starpu_replay.c 30 KB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171
  1. /* StarPU --- Runtime system for heterogeneous multicore architectures.
  2. *
  3. * Copyright (C) 2016-2020 Université de Bordeaux, CNRS (LaBRI UMR 5800), Inria
  4. * Copyright (C) 2017 Erwan Leria
  5. *
  6. * StarPU is free software; you can redistribute it and/or modify
  7. * it under the terms of the GNU Lesser General Public License as published by
  8. * the Free Software Foundation; either version 2.1 of the License, or (at
  9. * your option) any later version.
  10. *
  11. * StarPU is distributed in the hope that it will be useful, but
  12. * WITHOUT ANY WARRANTY; without even the implied warranty of
  13. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  14. *
  15. * See the GNU Lesser General Public License in COPYING.LGPL for more details.
  16. */
  17. /*
  18. * This reads a tasks.rec file and replays the recorded task graph.
  19. * Currently, this version is done to run with simgrid.
  20. *
  21. * For further information, contact erwan.leria@inria.fr
  22. */
  23. #include <starpu.h>
  24. #include <unistd.h>
  25. #include <stdio.h>
  26. #include <math.h>
  27. #include <common/uthash.h>
  28. #include <common/utils.h>
  29. #include <starpu_scheduler.h>
  30. #include <common/rbtree.h>
  31. #define REPLAY_NMAX_DEPENDENCIES 8
  32. #define ARRAY_DUP(in, out, n) memcpy(out, in, n * sizeof(*out))
  33. #define ARRAY_INIT(array, n) memset(array, 0, n * sizeof(*array))
  34. /* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
  35. * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
  36. * Declarations of global variables, structures, pointers, ... *
  37. * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
  38. * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
  39. static int static_workerid;
  40. /* TODO: move to core header while moving starpu_replay_sched to core */
  41. extern void schedRecInit(const char * filename);
  42. extern void applySchedRec(struct starpu_task * starpu_task, long submit_order);
  43. /* Enum for normal and "wontuse" tasks */
  44. enum task_type {NormalTask, WontUseTask};
  45. typedef unsigned long jobid_t;
  46. enum task_type control;
  47. static char *name = NULL;
  48. static char *model = NULL;
  49. static jobid_t jobid;
  50. static jobid_t *dependson;
  51. static long submitorder = -1;
  52. static starpu_tag_t tag;
  53. static int workerid;
  54. static uint32_t footprint;
  55. static double flops, total_flops = 0.;
  56. static double startTime; //start time (The instant when the task starts)
  57. static double endTime; //end time (The instant when the task ends)
  58. static int iteration = -1;
  59. static starpu_data_handle_t handles[STARPU_NMAXBUFS];
  60. static enum starpu_data_access_mode modes[STARPU_NMAXBUFS];
  61. static char normal_reg_signal[STARPU_NMAXBUFS];
  62. /* Use the following arrays when the number of data is greater than STARPU_NMAXBUFS */
  63. starpu_data_handle_t * handles_ptr;
  64. enum starpu_data_access_mode * modes_ptr;
  65. size_t * sizes_set;
  66. static size_t dependson_size;
  67. static size_t ndependson;
  68. static unsigned nb_parameters = 0; /* Number of parameters */
  69. static int alloc_mode; /* If alloc_mode value is 1, then the handles are stored in dyn_handles, else they are in handles */
  70. static int priority = 0;
  71. char * reg_signal = NULL; /* The register signal (0 or 1 coded on 8 bit) is used to know which handle of the task has to be registered in StarPU (in fact to avoid handle twice)*/
  72. /* Record all tasks, hashed by jobid. */
  73. static struct task
  74. {
  75. struct starpu_rbtree_node node;
  76. UT_hash_handle hh;
  77. jobid_t jobid;
  78. int iteration;
  79. long submit_order;
  80. jobid_t *deps;
  81. size_t ndependson;
  82. struct starpu_task task;
  83. enum task_type type;
  84. int reg_signal;
  85. } *tasks;
  86. /* Record handles */
  87. static struct handle
  88. {
  89. UT_hash_handle hh;
  90. starpu_data_handle_t mem_ptr; /* This value should be the registered handle */
  91. starpu_data_handle_t handle; /* The key is the original value of the handle in the file */
  92. } * handles_hash;
  93. /* Record models */
  94. static struct perfmodel
  95. {
  96. UT_hash_handle hh;
  97. struct starpu_perfmodel perfmodel;
  98. char * model_name;
  99. } * model_hash;
  100. /*
  101. * Replay data interface
  102. * We don't care about many things anyway, essentially only sizes.
  103. */
  104. struct replay_interface
  105. {
  106. enum starpu_data_interface_id id;
  107. starpu_data_handle_t orig_handle;
  108. size_t size;
  109. size_t alloc_size;
  110. size_t max_size;
  111. };
  112. static struct starpu_data_interface_ops replay_interface_ops;
  113. static void register_replay(starpu_data_handle_t handle, unsigned home_node, void *data_interface)
  114. {
  115. (void) home_node;
  116. struct replay_interface *replay_interface = data_interface;
  117. unsigned node;
  118. for (node = 0; node < STARPU_MAXNODES; node++)
  119. {
  120. struct replay_interface *local_interface =
  121. starpu_data_get_interface_on_node(handle, node);
  122. local_interface->id = replay_interface->id;
  123. local_interface->orig_handle = replay_interface->orig_handle;
  124. local_interface->size = replay_interface->size;
  125. local_interface->alloc_size = replay_interface->alloc_size;
  126. local_interface->max_size = replay_interface->max_size;
  127. }
  128. }
  129. static void replay_data_register(starpu_data_handle_t *handleptr, starpu_data_handle_t orig_handle, int home_node, size_t size, size_t alloc_size, size_t max_size)
  130. {
  131. if (replay_interface_ops.interfaceid == STARPU_UNKNOWN_INTERFACE_ID)
  132. {
  133. replay_interface_ops.interfaceid = starpu_data_interface_get_next_id();
  134. }
  135. struct replay_interface interface = {
  136. .id = replay_interface_ops.interfaceid,
  137. .orig_handle = orig_handle,
  138. .size = size,
  139. .alloc_size = alloc_size,
  140. .max_size = max_size,
  141. };
  142. starpu_data_register(handleptr, home_node, &interface, &replay_interface_ops);
  143. }
  144. static size_t replay_get_size(starpu_data_handle_t handle)
  145. {
  146. struct replay_interface *interface =
  147. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  148. return interface->size;
  149. }
  150. static size_t replay_get_alloc_size(starpu_data_handle_t handle)
  151. {
  152. struct replay_interface *interface =
  153. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  154. return interface->alloc_size;
  155. }
  156. static size_t replay_get_max_size(starpu_data_handle_t handle)
  157. {
  158. struct replay_interface *interface =
  159. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  160. return interface->max_size;
  161. }
  162. static uint32_t replay_footprint(starpu_data_handle_t handle)
  163. {
  164. return starpu_hash_crc32c_be(replay_get_size(handle), 0);
  165. }
  166. static int replay_compare(void *data_interface_a, void *data_interface_b)
  167. {
  168. struct replay_interface *replay_a = data_interface_a;
  169. struct replay_interface *replay_b = data_interface_b;
  170. /* Two variables are considered compatible if they have the same size */
  171. return replay_a->size == replay_b->size;
  172. }
  173. static void display_replay(starpu_data_handle_t handle, FILE *f)
  174. {
  175. struct replay_interface *replay_interface =
  176. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  177. fprintf(f, "%lu/%lu/%lu\t",
  178. (unsigned long) replay_interface->size,
  179. (unsigned long) replay_interface->alloc_size,
  180. (unsigned long) replay_interface->max_size);
  181. }
  182. static starpu_ssize_t describe_replay(void *data_interface, char *buf, size_t size)
  183. {
  184. struct replay_interface *replay_interface = data_interface;
  185. return snprintf(buf, size, "r%lu/%lu/%lu\t",
  186. (unsigned long) replay_interface->size,
  187. (unsigned long) replay_interface->alloc_size,
  188. (unsigned long) replay_interface->max_size);
  189. }
  190. static starpu_ssize_t allocate_replay_on_node(void *data_interface, unsigned dst_node)
  191. {
  192. struct replay_interface *replay_interface = data_interface;
  193. starpu_memory_allocate(dst_node, replay_interface->alloc_size, STARPU_MEMORY_OVERFLOW);
  194. return 0;
  195. }
  196. static void free_replay_on_node(void *data_interface, unsigned dst_node)
  197. {
  198. struct replay_interface *replay_interface = data_interface;
  199. starpu_memory_deallocate(dst_node, replay_interface->alloc_size);
  200. }
  201. static int replay_copy(void *src_interface, unsigned src_node, void *dst_interface, unsigned dst_node, void *async_data)
  202. {
  203. (void) dst_interface;
  204. struct replay_interface *src = src_interface;
  205. /* We don't care about pointers */
  206. return starpu_interface_copy(1, 0, src_node, 1, 0, dst_node, src->size, async_data);
  207. }
  208. static const struct starpu_data_copy_methods replay_copy_data_methods =
  209. {
  210. .any_to_any = replay_copy,
  211. };
  212. static struct starpu_data_interface_ops replay_interface_ops =
  213. {
  214. .register_data_handle = register_replay,
  215. .allocate_data_on_node = allocate_replay_on_node,
  216. .free_data_on_node = free_replay_on_node,
  217. .copy_methods = &replay_copy_data_methods,
  218. .get_size = replay_get_size,
  219. .get_alloc_size = replay_get_alloc_size,
  220. .get_max_size = replay_get_max_size,
  221. .footprint = replay_footprint,
  222. .compare = replay_compare,
  223. .interfaceid = STARPU_UNKNOWN_INTERFACE_ID,
  224. .interface_size = sizeof(struct replay_interface),
  225. .display = display_replay,
  226. .pack_data = NULL,
  227. .unpack_data = NULL,
  228. .describe = describe_replay,
  229. /* We want to observe actual allocations/deallocations */
  230. .dontcache = 1,
  231. };
  232. /* [SUBMITORDER] The tree of the submit order */
  233. static struct starpu_rbtree tree = STARPU_RBTREE_INITIALIZER;
  234. /* the cmp_fn arg for rb_tree_insert() */
  235. unsigned int diff(struct starpu_rbtree_node * left_elm, struct starpu_rbtree_node * right_elm)
  236. {
  237. long oleft = ((struct task *) left_elm)->submit_order;
  238. long oright = ((struct task *) right_elm)->submit_order;
  239. if (oleft == -1 && oright == -1)
  240. {
  241. if (left_elm < right_elm)
  242. return -1;
  243. else
  244. return 1;
  245. }
  246. return oleft - oright;
  247. }
  248. /* Settings for the perfmodel */
  249. struct task_arg
  250. {
  251. uint32_t footprint;
  252. unsigned narch;
  253. double perf[];
  254. };
  255. uint32_t get_footprint(struct starpu_task * task)
  256. {
  257. return ((struct task_arg*) (task->cl_arg))->footprint;
  258. }
  259. double arch_cost_function(struct starpu_task *task, struct starpu_perfmodel_arch *arch, unsigned nimpl)
  260. {
  261. int device = starpu_perfmodel_arch_comb_get(arch->ndevices, arch->devices);
  262. STARPU_ASSERT(device != -1);
  263. (void) nimpl;
  264. /* Then, get the pointer to the value of the expected time */
  265. struct task_arg *arg = task->cl_arg;
  266. if (device < (int) arg->narch)
  267. {
  268. double val = arg->perf[device];
  269. if (!(val == 0 || isnan(val)))
  270. return val;
  271. }
  272. fprintf(stderr, "[starpu] Error, expected_time is 0 or lower (replay.c line : %d)", __LINE__- 6);
  273. return 0.0;
  274. }
  275. /* End of settings */
  276. static unsigned long nexecuted_tasks;
  277. void dumb_kernel(void *buffers[], void *args) {
  278. (void) buffers;
  279. (void) args;
  280. nexecuted_tasks++;
  281. if (!(nexecuted_tasks % 1000))
  282. {
  283. fprintf(stderr, "\rExecuted task %lu...", nexecuted_tasks);
  284. fflush(stdout);
  285. }
  286. unsigned this_worker = starpu_worker_get_id_check();
  287. struct starpu_perfmodel_arch *perf_arch = starpu_worker_get_perf_archtype(this_worker, STARPU_NMAX_SCHED_CTXS);
  288. struct starpu_task *task = starpu_task_get_current();
  289. unsigned impl = starpu_task_get_implementation(task);
  290. double length = starpu_task_expected_length(task, perf_arch, impl);
  291. STARPU_ASSERT_MSG(!_STARPU_IS_ZERO(length) && !isnan(length),
  292. "Codelet %s does not have a perfmodel, or is not calibrated enough, please re-run in non-simgrid mode until it is calibrated",
  293. starpu_task_get_name(task));
  294. starpu_sleep(length / 1000000);
  295. }
  296. /* [CODELET] Initialization of an unique codelet for all the tasks*/
  297. static int can_execute(unsigned worker_id, struct starpu_task *task, unsigned nimpl)
  298. {
  299. struct starpu_perfmodel_arch * arch = starpu_worker_get_perf_archtype(worker_id, STARPU_NMAX_SCHED_CTXS);
  300. int device = starpu_perfmodel_arch_comb_get(arch->ndevices, arch->devices);
  301. if (device == -1)
  302. /* Doesn't exist yet, thus unknown, assuming it can not work there. */
  303. return 0;
  304. (void) nimpl;
  305. /* Then, get the pointer to the value of the expected time */
  306. struct task_arg *arg = task->cl_arg;
  307. if (device < (int) arg->narch)
  308. {
  309. double val = arg->perf[device];
  310. if (!(val == 0 || isnan(val)))
  311. return 1;
  312. }
  313. return 0;
  314. }
  315. static struct starpu_perfmodel myperfmodel =
  316. {
  317. .type = STARPU_PER_ARCH,
  318. .arch_cost_function = arch_cost_function,
  319. .footprint = get_footprint,
  320. };
  321. static struct starpu_codelet cl =
  322. {
  323. .cpu_funcs = { dumb_kernel },
  324. .cpu_funcs_name = { "dumb_kernel" },
  325. .cuda_funcs = { dumb_kernel },
  326. .opencl_funcs = { dumb_kernel },
  327. .nbuffers = STARPU_VARIABLE_NBUFFERS,
  328. .can_execute = can_execute,
  329. .model = &myperfmodel,
  330. .flags = STARPU_CODELET_SIMGRID_EXECUTE,
  331. };
  332. /* * * * * * * * * * * * * *
  333. * * * * * Functions * * * * *
  334. * * * * * * * * * * * * * * */
  335. /* The following function checks if the program has to use static or dynamic arrays*/
  336. static int set_alloc_mode(int total_parameters)
  337. {
  338. return total_parameters <= STARPU_NMAXBUFS;
  339. }
  340. /* According to the allocation mode, modify handles_ptr and modes_ptr in static or dynamic */
  341. static void arrays_managing(int mode)
  342. {
  343. if (mode)
  344. {
  345. handles_ptr = &handles[0];
  346. modes_ptr = &modes[0];
  347. reg_signal = &normal_reg_signal[0];
  348. }
  349. else
  350. {
  351. _STARPU_MALLOC(handles_ptr, sizeof(*handles_ptr) * nb_parameters);
  352. _STARPU_MALLOC(modes_ptr, sizeof(*modes_ptr) * nb_parameters);
  353. _STARPU_CALLOC(reg_signal, nb_parameters, sizeof(char *));
  354. }
  355. }
  356. /* Check if a handle hasn't been registered yet */
  357. static void variable_data_register_check(size_t * array_of_size, int nb_handles)
  358. {
  359. int h, i;
  360. starpu_data_handle_t orig_handles[nb_handles];
  361. ARRAY_DUP(handles_ptr, orig_handles, nb_handles);
  362. for (h = 0 ; h < nb_handles ; h++)
  363. {
  364. if(reg_signal[h]) /* Get the register signal, if it's 1 do ... */
  365. {
  366. struct handle * handles_cell;
  367. for (i = 0; i < h; i++)
  368. {
  369. /* Maybe we just registered it in this very h loop */
  370. if (handles_ptr[h] == orig_handles[i])
  371. {
  372. handles_ptr[h] = handles_ptr[i];
  373. break;
  374. }
  375. }
  376. if (i == h)
  377. {
  378. _STARPU_MALLOC(handles_cell, sizeof(*handles_cell));
  379. STARPU_ASSERT(handles_cell != NULL);
  380. handles_cell->handle = handles_ptr[h]; /* Get the hidden key (initial handle from the file) to store it as a key*/
  381. replay_data_register(handles_ptr+h, handles_ptr[h],
  382. modes_ptr[h] & STARPU_R ? STARPU_MAIN_RAM : -1,
  383. array_of_size[h], array_of_size[h], array_of_size[h]);
  384. handles_cell->mem_ptr = handles_ptr[h]; /* Store the new value of the handle into the hash table */
  385. HASH_ADD(hh, handles_hash, handle, sizeof(handles_ptr[h]), handles_cell);
  386. }
  387. }
  388. }
  389. }
  390. void reset(void)
  391. {
  392. control = NormalTask;
  393. if (name != NULL)
  394. {
  395. free(name);
  396. name = NULL;
  397. }
  398. if (model != NULL)
  399. {
  400. free(model);
  401. model = NULL;
  402. }
  403. if (sizes_set != NULL)
  404. {
  405. free(sizes_set);
  406. sizes_set = NULL;
  407. }
  408. if (reg_signal != NULL)
  409. {
  410. if (!alloc_mode)
  411. {
  412. free(reg_signal);
  413. reg_signal = NULL;
  414. }
  415. else
  416. {
  417. ARRAY_INIT(reg_signal, nb_parameters);
  418. }
  419. }
  420. jobid = 0;
  421. ndependson = 0;
  422. tag = -1;
  423. workerid = -1;
  424. footprint = 0;
  425. startTime = 0.0;
  426. endTime = 0.0;
  427. if (submitorder != -1)
  428. submitorder = -1;
  429. iteration = -1;
  430. nb_parameters = 0;
  431. alloc_mode = 1;
  432. }
  433. void fix_wontuse_handle(struct task * wontuseTask)
  434. {
  435. STARPU_ASSERT(wontuseTask);
  436. if (!wontuseTask->reg_signal)
  437. /* Data was already registered when we created this task, so it's already a handle */
  438. return;
  439. struct handle *handle_tmp;
  440. /* Data was not registered when we created this task, so this is the application pointer, look it up now */
  441. HASH_FIND(hh, handles_hash, &wontuseTask->task.handles[0], sizeof(wontuseTask->task.handles[0]), handle_tmp);
  442. if (handle_tmp)
  443. wontuseTask->task.handles[0] = handle_tmp->mem_ptr;
  444. else
  445. /* This data wasn't actually used, don't care about it */
  446. wontuseTask->task.handles[0] = NULL;
  447. }
  448. /* Function that submits all the tasks (used when the program reaches EOF) */
  449. int submit_tasks(void)
  450. {
  451. /* Add dependencies */
  452. const struct starpu_rbtree * tmptree = &tree;
  453. struct starpu_rbtree_node * currentNode = starpu_rbtree_first(tmptree);
  454. long last_submitorder = 0;
  455. while (currentNode != NULL)
  456. {
  457. struct task * currentTask = (struct task *) currentNode;
  458. if (currentTask->type == NormalTask)
  459. {
  460. if (currentTask->submit_order != -1)
  461. {
  462. STARPU_ASSERT(currentTask->submit_order >= last_submitorder + 1);
  463. while (currentTask->submit_order > last_submitorder + 1)
  464. {
  465. /* Oops, some tasks were not submitted by original application, fake some */
  466. struct starpu_task *task = starpu_task_create();
  467. int ret;
  468. task->cl = NULL;
  469. ret = starpu_task_submit(task);
  470. STARPU_ASSERT(ret == 0);
  471. last_submitorder++;
  472. }
  473. }
  474. if (currentTask->ndependson > 0)
  475. {
  476. struct starpu_task * taskdeps[currentTask->ndependson];
  477. unsigned i, j = 0;
  478. for (i = 0; i < currentTask->ndependson; i++)
  479. {
  480. struct task * taskdep;
  481. /* Get the ith jobid of deps_jobid */
  482. HASH_FIND(hh, tasks, &currentTask->deps[i], sizeof(jobid), taskdep);
  483. if(taskdep)
  484. {
  485. taskdeps[j] = &taskdep->task;
  486. j ++;
  487. }
  488. }
  489. starpu_task_declare_deps_array(&currentTask->task, j, taskdeps);
  490. }
  491. if (!(currentTask->iteration == -1))
  492. starpu_iteration_push(currentTask->iteration);
  493. applySchedRec(&currentTask->task, currentTask->submit_order);
  494. int ret_val = starpu_task_submit(&currentTask->task);
  495. if (!(currentTask->iteration == -1))
  496. starpu_iteration_pop();
  497. if (ret_val != 0)
  498. {
  499. fprintf(stderr, "\nWhile submitting task %ld (%s): return %d\n",
  500. currentTask->submit_order,
  501. currentTask->task.name? currentTask->task.name : "unknown",
  502. ret_val);
  503. return -1;
  504. }
  505. //fprintf(stderr, "submitting task %s (%lu, %llu)\n", currentTask->task.name?currentTask->task.name:"anonymous", currentTask->jobid, (unsigned long long) currentTask->task.tag_id);
  506. if (!(currentTask->submit_order % 1000))
  507. {
  508. fprintf(stderr, "\rSubmitted task order %ld...", currentTask->submit_order);
  509. fflush(stdout);
  510. }
  511. if (currentTask->submit_order != -1)
  512. last_submitorder++;
  513. }
  514. else
  515. {
  516. fix_wontuse_handle(currentTask); /* Add the handle in the wontuse task */
  517. /* FIXME: can not actually work properly since we have
  518. * disabled sequential consistency, so we don't have any
  519. * easy way to make this wait for the last task that
  520. * wrote to the handle. */
  521. if (currentTask->task.handles[0])
  522. starpu_data_wont_use(currentTask->task.handles[0]);
  523. }
  524. currentNode = starpu_rbtree_next(currentNode);
  525. }
  526. fprintf(stderr, " done.\n");
  527. return 1;
  528. }
  529. /* * * * * * * * * * * * * * * */
  530. /* * * * * * MAIN * * * * * * */
  531. /* * * * * * * * * * * * * * */
  532. static void usage(const char *program)
  533. {
  534. fprintf(stderr,"Usage: %s [--static-workerid] tasks.rec [sched.rec]\n", program);
  535. exit(EXIT_FAILURE);
  536. }
  537. int main(int argc, char **argv)
  538. {
  539. starpu_data_set_default_sequential_consistency_flag(0);
  540. FILE *rec;
  541. char *s;
  542. const char *tasks_rec = NULL;
  543. const char *sched_rec = NULL;
  544. unsigned i;
  545. size_t s_allocated = 128;
  546. unsigned long nread_tasks = 0;
  547. _STARPU_MALLOC(s, s_allocated);
  548. dependson_size = REPLAY_NMAX_DEPENDENCIES; /* Change the value of REPLAY_NMAX_DEPENCIES to modify the number of dependencies */
  549. _STARPU_MALLOC(dependson, dependson_size * sizeof (* dependson));
  550. alloc_mode = 1;
  551. for (i = 1; i < (unsigned) argc; i++)
  552. {
  553. if (!strcmp(argv[i], "--help") || !strcmp(argv[i], "-h"))
  554. {
  555. usage(argv[0]);
  556. }
  557. else if (!strcmp(argv[i], "--static-workerid"))
  558. {
  559. static_workerid = 1;
  560. }
  561. else
  562. {
  563. if (!tasks_rec)
  564. tasks_rec = argv[i];
  565. else if (!sched_rec)
  566. sched_rec = argv[i];
  567. else
  568. usage(argv[0]);
  569. }
  570. }
  571. if (!tasks_rec)
  572. usage(argv[0]);
  573. if (sched_rec)
  574. schedRecInit(sched_rec);
  575. rec = fopen(tasks_rec, "r");
  576. if (!rec)
  577. {
  578. fprintf(stderr,"unable to open file %s: %s\n", tasks_rec, strerror(errno));
  579. exit(EXIT_FAILURE);
  580. }
  581. int ret = starpu_init(NULL);
  582. if (ret == -ENODEV) goto enodev;
  583. /* Read line by line, and on empty line submit the task with the accumulated information */
  584. reset();
  585. double start = starpu_timing_now();
  586. int linenum = 0;
  587. while(1)
  588. {
  589. char *ln;
  590. if (!fgets(s, s_allocated, rec))
  591. {
  592. fprintf(stderr, " done.\n");
  593. int submitted = submit_tasks();
  594. if (submitted == -1)
  595. {
  596. goto enodev;
  597. }
  598. goto eof;
  599. }
  600. while (!(ln = strchr(s, '\n')))
  601. {
  602. /* fprintf(stderr,"buffer size %d too small, doubling it\n", s_allocated); */
  603. _STARPU_REALLOC(s, s_allocated * 2);
  604. if (!fgets(s + s_allocated-1, s_allocated+1, rec))
  605. {
  606. fprintf(stderr, "\n");
  607. int submitted = submit_tasks();
  608. if (submitted == -1)
  609. {
  610. goto enodev;
  611. }
  612. goto eof;
  613. }
  614. s_allocated *= 2;
  615. }
  616. linenum++;
  617. if (ln == s)
  618. {
  619. /* Empty line, do task */
  620. struct task * task;
  621. _STARPU_MALLOC(task, sizeof(*task));
  622. starpu_task_init(&task->task);
  623. task->deps = NULL;
  624. task->submit_order = submitorder;
  625. starpu_rbtree_node_init(&task->node);
  626. starpu_rbtree_insert(&tree, &task->node, diff);
  627. task->jobid = jobid;
  628. task->iteration = iteration;
  629. if (name != NULL)
  630. task->task.name = strdup(name);
  631. task->type = control;
  632. if (control == NormalTask)
  633. {
  634. if (workerid >= 0)
  635. {
  636. task->task.priority = priority;
  637. task->task.cl = &cl;
  638. if (static_workerid)
  639. {
  640. task->task.workerid = workerid;
  641. task->task.execute_on_a_specific_worker = 1;
  642. }
  643. if (alloc_mode)
  644. {
  645. /* Duplicating the handles stored (and registered in the current context) into the task */
  646. ARRAY_DUP(modes_ptr, task->task.modes, nb_parameters);
  647. ARRAY_DUP(modes_ptr, task->task.cl->modes, nb_parameters);
  648. variable_data_register_check(sizes_set, nb_parameters);
  649. ARRAY_DUP(handles_ptr, task->task.handles, nb_parameters);
  650. }
  651. else
  652. {
  653. task->task.dyn_modes = modes_ptr;
  654. _STARPU_MALLOC(task->task.cl->dyn_modes, (sizeof(*task->task.cl->dyn_modes) * nb_parameters));
  655. ARRAY_DUP(modes_ptr, task->task.cl->dyn_modes, nb_parameters);
  656. variable_data_register_check(sizes_set, nb_parameters);
  657. task->task.dyn_handles = handles_ptr;
  658. }
  659. task->task.nbuffers = nb_parameters;
  660. struct perfmodel * realmodel;
  661. HASH_FIND_STR(model_hash, model, realmodel);
  662. if (realmodel == NULL)
  663. {
  664. int len = strlen(model);
  665. _STARPU_CALLOC(realmodel, 1, sizeof(struct perfmodel));
  666. _STARPU_MALLOC(realmodel->model_name, sizeof(char) * (len+1));
  667. realmodel->model_name = strcpy(realmodel->model_name, model);
  668. starpu_perfmodel_init(&realmodel->perfmodel);
  669. int error = starpu_perfmodel_load_symbol(model, &realmodel->perfmodel);
  670. if (!error)
  671. {
  672. HASH_ADD_STR(model_hash, model_name, realmodel);
  673. }
  674. else
  675. {
  676. fprintf(stderr, "[starpu][Warning] Error loading perfmodel symbol %s\n", model);
  677. fprintf(stderr, "[starpu][Warning] Taking only measurements from the given execution, and forcing execution on worker %d\n", workerid);
  678. starpu_perfmodel_unload_model(&realmodel->perfmodel);
  679. free(realmodel->model_name);
  680. free(realmodel);
  681. realmodel = NULL;
  682. }
  683. }
  684. struct starpu_perfmodel_arch *arch = starpu_worker_get_perf_archtype(workerid, 0);
  685. unsigned comb = starpu_perfmodel_arch_comb_add(arch->ndevices, arch->devices);
  686. unsigned narch = starpu_perfmodel_get_narch_combs();
  687. struct task_arg *arg;
  688. _STARPU_MALLOC(arg, sizeof(struct task_arg) + sizeof(double) * narch);
  689. arg->footprint = footprint;
  690. arg->narch = narch;
  691. double * perfTime = arg->perf;
  692. if (realmodel == NULL)
  693. {
  694. /* Erf, do without perfmodel, for execution there */
  695. task->task.workerid = workerid;
  696. task->task.execute_on_a_specific_worker = 1;
  697. for (i = 0; i < narch ; i++)
  698. {
  699. if (i == comb)
  700. perfTime[i] = endTime - startTime;
  701. else
  702. perfTime[i] = NAN;
  703. }
  704. }
  705. else
  706. {
  707. int one = 0;
  708. for (i = 0; i < narch ; i++)
  709. {
  710. arch = starpu_perfmodel_arch_comb_fetch(i);
  711. perfTime[i] = starpu_perfmodel_history_based_expected_perf(&realmodel->perfmodel, arch, footprint);
  712. if (!(perfTime[i] == 0 || isnan(perfTime[i])))
  713. one = 1;
  714. }
  715. if (!one)
  716. {
  717. fprintf(stderr, "We do not have any performance measurement for symbol '%s' for footprint %x, we can not execute this", model, footprint);
  718. exit(EXIT_FAILURE);
  719. }
  720. }
  721. task->task.cl_arg = arg;
  722. task->task.flops = flops;
  723. total_flops += flops;
  724. }
  725. task->task.cl_arg_size = 0;
  726. task->task.tag_id = tag;
  727. task->task.use_tag = 1;
  728. task->ndependson = ndependson;
  729. if (ndependson > 0)
  730. {
  731. _STARPU_MALLOC(task->deps, ndependson * sizeof (* task->deps));
  732. ARRAY_DUP(dependson, task->deps, ndependson);
  733. }
  734. }
  735. else
  736. {
  737. STARPU_ASSERT(nb_parameters == 1);
  738. task->reg_signal = reg_signal[0];
  739. ARRAY_DUP(handles_ptr, task->task.handles, nb_parameters);
  740. }
  741. /* Add this task to task hash */
  742. HASH_ADD(hh, tasks, jobid, sizeof(jobid), task);
  743. nread_tasks++;
  744. if (!(nread_tasks % 1000))
  745. {
  746. fprintf(stderr, "\rRead task %lu...", nread_tasks);
  747. fflush(stdout);
  748. }
  749. reset();
  750. }
  751. /* Record various information */
  752. #define TEST(field) (!strncmp(s, field": ", strlen(field) + 2))
  753. else if(TEST("Control"))
  754. {
  755. char * c = s+9;
  756. if(!strncmp(c, "WontUse", 7))
  757. {
  758. control = WontUseTask;
  759. nb_parameters = 1;
  760. alloc_mode = set_alloc_mode(nb_parameters);
  761. arrays_managing(alloc_mode);
  762. }
  763. else
  764. control = NormalTask;
  765. }
  766. else if (TEST("Name"))
  767. {
  768. *ln = 0;
  769. name = strdup(s+6);
  770. }
  771. else if (TEST("Model"))
  772. {
  773. *ln = 0;
  774. model = strdup(s+7);
  775. }
  776. else if (TEST("JobId"))
  777. jobid = atol(s+7);
  778. else if(TEST("SubmitOrder"))
  779. submitorder = atoi(s+13);
  780. else if (TEST("DependsOn"))
  781. {
  782. char *c = s + 11;
  783. for (ndependson = 0; *c != '\n'; ndependson++)
  784. {
  785. if (ndependson >= dependson_size)
  786. {
  787. dependson_size *= 2;
  788. _STARPU_REALLOC(dependson, dependson_size * sizeof(*dependson));
  789. }
  790. dependson[ndependson] = strtol(c, &c, 10);
  791. }
  792. }
  793. else if (TEST("Tag"))
  794. {
  795. tag = strtol(s+5, NULL, 16);
  796. }
  797. else if (TEST("WorkerId"))
  798. {
  799. workerid = atoi(s+10);
  800. }
  801. else if (TEST("Footprint"))
  802. {
  803. footprint = strtoul(s+11, NULL, 16);
  804. }
  805. else if (TEST("Parameters"))
  806. {
  807. /* Parameters line format is PARAM1 PARAM2 (...)PARAMi (...)PARAMn */
  808. char * param_str = s + 12;
  809. int count = 0;
  810. for (i = 0 ; param_str[i] != '\n'; i++)
  811. {
  812. if (param_str[i] == ' ') /* Checking the number of ' ' (space), assuming that the file is not corrupted */
  813. {
  814. count++;
  815. }
  816. }
  817. nb_parameters = count + 1; /* There is one space per paramater except for the last one, that's why we have to add +1 (dirty programming) */
  818. /* This part of the algorithm will determine if it needs static or dynamic arrays */
  819. alloc_mode = set_alloc_mode(nb_parameters);
  820. arrays_managing(alloc_mode);
  821. }
  822. else if (TEST("Handles"))
  823. {
  824. char *buffer = s + 9;
  825. const char *delim = " ";
  826. char *token = strtok(buffer, delim);
  827. for (i = 0 ; i < nb_parameters ; i++)
  828. {
  829. STARPU_ASSERT(token);
  830. struct handle *handles_cell; /* A cell of the hash table for the handles */
  831. starpu_data_handle_t handle_value = (starpu_data_handle_t) strtol(token, NULL, 16); /* Get the ith handle on the line (in the file) */
  832. HASH_FIND(hh, handles_hash, &handle_value, sizeof(handle_value), handles_cell); /* Find if the handle_value was already registered as a key in the hash table */
  833. /* If it wasn't, then add it to the hash table */
  834. if (handles_cell == NULL)
  835. {
  836. /* Hide the initial handle from the file into the handles array to find it when necessary */
  837. handles_ptr[i] = handle_value;
  838. reg_signal[i] = 1;
  839. }
  840. else
  841. {
  842. handles_ptr[i] = handles_cell->mem_ptr;
  843. reg_signal[i] = 0;
  844. }
  845. token = strtok(NULL, delim);
  846. }
  847. }
  848. else if (TEST("Modes"))
  849. {
  850. char * buffer = s + 7;
  851. unsigned mode_i = 0;
  852. const char * delim = " ";
  853. char * token = strtok(buffer, delim);
  854. while (token != NULL && mode_i < nb_parameters)
  855. {
  856. /* Subject to the names of starpu modes enumerator are not modified */
  857. if (!strncmp(token, "RW", 2))
  858. {
  859. *(modes_ptr+mode_i) = STARPU_RW;
  860. mode_i++;
  861. }
  862. else if (!strncmp(token, "R", 1))
  863. {
  864. *(modes_ptr+mode_i) = STARPU_R;
  865. mode_i++;
  866. }
  867. else if (!strncmp(token, "W", 1))
  868. {
  869. *(modes_ptr+mode_i) = STARPU_W;
  870. mode_i++;
  871. }
  872. /* Other cases produce a warning*/
  873. else
  874. {
  875. fprintf(stderr, "[Warning] A mode is different from R/W (jobid task : %lu)", jobid);
  876. }
  877. token = strtok(NULL, delim);
  878. }
  879. }
  880. else if (TEST("Sizes"))
  881. {
  882. char * buffer = s + 7;
  883. const char * delim = " ";
  884. char * token = strtok(buffer, delim);
  885. unsigned k = 0;
  886. _STARPU_MALLOC(sizes_set, nb_parameters * sizeof(size_t));
  887. while (token != NULL && k < nb_parameters)
  888. {
  889. sizes_set[k] = strtol(token, NULL, 10);
  890. token = strtok(NULL, delim);
  891. k++;
  892. }
  893. }
  894. else if (TEST("StartTime"))
  895. {
  896. startTime = strtod(s+11, NULL);
  897. }
  898. else if (TEST("EndTime"))
  899. {
  900. endTime = strtod(s+9, NULL);
  901. }
  902. else if (TEST("GFlop"))
  903. {
  904. flops = 1000000000 * strtod(s+7, NULL);
  905. }
  906. else if (TEST("Iteration"))
  907. {
  908. iteration = (unsigned) strtol(s+11, NULL, 10);
  909. }
  910. else if (TEST("Priority"))
  911. {
  912. priority = strtol(s + 10, NULL, 10);
  913. }
  914. }
  915. eof:
  916. starpu_task_wait_for_all();
  917. fprintf(stderr, " done.\n");
  918. printf("%g ms", (starpu_timing_now() - start) / 1000.);
  919. if (total_flops != 0.)
  920. printf("\t%g GF/s", (total_flops / (starpu_timing_now() - start)) / 1000.);
  921. printf("\n");
  922. /* FREE allocated memory */
  923. free(dependson);
  924. free(s);
  925. /* End of FREE */
  926. struct handle *handle=NULL, *handletmp=NULL;
  927. HASH_ITER(hh, handles_hash, handle, handletmp)
  928. {
  929. starpu_data_unregister(handle->mem_ptr);
  930. HASH_DEL(handles_hash, handle);
  931. free(handle);
  932. }
  933. struct perfmodel *model_s=NULL, *modeltmp=NULL;
  934. HASH_ITER(hh, model_hash, model_s, modeltmp)
  935. {
  936. starpu_perfmodel_unload_model(&model_s->perfmodel);
  937. HASH_DEL(model_hash, model_s);
  938. free(model_s->model_name);
  939. free(model_s);
  940. }
  941. struct task *task=NULL, *tasktmp=NULL;
  942. HASH_ITER(hh, tasks, task, tasktmp)
  943. {
  944. free(task->task.cl_arg);
  945. free((char*)task->task.name);
  946. if (task->task.dyn_handles != NULL)
  947. {
  948. free(task->task.dyn_handles);
  949. free(task->task.dyn_modes);
  950. }
  951. HASH_DEL(tasks, task);
  952. starpu_task_clean(&task->task);
  953. free(task->deps);
  954. starpu_rbtree_remove(&tree, &task->node);
  955. free(task);
  956. }
  957. starpu_shutdown();
  958. return 0;
  959. enodev:
  960. starpu_shutdown();
  961. return 77;
  962. }