPORTextern void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d); |
4039 | KMP_EXPORTextern void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid, |
4040 | int is_ws, int num, |
4041 | void *data); |
4042 | KMP_EXPORTextern void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws, |
4043 | int num, void *data); |
4044 | KMP_EXPORTextern void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid, |
4045 | int is_ws); |
4046 | KMP_EXPORTextern kmp_int32 __kmpc_omp_reg_task_with_affinity( |
4047 | ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins, |
4048 | kmp_task_affinity_info_t *affin_list); |
4049 | KMP_EXPORTextern void __kmp_set_num_teams(int num_teams); |
4050 | KMP_EXPORTextern int __kmp_get_max_teams(void); |
4051 | KMP_EXPORTextern void __kmp_set_teams_thread_limit(int limit); |
4052 | KMP_EXPORTextern int __kmp_get_teams_thread_limit(void); |
4053 | |
4054 | /* Lock interface routines (fast versions with gtid passed in) */ |
4055 | KMP_EXPORTextern void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid, |
4056 | void **user_lock); |
4057 | KMP_EXPORTextern void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid, |
4058 | void **user_lock); |
4059 | KMP_EXPORTextern void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid, |
4060 | void **user_lock); |
4061 | KMP_EXPORTextern void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid, |
4062 | void **user_lock); |
4063 | KMP_EXPORTextern void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock); |
4064 | KMP_EXPORTextern void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid, |
4065 | void **user_lock); |
4066 | KMP_EXPORTextern void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid, |
4067 | void **user_lock); |
4068 | KMP_EXPORTextern void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid, |
4069 | void **user_lock); |
4070 | KMP_EXPORTextern int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock); |
4071 | KMP_EXPORTextern int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid, |
4072 | void **user_lock); |
4073 | |
4074 | KMP_EXPORTextern void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid, |
4075 | void **user_lock, uintptr_t hint); |
4076 | KMP_EXPORTextern void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid, |
4077 | void **user_lock, |
4078 | uintptr_t hint); |
4079 | |
4080 | /* Interface to fast scalable reduce methods routines */ |
4081 | |
4082 | KMP_EXPORTextern kmp_int32 __kmpc_reduce_nowait( |
4083 | ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, |
4084 | void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data), |
4085 | kmp_critical_name *lck); |
4086 | KMP_EXPORTextern void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid, |
4087 | kmp_critical_name *lck); |
4088 | KMP_EXPORTextern kmp_int32 __kmpc_reduce( |
4089 | ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, |
4090 | void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data), |
4091 | kmp_critical_name *lck); |
4092 | KMP_EXPORTextern void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid, |
4093 | kmp_critical_name *lck); |
4094 | |
4095 | /* Internal fast reduction routines */ |
4096 | |
4097 | extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method( |
4098 | ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, |
4099 | void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data), |
4100 | kmp_critical_name *lck); |
4101 | |
4102 | // this function is for testing set/get/determine reduce method |
4103 | KMP_EXPORTextern kmp_int32 __kmp_get_reduce_method(void); |
4104 | |
4105 | KMP_EXPORTextern kmp_uint64 __kmpc_get_taskid(); |
4106 | KMP_EXPORTextern kmp_uint64 __kmpc_get_parent_taskid(); |
4107 | |
4108 | // C++ port |
4109 | // missing 'extern "C"' declarations |
4110 | |
4111 | KMP_EXPORTextern kmp_int32 __kmpc_in_parallel(ident_t *loc); |
4112 | KMP_EXPORTextern void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid); |
4113 | KMP_EXPORTextern void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid, |
4114 | kmp_int32 num_threads); |
4115 | |
4116 | KMP_EXPORTextern void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid, |
4117 | int proc_bind); |
4118 | KMP_EXPORTextern void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid, |
4119 | kmp_int32 num_teams, |
4120 | kmp_int32 num_threads); |
4121 | /* Function for OpenMP 5.1 num_teams clause */ |
4122 | KMP_EXPORTextern void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid, |
4123 | kmp_int32 num_teams_lb, |
4124 | kmp_int32 num_teams_ub, |
4125 | kmp_int32 num_threads); |
4126 | KMP_EXPORTextern void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc, |
4127 | kmpc_micro microtask, ...); |
4128 | struct kmp_dim { // loop bounds info casted to kmp_int64 |
4129 | kmp_int64 lo; // lower |
4130 | kmp_int64 up; // upper |
4131 | kmp_int64 st; // stride |
4132 | }; |
4133 | KMP_EXPORTextern void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid, |
4134 | kmp_int32 num_dims, |
4135 | const struct kmp_dim *dims); |
4136 | KMP_EXPORTextern void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid, |
4137 | const kmp_int64 *vec); |
4138 | KMP_EXPORTextern void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid, |
4139 | const kmp_int64 *vec); |
4140 | KMP_EXPORTextern void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid); |
4141 | |
4142 | KMP_EXPORTextern void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid, |
4143 | void *data, size_t size, |
4144 | void ***cache); |
4145 | |
4146 | // Symbols for MS mutual detection. |
4147 | extern int _You_must_link_with_exactly_one_OpenMP_library; |
4148 | extern int _You_must_link_with_Intel_OpenMP_library; |
4149 | #if KMP_OS_WINDOWS0 && (KMP_VERSION_MAJOR5 > 4) |
4150 | extern int _You_must_link_with_Microsoft_OpenMP_library; |
4151 | #endif |
4152 | |
4153 | // The routines below are not exported. |
4154 | // Consider making them 'static' in corresponding source files. |
4155 | void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr, |
4156 | void *data_addr, size_t pc_size); |
4157 | struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr, |
4158 | void *data_addr, |
4159 | size_t pc_size); |
4160 | void __kmp_threadprivate_resize_cache(int newCapacity); |
4161 | void __kmp_cleanup_threadprivate_caches(); |
4162 | |
4163 | // ompc_, kmpc_ entries moved from omp.h. |
4164 | #if KMP_OS_WINDOWS0 |
4165 | #define KMPC_CONVENTION __cdecl |
4166 | #else |
4167 | #define KMPC_CONVENTION |
4168 | #endif |
4169 | |
4170 | #ifndef __OMP_H |
4171 | typedef enum omp_sched_t { |
4172 | omp_sched_static = 1, |
4173 | omp_sched_dynamic = 2, |
4174 | omp_sched_guided = 3, |
4175 | omp_sched_auto = 4 |
4176 | } omp_sched_t; |
4177 | typedef void *kmp_affinity_mask_t; |
4178 | #endif |
4179 | |
4180 | KMP_EXPORTextern void KMPC_CONVENTION ompc_set_max_active_levels(int); |
4181 | KMP_EXPORTextern void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int); |
4182 | KMP_EXPORTextern int KMPC_CONVENTION ompc_get_ancestor_thread_num(int); |
4183 | KMP_EXPORTextern int KMPC_CONVENTION ompc_get_team_size(int); |
4184 | KMP_EXPORTextern int KMPC_CONVENTION |
4185 | kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *); |
4186 | KMP_EXPORTextern int KMPC_CONVENTION |
4187 | kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *); |
4188 | KMP_EXPORTextern int KMPC_CONVENTION |
4189 | kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *); |
4190 | |
4191 | KMP_EXPORTextern void KMPC_CONVENTION kmpc_set_stacksize(int); |
4192 | KMP_EXPORTextern void KMPC_CONVENTION kmpc_set_stacksize_s(size_t); |
4193 | KMP_EXPORTextern void KMPC_CONVENTION kmpc_set_library(int); |
4194 | KMP_EXPORTextern void KMPC_CONVENTION kmpc_set_defaults(char const *); |
4195 | KMP_EXPORTextern void KMPC_CONVENTION kmpc_set_disp_num_buffers(int); |
4196 | void KMP_EXPAND_NAME(ompc_set_affinity_format)__kmp_api_ompc_set_affinity_format(char const *format); |
4197 | size_t KMP_EXPAND_NAME(ompc_get_affinity_format)__kmp_api_ompc_get_affinity_format(char *buffer, size_t size); |
4198 | void KMP_EXPAND_NAME(ompc_display_affinity)__kmp_api_ompc_display_affinity(char const *format); |
4199 | size_t KMP_EXPAND_NAME(ompc_capture_affinity)__kmp_api_ompc_capture_affinity(char *buffer, size_t buf_size, |
4200 | char const *format); |
4201 | |
4202 | enum kmp_target_offload_kind { |
4203 | tgt_disabled = 0, |
4204 | tgt_default = 1, |
4205 | tgt_mandatory = 2 |
4206 | }; |
4207 | typedef enum kmp_target_offload_kind kmp_target_offload_kind_t; |
4208 | // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise |
4209 | extern kmp_target_offload_kind_t __kmp_target_offload; |
4210 | extern int __kmpc_get_target_offload(); |
4211 | |
4212 | // Constants used in libomptarget |
4213 | #define KMP_DEVICE_DEFAULT-1 -1 // This is libomptarget's default device. |
4214 | #define KMP_DEVICE_ALL-11 -11 // This is libomptarget's "all devices". |
4215 | |
4216 | // OMP Pause Resource |
4217 | |
4218 | // The following enum is used both to set the status in __kmp_pause_status, and |
4219 | // as the internal equivalent of the externally-visible omp_pause_resource_t. |
4220 | typedef enum kmp_pause_status_t { |
4221 | kmp_not_paused = 0, // status is not paused, or, requesting resume |
4222 | kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause |
4223 | kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause |
4224 | } kmp_pause_status_t; |
4225 | |
4226 | // This stores the pause state of the runtime |
4227 | extern kmp_pause_status_t __kmp_pause_status; |
4228 | extern int __kmpc_pause_resource(kmp_pause_status_t level); |
4229 | extern int __kmp_pause_resource(kmp_pause_status_t level); |
4230 | // Soft resume sets __kmp_pause_status, and wakes up all threads. |
4231 | extern void __kmp_resume_if_soft_paused(); |
4232 | // Hard resume simply resets the status to not paused. Library will appear to |
4233 | // be uninitialized after hard pause. Let OMP constructs trigger required |
4234 | // initializations. |
4235 | static inline void __kmp_resume_if_hard_paused() { |
4236 | if (__kmp_pause_status == kmp_hard_paused) { |
4237 | __kmp_pause_status = kmp_not_paused; |
4238 | } |
4239 | } |
4240 | |
4241 | extern void __kmp_omp_display_env(int verbose); |
4242 | |
4243 | // 1: it is initializing hidden helper team |
4244 | extern volatile int __kmp_init_hidden_helper; |
4245 | // 1: the hidden helper team is done |
4246 | extern volatile int __kmp_hidden_helper_team_done; |
4247 | // 1: enable hidden helper task |
4248 | extern kmp_int32 __kmp_enable_hidden_helper; |
4249 | // Main thread of hidden helper team |
4250 | extern kmp_info_t *__kmp_hidden_helper_main_thread; |
4251 | // Descriptors for the hidden helper threads |
4252 | extern kmp_info_t **__kmp_hidden_helper_threads; |
4253 | // Number of hidden helper threads |
4254 | extern kmp_int32 __kmp_hidden_helper_threads_num; |
4255 | // Number of hidden helper tasks that have not been executed yet |
4256 | extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks; |
4257 | |
4258 | extern void __kmp_hidden_helper_initialize(); |
4259 | extern void __kmp_hidden_helper_threads_initz_routine(); |
4260 | extern void __kmp_do_initialize_hidden_helper_threads(); |
4261 | extern void __kmp_hidden_helper_threads_initz_wait(); |
4262 | extern void __kmp_hidden_helper_initz_release(); |
4263 | extern void __kmp_hidden_helper_threads_deinitz_wait(); |
4264 | extern void __kmp_hidden_helper_threads_deinitz_release(); |
4265 | extern void __kmp_hidden_helper_main_thread_wait(); |
4266 | extern void __kmp_hidden_helper_worker_thread_wait(); |
4267 | extern void __kmp_hidden_helper_worker_thread_signal(); |
4268 | extern void __kmp_hidden_helper_main_thread_release(); |
4269 | |
4270 | // Check whether a given thread is a hidden helper thread |
4271 | #define KMP_HIDDEN_HELPER_THREAD(gtid)((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num ) \ |
4272 | ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num) |
4273 | |
4274 | #define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid)((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num ) \ |
4275 | ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num) |
4276 | |
4277 | #define KMP_HIDDEN_HELPER_MAIN_THREAD(gtid)((gtid) == 1 && (gtid) <= __kmp_hidden_helper_threads_num ) \ |
4278 | ((gtid) == 1 && (gtid) <= __kmp_hidden_helper_threads_num) |
4279 | |
4280 | #define KMP_HIDDEN_HELPER_TEAM(team)(team->t.t_threads[0] == __kmp_hidden_helper_main_thread) \ |
4281 | (team->t.t_threads[0] == __kmp_hidden_helper_main_thread) |
4282 | |
4283 | // Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a |
4284 | // main thread, is skipped. |
4285 | #define KMP_GTID_TO_SHADOW_GTID(gtid)((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2) \ |
4286 | ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2) |
4287 | |
4288 | // Return the adjusted gtid value by subtracting from gtid the number |
4289 | // of hidden helper threads. This adjusted value is the gtid the thread would |
4290 | // have received if there were no hidden helper threads. |
4291 | static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) { |
4292 | int adjusted_gtid = gtid; |
4293 | if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 && |
4294 | gtid - __kmp_hidden_helper_threads_num >= 0) { |
4295 | adjusted_gtid -= __kmp_hidden_helper_threads_num; |
4296 | } |
4297 | return adjusted_gtid; |
4298 | } |
4299 | |
4300 | // Support for error directive |
4301 | typedef enum kmp_severity_t { |
4302 | severity_warning = 1, |
4303 | severity_fatal = 2 |
4304 | } kmp_severity_t; |
4305 | extern void __kmpc_error(ident_t *loc, int severity, const char *message); |
4306 | |
4307 | // Support for scope directive |
4308 | KMP_EXPORTextern void __kmpc_scope(ident_t *loc, kmp_int32 gtid, void *reserved); |
4309 | KMP_EXPORTextern void __kmpc_end_scope(ident_t *loc, kmp_int32 gtid, void *reserved); |
4310 | |
4311 | #ifdef __cplusplus201703L |
4312 | } |
4313 | #endif |
4314 | |
4315 | template <bool C, bool S> |
4316 | extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag); |
4317 | template <bool C, bool S> |
4318 | extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag); |
4319 | template <bool C, bool S> |
4320 | extern void __kmp_atomic_suspend_64(int th_gtid, |
4321 | kmp_atomic_flag_64<C, S> *flag); |
4322 | extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag); |
4323 | #if KMP_HAVE_MWAIT((0 || 1) && (1 || 0) && !0) || KMP_HAVE_UMWAIT((0 || 1) && (1 || 0) && !0) |
4324 | template <bool C, bool S> |
4325 | extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag); |
4326 | template <bool C, bool S> |
4327 | extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag); |
4328 | template <bool C, bool S> |
4329 | extern void __kmp_atomic_mwait_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag); |
4330 | extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag); |
4331 | #endif |
4332 | template <bool C, bool S> |
4333 | extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag); |
4334 | template <bool C, bool S> |
4335 | extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag); |
4336 | template <bool C, bool S> |
4337 | extern void __kmp_atomic_resume_64(int target_gtid, |
4338 | kmp_atomic_flag_64<C, S> *flag); |
4339 | extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag); |
4340 | |
4341 | template <bool C, bool S> |
4342 | int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid, |
4343 | kmp_flag_32<C, S> *flag, int final_spin, |
4344 | int *thread_finished, |
4345 | #if USE_ITT_BUILD1 |
4346 | void *itt_sync_obj, |
4347 | #endif /* USE_ITT_BUILD */ |
4348 | kmp_int32 is_constrained); |
4349 | template <bool C, bool S> |
4350 | int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid, |
4351 | kmp_flag_64<C, S> *flag, int final_spin, |
4352 | int *thread_finished, |
4353 | #if USE_ITT_BUILD1 |
4354 | void *itt_sync_obj, |
4355 | #endif /* USE_ITT_BUILD */ |
4356 | kmp_int32 is_constrained); |
4357 | template <bool C, bool S> |
4358 | int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid, |
4359 | kmp_atomic_flag_64<C, S> *flag, |
4360 | int final_spin, int *thread_finished, |
4361 | #if USE_ITT_BUILD1 |
4362 | void *itt_sync_obj, |
4363 | #endif /* USE_ITT_BUILD */ |
4364 | kmp_int32 is_constrained); |
4365 | int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid, |
4366 | kmp_flag_oncore *flag, int final_spin, |
4367 | int *thread_finished, |
4368 | #if USE_ITT_BUILD1 |
4369 | void *itt_sync_obj, |
4370 | #endif /* USE_ITT_BUILD */ |
4371 | kmp_int32 is_constrained); |
4372 | |
4373 | extern int __kmp_nesting_mode; |
4374 | extern int __kmp_nesting_mode_nlevels; |
4375 | extern int *__kmp_nesting_nth_level; |
4376 | extern void __kmp_init_nesting_mode(); |
4377 | extern void __kmp_set_nesting_mode_threads(); |
4378 | |
4379 | /// This class safely opens and closes a C-style FILE* object using RAII |
4380 | /// semantics. There are also methods which allow using stdout or stderr as |
4381 | /// the underlying FILE* object. With the implicit conversion operator to |
4382 | /// FILE*, an object with this type can be used in any function which takes |
4383 | /// a FILE* object e.g., fprintf(). |
4384 | /// No close method is needed at use sites. |
4385 | class kmp_safe_raii_file_t { |
4386 | FILE *f; |
4387 | |
4388 | void close() { |
4389 | if (f && f != stdoutstdout && f != stderrstderr) { |
4390 | fclose(f); |
4391 | f = nullptr; |
4392 | } |
4393 | } |
4394 | |
4395 | public: |
4396 | kmp_safe_raii_file_t() : f(nullptr) {} |
4397 | kmp_safe_raii_file_t(const char *filename, const char *mode, |
4398 | const char *env_var = nullptr) |
4399 | : f(nullptr) { |
4400 | open(filename, mode, env_var); |
4401 | } |
4402 | ~kmp_safe_raii_file_t() { close(); } |
4403 | |
4404 | /// Open filename using mode. This is automatically closed in the destructor. |
4405 | /// The env_var parameter indicates the environment variable the filename |
4406 | /// came from if != nullptr. |
4407 | void open(const char *filename, const char *mode, |
4408 | const char *env_var = nullptr) { |
4409 | KMP_ASSERT(!f)if (!(!f)) { __kmp_debug_assert("!f", "openmp/runtime/src/kmp.h" , 4409); }; |
4410 | f = fopen(filename, mode); |
4411 | if (!f) { |
4412 | int code = errno(*__errno_location ()); |
4413 | if (env_var) { |
4414 | __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename)__kmp_msg_format(kmp_i18n_msg_CantOpenFileForReading, filename ), KMP_ERR(code)__kmp_msg_error_code(code), |
4415 | KMP_HNT(CheckEnvVar, env_var, filename)__kmp_msg_format(kmp_i18n_hnt_CheckEnvVar, env_var, filename), __kmp_msg_null); |
4416 | } else { |
4417 | __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename)__kmp_msg_format(kmp_i18n_msg_CantOpenFileForReading, filename ), KMP_ERR(code)__kmp_msg_error_code(code), |
4418 | __kmp_msg_null); |
4419 | } |
4420 | } |
4421 | } |
4422 | /// Instead of erroring out, return non-zero when |
4423 | /// unsuccessful fopen() for any reason |
4424 | int try_open(const char *filename, const char *mode) { |
4425 | KMP_ASSERT(!f)if (!(!f)) { __kmp_debug_assert("!f", "openmp/runtime/src/kmp.h" , 4425); }; |
4426 | f = fopen(filename, mode); |
4427 | if (!f) |
4428 | return errno(*__errno_location ()); |
4429 | return 0; |
4430 | } |
4431 | /// Set the FILE* object to stdout and output there |
4432 | /// No open call should happen before this call. |
4433 | void set_stdout() { |
4434 | KMP_ASSERT(!f)if (!(!f)) { __kmp_debug_assert("!f", "openmp/runtime/src/kmp.h" , 4434); }; |
4435 | f = stdoutstdout; |
4436 | } |
4437 | /// Set the FILE* object to stderr and output there |
4438 | /// No open call should happen before this call. |
4439 | void set_stderr() { |
4440 | KMP_ASSERT(!f)if (!(!f)) { __kmp_debug_assert("!f", "openmp/runtime/src/kmp.h" , 4440); }; |
4441 | f = stderrstderr; |
4442 | } |
4443 | operator bool() { return bool(f); } |
4444 | operator FILE *() { return f; } |
4445 | }; |
4446 | |
4447 | template <typename SourceType, typename TargetType, |
4448 | bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)), |
4449 | bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)), |
4450 | bool isSourceSigned = std::is_signed<SourceType>::value, |
4451 | bool isTargetSigned = std::is_signed<TargetType>::value> |
4452 | struct kmp_convert {}; |
4453 | |
4454 | // Both types are signed; Source smaller |
4455 | template <typename SourceType, typename TargetType> |
4456 | struct kmp_convert<SourceType, TargetType, true, false, true, true> { |
4457 | static TargetType to(SourceType src) { return (TargetType)src; } |
4458 | }; |
4459 | // Source equal |
4460 | template <typename SourceType, typename TargetType> |
4461 | struct kmp_convert<SourceType, TargetType, false, true, true, true> { |
4462 | static TargetType to(SourceType src) { return src; } |
4463 | }; |
4464 | // Source bigger |
4465 | template <typename SourceType, typename TargetType> |
4466 | struct kmp_convert<SourceType, TargetType, false, false, true, true> { |
4467 | static TargetType to(SourceType src) { |
4468 | KMP_ASSERT(src <= static_cast<SourceType>(if (!(src <= static_cast<SourceType>( (std::numeric_limits <TargetType>::max)()))) { __kmp_debug_assert("src <= static_cast<SourceType>( (std::numeric_limits<TargetType>::max)())" , "openmp/runtime/src/kmp.h", 4469); } |
4469 | (std::numeric_limits<TargetType>::max)()))if (!(src <= static_cast<SourceType>( (std::numeric_limits <TargetType>::max)()))) { __kmp_debug_assert("src <= static_cast<SourceType>( (std::numeric_limits<TargetType>::max)())" , "openmp/runtime/src/kmp.h", 4469); }; |
4470 | KMP_ASSERT(src >= static_cast<SourceType>(if (!(src >= static_cast<SourceType>( (std::numeric_limits <TargetType>::min)()))) { __kmp_debug_assert("src >= static_cast<SourceType>( (std::numeric_limits<TargetType>::min)())" , "openmp/runtime/src/kmp.h", 4471); } |
4471 | (std::numeric_limits<TargetType>::min)()))if (!(src >= static_cast<SourceType>( (std::numeric_limits <TargetType>::min)()))) { __kmp_debug_assert("src >= static_cast<SourceType>( (std::numeric_limits<TargetType>::min)())" , "openmp/runtime/src/kmp.h", 4471); }; |
4472 | return (TargetType)src; |
4473 | } |
4474 | }; |
4475 | |
4476 | // Source signed, Target unsigned |
4477 | // Source smaller |
4478 | template <typename SourceType, typename TargetType> |
4479 | struct kmp_convert<SourceType, TargetType, true, false, true, false> { |
4480 | static TargetType to(SourceType src) { |
4481 | KMP_ASSERT(src >= 0)if (!(src >= 0)) { __kmp_debug_assert("src >= 0", "openmp/runtime/src/kmp.h" , 4481); }; |
4482 | return (TargetType)src; |
4483 | } |
4484 | }; |
4485 | // Source equal |
4486 | template <typename SourceType, typename TargetType> |
4487 | struct kmp_convert<SourceType, TargetType, false, true, true, false> { |
4488 | static TargetType to(SourceType src) { |
4489 | KMP_ASSERT(src >= 0)if (!(src >= 0)) { __kmp_debug_assert("src >= 0", "openmp/runtime/src/kmp.h" , 4489); }; |
4490 | return (TargetType)src; |
4491 | } |
4492 | }; |
4493 | // Source bigger |
4494 | template <typename SourceType, typename TargetType> |
4495 | struct kmp_convert<SourceType, TargetType, false, false, true, false> { |
4496 | static TargetType to(SourceType src) { |
4497 | KMP_ASSERT(src >= 0)if (!(src >= 0)) { __kmp_debug_assert("src >= 0", "openmp/runtime/src/kmp.h" , 4497); }; |
4498 | KMP_ASSERT(src <= static_cast<SourceType>(if (!(src <= static_cast<SourceType>( (std::numeric_limits <TargetType>::max)()))) { __kmp_debug_assert("src <= static_cast<SourceType>( (std::numeric_limits<TargetType>::max)())" , "openmp/runtime/src/kmp.h", 4499); } |
4499 | (std::numeric_limits<TargetType>::max)()))if (!(src <= static_cast<SourceType>( (std::numeric_limits <TargetType>::max)()))) { __kmp_debug_assert("src <= static_cast<SourceType>( (std::numeric_limits<TargetType>::max)())" , "openmp/runtime/src/kmp.h", 4499); }; |
4500 | return (TargetType)src; |
4501 | } |
4502 | }; |
4503 | |
4504 | // Source unsigned, Target signed |
4505 | // Source smaller |
4506 | template <typename SourceType, typename TargetType> |
4507 | struct kmp_convert<SourceType, TargetType, true, false, false, true> { |
4508 | static TargetType to(SourceType src) { return (TargetType)src; } |
4509 | }; |
4510 | // Source equal |
4511 | template <typename SourceType, typename TargetType> |
4512 | struct kmp_convert<SourceType, TargetType, false, true, false, true> { |
4513 | static TargetType to(SourceType src) { |
4514 | KMP_ASSERT(src <= static_cast<SourceType>(if (!(src <= static_cast<SourceType>( (std::numeric_limits <TargetType>::max)()))) { __kmp_debug_assert("src <= static_cast<SourceType>( (std::numeric_limits<TargetType>::max)())" , "openmp/runtime/src/kmp.h", 4515); } |
4515 | (std::numeric_limits<TargetType>::max)()))if (!(src <= static_cast<SourceType>( (std::numeric_limits <TargetType>::max)()))) { __kmp_debug_assert("src <= static_cast<SourceType>( (std::numeric_limits<TargetType>::max)())" , "openmp/runtime/src/kmp.h", 4515); }; |
4516 | return (TargetType)src; |
4517 | } |
4518 | }; |
4519 | // Source bigger |
4520 | template <typename SourceType, typename TargetType> |
4521 | struct kmp_convert<SourceType, TargetType, false, false, false, true> { |
4522 | static TargetType to(SourceType src) { |
4523 | KMP_ASSERT(src <= static_cast<SourceType>(if (!(src <= static_cast<SourceType>( (std::numeric_limits <TargetType>::max)()))) { __kmp_debug_assert("src <= static_cast<SourceType>( (std::numeric_limits<TargetType>::max)())" , "openmp/runtime/src/kmp.h", 4524); } |
4524 | (std::numeric_limits<TargetType>::max)()))if (!(src <= static_cast<SourceType>( (std::numeric_limits <TargetType>::max)()))) { __kmp_debug_assert("src <= static_cast<SourceType>( (std::numeric_limits<TargetType>::max)())" , "openmp/runtime/src/kmp.h", 4524); }; |
4525 | return (TargetType)src; |
4526 | } |
4527 | }; |
4528 | |
4529 | // Source unsigned, Target unsigned |
4530 | // Source smaller |
4531 | template <typename SourceType, typename TargetType> |
4532 | struct kmp_convert<SourceType, TargetType, true, false, false, false> { |
4533 | static TargetType to(SourceType src) { return (TargetType)src; } |
4534 | }; |
4535 | // Source equal |
4536 | template <typename SourceType, typename TargetType> |
4537 | struct kmp_convert<SourceType, TargetType, false, true, false, false> { |
4538 | static TargetType to(SourceType src) { return src; } |
4539 | }; |
4540 | // Source bigger |
4541 | template <typename SourceType, typename TargetType> |
4542 | struct kmp_convert<SourceType, TargetType, false, false, false, false> { |
4543 | static TargetType to(SourceType src) { |
4544 | KMP_ASSERT(src <= static_cast<SourceType>(if (!(src <= static_cast<SourceType>( (std::numeric_limits <TargetType>::max)()))) { __kmp_debug_assert("src <= static_cast<SourceType>( (std::numeric_limits<TargetType>::max)())" , "openmp/runtime/src/kmp.h", 4545); } |
4545 | (std::numeric_limits<TargetType>::max)()))if (!(src <= static_cast<SourceType>( (std::numeric_limits <TargetType>::max)()))) { __kmp_debug_assert("src <= static_cast<SourceType>( (std::numeric_limits<TargetType>::max)())" , "openmp/runtime/src/kmp.h", 4545); }; |
4546 | return (TargetType)src; |
4547 | } |
4548 | }; |
4549 | |
4550 | template <typename T1, typename T2> |
4551 | static inline void __kmp_type_convert(T1 src, T2 *dest) { |
4552 | *dest = kmp_convert<T1, T2>::to(src); |
4553 | } |
4554 | |
4555 | #endif /* KMP_H */ |