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 //----------------------------------------------------------
193 // thread support
194 //----------------------------------------------------------
195
__ompt_get_thread_data_internal()196 ompt_data_t *__ompt_get_thread_data_internal() {
197 if (__kmp_get_gtid() >= 0) {
198 kmp_info_t *thread = ompt_get_thread();
199 if (thread == NULL)
200 return NULL;
201 return &(thread->th.ompt_thread_info.thread_data);
202 }
203 return NULL;
204 }
205
206 //----------------------------------------------------------
207 // state support
208 //----------------------------------------------------------
209
__ompt_thread_assign_wait_id(void * variable)210 void __ompt_thread_assign_wait_id(void *variable) {
211 kmp_info_t *ti = ompt_get_thread();
212
213 if (ti)
214 ti->th.ompt_thread_info.wait_id = (ompt_wait_id_t)(uintptr_t)variable;
215 }
216
__ompt_get_state_internal(ompt_wait_id_t * omp_wait_id)217 int __ompt_get_state_internal(ompt_wait_id_t *omp_wait_id) {
218 kmp_info_t *ti = ompt_get_thread();
219
220 if (ti) {
221 if (omp_wait_id)
222 *omp_wait_id = ti->th.ompt_thread_info.wait_id;
223 return ti->th.ompt_thread_info.state;
224 }
225 return ompt_state_undefined;
226 }
227
228 //----------------------------------------------------------
229 // parallel region support
230 //----------------------------------------------------------
231
__ompt_get_parallel_info_internal(int ancestor_level,ompt_data_t ** parallel_data,int * team_size)232 int __ompt_get_parallel_info_internal(int ancestor_level,
233 ompt_data_t **parallel_data,
234 int *team_size) {
235 if (__kmp_get_gtid() >= 0) {
236 ompt_team_info_t *info;
237 if (team_size) {
238 info = __ompt_get_teaminfo(ancestor_level, team_size);
239 } else {
240 info = __ompt_get_teaminfo(ancestor_level, NULL);
241 }
242 if (parallel_data) {
243 *parallel_data = info ? &(info->parallel_data) : NULL;
244 }
245 return info ? 2 : 0;
246 } else {
247 return 0;
248 }
249 }
250
251 //----------------------------------------------------------
252 // lightweight task team support
253 //----------------------------------------------------------
254
__ompt_lw_taskteam_init(ompt_lw_taskteam_t * lwt,kmp_info_t * thr,int gtid,ompt_data_t * ompt_pid,void * codeptr)255 void __ompt_lw_taskteam_init(ompt_lw_taskteam_t *lwt, kmp_info_t *thr, int gtid,
256 ompt_data_t *ompt_pid, void *codeptr) {
257 // initialize parallel_data with input, return address to parallel_data on
258 // exit
259 lwt->ompt_team_info.parallel_data = *ompt_pid;
260 lwt->ompt_team_info.master_return_address = codeptr;
261 lwt->ompt_task_info.task_data.value = 0;
262 lwt->ompt_task_info.frame.enter_frame = ompt_data_none;
263 lwt->ompt_task_info.frame.exit_frame = ompt_data_none;
264 lwt->ompt_task_info.scheduling_parent = NULL;
265 lwt->heap = 0;
266 lwt->parent = 0;
267 }
268
__ompt_lw_taskteam_link(ompt_lw_taskteam_t * lwt,kmp_info_t * thr,int on_heap,bool always)269 void __ompt_lw_taskteam_link(ompt_lw_taskteam_t *lwt, kmp_info_t *thr,
270 int on_heap, bool always) {
271 ompt_lw_taskteam_t *link_lwt = lwt;
272 if (always ||
273 thr->th.th_team->t.t_serialized >
274 1) { // we already have a team, so link the new team and swap values
275 if (on_heap) { // the lw_taskteam cannot stay on stack, allocate it on heap
276 link_lwt =
277 (ompt_lw_taskteam_t *)__kmp_allocate(sizeof(ompt_lw_taskteam_t));
278 }
279 link_lwt->heap = on_heap;
280
281 // would be swap in the (on_stack) case.
282 ompt_team_info_t tmp_team = lwt->ompt_team_info;
283 link_lwt->ompt_team_info = *OMPT_CUR_TEAM_INFO(thr);
284 *OMPT_CUR_TEAM_INFO(thr) = tmp_team;
285
286 ompt_task_info_t tmp_task = lwt->ompt_task_info;
287 link_lwt->ompt_task_info = *OMPT_CUR_TASK_INFO(thr);
288 *OMPT_CUR_TASK_INFO(thr) = tmp_task;
289
290 // link the taskteam into the list of taskteams:
291 ompt_lw_taskteam_t *my_parent =
292 thr->th.th_team->t.ompt_serialized_team_info;
293 link_lwt->parent = my_parent;
294 thr->th.th_team->t.ompt_serialized_team_info = link_lwt;
295 #if OMPD_SUPPORT
296 if (ompd_state & OMPD_ENABLE_BP) {
297 ompd_bp_parallel_begin();
298 }
299 #endif
300 } else {
301 // this is the first serialized team, so we just store the values in the
302 // team and drop the taskteam-object
303 *OMPT_CUR_TEAM_INFO(thr) = lwt->ompt_team_info;
304 #if OMPD_SUPPORT
305 if (ompd_state & OMPD_ENABLE_BP) {
306 ompd_bp_parallel_begin();
307 }
308 #endif
309 *OMPT_CUR_TASK_INFO(thr) = lwt->ompt_task_info;
310 }
311 }
312
__ompt_lw_taskteam_unlink(kmp_info_t * thr)313 void __ompt_lw_taskteam_unlink(kmp_info_t *thr) {
314 ompt_lw_taskteam_t *lwtask = thr->th.th_team->t.ompt_serialized_team_info;
315 if (lwtask) {
316 #if OMPD_SUPPORT
317 if (ompd_state & OMPD_ENABLE_BP) {
318 ompd_bp_parallel_end();
319 }
320 #endif
321 thr->th.th_team->t.ompt_serialized_team_info = lwtask->parent;
322
323 ompt_team_info_t tmp_team = lwtask->ompt_team_info;
324 lwtask->ompt_team_info = *OMPT_CUR_TEAM_INFO(thr);
325 *OMPT_CUR_TEAM_INFO(thr) = tmp_team;
326
327 ompt_task_info_t tmp_task = lwtask->ompt_task_info;
328 lwtask->ompt_task_info = *OMPT_CUR_TASK_INFO(thr);
329 *OMPT_CUR_TASK_INFO(thr) = tmp_task;
330
331 if (lwtask->heap) {
332 __kmp_free(lwtask);
333 lwtask = NULL;
334 }
335 }
336 // return lwtask;
337 }
338
339 //----------------------------------------------------------
340 // task support
341 //----------------------------------------------------------
342
__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)343 int __ompt_get_task_info_internal(int ancestor_level, int *type,
344 ompt_data_t **task_data,
345 ompt_frame_t **task_frame,
346 ompt_data_t **parallel_data,
347 int *thread_num) {
348 if (__kmp_get_gtid() < 0)
349 return 0;
350
351 if (ancestor_level < 0)
352 return 0;
353
354 // copied from __ompt_get_scheduling_taskinfo
355 ompt_task_info_t *info = NULL;
356 ompt_team_info_t *team_info = NULL;
357 kmp_info_t *thr = ompt_get_thread();
358 int level = ancestor_level;
359
360 if (thr) {
361 kmp_taskdata_t *taskdata = thr->th.th_current_task;
362 if (taskdata == NULL)
363 return 0;
364 kmp_team *team = thr->th.th_team, *prev_team = NULL;
365 if (team == NULL)
366 return 0;
367 ompt_lw_taskteam_t *lwt = NULL,
368 *next_lwt = LWT_FROM_TEAM(taskdata->td_team),
369 *prev_lwt = NULL;
370
371 while (ancestor_level > 0) {
372 // needed for thread_num
373 prev_team = team;
374 prev_lwt = lwt;
375 // next lightweight team (if any)
376 if (lwt)
377 lwt = lwt->parent;
378
379 // next heavyweight team (if any) after
380 // lightweight teams are exhausted
381 if (!lwt && taskdata) {
382 // first try scheduling parent (for explicit task scheduling)
383 if (taskdata->ompt_task_info.scheduling_parent) {
384 taskdata = taskdata->ompt_task_info.scheduling_parent;
385 } else if (next_lwt) {
386 lwt = next_lwt;
387 next_lwt = NULL;
388 } else {
389 // then go for implicit tasks
390 taskdata = taskdata->td_parent;
391 if (team == NULL)
392 return 0;
393 team = team->t.t_parent;
394 if (taskdata) {
395 next_lwt = LWT_FROM_TEAM(taskdata->td_team);
396 }
397 }
398 }
399 ancestor_level--;
400 }
401
402 if (lwt) {
403 info = &lwt->ompt_task_info;
404 team_info = &lwt->ompt_team_info;
405 if (type) {
406 *type = ompt_task_implicit;
407 }
408 } else if (taskdata) {
409 info = &taskdata->ompt_task_info;
410 team_info = &team->t.ompt_team_info;
411 if (type) {
412 if (taskdata->td_parent) {
413 *type = (taskdata->td_flags.tasktype ? ompt_task_explicit
414 : ompt_task_implicit) |
415 TASK_TYPE_DETAILS_FORMAT(taskdata);
416 } else {
417 *type = ompt_task_initial;
418 }
419 }
420 }
421 if (task_data) {
422 *task_data = info ? &info->task_data : NULL;
423 }
424 if (task_frame) {
425 // OpenMP spec asks for the scheduling task to be returned.
426 *task_frame = info ? &info->frame : NULL;
427 }
428 if (parallel_data) {
429 *parallel_data = team_info ? &(team_info->parallel_data) : NULL;
430 }
431 if (thread_num) {
432 if (level == 0)
433 *thread_num = __kmp_get_tid();
434 else if (prev_lwt)
435 *thread_num = 0;
436 else
437 *thread_num = prev_team->t.t_master_tid;
438 // *thread_num = team->t.t_master_tid;
439 }
440 return info ? 2 : 0;
441 }
442 return 0;
443 }
444
__ompt_get_task_memory_internal(void ** addr,size_t * size,int blocknum)445 int __ompt_get_task_memory_internal(void **addr, size_t *size, int blocknum) {
446 if (blocknum != 0)
447 return 0; // support only a single block
448
449 kmp_info_t *thr = ompt_get_thread();
450 if (!thr)
451 return 0;
452
453 kmp_taskdata_t *taskdata = thr->th.th_current_task;
454 kmp_task_t *task = KMP_TASKDATA_TO_TASK(taskdata);
455
456 if (taskdata->td_flags.tasktype != TASK_EXPLICIT)
457 return 0; // support only explicit task
458
459 void *ret_addr;
460 int64_t ret_size = taskdata->td_size_alloc - sizeof(kmp_taskdata_t);
461
462 // kmp_task_t->data1 is an optional member
463 if (taskdata->td_flags.destructors_thunk)
464 ret_addr = &task->data1 + 1;
465 else
466 ret_addr = &task->part_id + 1;
467
468 ret_size -= (char *)(ret_addr) - (char *)(task);
469 if (ret_size < 0)
470 return 0;
471
472 *addr = ret_addr;
473 *size = (size_t)ret_size;
474 return 1;
475 }
476
477 //----------------------------------------------------------
478 // team support
479 //----------------------------------------------------------
480
__ompt_team_assign_id(kmp_team_t * team,ompt_data_t ompt_pid)481 void __ompt_team_assign_id(kmp_team_t *team, ompt_data_t ompt_pid) {
482 team->t.ompt_team_info.parallel_data = ompt_pid;
483 }
484
485 //----------------------------------------------------------
486 // misc
487 //----------------------------------------------------------
488
__ompt_get_unique_id_internal()489 static uint64_t __ompt_get_unique_id_internal() {
490 static uint64_t thread = 1;
491 static THREAD_LOCAL uint64_t ID = 0;
492 if (ID == 0) {
493 uint64_t new_thread = KMP_TEST_THEN_INC64((kmp_int64 *)&thread);
494 ID = new_thread << (sizeof(uint64_t) * 8 - OMPT_THREAD_ID_BITS);
495 }
496 return ++ID;
497 }
498
__ompt_get_barrier_kind(enum barrier_type bt,kmp_info_t * thr)499 ompt_sync_region_t __ompt_get_barrier_kind(enum barrier_type bt,
500 kmp_info_t *thr) {
501 if (bt == bs_forkjoin_barrier)
502 return ompt_sync_region_barrier_implicit;
503
504 if (bt != bs_plain_barrier)
505 return ompt_sync_region_barrier_implementation;
506
507 if (!thr->th.th_ident)
508 return ompt_sync_region_barrier;
509
510 kmp_int32 flags = thr->th.th_ident->flags;
511
512 if ((flags & KMP_IDENT_BARRIER_EXPL) != 0)
513 return ompt_sync_region_barrier_explicit;
514
515 if ((flags & KMP_IDENT_BARRIER_IMPL) != 0)
516 return ompt_sync_region_barrier_implicit;
517
518 return ompt_sync_region_barrier_implementation;
519 }
520