1 //===--- ompt-multiplex.h - header-only multiplexing of OMPT tools -- C -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This header file enables an OMPT tool to load another OMPT tool and
10 // automatically forwards OMPT event-callbacks to the nested tool.
11 //
12 // For details see openmp/tools/multiplex/README.md
13 //
14 //===----------------------------------------------------------------------===//
15
16 #ifndef OMPT_MULTIPLEX_H
17 #define OMPT_MULTIPLEX_H
18
19 #ifndef _GNU_SOURCE
20 #define _GNU_SOURCE
21 #endif
22 #include <dlfcn.h>
23 #include <errno.h>
24 #include <execinfo.h>
25 #include <inttypes.h>
26 #include <omp-tools.h>
27 #include <omp.h>
28 #include <stdio.h>
29 #include <string.h>
30
31 static ompt_set_callback_t ompt_multiplex_set_callback;
32 static ompt_get_task_info_t ompt_multiplex_get_task_info;
33 static ompt_get_thread_data_t ompt_multiplex_get_thread_data;
34 static ompt_get_parallel_info_t ompt_multiplex_get_parallel_info;
35
36 // If OMPT_MULTIPLEX_TOOL_NAME is defined, use the tool name as prefix
37 // contains name of the environment var in which the tool path is specified
38 // for TOOL_LIBRARIES and VERBOSE_INIT variables. Only overwrite, if
39 // they are not explicitly defined.
40 #ifdef OMPT_MULTIPLEX_TOOL_NAME
41 #ifndef CLIENT_TOOL_LIBRARIES_VAR
42 #define CLIENT_TOOL_LIBRARIES_VAR OMPT_MULTIPLEX_TOOL_NAME "_TOOL_LIBRARIES"
43 #endif
44 #ifndef CLIENT_TOOL_VERBOSE_INIT_VAR
45 #define CLIENT_TOOL_VERBOSE_INIT_VAR \
46 OMPT_MULTIPLEX_TOOL_NAME "_TOOL_VERBOSE_INIT"
47 #endif
48 #endif
49
50 // If CLIENT_TOOL_VERBOSE_INIT_VAR is still not defined, use the OMPT
51 // env var.
52 #ifndef CLIENT_TOOL_VERBOSE_INIT_VAR
53 #warning CLIENT_TOOL_VERBOSE_INIT_VAR redefined to OMP_TOOL_VERBOSE_INIT
54 #define CLIENT_TOOL_VERBOSE_INIT_VAR "OMP_TOOL_VERBOSE_INIT"
55 #endif
56
57 // contains name of the environment var in which the tool path is specified
58 #ifndef CLIENT_TOOL_LIBRARIES_VAR
59 #error CLIENT_TOOL_LIBRARIES_VAR should be defined before including of ompt-multiplex.h
60 #endif
61
62 #if defined(OMPT_MULTIPLEX_CUSTOM_DELETE_TASK_DATA) && \
63 !defined(OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_TASK_DATA)
64 #error OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_TASK_DATA must be set if OMPT_MULTIPLEX_CUSTOM_DELETE_TASK_DATA is set
65 #endif
66
67 #if defined(OMPT_MULTIPLEX_CUSTOM_DELETE_PARALLEL_DATA) && \
68 !defined(OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_PARALLEL_DATA)
69 #error OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_PARALLEL_DATA must be set if OMPT_MULTIPLEX_CUSTOM_DELETE_PARALLEL_DATA is set
70 #endif
71
72 #if defined(OMPT_MULTIPLEX_CUSTOM_DELETE_THREAD_DATA) && \
73 !defined(OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_THREAD_DATA)
74 #error OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_THREAD_DATA must be set if OMPT_MULTIPLEX_CUSTOM_DELETE_THREAD_DATA is set
75 #endif
76
77 #define OMPT_API_ROUTINE static
78
79 #ifndef OMPT_STR_MATCH
80 #define OMPT_STR_MATCH(haystack, needle) (!strcasecmp(haystack, needle))
81 #endif
82
83 // prints for an enabled OMP_TOOL_VERBOSE_INIT.
84 // In the future a prefix could be added in the first define, the second define
85 // omits the prefix to allow for continued lines. Example: "PREFIX: Start
86 // tool... Success." instead of "PREFIX: Start tool... PREFIX: Success."
87 #define OMPT_VERBOSE_INIT_PRINT(...) \
88 if (verbose_init) \
89 fprintf(verbose_file, __VA_ARGS__)
90 #define OMPT_VERBOSE_INIT_CONTINUED_PRINT(...) \
91 if (verbose_init) \
92 fprintf(verbose_file, __VA_ARGS__)
93
94 static FILE *verbose_file;
95 static int verbose_init;
96
setup_verbose_init()97 void setup_verbose_init() {
98 const char *ompt_env_verbose_init = getenv(CLIENT_TOOL_VERBOSE_INIT_VAR);
99 // possible options: disabled | stdout | stderr | <filename>
100 // if set, not empty and not disabled -> prepare for logging
101 if (ompt_env_verbose_init && strcmp(ompt_env_verbose_init, "") &&
102 !OMPT_STR_MATCH(ompt_env_verbose_init, "disabled")) {
103 verbose_init = 1;
104 if (OMPT_STR_MATCH(ompt_env_verbose_init, "STDERR"))
105 verbose_file = stderr;
106 else if (OMPT_STR_MATCH(ompt_env_verbose_init, "STDOUT"))
107 verbose_file = stdout;
108 else if (!OMPT_STR_MATCH(ompt_env_verbose_init,
109 getenv("OMP_TOOL_VERBOSE_INIT")))
110 verbose_file = fopen(ompt_env_verbose_init, "w");
111 else {
112 verbose_init = 0;
113 printf("Multiplex: Can not open file defined in OMP_TOOL_VERBOSE_INIT "
114 "twice.");
115 }
116 } else
117 verbose_init = 0;
118 }
119
120 #define OMPT_LOAD_CLIENT_FOREACH_OMPT_EVENT(macro) \
121 macro(callback_thread_begin, ompt_callback_thread_begin_t, 1); \
122 macro(callback_thread_end, ompt_callback_thread_end_t, 2); \
123 macro(callback_parallel_begin, ompt_callback_parallel_begin_t, 3); \
124 macro(callback_parallel_end, ompt_callback_parallel_end_t, 4); \
125 macro(callback_task_create, ompt_callback_task_create_t, 5); \
126 macro(callback_task_schedule, ompt_callback_task_schedule_t, 6); \
127 macro(callback_implicit_task, ompt_callback_implicit_task_t, 7); \
128 macro(callback_target, ompt_callback_target_t, 8); \
129 macro(callback_target_data_op, ompt_callback_target_data_op_t, 9); \
130 macro(callback_target_submit, ompt_callback_target_submit_t, 10); \
131 macro(callback_control_tool, ompt_callback_control_tool_t, 11); \
132 macro(callback_device_initialize, ompt_callback_device_initialize_t, 12); \
133 macro(callback_device_finalize, ompt_callback_device_finalize_t, 13); \
134 macro(callback_device_load, ompt_callback_device_load_t, 14); \
135 macro(callback_device_unload, ompt_callback_device_unload_t, 15); \
136 macro(callback_sync_region_wait, ompt_callback_sync_region_t, 16); \
137 macro(callback_mutex_released, ompt_callback_mutex_t, 17); \
138 macro(callback_dependences, ompt_callback_dependences_t, 18); \
139 macro(callback_task_dependence, ompt_callback_task_dependence_t, 19); \
140 macro(callback_work, ompt_callback_work_t, 20); \
141 macro(callback_masked, ompt_callback_masked_t, 21); \
142 macro(callback_target_map, ompt_callback_target_map_t, 22); \
143 macro(callback_sync_region, ompt_callback_sync_region_t, 23); \
144 macro(callback_lock_init, ompt_callback_mutex_acquire_t, 24); \
145 macro(callback_lock_destroy, ompt_callback_mutex_t, 25); \
146 macro(callback_mutex_acquire, ompt_callback_mutex_acquire_t, 26); \
147 macro(callback_mutex_acquired, ompt_callback_mutex_t, 27); \
148 macro(callback_nest_lock, ompt_callback_nest_lock_t, 28); \
149 macro(callback_flush, ompt_callback_flush_t, 29); \
150 macro(callback_cancel, ompt_callback_cancel_t, 30); \
151 macro(callback_reduction, ompt_callback_sync_region_t, 31); \
152 macro(callback_dispatch, ompt_callback_dispatch_t, 32);
153
154 typedef struct ompt_multiplex_callbacks_s {
155 #define ompt_event_macro(event, callback, eventid) callback ompt_##event
156
157 OMPT_LOAD_CLIENT_FOREACH_OMPT_EVENT(ompt_event_macro)
158
159 #undef ompt_event_macro
160 } ompt_multiplex_callbacks_t;
161
162 typedef struct ompt_multiplex_callback_implementation_status_s {
163 #define ompt_event_macro(event, callback, eventid) int ompt_##event
164
165 OMPT_LOAD_CLIENT_FOREACH_OMPT_EVENT(ompt_event_macro)
166
167 #undef ompt_event_macro
168 } ompt_multiplex_callback_implementation_status_t;
169
170 ompt_start_tool_result_t *ompt_multiplex_own_fns = NULL;
171 ompt_start_tool_result_t *ompt_multiplex_client_fns = NULL;
172 ompt_function_lookup_t ompt_multiplex_lookup_function;
173 ompt_multiplex_callbacks_t ompt_multiplex_own_callbacks,
174 ompt_multiplex_client_callbacks;
175 ompt_multiplex_callback_implementation_status_t
176 ompt_multiplex_implementation_status;
177
178 typedef struct ompt_multiplex_data_pair_s {
179 ompt_data_t own_data;
180 ompt_data_t client_data;
181 } ompt_multiplex_data_pair_t;
182
183 #if !defined(OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_THREAD_DATA) || \
184 !defined(OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_PARALLEL_DATA) || \
185 !defined(OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_TASK_DATA)
186 static ompt_multiplex_data_pair_t *
ompt_multiplex_allocate_data_pair(ompt_data_t * data_pointer)187 ompt_multiplex_allocate_data_pair(ompt_data_t *data_pointer) {
188 data_pointer->ptr = malloc(sizeof(ompt_multiplex_data_pair_t));
189 if (!data_pointer->ptr) {
190 printf("Malloc ERROR\n");
191 exit(-1);
192 }
193 ompt_multiplex_data_pair_t *data_pair =
194 (ompt_multiplex_data_pair_t *)data_pointer->ptr;
195 data_pair->own_data.ptr = NULL;
196 data_pair->client_data.ptr = NULL;
197 return data_pair;
198 }
199
ompt_multiplex_free_data_pair(ompt_data_t * data_pointer)200 static void ompt_multiplex_free_data_pair(ompt_data_t *data_pointer) {
201 free((*data_pointer).ptr);
202 }
203
ompt_multiplex_get_own_ompt_data(ompt_data_t * data)204 static ompt_data_t *ompt_multiplex_get_own_ompt_data(ompt_data_t *data) {
205 if (!data)
206 return NULL;
207 if (!data->ptr)
208 return NULL;
209 ompt_multiplex_data_pair_t *data_pair =
210 (ompt_multiplex_data_pair_t *)data->ptr;
211 return &(data_pair->own_data);
212 }
213
ompt_multiplex_get_client_ompt_data(ompt_data_t * data)214 static ompt_data_t *ompt_multiplex_get_client_ompt_data(ompt_data_t *data) {
215 if (!data)
216 return NULL;
217 if (!data->ptr)
218 return NULL;
219 ompt_multiplex_data_pair_t *data_pair =
220 (ompt_multiplex_data_pair_t *)data->ptr;
221 return &(data_pair->client_data);
222 }
223 #endif //! defined(OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_THREAD_DATA) ||
224 //! !defined(OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_PARALLEL_DATA) ||
225 //! !defined(OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_TASK_DATA)
226
ompt_multiplex_get_own_thread_data(ompt_data_t * data)227 static ompt_data_t *ompt_multiplex_get_own_thread_data(ompt_data_t *data) {
228 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_THREAD_DATA
229 return ompt_multiplex_get_own_ompt_data(data);
230 #else
231 return data;
232 #endif
233 }
234
ompt_multiplex_get_own_parallel_data(ompt_data_t * data)235 static ompt_data_t *ompt_multiplex_get_own_parallel_data(ompt_data_t *data) {
236 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_PARALLEL_DATA
237 return ompt_multiplex_get_own_ompt_data(data);
238 #else
239 return data;
240 #endif
241 }
242
ompt_multiplex_get_own_task_data(ompt_data_t * data)243 static ompt_data_t *ompt_multiplex_get_own_task_data(ompt_data_t *data) {
244 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_TASK_DATA
245 return ompt_multiplex_get_own_ompt_data(data);
246 #else
247 return data;
248 #endif
249 }
250
ompt_multiplex_get_client_thread_data(ompt_data_t * data)251 static ompt_data_t *ompt_multiplex_get_client_thread_data(ompt_data_t *data) {
252 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_THREAD_DATA
253 return ompt_multiplex_get_client_ompt_data(data);
254 #else
255 return OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_THREAD_DATA(data);
256 #endif
257 }
258
ompt_multiplex_get_client_parallel_data(ompt_data_t * data)259 static ompt_data_t *ompt_multiplex_get_client_parallel_data(ompt_data_t *data) {
260 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_PARALLEL_DATA
261 return ompt_multiplex_get_client_ompt_data(data);
262 #else
263 return OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_PARALLEL_DATA(data);
264 #endif
265 }
266
ompt_multiplex_get_client_task_data(ompt_data_t * data)267 static ompt_data_t *ompt_multiplex_get_client_task_data(ompt_data_t *data) {
268 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_TASK_DATA
269 return ompt_multiplex_get_client_ompt_data(data);
270 #else
271 return OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_TASK_DATA(data);
272 #endif
273 }
274
ompt_multiplex_callback_mutex_acquire(ompt_mutex_t kind,unsigned int hint,unsigned int impl,ompt_wait_id_t wait_id,const void * codeptr_ra)275 static void ompt_multiplex_callback_mutex_acquire(ompt_mutex_t kind,
276 unsigned int hint,
277 unsigned int impl,
278 ompt_wait_id_t wait_id,
279 const void *codeptr_ra) {
280 if (ompt_multiplex_own_callbacks.ompt_callback_mutex_acquire) {
281 ompt_multiplex_own_callbacks.ompt_callback_mutex_acquire(
282 kind, hint, impl, wait_id, codeptr_ra);
283 }
284 if (ompt_multiplex_client_callbacks.ompt_callback_mutex_acquire) {
285 ompt_multiplex_client_callbacks.ompt_callback_mutex_acquire(
286 kind, hint, impl, wait_id, codeptr_ra);
287 }
288 }
289
ompt_multiplex_callback_mutex_acquired(ompt_mutex_t kind,ompt_wait_id_t wait_id,const void * codeptr_ra)290 static void ompt_multiplex_callback_mutex_acquired(ompt_mutex_t kind,
291 ompt_wait_id_t wait_id,
292 const void *codeptr_ra) {
293 if (ompt_multiplex_own_callbacks.ompt_callback_mutex_acquired) {
294 ompt_multiplex_own_callbacks.ompt_callback_mutex_acquired(kind, wait_id,
295 codeptr_ra);
296 }
297 if (ompt_multiplex_client_callbacks.ompt_callback_mutex_acquired) {
298 ompt_multiplex_client_callbacks.ompt_callback_mutex_acquired(kind, wait_id,
299 codeptr_ra);
300 }
301 }
302
ompt_multiplex_callback_mutex_released(ompt_mutex_t kind,ompt_wait_id_t wait_id,const void * codeptr_ra)303 static void ompt_multiplex_callback_mutex_released(ompt_mutex_t kind,
304 ompt_wait_id_t wait_id,
305 const void *codeptr_ra) {
306 if (ompt_multiplex_own_callbacks.ompt_callback_mutex_released) {
307 ompt_multiplex_own_callbacks.ompt_callback_mutex_released(kind, wait_id,
308 codeptr_ra);
309 }
310 if (ompt_multiplex_client_callbacks.ompt_callback_mutex_released) {
311 ompt_multiplex_client_callbacks.ompt_callback_mutex_released(kind, wait_id,
312 codeptr_ra);
313 }
314 }
315
ompt_multiplex_callback_nest_lock(ompt_scope_endpoint_t endpoint,ompt_wait_id_t wait_id,const void * codeptr_ra)316 static void ompt_multiplex_callback_nest_lock(ompt_scope_endpoint_t endpoint,
317 ompt_wait_id_t wait_id,
318 const void *codeptr_ra) {
319 if (ompt_multiplex_own_callbacks.ompt_callback_nest_lock) {
320 ompt_multiplex_own_callbacks.ompt_callback_nest_lock(endpoint, wait_id,
321 codeptr_ra);
322 }
323 if (ompt_multiplex_client_callbacks.ompt_callback_nest_lock) {
324 ompt_multiplex_client_callbacks.ompt_callback_nest_lock(endpoint, wait_id,
325 codeptr_ra);
326 }
327 }
328
ompt_multiplex_callback_sync_region(ompt_sync_region_t kind,ompt_scope_endpoint_t endpoint,ompt_data_t * parallel_data,ompt_data_t * task_data,const void * codeptr_ra)329 static void ompt_multiplex_callback_sync_region(ompt_sync_region_t kind,
330 ompt_scope_endpoint_t endpoint,
331 ompt_data_t *parallel_data,
332 ompt_data_t *task_data,
333 const void *codeptr_ra) {
334 if (ompt_multiplex_own_callbacks.ompt_callback_sync_region) {
335 ompt_multiplex_own_callbacks.ompt_callback_sync_region(
336 kind, endpoint, ompt_multiplex_get_own_parallel_data(parallel_data),
337 ompt_multiplex_get_own_task_data(task_data), codeptr_ra);
338 }
339 if (ompt_multiplex_client_callbacks.ompt_callback_sync_region) {
340 ompt_multiplex_client_callbacks.ompt_callback_sync_region(
341 kind, endpoint, ompt_multiplex_get_client_parallel_data(parallel_data),
342 ompt_multiplex_get_client_task_data(task_data), codeptr_ra);
343 }
344 }
345
ompt_multiplex_callback_sync_region_wait(ompt_sync_region_t kind,ompt_scope_endpoint_t endpoint,ompt_data_t * parallel_data,ompt_data_t * task_data,const void * codeptr_ra)346 static void ompt_multiplex_callback_sync_region_wait(
347 ompt_sync_region_t kind, ompt_scope_endpoint_t endpoint,
348 ompt_data_t *parallel_data, ompt_data_t *task_data,
349 const void *codeptr_ra) {
350 if (ompt_multiplex_own_callbacks.ompt_callback_sync_region_wait) {
351 ompt_multiplex_own_callbacks.ompt_callback_sync_region_wait(
352 kind, endpoint, ompt_multiplex_get_own_parallel_data(parallel_data),
353 ompt_multiplex_get_own_task_data(task_data), codeptr_ra);
354 }
355 if (ompt_multiplex_client_callbacks.ompt_callback_sync_region_wait) {
356 ompt_multiplex_client_callbacks.ompt_callback_sync_region_wait(
357 kind, endpoint, ompt_multiplex_get_client_parallel_data(parallel_data),
358 ompt_multiplex_get_client_task_data(task_data), codeptr_ra);
359 }
360 }
361
ompt_multiplex_callback_flush(ompt_data_t * thread_data,const void * codeptr_ra)362 static void ompt_multiplex_callback_flush(ompt_data_t *thread_data,
363 const void *codeptr_ra) {
364 if (ompt_multiplex_own_callbacks.ompt_callback_flush) {
365 ompt_multiplex_own_callbacks.ompt_callback_flush(
366 ompt_multiplex_get_own_thread_data(thread_data), codeptr_ra);
367 }
368 if (ompt_multiplex_client_callbacks.ompt_callback_flush) {
369 ompt_multiplex_client_callbacks.ompt_callback_flush(
370 ompt_multiplex_get_client_thread_data(thread_data), codeptr_ra);
371 }
372 }
373
ompt_multiplex_callback_cancel(ompt_data_t * task_data,int flags,const void * codeptr_ra)374 static void ompt_multiplex_callback_cancel(ompt_data_t *task_data, int flags,
375 const void *codeptr_ra) {
376 if (ompt_multiplex_own_callbacks.ompt_callback_cancel) {
377 ompt_multiplex_own_callbacks.ompt_callback_cancel(
378 ompt_multiplex_get_own_task_data(task_data), flags, codeptr_ra);
379 }
380 if (ompt_multiplex_client_callbacks.ompt_callback_cancel) {
381 ompt_multiplex_client_callbacks.ompt_callback_cancel(
382 ompt_multiplex_get_client_task_data(task_data), flags, codeptr_ra);
383 }
384 }
385
ompt_multiplex_callback_implicit_task(ompt_scope_endpoint_t endpoint,ompt_data_t * parallel_data,ompt_data_t * task_data,unsigned int team_size,unsigned int thread_num,int flags)386 static void ompt_multiplex_callback_implicit_task(
387 ompt_scope_endpoint_t endpoint, ompt_data_t *parallel_data,
388 ompt_data_t *task_data, unsigned int team_size, unsigned int thread_num,
389 int flags) {
390 if (endpoint == ompt_scope_begin) {
391 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_TASK_DATA
392 ompt_multiplex_allocate_data_pair(task_data);
393 #endif
394 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_PARALLEL_DATA
395 if (flags & ompt_task_initial)
396 ompt_multiplex_allocate_data_pair(parallel_data);
397 #endif
398 if (ompt_multiplex_own_callbacks.ompt_callback_implicit_task) {
399 ompt_multiplex_own_callbacks.ompt_callback_implicit_task(
400 endpoint, ompt_multiplex_get_own_parallel_data(parallel_data),
401 ompt_multiplex_get_own_task_data(task_data), team_size, thread_num,
402 flags);
403 }
404 if (ompt_multiplex_client_callbacks.ompt_callback_implicit_task) {
405 ompt_multiplex_client_callbacks.ompt_callback_implicit_task(
406 endpoint, ompt_multiplex_get_client_parallel_data(parallel_data),
407 ompt_multiplex_get_client_task_data(task_data), team_size, thread_num,
408 flags);
409 }
410 } else {
411 // defines to make sure, callbacks are called in correct order depending on
412 // defines set by the user
413 #if defined(OMPT_MULTIPLEX_CUSTOM_DELETE_TASK_DATA) || \
414 !defined(OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_TASK_DATA)
415 if (ompt_multiplex_own_callbacks.ompt_callback_implicit_task) {
416 ompt_multiplex_own_callbacks.ompt_callback_implicit_task(
417 endpoint, ompt_multiplex_get_own_parallel_data(parallel_data),
418 ompt_multiplex_get_own_task_data(task_data), team_size, thread_num,
419 flags);
420 }
421 #endif
422
423 if (ompt_multiplex_client_callbacks.ompt_callback_implicit_task) {
424 ompt_multiplex_client_callbacks.ompt_callback_implicit_task(
425 endpoint, ompt_multiplex_get_client_parallel_data(parallel_data),
426 ompt_multiplex_get_client_task_data(task_data), team_size, thread_num,
427 flags);
428 }
429
430 #if defined(OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_TASK_DATA) && \
431 !defined(OMPT_MULTIPLEX_CUSTOM_DELETE_TASK_DATA)
432 if (ompt_multiplex_own_callbacks.ompt_callback_implicit_task) {
433 ompt_multiplex_own_callbacks.ompt_callback_implicit_task(
434 endpoint, ompt_multiplex_get_own_parallel_data(parallel_data),
435 ompt_multiplex_get_own_task_data(task_data), team_size, thread_num,
436 flags);
437 }
438 #endif
439
440 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_TASK_DATA
441 ompt_multiplex_free_data_pair(task_data);
442 #endif
443
444 #if defined(OMPT_MULTIPLEX_CUSTOM_DELETE_PARALLEL_DATA)
445 if (flags & ompt_task_initial)
446 OMPT_MULTIPLEX_CUSTOM_DELETE_PARALLEL_DATA(parallel_data);
447 #endif
448 #if defined(OMPT_MULTIPLEX_CUSTOM_DELETE_TASK_DATA)
449 OMPT_MULTIPLEX_CUSTOM_DELETE_TASK_DATA(task_data);
450 #endif
451 }
452 }
453
ompt_multiplex_callback_lock_init(ompt_mutex_t kind,unsigned int hint,unsigned int impl,ompt_wait_id_t wait_id,const void * codeptr_ra)454 static void ompt_multiplex_callback_lock_init(ompt_mutex_t kind,
455 unsigned int hint,
456 unsigned int impl,
457 ompt_wait_id_t wait_id,
458 const void *codeptr_ra) {
459 if (ompt_multiplex_own_callbacks.ompt_callback_lock_init) {
460 ompt_multiplex_own_callbacks.ompt_callback_lock_init(kind, hint, impl,
461 wait_id, codeptr_ra);
462 }
463 if (ompt_multiplex_client_callbacks.ompt_callback_lock_init) {
464 ompt_multiplex_client_callbacks.ompt_callback_lock_init(
465 kind, hint, impl, wait_id, codeptr_ra);
466 }
467 }
468
ompt_multiplex_callback_lock_destroy(ompt_mutex_t kind,ompt_wait_id_t wait_id,const void * codeptr_ra)469 static void ompt_multiplex_callback_lock_destroy(ompt_mutex_t kind,
470 ompt_wait_id_t wait_id,
471 const void *codeptr_ra) {
472 if (ompt_multiplex_own_callbacks.ompt_callback_lock_destroy) {
473 ompt_multiplex_own_callbacks.ompt_callback_lock_destroy(kind, wait_id,
474 codeptr_ra);
475 }
476 if (ompt_multiplex_client_callbacks.ompt_callback_lock_destroy) {
477 ompt_multiplex_client_callbacks.ompt_callback_lock_destroy(kind, wait_id,
478 codeptr_ra);
479 }
480 }
481
ompt_multiplex_callback_work(ompt_work_t wstype,ompt_scope_endpoint_t endpoint,ompt_data_t * parallel_data,ompt_data_t * task_data,uint64_t count,const void * codeptr_ra)482 static void ompt_multiplex_callback_work(ompt_work_t wstype,
483 ompt_scope_endpoint_t endpoint,
484 ompt_data_t *parallel_data,
485 ompt_data_t *task_data, uint64_t count,
486 const void *codeptr_ra) {
487 if (ompt_multiplex_own_callbacks.ompt_callback_work) {
488 ompt_multiplex_own_callbacks.ompt_callback_work(
489 wstype, endpoint, ompt_multiplex_get_own_parallel_data(parallel_data),
490 ompt_multiplex_get_own_task_data(task_data), count, codeptr_ra);
491 }
492 if (ompt_multiplex_client_callbacks.ompt_callback_work) {
493 ompt_multiplex_client_callbacks.ompt_callback_work(
494 wstype, endpoint,
495 ompt_multiplex_get_client_parallel_data(parallel_data),
496 ompt_multiplex_get_client_task_data(task_data), count, codeptr_ra);
497 }
498 }
499
ompt_multiplex_callback_masked(ompt_scope_endpoint_t endpoint,ompt_data_t * parallel_data,ompt_data_t * task_data,const void * codeptr_ra)500 static void ompt_multiplex_callback_masked(ompt_scope_endpoint_t endpoint,
501 ompt_data_t *parallel_data,
502 ompt_data_t *task_data,
503 const void *codeptr_ra) {
504 if (ompt_multiplex_own_callbacks.ompt_callback_masked) {
505 ompt_multiplex_own_callbacks.ompt_callback_masked(
506 endpoint, ompt_multiplex_get_own_parallel_data(parallel_data),
507 ompt_multiplex_get_own_task_data(task_data), codeptr_ra);
508 }
509 if (ompt_multiplex_client_callbacks.ompt_callback_masked) {
510 ompt_multiplex_client_callbacks.ompt_callback_masked(
511 endpoint, ompt_multiplex_get_client_parallel_data(parallel_data),
512 ompt_multiplex_get_client_task_data(task_data), codeptr_ra);
513 }
514 }
515
ompt_multiplex_callback_parallel_begin(ompt_data_t * parent_task_data,const ompt_frame_t * parent_task_frame,ompt_data_t * parallel_data,uint32_t requested_team_size,int flag,const void * codeptr_ra)516 static void ompt_multiplex_callback_parallel_begin(
517 ompt_data_t *parent_task_data, const ompt_frame_t *parent_task_frame,
518 ompt_data_t *parallel_data, uint32_t requested_team_size, int flag,
519 const void *codeptr_ra) {
520 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_PARALLEL_DATA
521 ompt_multiplex_allocate_data_pair(parallel_data);
522 #endif
523 if (ompt_multiplex_own_callbacks.ompt_callback_parallel_begin) {
524 ompt_multiplex_own_callbacks.ompt_callback_parallel_begin(
525 ompt_multiplex_get_own_task_data(parent_task_data), parent_task_frame,
526 ompt_multiplex_get_own_parallel_data(parallel_data),
527 requested_team_size, flag, codeptr_ra);
528 }
529 if (ompt_multiplex_client_callbacks.ompt_callback_parallel_begin) {
530 ompt_multiplex_client_callbacks.ompt_callback_parallel_begin(
531 ompt_multiplex_get_client_task_data(parent_task_data),
532 parent_task_frame,
533 ompt_multiplex_get_client_parallel_data(parallel_data),
534 requested_team_size, flag, codeptr_ra);
535 }
536 }
537
ompt_multiplex_callback_parallel_end(ompt_data_t * parallel_data,ompt_data_t * task_data,int flag,const void * codeptr_ra)538 static void ompt_multiplex_callback_parallel_end(ompt_data_t *parallel_data,
539 ompt_data_t *task_data,
540 int flag,
541 const void *codeptr_ra) {
542 // defines to make sure, callbacks are called in correct order depending on
543 // defines set by the user
544 #if defined(OMPT_MULTIPLEX_CUSTOM_DELETE_PARALLEL_DATA) || \
545 !defined(OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_PARALLEL_DATA)
546 if (ompt_multiplex_own_callbacks.ompt_callback_parallel_end) {
547 ompt_multiplex_own_callbacks.ompt_callback_parallel_end(
548 ompt_multiplex_get_own_parallel_data(parallel_data),
549 ompt_multiplex_get_own_task_data(task_data), flag, codeptr_ra);
550 }
551 #endif
552
553 if (ompt_multiplex_client_callbacks.ompt_callback_parallel_end) {
554 ompt_multiplex_client_callbacks.ompt_callback_parallel_end(
555 ompt_multiplex_get_client_parallel_data(parallel_data),
556 ompt_multiplex_get_client_task_data(task_data), flag, codeptr_ra);
557 }
558
559 #if defined(OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_PARALLEL_DATA) && \
560 !defined(OMPT_MULTIPLEX_CUSTOM_DELETE_PARALLEL_DATA)
561 if (ompt_multiplex_own_callbacks.ompt_callback_parallel_end) {
562 ompt_multiplex_own_callbacks.ompt_callback_parallel_end(
563 ompt_multiplex_get_own_parallel_data(parallel_data),
564 ompt_multiplex_get_own_task_data(task_data), flag, codeptr_ra);
565 }
566 #endif
567
568 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_PARALLEL_DATA
569 ompt_multiplex_free_data_pair(parallel_data);
570 #endif
571
572 #if defined(OMPT_MULTIPLEX_CUSTOM_DELETE_PARALLEL_DATA)
573 OMPT_MULTIPLEX_CUSTOM_DELETE_PARALLEL_DATA(parallel_data);
574 #endif
575 }
576
ompt_multiplex_callback_task_create(ompt_data_t * parent_task_data,const ompt_frame_t * parent_frame,ompt_data_t * new_task_data,int type,int has_dependences,const void * codeptr_ra)577 static void ompt_multiplex_callback_task_create(
578 ompt_data_t *parent_task_data, const ompt_frame_t *parent_frame,
579 ompt_data_t *new_task_data, int type, int has_dependences,
580 const void *codeptr_ra) {
581 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_TASK_DATA
582 ompt_multiplex_allocate_data_pair(new_task_data);
583 #endif
584
585 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_PARALLEL_DATA
586 if (type & ompt_task_initial) {
587 ompt_data_t *parallel_data;
588 ompt_multiplex_get_parallel_info(0, ¶llel_data, NULL);
589 ompt_multiplex_allocate_data_pair(parallel_data);
590 }
591 #endif
592
593 if (ompt_multiplex_own_callbacks.ompt_callback_task_create) {
594 ompt_multiplex_own_callbacks.ompt_callback_task_create(
595 ompt_multiplex_get_own_task_data(parent_task_data), parent_frame,
596 ompt_multiplex_get_own_task_data(new_task_data), type, has_dependences,
597 codeptr_ra);
598 }
599 if (ompt_multiplex_client_callbacks.ompt_callback_task_create) {
600 ompt_multiplex_client_callbacks.ompt_callback_task_create(
601 ompt_multiplex_get_client_task_data(parent_task_data), parent_frame,
602 ompt_multiplex_get_client_task_data(new_task_data), type,
603 has_dependences, codeptr_ra);
604 }
605 }
606
607 static void
ompt_multiplex_callback_task_schedule(ompt_data_t * first_task_data,ompt_task_status_t prior_task_status,ompt_data_t * second_task_data)608 ompt_multiplex_callback_task_schedule(ompt_data_t *first_task_data,
609 ompt_task_status_t prior_task_status,
610 ompt_data_t *second_task_data) {
611 if (prior_task_status != ompt_task_complete) {
612 if (ompt_multiplex_own_callbacks.ompt_callback_task_schedule) {
613 ompt_multiplex_own_callbacks.ompt_callback_task_schedule(
614 ompt_multiplex_get_own_task_data(first_task_data), prior_task_status,
615 ompt_multiplex_get_own_task_data(second_task_data));
616 }
617 if (ompt_multiplex_client_callbacks.ompt_callback_task_schedule) {
618 ompt_multiplex_client_callbacks.ompt_callback_task_schedule(
619 ompt_multiplex_get_client_task_data(first_task_data),
620 prior_task_status,
621 ompt_multiplex_get_client_task_data(second_task_data));
622 }
623 } else {
624 // defines to make sure, callbacks are called in correct order depending on
625 // defines set by the user
626 #if defined(OMPT_MULTIPLEX_CUSTOM_DELETE_TASK_DATA) || \
627 !defined(OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_TASK_DATA)
628 if (ompt_multiplex_own_callbacks.ompt_callback_task_schedule) {
629 ompt_multiplex_own_callbacks.ompt_callback_task_schedule(
630 ompt_multiplex_get_own_task_data(first_task_data), prior_task_status,
631 ompt_multiplex_get_own_task_data(second_task_data));
632 }
633 #endif
634
635 if (ompt_multiplex_client_callbacks.ompt_callback_task_schedule) {
636 ompt_multiplex_client_callbacks.ompt_callback_task_schedule(
637 ompt_multiplex_get_client_task_data(first_task_data),
638 prior_task_status,
639 ompt_multiplex_get_client_task_data(second_task_data));
640 }
641
642 #if defined(OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_TASK_DATA) && \
643 !defined(OMPT_MULTIPLEX_CUSTOM_DELETE_TASK_DATA)
644 if (ompt_multiplex_own_callbacks.ompt_callback_task_schedule) {
645 ompt_multiplex_own_callbacks.ompt_callback_task_schedule(
646 ompt_multiplex_get_own_task_data(first_task_data), prior_task_status,
647 ompt_multiplex_get_own_task_data(second_task_data));
648 }
649 #endif
650
651 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_TASK_DATA
652 ompt_multiplex_free_data_pair(first_task_data);
653 #endif
654
655 #if defined(OMPT_MULTIPLEX_CUSTOM_DELETE_TASK_DATA)
656 OMPT_MULTIPLEX_CUSTOM_DELETE_TASK_DATA(first_task_data);
657 #endif
658 }
659 }
660
ompt_multiplex_callback_dependences(ompt_data_t * task_data,const ompt_dependence_t * deps,int ndeps)661 static void ompt_multiplex_callback_dependences(ompt_data_t *task_data,
662 const ompt_dependence_t *deps,
663 int ndeps) {
664 if (ompt_multiplex_own_callbacks.ompt_callback_dependences) {
665 ompt_multiplex_own_callbacks.ompt_callback_dependences(
666 ompt_multiplex_get_own_task_data(task_data), deps, ndeps);
667 }
668 if (ompt_multiplex_client_callbacks.ompt_callback_dependences) {
669 ompt_multiplex_client_callbacks.ompt_callback_dependences(
670 ompt_multiplex_get_client_task_data(task_data), deps, ndeps);
671 }
672 }
673
674 static void
ompt_multiplex_callback_task_dependence(ompt_data_t * first_task_data,ompt_data_t * second_task_data)675 ompt_multiplex_callback_task_dependence(ompt_data_t *first_task_data,
676 ompt_data_t *second_task_data) {
677 if (ompt_multiplex_own_callbacks.ompt_callback_task_dependence) {
678 ompt_multiplex_own_callbacks.ompt_callback_task_dependence(
679 ompt_multiplex_get_own_task_data(first_task_data),
680 ompt_multiplex_get_own_task_data(second_task_data));
681 }
682 if (ompt_multiplex_client_callbacks.ompt_callback_task_dependence) {
683 ompt_multiplex_client_callbacks.ompt_callback_task_dependence(
684 ompt_multiplex_get_client_task_data(first_task_data),
685 ompt_multiplex_get_client_task_data(second_task_data));
686 }
687 }
688
ompt_multiplex_callback_thread_begin(ompt_thread_t thread_type,ompt_data_t * thread_data)689 static void ompt_multiplex_callback_thread_begin(ompt_thread_t thread_type,
690 ompt_data_t *thread_data) {
691 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_THREAD_DATA
692 ompt_multiplex_allocate_data_pair(thread_data);
693 #endif
694 if (ompt_multiplex_own_callbacks.ompt_callback_thread_begin) {
695 ompt_multiplex_own_callbacks.ompt_callback_thread_begin(
696 thread_type, ompt_multiplex_get_own_thread_data(thread_data));
697 }
698 if (ompt_multiplex_client_callbacks.ompt_callback_thread_begin) {
699 ompt_multiplex_client_callbacks.ompt_callback_thread_begin(
700 thread_type, ompt_multiplex_get_client_thread_data(thread_data));
701 }
702 }
703
ompt_multiplex_callback_thread_end(ompt_data_t * thread_data)704 static void ompt_multiplex_callback_thread_end(ompt_data_t *thread_data) {
705 // defines to make sure, callbacks are called in correct order depending on
706 // defines set by the user
707 #if defined(OMPT_MULTIPLEX_CUSTOM_DELETE_THREAD_DATA) || \
708 !defined(OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_THREAD_DATA)
709 if (ompt_multiplex_own_callbacks.ompt_callback_thread_end) {
710 ompt_multiplex_own_callbacks.ompt_callback_thread_end(
711 ompt_multiplex_get_own_thread_data(thread_data));
712 }
713 #endif
714
715 if (ompt_multiplex_client_callbacks.ompt_callback_thread_end) {
716 ompt_multiplex_client_callbacks.ompt_callback_thread_end(
717 ompt_multiplex_get_client_thread_data(thread_data));
718 }
719
720 #if defined(OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_THREAD_DATA) && \
721 !defined(OMPT_MULTIPLEX_CUSTOM_DELETE_THREAD_DATA)
722 if (ompt_multiplex_own_callbacks.ompt_callback_thread_end) {
723 ompt_multiplex_own_callbacks.ompt_callback_thread_end(
724 ompt_multiplex_get_own_thread_data(thread_data));
725 }
726 #endif
727
728 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_THREAD_DATA
729 ompt_multiplex_free_data_pair(thread_data);
730 #endif
731
732 #if defined(OMPT_MULTIPLEX_CUSTOM_DELETE_THREAD_DATA)
733 OMPT_MULTIPLEX_CUSTOM_DELETE_THREAD_DATA(thread_data);
734 #endif
735 }
736
ompt_multiplex_callback_control_tool(uint64_t command,uint64_t modifier,void * arg,const void * codeptr_ra)737 static int ompt_multiplex_callback_control_tool(uint64_t command,
738 uint64_t modifier, void *arg,
739 const void *codeptr_ra) {
740 int ownRet = 0, clientRet = 0;
741 if (ompt_multiplex_own_callbacks.ompt_callback_control_tool) {
742 ownRet = ompt_multiplex_own_callbacks.ompt_callback_control_tool(
743 command, modifier, arg, codeptr_ra);
744 }
745 if (ompt_multiplex_client_callbacks.ompt_callback_control_tool) {
746 clientRet = ompt_multiplex_client_callbacks.ompt_callback_control_tool(
747 command, modifier, arg, codeptr_ra);
748 }
749 return ownRet < clientRet ? ownRet : clientRet;
750 }
751
ompt_multiplex_callback_target(ompt_target_t kind,ompt_scope_endpoint_t endpoint,int device_num,ompt_data_t * task_data,ompt_id_t target_id,const void * codeptr_ra)752 static void ompt_multiplex_callback_target(
753 ompt_target_t kind, ompt_scope_endpoint_t endpoint, int device_num,
754 ompt_data_t *task_data, ompt_id_t target_id, const void *codeptr_ra) {
755 if (ompt_multiplex_own_callbacks.ompt_callback_target) {
756 ompt_multiplex_own_callbacks.ompt_callback_target(
757 kind, endpoint, device_num, ompt_multiplex_get_own_task_data(task_data),
758 target_id, codeptr_ra);
759 }
760 if (ompt_multiplex_client_callbacks.ompt_callback_target) {
761 ompt_multiplex_client_callbacks.ompt_callback_target(
762 kind, endpoint, device_num,
763 ompt_multiplex_get_client_task_data(task_data), target_id, codeptr_ra);
764 }
765 }
766
ompt_multiplex_callback_target_data_op(ompt_id_t target_id,ompt_id_t host_op_id,ompt_target_data_op_t optype,void * src_addr,int src_device_num,void * dest_addr,int dest_device_num,size_t bytes,const void * codeptr_ra)767 static void ompt_multiplex_callback_target_data_op(
768 ompt_id_t target_id, ompt_id_t host_op_id, ompt_target_data_op_t optype,
769 void *src_addr, int src_device_num, void *dest_addr, int dest_device_num,
770 size_t bytes, const void *codeptr_ra) {
771 if (ompt_multiplex_own_callbacks.ompt_callback_target_data_op) {
772 ompt_multiplex_own_callbacks.ompt_callback_target_data_op(
773 target_id, host_op_id, optype, src_addr, src_device_num, dest_addr,
774 dest_device_num, bytes, codeptr_ra);
775 }
776 if (ompt_multiplex_client_callbacks.ompt_callback_target_data_op) {
777 ompt_multiplex_client_callbacks.ompt_callback_target_data_op(
778 target_id, host_op_id, optype, src_addr, src_device_num, dest_addr,
779 dest_device_num, bytes, codeptr_ra);
780 }
781 }
782
783 static void
ompt_multiplex_callback_target_submit(ompt_id_t target_id,ompt_id_t host_op_id,unsigned int requested_num_teams)784 ompt_multiplex_callback_target_submit(ompt_id_t target_id, ompt_id_t host_op_id,
785 unsigned int requested_num_teams) {
786 if (ompt_multiplex_own_callbacks.ompt_callback_target_submit) {
787 ompt_multiplex_own_callbacks.ompt_callback_target_submit(
788 target_id, host_op_id, requested_num_teams);
789 }
790 if (ompt_multiplex_client_callbacks.ompt_callback_target_submit) {
791 ompt_multiplex_client_callbacks.ompt_callback_target_submit(
792 target_id, host_op_id, requested_num_teams);
793 }
794 }
795
ompt_multiplex_callback_device_initialize(int device_num,const char * type,ompt_device_t * device,ompt_function_lookup_t lookup,const char * documentation)796 static void ompt_multiplex_callback_device_initialize(
797 int device_num, const char *type, ompt_device_t *device,
798 ompt_function_lookup_t lookup, const char *documentation) {
799 if (ompt_multiplex_own_callbacks.ompt_callback_device_initialize) {
800 ompt_multiplex_own_callbacks.ompt_callback_device_initialize(
801 device_num, type, device, lookup, documentation);
802 }
803 if (ompt_multiplex_client_callbacks.ompt_callback_device_initialize) {
804 ompt_multiplex_client_callbacks.ompt_callback_device_initialize(
805 device_num, type, device, lookup, documentation);
806 }
807 }
808
ompt_multiplex_callback_device_finalize(int device_num)809 static void ompt_multiplex_callback_device_finalize(int device_num) {
810 if (ompt_multiplex_own_callbacks.ompt_callback_device_finalize) {
811 ompt_multiplex_own_callbacks.ompt_callback_device_finalize(device_num);
812 }
813 if (ompt_multiplex_client_callbacks.ompt_callback_device_finalize) {
814 ompt_multiplex_client_callbacks.ompt_callback_device_finalize(device_num);
815 }
816 }
817
818 static void
ompt_multiplex_callback_device_load(int device_num,const char * filename,int64_t offset_in_file,void * vma_in_file,size_t bytes,void * host_addr,void * device_addr,uint64_t module_id)819 ompt_multiplex_callback_device_load(int device_num, const char *filename,
820 int64_t offset_in_file, void *vma_in_file,
821 size_t bytes, void *host_addr,
822 void *device_addr, uint64_t module_id) {
823 if (ompt_multiplex_own_callbacks.ompt_callback_device_load) {
824 ompt_multiplex_own_callbacks.ompt_callback_device_load(
825 device_num, filename, offset_in_file, vma_in_file, bytes, host_addr,
826 device_addr, module_id);
827 }
828 if (ompt_multiplex_client_callbacks.ompt_callback_device_load) {
829 ompt_multiplex_client_callbacks.ompt_callback_device_load(
830 device_num, filename, offset_in_file, vma_in_file, bytes, host_addr,
831 device_addr, module_id);
832 }
833 }
834
ompt_multiplex_callback_device_unload(int device_num,uint64_t module_id)835 static void ompt_multiplex_callback_device_unload(int device_num,
836 uint64_t module_id) {
837 if (ompt_multiplex_own_callbacks.ompt_callback_device_unload) {
838 ompt_multiplex_own_callbacks.ompt_callback_device_unload(device_num,
839 module_id);
840 }
841 if (ompt_multiplex_client_callbacks.ompt_callback_device_unload) {
842 ompt_multiplex_client_callbacks.ompt_callback_device_unload(device_num,
843 module_id);
844 }
845 }
846
847 static void
ompt_multiplex_callback_target_map(ompt_id_t target_id,unsigned int nitems,void ** host_addr,void ** device_addr,size_t * bytes,unsigned int * mapping_flags,const void * codeptr_ra)848 ompt_multiplex_callback_target_map(ompt_id_t target_id, unsigned int nitems,
849 void **host_addr, void **device_addr,
850 size_t *bytes, unsigned int *mapping_flags,
851 const void *codeptr_ra) {
852 if (ompt_multiplex_own_callbacks.ompt_callback_target_map) {
853 ompt_multiplex_own_callbacks.ompt_callback_target_map(
854 target_id, nitems, host_addr, device_addr, bytes, mapping_flags,
855 codeptr_ra);
856 }
857 if (ompt_multiplex_client_callbacks.ompt_callback_target_map) {
858 ompt_multiplex_client_callbacks.ompt_callback_target_map(
859 target_id, nitems, host_addr, device_addr, bytes, mapping_flags,
860 codeptr_ra);
861 }
862 }
863
ompt_multiplex_callback_reduction(ompt_sync_region_t kind,ompt_scope_endpoint_t endpoint,ompt_data_t * parallel_data,ompt_data_t * task_data,const void * codeptr_ra)864 static void ompt_multiplex_callback_reduction(ompt_sync_region_t kind,
865 ompt_scope_endpoint_t endpoint,
866 ompt_data_t *parallel_data,
867 ompt_data_t *task_data,
868 const void *codeptr_ra) {
869 if (ompt_multiplex_own_callbacks.ompt_callback_reduction) {
870 ompt_multiplex_own_callbacks.ompt_callback_reduction(
871 kind, endpoint, ompt_multiplex_get_own_parallel_data(parallel_data),
872 ompt_multiplex_get_own_task_data(task_data), codeptr_ra);
873 }
874 if (ompt_multiplex_client_callbacks.ompt_callback_reduction) {
875 ompt_multiplex_client_callbacks.ompt_callback_reduction(
876 kind, endpoint, ompt_multiplex_get_client_parallel_data(parallel_data),
877 ompt_multiplex_get_client_task_data(task_data), codeptr_ra);
878 }
879 }
880
ompt_multiplex_callback_dispatch(ompt_data_t * parallel_data,ompt_data_t * task_data,ompt_dispatch_t kind,ompt_data_t instance)881 static void ompt_multiplex_callback_dispatch(ompt_data_t *parallel_data,
882 ompt_data_t *task_data,
883 ompt_dispatch_t kind,
884 ompt_data_t instance) {
885 if (ompt_multiplex_own_callbacks.ompt_callback_dispatch) {
886 ompt_multiplex_own_callbacks.ompt_callback_dispatch(
887 ompt_multiplex_get_own_parallel_data(parallel_data),
888 ompt_multiplex_get_own_task_data(task_data), kind, instance);
889 }
890 if (ompt_multiplex_client_callbacks.ompt_callback_dispatch) {
891 ompt_multiplex_client_callbacks.ompt_callback_dispatch(
892 ompt_multiplex_get_client_parallel_data(parallel_data),
893 ompt_multiplex_get_client_task_data(task_data), kind, instance);
894 }
895 }
896
897 // runtime entry functions
898
ompt_multiplex_own_get_task_info(int ancestor_level,int * type,ompt_data_t ** task_data,ompt_frame_t ** task_frame,ompt_data_t ** parallel_data,int * thread_num)899 int ompt_multiplex_own_get_task_info(int ancestor_level, int *type,
900 ompt_data_t **task_data,
901 ompt_frame_t **task_frame,
902 ompt_data_t **parallel_data,
903 int *thread_num) {
904 int ret = ompt_multiplex_get_task_info(ancestor_level, type, task_data,
905 task_frame, parallel_data, thread_num);
906
907 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_TASK_DATA
908 if (task_data)
909 *task_data = ompt_multiplex_get_own_ompt_data(*task_data);
910 #endif
911 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_PARALLEL_DATA
912 if (parallel_data)
913 *parallel_data = ompt_multiplex_get_own_ompt_data(*parallel_data);
914 #endif
915 return ret;
916 }
917
ompt_multiplex_client_get_task_info(int ancestor_level,int * type,ompt_data_t ** task_data,ompt_frame_t ** task_frame,ompt_data_t ** parallel_data,int * thread_num)918 int ompt_multiplex_client_get_task_info(int ancestor_level, int *type,
919 ompt_data_t **task_data,
920 ompt_frame_t **task_frame,
921 ompt_data_t **parallel_data,
922 int *thread_num) {
923 int ret = ompt_multiplex_get_task_info(ancestor_level, type, task_data,
924 task_frame, parallel_data, thread_num);
925
926 if (task_data)
927 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_TASK_DATA
928 *task_data = ompt_multiplex_get_client_ompt_data(*task_data);
929 #else
930 *task_data = OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_TASK_DATA(*task_data);
931 #endif
932
933 if (parallel_data)
934 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_PARALLEL_DATA
935 *parallel_data = ompt_multiplex_get_client_ompt_data(*parallel_data);
936 #else
937 *parallel_data =
938 OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_PARALLEL_DATA(*parallel_data);
939 #endif
940 return ret;
941 }
942
ompt_multiplex_own_get_thread_data()943 ompt_data_t *ompt_multiplex_own_get_thread_data() {
944 ompt_data_t *ret;
945 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_THREAD_DATA
946 ret = ompt_multiplex_get_own_ompt_data(ompt_multiplex_get_thread_data());
947 #else
948 ret = ompt_multiplex_get_thread_data();
949 #endif
950 return ret;
951 }
952
ompt_multiplex_client_get_thread_data()953 ompt_data_t *ompt_multiplex_client_get_thread_data() {
954 ompt_data_t *ret;
955 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_THREAD_DATA
956 ret = ompt_multiplex_get_client_ompt_data(ompt_multiplex_get_thread_data());
957 #else
958 ret = OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_THREAD_DATA(
959 ompt_multiplex_get_thread_data());
960 #endif
961 return ret;
962 }
963
ompt_multiplex_own_get_parallel_info(int ancestor_level,ompt_data_t ** parallel_data,int * team_size)964 int ompt_multiplex_own_get_parallel_info(int ancestor_level,
965 ompt_data_t **parallel_data,
966 int *team_size) {
967 int ret = ompt_multiplex_get_parallel_info(ancestor_level, parallel_data,
968 team_size);
969 if (parallel_data)
970 *parallel_data = ompt_multiplex_get_own_parallel_data(*parallel_data);
971 return ret;
972 }
973
ompt_multiplex_client_get_parallel_info(int ancestor_level,ompt_data_t ** parallel_data,int * team_size)974 int ompt_multiplex_client_get_parallel_info(int ancestor_level,
975 ompt_data_t **parallel_data,
976 int *team_size) {
977 int ret = ompt_multiplex_get_parallel_info(ancestor_level, parallel_data,
978 team_size);
979 if (parallel_data)
980 #ifndef OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_PARALLEL_DATA
981 *parallel_data = ompt_multiplex_get_client_ompt_data(*parallel_data);
982 #else
983 *parallel_data =
984 OMPT_MULTIPLEX_CUSTOM_GET_CLIENT_PARALLEL_DATA(*parallel_data);
985 #endif
986 return ret;
987 }
988
ompt_multiplex_own_set_callback(ompt_callbacks_t which,ompt_callback_t callback)989 OMPT_API_ROUTINE int ompt_multiplex_own_set_callback(ompt_callbacks_t which,
990 ompt_callback_t callback) {
991 switch (which) {
992
993 #define ompt_event_macro(event_name, callback_type, event_id) \
994 case ompt_##event_name: \
995 ompt_multiplex_own_callbacks.ompt_##event_name = (callback_type)callback; \
996 if (ompt_multiplex_implementation_status.ompt_##event_name == -1) \
997 return ompt_multiplex_implementation_status.ompt_##event_name = \
998 ompt_multiplex_set_callback( \
999 ompt_##event_name, \
1000 (ompt_callback_t)&ompt_multiplex_##event_name); \
1001 else \
1002 return ompt_multiplex_implementation_status.ompt_##event_name
1003
1004 OMPT_LOAD_CLIENT_FOREACH_OMPT_EVENT(ompt_event_macro)
1005
1006 #undef ompt_event_macro
1007
1008 default:
1009 return ompt_set_error;
1010 }
1011 }
1012
1013 OMPT_API_ROUTINE int
ompt_multiplex_client_set_callback(ompt_callbacks_t which,ompt_callback_t callback)1014 ompt_multiplex_client_set_callback(ompt_callbacks_t which,
1015 ompt_callback_t callback) {
1016 switch (which) {
1017
1018 #define ompt_event_macro(event_name, callback_type, event_id) \
1019 case ompt_##event_name: \
1020 ompt_multiplex_client_callbacks.ompt_##event_name = \
1021 (callback_type)callback; \
1022 if (ompt_multiplex_implementation_status.ompt_##event_name == -1) \
1023 return ompt_multiplex_implementation_status.ompt_##event_name = \
1024 ompt_multiplex_set_callback( \
1025 ompt_##event_name, \
1026 (ompt_callback_t)&ompt_multiplex_##event_name); \
1027 else \
1028 return ompt_multiplex_implementation_status.ompt_##event_name
1029
1030 OMPT_LOAD_CLIENT_FOREACH_OMPT_EVENT(ompt_event_macro)
1031
1032 #undef ompt_event_macro
1033
1034 default:
1035 return ompt_set_error;
1036 }
1037 }
1038
ompt_multiplex_own_lookup(const char * name)1039 ompt_interface_fn_t ompt_multiplex_own_lookup(const char *name) {
1040 if (!strcmp(name, "ompt_set_callback"))
1041 return (ompt_interface_fn_t)&ompt_multiplex_own_set_callback;
1042 else if (!strcmp(name, "ompt_get_task_info"))
1043 return (ompt_interface_fn_t)&ompt_multiplex_own_get_task_info;
1044 else if (!strcmp(name, "ompt_get_thread_data"))
1045 return (ompt_interface_fn_t)&ompt_multiplex_own_get_thread_data;
1046 else if (!strcmp(name, "ompt_get_parallel_info"))
1047 return (ompt_interface_fn_t)&ompt_multiplex_own_get_parallel_info;
1048 else
1049 return ompt_multiplex_lookup_function(name);
1050 }
1051
ompt_multiplex_client_lookup(const char * name)1052 ompt_interface_fn_t ompt_multiplex_client_lookup(const char *name) {
1053 if (!strcmp(name, "ompt_set_callback"))
1054 return (ompt_interface_fn_t)&ompt_multiplex_client_set_callback;
1055 else if (!strcmp(name, "ompt_get_task_info"))
1056 return (ompt_interface_fn_t)&ompt_multiplex_client_get_task_info;
1057 else if (!strcmp(name, "ompt_get_thread_data"))
1058 return (ompt_interface_fn_t)&ompt_multiplex_client_get_thread_data;
1059 else if (!strcmp(name, "ompt_get_parallel_info"))
1060 return (ompt_interface_fn_t)&ompt_multiplex_client_get_parallel_info;
1061 else
1062 return ompt_multiplex_lookup_function(name);
1063 }
1064
ompt_multiplex_initialize(ompt_function_lookup_t lookup,int initial_device_num,ompt_data_t * data)1065 int ompt_multiplex_initialize(ompt_function_lookup_t lookup,
1066 int initial_device_num, ompt_data_t *data) {
1067 ompt_multiplex_lookup_function = lookup;
1068 ompt_multiplex_set_callback =
1069 (ompt_set_callback_t)lookup("ompt_set_callback");
1070 ompt_multiplex_get_task_info =
1071 (ompt_get_task_info_t)lookup("ompt_get_task_info");
1072 ompt_multiplex_get_thread_data =
1073 (ompt_get_thread_data_t)lookup("ompt_get_thread_data");
1074 ompt_multiplex_get_parallel_info =
1075 (ompt_get_parallel_info_t)lookup("ompt_get_parallel_info");
1076
1077 // initialize ompt_multiplex_implementation_status
1078 #define ompt_event_macro(event_name, callback_type, event_id) \
1079 ompt_multiplex_implementation_status.ompt_##event_name = -1
1080
1081 OMPT_LOAD_CLIENT_FOREACH_OMPT_EVENT(ompt_event_macro)
1082
1083 #undef ompt_event_macro
1084
1085 int ownRet = ompt_multiplex_own_fns->initialize(
1086 ompt_multiplex_own_lookup, initial_device_num,
1087 &(ompt_multiplex_own_fns->tool_data));
1088 int clientRet = 0;
1089 if (ompt_multiplex_client_fns)
1090 clientRet = ompt_multiplex_client_fns->initialize(
1091 ompt_multiplex_client_lookup, initial_device_num,
1092 &(ompt_multiplex_client_fns->tool_data));
1093
1094 return ownRet > clientRet ? ownRet : clientRet;
1095 }
1096
ompt_multiplex_finalize(ompt_data_t * fns)1097 void ompt_multiplex_finalize(ompt_data_t *fns) {
1098 if (ompt_multiplex_client_fns)
1099 ompt_multiplex_client_fns->finalize(
1100 &(ompt_multiplex_client_fns->tool_data));
1101 ompt_multiplex_own_fns->finalize(&(ompt_multiplex_own_fns->tool_data));
1102 }
1103
1104 #ifdef __cplusplus
1105 extern "C" {
1106 #endif
1107
1108 // forward declaration because of name shifting from ompt_start_tool
1109 // to ompt_multiplex_own_start_tool below
1110 ompt_start_tool_result_t *
1111 ompt_multiplex_own_start_tool(unsigned int omp_version,
1112 const char *runtime_version);
1113
ompt_start_tool(unsigned int omp_version,const char * runtime_version)1114 ompt_start_tool_result_t *ompt_start_tool(unsigned int omp_version,
1115 const char *runtime_version) {
1116 setup_verbose_init();
1117 OMPT_VERBOSE_INIT_PRINT(
1118 "----- START LOGGING OF CLIENT TOOL REGISTRATION -----\n");
1119 // try loading client tool
1120 OMPT_VERBOSE_INIT_PRINT("Search for " CLIENT_TOOL_LIBRARIES_VAR
1121 " env var... ");
1122 const char *tool_libs = getenv(CLIENT_TOOL_LIBRARIES_VAR);
1123 if (tool_libs) {
1124 OMPT_VERBOSE_INIT_CONTINUED_PRINT("Sucess.\n");
1125 OMPT_VERBOSE_INIT_PRINT(CLIENT_TOOL_LIBRARIES_VAR " = %s\n", tool_libs);
1126 // copy environement variable
1127 char *tool_libs_buffer = strdup(tool_libs);
1128 if (!tool_libs_buffer) {
1129 printf("strdup Error (%i)\n", errno);
1130 exit(-1);
1131 }
1132
1133 int progress = 0;
1134 // Reset dl-error
1135 dlerror();
1136 while (progress < strlen(tool_libs)) {
1137 ompt_multiplex_client_fns = NULL;
1138 ompt_start_tool_result_t *(*client_start_tool)(unsigned int,
1139 const char *) = NULL;
1140 OMPT_VERBOSE_INIT_PRINT(
1141 "Look for candidates within " CLIENT_TOOL_LIBRARIES_VAR "...\n");
1142 int tmp_progress = progress;
1143 while (tmp_progress < strlen(tool_libs) &&
1144 tool_libs_buffer[tmp_progress] != ':')
1145 tmp_progress++;
1146 if (tmp_progress < strlen(tool_libs))
1147 tool_libs_buffer[tmp_progress] = 0;
1148 OMPT_VERBOSE_INIT_PRINT("Try out one candidate...\n");
1149 char *fname = tool_libs_buffer + progress;
1150 OMPT_VERBOSE_INIT_PRINT("Opening %s... ", fname);
1151 void *h = dlopen(fname, RTLD_LAZY);
1152 if (h) {
1153 client_start_tool =
1154 (ompt_start_tool_result_t * (*)(unsigned int, const char *))
1155 dlsym(h, "ompt_start_tool");
1156 if (client_start_tool &&
1157 (ompt_multiplex_client_fns =
1158 (*client_start_tool)(omp_version, runtime_version))) {
1159 OMPT_VERBOSE_INIT_CONTINUED_PRINT("Sucess.\n");
1160 OMPT_VERBOSE_INIT_PRINT(
1161 "Tool was started and is using the OMPT interface.\n");
1162 break;
1163 } else {
1164 OMPT_VERBOSE_INIT_CONTINUED_PRINT(
1165 "Failed: client_start_tool = %p, ompt_multiplex_client_fns = %p, "
1166 "%s\n",
1167 client_start_tool, ompt_multiplex_client_fns, dlerror());
1168 }
1169 } else {
1170 OMPT_VERBOSE_INIT_CONTINUED_PRINT("Failed: %s\n", dlerror());
1171 printf("Loading %s from %s failed with: %s\n",
1172 tool_libs_buffer + progress, CLIENT_TOOL_LIBRARIES_VAR,
1173 dlerror());
1174 }
1175 progress = tmp_progress + 1;
1176 }
1177 free(tool_libs_buffer);
1178 OMPT_VERBOSE_INIT_PRINT(
1179 "----- END LOGGING OF CLIENT TOOL REGISTRATION -----\n");
1180 }
1181 // load own tool
1182 OMPT_VERBOSE_INIT_PRINT(
1183 "----- START LOGGING OF OWN TOOL REGISTRATION -----\n");
1184 ompt_multiplex_own_fns =
1185 ompt_multiplex_own_start_tool(omp_version, runtime_version);
1186 OMPT_VERBOSE_INIT_PRINT("ompt_multiplex_own_fns = %p\n",
1187 ompt_multiplex_own_fns);
1188 OMPT_VERBOSE_INIT_PRINT("----- END LOGGING OF OWN TOOL REGISTRATION -----\n");
1189 // return multiplexed versions
1190 static ompt_start_tool_result_t ompt_start_tool_result = {
1191 &ompt_multiplex_initialize, &ompt_multiplex_finalize, {0}};
1192 if (verbose_init && verbose_file != stderr && verbose_file != stdout)
1193 fclose(verbose_file);
1194 if (!ompt_multiplex_client_fns)
1195 return ompt_multiplex_own_fns;
1196 if (!ompt_multiplex_own_fns)
1197 return ompt_multiplex_client_fns;
1198 return &ompt_start_tool_result;
1199 }
1200 #ifdef __cplusplus
1201 }
1202 #endif
1203
1204 // We rename the ompt_start_tool function of the OMPT tool and call the
1205 // renamed function from the ompt_start_tool function defined above.
1206 #define ompt_start_tool ompt_multiplex_own_start_tool
1207
1208 #endif /* OMPT_MULTIPLEX_H */
1209