1 /*
2 * ompt-specific.cpp -- OMPT internal functions
3 */
4
5 //===----------------------------------------------------------------------===//
6 //
7 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
8 // See https://llvm.org/LICENSE.txt for license information.
9 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
10 //
11 //===----------------------------------------------------------------------===//
12
13 //******************************************************************************
14 // include files
15 //******************************************************************************
16
17 #include "kmp.h"
18 #include "ompt-specific.h"
19
20 #if KMP_OS_UNIX
21 #include <dlfcn.h>
22 #endif
23
24 #if KMP_OS_WINDOWS
25 #define THREAD_LOCAL __declspec(thread)
26 #else
27 #define THREAD_LOCAL __thread
28 #endif
29
30 #define OMPT_WEAK_ATTRIBUTE KMP_WEAK_ATTRIBUTE_INTERNAL
31
32 //******************************************************************************
33 // macros
34 //******************************************************************************
35
36 #define LWT_FROM_TEAM(team) (team)->t.ompt_serialized_team_info
37
38 #define OMPT_THREAD_ID_BITS 16
39
40 //******************************************************************************
41 // private operations
42 //******************************************************************************
43
44 //----------------------------------------------------------
45 // traverse the team and task hierarchy
46 // note: __ompt_get_teaminfo and __ompt_get_task_info_object
47 // traverse the hierarchy similarly and need to be
48 // kept consistent
49 //----------------------------------------------------------
50
__ompt_get_teaminfo(int depth,int * size)51 ompt_team_info_t *__ompt_get_teaminfo(int depth, int *size) {
52 kmp_info_t *thr = ompt_get_thread();
53
54 if (thr) {
55 kmp_team *team = thr->th.th_team;
56 if (team == NULL)
57 return NULL;
58
59 ompt_lw_taskteam_t *next_lwt = LWT_FROM_TEAM(team), *lwt = NULL;
60
61 while (depth > 0) {
62 // next lightweight team (if any)
63 if (lwt)
64 lwt = lwt->parent;
65
66 // next heavyweight team (if any) after
67 // lightweight teams are exhausted
68 if (!lwt && team) {
69 if (next_lwt) {
70 lwt = next_lwt;
71 next_lwt = NULL;
72 } else {
73 team = team->t.t_parent;
74 if (team) {
75 next_lwt = LWT_FROM_TEAM(team);
76 }
77 }
78 }
79
80 depth--;
81 }
82
83 if (lwt) {
84 // lightweight teams have one task
85 if (size)
86 *size = 1;
87
88 // return team info for lightweight team
89 return &lwt->ompt_team_info;
90 } else if (team) {
91 // extract size from heavyweight team
92 if (size)
93 *size = team->t.t_nproc;
94
95 // return team info for heavyweight team
96 return &team->t.ompt_team_info;
97 }
98 }
99
100 return NULL;
101 }
102
__ompt_get_task_info_object(int depth)103 ompt_task_info_t *__ompt_get_task_info_object(int depth) {
104 ompt_task_info_t *info = NULL;
105 kmp_info_t *thr = ompt_get_thread();
106
107 if (thr) {
108 kmp_taskdata_t *taskdata = thr->th.th_current_task;
109 ompt_lw_taskteam_t *lwt = NULL,
110 *next_lwt = LWT_FROM_TEAM(taskdata->td_team);
111
112 while (depth > 0) {
113 // next lightweight team (if any)
114 if (lwt)
115 lwt = lwt->parent;
116
117 // next heavyweight team (if any) after
118 // lightweight teams are exhausted
119 if (!lwt && taskdata) {
120 if (next_lwt) {
121 lwt = next_lwt;
122 next_lwt = NULL;
123 } else {
124 taskdata = taskdata->td_parent;
125 if (taskdata) {
126 next_lwt = LWT_FROM_TEAM(taskdata->td_team);
127 }
128 }
129 }
130 depth--;
131 }
132
133 if (lwt) {
134 info = &lwt->ompt_task_info;
135 } else if (taskdata) {
136 info = &taskdata->ompt_task_info;
137 }
138 }
139
140 return info;
141 }
142
__ompt_get_scheduling_taskinfo(int depth)143 ompt_task_info_t *__ompt_get_scheduling_taskinfo(int depth) {
144 ompt_task_info_t *info = NULL;
145 kmp_info_t *thr = ompt_get_thread();
146
147 if (thr) {
148 kmp_taskdata_t *taskdata = thr->th.th_current_task;
149
150 ompt_lw_taskteam_t *lwt = NULL,
151 *next_lwt = LWT_FROM_TEAM(taskdata->td_team);
152
153 while (depth > 0) {
154 // next lightweight team (if any)
155 if (lwt)
156 lwt = lwt->parent;
157
158 // next heavyweight team (if any) after
159 // lightweight teams are exhausted
160 if (!lwt && taskdata) {
161 // first try scheduling parent (for explicit task scheduling)
162 if (taskdata->ompt_task_info.scheduling_parent) {
163 taskdata = taskdata->ompt_task_info.scheduling_parent;
164 } else if (next_lwt) {
165 lwt = next_lwt;
166 next_lwt = NULL;
167 } else {
168 // then go for implicit tasks
169 taskdata = taskdata->td_parent;
170 if (taskdata) {
171 next_lwt = LWT_FROM_TEAM(taskdata->td_team);
172 }
173 }
174 }
175 depth--;
176 }
177
178 if (lwt) {
179 info = &lwt->ompt_task_info;
180 } else if (taskdata) {
181 info = &taskdata->ompt_task_info;
182 }
183 }
184
185 return info;
186 }
187
188 //******************************************************************************
189 // interface operations
190 //******************************************************************************
191 //----------------------------------------------------------
192 // initialization support
193 //----------------------------------------------------------
194
__ompt_force_initialization()195 void __ompt_force_initialization() { __kmp_serial_initialize(); }
196
197 //----------------------------------------------------------
198 // thread support
199 //----------------------------------------------------------
200
__ompt_get_thread_data_internal()201 ompt_data_t *__ompt_get_thread_data_internal() {
202 if (__kmp_get_gtid() >= 0) {
203 kmp_info_t *thread = ompt_get_thread();
204 if (thread == NULL)
205 return NULL;
206 return &(thread->th.ompt_thread_info.thread_data);
207 }
208 return NULL;
209 }
210
211 //----------------------------------------------------------
212 // state support
213 //----------------------------------------------------------
214
__ompt_thread_assign_wait_id(void * variable)215 void __ompt_thread_assign_wait_id(void *variable) {
216 kmp_info_t *ti = ompt_get_thread();
217
218 if (ti)
219 ti->th.ompt_thread_info.wait_id = (ompt_wait_id_t)(uintptr_t)variable;
220 }
221
__ompt_get_state_internal(ompt_wait_id_t * omp_wait_id)222 int __ompt_get_state_internal(ompt_wait_id_t *omp_wait_id) {
223 kmp_info_t *ti = ompt_get_thread();
224
225 if (ti) {
226 if (omp_wait_id)
227 *omp_wait_id = ti->th.ompt_thread_info.wait_id;
228 return ti->th.ompt_thread_info.state;
229 }
230 return ompt_state_undefined;
231 }
232
233 //----------------------------------------------------------
234 // parallel region support
235 //----------------------------------------------------------
236
__ompt_get_parallel_info_internal(int ancestor_level,ompt_data_t ** parallel_data,int * team_size)237 int __ompt_get_parallel_info_internal(int ancestor_level,
238 ompt_data_t **parallel_data,
239 int *team_size) {
240 if (__kmp_get_gtid() >= 0) {
241 ompt_team_info_t *info;
242 if (team_size) {
243 info = __ompt_get_teaminfo(ancestor_level, team_size);
244 } else {
245 info = __ompt_get_teaminfo(ancestor_level, NULL);
246 }
247 if (parallel_data) {
248 *parallel_data = info ? &(info->parallel_data) : NULL;
249 }
250 return info ? 2 : 0;
251 } else {
252 return 0;
253 }
254 }
255
256 //----------------------------------------------------------
257 // lightweight task team support
258 //----------------------------------------------------------
259
__ompt_lw_taskteam_init(ompt_lw_taskteam_t * lwt,kmp_info_t * thr,int gtid,ompt_data_t * ompt_pid,void * codeptr)260 void __ompt_lw_taskteam_init(ompt_lw_taskteam_t *lwt, kmp_info_t *thr, int gtid,
261 ompt_data_t *ompt_pid, void *codeptr) {
262 // initialize parallel_data with input, return address to parallel_data on
263 // exit
264 lwt->ompt_team_info.parallel_data = *ompt_pid;
265 lwt->ompt_team_info.master_return_address = codeptr;
266 lwt->ompt_task_info.task_data.value = 0;
267 lwt->ompt_task_info.frame.enter_frame = ompt_data_none;
268 lwt->ompt_task_info.frame.exit_frame = ompt_data_none;
269 lwt->ompt_task_info.scheduling_parent = NULL;
270 lwt->heap = 0;
271 lwt->parent = 0;
272 }
273
__ompt_lw_taskteam_link(ompt_lw_taskteam_t * lwt,kmp_info_t * thr,int on_heap,bool always)274 void __ompt_lw_taskteam_link(ompt_lw_taskteam_t *lwt, kmp_info_t *thr,
275 int on_heap, bool always) {
276 ompt_lw_taskteam_t *link_lwt = lwt;
277 if (always ||
278 thr->th.th_team->t.t_serialized >
279 1) { // we already have a team, so link the new team and swap values
280 if (on_heap) { // the lw_taskteam cannot stay on stack, allocate it on heap
281 link_lwt =
282 (ompt_lw_taskteam_t *)__kmp_allocate(sizeof(ompt_lw_taskteam_t));
283 }
284 link_lwt->heap = on_heap;
285
286 // would be swap in the (on_stack) case.
287 ompt_team_info_t tmp_team = lwt->ompt_team_info;
288 link_lwt->ompt_team_info = *OMPT_CUR_TEAM_INFO(thr);
289 *OMPT_CUR_TEAM_INFO(thr) = tmp_team;
290
291 // link the taskteam into the list of taskteams:
292 ompt_lw_taskteam_t *my_parent =
293 thr->th.th_team->t.ompt_serialized_team_info;
294 link_lwt->parent = my_parent;
295 thr->th.th_team->t.ompt_serialized_team_info = link_lwt;
296 #if OMPD_SUPPORT
297 if (ompd_state & OMPD_ENABLE_BP) {
298 ompd_bp_parallel_begin();
299 }
300 #endif
301
302 ompt_task_info_t tmp_task = lwt->ompt_task_info;
303 link_lwt->ompt_task_info = *OMPT_CUR_TASK_INFO(thr);
304 *OMPT_CUR_TASK_INFO(thr) = tmp_task;
305 } else {
306 // this is the first serialized team, so we just store the values in the
307 // team and drop the taskteam-object
308 *OMPT_CUR_TEAM_INFO(thr) = lwt->ompt_team_info;
309 #if OMPD_SUPPORT
310 if (ompd_state & OMPD_ENABLE_BP) {
311 ompd_bp_parallel_begin();
312 }
313 #endif
314 *OMPT_CUR_TASK_INFO(thr) = lwt->ompt_task_info;
315 }
316 }
317
__ompt_lw_taskteam_unlink(kmp_info_t * thr)318 void __ompt_lw_taskteam_unlink(kmp_info_t *thr) {
319 ompt_lw_taskteam_t *lwtask = thr->th.th_team->t.ompt_serialized_team_info;
320 if (lwtask) {
321 ompt_task_info_t tmp_task = lwtask->ompt_task_info;
322 lwtask->ompt_task_info = *OMPT_CUR_TASK_INFO(thr);
323 *OMPT_CUR_TASK_INFO(thr) = tmp_task;
324 #if OMPD_SUPPORT
325 if (ompd_state & OMPD_ENABLE_BP) {
326 ompd_bp_parallel_end();
327 }
328 #endif
329 thr->th.th_team->t.ompt_serialized_team_info = lwtask->parent;
330
331 ompt_team_info_t tmp_team = lwtask->ompt_team_info;
332 lwtask->ompt_team_info = *OMPT_CUR_TEAM_INFO(thr);
333 *OMPT_CUR_TEAM_INFO(thr) = tmp_team;
334
335 if (lwtask->heap) {
336 __kmp_free(lwtask);
337 lwtask = NULL;
338 }
339 }
340 // return lwtask;
341 }
342
343 //----------------------------------------------------------
344 // task support
345 //----------------------------------------------------------
346
__ompt_get_task_data()347 ompt_data_t *__ompt_get_task_data() {
348 kmp_info_t *thr = ompt_get_thread();
349 ompt_data_t *task_data = thr ? OMPT_CUR_TASK_DATA(thr) : NULL;
350 return task_data;
351 }
352
__ompt_get_target_task_data()353 ompt_data_t *__ompt_get_target_task_data() {
354 return &__kmp_threads[__kmp_get_gtid()]->th.ompt_thread_info.target_task_data;
355 }
356
__ompt_get_task_info_internal(int ancestor_level,int * type,ompt_data_t ** task_data,ompt_frame_t ** task_frame,ompt_data_t ** parallel_data,int * thread_num)357 int __ompt_get_task_info_internal(int ancestor_level, int *type,
358 ompt_data_t **task_data,
359 ompt_frame_t **task_frame,
360 ompt_data_t **parallel_data,
361 int *thread_num) {
362 if (__kmp_get_gtid() < 0)
363 return 0;
364
365 if (ancestor_level < 0)
366 return 0;
367
368 // copied from __ompt_get_scheduling_taskinfo
369 ompt_task_info_t *info = NULL;
370 ompt_team_info_t *team_info = NULL;
371 kmp_info_t *thr = ompt_get_thread();
372 int level = ancestor_level;
373
374 if (thr) {
375 kmp_taskdata_t *taskdata = thr->th.th_current_task;
376 if (taskdata == NULL)
377 return 0;
378 kmp_team *team = thr->th.th_team, *prev_team = NULL;
379 if (team == NULL)
380 return 0;
381 ompt_lw_taskteam_t *lwt = NULL,
382 *next_lwt = LWT_FROM_TEAM(taskdata->td_team);
383
384 while (ancestor_level > 0) {
385 // next lightweight team (if any)
386 if (lwt)
387 lwt = lwt->parent;
388
389 // next heavyweight team (if any) after
390 // lightweight teams are exhausted
391 if (!lwt && taskdata) {
392 // first try scheduling parent (for explicit task scheduling)
393 if (taskdata->ompt_task_info.scheduling_parent) {
394 taskdata = taskdata->ompt_task_info.scheduling_parent;
395 } else if (next_lwt) {
396 lwt = next_lwt;
397 next_lwt = NULL;
398 } else {
399 // then go for implicit tasks
400 taskdata = taskdata->td_parent;
401 if (team == NULL)
402 return 0;
403 prev_team = team;
404 team = team->t.t_parent;
405 if (taskdata) {
406 next_lwt = LWT_FROM_TEAM(taskdata->td_team);
407 }
408 }
409 }
410 ancestor_level--;
411 }
412
413 if (lwt) {
414 info = &lwt->ompt_task_info;
415 team_info = &lwt->ompt_team_info;
416 if (type) {
417 *type = ompt_task_implicit;
418 }
419 } else if (taskdata) {
420 info = &taskdata->ompt_task_info;
421 team_info = &team->t.ompt_team_info;
422 if (type) {
423 if (taskdata->td_parent) {
424 *type = TASK_TYPE_DETAILS_FORMAT(taskdata);
425 } else {
426 *type = ompt_task_initial;
427 }
428 }
429 }
430 if (task_data) {
431 *task_data = info ? &info->task_data : NULL;
432 }
433 if (task_frame) {
434 // OpenMP spec asks for the scheduling task to be returned.
435 *task_frame = info ? &info->frame : NULL;
436 }
437 if (parallel_data) {
438 *parallel_data = team_info ? &(team_info->parallel_data) : NULL;
439 }
440 if (thread_num) {
441 if (level == 0)
442 *thread_num = __kmp_get_tid();
443 else if (lwt)
444 *thread_num = 0;
445 else if (!prev_team) {
446 // The innermost parallel region contains at least one explicit task.
447 // The task at level > 0 is either an implicit task that
448 // corresponds to the mentioned region or one of the explicit tasks
449 // nested inside the same region. Note that the task isn't the
450 // innermost explicit tasks (because of condition level > 0).
451 // Since the task at this level still belongs to the innermost parallel
452 // region, thread_num is determined the same way as for level==0.
453 *thread_num = __kmp_get_tid();
454 } else
455 *thread_num = prev_team->t.t_master_tid;
456 // *thread_num = team->t.t_master_tid;
457 }
458 return info ? 2 : 0;
459 }
460 return 0;
461 }
462
__ompt_get_task_memory_internal(void ** addr,size_t * size,int blocknum)463 int __ompt_get_task_memory_internal(void **addr, size_t *size, int blocknum) {
464 *size = 0;
465 if (blocknum != 0)
466 return 0; // support only a single block
467
468 kmp_info_t *thr = ompt_get_thread();
469 if (!thr)
470 return 0;
471
472 kmp_taskdata_t *taskdata = thr->th.th_current_task;
473
474 if (taskdata->td_flags.tasktype != TASK_EXPLICIT)
475 return 0; // support only explicit task
476
477 *addr = taskdata;
478 *size = taskdata->td_size_alloc;
479 return 0;
480 }
481
482 //----------------------------------------------------------
483 // team support
484 //----------------------------------------------------------
485
__ompt_team_assign_id(kmp_team_t * team,ompt_data_t ompt_pid)486 void __ompt_team_assign_id(kmp_team_t *team, ompt_data_t ompt_pid) {
487 team->t.ompt_team_info.parallel_data = ompt_pid;
488 }
489
490 //----------------------------------------------------------
491 // misc
492 //----------------------------------------------------------
493
__ompt_get_unique_id_internal()494 static uint64_t __ompt_get_unique_id_internal() {
495 static uint64_t thread = 1;
496 static THREAD_LOCAL uint64_t ID = 0;
497 if (ID == 0) {
498 uint64_t new_thread = KMP_TEST_THEN_INC64((kmp_int64 *)&thread);
499 ID = new_thread << (sizeof(uint64_t) * 8 - OMPT_THREAD_ID_BITS);
500 }
501 return ++ID;
502 }
503
__ompt_get_barrier_kind(enum barrier_type bt,kmp_info_t * thr)504 ompt_sync_region_t __ompt_get_barrier_kind(enum barrier_type bt,
505 kmp_info_t *thr) {
506 if (bt == bs_forkjoin_barrier) {
507 if (thr->th.ompt_thread_info.parallel_flags & ompt_parallel_league)
508 return ompt_sync_region_barrier_teams;
509 else
510 return ompt_sync_region_barrier_implicit_parallel;
511 }
512
513 if (bt != bs_plain_barrier || !thr->th.th_ident)
514 return ompt_sync_region_barrier_implementation;
515
516 kmp_int32 flags = thr->th.th_ident->flags;
517
518 if ((flags & KMP_IDENT_BARRIER_EXPL) != 0)
519 return ompt_sync_region_barrier_explicit;
520
521 if ((flags & KMP_IDENT_BARRIER_IMPL) != 0)
522 return ompt_sync_region_barrier_implicit_workshare;
523
524 return ompt_sync_region_barrier_implementation;
525 }
526