1 | /* |
2 | * ompt-general.cpp -- OMPT implementation of interface functions |
3 | */ |
4 | |
5 | //===----------------------------------------------------------------------===// |
6 | // |
7 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
8 | // See https://llvm.org/LICENSE.txt for license information. |
9 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | /***************************************************************************** |
14 | * system include files |
15 | ****************************************************************************/ |
16 | |
17 | #include <assert.h> |
18 | |
19 | #include <stdint.h> |
20 | #include <stdio.h> |
21 | #include <stdlib.h> |
22 | #include <string.h> |
23 | #if KMP_OS_UNIX |
24 | #include <dlfcn.h> |
25 | #endif |
26 | |
27 | /***************************************************************************** |
28 | * ompt include files |
29 | ****************************************************************************/ |
30 | |
31 | #include "ompt-specific.cpp" |
32 | |
33 | /***************************************************************************** |
34 | * macros |
35 | ****************************************************************************/ |
36 | |
37 | #define ompt_get_callback_success 1 |
38 | #define ompt_get_callback_failure 0 |
39 | |
40 | #define no_tool_present 0 |
41 | |
42 | #define OMPT_API_ROUTINE static |
43 | |
44 | #ifndef OMPT_STR_MATCH |
45 | #define OMPT_STR_MATCH(haystack, needle) (!strcasecmp(haystack, needle)) |
46 | #endif |
47 | |
48 | // prints for an enabled OMP_TOOL_VERBOSE_INIT. |
49 | // In the future a prefix could be added in the first define, the second define |
50 | // omits the prefix to allow for continued lines. Example: "PREFIX: Start |
51 | // tool... Success." instead of "PREFIX: Start tool... PREFIX: Success." |
52 | #define OMPT_VERBOSE_INIT_PRINT(...) \ |
53 | if (verbose_init) \ |
54 | fprintf(verbose_file, __VA_ARGS__) |
55 | #define OMPT_VERBOSE_INIT_CONTINUED_PRINT(...) \ |
56 | if (verbose_init) \ |
57 | fprintf(verbose_file, __VA_ARGS__) |
58 | |
59 | static FILE *verbose_file; |
60 | static int verbose_init; |
61 | |
62 | /***************************************************************************** |
63 | * types |
64 | ****************************************************************************/ |
65 | |
66 | typedef struct { |
67 | const char *state_name; |
68 | ompt_state_t state_id; |
69 | } ompt_state_info_t; |
70 | |
71 | typedef struct { |
72 | const char *name; |
73 | kmp_mutex_impl_t id; |
74 | } kmp_mutex_impl_info_t; |
75 | |
76 | enum tool_setting_e { |
77 | omp_tool_error, |
78 | omp_tool_unset, |
79 | omp_tool_disabled, |
80 | omp_tool_enabled |
81 | }; |
82 | |
83 | /***************************************************************************** |
84 | * global variables |
85 | ****************************************************************************/ |
86 | |
87 | ompt_callbacks_active_t ompt_enabled; |
88 | |
89 | ompt_state_info_t ompt_state_info[] = { |
90 | #define ompt_state_macro(state, code) {#state, state}, |
91 | FOREACH_OMPT_STATE(ompt_state_macro) |
92 | #undef ompt_state_macro |
93 | }; |
94 | |
95 | kmp_mutex_impl_info_t kmp_mutex_impl_info[] = { |
96 | #define kmp_mutex_impl_macro(name, id) {#name, name}, |
97 | FOREACH_KMP_MUTEX_IMPL(kmp_mutex_impl_macro) |
98 | #undef kmp_mutex_impl_macro |
99 | }; |
100 | |
101 | ompt_callbacks_internal_t ompt_callbacks; |
102 | |
103 | static ompt_start_tool_result_t *ompt_start_tool_result = NULL; |
104 | |
105 | #if KMP_OS_WINDOWS |
106 | static HMODULE ompt_tool_module = NULL; |
107 | #define OMPT_DLCLOSE(Lib) FreeLibrary(Lib) |
108 | #else |
109 | static void *ompt_tool_module = NULL; |
110 | #define OMPT_DLCLOSE(Lib) dlclose(Lib) |
111 | #endif |
112 | |
113 | /// Used to track the initializer and the finalizer provided by libomptarget |
114 | static ompt_start_tool_result_t *libomptarget_ompt_result = NULL; |
115 | |
116 | /***************************************************************************** |
117 | * forward declarations |
118 | ****************************************************************************/ |
119 | |
120 | static ompt_interface_fn_t ompt_fn_lookup(const char *s); |
121 | |
122 | OMPT_API_ROUTINE ompt_data_t *ompt_get_thread_data(void); |
123 | |
124 | /***************************************************************************** |
125 | * initialization and finalization (private operations) |
126 | ****************************************************************************/ |
127 | |
128 | typedef ompt_start_tool_result_t *(*ompt_start_tool_t)(unsigned int, |
129 | const char *); |
130 | |
131 | #if KMP_OS_DARWIN |
132 | |
133 | // While Darwin supports weak symbols, the library that wishes to provide a new |
134 | // implementation has to link against this runtime which defeats the purpose |
135 | // of having tools that are agnostic of the underlying runtime implementation. |
136 | // |
137 | // Fortunately, the linker includes all symbols of an executable in the global |
138 | // symbol table by default so dlsym() even finds static implementations of |
139 | // ompt_start_tool. For this to work on Linux, -Wl,--export-dynamic needs to be |
140 | // passed when building the application which we don't want to rely on. |
141 | |
142 | static ompt_start_tool_result_t *ompt_tool_darwin(unsigned int omp_version, |
143 | const char *runtime_version) { |
144 | ompt_start_tool_result_t *ret = NULL; |
145 | // Search symbol in the current address space. |
146 | ompt_start_tool_t start_tool = |
147 | (ompt_start_tool_t)dlsym(RTLD_DEFAULT, "ompt_start_tool" ); |
148 | if (start_tool) { |
149 | ret = start_tool(omp_version, runtime_version); |
150 | } |
151 | return ret; |
152 | } |
153 | |
154 | #elif OMPT_HAVE_WEAK_ATTRIBUTE |
155 | |
156 | // On Unix-like systems that support weak symbols the following implementation |
157 | // of ompt_start_tool() will be used in case no tool-supplied implementation of |
158 | // this function is present in the address space of a process. |
159 | |
160 | _OMP_EXTERN OMPT_WEAK_ATTRIBUTE ompt_start_tool_result_t * |
161 | ompt_start_tool(unsigned int omp_version, const char *runtime_version) { |
162 | ompt_start_tool_result_t *ret = NULL; |
163 | // Search next symbol in the current address space. This can happen if the |
164 | // runtime library is linked before the tool. Since glibc 2.2 strong symbols |
165 | // don't override weak symbols that have been found before unless the user |
166 | // sets the environment variable LD_DYNAMIC_WEAK. |
167 | ompt_start_tool_t next_tool = |
168 | (ompt_start_tool_t)dlsym(RTLD_NEXT, name: "ompt_start_tool" ); |
169 | if (next_tool) { |
170 | ret = next_tool(omp_version, runtime_version); |
171 | } |
172 | return ret; |
173 | } |
174 | |
175 | #elif OMPT_HAVE_PSAPI |
176 | |
177 | // On Windows, the ompt_tool_windows function is used to find the |
178 | // ompt_start_tool symbol across all modules loaded by a process. If |
179 | // ompt_start_tool is found, ompt_start_tool's return value is used to |
180 | // initialize the tool. Otherwise, NULL is returned and OMPT won't be enabled. |
181 | |
182 | #include <psapi.h> |
183 | #pragma comment(lib, "psapi.lib") |
184 | |
185 | // The number of loaded modules to start enumeration with EnumProcessModules() |
186 | #define NUM_MODULES 128 |
187 | |
188 | static ompt_start_tool_result_t * |
189 | ompt_tool_windows(unsigned int omp_version, const char *runtime_version) { |
190 | int i; |
191 | DWORD needed, new_size; |
192 | HMODULE *modules; |
193 | HANDLE process = GetCurrentProcess(); |
194 | modules = (HMODULE *)malloc(NUM_MODULES * sizeof(HMODULE)); |
195 | ompt_start_tool_t ompt_tool_p = NULL; |
196 | |
197 | #if OMPT_DEBUG |
198 | printf("ompt_tool_windows(): looking for ompt_start_tool\n" ); |
199 | #endif |
200 | if (!EnumProcessModules(process, modules, NUM_MODULES * sizeof(HMODULE), |
201 | &needed)) { |
202 | // Regardless of the error reason use the stub initialization function |
203 | free(modules); |
204 | return NULL; |
205 | } |
206 | // Check if NUM_MODULES is enough to list all modules |
207 | new_size = needed / sizeof(HMODULE); |
208 | if (new_size > NUM_MODULES) { |
209 | #if OMPT_DEBUG |
210 | printf("ompt_tool_windows(): resize buffer to %d bytes\n" , needed); |
211 | #endif |
212 | modules = (HMODULE *)realloc(modules, needed); |
213 | // If resizing failed use the stub function. |
214 | if (!EnumProcessModules(process, modules, needed, &needed)) { |
215 | free(modules); |
216 | return NULL; |
217 | } |
218 | } |
219 | for (i = 0; i < new_size; ++i) { |
220 | (FARPROC &)ompt_tool_p = GetProcAddress(modules[i], "ompt_start_tool" ); |
221 | if (ompt_tool_p) { |
222 | #if OMPT_DEBUG |
223 | TCHAR modName[MAX_PATH]; |
224 | if (GetModuleFileName(modules[i], modName, MAX_PATH)) |
225 | printf("ompt_tool_windows(): ompt_start_tool found in module %s\n" , |
226 | modName); |
227 | #endif |
228 | free(modules); |
229 | return (*ompt_tool_p)(omp_version, runtime_version); |
230 | } |
231 | #if OMPT_DEBUG |
232 | else { |
233 | TCHAR modName[MAX_PATH]; |
234 | if (GetModuleFileName(modules[i], modName, MAX_PATH)) |
235 | printf("ompt_tool_windows(): ompt_start_tool not found in module %s\n" , |
236 | modName); |
237 | } |
238 | #endif |
239 | } |
240 | free(modules); |
241 | return NULL; |
242 | } |
243 | #else |
244 | #error Activation of OMPT is not supported on this platform. |
245 | #endif |
246 | |
247 | static ompt_start_tool_result_t * |
248 | ompt_try_start_tool(unsigned int omp_version, const char *runtime_version) { |
249 | ompt_start_tool_result_t *ret = NULL; |
250 | ompt_start_tool_t start_tool = NULL; |
251 | #if KMP_OS_WINDOWS |
252 | // Cannot use colon to describe a list of absolute paths on Windows |
253 | const char *sep = ";" ; |
254 | #else |
255 | const char *sep = ":" ; |
256 | #endif |
257 | |
258 | OMPT_VERBOSE_INIT_PRINT("----- START LOGGING OF TOOL REGISTRATION -----\n" ); |
259 | OMPT_VERBOSE_INIT_PRINT("Search for OMP tool in current address space... " ); |
260 | |
261 | #if KMP_OS_DARWIN |
262 | // Try in the current address space |
263 | ret = ompt_tool_darwin(omp_version, runtime_version); |
264 | #elif OMPT_HAVE_WEAK_ATTRIBUTE |
265 | ret = ompt_start_tool(omp_version, runtime_version); |
266 | #elif OMPT_HAVE_PSAPI |
267 | ret = ompt_tool_windows(omp_version, runtime_version); |
268 | #else |
269 | #error Activation of OMPT is not supported on this platform. |
270 | #endif |
271 | if (ret) { |
272 | OMPT_VERBOSE_INIT_CONTINUED_PRINT("Success.\n" ); |
273 | OMPT_VERBOSE_INIT_PRINT( |
274 | "Tool was started and is using the OMPT interface.\n" ); |
275 | OMPT_VERBOSE_INIT_PRINT("----- END LOGGING OF TOOL REGISTRATION -----\n" ); |
276 | return ret; |
277 | } |
278 | |
279 | // Try tool-libraries-var ICV |
280 | OMPT_VERBOSE_INIT_CONTINUED_PRINT("Failed.\n" ); |
281 | const char *tool_libs = getenv(name: "OMP_TOOL_LIBRARIES" ); |
282 | if (tool_libs) { |
283 | OMPT_VERBOSE_INIT_PRINT("Searching tool libraries...\n" ); |
284 | OMPT_VERBOSE_INIT_PRINT("OMP_TOOL_LIBRARIES = %s\n" , tool_libs); |
285 | char *libs = __kmp_str_format(format: "%s" , tool_libs); |
286 | char *buf; |
287 | char *fname = __kmp_str_token(str: libs, delim: sep, buf: &buf); |
288 | // Reset dl-error |
289 | dlerror(); |
290 | |
291 | while (fname) { |
292 | #if KMP_OS_UNIX |
293 | OMPT_VERBOSE_INIT_PRINT("Opening %s... " , fname); |
294 | void *h = dlopen(file: fname, RTLD_LAZY); |
295 | if (!h) { |
296 | OMPT_VERBOSE_INIT_CONTINUED_PRINT("Failed: %s\n" , dlerror()); |
297 | } else { |
298 | OMPT_VERBOSE_INIT_CONTINUED_PRINT("Success. \n" ); |
299 | OMPT_VERBOSE_INIT_PRINT("Searching for ompt_start_tool in %s... " , |
300 | fname); |
301 | dlerror(); // Clear any existing error |
302 | start_tool = (ompt_start_tool_t)dlsym(handle: h, name: "ompt_start_tool" ); |
303 | if (!start_tool) { |
304 | char *error = dlerror(); |
305 | if (error != NULL) { |
306 | OMPT_VERBOSE_INIT_CONTINUED_PRINT("Failed: %s\n" , error); |
307 | } else { |
308 | OMPT_VERBOSE_INIT_CONTINUED_PRINT("Failed: %s\n" , |
309 | "ompt_start_tool = NULL" ); |
310 | } |
311 | } else |
312 | #elif KMP_OS_WINDOWS |
313 | OMPT_VERBOSE_INIT_PRINT("Opening %s... " , fname); |
314 | HMODULE h = LoadLibrary(fname); |
315 | if (!h) { |
316 | OMPT_VERBOSE_INIT_CONTINUED_PRINT("Failed: Error %u\n" , |
317 | (unsigned)GetLastError()); |
318 | } else { |
319 | OMPT_VERBOSE_INIT_CONTINUED_PRINT("Success. \n" ); |
320 | OMPT_VERBOSE_INIT_PRINT("Searching for ompt_start_tool in %s... " , |
321 | fname); |
322 | start_tool = (ompt_start_tool_t)GetProcAddress(h, "ompt_start_tool" ); |
323 | if (!start_tool) { |
324 | OMPT_VERBOSE_INIT_CONTINUED_PRINT("Failed: Error %u\n" , |
325 | (unsigned)GetLastError()); |
326 | } else |
327 | #else |
328 | #error Activation of OMPT is not supported on this platform. |
329 | #endif |
330 | { // if (start_tool) |
331 | ret = (*start_tool)(omp_version, runtime_version); |
332 | if (ret) { |
333 | OMPT_VERBOSE_INIT_CONTINUED_PRINT("Success.\n" ); |
334 | OMPT_VERBOSE_INIT_PRINT( |
335 | "Tool was started and is using the OMPT interface.\n" ); |
336 | ompt_tool_module = h; |
337 | break; |
338 | } |
339 | OMPT_VERBOSE_INIT_CONTINUED_PRINT( |
340 | "Found but not using the OMPT interface.\n" ); |
341 | OMPT_VERBOSE_INIT_PRINT("Continuing search...\n" ); |
342 | } |
343 | OMPT_DLCLOSE(h); |
344 | } |
345 | fname = __kmp_str_token(NULL, delim: sep, buf: &buf); |
346 | } |
347 | __kmp_str_free(str: &libs); |
348 | } else { |
349 | OMPT_VERBOSE_INIT_PRINT("No OMP_TOOL_LIBRARIES defined.\n" ); |
350 | } |
351 | |
352 | // usable tool found in tool-libraries |
353 | if (ret) { |
354 | OMPT_VERBOSE_INIT_PRINT("----- END LOGGING OF TOOL REGISTRATION -----\n" ); |
355 | return ret; |
356 | } |
357 | |
358 | #if KMP_OS_UNIX |
359 | { // Non-standard: load archer tool if application is built with TSan |
360 | const char *fname = "libarcher.so" ; |
361 | OMPT_VERBOSE_INIT_PRINT( |
362 | "...searching tool libraries failed. Using archer tool.\n" ); |
363 | OMPT_VERBOSE_INIT_PRINT("Opening %s... " , fname); |
364 | void *h = dlopen(file: fname, RTLD_LAZY); |
365 | if (h) { |
366 | OMPT_VERBOSE_INIT_CONTINUED_PRINT("Success.\n" ); |
367 | OMPT_VERBOSE_INIT_PRINT("Searching for ompt_start_tool in %s... " , fname); |
368 | start_tool = (ompt_start_tool_t)dlsym(handle: h, name: "ompt_start_tool" ); |
369 | if (start_tool) { |
370 | ret = (*start_tool)(omp_version, runtime_version); |
371 | if (ret) { |
372 | OMPT_VERBOSE_INIT_CONTINUED_PRINT("Success.\n" ); |
373 | OMPT_VERBOSE_INIT_PRINT( |
374 | "Tool was started and is using the OMPT interface.\n" ); |
375 | OMPT_VERBOSE_INIT_PRINT( |
376 | "----- END LOGGING OF TOOL REGISTRATION -----\n" ); |
377 | return ret; |
378 | } |
379 | OMPT_VERBOSE_INIT_CONTINUED_PRINT( |
380 | "Found but not using the OMPT interface.\n" ); |
381 | } else { |
382 | OMPT_VERBOSE_INIT_CONTINUED_PRINT("Failed: %s\n" , dlerror()); |
383 | } |
384 | } |
385 | } |
386 | #endif |
387 | OMPT_VERBOSE_INIT_PRINT("No OMP tool loaded.\n" ); |
388 | OMPT_VERBOSE_INIT_PRINT("----- END LOGGING OF TOOL REGISTRATION -----\n" ); |
389 | return ret; |
390 | } |
391 | |
392 | void ompt_pre_init() { |
393 | //-------------------------------------------------- |
394 | // Execute the pre-initialization logic only once. |
395 | //-------------------------------------------------- |
396 | static int ompt_pre_initialized = 0; |
397 | |
398 | if (ompt_pre_initialized) |
399 | return; |
400 | |
401 | ompt_pre_initialized = 1; |
402 | |
403 | //-------------------------------------------------- |
404 | // Use a tool iff a tool is enabled and available. |
405 | //-------------------------------------------------- |
406 | const char *ompt_env_var = getenv(name: "OMP_TOOL" ); |
407 | tool_setting_e tool_setting = omp_tool_error; |
408 | |
409 | if (!ompt_env_var || !strcmp(s1: ompt_env_var, s2: "" )) |
410 | tool_setting = omp_tool_unset; |
411 | else if (OMPT_STR_MATCH(ompt_env_var, "disabled" )) |
412 | tool_setting = omp_tool_disabled; |
413 | else if (OMPT_STR_MATCH(ompt_env_var, "enabled" )) |
414 | tool_setting = omp_tool_enabled; |
415 | |
416 | const char *ompt_env_verbose_init = getenv(name: "OMP_TOOL_VERBOSE_INIT" ); |
417 | // possible options: disabled | stdout | stderr | <filename> |
418 | // if set, not empty and not disabled -> prepare for logging |
419 | if (ompt_env_verbose_init && strcmp(s1: ompt_env_verbose_init, s2: "" ) && |
420 | !OMPT_STR_MATCH(ompt_env_verbose_init, "disabled" )) { |
421 | verbose_init = 1; |
422 | if (OMPT_STR_MATCH(ompt_env_verbose_init, "STDERR" )) |
423 | verbose_file = stderr; |
424 | else if (OMPT_STR_MATCH(ompt_env_verbose_init, "STDOUT" )) |
425 | verbose_file = stdout; |
426 | else |
427 | verbose_file = fopen(filename: ompt_env_verbose_init, modes: "w" ); |
428 | } else |
429 | verbose_init = 0; |
430 | |
431 | #if OMPT_DEBUG |
432 | printf("ompt_pre_init(): tool_setting = %d\n" , tool_setting); |
433 | #endif |
434 | switch (tool_setting) { |
435 | case omp_tool_disabled: |
436 | OMPT_VERBOSE_INIT_PRINT("OMP tool disabled. \n" ); |
437 | break; |
438 | |
439 | case omp_tool_unset: |
440 | case omp_tool_enabled: |
441 | |
442 | //-------------------------------------------------- |
443 | // Load tool iff specified in environment variable |
444 | //-------------------------------------------------- |
445 | ompt_start_tool_result = |
446 | ompt_try_start_tool(omp_version: __kmp_openmp_version, runtime_version: ompt_get_runtime_version()); |
447 | |
448 | memset(s: &ompt_enabled, c: 0, n: sizeof(ompt_enabled)); |
449 | break; |
450 | |
451 | case omp_tool_error: |
452 | fprintf(stderr, |
453 | format: "Warning: OMP_TOOL has invalid value \"%s\".\n" |
454 | " legal values are (NULL,\"\",\"disabled\"," |
455 | "\"enabled\").\n" , |
456 | ompt_env_var); |
457 | break; |
458 | } |
459 | if (verbose_init && verbose_file != stderr && verbose_file != stdout) |
460 | fclose(stream: verbose_file); |
461 | #if OMPT_DEBUG |
462 | printf("ompt_pre_init(): ompt_enabled = %d\n" , ompt_enabled.enabled); |
463 | #endif |
464 | } |
465 | |
466 | extern "C" int omp_get_initial_device(void); |
467 | |
468 | void ompt_post_init() { |
469 | //-------------------------------------------------- |
470 | // Execute the post-initialization logic only once. |
471 | //-------------------------------------------------- |
472 | static int ompt_post_initialized = 0; |
473 | |
474 | if (ompt_post_initialized) |
475 | return; |
476 | |
477 | ompt_post_initialized = 1; |
478 | |
479 | //-------------------------------------------------- |
480 | // Initialize the tool if so indicated. |
481 | //-------------------------------------------------- |
482 | if (ompt_start_tool_result) { |
483 | ompt_enabled.enabled = !!ompt_start_tool_result->initialize( |
484 | ompt_fn_lookup, omp_get_initial_device(), |
485 | &(ompt_start_tool_result->tool_data)); |
486 | |
487 | if (!ompt_enabled.enabled) { |
488 | // tool not enabled, zero out the bitmap, and done |
489 | memset(s: &ompt_enabled, c: 0, n: sizeof(ompt_enabled)); |
490 | return; |
491 | } |
492 | |
493 | kmp_info_t *root_thread = ompt_get_thread(); |
494 | |
495 | ompt_set_thread_state(thread: root_thread, state: ompt_state_overhead); |
496 | |
497 | if (ompt_enabled.ompt_callback_thread_begin) { |
498 | ompt_callbacks.ompt_callback(ompt_callback_thread_begin)( |
499 | ompt_thread_initial, __ompt_get_thread_data_internal()); |
500 | } |
501 | ompt_data_t *task_data = nullptr; |
502 | ompt_data_t *parallel_data = nullptr; |
503 | __ompt_get_task_info_internal(ancestor_level: 0, NULL, task_data: &task_data, NULL, parallel_data: ¶llel_data, |
504 | NULL); |
505 | if (ompt_enabled.ompt_callback_implicit_task) { |
506 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)( |
507 | ompt_scope_begin, parallel_data, task_data, 1, 1, ompt_task_initial); |
508 | } |
509 | |
510 | ompt_set_thread_state(thread: root_thread, state: ompt_state_work_serial); |
511 | } |
512 | } |
513 | |
514 | void ompt_fini() { |
515 | if (ompt_enabled.enabled) { |
516 | if (ompt_start_tool_result && ompt_start_tool_result->finalize) { |
517 | ompt_start_tool_result->finalize(&(ompt_start_tool_result->tool_data)); |
518 | } |
519 | if (libomptarget_ompt_result && libomptarget_ompt_result->finalize) { |
520 | libomptarget_ompt_result->finalize(NULL); |
521 | } |
522 | } |
523 | |
524 | if (ompt_tool_module) |
525 | OMPT_DLCLOSE(ompt_tool_module); |
526 | memset(s: &ompt_enabled, c: 0, n: sizeof(ompt_enabled)); |
527 | } |
528 | |
529 | /***************************************************************************** |
530 | * interface operations |
531 | ****************************************************************************/ |
532 | |
533 | /***************************************************************************** |
534 | * state |
535 | ****************************************************************************/ |
536 | |
537 | OMPT_API_ROUTINE int ompt_enumerate_states(int current_state, int *next_state, |
538 | const char **next_state_name) { |
539 | const static int len = sizeof(ompt_state_info) / sizeof(ompt_state_info_t); |
540 | int i = 0; |
541 | |
542 | for (i = 0; i < len - 1; i++) { |
543 | if (ompt_state_info[i].state_id == current_state) { |
544 | *next_state = ompt_state_info[i + 1].state_id; |
545 | *next_state_name = ompt_state_info[i + 1].state_name; |
546 | return 1; |
547 | } |
548 | } |
549 | |
550 | return 0; |
551 | } |
552 | |
553 | OMPT_API_ROUTINE int ompt_enumerate_mutex_impls(int current_impl, |
554 | int *next_impl, |
555 | const char **next_impl_name) { |
556 | const static int len = |
557 | sizeof(kmp_mutex_impl_info) / sizeof(kmp_mutex_impl_info_t); |
558 | int i = 0; |
559 | for (i = 0; i < len - 1; i++) { |
560 | if (kmp_mutex_impl_info[i].id != current_impl) |
561 | continue; |
562 | *next_impl = kmp_mutex_impl_info[i + 1].id; |
563 | *next_impl_name = kmp_mutex_impl_info[i + 1].name; |
564 | return 1; |
565 | } |
566 | return 0; |
567 | } |
568 | |
569 | /***************************************************************************** |
570 | * callbacks |
571 | ****************************************************************************/ |
572 | |
573 | OMPT_API_ROUTINE ompt_set_result_t ompt_set_callback(ompt_callbacks_t which, |
574 | ompt_callback_t callback) { |
575 | switch (which) { |
576 | |
577 | #define ompt_event_macro(event_name, callback_type, event_id) \ |
578 | case event_name: \ |
579 | ompt_callbacks.ompt_callback(event_name) = (callback_type)callback; \ |
580 | ompt_enabled.event_name = (callback != 0); \ |
581 | if (callback) \ |
582 | return ompt_event_implementation_status(event_name); \ |
583 | else \ |
584 | return ompt_set_always; |
585 | |
586 | FOREACH_OMPT_EVENT(ompt_event_macro) |
587 | |
588 | #undef ompt_event_macro |
589 | |
590 | default: |
591 | return ompt_set_error; |
592 | } |
593 | } |
594 | |
595 | OMPT_API_ROUTINE int ompt_get_callback(ompt_callbacks_t which, |
596 | ompt_callback_t *callback) { |
597 | if (!ompt_enabled.enabled) |
598 | return ompt_get_callback_failure; |
599 | |
600 | switch (which) { |
601 | |
602 | #define ompt_event_macro(event_name, callback_type, event_id) \ |
603 | case event_name: { \ |
604 | ompt_callback_t mycb = \ |
605 | (ompt_callback_t)ompt_callbacks.ompt_callback(event_name); \ |
606 | if (ompt_enabled.event_name && mycb) { \ |
607 | *callback = mycb; \ |
608 | return ompt_get_callback_success; \ |
609 | } \ |
610 | return ompt_get_callback_failure; \ |
611 | } |
612 | |
613 | FOREACH_OMPT_EVENT(ompt_event_macro) |
614 | |
615 | #undef ompt_event_macro |
616 | |
617 | default: |
618 | return ompt_get_callback_failure; |
619 | } |
620 | } |
621 | |
622 | /***************************************************************************** |
623 | * parallel regions |
624 | ****************************************************************************/ |
625 | |
626 | OMPT_API_ROUTINE int ompt_get_parallel_info(int ancestor_level, |
627 | ompt_data_t **parallel_data, |
628 | int *team_size) { |
629 | if (!ompt_enabled.enabled) |
630 | return 0; |
631 | return __ompt_get_parallel_info_internal(ancestor_level, parallel_data, |
632 | team_size); |
633 | } |
634 | |
635 | OMPT_API_ROUTINE int ompt_get_state(ompt_wait_id_t *wait_id) { |
636 | if (!ompt_enabled.enabled) |
637 | return ompt_state_work_serial; |
638 | int thread_state = __ompt_get_state_internal(omp_wait_id: wait_id); |
639 | |
640 | if (thread_state == ompt_state_undefined) { |
641 | thread_state = ompt_state_work_serial; |
642 | } |
643 | |
644 | return thread_state; |
645 | } |
646 | |
647 | /***************************************************************************** |
648 | * tasks |
649 | ****************************************************************************/ |
650 | |
651 | OMPT_API_ROUTINE ompt_data_t *ompt_get_thread_data(void) { |
652 | if (!ompt_enabled.enabled) |
653 | return NULL; |
654 | return __ompt_get_thread_data_internal(); |
655 | } |
656 | |
657 | OMPT_API_ROUTINE int ompt_get_task_info(int ancestor_level, int *type, |
658 | ompt_data_t **task_data, |
659 | ompt_frame_t **task_frame, |
660 | ompt_data_t **parallel_data, |
661 | int *thread_num) { |
662 | if (!ompt_enabled.enabled) |
663 | return 0; |
664 | return __ompt_get_task_info_internal(ancestor_level, type, task_data, |
665 | task_frame, parallel_data, thread_num); |
666 | } |
667 | |
668 | OMPT_API_ROUTINE int ompt_get_task_memory(void **addr, size_t *size, |
669 | int block) { |
670 | return __ompt_get_task_memory_internal(addr, size, blocknum: block); |
671 | } |
672 | |
673 | /***************************************************************************** |
674 | * num_procs |
675 | ****************************************************************************/ |
676 | |
677 | OMPT_API_ROUTINE int ompt_get_num_procs(void) { |
678 | // copied from kmp_ftn_entry.h (but modified: OMPT can only be called when |
679 | // runtime is initialized) |
680 | return __kmp_avail_proc; |
681 | } |
682 | |
683 | /***************************************************************************** |
684 | * places |
685 | ****************************************************************************/ |
686 | |
687 | OMPT_API_ROUTINE int ompt_get_num_places(void) { |
688 | // copied from kmp_ftn_entry.h (but modified) |
689 | #if !KMP_AFFINITY_SUPPORTED |
690 | return 0; |
691 | #else |
692 | if (!KMP_AFFINITY_CAPABLE()) |
693 | return 0; |
694 | return __kmp_affinity.num_masks; |
695 | #endif |
696 | } |
697 | |
698 | OMPT_API_ROUTINE int ompt_get_place_proc_ids(int place_num, int ids_size, |
699 | int *ids) { |
700 | // copied from kmp_ftn_entry.h (but modified) |
701 | #if !KMP_AFFINITY_SUPPORTED |
702 | return 0; |
703 | #else |
704 | int i, count; |
705 | int tmp_ids[ids_size]; |
706 | for (int j = 0; j < ids_size; j++) |
707 | tmp_ids[j] = 0; |
708 | if (!KMP_AFFINITY_CAPABLE()) |
709 | return 0; |
710 | if (place_num < 0 || place_num >= (int)__kmp_affinity.num_masks) |
711 | return 0; |
712 | /* TODO: Is this safe for asynchronous call from signal handler during runtime |
713 | * shutdown? */ |
714 | kmp_affin_mask_t *mask = KMP_CPU_INDEX(__kmp_affinity.masks, place_num); |
715 | count = 0; |
716 | KMP_CPU_SET_ITERATE(i, mask) { |
717 | if ((!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) || |
718 | (!KMP_CPU_ISSET(i, mask))) { |
719 | continue; |
720 | } |
721 | if (count < ids_size) |
722 | tmp_ids[count] = i; |
723 | count++; |
724 | } |
725 | if (ids_size >= count) { |
726 | for (i = 0; i < count; i++) { |
727 | ids[i] = tmp_ids[i]; |
728 | } |
729 | } |
730 | return count; |
731 | #endif |
732 | } |
733 | |
734 | OMPT_API_ROUTINE int ompt_get_place_num(void) { |
735 | // copied from kmp_ftn_entry.h (but modified) |
736 | #if !KMP_AFFINITY_SUPPORTED |
737 | return -1; |
738 | #else |
739 | if (!ompt_enabled.enabled || __kmp_get_gtid() < 0) |
740 | return -1; |
741 | |
742 | int gtid; |
743 | kmp_info_t *thread; |
744 | if (!KMP_AFFINITY_CAPABLE()) |
745 | return -1; |
746 | gtid = __kmp_entry_gtid(); |
747 | thread = __kmp_thread_from_gtid(gtid); |
748 | if (thread == NULL || thread->th.th_current_place < 0) |
749 | return -1; |
750 | return thread->th.th_current_place; |
751 | #endif |
752 | } |
753 | |
754 | OMPT_API_ROUTINE int ompt_get_partition_place_nums(int place_nums_size, |
755 | int *place_nums) { |
756 | // copied from kmp_ftn_entry.h (but modified) |
757 | #if !KMP_AFFINITY_SUPPORTED |
758 | return 0; |
759 | #else |
760 | if (!ompt_enabled.enabled || __kmp_get_gtid() < 0) |
761 | return 0; |
762 | |
763 | int i, gtid, place_num, first_place, last_place, start, end; |
764 | kmp_info_t *thread; |
765 | if (!KMP_AFFINITY_CAPABLE()) |
766 | return 0; |
767 | gtid = __kmp_entry_gtid(); |
768 | thread = __kmp_thread_from_gtid(gtid); |
769 | if (thread == NULL) |
770 | return 0; |
771 | first_place = thread->th.th_first_place; |
772 | last_place = thread->th.th_last_place; |
773 | if (first_place < 0 || last_place < 0) |
774 | return 0; |
775 | if (first_place <= last_place) { |
776 | start = first_place; |
777 | end = last_place; |
778 | } else { |
779 | start = last_place; |
780 | end = first_place; |
781 | } |
782 | if (end - start <= place_nums_size) |
783 | for (i = 0, place_num = start; place_num <= end; ++place_num, ++i) { |
784 | place_nums[i] = place_num; |
785 | } |
786 | return end - start + 1; |
787 | #endif |
788 | } |
789 | |
790 | /***************************************************************************** |
791 | * places |
792 | ****************************************************************************/ |
793 | |
794 | OMPT_API_ROUTINE int ompt_get_proc_id(void) { |
795 | if (!ompt_enabled.enabled || __kmp_get_gtid() < 0) |
796 | return -1; |
797 | #if KMP_HAVE_SCHED_GETCPU |
798 | return sched_getcpu(); |
799 | #elif KMP_OS_WINDOWS |
800 | PROCESSOR_NUMBER pn; |
801 | GetCurrentProcessorNumberEx(&pn); |
802 | return 64 * pn.Group + pn.Number; |
803 | #else |
804 | return -1; |
805 | #endif |
806 | } |
807 | |
808 | /***************************************************************************** |
809 | * compatability |
810 | ****************************************************************************/ |
811 | |
812 | /* |
813 | * Currently unused function |
814 | OMPT_API_ROUTINE int ompt_get_ompt_version() { return OMPT_VERSION; } |
815 | */ |
816 | |
817 | /***************************************************************************** |
818 | * application-facing API |
819 | ****************************************************************************/ |
820 | |
821 | /*---------------------------------------------------------------------------- |
822 | | control |
823 | ---------------------------------------------------------------------------*/ |
824 | |
825 | int __kmp_control_tool(uint64_t command, uint64_t modifier, void *arg) { |
826 | |
827 | if (ompt_enabled.enabled) { |
828 | if (ompt_enabled.ompt_callback_control_tool) { |
829 | return ompt_callbacks.ompt_callback(ompt_callback_control_tool)( |
830 | command, modifier, arg, OMPT_LOAD_RETURN_ADDRESS(__kmp_entry_gtid())); |
831 | } else { |
832 | return -1; |
833 | } |
834 | } else { |
835 | return -2; |
836 | } |
837 | } |
838 | |
839 | /***************************************************************************** |
840 | * misc |
841 | ****************************************************************************/ |
842 | |
843 | OMPT_API_ROUTINE uint64_t ompt_get_unique_id(void) { |
844 | return __ompt_get_unique_id_internal(); |
845 | } |
846 | |
847 | OMPT_API_ROUTINE void ompt_finalize_tool(void) { __kmp_internal_end_atexit(); } |
848 | |
849 | /***************************************************************************** |
850 | * Target |
851 | ****************************************************************************/ |
852 | |
853 | OMPT_API_ROUTINE int ompt_get_target_info(uint64_t *device_num, |
854 | ompt_id_t *target_id, |
855 | ompt_id_t *host_op_id) { |
856 | return 0; // thread is not in a target region |
857 | } |
858 | |
859 | OMPT_API_ROUTINE int ompt_get_num_devices(void) { |
860 | return 1; // only one device (the current device) is available |
861 | } |
862 | |
863 | /***************************************************************************** |
864 | * API inquiry for tool |
865 | ****************************************************************************/ |
866 | |
867 | static ompt_interface_fn_t ompt_fn_lookup(const char *s) { |
868 | |
869 | #define ompt_interface_fn(fn) \ |
870 | fn##_t fn##_f = fn; \ |
871 | if (strcmp(s, #fn) == 0) \ |
872 | return (ompt_interface_fn_t)fn##_f; |
873 | |
874 | FOREACH_OMPT_INQUIRY_FN(ompt_interface_fn) |
875 | |
876 | #undef ompt_interface_fn |
877 | |
878 | return NULL; |
879 | } |
880 | |
881 | static ompt_data_t *ompt_get_task_data() { return __ompt_get_task_data(); } |
882 | |
883 | static ompt_data_t *ompt_get_target_task_data() { |
884 | return __ompt_get_target_task_data(); |
885 | } |
886 | |
887 | /// Lookup function to query libomp callbacks registered by the tool |
888 | static ompt_interface_fn_t ompt_libomp_target_fn_lookup(const char *s) { |
889 | #define provide_fn(fn) \ |
890 | if (strcmp(s, #fn) == 0) \ |
891 | return (ompt_interface_fn_t)fn; |
892 | |
893 | provide_fn(ompt_get_callback); |
894 | provide_fn(ompt_get_task_data); |
895 | provide_fn(ompt_get_target_task_data); |
896 | #undef provide_fn |
897 | |
898 | #define ompt_interface_fn(fn, type, code) \ |
899 | if (strcmp(s, #fn) == 0) \ |
900 | return (ompt_interface_fn_t)ompt_callbacks.ompt_callback(fn); |
901 | |
902 | FOREACH_OMPT_DEVICE_EVENT(ompt_interface_fn) |
903 | FOREACH_OMPT_EMI_EVENT(ompt_interface_fn) |
904 | FOREACH_OMPT_NOEMI_EVENT(ompt_interface_fn) |
905 | #undef ompt_interface_fn |
906 | |
907 | return (ompt_interface_fn_t)0; |
908 | } |
909 | |
910 | /// This function is called by the libomptarget connector to assign |
911 | /// callbacks already registered with libomp. |
912 | _OMP_EXTERN void ompt_libomp_connect(ompt_start_tool_result_t *result) { |
913 | OMPT_VERBOSE_INIT_PRINT("libomp --> OMPT: Enter ompt_libomp_connect\n" ); |
914 | |
915 | // Ensure libomp callbacks have been added if not already |
916 | __ompt_force_initialization(); |
917 | |
918 | if (ompt_enabled.enabled && |
919 | // Callbacks are initiated only if the device initialize callback |
920 | // has been registered by the tool |
921 | ompt_callbacks.ompt_callback(ompt_callback_device_initialize)) { |
922 | if (result) { |
923 | OMPT_VERBOSE_INIT_PRINT( |
924 | "libomp --> OMPT: Connecting with libomptarget\n" ); |
925 | // Pass in the libomp lookup function so that the already registered |
926 | // functions can be extracted and assigned to the callbacks in |
927 | // libomptarget |
928 | result->initialize(ompt_libomp_target_fn_lookup, |
929 | /* initial_device_num */ 0, /* tool_data */ nullptr); |
930 | // Track the object provided by libomptarget so that the finalizer can be |
931 | // called during OMPT finalization |
932 | libomptarget_ompt_result = result; |
933 | } |
934 | } |
935 | OMPT_VERBOSE_INIT_PRINT("libomp --> OMPT: Exit ompt_libomp_connect\n" ); |
936 | } |
937 | |