2 * ompt-general.cpp -- OMPT implementation of interface functions
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 //===----------------------------------------------------------------------===//
13 /*****************************************************************************
14 * system include files
15 ****************************************************************************/
27 /*****************************************************************************
29 ****************************************************************************/
31 #include "ompt-specific.cpp"
33 /*****************************************************************************
35 ****************************************************************************/
37 #define ompt_get_callback_success 1
38 #define ompt_get_callback_failure 0
40 #define no_tool_present 0
42 #define OMPT_API_ROUTINE static
44 #ifndef OMPT_STR_MATCH
45 #define OMPT_STR_MATCH(haystack, needle) (!strcasecmp(haystack, needle))
48 /*****************************************************************************
50 ****************************************************************************/
53 const char *state_name;
54 ompt_state_t state_id;
60 } kmp_mutex_impl_info_t;
69 /*****************************************************************************
71 ****************************************************************************/
73 ompt_callbacks_active_t ompt_enabled;
75 ompt_state_info_t ompt_state_info[] = {
76 #define ompt_state_macro(state, code) {#state, state},
77 FOREACH_OMPT_STATE(ompt_state_macro)
78 #undef ompt_state_macro
81 kmp_mutex_impl_info_t kmp_mutex_impl_info[] = {
82 #define kmp_mutex_impl_macro(name, id) {#name, name},
83 FOREACH_KMP_MUTEX_IMPL(kmp_mutex_impl_macro)
84 #undef kmp_mutex_impl_macro
87 ompt_callbacks_internal_t ompt_callbacks;
89 static ompt_start_tool_result_t *ompt_start_tool_result = NULL;
91 /*****************************************************************************
92 * forward declarations
93 ****************************************************************************/
95 static ompt_interface_fn_t ompt_fn_lookup(const char *s);
97 OMPT_API_ROUTINE ompt_data_t *ompt_get_thread_data(void);
99 /*****************************************************************************
100 * initialization and finalization (private operations)
101 ****************************************************************************/
103 typedef ompt_start_tool_result_t *(*ompt_start_tool_t)(unsigned int,
108 // While Darwin supports weak symbols, the library that wishes to provide a new
109 // implementation has to link against this runtime which defeats the purpose
110 // of having tools that are agnostic of the underlying runtime implementation.
112 // Fortunately, the linker includes all symbols of an executable in the global
113 // symbol table by default so dlsym() even finds static implementations of
114 // ompt_start_tool. For this to work on Linux, -Wl,--export-dynamic needs to be
115 // passed when building the application which we don't want to rely on.
117 static ompt_start_tool_result_t *ompt_tool_darwin(unsigned int omp_version,
118 const char *runtime_version) {
119 ompt_start_tool_result_t *ret = NULL;
120 // Search symbol in the current address space.
121 ompt_start_tool_t start_tool =
122 (ompt_start_tool_t)dlsym(RTLD_DEFAULT, "ompt_start_tool");
124 ret = start_tool(omp_version, runtime_version);
129 #elif OMPT_HAVE_WEAK_ATTRIBUTE
131 // On Unix-like systems that support weak symbols the following implementation
132 // of ompt_start_tool() will be used in case no tool-supplied implementation of
133 // this function is present in the address space of a process.
135 _OMP_EXTERN OMPT_WEAK_ATTRIBUTE ompt_start_tool_result_t *
136 ompt_start_tool(unsigned int omp_version, const char *runtime_version) {
137 ompt_start_tool_result_t *ret = NULL;
138 // Search next symbol in the current address space. This can happen if the
139 // runtime library is linked before the tool. Since glibc 2.2 strong symbols
140 // don't override weak symbols that have been found before unless the user
141 // sets the environment variable LD_DYNAMIC_WEAK.
142 ompt_start_tool_t next_tool =
143 (ompt_start_tool_t)dlsym(RTLD_NEXT, "ompt_start_tool");
145 ret = next_tool(omp_version, runtime_version);
150 #elif OMPT_HAVE_PSAPI
152 // On Windows, the ompt_tool_windows function is used to find the
153 // ompt_start_tool symbol across all modules loaded by a process. If
154 // ompt_start_tool is found, ompt_start_tool's return value is used to
155 // initialize the tool. Otherwise, NULL is returned and OMPT won't be enabled.
158 #pragma comment(lib, "psapi.lib")
160 // The number of loaded modules to start enumeration with EnumProcessModules()
161 #define NUM_MODULES 128
163 static ompt_start_tool_result_t *
164 ompt_tool_windows(unsigned int omp_version, const char *runtime_version) {
166 DWORD needed, new_size;
168 HANDLE process = GetCurrentProcess();
169 modules = (HMODULE *)malloc(NUM_MODULES * sizeof(HMODULE));
170 ompt_start_tool_t ompt_tool_p = NULL;
173 printf("ompt_tool_windows(): looking for ompt_start_tool\n");
175 if (!EnumProcessModules(process, modules, NUM_MODULES * sizeof(HMODULE),
177 // Regardless of the error reason use the stub initialization function
181 // Check if NUM_MODULES is enough to list all modules
182 new_size = needed / sizeof(HMODULE);
183 if (new_size > NUM_MODULES) {
185 printf("ompt_tool_windows(): resize buffer to %d bytes\n", needed);
187 modules = (HMODULE *)realloc(modules, needed);
188 // If resizing failed use the stub function.
189 if (!EnumProcessModules(process, modules, needed, &needed)) {
194 for (i = 0; i < new_size; ++i) {
195 (FARPROC &)ompt_tool_p = GetProcAddress(modules[i], "ompt_start_tool");
198 TCHAR modName[MAX_PATH];
199 if (GetModuleFileName(modules[i], modName, MAX_PATH))
200 printf("ompt_tool_windows(): ompt_start_tool found in module %s\n",
204 return (*ompt_tool_p)(omp_version, runtime_version);
208 TCHAR modName[MAX_PATH];
209 if (GetModuleFileName(modules[i], modName, MAX_PATH))
210 printf("ompt_tool_windows(): ompt_start_tool not found in module %s\n",
219 #error Activation of OMPT is not supported on this platform.
222 static ompt_start_tool_result_t *
223 ompt_try_start_tool(unsigned int omp_version, const char *runtime_version) {
224 ompt_start_tool_result_t *ret = NULL;
225 ompt_start_tool_t start_tool = NULL;
227 // Cannot use colon to describe a list of absolute paths on Windows
228 const char *sep = ";";
230 const char *sep = ":";
234 // Try in the current address space
235 ret = ompt_tool_darwin(omp_version, runtime_version);
236 #elif OMPT_HAVE_WEAK_ATTRIBUTE
237 ret = ompt_start_tool(omp_version, runtime_version);
238 #elif OMPT_HAVE_PSAPI
239 ret = ompt_tool_windows(omp_version, runtime_version);
241 #error Activation of OMPT is not supported on this platform.
246 // Try tool-libraries-var ICV
247 const char *tool_libs = getenv("OMP_TOOL_LIBRARIES");
249 char *libs = __kmp_str_format("%s", tool_libs);
251 char *fname = __kmp_str_token(libs, sep, &buf);
254 void *h = dlopen(fname, RTLD_LAZY);
256 start_tool = (ompt_start_tool_t)dlsym(h, "ompt_start_tool");
258 HMODULE h = LoadLibrary(fname);
260 start_tool = (ompt_start_tool_t)GetProcAddress(h, "ompt_start_tool");
262 #error Activation of OMPT is not supported on this platform.
264 if (start_tool && (ret = (*start_tool)(omp_version, runtime_version)))
267 fname = __kmp_str_token(NULL, sep, &buf);
269 __kmp_str_free(&libs);
274 void ompt_pre_init() {
275 //--------------------------------------------------
276 // Execute the pre-initialization logic only once.
277 //--------------------------------------------------
278 static int ompt_pre_initialized = 0;
280 if (ompt_pre_initialized)
283 ompt_pre_initialized = 1;
285 //--------------------------------------------------
286 // Use a tool iff a tool is enabled and available.
287 //--------------------------------------------------
288 const char *ompt_env_var = getenv("OMP_TOOL");
289 tool_setting_e tool_setting = omp_tool_error;
291 if (!ompt_env_var || !strcmp(ompt_env_var, ""))
292 tool_setting = omp_tool_unset;
293 else if (OMPT_STR_MATCH(ompt_env_var, "disabled"))
294 tool_setting = omp_tool_disabled;
295 else if (OMPT_STR_MATCH(ompt_env_var, "enabled"))
296 tool_setting = omp_tool_enabled;
299 printf("ompt_pre_init(): tool_setting = %d\n", tool_setting);
301 switch (tool_setting) {
302 case omp_tool_disabled:
306 case omp_tool_enabled:
308 //--------------------------------------------------
309 // Load tool iff specified in environment variable
310 //--------------------------------------------------
311 ompt_start_tool_result =
312 ompt_try_start_tool(__kmp_openmp_version, ompt_get_runtime_version());
314 memset(&ompt_enabled, 0, sizeof(ompt_enabled));
318 fprintf(stderr, "Warning: OMP_TOOL has invalid value \"%s\".\n"
319 " legal values are (NULL,\"\",\"disabled\","
325 printf("ompt_pre_init(): ompt_enabled = %d\n", ompt_enabled);
329 extern "C" int omp_get_initial_device(void);
331 void ompt_post_init() {
332 //--------------------------------------------------
333 // Execute the post-initialization logic only once.
334 //--------------------------------------------------
335 static int ompt_post_initialized = 0;
337 if (ompt_post_initialized)
340 ompt_post_initialized = 1;
342 //--------------------------------------------------
343 // Initialize the tool if so indicated.
344 //--------------------------------------------------
345 if (ompt_start_tool_result) {
346 ompt_enabled.enabled = !!ompt_start_tool_result->initialize(
347 ompt_fn_lookup, omp_get_initial_device(), &(ompt_start_tool_result->tool_data));
349 if (!ompt_enabled.enabled) {
350 // tool not enabled, zero out the bitmap, and done
351 memset(&ompt_enabled, 0, sizeof(ompt_enabled));
355 kmp_info_t *root_thread = ompt_get_thread();
357 ompt_set_thread_state(root_thread, ompt_state_overhead);
359 if (ompt_enabled.ompt_callback_thread_begin) {
360 ompt_callbacks.ompt_callback(ompt_callback_thread_begin)(
361 ompt_thread_initial, __ompt_get_thread_data_internal());
363 ompt_data_t *task_data;
364 ompt_data_t *parallel_data;
365 __ompt_get_task_info_internal(0, NULL, &task_data, NULL, ¶llel_data, NULL);
366 if (ompt_enabled.ompt_callback_implicit_task) {
367 ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
368 ompt_scope_begin, parallel_data, task_data, 1, 1, ompt_task_initial);
371 ompt_set_thread_state(root_thread, ompt_state_work_serial);
376 if (ompt_enabled.enabled) {
377 ompt_start_tool_result->finalize(&(ompt_start_tool_result->tool_data));
380 memset(&ompt_enabled, 0, sizeof(ompt_enabled));
383 /*****************************************************************************
384 * interface operations
385 ****************************************************************************/
387 /*****************************************************************************
389 ****************************************************************************/
391 OMPT_API_ROUTINE int ompt_enumerate_states(int current_state, int *next_state,
392 const char **next_state_name) {
393 const static int len = sizeof(ompt_state_info) / sizeof(ompt_state_info_t);
396 for (i = 0; i < len - 1; i++) {
397 if (ompt_state_info[i].state_id == current_state) {
398 *next_state = ompt_state_info[i + 1].state_id;
399 *next_state_name = ompt_state_info[i + 1].state_name;
407 OMPT_API_ROUTINE int ompt_enumerate_mutex_impls(int current_impl,
409 const char **next_impl_name) {
410 const static int len =
411 sizeof(kmp_mutex_impl_info) / sizeof(kmp_mutex_impl_info_t);
413 for (i = 0; i < len - 1; i++) {
414 if (kmp_mutex_impl_info[i].id != current_impl)
416 *next_impl = kmp_mutex_impl_info[i + 1].id;
417 *next_impl_name = kmp_mutex_impl_info[i + 1].name;
423 /*****************************************************************************
425 ****************************************************************************/
427 OMPT_API_ROUTINE ompt_set_result_t ompt_set_callback(ompt_callbacks_t which,
428 ompt_callback_t callback) {
431 #define ompt_event_macro(event_name, callback_type, event_id) \
433 if (ompt_event_implementation_status(event_name)) { \
434 ompt_callbacks.ompt_callback(event_name) = (callback_type)callback; \
435 ompt_enabled.event_name = (callback != 0); \
438 return ompt_event_implementation_status(event_name); \
440 return ompt_set_always;
442 FOREACH_OMPT_EVENT(ompt_event_macro)
444 #undef ompt_event_macro
447 return ompt_set_error;
451 OMPT_API_ROUTINE int ompt_get_callback(ompt_callbacks_t which,
452 ompt_callback_t *callback) {
453 if (!ompt_enabled.enabled)
454 return ompt_get_callback_failure;
458 #define ompt_event_macro(event_name, callback_type, event_id) \
460 if (ompt_event_implementation_status(event_name)) { \
461 ompt_callback_t mycb = \
462 (ompt_callback_t)ompt_callbacks.ompt_callback(event_name); \
463 if (ompt_enabled.event_name && mycb) { \
465 return ompt_get_callback_success; \
468 return ompt_get_callback_failure;
470 FOREACH_OMPT_EVENT(ompt_event_macro)
472 #undef ompt_event_macro
475 return ompt_get_callback_failure;
479 /*****************************************************************************
481 ****************************************************************************/
483 OMPT_API_ROUTINE int ompt_get_parallel_info(int ancestor_level,
484 ompt_data_t **parallel_data,
486 if (!ompt_enabled.enabled)
488 return __ompt_get_parallel_info_internal(ancestor_level, parallel_data,
492 OMPT_API_ROUTINE int ompt_get_state(ompt_wait_id_t *wait_id) {
493 if (!ompt_enabled.enabled)
494 return ompt_state_work_serial;
495 int thread_state = __ompt_get_state_internal(wait_id);
497 if (thread_state == ompt_state_undefined) {
498 thread_state = ompt_state_work_serial;
504 /*****************************************************************************
506 ****************************************************************************/
508 OMPT_API_ROUTINE ompt_data_t *ompt_get_thread_data(void) {
509 if (!ompt_enabled.enabled)
511 return __ompt_get_thread_data_internal();
514 OMPT_API_ROUTINE int ompt_get_task_info(int ancestor_level, int *type,
515 ompt_data_t **task_data,
516 ompt_frame_t **task_frame,
517 ompt_data_t **parallel_data,
519 if (!ompt_enabled.enabled)
521 return __ompt_get_task_info_internal(ancestor_level, type, task_data,
522 task_frame, parallel_data, thread_num);
525 OMPT_API_ROUTINE int ompt_get_task_memory(void **addr, size_t *size,
527 return __ompt_get_task_memory_internal(addr, size, block);
530 /*****************************************************************************
532 ****************************************************************************/
534 OMPT_API_ROUTINE int ompt_get_num_procs(void) {
535 // copied from kmp_ftn_entry.h (but modified: OMPT can only be called when
536 // runtime is initialized)
537 return __kmp_avail_proc;
540 /*****************************************************************************
542 ****************************************************************************/
544 OMPT_API_ROUTINE int ompt_get_num_places(void) {
545 // copied from kmp_ftn_entry.h (but modified)
546 #if !KMP_AFFINITY_SUPPORTED
549 if (!KMP_AFFINITY_CAPABLE())
551 return __kmp_affinity_num_masks;
555 OMPT_API_ROUTINE int ompt_get_place_proc_ids(int place_num, int ids_size,
557 // copied from kmp_ftn_entry.h (but modified)
558 #if !KMP_AFFINITY_SUPPORTED
562 int tmp_ids[ids_size];
563 if (!KMP_AFFINITY_CAPABLE())
565 if (place_num < 0 || place_num >= (int)__kmp_affinity_num_masks)
567 /* TODO: Is this safe for asynchronous call from signal handler during runtime
569 kmp_affin_mask_t *mask = KMP_CPU_INDEX(__kmp_affinity_masks, place_num);
571 KMP_CPU_SET_ITERATE(i, mask) {
572 if ((!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) ||
573 (!KMP_CPU_ISSET(i, mask))) {
576 if (count < ids_size)
580 if (ids_size >= count) {
581 for (i = 0; i < count; i++) {
589 OMPT_API_ROUTINE int ompt_get_place_num(void) {
590 // copied from kmp_ftn_entry.h (but modified)
591 #if !KMP_AFFINITY_SUPPORTED
594 if (!ompt_enabled.enabled || __kmp_get_gtid() < 0)
599 if (!KMP_AFFINITY_CAPABLE())
601 gtid = __kmp_entry_gtid();
602 thread = __kmp_thread_from_gtid(gtid);
603 if (thread == NULL || thread->th.th_current_place < 0)
605 return thread->th.th_current_place;
609 OMPT_API_ROUTINE int ompt_get_partition_place_nums(int place_nums_size,
611 // copied from kmp_ftn_entry.h (but modified)
612 #if !KMP_AFFINITY_SUPPORTED
615 if (!ompt_enabled.enabled || __kmp_get_gtid() < 0)
618 int i, gtid, place_num, first_place, last_place, start, end;
620 if (!KMP_AFFINITY_CAPABLE())
622 gtid = __kmp_entry_gtid();
623 thread = __kmp_thread_from_gtid(gtid);
626 first_place = thread->th.th_first_place;
627 last_place = thread->th.th_last_place;
628 if (first_place < 0 || last_place < 0)
630 if (first_place <= last_place) {
637 if (end - start <= place_nums_size)
638 for (i = 0, place_num = start; place_num <= end; ++place_num, ++i) {
639 place_nums[i] = place_num;
641 return end - start + 1;
645 /*****************************************************************************
647 ****************************************************************************/
649 OMPT_API_ROUTINE int ompt_get_proc_id(void) {
650 if (!ompt_enabled.enabled || __kmp_get_gtid() < 0)
653 return sched_getcpu();
656 GetCurrentProcessorNumberEx(&pn);
657 return 64 * pn.Group + pn.Number;
663 /*****************************************************************************
665 ****************************************************************************/
668 * Currently unused function
669 OMPT_API_ROUTINE int ompt_get_ompt_version() { return OMPT_VERSION; }
672 /*****************************************************************************
673 * application-facing API
674 ****************************************************************************/
676 /*----------------------------------------------------------------------------
678 ---------------------------------------------------------------------------*/
680 int __kmp_control_tool(uint64_t command, uint64_t modifier, void *arg) {
682 if (ompt_enabled.enabled) {
683 if (ompt_enabled.ompt_callback_control_tool) {
684 return ompt_callbacks.ompt_callback(ompt_callback_control_tool)(
685 command, modifier, arg, OMPT_LOAD_RETURN_ADDRESS(__kmp_entry_gtid()));
694 /*****************************************************************************
696 ****************************************************************************/
698 OMPT_API_ROUTINE uint64_t ompt_get_unique_id(void) {
699 return __ompt_get_unique_id_internal();
702 OMPT_API_ROUTINE void ompt_finalize_tool(void) { __kmp_internal_end_atexit(); }
704 /*****************************************************************************
706 ****************************************************************************/
708 OMPT_API_ROUTINE int ompt_get_target_info(uint64_t *device_num,
709 ompt_id_t *target_id,
710 ompt_id_t *host_op_id) {
711 return 0; // thread is not in a target region
714 OMPT_API_ROUTINE int ompt_get_num_devices(void) {
715 return 1; // only one device (the current device) is available
718 /*****************************************************************************
719 * API inquiry for tool
720 ****************************************************************************/
722 static ompt_interface_fn_t ompt_fn_lookup(const char *s) {
724 #define ompt_interface_fn(fn) \
725 fn##_t fn##_f = fn; \
726 if (strcmp(s, #fn) == 0) \
727 return (ompt_interface_fn_t)fn##_f;
729 FOREACH_OMPT_INQUIRY_FN(ompt_interface_fn)
731 return (ompt_interface_fn_t)0;