fxt-tool.c 15 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689
  1. /*
  2. * StarPU
  3. * Copyright (C) INRIA 2008-2009 (see AUTHORS file)
  4. *
  5. * This program is free software; you can redistribute it and/or modify
  6. * it under the terms of the GNU Lesser General Public License as published by
  7. * the Free Software Foundation; either version 2.1 of the License, or (at
  8. * your option) any later version.
  9. *
  10. * This program is distributed in the hope that it will be useful, but
  11. * WITHOUT ANY WARRANTY; without even the implied warranty of
  12. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  13. *
  14. * See the GNU Lesser General Public License in COPYING.LGPL for more details.
  15. */
  16. #include "fxt-tool.h"
  17. event_list_t events[MAXWORKERS];
  18. workq_list_t taskq;
  19. char *worker_name[MAXWORKERS];
  20. static char *cuda_worker_colors[MAXWORKERS] = {"/greens9/7", "/greens9/6", "/greens9/5", "/greens9/4"};
  21. static char *cpus_worker_colors[MAXWORKERS] = {"/ylorrd9/9", "/ylorrd9/6", "/ylorrd9/3", "/ylorrd9/1"};
  22. static char *other_worker_colors[MAXWORKERS] = {"/greys9/9", "/greys9/8", "/greys9/7", "/greys9/6"};
  23. static char *worker_colors[MAXWORKERS];
  24. static fxt_t fut;
  25. struct fxt_ev_64 ev;
  26. unsigned first_event = 1;
  27. uint64_t start_time = 0;
  28. uint64_t end_time = 0;
  29. unsigned nworkers = 0;
  30. /*
  31. * Paje trace file tools
  32. */
  33. static char *out_paje_path = "paje.trace";
  34. static FILE *out_paje_file;
  35. void paje_output_file_init(void)
  36. {
  37. /* create a new file */
  38. out_paje_file = fopen(out_paje_path, "w+");
  39. write_paje_header(out_paje_file);
  40. fprintf(out_paje_file, " \n \
  41. 1 P 0 \"Program\" \n \
  42. 1 Mn P \"Memory Node\" \n \
  43. 1 T Mn \"Worker\" \n \
  44. 3 S T \"Thread State\" \n \
  45. 3 MS Mn \"Memory Node State\" \n \
  46. 6 Fi S FetchingInput \"1.0 .1 1.0\" \n \
  47. 6 Po S PushingOutput \"0.1 1.0 1.0\" \n \
  48. 6 E S Executing \".0 .6 .4\" \n \
  49. 6 C S Callback \".0 .3 .8\" \n \
  50. 6 B S Blocked \".9 .1 .0\" \n \
  51. 6 A MS Allocating \".4 .1 .0\" \n \
  52. 6 Ar MS AllocatingReuse \".1 .1 .8\" \n \
  53. 6 R MS Reclaiming \".0 .1 .4\" \n \
  54. 6 Co MS DriverCopy \".3 .5 .1\" \n \
  55. 6 No MS Nothing \".0 .0 .0\" \n \
  56. 5 L P Mn Mn L\n");
  57. }
  58. void paje_output_file_terminate(void)
  59. {
  60. /* close the file */
  61. fclose(out_paje_file);
  62. }
  63. /*
  64. * Generic tools
  65. */
  66. void handle_new_mem_node(void)
  67. {
  68. char *memnodestr = malloc(16*sizeof(char));
  69. sprintf(memnodestr, "%ld", ev.param[0]);
  70. fprintf(out_paje_file, "7 %f %s Mn p MEMNODE%s\n", (float)((ev.time-start_time)/1000000.0), memnodestr, memnodestr);
  71. }
  72. static unsigned cuda_index = 0;
  73. static unsigned cpus_index = 0;
  74. static unsigned other_index = 0;
  75. void handle_new_worker(void)
  76. {
  77. /*
  78. arg0 : type of worker (cuda, core ..)
  79. arg1 : memory node
  80. arg2 : thread id
  81. */
  82. char *str = malloc(20*sizeof(char));
  83. char *color = NULL;
  84. strcpy(str, "unknown");
  85. switch (ev.param[0]) {
  86. case FUT_APPS_KEY:
  87. str = "apps";
  88. color = other_worker_colors[other_index++];
  89. break;
  90. case FUT_CORE_KEY:
  91. str = "core";
  92. color = cpus_worker_colors[cpus_index++];
  93. break;
  94. case FUT_CUDA_KEY:
  95. str = "cuda";
  96. color = cuda_worker_colors[cuda_index++];
  97. break;
  98. }
  99. // fprintf(stderr, "new %s worker (tid = %d)\n", str, ev.param[1]);
  100. char *memnodestr = malloc(16*sizeof(char));
  101. sprintf(memnodestr, "%ld", ev.param[1]);
  102. char *tidstr = malloc(16*sizeof(char));
  103. sprintf(tidstr, "%ld", ev.param[2]);
  104. fprintf(out_paje_file, "7 %f %s T MEMNODE%s %s \n", (float)((ev.time-start_time)/1000000.0), tidstr, memnodestr, tidstr);
  105. /* create a new key in the htable */
  106. uint64_t workerid = nworkers++;
  107. ENTRY item;
  108. item.key = tidstr;
  109. item.data = (void *)workerid;
  110. worker_colors[workerid] = color;
  111. ENTRY *res;
  112. res = hsearch(item, FIND);
  113. worker_name[workerid] = str;
  114. /* only register a thread once */
  115. STARPU_ASSERT(res == NULL);
  116. res = hsearch(item, ENTER);
  117. STARPU_ASSERT(res);
  118. events[workerid] = event_list_new();
  119. }
  120. void handle_worker_terminated(void)
  121. {
  122. char *tidstr = malloc(16*sizeof(char));
  123. sprintf(tidstr, "%ld", ev.param[1]);
  124. fprintf(out_paje_file, "8 %f %s\n", (float)((ev.time-start_time)/1000000.0), tidstr);
  125. }
  126. int find_workder_id(unsigned long tid)
  127. {
  128. char tidstr[16];
  129. sprintf(tidstr, "%ld", tid);
  130. ENTRY item;
  131. item.key = tidstr;
  132. item.data = NULL;
  133. ENTRY *res;
  134. res = hsearch(item, FIND);
  135. //STARPU_ASSERT(res);
  136. if (!res)
  137. return -1;
  138. int id = (uintptr_t)(res->data);
  139. return id;
  140. }
  141. void handle_start_codelet_body(void)
  142. {
  143. //fprintf(stderr, "start codelet %p on tid %d\n", (void *)ev.param[0], ev.param[1]);
  144. int worker;
  145. worker = find_workder_id(ev.param[1]);
  146. if (worker < 0) return;
  147. // printf("-> worker %d\n", worker);
  148. fprintf(out_paje_file, "10 %f S %ld E\n", (float)((ev.time-start_time)/1000000.0), ev.param[1] );
  149. event_t e = event_new();
  150. e->time = ev.time;
  151. e->mode = WORKING;
  152. event_list_push_back(events[worker], e);
  153. end_time = STARPU_MAX(end_time, ev.time);
  154. }
  155. void handle_end_codelet_body(void)
  156. {
  157. //fprintf(stderr, "end codelet %p on tid %d\n", (void *)ev.param[0], ev.param[1]);
  158. int worker;
  159. worker = find_workder_id(ev.param[1]);
  160. if (worker < 0) return;
  161. // printf("<- worker %d\n", worker);
  162. fprintf(out_paje_file, "10 %f S %ld B\n", (float)((ev.time-start_time)/1000000.0), ev.param[1] );
  163. event_t e = event_new();
  164. e->time = ev.time;
  165. e->mode = IDLE;
  166. event_list_push_back(events[worker], e);
  167. end_time = STARPU_MAX(end_time, ev.time);
  168. }
  169. void handle_start_callback(void)
  170. {
  171. int worker;
  172. worker = find_workder_id(ev.param[1]);
  173. if (worker < 0) return;
  174. fprintf(out_paje_file, "10 %f S %ld C\n", (float)((ev.time-start_time)/1000000.0), ev.param[1] );
  175. }
  176. void handle_end_callback(void)
  177. {
  178. int worker;
  179. worker = find_workder_id(ev.param[1]);
  180. if (worker < 0) return;
  181. fprintf(out_paje_file, "10 %f S %ld B\n", (float)((ev.time-start_time)/1000000.0), ev.param[1] );
  182. }
  183. void handle_start_fetch_input(void)
  184. {
  185. int worker;
  186. worker = find_workder_id(ev.param[1]);
  187. if (worker < 0) return;
  188. fprintf(out_paje_file, "10 %f S %ld Fi\n", (float)((ev.time-start_time)/1000000.0), ev.param[1] );
  189. event_t e = event_new();
  190. e->time = ev.time;
  191. e->mode = FETCHING;
  192. event_list_push_back(events[worker], e);
  193. end_time = STARPU_MAX(end_time, ev.time);
  194. }
  195. void handle_end_fetch_input(void)
  196. {
  197. int worker;
  198. worker = find_workder_id(ev.param[1]);
  199. if (worker < 0) return;
  200. fprintf(out_paje_file, "10 %f S %ld B\n", (float)((ev.time-start_time)/1000000.0), ev.param[1] );
  201. event_t e = event_new();
  202. e->time = ev.time;
  203. e->mode = IDLE;
  204. event_list_push_back(events[worker], e);
  205. end_time = STARPU_MAX(end_time, ev.time);
  206. }
  207. void handle_start_push_output(void)
  208. {
  209. int worker;
  210. worker = find_workder_id(ev.param[1]);
  211. if (worker < 0) return;
  212. fprintf(out_paje_file, "10 %f S %ld Po\n", (float)((ev.time-start_time)/1000000.0), ev.param[1] );
  213. event_t e = event_new();
  214. e->time = ev.time;
  215. e->mode = PUSHING;
  216. event_list_push_back(events[worker], e);
  217. end_time = STARPU_MAX(end_time, ev.time);
  218. }
  219. void handle_end_push_output(void)
  220. {
  221. int worker;
  222. worker = find_workder_id(ev.param[1]);
  223. if (worker < 0) return;
  224. fprintf(out_paje_file, "10 %f S %ld B\n", (float)((ev.time-start_time)/1000000.0), ev.param[1] );
  225. event_t e = event_new();
  226. e->time = ev.time;
  227. e->mode = IDLE;
  228. event_list_push_back(events[worker], e);
  229. end_time = STARPU_MAX(end_time, ev.time);
  230. }
  231. void handle_data_copy(void)
  232. {
  233. }
  234. void handle_start_driver_copy(void)
  235. {
  236. unsigned src = ev.param[0];
  237. unsigned dst = ev.param[1];
  238. unsigned size = ev.param[2];
  239. unsigned comid = ev.param[3];
  240. fprintf(out_paje_file, "10 %f MS MEMNODE%d Co\n", (float)((ev.time-start_time)/1000000.0), dst);
  241. fprintf(out_paje_file, "18 %f L p %d MEMNODE%d com_%d\n", (float)((ev.time-start_time)/1000000.0), size, src, comid);
  242. }
  243. void handle_end_driver_copy(void)
  244. {
  245. unsigned dst = ev.param[1];
  246. unsigned size = ev.param[2];
  247. unsigned comid = ev.param[3];
  248. fprintf(out_paje_file, "10 %f MS MEMNODE%d No\n", (float)((ev.time-start_time)/1000000.0), dst);
  249. fprintf(out_paje_file, "19 %f L p %d MEMNODE%d com_%d\n", (float)((ev.time-start_time)/1000000.0), size, dst, comid);
  250. }
  251. void handle_start_alloc(void)
  252. {
  253. unsigned memnode = ev.param[0];
  254. fprintf(out_paje_file, "10 %f MS MEMNODE%d A\n", (float)((ev.time-start_time)/1000000.0), memnode);
  255. }
  256. void handle_end_alloc(void)
  257. {
  258. unsigned memnode = ev.param[0];
  259. fprintf(out_paje_file, "10 %f MS MEMNODE%d No\n", (float)((ev.time-start_time)/1000000.0), memnode);
  260. }
  261. void handle_start_alloc_reuse(void)
  262. {
  263. unsigned memnode = ev.param[0];
  264. fprintf(out_paje_file, "10 %f MS MEMNODE%d Ar\n", (float)((ev.time-start_time)/1000000.0), memnode);
  265. }
  266. void handle_end_alloc_reuse(void)
  267. {
  268. unsigned memnode = ev.param[0];
  269. fprintf(out_paje_file, "10 %f MS MEMNODE%d No\n", (float)((ev.time-start_time)/1000000.0), memnode);
  270. }
  271. void handle_start_memreclaim(void)
  272. {
  273. unsigned memnode = ev.param[0];
  274. fprintf(out_paje_file, "10 %f MS MEMNODE%d R\n", (float)((ev.time-start_time)/1000000.0), memnode);
  275. }
  276. void handle_end_memreclaim(void)
  277. {
  278. unsigned memnode = ev.param[0];
  279. fprintf(out_paje_file, "10 %f MS MEMNODE%d No\n", (float)((ev.time-start_time)/1000000.0), memnode);
  280. }
  281. int maxq_size = 0;
  282. int curq_size = 0;
  283. void handle_job_push(void)
  284. {
  285. curq_size++;
  286. maxq_size = STARPU_MAX(maxq_size, curq_size);
  287. workq_t e = workq_new();
  288. e->time = ev.time;
  289. e->diff = +1;
  290. e->current_size = curq_size;
  291. workq_list_push_back(taskq, e);
  292. }
  293. void handle_job_pop(void)
  294. {
  295. curq_size--;
  296. workq_t e = workq_new();
  297. e->time = ev.time;
  298. e->diff = -1;
  299. e->current_size = curq_size;
  300. workq_list_push_back(taskq, e);
  301. }
  302. void handle_codelet_tag_deps(void)
  303. {
  304. uint64_t child;
  305. uint64_t father;
  306. child = ev.param[0];
  307. father = ev.param[1];
  308. add_deps(child, father);
  309. }
  310. void handle_task_done(void)
  311. {
  312. uint64_t tag_id;
  313. tag_id = ev.param[0];
  314. int worker;
  315. worker = find_workder_id(ev.param[1]);
  316. dot_set_tag_done(tag_id, worker_colors[worker]);
  317. }
  318. #ifdef FLASH_RENDER
  319. void generate_flash_output(void)
  320. {
  321. flash_engine_init();
  322. flash_engine_generate_output(events, taskq, worker_name, nworkers, maxq_size, start_time, end_time, "toto.swf");
  323. }
  324. #endif
  325. void generate_svg_output(void)
  326. {
  327. svg_engine_generate_output(events, taskq, worker_name, nworkers, maxq_size, start_time, end_time, "toto.svg");
  328. }
  329. void generate_gnuplot_output(void)
  330. {
  331. FILE *output;
  332. output = fopen("data", "w+");
  333. STARPU_ASSERT(output);
  334. unsigned linesize;
  335. unsigned maxline = 0;
  336. unsigned worker;
  337. for (worker = 0; worker < nworkers; worker++)
  338. {
  339. linesize = 0;
  340. event_itor_t i;
  341. for (i = event_list_begin(events[worker]);
  342. i != event_list_end(events[worker]);
  343. i = event_list_next(i))
  344. {
  345. linesize++;
  346. }
  347. maxline = STARPU_MAX(maxline, linesize);
  348. }
  349. unsigned i;
  350. for (i = 0; i < maxline + 1; i++)
  351. {
  352. fprintf(output, "bla\t");
  353. }
  354. fprintf(output,"\n");
  355. for (worker = 0; worker < nworkers; worker++)
  356. {
  357. unsigned long prev = start_time;
  358. fprintf(output, "%d\t", 0);
  359. event_itor_t i;
  360. for (i = event_list_begin(events[worker]);
  361. i != event_list_end(events[worker]);
  362. i = event_list_next(i))
  363. {
  364. fprintf(output, "%lu\t", (i->time - prev)/FACTOR);
  365. prev = i->time;
  366. }
  367. fprintf(output, "\n");
  368. }
  369. fclose(output);
  370. }
  371. #ifdef USE_GTK
  372. void gtk_viewer(int argc, char **argv)
  373. {
  374. gtk_viewer_apps(argc, argv, events, taskq, worker_name, nworkers, maxq_size, start_time, end_time);
  375. }
  376. #endif
  377. /*
  378. * This program should be used to parse the log generated by FxT
  379. */
  380. int main(int argc, char **argv)
  381. {
  382. char *filename, *filenameout = NULL;
  383. int ret;
  384. int fd_in, fd_out;
  385. int use_stdout = 1;
  386. init_dag_dot();
  387. if (argc < 2) {
  388. fprintf(stderr, "Usage : %s input_filename [-o output_filename]\n", argv[0]);
  389. exit(-1);
  390. }
  391. filename = argv[1];
  392. fd_in = open(filename, O_RDONLY);
  393. if (fd_in < 0) {
  394. perror("open failed :");
  395. exit(-1);
  396. }
  397. if (argc > 2) {
  398. filenameout = argv[2];
  399. use_stdout = 0;
  400. fd_out = open(filenameout, O_RDWR);
  401. if (fd_out < 0) {
  402. perror("open (out) failed :");
  403. exit(-1);
  404. }
  405. }
  406. fut = fxt_fdopen(fd_in);
  407. if (!fut) {
  408. perror("fxt_fdopen :");
  409. exit(-1);
  410. }
  411. fxt_blockev_t block;
  412. block = fxt_blockev_enter(fut);
  413. /* create a htable to identify each worker(tid) */
  414. hcreate(MAXWORKERS);
  415. paje_output_file_init();
  416. taskq = workq_list_new();
  417. while(1) {
  418. ret = fxt_next_ev(block, FXT_EV_TYPE_64, (struct fxt_ev *)&ev);
  419. if (ret != FXT_EV_OK) {
  420. fprintf(stderr, "no more block ...\n");
  421. break;
  422. }
  423. __attribute__ ((unused)) int nbparam = ev.nb_params;
  424. if (first_event)
  425. {
  426. first_event = 0;
  427. start_time = ev.time;
  428. fprintf(out_paje_file, "7 %f p P 0 program \n", (float)(start_time-start_time));
  429. }
  430. switch (ev.code) {
  431. case FUT_NEW_WORKER_KEY:
  432. handle_new_worker();
  433. break;
  434. case FUT_NEW_MEM_NODE:
  435. handle_new_mem_node();
  436. break;
  437. /* detect when the workers were idling or not */
  438. case FUT_START_CODELET_BODY:
  439. handle_start_codelet_body();
  440. break;
  441. case FUT_END_CODELET_BODY:
  442. handle_end_codelet_body();
  443. break;
  444. case FUT_START_CALLBACK:
  445. handle_start_callback();
  446. break;
  447. case FUT_END_CALLBACK:
  448. handle_end_callback();
  449. break;
  450. /* monitor stack size */
  451. case FUT_JOB_PUSH:
  452. handle_job_push();
  453. break;
  454. case FUT_JOB_POP:
  455. handle_job_pop();
  456. break;
  457. /* check the memory transfer overhead */
  458. case FUT_START_FETCH_INPUT:
  459. handle_start_fetch_input();
  460. break;
  461. case FUT_END_FETCH_INPUT:
  462. handle_end_fetch_input();
  463. break;
  464. case FUT_START_PUSH_OUTPUT:
  465. handle_start_push_output();
  466. break;
  467. case FUT_END_PUSH_OUTPUT:
  468. handle_end_push_output();
  469. break;
  470. case FUT_CODELET_TAG:
  471. //handle_codelet_tag();
  472. break;
  473. case FUT_CODELET_TAG_DEPS:
  474. handle_codelet_tag_deps();
  475. break;
  476. case FUT_TASK_DONE:
  477. handle_task_done();
  478. break;
  479. case FUT_DATA_COPY:
  480. handle_data_copy();
  481. break;
  482. case FUT_START_DRIVER_COPY:
  483. handle_start_driver_copy();
  484. break;
  485. case FUT_END_DRIVER_COPY:
  486. handle_end_driver_copy();
  487. break;
  488. case FUT_WORK_STEALING:
  489. /* XXX */
  490. break;
  491. case FUT_WORKER_TERMINATED:
  492. handle_worker_terminated();
  493. break;
  494. case FUT_START_ALLOC:
  495. handle_start_alloc();
  496. break;
  497. case FUT_END_ALLOC:
  498. handle_end_alloc();
  499. break;
  500. case FUT_START_ALLOC_REUSE:
  501. handle_start_alloc_reuse();
  502. break;
  503. case FUT_END_ALLOC_REUSE:
  504. handle_end_alloc_reuse();
  505. break;
  506. case FUT_START_MEMRECLAIM:
  507. handle_start_memreclaim();
  508. break;
  509. case FUT_END_MEMRECLAIM:
  510. handle_end_memreclaim();
  511. break;
  512. default:
  513. fprintf(stderr, "unknown event.. %x at time %llx\n", (unsigned)ev.code, (long long unsigned)ev.time);
  514. break;
  515. }
  516. }
  517. generate_gnuplot_output();
  518. //generate_flash_output();
  519. generate_svg_output();
  520. paje_output_file_terminate();
  521. terminate_dat_dot();
  522. #ifdef USE_GTK
  523. gtk_viewer(argc, argv);
  524. #endif
  525. return 0;
  526. }