2 * include/omp-tools.h.var
5 //===----------------------------------------------------------------------===//
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
11 //===----------------------------------------------------------------------===//
16 /*****************************************************************************
17 * system include files
18 *****************************************************************************/
23 #ifdef DEPRECATION_WARNINGS
25 # define DEPRECATED_51 [[deprecated("as of 5.1")]]
27 # define DEPRECATED_51 __attribute__((deprecated("as of 5.1")))
33 /*****************************************************************************
35 *****************************************************************************/
37 #define FOREACH_OMPT_INQUIRY_FN(macro) \
38 macro (ompt_enumerate_states) \
39 macro (ompt_enumerate_mutex_impls) \
41 macro (ompt_set_callback) \
42 macro (ompt_get_callback) \
44 macro (ompt_get_state) \
46 macro (ompt_get_parallel_info) \
47 macro (ompt_get_task_info) \
48 macro (ompt_get_task_memory) \
49 macro (ompt_get_thread_data) \
50 macro (ompt_get_unique_id) \
51 macro (ompt_finalize_tool) \
53 macro(ompt_get_num_procs) \
54 macro(ompt_get_num_places) \
55 macro(ompt_get_place_proc_ids) \
56 macro(ompt_get_place_num) \
57 macro(ompt_get_partition_place_nums) \
58 macro(ompt_get_proc_id) \
60 macro(ompt_get_target_info) \
61 macro(ompt_get_num_devices)
63 #define FOREACH_OMPT_STATE(macro) \
65 /* first available state */ \
66 macro (ompt_state_undefined, 0x102) /* undefined thread state */ \
68 /* work states (0..15) */ \
69 macro (ompt_state_work_serial, 0x000) /* working outside parallel */ \
70 macro (ompt_state_work_parallel, 0x001) /* working within parallel */ \
71 macro (ompt_state_work_reduction, 0x002) /* performing a reduction */ \
73 /* barrier wait states (16..31) */ \
74 macro (ompt_state_wait_barrier, 0x010) /* waiting at a barrier */ \
75 macro (ompt_state_wait_barrier_implicit_parallel, 0x011) \
76 /* implicit barrier at the end of parallel region */\
77 macro (ompt_state_wait_barrier_implicit_workshare, 0x012) \
78 /* implicit barrier at the end of worksharing */ \
79 macro (ompt_state_wait_barrier_implicit, 0x013) /* implicit barrier */ \
80 macro (ompt_state_wait_barrier_explicit, 0x014) /* explicit barrier */ \
82 /* task wait states (32..63) */ \
83 macro (ompt_state_wait_taskwait, 0x020) /* waiting at a taskwait */ \
84 macro (ompt_state_wait_taskgroup, 0x021) /* waiting at a taskgroup */ \
86 /* mutex wait states (64..127) */ \
87 macro (ompt_state_wait_mutex, 0x040) \
88 macro (ompt_state_wait_lock, 0x041) /* waiting for lock */ \
89 macro (ompt_state_wait_critical, 0x042) /* waiting for critical */ \
90 macro (ompt_state_wait_atomic, 0x043) /* waiting for atomic */ \
91 macro (ompt_state_wait_ordered, 0x044) /* waiting for ordered */ \
93 /* target wait states (128..255) */ \
94 macro (ompt_state_wait_target, 0x080) /* waiting for target region */ \
95 macro (ompt_state_wait_target_map, 0x081) /* waiting for target data mapping operation */ \
96 macro (ompt_state_wait_target_update, 0x082) /* waiting for target update operation */ \
98 /* misc (256..511) */ \
99 macro (ompt_state_idle, 0x100) /* waiting for work */ \
100 macro (ompt_state_overhead, 0x101) /* overhead excluding wait states */ \
102 /* implementation-specific states (512..) */
105 #define FOREACH_KMP_MUTEX_IMPL(macro) \
106 macro (kmp_mutex_impl_none, 0) /* unknown implementation */ \
107 macro (kmp_mutex_impl_spin, 1) /* based on spin */ \
108 macro (kmp_mutex_impl_queuing, 2) /* based on some fair policy */ \
109 macro (kmp_mutex_impl_speculative, 3) /* based on HW-supported speculation */
111 #define FOREACH_OMPT_HOST_EVENT(macro) \
113 /*--- Mandatory Events ---*/ \
114 macro (ompt_callback_thread_begin, ompt_callback_thread_begin_t, 1) /* thread begin */ \
115 macro (ompt_callback_thread_end, ompt_callback_thread_end_t, 2) /* thread end */ \
117 macro (ompt_callback_parallel_begin, ompt_callback_parallel_begin_t, 3) /* parallel begin */ \
118 macro (ompt_callback_parallel_end, ompt_callback_parallel_end_t, 4) /* parallel end */ \
120 macro (ompt_callback_task_create, ompt_callback_task_create_t, 5) /* task begin */ \
121 macro (ompt_callback_task_schedule, ompt_callback_task_schedule_t, 6) /* task schedule */ \
122 macro (ompt_callback_implicit_task, ompt_callback_implicit_task_t, 7) /* implicit task */ \
124 macro (ompt_callback_control_tool, ompt_callback_control_tool_t, 11) /* control tool */ \
126 /* Optional Events */ \
127 macro (ompt_callback_sync_region_wait, ompt_callback_sync_region_t, 16) /* sync region wait begin or end */ \
129 macro (ompt_callback_mutex_released, ompt_callback_mutex_t, 17) /* mutex released */ \
131 macro (ompt_callback_dependences, ompt_callback_dependences_t, 18) /* report task dependences */ \
132 macro (ompt_callback_task_dependence, ompt_callback_task_dependence_t, 19) /* report task dependence */ \
134 macro (ompt_callback_work, ompt_callback_work_t, 20) /* task at work begin or end */ \
136 macro (ompt_callback_masked, ompt_callback_masked_t, 21) /* task at masked begin or end */ \
138 macro (ompt_callback_sync_region, ompt_callback_sync_region_t, 23) /* sync region begin or end */ \
140 macro (ompt_callback_lock_init, ompt_callback_mutex_acquire_t, 24) /* lock init */ \
141 macro (ompt_callback_lock_destroy, ompt_callback_mutex_t, 25) /* lock destroy */ \
143 macro (ompt_callback_mutex_acquire, ompt_callback_mutex_acquire_t, 26) /* mutex acquire */ \
144 macro (ompt_callback_mutex_acquired, ompt_callback_mutex_t, 27) /* mutex acquired */ \
146 macro (ompt_callback_nest_lock, ompt_callback_nest_lock_t, 28) /* nest lock */ \
148 macro (ompt_callback_flush, ompt_callback_flush_t, 29) /* after executing flush */ \
150 macro (ompt_callback_cancel, ompt_callback_cancel_t, 30) /* cancel innermost binding region */ \
152 macro (ompt_callback_reduction, ompt_callback_sync_region_t, 31) /* reduction */ \
154 macro (ompt_callback_dispatch, ompt_callback_dispatch_t, 32) /* dispatch of work */ \
155 macro (ompt_callback_error, ompt_callback_error_t, 37) /* error */
157 #define FOREACH_OMPT_DEVICE_EVENT(macro) \
158 /*--- Mandatory Events ---*/ \
159 macro (ompt_callback_device_initialize, ompt_callback_device_initialize_t, 12) /* device initialize */ \
160 macro (ompt_callback_device_finalize, ompt_callback_device_finalize_t, 13) /* device finalize */ \
162 macro (ompt_callback_device_load, ompt_callback_device_load_t, 14) /* device load */ \
163 macro (ompt_callback_device_unload, ompt_callback_device_unload_t, 15) /* device unload */
165 #define FOREACH_OMPT_NOEMI_EVENT(macro) \
166 /*--- Mandatory Events ---*/ \
167 macro (ompt_callback_target, ompt_callback_target_t, 8) /* target */ \
168 macro (ompt_callback_target_data_op, ompt_callback_target_data_op_t, 9) /* target data op */ \
169 macro (ompt_callback_target_submit, ompt_callback_target_submit_t, 10) /* target submit */ \
170 /* Optional Events */ \
171 macro (ompt_callback_target_map, ompt_callback_target_map_t, 22) /* target map */
173 #define FOREACH_OMPT_EMI_EVENT(macro) \
174 /*--- Mandatory Events ---*/ \
175 macro (ompt_callback_target_emi, ompt_callback_target_emi_t, 33) /* target */ \
176 macro (ompt_callback_target_data_op_emi,ompt_callback_target_data_op_emi_t,34) /* target data op */ \
177 macro (ompt_callback_target_submit_emi, ompt_callback_target_submit_emi_t, 35) /* target submit */ \
178 /* Optional Events */ \
179 macro (ompt_callback_target_map_emi, ompt_callback_target_map_emi_t, 36) /* target map */
181 #define FOREACH_OMPT_50_TARGET_EVENT(macro) \
182 FOREACH_OMPT_DEVICE_EVENT(macro) \
183 FOREACH_OMPT_NOEMI_EVENT(macro)
185 #define FOREACH_OMPT_51_TARGET_EVENT(macro) \
186 FOREACH_OMPT_DEVICE_EVENT(macro) \
187 FOREACH_OMPT_EMI_EVENT(macro)
189 #define FOREACH_OMPT_EVENT(macro) \
190 FOREACH_OMPT_HOST_EVENT(macro) \
191 FOREACH_OMPT_DEVICE_EVENT(macro) \
192 FOREACH_OMPT_NOEMI_EVENT(macro) \
193 FOREACH_OMPT_EMI_EVENT(macro)
195 #define FOREACH_OMPT_51_EVENT(macro) \
196 FOREACH_OMPT_HOST_EVENT(macro) \
197 FOREACH_OMPT_DEVICE_EVENT(macro) \
198 FOREACH_OMPT_EMI_EVENT(macro)
200 /*****************************************************************************
201 * implementation specific types
202 *****************************************************************************/
204 typedef enum kmp_mutex_impl_t {
205 #define kmp_mutex_impl_macro(impl, code) impl = code,
206 FOREACH_KMP_MUTEX_IMPL(kmp_mutex_impl_macro)
207 #undef kmp_mutex_impl_macro
210 /*****************************************************************************
211 * definitions generated from spec
212 *****************************************************************************/
214 typedef enum ompt_callbacks_t {
215 ompt_callback_thread_begin = 1,
216 ompt_callback_thread_end = 2,
217 ompt_callback_parallel_begin = 3,
218 ompt_callback_parallel_end = 4,
219 ompt_callback_task_create = 5,
220 ompt_callback_task_schedule = 6,
221 ompt_callback_implicit_task = 7,
222 ompt_callback_target = 8,
223 ompt_callback_target_data_op = 9,
224 ompt_callback_target_submit = 10,
225 ompt_callback_control_tool = 11,
226 ompt_callback_device_initialize = 12,
227 ompt_callback_device_finalize = 13,
228 ompt_callback_device_load = 14,
229 ompt_callback_device_unload = 15,
230 ompt_callback_sync_region_wait = 16,
231 ompt_callback_mutex_released = 17,
232 ompt_callback_dependences = 18,
233 ompt_callback_task_dependence = 19,
234 ompt_callback_work = 20,
235 ompt_callback_master DEPRECATED_51 = 21,
236 ompt_callback_masked = 21,
237 ompt_callback_target_map = 22,
238 ompt_callback_sync_region = 23,
239 ompt_callback_lock_init = 24,
240 ompt_callback_lock_destroy = 25,
241 ompt_callback_mutex_acquire = 26,
242 ompt_callback_mutex_acquired = 27,
243 ompt_callback_nest_lock = 28,
244 ompt_callback_flush = 29,
245 ompt_callback_cancel = 30,
246 ompt_callback_reduction = 31,
247 ompt_callback_dispatch = 32,
248 ompt_callback_target_emi = 33,
249 ompt_callback_target_data_op_emi = 34,
250 ompt_callback_target_submit_emi = 35,
251 ompt_callback_target_map_emi = 36,
252 ompt_callback_error = 37
255 typedef enum ompt_record_t {
256 ompt_record_ompt = 1,
257 ompt_record_native = 2,
258 ompt_record_invalid = 3
261 typedef enum ompt_record_native_t {
262 ompt_record_native_info = 1,
263 ompt_record_native_event = 2
264 } ompt_record_native_t;
266 typedef enum ompt_set_result_t {
269 ompt_set_impossible = 2,
270 ompt_set_sometimes = 3,
271 ompt_set_sometimes_paired = 4,
275 typedef uint64_t ompt_id_t;
277 typedef uint64_t ompt_device_time_t;
279 typedef uint64_t ompt_buffer_cursor_t;
281 typedef enum ompt_thread_t {
282 ompt_thread_initial = 1,
283 ompt_thread_worker = 2,
284 ompt_thread_other = 3,
285 ompt_thread_unknown = 4
288 typedef enum ompt_scope_endpoint_t {
289 ompt_scope_begin = 1,
291 ompt_scope_beginend = 3
292 } ompt_scope_endpoint_t;
294 typedef enum ompt_dispatch_t {
295 ompt_dispatch_iteration = 1,
296 ompt_dispatch_section = 2,
297 ompt_dispatch_ws_loop_chunk = 3,
298 ompt_dispatch_taskloop_chunk = 4,
299 ompt_dispatch_distribute_chunk = 5
302 typedef enum ompt_sync_region_t {
303 ompt_sync_region_barrier DEPRECATED_51 = 1,
304 ompt_sync_region_barrier_implicit DEPRECATED_51 = 2,
305 ompt_sync_region_barrier_explicit = 3,
306 ompt_sync_region_barrier_implementation = 4,
307 ompt_sync_region_taskwait = 5,
308 ompt_sync_region_taskgroup = 6,
309 ompt_sync_region_reduction = 7,
310 ompt_sync_region_barrier_implicit_workshare = 8,
311 ompt_sync_region_barrier_implicit_parallel = 9,
312 ompt_sync_region_barrier_teams = 10
313 } ompt_sync_region_t;
315 typedef enum ompt_target_data_op_t {
316 ompt_target_data_alloc = 1,
317 ompt_target_data_transfer_to_device = 2,
318 ompt_target_data_transfer_from_device = 3,
319 ompt_target_data_delete = 4,
320 ompt_target_data_associate = 5,
321 ompt_target_data_disassociate = 6,
322 ompt_target_data_alloc_async = 17,
323 ompt_target_data_transfer_to_device_async = 18,
324 ompt_target_data_transfer_from_device_async = 19,
325 ompt_target_data_delete_async = 20
326 } ompt_target_data_op_t;
328 typedef enum ompt_work_t {
330 ompt_work_sections = 2,
331 ompt_work_single_executor = 3,
332 ompt_work_single_other = 4,
333 ompt_work_workshare = 5,
334 ompt_work_distribute = 6,
335 ompt_work_taskloop = 7,
337 ompt_work_loop_static = 10,
338 ompt_work_loop_dynamic = 11,
339 ompt_work_loop_guided = 12,
340 ompt_work_loop_other = 13
343 typedef enum ompt_mutex_t {
345 ompt_mutex_test_lock = 2,
346 ompt_mutex_nest_lock = 3,
347 ompt_mutex_test_nest_lock = 4,
348 ompt_mutex_critical = 5,
349 ompt_mutex_atomic = 6,
350 ompt_mutex_ordered = 7
353 typedef enum ompt_native_mon_flag_t {
354 ompt_native_data_motion_explicit = 0x01,
355 ompt_native_data_motion_implicit = 0x02,
356 ompt_native_kernel_invocation = 0x04,
357 ompt_native_kernel_execution = 0x08,
358 ompt_native_driver = 0x10,
359 ompt_native_runtime = 0x20,
360 ompt_native_overhead = 0x40,
361 ompt_native_idleness = 0x80
362 } ompt_native_mon_flag_t;
364 typedef enum ompt_task_flag_t {
365 ompt_task_initial = 0x00000001,
366 ompt_task_implicit = 0x00000002,
367 ompt_task_explicit = 0x00000004,
368 ompt_task_target = 0x00000008,
369 ompt_task_taskwait = 0x00000010,
370 ompt_task_undeferred = 0x08000000,
371 ompt_task_untied = 0x10000000,
372 ompt_task_final = 0x20000000,
373 ompt_task_mergeable = 0x40000000,
374 ompt_task_merged = 0x80000000
377 typedef enum ompt_task_status_t {
378 ompt_task_complete = 1,
380 ompt_task_cancel = 3,
381 ompt_task_detach = 4,
382 ompt_task_early_fulfill = 5,
383 ompt_task_late_fulfill = 6,
384 ompt_task_switch = 7,
385 ompt_taskwait_complete = 8
386 } ompt_task_status_t;
388 typedef enum ompt_target_t {
390 ompt_target_enter_data = 2,
391 ompt_target_exit_data = 3,
392 ompt_target_update = 4,
393 ompt_target_nowait = 9,
394 ompt_target_enter_data_nowait = 10,
395 ompt_target_exit_data_nowait = 11,
396 ompt_target_update_nowait = 12
399 typedef enum ompt_parallel_flag_t {
400 ompt_parallel_invoker_program = 0x00000001,
401 ompt_parallel_invoker_runtime = 0x00000002,
402 ompt_parallel_league = 0x40000000,
403 ompt_parallel_team = 0x80000000
404 } ompt_parallel_flag_t;
406 typedef enum ompt_target_map_flag_t {
407 ompt_target_map_flag_to = 0x01,
408 ompt_target_map_flag_from = 0x02,
409 ompt_target_map_flag_alloc = 0x04,
410 ompt_target_map_flag_release = 0x08,
411 ompt_target_map_flag_delete = 0x10,
412 ompt_target_map_flag_implicit = 0x20
413 } ompt_target_map_flag_t;
415 typedef enum ompt_dependence_type_t {
416 ompt_dependence_type_in = 1,
417 ompt_dependence_type_out = 2,
418 ompt_dependence_type_inout = 3,
419 ompt_dependence_type_mutexinoutset = 4,
420 ompt_dependence_type_source = 5,
421 ompt_dependence_type_sink = 6,
422 ompt_dependence_type_inoutset = 7
423 } ompt_dependence_type_t;
425 typedef enum ompt_severity_t {
430 typedef enum ompt_cancel_flag_t {
431 ompt_cancel_parallel = 0x01,
432 ompt_cancel_sections = 0x02,
433 ompt_cancel_loop = 0x04,
434 ompt_cancel_taskgroup = 0x08,
435 ompt_cancel_activated = 0x10,
436 ompt_cancel_detected = 0x20,
437 ompt_cancel_discarded_task = 0x40
438 } ompt_cancel_flag_t;
440 typedef uint64_t ompt_hwid_t;
442 typedef uint64_t ompt_wait_id_t;
444 typedef enum ompt_frame_flag_t {
445 ompt_frame_runtime = 0x00,
446 ompt_frame_application = 0x01,
447 ompt_frame_cfa = 0x10,
448 ompt_frame_framepointer = 0x20,
449 ompt_frame_stackaddress = 0x30
452 typedef enum ompt_state_t {
453 ompt_state_work_serial = 0x000,
454 ompt_state_work_parallel = 0x001,
455 ompt_state_work_reduction = 0x002,
457 ompt_state_wait_barrier DEPRECATED_51 = 0x010,
458 ompt_state_wait_barrier_implicit_parallel = 0x011,
459 ompt_state_wait_barrier_implicit_workshare = 0x012,
460 ompt_state_wait_barrier_implicit DEPRECATED_51 = 0x013,
461 ompt_state_wait_barrier_explicit = 0x014,
462 ompt_state_wait_barrier_implementation = 0x015,
463 ompt_state_wait_barrier_teams = 0x016,
465 ompt_state_wait_taskwait = 0x020,
466 ompt_state_wait_taskgroup = 0x021,
468 ompt_state_wait_mutex = 0x040,
469 ompt_state_wait_lock = 0x041,
470 ompt_state_wait_critical = 0x042,
471 ompt_state_wait_atomic = 0x043,
472 ompt_state_wait_ordered = 0x044,
474 ompt_state_wait_target = 0x080,
475 ompt_state_wait_target_map = 0x081,
476 ompt_state_wait_target_update = 0x082,
478 ompt_state_idle = 0x100,
479 ompt_state_overhead = 0x101,
480 ompt_state_undefined = 0x102
483 typedef uint64_t (*ompt_get_unique_id_t) (void);
485 typedef uint64_t ompd_size_t;
487 typedef uint64_t ompd_wait_id_t;
489 typedef uint64_t ompd_addr_t;
490 typedef int64_t ompd_word_t;
491 typedef uint64_t ompd_seg_t;
493 typedef uint64_t ompd_device_t;
495 typedef uint64_t ompd_thread_id_t;
497 typedef enum ompd_scope_t {
498 ompd_scope_global = 1,
499 ompd_scope_address_space = 2,
500 ompd_scope_thread = 3,
501 ompd_scope_parallel = 4,
502 ompd_scope_implicit_task = 5,
506 typedef uint64_t ompd_icv_id_t;
508 typedef enum ompd_rc_t {
510 ompd_rc_unavailable = 1,
511 ompd_rc_stale_handle = 2,
512 ompd_rc_bad_input = 3,
514 ompd_rc_unsupported = 5,
515 ompd_rc_needs_state_tracking = 6,
516 ompd_rc_incompatible = 7,
517 ompd_rc_device_read_error = 8,
518 ompd_rc_device_write_error = 9,
520 ompd_rc_incomplete = 11,
521 ompd_rc_callback_error = 12
524 typedef void (*ompt_interface_fn_t) (void);
526 typedef ompt_interface_fn_t (*ompt_function_lookup_t) (
527 const char *interface_function_name
530 typedef union ompt_data_t {
535 typedef struct ompt_frame_t {
536 ompt_data_t exit_frame;
537 ompt_data_t enter_frame;
538 int exit_frame_flags;
539 int enter_frame_flags;
542 typedef void (*ompt_callback_t) (void);
544 typedef void ompt_device_t;
546 typedef void ompt_buffer_t;
548 typedef void (*ompt_callback_buffer_request_t) (
550 ompt_buffer_t **buffer,
554 typedef void (*ompt_callback_buffer_complete_t) (
556 ompt_buffer_t *buffer,
558 ompt_buffer_cursor_t begin,
562 typedef void (*ompt_finalize_t) (
563 ompt_data_t *tool_data
566 typedef int (*ompt_initialize_t) (
567 ompt_function_lookup_t lookup,
568 int initial_device_num,
569 ompt_data_t *tool_data
572 typedef struct ompt_start_tool_result_t {
573 ompt_initialize_t initialize;
574 ompt_finalize_t finalize;
575 ompt_data_t tool_data;
576 } ompt_start_tool_result_t;
578 typedef struct ompt_record_abstract_t {
579 ompt_record_native_t rclass;
581 ompt_device_time_t start_time;
582 ompt_device_time_t end_time;
584 } ompt_record_abstract_t;
586 typedef struct ompt_dependence_t {
587 ompt_data_t variable;
588 ompt_dependence_type_t dependence_type;
591 typedef struct ompt_dispatch_chunk_t {
594 } ompt_dispatch_chunk_t;
596 typedef int (*ompt_enumerate_states_t) (
599 const char **next_state_name
602 typedef int (*ompt_enumerate_mutex_impls_t) (
605 const char **next_impl_name
608 typedef ompt_set_result_t (*ompt_set_callback_t) (
609 ompt_callbacks_t event,
610 ompt_callback_t callback
613 typedef int (*ompt_get_callback_t) (
614 ompt_callbacks_t event,
615 ompt_callback_t *callback
618 typedef ompt_data_t *(*ompt_get_thread_data_t) (void);
620 typedef int (*ompt_get_num_procs_t) (void);
622 typedef int (*ompt_get_num_places_t) (void);
624 typedef int (*ompt_get_place_proc_ids_t) (
630 typedef int (*ompt_get_place_num_t) (void);
632 typedef int (*ompt_get_partition_place_nums_t) (
637 typedef int (*ompt_get_proc_id_t) (void);
639 typedef int (*ompt_get_state_t) (
640 ompt_wait_id_t *wait_id
643 typedef int (*ompt_get_parallel_info_t) (
645 ompt_data_t **parallel_data,
649 typedef int (*ompt_get_task_info_t) (
652 ompt_data_t **task_data,
653 ompt_frame_t **task_frame,
654 ompt_data_t **parallel_data,
658 typedef int (*ompt_get_task_memory_t)(
664 typedef int (*ompt_get_target_info_t) (
665 uint64_t *device_num,
666 ompt_id_t *target_id,
667 ompt_id_t *host_op_id
670 typedef int (*ompt_get_num_devices_t) (void);
672 typedef void (*ompt_finalize_tool_t) (void);
674 typedef int (*ompt_get_device_num_procs_t) (
675 ompt_device_t *device
678 typedef ompt_device_time_t (*ompt_get_device_time_t) (
679 ompt_device_t *device
682 typedef double (*ompt_translate_time_t) (
683 ompt_device_t *device,
684 ompt_device_time_t time
687 typedef ompt_set_result_t (*ompt_set_trace_ompt_t) (
688 ompt_device_t *device,
693 typedef ompt_set_result_t (*ompt_set_trace_native_t) (
694 ompt_device_t *device,
699 typedef int (*ompt_start_trace_t) (
700 ompt_device_t *device,
701 ompt_callback_buffer_request_t request,
702 ompt_callback_buffer_complete_t complete
705 typedef int (*ompt_pause_trace_t) (
706 ompt_device_t *device,
710 typedef int (*ompt_flush_trace_t) (
711 ompt_device_t *device
714 typedef int (*ompt_stop_trace_t) (
715 ompt_device_t *device
718 typedef int (*ompt_advance_buffer_cursor_t) (
719 ompt_device_t *device,
720 ompt_buffer_t *buffer,
722 ompt_buffer_cursor_t current,
723 ompt_buffer_cursor_t *next
726 typedef ompt_record_t (*ompt_get_record_type_t) (
727 ompt_buffer_t *buffer,
728 ompt_buffer_cursor_t current
731 typedef void *(*ompt_get_record_native_t) (
732 ompt_buffer_t *buffer,
733 ompt_buffer_cursor_t current,
734 ompt_id_t *host_op_id
737 typedef ompt_record_abstract_t *
738 (*ompt_get_record_abstract_t) (
742 typedef void (*ompt_callback_thread_begin_t) (
743 ompt_thread_t thread_type,
744 ompt_data_t *thread_data
747 typedef struct ompt_record_thread_begin_t {
748 ompt_thread_t thread_type;
749 } ompt_record_thread_begin_t;
751 typedef void (*ompt_callback_thread_end_t) (
752 ompt_data_t *thread_data
755 typedef void (*ompt_callback_parallel_begin_t) (
756 ompt_data_t *encountering_task_data,
757 const ompt_frame_t *encountering_task_frame,
758 ompt_data_t *parallel_data,
759 unsigned int requested_parallelism,
761 const void *codeptr_ra
764 typedef struct ompt_record_parallel_begin_t {
765 ompt_id_t encountering_task_id;
766 ompt_id_t parallel_id;
767 unsigned int requested_parallelism;
769 const void *codeptr_ra;
770 } ompt_record_parallel_begin_t;
772 typedef void (*ompt_callback_parallel_end_t) (
773 ompt_data_t *parallel_data,
774 ompt_data_t *encountering_task_data,
776 const void *codeptr_ra
779 typedef struct ompt_record_parallel_end_t {
780 ompt_id_t parallel_id;
781 ompt_id_t encountering_task_id;
783 const void *codeptr_ra;
784 } ompt_record_parallel_end_t;
786 typedef void (*ompt_callback_work_t) (
787 ompt_work_t work_type,
788 ompt_scope_endpoint_t endpoint,
789 ompt_data_t *parallel_data,
790 ompt_data_t *task_data,
792 const void *codeptr_ra
795 typedef struct ompt_record_work_t {
796 ompt_work_t work_type;
797 ompt_scope_endpoint_t endpoint;
798 ompt_id_t parallel_id;
801 const void *codeptr_ra;
802 } ompt_record_work_t;
804 typedef void (*ompt_callback_dispatch_t) (
805 ompt_data_t *parallel_data,
806 ompt_data_t *task_data,
807 ompt_dispatch_t kind,
811 typedef struct ompt_record_dispatch_t {
812 ompt_id_t parallel_id;
814 ompt_dispatch_t kind;
815 ompt_data_t instance;
816 } ompt_record_dispatch_t;
818 typedef void (*ompt_callback_task_create_t) (
819 ompt_data_t *encountering_task_data,
820 const ompt_frame_t *encountering_task_frame,
821 ompt_data_t *new_task_data,
824 const void *codeptr_ra
827 typedef struct ompt_record_task_create_t {
828 ompt_id_t encountering_task_id;
829 ompt_id_t new_task_id;
832 const void *codeptr_ra;
833 } ompt_record_task_create_t;
835 typedef void (*ompt_callback_dependences_t) (
836 ompt_data_t *task_data,
837 const ompt_dependence_t *deps,
841 typedef struct ompt_record_dependences_t {
843 ompt_dependence_t dep;
845 } ompt_record_dependences_t;
847 typedef void (*ompt_callback_task_dependence_t) (
848 ompt_data_t *src_task_data,
849 ompt_data_t *sink_task_data
852 typedef struct ompt_record_task_dependence_t {
853 ompt_id_t src_task_id;
854 ompt_id_t sink_task_id;
855 } ompt_record_task_dependence_t;
857 typedef void (*ompt_callback_task_schedule_t) (
858 ompt_data_t *prior_task_data,
859 ompt_task_status_t prior_task_status,
860 ompt_data_t *next_task_data
863 typedef struct ompt_record_task_schedule_t {
864 ompt_id_t prior_task_id;
865 ompt_task_status_t prior_task_status;
866 ompt_id_t next_task_id;
867 } ompt_record_task_schedule_t;
869 typedef void (*ompt_callback_implicit_task_t) (
870 ompt_scope_endpoint_t endpoint,
871 ompt_data_t *parallel_data,
872 ompt_data_t *task_data,
873 unsigned int actual_parallelism,
878 typedef struct ompt_record_implicit_task_t {
879 ompt_scope_endpoint_t endpoint;
880 ompt_id_t parallel_id;
882 unsigned int actual_parallelism;
885 } ompt_record_implicit_task_t;
887 typedef void (*ompt_callback_masked_t) (
888 ompt_scope_endpoint_t endpoint,
889 ompt_data_t *parallel_data,
890 ompt_data_t *task_data,
891 const void *codeptr_ra
894 typedef ompt_callback_masked_t ompt_callback_master_t DEPRECATED_51;
896 typedef struct ompt_record_masked_t {
897 ompt_scope_endpoint_t endpoint;
898 ompt_id_t parallel_id;
900 const void *codeptr_ra;
901 } ompt_record_masked_t;
903 typedef void (*ompt_callback_sync_region_t) (
904 ompt_sync_region_t kind,
905 ompt_scope_endpoint_t endpoint,
906 ompt_data_t *parallel_data,
907 ompt_data_t *task_data,
908 const void *codeptr_ra
911 typedef struct ompt_record_sync_region_t {
912 ompt_sync_region_t kind;
913 ompt_scope_endpoint_t endpoint;
914 ompt_id_t parallel_id;
916 const void *codeptr_ra;
917 } ompt_record_sync_region_t;
919 typedef void (*ompt_callback_mutex_acquire_t) (
923 ompt_wait_id_t wait_id,
924 const void *codeptr_ra
927 typedef struct ompt_record_mutex_acquire_t {
931 ompt_wait_id_t wait_id;
932 const void *codeptr_ra;
933 } ompt_record_mutex_acquire_t;
935 typedef void (*ompt_callback_mutex_t) (
937 ompt_wait_id_t wait_id,
938 const void *codeptr_ra
941 typedef struct ompt_record_mutex_t {
943 ompt_wait_id_t wait_id;
944 const void *codeptr_ra;
945 } ompt_record_mutex_t;
947 typedef void (*ompt_callback_nest_lock_t) (
948 ompt_scope_endpoint_t endpoint,
949 ompt_wait_id_t wait_id,
950 const void *codeptr_ra
953 typedef struct ompt_record_nest_lock_t {
954 ompt_scope_endpoint_t endpoint;
955 ompt_wait_id_t wait_id;
956 const void *codeptr_ra;
957 } ompt_record_nest_lock_t;
959 typedef void (*ompt_callback_flush_t) (
960 ompt_data_t *thread_data,
961 const void *codeptr_ra
964 typedef struct ompt_record_flush_t {
965 const void *codeptr_ra;
966 } ompt_record_flush_t;
968 typedef void (*ompt_callback_cancel_t) (
969 ompt_data_t *task_data,
971 const void *codeptr_ra
974 typedef struct ompt_record_cancel_t {
977 const void *codeptr_ra;
978 } ompt_record_cancel_t;
980 typedef void (*ompt_callback_device_initialize_t) (
983 ompt_device_t *device,
984 ompt_function_lookup_t lookup,
985 const char *documentation
988 typedef void (*ompt_callback_device_finalize_t) (
992 typedef void (*ompt_callback_device_load_t) (
994 const char *filename,
995 int64_t offset_in_file,
1003 typedef void (*ompt_callback_device_unload_t) (
1008 typedef void (*ompt_callback_target_data_op_emi_t) (
1009 ompt_scope_endpoint_t endpoint,
1010 ompt_data_t *target_task_data,
1011 ompt_data_t *target_data,
1012 ompt_id_t *host_op_id,
1013 ompt_target_data_op_t optype,
1017 int dest_device_num,
1019 const void *codeptr_ra
1022 typedef void (*ompt_callback_target_data_op_t) (
1023 ompt_id_t target_id,
1024 ompt_id_t host_op_id,
1025 ompt_target_data_op_t optype,
1029 int dest_device_num,
1031 const void *codeptr_ra
1034 typedef struct ompt_record_target_data_op_t {
1035 ompt_id_t host_op_id;
1036 ompt_target_data_op_t optype;
1040 int dest_device_num;
1042 ompt_device_time_t end_time;
1043 const void *codeptr_ra;
1044 } ompt_record_target_data_op_t;
1046 typedef void (*ompt_callback_target_emi_t) (
1048 ompt_scope_endpoint_t endpoint,
1050 ompt_data_t *task_data,
1051 ompt_data_t *target_task_data,
1052 ompt_data_t *target_data,
1053 const void *codeptr_ra
1056 typedef void (*ompt_callback_target_t) (
1058 ompt_scope_endpoint_t endpoint,
1060 ompt_data_t *task_data,
1061 ompt_id_t target_id,
1062 const void *codeptr_ra
1065 typedef struct ompt_record_target_t {
1067 ompt_scope_endpoint_t endpoint;
1070 ompt_id_t target_id;
1071 const void *codeptr_ra;
1072 } ompt_record_target_t;
1074 typedef void (*ompt_callback_target_map_emi_t) (
1075 ompt_data_t *target_data,
1076 unsigned int nitems,
1080 unsigned int *mapping_flags,
1081 const void *codeptr_ra
1084 typedef void (*ompt_callback_target_map_t) (
1085 ompt_id_t target_id,
1086 unsigned int nitems,
1090 unsigned int *mapping_flags,
1091 const void *codeptr_ra
1094 typedef struct ompt_record_target_map_t {
1095 ompt_id_t target_id;
1096 unsigned int nitems;
1100 unsigned int *mapping_flags;
1101 const void *codeptr_ra;
1102 } ompt_record_target_map_t;
1104 typedef void (*ompt_callback_target_submit_emi_t) (
1105 ompt_scope_endpoint_t endpoint,
1106 ompt_data_t *target_data,
1107 ompt_id_t *host_op_id,
1108 unsigned int requested_num_teams
1111 typedef void (*ompt_callback_target_submit_t) (
1112 ompt_id_t target_id,
1113 ompt_id_t host_op_id,
1114 unsigned int requested_num_teams
1117 typedef struct ompt_record_target_kernel_t {
1118 ompt_id_t host_op_id;
1119 unsigned int requested_num_teams;
1120 unsigned int granted_num_teams;
1121 ompt_device_time_t end_time;
1122 } ompt_record_target_kernel_t;
1124 typedef int (*ompt_callback_control_tool_t) (
1128 const void *codeptr_ra
1131 typedef struct ompt_record_control_tool_t {
1134 const void *codeptr_ra;
1135 } ompt_record_control_tool_t;
1137 typedef void (*ompt_callback_error_t) (
1138 ompt_severity_t severity,
1139 const char *message, size_t length,
1140 const void *codeptr_ra
1143 typedef struct ompt_record_error_t {
1144 ompt_severity_t severity;
1145 const char *message;
1147 const void *codeptr_ra;
1148 } ompt_record_error_t;
1150 typedef struct ompd_address_t {
1152 ompd_addr_t address;
1155 typedef struct ompd_frame_info_t {
1156 ompd_address_t frame_address;
1157 ompd_word_t frame_flag;
1158 } ompd_frame_info_t;
1160 typedef struct _ompd_aspace_handle ompd_address_space_handle_t;
1161 typedef struct _ompd_thread_handle ompd_thread_handle_t;
1162 typedef struct _ompd_parallel_handle ompd_parallel_handle_t;
1163 typedef struct _ompd_task_handle ompd_task_handle_t;
1165 typedef struct _ompd_aspace_cont ompd_address_space_context_t;
1166 typedef struct _ompd_thread_cont ompd_thread_context_t;
1168 typedef struct ompd_device_type_sizes_t {
1169 uint8_t sizeof_char;
1170 uint8_t sizeof_short;
1172 uint8_t sizeof_long;
1173 uint8_t sizeof_long_long;
1174 uint8_t sizeof_pointer;
1175 } ompd_device_type_sizes_t;
1177 void ompd_dll_locations_valid(void);
1179 typedef ompd_rc_t (*ompd_callback_memory_alloc_fn_t)(ompd_size_t nbytes,
1182 typedef ompd_rc_t (*ompd_callback_memory_free_fn_t)(void *ptr);
1184 typedef ompd_rc_t (*ompd_callback_get_thread_context_for_thread_id_fn_t)(
1185 ompd_address_space_context_t *address_space_context, ompd_thread_id_t kind,
1186 ompd_size_t sizeof_thread_id, const void *thread_id,
1187 ompd_thread_context_t **thread_context);
1189 typedef ompd_rc_t (*ompd_callback_sizeof_fn_t)(
1190 ompd_address_space_context_t *address_space_context,
1191 ompd_device_type_sizes_t *sizes);
1193 typedef ompd_rc_t (*ompd_callback_symbol_addr_fn_t)(
1194 ompd_address_space_context_t *address_space_context,
1195 ompd_thread_context_t *thread_context, const char *symbol_name,
1196 ompd_address_t *symbol_addr, const char *file_name);
1198 typedef ompd_rc_t (*ompd_callback_memory_read_fn_t)(
1199 ompd_address_space_context_t *address_space_context,
1200 ompd_thread_context_t *thread_context, const ompd_address_t *addr,
1201 ompd_size_t nbytes, void *buffer);
1203 typedef ompd_rc_t (*ompd_callback_memory_write_fn_t)(
1204 ompd_address_space_context_t *address_space_context,
1205 ompd_thread_context_t *thread_context, const ompd_address_t *addr,
1206 ompd_size_t nbytes, const void *buffer);
1208 typedef ompd_rc_t (*ompd_callback_device_host_fn_t)(
1209 ompd_address_space_context_t *address_space_context, const void *input,
1210 ompd_size_t unit_size, ompd_size_t count, void *output);
1212 typedef ompd_rc_t (*ompd_callback_print_string_fn_t)(const char *string,
1215 typedef struct ompd_callbacks_t {
1216 ompd_callback_memory_alloc_fn_t alloc_memory;
1217 ompd_callback_memory_free_fn_t free_memory;
1218 ompd_callback_print_string_fn_t print_string;
1219 ompd_callback_sizeof_fn_t sizeof_type;
1220 ompd_callback_symbol_addr_fn_t symbol_addr_lookup;
1221 ompd_callback_memory_read_fn_t read_memory;
1222 ompd_callback_memory_write_fn_t write_memory;
1223 ompd_callback_memory_read_fn_t read_string;
1224 ompd_callback_device_host_fn_t device_to_host;
1225 ompd_callback_device_host_fn_t host_to_device;
1226 ompd_callback_get_thread_context_for_thread_id_fn_t
1227 get_thread_context_for_thread_id;
1230 void ompd_bp_parallel_begin(void);
1232 void ompd_bp_parallel_end(void);
1234 void ompd_bp_task_begin(void);
1236 void ompd_bp_task_end(void);
1238 void ompd_bp_thread_begin(void);
1240 void ompd_bp_thread_end(void);
1242 void ompd_bp_device_begin(void);
1244 void ompd_bp_device_end(void);
1246 ompd_rc_t ompd_initialize(ompd_word_t api_version,
1247 const ompd_callbacks_t *callbacks);
1249 ompd_rc_t ompd_get_api_version(ompd_word_t *version);
1251 ompd_rc_t ompd_get_version_string(const char **string);
1253 ompd_rc_t ompd_finalize(void);
1255 ompd_rc_t ompd_process_initialize(ompd_address_space_context_t *context,
1256 ompd_address_space_handle_t **handle);
1258 ompd_rc_t ompd_device_initialize(ompd_address_space_handle_t *process_handle,
1259 ompd_address_space_context_t *device_context,
1260 ompd_device_t kind, ompd_size_t sizeof_id,
1262 ompd_address_space_handle_t **device_handle);
1264 ompd_rc_t ompd_rel_address_space_handle(ompd_address_space_handle_t *handle);
1266 ompd_rc_t ompd_get_omp_version(ompd_address_space_handle_t *address_space,
1267 ompd_word_t *omp_version);
1270 ompd_get_omp_version_string(ompd_address_space_handle_t *address_space,
1271 const char **string);
1273 ompd_rc_t ompd_get_thread_in_parallel(ompd_parallel_handle_t *parallel_handle,
1275 ompd_thread_handle_t **thread_handle);
1277 ompd_rc_t ompd_get_thread_handle(ompd_address_space_handle_t *handle,
1278 ompd_thread_id_t kind,
1279 ompd_size_t sizeof_thread_id,
1280 const void *thread_id,
1281 ompd_thread_handle_t **thread_handle);
1283 ompd_rc_t ompd_rel_thread_handle(ompd_thread_handle_t *thread_handle);
1285 ompd_rc_t ompd_thread_handle_compare(ompd_thread_handle_t *thread_handle_1,
1286 ompd_thread_handle_t *thread_handle_2,
1289 ompd_rc_t ompd_get_thread_id(ompd_thread_handle_t *thread_handle,
1290 ompd_thread_id_t kind,
1291 ompd_size_t sizeof_thread_id, void *thread_id);
1294 ompd_get_curr_parallel_handle(ompd_thread_handle_t *thread_handle,
1295 ompd_parallel_handle_t **parallel_handle);
1297 ompd_rc_t ompd_get_enclosing_parallel_handle(
1298 ompd_parallel_handle_t *parallel_handle,
1299 ompd_parallel_handle_t **enclosing_parallel_handle);
1302 ompd_get_task_parallel_handle(ompd_task_handle_t *task_handle,
1303 ompd_parallel_handle_t **task_parallel_handle);
1305 ompd_rc_t ompd_rel_parallel_handle(ompd_parallel_handle_t *parallel_handle);
1308 ompd_parallel_handle_compare(ompd_parallel_handle_t *parallel_handle_1,
1309 ompd_parallel_handle_t *parallel_handle_2,
1312 ompd_rc_t ompd_get_curr_task_handle(ompd_thread_handle_t *thread_handle,
1313 ompd_task_handle_t **task_handle);
1316 ompd_get_generating_task_handle(ompd_task_handle_t *task_handle,
1317 ompd_task_handle_t **generating_task_handle);
1320 ompd_get_scheduling_task_handle(ompd_task_handle_t *task_handle,
1321 ompd_task_handle_t **scheduling_task_handle);
1323 ompd_rc_t ompd_get_task_in_parallel(ompd_parallel_handle_t *parallel_handle,
1325 ompd_task_handle_t **task_handle);
1327 ompd_rc_t ompd_rel_task_handle(ompd_task_handle_t *task_handle);
1329 ompd_rc_t ompd_task_handle_compare(ompd_task_handle_t *task_handle_1,
1330 ompd_task_handle_t *task_handle_2,
1333 ompd_rc_t ompd_get_task_function(ompd_task_handle_t *task_handle,
1334 ompd_address_t *entry_point);
1336 ompd_rc_t ompd_get_task_frame(ompd_task_handle_t *task_handle,
1337 ompd_frame_info_t *exit_frame,
1338 ompd_frame_info_t *enter_frame);
1341 ompd_enumerate_states(ompd_address_space_handle_t *address_space_handle,
1342 ompd_word_t current_state, ompd_word_t *next_state,
1343 const char **next_state_name, ompd_word_t *more_enums);
1345 ompd_rc_t ompd_get_state(ompd_thread_handle_t *thread_handle,
1346 ompd_word_t *state, ompd_wait_id_t *wait_id);
1349 ompd_get_display_control_vars(ompd_address_space_handle_t *address_space_handle,
1350 const char *const **control_vars);
1352 ompd_rc_t ompd_rel_display_control_vars(const char *const **control_vars);
1354 ompd_rc_t ompd_enumerate_icvs(ompd_address_space_handle_t *handle,
1355 ompd_icv_id_t current, ompd_icv_id_t *next_id,
1356 const char **next_icv_name,
1357 ompd_scope_t *next_scope, int *more);
1359 ompd_rc_t ompd_get_icv_from_scope(void *handle, ompd_scope_t scope,
1360 ompd_icv_id_t icv_id, ompd_word_t *icv_value);
1362 ompd_rc_t ompd_get_icv_string_from_scope(void *handle, ompd_scope_t scope,
1363 ompd_icv_id_t icv_id,
1364 const char **icv_string);
1366 ompd_rc_t ompd_get_tool_data(void *handle, ompd_scope_t scope,
1367 ompd_word_t *value, ompd_address_t *ptr);
1369 typedef struct ompt_record_ompt_t {
1370 ompt_callbacks_t type;
1371 ompt_device_time_t time;
1372 ompt_id_t thread_id;
1373 ompt_id_t target_id;
1375 ompt_record_thread_begin_t thread_begin;
1376 ompt_record_parallel_begin_t parallel_begin;
1377 ompt_record_parallel_end_t parallel_end;
1378 ompt_record_work_t work;
1379 ompt_record_dispatch_t dispatch;
1380 ompt_record_task_create_t task_create;
1381 ompt_record_dependences_t dependences;
1382 ompt_record_task_dependence_t task_dependence;
1383 ompt_record_task_schedule_t task_schedule;
1384 ompt_record_implicit_task_t implicit_task;
1385 ompt_record_masked_t masked;
1386 ompt_record_sync_region_t sync_region;
1387 ompt_record_mutex_acquire_t mutex_acquire;
1388 ompt_record_mutex_t mutex;
1389 ompt_record_nest_lock_t nest_lock;
1390 ompt_record_flush_t flush;
1391 ompt_record_cancel_t cancel;
1392 ompt_record_target_t target;
1393 ompt_record_target_data_op_t target_data_op;
1394 ompt_record_target_map_t target_map;
1395 ompt_record_target_kernel_t target_kernel;
1396 ompt_record_control_tool_t control_tool;
1398 } ompt_record_ompt_t;
1400 typedef ompt_record_ompt_t *(*ompt_get_record_ompt_t) (
1401 ompt_buffer_t *buffer,
1402 ompt_buffer_cursor_t current
1405 #define ompt_id_none 0
1406 #define ompt_data_none {0}
1407 #define ompt_time_none 0
1408 #define ompt_hwid_none 0
1409 #define ompt_addr_none ~0
1410 #define ompt_mutex_impl_none 0
1411 #define ompt_wait_id_none 0
1413 #define ompd_segment_none 0
1415 #endif /* __OMPT__ */