starpu_replay.c 29 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167
  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. if (currentTask->task.handles[0])
  518. starpu_data_wont_use(currentTask->task.handles[0]);
  519. }
  520. currentNode = starpu_rbtree_next(currentNode);
  521. }
  522. fprintf(stderr, " done.\n");
  523. return 1;
  524. }
  525. /* * * * * * * * * * * * * * * */
  526. /* * * * * * MAIN * * * * * * */
  527. /* * * * * * * * * * * * * * */
  528. static void usage(const char *program)
  529. {
  530. fprintf(stderr,"Usage: %s [--static-workerid] tasks.rec [sched.rec]\n", program);
  531. exit(EXIT_FAILURE);
  532. }
  533. int main(int argc, char **argv)
  534. {
  535. FILE *rec;
  536. char *s;
  537. const char *tasks_rec = NULL;
  538. const char *sched_rec = NULL;
  539. unsigned i;
  540. size_t s_allocated = 128;
  541. unsigned long nread_tasks = 0;
  542. /* FIXME: we do not support data with sequential consistency disabled */
  543. _STARPU_MALLOC(s, s_allocated);
  544. dependson_size = REPLAY_NMAX_DEPENDENCIES; /* Change the value of REPLAY_NMAX_DEPENCIES to modify the number of dependencies */
  545. _STARPU_MALLOC(dependson, dependson_size * sizeof (* dependson));
  546. alloc_mode = 1;
  547. for (i = 1; i < (unsigned) argc; i++)
  548. {
  549. if (!strcmp(argv[i], "--help") || !strcmp(argv[i], "-h"))
  550. {
  551. usage(argv[0]);
  552. }
  553. else if (!strcmp(argv[i], "--static-workerid"))
  554. {
  555. static_workerid = 1;
  556. }
  557. else
  558. {
  559. if (!tasks_rec)
  560. tasks_rec = argv[i];
  561. else if (!sched_rec)
  562. sched_rec = argv[i];
  563. else
  564. usage(argv[0]);
  565. }
  566. }
  567. if (!tasks_rec)
  568. usage(argv[0]);
  569. if (sched_rec)
  570. schedRecInit(sched_rec);
  571. rec = fopen(tasks_rec, "r");
  572. if (!rec)
  573. {
  574. fprintf(stderr,"unable to open file %s: %s\n", tasks_rec, strerror(errno));
  575. exit(EXIT_FAILURE);
  576. }
  577. int ret = starpu_init(NULL);
  578. if (ret == -ENODEV) goto enodev;
  579. /* Read line by line, and on empty line submit the task with the accumulated information */
  580. reset();
  581. double start = starpu_timing_now();
  582. int linenum = 0;
  583. while(1)
  584. {
  585. char *ln;
  586. if (!fgets(s, s_allocated, rec))
  587. {
  588. fprintf(stderr, " done.\n");
  589. int submitted = submit_tasks();
  590. if (submitted == -1)
  591. {
  592. goto enodev;
  593. }
  594. goto eof;
  595. }
  596. while (!(ln = strchr(s, '\n')))
  597. {
  598. /* fprintf(stderr,"buffer size %d too small, doubling it\n", s_allocated); */
  599. _STARPU_REALLOC(s, s_allocated * 2);
  600. if (!fgets(s + s_allocated-1, s_allocated+1, rec))
  601. {
  602. fprintf(stderr, "\n");
  603. int submitted = submit_tasks();
  604. if (submitted == -1)
  605. {
  606. goto enodev;
  607. }
  608. goto eof;
  609. }
  610. s_allocated *= 2;
  611. }
  612. linenum++;
  613. if (ln == s)
  614. {
  615. /* Empty line, do task */
  616. struct task * task;
  617. _STARPU_MALLOC(task, sizeof(*task));
  618. starpu_task_init(&task->task);
  619. task->deps = NULL;
  620. task->submit_order = submitorder;
  621. starpu_rbtree_node_init(&task->node);
  622. starpu_rbtree_insert(&tree, &task->node, diff);
  623. task->jobid = jobid;
  624. task->iteration = iteration;
  625. if (name != NULL)
  626. task->task.name = strdup(name);
  627. task->type = control;
  628. if (control == NormalTask)
  629. {
  630. if (workerid >= 0)
  631. {
  632. task->task.priority = priority;
  633. task->task.cl = &cl;
  634. if (static_workerid)
  635. {
  636. task->task.workerid = workerid;
  637. task->task.execute_on_a_specific_worker = 1;
  638. }
  639. if (alloc_mode)
  640. {
  641. /* Duplicating the handles stored (and registered in the current context) into the task */
  642. ARRAY_DUP(modes_ptr, task->task.modes, nb_parameters);
  643. ARRAY_DUP(modes_ptr, task->task.cl->modes, nb_parameters);
  644. variable_data_register_check(sizes_set, nb_parameters);
  645. ARRAY_DUP(handles_ptr, task->task.handles, nb_parameters);
  646. }
  647. else
  648. {
  649. task->task.dyn_modes = modes_ptr;
  650. _STARPU_MALLOC(task->task.cl->dyn_modes, (sizeof(*task->task.cl->dyn_modes) * nb_parameters));
  651. ARRAY_DUP(modes_ptr, task->task.cl->dyn_modes, nb_parameters);
  652. variable_data_register_check(sizes_set, nb_parameters);
  653. task->task.dyn_handles = handles_ptr;
  654. }
  655. task->task.nbuffers = nb_parameters;
  656. struct perfmodel * realmodel;
  657. HASH_FIND_STR(model_hash, model, realmodel);
  658. if (realmodel == NULL)
  659. {
  660. int len = strlen(model);
  661. _STARPU_CALLOC(realmodel, 1, sizeof(struct perfmodel));
  662. _STARPU_MALLOC(realmodel->model_name, sizeof(char) * (len+1));
  663. realmodel->model_name = strcpy(realmodel->model_name, model);
  664. starpu_perfmodel_init(&realmodel->perfmodel);
  665. int error = starpu_perfmodel_load_symbol(model, &realmodel->perfmodel);
  666. if (!error)
  667. {
  668. HASH_ADD_STR(model_hash, model_name, realmodel);
  669. }
  670. else
  671. {
  672. fprintf(stderr, "[starpu][Warning] Error loading perfmodel symbol %s\n", model);
  673. fprintf(stderr, "[starpu][Warning] Taking only measurements from the given execution, and forcing execution on worker %d\n", workerid);
  674. starpu_perfmodel_unload_model(&realmodel->perfmodel);
  675. free(realmodel->model_name);
  676. free(realmodel);
  677. realmodel = NULL;
  678. }
  679. }
  680. struct starpu_perfmodel_arch *arch = starpu_worker_get_perf_archtype(workerid, 0);
  681. unsigned comb = starpu_perfmodel_arch_comb_add(arch->ndevices, arch->devices);
  682. unsigned narch = starpu_perfmodel_get_narch_combs();
  683. struct task_arg *arg;
  684. _STARPU_MALLOC(arg, sizeof(struct task_arg) + sizeof(double) * narch);
  685. arg->footprint = footprint;
  686. arg->narch = narch;
  687. double * perfTime = arg->perf;
  688. if (realmodel == NULL)
  689. {
  690. /* Erf, do without perfmodel, for execution there */
  691. task->task.workerid = workerid;
  692. task->task.execute_on_a_specific_worker = 1;
  693. for (i = 0; i < narch ; i++)
  694. {
  695. if (i == comb)
  696. perfTime[i] = endTime - startTime;
  697. else
  698. perfTime[i] = NAN;
  699. }
  700. }
  701. else
  702. {
  703. int one = 0;
  704. for (i = 0; i < narch ; i++)
  705. {
  706. arch = starpu_perfmodel_arch_comb_fetch(i);
  707. perfTime[i] = starpu_perfmodel_history_based_expected_perf(&realmodel->perfmodel, arch, footprint);
  708. if (!(perfTime[i] == 0 || isnan(perfTime[i])))
  709. one = 1;
  710. }
  711. if (!one)
  712. {
  713. fprintf(stderr, "We do not have any performance measurement for symbol '%s' for footprint %x, we can not execute this", model, footprint);
  714. exit(EXIT_FAILURE);
  715. }
  716. }
  717. task->task.cl_arg = arg;
  718. task->task.flops = flops;
  719. total_flops += flops;
  720. }
  721. task->task.cl_arg_size = 0;
  722. task->task.tag_id = tag;
  723. task->task.use_tag = 1;
  724. task->ndependson = ndependson;
  725. if (ndependson > 0)
  726. {
  727. _STARPU_MALLOC(task->deps, ndependson * sizeof (* task->deps));
  728. ARRAY_DUP(dependson, task->deps, ndependson);
  729. }
  730. }
  731. else
  732. {
  733. STARPU_ASSERT(nb_parameters == 1);
  734. task->reg_signal = reg_signal[0];
  735. ARRAY_DUP(handles_ptr, task->task.handles, nb_parameters);
  736. }
  737. /* Add this task to task hash */
  738. HASH_ADD(hh, tasks, jobid, sizeof(jobid), task);
  739. nread_tasks++;
  740. if (!(nread_tasks % 1000))
  741. {
  742. fprintf(stderr, "\rRead task %lu...", nread_tasks);
  743. fflush(stdout);
  744. }
  745. reset();
  746. }
  747. /* Record various information */
  748. #define TEST(field) (!strncmp(s, field": ", strlen(field) + 2))
  749. else if(TEST("Control"))
  750. {
  751. char * c = s+9;
  752. if(!strncmp(c, "WontUse", 7))
  753. {
  754. control = WontUseTask;
  755. nb_parameters = 1;
  756. alloc_mode = set_alloc_mode(nb_parameters);
  757. arrays_managing(alloc_mode);
  758. }
  759. else
  760. control = NormalTask;
  761. }
  762. else if (TEST("Name"))
  763. {
  764. *ln = 0;
  765. name = strdup(s+6);
  766. }
  767. else if (TEST("Model"))
  768. {
  769. *ln = 0;
  770. model = strdup(s+7);
  771. }
  772. else if (TEST("JobId"))
  773. jobid = atol(s+7);
  774. else if(TEST("SubmitOrder"))
  775. submitorder = atoi(s+13);
  776. else if (TEST("DependsOn"))
  777. {
  778. char *c = s + 11;
  779. for (ndependson = 0; *c != '\n'; ndependson++)
  780. {
  781. if (ndependson >= dependson_size)
  782. {
  783. dependson_size *= 2;
  784. _STARPU_REALLOC(dependson, dependson_size * sizeof(*dependson));
  785. }
  786. dependson[ndependson] = strtol(c, &c, 10);
  787. }
  788. }
  789. else if (TEST("Tag"))
  790. {
  791. tag = strtol(s+5, NULL, 16);
  792. }
  793. else if (TEST("WorkerId"))
  794. {
  795. workerid = atoi(s+10);
  796. }
  797. else if (TEST("Footprint"))
  798. {
  799. footprint = strtoul(s+11, NULL, 16);
  800. }
  801. else if (TEST("Parameters"))
  802. {
  803. /* Parameters line format is PARAM1 PARAM2 (...)PARAMi (...)PARAMn */
  804. char * param_str = s + 12;
  805. int count = 0;
  806. for (i = 0 ; param_str[i] != '\n'; i++)
  807. {
  808. if (param_str[i] == ' ') /* Checking the number of ' ' (space), assuming that the file is not corrupted */
  809. {
  810. count++;
  811. }
  812. }
  813. 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) */
  814. /* This part of the algorithm will determine if it needs static or dynamic arrays */
  815. alloc_mode = set_alloc_mode(nb_parameters);
  816. arrays_managing(alloc_mode);
  817. }
  818. else if (TEST("Handles"))
  819. {
  820. char *buffer = s + 9;
  821. const char *delim = " ";
  822. char *token = strtok(buffer, delim);
  823. for (i = 0 ; i < nb_parameters ; i++)
  824. {
  825. STARPU_ASSERT(token);
  826. struct handle *handles_cell; /* A cell of the hash table for the handles */
  827. starpu_data_handle_t handle_value = (starpu_data_handle_t) strtol(token, NULL, 16); /* Get the ith handle on the line (in the file) */
  828. 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 */
  829. /* If it wasn't, then add it to the hash table */
  830. if (handles_cell == NULL)
  831. {
  832. /* Hide the initial handle from the file into the handles array to find it when necessary */
  833. handles_ptr[i] = handle_value;
  834. reg_signal[i] = 1;
  835. }
  836. else
  837. {
  838. handles_ptr[i] = handles_cell->mem_ptr;
  839. reg_signal[i] = 0;
  840. }
  841. token = strtok(NULL, delim);
  842. }
  843. }
  844. else if (TEST("Modes"))
  845. {
  846. char * buffer = s + 7;
  847. unsigned mode_i = 0;
  848. const char * delim = " ";
  849. char * token = strtok(buffer, delim);
  850. while (token != NULL && mode_i < nb_parameters)
  851. {
  852. /* Subject to the names of starpu modes enumerator are not modified */
  853. if (!strncmp(token, "RW", 2))
  854. {
  855. *(modes_ptr+mode_i) = STARPU_RW;
  856. mode_i++;
  857. }
  858. else if (!strncmp(token, "R", 1))
  859. {
  860. *(modes_ptr+mode_i) = STARPU_R;
  861. mode_i++;
  862. }
  863. else if (!strncmp(token, "W", 1))
  864. {
  865. *(modes_ptr+mode_i) = STARPU_W;
  866. mode_i++;
  867. }
  868. /* Other cases produce a warning*/
  869. else
  870. {
  871. fprintf(stderr, "[Warning] A mode is different from R/W (jobid task : %lu)", jobid);
  872. }
  873. token = strtok(NULL, delim);
  874. }
  875. }
  876. else if (TEST("Sizes"))
  877. {
  878. char * buffer = s + 7;
  879. const char * delim = " ";
  880. char * token = strtok(buffer, delim);
  881. unsigned k = 0;
  882. _STARPU_MALLOC(sizes_set, nb_parameters * sizeof(size_t));
  883. while (token != NULL && k < nb_parameters)
  884. {
  885. sizes_set[k] = strtol(token, NULL, 10);
  886. token = strtok(NULL, delim);
  887. k++;
  888. }
  889. }
  890. else if (TEST("StartTime"))
  891. {
  892. startTime = strtod(s+11, NULL);
  893. }
  894. else if (TEST("EndTime"))
  895. {
  896. endTime = strtod(s+9, NULL);
  897. }
  898. else if (TEST("GFlop"))
  899. {
  900. flops = 1000000000 * strtod(s+7, NULL);
  901. }
  902. else if (TEST("Iteration"))
  903. {
  904. iteration = (unsigned) strtol(s+11, NULL, 10);
  905. }
  906. else if (TEST("Priority"))
  907. {
  908. priority = strtol(s + 10, NULL, 10);
  909. }
  910. }
  911. eof:
  912. starpu_task_wait_for_all();
  913. fprintf(stderr, " done.\n");
  914. printf("%g ms", (starpu_timing_now() - start) / 1000.);
  915. if (total_flops != 0.)
  916. printf("\t%g GF/s", (total_flops / (starpu_timing_now() - start)) / 1000.);
  917. printf("\n");
  918. /* FREE allocated memory */
  919. free(dependson);
  920. free(s);
  921. /* End of FREE */
  922. struct handle *handle=NULL, *handletmp=NULL;
  923. HASH_ITER(hh, handles_hash, handle, handletmp)
  924. {
  925. starpu_data_unregister(handle->mem_ptr);
  926. HASH_DEL(handles_hash, handle);
  927. free(handle);
  928. }
  929. struct perfmodel *model_s=NULL, *modeltmp=NULL;
  930. HASH_ITER(hh, model_hash, model_s, modeltmp)
  931. {
  932. starpu_perfmodel_unload_model(&model_s->perfmodel);
  933. HASH_DEL(model_hash, model_s);
  934. free(model_s->model_name);
  935. free(model_s);
  936. }
  937. struct task *task=NULL, *tasktmp=NULL;
  938. HASH_ITER(hh, tasks, task, tasktmp)
  939. {
  940. free(task->task.cl_arg);
  941. free((char*)task->task.name);
  942. if (task->task.dyn_handles != NULL)
  943. {
  944. free(task->task.dyn_handles);
  945. free(task->task.dyn_modes);
  946. }
  947. HASH_DEL(tasks, task);
  948. starpu_task_clean(&task->task);
  949. free(task->deps);
  950. starpu_rbtree_remove(&tree, &task->node);
  951. free(task);
  952. }
  953. starpu_shutdown();
  954. return 0;
  955. enodev:
  956. starpu_shutdown();
  957. return 77;
  958. }