blob: c993be4192df91df786f6a6db940716f9c461671 [file] [log] [blame]
Jim Cownie5e8470a2013-09-27 10:38:44 +00001/*
Jonathan Peytonde4749b2016-12-14 23:01:24 +00002 * kmp_runtime.cpp -- KPTS runtime support library
Jim Cownie5e8470a2013-09-27 10:38:44 +00003 */
4
5
6//===----------------------------------------------------------------------===//
7//
8// The LLVM Compiler Infrastructure
9//
10// This file is dual licensed under the MIT and the University of Illinois Open
11// Source Licenses. See LICENSE.txt for details.
12//
13//===----------------------------------------------------------------------===//
14
15
16#include "kmp.h"
Jonathan Peyton30419822017-05-12 18:01:32 +000017#include "kmp_affinity.h"
Jim Cownie5e8470a2013-09-27 10:38:44 +000018#include "kmp_atomic.h"
Jim Cownie5e8470a2013-09-27 10:38:44 +000019#include "kmp_environment.h"
Jonathan Peyton30419822017-05-12 18:01:32 +000020#include "kmp_error.h"
Jim Cownie5e8470a2013-09-27 10:38:44 +000021#include "kmp_i18n.h"
22#include "kmp_io.h"
Jonathan Peyton30419822017-05-12 18:01:32 +000023#include "kmp_itt.h"
24#include "kmp_settings.h"
Jim Cownie4cc4bb42014-10-07 16:25:50 +000025#include "kmp_stats.h"
Jonathan Peyton30419822017-05-12 18:01:32 +000026#include "kmp_str.h"
Jim Cownie4cc4bb42014-10-07 16:25:50 +000027#include "kmp_wait_release.h"
Jonathan Peyton30419822017-05-12 18:01:32 +000028#include "kmp_wrapper_getpid.h"
Jim Cownie5e8470a2013-09-27 10:38:44 +000029
Andrey Churbanovd7d088f2015-04-29 16:42:24 +000030#if OMPT_SUPPORT
31#include "ompt-specific.h"
32#endif
33
Jim Cownie5e8470a2013-09-27 10:38:44 +000034/* these are temporary issues to be dealt with */
35#define KMP_USE_PRCTL 0
Jim Cownie5e8470a2013-09-27 10:38:44 +000036
Jim Cownie5e8470a2013-09-27 10:38:44 +000037#if KMP_OS_WINDOWS
38#include <process.h>
39#endif
40
Jonas Hahnfeld50fed042016-11-07 15:58:36 +000041#include "tsan_annotations.h"
Jim Cownie5e8470a2013-09-27 10:38:44 +000042
43#if defined(KMP_GOMP_COMPAT)
Jonathan Peyton30419822017-05-12 18:01:32 +000044char const __kmp_version_alt_comp[] =
45 KMP_VERSION_PREFIX "alternative compiler support: yes";
Jim Cownie5e8470a2013-09-27 10:38:44 +000046#endif /* defined(KMP_GOMP_COMPAT) */
47
48char const __kmp_version_omp_api[] = KMP_VERSION_PREFIX "API version: "
Jonathan Peytone844a542017-03-06 22:07:40 +000049#if OMP_50_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +000050 "5.0 (201611)";
Jonathan Peytone844a542017-03-06 22:07:40 +000051#elif OMP_45_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +000052 "4.5 (201511)";
Jonathan Peyton74f3ffc2016-09-30 15:50:14 +000053#elif OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +000054 "4.0 (201307)";
Jim Cownie5e8470a2013-09-27 10:38:44 +000055#else
Jonathan Peyton30419822017-05-12 18:01:32 +000056 "3.1 (201107)";
Jim Cownie5e8470a2013-09-27 10:38:44 +000057#endif
58
59#ifdef KMP_DEBUG
Jonathan Peyton30419822017-05-12 18:01:32 +000060char const __kmp_version_lock[] =
61 KMP_VERSION_PREFIX "lock type: run time selectable";
Jim Cownie5e8470a2013-09-27 10:38:44 +000062#endif /* KMP_DEBUG */
63
Jonathan Peyton30419822017-05-12 18:01:32 +000064#define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
Jim Cownie181b4bb2013-12-23 17:28:57 +000065
Jim Cownie5e8470a2013-09-27 10:38:44 +000066/* ------------------------------------------------------------------------ */
Jim Cownie5e8470a2013-09-27 10:38:44 +000067
68kmp_info_t __kmp_monitor;
69
Jim Cownie5e8470a2013-09-27 10:38:44 +000070/* Forward declarations */
71
Jonathan Peyton30419822017-05-12 18:01:32 +000072void __kmp_cleanup(void);
Jim Cownie5e8470a2013-09-27 10:38:44 +000073
Jonathan Peyton30419822017-05-12 18:01:32 +000074static void __kmp_initialize_info(kmp_info_t *, kmp_team_t *, int tid,
75 int gtid);
76static void __kmp_initialize_team(kmp_team_t *team, int new_nproc,
77 kmp_internal_control_t *new_icvs,
78 ident_t *loc);
Jonathan Peyton2321d572015-06-08 19:25:25 +000079#if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
Jonathan Peyton30419822017-05-12 18:01:32 +000080static void __kmp_partition_places(kmp_team_t *team,
81 int update_master_only = 0);
Jonathan Peyton2321d572015-06-08 19:25:25 +000082#endif
Jonathan Peyton30419822017-05-12 18:01:32 +000083static void __kmp_do_serial_initialize(void);
84void __kmp_fork_barrier(int gtid, int tid);
85void __kmp_join_barrier(int gtid);
86void __kmp_setup_icv_copy(kmp_team_t *team, int new_nproc,
87 kmp_internal_control_t *new_icvs, ident_t *loc);
Jim Cownie5e8470a2013-09-27 10:38:44 +000088
Jim Cownie5e8470a2013-09-27 10:38:44 +000089#ifdef USE_LOAD_BALANCE
Jonathan Peyton30419822017-05-12 18:01:32 +000090static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc);
Jim Cownie5e8470a2013-09-27 10:38:44 +000091#endif
92
93static int __kmp_expand_threads(int nWish, int nNeed);
Jonathan Peyton2321d572015-06-08 19:25:25 +000094#if KMP_OS_WINDOWS
Jonathan Peyton30419822017-05-12 18:01:32 +000095static int __kmp_unregister_root_other_thread(int gtid);
Jonathan Peyton2321d572015-06-08 19:25:25 +000096#endif
Jonathan Peyton30419822017-05-12 18:01:32 +000097static void __kmp_unregister_library(void); // called by __kmp_internal_end()
98static void __kmp_reap_thread(kmp_info_t *thread, int is_root);
Jim Cownie5e8470a2013-09-27 10:38:44 +000099static kmp_info_t *__kmp_thread_pool_insert_pt = NULL;
100
Jim Cownie5e8470a2013-09-27 10:38:44 +0000101/* Calculate the identifier of the current thread */
Jonathan Peyton30419822017-05-12 18:01:32 +0000102/* fast (and somewhat portable) way to get unique identifier of executing
103 thread. Returns KMP_GTID_DNE if we haven't been assigned a gtid. */
Jonathan Peyton30419822017-05-12 18:01:32 +0000104int __kmp_get_global_thread_id() {
105 int i;
106 kmp_info_t **other_threads;
107 size_t stack_data;
108 char *stack_addr;
109 size_t stack_size;
110 char *stack_base;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000111
Jonathan Peyton30419822017-05-12 18:01:32 +0000112 KA_TRACE(
113 1000,
114 ("*** __kmp_get_global_thread_id: entering, nproc=%d all_nproc=%d\n",
115 __kmp_nth, __kmp_all_nth));
Jim Cownie5e8470a2013-09-27 10:38:44 +0000116
Jonathan Peyton30419822017-05-12 18:01:32 +0000117 /* JPH - to handle the case where __kmpc_end(0) is called immediately prior to
118 a parallel region, made it return KMP_GTID_DNE to force serial_initialize
119 by caller. Had to handle KMP_GTID_DNE at all call-sites, or else guarantee
120 __kmp_init_gtid for this to work. */
Jim Cownie5e8470a2013-09-27 10:38:44 +0000121
Jonathan Peyton30419822017-05-12 18:01:32 +0000122 if (!TCR_4(__kmp_init_gtid))
123 return KMP_GTID_DNE;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000124
125#ifdef KMP_TDATA_GTID
Jonathan Peyton30419822017-05-12 18:01:32 +0000126 if (TCR_4(__kmp_gtid_mode) >= 3) {
127 KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using TDATA\n"));
128 return __kmp_gtid;
129 }
Jim Cownie5e8470a2013-09-27 10:38:44 +0000130#endif
Jonathan Peyton30419822017-05-12 18:01:32 +0000131 if (TCR_4(__kmp_gtid_mode) >= 2) {
132 KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using keyed TLS\n"));
133 return __kmp_gtid_get_specific();
134 }
135 KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using internal alg.\n"));
136
137 stack_addr = (char *)&stack_data;
138 other_threads = __kmp_threads;
139
140 /* ATT: The code below is a source of potential bugs due to unsynchronized
141 access to __kmp_threads array. For example:
142 1. Current thread loads other_threads[i] to thr and checks it, it is
143 non-NULL.
144 2. Current thread is suspended by OS.
145 3. Another thread unregisters and finishes (debug versions of free()
146 may fill memory with something like 0xEF).
147 4. Current thread is resumed.
148 5. Current thread reads junk from *thr.
149 TODO: Fix it. --ln */
150
151 for (i = 0; i < __kmp_threads_capacity; i++) {
152
153 kmp_info_t *thr = (kmp_info_t *)TCR_SYNC_PTR(other_threads[i]);
154 if (!thr)
155 continue;
156
157 stack_size = (size_t)TCR_PTR(thr->th.th_info.ds.ds_stacksize);
158 stack_base = (char *)TCR_PTR(thr->th.th_info.ds.ds_stackbase);
159
160 /* stack grows down -- search through all of the active threads */
161
162 if (stack_addr <= stack_base) {
163 size_t stack_diff = stack_base - stack_addr;
164
165 if (stack_diff <= stack_size) {
166 /* The only way we can be closer than the allocated */
167 /* stack size is if we are running on this thread. */
168 KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == i);
169 return i;
170 }
Jim Cownie5e8470a2013-09-27 10:38:44 +0000171 }
Jonathan Peyton30419822017-05-12 18:01:32 +0000172 }
Jim Cownie5e8470a2013-09-27 10:38:44 +0000173
Jonathan Peyton30419822017-05-12 18:01:32 +0000174 /* get specific to try and determine our gtid */
175 KA_TRACE(1000,
176 ("*** __kmp_get_global_thread_id: internal alg. failed to find "
177 "thread, using TLS\n"));
178 i = __kmp_gtid_get_specific();
Jim Cownie5e8470a2013-09-27 10:38:44 +0000179
Jonathan Peyton30419822017-05-12 18:01:32 +0000180 /*fprintf( stderr, "=== %d\n", i ); */ /* GROO */
Jim Cownie5e8470a2013-09-27 10:38:44 +0000181
Jonathan Peyton30419822017-05-12 18:01:32 +0000182 /* if we havn't been assigned a gtid, then return code */
183 if (i < 0)
Jim Cownie5e8470a2013-09-27 10:38:44 +0000184 return i;
Jonathan Peyton30419822017-05-12 18:01:32 +0000185
186 /* dynamically updated stack window for uber threads to avoid get_specific
187 call */
188 if (!TCR_4(other_threads[i]->th.th_info.ds.ds_stackgrow)) {
189 KMP_FATAL(StackOverflow, i);
190 }
191
192 stack_base = (char *)other_threads[i]->th.th_info.ds.ds_stackbase;
193 if (stack_addr > stack_base) {
194 TCW_PTR(other_threads[i]->th.th_info.ds.ds_stackbase, stack_addr);
195 TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize,
196 other_threads[i]->th.th_info.ds.ds_stacksize + stack_addr -
197 stack_base);
198 } else {
199 TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize,
200 stack_base - stack_addr);
201 }
202
203 /* Reprint stack bounds for ubermaster since they have been refined */
204 if (__kmp_storage_map) {
205 char *stack_end = (char *)other_threads[i]->th.th_info.ds.ds_stackbase;
206 char *stack_beg = stack_end - other_threads[i]->th.th_info.ds.ds_stacksize;
207 __kmp_print_storage_map_gtid(i, stack_beg, stack_end,
208 other_threads[i]->th.th_info.ds.ds_stacksize,
209 "th_%d stack (refinement)", i);
210 }
211 return i;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000212}
213
Jonathan Peyton30419822017-05-12 18:01:32 +0000214int __kmp_get_global_thread_id_reg() {
215 int gtid;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000216
Jonathan Peyton30419822017-05-12 18:01:32 +0000217 if (!__kmp_init_serial) {
218 gtid = KMP_GTID_DNE;
219 } else
Jim Cownie5e8470a2013-09-27 10:38:44 +0000220#ifdef KMP_TDATA_GTID
Jonathan Peyton30419822017-05-12 18:01:32 +0000221 if (TCR_4(__kmp_gtid_mode) >= 3) {
222 KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using TDATA\n"));
223 gtid = __kmp_gtid;
224 } else
Jim Cownie5e8470a2013-09-27 10:38:44 +0000225#endif
Jonathan Peyton30419822017-05-12 18:01:32 +0000226 if (TCR_4(__kmp_gtid_mode) >= 2) {
227 KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using keyed TLS\n"));
228 gtid = __kmp_gtid_get_specific();
229 } else {
230 KA_TRACE(1000,
231 ("*** __kmp_get_global_thread_id_reg: using internal alg.\n"));
232 gtid = __kmp_get_global_thread_id();
233 }
234
235 /* we must be a new uber master sibling thread */
236 if (gtid == KMP_GTID_DNE) {
237 KA_TRACE(10,
238 ("__kmp_get_global_thread_id_reg: Encountered new root thread. "
239 "Registering a new gtid.\n"));
240 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
241 if (!__kmp_init_serial) {
242 __kmp_do_serial_initialize();
243 gtid = __kmp_gtid_get_specific();
Jim Cownie5e8470a2013-09-27 10:38:44 +0000244 } else {
Jonathan Peyton30419822017-05-12 18:01:32 +0000245 gtid = __kmp_register_root(FALSE);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000246 }
Jonathan Peyton30419822017-05-12 18:01:32 +0000247 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
248 /*__kmp_printf( "+++ %d\n", gtid ); */ /* GROO */
249 }
Jim Cownie5e8470a2013-09-27 10:38:44 +0000250
Jonathan Peyton30419822017-05-12 18:01:32 +0000251 KMP_DEBUG_ASSERT(gtid >= 0);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000252
Jonathan Peyton30419822017-05-12 18:01:32 +0000253 return gtid;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000254}
255
256/* caller must hold forkjoin_lock */
Jonathan Peyton30419822017-05-12 18:01:32 +0000257void __kmp_check_stack_overlap(kmp_info_t *th) {
258 int f;
259 char *stack_beg = NULL;
260 char *stack_end = NULL;
261 int gtid;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000262
Jonathan Peyton30419822017-05-12 18:01:32 +0000263 KA_TRACE(10, ("__kmp_check_stack_overlap: called\n"));
264 if (__kmp_storage_map) {
265 stack_end = (char *)th->th.th_info.ds.ds_stackbase;
266 stack_beg = stack_end - th->th.th_info.ds.ds_stacksize;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000267
Jonathan Peyton30419822017-05-12 18:01:32 +0000268 gtid = __kmp_gtid_from_thread(th);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000269
Jonathan Peyton30419822017-05-12 18:01:32 +0000270 if (gtid == KMP_GTID_MONITOR) {
271 __kmp_print_storage_map_gtid(
272 gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize,
273 "th_%s stack (%s)", "mon",
274 (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual");
Jim Cownie5e8470a2013-09-27 10:38:44 +0000275 } else {
Jonathan Peyton30419822017-05-12 18:01:32 +0000276 __kmp_print_storage_map_gtid(
277 gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize,
278 "th_%d stack (%s)", gtid,
279 (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual");
280 }
281 }
Jim Cownie5e8470a2013-09-27 10:38:44 +0000282
Jonathan Peyton30419822017-05-12 18:01:32 +0000283 /* No point in checking ubermaster threads since they use refinement and
284 * cannot overlap */
285 gtid = __kmp_gtid_from_thread(th);
286 if (__kmp_env_checks == TRUE && !KMP_UBER_GTID(gtid)) {
287 KA_TRACE(10,
288 ("__kmp_check_stack_overlap: performing extensive checking\n"));
289 if (stack_beg == NULL) {
290 stack_end = (char *)th->th.th_info.ds.ds_stackbase;
291 stack_beg = stack_end - th->th.th_info.ds.ds_stacksize;
292 }
293
294 for (f = 0; f < __kmp_threads_capacity; f++) {
295 kmp_info_t *f_th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[f]);
296
297 if (f_th && f_th != th) {
298 char *other_stack_end =
299 (char *)TCR_PTR(f_th->th.th_info.ds.ds_stackbase);
300 char *other_stack_beg =
301 other_stack_end - (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize);
302 if ((stack_beg > other_stack_beg && stack_beg < other_stack_end) ||
303 (stack_end > other_stack_beg && stack_end < other_stack_end)) {
304
305 /* Print the other stack values before the abort */
306 if (__kmp_storage_map)
307 __kmp_print_storage_map_gtid(
308 -1, other_stack_beg, other_stack_end,
309 (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize),
310 "th_%d stack (overlapped)", __kmp_gtid_from_thread(f_th));
311
312 __kmp_msg(kmp_ms_fatal, KMP_MSG(StackOverlap),
313 KMP_HNT(ChangeStackLimit), __kmp_msg_null);
314 }
315 }
316 }
317 }
318 KA_TRACE(10, ("__kmp_check_stack_overlap: returning\n"));
319}
320
321/* ------------------------------------------------------------------------ */
322
323void __kmp_infinite_loop(void) {
324 static int done = FALSE;
325
326 while (!done) {
327 KMP_YIELD(1);
328 }
329}
330
331#define MAX_MESSAGE 512
332
333void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2, size_t size,
334 char const *format, ...) {
335 char buffer[MAX_MESSAGE];
336 va_list ap;
337
338 va_start(ap, format);
339 KMP_SNPRINTF(buffer, sizeof(buffer), "OMP storage map: %p %p%8lu %s\n", p1,
340 p2, (unsigned long)size, format);
341 __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock);
342 __kmp_vprintf(kmp_err, buffer, ap);
343#if KMP_PRINT_DATA_PLACEMENT
344 int node;
345 if (gtid >= 0) {
346 if (p1 <= p2 && (char *)p2 - (char *)p1 == size) {
347 if (__kmp_storage_map_verbose) {
348 node = __kmp_get_host_node(p1);
349 if (node < 0) /* doesn't work, so don't try this next time */
350 __kmp_storage_map_verbose = FALSE;
351 else {
352 char *last;
353 int lastNode;
354 int localProc = __kmp_get_cpu_from_gtid(gtid);
355
356 const int page_size = KMP_GET_PAGE_SIZE();
357
358 p1 = (void *)((size_t)p1 & ~((size_t)page_size - 1));
359 p2 = (void *)(((size_t)p2 - 1) & ~((size_t)page_size - 1));
360 if (localProc >= 0)
361 __kmp_printf_no_lock(" GTID %d localNode %d\n", gtid,
362 localProc >> 1);
363 else
364 __kmp_printf_no_lock(" GTID %d\n", gtid);
365#if KMP_USE_PRCTL
366 /* The more elaborate format is disabled for now because of the prctl
367 * hanging bug. */
368 do {
369 last = p1;
370 lastNode = node;
371 /* This loop collates adjacent pages with the same host node. */
372 do {
373 (char *)p1 += page_size;
374 } while (p1 <= p2 && (node = __kmp_get_host_node(p1)) == lastNode);
375 __kmp_printf_no_lock(" %p-%p memNode %d\n", last, (char *)p1 - 1,
376 lastNode);
377 } while (p1 <= p2);
378#else
379 __kmp_printf_no_lock(" %p-%p memNode %d\n", p1,
380 (char *)p1 + (page_size - 1),
381 __kmp_get_host_node(p1));
382 if (p1 < p2) {
383 __kmp_printf_no_lock(" %p-%p memNode %d\n", p2,
384 (char *)p2 + (page_size - 1),
385 __kmp_get_host_node(p2));
386 }
387#endif
388 }
389 }
390 } else
391 __kmp_printf_no_lock(" %s\n", KMP_I18N_STR(StorageMapWarning));
392 }
393#endif /* KMP_PRINT_DATA_PLACEMENT */
394 __kmp_release_bootstrap_lock(&__kmp_stdio_lock);
395}
396
397void __kmp_warn(char const *format, ...) {
398 char buffer[MAX_MESSAGE];
399 va_list ap;
400
401 if (__kmp_generate_warnings == kmp_warnings_off) {
402 return;
403 }
404
405 va_start(ap, format);
406
407 KMP_SNPRINTF(buffer, sizeof(buffer), "OMP warning: %s\n", format);
408 __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock);
409 __kmp_vprintf(kmp_err, buffer, ap);
410 __kmp_release_bootstrap_lock(&__kmp_stdio_lock);
411
412 va_end(ap);
413}
414
415void __kmp_abort_process() {
416 // Later threads may stall here, but that's ok because abort() will kill them.
417 __kmp_acquire_bootstrap_lock(&__kmp_exit_lock);
418
419 if (__kmp_debug_buf) {
420 __kmp_dump_debug_buffer();
421 }; // if
422
423 if (KMP_OS_WINDOWS) {
424 // Let other threads know of abnormal termination and prevent deadlock
425 // if abort happened during library initialization or shutdown
426 __kmp_global.g.g_abort = SIGABRT;
427
428 /* On Windows* OS by default abort() causes pop-up error box, which stalls
429 nightly testing. Unfortunately, we cannot reliably suppress pop-up error
430 boxes. _set_abort_behavior() works well, but this function is not
431 available in VS7 (this is not problem for DLL, but it is a problem for
432 static OpenMP RTL). SetErrorMode (and so, timelimit utility) does not
433 help, at least in some versions of MS C RTL.
434
435 It seems following sequence is the only way to simulate abort() and
436 avoid pop-up error box. */
437 raise(SIGABRT);
438 _exit(3); // Just in case, if signal ignored, exit anyway.
439 } else {
440 abort();
441 }; // if
442
443 __kmp_infinite_loop();
444 __kmp_release_bootstrap_lock(&__kmp_exit_lock);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000445
446} // __kmp_abort_process
447
Jonathan Peyton30419822017-05-12 18:01:32 +0000448void __kmp_abort_thread(void) {
449 // TODO: Eliminate g_abort global variable and this function.
450 // In case of abort just call abort(), it will kill all the threads.
451 __kmp_infinite_loop();
Jim Cownie5e8470a2013-09-27 10:38:44 +0000452} // __kmp_abort_thread
453
Jonathan Peyton30419822017-05-12 18:01:32 +0000454/* Print out the storage map for the major kmp_info_t thread data structures
455 that are allocated together. */
Jim Cownie5e8470a2013-09-27 10:38:44 +0000456
Jonathan Peyton30419822017-05-12 18:01:32 +0000457static void __kmp_print_thread_storage_map(kmp_info_t *thr, int gtid) {
458 __kmp_print_storage_map_gtid(gtid, thr, thr + 1, sizeof(kmp_info_t), "th_%d",
459 gtid);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000460
Jonathan Peyton30419822017-05-12 18:01:32 +0000461 __kmp_print_storage_map_gtid(gtid, &thr->th.th_info, &thr->th.th_team,
462 sizeof(kmp_desc_t), "th_%d.th_info", gtid);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000463
Jonathan Peyton30419822017-05-12 18:01:32 +0000464 __kmp_print_storage_map_gtid(gtid, &thr->th.th_local, &thr->th.th_pri_head,
465 sizeof(kmp_local_t), "th_%d.th_local", gtid);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000466
Jonathan Peyton30419822017-05-12 18:01:32 +0000467 __kmp_print_storage_map_gtid(
468 gtid, &thr->th.th_bar[0], &thr->th.th_bar[bs_last_barrier],
469 sizeof(kmp_balign_t) * bs_last_barrier, "th_%d.th_bar", gtid);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000470
Jonathan Peyton30419822017-05-12 18:01:32 +0000471 __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_plain_barrier],
472 &thr->th.th_bar[bs_plain_barrier + 1],
473 sizeof(kmp_balign_t), "th_%d.th_bar[plain]",
474 gtid);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000475
Jonathan Peyton30419822017-05-12 18:01:32 +0000476 __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_forkjoin_barrier],
477 &thr->th.th_bar[bs_forkjoin_barrier + 1],
478 sizeof(kmp_balign_t), "th_%d.th_bar[forkjoin]",
479 gtid);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000480
Jonathan Peyton30419822017-05-12 18:01:32 +0000481#if KMP_FAST_REDUCTION_BARRIER
482 __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_reduction_barrier],
483 &thr->th.th_bar[bs_reduction_barrier + 1],
484 sizeof(kmp_balign_t), "th_%d.th_bar[reduction]",
485 gtid);
486#endif // KMP_FAST_REDUCTION_BARRIER
Jim Cownie5e8470a2013-09-27 10:38:44 +0000487}
488
Jonathan Peyton30419822017-05-12 18:01:32 +0000489/* Print out the storage map for the major kmp_team_t team data structures
490 that are allocated together. */
Jim Cownie5e8470a2013-09-27 10:38:44 +0000491
Jonathan Peyton30419822017-05-12 18:01:32 +0000492static void __kmp_print_team_storage_map(const char *header, kmp_team_t *team,
493 int team_id, int num_thr) {
494 int num_disp_buff = team->t.t_max_nproc > 1 ? __kmp_dispatch_num_buffers : 2;
495 __kmp_print_storage_map_gtid(-1, team, team + 1, sizeof(kmp_team_t), "%s_%d",
496 header, team_id);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000497
Jonathan Peyton30419822017-05-12 18:01:32 +0000498 __kmp_print_storage_map_gtid(-1, &team->t.t_bar[0],
499 &team->t.t_bar[bs_last_barrier],
500 sizeof(kmp_balign_team_t) * bs_last_barrier,
501 "%s_%d.t_bar", header, team_id);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000502
Jonathan Peyton30419822017-05-12 18:01:32 +0000503 __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_plain_barrier],
504 &team->t.t_bar[bs_plain_barrier + 1],
505 sizeof(kmp_balign_team_t), "%s_%d.t_bar[plain]",
506 header, team_id);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000507
Jonathan Peyton30419822017-05-12 18:01:32 +0000508 __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_forkjoin_barrier],
509 &team->t.t_bar[bs_forkjoin_barrier + 1],
510 sizeof(kmp_balign_team_t),
511 "%s_%d.t_bar[forkjoin]", header, team_id);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000512
Jonathan Peyton30419822017-05-12 18:01:32 +0000513#if KMP_FAST_REDUCTION_BARRIER
514 __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_reduction_barrier],
515 &team->t.t_bar[bs_reduction_barrier + 1],
516 sizeof(kmp_balign_team_t),
517 "%s_%d.t_bar[reduction]", header, team_id);
518#endif // KMP_FAST_REDUCTION_BARRIER
Jim Cownie5e8470a2013-09-27 10:38:44 +0000519
Jonathan Peyton30419822017-05-12 18:01:32 +0000520 __kmp_print_storage_map_gtid(
521 -1, &team->t.t_dispatch[0], &team->t.t_dispatch[num_thr],
522 sizeof(kmp_disp_t) * num_thr, "%s_%d.t_dispatch", header, team_id);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000523
Jonathan Peyton30419822017-05-12 18:01:32 +0000524 __kmp_print_storage_map_gtid(
525 -1, &team->t.t_threads[0], &team->t.t_threads[num_thr],
526 sizeof(kmp_info_t *) * num_thr, "%s_%d.t_threads", header, team_id);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000527
Jonathan Peyton30419822017-05-12 18:01:32 +0000528 __kmp_print_storage_map_gtid(-1, &team->t.t_disp_buffer[0],
529 &team->t.t_disp_buffer[num_disp_buff],
530 sizeof(dispatch_shared_info_t) * num_disp_buff,
531 "%s_%d.t_disp_buffer", header, team_id);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000532
Jonathan Peyton30419822017-05-12 18:01:32 +0000533 __kmp_print_storage_map_gtid(-1, &team->t.t_taskq, &team->t.t_copypriv_data,
534 sizeof(kmp_taskq_t), "%s_%d.t_taskq", header,
535 team_id);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000536}
537
538static void __kmp_init_allocator() {}
539static void __kmp_fini_allocator() {}
Jim Cownie5e8470a2013-09-27 10:38:44 +0000540
541/* ------------------------------------------------------------------------ */
542
Jonathan Peyton99016992015-05-26 17:32:53 +0000543#ifdef KMP_DYNAMIC_LIB
Jonathan Peyton30419822017-05-12 18:01:32 +0000544#if KMP_OS_WINDOWS
Jim Cownie5e8470a2013-09-27 10:38:44 +0000545
Jonathan Peyton30419822017-05-12 18:01:32 +0000546static void __kmp_reset_lock(kmp_bootstrap_lock_t *lck) {
547 // TODO: Change to __kmp_break_bootstrap_lock().
548 __kmp_init_bootstrap_lock(lck); // make the lock released
Jim Cownie5e8470a2013-09-27 10:38:44 +0000549}
550
Jonathan Peyton30419822017-05-12 18:01:32 +0000551static void __kmp_reset_locks_on_process_detach(int gtid_req) {
552 int i;
553 int thread_count;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000554
Jonathan Peyton30419822017-05-12 18:01:32 +0000555 // PROCESS_DETACH is expected to be called by a thread that executes
556 // ProcessExit() or FreeLibrary(). OS terminates other threads (except the one
557 // calling ProcessExit or FreeLibrary). So, it might be safe to access the
558 // __kmp_threads[] without taking the forkjoin_lock. However, in fact, some
559 // threads can be still alive here, although being about to be terminated. The
560 // threads in the array with ds_thread==0 are most suspicious. Actually, it
561 // can be not safe to access the __kmp_threads[].
Jim Cownie5e8470a2013-09-27 10:38:44 +0000562
Jonathan Peyton30419822017-05-12 18:01:32 +0000563 // TODO: does it make sense to check __kmp_roots[] ?
Jim Cownie5e8470a2013-09-27 10:38:44 +0000564
Jonathan Peyton30419822017-05-12 18:01:32 +0000565 // Let's check that there are no other alive threads registered with the OMP
566 // lib.
567 while (1) {
568 thread_count = 0;
569 for (i = 0; i < __kmp_threads_capacity; ++i) {
570 if (!__kmp_threads)
571 continue;
572 kmp_info_t *th = __kmp_threads[i];
573 if (th == NULL)
574 continue;
575 int gtid = th->th.th_info.ds.ds_gtid;
576 if (gtid == gtid_req)
577 continue;
578 if (gtid < 0)
579 continue;
580 DWORD exit_val;
581 int alive = __kmp_is_thread_alive(th, &exit_val);
582 if (alive) {
583 ++thread_count;
584 }
Jim Cownie5e8470a2013-09-27 10:38:44 +0000585 }
Jonathan Peyton30419822017-05-12 18:01:32 +0000586 if (thread_count == 0)
587 break; // success
588 }
Jim Cownie5e8470a2013-09-27 10:38:44 +0000589
Jonathan Peyton30419822017-05-12 18:01:32 +0000590 // Assume that I'm alone. Now it might be safe to check and reset locks.
591 // __kmp_forkjoin_lock and __kmp_stdio_lock are expected to be reset.
592 __kmp_reset_lock(&__kmp_forkjoin_lock);
593#ifdef KMP_DEBUG
594 __kmp_reset_lock(&__kmp_stdio_lock);
595#endif // KMP_DEBUG
Jim Cownie5e8470a2013-09-27 10:38:44 +0000596}
597
Jonathan Peyton30419822017-05-12 18:01:32 +0000598BOOL WINAPI DllMain(HINSTANCE hInstDLL, DWORD fdwReason, LPVOID lpReserved) {
599 //__kmp_acquire_bootstrap_lock( &__kmp_initz_lock );
Jim Cownie5e8470a2013-09-27 10:38:44 +0000600
Jonathan Peyton30419822017-05-12 18:01:32 +0000601 switch (fdwReason) {
Jim Cownie5e8470a2013-09-27 10:38:44 +0000602
Jonathan Peyton30419822017-05-12 18:01:32 +0000603 case DLL_PROCESS_ATTACH:
604 KA_TRACE(10, ("DllMain: PROCESS_ATTACH\n"));
Jim Cownie5e8470a2013-09-27 10:38:44 +0000605
606 return TRUE;
Jonathan Peyton30419822017-05-12 18:01:32 +0000607
608 case DLL_PROCESS_DETACH:
609 KA_TRACE(10, ("DllMain: PROCESS_DETACH T#%d\n", __kmp_gtid_get_specific()));
610
611 if (lpReserved != NULL) {
612 // lpReserved is used for telling the difference:
613 // lpReserved == NULL when FreeLibrary() was called,
614 // lpReserved != NULL when the process terminates.
615 // When FreeLibrary() is called, worker threads remain alive. So they will
616 // release the forkjoin lock by themselves. When the process terminates,
617 // worker threads disappear triggering the problem of unreleased forkjoin
618 // lock as described below.
619
620 // A worker thread can take the forkjoin lock. The problem comes up if
621 // that worker thread becomes dead before it releases the forkjoin lock.
622 // The forkjoin lock remains taken, while the thread executing
623 // DllMain()->PROCESS_DETACH->__kmp_internal_end_library() below will try
624 // to take the forkjoin lock and will always fail, so that the application
625 // will never finish [normally]. This scenario is possible if
626 // __kmpc_end() has not been executed. It looks like it's not a corner
627 // case, but common cases:
628 // - the main function was compiled by an alternative compiler;
629 // - the main function was compiled by icl but without /Qopenmp
630 // (application with plugins);
631 // - application terminates by calling C exit(), Fortran CALL EXIT() or
632 // Fortran STOP.
633 // - alive foreign thread prevented __kmpc_end from doing cleanup.
634 //
635 // This is a hack to work around the problem.
636 // TODO: !!! figure out something better.
637 __kmp_reset_locks_on_process_detach(__kmp_gtid_get_specific());
638 }
639
640 __kmp_internal_end_library(__kmp_gtid_get_specific());
641
642 return TRUE;
643
644 case DLL_THREAD_ATTACH:
645 KA_TRACE(10, ("DllMain: THREAD_ATTACH\n"));
646
647 /* if we want to register new siblings all the time here call
648 * __kmp_get_gtid(); */
649 return TRUE;
650
651 case DLL_THREAD_DETACH:
652 KA_TRACE(10, ("DllMain: THREAD_DETACH T#%d\n", __kmp_gtid_get_specific()));
653
654 __kmp_internal_end_thread(__kmp_gtid_get_specific());
655 return TRUE;
656 }
657
658 return TRUE;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000659}
660
Jonathan Peyton30419822017-05-12 18:01:32 +0000661#endif /* KMP_OS_WINDOWS */
Jonathan Peyton99016992015-05-26 17:32:53 +0000662#endif /* KMP_DYNAMIC_LIB */
Jim Cownie5e8470a2013-09-27 10:38:44 +0000663
Jim Cownie5e8470a2013-09-27 10:38:44 +0000664/* Change the library type to "status" and return the old type */
665/* called from within initialization routines where __kmp_initz_lock is held */
Jonathan Peyton30419822017-05-12 18:01:32 +0000666int __kmp_change_library(int status) {
667 int old_status;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000668
Jonathan Peyton30419822017-05-12 18:01:32 +0000669 old_status = __kmp_yield_init &
670 1; // check whether KMP_LIBRARY=throughput (even init count)
Jim Cownie5e8470a2013-09-27 10:38:44 +0000671
Jonathan Peyton30419822017-05-12 18:01:32 +0000672 if (status) {
673 __kmp_yield_init |= 1; // throughput => turnaround (odd init count)
674 } else {
675 __kmp_yield_init &= ~1; // turnaround => throughput (even init count)
676 }
Jim Cownie5e8470a2013-09-27 10:38:44 +0000677
Jonathan Peyton30419822017-05-12 18:01:32 +0000678 return old_status; // return previous setting of whether
679 // KMP_LIBRARY=throughput
Jim Cownie5e8470a2013-09-27 10:38:44 +0000680}
681
Jonathan Peyton30419822017-05-12 18:01:32 +0000682/* __kmp_parallel_deo -- Wait until it's our turn. */
683void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) {
684 int gtid = *gtid_ref;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000685#ifdef BUILD_PARALLEL_ORDERED
Jonathan Peyton30419822017-05-12 18:01:32 +0000686 kmp_team_t *team = __kmp_team_from_gtid(gtid);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000687#endif /* BUILD_PARALLEL_ORDERED */
688
Jonathan Peyton30419822017-05-12 18:01:32 +0000689 if (__kmp_env_consistency_check) {
690 if (__kmp_threads[gtid]->th.th_root->r.r_active)
Andrey Churbanov5c56fb52015-02-20 18:05:17 +0000691#if KMP_USE_DYNAMIC_LOCK
Jonathan Peyton30419822017-05-12 18:01:32 +0000692 __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL, 0);
Andrey Churbanov5c56fb52015-02-20 18:05:17 +0000693#else
Jonathan Peyton30419822017-05-12 18:01:32 +0000694 __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL);
Andrey Churbanov5c56fb52015-02-20 18:05:17 +0000695#endif
Jonathan Peyton30419822017-05-12 18:01:32 +0000696 }
Jim Cownie5e8470a2013-09-27 10:38:44 +0000697#ifdef BUILD_PARALLEL_ORDERED
Jonathan Peyton30419822017-05-12 18:01:32 +0000698 if (!team->t.t_serialized) {
699 KMP_MB();
700 KMP_WAIT_YIELD(&team->t.t_ordered.dt.t_value, __kmp_tid_from_gtid(gtid),
701 KMP_EQ, NULL);
702 KMP_MB();
703 }
Jim Cownie5e8470a2013-09-27 10:38:44 +0000704#endif /* BUILD_PARALLEL_ORDERED */
705}
706
Jonathan Peyton30419822017-05-12 18:01:32 +0000707/* __kmp_parallel_dxo -- Signal the next task. */
708void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) {
709 int gtid = *gtid_ref;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000710#ifdef BUILD_PARALLEL_ORDERED
Jonathan Peyton30419822017-05-12 18:01:32 +0000711 int tid = __kmp_tid_from_gtid(gtid);
712 kmp_team_t *team = __kmp_team_from_gtid(gtid);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000713#endif /* BUILD_PARALLEL_ORDERED */
714
Jonathan Peyton30419822017-05-12 18:01:32 +0000715 if (__kmp_env_consistency_check) {
716 if (__kmp_threads[gtid]->th.th_root->r.r_active)
717 __kmp_pop_sync(gtid, ct_ordered_in_parallel, loc_ref);
718 }
Jim Cownie5e8470a2013-09-27 10:38:44 +0000719#ifdef BUILD_PARALLEL_ORDERED
Jonathan Peyton30419822017-05-12 18:01:32 +0000720 if (!team->t.t_serialized) {
721 KMP_MB(); /* Flush all pending memory write invalidates. */
Jim Cownie5e8470a2013-09-27 10:38:44 +0000722
Jonathan Peyton30419822017-05-12 18:01:32 +0000723 /* use the tid of the next thread in this team */
724 /* TODO replace with general release procedure */
725 team->t.t_ordered.dt.t_value = ((tid + 1) % team->t.t_nproc);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000726
Andrey Churbanovd7d088f2015-04-29 16:42:24 +0000727#if OMPT_SUPPORT && OMPT_BLAME
Jonathan Peyton30419822017-05-12 18:01:32 +0000728 if (ompt_enabled &&
729 ompt_callbacks.ompt_callback(ompt_event_release_ordered)) {
730 /* accept blame for "ordered" waiting */
731 kmp_info_t *this_thread = __kmp_threads[gtid];
732 ompt_callbacks.ompt_callback(ompt_event_release_ordered)(
733 this_thread->th.ompt_thread_info.wait_id);
734 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +0000735#endif
736
Jonathan Peyton30419822017-05-12 18:01:32 +0000737 KMP_MB(); /* Flush all pending memory write invalidates. */
738 }
Jim Cownie5e8470a2013-09-27 10:38:44 +0000739#endif /* BUILD_PARALLEL_ORDERED */
740}
741
742/* ------------------------------------------------------------------------ */
Jim Cownie5e8470a2013-09-27 10:38:44 +0000743/* The BARRIER for a SINGLE process section is always explicit */
744
Jonathan Peyton30419822017-05-12 18:01:32 +0000745int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws) {
746 int status;
747 kmp_info_t *th;
748 kmp_team_t *team;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000749
Jonathan Peyton30419822017-05-12 18:01:32 +0000750 if (!TCR_4(__kmp_init_parallel))
751 __kmp_parallel_initialize();
Jim Cownie5e8470a2013-09-27 10:38:44 +0000752
Jonathan Peyton30419822017-05-12 18:01:32 +0000753 th = __kmp_threads[gtid];
754 team = th->th.th_team;
755 status = 0;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000756
Jonathan Peyton30419822017-05-12 18:01:32 +0000757 th->th.th_ident = id_ref;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000758
Jonathan Peyton30419822017-05-12 18:01:32 +0000759 if (team->t.t_serialized) {
760 status = 1;
761 } else {
762 kmp_int32 old_this = th->th.th_local.this_construct;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000763
Jonathan Peyton30419822017-05-12 18:01:32 +0000764 ++th->th.th_local.this_construct;
765 /* try to set team count to thread count--success means thread got the
766 single block */
767 /* TODO: Should this be acquire or release? */
768 if (team->t.t_construct == old_this) {
769 status = KMP_COMPARE_AND_STORE_ACQ32(&team->t.t_construct, old_this,
770 th->th.th_local.this_construct);
771 }
Andrey Churbanov51aecb82015-05-06 19:22:36 +0000772#if USE_ITT_BUILD
Jonathan Peyton30419822017-05-12 18:01:32 +0000773 if (__itt_metadata_add_ptr && __kmp_forkjoin_frames_mode == 3 &&
774 KMP_MASTER_GTID(gtid) &&
Andrey Churbanov51aecb82015-05-06 19:22:36 +0000775#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +0000776 th->th.th_teams_microtask == NULL &&
Andrey Churbanov51aecb82015-05-06 19:22:36 +0000777#endif
Jonathan Peyton30419822017-05-12 18:01:32 +0000778 team->t.t_active_level ==
779 1) { // Only report metadata by master of active team at level 1
780 __kmp_itt_metadata_single(id_ref);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000781 }
Jonathan Peyton30419822017-05-12 18:01:32 +0000782#endif /* USE_ITT_BUILD */
783 }
Jim Cownie5e8470a2013-09-27 10:38:44 +0000784
Jonathan Peyton30419822017-05-12 18:01:32 +0000785 if (__kmp_env_consistency_check) {
786 if (status && push_ws) {
787 __kmp_push_workshare(gtid, ct_psingle, id_ref);
788 } else {
789 __kmp_check_workshare(gtid, ct_psingle, id_ref);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000790 }
Jonathan Peyton30419822017-05-12 18:01:32 +0000791 }
Jim Cownie5e8470a2013-09-27 10:38:44 +0000792#if USE_ITT_BUILD
Jonathan Peyton30419822017-05-12 18:01:32 +0000793 if (status) {
794 __kmp_itt_single_start(gtid);
795 }
Jim Cownie5e8470a2013-09-27 10:38:44 +0000796#endif /* USE_ITT_BUILD */
Jonathan Peyton30419822017-05-12 18:01:32 +0000797 return status;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000798}
799
Jonathan Peyton30419822017-05-12 18:01:32 +0000800void __kmp_exit_single(int gtid) {
Jim Cownie5e8470a2013-09-27 10:38:44 +0000801#if USE_ITT_BUILD
Jonathan Peyton30419822017-05-12 18:01:32 +0000802 __kmp_itt_single_end(gtid);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000803#endif /* USE_ITT_BUILD */
Jonathan Peyton30419822017-05-12 18:01:32 +0000804 if (__kmp_env_consistency_check)
805 __kmp_pop_workshare(gtid, ct_psingle, NULL);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000806}
807
Jonathan Peyton30419822017-05-12 18:01:32 +0000808/* determine if we can go parallel or must use a serialized parallel region and
Jim Cownie5e8470a2013-09-27 10:38:44 +0000809 * how many threads we can use
810 * set_nproc is the number of threads requested for the team
811 * returns 0 if we should serialize or only use one thread,
812 * otherwise the number of threads to use
Jonathan Peyton30419822017-05-12 18:01:32 +0000813 * The forkjoin lock is held by the caller. */
814static int __kmp_reserve_threads(kmp_root_t *root, kmp_team_t *parent_team,
815 int master_tid, int set_nthreads
Jim Cownie5e8470a2013-09-27 10:38:44 +0000816#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +0000817 ,
818 int enter_teams
Jim Cownie5e8470a2013-09-27 10:38:44 +0000819#endif /* OMP_40_ENABLED */
Jonathan Peyton30419822017-05-12 18:01:32 +0000820 ) {
821 int capacity;
822 int new_nthreads;
823 KMP_DEBUG_ASSERT(__kmp_init_serial);
824 KMP_DEBUG_ASSERT(root && parent_team);
Jim Cownie5e8470a2013-09-27 10:38:44 +0000825
Jonathan Peyton30419822017-05-12 18:01:32 +0000826 // If dyn-var is set, dynamically adjust the number of desired threads,
827 // according to the method specified by dynamic_mode.
828 new_nthreads = set_nthreads;
829 if (!get__dynamic_2(parent_team, master_tid)) {
830 ;
831 }
Jim Cownie5e8470a2013-09-27 10:38:44 +0000832#ifdef USE_LOAD_BALANCE
Jonathan Peyton30419822017-05-12 18:01:32 +0000833 else if (__kmp_global.g.g_dynamic_mode == dynamic_load_balance) {
834 new_nthreads = __kmp_load_balance_nproc(root, set_nthreads);
835 if (new_nthreads == 1) {
836 KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced "
837 "reservation to 1 thread\n",
838 master_tid));
839 return 1;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000840 }
Jonathan Peyton30419822017-05-12 18:01:32 +0000841 if (new_nthreads < set_nthreads) {
842 KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced "
843 "reservation to %d threads\n",
844 master_tid, new_nthreads));
845 }
846 }
Jim Cownie5e8470a2013-09-27 10:38:44 +0000847#endif /* USE_LOAD_BALANCE */
Jonathan Peyton30419822017-05-12 18:01:32 +0000848 else if (__kmp_global.g.g_dynamic_mode == dynamic_thread_limit) {
849 new_nthreads = __kmp_avail_proc - __kmp_nth +
850 (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
851 if (new_nthreads <= 1) {
852 KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced "
853 "reservation to 1 thread\n",
854 master_tid));
855 return 1;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000856 }
Jonathan Peyton30419822017-05-12 18:01:32 +0000857 if (new_nthreads < set_nthreads) {
858 KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced "
859 "reservation to %d threads\n",
860 master_tid, new_nthreads));
861 } else {
862 new_nthreads = set_nthreads;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000863 }
Jonathan Peyton30419822017-05-12 18:01:32 +0000864 } else if (__kmp_global.g.g_dynamic_mode == dynamic_random) {
865 if (set_nthreads > 2) {
866 new_nthreads = __kmp_get_random(parent_team->t.t_threads[master_tid]);
867 new_nthreads = (new_nthreads % set_nthreads) + 1;
868 if (new_nthreads == 1) {
869 KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced "
870 "reservation to 1 thread\n",
871 master_tid));
Jim Cownie5e8470a2013-09-27 10:38:44 +0000872 return 1;
Jonathan Peyton30419822017-05-12 18:01:32 +0000873 }
874 if (new_nthreads < set_nthreads) {
875 KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced "
876 "reservation to %d threads\n",
877 master_tid, new_nthreads));
878 }
879 }
880 } else {
881 KMP_ASSERT(0);
882 }
883
Jonathan Peytonf4392462017-07-27 20:58:41 +0000884 // Respect KMP_ALL_THREADS/KMP_DEVICE_THREAD_LIMIT.
Jonathan Peyton30419822017-05-12 18:01:32 +0000885 if (__kmp_nth + new_nthreads -
886 (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
887 __kmp_max_nth) {
888 int tl_nthreads = __kmp_max_nth - __kmp_nth +
889 (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
890 if (tl_nthreads <= 0) {
891 tl_nthreads = 1;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000892 }
893
Jonathan Peyton30419822017-05-12 18:01:32 +0000894 // If dyn-var is false, emit a 1-time warning.
895 if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
896 __kmp_reserve_warn = 1;
897 __kmp_msg(kmp_ms_warning,
898 KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads),
899 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
900 }
901 if (tl_nthreads == 1) {
Jonathan Peytonf4392462017-07-27 20:58:41 +0000902 KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT "
903 "reduced reservation to 1 thread\n",
Jonathan Peyton30419822017-05-12 18:01:32 +0000904 master_tid));
905 return 1;
906 }
Jonathan Peytonf4392462017-07-27 20:58:41 +0000907 KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT reduced "
908 "reservation to %d threads\n",
909 master_tid, tl_nthreads));
910 new_nthreads = tl_nthreads;
911 }
912
913 // Respect OMP_THREAD_LIMIT
914 if (root->r.r_cg_nthreads + new_nthreads -
915 (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
916 __kmp_cg_max_nth) {
917 int tl_nthreads = __kmp_cg_max_nth - root->r.r_cg_nthreads +
918 (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
919 if (tl_nthreads <= 0) {
920 tl_nthreads = 1;
921 }
922
923 // If dyn-var is false, emit a 1-time warning.
924 if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
925 __kmp_reserve_warn = 1;
926 __kmp_msg(kmp_ms_warning,
927 KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads),
928 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
929 }
930 if (tl_nthreads == 1) {
931 KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT "
932 "reduced reservation to 1 thread\n",
933 master_tid));
934 return 1;
935 }
936 KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT reduced "
Jonathan Peyton30419822017-05-12 18:01:32 +0000937 "reservation to %d threads\n",
938 master_tid, tl_nthreads));
939 new_nthreads = tl_nthreads;
940 }
941
942 // Check if the threads array is large enough, or needs expanding.
Jonathan Peyton30419822017-05-12 18:01:32 +0000943 // See comment in __kmp_register_root() about the adjustment if
944 // __kmp_threads[0] == NULL.
945 capacity = __kmp_threads_capacity;
946 if (TCR_PTR(__kmp_threads[0]) == NULL) {
947 --capacity;
948 }
949 if (__kmp_nth + new_nthreads -
950 (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
951 capacity) {
952 // Expand the threads array.
953 int slotsRequired = __kmp_nth + new_nthreads -
954 (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) -
955 capacity;
956 int slotsAdded = __kmp_expand_threads(slotsRequired, slotsRequired);
957 if (slotsAdded < slotsRequired) {
958 // The threads array was not expanded enough.
959 new_nthreads -= (slotsRequired - slotsAdded);
960 KMP_ASSERT(new_nthreads >= 1);
961
962 // If dyn-var is false, emit a 1-time warning.
963 if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
964 __kmp_reserve_warn = 1;
965 if (__kmp_tp_cached) {
966 __kmp_msg(kmp_ms_warning,
967 KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads),
968 KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity),
969 KMP_HNT(PossibleSystemLimitOnThreads), __kmp_msg_null);
970 } else {
971 __kmp_msg(kmp_ms_warning,
972 KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads),
973 KMP_HNT(SystemLimitOnThreads), __kmp_msg_null);
974 }
975 }
976 }
977 }
978
Jonathan Peyton642688b2017-06-01 16:46:36 +0000979#ifdef KMP_DEBUG
Jonathan Peyton30419822017-05-12 18:01:32 +0000980 if (new_nthreads == 1) {
981 KC_TRACE(10,
982 ("__kmp_reserve_threads: T#%d serializing team after reclaiming "
983 "dead roots and rechecking; requested %d threads\n",
984 __kmp_get_gtid(), set_nthreads));
Jonathan Peyton642688b2017-06-01 16:46:36 +0000985 } else {
986 KC_TRACE(10, ("__kmp_reserve_threads: T#%d allocating %d threads; requested"
987 " %d threads\n",
988 __kmp_get_gtid(), new_nthreads, set_nthreads));
Jonathan Peyton30419822017-05-12 18:01:32 +0000989 }
Jonathan Peyton642688b2017-06-01 16:46:36 +0000990#endif // KMP_DEBUG
Jonathan Peyton30419822017-05-12 18:01:32 +0000991 return new_nthreads;
Jim Cownie5e8470a2013-09-27 10:38:44 +0000992}
993
Jonathan Peyton30419822017-05-12 18:01:32 +0000994/* Allocate threads from the thread pool and assign them to the new team. We are
995 assured that there are enough threads available, because we checked on that
996 earlier within critical section forkjoin */
997static void __kmp_fork_team_threads(kmp_root_t *root, kmp_team_t *team,
998 kmp_info_t *master_th, int master_gtid) {
999 int i;
1000 int use_hot_team;
Jim Cownie5e8470a2013-09-27 10:38:44 +00001001
Jonathan Peyton30419822017-05-12 18:01:32 +00001002 KA_TRACE(10, ("__kmp_fork_team_threads: new_nprocs = %d\n", team->t.t_nproc));
1003 KMP_DEBUG_ASSERT(master_gtid == __kmp_get_gtid());
1004 KMP_MB();
Jim Cownie5e8470a2013-09-27 10:38:44 +00001005
Jonathan Peyton30419822017-05-12 18:01:32 +00001006 /* first, let's setup the master thread */
1007 master_th->th.th_info.ds.ds_tid = 0;
1008 master_th->th.th_team = team;
1009 master_th->th.th_team_nproc = team->t.t_nproc;
1010 master_th->th.th_team_master = master_th;
1011 master_th->th.th_team_serialized = FALSE;
1012 master_th->th.th_dispatch = &team->t.t_dispatch[0];
Jim Cownie5e8470a2013-09-27 10:38:44 +00001013
Jonathan Peyton30419822017-05-12 18:01:32 +00001014/* make sure we are not the optimized hot team */
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001015#if KMP_NESTED_HOT_TEAMS
Jonathan Peyton30419822017-05-12 18:01:32 +00001016 use_hot_team = 0;
1017 kmp_hot_team_ptr_t *hot_teams = master_th->th.th_hot_teams;
1018 if (hot_teams) { // hot teams array is not allocated if
1019 // KMP_HOT_TEAMS_MAX_LEVEL=0
1020 int level = team->t.t_active_level - 1; // index in array of hot teams
1021 if (master_th->th.th_teams_microtask) { // are we inside the teams?
1022 if (master_th->th.th_teams_size.nteams > 1) {
1023 ++level; // level was not increased in teams construct for
1024 // team_of_masters
1025 }
1026 if (team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
1027 master_th->th.th_teams_level == team->t.t_level) {
1028 ++level; // level was not increased in teams construct for
1029 // team_of_workers before the parallel
1030 } // team->t.t_level will be increased inside parallel
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001031 }
Jonathan Peyton30419822017-05-12 18:01:32 +00001032 if (level < __kmp_hot_teams_max_level) {
1033 if (hot_teams[level].hot_team) {
1034 // hot team has already been allocated for given level
1035 KMP_DEBUG_ASSERT(hot_teams[level].hot_team == team);
1036 use_hot_team = 1; // the team is ready to use
1037 } else {
1038 use_hot_team = 0; // AC: threads are not allocated yet
1039 hot_teams[level].hot_team = team; // remember new hot team
1040 hot_teams[level].hot_team_nth = team->t.t_nproc;
1041 }
1042 } else {
1043 use_hot_team = 0;
1044 }
1045 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001046#else
Jonathan Peyton30419822017-05-12 18:01:32 +00001047 use_hot_team = team == root->r.r_hot_team;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001048#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00001049 if (!use_hot_team) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00001050
Jonathan Peyton30419822017-05-12 18:01:32 +00001051 /* install the master thread */
1052 team->t.t_threads[0] = master_th;
1053 __kmp_initialize_info(master_th, team, 0, master_gtid);
Jim Cownie5e8470a2013-09-27 10:38:44 +00001054
Jonathan Peyton30419822017-05-12 18:01:32 +00001055 /* now, install the worker threads */
1056 for (i = 1; i < team->t.t_nproc; i++) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00001057
Jonathan Peyton30419822017-05-12 18:01:32 +00001058 /* fork or reallocate a new thread and install it in team */
1059 kmp_info_t *thr = __kmp_allocate_thread(root, team, i);
1060 team->t.t_threads[i] = thr;
1061 KMP_DEBUG_ASSERT(thr);
1062 KMP_DEBUG_ASSERT(thr->th.th_team == team);
1063 /* align team and thread arrived states */
1064 KA_TRACE(20, ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived "
1065 "T#%d(%d:%d) join =%llu, plain=%llu\n",
1066 __kmp_gtid_from_tid(0, team), team->t.t_id, 0,
1067 __kmp_gtid_from_tid(i, team), team->t.t_id, i,
1068 team->t.t_bar[bs_forkjoin_barrier].b_arrived,
1069 team->t.t_bar[bs_plain_barrier].b_arrived));
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001070#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00001071 thr->th.th_teams_microtask = master_th->th.th_teams_microtask;
1072 thr->th.th_teams_level = master_th->th.th_teams_level;
1073 thr->th.th_teams_size = master_th->th.th_teams_size;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001074#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00001075 { // Initialize threads' barrier data.
1076 int b;
1077 kmp_balign_t *balign = team->t.t_threads[i]->th.th_bar;
1078 for (b = 0; b < bs_last_barrier; ++b) {
1079 balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
1080 KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
Jonathan Peyton8fbb49a2015-07-09 18:16:58 +00001081#if USE_DEBUGGER
Jonathan Peyton30419822017-05-12 18:01:32 +00001082 balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
Jonathan Peyton8fbb49a2015-07-09 18:16:58 +00001083#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00001084 }; // for b
1085 }
1086 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00001087
Alp Toker98758b02014-03-02 04:12:06 +00001088#if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
Jonathan Peyton30419822017-05-12 18:01:32 +00001089 __kmp_partition_places(team);
Jim Cownie5e8470a2013-09-27 10:38:44 +00001090#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00001091 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00001092
Jonathan Peyton30419822017-05-12 18:01:32 +00001093 KMP_MB();
Jim Cownie5e8470a2013-09-27 10:38:44 +00001094}
1095
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001096#if KMP_ARCH_X86 || KMP_ARCH_X86_64
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001097// Propagate any changes to the floating point control registers out to the team
Jonathan Peyton30419822017-05-12 18:01:32 +00001098// We try to avoid unnecessary writes to the relevant cache line in the team
1099// structure, so we don't make changes unless they are needed.
1100inline static void propagateFPControl(kmp_team_t *team) {
1101 if (__kmp_inherit_fp_control) {
1102 kmp_int16 x87_fpu_control_word;
1103 kmp_uint32 mxcsr;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001104
Jonathan Peyton30419822017-05-12 18:01:32 +00001105 // Get master values of FPU control flags (both X87 and vector)
1106 __kmp_store_x87_fpu_control_word(&x87_fpu_control_word);
1107 __kmp_store_mxcsr(&mxcsr);
1108 mxcsr &= KMP_X86_MXCSR_MASK;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001109
Jonathan Peyton30419822017-05-12 18:01:32 +00001110// There is no point looking at t_fp_control_saved here.
1111// If it is TRUE, we still have to update the values if they are different from
1112// those we now have.
1113// If it is FALSE we didn't save anything yet, but our objective is the same. We
1114// have to ensure that the values in the team are the same as those we have.
1115// So, this code achieves what we need whether or not t_fp_control_saved is
1116// true. By checking whether the value needs updating we avoid unnecessary
1117// writes that would put the cache-line into a written state, causing all
1118// threads in the team to have to read it again.
1119 KMP_CHECK_UPDATE(team->t.t_x87_fpu_control_word, x87_fpu_control_word);
1120 KMP_CHECK_UPDATE(team->t.t_mxcsr, mxcsr);
1121 // Although we don't use this value, other code in the runtime wants to know
1122 // whether it should restore them. So we must ensure it is correct.
1123 KMP_CHECK_UPDATE(team->t.t_fp_control_saved, TRUE);
1124 } else {
1125 // Similarly here. Don't write to this cache-line in the team structure
1126 // unless we have to.
1127 KMP_CHECK_UPDATE(team->t.t_fp_control_saved, FALSE);
1128 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001129}
1130
Jonathan Peyton30419822017-05-12 18:01:32 +00001131// Do the opposite, setting the hardware registers to the updated values from
1132// the team.
1133inline static void updateHWFPControl(kmp_team_t *team) {
1134 if (__kmp_inherit_fp_control && team->t.t_fp_control_saved) {
1135 // Only reset the fp control regs if they have been changed in the team.
1136 // the parallel region that we are exiting.
1137 kmp_int16 x87_fpu_control_word;
1138 kmp_uint32 mxcsr;
1139 __kmp_store_x87_fpu_control_word(&x87_fpu_control_word);
1140 __kmp_store_mxcsr(&mxcsr);
1141 mxcsr &= KMP_X86_MXCSR_MASK;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001142
Jonathan Peyton30419822017-05-12 18:01:32 +00001143 if (team->t.t_x87_fpu_control_word != x87_fpu_control_word) {
1144 __kmp_clear_x87_fpu_status_word();
1145 __kmp_load_x87_fpu_control_word(&team->t.t_x87_fpu_control_word);
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001146 }
Jonathan Peyton30419822017-05-12 18:01:32 +00001147
1148 if (team->t.t_mxcsr != mxcsr) {
1149 __kmp_load_mxcsr(&team->t.t_mxcsr);
1150 }
1151 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001152}
1153#else
Jonathan Peyton30419822017-05-12 18:01:32 +00001154#define propagateFPControl(x) ((void)0)
1155#define updateHWFPControl(x) ((void)0)
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001156#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
1157
Jonathan Peyton30419822017-05-12 18:01:32 +00001158static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team,
1159 int realloc); // forward declaration
Jim Cownie5e8470a2013-09-27 10:38:44 +00001160
Jonathan Peyton30419822017-05-12 18:01:32 +00001161/* Run a parallel region that has been serialized, so runs only in a team of the
1162 single master thread. */
1163void __kmp_serialized_parallel(ident_t *loc, kmp_int32 global_tid) {
1164 kmp_info_t *this_thr;
1165 kmp_team_t *serial_team;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001166
Jonathan Peyton30419822017-05-12 18:01:32 +00001167 KC_TRACE(10, ("__kmpc_serialized_parallel: called by T#%d\n", global_tid));
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001168
Jonathan Peyton30419822017-05-12 18:01:32 +00001169 /* Skip all this code for autopar serialized loops since it results in
1170 unacceptable overhead */
1171 if (loc != NULL && (loc->flags & KMP_IDENT_AUTOPAR))
1172 return;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001173
Jonathan Peyton30419822017-05-12 18:01:32 +00001174 if (!TCR_4(__kmp_init_parallel))
1175 __kmp_parallel_initialize();
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001176
Jonathan Peyton30419822017-05-12 18:01:32 +00001177 this_thr = __kmp_threads[global_tid];
1178 serial_team = this_thr->th.th_serial_team;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001179
Jonathan Peyton30419822017-05-12 18:01:32 +00001180 /* utilize the serialized team held by this thread */
1181 KMP_DEBUG_ASSERT(serial_team);
1182 KMP_MB();
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001183
Jonathan Peyton30419822017-05-12 18:01:32 +00001184 if (__kmp_tasking_mode != tskm_immediate_exec) {
1185 KMP_DEBUG_ASSERT(
1186 this_thr->th.th_task_team ==
1187 this_thr->th.th_team->t.t_task_team[this_thr->th.th_task_state]);
1188 KMP_DEBUG_ASSERT(serial_team->t.t_task_team[this_thr->th.th_task_state] ==
1189 NULL);
1190 KA_TRACE(20, ("__kmpc_serialized_parallel: T#%d pushing task_team %p / "
1191 "team %p, new task_team = NULL\n",
1192 global_tid, this_thr->th.th_task_team, this_thr->th.th_team));
1193 this_thr->th.th_task_team = NULL;
1194 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001195
1196#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00001197 kmp_proc_bind_t proc_bind = this_thr->th.th_set_proc_bind;
1198 if (this_thr->th.th_current_task->td_icvs.proc_bind == proc_bind_false) {
1199 proc_bind = proc_bind_false;
1200 } else if (proc_bind == proc_bind_default) {
1201 // No proc_bind clause was specified, so use the current value
1202 // of proc-bind-var for this parallel region.
1203 proc_bind = this_thr->th.th_current_task->td_icvs.proc_bind;
1204 }
1205 // Reset for next parallel region
1206 this_thr->th.th_set_proc_bind = proc_bind_default;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001207#endif /* OMP_40_ENABLED */
1208
Jonathan Peyton30419822017-05-12 18:01:32 +00001209 if (this_thr->th.th_team != serial_team) {
1210 // Nested level will be an index in the nested nthreads array
1211 int level = this_thr->th.th_team->t.t_level;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001212
Jonathan Peyton30419822017-05-12 18:01:32 +00001213 if (serial_team->t.t_serialized) {
1214 /* this serial team was already used
1215 TODO increase performance by making this locks more specific */
1216 kmp_team_t *new_team;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001217
Jonathan Peyton30419822017-05-12 18:01:32 +00001218 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001219
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001220#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00001221 ompt_parallel_id_t ompt_parallel_id = __ompt_parallel_id_new(global_tid);
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001222#endif
1223
Jonathan Peyton30419822017-05-12 18:01:32 +00001224 new_team = __kmp_allocate_team(this_thr->th.th_root, 1, 1,
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001225#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00001226 ompt_parallel_id,
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001227#endif
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001228#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00001229 proc_bind,
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001230#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00001231 &this_thr->th.th_current_task->td_icvs,
1232 0 USE_NESTED_HOT_ARG(NULL));
1233 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1234 KMP_ASSERT(new_team);
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001235
Jonathan Peyton30419822017-05-12 18:01:32 +00001236 /* setup new serialized team and install it */
1237 new_team->t.t_threads[0] = this_thr;
1238 new_team->t.t_parent = this_thr->th.th_team;
1239 serial_team = new_team;
1240 this_thr->th.th_serial_team = serial_team;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001241
Jonathan Peyton30419822017-05-12 18:01:32 +00001242 KF_TRACE(
1243 10,
1244 ("__kmpc_serialized_parallel: T#%d allocated new serial team %p\n",
1245 global_tid, serial_team));
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001246
Jonathan Peyton30419822017-05-12 18:01:32 +00001247 /* TODO the above breaks the requirement that if we run out of resources,
1248 then we can still guarantee that serialized teams are ok, since we may
1249 need to allocate a new one */
1250 } else {
1251 KF_TRACE(
1252 10,
1253 ("__kmpc_serialized_parallel: T#%d reusing cached serial team %p\n",
1254 global_tid, serial_team));
1255 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001256
Jonathan Peyton30419822017-05-12 18:01:32 +00001257 /* we have to initialize this serial team */
1258 KMP_DEBUG_ASSERT(serial_team->t.t_threads);
1259 KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr);
1260 KMP_DEBUG_ASSERT(this_thr->th.th_team != serial_team);
1261 serial_team->t.t_ident = loc;
1262 serial_team->t.t_serialized = 1;
1263 serial_team->t.t_nproc = 1;
1264 serial_team->t.t_parent = this_thr->th.th_team;
1265 serial_team->t.t_sched = this_thr->th.th_team->t.t_sched;
1266 this_thr->th.th_team = serial_team;
1267 serial_team->t.t_master_tid = this_thr->th.th_info.ds.ds_tid;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001268
Jonathan Peyton30419822017-05-12 18:01:32 +00001269 KF_TRACE(10, ("__kmpc_serialized_parallel: T#d curtask=%p\n", global_tid,
1270 this_thr->th.th_current_task));
1271 KMP_ASSERT(this_thr->th.th_current_task->td_flags.executing == 1);
1272 this_thr->th.th_current_task->td_flags.executing = 0;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001273
Jonathan Peyton30419822017-05-12 18:01:32 +00001274 __kmp_push_current_task_to_thread(this_thr, serial_team, 0);
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001275
Jonathan Peyton30419822017-05-12 18:01:32 +00001276 /* TODO: GEH: do ICVs work for nested serialized teams? Don't we need an
1277 implicit task for each serialized task represented by
1278 team->t.t_serialized? */
1279 copy_icvs(&this_thr->th.th_current_task->td_icvs,
1280 &this_thr->th.th_current_task->td_parent->td_icvs);
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001281
Jonathan Peyton30419822017-05-12 18:01:32 +00001282 // Thread value exists in the nested nthreads array for the next nested
1283 // level
1284 if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) {
1285 this_thr->th.th_current_task->td_icvs.nproc =
1286 __kmp_nested_nth.nth[level + 1];
1287 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001288
1289#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00001290 if (__kmp_nested_proc_bind.used &&
1291 (level + 1 < __kmp_nested_proc_bind.used)) {
1292 this_thr->th.th_current_task->td_icvs.proc_bind =
1293 __kmp_nested_proc_bind.bind_types[level + 1];
1294 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001295#endif /* OMP_40_ENABLED */
1296
Jonathan Peyton8fbb49a2015-07-09 18:16:58 +00001297#if USE_DEBUGGER
Jonathan Peyton30419822017-05-12 18:01:32 +00001298 serial_team->t.t_pkfn = (microtask_t)(~0); // For the debugger.
Jonathan Peyton8fbb49a2015-07-09 18:16:58 +00001299#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00001300 this_thr->th.th_info.ds.ds_tid = 0;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001301
Jonathan Peyton30419822017-05-12 18:01:32 +00001302 /* set thread cache values */
1303 this_thr->th.th_team_nproc = 1;
1304 this_thr->th.th_team_master = this_thr;
1305 this_thr->th.th_team_serialized = 1;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001306
Jonathan Peyton30419822017-05-12 18:01:32 +00001307 serial_team->t.t_level = serial_team->t.t_parent->t.t_level + 1;
1308 serial_team->t.t_active_level = serial_team->t.t_parent->t.t_active_level;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001309
Jonathan Peyton30419822017-05-12 18:01:32 +00001310 propagateFPControl(serial_team);
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001311
Jonathan Peyton30419822017-05-12 18:01:32 +00001312 /* check if we need to allocate dispatch buffers stack */
1313 KMP_DEBUG_ASSERT(serial_team->t.t_dispatch);
1314 if (!serial_team->t.t_dispatch->th_disp_buffer) {
1315 serial_team->t.t_dispatch->th_disp_buffer =
1316 (dispatch_private_info_t *)__kmp_allocate(
1317 sizeof(dispatch_private_info_t));
1318 }
1319 this_thr->th.th_dispatch = serial_team->t.t_dispatch;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001320
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001321#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00001322 ompt_parallel_id_t ompt_parallel_id = __ompt_parallel_id_new(global_tid);
1323 __ompt_team_assign_id(serial_team, ompt_parallel_id);
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001324#endif
1325
Jonathan Peyton30419822017-05-12 18:01:32 +00001326 KMP_MB();
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001327
Jonathan Peyton30419822017-05-12 18:01:32 +00001328 } else {
1329 /* this serialized team is already being used,
1330 * that's fine, just add another nested level */
1331 KMP_DEBUG_ASSERT(this_thr->th.th_team == serial_team);
1332 KMP_DEBUG_ASSERT(serial_team->t.t_threads);
1333 KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr);
1334 ++serial_team->t.t_serialized;
1335 this_thr->th.th_team_serialized = serial_team->t.t_serialized;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001336
Jonathan Peyton30419822017-05-12 18:01:32 +00001337 // Nested level will be an index in the nested nthreads array
1338 int level = this_thr->th.th_team->t.t_level;
1339 // Thread value exists in the nested nthreads array for the next nested
1340 // level
1341 if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) {
1342 this_thr->th.th_current_task->td_icvs.nproc =
1343 __kmp_nested_nth.nth[level + 1];
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001344 }
Jonathan Peyton30419822017-05-12 18:01:32 +00001345 serial_team->t.t_level++;
1346 KF_TRACE(10, ("__kmpc_serialized_parallel: T#%d increasing nesting level "
1347 "of serial team %p to %d\n",
1348 global_tid, serial_team, serial_team->t.t_level));
1349
1350 /* allocate/push dispatch buffers stack */
1351 KMP_DEBUG_ASSERT(serial_team->t.t_dispatch);
1352 {
1353 dispatch_private_info_t *disp_buffer =
1354 (dispatch_private_info_t *)__kmp_allocate(
1355 sizeof(dispatch_private_info_t));
1356 disp_buffer->next = serial_team->t.t_dispatch->th_disp_buffer;
1357 serial_team->t.t_dispatch->th_disp_buffer = disp_buffer;
1358 }
1359 this_thr->th.th_dispatch = serial_team->t.t_dispatch;
1360
1361 KMP_MB();
1362 }
Olga Malyshevadbdcfa12017-04-04 13:56:50 +00001363#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00001364 KMP_CHECK_UPDATE(serial_team->t.t_cancel_request, cancel_noreq);
Olga Malyshevadbdcfa12017-04-04 13:56:50 +00001365#endif
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001366
Jonathan Peyton30419822017-05-12 18:01:32 +00001367 if (__kmp_env_consistency_check)
1368 __kmp_push_parallel(global_tid, NULL);
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001369}
Jim Cownie181b4bb2013-12-23 17:28:57 +00001370
Jim Cownie5e8470a2013-09-27 10:38:44 +00001371/* most of the work for a fork */
1372/* return true if we really went parallel, false if serialized */
Jonathan Peyton30419822017-05-12 18:01:32 +00001373int __kmp_fork_call(ident_t *loc, int gtid,
1374 enum fork_context_e call_context, // Intel, GNU, ...
1375 kmp_int32 argc,
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001376#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00001377 void *unwrapped_task,
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001378#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00001379 microtask_t microtask, launch_t invoker,
Jim Cownie5e8470a2013-09-27 10:38:44 +00001380/* TODO: revert workaround for Intel(R) 64 tracker #96 */
Andrey Churbanovcbda8682015-01-13 14:43:35 +00001381#if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
Jonathan Peyton30419822017-05-12 18:01:32 +00001382 va_list *ap
Jim Cownie5e8470a2013-09-27 10:38:44 +00001383#else
Jonathan Peyton30419822017-05-12 18:01:32 +00001384 va_list ap
Jim Cownie5e8470a2013-09-27 10:38:44 +00001385#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00001386 ) {
1387 void **argv;
1388 int i;
1389 int master_tid;
1390 int master_this_cons;
1391 kmp_team_t *team;
1392 kmp_team_t *parent_team;
1393 kmp_info_t *master_th;
1394 kmp_root_t *root;
1395 int nthreads;
1396 int master_active;
1397 int master_set_numthreads;
1398 int level;
Jim Cownie5e8470a2013-09-27 10:38:44 +00001399#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00001400 int active_level;
1401 int teams_level;
Jim Cownie5e8470a2013-09-27 10:38:44 +00001402#endif
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001403#if KMP_NESTED_HOT_TEAMS
Jonathan Peyton30419822017-05-12 18:01:32 +00001404 kmp_hot_team_ptr_t **p_hot_teams;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001405#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00001406 { // KMP_TIME_BLOCK
Jonathan Peyton5375fe82016-11-14 21:13:44 +00001407 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_fork_call);
Jonathan Peyton45be4502015-08-11 21:36:41 +00001408 KMP_COUNT_VALUE(OMP_PARALLEL_args, argc);
Jim Cownie5e8470a2013-09-27 10:38:44 +00001409
Jonathan Peyton30419822017-05-12 18:01:32 +00001410 KA_TRACE(20, ("__kmp_fork_call: enter T#%d\n", gtid));
1411 if (__kmp_stkpadding > 0 && __kmp_root[gtid] != NULL) {
1412 /* Some systems prefer the stack for the root thread(s) to start with */
1413 /* some gap from the parent stack to prevent false sharing. */
1414 void *dummy = KMP_ALLOCA(__kmp_stkpadding);
1415 /* These 2 lines below are so this does not get optimized out */
1416 if (__kmp_stkpadding > KMP_MAX_STKPADDING)
1417 __kmp_stkpadding += (short)((kmp_int64)dummy);
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001418 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00001419
1420 /* initialize if needed */
Jonathan Peyton30419822017-05-12 18:01:32 +00001421 KMP_DEBUG_ASSERT(
1422 __kmp_init_serial); // AC: potentially unsafe, not in sync with shutdown
1423 if (!TCR_4(__kmp_init_parallel))
1424 __kmp_parallel_initialize();
Jim Cownie5e8470a2013-09-27 10:38:44 +00001425
1426 /* setup current data */
Jonathan Peyton30419822017-05-12 18:01:32 +00001427 master_th = __kmp_threads[gtid]; // AC: potentially unsafe, not in sync with
1428 // shutdown
1429 parent_team = master_th->th.th_team;
1430 master_tid = master_th->th.th_info.ds.ds_tid;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001431 master_this_cons = master_th->th.th_local.this_construct;
Jonathan Peyton30419822017-05-12 18:01:32 +00001432 root = master_th->th.th_root;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001433 master_active = root->r.r_active;
1434 master_set_numthreads = master_th->th.th_set_nproc;
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001435
1436#if OMPT_SUPPORT
1437 ompt_parallel_id_t ompt_parallel_id;
1438 ompt_task_id_t ompt_task_id;
1439 ompt_frame_t *ompt_frame;
1440 ompt_task_id_t my_task_id;
1441 ompt_parallel_id_t my_parallel_id;
1442
Jonathan Peytonb68a85d2015-09-21 18:11:22 +00001443 if (ompt_enabled) {
Jonathan Peyton30419822017-05-12 18:01:32 +00001444 ompt_parallel_id = __ompt_parallel_id_new(gtid);
1445 ompt_task_id = __ompt_get_task_id_internal(0);
1446 ompt_frame = __ompt_get_task_frame_internal(0);
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001447 }
1448#endif
1449
Jim Cownie5e8470a2013-09-27 10:38:44 +00001450 // Nested level will be an index in the nested nthreads array
Jonathan Peyton30419822017-05-12 18:01:32 +00001451 level = parent_team->t.t_level;
1452 // used to launch non-serial teams even if nested is not allowed
1453 active_level = parent_team->t.t_active_level;
Jonathan Peytonc76f9f02016-06-21 19:12:07 +00001454#if OMP_40_ENABLED
Jonathan Peyton642688b2017-06-01 16:46:36 +00001455 // needed to check nesting inside the teams
1456 teams_level = master_th->th.th_teams_level;
Jim Cownie5e8470a2013-09-27 10:38:44 +00001457#endif
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001458#if KMP_NESTED_HOT_TEAMS
Jonathan Peyton30419822017-05-12 18:01:32 +00001459 p_hot_teams = &master_th->th.th_hot_teams;
1460 if (*p_hot_teams == NULL && __kmp_hot_teams_max_level > 0) {
1461 *p_hot_teams = (kmp_hot_team_ptr_t *)__kmp_allocate(
1462 sizeof(kmp_hot_team_ptr_t) * __kmp_hot_teams_max_level);
1463 (*p_hot_teams)[0].hot_team = root->r.r_hot_team;
Jonathan Peyton642688b2017-06-01 16:46:36 +00001464 // it is either actual or not needed (when active_level > 0)
1465 (*p_hot_teams)[0].hot_team_nth = 1;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001466 }
1467#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00001468
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001469#if OMPT_SUPPORT
Jonathan Peytonb68a85d2015-09-21 18:11:22 +00001470 if (ompt_enabled &&
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001471 ompt_callbacks.ompt_callback(ompt_event_parallel_begin)) {
Jonathan Peyton30419822017-05-12 18:01:32 +00001472 int team_size = master_set_numthreads;
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001473
Jonathan Peyton30419822017-05-12 18:01:32 +00001474 ompt_callbacks.ompt_callback(ompt_event_parallel_begin)(
1475 ompt_task_id, ompt_frame, ompt_parallel_id, team_size, unwrapped_task,
1476 OMPT_INVOKER(call_context));
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001477 }
1478#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00001479
Jim Cownie5e8470a2013-09-27 10:38:44 +00001480 master_th->th.th_ident = loc;
1481
1482#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00001483 if (master_th->th.th_teams_microtask && ap &&
1484 microtask != (microtask_t)__kmp_teams_master && level == teams_level) {
1485 // AC: This is start of parallel that is nested inside teams construct.
1486 // The team is actual (hot), all workers are ready at the fork barrier.
1487 // No lock needed to initialize the team a bit, then free workers.
1488 parent_team->t.t_ident = loc;
1489 __kmp_alloc_argv_entries(argc, parent_team, TRUE);
1490 parent_team->t.t_argc = argc;
1491 argv = (void **)parent_team->t.t_argv;
1492 for (i = argc - 1; i >= 0; --i)
Jim Cownie5e8470a2013-09-27 10:38:44 +00001493/* TODO: revert workaround for Intel(R) 64 tracker #96 */
Andrey Churbanovcbda8682015-01-13 14:43:35 +00001494#if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
Jonathan Peyton30419822017-05-12 18:01:32 +00001495 *argv++ = va_arg(*ap, void *);
Jim Cownie5e8470a2013-09-27 10:38:44 +00001496#else
Jonathan Peyton30419822017-05-12 18:01:32 +00001497 *argv++ = va_arg(ap, void *);
Jim Cownie5e8470a2013-09-27 10:38:44 +00001498#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00001499 // Increment our nested depth levels, but not increase the serialization
1500 if (parent_team == master_th->th.th_serial_team) {
1501 // AC: we are in serialized parallel
1502 __kmpc_serialized_parallel(loc, gtid);
1503 KMP_DEBUG_ASSERT(parent_team->t.t_serialized > 1);
1504 // AC: need this in order enquiry functions work
1505 // correctly, will restore at join time
1506 parent_team->t.t_serialized--;
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001507#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00001508 void *dummy;
1509 void **exit_runtime_p;
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001510
Jonathan Peyton30419822017-05-12 18:01:32 +00001511 ompt_lw_taskteam_t lw_taskteam;
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001512
Jonathan Peyton30419822017-05-12 18:01:32 +00001513 if (ompt_enabled) {
1514 __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid, unwrapped_task,
1515 ompt_parallel_id);
1516 lw_taskteam.ompt_task_info.task_id = __ompt_task_id_new(gtid);
1517 exit_runtime_p =
1518 &(lw_taskteam.ompt_task_info.frame.exit_runtime_frame);
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001519
Jonathan Peyton30419822017-05-12 18:01:32 +00001520 __ompt_lw_taskteam_link(&lw_taskteam, master_th);
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001521
1522#if OMPT_TRACE
Jonathan Peyton30419822017-05-12 18:01:32 +00001523 /* OMPT implicit task begin */
1524 my_task_id = lw_taskteam.ompt_task_info.task_id;
1525 my_parallel_id = parent_team->t.ompt_team_info.parallel_id;
1526 if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)) {
1527 ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)(
1528 my_parallel_id, my_task_id);
1529 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001530#endif
1531
Jonathan Peyton30419822017-05-12 18:01:32 +00001532 /* OMPT state */
1533 master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
1534 } else {
1535 exit_runtime_p = &dummy;
Jim Cownie5e8470a2013-09-27 10:38:44 +00001536 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001537#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00001538
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001539 {
Jonathan Peyton30419822017-05-12 18:01:32 +00001540 KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1541 KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1542 __kmp_invoke_microtask(microtask, gtid, 0, argc, parent_team->t.t_argv
1543#if OMPT_SUPPORT
1544 ,
1545 exit_runtime_p
1546#endif
1547 );
Jim Cownie5e8470a2013-09-27 10:38:44 +00001548 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00001549
Jonathan Peyton30419822017-05-12 18:01:32 +00001550#if OMPT_SUPPORT
1551 *exit_runtime_p = NULL;
1552 if (ompt_enabled) {
1553#if OMPT_TRACE
1554 lw_taskteam.ompt_task_info.frame.exit_runtime_frame = NULL;
Jim Cownie5e8470a2013-09-27 10:38:44 +00001555
Jonathan Peyton30419822017-05-12 18:01:32 +00001556 if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) {
1557 ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)(
1558 ompt_parallel_id, ompt_task_id);
1559 }
1560
1561 __ompt_lw_taskteam_unlink(master_th);
1562 // reset clear the task id only after unlinking the task
1563 lw_taskteam.ompt_task_info.task_id = ompt_task_id_none;
1564#endif
1565
1566 if (ompt_callbacks.ompt_callback(ompt_event_parallel_end)) {
1567 ompt_callbacks.ompt_callback(ompt_event_parallel_end)(
1568 ompt_parallel_id, ompt_task_id, OMPT_INVOKER(call_context));
1569 }
1570 master_th->th.ompt_thread_info.state = ompt_state_overhead;
1571 }
1572#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00001573 return TRUE;
Jonathan Peyton30419822017-05-12 18:01:32 +00001574 }
1575
1576 parent_team->t.t_pkfn = microtask;
1577#if OMPT_SUPPORT
1578 parent_team->t.ompt_team_info.microtask = unwrapped_task;
1579#endif
1580 parent_team->t.t_invoke = invoker;
1581 KMP_TEST_THEN_INC32((kmp_int32 *)&root->r.r_in_parallel);
1582 parent_team->t.t_active_level++;
1583 parent_team->t.t_level++;
1584
1585 /* Change number of threads in the team if requested */
1586 if (master_set_numthreads) { // The parallel has num_threads clause
1587 if (master_set_numthreads < master_th->th.th_teams_size.nth) {
1588 // AC: only can reduce number of threads dynamically, can't increase
1589 kmp_info_t **other_threads = parent_team->t.t_threads;
1590 parent_team->t.t_nproc = master_set_numthreads;
1591 for (i = 0; i < master_set_numthreads; ++i) {
1592 other_threads[i]->th.th_team_nproc = master_set_numthreads;
1593 }
1594 // Keep extra threads hot in the team for possible next parallels
1595 }
1596 master_th->th.th_set_nproc = 0;
1597 }
1598
1599#if USE_DEBUGGER
1600 if (__kmp_debugging) { // Let debugger override number of threads.
1601 int nth = __kmp_omp_num_threads(loc);
Jonathan Peyton642688b2017-06-01 16:46:36 +00001602 if (nth > 0) { // 0 means debugger doesn't want to change num threads
Jonathan Peyton30419822017-05-12 18:01:32 +00001603 master_set_numthreads = nth;
1604 }; // if
1605 }; // if
1606#endif
1607
1608 KF_TRACE(10, ("__kmp_fork_call: before internal fork: root=%p, team=%p, "
1609 "master_th=%p, gtid=%d\n",
1610 root, parent_team, master_th, gtid));
1611 __kmp_internal_fork(loc, gtid, parent_team);
1612 KF_TRACE(10, ("__kmp_fork_call: after internal fork: root=%p, team=%p, "
1613 "master_th=%p, gtid=%d\n",
1614 root, parent_team, master_th, gtid));
1615
1616 /* Invoke microtask for MASTER thread */
1617 KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n", gtid,
1618 parent_team->t.t_id, parent_team->t.t_pkfn));
1619
1620 {
1621 KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1622 KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1623 if (!parent_team->t.t_invoke(gtid)) {
1624 KMP_ASSERT2(0, "cannot invoke microtask for MASTER thread");
1625 }
1626 }
1627 KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid,
1628 parent_team->t.t_id, parent_team->t.t_pkfn));
1629 KMP_MB(); /* Flush all pending memory write invalidates. */
1630
1631 KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
1632
1633 return TRUE;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001634 } // Parallel closely nested in teams construct
Jim Cownie5e8470a2013-09-27 10:38:44 +00001635#endif /* OMP_40_ENABLED */
1636
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001637#if KMP_DEBUG
Jonathan Peyton30419822017-05-12 18:01:32 +00001638 if (__kmp_tasking_mode != tskm_immediate_exec) {
1639 KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
1640 parent_team->t.t_task_team[master_th->th.th_task_state]);
Jim Cownie5e8470a2013-09-27 10:38:44 +00001641 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001642#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00001643
Jonathan Peyton30419822017-05-12 18:01:32 +00001644 if (parent_team->t.t_active_level >=
1645 master_th->th.th_current_task->td_icvs.max_active_levels) {
1646 nthreads = 1;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001647 } else {
Andrey Churbanov92effc42015-08-18 10:08:27 +00001648#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00001649 int enter_teams = ((ap == NULL && active_level == 0) ||
1650 (ap && teams_level > 0 && teams_level == level));
Andrey Churbanov92effc42015-08-18 10:08:27 +00001651#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00001652 nthreads =
1653 master_set_numthreads
1654 ? master_set_numthreads
1655 : get__nproc_2(
1656 parent_team,
1657 master_tid); // TODO: get nproc directly from current task
Andrey Churbanov92effc42015-08-18 10:08:27 +00001658
Jonathan Peyton30419822017-05-12 18:01:32 +00001659 // Check if we need to take forkjoin lock? (no need for serialized
1660 // parallel out of teams construct). This code moved here from
1661 // __kmp_reserve_threads() to speedup nested serialized parallels.
1662 if (nthreads > 1) {
1663 if ((!get__nested(master_th) && (root->r.r_in_parallel
Andrey Churbanov92effc42015-08-18 10:08:27 +00001664#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00001665 && !enter_teams
Andrey Churbanov92effc42015-08-18 10:08:27 +00001666#endif /* OMP_40_ENABLED */
Jonathan Peyton30419822017-05-12 18:01:32 +00001667 )) ||
1668 (__kmp_library == library_serial)) {
Jonathan Peyton642688b2017-06-01 16:46:36 +00001669 KC_TRACE(10, ("__kmp_fork_call: T#%d serializing team; requested %d"
1670 " threads\n",
1671 gtid, nthreads));
Jonathan Peyton30419822017-05-12 18:01:32 +00001672 nthreads = 1;
Andrey Churbanov92effc42015-08-18 10:08:27 +00001673 }
Jonathan Peyton30419822017-05-12 18:01:32 +00001674 }
1675 if (nthreads > 1) {
1676 /* determine how many new threads we can use */
1677 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
Jonathan Peyton30419822017-05-12 18:01:32 +00001678 nthreads = __kmp_reserve_threads(
1679 root, parent_team, master_tid, nthreads
Jim Cownie5e8470a2013-09-27 10:38:44 +00001680#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00001681 /* AC: If we execute teams from parallel region (on host), then
1682 teams should be created but each can only have 1 thread if
1683 nesting is disabled. If teams called from serial region, then
1684 teams and their threads should be created regardless of the
1685 nesting setting. */
1686 ,
1687 enter_teams
Jim Cownie5e8470a2013-09-27 10:38:44 +00001688#endif /* OMP_40_ENABLED */
Jonathan Peyton30419822017-05-12 18:01:32 +00001689 );
1690 if (nthreads == 1) {
1691 // Free lock for single thread execution here; for multi-thread
1692 // execution it will be freed later after team of threads created
1693 // and initialized
1694 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
Andrey Churbanov92effc42015-08-18 10:08:27 +00001695 }
Jonathan Peyton30419822017-05-12 18:01:32 +00001696 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00001697 }
Jonathan Peyton30419822017-05-12 18:01:32 +00001698 KMP_DEBUG_ASSERT(nthreads > 0);
Jim Cownie5e8470a2013-09-27 10:38:44 +00001699
Jonathan Peyton30419822017-05-12 18:01:32 +00001700 // If we temporarily changed the set number of threads then restore it now
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001701 master_th->th.th_set_nproc = 0;
Jim Cownie5e8470a2013-09-27 10:38:44 +00001702
Jim Cownie5e8470a2013-09-27 10:38:44 +00001703 /* create a serialized parallel region? */
Jonathan Peyton30419822017-05-12 18:01:32 +00001704 if (nthreads == 1) {
1705/* josh todo: hypothetical question: what do we do for OS X*? */
1706#if KMP_OS_LINUX && \
1707 (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64)
1708 void *args[argc];
Jim Cownie5e8470a2013-09-27 10:38:44 +00001709#else
Jonathan Peyton30419822017-05-12 18:01:32 +00001710 void **args = (void **)KMP_ALLOCA(argc * sizeof(void *));
1711#endif /* KMP_OS_LINUX && ( KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || \
1712 KMP_ARCH_AARCH64) */
Jim Cownie5e8470a2013-09-27 10:38:44 +00001713
Jonathan Peyton30419822017-05-12 18:01:32 +00001714 KA_TRACE(20,
1715 ("__kmp_fork_call: T#%d serializing parallel region\n", gtid));
Jim Cownie5e8470a2013-09-27 10:38:44 +00001716
Jonathan Peyton30419822017-05-12 18:01:32 +00001717 __kmpc_serialized_parallel(loc, gtid);
Jim Cownie5e8470a2013-09-27 10:38:44 +00001718
Jonathan Peyton30419822017-05-12 18:01:32 +00001719 if (call_context == fork_context_intel) {
1720 /* TODO this sucks, use the compiler itself to pass args! :) */
1721 master_th->th.th_serial_team->t.t_ident = loc;
Jim Cownie5e8470a2013-09-27 10:38:44 +00001722#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00001723 if (!ap) {
1724 // revert change made in __kmpc_serialized_parallel()
1725 master_th->th.th_serial_team->t.t_level--;
1726// Get args from parent team for teams construct
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001727
1728#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00001729 void *dummy;
1730 void **exit_runtime_p;
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001731
Jonathan Peyton30419822017-05-12 18:01:32 +00001732 ompt_lw_taskteam_t lw_taskteam;
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001733
Jonathan Peyton30419822017-05-12 18:01:32 +00001734 if (ompt_enabled) {
1735 __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1736 unwrapped_task, ompt_parallel_id);
1737 lw_taskteam.ompt_task_info.task_id = __ompt_task_id_new(gtid);
1738 exit_runtime_p =
1739 &(lw_taskteam.ompt_task_info.frame.exit_runtime_frame);
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001740
Jonathan Peyton30419822017-05-12 18:01:32 +00001741 __ompt_lw_taskteam_link(&lw_taskteam, master_th);
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001742
1743#if OMPT_TRACE
Jonathan Peyton30419822017-05-12 18:01:32 +00001744 my_task_id = lw_taskteam.ompt_task_info.task_id;
1745 if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)) {
1746 ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)(
1747 ompt_parallel_id, my_task_id);
1748 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001749#endif
1750
Jonathan Peyton30419822017-05-12 18:01:32 +00001751 /* OMPT state */
1752 master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
1753 } else {
1754 exit_runtime_p = &dummy;
1755 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001756#endif
1757
Jonathan Peyton30419822017-05-12 18:01:32 +00001758 {
1759 KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1760 KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1761 __kmp_invoke_microtask(microtask, gtid, 0, argc,
1762 parent_team->t.t_argv
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001763#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00001764 ,
1765 exit_runtime_p
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001766#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00001767 );
1768 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001769
1770#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00001771 *exit_runtime_p = NULL;
1772 if (ompt_enabled) {
1773 lw_taskteam.ompt_task_info.frame.exit_runtime_frame = NULL;
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001774
1775#if OMPT_TRACE
Jonathan Peyton30419822017-05-12 18:01:32 +00001776 if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) {
1777 ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)(
1778 ompt_parallel_id, ompt_task_id);
1779 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001780#endif
1781
Jonathan Peyton30419822017-05-12 18:01:32 +00001782 __ompt_lw_taskteam_unlink(master_th);
1783 // reset clear the task id only after unlinking the task
1784 lw_taskteam.ompt_task_info.task_id = ompt_task_id_none;
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001785
Jonathan Peyton30419822017-05-12 18:01:32 +00001786 if (ompt_callbacks.ompt_callback(ompt_event_parallel_end)) {
1787 ompt_callbacks.ompt_callback(ompt_event_parallel_end)(
1788 ompt_parallel_id, ompt_task_id, OMPT_INVOKER(call_context));
1789 }
1790 master_th->th.ompt_thread_info.state = ompt_state_overhead;
1791 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001792#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00001793 } else if (microtask == (microtask_t)__kmp_teams_master) {
1794 KMP_DEBUG_ASSERT(master_th->th.th_team ==
1795 master_th->th.th_serial_team);
1796 team = master_th->th.th_team;
1797 // team->t.t_pkfn = microtask;
1798 team->t.t_invoke = invoker;
1799 __kmp_alloc_argv_entries(argc, team, TRUE);
1800 team->t.t_argc = argc;
1801 argv = (void **)team->t.t_argv;
1802 if (ap) {
1803 for (i = argc - 1; i >= 0; --i)
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001804// TODO: revert workaround for Intel(R) 64 tracker #96
Andrey Churbanovcbda8682015-01-13 14:43:35 +00001805#if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
Jonathan Peyton30419822017-05-12 18:01:32 +00001806 *argv++ = va_arg(*ap, void *);
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001807#else
Jonathan Peyton30419822017-05-12 18:01:32 +00001808 *argv++ = va_arg(ap, void *);
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001809#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00001810 } else {
1811 for (i = 0; i < argc; ++i)
1812 // Get args from parent team for teams construct
1813 argv[i] = parent_team->t.t_argv[i];
1814 }
1815 // AC: revert change made in __kmpc_serialized_parallel()
1816 // because initial code in teams should have level=0
1817 team->t.t_level--;
1818 // AC: call special invoker for outer "parallel" of teams construct
1819 {
1820 KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1821 KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1822 invoker(gtid);
1823 }
1824 } else {
1825#endif /* OMP_40_ENABLED */
1826 argv = args;
1827 for (i = argc - 1; i >= 0; --i)
1828// TODO: revert workaround for Intel(R) 64 tracker #96
1829#if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
1830 *argv++ = va_arg(*ap, void *);
1831#else
1832 *argv++ = va_arg(ap, void *);
1833#endif
1834 KMP_MB();
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001835
1836#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00001837 void *dummy;
1838 void **exit_runtime_p;
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001839
Jonathan Peyton30419822017-05-12 18:01:32 +00001840 ompt_lw_taskteam_t lw_taskteam;
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001841
Jonathan Peyton30419822017-05-12 18:01:32 +00001842 if (ompt_enabled) {
1843 __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1844 unwrapped_task, ompt_parallel_id);
1845 lw_taskteam.ompt_task_info.task_id = __ompt_task_id_new(gtid);
1846 exit_runtime_p =
1847 &(lw_taskteam.ompt_task_info.frame.exit_runtime_frame);
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001848
Jonathan Peyton30419822017-05-12 18:01:32 +00001849 __ompt_lw_taskteam_link(&lw_taskteam, master_th);
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001850
1851#if OMPT_TRACE
Jonathan Peyton30419822017-05-12 18:01:32 +00001852 /* OMPT implicit task begin */
1853 my_task_id = lw_taskteam.ompt_task_info.task_id;
1854 my_parallel_id = ompt_parallel_id;
1855 if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)) {
1856 ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)(
1857 my_parallel_id, my_task_id);
1858 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001859#endif
1860
Jonathan Peyton30419822017-05-12 18:01:32 +00001861 /* OMPT state */
1862 master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
1863 } else {
1864 exit_runtime_p = &dummy;
1865 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001866#endif
1867
Jonathan Peyton30419822017-05-12 18:01:32 +00001868 {
1869 KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1870 KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1871 __kmp_invoke_microtask(microtask, gtid, 0, argc, args
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001872#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00001873 ,
1874 exit_runtime_p
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001875#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00001876 );
1877 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001878
1879#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00001880 *exit_runtime_p = NULL;
1881 if (ompt_enabled) {
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001882#if OMPT_TRACE
Jonathan Peyton30419822017-05-12 18:01:32 +00001883 lw_taskteam.ompt_task_info.frame.exit_runtime_frame = NULL;
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001884
Jonathan Peyton30419822017-05-12 18:01:32 +00001885 if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) {
1886 ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)(
1887 my_parallel_id, my_task_id);
1888 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001889#endif
1890
Jonathan Peyton30419822017-05-12 18:01:32 +00001891 __ompt_lw_taskteam_unlink(master_th);
1892 // reset clear the task id only after unlinking the task
1893 lw_taskteam.ompt_task_info.task_id = ompt_task_id_none;
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001894
Jonathan Peyton30419822017-05-12 18:01:32 +00001895 if (ompt_callbacks.ompt_callback(ompt_event_parallel_end)) {
1896 ompt_callbacks.ompt_callback(ompt_event_parallel_end)(
1897 ompt_parallel_id, ompt_task_id, OMPT_INVOKER(call_context));
1898 }
1899 master_th->th.ompt_thread_info.state = ompt_state_overhead;
1900 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001901#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00001902#if OMP_40_ENABLED
Jim Cownie5e8470a2013-09-27 10:38:44 +00001903 }
Jonathan Peyton30419822017-05-12 18:01:32 +00001904#endif /* OMP_40_ENABLED */
1905 } else if (call_context == fork_context_gnu) {
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001906#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00001907 ompt_lw_taskteam_t *lwt =
1908 (ompt_lw_taskteam_t *)__kmp_allocate(sizeof(ompt_lw_taskteam_t));
1909 __ompt_lw_taskteam_init(lwt, master_th, gtid, unwrapped_task,
1910 ompt_parallel_id);
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001911
Jonathan Peyton30419822017-05-12 18:01:32 +00001912 lwt->ompt_task_info.task_id = __ompt_task_id_new(gtid);
1913 lwt->ompt_task_info.frame.exit_runtime_frame = NULL;
1914 __ompt_lw_taskteam_link(lwt, master_th);
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00001915#endif
1916
Jonathan Peyton30419822017-05-12 18:01:32 +00001917 // we were called from GNU native code
1918 KA_TRACE(20, ("__kmp_fork_call: T#%d serial exit\n", gtid));
Jim Cownie5e8470a2013-09-27 10:38:44 +00001919 return FALSE;
Andrey Churbanovc47afcd2017-07-03 11:24:08 +00001920 } else {
Jonathan Peyton30419822017-05-12 18:01:32 +00001921 KMP_ASSERT2(call_context < fork_context_last,
1922 "__kmp_fork_call: unknown fork_context parameter");
1923 }
1924
1925 KA_TRACE(20, ("__kmp_fork_call: T#%d serial exit\n", gtid));
1926 KMP_MB();
1927 return FALSE;
Jim Cownie5e8470a2013-09-27 10:38:44 +00001928 }
1929
Jim Cownie5e8470a2013-09-27 10:38:44 +00001930 // GEH: only modify the executing flag in the case when not serialized
1931 // serialized case is handled in kmpc_serialized_parallel
Jonathan Peyton30419822017-05-12 18:01:32 +00001932 KF_TRACE(10, ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, "
1933 "curtask=%p, curtask_max_aclevel=%d\n",
1934 parent_team->t.t_active_level, master_th,
1935 master_th->th.th_current_task,
1936 master_th->th.th_current_task->td_icvs.max_active_levels));
1937 // TODO: GEH - cannot do this assertion because root thread not set up as
1938 // executing
Jim Cownie5e8470a2013-09-27 10:38:44 +00001939 // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 1 );
1940 master_th->th.th_current_task->td_flags.executing = 0;
Jim Cownie5e8470a2013-09-27 10:38:44 +00001941
1942#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00001943 if (!master_th->th.th_teams_microtask || level > teams_level)
Jim Cownie5e8470a2013-09-27 10:38:44 +00001944#endif /* OMP_40_ENABLED */
1945 {
Jonathan Peyton30419822017-05-12 18:01:32 +00001946 /* Increment our nested depth level */
1947 KMP_TEST_THEN_INC32((kmp_int32 *)&root->r.r_in_parallel);
Jim Cownie5e8470a2013-09-27 10:38:44 +00001948 }
1949
Jim Cownie5e8470a2013-09-27 10:38:44 +00001950 // See if we need to make a copy of the ICVs.
Jim Cownie5e8470a2013-09-27 10:38:44 +00001951 int nthreads_icv = master_th->th.th_current_task->td_icvs.nproc;
Jonathan Peyton30419822017-05-12 18:01:32 +00001952 if ((level + 1 < __kmp_nested_nth.used) &&
1953 (__kmp_nested_nth.nth[level + 1] != nthreads_icv)) {
1954 nthreads_icv = __kmp_nested_nth.nth[level + 1];
1955 } else {
1956 nthreads_icv = 0; // don't update
Jim Cownie5e8470a2013-09-27 10:38:44 +00001957 }
1958
1959#if OMP_40_ENABLED
Jim Cownie5e8470a2013-09-27 10:38:44 +00001960 // Figure out the proc_bind_policy for the new team.
Jim Cownie5e8470a2013-09-27 10:38:44 +00001961 kmp_proc_bind_t proc_bind = master_th->th.th_set_proc_bind;
Jonathan Peyton30419822017-05-12 18:01:32 +00001962 kmp_proc_bind_t proc_bind_icv =
1963 proc_bind_default; // proc_bind_default means don't update
1964 if (master_th->th.th_current_task->td_icvs.proc_bind == proc_bind_false) {
1965 proc_bind = proc_bind_false;
1966 } else {
1967 if (proc_bind == proc_bind_default) {
1968 // No proc_bind clause specified; use current proc-bind-var for this
1969 // parallel region
1970 proc_bind = master_th->th.th_current_task->td_icvs.proc_bind;
1971 }
1972 /* else: The proc_bind policy was specified explicitly on parallel clause.
1973 This overrides proc-bind-var for this parallel region, but does not
1974 change proc-bind-var. */
1975 // Figure the value of proc-bind-var for the child threads.
1976 if ((level + 1 < __kmp_nested_proc_bind.used) &&
1977 (__kmp_nested_proc_bind.bind_types[level + 1] !=
1978 master_th->th.th_current_task->td_icvs.proc_bind)) {
1979 proc_bind_icv = __kmp_nested_proc_bind.bind_types[level + 1];
1980 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00001981 }
1982
Jim Cownie5e8470a2013-09-27 10:38:44 +00001983 // Reset for next parallel region
Jim Cownie5e8470a2013-09-27 10:38:44 +00001984 master_th->th.th_set_proc_bind = proc_bind_default;
1985#endif /* OMP_40_ENABLED */
1986
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001987 if ((nthreads_icv > 0)
Jim Cownie5e8470a2013-09-27 10:38:44 +00001988#if OMP_40_ENABLED
Jim Cownie4cc4bb42014-10-07 16:25:50 +00001989 || (proc_bind_icv != proc_bind_default)
Jim Cownie5e8470a2013-09-27 10:38:44 +00001990#endif /* OMP_40_ENABLED */
Jonathan Peyton30419822017-05-12 18:01:32 +00001991 ) {
1992 kmp_internal_control_t new_icvs;
1993 copy_icvs(&new_icvs, &master_th->th.th_current_task->td_icvs);
1994 new_icvs.next = NULL;
1995 if (nthreads_icv > 0) {
1996 new_icvs.nproc = nthreads_icv;
1997 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00001998
1999#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00002000 if (proc_bind_icv != proc_bind_default) {
2001 new_icvs.proc_bind = proc_bind_icv;
2002 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00002003#endif /* OMP_40_ENABLED */
2004
Jonathan Peyton30419822017-05-12 18:01:32 +00002005 /* allocate a new parallel team */
2006 KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n"));
2007 team = __kmp_allocate_team(root, nthreads, nthreads,
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00002008#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00002009 ompt_parallel_id,
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00002010#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00002011#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00002012 proc_bind,
Jim Cownie5e8470a2013-09-27 10:38:44 +00002013#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00002014 &new_icvs, argc USE_NESTED_HOT_ARG(master_th));
Jim Cownie4cc4bb42014-10-07 16:25:50 +00002015 } else {
Jonathan Peyton30419822017-05-12 18:01:32 +00002016 /* allocate a new parallel team */
2017 KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n"));
2018 team = __kmp_allocate_team(root, nthreads, nthreads,
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00002019#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00002020 ompt_parallel_id,
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00002021#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00002022#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00002023 proc_bind,
Jim Cownie5e8470a2013-09-27 10:38:44 +00002024#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00002025 &master_th->th.th_current_task->td_icvs,
2026 argc USE_NESTED_HOT_ARG(master_th));
Jim Cownie5e8470a2013-09-27 10:38:44 +00002027 }
Jonathan Peyton30419822017-05-12 18:01:32 +00002028 KF_TRACE(
2029 10, ("__kmp_fork_call: after __kmp_allocate_team - team = %p\n", team));
Jim Cownie5e8470a2013-09-27 10:38:44 +00002030
2031 /* setup the new team */
Jonathan Peytonb044e4f2016-05-23 18:01:19 +00002032 KMP_CHECK_UPDATE(team->t.t_master_tid, master_tid);
2033 KMP_CHECK_UPDATE(team->t.t_master_this_cons, master_this_cons);
2034 KMP_CHECK_UPDATE(team->t.t_ident, loc);
2035 KMP_CHECK_UPDATE(team->t.t_parent, parent_team);
2036 KMP_CHECK_UPDATE_SYNC(team->t.t_pkfn, microtask);
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00002037#if OMPT_SUPPORT
Jonathan Peytonb044e4f2016-05-23 18:01:19 +00002038 KMP_CHECK_UPDATE_SYNC(team->t.ompt_team_info.microtask, unwrapped_task);
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00002039#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00002040 KMP_CHECK_UPDATE(team->t.t_invoke, invoker); // TODO move to root, maybe
2041// TODO: parent_team->t.t_level == INT_MAX ???
Jim Cownie5e8470a2013-09-27 10:38:44 +00002042#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00002043 if (!master_th->th.th_teams_microtask || level > teams_level) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00002044#endif /* OMP_40_ENABLED */
Jonathan Peyton30419822017-05-12 18:01:32 +00002045 int new_level = parent_team->t.t_level + 1;
2046 KMP_CHECK_UPDATE(team->t.t_level, new_level);
2047 new_level = parent_team->t.t_active_level + 1;
2048 KMP_CHECK_UPDATE(team->t.t_active_level, new_level);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002049#if OMP_40_ENABLED
2050 } else {
Jonathan Peyton30419822017-05-12 18:01:32 +00002051 // AC: Do not increase parallel level at start of the teams construct
2052 int new_level = parent_team->t.t_level;
2053 KMP_CHECK_UPDATE(team->t.t_level, new_level);
2054 new_level = parent_team->t.t_active_level;
2055 KMP_CHECK_UPDATE(team->t.t_active_level, new_level);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002056 }
2057#endif /* OMP_40_ENABLED */
Jonathan Peytonb044e4f2016-05-23 18:01:19 +00002058 kmp_r_sched_t new_sched = get__sched_2(parent_team, master_tid);
Jonathan Peyton30419822017-05-12 18:01:32 +00002059 if (team->t.t_sched.r_sched_type != new_sched.r_sched_type ||
2060 team->t.t_sched.chunk != new_sched.chunk)
2061 team->t.t_sched =
2062 new_sched; // set master's schedule as new run-time schedule
Jim Cownie5e8470a2013-09-27 10:38:44 +00002063
Jonathan Peyton45ca5da2015-10-19 19:33:38 +00002064#if OMP_40_ENABLED
Jonathan Peytonb044e4f2016-05-23 18:01:19 +00002065 KMP_CHECK_UPDATE(team->t.t_cancel_request, cancel_noreq);
Jonathan Peyton45ca5da2015-10-19 19:33:38 +00002066#endif
2067
Jim Cownie4cc4bb42014-10-07 16:25:50 +00002068 // Update the floating point rounding in the team if required.
2069 propagateFPControl(team);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002070
Jonathan Peyton30419822017-05-12 18:01:32 +00002071 if (__kmp_tasking_mode != tskm_immediate_exec) {
2072 // Set master's task team to team's task team. Unless this is hot team, it
2073 // should be NULL.
Jonathan Peyton30419822017-05-12 18:01:32 +00002074 KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
2075 parent_team->t.t_task_team[master_th->th.th_task_state]);
Jonathan Peyton30419822017-05-12 18:01:32 +00002076 KA_TRACE(20, ("__kmp_fork_call: Master T#%d pushing task_team %p / team "
2077 "%p, new task_team %p / team %p\n",
2078 __kmp_gtid_from_thread(master_th),
2079 master_th->th.th_task_team, parent_team,
2080 team->t.t_task_team[master_th->th.th_task_state], team));
Jonathan Peytond3f2b942016-02-09 22:32:41 +00002081
Jonathan Peyton30419822017-05-12 18:01:32 +00002082 if (active_level || master_th->th.th_task_team) {
2083 // Take a memo of master's task_state
2084 KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack);
2085 if (master_th->th.th_task_state_top >=
2086 master_th->th.th_task_state_stack_sz) { // increase size
2087 kmp_uint32 new_size = 2 * master_th->th.th_task_state_stack_sz;
2088 kmp_uint8 *old_stack, *new_stack;
2089 kmp_uint32 i;
2090 new_stack = (kmp_uint8 *)__kmp_allocate(new_size);
2091 for (i = 0; i < master_th->th.th_task_state_stack_sz; ++i) {
2092 new_stack[i] = master_th->th.th_task_state_memo_stack[i];
2093 }
2094 for (i = master_th->th.th_task_state_stack_sz; i < new_size;
2095 ++i) { // zero-init rest of stack
2096 new_stack[i] = 0;
2097 }
2098 old_stack = master_th->th.th_task_state_memo_stack;
2099 master_th->th.th_task_state_memo_stack = new_stack;
2100 master_th->th.th_task_state_stack_sz = new_size;
2101 __kmp_free(old_stack);
Andrey Churbanov6d224db2015-02-10 18:37:43 +00002102 }
Jonathan Peyton30419822017-05-12 18:01:32 +00002103 // Store master's task_state on stack
2104 master_th->th
2105 .th_task_state_memo_stack[master_th->th.th_task_state_top] =
2106 master_th->th.th_task_state;
2107 master_th->th.th_task_state_top++;
2108#if KMP_NESTED_HOT_TEAMS
Jonathan Peyton642688b2017-06-01 16:46:36 +00002109 if (team == master_th->th.th_hot_teams[active_level].hot_team) {
2110 // Restore master's nested state if nested hot team
Jonathan Peyton30419822017-05-12 18:01:32 +00002111 master_th->th.th_task_state =
2112 master_th->th
2113 .th_task_state_memo_stack[master_th->th.th_task_state_top];
2114 } else {
2115#endif
2116 master_th->th.th_task_state = 0;
2117#if KMP_NESTED_HOT_TEAMS
2118 }
2119#endif
2120 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00002121#if !KMP_NESTED_HOT_TEAMS
Jonathan Peyton30419822017-05-12 18:01:32 +00002122 KMP_DEBUG_ASSERT((master_th->th.th_task_team == NULL) ||
2123 (team == root->r.r_hot_team));
Jim Cownie4cc4bb42014-10-07 16:25:50 +00002124#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00002125 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00002126
Jonathan Peyton30419822017-05-12 18:01:32 +00002127 KA_TRACE(
2128 20,
2129 ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n",
2130 gtid, parent_team->t.t_id, team->t.t_master_tid, team->t.t_id,
2131 team->t.t_nproc));
2132 KMP_DEBUG_ASSERT(team != root->r.r_hot_team ||
2133 (team->t.t_master_tid == 0 &&
2134 (team->t.t_parent == root->r.r_root_team ||
2135 team->t.t_parent->t.t_serialized)));
Jim Cownie5e8470a2013-09-27 10:38:44 +00002136 KMP_MB();
2137
2138 /* now, setup the arguments */
Jonathan Peyton30419822017-05-12 18:01:32 +00002139 argv = (void **)team->t.t_argv;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002140#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00002141 if (ap) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00002142#endif /* OMP_40_ENABLED */
Jonathan Peyton30419822017-05-12 18:01:32 +00002143 for (i = argc - 1; i >= 0; --i) {
Jim Cownie4cc4bb42014-10-07 16:25:50 +00002144// TODO: revert workaround for Intel(R) 64 tracker #96
Andrey Churbanovcbda8682015-01-13 14:43:35 +00002145#if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
Jonathan Peyton30419822017-05-12 18:01:32 +00002146 void *new_argv = va_arg(*ap, void *);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002147#else
Jonathan Peyton30419822017-05-12 18:01:32 +00002148 void *new_argv = va_arg(ap, void *);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002149#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00002150 KMP_CHECK_UPDATE(*argv, new_argv);
2151 argv++;
2152 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00002153#if OMP_40_ENABLED
2154 } else {
Jonathan Peyton30419822017-05-12 18:01:32 +00002155 for (i = 0; i < argc; ++i) {
2156 // Get args from parent team for teams construct
2157 KMP_CHECK_UPDATE(argv[i], team->t.t_parent->t.t_argv[i]);
2158 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00002159 }
2160#endif /* OMP_40_ENABLED */
2161
2162 /* now actually fork the threads */
Jonathan Peytonb044e4f2016-05-23 18:01:19 +00002163 KMP_CHECK_UPDATE(team->t.t_master_active, master_active);
Jim Cownie4cc4bb42014-10-07 16:25:50 +00002164 if (!root->r.r_active) // Only do assignment if it prevents cache ping-pong
Jonathan Peyton30419822017-05-12 18:01:32 +00002165 root->r.r_active = TRUE;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002166
Jonathan Peyton30419822017-05-12 18:01:32 +00002167 __kmp_fork_team_threads(root, team, master_th, gtid);
2168 __kmp_setup_icv_copy(team, nthreads,
2169 &master_th->th.th_current_task->td_icvs, loc);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002170
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00002171#if OMPT_SUPPORT
2172 master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
2173#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00002174
Jonathan Peyton30419822017-05-12 18:01:32 +00002175 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002176
Jim Cownie5e8470a2013-09-27 10:38:44 +00002177#if USE_ITT_BUILD
Jonathan Peyton30419822017-05-12 18:01:32 +00002178 if (team->t.t_active_level == 1 // only report frames at level 1
2179#if OMP_40_ENABLED
Andrey Churbanov51aecb82015-05-06 19:22:36 +00002180 && !master_th->th.th_teams_microtask // not in teams construct
Jonathan Peyton30419822017-05-12 18:01:32 +00002181#endif /* OMP_40_ENABLED */
2182 ) {
Andrey Churbanov51aecb82015-05-06 19:22:36 +00002183#if USE_ITT_NOTIFY
Jonathan Peyton30419822017-05-12 18:01:32 +00002184 if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) &&
2185 (__kmp_forkjoin_frames_mode == 3 ||
2186 __kmp_forkjoin_frames_mode == 1)) {
2187 kmp_uint64 tmp_time = 0;
2188 if (__itt_get_timestamp_ptr)
2189 tmp_time = __itt_get_timestamp();
2190 // Internal fork - report frame begin
2191 master_th->th.th_frame_time = tmp_time;
2192 if (__kmp_forkjoin_frames_mode == 3)
2193 team->t.t_region_time = tmp_time;
Jonathan Peyton642688b2017-06-01 16:46:36 +00002194 } else
2195// only one notification scheme (either "submit" or "forking/joined", not both)
Andrey Churbanov51aecb82015-05-06 19:22:36 +00002196#endif /* USE_ITT_NOTIFY */
Jonathan Peyton30419822017-05-12 18:01:32 +00002197 if ((__itt_frame_begin_v3_ptr || KMP_ITT_DEBUG) &&
2198 __kmp_forkjoin_frames && !__kmp_forkjoin_frames_mode) {
2199 // Mark start of "parallel" region for VTune.
2200 __kmp_itt_region_forking(gtid, team->t.t_nproc, 0);
2201 }
Andrey Churbanovf6451d92015-01-16 15:58:03 +00002202 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00002203#endif /* USE_ITT_BUILD */
2204
2205 /* now go on and do the work */
Jonathan Peyton30419822017-05-12 18:01:32 +00002206 KMP_DEBUG_ASSERT(team == __kmp_threads[gtid]->th.th_team);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002207 KMP_MB();
Jonathan Peyton30419822017-05-12 18:01:32 +00002208 KF_TRACE(10,
2209 ("__kmp_internal_fork : root=%p, team=%p, master_th=%p, gtid=%d\n",
2210 root, team, master_th, gtid));
Jim Cownie5e8470a2013-09-27 10:38:44 +00002211
2212#if USE_ITT_BUILD
Jonathan Peyton30419822017-05-12 18:01:32 +00002213 if (__itt_stack_caller_create_ptr) {
2214 team->t.t_stack_id =
2215 __kmp_itt_stack_caller_create(); // create new stack stitching id
2216 // before entering fork barrier
Jim Cownie5e8470a2013-09-27 10:38:44 +00002217 }
2218#endif /* USE_ITT_BUILD */
2219
2220#if OMP_40_ENABLED
Jonathan Peyton642688b2017-06-01 16:46:36 +00002221 // AC: skip __kmp_internal_fork at teams construct, let only master
2222 // threads execute
2223 if (ap)
Jim Cownie5e8470a2013-09-27 10:38:44 +00002224#endif /* OMP_40_ENABLED */
2225 {
Jonathan Peyton30419822017-05-12 18:01:32 +00002226 __kmp_internal_fork(loc, gtid, team);
2227 KF_TRACE(10, ("__kmp_internal_fork : after : root=%p, team=%p, "
2228 "master_th=%p, gtid=%d\n",
2229 root, team, master_th, gtid));
Jim Cownie5e8470a2013-09-27 10:38:44 +00002230 }
2231
Jim Cownie4cc4bb42014-10-07 16:25:50 +00002232 if (call_context == fork_context_gnu) {
Jonathan Peyton30419822017-05-12 18:01:32 +00002233 KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
2234 return TRUE;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002235 }
2236
2237 /* Invoke microtask for MASTER thread */
Jonathan Peyton30419822017-05-12 18:01:32 +00002238 KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n", gtid,
2239 team->t.t_id, team->t.t_pkfn));
2240 } // END of timer KMP_fork_call block
Jim Cownie5e8470a2013-09-27 10:38:44 +00002241
Jonathan Peyton30419822017-05-12 18:01:32 +00002242 {
2243 KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
2244 KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
2245 if (!team->t.t_invoke(gtid)) {
2246 KMP_ASSERT2(0, "cannot invoke microtask for MASTER thread");
Jim Cownie5e8470a2013-09-27 10:38:44 +00002247 }
Jonathan Peyton30419822017-05-12 18:01:32 +00002248 }
2249 KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid,
2250 team->t.t_id, team->t.t_pkfn));
2251 KMP_MB(); /* Flush all pending memory write invalidates. */
Jim Cownie5e8470a2013-09-27 10:38:44 +00002252
Jonathan Peyton30419822017-05-12 18:01:32 +00002253 KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
Jim Cownie5e8470a2013-09-27 10:38:44 +00002254
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00002255#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00002256 if (ompt_enabled) {
2257 master_th->th.ompt_thread_info.state = ompt_state_overhead;
2258 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00002259#endif
2260
Jonathan Peyton30419822017-05-12 18:01:32 +00002261 return TRUE;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002262}
2263
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00002264#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00002265static inline void __kmp_join_restore_state(kmp_info_t *thread,
2266 kmp_team_t *team) {
2267 // restore state outside the region
2268 thread->th.ompt_thread_info.state =
2269 ((team->t.t_serialized) ? ompt_state_work_serial
2270 : ompt_state_work_parallel);
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00002271}
2272
Jonathan Peyton30419822017-05-12 18:01:32 +00002273static inline void __kmp_join_ompt(kmp_info_t *thread, kmp_team_t *team,
2274 ompt_parallel_id_t parallel_id,
2275 fork_context_e fork_context) {
2276 ompt_task_info_t *task_info = __ompt_get_taskinfo(0);
2277 if (ompt_callbacks.ompt_callback(ompt_event_parallel_end)) {
2278 ompt_callbacks.ompt_callback(ompt_event_parallel_end)(
2279 parallel_id, task_info->task_id, OMPT_INVOKER(fork_context));
2280 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00002281
Jonathan Peyton30419822017-05-12 18:01:32 +00002282 task_info->frame.reenter_runtime_frame = NULL;
2283 __kmp_join_restore_state(thread, team);
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00002284}
2285#endif
2286
Jonathan Peyton30419822017-05-12 18:01:32 +00002287void __kmp_join_call(ident_t *loc, int gtid
Jonathan Peytonf89fbbb2015-08-31 18:15:00 +00002288#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00002289 ,
2290 enum fork_context_e fork_context
Jonathan Peytonf89fbbb2015-08-31 18:15:00 +00002291#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00002292#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00002293 ,
2294 int exit_teams
Jim Cownie5e8470a2013-09-27 10:38:44 +00002295#endif /* OMP_40_ENABLED */
Jonathan Peyton30419822017-05-12 18:01:32 +00002296 ) {
2297 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_join_call);
2298 kmp_team_t *team;
2299 kmp_team_t *parent_team;
2300 kmp_info_t *master_th;
2301 kmp_root_t *root;
2302 int master_active;
2303 int i;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002304
Jonathan Peyton30419822017-05-12 18:01:32 +00002305 KA_TRACE(20, ("__kmp_join_call: enter T#%d\n", gtid));
Jim Cownie5e8470a2013-09-27 10:38:44 +00002306
Jonathan Peyton30419822017-05-12 18:01:32 +00002307 /* setup current data */
2308 master_th = __kmp_threads[gtid];
2309 root = master_th->th.th_root;
2310 team = master_th->th.th_team;
2311 parent_team = team->t.t_parent;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002312
Jonathan Peyton30419822017-05-12 18:01:32 +00002313 master_th->th.th_ident = loc;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002314
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00002315#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00002316 if (ompt_enabled) {
2317 master_th->th.ompt_thread_info.state = ompt_state_overhead;
2318 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00002319#endif
2320
Jim Cownie4cc4bb42014-10-07 16:25:50 +00002321#if KMP_DEBUG
Jonathan Peyton30419822017-05-12 18:01:32 +00002322 if (__kmp_tasking_mode != tskm_immediate_exec && !exit_teams) {
2323 KA_TRACE(20, ("__kmp_join_call: T#%d, old team = %p old task_team = %p, "
2324 "th_task_team = %p\n",
2325 __kmp_gtid_from_thread(master_th), team,
2326 team->t.t_task_team[master_th->th.th_task_state],
2327 master_th->th.th_task_team));
2328 KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
2329 team->t.t_task_team[master_th->th.th_task_state]);
2330 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00002331#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00002332
Jonathan Peyton30419822017-05-12 18:01:32 +00002333 if (team->t.t_serialized) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00002334#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00002335 if (master_th->th.th_teams_microtask) {
2336 // We are in teams construct
2337 int level = team->t.t_level;
2338 int tlevel = master_th->th.th_teams_level;
2339 if (level == tlevel) {
2340 // AC: we haven't incremented it earlier at start of teams construct,
2341 // so do it here - at the end of teams construct
2342 team->t.t_level++;
2343 } else if (level == tlevel + 1) {
2344 // AC: we are exiting parallel inside teams, need to increment
2345 // serialization in order to restore it in the next call to
2346 // __kmpc_end_serialized_parallel
2347 team->t.t_serialized++;
2348 }
Andrey Churbanov6d224db2015-02-10 18:37:43 +00002349 }
Jonathan Peyton441f3372015-09-21 17:24:46 +00002350#endif /* OMP_40_ENABLED */
Jonathan Peyton30419822017-05-12 18:01:32 +00002351 __kmpc_end_serialized_parallel(loc, gtid);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002352
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00002353#if OMPT_SUPPORT
Jonathan Peytonb68a85d2015-09-21 18:11:22 +00002354 if (ompt_enabled) {
Jonathan Peyton30419822017-05-12 18:01:32 +00002355 __kmp_join_restore_state(master_th, parent_team);
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00002356 }
2357#endif
2358
Jonathan Peyton30419822017-05-12 18:01:32 +00002359 return;
2360 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00002361
Jonathan Peyton30419822017-05-12 18:01:32 +00002362 master_active = team->t.t_master_active;
2363
2364#if OMP_40_ENABLED
2365 if (!exit_teams)
2366#endif /* OMP_40_ENABLED */
2367 {
2368 // AC: No barrier for internal teams at exit from teams construct.
2369 // But there is barrier for external team (league).
2370 __kmp_internal_join(loc, gtid, team);
2371 }
2372#if OMP_40_ENABLED
2373 else {
2374 master_th->th.th_task_state =
2375 0; // AC: no tasking in teams (out of any parallel)
2376 }
2377#endif /* OMP_40_ENABLED */
2378
2379 KMP_MB();
2380
2381#if OMPT_SUPPORT
2382 ompt_parallel_id_t parallel_id = team->t.ompt_team_info.parallel_id;
2383#endif
2384
2385#if USE_ITT_BUILD
2386 if (__itt_stack_caller_create_ptr) {
2387 __kmp_itt_stack_caller_destroy(
2388 (__itt_caller)team->t
2389 .t_stack_id); // destroy the stack stitching id after join barrier
2390 }
2391
2392 // Mark end of "parallel" region for VTune.
2393 if (team->t.t_active_level == 1
2394#if OMP_40_ENABLED
2395 && !master_th->th.th_teams_microtask /* not in teams construct */
2396#endif /* OMP_40_ENABLED */
2397 ) {
2398 master_th->th.th_ident = loc;
2399 // only one notification scheme (either "submit" or "forking/joined", not
2400 // both)
2401 if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) &&
2402 __kmp_forkjoin_frames_mode == 3)
2403 __kmp_itt_frame_submit(gtid, team->t.t_region_time,
2404 master_th->th.th_frame_time, 0, loc,
2405 master_th->th.th_team_nproc, 1);
2406 else if ((__itt_frame_end_v3_ptr || KMP_ITT_DEBUG) &&
2407 !__kmp_forkjoin_frames_mode && __kmp_forkjoin_frames)
2408 __kmp_itt_region_joined(gtid);
2409 } // active_level == 1
2410#endif /* USE_ITT_BUILD */
2411
2412#if OMP_40_ENABLED
2413 if (master_th->th.th_teams_microtask && !exit_teams &&
2414 team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
2415 team->t.t_level == master_th->th.th_teams_level + 1) {
2416 // AC: We need to leave the team structure intact at the end of parallel
2417 // inside the teams construct, so that at the next parallel same (hot) team
2418 // works, only adjust nesting levels
2419
2420 /* Decrement our nested depth level */
2421 team->t.t_level--;
2422 team->t.t_active_level--;
2423 KMP_TEST_THEN_DEC32((kmp_int32 *)&root->r.r_in_parallel);
2424
2425 /* Restore number of threads in the team if needed */
2426 if (master_th->th.th_team_nproc < master_th->th.th_teams_size.nth) {
2427 int old_num = master_th->th.th_team_nproc;
2428 int new_num = master_th->th.th_teams_size.nth;
2429 kmp_info_t **other_threads = team->t.t_threads;
2430 team->t.t_nproc = new_num;
2431 for (i = 0; i < old_num; ++i) {
2432 other_threads[i]->th.th_team_nproc = new_num;
2433 }
2434 // Adjust states of non-used threads of the team
2435 for (i = old_num; i < new_num; ++i) {
2436 // Re-initialize thread's barrier data.
2437 int b;
2438 kmp_balign_t *balign = other_threads[i]->th.th_bar;
2439 for (b = 0; b < bs_last_barrier; ++b) {
2440 balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
2441 KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
2442#if USE_DEBUGGER
2443 balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
2444#endif
2445 }
2446 if (__kmp_tasking_mode != tskm_immediate_exec) {
2447 // Synchronize thread's task state
2448 other_threads[i]->th.th_task_state = master_th->th.th_task_state;
2449 }
2450 }
2451 }
2452
2453#if OMPT_SUPPORT
2454 if (ompt_enabled) {
2455 __kmp_join_ompt(master_th, parent_team, parallel_id, fork_context);
2456 }
2457#endif
2458
2459 return;
2460 }
2461#endif /* OMP_40_ENABLED */
2462
2463 /* do cleanup and restore the parent team */
2464 master_th->th.th_info.ds.ds_tid = team->t.t_master_tid;
2465 master_th->th.th_local.this_construct = team->t.t_master_this_cons;
2466
2467 master_th->th.th_dispatch = &parent_team->t.t_dispatch[team->t.t_master_tid];
2468
2469 /* jc: The following lock has instructions with REL and ACQ semantics,
2470 separating the parallel user code called in this parallel region
2471 from the serial user code called after this function returns. */
2472 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
2473
2474#if OMP_40_ENABLED
2475 if (!master_th->th.th_teams_microtask ||
2476 team->t.t_level > master_th->th.th_teams_level)
2477#endif /* OMP_40_ENABLED */
2478 {
2479 /* Decrement our nested depth level */
2480 KMP_TEST_THEN_DEC32((kmp_int32 *)&root->r.r_in_parallel);
2481 }
2482 KMP_DEBUG_ASSERT(root->r.r_in_parallel >= 0);
2483
2484#if OMPT_SUPPORT && OMPT_TRACE
2485 if (ompt_enabled) {
2486 ompt_task_info_t *task_info = __ompt_get_taskinfo(0);
2487 if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) {
2488 ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)(
2489 parallel_id, task_info->task_id);
2490 }
2491 task_info->frame.exit_runtime_frame = NULL;
2492 task_info->task_id = 0;
2493 }
2494#endif
2495
2496 KF_TRACE(10, ("__kmp_join_call1: T#%d, this_thread=%p team=%p\n", 0,
2497 master_th, team));
2498 __kmp_pop_current_task_from_thread(master_th);
2499
2500#if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
2501 // Restore master thread's partition.
2502 master_th->th.th_first_place = team->t.t_first_place;
2503 master_th->th.th_last_place = team->t.t_last_place;
2504#endif /* OMP_40_ENABLED */
2505
2506 updateHWFPControl(team);
2507
2508 if (root->r.r_active != master_active)
2509 root->r.r_active = master_active;
2510
2511 __kmp_free_team(root, team USE_NESTED_HOT_ARG(
2512 master_th)); // this will free worker threads
2513
2514 /* this race was fun to find. make sure the following is in the critical
2515 region otherwise assertions may fail occasionally since the old team may be
2516 reallocated and the hierarchy appears inconsistent. it is actually safe to
2517 run and won't cause any bugs, but will cause those assertion failures. it's
2518 only one deref&assign so might as well put this in the critical region */
2519 master_th->th.th_team = parent_team;
2520 master_th->th.th_team_nproc = parent_team->t.t_nproc;
2521 master_th->th.th_team_master = parent_team->t.t_threads[0];
2522 master_th->th.th_team_serialized = parent_team->t.t_serialized;
2523
2524 /* restore serialized team, if need be */
2525 if (parent_team->t.t_serialized &&
2526 parent_team != master_th->th.th_serial_team &&
2527 parent_team != root->r.r_root_team) {
2528 __kmp_free_team(root,
2529 master_th->th.th_serial_team USE_NESTED_HOT_ARG(NULL));
2530 master_th->th.th_serial_team = parent_team;
2531 }
2532
2533 if (__kmp_tasking_mode != tskm_immediate_exec) {
2534 if (master_th->th.th_task_state_top >
2535 0) { // Restore task state from memo stack
2536 KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack);
2537 // Remember master's state if we re-use this nested hot team
2538 master_th->th.th_task_state_memo_stack[master_th->th.th_task_state_top] =
2539 master_th->th.th_task_state;
2540 --master_th->th.th_task_state_top; // pop
2541 // Now restore state at this level
2542 master_th->th.th_task_state =
2543 master_th->th
2544 .th_task_state_memo_stack[master_th->th.th_task_state_top];
2545 }
2546 // Copy the task team from the parent team to the master thread
2547 master_th->th.th_task_team =
2548 parent_team->t.t_task_team[master_th->th.th_task_state];
2549 KA_TRACE(20,
2550 ("__kmp_join_call: Master T#%d restoring task_team %p / team %p\n",
2551 __kmp_gtid_from_thread(master_th), master_th->th.th_task_team,
2552 parent_team));
2553 }
2554
2555 // TODO: GEH - cannot do this assertion because root thread not set up as
2556 // executing
2557 // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 0 );
2558 master_th->th.th_current_task->td_flags.executing = 1;
2559
2560 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2561
2562#if OMPT_SUPPORT
2563 if (ompt_enabled) {
2564 __kmp_join_ompt(master_th, parent_team, parallel_id, fork_context);
2565 }
2566#endif
2567
2568 KMP_MB();
2569 KA_TRACE(20, ("__kmp_join_call: exit T#%d\n", gtid));
2570}
Jim Cownie5e8470a2013-09-27 10:38:44 +00002571
2572/* Check whether we should push an internal control record onto the
2573 serial team stack. If so, do it. */
Jonathan Peyton30419822017-05-12 18:01:32 +00002574void __kmp_save_internal_controls(kmp_info_t *thread) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00002575
Jonathan Peyton30419822017-05-12 18:01:32 +00002576 if (thread->th.th_team != thread->th.th_serial_team) {
2577 return;
2578 }
2579 if (thread->th.th_team->t.t_serialized > 1) {
2580 int push = 0;
2581
2582 if (thread->th.th_team->t.t_control_stack_top == NULL) {
2583 push = 1;
2584 } else {
2585 if (thread->th.th_team->t.t_control_stack_top->serial_nesting_level !=
2586 thread->th.th_team->t.t_serialized) {
2587 push = 1;
2588 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00002589 }
Jonathan Peyton30419822017-05-12 18:01:32 +00002590 if (push) { /* push a record on the serial team's stack */
2591 kmp_internal_control_t *control =
2592 (kmp_internal_control_t *)__kmp_allocate(
2593 sizeof(kmp_internal_control_t));
Jim Cownie5e8470a2013-09-27 10:38:44 +00002594
Jonathan Peyton30419822017-05-12 18:01:32 +00002595 copy_icvs(control, &thread->th.th_current_task->td_icvs);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002596
Jonathan Peyton30419822017-05-12 18:01:32 +00002597 control->serial_nesting_level = thread->th.th_team->t.t_serialized;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002598
Jonathan Peyton30419822017-05-12 18:01:32 +00002599 control->next = thread->th.th_team->t.t_control_stack_top;
2600 thread->th.th_team->t.t_control_stack_top = control;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002601 }
Jonathan Peyton30419822017-05-12 18:01:32 +00002602 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00002603}
2604
2605/* Changes set_nproc */
Jonathan Peyton30419822017-05-12 18:01:32 +00002606void __kmp_set_num_threads(int new_nth, int gtid) {
2607 kmp_info_t *thread;
2608 kmp_root_t *root;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002609
Jonathan Peyton30419822017-05-12 18:01:32 +00002610 KF_TRACE(10, ("__kmp_set_num_threads: new __kmp_nth = %d\n", new_nth));
2611 KMP_DEBUG_ASSERT(__kmp_init_serial);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002612
Jonathan Peyton30419822017-05-12 18:01:32 +00002613 if (new_nth < 1)
2614 new_nth = 1;
2615 else if (new_nth > __kmp_max_nth)
2616 new_nth = __kmp_max_nth;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002617
Jonathan Peyton30419822017-05-12 18:01:32 +00002618 KMP_COUNT_VALUE(OMP_set_numthreads, new_nth);
2619 thread = __kmp_threads[gtid];
Jim Cownie5e8470a2013-09-27 10:38:44 +00002620
Jonathan Peyton30419822017-05-12 18:01:32 +00002621 __kmp_save_internal_controls(thread);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002622
Jonathan Peyton30419822017-05-12 18:01:32 +00002623 set__nproc(thread, new_nth);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002624
Jonathan Peyton30419822017-05-12 18:01:32 +00002625 // If this omp_set_num_threads() call will cause the hot team size to be
2626 // reduced (in the absence of a num_threads clause), then reduce it now,
2627 // rather than waiting for the next parallel region.
2628 root = thread->th.th_root;
2629 if (__kmp_init_parallel && (!root->r.r_active) &&
2630 (root->r.r_hot_team->t.t_nproc > new_nth)
Jim Cownie4cc4bb42014-10-07 16:25:50 +00002631#if KMP_NESTED_HOT_TEAMS
2632 && __kmp_hot_teams_max_level && !__kmp_hot_teams_mode
2633#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00002634 ) {
2635 kmp_team_t *hot_team = root->r.r_hot_team;
2636 int f;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002637
Jonathan Peyton30419822017-05-12 18:01:32 +00002638 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002639
Jonathan Peyton30419822017-05-12 18:01:32 +00002640 // Release the extra threads we don't need any more.
2641 for (f = new_nth; f < hot_team->t.t_nproc; f++) {
2642 KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL);
2643 if (__kmp_tasking_mode != tskm_immediate_exec) {
2644 // When decreasing team size, threads no longer in the team should unref
2645 // task team.
2646 hot_team->t.t_threads[f]->th.th_task_team = NULL;
2647 }
2648 __kmp_free_thread(hot_team->t.t_threads[f]);
2649 hot_team->t.t_threads[f] = NULL;
2650 }
2651 hot_team->t.t_nproc = new_nth;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00002652#if KMP_NESTED_HOT_TEAMS
Jonathan Peyton30419822017-05-12 18:01:32 +00002653 if (thread->th.th_hot_teams) {
2654 KMP_DEBUG_ASSERT(hot_team == thread->th.th_hot_teams[0].hot_team);
2655 thread->th.th_hot_teams[0].hot_team_nth = new_nth;
2656 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00002657#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00002658
Jonathan Peyton30419822017-05-12 18:01:32 +00002659 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002660
Jonathan Peyton30419822017-05-12 18:01:32 +00002661 // Update the t_nproc field in the threads that are still active.
2662 for (f = 0; f < new_nth; f++) {
2663 KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL);
2664 hot_team->t.t_threads[f]->th.th_team_nproc = new_nth;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002665 }
Jonathan Peyton30419822017-05-12 18:01:32 +00002666 // Special flag in case omp_set_num_threads() call
2667 hot_team->t.t_size_changed = -1;
2668 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00002669}
2670
Jim Cownie5e8470a2013-09-27 10:38:44 +00002671/* Changes max_active_levels */
Jonathan Peyton30419822017-05-12 18:01:32 +00002672void __kmp_set_max_active_levels(int gtid, int max_active_levels) {
2673 kmp_info_t *thread;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002674
Jonathan Peyton30419822017-05-12 18:01:32 +00002675 KF_TRACE(10, ("__kmp_set_max_active_levels: new max_active_levels for thread "
2676 "%d = (%d)\n",
2677 gtid, max_active_levels));
2678 KMP_DEBUG_ASSERT(__kmp_init_serial);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002679
Jonathan Peyton30419822017-05-12 18:01:32 +00002680 // validate max_active_levels
2681 if (max_active_levels < 0) {
2682 KMP_WARNING(ActiveLevelsNegative, max_active_levels);
2683 // We ignore this call if the user has specified a negative value.
2684 // The current setting won't be changed. The last valid setting will be
2685 // used. A warning will be issued (if warnings are allowed as controlled by
2686 // the KMP_WARNINGS env var).
2687 KF_TRACE(10, ("__kmp_set_max_active_levels: the call is ignored: new "
2688 "max_active_levels for thread %d = (%d)\n",
2689 gtid, max_active_levels));
2690 return;
2691 }
2692 if (max_active_levels <= KMP_MAX_ACTIVE_LEVELS_LIMIT) {
2693 // it's OK, the max_active_levels is within the valid range: [ 0;
2694 // KMP_MAX_ACTIVE_LEVELS_LIMIT ]
2695 // We allow a zero value. (implementation defined behavior)
2696 } else {
2697 KMP_WARNING(ActiveLevelsExceedLimit, max_active_levels,
2698 KMP_MAX_ACTIVE_LEVELS_LIMIT);
2699 max_active_levels = KMP_MAX_ACTIVE_LEVELS_LIMIT;
2700 // Current upper limit is MAX_INT. (implementation defined behavior)
2701 // If the input exceeds the upper limit, we correct the input to be the
2702 // upper limit. (implementation defined behavior)
2703 // Actually, the flow should never get here until we use MAX_INT limit.
2704 }
2705 KF_TRACE(10, ("__kmp_set_max_active_levels: after validation: new "
2706 "max_active_levels for thread %d = (%d)\n",
2707 gtid, max_active_levels));
Jim Cownie5e8470a2013-09-27 10:38:44 +00002708
Jonathan Peyton30419822017-05-12 18:01:32 +00002709 thread = __kmp_threads[gtid];
Jim Cownie5e8470a2013-09-27 10:38:44 +00002710
Jonathan Peyton30419822017-05-12 18:01:32 +00002711 __kmp_save_internal_controls(thread);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002712
Jonathan Peyton30419822017-05-12 18:01:32 +00002713 set__max_active_levels(thread, max_active_levels);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002714}
2715
2716/* Gets max_active_levels */
Jonathan Peyton30419822017-05-12 18:01:32 +00002717int __kmp_get_max_active_levels(int gtid) {
2718 kmp_info_t *thread;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002719
Jonathan Peyton30419822017-05-12 18:01:32 +00002720 KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d\n", gtid));
2721 KMP_DEBUG_ASSERT(__kmp_init_serial);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002722
Jonathan Peyton30419822017-05-12 18:01:32 +00002723 thread = __kmp_threads[gtid];
2724 KMP_DEBUG_ASSERT(thread->th.th_current_task);
2725 KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d, curtask=%p, "
2726 "curtask_maxaclevel=%d\n",
2727 gtid, thread->th.th_current_task,
2728 thread->th.th_current_task->td_icvs.max_active_levels));
2729 return thread->th.th_current_task->td_icvs.max_active_levels;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002730}
2731
2732/* Changes def_sched_var ICV values (run-time schedule kind and chunk) */
Jonathan Peyton30419822017-05-12 18:01:32 +00002733void __kmp_set_schedule(int gtid, kmp_sched_t kind, int chunk) {
2734 kmp_info_t *thread;
2735 // kmp_team_t *team;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002736
Jonathan Peyton30419822017-05-12 18:01:32 +00002737 KF_TRACE(10, ("__kmp_set_schedule: new schedule for thread %d = (%d, %d)\n",
2738 gtid, (int)kind, chunk));
2739 KMP_DEBUG_ASSERT(__kmp_init_serial);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002740
Jonathan Peyton30419822017-05-12 18:01:32 +00002741 // Check if the kind parameter is valid, correct if needed.
2742 // Valid parameters should fit in one of two intervals - standard or extended:
2743 // <lower>, <valid>, <upper_std>, <lower_ext>, <valid>, <upper>
2744 // 2008-01-25: 0, 1 - 4, 5, 100, 101 - 102, 103
2745 if (kind <= kmp_sched_lower || kind >= kmp_sched_upper ||
2746 (kind <= kmp_sched_lower_ext && kind >= kmp_sched_upper_std)) {
2747 // TODO: Hint needs attention in case we change the default schedule.
2748 __kmp_msg(kmp_ms_warning, KMP_MSG(ScheduleKindOutOfRange, kind),
2749 KMP_HNT(DefaultScheduleKindUsed, "static, no chunk"),
2750 __kmp_msg_null);
2751 kind = kmp_sched_default;
2752 chunk = 0; // ignore chunk value in case of bad kind
2753 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00002754
Jonathan Peyton30419822017-05-12 18:01:32 +00002755 thread = __kmp_threads[gtid];
Jim Cownie5e8470a2013-09-27 10:38:44 +00002756
Jonathan Peyton30419822017-05-12 18:01:32 +00002757 __kmp_save_internal_controls(thread);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002758
Jonathan Peyton30419822017-05-12 18:01:32 +00002759 if (kind < kmp_sched_upper_std) {
2760 if (kind == kmp_sched_static && chunk < KMP_DEFAULT_CHUNK) {
2761 // differ static chunked vs. unchunked: chunk should be invalid to
2762 // indicate unchunked schedule (which is the default)
2763 thread->th.th_current_task->td_icvs.sched.r_sched_type = kmp_sch_static;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002764 } else {
Jonathan Peyton30419822017-05-12 18:01:32 +00002765 thread->th.th_current_task->td_icvs.sched.r_sched_type =
2766 __kmp_sch_map[kind - kmp_sched_lower - 1];
Jim Cownie5e8470a2013-09-27 10:38:44 +00002767 }
Jonathan Peyton30419822017-05-12 18:01:32 +00002768 } else {
2769 // __kmp_sch_map[ kind - kmp_sched_lower_ext + kmp_sched_upper_std -
2770 // kmp_sched_lower - 2 ];
2771 thread->th.th_current_task->td_icvs.sched.r_sched_type =
2772 __kmp_sch_map[kind - kmp_sched_lower_ext + kmp_sched_upper_std -
2773 kmp_sched_lower - 2];
2774 }
Andrey Churbanovd454c732017-06-05 17:17:33 +00002775 if (kind == kmp_sched_auto || chunk < 1) {
Jonathan Peyton30419822017-05-12 18:01:32 +00002776 // ignore parameter chunk for schedule auto
2777 thread->th.th_current_task->td_icvs.sched.chunk = KMP_DEFAULT_CHUNK;
2778 } else {
2779 thread->th.th_current_task->td_icvs.sched.chunk = chunk;
2780 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00002781}
2782
2783/* Gets def_sched_var ICV values */
Jonathan Peyton30419822017-05-12 18:01:32 +00002784void __kmp_get_schedule(int gtid, kmp_sched_t *kind, int *chunk) {
2785 kmp_info_t *thread;
2786 enum sched_type th_type;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002787
Jonathan Peyton30419822017-05-12 18:01:32 +00002788 KF_TRACE(10, ("__kmp_get_schedule: thread %d\n", gtid));
2789 KMP_DEBUG_ASSERT(__kmp_init_serial);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002790
Jonathan Peyton30419822017-05-12 18:01:32 +00002791 thread = __kmp_threads[gtid];
Jim Cownie5e8470a2013-09-27 10:38:44 +00002792
Jonathan Peyton30419822017-05-12 18:01:32 +00002793 th_type = thread->th.th_current_task->td_icvs.sched.r_sched_type;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002794
Jonathan Peyton30419822017-05-12 18:01:32 +00002795 switch (th_type) {
2796 case kmp_sch_static:
2797 case kmp_sch_static_greedy:
2798 case kmp_sch_static_balanced:
2799 *kind = kmp_sched_static;
2800 *chunk = 0; // chunk was not set, try to show this fact via zero value
2801 return;
2802 case kmp_sch_static_chunked:
2803 *kind = kmp_sched_static;
2804 break;
2805 case kmp_sch_dynamic_chunked:
2806 *kind = kmp_sched_dynamic;
2807 break;
2808 case kmp_sch_guided_chunked:
2809 case kmp_sch_guided_iterative_chunked:
2810 case kmp_sch_guided_analytical_chunked:
2811 *kind = kmp_sched_guided;
2812 break;
2813 case kmp_sch_auto:
2814 *kind = kmp_sched_auto;
2815 break;
2816 case kmp_sch_trapezoidal:
2817 *kind = kmp_sched_trapezoidal;
2818 break;
Jonathan Peytona1234cf2016-10-07 18:01:35 +00002819#if KMP_STATIC_STEAL_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00002820 case kmp_sch_static_steal:
2821 *kind = kmp_sched_static_steal;
2822 break;
Jonathan Peytona1234cf2016-10-07 18:01:35 +00002823#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00002824 default:
2825 KMP_FATAL(UnknownSchedulingType, th_type);
2826 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00002827
Jonathan Peyton30419822017-05-12 18:01:32 +00002828 *chunk = thread->th.th_current_task->td_icvs.sched.chunk;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002829}
2830
Jonathan Peyton30419822017-05-12 18:01:32 +00002831int __kmp_get_ancestor_thread_num(int gtid, int level) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00002832
Jonathan Peyton30419822017-05-12 18:01:32 +00002833 int ii, dd;
2834 kmp_team_t *team;
2835 kmp_info_t *thr;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002836
Jonathan Peyton30419822017-05-12 18:01:32 +00002837 KF_TRACE(10, ("__kmp_get_ancestor_thread_num: thread %d %d\n", gtid, level));
2838 KMP_DEBUG_ASSERT(__kmp_init_serial);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002839
Jonathan Peyton30419822017-05-12 18:01:32 +00002840 // validate level
2841 if (level == 0)
2842 return 0;
2843 if (level < 0)
2844 return -1;
2845 thr = __kmp_threads[gtid];
2846 team = thr->th.th_team;
2847 ii = team->t.t_level;
2848 if (level > ii)
2849 return -1;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002850
2851#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00002852 if (thr->th.th_teams_microtask) {
2853 // AC: we are in teams region where multiple nested teams have same level
2854 int tlevel = thr->th.th_teams_level; // the level of the teams construct
2855 if (level <=
2856 tlevel) { // otherwise usual algorithm works (will not touch the teams)
2857 KMP_DEBUG_ASSERT(ii >= tlevel);
2858 // AC: As we need to pass by the teams league, we need to artificially
2859 // increase ii
2860 if (ii == tlevel) {
2861 ii += 2; // three teams have same level
2862 } else {
2863 ii++; // two teams have same level
2864 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00002865 }
Jonathan Peyton30419822017-05-12 18:01:32 +00002866 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00002867#endif
2868
Jonathan Peyton30419822017-05-12 18:01:32 +00002869 if (ii == level)
2870 return __kmp_tid_from_gtid(gtid);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002871
Jonathan Peyton30419822017-05-12 18:01:32 +00002872 dd = team->t.t_serialized;
2873 level++;
2874 while (ii > level) {
2875 for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00002876 }
Jonathan Peyton30419822017-05-12 18:01:32 +00002877 if ((team->t.t_serialized) && (!dd)) {
2878 team = team->t.t_parent;
2879 continue;
2880 }
2881 if (ii > level) {
2882 team = team->t.t_parent;
2883 dd = team->t.t_serialized;
2884 ii--;
2885 }
2886 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00002887
Jonathan Peyton30419822017-05-12 18:01:32 +00002888 return (dd > 1) ? (0) : (team->t.t_master_tid);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002889}
2890
Jonathan Peyton30419822017-05-12 18:01:32 +00002891int __kmp_get_team_size(int gtid, int level) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00002892
Jonathan Peyton30419822017-05-12 18:01:32 +00002893 int ii, dd;
2894 kmp_team_t *team;
2895 kmp_info_t *thr;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002896
Jonathan Peyton30419822017-05-12 18:01:32 +00002897 KF_TRACE(10, ("__kmp_get_team_size: thread %d %d\n", gtid, level));
2898 KMP_DEBUG_ASSERT(__kmp_init_serial);
Jim Cownie5e8470a2013-09-27 10:38:44 +00002899
Jonathan Peyton30419822017-05-12 18:01:32 +00002900 // validate level
2901 if (level == 0)
2902 return 1;
2903 if (level < 0)
2904 return -1;
2905 thr = __kmp_threads[gtid];
2906 team = thr->th.th_team;
2907 ii = team->t.t_level;
2908 if (level > ii)
2909 return -1;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002910
2911#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00002912 if (thr->th.th_teams_microtask) {
2913 // AC: we are in teams region where multiple nested teams have same level
2914 int tlevel = thr->th.th_teams_level; // the level of the teams construct
2915 if (level <=
2916 tlevel) { // otherwise usual algorithm works (will not touch the teams)
2917 KMP_DEBUG_ASSERT(ii >= tlevel);
2918 // AC: As we need to pass by the teams league, we need to artificially
2919 // increase ii
2920 if (ii == tlevel) {
2921 ii += 2; // three teams have same level
2922 } else {
2923 ii++; // two teams have same level
2924 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00002925 }
Jonathan Peyton30419822017-05-12 18:01:32 +00002926 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00002927#endif
2928
Jonathan Peyton30419822017-05-12 18:01:32 +00002929 while (ii > level) {
2930 for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00002931 }
Jonathan Peyton30419822017-05-12 18:01:32 +00002932 if (team->t.t_serialized && (!dd)) {
2933 team = team->t.t_parent;
2934 continue;
2935 }
2936 if (ii > level) {
2937 team = team->t.t_parent;
2938 ii--;
2939 }
2940 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00002941
Jonathan Peyton30419822017-05-12 18:01:32 +00002942 return team->t.t_nproc;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002943}
2944
Jonathan Peyton30419822017-05-12 18:01:32 +00002945kmp_r_sched_t __kmp_get_schedule_global() {
2946 // This routine created because pairs (__kmp_sched, __kmp_chunk) and
2947 // (__kmp_static, __kmp_guided) may be changed by kmp_set_defaults
2948 // independently. So one can get the updated schedule here.
Jim Cownie5e8470a2013-09-27 10:38:44 +00002949
Jonathan Peyton30419822017-05-12 18:01:32 +00002950 kmp_r_sched_t r_sched;
Jim Cownie5e8470a2013-09-27 10:38:44 +00002951
Jonathan Peyton30419822017-05-12 18:01:32 +00002952 // create schedule from 4 globals: __kmp_sched, __kmp_chunk, __kmp_static,
2953 // __kmp_guided. __kmp_sched should keep original value, so that user can set
2954 // KMP_SCHEDULE multiple times, and thus have different run-time schedules in
2955 // different roots (even in OMP 2.5)
2956 if (__kmp_sched == kmp_sch_static) {
2957 r_sched.r_sched_type = __kmp_static; // replace STATIC with more detailed
2958 // schedule (balanced or greedy)
2959 } else if (__kmp_sched == kmp_sch_guided_chunked) {
2960 r_sched.r_sched_type = __kmp_guided; // replace GUIDED with more detailed
2961 // schedule (iterative or analytical)
2962 } else {
2963 r_sched.r_sched_type =
2964 __kmp_sched; // (STATIC_CHUNKED), or (DYNAMIC_CHUNKED), or other
2965 }
2966
2967 if (__kmp_chunk < KMP_DEFAULT_CHUNK) { // __kmp_chunk may be wrong here (if it
2968 // was not ever set)
2969 r_sched.chunk = KMP_DEFAULT_CHUNK;
2970 } else {
2971 r_sched.chunk = __kmp_chunk;
2972 }
2973
2974 return r_sched;
2975}
2976
2977/* Allocate (realloc == FALSE) * or reallocate (realloc == TRUE)
2978 at least argc number of *t_argv entries for the requested team. */
2979static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team, int realloc) {
2980
2981 KMP_DEBUG_ASSERT(team);
2982 if (!realloc || argc > team->t.t_max_argc) {
2983
2984 KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: needed entries=%d, "
2985 "current entries=%d\n",
2986 team->t.t_id, argc, (realloc) ? team->t.t_max_argc : 0));
2987 /* if previously allocated heap space for args, free them */
2988 if (realloc && team->t.t_argv != &team->t.t_inline_argv[0])
2989 __kmp_free((void *)team->t.t_argv);
2990
2991 if (argc <= KMP_INLINE_ARGV_ENTRIES) {
2992 /* use unused space in the cache line for arguments */
2993 team->t.t_max_argc = KMP_INLINE_ARGV_ENTRIES;
2994 KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: inline allocate %d "
2995 "argv entries\n",
2996 team->t.t_id, team->t.t_max_argc));
2997 team->t.t_argv = &team->t.t_inline_argv[0];
2998 if (__kmp_storage_map) {
2999 __kmp_print_storage_map_gtid(
3000 -1, &team->t.t_inline_argv[0],
3001 &team->t.t_inline_argv[KMP_INLINE_ARGV_ENTRIES],
3002 (sizeof(void *) * KMP_INLINE_ARGV_ENTRIES), "team_%d.t_inline_argv",
3003 team->t.t_id);
3004 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00003005 } else {
Jonathan Peyton30419822017-05-12 18:01:32 +00003006 /* allocate space for arguments in the heap */
3007 team->t.t_max_argc = (argc <= (KMP_MIN_MALLOC_ARGV_ENTRIES >> 1))
3008 ? KMP_MIN_MALLOC_ARGV_ENTRIES
3009 : 2 * argc;
3010 KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: dynamic allocate %d "
3011 "argv entries\n",
3012 team->t.t_id, team->t.t_max_argc));
3013 team->t.t_argv =
3014 (void **)__kmp_page_allocate(sizeof(void *) * team->t.t_max_argc);
3015 if (__kmp_storage_map) {
3016 __kmp_print_storage_map_gtid(-1, &team->t.t_argv[0],
3017 &team->t.t_argv[team->t.t_max_argc],
3018 sizeof(void *) * team->t.t_max_argc,
3019 "team_%d.t_argv", team->t.t_id);
3020 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00003021 }
Jonathan Peyton30419822017-05-12 18:01:32 +00003022 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00003023}
3024
Jonathan Peyton30419822017-05-12 18:01:32 +00003025static void __kmp_allocate_team_arrays(kmp_team_t *team, int max_nth) {
3026 int i;
3027 int num_disp_buff = max_nth > 1 ? __kmp_dispatch_num_buffers : 2;
3028 team->t.t_threads =
3029 (kmp_info_t **)__kmp_allocate(sizeof(kmp_info_t *) * max_nth);
3030 team->t.t_disp_buffer = (dispatch_shared_info_t *)__kmp_allocate(
3031 sizeof(dispatch_shared_info_t) * num_disp_buff);
3032 team->t.t_dispatch =
3033 (kmp_disp_t *)__kmp_allocate(sizeof(kmp_disp_t) * max_nth);
3034 team->t.t_implicit_task_taskdata =
3035 (kmp_taskdata_t *)__kmp_allocate(sizeof(kmp_taskdata_t) * max_nth);
3036 team->t.t_max_nproc = max_nth;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003037
Jonathan Peyton30419822017-05-12 18:01:32 +00003038 /* setup dispatch buffers */
3039 for (i = 0; i < num_disp_buff; ++i) {
3040 team->t.t_disp_buffer[i].buffer_index = i;
Jonathan Peytondf6818b2016-06-14 17:57:47 +00003041#if OMP_45_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00003042 team->t.t_disp_buffer[i].doacross_buf_idx = i;
Jonathan Peyton71909c52016-03-02 22:42:06 +00003043#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00003044 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00003045}
3046
Jonathan Peyton30419822017-05-12 18:01:32 +00003047static void __kmp_free_team_arrays(kmp_team_t *team) {
3048 /* Note: this does not free the threads in t_threads (__kmp_free_threads) */
3049 int i;
3050 for (i = 0; i < team->t.t_max_nproc; ++i) {
3051 if (team->t.t_dispatch[i].th_disp_buffer != NULL) {
3052 __kmp_free(team->t.t_dispatch[i].th_disp_buffer);
3053 team->t.t_dispatch[i].th_disp_buffer = NULL;
3054 }; // if
3055 }; // for
3056 __kmp_free(team->t.t_threads);
3057 __kmp_free(team->t.t_disp_buffer);
3058 __kmp_free(team->t.t_dispatch);
3059 __kmp_free(team->t.t_implicit_task_taskdata);
3060 team->t.t_threads = NULL;
3061 team->t.t_disp_buffer = NULL;
3062 team->t.t_dispatch = NULL;
3063 team->t.t_implicit_task_taskdata = 0;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003064}
3065
Jonathan Peyton30419822017-05-12 18:01:32 +00003066static void __kmp_reallocate_team_arrays(kmp_team_t *team, int max_nth) {
3067 kmp_info_t **oldThreads = team->t.t_threads;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003068
Jonathan Peyton30419822017-05-12 18:01:32 +00003069 __kmp_free(team->t.t_disp_buffer);
3070 __kmp_free(team->t.t_dispatch);
3071 __kmp_free(team->t.t_implicit_task_taskdata);
3072 __kmp_allocate_team_arrays(team, max_nth);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003073
Jonathan Peyton30419822017-05-12 18:01:32 +00003074 KMP_MEMCPY(team->t.t_threads, oldThreads,
3075 team->t.t_nproc * sizeof(kmp_info_t *));
Jim Cownie5e8470a2013-09-27 10:38:44 +00003076
Jonathan Peyton30419822017-05-12 18:01:32 +00003077 __kmp_free(oldThreads);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003078}
3079
Jonathan Peyton30419822017-05-12 18:01:32 +00003080static kmp_internal_control_t __kmp_get_global_icvs(void) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00003081
Jonathan Peyton30419822017-05-12 18:01:32 +00003082 kmp_r_sched_t r_sched =
3083 __kmp_get_schedule_global(); // get current state of scheduling globals
Jim Cownie5e8470a2013-09-27 10:38:44 +00003084
3085#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00003086 KMP_DEBUG_ASSERT(__kmp_nested_proc_bind.used > 0);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003087#endif /* OMP_40_ENABLED */
3088
Jonathan Peyton30419822017-05-12 18:01:32 +00003089 kmp_internal_control_t g_icvs = {
3090 0, // int serial_nesting_level; //corresponds to value of th_team_serialized
3091 (kmp_int8)__kmp_dflt_nested, // int nested; //internal control
3092 // for nested parallelism (per thread)
3093 (kmp_int8)__kmp_global.g.g_dynamic, // internal control for dynamic
3094 // adjustment of threads (per thread)
3095 (kmp_int8)__kmp_env_blocktime, // int bt_set; //internal control for
3096 // whether blocktime is explicitly set
3097 __kmp_dflt_blocktime, // int blocktime; //internal control for blocktime
Jonathan Peytone1c7c132016-10-07 18:12:19 +00003098#if KMP_USE_MONITOR
Jonathan Peyton30419822017-05-12 18:01:32 +00003099 __kmp_bt_intervals, // int bt_intervals; //internal control for blocktime
3100// intervals
Jonathan Peytone1c7c132016-10-07 18:12:19 +00003101#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00003102 __kmp_dflt_team_nth, // int nproc; //internal control for # of threads for
3103 // next parallel region (per thread)
3104 // (use a max ub on value if __kmp_parallel_initialize not called yet)
3105 __kmp_dflt_max_active_levels, // int max_active_levels; //internal control
3106 // for max_active_levels
3107 r_sched, // kmp_r_sched_t sched; //internal control for runtime schedule
3108// {sched,chunk} pair
Jim Cownie5e8470a2013-09-27 10:38:44 +00003109#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00003110 __kmp_nested_proc_bind.bind_types[0],
3111 __kmp_default_device,
Jim Cownie5e8470a2013-09-27 10:38:44 +00003112#endif /* OMP_40_ENABLED */
Jonathan Peyton30419822017-05-12 18:01:32 +00003113 NULL // struct kmp_internal_control *next;
3114 };
Jim Cownie5e8470a2013-09-27 10:38:44 +00003115
Jonathan Peyton30419822017-05-12 18:01:32 +00003116 return g_icvs;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003117}
3118
Jonathan Peyton30419822017-05-12 18:01:32 +00003119static kmp_internal_control_t __kmp_get_x_global_icvs(const kmp_team_t *team) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00003120
Jonathan Peyton30419822017-05-12 18:01:32 +00003121 kmp_internal_control_t gx_icvs;
3122 gx_icvs.serial_nesting_level =
3123 0; // probably =team->t.t_serial like in save_inter_controls
3124 copy_icvs(&gx_icvs, &team->t.t_threads[0]->th.th_current_task->td_icvs);
3125 gx_icvs.next = NULL;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003126
Jonathan Peyton30419822017-05-12 18:01:32 +00003127 return gx_icvs;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003128}
3129
Jonathan Peyton30419822017-05-12 18:01:32 +00003130static void __kmp_initialize_root(kmp_root_t *root) {
3131 int f;
3132 kmp_team_t *root_team;
3133 kmp_team_t *hot_team;
3134 int hot_team_max_nth;
3135 kmp_r_sched_t r_sched =
3136 __kmp_get_schedule_global(); // get current state of scheduling globals
3137 kmp_internal_control_t r_icvs = __kmp_get_global_icvs();
3138 KMP_DEBUG_ASSERT(root);
3139 KMP_ASSERT(!root->r.r_begin);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003140
Jonathan Peyton30419822017-05-12 18:01:32 +00003141 /* setup the root state structure */
3142 __kmp_init_lock(&root->r.r_begin_lock);
3143 root->r.r_begin = FALSE;
3144 root->r.r_active = FALSE;
3145 root->r.r_in_parallel = 0;
3146 root->r.r_blocktime = __kmp_dflt_blocktime;
3147 root->r.r_nested = __kmp_dflt_nested;
Jonathan Peytonf4392462017-07-27 20:58:41 +00003148 root->r.r_cg_nthreads = 1;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003149
Jonathan Peyton30419822017-05-12 18:01:32 +00003150 /* setup the root team for this task */
3151 /* allocate the root team structure */
3152 KF_TRACE(10, ("__kmp_initialize_root: before root_team\n"));
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00003153
Jonathan Peyton30419822017-05-12 18:01:32 +00003154 root_team =
3155 __kmp_allocate_team(root,
3156 1, // new_nproc
3157 1, // max_nproc
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00003158#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00003159 0, // root parallel id
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00003160#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00003161#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00003162 __kmp_nested_proc_bind.bind_types[0],
Jim Cownie5e8470a2013-09-27 10:38:44 +00003163#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00003164 &r_icvs,
3165 0 // argc
3166 USE_NESTED_HOT_ARG(NULL) // master thread is unknown
3167 );
Jonathan Peyton8fbb49a2015-07-09 18:16:58 +00003168#if USE_DEBUGGER
Jonathan Peyton30419822017-05-12 18:01:32 +00003169 // Non-NULL value should be assigned to make the debugger display the root
3170 // team.
3171 TCW_SYNC_PTR(root_team->t.t_pkfn, (microtask_t)(~0));
Jonathan Peyton8fbb49a2015-07-09 18:16:58 +00003172#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00003173
Jonathan Peyton30419822017-05-12 18:01:32 +00003174 KF_TRACE(10, ("__kmp_initialize_root: after root_team = %p\n", root_team));
Jim Cownie5e8470a2013-09-27 10:38:44 +00003175
Jonathan Peyton30419822017-05-12 18:01:32 +00003176 root->r.r_root_team = root_team;
3177 root_team->t.t_control_stack_top = NULL;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003178
Jonathan Peyton30419822017-05-12 18:01:32 +00003179 /* initialize root team */
3180 root_team->t.t_threads[0] = NULL;
3181 root_team->t.t_nproc = 1;
3182 root_team->t.t_serialized = 1;
3183 // TODO???: root_team->t.t_max_active_levels = __kmp_dflt_max_active_levels;
3184 root_team->t.t_sched.r_sched_type = r_sched.r_sched_type;
3185 root_team->t.t_sched.chunk = r_sched.chunk;
3186 KA_TRACE(
3187 20,
3188 ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n",
3189 root_team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
Jim Cownie5e8470a2013-09-27 10:38:44 +00003190
Jonathan Peyton30419822017-05-12 18:01:32 +00003191 /* setup the hot team for this task */
3192 /* allocate the hot team structure */
3193 KF_TRACE(10, ("__kmp_initialize_root: before hot_team\n"));
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00003194
Jonathan Peyton30419822017-05-12 18:01:32 +00003195 hot_team =
3196 __kmp_allocate_team(root,
3197 1, // new_nproc
3198 __kmp_dflt_team_nth_ub * 2, // max_nproc
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00003199#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00003200 0, // root parallel id
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00003201#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00003202#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00003203 __kmp_nested_proc_bind.bind_types[0],
Jim Cownie5e8470a2013-09-27 10:38:44 +00003204#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00003205 &r_icvs,
3206 0 // argc
3207 USE_NESTED_HOT_ARG(NULL) // master thread is unknown
3208 );
3209 KF_TRACE(10, ("__kmp_initialize_root: after hot_team = %p\n", hot_team));
Jim Cownie5e8470a2013-09-27 10:38:44 +00003210
Jonathan Peyton30419822017-05-12 18:01:32 +00003211 root->r.r_hot_team = hot_team;
3212 root_team->t.t_control_stack_top = NULL;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003213
Jonathan Peyton30419822017-05-12 18:01:32 +00003214 /* first-time initialization */
3215 hot_team->t.t_parent = root_team;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003216
Jonathan Peyton30419822017-05-12 18:01:32 +00003217 /* initialize hot team */
3218 hot_team_max_nth = hot_team->t.t_max_nproc;
3219 for (f = 0; f < hot_team_max_nth; ++f) {
3220 hot_team->t.t_threads[f] = NULL;
3221 }; // for
3222 hot_team->t.t_nproc = 1;
3223 // TODO???: hot_team->t.t_max_active_levels = __kmp_dflt_max_active_levels;
3224 hot_team->t.t_sched.r_sched_type = r_sched.r_sched_type;
3225 hot_team->t.t_sched.chunk = r_sched.chunk;
3226 hot_team->t.t_size_changed = 0;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003227}
3228
3229#ifdef KMP_DEBUG
3230
Jim Cownie5e8470a2013-09-27 10:38:44 +00003231typedef struct kmp_team_list_item {
Jonathan Peyton30419822017-05-12 18:01:32 +00003232 kmp_team_p const *entry;
3233 struct kmp_team_list_item *next;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003234} kmp_team_list_item_t;
Jonathan Peyton30419822017-05-12 18:01:32 +00003235typedef kmp_team_list_item_t *kmp_team_list_t;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003236
Jonathan Peyton30419822017-05-12 18:01:32 +00003237static void __kmp_print_structure_team_accum( // Add team to list of teams.
3238 kmp_team_list_t list, // List of teams.
3239 kmp_team_p const *team // Team to add.
3240 ) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00003241
Jonathan Peyton30419822017-05-12 18:01:32 +00003242 // List must terminate with item where both entry and next are NULL.
3243 // Team is added to the list only once.
3244 // List is sorted in ascending order by team id.
3245 // Team id is *not* a key.
Jim Cownie5e8470a2013-09-27 10:38:44 +00003246
Jonathan Peyton30419822017-05-12 18:01:32 +00003247 kmp_team_list_t l;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003248
Jonathan Peyton30419822017-05-12 18:01:32 +00003249 KMP_DEBUG_ASSERT(list != NULL);
3250 if (team == NULL) {
3251 return;
3252 }; // if
Jim Cownie5e8470a2013-09-27 10:38:44 +00003253
Jonathan Peyton30419822017-05-12 18:01:32 +00003254 __kmp_print_structure_team_accum(list, team->t.t_parent);
3255 __kmp_print_structure_team_accum(list, team->t.t_next_pool);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003256
Jonathan Peyton30419822017-05-12 18:01:32 +00003257 // Search list for the team.
3258 l = list;
3259 while (l->next != NULL && l->entry != team) {
3260 l = l->next;
3261 }; // while
3262 if (l->next != NULL) {
3263 return; // Team has been added before, exit.
3264 }; // if
Jim Cownie5e8470a2013-09-27 10:38:44 +00003265
Jonathan Peyton30419822017-05-12 18:01:32 +00003266 // Team is not found. Search list again for insertion point.
3267 l = list;
3268 while (l->next != NULL && l->entry->t.t_id <= team->t.t_id) {
3269 l = l->next;
3270 }; // while
Jim Cownie5e8470a2013-09-27 10:38:44 +00003271
Jonathan Peyton30419822017-05-12 18:01:32 +00003272 // Insert team.
3273 {
3274 kmp_team_list_item_t *item = (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(
3275 sizeof(kmp_team_list_item_t));
3276 *item = *l;
3277 l->entry = team;
3278 l->next = item;
3279 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00003280}
3281
Jonathan Peyton30419822017-05-12 18:01:32 +00003282static void __kmp_print_structure_team(char const *title, kmp_team_p const *team
Jim Cownie5e8470a2013-09-27 10:38:44 +00003283
Jonathan Peyton30419822017-05-12 18:01:32 +00003284 ) {
3285 __kmp_printf("%s", title);
3286 if (team != NULL) {
3287 __kmp_printf("%2x %p\n", team->t.t_id, team);
3288 } else {
3289 __kmp_printf(" - (nil)\n");
3290 }; // if
Jim Cownie5e8470a2013-09-27 10:38:44 +00003291}
3292
Jonathan Peyton30419822017-05-12 18:01:32 +00003293static void __kmp_print_structure_thread(char const *title,
3294 kmp_info_p const *thread) {
3295 __kmp_printf("%s", title);
3296 if (thread != NULL) {
3297 __kmp_printf("%2d %p\n", thread->th.th_info.ds.ds_gtid, thread);
3298 } else {
3299 __kmp_printf(" - (nil)\n");
3300 }; // if
Jim Cownie5e8470a2013-09-27 10:38:44 +00003301}
3302
Jonathan Peyton30419822017-05-12 18:01:32 +00003303void __kmp_print_structure(void) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00003304
Jonathan Peyton30419822017-05-12 18:01:32 +00003305 kmp_team_list_t list;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003306
Jonathan Peyton30419822017-05-12 18:01:32 +00003307 // Initialize list of teams.
3308 list =
3309 (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(sizeof(kmp_team_list_item_t));
3310 list->entry = NULL;
3311 list->next = NULL;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003312
Jonathan Peyton30419822017-05-12 18:01:32 +00003313 __kmp_printf("\n------------------------------\nGlobal Thread "
3314 "Table\n------------------------------\n");
3315 {
3316 int gtid;
3317 for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3318 __kmp_printf("%2d", gtid);
3319 if (__kmp_threads != NULL) {
3320 __kmp_printf(" %p", __kmp_threads[gtid]);
3321 }; // if
3322 if (__kmp_root != NULL) {
3323 __kmp_printf(" %p", __kmp_root[gtid]);
3324 }; // if
3325 __kmp_printf("\n");
3326 }; // for gtid
3327 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00003328
Jonathan Peyton30419822017-05-12 18:01:32 +00003329 // Print out __kmp_threads array.
3330 __kmp_printf("\n------------------------------\nThreads\n--------------------"
3331 "----------\n");
3332 if (__kmp_threads != NULL) {
3333 int gtid;
3334 for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3335 kmp_info_t const *thread = __kmp_threads[gtid];
3336 if (thread != NULL) {
3337 __kmp_printf("GTID %2d %p:\n", gtid, thread);
3338 __kmp_printf(" Our Root: %p\n", thread->th.th_root);
3339 __kmp_print_structure_team(" Our Team: ", thread->th.th_team);
3340 __kmp_print_structure_team(" Serial Team: ",
3341 thread->th.th_serial_team);
3342 __kmp_printf(" Threads: %2d\n", thread->th.th_team_nproc);
3343 __kmp_print_structure_thread(" Master: ",
3344 thread->th.th_team_master);
3345 __kmp_printf(" Serialized?: %2d\n", thread->th.th_team_serialized);
3346 __kmp_printf(" Set NProc: %2d\n", thread->th.th_set_nproc);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003347#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00003348 __kmp_printf(" Set Proc Bind: %2d\n", thread->th.th_set_proc_bind);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003349#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00003350 __kmp_print_structure_thread(" Next in pool: ",
3351 thread->th.th_next_pool);
3352 __kmp_printf("\n");
3353 __kmp_print_structure_team_accum(list, thread->th.th_team);
3354 __kmp_print_structure_team_accum(list, thread->th.th_serial_team);
3355 }; // if
3356 }; // for gtid
3357 } else {
3358 __kmp_printf("Threads array is not allocated.\n");
3359 }; // if
Jim Cownie5e8470a2013-09-27 10:38:44 +00003360
Jonathan Peyton30419822017-05-12 18:01:32 +00003361 // Print out __kmp_root array.
3362 __kmp_printf("\n------------------------------\nUbers\n----------------------"
3363 "--------\n");
3364 if (__kmp_root != NULL) {
3365 int gtid;
3366 for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3367 kmp_root_t const *root = __kmp_root[gtid];
3368 if (root != NULL) {
3369 __kmp_printf("GTID %2d %p:\n", gtid, root);
3370 __kmp_print_structure_team(" Root Team: ", root->r.r_root_team);
3371 __kmp_print_structure_team(" Hot Team: ", root->r.r_hot_team);
3372 __kmp_print_structure_thread(" Uber Thread: ",
3373 root->r.r_uber_thread);
3374 __kmp_printf(" Active?: %2d\n", root->r.r_active);
3375 __kmp_printf(" Nested?: %2d\n", root->r.r_nested);
3376 __kmp_printf(" In Parallel: %2d\n", root->r.r_in_parallel);
3377 __kmp_printf("\n");
3378 __kmp_print_structure_team_accum(list, root->r.r_root_team);
3379 __kmp_print_structure_team_accum(list, root->r.r_hot_team);
3380 }; // if
3381 }; // for gtid
3382 } else {
3383 __kmp_printf("Ubers array is not allocated.\n");
3384 }; // if
Jim Cownie5e8470a2013-09-27 10:38:44 +00003385
Jonathan Peyton30419822017-05-12 18:01:32 +00003386 __kmp_printf("\n------------------------------\nTeams\n----------------------"
3387 "--------\n");
3388 while (list->next != NULL) {
3389 kmp_team_p const *team = list->entry;
3390 int i;
3391 __kmp_printf("Team %2x %p:\n", team->t.t_id, team);
3392 __kmp_print_structure_team(" Parent Team: ", team->t.t_parent);
3393 __kmp_printf(" Master TID: %2d\n", team->t.t_master_tid);
3394 __kmp_printf(" Max threads: %2d\n", team->t.t_max_nproc);
3395 __kmp_printf(" Levels of serial: %2d\n", team->t.t_serialized);
3396 __kmp_printf(" Number threads: %2d\n", team->t.t_nproc);
3397 for (i = 0; i < team->t.t_nproc; ++i) {
3398 __kmp_printf(" Thread %2d: ", i);
3399 __kmp_print_structure_thread("", team->t.t_threads[i]);
3400 }; // for i
3401 __kmp_print_structure_team(" Next in pool: ", team->t.t_next_pool);
3402 __kmp_printf("\n");
3403 list = list->next;
3404 }; // while
Jim Cownie5e8470a2013-09-27 10:38:44 +00003405
Jonathan Peyton30419822017-05-12 18:01:32 +00003406 // Print out __kmp_thread_pool and __kmp_team_pool.
3407 __kmp_printf("\n------------------------------\nPools\n----------------------"
3408 "--------\n");
3409 __kmp_print_structure_thread("Thread pool: ",
Andrey Churbanovc47afcd2017-07-03 11:24:08 +00003410 CCAST(kmp_info_t *, __kmp_thread_pool));
Jonathan Peyton30419822017-05-12 18:01:32 +00003411 __kmp_print_structure_team("Team pool: ",
Andrey Churbanovc47afcd2017-07-03 11:24:08 +00003412 CCAST(kmp_team_t *, __kmp_team_pool));
Jonathan Peyton30419822017-05-12 18:01:32 +00003413 __kmp_printf("\n");
Jim Cownie5e8470a2013-09-27 10:38:44 +00003414
Jonathan Peyton30419822017-05-12 18:01:32 +00003415 // Free team list.
3416 while (list != NULL) {
3417 kmp_team_list_item_t *item = list;
3418 list = list->next;
3419 KMP_INTERNAL_FREE(item);
3420 }; // while
Jim Cownie5e8470a2013-09-27 10:38:44 +00003421}
3422
3423#endif
3424
Jim Cownie5e8470a2013-09-27 10:38:44 +00003425//---------------------------------------------------------------------------
3426// Stuff for per-thread fast random number generator
3427// Table of primes
Jim Cownie5e8470a2013-09-27 10:38:44 +00003428static const unsigned __kmp_primes[] = {
Jonathan Peyton30419822017-05-12 18:01:32 +00003429 0x9e3779b1, 0xffe6cc59, 0x2109f6dd, 0x43977ab5, 0xba5703f5, 0xb495a877,
3430 0xe1626741, 0x79695e6b, 0xbc98c09f, 0xd5bee2b3, 0x287488f9, 0x3af18231,
3431 0x9677cd4d, 0xbe3a6929, 0xadc6a877, 0xdcf0674b, 0xbe4d6fe9, 0x5f15e201,
3432 0x99afc3fd, 0xf3f16801, 0xe222cfff, 0x24ba5fdb, 0x0620452d, 0x79f149e3,
3433 0xc8b93f49, 0x972702cd, 0xb07dd827, 0x6c97d5ed, 0x085a3d61, 0x46eb5ea7,
3434 0x3d9910ed, 0x2e687b5b, 0x29609227, 0x6eb081f1, 0x0954c4e1, 0x9d114db9,
3435 0x542acfa9, 0xb3e6bd7b, 0x0742d917, 0xe9f3ffa7, 0x54581edb, 0xf2480f45,
3436 0x0bb9288f, 0xef1affc7, 0x85fa0ca7, 0x3ccc14db, 0xe6baf34b, 0x343377f7,
3437 0x5ca19031, 0xe6d9293b, 0xf0a9f391, 0x5d2e980b, 0xfc411073, 0xc3749363,
3438 0xb892d829, 0x3549366b, 0x629750ad, 0xb98294e5, 0x892d9483, 0xc235baf3,
3439 0x3d2402a3, 0x6bdef3c9, 0xbec333cd, 0x40c9520f};
Jim Cownie5e8470a2013-09-27 10:38:44 +00003440
3441//---------------------------------------------------------------------------
3442// __kmp_get_random: Get a random number using a linear congruential method.
Jonathan Peyton30419822017-05-12 18:01:32 +00003443unsigned short __kmp_get_random(kmp_info_t *thread) {
Jim Cownie4cc4bb42014-10-07 16:25:50 +00003444 unsigned x = thread->th.th_x;
Jonathan Peyton30419822017-05-12 18:01:32 +00003445 unsigned short r = x >> 16;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003446
Jonathan Peyton30419822017-05-12 18:01:32 +00003447 thread->th.th_x = x * thread->th.th_a + 1;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003448
3449 KA_TRACE(30, ("__kmp_get_random: THREAD: %d, RETURN: %u\n",
Jonathan Peyton30419822017-05-12 18:01:32 +00003450 thread->th.th_info.ds.ds_tid, r));
Jim Cownie5e8470a2013-09-27 10:38:44 +00003451
3452 return r;
3453}
3454//--------------------------------------------------------
3455// __kmp_init_random: Initialize a random number generator
Jonathan Peyton30419822017-05-12 18:01:32 +00003456void __kmp_init_random(kmp_info_t *thread) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00003457 unsigned seed = thread->th.th_info.ds.ds_tid;
3458
Jonathan Peyton30419822017-05-12 18:01:32 +00003459 thread->th.th_a =
3460 __kmp_primes[seed % (sizeof(__kmp_primes) / sizeof(__kmp_primes[0]))];
3461 thread->th.th_x = (seed + 1) * thread->th.th_a + 1;
3462 KA_TRACE(30,
3463 ("__kmp_init_random: THREAD: %u; A: %u\n", seed, thread->th.th_a));
Jim Cownie5e8470a2013-09-27 10:38:44 +00003464}
3465
Jim Cownie5e8470a2013-09-27 10:38:44 +00003466#if KMP_OS_WINDOWS
Jonathan Peyton30419822017-05-12 18:01:32 +00003467/* reclaim array entries for root threads that are already dead, returns number
3468 * reclaimed */
3469static int __kmp_reclaim_dead_roots(void) {
3470 int i, r = 0;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003471
Jonathan Peyton30419822017-05-12 18:01:32 +00003472 for (i = 0; i < __kmp_threads_capacity; ++i) {
3473 if (KMP_UBER_GTID(i) &&
3474 !__kmp_still_running((kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[i])) &&
3475 !__kmp_root[i]
3476 ->r.r_active) { // AC: reclaim only roots died in non-active state
3477 r += __kmp_unregister_root_other_thread(i);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003478 }
Jonathan Peyton30419822017-05-12 18:01:32 +00003479 }
3480 return r;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003481}
3482#endif
3483
Jonathan Peyton30419822017-05-12 18:01:32 +00003484/* This function attempts to create free entries in __kmp_threads and
3485 __kmp_root, and returns the number of free entries generated.
Jim Cownie5e8470a2013-09-27 10:38:44 +00003486
Jonathan Peyton30419822017-05-12 18:01:32 +00003487 For Windows* OS static library, the first mechanism used is to reclaim array
3488 entries for root threads that are already dead.
Jim Cownie5e8470a2013-09-27 10:38:44 +00003489
Jonathan Peyton30419822017-05-12 18:01:32 +00003490 On all platforms, expansion is attempted on the arrays __kmp_threads_ and
3491 __kmp_root, with appropriate update to __kmp_threads_capacity. Array
3492 capacity is increased by doubling with clipping to __kmp_tp_capacity, if
3493 threadprivate cache array has been created. Synchronization with
3494 __kmpc_threadprivate_cached is done using __kmp_tp_cached_lock.
Jim Cownie5e8470a2013-09-27 10:38:44 +00003495
Jonathan Peyton30419822017-05-12 18:01:32 +00003496 After any dead root reclamation, if the clipping value allows array expansion
3497 to result in the generation of a total of nWish free slots, the function does
3498 that expansion. If not, but the clipping value allows array expansion to
3499 result in the generation of a total of nNeed free slots, the function does
3500 that expansion. Otherwise, nothing is done beyond the possible initial root
3501 thread reclamation. However, if nNeed is zero, a best-effort attempt is made
3502 to fulfil nWish as far as possible, i.e. the function will attempt to create
Jim Cownie5e8470a2013-09-27 10:38:44 +00003503 as many free slots as possible up to nWish.
3504
Jonathan Peyton30419822017-05-12 18:01:32 +00003505 If any argument is negative, the behavior is undefined. */
3506static int __kmp_expand_threads(int nWish, int nNeed) {
3507 int added = 0;
3508 int old_tp_cached;
3509 int __kmp_actual_max_nth;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003510
Jonathan Peyton30419822017-05-12 18:01:32 +00003511 if (nNeed > nWish) /* normalize the arguments */
3512 nWish = nNeed;
Jonathan Peyton99016992015-05-26 17:32:53 +00003513#if KMP_OS_WINDOWS && !defined KMP_DYNAMIC_LIB
Jonathan Peyton30419822017-05-12 18:01:32 +00003514 /* only for Windows static library */
3515 /* reclaim array entries for root threads that are already dead */
3516 added = __kmp_reclaim_dead_roots();
Jim Cownie5e8470a2013-09-27 10:38:44 +00003517
Jonathan Peyton30419822017-05-12 18:01:32 +00003518 if (nNeed) {
3519 nNeed -= added;
3520 if (nNeed < 0)
3521 nNeed = 0;
3522 }
3523 if (nWish) {
3524 nWish -= added;
3525 if (nWish < 0)
3526 nWish = 0;
3527 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00003528#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00003529 if (nWish <= 0)
Jim Cownie5e8470a2013-09-27 10:38:44 +00003530 return added;
Jonathan Peyton30419822017-05-12 18:01:32 +00003531
3532 while (1) {
3533 int nTarget;
3534 int minimumRequiredCapacity;
3535 int newCapacity;
3536 kmp_info_t **newThreads;
3537 kmp_root_t **newRoot;
3538
3539 // Note that __kmp_threads_capacity is not bounded by __kmp_max_nth. If
3540 // __kmp_max_nth is set to some value less than __kmp_sys_max_nth by the
Jonathan Peytonf4392462017-07-27 20:58:41 +00003541 // user via KMP_DEVICE_THREAD_LIMIT, then __kmp_threads_capacity may become
Jonathan Peyton30419822017-05-12 18:01:32 +00003542 // > __kmp_max_nth in one of two ways:
3543 //
3544 // 1) The initialization thread (gtid = 0) exits. __kmp_threads[0]
3545 // may not be resused by another thread, so we may need to increase
Jonathan Peyton09244f32017-07-26 20:07:58 +00003546 // __kmp_threads_capacity to __kmp_max_nth + 1.
Jonathan Peyton30419822017-05-12 18:01:32 +00003547 //
3548 // 2) New foreign root(s) are encountered. We always register new foreign
3549 // roots. This may cause a smaller # of threads to be allocated at
3550 // subsequent parallel regions, but the worker threads hang around (and
3551 // eventually go to sleep) and need slots in the __kmp_threads[] array.
3552 //
3553 // Anyway, that is the reason for moving the check to see if
Jonathan Peyton09244f32017-07-26 20:07:58 +00003554 // __kmp_max_nth was exceeded into __kmp_reserve_threads()
Jonathan Peyton30419822017-05-12 18:01:32 +00003555 // instead of having it performed here. -BB
3556 old_tp_cached = __kmp_tp_cached;
3557 __kmp_actual_max_nth =
3558 old_tp_cached ? __kmp_tp_capacity : __kmp_sys_max_nth;
3559 KMP_DEBUG_ASSERT(__kmp_actual_max_nth >= __kmp_threads_capacity);
3560
3561 /* compute expansion headroom to check if we can expand and whether to aim
3562 for nWish or nNeed */
3563 nTarget = nWish;
3564 if (__kmp_actual_max_nth - __kmp_threads_capacity < nTarget) {
3565 /* can't fulfil nWish, so try nNeed */
3566 if (nNeed) {
3567 nTarget = nNeed;
3568 if (__kmp_actual_max_nth - __kmp_threads_capacity < nTarget) {
3569 /* possible expansion too small -- give up */
3570 break;
3571 }
3572 } else {
3573 /* best-effort */
3574 nTarget = __kmp_actual_max_nth - __kmp_threads_capacity;
3575 if (!nTarget) {
3576 /* can expand at all -- give up */
3577 break;
3578 }
3579 }
3580 }
3581 minimumRequiredCapacity = __kmp_threads_capacity + nTarget;
3582
3583 newCapacity = __kmp_threads_capacity;
3584 do {
3585 newCapacity = newCapacity <= (__kmp_actual_max_nth >> 1)
3586 ? (newCapacity << 1)
3587 : __kmp_actual_max_nth;
3588 } while (newCapacity < minimumRequiredCapacity);
3589 newThreads = (kmp_info_t **)__kmp_allocate(
3590 (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * newCapacity +
3591 CACHE_LINE);
3592 newRoot = (kmp_root_t **)((char *)newThreads +
3593 sizeof(kmp_info_t *) * newCapacity);
3594 KMP_MEMCPY(newThreads, __kmp_threads,
3595 __kmp_threads_capacity * sizeof(kmp_info_t *));
3596 KMP_MEMCPY(newRoot, __kmp_root,
3597 __kmp_threads_capacity * sizeof(kmp_root_t *));
3598 memset(newThreads + __kmp_threads_capacity, 0,
3599 (newCapacity - __kmp_threads_capacity) * sizeof(kmp_info_t *));
3600 memset(newRoot + __kmp_threads_capacity, 0,
3601 (newCapacity - __kmp_threads_capacity) * sizeof(kmp_root_t *));
3602
3603 if (!old_tp_cached && __kmp_tp_cached && newCapacity > __kmp_tp_capacity) {
3604 /* __kmp_tp_cached has changed, i.e. __kmpc_threadprivate_cached has
3605 allocated a threadprivate cache while we were allocating the expanded
3606 array, and our new capacity is larger than the threadprivate cache
3607 capacity, so we should deallocate the expanded arrays and try again.
3608 This is the first check of a double-check pair. */
3609 __kmp_free(newThreads);
3610 continue; /* start over and try again */
3611 }
3612 __kmp_acquire_bootstrap_lock(&__kmp_tp_cached_lock);
3613 if (!old_tp_cached && __kmp_tp_cached && newCapacity > __kmp_tp_capacity) {
3614 /* Same check as above, but this time with the lock so we can be sure if
3615 we can succeed. */
3616 __kmp_release_bootstrap_lock(&__kmp_tp_cached_lock);
3617 __kmp_free(newThreads);
3618 continue; /* start over and try again */
3619 } else {
3620 /* success */
3621 // __kmp_free( __kmp_threads ); // ATT: It leads to crash. Need to be
3622 // investigated.
3623 *(kmp_info_t * *volatile *)&__kmp_threads = newThreads;
3624 *(kmp_root_t * *volatile *)&__kmp_root = newRoot;
3625 added += newCapacity - __kmp_threads_capacity;
3626 *(volatile int *)&__kmp_threads_capacity = newCapacity;
3627 __kmp_release_bootstrap_lock(&__kmp_tp_cached_lock);
3628 break; /* succeeded, so we can exit the loop */
3629 }
3630 }
3631 return added;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003632}
3633
Jonathan Peyton30419822017-05-12 18:01:32 +00003634/* Register the current thread as a root thread and obtain our gtid. We must
3635 have the __kmp_initz_lock held at this point. Argument TRUE only if are the
3636 thread that calls from __kmp_do_serial_initialize() */
3637int __kmp_register_root(int initial_thread) {
3638 kmp_info_t *root_thread;
3639 kmp_root_t *root;
3640 int gtid;
3641 int capacity;
3642 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
3643 KA_TRACE(20, ("__kmp_register_root: entered\n"));
3644 KMP_MB();
Jim Cownie5e8470a2013-09-27 10:38:44 +00003645
Jonathan Peyton30419822017-05-12 18:01:32 +00003646 /* 2007-03-02:
3647 If initial thread did not invoke OpenMP RTL yet, and this thread is not an
3648 initial one, "__kmp_all_nth >= __kmp_threads_capacity" condition does not
3649 work as expected -- it may return false (that means there is at least one
3650 empty slot in __kmp_threads array), but it is possible the only free slot
3651 is #0, which is reserved for initial thread and so cannot be used for this
3652 one. Following code workarounds this bug.
Jim Cownie5e8470a2013-09-27 10:38:44 +00003653
Jonathan Peyton30419822017-05-12 18:01:32 +00003654 However, right solution seems to be not reserving slot #0 for initial
3655 thread because:
3656 (1) there is no magic in slot #0,
3657 (2) we cannot detect initial thread reliably (the first thread which does
3658 serial initialization may be not a real initial thread).
3659 */
3660 capacity = __kmp_threads_capacity;
3661 if (!initial_thread && TCR_PTR(__kmp_threads[0]) == NULL) {
3662 --capacity;
3663 }; // if
Jim Cownie5e8470a2013-09-27 10:38:44 +00003664
Jonathan Peyton30419822017-05-12 18:01:32 +00003665 /* see if there are too many threads */
3666 if (__kmp_all_nth >= capacity && !__kmp_expand_threads(1, 1)) {
3667 if (__kmp_tp_cached) {
3668 __kmp_msg(kmp_ms_fatal, KMP_MSG(CantRegisterNewThread),
3669 KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity),
3670 KMP_HNT(PossibleSystemLimitOnThreads), __kmp_msg_null);
3671 } else {
3672 __kmp_msg(kmp_ms_fatal, KMP_MSG(CantRegisterNewThread),
3673 KMP_HNT(SystemLimitOnThreads), __kmp_msg_null);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003674 }
Jonathan Peyton30419822017-05-12 18:01:32 +00003675 }; // if
3676
3677 /* find an available thread slot */
3678 /* Don't reassign the zero slot since we need that to only be used by initial
3679 thread */
3680 for (gtid = (initial_thread ? 0 : 1); TCR_PTR(__kmp_threads[gtid]) != NULL;
3681 gtid++)
3682 ;
3683 KA_TRACE(1,
3684 ("__kmp_register_root: found slot in threads array: T#%d\n", gtid));
3685 KMP_ASSERT(gtid < __kmp_threads_capacity);
3686
3687 /* update global accounting */
3688 __kmp_all_nth++;
3689 TCW_4(__kmp_nth, __kmp_nth + 1);
3690
3691 // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low
3692 // numbers of procs, and method #2 (keyed API call) for higher numbers.
3693 if (__kmp_adjust_gtid_mode) {
3694 if (__kmp_all_nth >= __kmp_tls_gtid_min) {
3695 if (TCR_4(__kmp_gtid_mode) != 2) {
3696 TCW_4(__kmp_gtid_mode, 2);
3697 }
3698 } else {
3699 if (TCR_4(__kmp_gtid_mode) != 1) {
3700 TCW_4(__kmp_gtid_mode, 1);
3701 }
3702 }
3703 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00003704
3705#ifdef KMP_ADJUST_BLOCKTIME
Jonathan Peyton30419822017-05-12 18:01:32 +00003706 /* Adjust blocktime to zero if necessary */
3707 /* Middle initialization might not have occurred yet */
3708 if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
3709 if (__kmp_nth > __kmp_avail_proc) {
3710 __kmp_zero_bt = TRUE;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003711 }
Jonathan Peyton30419822017-05-12 18:01:32 +00003712 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00003713#endif /* KMP_ADJUST_BLOCKTIME */
3714
Jonathan Peyton30419822017-05-12 18:01:32 +00003715 /* setup this new hierarchy */
3716 if (!(root = __kmp_root[gtid])) {
3717 root = __kmp_root[gtid] = (kmp_root_t *)__kmp_allocate(sizeof(kmp_root_t));
3718 KMP_DEBUG_ASSERT(!root->r.r_root_team);
3719 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00003720
Jonathan Peyton5375fe82016-11-14 21:13:44 +00003721#if KMP_STATS_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00003722 // Initialize stats as soon as possible (right after gtid assignment).
3723 __kmp_stats_thread_ptr = __kmp_stats_list->push_back(gtid);
3724 KMP_START_EXPLICIT_TIMER(OMP_worker_thread_life);
3725 KMP_SET_THREAD_STATE(SERIAL_REGION);
3726 KMP_INIT_PARTITIONED_TIMERS(OMP_serial);
Jonathan Peyton5375fe82016-11-14 21:13:44 +00003727#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00003728 __kmp_initialize_root(root);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003729
Jonathan Peyton30419822017-05-12 18:01:32 +00003730 /* setup new root thread structure */
3731 if (root->r.r_uber_thread) {
3732 root_thread = root->r.r_uber_thread;
3733 } else {
3734 root_thread = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t));
3735 if (__kmp_storage_map) {
3736 __kmp_print_thread_storage_map(root_thread, gtid);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003737 }
Jonathan Peyton30419822017-05-12 18:01:32 +00003738 root_thread->th.th_info.ds.ds_gtid = gtid;
3739 root_thread->th.th_root = root;
3740 if (__kmp_env_consistency_check) {
3741 root_thread->th.th_cons = __kmp_allocate_cons_stack(gtid);
3742 }
3743#if USE_FAST_MEMORY
3744 __kmp_initialize_fast_memory(root_thread);
3745#endif /* USE_FAST_MEMORY */
Jim Cownie5e8470a2013-09-27 10:38:44 +00003746
Jonathan Peyton30419822017-05-12 18:01:32 +00003747#if KMP_USE_BGET
3748 KMP_DEBUG_ASSERT(root_thread->th.th_local.bget_data == NULL);
3749 __kmp_initialize_bget(root_thread);
3750#endif
3751 __kmp_init_random(root_thread); // Initialize random number generator
3752 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00003753
Jonathan Peyton30419822017-05-12 18:01:32 +00003754 /* setup the serial team held in reserve by the root thread */
3755 if (!root_thread->th.th_serial_team) {
3756 kmp_internal_control_t r_icvs = __kmp_get_global_icvs();
3757 KF_TRACE(10, ("__kmp_register_root: before serial_team\n"));
3758 root_thread->th.th_serial_team =
3759 __kmp_allocate_team(root, 1, 1,
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00003760#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00003761 0, // root parallel id
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00003762#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00003763#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00003764 proc_bind_default,
Jim Cownie5e8470a2013-09-27 10:38:44 +00003765#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00003766 &r_icvs, 0 USE_NESTED_HOT_ARG(NULL));
3767 }
3768 KMP_ASSERT(root_thread->th.th_serial_team);
3769 KF_TRACE(10, ("__kmp_register_root: after serial_team = %p\n",
3770 root_thread->th.th_serial_team));
Jim Cownie5e8470a2013-09-27 10:38:44 +00003771
Jonathan Peyton30419822017-05-12 18:01:32 +00003772 /* drop root_thread into place */
3773 TCW_SYNC_PTR(__kmp_threads[gtid], root_thread);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003774
Jonathan Peyton30419822017-05-12 18:01:32 +00003775 root->r.r_root_team->t.t_threads[0] = root_thread;
3776 root->r.r_hot_team->t.t_threads[0] = root_thread;
3777 root_thread->th.th_serial_team->t.t_threads[0] = root_thread;
3778 // AC: the team created in reserve, not for execution (it is unused for now).
3779 root_thread->th.th_serial_team->t.t_serialized = 0;
3780 root->r.r_uber_thread = root_thread;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003781
Jonathan Peyton30419822017-05-12 18:01:32 +00003782 /* initialize the thread, get it ready to go */
3783 __kmp_initialize_info(root_thread, root->r.r_root_team, 0, gtid);
3784 TCW_4(__kmp_init_gtid, TRUE);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003785
Jonathan Peyton30419822017-05-12 18:01:32 +00003786 /* prepare the master thread for get_gtid() */
3787 __kmp_gtid_set_specific(gtid);
Jim Cownie4cc4bb42014-10-07 16:25:50 +00003788
Jonathan Peyton7abf9d52016-05-26 18:19:10 +00003789#if USE_ITT_BUILD
Jonathan Peyton30419822017-05-12 18:01:32 +00003790 __kmp_itt_thread_name(gtid);
Jonathan Peyton7abf9d52016-05-26 18:19:10 +00003791#endif /* USE_ITT_BUILD */
Jim Cownie4cc4bb42014-10-07 16:25:50 +00003792
Jonathan Peyton30419822017-05-12 18:01:32 +00003793#ifdef KMP_TDATA_GTID
3794 __kmp_gtid = gtid;
Jonathan Peyton8fbb49a2015-07-09 18:16:58 +00003795#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00003796 __kmp_create_worker(gtid, root_thread, __kmp_stksize);
3797 KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == gtid);
3798
3799 KA_TRACE(20, ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, "
3800 "plain=%u\n",
3801 gtid, __kmp_gtid_from_tid(0, root->r.r_hot_team),
3802 root->r.r_hot_team->t.t_id, 0, KMP_INIT_BARRIER_STATE,
3803 KMP_INIT_BARRIER_STATE));
3804 { // Initialize barrier data.
3805 int b;
3806 for (b = 0; b < bs_last_barrier; ++b) {
3807 root_thread->th.th_bar[b].bb.b_arrived = KMP_INIT_BARRIER_STATE;
3808#if USE_DEBUGGER
3809 root_thread->th.th_bar[b].bb.b_worker_arrived = 0;
3810#endif
3811 }; // for
3812 }
3813 KMP_DEBUG_ASSERT(root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier].b_arrived ==
3814 KMP_INIT_BARRIER_STATE);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003815
Alp Toker763b9392014-02-28 09:42:41 +00003816#if KMP_AFFINITY_SUPPORTED
Jonathan Peyton30419822017-05-12 18:01:32 +00003817#if OMP_40_ENABLED
3818 root_thread->th.th_current_place = KMP_PLACE_UNDEFINED;
3819 root_thread->th.th_new_place = KMP_PLACE_UNDEFINED;
3820 root_thread->th.th_first_place = KMP_PLACE_UNDEFINED;
3821 root_thread->th.th_last_place = KMP_PLACE_UNDEFINED;
3822#endif
Jonathan Peyton2f7c0772016-02-25 18:49:52 +00003823
Jonathan Peyton30419822017-05-12 18:01:32 +00003824 if (TCR_4(__kmp_init_middle)) {
3825 __kmp_affinity_set_init_mask(gtid, TRUE);
3826 }
Alp Toker763b9392014-02-28 09:42:41 +00003827#endif /* KMP_AFFINITY_SUPPORTED */
Jim Cownie5e8470a2013-09-27 10:38:44 +00003828
Jonathan Peyton30419822017-05-12 18:01:32 +00003829 __kmp_root_counter++;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003830
Jonathan Peyton30419822017-05-12 18:01:32 +00003831 KMP_MB();
3832 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003833
Jonathan Peyton30419822017-05-12 18:01:32 +00003834 return gtid;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003835}
3836
Jim Cownie4cc4bb42014-10-07 16:25:50 +00003837#if KMP_NESTED_HOT_TEAMS
Jonathan Peyton30419822017-05-12 18:01:32 +00003838static int __kmp_free_hot_teams(kmp_root_t *root, kmp_info_t *thr, int level,
3839 const int max_level) {
3840 int i, n, nth;
3841 kmp_hot_team_ptr_t *hot_teams = thr->th.th_hot_teams;
3842 if (!hot_teams || !hot_teams[level].hot_team) {
3843 return 0;
3844 }
3845 KMP_DEBUG_ASSERT(level < max_level);
3846 kmp_team_t *team = hot_teams[level].hot_team;
3847 nth = hot_teams[level].hot_team_nth;
3848 n = nth - 1; // master is not freed
3849 if (level < max_level - 1) {
3850 for (i = 0; i < nth; ++i) {
3851 kmp_info_t *th = team->t.t_threads[i];
3852 n += __kmp_free_hot_teams(root, th, level + 1, max_level);
3853 if (i > 0 && th->th.th_hot_teams) {
3854 __kmp_free(th->th.th_hot_teams);
3855 th->th.th_hot_teams = NULL;
3856 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00003857 }
Jonathan Peyton30419822017-05-12 18:01:32 +00003858 }
3859 __kmp_free_team(root, team, NULL);
3860 return n;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00003861}
3862#endif
3863
Jonathan Peyton30419822017-05-12 18:01:32 +00003864// Resets a root thread and clear its root and hot teams.
3865// Returns the number of __kmp_threads entries directly and indirectly freed.
3866static int __kmp_reset_root(int gtid, kmp_root_t *root) {
3867 kmp_team_t *root_team = root->r.r_root_team;
3868 kmp_team_t *hot_team = root->r.r_hot_team;
3869 int n = hot_team->t.t_nproc;
3870 int i;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003871
Jonathan Peyton30419822017-05-12 18:01:32 +00003872 KMP_DEBUG_ASSERT(!root->r.r_active);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003873
Jonathan Peyton30419822017-05-12 18:01:32 +00003874 root->r.r_root_team = NULL;
3875 root->r.r_hot_team = NULL;
3876 // __kmp_free_team() does not free hot teams, so we have to clear r_hot_team
3877 // before call to __kmp_free_team().
3878 __kmp_free_team(root, root_team USE_NESTED_HOT_ARG(NULL));
Jim Cownie4cc4bb42014-10-07 16:25:50 +00003879#if KMP_NESTED_HOT_TEAMS
Jonathan Peyton30419822017-05-12 18:01:32 +00003880 if (__kmp_hot_teams_max_level >
3881 0) { // need to free nested hot teams and their threads if any
3882 for (i = 0; i < hot_team->t.t_nproc; ++i) {
3883 kmp_info_t *th = hot_team->t.t_threads[i];
3884 if (__kmp_hot_teams_max_level > 1) {
3885 n += __kmp_free_hot_teams(root, th, 1, __kmp_hot_teams_max_level);
3886 }
3887 if (th->th.th_hot_teams) {
3888 __kmp_free(th->th.th_hot_teams);
3889 th->th.th_hot_teams = NULL;
3890 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00003891 }
Jonathan Peyton30419822017-05-12 18:01:32 +00003892 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00003893#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00003894 __kmp_free_team(root, hot_team USE_NESTED_HOT_ARG(NULL));
Jim Cownie5e8470a2013-09-27 10:38:44 +00003895
Jonathan Peyton30419822017-05-12 18:01:32 +00003896 // Before we can reap the thread, we need to make certain that all other
3897 // threads in the teams that had this root as ancestor have stopped trying to
3898 // steal tasks.
3899 if (__kmp_tasking_mode != tskm_immediate_exec) {
3900 __kmp_wait_to_unref_task_teams();
3901 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00003902
Jonathan Peyton30419822017-05-12 18:01:32 +00003903#if KMP_OS_WINDOWS
3904 /* Close Handle of root duplicated in __kmp_create_worker (tr #62919) */
3905 KA_TRACE(
3906 10, ("__kmp_reset_root: free handle, th = %p, handle = %" KMP_UINTPTR_SPEC
3907 "\n",
3908 (LPVOID) & (root->r.r_uber_thread->th),
3909 root->r.r_uber_thread->th.th_info.ds.ds_thread));
3910 __kmp_free_handle(root->r.r_uber_thread->th.th_info.ds.ds_thread);
3911#endif /* KMP_OS_WINDOWS */
Jim Cownie5e8470a2013-09-27 10:38:44 +00003912
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00003913#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00003914 if (ompt_enabled && ompt_callbacks.ompt_callback(ompt_event_thread_end)) {
3915 int gtid = __kmp_get_gtid();
3916 __ompt_thread_end(ompt_thread_initial, gtid);
3917 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00003918#endif
3919
Jonathan Peyton30419822017-05-12 18:01:32 +00003920 TCW_4(__kmp_nth,
3921 __kmp_nth - 1); // __kmp_reap_thread will decrement __kmp_all_nth.
Jonathan Peytonf4392462017-07-27 20:58:41 +00003922 root->r.r_cg_nthreads--;
3923
Jonathan Peyton30419822017-05-12 18:01:32 +00003924 __kmp_reap_thread(root->r.r_uber_thread, 1);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003925
Jonathan Peyton30419822017-05-12 18:01:32 +00003926 // We canot put root thread to __kmp_thread_pool, so we have to reap it istead
3927 // of freeing.
3928 root->r.r_uber_thread = NULL;
3929 /* mark root as no longer in use */
3930 root->r.r_begin = FALSE;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003931
Jonathan Peyton30419822017-05-12 18:01:32 +00003932 return n;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003933}
3934
Jonathan Peyton30419822017-05-12 18:01:32 +00003935void __kmp_unregister_root_current_thread(int gtid) {
3936 KA_TRACE(1, ("__kmp_unregister_root_current_thread: enter T#%d\n", gtid));
3937 /* this lock should be ok, since unregister_root_current_thread is never
3938 called during an abort, only during a normal close. furthermore, if you
3939 have the forkjoin lock, you should never try to get the initz lock */
3940 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
3941 if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
3942 KC_TRACE(10, ("__kmp_unregister_root_current_thread: already finished, "
3943 "exiting T#%d\n",
3944 gtid));
3945 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
3946 return;
3947 }
3948 kmp_root_t *root = __kmp_root[gtid];
Jim Cownie77c2a632014-09-03 11:34:33 +00003949
Jonathan Peyton30419822017-05-12 18:01:32 +00003950 KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
3951 KMP_ASSERT(KMP_UBER_GTID(gtid));
3952 KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root);
3953 KMP_ASSERT(root->r.r_active == FALSE);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003954
Jonathan Peyton30419822017-05-12 18:01:32 +00003955 KMP_MB();
Jim Cownie5e8470a2013-09-27 10:38:44 +00003956
Jonathan Peytondf6818b2016-06-14 17:57:47 +00003957#if OMP_45_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00003958 kmp_info_t *thread = __kmp_threads[gtid];
3959 kmp_team_t *team = thread->th.th_team;
3960 kmp_task_team_t *task_team = thread->th.th_task_team;
Andrey Churbanov535b6fa2015-05-07 17:41:51 +00003961
Jonathan Peyton30419822017-05-12 18:01:32 +00003962 // we need to wait for the proxy tasks before finishing the thread
3963 if (task_team != NULL && task_team->tt.tt_found_proxy_tasks) {
Jonathan Peyton6d247f72015-09-10 21:33:50 +00003964#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00003965 // the runtime is shutting down so we won't report any events
3966 thread->th.ompt_thread_info.state = ompt_state_undefined;
Jonathan Peyton6d247f72015-09-10 21:33:50 +00003967#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00003968 __kmp_task_team_wait(thread, team USE_ITT_BUILD_ARG(NULL));
3969 }
Andrey Churbanov535b6fa2015-05-07 17:41:51 +00003970#endif
3971
Jonathan Peyton30419822017-05-12 18:01:32 +00003972 __kmp_reset_root(gtid, root);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003973
Jonathan Peyton30419822017-05-12 18:01:32 +00003974 /* free up this thread slot */
3975 __kmp_gtid_set_specific(KMP_GTID_DNE);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003976#ifdef KMP_TDATA_GTID
Jonathan Peyton30419822017-05-12 18:01:32 +00003977 __kmp_gtid = KMP_GTID_DNE;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003978#endif
3979
Jonathan Peyton30419822017-05-12 18:01:32 +00003980 KMP_MB();
3981 KC_TRACE(10,
3982 ("__kmp_unregister_root_current_thread: T#%d unregistered\n", gtid));
Jim Cownie5e8470a2013-09-27 10:38:44 +00003983
Jonathan Peyton30419822017-05-12 18:01:32 +00003984 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
Jim Cownie5e8470a2013-09-27 10:38:44 +00003985}
3986
Jonathan Peyton2321d572015-06-08 19:25:25 +00003987#if KMP_OS_WINDOWS
Jim Cownie5e8470a2013-09-27 10:38:44 +00003988/* __kmp_forkjoin_lock must be already held
Jonathan Peyton30419822017-05-12 18:01:32 +00003989 Unregisters a root thread that is not the current thread. Returns the number
3990 of __kmp_threads entries freed as a result. */
3991static int __kmp_unregister_root_other_thread(int gtid) {
3992 kmp_root_t *root = __kmp_root[gtid];
3993 int r;
Jim Cownie5e8470a2013-09-27 10:38:44 +00003994
Jonathan Peyton30419822017-05-12 18:01:32 +00003995 KA_TRACE(1, ("__kmp_unregister_root_other_thread: enter T#%d\n", gtid));
3996 KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
3997 KMP_ASSERT(KMP_UBER_GTID(gtid));
3998 KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root);
3999 KMP_ASSERT(root->r.r_active == FALSE);
Jim Cownie5e8470a2013-09-27 10:38:44 +00004000
Jonathan Peyton30419822017-05-12 18:01:32 +00004001 r = __kmp_reset_root(gtid, root);
4002 KC_TRACE(10,
4003 ("__kmp_unregister_root_other_thread: T#%d unregistered\n", gtid));
4004 return r;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004005}
Jonathan Peyton2321d572015-06-08 19:25:25 +00004006#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00004007
Jim Cownie5e8470a2013-09-27 10:38:44 +00004008#if KMP_DEBUG
4009void __kmp_task_info() {
4010
Jonathan Peyton30419822017-05-12 18:01:32 +00004011 kmp_int32 gtid = __kmp_entry_gtid();
4012 kmp_int32 tid = __kmp_tid_from_gtid(gtid);
4013 kmp_info_t *this_thr = __kmp_threads[gtid];
4014 kmp_team_t *steam = this_thr->th.th_serial_team;
4015 kmp_team_t *team = this_thr->th.th_team;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004016
Jonathan Peyton30419822017-05-12 18:01:32 +00004017 __kmp_printf("__kmp_task_info: gtid=%d tid=%d t_thread=%p team=%p curtask=%p "
4018 "ptask=%p\n",
4019 gtid, tid, this_thr, team, this_thr->th.th_current_task,
4020 team->t.t_implicit_task_taskdata[tid].td_parent);
Jim Cownie5e8470a2013-09-27 10:38:44 +00004021}
4022#endif // KMP_DEBUG
4023
Jonathan Peyton30419822017-05-12 18:01:32 +00004024/* TODO optimize with one big memclr, take out what isn't needed, split
4025 responsibility to workers as much as possible, and delay initialization of
4026 features as much as possible */
4027static void __kmp_initialize_info(kmp_info_t *this_thr, kmp_team_t *team,
4028 int tid, int gtid) {
4029 /* this_thr->th.th_info.ds.ds_gtid is setup in
4030 kmp_allocate_thread/create_worker.
4031 this_thr->th.th_serial_team is setup in __kmp_allocate_thread */
4032 kmp_info_t *master = team->t.t_threads[0];
4033 KMP_DEBUG_ASSERT(this_thr != NULL);
4034 KMP_DEBUG_ASSERT(this_thr->th.th_serial_team);
4035 KMP_DEBUG_ASSERT(team);
4036 KMP_DEBUG_ASSERT(team->t.t_threads);
4037 KMP_DEBUG_ASSERT(team->t.t_dispatch);
4038 KMP_DEBUG_ASSERT(master);
4039 KMP_DEBUG_ASSERT(master->th.th_root);
Jim Cownie5e8470a2013-09-27 10:38:44 +00004040
Jonathan Peyton30419822017-05-12 18:01:32 +00004041 KMP_MB();
Jim Cownie5e8470a2013-09-27 10:38:44 +00004042
Jonathan Peyton30419822017-05-12 18:01:32 +00004043 TCW_SYNC_PTR(this_thr->th.th_team, team);
Jim Cownie5e8470a2013-09-27 10:38:44 +00004044
Jonathan Peyton30419822017-05-12 18:01:32 +00004045 this_thr->th.th_info.ds.ds_tid = tid;
4046 this_thr->th.th_set_nproc = 0;
4047 if (__kmp_tasking_mode != tskm_immediate_exec)
4048 // When tasking is possible, threads are not safe to reap until they are
4049 // done tasking; this will be set when tasking code is exited in wait
4050 this_thr->th.th_reap_state = KMP_NOT_SAFE_TO_REAP;
4051 else // no tasking --> always safe to reap
4052 this_thr->th.th_reap_state = KMP_SAFE_TO_REAP;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004053#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00004054 this_thr->th.th_set_proc_bind = proc_bind_default;
4055#if KMP_AFFINITY_SUPPORTED
4056 this_thr->th.th_new_place = this_thr->th.th_current_place;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004057#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00004058#endif
4059 this_thr->th.th_root = master->th.th_root;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004060
Jonathan Peyton30419822017-05-12 18:01:32 +00004061 /* setup the thread's cache of the team structure */
4062 this_thr->th.th_team_nproc = team->t.t_nproc;
4063 this_thr->th.th_team_master = master;
4064 this_thr->th.th_team_serialized = team->t.t_serialized;
4065 TCW_PTR(this_thr->th.th_sleep_loc, NULL);
Jim Cownie5e8470a2013-09-27 10:38:44 +00004066
Jonathan Peyton30419822017-05-12 18:01:32 +00004067 KMP_DEBUG_ASSERT(team->t.t_implicit_task_taskdata);
Jim Cownie5e8470a2013-09-27 10:38:44 +00004068
Jonathan Peyton30419822017-05-12 18:01:32 +00004069 KF_TRACE(10, ("__kmp_initialize_info1: T#%d:%d this_thread=%p curtask=%p\n",
4070 tid, gtid, this_thr, this_thr->th.th_current_task));
Jim Cownie5e8470a2013-09-27 10:38:44 +00004071
Jonathan Peyton30419822017-05-12 18:01:32 +00004072 __kmp_init_implicit_task(this_thr->th.th_team_master->th.th_ident, this_thr,
4073 team, tid, TRUE);
Jim Cownie5e8470a2013-09-27 10:38:44 +00004074
Jonathan Peyton30419822017-05-12 18:01:32 +00004075 KF_TRACE(10, ("__kmp_initialize_info2: T#%d:%d this_thread=%p curtask=%p\n",
4076 tid, gtid, this_thr, this_thr->th.th_current_task));
4077 // TODO: Initialize ICVs from parent; GEH - isn't that already done in
4078 // __kmp_initialize_team()?
Jim Cownie5e8470a2013-09-27 10:38:44 +00004079
Jonathan Peyton30419822017-05-12 18:01:32 +00004080 /* TODO no worksharing in speculative threads */
4081 this_thr->th.th_dispatch = &team->t.t_dispatch[tid];
Jim Cownie5e8470a2013-09-27 10:38:44 +00004082
Jonathan Peyton30419822017-05-12 18:01:32 +00004083 this_thr->th.th_local.this_construct = 0;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004084
4085#ifdef BUILD_TV
Jonathan Peyton30419822017-05-12 18:01:32 +00004086 this_thr->th.th_local.tv_data = 0;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004087#endif
4088
Jonathan Peyton30419822017-05-12 18:01:32 +00004089 if (!this_thr->th.th_pri_common) {
4090 this_thr->th.th_pri_common =
4091 (struct common_table *)__kmp_allocate(sizeof(struct common_table));
4092 if (__kmp_storage_map) {
4093 __kmp_print_storage_map_gtid(
4094 gtid, this_thr->th.th_pri_common, this_thr->th.th_pri_common + 1,
4095 sizeof(struct common_table), "th_%d.th_pri_common\n", gtid);
Jim Cownie5e8470a2013-09-27 10:38:44 +00004096 }; // if
Jonathan Peyton30419822017-05-12 18:01:32 +00004097 this_thr->th.th_pri_head = NULL;
4098 }; // if
Jim Cownie5e8470a2013-09-27 10:38:44 +00004099
Jonathan Peyton30419822017-05-12 18:01:32 +00004100 /* Initialize dynamic dispatch */
4101 {
4102 volatile kmp_disp_t *dispatch = this_thr->th.th_dispatch;
4103 // Use team max_nproc since this will never change for the team.
4104 size_t disp_size =
4105 sizeof(dispatch_private_info_t) *
4106 (team->t.t_max_nproc == 1 ? 1 : __kmp_dispatch_num_buffers);
4107 KD_TRACE(10, ("__kmp_initialize_info: T#%d max_nproc: %d\n", gtid,
4108 team->t.t_max_nproc));
4109 KMP_ASSERT(dispatch);
4110 KMP_DEBUG_ASSERT(team->t.t_dispatch);
4111 KMP_DEBUG_ASSERT(dispatch == &team->t.t_dispatch[tid]);
Jim Cownie5e8470a2013-09-27 10:38:44 +00004112
Jonathan Peyton30419822017-05-12 18:01:32 +00004113 dispatch->th_disp_index = 0;
Jonathan Peytondf6818b2016-06-14 17:57:47 +00004114#if OMP_45_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00004115 dispatch->th_doacross_buf_idx = 0;
Jonathan Peyton71909c52016-03-02 22:42:06 +00004116#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00004117 if (!dispatch->th_disp_buffer) {
4118 dispatch->th_disp_buffer =
4119 (dispatch_private_info_t *)__kmp_allocate(disp_size);
Jim Cownie5e8470a2013-09-27 10:38:44 +00004120
Jonathan Peyton30419822017-05-12 18:01:32 +00004121 if (__kmp_storage_map) {
4122 __kmp_print_storage_map_gtid(
4123 gtid, &dispatch->th_disp_buffer[0],
4124 &dispatch->th_disp_buffer[team->t.t_max_nproc == 1
4125 ? 1
4126 : __kmp_dispatch_num_buffers],
4127 disp_size, "th_%d.th_dispatch.th_disp_buffer "
4128 "(team_%d.t_dispatch[%d].th_disp_buffer)",
4129 gtid, team->t.t_id, gtid);
4130 }
4131 } else {
4132 memset(&dispatch->th_disp_buffer[0], '\0', disp_size);
Jim Cownie5e8470a2013-09-27 10:38:44 +00004133 }
4134
Jonathan Peyton30419822017-05-12 18:01:32 +00004135 dispatch->th_dispatch_pr_current = 0;
4136 dispatch->th_dispatch_sh_current = 0;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004137
Jonathan Peyton30419822017-05-12 18:01:32 +00004138 dispatch->th_deo_fcn = 0; /* ORDERED */
4139 dispatch->th_dxo_fcn = 0; /* END ORDERED */
4140 }
Andrey Churbanov6d224db2015-02-10 18:37:43 +00004141
Jonathan Peyton30419822017-05-12 18:01:32 +00004142 this_thr->th.th_next_pool = NULL;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004143
Jonathan Peyton30419822017-05-12 18:01:32 +00004144 if (!this_thr->th.th_task_state_memo_stack) {
4145 size_t i;
4146 this_thr->th.th_task_state_memo_stack =
4147 (kmp_uint8 *)__kmp_allocate(4 * sizeof(kmp_uint8));
4148 this_thr->th.th_task_state_top = 0;
4149 this_thr->th.th_task_state_stack_sz = 4;
4150 for (i = 0; i < this_thr->th.th_task_state_stack_sz;
4151 ++i) // zero init the stack
4152 this_thr->th.th_task_state_memo_stack[i] = 0;
4153 }
4154
4155 KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
4156 KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
4157
4158 KMP_MB();
Jim Cownie5e8470a2013-09-27 10:38:44 +00004159}
4160
Jonathan Peyton30419822017-05-12 18:01:32 +00004161/* allocate a new thread for the requesting team. this is only called from
4162 within a forkjoin critical section. we will first try to get an available
4163 thread from the thread pool. if none is available, we will fork a new one
4164 assuming we are able to create a new one. this should be assured, as the
4165 caller should check on this first. */
4166kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
4167 int new_tid) {
4168 kmp_team_t *serial_team;
4169 kmp_info_t *new_thr;
4170 int new_gtid;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004171
Jonathan Peyton30419822017-05-12 18:01:32 +00004172 KA_TRACE(20, ("__kmp_allocate_thread: T#%d\n", __kmp_get_gtid()));
4173 KMP_DEBUG_ASSERT(root && team);
Jim Cownie4cc4bb42014-10-07 16:25:50 +00004174#if !KMP_NESTED_HOT_TEAMS
Jonathan Peyton30419822017-05-12 18:01:32 +00004175 KMP_DEBUG_ASSERT(KMP_MASTER_GTID(__kmp_get_gtid()));
Jim Cownie4cc4bb42014-10-07 16:25:50 +00004176#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00004177 KMP_MB();
Jim Cownie5e8470a2013-09-27 10:38:44 +00004178
Jonathan Peyton30419822017-05-12 18:01:32 +00004179 /* first, try to get one from the thread pool */
4180 if (__kmp_thread_pool) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00004181
Andrey Churbanovc47afcd2017-07-03 11:24:08 +00004182 new_thr = CCAST(kmp_info_t *, __kmp_thread_pool);
Jonathan Peyton30419822017-05-12 18:01:32 +00004183 __kmp_thread_pool = (volatile kmp_info_t *)new_thr->th.th_next_pool;
4184 if (new_thr == __kmp_thread_pool_insert_pt) {
4185 __kmp_thread_pool_insert_pt = NULL;
4186 }
4187 TCW_4(new_thr->th.th_in_pool, FALSE);
4188 // Don't touch th_active_in_pool or th_active.
4189 // The worker thread adjusts those flags as it sleeps/awakens.
4190 __kmp_thread_pool_nth--;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004191
Jonathan Peyton30419822017-05-12 18:01:32 +00004192 KA_TRACE(20, ("__kmp_allocate_thread: T#%d using thread T#%d\n",
4193 __kmp_get_gtid(), new_thr->th.th_info.ds.ds_gtid));
4194 KMP_ASSERT(!new_thr->th.th_team);
4195 KMP_DEBUG_ASSERT(__kmp_nth < __kmp_threads_capacity);
4196 KMP_DEBUG_ASSERT(__kmp_thread_pool_nth >= 0);
Jim Cownie5e8470a2013-09-27 10:38:44 +00004197
Jonathan Peyton30419822017-05-12 18:01:32 +00004198 /* setup the thread structure */
4199 __kmp_initialize_info(new_thr, team, new_tid,
4200 new_thr->th.th_info.ds.ds_gtid);
4201 KMP_DEBUG_ASSERT(new_thr->th.th_serial_team);
Jim Cownie5e8470a2013-09-27 10:38:44 +00004202
Jonathan Peyton30419822017-05-12 18:01:32 +00004203 TCW_4(__kmp_nth, __kmp_nth + 1);
Jonathan Peytonf4392462017-07-27 20:58:41 +00004204 root->r.r_cg_nthreads++;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004205
Jonathan Peyton30419822017-05-12 18:01:32 +00004206 new_thr->th.th_task_state = 0;
4207 new_thr->th.th_task_state_top = 0;
4208 new_thr->th.th_task_state_stack_sz = 4;
Andrey Churbanov6d224db2015-02-10 18:37:43 +00004209
Jim Cownie5e8470a2013-09-27 10:38:44 +00004210#ifdef KMP_ADJUST_BLOCKTIME
Jonathan Peyton30419822017-05-12 18:01:32 +00004211 /* Adjust blocktime back to zero if necessary */
4212 /* Middle initialization might not have occurred yet */
4213 if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
4214 if (__kmp_nth > __kmp_avail_proc) {
4215 __kmp_zero_bt = TRUE;
4216 }
4217 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00004218#endif /* KMP_ADJUST_BLOCKTIME */
4219
Jim Cownie4cc4bb42014-10-07 16:25:50 +00004220#if KMP_DEBUG
Jonathan Peyton30419822017-05-12 18:01:32 +00004221 // If thread entered pool via __kmp_free_thread, wait_flag should !=
4222 // KMP_BARRIER_PARENT_FLAG.
Jim Cownie4cc4bb42014-10-07 16:25:50 +00004223 int b;
Jonathan Peyton30419822017-05-12 18:01:32 +00004224 kmp_balign_t *balign = new_thr->th.th_bar;
4225 for (b = 0; b < bs_last_barrier; ++b)
4226 KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
Jim Cownie5e8470a2013-09-27 10:38:44 +00004227#endif
4228
Jonathan Peyton30419822017-05-12 18:01:32 +00004229 KF_TRACE(10, ("__kmp_allocate_thread: T#%d using thread %p T#%d\n",
4230 __kmp_get_gtid(), new_thr, new_thr->th.th_info.ds.ds_gtid));
Jim Cownie5e8470a2013-09-27 10:38:44 +00004231
Jim Cownie5e8470a2013-09-27 10:38:44 +00004232 KMP_MB();
4233 return new_thr;
Jonathan Peyton30419822017-05-12 18:01:32 +00004234 }
4235
4236 /* no, well fork a new one */
4237 KMP_ASSERT(__kmp_nth == __kmp_all_nth);
4238 KMP_ASSERT(__kmp_all_nth < __kmp_threads_capacity);
4239
4240#if KMP_USE_MONITOR
4241 // If this is the first worker thread the RTL is creating, then also
4242 // launch the monitor thread. We try to do this as early as possible.
4243 if (!TCR_4(__kmp_init_monitor)) {
4244 __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
4245 if (!TCR_4(__kmp_init_monitor)) {
4246 KF_TRACE(10, ("before __kmp_create_monitor\n"));
4247 TCW_4(__kmp_init_monitor, 1);
4248 __kmp_create_monitor(&__kmp_monitor);
4249 KF_TRACE(10, ("after __kmp_create_monitor\n"));
4250#if KMP_OS_WINDOWS
4251 // AC: wait until monitor has started. This is a fix for CQ232808.
4252 // The reason is that if the library is loaded/unloaded in a loop with
4253 // small (parallel) work in between, then there is high probability that
4254 // monitor thread started after the library shutdown. At shutdown it is
4255 // too late to cope with the problem, because when the master is in
4256 // DllMain (process detach) the monitor has no chances to start (it is
4257 // blocked), and master has no means to inform the monitor that the
4258 // library has gone, because all the memory which the monitor can access
4259 // is going to be released/reset.
4260 while (TCR_4(__kmp_init_monitor) < 2) {
4261 KMP_YIELD(TRUE);
4262 }
4263 KF_TRACE(10, ("after monitor thread has started\n"));
4264#endif
4265 }
4266 __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
4267 }
4268#endif
4269
4270 KMP_MB();
4271 for (new_gtid = 1; TCR_PTR(__kmp_threads[new_gtid]) != NULL; ++new_gtid) {
4272 KMP_DEBUG_ASSERT(new_gtid < __kmp_threads_capacity);
4273 }
4274
4275 /* allocate space for it. */
4276 new_thr = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t));
4277
4278 TCW_SYNC_PTR(__kmp_threads[new_gtid], new_thr);
4279
4280 if (__kmp_storage_map) {
4281 __kmp_print_thread_storage_map(new_thr, new_gtid);
4282 }
4283
4284 // add the reserve serialized team, initialized from the team's master thread
4285 {
4286 kmp_internal_control_t r_icvs = __kmp_get_x_global_icvs(team);
4287 KF_TRACE(10, ("__kmp_allocate_thread: before th_serial/serial_team\n"));
4288 new_thr->th.th_serial_team = serial_team =
4289 (kmp_team_t *)__kmp_allocate_team(root, 1, 1,
4290#if OMPT_SUPPORT
4291 0, // root parallel id
4292#endif
4293#if OMP_40_ENABLED
4294 proc_bind_default,
4295#endif
4296 &r_icvs, 0 USE_NESTED_HOT_ARG(NULL));
4297 }
4298 KMP_ASSERT(serial_team);
4299 serial_team->t.t_serialized = 0; // AC: the team created in reserve, not for
4300 // execution (it is unused for now).
4301 serial_team->t.t_threads[0] = new_thr;
4302 KF_TRACE(10,
4303 ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n",
4304 new_thr));
4305
4306 /* setup the thread structures */
4307 __kmp_initialize_info(new_thr, team, new_tid, new_gtid);
4308
4309#if USE_FAST_MEMORY
4310 __kmp_initialize_fast_memory(new_thr);
4311#endif /* USE_FAST_MEMORY */
4312
4313#if KMP_USE_BGET
4314 KMP_DEBUG_ASSERT(new_thr->th.th_local.bget_data == NULL);
4315 __kmp_initialize_bget(new_thr);
4316#endif
4317
4318 __kmp_init_random(new_thr); // Initialize random number generator
4319
4320 /* Initialize these only once when thread is grabbed for a team allocation */
4321 KA_TRACE(20,
4322 ("__kmp_allocate_thread: T#%d init go fork=%u, plain=%u\n",
4323 __kmp_get_gtid(), KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
4324
4325 int b;
4326 kmp_balign_t *balign = new_thr->th.th_bar;
4327 for (b = 0; b < bs_last_barrier; ++b) {
4328 balign[b].bb.b_go = KMP_INIT_BARRIER_STATE;
4329 balign[b].bb.team = NULL;
4330 balign[b].bb.wait_flag = KMP_BARRIER_NOT_WAITING;
4331 balign[b].bb.use_oncore_barrier = 0;
4332 }
4333
4334 new_thr->th.th_spin_here = FALSE;
4335 new_thr->th.th_next_waiting = 0;
4336
4337#if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
4338 new_thr->th.th_current_place = KMP_PLACE_UNDEFINED;
4339 new_thr->th.th_new_place = KMP_PLACE_UNDEFINED;
4340 new_thr->th.th_first_place = KMP_PLACE_UNDEFINED;
4341 new_thr->th.th_last_place = KMP_PLACE_UNDEFINED;
4342#endif
4343
4344 TCW_4(new_thr->th.th_in_pool, FALSE);
4345 new_thr->th.th_active_in_pool = FALSE;
4346 TCW_4(new_thr->th.th_active, TRUE);
4347
4348 /* adjust the global counters */
4349 __kmp_all_nth++;
4350 __kmp_nth++;
4351
Jonathan Peytonf4392462017-07-27 20:58:41 +00004352 root->r.r_cg_nthreads++;
4353
Jonathan Peyton30419822017-05-12 18:01:32 +00004354 // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low
4355 // numbers of procs, and method #2 (keyed API call) for higher numbers.
4356 if (__kmp_adjust_gtid_mode) {
4357 if (__kmp_all_nth >= __kmp_tls_gtid_min) {
4358 if (TCR_4(__kmp_gtid_mode) != 2) {
4359 TCW_4(__kmp_gtid_mode, 2);
4360 }
4361 } else {
4362 if (TCR_4(__kmp_gtid_mode) != 1) {
4363 TCW_4(__kmp_gtid_mode, 1);
4364 }
4365 }
4366 }
4367
4368#ifdef KMP_ADJUST_BLOCKTIME
4369 /* Adjust blocktime back to zero if necessary */
4370 /* Middle initialization might not have occurred yet */
4371 if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
4372 if (__kmp_nth > __kmp_avail_proc) {
4373 __kmp_zero_bt = TRUE;
4374 }
4375 }
4376#endif /* KMP_ADJUST_BLOCKTIME */
4377
4378 /* actually fork it and create the new worker thread */
4379 KF_TRACE(
4380 10, ("__kmp_allocate_thread: before __kmp_create_worker: %p\n", new_thr));
4381 __kmp_create_worker(new_gtid, new_thr, __kmp_stksize);
4382 KF_TRACE(10,
4383 ("__kmp_allocate_thread: after __kmp_create_worker: %p\n", new_thr));
4384
4385 KA_TRACE(20, ("__kmp_allocate_thread: T#%d forked T#%d\n", __kmp_get_gtid(),
4386 new_gtid));
4387 KMP_MB();
4388 return new_thr;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004389}
4390
Jonathan Peyton30419822017-05-12 18:01:32 +00004391/* Reinitialize team for reuse.
4392 The hot team code calls this case at every fork barrier, so EPCC barrier
4393 test are extremely sensitive to changes in it, esp. writes to the team
4394 struct, which cause a cache invalidation in all threads.
4395 IF YOU TOUCH THIS ROUTINE, RUN EPCC C SYNCBENCH ON A BIG-IRON MACHINE!!! */
4396static void __kmp_reinitialize_team(kmp_team_t *team,
4397 kmp_internal_control_t *new_icvs,
4398 ident_t *loc) {
4399 KF_TRACE(10, ("__kmp_reinitialize_team: enter this_thread=%p team=%p\n",
4400 team->t.t_threads[0], team));
4401 KMP_DEBUG_ASSERT(team && new_icvs);
4402 KMP_DEBUG_ASSERT((!TCR_4(__kmp_init_parallel)) || new_icvs->nproc);
4403 KMP_CHECK_UPDATE(team->t.t_ident, loc);
Jim Cownie5e8470a2013-09-27 10:38:44 +00004404
Jonathan Peyton30419822017-05-12 18:01:32 +00004405 KMP_CHECK_UPDATE(team->t.t_id, KMP_GEN_TEAM_ID());
Jonathan Peyton30419822017-05-12 18:01:32 +00004406 // Copy ICVs to the master thread's implicit taskdata
4407 __kmp_init_implicit_task(loc, team->t.t_threads[0], team, 0, FALSE);
4408 copy_icvs(&team->t.t_implicit_task_taskdata[0].td_icvs, new_icvs);
Jim Cownie181b4bb2013-12-23 17:28:57 +00004409
Jonathan Peyton30419822017-05-12 18:01:32 +00004410 KF_TRACE(10, ("__kmp_reinitialize_team: exit this_thread=%p team=%p\n",
4411 team->t.t_threads[0], team));
Jim Cownie181b4bb2013-12-23 17:28:57 +00004412}
4413
Jonathan Peyton30419822017-05-12 18:01:32 +00004414/* Initialize the team data structure.
4415 This assumes the t_threads and t_max_nproc are already set.
4416 Also, we don't touch the arguments */
4417static void __kmp_initialize_team(kmp_team_t *team, int new_nproc,
4418 kmp_internal_control_t *new_icvs,
4419 ident_t *loc) {
4420 KF_TRACE(10, ("__kmp_initialize_team: enter: team=%p\n", team));
Jim Cownie5e8470a2013-09-27 10:38:44 +00004421
Jonathan Peyton30419822017-05-12 18:01:32 +00004422 /* verify */
4423 KMP_DEBUG_ASSERT(team);
4424 KMP_DEBUG_ASSERT(new_nproc <= team->t.t_max_nproc);
4425 KMP_DEBUG_ASSERT(team->t.t_threads);
4426 KMP_MB();
Jim Cownie181b4bb2013-12-23 17:28:57 +00004427
Jonathan Peyton30419822017-05-12 18:01:32 +00004428 team->t.t_master_tid = 0; /* not needed */
4429 /* team->t.t_master_bar; not needed */
4430 team->t.t_serialized = new_nproc > 1 ? 0 : 1;
4431 team->t.t_nproc = new_nproc;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004432
Jonathan Peyton30419822017-05-12 18:01:32 +00004433 /* team->t.t_parent = NULL; TODO not needed & would mess up hot team */
4434 team->t.t_next_pool = NULL;
4435 /* memset( team->t.t_threads, 0, sizeof(kmp_info_t*)*new_nproc ); would mess
4436 * up hot team */
Jim Cownie5e8470a2013-09-27 10:38:44 +00004437
Jonathan Peyton30419822017-05-12 18:01:32 +00004438 TCW_SYNC_PTR(team->t.t_pkfn, NULL); /* not needed */
4439 team->t.t_invoke = NULL; /* not needed */
Jim Cownie5e8470a2013-09-27 10:38:44 +00004440
Jonathan Peyton30419822017-05-12 18:01:32 +00004441 // TODO???: team->t.t_max_active_levels = new_max_active_levels;
4442 team->t.t_sched = new_icvs->sched;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004443
4444#if KMP_ARCH_X86 || KMP_ARCH_X86_64
Jonathan Peyton30419822017-05-12 18:01:32 +00004445 team->t.t_fp_control_saved = FALSE; /* not needed */
4446 team->t.t_x87_fpu_control_word = 0; /* not needed */
4447 team->t.t_mxcsr = 0; /* not needed */
Jim Cownie5e8470a2013-09-27 10:38:44 +00004448#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4449
Jonathan Peyton30419822017-05-12 18:01:32 +00004450 team->t.t_construct = 0;
4451 __kmp_init_lock(&team->t.t_single_lock);
Jim Cownie5e8470a2013-09-27 10:38:44 +00004452
Jonathan Peyton30419822017-05-12 18:01:32 +00004453 team->t.t_ordered.dt.t_value = 0;
4454 team->t.t_master_active = FALSE;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004455
Jonathan Peyton30419822017-05-12 18:01:32 +00004456 memset(&team->t.t_taskq, '\0', sizeof(kmp_taskq_t));
Jim Cownie5e8470a2013-09-27 10:38:44 +00004457
4458#ifdef KMP_DEBUG
Jonathan Peyton30419822017-05-12 18:01:32 +00004459 team->t.t_copypriv_data = NULL; /* not necessary, but nice for debugging */
Jim Cownie5e8470a2013-09-27 10:38:44 +00004460#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00004461 team->t.t_copyin_counter = 0; /* for barrier-free copyin implementation */
Jim Cownie5e8470a2013-09-27 10:38:44 +00004462
Jonathan Peyton30419822017-05-12 18:01:32 +00004463 team->t.t_control_stack_top = NULL;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004464
Jonathan Peyton30419822017-05-12 18:01:32 +00004465 __kmp_reinitialize_team(team, new_icvs, loc);
Jim Cownie5e8470a2013-09-27 10:38:44 +00004466
Jonathan Peyton30419822017-05-12 18:01:32 +00004467 KMP_MB();
4468 KF_TRACE(10, ("__kmp_initialize_team: exit: team=%p\n", team));
Jim Cownie5e8470a2013-09-27 10:38:44 +00004469}
4470
Alp Toker98758b02014-03-02 04:12:06 +00004471#if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
Jim Cownie5e8470a2013-09-27 10:38:44 +00004472/* Sets full mask for thread and returns old mask, no changes to structures. */
4473static void
Jonathan Peyton30419822017-05-12 18:01:32 +00004474__kmp_set_thread_affinity_mask_full_tmp(kmp_affin_mask_t *old_mask) {
4475 if (KMP_AFFINITY_CAPABLE()) {
4476 int status;
4477 if (old_mask != NULL) {
4478 status = __kmp_get_system_affinity(old_mask, TRUE);
4479 int error = errno;
4480 if (status != 0) {
4481 __kmp_msg(kmp_ms_fatal, KMP_MSG(ChangeThreadAffMaskError),
4482 KMP_ERR(error), __kmp_msg_null);
4483 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00004484 }
Jonathan Peyton30419822017-05-12 18:01:32 +00004485 __kmp_set_system_affinity(__kmp_affin_fullMask, TRUE);
4486 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00004487}
4488#endif
4489
Alp Toker98758b02014-03-02 04:12:06 +00004490#if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
Jim Cownie5e8470a2013-09-27 10:38:44 +00004491
Jim Cownie5e8470a2013-09-27 10:38:44 +00004492// __kmp_partition_places() is the heart of the OpenMP 4.0 affinity mechanism.
4493// It calculats the worker + master thread's partition based upon the parent
Alp Toker8f2d3f02014-02-24 10:40:15 +00004494// thread's partition, and binds each worker to a thread in their partition.
Jim Cownie5e8470a2013-09-27 10:38:44 +00004495// The master thread's partition should already include its current binding.
Jonathan Peyton30419822017-05-12 18:01:32 +00004496static void __kmp_partition_places(kmp_team_t *team, int update_master_only) {
4497 // Copy the master thread's place partion to the team struct
4498 kmp_info_t *master_th = team->t.t_threads[0];
4499 KMP_DEBUG_ASSERT(master_th != NULL);
4500 kmp_proc_bind_t proc_bind = team->t.t_proc_bind;
4501 int first_place = master_th->th.th_first_place;
4502 int last_place = master_th->th.th_last_place;
4503 int masters_place = master_th->th.th_current_place;
4504 team->t.t_first_place = first_place;
4505 team->t.t_last_place = last_place;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004506
Jonathan Peyton30419822017-05-12 18:01:32 +00004507 KA_TRACE(20, ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) "
4508 "bound to place %d partition = [%d,%d]\n",
4509 proc_bind, __kmp_gtid_from_thread(team->t.t_threads[0]),
4510 team->t.t_id, masters_place, first_place, last_place));
Jim Cownie5e8470a2013-09-27 10:38:44 +00004511
Jonathan Peyton30419822017-05-12 18:01:32 +00004512 switch (proc_bind) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00004513
Jonathan Peyton30419822017-05-12 18:01:32 +00004514 case proc_bind_default:
4515 // serial teams might have the proc_bind policy set to proc_bind_default. It
4516 // doesn't matter, as we don't rebind master thread for any proc_bind policy
4517 KMP_DEBUG_ASSERT(team->t.t_nproc == 1);
4518 break;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004519
Jonathan Peyton30419822017-05-12 18:01:32 +00004520 case proc_bind_master: {
4521 int f;
4522 int n_th = team->t.t_nproc;
4523 for (f = 1; f < n_th; f++) {
4524 kmp_info_t *th = team->t.t_threads[f];
4525 KMP_DEBUG_ASSERT(th != NULL);
4526 th->th.th_first_place = first_place;
4527 th->th.th_last_place = last_place;
4528 th->th.th_new_place = masters_place;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004529
Jonathan Peyton30419822017-05-12 18:01:32 +00004530 KA_TRACE(100, ("__kmp_partition_places: master: T#%d(%d:%d) place %d "
4531 "partition = [%d,%d]\n",
4532 __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,
4533 f, masters_place, first_place, last_place));
Jim Cownie5e8470a2013-09-27 10:38:44 +00004534 }
Jonathan Peyton30419822017-05-12 18:01:32 +00004535 } break;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004536
Jonathan Peyton30419822017-05-12 18:01:32 +00004537 case proc_bind_close: {
4538 int f;
4539 int n_th = team->t.t_nproc;
4540 int n_places;
4541 if (first_place <= last_place) {
4542 n_places = last_place - first_place + 1;
4543 } else {
4544 n_places = __kmp_affinity_num_masks - first_place + last_place + 1;
4545 }
4546 if (n_th <= n_places) {
4547 int place = masters_place;
4548 for (f = 1; f < n_th; f++) {
4549 kmp_info_t *th = team->t.t_threads[f];
4550 KMP_DEBUG_ASSERT(th != NULL);
4551
4552 if (place == last_place) {
4553 place = first_place;
4554 } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4555 place = 0;
4556 } else {
4557 place++;
4558 }
4559 th->th.th_first_place = first_place;
4560 th->th.th_last_place = last_place;
4561 th->th.th_new_place = place;
4562
4563 KA_TRACE(100, ("__kmp_partition_places: close: T#%d(%d:%d) place %d "
4564 "partition = [%d,%d]\n",
4565 __kmp_gtid_from_thread(team->t.t_threads[f]),
4566 team->t.t_id, f, place, first_place, last_place));
4567 }
4568 } else {
4569 int S, rem, gap, s_count;
4570 S = n_th / n_places;
4571 s_count = 0;
4572 rem = n_th - (S * n_places);
4573 gap = rem > 0 ? n_places / rem : n_places;
4574 int place = masters_place;
4575 int gap_ct = gap;
4576 for (f = 0; f < n_th; f++) {
4577 kmp_info_t *th = team->t.t_threads[f];
4578 KMP_DEBUG_ASSERT(th != NULL);
4579
4580 th->th.th_first_place = first_place;
4581 th->th.th_last_place = last_place;
4582 th->th.th_new_place = place;
4583 s_count++;
4584
4585 if ((s_count == S) && rem && (gap_ct == gap)) {
4586 // do nothing, add an extra thread to place on next iteration
4587 } else if ((s_count == S + 1) && rem && (gap_ct == gap)) {
4588 // we added an extra thread to this place; move to next place
4589 if (place == last_place) {
4590 place = first_place;
4591 } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4592 place = 0;
4593 } else {
4594 place++;
4595 }
4596 s_count = 0;
4597 gap_ct = 1;
4598 rem--;
4599 } else if (s_count == S) { // place full; don't add extra
4600 if (place == last_place) {
4601 place = first_place;
4602 } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4603 place = 0;
4604 } else {
4605 place++;
4606 }
4607 gap_ct++;
4608 s_count = 0;
4609 }
4610
4611 KA_TRACE(100,
4612 ("__kmp_partition_places: close: T#%d(%d:%d) place %d "
4613 "partition = [%d,%d]\n",
4614 __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id, f,
4615 th->th.th_new_place, first_place, last_place));
4616 }
4617 KMP_DEBUG_ASSERT(place == masters_place);
4618 }
4619 } break;
4620
4621 case proc_bind_spread: {
4622 int f;
4623 int n_th = team->t.t_nproc;
4624 int n_places;
4625 int thidx;
4626 if (first_place <= last_place) {
4627 n_places = last_place - first_place + 1;
4628 } else {
4629 n_places = __kmp_affinity_num_masks - first_place + last_place + 1;
4630 }
4631 if (n_th <= n_places) {
Paul Osmialowskia0162792017-08-10 23:04:11 +00004632 int place = -1;
Jonathan Peyton30419822017-05-12 18:01:32 +00004633
Paul Osmialowskia0162792017-08-10 23:04:11 +00004634 if (n_places != static_cast<int>(__kmp_affinity_num_masks)) {
4635 int S = n_places / n_th;
4636 int s_count, rem, gap, gap_ct;
4637
4638 place = masters_place;
4639 rem = n_places - n_th * S;
4640 gap = rem ? n_th / rem : 1;
4641 gap_ct = gap;
4642 thidx = n_th;
4643 if (update_master_only == 1)
4644 thidx = 1;
4645 for (f = 0; f < thidx; f++) {
4646 kmp_info_t *th = team->t.t_threads[f];
4647 KMP_DEBUG_ASSERT(th != NULL);
4648
4649 th->th.th_first_place = place;
4650 th->th.th_new_place = place;
4651 s_count = 1;
4652 while (s_count < S) {
4653 if (place == last_place) {
4654 place = first_place;
4655 } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4656 place = 0;
4657 } else {
4658 place++;
4659 }
4660 s_count++;
4661 }
4662 if (rem && (gap_ct == gap)) {
4663 if (place == last_place) {
4664 place = first_place;
4665 } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4666 place = 0;
4667 } else {
4668 place++;
4669 }
4670 rem--;
4671 gap_ct = 0;
4672 }
4673 th->th.th_last_place = place;
4674 gap_ct++;
4675
Jonathan Peyton30419822017-05-12 18:01:32 +00004676 if (place == last_place) {
4677 place = first_place;
4678 } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4679 place = 0;
4680 } else {
4681 place++;
4682 }
Paul Osmialowskia0162792017-08-10 23:04:11 +00004683
4684 KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4685 "partition = [%d,%d], __kmp_affinity_num_masks: %u\n",
4686 __kmp_gtid_from_thread(team->t.t_threads[f]),
4687 team->t.t_id, f, th->th.th_new_place,
4688 th->th.th_first_place, th->th.th_last_place,
4689 __kmp_affinity_num_masks));
Jonathan Peyton30419822017-05-12 18:01:32 +00004690 }
Paul Osmialowskia0162792017-08-10 23:04:11 +00004691 } else {
4692 /* Having uniform space of available computation places I can create
4693 T partitions of round(P/T) size and put threads into the first
4694 place of each partition. */
4695 double current = static_cast<double>(masters_place);
4696 double spacing =
4697 (static_cast<double>(n_places + 1) / static_cast<double>(n_th));
4698 int first, last;
4699 kmp_info_t *th;
4700
4701 thidx = n_th + 1;
4702 if (update_master_only == 1)
4703 thidx = 1;
4704 for (f = 0; f < thidx; f++) {
4705 first = static_cast<int>(current);
4706 last = static_cast<int>(current + spacing) - 1;
4707 KMP_DEBUG_ASSERT(last >= first);
4708 if (first >= n_places) {
4709 if (masters_place) {
4710 first -= n_places;
4711 last -= n_places;
4712 if (first == (masters_place + 1)) {
4713 KMP_DEBUG_ASSERT(f == n_th);
4714 first--;
4715 }
4716 if (last == masters_place) {
4717 KMP_DEBUG_ASSERT(f == (n_th - 1));
4718 last--;
4719 }
4720 } else {
4721 KMP_DEBUG_ASSERT(f == n_th);
4722 first = 0;
4723 last = 0;
4724 }
Jonathan Peyton30419822017-05-12 18:01:32 +00004725 }
Paul Osmialowskia0162792017-08-10 23:04:11 +00004726 if (last >= n_places) {
4727 last = (n_places - 1);
4728 }
4729 place = first;
4730 current += spacing;
4731 if (f < n_th) {
4732 KMP_DEBUG_ASSERT(0 <= first);
4733 KMP_DEBUG_ASSERT(n_places > first);
4734 KMP_DEBUG_ASSERT(0 <= last);
4735 KMP_DEBUG_ASSERT(n_places > last);
4736 KMP_DEBUG_ASSERT(last_place >= first_place);
4737 th = team->t.t_threads[f];
4738 KMP_DEBUG_ASSERT(th);
4739 th->th.th_first_place = first;
4740 th->th.th_new_place = place;
4741 th->th.th_last_place = last;
Jonathan Peyton30419822017-05-12 18:01:32 +00004742
Paul Osmialowskia0162792017-08-10 23:04:11 +00004743 KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4744 "partition = [%d,%d], spacing = %.4f\n",
4745 __kmp_gtid_from_thread(team->t.t_threads[f]),
4746 team->t.t_id, f, th->th.th_new_place,
4747 th->th.th_first_place, th->th.th_last_place,
4748 spacing));
4749 }
Jonathan Peyton30419822017-05-12 18:01:32 +00004750 }
Jonathan Peyton30419822017-05-12 18:01:32 +00004751 }
4752 KMP_DEBUG_ASSERT(update_master_only || place == masters_place);
4753 } else {
4754 int S, rem, gap, s_count;
4755 S = n_th / n_places;
4756 s_count = 0;
4757 rem = n_th - (S * n_places);
4758 gap = rem > 0 ? n_places / rem : n_places;
4759 int place = masters_place;
4760 int gap_ct = gap;
4761 thidx = n_th;
4762 if (update_master_only == 1)
4763 thidx = 1;
4764 for (f = 0; f < thidx; f++) {
4765 kmp_info_t *th = team->t.t_threads[f];
4766 KMP_DEBUG_ASSERT(th != NULL);
4767
4768 th->th.th_first_place = place;
4769 th->th.th_last_place = place;
4770 th->th.th_new_place = place;
4771 s_count++;
4772
4773 if ((s_count == S) && rem && (gap_ct == gap)) {
4774 // do nothing, add an extra thread to place on next iteration
4775 } else if ((s_count == S + 1) && rem && (gap_ct == gap)) {
4776 // we added an extra thread to this place; move on to next place
4777 if (place == last_place) {
4778 place = first_place;
4779 } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4780 place = 0;
4781 } else {
4782 place++;
4783 }
4784 s_count = 0;
4785 gap_ct = 1;
4786 rem--;
4787 } else if (s_count == S) { // place is full; don't add extra thread
4788 if (place == last_place) {
4789 place = first_place;
4790 } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4791 place = 0;
4792 } else {
4793 place++;
4794 }
4795 gap_ct++;
4796 s_count = 0;
4797 }
4798
4799 KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4800 "partition = [%d,%d]\n",
4801 __kmp_gtid_from_thread(team->t.t_threads[f]),
4802 team->t.t_id, f, th->th.th_new_place,
4803 th->th.th_first_place, th->th.th_last_place));
4804 }
4805 KMP_DEBUG_ASSERT(update_master_only || place == masters_place);
4806 }
4807 } break;
4808
4809 default:
4810 break;
4811 }
4812
4813 KA_TRACE(20, ("__kmp_partition_places: exit T#%d\n", team->t.t_id));
Jim Cownie5e8470a2013-09-27 10:38:44 +00004814}
4815
Alp Toker98758b02014-03-02 04:12:06 +00004816#endif /* OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED */
Jim Cownie5e8470a2013-09-27 10:38:44 +00004817
Jonathan Peyton30419822017-05-12 18:01:32 +00004818/* allocate a new team data structure to use. take one off of the free pool if
4819 available */
Jim Cownie5e8470a2013-09-27 10:38:44 +00004820kmp_team_t *
Jonathan Peyton30419822017-05-12 18:01:32 +00004821__kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00004822#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00004823 ompt_parallel_id_t ompt_parallel_id,
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00004824#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00004825#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00004826 kmp_proc_bind_t new_proc_bind,
Jim Cownie5e8470a2013-09-27 10:38:44 +00004827#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00004828 kmp_internal_control_t *new_icvs,
4829 int argc USE_NESTED_HOT_ARG(kmp_info_t *master)) {
4830 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_allocate_team);
4831 int f;
4832 kmp_team_t *team;
4833 int use_hot_team = !root->r.r_active;
4834 int level = 0;
Jim Cownie5e8470a2013-09-27 10:38:44 +00004835
Jonathan Peyton30419822017-05-12 18:01:32 +00004836 KA_TRACE(20, ("__kmp_allocate_team: called\n"));
4837 KMP_DEBUG_ASSERT(new_nproc >= 1 && argc >= 0);
4838 KMP_DEBUG_ASSERT(max_nproc >= new_nproc);
4839 KMP_MB();
Jim Cownie5e8470a2013-09-27 10:38:44 +00004840
Jim Cownie4cc4bb42014-10-07 16:25:50 +00004841#if KMP_NESTED_HOT_TEAMS
Jonathan Peyton30419822017-05-12 18:01:32 +00004842 kmp_hot_team_ptr_t *hot_teams;
4843 if (master) {
4844 team = master->th.th_team;
4845 level = team->t.t_active_level;
4846 if (master->th.th_teams_microtask) { // in teams construct?
4847 if (master->th.th_teams_size.nteams > 1 &&
4848 ( // #teams > 1
4849 team->t.t_pkfn ==
4850 (microtask_t)__kmp_teams_master || // inner fork of the teams
4851 master->th.th_teams_level <
4852 team->t.t_level)) { // or nested parallel inside the teams
4853 ++level; // not increment if #teams==1, or for outer fork of the teams;
4854 // increment otherwise
4855 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00004856 }
Jonathan Peyton30419822017-05-12 18:01:32 +00004857 hot_teams = master->th.th_hot_teams;
4858 if (level < __kmp_hot_teams_max_level && hot_teams &&
4859 hot_teams[level]
4860 .hot_team) { // hot team has already been allocated for given level
4861 use_hot_team = 1;
4862 } else {
4863 use_hot_team = 0;
4864 }
4865 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00004866#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00004867 // Optimization to use a "hot" team
4868 if (use_hot_team && new_nproc > 1) {
4869 KMP_DEBUG_ASSERT(new_nproc == max_nproc);
Jim Cownie4cc4bb42014-10-07 16:25:50 +00004870#if KMP_NESTED_HOT_TEAMS
Jonathan Peyton30419822017-05-12 18:01:32 +00004871 team = hot_teams[level].hot_team;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00004872#else
Jonathan Peyton30419822017-05-12 18:01:32 +00004873 team = root->r.r_hot_team;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00004874#endif
4875#if KMP_DEBUG
Jonathan Peyton30419822017-05-12 18:01:32 +00004876 if (__kmp_tasking_mode != tskm_immediate_exec) {
4877 KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
4878 "task_team[1] = %p before reinit\n",
4879 team->t.t_task_team[0], team->t.t_task_team[1]));
4880 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00004881#endif
4882
Jonathan Peyton30419822017-05-12 18:01:32 +00004883 // Has the number of threads changed?
4884 /* Let's assume the most common case is that the number of threads is
4885 unchanged, and put that case first. */
4886 if (team->t.t_nproc == new_nproc) { // Check changes in number of threads
4887 KA_TRACE(20, ("__kmp_allocate_team: reusing hot team\n"));
4888 // This case can mean that omp_set_num_threads() was called and the hot
Jonathan Peyton642688b2017-06-01 16:46:36 +00004889 // team size was already reduced, so we check the special flag
Jonathan Peyton30419822017-05-12 18:01:32 +00004890 if (team->t.t_size_changed == -1) {
4891 team->t.t_size_changed = 1;
4892 } else {
4893 KMP_CHECK_UPDATE(team->t.t_size_changed, 0);
4894 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00004895
Jonathan Peyton30419822017-05-12 18:01:32 +00004896 // TODO???: team->t.t_max_active_levels = new_max_active_levels;
4897 kmp_r_sched_t new_sched = new_icvs->sched;
4898 if (team->t.t_sched.r_sched_type != new_sched.r_sched_type ||
4899 team->t.t_sched.chunk != new_sched.chunk)
4900 team->t.t_sched =
4901 new_sched; // set master's schedule as new run-time schedule
Jim Cownie4cc4bb42014-10-07 16:25:50 +00004902
Jonathan Peyton30419822017-05-12 18:01:32 +00004903 __kmp_reinitialize_team(team, new_icvs,
4904 root->r.r_uber_thread->th.th_ident);
Jim Cownie4cc4bb42014-10-07 16:25:50 +00004905
Jonathan Peyton30419822017-05-12 18:01:32 +00004906 KF_TRACE(10, ("__kmp_allocate_team2: T#%d, this_thread=%p team=%p\n", 0,
4907 team->t.t_threads[0], team));
4908 __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
Jim Cownie4cc4bb42014-10-07 16:25:50 +00004909
4910#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00004911#if KMP_AFFINITY_SUPPORTED
4912 if ((team->t.t_size_changed == 0) &&
4913 (team->t.t_proc_bind == new_proc_bind)) {
4914 if (new_proc_bind == proc_bind_spread) {
4915 __kmp_partition_places(
4916 team, 1); // add flag to update only master for spread
Jim Cownie4cc4bb42014-10-07 16:25:50 +00004917 }
Jonathan Peyton30419822017-05-12 18:01:32 +00004918 KA_TRACE(200, ("__kmp_allocate_team: reusing hot team #%d bindings: "
4919 "proc_bind = %d, partition = [%d,%d]\n",
4920 team->t.t_id, new_proc_bind, team->t.t_first_place,
4921 team->t.t_last_place));
4922 } else {
4923 KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
4924 __kmp_partition_places(team);
4925 }
4926#else
4927 KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
4928#endif /* KMP_AFFINITY_SUPPORTED */
4929#endif /* OMP_40_ENABLED */
4930 } else if (team->t.t_nproc > new_nproc) {
4931 KA_TRACE(20,
4932 ("__kmp_allocate_team: decreasing hot team thread count to %d\n",
4933 new_nproc));
Jim Cownie5e8470a2013-09-27 10:38:44 +00004934
Jonathan Peyton30419822017-05-12 18:01:32 +00004935 team->t.t_size_changed = 1;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00004936#if KMP_NESTED_HOT_TEAMS
Jonathan Peyton30419822017-05-12 18:01:32 +00004937 if (__kmp_hot_teams_mode == 0) {
4938 // AC: saved number of threads should correspond to team's value in this
4939 // mode, can be bigger in mode 1, when hot team has threads in reserve
4940 KMP_DEBUG_ASSERT(hot_teams[level].hot_team_nth == team->t.t_nproc);
4941 hot_teams[level].hot_team_nth = new_nproc;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00004942#endif // KMP_NESTED_HOT_TEAMS
Jonathan Peyton30419822017-05-12 18:01:32 +00004943 /* release the extra threads we don't need any more */
4944 for (f = new_nproc; f < team->t.t_nproc; f++) {
4945 KMP_DEBUG_ASSERT(team->t.t_threads[f]);
4946 if (__kmp_tasking_mode != tskm_immediate_exec) {
4947 // When decreasing team size, threads no longer in the team should
4948 // unref task team.
4949 team->t.t_threads[f]->th.th_task_team = NULL;
4950 }
4951 __kmp_free_thread(team->t.t_threads[f]);
4952 team->t.t_threads[f] = NULL;
4953 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00004954#if KMP_NESTED_HOT_TEAMS
Jonathan Peyton30419822017-05-12 18:01:32 +00004955 } // (__kmp_hot_teams_mode == 0)
4956 else {
4957 // When keeping extra threads in team, switch threads to wait on own
4958 // b_go flag
4959 for (f = new_nproc; f < team->t.t_nproc; ++f) {
4960 KMP_DEBUG_ASSERT(team->t.t_threads[f]);
4961 kmp_balign_t *balign = team->t.t_threads[f]->th.th_bar;
4962 for (int b = 0; b < bs_last_barrier; ++b) {
4963 if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG) {
4964 balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
Andrey Churbanovd6e1d7e2016-08-11 13:04:00 +00004965 }
Jonathan Peyton30419822017-05-12 18:01:32 +00004966 KMP_CHECK_UPDATE(balign[b].bb.leaf_kids, 0);
4967 }
4968 }
4969 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00004970#endif // KMP_NESTED_HOT_TEAMS
Jonathan Peyton30419822017-05-12 18:01:32 +00004971 team->t.t_nproc = new_nproc;
4972 // TODO???: team->t.t_max_active_levels = new_max_active_levels;
4973 if (team->t.t_sched.r_sched_type != new_icvs->sched.r_sched_type ||
4974 team->t.t_sched.chunk != new_icvs->sched.chunk)
4975 team->t.t_sched = new_icvs->sched;
4976 __kmp_reinitialize_team(team, new_icvs,
4977 root->r.r_uber_thread->th.th_ident);
Jim Cownie5e8470a2013-09-27 10:38:44 +00004978
Jonathan Peyton30419822017-05-12 18:01:32 +00004979 /* update the remaining threads */
4980 for (f = 0; f < new_nproc; ++f) {
4981 team->t.t_threads[f]->th.th_team_nproc = new_nproc;
4982 }
4983 // restore the current task state of the master thread: should be the
4984 // implicit task
4985 KF_TRACE(10, ("__kmp_allocate_team: T#%d, this_thread=%p team=%p\n", 0,
4986 team->t.t_threads[0], team));
Jim Cownie5e8470a2013-09-27 10:38:44 +00004987
Jonathan Peyton30419822017-05-12 18:01:32 +00004988 __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
Jim Cownie5e8470a2013-09-27 10:38:44 +00004989
4990#ifdef KMP_DEBUG
Jonathan Peyton30419822017-05-12 18:01:32 +00004991 for (f = 0; f < team->t.t_nproc; f++) {
4992 KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
4993 team->t.t_threads[f]->th.th_team_nproc ==
4994 team->t.t_nproc);
4995 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00004996#endif
4997
4998#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00004999 KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5000#if KMP_AFFINITY_SUPPORTED
5001 __kmp_partition_places(team);
Jim Cownie5e8470a2013-09-27 10:38:44 +00005002#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00005003#endif
5004 } else { // team->t.t_nproc < new_nproc
Alp Toker98758b02014-03-02 04:12:06 +00005005#if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
Jonathan Peyton30419822017-05-12 18:01:32 +00005006 kmp_affin_mask_t *old_mask;
5007 if (KMP_AFFINITY_CAPABLE()) {
5008 KMP_CPU_ALLOC(old_mask);
5009 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00005010#endif
5011
Jonathan Peyton30419822017-05-12 18:01:32 +00005012 KA_TRACE(20,
5013 ("__kmp_allocate_team: increasing hot team thread count to %d\n",
5014 new_nproc));
Jim Cownie5e8470a2013-09-27 10:38:44 +00005015
Jonathan Peyton30419822017-05-12 18:01:32 +00005016 team->t.t_size_changed = 1;
Jim Cownie5e8470a2013-09-27 10:38:44 +00005017
Jim Cownie4cc4bb42014-10-07 16:25:50 +00005018#if KMP_NESTED_HOT_TEAMS
Jonathan Peyton30419822017-05-12 18:01:32 +00005019 int avail_threads = hot_teams[level].hot_team_nth;
5020 if (new_nproc < avail_threads)
5021 avail_threads = new_nproc;
5022 kmp_info_t **other_threads = team->t.t_threads;
5023 for (f = team->t.t_nproc; f < avail_threads; ++f) {
5024 // Adjust barrier data of reserved threads (if any) of the team
5025 // Other data will be set in __kmp_initialize_info() below.
Jim Cownie5e8470a2013-09-27 10:38:44 +00005026 int b;
Jonathan Peyton30419822017-05-12 18:01:32 +00005027 kmp_balign_t *balign = other_threads[f]->th.th_bar;
5028 for (b = 0; b < bs_last_barrier; ++b) {
5029 balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5030 KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
Jonathan Peyton8fbb49a2015-07-09 18:16:58 +00005031#if USE_DEBUGGER
Jonathan Peyton30419822017-05-12 18:01:32 +00005032 balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
Jonathan Peyton8fbb49a2015-07-09 18:16:58 +00005033#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00005034 }
Jonathan Peyton30419822017-05-12 18:01:32 +00005035 }
5036 if (hot_teams[level].hot_team_nth >= new_nproc) {
5037 // we have all needed threads in reserve, no need to allocate any
5038 // this only possible in mode 1, cannot have reserved threads in mode 0
5039 KMP_DEBUG_ASSERT(__kmp_hot_teams_mode == 1);
5040 team->t.t_nproc = new_nproc; // just get reserved threads involved
5041 } else {
5042 // we may have some threads in reserve, but not enough
5043 team->t.t_nproc =
5044 hot_teams[level]
5045 .hot_team_nth; // get reserved threads involved if any
5046 hot_teams[level].hot_team_nth = new_nproc; // adjust hot team max size
5047#endif // KMP_NESTED_HOT_TEAMS
5048 if (team->t.t_max_nproc < new_nproc) {
5049 /* reallocate larger arrays */
5050 __kmp_reallocate_team_arrays(team, new_nproc);
5051 __kmp_reinitialize_team(team, new_icvs, NULL);
5052 }
5053
5054#if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5055 /* Temporarily set full mask for master thread before creation of
5056 workers. The reason is that workers inherit the affinity from master,
5057 so if a lot of workers are created on the single core quickly, they
5058 don't get a chance to set their own affinity for a long time. */
5059 __kmp_set_thread_affinity_mask_full_tmp(old_mask);
5060#endif
5061
5062 /* allocate new threads for the hot team */
5063 for (f = team->t.t_nproc; f < new_nproc; f++) {
5064 kmp_info_t *new_worker = __kmp_allocate_thread(root, team, f);
5065 KMP_DEBUG_ASSERT(new_worker);
5066 team->t.t_threads[f] = new_worker;
5067
5068 KA_TRACE(20,
5069 ("__kmp_allocate_team: team %d init T#%d arrived: "
5070 "join=%llu, plain=%llu\n",
5071 team->t.t_id, __kmp_gtid_from_tid(f, team), team->t.t_id, f,
5072 team->t.t_bar[bs_forkjoin_barrier].b_arrived,
5073 team->t.t_bar[bs_plain_barrier].b_arrived));
5074
5075 { // Initialize barrier data for new threads.
5076 int b;
5077 kmp_balign_t *balign = new_worker->th.th_bar;
5078 for (b = 0; b < bs_last_barrier; ++b) {
5079 balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5080 KMP_DEBUG_ASSERT(balign[b].bb.wait_flag !=
5081 KMP_BARRIER_PARENT_FLAG);
5082#if USE_DEBUGGER
5083 balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5084#endif
5085 }
5086 }
5087 }
5088
5089#if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5090 if (KMP_AFFINITY_CAPABLE()) {
5091 /* Restore initial master thread's affinity mask */
5092 __kmp_set_system_affinity(old_mask, TRUE);
5093 KMP_CPU_FREE(old_mask);
5094 }
5095#endif
5096#if KMP_NESTED_HOT_TEAMS
5097 } // end of check of t_nproc vs. new_nproc vs. hot_team_nth
5098#endif // KMP_NESTED_HOT_TEAMS
5099 /* make sure everyone is syncronized */
5100 int old_nproc = team->t.t_nproc; // save old value and use to update only
5101 // new threads below
5102 __kmp_initialize_team(team, new_nproc, new_icvs,
5103 root->r.r_uber_thread->th.th_ident);
5104
5105 /* reinitialize the threads */
5106 KMP_DEBUG_ASSERT(team->t.t_nproc == new_nproc);
5107 for (f = 0; f < team->t.t_nproc; ++f)
5108 __kmp_initialize_info(team->t.t_threads[f], team, f,
5109 __kmp_gtid_from_tid(f, team));
5110 if (level) { // set th_task_state for new threads in nested hot team
5111 // __kmp_initialize_info() no longer zeroes th_task_state, so we should
5112 // only need to set the th_task_state for the new threads. th_task_state
5113 // for master thread will not be accurate until after this in
5114 // __kmp_fork_call(), so we look to the master's memo_stack to get the
5115 // correct value.
5116 for (f = old_nproc; f < team->t.t_nproc; ++f)
5117 team->t.t_threads[f]->th.th_task_state =
5118 team->t.t_threads[0]->th.th_task_state_memo_stack[level];
5119 } else { // set th_task_state for new threads in non-nested hot team
5120 int old_state =
5121 team->t.t_threads[0]->th.th_task_state; // copy master's state
5122 for (f = old_nproc; f < team->t.t_nproc; ++f)
5123 team->t.t_threads[f]->th.th_task_state = old_state;
5124 }
5125
5126#ifdef KMP_DEBUG
5127 for (f = 0; f < team->t.t_nproc; ++f) {
5128 KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
5129 team->t.t_threads[f]->th.th_team_nproc ==
5130 team->t.t_nproc);
5131 }
5132#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00005133
5134#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00005135 KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5136#if KMP_AFFINITY_SUPPORTED
5137 __kmp_partition_places(team);
5138#endif
5139#endif
5140 } // Check changes in number of threads
5141
5142#if OMP_40_ENABLED
5143 kmp_info_t *master = team->t.t_threads[0];
5144 if (master->th.th_teams_microtask) {
5145 for (f = 1; f < new_nproc; ++f) {
5146 // propagate teams construct specific info to workers
5147 kmp_info_t *thr = team->t.t_threads[f];
5148 thr->th.th_teams_microtask = master->th.th_teams_microtask;
5149 thr->th.th_teams_level = master->th.th_teams_level;
5150 thr->th.th_teams_size = master->th.th_teams_size;
5151 }
5152 }
5153#endif /* OMP_40_ENABLED */
5154#if KMP_NESTED_HOT_TEAMS
5155 if (level) {
5156 // Sync barrier state for nested hot teams, not needed for outermost hot
5157 // team.
5158 for (f = 1; f < new_nproc; ++f) {
5159 kmp_info_t *thr = team->t.t_threads[f];
5160 int b;
5161 kmp_balign_t *balign = thr->th.th_bar;
5162 for (b = 0; b < bs_last_barrier; ++b) {
5163 balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5164 KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
5165#if USE_DEBUGGER
5166 balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5167#endif
5168 }
5169 }
5170 }
5171#endif // KMP_NESTED_HOT_TEAMS
5172
5173 /* reallocate space for arguments if necessary */
5174 __kmp_alloc_argv_entries(argc, team, TRUE);
5175 KMP_CHECK_UPDATE(team->t.t_argc, argc);
5176 // The hot team re-uses the previous task team,
5177 // if untouched during the previous release->gather phase.
5178
5179 KF_TRACE(10, (" hot_team = %p\n", team));
5180
5181#if KMP_DEBUG
5182 if (__kmp_tasking_mode != tskm_immediate_exec) {
5183 KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
5184 "task_team[1] = %p after reinit\n",
5185 team->t.t_task_team[0], team->t.t_task_team[1]));
5186 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00005187#endif
5188
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00005189#if OMPT_SUPPORT
5190 __ompt_team_assign_id(team, ompt_parallel_id);
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00005191#endif
5192
Jim Cownie5e8470a2013-09-27 10:38:44 +00005193 KMP_MB();
5194
Jim Cownie5e8470a2013-09-27 10:38:44 +00005195 return team;
Jonathan Peyton30419822017-05-12 18:01:32 +00005196 }
5197
5198 /* next, let's try to take one from the team pool */
5199 KMP_MB();
Andrey Churbanovc47afcd2017-07-03 11:24:08 +00005200 for (team = CCAST(kmp_team_t *, __kmp_team_pool); (team);) {
Jonathan Peyton30419822017-05-12 18:01:32 +00005201 /* TODO: consider resizing undersized teams instead of reaping them, now
5202 that we have a resizing mechanism */
5203 if (team->t.t_max_nproc >= max_nproc) {
5204 /* take this team from the team pool */
5205 __kmp_team_pool = team->t.t_next_pool;
5206
5207 /* setup the team for fresh use */
5208 __kmp_initialize_team(team, new_nproc, new_icvs, NULL);
5209
5210 KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and "
5211 "task_team[1] %p to NULL\n",
5212 &team->t.t_task_team[0], &team->t.t_task_team[1]));
5213 team->t.t_task_team[0] = NULL;
5214 team->t.t_task_team[1] = NULL;
5215
5216 /* reallocate space for arguments if necessary */
5217 __kmp_alloc_argv_entries(argc, team, TRUE);
5218 KMP_CHECK_UPDATE(team->t.t_argc, argc);
5219
5220 KA_TRACE(
5221 20, ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",
5222 team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
5223 { // Initialize barrier data.
5224 int b;
5225 for (b = 0; b < bs_last_barrier; ++b) {
5226 team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE;
5227#if USE_DEBUGGER
5228 team->t.t_bar[b].b_master_arrived = 0;
5229 team->t.t_bar[b].b_team_arrived = 0;
5230#endif
5231 }
5232 }
5233
5234#if OMP_40_ENABLED
5235 team->t.t_proc_bind = new_proc_bind;
5236#endif
5237
5238 KA_TRACE(20, ("__kmp_allocate_team: using team from pool %d.\n",
5239 team->t.t_id));
5240
5241#if OMPT_SUPPORT
5242 __ompt_team_assign_id(team, ompt_parallel_id);
5243#endif
5244
5245 KMP_MB();
5246
5247 return team;
5248 }
5249
5250/* reap team if it is too small, then loop back and check the next one */
5251// not sure if this is wise, but, will be redone during the hot-teams rewrite.
5252/* TODO: Use technique to find the right size hot-team, don't reap them */
5253 team = __kmp_reap_team(team);
5254 __kmp_team_pool = team;
5255 }
5256
5257 /* nothing available in the pool, no matter, make a new team! */
5258 KMP_MB();
5259 team = (kmp_team_t *)__kmp_allocate(sizeof(kmp_team_t));
5260
5261 /* and set it up */
5262 team->t.t_max_nproc = max_nproc;
5263 /* NOTE well, for some reason allocating one big buffer and dividing it up
5264 seems to really hurt performance a lot on the P4, so, let's not use this */
5265 __kmp_allocate_team_arrays(team, max_nproc);
5266
5267 KA_TRACE(20, ("__kmp_allocate_team: making a new team\n"));
5268 __kmp_initialize_team(team, new_nproc, new_icvs, NULL);
5269
5270 KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and task_team[1] "
5271 "%p to NULL\n",
5272 &team->t.t_task_team[0], &team->t.t_task_team[1]));
5273 team->t.t_task_team[0] = NULL; // to be removed, as __kmp_allocate zeroes
5274 // memory, no need to duplicate
5275 team->t.t_task_team[1] = NULL; // to be removed, as __kmp_allocate zeroes
5276 // memory, no need to duplicate
5277
5278 if (__kmp_storage_map) {
5279 __kmp_print_team_storage_map("team", team, team->t.t_id, new_nproc);
5280 }
5281
5282 /* allocate space for arguments */
5283 __kmp_alloc_argv_entries(argc, team, FALSE);
5284 team->t.t_argc = argc;
5285
5286 KA_TRACE(20,
5287 ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",
5288 team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
5289 { // Initialize barrier data.
5290 int b;
5291 for (b = 0; b < bs_last_barrier; ++b) {
5292 team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE;
5293#if USE_DEBUGGER
5294 team->t.t_bar[b].b_master_arrived = 0;
5295 team->t.t_bar[b].b_team_arrived = 0;
5296#endif
5297 }
5298 }
5299
5300#if OMP_40_ENABLED
5301 team->t.t_proc_bind = new_proc_bind;
5302#endif
5303
5304#if OMPT_SUPPORT
5305 __ompt_team_assign_id(team, ompt_parallel_id);
5306 team->t.ompt_serialized_team_info = NULL;
5307#endif
5308
5309 KMP_MB();
5310
5311 KA_TRACE(20, ("__kmp_allocate_team: done creating a new team %d.\n",
5312 team->t.t_id));
5313
5314 return team;
Jim Cownie5e8470a2013-09-27 10:38:44 +00005315}
5316
5317/* TODO implement hot-teams at all levels */
5318/* TODO implement lazy thread release on demand (disband request) */
5319
5320/* free the team. return it to the team pool. release all the threads
5321 * associated with it */
Jonathan Peyton30419822017-05-12 18:01:32 +00005322void __kmp_free_team(kmp_root_t *root,
5323 kmp_team_t *team USE_NESTED_HOT_ARG(kmp_info_t *master)) {
5324 int f;
5325 KA_TRACE(20, ("__kmp_free_team: T#%d freeing team %d\n", __kmp_get_gtid(),
5326 team->t.t_id));
Jim Cownie5e8470a2013-09-27 10:38:44 +00005327
Jonathan Peyton30419822017-05-12 18:01:32 +00005328 /* verify state */
5329 KMP_DEBUG_ASSERT(root);
5330 KMP_DEBUG_ASSERT(team);
5331 KMP_DEBUG_ASSERT(team->t.t_nproc <= team->t.t_max_nproc);
5332 KMP_DEBUG_ASSERT(team->t.t_threads);
Jim Cownie5e8470a2013-09-27 10:38:44 +00005333
Jonathan Peyton30419822017-05-12 18:01:32 +00005334 int use_hot_team = team == root->r.r_hot_team;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00005335#if KMP_NESTED_HOT_TEAMS
Jonathan Peyton30419822017-05-12 18:01:32 +00005336 int level;
5337 kmp_hot_team_ptr_t *hot_teams;
5338 if (master) {
5339 level = team->t.t_active_level - 1;
5340 if (master->th.th_teams_microtask) { // in teams construct?
5341 if (master->th.th_teams_size.nteams > 1) {
5342 ++level; // level was not increased in teams construct for
5343 // team_of_masters
5344 }
5345 if (team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
5346 master->th.th_teams_level == team->t.t_level) {
5347 ++level; // level was not increased in teams construct for
5348 // team_of_workers before the parallel
5349 } // team->t.t_level will be increased inside parallel
Jim Cownie4cc4bb42014-10-07 16:25:50 +00005350 }
Jonathan Peyton30419822017-05-12 18:01:32 +00005351 hot_teams = master->th.th_hot_teams;
5352 if (level < __kmp_hot_teams_max_level) {
5353 KMP_DEBUG_ASSERT(team == hot_teams[level].hot_team);
5354 use_hot_team = 1;
5355 }
5356 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00005357#endif // KMP_NESTED_HOT_TEAMS
5358
Jonathan Peyton30419822017-05-12 18:01:32 +00005359 /* team is done working */
5360 TCW_SYNC_PTR(team->t.t_pkfn,
5361 NULL); // Important for Debugging Support Library.
5362 team->t.t_copyin_counter = 0; // init counter for possible reuse
5363 // Do not reset pointer to parent team to NULL for hot teams.
Jim Cownie5e8470a2013-09-27 10:38:44 +00005364
Jonathan Peyton30419822017-05-12 18:01:32 +00005365 /* if we are non-hot team, release our threads */
5366 if (!use_hot_team) {
5367 if (__kmp_tasking_mode != tskm_immediate_exec) {
5368 // Wait for threads to reach reapable state
5369 for (f = 1; f < team->t.t_nproc; ++f) {
5370 KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5371 kmp_info_t *th = team->t.t_threads[f];
5372 volatile kmp_uint32 *state = &th->th.th_reap_state;
5373 while (*state != KMP_SAFE_TO_REAP) {
Andrey Churbanov581490e2017-02-06 18:53:32 +00005374#if KMP_OS_WINDOWS
Jonathan Peyton30419822017-05-12 18:01:32 +00005375 // On Windows a thread can be killed at any time, check this
5376 DWORD ecode;
5377 if (!__kmp_is_thread_alive(th, &ecode)) {
5378 *state = KMP_SAFE_TO_REAP; // reset the flag for dead thread
5379 break;
5380 }
Andrey Churbanov581490e2017-02-06 18:53:32 +00005381#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00005382 // first check if thread is sleeping
5383 kmp_flag_64 fl(&th->th.th_bar[bs_forkjoin_barrier].bb.b_go, th);
5384 if (fl.is_sleeping())
5385 fl.resume(__kmp_gtid_from_thread(th));
5386 KMP_CPU_PAUSE();
5387 }
5388 }
Andrey Churbanov581490e2017-02-06 18:53:32 +00005389
Jonathan Peyton30419822017-05-12 18:01:32 +00005390 // Delete task teams
5391 int tt_idx;
5392 for (tt_idx = 0; tt_idx < 2; ++tt_idx) {
5393 kmp_task_team_t *task_team = team->t.t_task_team[tt_idx];
5394 if (task_team != NULL) {
5395 for (f = 0; f < team->t.t_nproc;
5396 ++f) { // Have all threads unref task teams
5397 team->t.t_threads[f]->th.th_task_team = NULL;
5398 }
5399 KA_TRACE(
5400 20,
5401 ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n",
5402 __kmp_get_gtid(), task_team, team->t.t_id));
Jonathan Peytonbaaccfa2015-11-16 22:48:41 +00005403#if KMP_NESTED_HOT_TEAMS
Jonathan Peyton30419822017-05-12 18:01:32 +00005404 __kmp_free_task_team(master, task_team);
Jonathan Peytonbaaccfa2015-11-16 22:48:41 +00005405#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00005406 team->t.t_task_team[tt_idx] = NULL;
Jim Cownie5e8470a2013-09-27 10:38:44 +00005407 }
Jonathan Peyton30419822017-05-12 18:01:32 +00005408 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00005409 }
5410
Jonathan Peyton30419822017-05-12 18:01:32 +00005411 // Reset pointer to parent team only for non-hot teams.
5412 team->t.t_parent = NULL;
5413 team->t.t_level = 0;
5414 team->t.t_active_level = 0;
Jim Cownie5e8470a2013-09-27 10:38:44 +00005415
Jonathan Peyton30419822017-05-12 18:01:32 +00005416 /* free the worker threads */
5417 for (f = 1; f < team->t.t_nproc; ++f) {
5418 KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5419 __kmp_free_thread(team->t.t_threads[f]);
5420 team->t.t_threads[f] = NULL;
5421 }
5422
5423 /* put the team back in the team pool */
5424 /* TODO limit size of team pool, call reap_team if pool too large */
Andrey Churbanovc47afcd2017-07-03 11:24:08 +00005425 team->t.t_next_pool = CCAST(kmp_team_t *, __kmp_team_pool);
Jonathan Peyton30419822017-05-12 18:01:32 +00005426 __kmp_team_pool = (volatile kmp_team_t *)team;
5427 }
5428
5429 KMP_MB();
5430}
Jim Cownie5e8470a2013-09-27 10:38:44 +00005431
5432/* reap the team. destroy it, reclaim all its resources and free its memory */
Jonathan Peyton30419822017-05-12 18:01:32 +00005433kmp_team_t *__kmp_reap_team(kmp_team_t *team) {
5434 kmp_team_t *next_pool = team->t.t_next_pool;
Jim Cownie5e8470a2013-09-27 10:38:44 +00005435
Jonathan Peyton30419822017-05-12 18:01:32 +00005436 KMP_DEBUG_ASSERT(team);
5437 KMP_DEBUG_ASSERT(team->t.t_dispatch);
5438 KMP_DEBUG_ASSERT(team->t.t_disp_buffer);
5439 KMP_DEBUG_ASSERT(team->t.t_threads);
5440 KMP_DEBUG_ASSERT(team->t.t_argv);
Jim Cownie5e8470a2013-09-27 10:38:44 +00005441
Jonathan Peyton30419822017-05-12 18:01:32 +00005442 /* TODO clean the threads that are a part of this? */
Jim Cownie5e8470a2013-09-27 10:38:44 +00005443
Jonathan Peyton30419822017-05-12 18:01:32 +00005444 /* free stuff */
5445 __kmp_free_team_arrays(team);
5446 if (team->t.t_argv != &team->t.t_inline_argv[0])
5447 __kmp_free((void *)team->t.t_argv);
5448 __kmp_free(team);
Jim Cownie5e8470a2013-09-27 10:38:44 +00005449
Jonathan Peyton30419822017-05-12 18:01:32 +00005450 KMP_MB();
5451 return next_pool;
Jim Cownie5e8470a2013-09-27 10:38:44 +00005452}
5453
Jim Cownie5e8470a2013-09-27 10:38:44 +00005454// Free the thread. Don't reap it, just place it on the pool of available
5455// threads.
5456//
5457// Changes for Quad issue 527845: We need a predictable OMP tid <-> gtid
5458// binding for the affinity mechanism to be useful.
5459//
5460// Now, we always keep the free list (__kmp_thread_pool) sorted by gtid.
5461// However, we want to avoid a potential performance problem by always
5462// scanning through the list to find the correct point at which to insert
5463// the thread (potential N**2 behavior). To do this we keep track of the
5464// last place a thread struct was inserted (__kmp_thread_pool_insert_pt).
5465// With single-level parallelism, threads will always be added to the tail
5466// of the list, kept track of by __kmp_thread_pool_insert_pt. With nested
5467// parallelism, all bets are off and we may need to scan through the entire
5468// free list.
5469//
5470// This change also has a potentially large performance benefit, for some
5471// applications. Previously, as threads were freed from the hot team, they
5472// would be placed back on the free list in inverse order. If the hot team
5473// grew back to it's original size, then the freed thread would be placed
5474// back on the hot team in reverse order. This could cause bad cache
5475// locality problems on programs where the size of the hot team regularly
5476// grew and shrunk.
5477//
5478// Now, for single-level parallelism, the OMP tid is alway == gtid.
Jonathan Peyton30419822017-05-12 18:01:32 +00005479void __kmp_free_thread(kmp_info_t *this_th) {
5480 int gtid;
5481 kmp_info_t **scan;
Jonathan Peytonf4392462017-07-27 20:58:41 +00005482 kmp_root_t *root = this_th->th.th_root;
Jim Cownie5e8470a2013-09-27 10:38:44 +00005483
Jonathan Peyton30419822017-05-12 18:01:32 +00005484 KA_TRACE(20, ("__kmp_free_thread: T#%d putting T#%d back on free pool.\n",
5485 __kmp_get_gtid(), this_th->th.th_info.ds.ds_gtid));
Jim Cownie5e8470a2013-09-27 10:38:44 +00005486
Jonathan Peyton30419822017-05-12 18:01:32 +00005487 KMP_DEBUG_ASSERT(this_th);
Jim Cownie5e8470a2013-09-27 10:38:44 +00005488
Jonathan Peyton30419822017-05-12 18:01:32 +00005489 // When moving thread to pool, switch thread to wait on own b_go flag, and
5490 // uninitialized (NULL team).
5491 int b;
5492 kmp_balign_t *balign = this_th->th.th_bar;
5493 for (b = 0; b < bs_last_barrier; ++b) {
5494 if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG)
5495 balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
5496 balign[b].bb.team = NULL;
5497 balign[b].bb.leaf_kids = 0;
5498 }
5499 this_th->th.th_task_state = 0;
5500
5501 /* put thread back on the free pool */
5502 TCW_PTR(this_th->th.th_team, NULL);
5503 TCW_PTR(this_th->th.th_root, NULL);
5504 TCW_PTR(this_th->th.th_dispatch, NULL); /* NOT NEEDED */
5505
5506 // If the __kmp_thread_pool_insert_pt is already past the new insert
5507 // point, then we need to re-scan the entire list.
5508 gtid = this_th->th.th_info.ds.ds_gtid;
5509 if (__kmp_thread_pool_insert_pt != NULL) {
5510 KMP_DEBUG_ASSERT(__kmp_thread_pool != NULL);
5511 if (__kmp_thread_pool_insert_pt->th.th_info.ds.ds_gtid > gtid) {
5512 __kmp_thread_pool_insert_pt = NULL;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00005513 }
Jonathan Peyton30419822017-05-12 18:01:32 +00005514 }
Jim Cownie4cc4bb42014-10-07 16:25:50 +00005515
Jonathan Peyton30419822017-05-12 18:01:32 +00005516 // Scan down the list to find the place to insert the thread.
5517 // scan is the address of a link in the list, possibly the address of
5518 // __kmp_thread_pool itself.
5519 //
5520 // In the absence of nested parallism, the for loop will have 0 iterations.
5521 if (__kmp_thread_pool_insert_pt != NULL) {
5522 scan = &(__kmp_thread_pool_insert_pt->th.th_next_pool);
5523 } else {
Andrey Churbanovc47afcd2017-07-03 11:24:08 +00005524 scan = CCAST(kmp_info_t **, &__kmp_thread_pool);
Jonathan Peyton30419822017-05-12 18:01:32 +00005525 }
5526 for (; (*scan != NULL) && ((*scan)->th.th_info.ds.ds_gtid < gtid);
5527 scan = &((*scan)->th.th_next_pool))
5528 ;
Jim Cownie5e8470a2013-09-27 10:38:44 +00005529
Jonathan Peyton30419822017-05-12 18:01:32 +00005530 // Insert the new element on the list, and set __kmp_thread_pool_insert_pt
5531 // to its address.
5532 TCW_PTR(this_th->th.th_next_pool, *scan);
5533 __kmp_thread_pool_insert_pt = *scan = this_th;
5534 KMP_DEBUG_ASSERT((this_th->th.th_next_pool == NULL) ||
5535 (this_th->th.th_info.ds.ds_gtid <
5536 this_th->th.th_next_pool->th.th_info.ds.ds_gtid));
5537 TCW_4(this_th->th.th_in_pool, TRUE);
5538 __kmp_thread_pool_nth++;
Jim Cownie5e8470a2013-09-27 10:38:44 +00005539
Jonathan Peyton30419822017-05-12 18:01:32 +00005540 TCW_4(__kmp_nth, __kmp_nth - 1);
Jonathan Peytonf4392462017-07-27 20:58:41 +00005541 root->r.r_cg_nthreads--;
Jim Cownie5e8470a2013-09-27 10:38:44 +00005542
5543#ifdef KMP_ADJUST_BLOCKTIME
Jonathan Peyton30419822017-05-12 18:01:32 +00005544 /* Adjust blocktime back to user setting or default if necessary */
5545 /* Middle initialization might never have occurred */
5546 if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
5547 KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
5548 if (__kmp_nth <= __kmp_avail_proc) {
5549 __kmp_zero_bt = FALSE;
Jim Cownie5e8470a2013-09-27 10:38:44 +00005550 }
Jonathan Peyton30419822017-05-12 18:01:32 +00005551 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00005552#endif /* KMP_ADJUST_BLOCKTIME */
5553
Jonathan Peyton30419822017-05-12 18:01:32 +00005554 KMP_MB();
Jim Cownie5e8470a2013-09-27 10:38:44 +00005555}
5556
Jim Cownie5e8470a2013-09-27 10:38:44 +00005557/* ------------------------------------------------------------------------ */
5558
Jonathan Peyton30419822017-05-12 18:01:32 +00005559void *__kmp_launch_thread(kmp_info_t *this_thr) {
5560 int gtid = this_thr->th.th_info.ds.ds_gtid;
5561 /* void *stack_data;*/
5562 kmp_team_t *(*volatile pteam);
Jim Cownie5e8470a2013-09-27 10:38:44 +00005563
Jonathan Peyton30419822017-05-12 18:01:32 +00005564 KMP_MB();
5565 KA_TRACE(10, ("__kmp_launch_thread: T#%d start\n", gtid));
Jim Cownie5e8470a2013-09-27 10:38:44 +00005566
Jonathan Peyton30419822017-05-12 18:01:32 +00005567 if (__kmp_env_consistency_check) {
5568 this_thr->th.th_cons = __kmp_allocate_cons_stack(gtid); // ATT: Memory leak?
5569 }
5570
5571#if OMPT_SUPPORT
5572 if (ompt_enabled) {
5573 this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5574 this_thr->th.ompt_thread_info.wait_id = 0;
5575 this_thr->th.ompt_thread_info.idle_frame = __builtin_frame_address(0);
5576 if (ompt_callbacks.ompt_callback(ompt_event_thread_begin)) {
5577 __ompt_thread_begin(ompt_thread_worker, gtid);
Jim Cownie5e8470a2013-09-27 10:38:44 +00005578 }
Jonathan Peyton30419822017-05-12 18:01:32 +00005579 }
5580#endif
5581
5582 /* This is the place where threads wait for work */
5583 while (!TCR_4(__kmp_global.g.g_done)) {
5584 KMP_DEBUG_ASSERT(this_thr == __kmp_threads[gtid]);
5585 KMP_MB();
5586
5587 /* wait for work to do */
5588 KA_TRACE(20, ("__kmp_launch_thread: T#%d waiting for work\n", gtid));
Jim Cownie5e8470a2013-09-27 10:38:44 +00005589
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00005590#if OMPT_SUPPORT
Jonathan Peytonb68a85d2015-09-21 18:11:22 +00005591 if (ompt_enabled) {
Jonathan Peyton30419822017-05-12 18:01:32 +00005592 this_thr->th.ompt_thread_info.state = ompt_state_idle;
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00005593 }
5594#endif
5595
Jonathan Peyton30419822017-05-12 18:01:32 +00005596 /* No tid yet since not part of a team */
5597 __kmp_fork_barrier(gtid, KMP_GTID_DNE);
5598
5599#if OMPT_SUPPORT
5600 if (ompt_enabled) {
5601 this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5602 }
5603#endif
5604
5605 pteam = (kmp_team_t * (*))(&this_thr->th.th_team);
5606
5607 /* have we been allocated? */
5608 if (TCR_SYNC_PTR(*pteam) && !TCR_4(__kmp_global.g.g_done)) {
5609#if OMPT_SUPPORT
5610 ompt_task_info_t *task_info;
5611 ompt_parallel_id_t my_parallel_id;
5612 if (ompt_enabled) {
5613 task_info = __ompt_get_taskinfo(0);
5614 my_parallel_id = (*pteam)->t.ompt_team_info.parallel_id;
5615 }
5616#endif
5617 /* we were just woken up, so run our new task */
5618 if (TCR_SYNC_PTR((*pteam)->t.t_pkfn) != NULL) {
5619 int rc;
5620 KA_TRACE(20,
5621 ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n",
5622 gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),
5623 (*pteam)->t.t_pkfn));
5624
5625 updateHWFPControl(*pteam);
5626
5627#if OMPT_SUPPORT
5628 if (ompt_enabled) {
5629 this_thr->th.ompt_thread_info.state = ompt_state_work_parallel;
5630 // Initialize OMPT task id for implicit task.
5631 int tid = __kmp_tid_from_gtid(gtid);
5632 task_info->task_id = __ompt_task_id_new(tid);
5633 }
5634#endif
5635
5636 {
5637 KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
5638 KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
5639 rc = (*pteam)->t.t_invoke(gtid);
5640 }
5641 KMP_ASSERT(rc);
5642
5643#if OMPT_SUPPORT
5644 if (ompt_enabled) {
5645 /* no frame set while outside task */
5646 task_info->frame.exit_runtime_frame = NULL;
5647
5648 this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5649 }
5650#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00005651 KMP_MB();
Jonathan Peyton30419822017-05-12 18:01:32 +00005652 KA_TRACE(20, ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n",
5653 gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),
5654 (*pteam)->t.t_pkfn));
5655 }
5656 /* join barrier after parallel region */
5657 __kmp_join_barrier(gtid);
Jonathan Peytonb4c73d82016-01-26 21:45:21 +00005658#if OMPT_SUPPORT && OMPT_TRACE
Jonathan Peyton30419822017-05-12 18:01:32 +00005659 if (ompt_enabled) {
5660 if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) {
5661 // don't access *pteam here: it may have already been freed
5662 // by the master thread behind the barrier (possible race)
5663 ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)(
5664 my_parallel_id, task_info->task_id);
Jim Cownie5e8470a2013-09-27 10:38:44 +00005665 }
Jonathan Peyton30419822017-05-12 18:01:32 +00005666 task_info->frame.exit_runtime_frame = NULL;
5667 task_info->task_id = 0;
5668 }
5669#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00005670 }
Jonathan Peyton30419822017-05-12 18:01:32 +00005671 }
5672 TCR_SYNC_PTR((intptr_t)__kmp_global.g.g_done);
Jim Cownie5e8470a2013-09-27 10:38:44 +00005673
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00005674#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00005675 if (ompt_enabled && ompt_callbacks.ompt_callback(ompt_event_thread_end)) {
5676 __ompt_thread_end(ompt_thread_worker, gtid);
5677 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00005678#endif
5679
Jonathan Peyton30419822017-05-12 18:01:32 +00005680 this_thr->th.th_task_team = NULL;
5681 /* run the destructors for the threadprivate data for this thread */
5682 __kmp_common_destroy_gtid(gtid);
Jim Cownie5e8470a2013-09-27 10:38:44 +00005683
Jonathan Peyton30419822017-05-12 18:01:32 +00005684 KA_TRACE(10, ("__kmp_launch_thread: T#%d done\n", gtid));
5685 KMP_MB();
5686 return this_thr;
Jim Cownie5e8470a2013-09-27 10:38:44 +00005687}
5688
5689/* ------------------------------------------------------------------------ */
Jim Cownie5e8470a2013-09-27 10:38:44 +00005690
Jonathan Peyton30419822017-05-12 18:01:32 +00005691void __kmp_internal_end_dest(void *specific_gtid) {
5692#if KMP_COMPILER_ICC
5693#pragma warning(push)
5694#pragma warning(disable : 810) // conversion from "void *" to "int" may lose
5695// significant bits
5696#endif
5697 // Make sure no significant bits are lost
5698 int gtid = (kmp_intptr_t)specific_gtid - 1;
5699#if KMP_COMPILER_ICC
5700#pragma warning(pop)
5701#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00005702
Jonathan Peyton30419822017-05-12 18:01:32 +00005703 KA_TRACE(30, ("__kmp_internal_end_dest: T#%d\n", gtid));
5704 /* NOTE: the gtid is stored as gitd+1 in the thread-local-storage
5705 * this is because 0 is reserved for the nothing-stored case */
Jim Cownie5e8470a2013-09-27 10:38:44 +00005706
Jonathan Peyton30419822017-05-12 18:01:32 +00005707 /* josh: One reason for setting the gtid specific data even when it is being
5708 destroyed by pthread is to allow gtid lookup through thread specific data
5709 (__kmp_gtid_get_specific). Some of the code, especially stat code,
5710 that gets executed in the call to __kmp_internal_end_thread, actually
5711 gets the gtid through the thread specific data. Setting it here seems
5712 rather inelegant and perhaps wrong, but allows __kmp_internal_end_thread
5713 to run smoothly.
5714 todo: get rid of this after we remove the dependence on
5715 __kmp_gtid_get_specific */
5716 if (gtid >= 0 && KMP_UBER_GTID(gtid))
5717 __kmp_gtid_set_specific(gtid);
5718#ifdef KMP_TDATA_GTID
5719 __kmp_gtid = gtid;
5720#endif
5721 __kmp_internal_end_thread(gtid);
Jim Cownie5e8470a2013-09-27 10:38:44 +00005722}
5723
Jonathan Peyton99016992015-05-26 17:32:53 +00005724#if KMP_OS_UNIX && KMP_DYNAMIC_LIB
Jim Cownie5e8470a2013-09-27 10:38:44 +00005725
Jonathan Peyton30419822017-05-12 18:01:32 +00005726// 2009-09-08 (lev): It looks the destructor does not work. In simple test cases
5727// destructors work perfectly, but in real libomp.so I have no evidence it is
5728// ever called. However, -fini linker option in makefile.mk works fine.
Jim Cownie5e8470a2013-09-27 10:38:44 +00005729
Jonathan Peyton30419822017-05-12 18:01:32 +00005730__attribute__((destructor)) void __kmp_internal_end_dtor(void) {
5731 __kmp_internal_end_atexit();
Jim Cownie5e8470a2013-09-27 10:38:44 +00005732}
5733
Jonathan Peyton30419822017-05-12 18:01:32 +00005734void __kmp_internal_end_fini(void) { __kmp_internal_end_atexit(); }
Jim Cownie5e8470a2013-09-27 10:38:44 +00005735
5736#endif
5737
Jonathan Peyton30419822017-05-12 18:01:32 +00005738/* [Windows] josh: when the atexit handler is called, there may still be more
5739 than one thread alive */
5740void __kmp_internal_end_atexit(void) {
5741 KA_TRACE(30, ("__kmp_internal_end_atexit\n"));
5742 /* [Windows]
5743 josh: ideally, we want to completely shutdown the library in this atexit
5744 handler, but stat code that depends on thread specific data for gtid fails
5745 because that data becomes unavailable at some point during the shutdown, so
5746 we call __kmp_internal_end_thread instead. We should eventually remove the
5747 dependency on __kmp_get_specific_gtid in the stat code and use
5748 __kmp_internal_end_library to cleanly shutdown the library.
Jim Cownie5e8470a2013-09-27 10:38:44 +00005749
Jonathan Peyton30419822017-05-12 18:01:32 +00005750 // TODO: Can some of this comment about GVS be removed?
5751 I suspect that the offending stat code is executed when the calling thread
5752 tries to clean up a dead root thread's data structures, resulting in GVS
5753 code trying to close the GVS structures for that thread, but since the stat
5754 code uses __kmp_get_specific_gtid to get the gtid with the assumption that
5755 the calling thread is cleaning up itself instead of another thread, it get
5756 confused. This happens because allowing a thread to unregister and cleanup
5757 another thread is a recent modification for addressing an issue.
5758 Based on the current design (20050722), a thread may end up
5759 trying to unregister another thread only if thread death does not trigger
5760 the calling of __kmp_internal_end_thread. For Linux* OS, there is the
5761 thread specific data destructor function to detect thread death. For
5762 Windows dynamic, there is DllMain(THREAD_DETACH). For Windows static, there
5763 is nothing. Thus, the workaround is applicable only for Windows static
5764 stat library. */
5765 __kmp_internal_end_library(-1);
5766#if KMP_OS_WINDOWS
5767 __kmp_close_console();
5768#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00005769}
5770
Jonathan Peyton30419822017-05-12 18:01:32 +00005771static void __kmp_reap_thread(kmp_info_t *thread, int is_root) {
5772 // It is assumed __kmp_forkjoin_lock is acquired.
Jim Cownie5e8470a2013-09-27 10:38:44 +00005773
Jonathan Peyton30419822017-05-12 18:01:32 +00005774 int gtid;
Jim Cownie5e8470a2013-09-27 10:38:44 +00005775
Jonathan Peyton30419822017-05-12 18:01:32 +00005776 KMP_DEBUG_ASSERT(thread != NULL);
Jim Cownie5e8470a2013-09-27 10:38:44 +00005777
Jonathan Peyton30419822017-05-12 18:01:32 +00005778 gtid = thread->th.th_info.ds.ds_gtid;
Jim Cownie5e8470a2013-09-27 10:38:44 +00005779
Jonathan Peyton30419822017-05-12 18:01:32 +00005780 if (!is_root) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00005781
Jonathan Peyton30419822017-05-12 18:01:32 +00005782 if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
5783 /* Assume the threads are at the fork barrier here */
5784 KA_TRACE(
5785 20, ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n",
5786 gtid));
5787 /* Need release fence here to prevent seg faults for tree forkjoin barrier
5788 * (GEH) */
5789 ANNOTATE_HAPPENS_BEFORE(thread);
5790 kmp_flag_64 flag(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, thread);
5791 __kmp_release_64(&flag);
Jim Cownie5e8470a2013-09-27 10:38:44 +00005792 }; // if
5793
Jonathan Peyton30419822017-05-12 18:01:32 +00005794 // Terminate OS thread.
5795 __kmp_reap_worker(thread);
Jonathan Peyton7ca7ef02016-11-21 16:18:57 +00005796
Jonathan Peyton30419822017-05-12 18:01:32 +00005797 // The thread was killed asynchronously. If it was actively
5798 // spinning in the thread pool, decrement the global count.
5799 //
5800 // There is a small timing hole here - if the worker thread was just waking
5801 // up after sleeping in the pool, had reset it's th_active_in_pool flag but
5802 // not decremented the global counter __kmp_thread_pool_active_nth yet, then
5803 // the global counter might not get updated.
5804 //
5805 // Currently, this can only happen as the library is unloaded,
5806 // so there are no harmful side effects.
5807 if (thread->th.th_active_in_pool) {
5808 thread->th.th_active_in_pool = FALSE;
Andrey Churbanov5ba90c72017-07-17 09:03:14 +00005809 KMP_TEST_THEN_DEC32(&__kmp_thread_pool_active_nth);
Jonathan Peyton30419822017-05-12 18:01:32 +00005810 KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0);
5811 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00005812
Jonathan Peyton30419822017-05-12 18:01:32 +00005813 // Decrement # of [worker] threads in the pool.
5814 KMP_DEBUG_ASSERT(__kmp_thread_pool_nth > 0);
5815 --__kmp_thread_pool_nth;
5816 }; // if
Jim Cownie5e8470a2013-09-27 10:38:44 +00005817
Jonathan Peyton30419822017-05-12 18:01:32 +00005818 __kmp_free_implicit_task(thread);
Jim Cownie5e8470a2013-09-27 10:38:44 +00005819
Jonathan Peyton30419822017-05-12 18:01:32 +00005820// Free the fast memory for tasking
5821#if USE_FAST_MEMORY
5822 __kmp_free_fast_memory(thread);
5823#endif /* USE_FAST_MEMORY */
5824
5825 __kmp_suspend_uninitialize_thread(thread);
5826
5827 KMP_DEBUG_ASSERT(__kmp_threads[gtid] == thread);
5828 TCW_SYNC_PTR(__kmp_threads[gtid], NULL);
5829
5830 --__kmp_all_nth;
5831// __kmp_nth was decremented when thread is added to the pool.
Jim Cownie5e8470a2013-09-27 10:38:44 +00005832
5833#ifdef KMP_ADJUST_BLOCKTIME
Jonathan Peyton30419822017-05-12 18:01:32 +00005834 /* Adjust blocktime back to user setting or default if necessary */
5835 /* Middle initialization might never have occurred */
5836 if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
5837 KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
5838 if (__kmp_nth <= __kmp_avail_proc) {
5839 __kmp_zero_bt = FALSE;
Jim Cownie5e8470a2013-09-27 10:38:44 +00005840 }
Jonathan Peyton30419822017-05-12 18:01:32 +00005841 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00005842#endif /* KMP_ADJUST_BLOCKTIME */
5843
Jonathan Peyton30419822017-05-12 18:01:32 +00005844 /* free the memory being used */
5845 if (__kmp_env_consistency_check) {
5846 if (thread->th.th_cons) {
5847 __kmp_free_cons_stack(thread->th.th_cons);
5848 thread->th.th_cons = NULL;
Jim Cownie5e8470a2013-09-27 10:38:44 +00005849 }; // if
Jonathan Peyton30419822017-05-12 18:01:32 +00005850 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00005851
Jonathan Peyton30419822017-05-12 18:01:32 +00005852 if (thread->th.th_pri_common != NULL) {
5853 __kmp_free(thread->th.th_pri_common);
5854 thread->th.th_pri_common = NULL;
5855 }; // if
Andrey Churbanov6d224db2015-02-10 18:37:43 +00005856
Jonathan Peyton30419822017-05-12 18:01:32 +00005857 if (thread->th.th_task_state_memo_stack != NULL) {
5858 __kmp_free(thread->th.th_task_state_memo_stack);
5859 thread->th.th_task_state_memo_stack = NULL;
5860 }
5861
5862#if KMP_USE_BGET
5863 if (thread->th.th_local.bget_data != NULL) {
5864 __kmp_finalize_bget(thread);
5865 }; // if
5866#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00005867
Alp Toker98758b02014-03-02 04:12:06 +00005868#if KMP_AFFINITY_SUPPORTED
Jonathan Peyton30419822017-05-12 18:01:32 +00005869 if (thread->th.th_affin_mask != NULL) {
5870 KMP_CPU_FREE(thread->th.th_affin_mask);
5871 thread->th.th_affin_mask = NULL;
5872 }; // if
Alp Toker98758b02014-03-02 04:12:06 +00005873#endif /* KMP_AFFINITY_SUPPORTED */
Jim Cownie5e8470a2013-09-27 10:38:44 +00005874
Jonathan Peyton30419822017-05-12 18:01:32 +00005875 __kmp_reap_team(thread->th.th_serial_team);
5876 thread->th.th_serial_team = NULL;
5877 __kmp_free(thread);
Jim Cownie5e8470a2013-09-27 10:38:44 +00005878
Jonathan Peyton30419822017-05-12 18:01:32 +00005879 KMP_MB();
Jim Cownie5e8470a2013-09-27 10:38:44 +00005880
5881} // __kmp_reap_thread
5882
Jonathan Peyton30419822017-05-12 18:01:32 +00005883static void __kmp_internal_end(void) {
5884 int i;
Jim Cownie5e8470a2013-09-27 10:38:44 +00005885
Jonathan Peyton30419822017-05-12 18:01:32 +00005886 /* First, unregister the library */
5887 __kmp_unregister_library();
Jim Cownie5e8470a2013-09-27 10:38:44 +00005888
Jonathan Peyton30419822017-05-12 18:01:32 +00005889#if KMP_OS_WINDOWS
5890 /* In Win static library, we can't tell when a root actually dies, so we
5891 reclaim the data structures for any root threads that have died but not
5892 unregistered themselves, in order to shut down cleanly.
5893 In Win dynamic library we also can't tell when a thread dies. */
5894 __kmp_reclaim_dead_roots(); // AC: moved here to always clean resources of
5895// dead roots
Jonathan Peytonb66d1aa2016-09-27 17:11:17 +00005896#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00005897
Jonathan Peyton30419822017-05-12 18:01:32 +00005898 for (i = 0; i < __kmp_threads_capacity; i++)
5899 if (__kmp_root[i])
5900 if (__kmp_root[i]->r.r_active)
5901 break;
5902 KMP_MB(); /* Flush all pending memory write invalidates. */
5903 TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
5904
5905 if (i < __kmp_threads_capacity) {
5906#if KMP_USE_MONITOR
5907 // 2009-09-08 (lev): Other alive roots found. Why do we kill the monitor??
5908 KMP_MB(); /* Flush all pending memory write invalidates. */
5909
5910// Need to check that monitor was initialized before reaping it. If we are
5911// called form __kmp_atfork_child (which sets __kmp_init_parallel = 0), then
5912// __kmp_monitor will appear to contain valid data, but it is only valid in the
5913// parent process, not the child.
5914 // New behavior (201008): instead of keying off of the flag
5915 // __kmp_init_parallel, the monitor thread creation is keyed off
5916 // of the new flag __kmp_init_monitor.
5917 __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
5918 if (TCR_4(__kmp_init_monitor)) {
5919 __kmp_reap_monitor(&__kmp_monitor);
5920 TCW_4(__kmp_init_monitor, 0);
5921 }
5922 __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
5923 KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"));
5924#endif // KMP_USE_MONITOR
5925 } else {
5926/* TODO move this to cleanup code */
5927#ifdef KMP_DEBUG
5928 /* make sure that everything has properly ended */
5929 for (i = 0; i < __kmp_threads_capacity; i++) {
5930 if (__kmp_root[i]) {
5931 // KMP_ASSERT( ! KMP_UBER_GTID( i ) ); // AC:
5932 // there can be uber threads alive here
5933 KMP_ASSERT(!__kmp_root[i]->r.r_active); // TODO: can they be active?
5934 }
5935 }
5936#endif
5937
5938 KMP_MB();
5939
5940 // Reap the worker threads.
5941 // This is valid for now, but be careful if threads are reaped sooner.
5942 while (__kmp_thread_pool != NULL) { // Loop thru all the thread in the pool.
5943 // Get the next thread from the pool.
Andrey Churbanovc47afcd2017-07-03 11:24:08 +00005944 kmp_info_t *thread = CCAST(kmp_info_t *, __kmp_thread_pool);
Jonathan Peyton30419822017-05-12 18:01:32 +00005945 __kmp_thread_pool = thread->th.th_next_pool;
5946 // Reap it.
5947 KMP_DEBUG_ASSERT(thread->th.th_reap_state == KMP_SAFE_TO_REAP);
5948 thread->th.th_next_pool = NULL;
5949 thread->th.th_in_pool = FALSE;
5950 __kmp_reap_thread(thread, 0);
5951 }; // while
5952 __kmp_thread_pool_insert_pt = NULL;
5953
5954 // Reap teams.
5955 while (__kmp_team_pool != NULL) { // Loop thru all the teams in the pool.
5956 // Get the next team from the pool.
Andrey Churbanovc47afcd2017-07-03 11:24:08 +00005957 kmp_team_t *team = CCAST(kmp_team_t *, __kmp_team_pool);
Jonathan Peyton30419822017-05-12 18:01:32 +00005958 __kmp_team_pool = team->t.t_next_pool;
5959 // Reap it.
5960 team->t.t_next_pool = NULL;
5961 __kmp_reap_team(team);
5962 }; // while
5963
5964 __kmp_reap_task_teams();
5965
5966 for (i = 0; i < __kmp_threads_capacity; ++i) {
5967 // TBD: Add some checking...
5968 // Something like KMP_DEBUG_ASSERT( __kmp_thread[ i ] == NULL );
5969 }
5970
5971 /* Make sure all threadprivate destructors get run by joining with all
5972 worker threads before resetting this flag */
5973 TCW_SYNC_4(__kmp_init_common, FALSE);
5974
5975 KA_TRACE(10, ("__kmp_internal_end: all workers reaped\n"));
5976 KMP_MB();
5977
5978#if KMP_USE_MONITOR
5979 // See note above: One of the possible fixes for CQ138434 / CQ140126
5980 //
5981 // FIXME: push both code fragments down and CSE them?
5982 // push them into __kmp_cleanup() ?
5983 __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
5984 if (TCR_4(__kmp_init_monitor)) {
5985 __kmp_reap_monitor(&__kmp_monitor);
5986 TCW_4(__kmp_init_monitor, 0);
5987 }
5988 __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
5989 KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"));
5990#endif
5991 } /* else !__kmp_global.t_active */
5992 TCW_4(__kmp_init_gtid, FALSE);
5993 KMP_MB(); /* Flush all pending memory write invalidates. */
5994
5995 __kmp_cleanup();
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00005996#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00005997 ompt_fini();
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00005998#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00005999}
6000
Jonathan Peyton30419822017-05-12 18:01:32 +00006001void __kmp_internal_end_library(int gtid_req) {
6002 /* if we have already cleaned up, don't try again, it wouldn't be pretty */
6003 /* this shouldn't be a race condition because __kmp_internal_end() is the
6004 only place to clear __kmp_serial_init */
6005 /* we'll check this later too, after we get the lock */
6006 // 2009-09-06: We do not set g_abort without setting g_done. This check looks
6007 // redundaant, because the next check will work in any case.
6008 if (__kmp_global.g.g_abort) {
6009 KA_TRACE(11, ("__kmp_internal_end_library: abort, exiting\n"));
6010 /* TODO abort? */
6011 return;
6012 }
6013 if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6014 KA_TRACE(10, ("__kmp_internal_end_library: already finished\n"));
6015 return;
6016 }
6017
6018 KMP_MB(); /* Flush all pending memory write invalidates. */
6019
6020 /* find out who we are and what we should do */
6021 {
6022 int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific();
6023 KA_TRACE(
6024 10, ("__kmp_internal_end_library: enter T#%d (%d)\n", gtid, gtid_req));
6025 if (gtid == KMP_GTID_SHUTDOWN) {
6026 KA_TRACE(10, ("__kmp_internal_end_library: !__kmp_init_runtime, system "
6027 "already shutdown\n"));
6028 return;
6029 } else if (gtid == KMP_GTID_MONITOR) {
6030 KA_TRACE(10, ("__kmp_internal_end_library: monitor thread, gtid not "
6031 "registered, or system shutdown\n"));
6032 return;
6033 } else if (gtid == KMP_GTID_DNE) {
6034 KA_TRACE(10, ("__kmp_internal_end_library: gtid not registered or system "
6035 "shutdown\n"));
6036 /* we don't know who we are, but we may still shutdown the library */
6037 } else if (KMP_UBER_GTID(gtid)) {
6038 /* unregister ourselves as an uber thread. gtid is no longer valid */
6039 if (__kmp_root[gtid]->r.r_active) {
6040 __kmp_global.g.g_abort = -1;
6041 TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6042 KA_TRACE(10,
6043 ("__kmp_internal_end_library: root still active, abort T#%d\n",
6044 gtid));
Jim Cownie5e8470a2013-09-27 10:38:44 +00006045 return;
Jonathan Peyton30419822017-05-12 18:01:32 +00006046 } else {
6047 KA_TRACE(
6048 10,
6049 ("__kmp_internal_end_library: unregistering sibling T#%d\n", gtid));
6050 __kmp_unregister_root_current_thread(gtid);
6051 }
6052 } else {
6053/* worker threads may call this function through the atexit handler, if they
6054 * call exit() */
6055/* For now, skip the usual subsequent processing and just dump the debug buffer.
6056 TODO: do a thorough shutdown instead */
6057#ifdef DUMP_DEBUG_ON_EXIT
6058 if (__kmp_debug_buf)
6059 __kmp_dump_debug_buffer();
6060#endif
6061 return;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006062 }
Jonathan Peyton30419822017-05-12 18:01:32 +00006063 }
6064 /* synchronize the termination process */
6065 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006066
Jonathan Peyton30419822017-05-12 18:01:32 +00006067 /* have we already finished */
6068 if (__kmp_global.g.g_abort) {
6069 KA_TRACE(10, ("__kmp_internal_end_library: abort, exiting\n"));
6070 /* TODO abort? */
6071 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6072 return;
6073 }
6074 if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6075 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6076 return;
6077 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00006078
Jonathan Peyton30419822017-05-12 18:01:32 +00006079 /* We need this lock to enforce mutex between this reading of
6080 __kmp_threads_capacity and the writing by __kmp_register_root.
6081 Alternatively, we can use a counter of roots that is atomically updated by
6082 __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and
6083 __kmp_internal_end_*. */
6084 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006085
Jonathan Peyton30419822017-05-12 18:01:32 +00006086 /* now we can safely conduct the actual termination */
6087 __kmp_internal_end();
Jim Cownie5e8470a2013-09-27 10:38:44 +00006088
Jonathan Peyton30419822017-05-12 18:01:32 +00006089 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6090 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006091
Jonathan Peyton30419822017-05-12 18:01:32 +00006092 KA_TRACE(10, ("__kmp_internal_end_library: exit\n"));
Jim Cownie5e8470a2013-09-27 10:38:44 +00006093
Jonathan Peyton30419822017-05-12 18:01:32 +00006094#ifdef DUMP_DEBUG_ON_EXIT
6095 if (__kmp_debug_buf)
6096 __kmp_dump_debug_buffer();
6097#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00006098
Jonathan Peyton30419822017-05-12 18:01:32 +00006099#if KMP_OS_WINDOWS
6100 __kmp_close_console();
6101#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00006102
Jonathan Peyton30419822017-05-12 18:01:32 +00006103 __kmp_fini_allocator();
Jim Cownie5e8470a2013-09-27 10:38:44 +00006104
6105} // __kmp_internal_end_library
6106
Jonathan Peyton30419822017-05-12 18:01:32 +00006107void __kmp_internal_end_thread(int gtid_req) {
6108 int i;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006109
Jonathan Peyton30419822017-05-12 18:01:32 +00006110 /* if we have already cleaned up, don't try again, it wouldn't be pretty */
6111 /* this shouldn't be a race condition because __kmp_internal_end() is the
6112 * only place to clear __kmp_serial_init */
6113 /* we'll check this later too, after we get the lock */
6114 // 2009-09-06: We do not set g_abort without setting g_done. This check looks
6115 // redundant, because the next check will work in any case.
6116 if (__kmp_global.g.g_abort) {
6117 KA_TRACE(11, ("__kmp_internal_end_thread: abort, exiting\n"));
6118 /* TODO abort? */
6119 return;
6120 }
6121 if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6122 KA_TRACE(10, ("__kmp_internal_end_thread: already finished\n"));
6123 return;
6124 }
6125
6126 KMP_MB(); /* Flush all pending memory write invalidates. */
6127
6128 /* find out who we are and what we should do */
6129 {
6130 int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific();
6131 KA_TRACE(10,
6132 ("__kmp_internal_end_thread: enter T#%d (%d)\n", gtid, gtid_req));
6133 if (gtid == KMP_GTID_SHUTDOWN) {
6134 KA_TRACE(10, ("__kmp_internal_end_thread: !__kmp_init_runtime, system "
6135 "already shutdown\n"));
6136 return;
6137 } else if (gtid == KMP_GTID_MONITOR) {
6138 KA_TRACE(10, ("__kmp_internal_end_thread: monitor thread, gtid not "
6139 "registered, or system shutdown\n"));
6140 return;
6141 } else if (gtid == KMP_GTID_DNE) {
6142 KA_TRACE(10, ("__kmp_internal_end_thread: gtid not registered or system "
6143 "shutdown\n"));
6144 return;
6145 /* we don't know who we are */
6146 } else if (KMP_UBER_GTID(gtid)) {
6147 /* unregister ourselves as an uber thread. gtid is no longer valid */
6148 if (__kmp_root[gtid]->r.r_active) {
6149 __kmp_global.g.g_abort = -1;
6150 TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6151 KA_TRACE(10,
6152 ("__kmp_internal_end_thread: root still active, abort T#%d\n",
6153 gtid));
Jim Cownie5e8470a2013-09-27 10:38:44 +00006154 return;
Jonathan Peyton30419822017-05-12 18:01:32 +00006155 } else {
6156 KA_TRACE(10, ("__kmp_internal_end_thread: unregistering sibling T#%d\n",
6157 gtid));
6158 __kmp_unregister_root_current_thread(gtid);
6159 }
6160 } else {
6161 /* just a worker thread, let's leave */
6162 KA_TRACE(10, ("__kmp_internal_end_thread: worker thread T#%d\n", gtid));
6163
6164 if (gtid >= 0) {
6165 __kmp_threads[gtid]->th.th_task_team = NULL;
6166 }
6167
6168 KA_TRACE(10,
6169 ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n",
6170 gtid));
6171 return;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006172 }
Jonathan Peyton30419822017-05-12 18:01:32 +00006173 }
6174#if defined KMP_DYNAMIC_LIB
6175 // AC: lets not shutdown the Linux* OS dynamic library at the exit of uber
6176 // thread, because we will better shutdown later in the library destructor.
6177 // The reason of this change is performance problem when non-openmp thread in
6178 // a loop forks and joins many openmp threads. We can save a lot of time
6179 // keeping worker threads alive until the program shutdown.
6180 // OM: Removed Linux* OS restriction to fix the crash on OS X* (DPD200239966)
6181 // and Windows(DPD200287443) that occurs when using critical sections from
6182 // foreign threads.
6183 KA_TRACE(10, ("__kmp_internal_end_thread: exiting T#%d\n", gtid_req));
6184 return;
6185#endif
6186 /* synchronize the termination process */
6187 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006188
Jonathan Peyton30419822017-05-12 18:01:32 +00006189 /* have we already finished */
6190 if (__kmp_global.g.g_abort) {
6191 KA_TRACE(10, ("__kmp_internal_end_thread: abort, exiting\n"));
6192 /* TODO abort? */
6193 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6194 return;
6195 }
6196 if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6197 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6198 return;
6199 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00006200
Jonathan Peyton30419822017-05-12 18:01:32 +00006201 /* We need this lock to enforce mutex between this reading of
6202 __kmp_threads_capacity and the writing by __kmp_register_root.
6203 Alternatively, we can use a counter of roots that is atomically updated by
6204 __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and
6205 __kmp_internal_end_*. */
Jim Cownie5e8470a2013-09-27 10:38:44 +00006206
Jonathan Peyton30419822017-05-12 18:01:32 +00006207 /* should we finish the run-time? are all siblings done? */
6208 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006209
Jonathan Peyton30419822017-05-12 18:01:32 +00006210 for (i = 0; i < __kmp_threads_capacity; ++i) {
6211 if (KMP_UBER_GTID(i)) {
6212 KA_TRACE(
6213 10,
6214 ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n", i));
6215 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6216 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6217 return;
6218 };
6219 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00006220
Jonathan Peyton30419822017-05-12 18:01:32 +00006221 /* now we can safely conduct the actual termination */
Jim Cownie5e8470a2013-09-27 10:38:44 +00006222
Jonathan Peyton30419822017-05-12 18:01:32 +00006223 __kmp_internal_end();
Jim Cownie5e8470a2013-09-27 10:38:44 +00006224
Jonathan Peyton30419822017-05-12 18:01:32 +00006225 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6226 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006227
Jonathan Peyton30419822017-05-12 18:01:32 +00006228 KA_TRACE(10, ("__kmp_internal_end_thread: exit T#%d\n", gtid_req));
Jim Cownie5e8470a2013-09-27 10:38:44 +00006229
Jonathan Peyton30419822017-05-12 18:01:32 +00006230#ifdef DUMP_DEBUG_ON_EXIT
6231 if (__kmp_debug_buf)
6232 __kmp_dump_debug_buffer();
6233#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00006234} // __kmp_internal_end_thread
6235
Jonathan Peyton30419822017-05-12 18:01:32 +00006236// -----------------------------------------------------------------------------
Jim Cownie5e8470a2013-09-27 10:38:44 +00006237// Library registration stuff.
6238
Jonathan Peyton30419822017-05-12 18:01:32 +00006239static long __kmp_registration_flag = 0;
6240// Random value used to indicate library initialization.
6241static char *__kmp_registration_str = NULL;
6242// Value to be saved in env var __KMP_REGISTERED_LIB_<pid>.
Jim Cownie5e8470a2013-09-27 10:38:44 +00006243
Jonathan Peyton30419822017-05-12 18:01:32 +00006244static inline char *__kmp_reg_status_name() {
6245 /* On RHEL 3u5 if linked statically, getpid() returns different values in
6246 each thread. If registration and unregistration go in different threads
6247 (omp_misc_other_root_exit.cpp test case), the name of registered_lib_env
6248 env var can not be found, because the name will contain different pid. */
6249 return __kmp_str_format("__KMP_REGISTERED_LIB_%d", (int)getpid());
Jim Cownie5e8470a2013-09-27 10:38:44 +00006250} // __kmp_reg_status_get
6251
Jonathan Peyton30419822017-05-12 18:01:32 +00006252void __kmp_register_library_startup(void) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00006253
Jonathan Peyton30419822017-05-12 18:01:32 +00006254 char *name = __kmp_reg_status_name(); // Name of the environment variable.
6255 int done = 0;
6256 union {
6257 double dtime;
6258 long ltime;
6259 } time;
6260#if KMP_ARCH_X86 || KMP_ARCH_X86_64
6261 __kmp_initialize_system_tick();
6262#endif
6263 __kmp_read_system_time(&time.dtime);
6264 __kmp_registration_flag = 0xCAFE0000L | (time.ltime & 0x0000FFFFL);
6265 __kmp_registration_str =
6266 __kmp_str_format("%p-%lx-%s", &__kmp_registration_flag,
6267 __kmp_registration_flag, KMP_LIBRARY_FILE);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006268
Jonathan Peyton30419822017-05-12 18:01:32 +00006269 KA_TRACE(50, ("__kmp_register_library_startup: %s=\"%s\"\n", name,
6270 __kmp_registration_str));
Jim Cownie5e8470a2013-09-27 10:38:44 +00006271
Jonathan Peyton30419822017-05-12 18:01:32 +00006272 while (!done) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00006273
Jonathan Peyton30419822017-05-12 18:01:32 +00006274 char *value = NULL; // Actual value of the environment variable.
Jim Cownie5e8470a2013-09-27 10:38:44 +00006275
Jonathan Peyton30419822017-05-12 18:01:32 +00006276 // Set environment variable, but do not overwrite if it is exist.
6277 __kmp_env_set(name, __kmp_registration_str, 0);
6278 // Check the variable is written.
6279 value = __kmp_env_get(name);
6280 if (value != NULL && strcmp(value, __kmp_registration_str) == 0) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00006281
Jonathan Peyton30419822017-05-12 18:01:32 +00006282 done = 1; // Ok, environment variable set successfully, exit the loop.
Jim Cownie5e8470a2013-09-27 10:38:44 +00006283
Jonathan Peyton30419822017-05-12 18:01:32 +00006284 } else {
Jim Cownie5e8470a2013-09-27 10:38:44 +00006285
Jonathan Peyton30419822017-05-12 18:01:32 +00006286 // Oops. Write failed. Another copy of OpenMP RTL is in memory.
6287 // Check whether it alive or dead.
6288 int neighbor = 0; // 0 -- unknown status, 1 -- alive, 2 -- dead.
6289 char *tail = value;
6290 char *flag_addr_str = NULL;
6291 char *flag_val_str = NULL;
6292 char const *file_name = NULL;
6293 __kmp_str_split(tail, '-', &flag_addr_str, &tail);
6294 __kmp_str_split(tail, '-', &flag_val_str, &tail);
6295 file_name = tail;
6296 if (tail != NULL) {
6297 long *flag_addr = 0;
6298 long flag_val = 0;
6299 KMP_SSCANF(flag_addr_str, "%p", &flag_addr);
6300 KMP_SSCANF(flag_val_str, "%lx", &flag_val);
6301 if (flag_addr != 0 && flag_val != 0 && strcmp(file_name, "") != 0) {
6302 // First, check whether environment-encoded address is mapped into
6303 // addr space.
6304 // If so, dereference it to see if it still has the right value.
6305 if (__kmp_is_address_mapped(flag_addr) && *flag_addr == flag_val) {
6306 neighbor = 1;
6307 } else {
6308 // If not, then we know the other copy of the library is no longer
6309 // running.
6310 neighbor = 2;
6311 }; // if
Jim Cownie5e8470a2013-09-27 10:38:44 +00006312 }; // if
Jonathan Peyton30419822017-05-12 18:01:32 +00006313 }; // if
6314 switch (neighbor) {
6315 case 0: // Cannot parse environment variable -- neighbor status unknown.
6316 // Assume it is the incompatible format of future version of the
6317 // library. Assume the other library is alive.
6318 // WARN( ... ); // TODO: Issue a warning.
6319 file_name = "unknown library";
6320 // Attention! Falling to the next case. That's intentional.
6321 case 1: { // Neighbor is alive.
6322 // Check it is allowed.
6323 char *duplicate_ok = __kmp_env_get("KMP_DUPLICATE_LIB_OK");
6324 if (!__kmp_str_match_true(duplicate_ok)) {
6325 // That's not allowed. Issue fatal error.
6326 __kmp_msg(kmp_ms_fatal,
6327 KMP_MSG(DuplicateLibrary, KMP_LIBRARY_FILE, file_name),
6328 KMP_HNT(DuplicateLibrary), __kmp_msg_null);
6329 }; // if
6330 KMP_INTERNAL_FREE(duplicate_ok);
6331 __kmp_duplicate_library_ok = 1;
6332 done = 1; // Exit the loop.
6333 } break;
6334 case 2: { // Neighbor is dead.
6335 // Clear the variable and try to register library again.
6336 __kmp_env_unset(name);
6337 } break;
6338 default: { KMP_DEBUG_ASSERT(0); } break;
6339 }; // switch
Jim Cownie5e8470a2013-09-27 10:38:44 +00006340
Jonathan Peyton30419822017-05-12 18:01:32 +00006341 }; // if
6342 KMP_INTERNAL_FREE((void *)value);
6343
6344 }; // while
6345 KMP_INTERNAL_FREE((void *)name);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006346
6347} // func __kmp_register_library_startup
6348
Jonathan Peyton30419822017-05-12 18:01:32 +00006349void __kmp_unregister_library(void) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00006350
Jonathan Peyton30419822017-05-12 18:01:32 +00006351 char *name = __kmp_reg_status_name();
6352 char *value = __kmp_env_get(name);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006353
Jonathan Peyton30419822017-05-12 18:01:32 +00006354 KMP_DEBUG_ASSERT(__kmp_registration_flag != 0);
6355 KMP_DEBUG_ASSERT(__kmp_registration_str != NULL);
6356 if (value != NULL && strcmp(value, __kmp_registration_str) == 0) {
6357 // Ok, this is our variable. Delete it.
6358 __kmp_env_unset(name);
6359 }; // if
Jim Cownie5e8470a2013-09-27 10:38:44 +00006360
Jonathan Peyton30419822017-05-12 18:01:32 +00006361 KMP_INTERNAL_FREE(__kmp_registration_str);
6362 KMP_INTERNAL_FREE(value);
6363 KMP_INTERNAL_FREE(name);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006364
Jonathan Peyton30419822017-05-12 18:01:32 +00006365 __kmp_registration_flag = 0;
6366 __kmp_registration_str = NULL;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006367
6368} // __kmp_unregister_library
6369
Jim Cownie5e8470a2013-09-27 10:38:44 +00006370// End of Library registration stuff.
Jonathan Peyton30419822017-05-12 18:01:32 +00006371// -----------------------------------------------------------------------------
Jim Cownie5e8470a2013-09-27 10:38:44 +00006372
Jonathan Peyton492e0a32017-06-13 17:17:26 +00006373#if KMP_MIC_SUPPORTED
Andrey Churbanov613edeb2015-02-20 18:14:43 +00006374
Jonathan Peyton30419822017-05-12 18:01:32 +00006375static void __kmp_check_mic_type() {
6376 kmp_cpuid_t cpuid_state = {0};
6377 kmp_cpuid_t *cs_p = &cpuid_state;
6378 __kmp_x86_cpuid(1, 0, cs_p);
6379 // We don't support mic1 at the moment
6380 if ((cs_p->eax & 0xff0) == 0xB10) {
6381 __kmp_mic_type = mic2;
6382 } else if ((cs_p->eax & 0xf0ff0) == 0x50670) {
6383 __kmp_mic_type = mic3;
6384 } else {
6385 __kmp_mic_type = non_mic;
6386 }
Andrey Churbanov613edeb2015-02-20 18:14:43 +00006387}
6388
Jonathan Peyton492e0a32017-06-13 17:17:26 +00006389#endif /* KMP_MIC_SUPPORTED */
Andrey Churbanov613edeb2015-02-20 18:14:43 +00006390
Jonathan Peyton30419822017-05-12 18:01:32 +00006391static void __kmp_do_serial_initialize(void) {
6392 int i, gtid;
6393 int size;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006394
Jonathan Peyton30419822017-05-12 18:01:32 +00006395 KA_TRACE(10, ("__kmp_do_serial_initialize: enter\n"));
Jim Cownie5e8470a2013-09-27 10:38:44 +00006396
Jonathan Peyton30419822017-05-12 18:01:32 +00006397 KMP_DEBUG_ASSERT(sizeof(kmp_int32) == 4);
6398 KMP_DEBUG_ASSERT(sizeof(kmp_uint32) == 4);
6399 KMP_DEBUG_ASSERT(sizeof(kmp_int64) == 8);
6400 KMP_DEBUG_ASSERT(sizeof(kmp_uint64) == 8);
6401 KMP_DEBUG_ASSERT(sizeof(kmp_intptr_t) == sizeof(void *));
Jim Cownie5e8470a2013-09-27 10:38:44 +00006402
Jonathan Peyton82a13bf2015-09-21 18:01:02 +00006403#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00006404 ompt_pre_init();
Jonathan Peyton82a13bf2015-09-21 18:01:02 +00006405#endif
6406
Jonathan Peyton30419822017-05-12 18:01:32 +00006407 __kmp_validate_locks();
Jim Cownie5e8470a2013-09-27 10:38:44 +00006408
Jonathan Peyton30419822017-05-12 18:01:32 +00006409 /* Initialize internal memory allocator */
6410 __kmp_init_allocator();
Jim Cownie5e8470a2013-09-27 10:38:44 +00006411
Jonathan Peyton30419822017-05-12 18:01:32 +00006412 /* Register the library startup via an environment variable and check to see
6413 whether another copy of the library is already registered. */
Jim Cownie5e8470a2013-09-27 10:38:44 +00006414
Jonathan Peyton30419822017-05-12 18:01:32 +00006415 __kmp_register_library_startup();
Jim Cownie5e8470a2013-09-27 10:38:44 +00006416
Jonathan Peyton30419822017-05-12 18:01:32 +00006417 /* TODO reinitialization of library */
6418 if (TCR_4(__kmp_global.g.g_done)) {
6419 KA_TRACE(10, ("__kmp_do_serial_initialize: reinitialization of library\n"));
6420 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00006421
Jonathan Peyton30419822017-05-12 18:01:32 +00006422 __kmp_global.g.g_abort = 0;
6423 TCW_SYNC_4(__kmp_global.g.g_done, FALSE);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006424
Jonathan Peyton30419822017-05-12 18:01:32 +00006425/* initialize the locks */
Jim Cownie5e8470a2013-09-27 10:38:44 +00006426#if KMP_USE_ADAPTIVE_LOCKS
6427#if KMP_DEBUG_ADAPTIVE_LOCKS
Jonathan Peyton30419822017-05-12 18:01:32 +00006428 __kmp_init_speculative_stats();
Jim Cownie5e8470a2013-09-27 10:38:44 +00006429#endif
6430#endif
Jonathan Peytonad579922015-12-17 16:19:05 +00006431#if KMP_STATS_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00006432 __kmp_stats_init();
Jonathan Peytonad579922015-12-17 16:19:05 +00006433#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00006434 __kmp_init_lock(&__kmp_global_lock);
6435 __kmp_init_queuing_lock(&__kmp_dispatch_lock);
6436 __kmp_init_lock(&__kmp_debug_lock);
6437 __kmp_init_atomic_lock(&__kmp_atomic_lock);
6438 __kmp_init_atomic_lock(&__kmp_atomic_lock_1i);
6439 __kmp_init_atomic_lock(&__kmp_atomic_lock_2i);
6440 __kmp_init_atomic_lock(&__kmp_atomic_lock_4i);
6441 __kmp_init_atomic_lock(&__kmp_atomic_lock_4r);
6442 __kmp_init_atomic_lock(&__kmp_atomic_lock_8i);
6443 __kmp_init_atomic_lock(&__kmp_atomic_lock_8r);
6444 __kmp_init_atomic_lock(&__kmp_atomic_lock_8c);
6445 __kmp_init_atomic_lock(&__kmp_atomic_lock_10r);
6446 __kmp_init_atomic_lock(&__kmp_atomic_lock_16r);
6447 __kmp_init_atomic_lock(&__kmp_atomic_lock_16c);
6448 __kmp_init_atomic_lock(&__kmp_atomic_lock_20c);
6449 __kmp_init_atomic_lock(&__kmp_atomic_lock_32c);
6450 __kmp_init_bootstrap_lock(&__kmp_forkjoin_lock);
6451 __kmp_init_bootstrap_lock(&__kmp_exit_lock);
Jonathan Peytonb66d1aa2016-09-27 17:11:17 +00006452#if KMP_USE_MONITOR
Jonathan Peyton30419822017-05-12 18:01:32 +00006453 __kmp_init_bootstrap_lock(&__kmp_monitor_lock);
Jonathan Peytonb66d1aa2016-09-27 17:11:17 +00006454#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00006455 __kmp_init_bootstrap_lock(&__kmp_tp_cached_lock);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006456
Jonathan Peyton30419822017-05-12 18:01:32 +00006457 /* conduct initialization and initial setup of configuration */
Jim Cownie5e8470a2013-09-27 10:38:44 +00006458
Jonathan Peyton30419822017-05-12 18:01:32 +00006459 __kmp_runtime_initialize();
Jim Cownie5e8470a2013-09-27 10:38:44 +00006460
Jonathan Peyton492e0a32017-06-13 17:17:26 +00006461#if KMP_MIC_SUPPORTED
Jonathan Peyton30419822017-05-12 18:01:32 +00006462 __kmp_check_mic_type();
Andrey Churbanov613edeb2015-02-20 18:14:43 +00006463#endif
6464
Jonathan Peyton30419822017-05-12 18:01:32 +00006465// Some global variable initialization moved here from kmp_env_initialize()
Jim Cownie5e8470a2013-09-27 10:38:44 +00006466#ifdef KMP_DEBUG
Jonathan Peyton30419822017-05-12 18:01:32 +00006467 kmp_diag = 0;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006468#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00006469 __kmp_abort_delay = 0;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006470
Jonathan Peyton30419822017-05-12 18:01:32 +00006471 // From __kmp_init_dflt_team_nth()
6472 /* assume the entire machine will be used */
6473 __kmp_dflt_team_nth_ub = __kmp_xproc;
6474 if (__kmp_dflt_team_nth_ub < KMP_MIN_NTH) {
6475 __kmp_dflt_team_nth_ub = KMP_MIN_NTH;
6476 }
6477 if (__kmp_dflt_team_nth_ub > __kmp_sys_max_nth) {
6478 __kmp_dflt_team_nth_ub = __kmp_sys_max_nth;
6479 }
6480 __kmp_max_nth = __kmp_sys_max_nth;
Jonathan Peytonf4392462017-07-27 20:58:41 +00006481 __kmp_cg_max_nth = __kmp_sys_max_nth;
Jonathan Peyton4f90c822017-08-02 20:04:45 +00006482 __kmp_teams_max_nth = __kmp_xproc; // set a "reasonable" default
6483 if (__kmp_teams_max_nth > __kmp_sys_max_nth) {
6484 __kmp_teams_max_nth = __kmp_sys_max_nth;
6485 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00006486
Jonathan Peyton30419822017-05-12 18:01:32 +00006487 // Three vars below moved here from __kmp_env_initialize() "KMP_BLOCKTIME"
6488 // part
6489 __kmp_dflt_blocktime = KMP_DEFAULT_BLOCKTIME;
Jonathan Peytone1c7c132016-10-07 18:12:19 +00006490#if KMP_USE_MONITOR
Jonathan Peyton30419822017-05-12 18:01:32 +00006491 __kmp_monitor_wakeups =
6492 KMP_WAKEUPS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups);
6493 __kmp_bt_intervals =
6494 KMP_INTERVALS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups);
Jonathan Peytone1c7c132016-10-07 18:12:19 +00006495#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00006496 // From "KMP_LIBRARY" part of __kmp_env_initialize()
6497 __kmp_library = library_throughput;
6498 // From KMP_SCHEDULE initialization
6499 __kmp_static = kmp_sch_static_balanced;
6500// AC: do not use analytical here, because it is non-monotonous
6501//__kmp_guided = kmp_sch_guided_iterative_chunked;
6502//__kmp_auto = kmp_sch_guided_analytical_chunked; // AC: it is the default, no
6503// need to repeat assignment
6504// Barrier initialization. Moved here from __kmp_env_initialize() Barrier branch
6505// bit control and barrier method control parts
Jim Cownie4cc4bb42014-10-07 16:25:50 +00006506#if KMP_FAST_REDUCTION_BARRIER
Jonathan Peyton30419822017-05-12 18:01:32 +00006507#define kmp_reduction_barrier_gather_bb ((int)1)
6508#define kmp_reduction_barrier_release_bb ((int)1)
6509#define kmp_reduction_barrier_gather_pat bp_hyper_bar
6510#define kmp_reduction_barrier_release_pat bp_hyper_bar
6511#endif // KMP_FAST_REDUCTION_BARRIER
6512 for (i = bs_plain_barrier; i < bs_last_barrier; i++) {
6513 __kmp_barrier_gather_branch_bits[i] = __kmp_barrier_gather_bb_dflt;
6514 __kmp_barrier_release_branch_bits[i] = __kmp_barrier_release_bb_dflt;
6515 __kmp_barrier_gather_pattern[i] = __kmp_barrier_gather_pat_dflt;
6516 __kmp_barrier_release_pattern[i] = __kmp_barrier_release_pat_dflt;
6517#if KMP_FAST_REDUCTION_BARRIER
6518 if (i == bs_reduction_barrier) { // tested and confirmed on ALTIX only (
6519 // lin_64 ): hyper,1
6520 __kmp_barrier_gather_branch_bits[i] = kmp_reduction_barrier_gather_bb;
6521 __kmp_barrier_release_branch_bits[i] = kmp_reduction_barrier_release_bb;
6522 __kmp_barrier_gather_pattern[i] = kmp_reduction_barrier_gather_pat;
6523 __kmp_barrier_release_pattern[i] = kmp_reduction_barrier_release_pat;
Andrey Churbanov613edeb2015-02-20 18:14:43 +00006524 }
Jonathan Peyton30419822017-05-12 18:01:32 +00006525#endif // KMP_FAST_REDUCTION_BARRIER
6526 }
6527#if KMP_FAST_REDUCTION_BARRIER
6528#undef kmp_reduction_barrier_release_pat
6529#undef kmp_reduction_barrier_gather_pat
6530#undef kmp_reduction_barrier_release_bb
6531#undef kmp_reduction_barrier_gather_bb
6532#endif // KMP_FAST_REDUCTION_BARRIER
Jonathan Peyton492e0a32017-06-13 17:17:26 +00006533#if KMP_MIC_SUPPORTED
Jonathan Peyton30419822017-05-12 18:01:32 +00006534 if (__kmp_mic_type == mic2) { // KNC
6535 // AC: plane=3,2, forkjoin=2,1 are optimal for 240 threads on KNC
6536 __kmp_barrier_gather_branch_bits[bs_plain_barrier] = 3; // plain gather
6537 __kmp_barrier_release_branch_bits[bs_forkjoin_barrier] =
6538 1; // forkjoin release
6539 __kmp_barrier_gather_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar;
6540 __kmp_barrier_release_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar;
6541 }
6542#if KMP_FAST_REDUCTION_BARRIER
6543 if (__kmp_mic_type == mic2) { // KNC
6544 __kmp_barrier_gather_pattern[bs_reduction_barrier] = bp_hierarchical_bar;
6545 __kmp_barrier_release_pattern[bs_reduction_barrier] = bp_hierarchical_bar;
6546 }
Jonathan Peyton492e0a32017-06-13 17:17:26 +00006547#endif // KMP_FAST_REDUCTION_BARRIER
6548#endif // KMP_MIC_SUPPORTED
Jim Cownie5e8470a2013-09-27 10:38:44 +00006549
Jonathan Peyton30419822017-05-12 18:01:32 +00006550// From KMP_CHECKS initialization
Jim Cownie5e8470a2013-09-27 10:38:44 +00006551#ifdef KMP_DEBUG
Jonathan Peyton30419822017-05-12 18:01:32 +00006552 __kmp_env_checks = TRUE; /* development versions have the extra checks */
Jim Cownie5e8470a2013-09-27 10:38:44 +00006553#else
Jonathan Peyton30419822017-05-12 18:01:32 +00006554 __kmp_env_checks = FALSE; /* port versions do not have the extra checks */
Jim Cownie5e8470a2013-09-27 10:38:44 +00006555#endif
6556
Jonathan Peyton30419822017-05-12 18:01:32 +00006557 // From "KMP_FOREIGN_THREADS_THREADPRIVATE" initialization
6558 __kmp_foreign_tp = TRUE;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006559
Jonathan Peyton30419822017-05-12 18:01:32 +00006560 __kmp_global.g.g_dynamic = FALSE;
6561 __kmp_global.g.g_dynamic_mode = dynamic_default;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006562
Jonathan Peyton30419822017-05-12 18:01:32 +00006563 __kmp_env_initialize(NULL);
Jim Cownie4cc4bb42014-10-07 16:25:50 +00006564
Jonathan Peyton30419822017-05-12 18:01:32 +00006565// Print all messages in message catalog for testing purposes.
6566#ifdef KMP_DEBUG
6567 char const *val = __kmp_env_get("KMP_DUMP_CATALOG");
6568 if (__kmp_str_match_true(val)) {
6569 kmp_str_buf_t buffer;
6570 __kmp_str_buf_init(&buffer);
6571 __kmp_i18n_dump_catalog(&buffer);
6572 __kmp_printf("%s", buffer.str);
6573 __kmp_str_buf_free(&buffer);
6574 }; // if
6575 __kmp_env_free(&val);
6576#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00006577
Jonathan Peyton30419822017-05-12 18:01:32 +00006578 __kmp_threads_capacity =
6579 __kmp_initial_threads_capacity(__kmp_dflt_team_nth_ub);
6580 // Moved here from __kmp_env_initialize() "KMP_ALL_THREADPRIVATE" part
6581 __kmp_tp_capacity = __kmp_default_tp_capacity(
6582 __kmp_dflt_team_nth_ub, __kmp_max_nth, __kmp_allThreadsSpecified);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006583
Jonathan Peyton30419822017-05-12 18:01:32 +00006584 // If the library is shut down properly, both pools must be NULL. Just in
6585 // case, set them to NULL -- some memory may leak, but subsequent code will
6586 // work even if pools are not freed.
6587 KMP_DEBUG_ASSERT(__kmp_thread_pool == NULL);
6588 KMP_DEBUG_ASSERT(__kmp_thread_pool_insert_pt == NULL);
6589 KMP_DEBUG_ASSERT(__kmp_team_pool == NULL);
6590 __kmp_thread_pool = NULL;
6591 __kmp_thread_pool_insert_pt = NULL;
6592 __kmp_team_pool = NULL;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006593
Jonathan Peyton30419822017-05-12 18:01:32 +00006594 /* Allocate all of the variable sized records */
6595 /* NOTE: __kmp_threads_capacity entries are allocated, but the arrays are
6596 * expandable */
6597 /* Since allocation is cache-aligned, just add extra padding at the end */
6598 size =
6599 (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * __kmp_threads_capacity +
6600 CACHE_LINE;
6601 __kmp_threads = (kmp_info_t **)__kmp_allocate(size);
6602 __kmp_root = (kmp_root_t **)((char *)__kmp_threads +
6603 sizeof(kmp_info_t *) * __kmp_threads_capacity);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006604
Jonathan Peyton30419822017-05-12 18:01:32 +00006605 /* init thread counts */
6606 KMP_DEBUG_ASSERT(__kmp_all_nth ==
6607 0); // Asserts fail if the library is reinitializing and
6608 KMP_DEBUG_ASSERT(__kmp_nth == 0); // something was wrong in termination.
6609 __kmp_all_nth = 0;
6610 __kmp_nth = 0;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006611
Jonathan Peyton30419822017-05-12 18:01:32 +00006612 /* setup the uber master thread and hierarchy */
6613 gtid = __kmp_register_root(TRUE);
6614 KA_TRACE(10, ("__kmp_do_serial_initialize T#%d\n", gtid));
6615 KMP_ASSERT(KMP_UBER_GTID(gtid));
6616 KMP_ASSERT(KMP_INITIAL_GTID(gtid));
Jim Cownie5e8470a2013-09-27 10:38:44 +00006617
Jonathan Peyton30419822017-05-12 18:01:32 +00006618 KMP_MB(); /* Flush all pending memory write invalidates. */
Jim Cownie5e8470a2013-09-27 10:38:44 +00006619
Jonathan Peyton30419822017-05-12 18:01:32 +00006620 __kmp_common_initialize();
Jim Cownie5e8470a2013-09-27 10:38:44 +00006621
Jonathan Peyton30419822017-05-12 18:01:32 +00006622#if KMP_OS_UNIX
6623 /* invoke the child fork handler */
6624 __kmp_register_atfork();
6625#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00006626
Jonathan Peyton30419822017-05-12 18:01:32 +00006627#if !defined KMP_DYNAMIC_LIB
6628 {
6629 /* Invoke the exit handler when the program finishes, only for static
6630 library. For dynamic library, we already have _fini and DllMain. */
6631 int rc = atexit(__kmp_internal_end_atexit);
6632 if (rc != 0) {
6633 __kmp_msg(kmp_ms_fatal, KMP_MSG(FunctionError, "atexit()"), KMP_ERR(rc),
6634 __kmp_msg_null);
6635 }; // if
6636 }
6637#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00006638
Jonathan Peyton30419822017-05-12 18:01:32 +00006639#if KMP_HANDLE_SIGNALS
6640#if KMP_OS_UNIX
6641 /* NOTE: make sure that this is called before the user installs their own
6642 signal handlers so that the user handlers are called first. this way they
6643 can return false, not call our handler, avoid terminating the library, and
6644 continue execution where they left off. */
6645 __kmp_install_signals(FALSE);
6646#endif /* KMP_OS_UNIX */
6647#if KMP_OS_WINDOWS
6648 __kmp_install_signals(TRUE);
6649#endif /* KMP_OS_WINDOWS */
6650#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00006651
Jonathan Peyton30419822017-05-12 18:01:32 +00006652 /* we have finished the serial initialization */
6653 __kmp_init_counter++;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006654
Jonathan Peyton30419822017-05-12 18:01:32 +00006655 __kmp_init_serial = TRUE;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006656
Jonathan Peyton30419822017-05-12 18:01:32 +00006657 if (__kmp_settings) {
6658 __kmp_env_print();
6659 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00006660
6661#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00006662 if (__kmp_display_env || __kmp_display_env_verbose) {
6663 __kmp_env_print_2();
6664 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00006665#endif // OMP_40_ENABLED
6666
Jonathan Peyton82a13bf2015-09-21 18:01:02 +00006667#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00006668 ompt_post_init();
Jonathan Peyton82a13bf2015-09-21 18:01:02 +00006669#endif
6670
Jonathan Peyton30419822017-05-12 18:01:32 +00006671 KMP_MB();
Jim Cownie5e8470a2013-09-27 10:38:44 +00006672
Jonathan Peyton30419822017-05-12 18:01:32 +00006673 KA_TRACE(10, ("__kmp_do_serial_initialize: exit\n"));
Jim Cownie5e8470a2013-09-27 10:38:44 +00006674}
6675
Jonathan Peyton30419822017-05-12 18:01:32 +00006676void __kmp_serial_initialize(void) {
6677 if (__kmp_init_serial) {
6678 return;
6679 }
6680 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6681 if (__kmp_init_serial) {
6682 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6683 return;
6684 }
6685 __kmp_do_serial_initialize();
6686 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6687}
6688
6689static void __kmp_do_middle_initialize(void) {
6690 int i, j;
6691 int prev_dflt_team_nth;
6692
6693 if (!__kmp_init_serial) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00006694 __kmp_do_serial_initialize();
Jonathan Peyton30419822017-05-12 18:01:32 +00006695 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00006696
Jonathan Peyton30419822017-05-12 18:01:32 +00006697 KA_TRACE(10, ("__kmp_middle_initialize: enter\n"));
Jim Cownie5e8470a2013-09-27 10:38:44 +00006698
Jonathan Peyton30419822017-05-12 18:01:32 +00006699 // Save the previous value for the __kmp_dflt_team_nth so that
6700 // we can avoid some reinitialization if it hasn't changed.
6701 prev_dflt_team_nth = __kmp_dflt_team_nth;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006702
Alp Toker98758b02014-03-02 04:12:06 +00006703#if KMP_AFFINITY_SUPPORTED
Jonathan Peyton30419822017-05-12 18:01:32 +00006704 // __kmp_affinity_initialize() will try to set __kmp_ncores to the
6705 // number of cores on the machine.
6706 __kmp_affinity_initialize();
Jim Cownie5e8470a2013-09-27 10:38:44 +00006707
Jonathan Peyton30419822017-05-12 18:01:32 +00006708 // Run through the __kmp_threads array and set the affinity mask
6709 // for each root thread that is currently registered with the RTL.
6710 for (i = 0; i < __kmp_threads_capacity; i++) {
6711 if (TCR_PTR(__kmp_threads[i]) != NULL) {
6712 __kmp_affinity_set_init_mask(i, TRUE);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006713 }
Jonathan Peyton30419822017-05-12 18:01:32 +00006714 }
Alp Toker98758b02014-03-02 04:12:06 +00006715#endif /* KMP_AFFINITY_SUPPORTED */
Jim Cownie5e8470a2013-09-27 10:38:44 +00006716
Jonathan Peyton30419822017-05-12 18:01:32 +00006717 KMP_ASSERT(__kmp_xproc > 0);
6718 if (__kmp_avail_proc == 0) {
6719 __kmp_avail_proc = __kmp_xproc;
6720 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00006721
Jonathan Peyton30419822017-05-12 18:01:32 +00006722 // If there were empty places in num_threads list (OMP_NUM_THREADS=,,2,3),
6723 // correct them now
6724 j = 0;
6725 while ((j < __kmp_nested_nth.used) && !__kmp_nested_nth.nth[j]) {
6726 __kmp_nested_nth.nth[j] = __kmp_dflt_team_nth = __kmp_dflt_team_nth_ub =
6727 __kmp_avail_proc;
6728 j++;
6729 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00006730
Jonathan Peyton30419822017-05-12 18:01:32 +00006731 if (__kmp_dflt_team_nth == 0) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00006732#ifdef KMP_DFLT_NTH_CORES
Jonathan Peyton30419822017-05-12 18:01:32 +00006733 // Default #threads = #cores
6734 __kmp_dflt_team_nth = __kmp_ncores;
6735 KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "
6736 "__kmp_ncores (%d)\n",
6737 __kmp_dflt_team_nth));
Jim Cownie5e8470a2013-09-27 10:38:44 +00006738#else
Jonathan Peyton30419822017-05-12 18:01:32 +00006739 // Default #threads = #available OS procs
6740 __kmp_dflt_team_nth = __kmp_avail_proc;
6741 KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "
6742 "__kmp_avail_proc(%d)\n",
6743 __kmp_dflt_team_nth));
Jim Cownie5e8470a2013-09-27 10:38:44 +00006744#endif /* KMP_DFLT_NTH_CORES */
Jonathan Peyton30419822017-05-12 18:01:32 +00006745 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00006746
Jonathan Peyton30419822017-05-12 18:01:32 +00006747 if (__kmp_dflt_team_nth < KMP_MIN_NTH) {
6748 __kmp_dflt_team_nth = KMP_MIN_NTH;
6749 }
6750 if (__kmp_dflt_team_nth > __kmp_sys_max_nth) {
6751 __kmp_dflt_team_nth = __kmp_sys_max_nth;
6752 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00006753
Jonathan Peyton30419822017-05-12 18:01:32 +00006754 // There's no harm in continuing if the following check fails,
6755 // but it indicates an error in the previous logic.
6756 KMP_DEBUG_ASSERT(__kmp_dflt_team_nth <= __kmp_dflt_team_nth_ub);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006757
Jonathan Peyton30419822017-05-12 18:01:32 +00006758 if (__kmp_dflt_team_nth != prev_dflt_team_nth) {
6759 // Run through the __kmp_threads array and set the num threads icv for each
6760 // root thread that is currently registered with the RTL (which has not
6761 // already explicitly set its nthreads-var with a call to
6762 // omp_set_num_threads()).
6763 for (i = 0; i < __kmp_threads_capacity; i++) {
6764 kmp_info_t *thread = __kmp_threads[i];
6765 if (thread == NULL)
6766 continue;
6767 if (thread->th.th_current_task->td_icvs.nproc != 0)
6768 continue;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006769
Jonathan Peyton30419822017-05-12 18:01:32 +00006770 set__nproc(__kmp_threads[i], __kmp_dflt_team_nth);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006771 }
Jonathan Peyton30419822017-05-12 18:01:32 +00006772 }
6773 KA_TRACE(
6774 20,
6775 ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n",
6776 __kmp_dflt_team_nth));
Jim Cownie5e8470a2013-09-27 10:38:44 +00006777
6778#ifdef KMP_ADJUST_BLOCKTIME
Jonathan Peyton30419822017-05-12 18:01:32 +00006779 /* Adjust blocktime to zero if necessary now that __kmp_avail_proc is set */
6780 if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
6781 KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
6782 if (__kmp_nth > __kmp_avail_proc) {
6783 __kmp_zero_bt = TRUE;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006784 }
Jonathan Peyton30419822017-05-12 18:01:32 +00006785 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00006786#endif /* KMP_ADJUST_BLOCKTIME */
6787
Jonathan Peyton30419822017-05-12 18:01:32 +00006788 /* we have finished middle initialization */
6789 TCW_SYNC_4(__kmp_init_middle, TRUE);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006790
Jonathan Peyton30419822017-05-12 18:01:32 +00006791 KA_TRACE(10, ("__kmp_do_middle_initialize: exit\n"));
Jim Cownie5e8470a2013-09-27 10:38:44 +00006792}
6793
Jonathan Peyton30419822017-05-12 18:01:32 +00006794void __kmp_middle_initialize(void) {
6795 if (__kmp_init_middle) {
6796 return;
6797 }
6798 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6799 if (__kmp_init_middle) {
6800 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6801 return;
6802 }
6803 __kmp_do_middle_initialize();
6804 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6805}
6806
6807void __kmp_parallel_initialize(void) {
6808 int gtid = __kmp_entry_gtid(); // this might be a new root
6809
6810 /* synchronize parallel initialization (for sibling) */
6811 if (TCR_4(__kmp_init_parallel))
6812 return;
6813 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6814 if (TCR_4(__kmp_init_parallel)) {
6815 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6816 return;
6817 }
6818
6819 /* TODO reinitialization after we have already shut down */
6820 if (TCR_4(__kmp_global.g.g_done)) {
6821 KA_TRACE(
6822 10,
6823 ("__kmp_parallel_initialize: attempt to init while shutting down\n"));
6824 __kmp_infinite_loop();
6825 }
6826
6827 /* jc: The lock __kmp_initz_lock is already held, so calling
6828 __kmp_serial_initialize would cause a deadlock. So we call
6829 __kmp_do_serial_initialize directly. */
6830 if (!__kmp_init_middle) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00006831 __kmp_do_middle_initialize();
Jonathan Peyton30419822017-05-12 18:01:32 +00006832 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00006833
Jonathan Peyton30419822017-05-12 18:01:32 +00006834 /* begin initialization */
6835 KA_TRACE(10, ("__kmp_parallel_initialize: enter\n"));
6836 KMP_ASSERT(KMP_UBER_GTID(gtid));
Jim Cownie5e8470a2013-09-27 10:38:44 +00006837
6838#if KMP_ARCH_X86 || KMP_ARCH_X86_64
Jonathan Peyton30419822017-05-12 18:01:32 +00006839 // Save the FP control regs.
6840 // Worker threads will set theirs to these values at thread startup.
6841 __kmp_store_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
6842 __kmp_store_mxcsr(&__kmp_init_mxcsr);
6843 __kmp_init_mxcsr &= KMP_X86_MXCSR_MASK;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006844#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
6845
6846#if KMP_OS_UNIX
Jonathan Peyton30419822017-05-12 18:01:32 +00006847#if KMP_HANDLE_SIGNALS
6848 /* must be after __kmp_serial_initialize */
6849 __kmp_install_signals(TRUE);
6850#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00006851#endif
6852
Jonathan Peyton30419822017-05-12 18:01:32 +00006853 __kmp_suspend_initialize();
Jim Cownie5e8470a2013-09-27 10:38:44 +00006854
Jonathan Peyton749b4d52016-01-27 21:02:04 +00006855#if defined(USE_LOAD_BALANCE)
Jonathan Peyton30419822017-05-12 18:01:32 +00006856 if (__kmp_global.g.g_dynamic_mode == dynamic_default) {
6857 __kmp_global.g.g_dynamic_mode = dynamic_load_balance;
6858 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00006859#else
Jonathan Peyton30419822017-05-12 18:01:32 +00006860 if (__kmp_global.g.g_dynamic_mode == dynamic_default) {
6861 __kmp_global.g.g_dynamic_mode = dynamic_thread_limit;
6862 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00006863#endif
6864
Jonathan Peyton30419822017-05-12 18:01:32 +00006865 if (__kmp_version) {
6866 __kmp_print_version_2();
6867 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00006868
Jonathan Peyton30419822017-05-12 18:01:32 +00006869 /* we have finished parallel initialization */
6870 TCW_SYNC_4(__kmp_init_parallel, TRUE);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006871
Jonathan Peyton30419822017-05-12 18:01:32 +00006872 KMP_MB();
6873 KA_TRACE(10, ("__kmp_parallel_initialize: exit\n"));
Jim Cownie5e8470a2013-09-27 10:38:44 +00006874
Jonathan Peyton30419822017-05-12 18:01:32 +00006875 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006876}
6877
Jim Cownie5e8470a2013-09-27 10:38:44 +00006878/* ------------------------------------------------------------------------ */
6879
Jonathan Peyton30419822017-05-12 18:01:32 +00006880void __kmp_run_before_invoked_task(int gtid, int tid, kmp_info_t *this_thr,
6881 kmp_team_t *team) {
6882 kmp_disp_t *dispatch;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006883
Jonathan Peyton30419822017-05-12 18:01:32 +00006884 KMP_MB();
Jim Cownie5e8470a2013-09-27 10:38:44 +00006885
Jonathan Peyton30419822017-05-12 18:01:32 +00006886 /* none of the threads have encountered any constructs, yet. */
6887 this_thr->th.th_local.this_construct = 0;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006888#if KMP_CACHE_MANAGE
Jonathan Peyton30419822017-05-12 18:01:32 +00006889 KMP_CACHE_PREFETCH(&this_thr->th.th_bar[bs_forkjoin_barrier].bb.b_arrived);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006890#endif /* KMP_CACHE_MANAGE */
Jonathan Peyton30419822017-05-12 18:01:32 +00006891 dispatch = (kmp_disp_t *)TCR_PTR(this_thr->th.th_dispatch);
6892 KMP_DEBUG_ASSERT(dispatch);
6893 KMP_DEBUG_ASSERT(team->t.t_dispatch);
6894 // KMP_DEBUG_ASSERT( this_thr->th.th_dispatch == &team->t.t_dispatch[
6895 // this_thr->th.th_info.ds.ds_tid ] );
Jim Cownie5e8470a2013-09-27 10:38:44 +00006896
Jonathan Peyton30419822017-05-12 18:01:32 +00006897 dispatch->th_disp_index = 0; /* reset the dispatch buffer counter */
Jonathan Peytondf6818b2016-06-14 17:57:47 +00006898#if OMP_45_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00006899 dispatch->th_doacross_buf_idx =
6900 0; /* reset the doacross dispatch buffer counter */
Jonathan Peyton71909c52016-03-02 22:42:06 +00006901#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00006902 if (__kmp_env_consistency_check)
6903 __kmp_push_parallel(gtid, team->t.t_ident);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006904
Jonathan Peyton30419822017-05-12 18:01:32 +00006905 KMP_MB(); /* Flush all pending memory write invalidates. */
Jim Cownie5e8470a2013-09-27 10:38:44 +00006906}
6907
Jonathan Peyton30419822017-05-12 18:01:32 +00006908void __kmp_run_after_invoked_task(int gtid, int tid, kmp_info_t *this_thr,
6909 kmp_team_t *team) {
6910 if (__kmp_env_consistency_check)
6911 __kmp_pop_parallel(gtid, team->t.t_ident);
Andrey Churbanovdf0d75e2016-10-27 11:43:07 +00006912
Jonathan Peyton30419822017-05-12 18:01:32 +00006913 __kmp_finish_implicit_task(this_thr);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006914}
6915
Jonathan Peyton30419822017-05-12 18:01:32 +00006916int __kmp_invoke_task_func(int gtid) {
6917 int rc;
6918 int tid = __kmp_tid_from_gtid(gtid);
6919 kmp_info_t *this_thr = __kmp_threads[gtid];
6920 kmp_team_t *team = this_thr->th.th_team;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006921
Jonathan Peyton30419822017-05-12 18:01:32 +00006922 __kmp_run_before_invoked_task(gtid, tid, this_thr, team);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006923#if USE_ITT_BUILD
Jonathan Peyton30419822017-05-12 18:01:32 +00006924 if (__itt_stack_caller_create_ptr) {
6925 __kmp_itt_stack_callee_enter(
6926 (__itt_caller)
6927 team->t.t_stack_id); // inform ittnotify about entering user's code
6928 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00006929#endif /* USE_ITT_BUILD */
Jim Cownie4cc4bb42014-10-07 16:25:50 +00006930#if INCLUDE_SSC_MARKS
Jonathan Peyton30419822017-05-12 18:01:32 +00006931 SSC_MARK_INVOKING();
Jim Cownie4cc4bb42014-10-07 16:25:50 +00006932#endif
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00006933
6934#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00006935 void *dummy;
6936 void **exit_runtime_p;
6937 ompt_task_id_t my_task_id;
6938 ompt_parallel_id_t my_parallel_id;
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00006939
Jonathan Peyton30419822017-05-12 18:01:32 +00006940 if (ompt_enabled) {
6941 exit_runtime_p = &(team->t.t_implicit_task_taskdata[tid]
6942 .ompt_task_info.frame.exit_runtime_frame);
6943 } else {
6944 exit_runtime_p = &dummy;
6945 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00006946
6947#if OMPT_TRACE
Jonathan Peyton30419822017-05-12 18:01:32 +00006948 my_task_id = team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_id;
6949 my_parallel_id = team->t.ompt_team_info.parallel_id;
6950 if (ompt_enabled &&
6951 ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)) {
6952 ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)(my_parallel_id,
6953 my_task_id);
6954 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00006955#endif
6956#endif
6957
Jonathan Peyton30419822017-05-12 18:01:32 +00006958 {
6959 KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
6960 KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
6961 rc =
6962 __kmp_invoke_microtask((microtask_t)TCR_SYNC_PTR(team->t.t_pkfn), gtid,
6963 tid, (int)team->t.t_argc, (void **)team->t.t_argv
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00006964#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00006965 ,
6966 exit_runtime_p
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00006967#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00006968 );
Jonas Hahnfeld8a270642016-09-14 13:59:19 +00006969#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00006970 *exit_runtime_p = NULL;
Jonas Hahnfeld8a270642016-09-14 13:59:19 +00006971#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00006972 }
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00006973
Jim Cownie5e8470a2013-09-27 10:38:44 +00006974#if USE_ITT_BUILD
Jonathan Peyton30419822017-05-12 18:01:32 +00006975 if (__itt_stack_caller_create_ptr) {
6976 __kmp_itt_stack_callee_leave(
6977 (__itt_caller)
6978 team->t.t_stack_id); // inform ittnotify about leaving user's code
6979 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00006980#endif /* USE_ITT_BUILD */
Jonathan Peyton30419822017-05-12 18:01:32 +00006981 __kmp_run_after_invoked_task(gtid, tid, this_thr, team);
Jim Cownie5e8470a2013-09-27 10:38:44 +00006982
Jonathan Peyton30419822017-05-12 18:01:32 +00006983 return rc;
Jim Cownie5e8470a2013-09-27 10:38:44 +00006984}
6985
6986#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00006987void __kmp_teams_master(int gtid) {
6988 // This routine is called by all master threads in teams construct
6989 kmp_info_t *thr = __kmp_threads[gtid];
6990 kmp_team_t *team = thr->th.th_team;
6991 ident_t *loc = team->t.t_ident;
6992 thr->th.th_set_nproc = thr->th.th_teams_size.nth;
6993 KMP_DEBUG_ASSERT(thr->th.th_teams_microtask);
6994 KMP_DEBUG_ASSERT(thr->th.th_set_nproc);
6995 KA_TRACE(20, ("__kmp_teams_master: T#%d, Tid %d, microtask %p\n", gtid,
6996 __kmp_tid_from_gtid(gtid), thr->th.th_teams_microtask));
6997// Launch league of teams now, but not let workers execute
6998// (they hang on fork barrier until next parallel)
Jim Cownie4cc4bb42014-10-07 16:25:50 +00006999#if INCLUDE_SSC_MARKS
Jonathan Peyton30419822017-05-12 18:01:32 +00007000 SSC_MARK_FORKING();
Jim Cownie4cc4bb42014-10-07 16:25:50 +00007001#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00007002 __kmp_fork_call(loc, gtid, fork_context_intel, team->t.t_argc,
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00007003#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00007004 (void *)thr->th.th_teams_microtask, // "unwrapped" task
Andrey Churbanovd7d088f2015-04-29 16:42:24 +00007005#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00007006 (microtask_t)thr->th.th_teams_microtask, // "wrapped" task
7007 VOLATILE_CAST(launch_t) __kmp_invoke_task_func, NULL);
Jim Cownie4cc4bb42014-10-07 16:25:50 +00007008#if INCLUDE_SSC_MARKS
Jonathan Peyton30419822017-05-12 18:01:32 +00007009 SSC_MARK_JOINING();
Jim Cownie4cc4bb42014-10-07 16:25:50 +00007010#endif
Jonathan Peyton61118492016-05-20 19:03:38 +00007011
Jonathan Peyton30419822017-05-12 18:01:32 +00007012 // AC: last parameter "1" eliminates join barrier which won't work because
7013 // worker threads are in a fork barrier waiting for more parallel regions
7014 __kmp_join_call(loc, gtid
Jonathan Peytonf89fbbb2015-08-31 18:15:00 +00007015#if OMPT_SUPPORT
Jonathan Peyton30419822017-05-12 18:01:32 +00007016 ,
7017 fork_context_intel
Jonathan Peytonf89fbbb2015-08-31 18:15:00 +00007018#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00007019 ,
7020 1);
Jim Cownie5e8470a2013-09-27 10:38:44 +00007021}
7022
Jonathan Peyton30419822017-05-12 18:01:32 +00007023int __kmp_invoke_teams_master(int gtid) {
7024 kmp_info_t *this_thr = __kmp_threads[gtid];
7025 kmp_team_t *team = this_thr->th.th_team;
7026#if KMP_DEBUG
7027 if (!__kmp_threads[gtid]->th.th_team->t.t_serialized)
7028 KMP_DEBUG_ASSERT((void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn ==
7029 (void *)__kmp_teams_master);
7030#endif
7031 __kmp_run_before_invoked_task(gtid, 0, this_thr, team);
7032 __kmp_teams_master(gtid);
7033 __kmp_run_after_invoked_task(gtid, 0, this_thr, team);
7034 return 1;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007035}
7036#endif /* OMP_40_ENABLED */
7037
7038/* this sets the requested number of threads for the next parallel region
Jonathan Peyton30419822017-05-12 18:01:32 +00007039 encountered by this team. since this should be enclosed in the forkjoin
7040 critical section it should avoid race conditions with assymmetrical nested
7041 parallelism */
Jim Cownie5e8470a2013-09-27 10:38:44 +00007042
Jonathan Peyton30419822017-05-12 18:01:32 +00007043void __kmp_push_num_threads(ident_t *id, int gtid, int num_threads) {
7044 kmp_info_t *thr = __kmp_threads[gtid];
Jim Cownie5e8470a2013-09-27 10:38:44 +00007045
Jonathan Peyton30419822017-05-12 18:01:32 +00007046 if (num_threads > 0)
7047 thr->th.th_set_nproc = num_threads;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007048}
7049
7050#if OMP_40_ENABLED
7051
7052/* this sets the requested number of teams for the teams region and/or
Jonathan Peyton30419822017-05-12 18:01:32 +00007053 the number of threads for the next parallel region encountered */
7054void __kmp_push_num_teams(ident_t *id, int gtid, int num_teams,
7055 int num_threads) {
7056 kmp_info_t *thr = __kmp_threads[gtid];
7057 KMP_DEBUG_ASSERT(num_teams >= 0);
7058 KMP_DEBUG_ASSERT(num_threads >= 0);
Jonathan Peyton1be692e2015-11-30 20:14:05 +00007059
Jonathan Peyton30419822017-05-12 18:01:32 +00007060 if (num_teams == 0)
7061 num_teams = 1; // default number of teams is 1.
Jonathan Peyton4f90c822017-08-02 20:04:45 +00007062 if (num_teams > __kmp_teams_max_nth) { // if too many teams requested?
Jonathan Peyton30419822017-05-12 18:01:32 +00007063 if (!__kmp_reserve_warn) {
7064 __kmp_reserve_warn = 1;
7065 __kmp_msg(kmp_ms_warning,
Jonathan Peyton4f90c822017-08-02 20:04:45 +00007066 KMP_MSG(CantFormThrTeam, num_teams, __kmp_teams_max_nth),
Jonathan Peyton30419822017-05-12 18:01:32 +00007067 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
Jim Cownie5e8470a2013-09-27 10:38:44 +00007068 }
Jonathan Peyton4f90c822017-08-02 20:04:45 +00007069 num_teams = __kmp_teams_max_nth;
Jonathan Peyton30419822017-05-12 18:01:32 +00007070 }
7071 // Set number of teams (number of threads in the outer "parallel" of the
7072 // teams)
7073 thr->th.th_set_nproc = thr->th.th_teams_size.nteams = num_teams;
Jim Cownie4cc4bb42014-10-07 16:25:50 +00007074
Jonathan Peyton30419822017-05-12 18:01:32 +00007075 // Remember the number of threads for inner parallel regions
7076 if (num_threads == 0) {
7077 if (!TCR_4(__kmp_init_middle))
7078 __kmp_middle_initialize(); // get __kmp_avail_proc calculated
7079 num_threads = __kmp_avail_proc / num_teams;
Jonathan Peyton4f90c822017-08-02 20:04:45 +00007080 if (num_teams * num_threads > __kmp_teams_max_nth) {
Jonathan Peyton30419822017-05-12 18:01:32 +00007081 // adjust num_threads w/o warning as it is not user setting
Jonathan Peyton4f90c822017-08-02 20:04:45 +00007082 num_threads = __kmp_teams_max_nth / num_teams;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007083 }
Jonathan Peyton30419822017-05-12 18:01:32 +00007084 } else {
Jonathan Peyton4f90c822017-08-02 20:04:45 +00007085 if (num_teams * num_threads > __kmp_teams_max_nth) {
7086 int new_threads = __kmp_teams_max_nth / num_teams;
Jonathan Peyton30419822017-05-12 18:01:32 +00007087 if (!__kmp_reserve_warn) { // user asked for too many threads
Jonathan Peyton4f90c822017-08-02 20:04:45 +00007088 __kmp_reserve_warn = 1; // that conflicts with KMP_TEAMS_THREAD_LIMIT
Jonathan Peyton30419822017-05-12 18:01:32 +00007089 __kmp_msg(kmp_ms_warning,
7090 KMP_MSG(CantFormThrTeam, num_threads, new_threads),
7091 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
7092 }
7093 num_threads = new_threads;
7094 }
7095 }
7096 thr->th.th_teams_size.nth = num_threads;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007097}
7098
Jim Cownie5e8470a2013-09-27 10:38:44 +00007099// Set the proc_bind var to use in the following parallel region.
Jonathan Peyton30419822017-05-12 18:01:32 +00007100void __kmp_push_proc_bind(ident_t *id, int gtid, kmp_proc_bind_t proc_bind) {
7101 kmp_info_t *thr = __kmp_threads[gtid];
7102 thr->th.th_set_proc_bind = proc_bind;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007103}
7104
7105#endif /* OMP_40_ENABLED */
7106
7107/* Launch the worker threads into the microtask. */
7108
Jonathan Peyton30419822017-05-12 18:01:32 +00007109void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team) {
7110 kmp_info_t *this_thr = __kmp_threads[gtid];
Jim Cownie5e8470a2013-09-27 10:38:44 +00007111
7112#ifdef KMP_DEBUG
Jonathan Peyton30419822017-05-12 18:01:32 +00007113 int f;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007114#endif /* KMP_DEBUG */
7115
Jonathan Peyton30419822017-05-12 18:01:32 +00007116 KMP_DEBUG_ASSERT(team);
7117 KMP_DEBUG_ASSERT(this_thr->th.th_team == team);
7118 KMP_ASSERT(KMP_MASTER_GTID(gtid));
7119 KMP_MB(); /* Flush all pending memory write invalidates. */
Jim Cownie5e8470a2013-09-27 10:38:44 +00007120
Jonathan Peyton30419822017-05-12 18:01:32 +00007121 team->t.t_construct = 0; /* no single directives seen yet */
7122 team->t.t_ordered.dt.t_value =
7123 0; /* thread 0 enters the ordered section first */
Jim Cownie5e8470a2013-09-27 10:38:44 +00007124
Jonathan Peyton30419822017-05-12 18:01:32 +00007125 /* Reset the identifiers on the dispatch buffer */
7126 KMP_DEBUG_ASSERT(team->t.t_disp_buffer);
7127 if (team->t.t_max_nproc > 1) {
7128 int i;
7129 for (i = 0; i < __kmp_dispatch_num_buffers; ++i) {
7130 team->t.t_disp_buffer[i].buffer_index = i;
Jonathan Peytondf6818b2016-06-14 17:57:47 +00007131#if OMP_45_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00007132 team->t.t_disp_buffer[i].doacross_buf_idx = i;
Jonathan Peyton71909c52016-03-02 22:42:06 +00007133#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00007134 }
Jonathan Peyton30419822017-05-12 18:01:32 +00007135 } else {
7136 team->t.t_disp_buffer[0].buffer_index = 0;
7137#if OMP_45_ENABLED
7138 team->t.t_disp_buffer[0].doacross_buf_idx = 0;
7139#endif
7140 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00007141
Jonathan Peyton30419822017-05-12 18:01:32 +00007142 KMP_MB(); /* Flush all pending memory write invalidates. */
7143 KMP_ASSERT(this_thr->th.th_team == team);
Jim Cownie5e8470a2013-09-27 10:38:44 +00007144
7145#ifdef KMP_DEBUG
Jonathan Peyton30419822017-05-12 18:01:32 +00007146 for (f = 0; f < team->t.t_nproc; f++) {
7147 KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
7148 team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc);
7149 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00007150#endif /* KMP_DEBUG */
7151
Jonathan Peyton30419822017-05-12 18:01:32 +00007152 /* release the worker threads so they may begin working */
7153 __kmp_fork_barrier(gtid, 0);
Jim Cownie5e8470a2013-09-27 10:38:44 +00007154}
7155
Jonathan Peyton30419822017-05-12 18:01:32 +00007156void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team) {
7157 kmp_info_t *this_thr = __kmp_threads[gtid];
Jim Cownie5e8470a2013-09-27 10:38:44 +00007158
Jonathan Peyton30419822017-05-12 18:01:32 +00007159 KMP_DEBUG_ASSERT(team);
7160 KMP_DEBUG_ASSERT(this_thr->th.th_team == team);
7161 KMP_ASSERT(KMP_MASTER_GTID(gtid));
7162 KMP_MB(); /* Flush all pending memory write invalidates. */
Jim Cownie5e8470a2013-09-27 10:38:44 +00007163
Jonathan Peyton30419822017-05-12 18:01:32 +00007164/* Join barrier after fork */
Jim Cownie5e8470a2013-09-27 10:38:44 +00007165
7166#ifdef KMP_DEBUG
Jonathan Peyton30419822017-05-12 18:01:32 +00007167 if (__kmp_threads[gtid] &&
7168 __kmp_threads[gtid]->th.th_team_nproc != team->t.t_nproc) {
7169 __kmp_printf("GTID: %d, __kmp_threads[%d]=%p\n", gtid, gtid,
7170 __kmp_threads[gtid]);
7171 __kmp_printf("__kmp_threads[%d]->th.th_team_nproc=%d, TEAM: %p, "
7172 "team->t.t_nproc=%d\n",
7173 gtid, __kmp_threads[gtid]->th.th_team_nproc, team,
7174 team->t.t_nproc);
7175 __kmp_print_structure();
7176 }
7177 KMP_DEBUG_ASSERT(__kmp_threads[gtid] &&
7178 __kmp_threads[gtid]->th.th_team_nproc == team->t.t_nproc);
Jim Cownie5e8470a2013-09-27 10:38:44 +00007179#endif /* KMP_DEBUG */
7180
Jonathan Peyton30419822017-05-12 18:01:32 +00007181 __kmp_join_barrier(gtid); /* wait for everyone */
Jim Cownie5e8470a2013-09-27 10:38:44 +00007182
Jonathan Peyton30419822017-05-12 18:01:32 +00007183 KMP_MB(); /* Flush all pending memory write invalidates. */
7184 KMP_ASSERT(this_thr->th.th_team == team);
Jim Cownie5e8470a2013-09-27 10:38:44 +00007185}
7186
Jim Cownie5e8470a2013-09-27 10:38:44 +00007187/* ------------------------------------------------------------------------ */
7188
7189#ifdef USE_LOAD_BALANCE
7190
Jim Cownie5e8470a2013-09-27 10:38:44 +00007191// Return the worker threads actively spinning in the hot team, if we
7192// are at the outermost level of parallelism. Otherwise, return 0.
Jonathan Peyton30419822017-05-12 18:01:32 +00007193static int __kmp_active_hot_team_nproc(kmp_root_t *root) {
7194 int i;
7195 int retval;
7196 kmp_team_t *hot_team;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007197
Jonathan Peyton30419822017-05-12 18:01:32 +00007198 if (root->r.r_active) {
7199 return 0;
7200 }
7201 hot_team = root->r.r_hot_team;
7202 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
7203 return hot_team->t.t_nproc - 1; // Don't count master thread
7204 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00007205
Jonathan Peyton30419822017-05-12 18:01:32 +00007206 // Skip the master thread - it is accounted for elsewhere.
7207 retval = 0;
7208 for (i = 1; i < hot_team->t.t_nproc; i++) {
7209 if (hot_team->t.t_threads[i]->th.th_active) {
7210 retval++;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007211 }
Jonathan Peyton30419822017-05-12 18:01:32 +00007212 }
7213 return retval;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007214}
7215
Jim Cownie5e8470a2013-09-27 10:38:44 +00007216// Perform an automatic adjustment to the number of
7217// threads used by the next parallel region.
Jonathan Peyton30419822017-05-12 18:01:32 +00007218static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc) {
7219 int retval;
7220 int pool_active;
7221 int hot_team_active;
7222 int team_curr_active;
7223 int system_active;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007224
Jonathan Peyton30419822017-05-12 18:01:32 +00007225 KB_TRACE(20, ("__kmp_load_balance_nproc: called root:%p set_nproc:%d\n", root,
7226 set_nproc));
7227 KMP_DEBUG_ASSERT(root);
7228 KMP_DEBUG_ASSERT(root->r.r_root_team->t.t_threads[0]
7229 ->th.th_current_task->td_icvs.dynamic == TRUE);
7230 KMP_DEBUG_ASSERT(set_nproc > 1);
Jim Cownie5e8470a2013-09-27 10:38:44 +00007231
Jonathan Peyton30419822017-05-12 18:01:32 +00007232 if (set_nproc == 1) {
7233 KB_TRACE(20, ("__kmp_load_balance_nproc: serial execution.\n"));
7234 return 1;
7235 }
7236
7237 // Threads that are active in the thread pool, active in the hot team for this
7238 // particular root (if we are at the outer par level), and the currently
7239 // executing thread (to become the master) are available to add to the new
7240 // team, but are currently contributing to the system load, and must be
7241 // accounted for.
7242 pool_active = TCR_4(__kmp_thread_pool_active_nth);
7243 hot_team_active = __kmp_active_hot_team_nproc(root);
7244 team_curr_active = pool_active + hot_team_active + 1;
7245
7246 // Check the system load.
7247 system_active = __kmp_get_load_balance(__kmp_avail_proc + team_curr_active);
7248 KB_TRACE(30, ("__kmp_load_balance_nproc: system active = %d pool active = %d "
7249 "hot team active = %d\n",
7250 system_active, pool_active, hot_team_active));
7251
7252 if (system_active < 0) {
7253 // There was an error reading the necessary info from /proc, so use the
7254 // thread limit algorithm instead. Once we set __kmp_global.g.g_dynamic_mode
7255 // = dynamic_thread_limit, we shouldn't wind up getting back here.
7256 __kmp_global.g.g_dynamic_mode = dynamic_thread_limit;
7257 KMP_WARNING(CantLoadBalUsing, "KMP_DYNAMIC_MODE=thread limit");
7258
7259 // Make this call behave like the thread limit algorithm.
7260 retval = __kmp_avail_proc - __kmp_nth +
7261 (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
7262 if (retval > set_nproc) {
7263 retval = set_nproc;
7264 }
7265 if (retval < KMP_MIN_NTH) {
7266 retval = KMP_MIN_NTH;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007267 }
7268
Jonathan Peyton30419822017-05-12 18:01:32 +00007269 KB_TRACE(20, ("__kmp_load_balance_nproc: thread limit exit. retval:%d\n",
7270 retval));
Jim Cownie5e8470a2013-09-27 10:38:44 +00007271 return retval;
Jonathan Peyton30419822017-05-12 18:01:32 +00007272 }
7273
7274 // There is a slight delay in the load balance algorithm in detecting new
7275 // running procs. The real system load at this instant should be at least as
7276 // large as the #active omp thread that are available to add to the team.
7277 if (system_active < team_curr_active) {
7278 system_active = team_curr_active;
7279 }
7280 retval = __kmp_avail_proc - system_active + team_curr_active;
7281 if (retval > set_nproc) {
7282 retval = set_nproc;
7283 }
7284 if (retval < KMP_MIN_NTH) {
7285 retval = KMP_MIN_NTH;
7286 }
7287
7288 KB_TRACE(20, ("__kmp_load_balance_nproc: exit. retval:%d\n", retval));
7289 return retval;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007290} // __kmp_load_balance_nproc()
7291
7292#endif /* USE_LOAD_BALANCE */
7293
Jim Cownie5e8470a2013-09-27 10:38:44 +00007294/* ------------------------------------------------------------------------ */
Jim Cownie5e8470a2013-09-27 10:38:44 +00007295
7296/* NOTE: this is called with the __kmp_init_lock held */
Jonathan Peyton30419822017-05-12 18:01:32 +00007297void __kmp_cleanup(void) {
7298 int f;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007299
Jonathan Peyton30419822017-05-12 18:01:32 +00007300 KA_TRACE(10, ("__kmp_cleanup: enter\n"));
Jim Cownie5e8470a2013-09-27 10:38:44 +00007301
Jonathan Peyton30419822017-05-12 18:01:32 +00007302 if (TCR_4(__kmp_init_parallel)) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00007303#if KMP_HANDLE_SIGNALS
Jonathan Peyton30419822017-05-12 18:01:32 +00007304 __kmp_remove_signals();
Jim Cownie5e8470a2013-09-27 10:38:44 +00007305#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00007306 TCW_4(__kmp_init_parallel, FALSE);
7307 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00007308
Jonathan Peyton30419822017-05-12 18:01:32 +00007309 if (TCR_4(__kmp_init_middle)) {
Alp Toker763b9392014-02-28 09:42:41 +00007310#if KMP_AFFINITY_SUPPORTED
Jonathan Peyton30419822017-05-12 18:01:32 +00007311 __kmp_affinity_uninitialize();
Alp Toker763b9392014-02-28 09:42:41 +00007312#endif /* KMP_AFFINITY_SUPPORTED */
Jonathan Peyton30419822017-05-12 18:01:32 +00007313 __kmp_cleanup_hierarchy();
7314 TCW_4(__kmp_init_middle, FALSE);
7315 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00007316
Jonathan Peyton30419822017-05-12 18:01:32 +00007317 KA_TRACE(10, ("__kmp_cleanup: go serial cleanup\n"));
Jim Cownie5e8470a2013-09-27 10:38:44 +00007318
Jonathan Peyton30419822017-05-12 18:01:32 +00007319 if (__kmp_init_serial) {
7320 __kmp_runtime_destroy();
7321 __kmp_init_serial = FALSE;
7322 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00007323
Jonathan Peyton30419822017-05-12 18:01:32 +00007324 for (f = 0; f < __kmp_threads_capacity; f++) {
7325 if (__kmp_root[f] != NULL) {
7326 __kmp_free(__kmp_root[f]);
7327 __kmp_root[f] = NULL;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007328 }
Jonathan Peyton30419822017-05-12 18:01:32 +00007329 }
7330 __kmp_free(__kmp_threads);
7331 // __kmp_threads and __kmp_root were allocated at once, as single block, so
7332 // there is no need in freeing __kmp_root.
7333 __kmp_threads = NULL;
7334 __kmp_root = NULL;
7335 __kmp_threads_capacity = 0;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007336
Andrey Churbanov5c56fb52015-02-20 18:05:17 +00007337#if KMP_USE_DYNAMIC_LOCK
Jonathan Peyton30419822017-05-12 18:01:32 +00007338 __kmp_cleanup_indirect_user_locks();
Andrey Churbanov5c56fb52015-02-20 18:05:17 +00007339#else
Jonathan Peyton30419822017-05-12 18:01:32 +00007340 __kmp_cleanup_user_locks();
Andrey Churbanov5c56fb52015-02-20 18:05:17 +00007341#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00007342
Jonathan Peyton30419822017-05-12 18:01:32 +00007343#if KMP_AFFINITY_SUPPORTED
Andrey Churbanovc47afcd2017-07-03 11:24:08 +00007344 KMP_INTERNAL_FREE(CCAST(char *, __kmp_cpuinfo_file));
Jonathan Peyton30419822017-05-12 18:01:32 +00007345 __kmp_cpuinfo_file = NULL;
7346#endif /* KMP_AFFINITY_SUPPORTED */
Jim Cownie5e8470a2013-09-27 10:38:44 +00007347
Jonathan Peyton30419822017-05-12 18:01:32 +00007348#if KMP_USE_ADAPTIVE_LOCKS
7349#if KMP_DEBUG_ADAPTIVE_LOCKS
7350 __kmp_print_speculative_stats();
7351#endif
7352#endif
7353 KMP_INTERNAL_FREE(__kmp_nested_nth.nth);
7354 __kmp_nested_nth.nth = NULL;
7355 __kmp_nested_nth.size = 0;
7356 __kmp_nested_nth.used = 0;
7357 KMP_INTERNAL_FREE(__kmp_nested_proc_bind.bind_types);
7358 __kmp_nested_proc_bind.bind_types = NULL;
7359 __kmp_nested_proc_bind.size = 0;
7360 __kmp_nested_proc_bind.used = 0;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007361
Jonathan Peyton30419822017-05-12 18:01:32 +00007362 __kmp_i18n_catclose();
Jim Cownie5e8470a2013-09-27 10:38:44 +00007363
Jim Cownie4cc4bb42014-10-07 16:25:50 +00007364#if KMP_STATS_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00007365 __kmp_stats_fini();
Jim Cownie4cc4bb42014-10-07 16:25:50 +00007366#endif
7367
Jonathan Peyton30419822017-05-12 18:01:32 +00007368 KA_TRACE(10, ("__kmp_cleanup: exit\n"));
Jim Cownie5e8470a2013-09-27 10:38:44 +00007369}
7370
7371/* ------------------------------------------------------------------------ */
Jonathan Peyton30419822017-05-12 18:01:32 +00007372
7373int __kmp_ignore_mppbeg(void) {
7374 char *env;
7375
7376 if ((env = getenv("KMP_IGNORE_MPPBEG")) != NULL) {
7377 if (__kmp_str_match_false(env))
7378 return FALSE;
7379 }
7380 // By default __kmpc_begin() is no-op.
7381 return TRUE;
7382}
7383
7384int __kmp_ignore_mppend(void) {
7385 char *env;
7386
7387 if ((env = getenv("KMP_IGNORE_MPPEND")) != NULL) {
7388 if (__kmp_str_match_false(env))
7389 return FALSE;
7390 }
7391 // By default __kmpc_end() is no-op.
7392 return TRUE;
7393}
7394
7395void __kmp_internal_begin(void) {
7396 int gtid;
7397 kmp_root_t *root;
7398
7399 /* this is a very important step as it will register new sibling threads
7400 and assign these new uber threads a new gtid */
7401 gtid = __kmp_entry_gtid();
7402 root = __kmp_threads[gtid]->th.th_root;
7403 KMP_ASSERT(KMP_UBER_GTID(gtid));
7404
7405 if (root->r.r_begin)
7406 return;
7407 __kmp_acquire_lock(&root->r.r_begin_lock, gtid);
7408 if (root->r.r_begin) {
7409 __kmp_release_lock(&root->r.r_begin_lock, gtid);
7410 return;
7411 }
7412
7413 root->r.r_begin = TRUE;
7414
7415 __kmp_release_lock(&root->r.r_begin_lock, gtid);
7416}
7417
Jim Cownie5e8470a2013-09-27 10:38:44 +00007418/* ------------------------------------------------------------------------ */
7419
Jonathan Peyton30419822017-05-12 18:01:32 +00007420void __kmp_user_set_library(enum library_type arg) {
7421 int gtid;
7422 kmp_root_t *root;
7423 kmp_info_t *thread;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007424
Jonathan Peyton30419822017-05-12 18:01:32 +00007425 /* first, make sure we are initialized so we can get our gtid */
7426
7427 gtid = __kmp_entry_gtid();
7428 thread = __kmp_threads[gtid];
7429
7430 root = thread->th.th_root;
7431
7432 KA_TRACE(20, ("__kmp_user_set_library: enter T#%d, arg: %d, %d\n", gtid, arg,
7433 library_serial));
7434 if (root->r.r_in_parallel) { /* Must be called in serial section of top-level
7435 thread */
7436 KMP_WARNING(SetLibraryIncorrectCall);
7437 return;
7438 }
7439
7440 switch (arg) {
7441 case library_serial:
7442 thread->th.th_set_nproc = 0;
7443 set__nproc(thread, 1);
7444 break;
7445 case library_turnaround:
7446 thread->th.th_set_nproc = 0;
7447 set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth
7448 : __kmp_dflt_team_nth_ub);
7449 break;
7450 case library_throughput:
7451 thread->th.th_set_nproc = 0;
7452 set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth
7453 : __kmp_dflt_team_nth_ub);
7454 break;
7455 default:
7456 KMP_FATAL(UnknownLibraryType, arg);
7457 }
7458
7459 __kmp_aux_set_library(arg);
Jim Cownie5e8470a2013-09-27 10:38:44 +00007460}
7461
Jonathan Peyton30419822017-05-12 18:01:32 +00007462void __kmp_aux_set_stacksize(size_t arg) {
7463 if (!__kmp_init_serial)
7464 __kmp_serial_initialize();
Jim Cownie5e8470a2013-09-27 10:38:44 +00007465
7466#if KMP_OS_DARWIN
Jonathan Peyton30419822017-05-12 18:01:32 +00007467 if (arg & (0x1000 - 1)) {
7468 arg &= ~(0x1000 - 1);
7469 if (arg + 0x1000) /* check for overflow if we round up */
7470 arg += 0x1000;
7471 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00007472#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00007473 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
Jim Cownie5e8470a2013-09-27 10:38:44 +00007474
Jonathan Peyton30419822017-05-12 18:01:32 +00007475 /* only change the default stacksize before the first parallel region */
7476 if (!TCR_4(__kmp_init_parallel)) {
7477 size_t value = arg; /* argument is in bytes */
Jim Cownie5e8470a2013-09-27 10:38:44 +00007478
Jonathan Peyton30419822017-05-12 18:01:32 +00007479 if (value < __kmp_sys_min_stksize)
7480 value = __kmp_sys_min_stksize;
7481 else if (value > KMP_MAX_STKSIZE)
7482 value = KMP_MAX_STKSIZE;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007483
Jonathan Peyton30419822017-05-12 18:01:32 +00007484 __kmp_stksize = value;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007485
Jonathan Peyton30419822017-05-12 18:01:32 +00007486 __kmp_env_stksize = TRUE; /* was KMP_STACKSIZE specified? */
7487 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00007488
Jonathan Peyton30419822017-05-12 18:01:32 +00007489 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
Jim Cownie5e8470a2013-09-27 10:38:44 +00007490}
7491
7492/* set the behaviour of the runtime library */
7493/* TODO this can cause some odd behaviour with sibling parallelism... */
Jonathan Peyton30419822017-05-12 18:01:32 +00007494void __kmp_aux_set_library(enum library_type arg) {
7495 __kmp_library = arg;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007496
Jonathan Peyton30419822017-05-12 18:01:32 +00007497 switch (__kmp_library) {
7498 case library_serial: {
7499 KMP_INFORM(LibraryIsSerial);
7500 (void)__kmp_change_library(TRUE);
7501 } break;
7502 case library_turnaround:
7503 (void)__kmp_change_library(TRUE);
7504 break;
7505 case library_throughput:
7506 (void)__kmp_change_library(FALSE);
7507 break;
7508 default:
7509 KMP_FATAL(UnknownLibraryType, arg);
7510 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00007511}
7512
7513/* ------------------------------------------------------------------------ */
Jim Cownie5e8470a2013-09-27 10:38:44 +00007514
Jonathan Peyton30419822017-05-12 18:01:32 +00007515void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid) {
7516 int blocktime = arg; /* argument is in milliseconds */
Jonathan Peytone1c7c132016-10-07 18:12:19 +00007517#if KMP_USE_MONITOR
Jonathan Peyton30419822017-05-12 18:01:32 +00007518 int bt_intervals;
Jonathan Peytone1c7c132016-10-07 18:12:19 +00007519#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00007520 int bt_set;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007521
Jonathan Peyton30419822017-05-12 18:01:32 +00007522 __kmp_save_internal_controls(thread);
Jim Cownie5e8470a2013-09-27 10:38:44 +00007523
Jonathan Peyton30419822017-05-12 18:01:32 +00007524 /* Normalize and set blocktime for the teams */
7525 if (blocktime < KMP_MIN_BLOCKTIME)
7526 blocktime = KMP_MIN_BLOCKTIME;
7527 else if (blocktime > KMP_MAX_BLOCKTIME)
7528 blocktime = KMP_MAX_BLOCKTIME;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007529
Jonathan Peyton30419822017-05-12 18:01:32 +00007530 set__blocktime_team(thread->th.th_team, tid, blocktime);
7531 set__blocktime_team(thread->th.th_serial_team, 0, blocktime);
Jim Cownie5e8470a2013-09-27 10:38:44 +00007532
Jonathan Peytone1c7c132016-10-07 18:12:19 +00007533#if KMP_USE_MONITOR
Jonathan Peyton30419822017-05-12 18:01:32 +00007534 /* Calculate and set blocktime intervals for the teams */
7535 bt_intervals = KMP_INTERVALS_FROM_BLOCKTIME(blocktime, __kmp_monitor_wakeups);
Jim Cownie5e8470a2013-09-27 10:38:44 +00007536
Jonathan Peyton30419822017-05-12 18:01:32 +00007537 set__bt_intervals_team(thread->th.th_team, tid, bt_intervals);
7538 set__bt_intervals_team(thread->th.th_serial_team, 0, bt_intervals);
Jonathan Peytone1c7c132016-10-07 18:12:19 +00007539#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00007540
Jonathan Peyton30419822017-05-12 18:01:32 +00007541 /* Set whether blocktime has been set to "TRUE" */
7542 bt_set = TRUE;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007543
Jonathan Peyton30419822017-05-12 18:01:32 +00007544 set__bt_set_team(thread->th.th_team, tid, bt_set);
7545 set__bt_set_team(thread->th.th_serial_team, 0, bt_set);
Jonathan Peytone1c7c132016-10-07 18:12:19 +00007546#if KMP_USE_MONITOR
Jonathan Peyton30419822017-05-12 18:01:32 +00007547 KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, "
7548 "bt_intervals=%d, monitor_updates=%d\n",
7549 __kmp_gtid_from_tid(tid, thread->th.th_team),
7550 thread->th.th_team->t.t_id, tid, blocktime, bt_intervals,
7551 __kmp_monitor_wakeups));
Samuel Antao33515192016-10-20 13:20:17 +00007552#else
Jonathan Peyton30419822017-05-12 18:01:32 +00007553 KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d\n",
7554 __kmp_gtid_from_tid(tid, thread->th.th_team),
7555 thread->th.th_team->t.t_id, tid, blocktime));
Jonathan Peytone1c7c132016-10-07 18:12:19 +00007556#endif
Jim Cownie5e8470a2013-09-27 10:38:44 +00007557}
7558
Jonathan Peyton30419822017-05-12 18:01:32 +00007559void __kmp_aux_set_defaults(char const *str, int len) {
7560 if (!__kmp_init_serial) {
7561 __kmp_serial_initialize();
7562 };
7563 __kmp_env_initialize(str);
Jim Cownie5e8470a2013-09-27 10:38:44 +00007564
Jonathan Peyton30419822017-05-12 18:01:32 +00007565 if (__kmp_settings
Jim Cownie5e8470a2013-09-27 10:38:44 +00007566#if OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00007567 || __kmp_display_env || __kmp_display_env_verbose
Jim Cownie5e8470a2013-09-27 10:38:44 +00007568#endif // OMP_40_ENABLED
Jonathan Peyton30419822017-05-12 18:01:32 +00007569 ) {
7570 __kmp_env_print();
7571 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00007572} // __kmp_aux_set_defaults
7573
7574/* ------------------------------------------------------------------------ */
Jonathan Peyton30419822017-05-12 18:01:32 +00007575/* internal fast reduction routines */
Jim Cownie5e8470a2013-09-27 10:38:44 +00007576
Jim Cownie5e8470a2013-09-27 10:38:44 +00007577PACKED_REDUCTION_METHOD_T
Jonathan Peyton30419822017-05-12 18:01:32 +00007578__kmp_determine_reduction_method(
7579 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
7580 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
7581 kmp_critical_name *lck) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00007582
Jonathan Peyton30419822017-05-12 18:01:32 +00007583 // Default reduction method: critical construct ( lck != NULL, like in current
7584 // PAROPT )
7585 // If ( reduce_data!=NULL && reduce_func!=NULL ): the tree-reduction method
7586 // can be selected by RTL
7587 // If loc->flags contains KMP_IDENT_ATOMIC_REDUCE, the atomic reduce method
7588 // can be selected by RTL
7589 // Finally, it's up to OpenMP RTL to make a decision on which method to select
7590 // among generated by PAROPT.
Jim Cownie5e8470a2013-09-27 10:38:44 +00007591
Jonathan Peyton30419822017-05-12 18:01:32 +00007592 PACKED_REDUCTION_METHOD_T retval;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007593
Jonathan Peyton30419822017-05-12 18:01:32 +00007594 int team_size;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007595
Jonathan Peyton30419822017-05-12 18:01:32 +00007596 KMP_DEBUG_ASSERT(loc); // it would be nice to test ( loc != 0 )
7597 KMP_DEBUG_ASSERT(lck); // it would be nice to test ( lck != 0 )
Jim Cownie5e8470a2013-09-27 10:38:44 +00007598
Jonathan Peyton30419822017-05-12 18:01:32 +00007599#define FAST_REDUCTION_ATOMIC_METHOD_GENERATED \
7600 ((loc->flags & (KMP_IDENT_ATOMIC_REDUCE)) == (KMP_IDENT_ATOMIC_REDUCE))
7601#define FAST_REDUCTION_TREE_METHOD_GENERATED ((reduce_data) && (reduce_func))
Jim Cownie5e8470a2013-09-27 10:38:44 +00007602
Jonathan Peyton30419822017-05-12 18:01:32 +00007603 retval = critical_reduce_block;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007604
Jonathan Peyton30419822017-05-12 18:01:32 +00007605 // another choice of getting a team size (with 1 dynamic deference) is slower
7606 team_size = __kmp_get_team_num_threads(global_tid);
7607 if (team_size == 1) {
Jim Cownie5e8470a2013-09-27 10:38:44 +00007608
Jonathan Peyton30419822017-05-12 18:01:32 +00007609 retval = empty_reduce_block;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007610
Jonathan Peyton30419822017-05-12 18:01:32 +00007611 } else {
Jim Cownie5e8470a2013-09-27 10:38:44 +00007612
Jonathan Peyton30419822017-05-12 18:01:32 +00007613 int atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED;
7614 int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007615
Jonathan Peyton30419822017-05-12 18:01:32 +00007616#if KMP_ARCH_X86_64 || KMP_ARCH_PPC64 || KMP_ARCH_AARCH64 || KMP_ARCH_MIPS64
Jim Cownie5e8470a2013-09-27 10:38:44 +00007617
Jonathan Peyton30419822017-05-12 18:01:32 +00007618#if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_WINDOWS || \
7619 KMP_OS_DARWIN
Jim Cownie5e8470a2013-09-27 10:38:44 +00007620
Jonathan Peyton30419822017-05-12 18:01:32 +00007621 int teamsize_cutoff = 4;
Jonathan Peyton91b78702015-06-08 19:39:07 +00007622
Jonathan Peyton492e0a32017-06-13 17:17:26 +00007623#if KMP_MIC_SUPPORTED
Jonathan Peyton30419822017-05-12 18:01:32 +00007624 if (__kmp_mic_type != non_mic) {
7625 teamsize_cutoff = 8;
7626 }
Andrey Churbanov613edeb2015-02-20 18:14:43 +00007627#endif
Jonathan Peyton30419822017-05-12 18:01:32 +00007628 if (tree_available) {
7629 if (team_size <= teamsize_cutoff) {
7630 if (atomic_available) {
7631 retval = atomic_reduce_block;
Jim Cownie5e8470a2013-09-27 10:38:44 +00007632 }
Jonathan Peyton30419822017-05-12 18:01:32 +00007633 } else {
7634 retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER;
7635 }
7636 } else if (atomic_available) {
7637 retval = atomic_reduce_block;
7638 }
7639#else
7640#error "Unknown or unsupported OS"
7641#endif // KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_WINDOWS ||
7642// KMP_OS_DARWIN
Jim Cownie5e8470a2013-09-27 10:38:44 +00007643
Jonathan Peyton30419822017-05-12 18:01:32 +00007644#elif KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_AARCH || KMP_ARCH_MIPS
7645
7646#if KMP_OS_LINUX || KMP_OS_WINDOWS
7647
7648 // basic tuning
7649
7650 if (atomic_available) {
7651 if (num_vars <= 2) { // && ( team_size <= 8 ) due to false-sharing ???
7652 retval = atomic_reduce_block;
7653 }
7654 } // otherwise: use critical section
7655
7656#elif KMP_OS_DARWIN
7657
7658 if (atomic_available && (num_vars <= 3)) {
7659 retval = atomic_reduce_block;
7660 } else if (tree_available) {
7661 if ((reduce_size > (9 * sizeof(kmp_real64))) &&
7662 (reduce_size < (2000 * sizeof(kmp_real64)))) {
7663 retval = TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER;
7664 }
7665 } // otherwise: use critical section
7666
7667#else
7668#error "Unknown or unsupported OS"
7669#endif
7670
7671#else
7672#error "Unknown or unsupported architecture"
7673#endif
7674 }
7675
7676 // KMP_FORCE_REDUCTION
7677
7678 // If the team is serialized (team_size == 1), ignore the forced reduction
7679 // method and stay with the unsynchronized method (empty_reduce_block)
7680 if (__kmp_force_reduction_method != reduction_method_not_defined &&
7681 team_size != 1) {
7682
7683 PACKED_REDUCTION_METHOD_T forced_retval = critical_reduce_block;
7684
7685 int atomic_available, tree_available;
7686
7687 switch ((forced_retval = __kmp_force_reduction_method)) {
7688 case critical_reduce_block:
7689 KMP_ASSERT(lck); // lck should be != 0
7690 break;
7691
7692 case atomic_reduce_block:
7693 atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED;
7694 if (!atomic_available) {
7695 KMP_WARNING(RedMethodNotSupported, "atomic");
7696 forced_retval = critical_reduce_block;
7697 }
7698 break;
7699
7700 case tree_reduce_block:
7701 tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
7702 if (!tree_available) {
7703 KMP_WARNING(RedMethodNotSupported, "tree");
7704 forced_retval = critical_reduce_block;
7705 } else {
7706#if KMP_FAST_REDUCTION_BARRIER
7707 forced_retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER;
7708#endif
7709 }
7710 break;
7711
7712 default:
7713 KMP_ASSERT(0); // "unsupported method specified"
Jim Cownie5e8470a2013-09-27 10:38:44 +00007714 }
7715
Jonathan Peyton30419822017-05-12 18:01:32 +00007716 retval = forced_retval;
7717 }
Jim Cownie5e8470a2013-09-27 10:38:44 +00007718
Jonathan Peyton30419822017-05-12 18:01:32 +00007719 KA_TRACE(10, ("reduction method selected=%08x\n", retval));
Jim Cownie5e8470a2013-09-27 10:38:44 +00007720
Jonathan Peyton30419822017-05-12 18:01:32 +00007721#undef FAST_REDUCTION_TREE_METHOD_GENERATED
7722#undef FAST_REDUCTION_ATOMIC_METHOD_GENERATED
7723
7724 return (retval);
Jim Cownie5e8470a2013-09-27 10:38:44 +00007725}
7726
7727// this function is for testing set/get/determine reduce method
Jonathan Peyton30419822017-05-12 18:01:32 +00007728kmp_int32 __kmp_get_reduce_method(void) {
7729 return ((__kmp_entry_thread()->th.th_local.packed_reduction_method) >> 8);
Jim Cownie5e8470a2013-09-27 10:38:44 +00007730}