starpu_mpi_insert_task.c 20 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468
  1. /* StarPU --- Runtime system for heterogeneous multicore architectures.
  2. *
  3. * Copyright (C) 2011 Centre National de la Recherche Scientifique
  4. * Copyright (C) 2011 Université de Bordeaux 1
  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. #include <stdarg.h>
  18. #include <mpi.h>
  19. #include <starpu.h>
  20. #include <starpu_data.h>
  21. #include <common/utils.h>
  22. #include <common/hash.h>
  23. #include <common/htable32.h>
  24. #include <util/starpu_insert_task_utils.h>
  25. //#define STARPU_MPI_VERBOSE 1
  26. #include <starpu_mpi_private.h>
  27. /* Whether we are allowed to keep copies of remote data. Does not work
  28. * yet: the sender has to know whether the receiver has it, keeping it
  29. * in an array indexed by node numbers. */
  30. #define MPI_CACHE
  31. #ifdef MPI_CACHE
  32. static struct starpu_htbl32_node_s **sent_data = NULL;
  33. static struct starpu_htbl32_node_s **received_data = NULL;
  34. static void _starpu_mpi_task_init(int nb_nodes)
  35. {
  36. int i;
  37. _STARPU_MPI_DEBUG("Initialising hash table for cache\n");
  38. sent_data = malloc(nb_nodes * sizeof(struct starpu_htbl32_node_s *));
  39. for(i=0 ; i<nb_nodes ; i++) sent_data[i] = NULL;
  40. received_data = malloc(nb_nodes * sizeof(struct starpu_htbl32_node_s *));
  41. for(i=0 ; i<nb_nodes ; i++) received_data[i] = NULL;
  42. }
  43. typedef struct _starpu_mpi_clear_cache_s {
  44. starpu_data_handle data;
  45. int rank;
  46. int mode;
  47. } _starpu_mpi_clear_cache_t;
  48. #define _STARPU_MPI_CLEAR_SENT_DATA 0
  49. #define _STARPU_MPI_CLEAR_RECEIVED_DATA 1
  50. void _starpu_mpi_clear_cache_callback(void *callback_arg)
  51. {
  52. _starpu_mpi_clear_cache_t *clear_cache = (_starpu_mpi_clear_cache_t *)callback_arg;
  53. uint32_t key = _starpu_crc32_be((uintptr_t)clear_cache->data, 0);
  54. if (clear_cache->mode == _STARPU_MPI_CLEAR_SENT_DATA) {
  55. _STARPU_MPI_DEBUG("Clearing sent cache for data %p and rank %d\n", clear_cache->data, clear_cache->rank);
  56. _starpu_htbl_insert_32(&sent_data[clear_cache->rank], key, NULL);
  57. }
  58. else if (clear_cache->mode == _STARPU_MPI_CLEAR_RECEIVED_DATA) {
  59. _STARPU_MPI_DEBUG("Clearing received cache for data %p and rank %d\n", clear_cache->data, clear_cache->rank);
  60. _starpu_htbl_insert_32(&received_data[clear_cache->rank], key, NULL);
  61. }
  62. free(clear_cache);
  63. }
  64. void _starpu_mpi_clear_cache_request(starpu_data_handle data_handle, int rank, int mode)
  65. {
  66. struct starpu_task *task = starpu_task_create();
  67. task->cl = NULL;
  68. task->buffers[0].handle = data_handle;
  69. task->buffers[0].mode = STARPU_RW;
  70. _starpu_mpi_clear_cache_t *clear_cache = malloc(sizeof(_starpu_mpi_clear_cache_t));
  71. clear_cache->data = data_handle;
  72. clear_cache->rank = rank;
  73. clear_cache->mode = mode;
  74. task->callback_func = _starpu_mpi_clear_cache_callback;
  75. task->callback_arg = clear_cache;
  76. starpu_task_submit(task);
  77. }
  78. #endif
  79. void _starpu_data_deallocate(starpu_data_handle data_handle)
  80. {
  81. #ifdef STARPU_DEVEL
  82. #warning _starpu_data_deallocate not implemented yet
  83. #endif
  84. }
  85. int starpu_mpi_insert_task(MPI_Comm comm, starpu_codelet *codelet, ...)
  86. {
  87. int arg_type;
  88. va_list varg_list;
  89. int me, do_execute;
  90. size_t arg_buffer_size = 0;
  91. char *arg_buffer;
  92. int dest=0, execute, inconsistent_execute;
  93. _STARPU_MPI_LOG_IN();
  94. MPI_Comm_rank(comm, &me);
  95. #ifdef MPI_CACHE
  96. if (sent_data == NULL) {
  97. int size;
  98. MPI_Comm_size(comm, &size);
  99. _starpu_mpi_task_init(size);
  100. }
  101. #endif
  102. /* Get the number of buffers and the size of the arguments */
  103. va_start(varg_list, codelet);
  104. arg_buffer_size = _starpu_insert_task_get_arg_size(varg_list);
  105. va_start(varg_list, codelet);
  106. _starpu_pack_cl_args(arg_buffer_size, &arg_buffer, varg_list);
  107. /* Finds out if the property STARPU_EXECUTE_ON_NODE or STARPU_EXECUTE_ON_DATA is specified */
  108. execute = -1;
  109. va_start(varg_list, codelet);
  110. while ((arg_type = va_arg(varg_list, int)) != 0) {
  111. if (arg_type==STARPU_EXECUTE_ON_NODE) {
  112. execute = va_arg(varg_list, int);
  113. }
  114. else if (arg_type==STARPU_EXECUTE_ON_DATA) {
  115. starpu_data_handle data = va_arg(varg_list, starpu_data_handle);
  116. execute = starpu_data_get_rank(data);
  117. }
  118. else if (arg_type==STARPU_R || arg_type==STARPU_W || arg_type==STARPU_RW || arg_type == STARPU_SCRATCH) {
  119. va_arg(varg_list, starpu_data_handle);
  120. }
  121. else if (arg_type==STARPU_VALUE) {
  122. va_arg(varg_list, void *);
  123. va_arg(varg_list, size_t);
  124. }
  125. else if (arg_type==STARPU_CALLBACK) {
  126. va_arg(varg_list, void (*)(void *));
  127. }
  128. else if (arg_type==STARPU_CALLBACK_WITH_ARG) {
  129. va_arg(varg_list, void (*)(void *));
  130. va_arg(varg_list, void *);
  131. }
  132. else if (arg_type==STARPU_CALLBACK_ARG) {
  133. va_arg(varg_list, void *);
  134. }
  135. else if (arg_type==STARPU_PRIORITY) {
  136. va_arg(varg_list, int);
  137. }
  138. }
  139. va_end(varg_list);
  140. /* Find out whether we are to execute the data because we own the data to be written to. */
  141. inconsistent_execute = 0;
  142. do_execute = -1;
  143. va_start(varg_list, codelet);
  144. while ((arg_type = va_arg(varg_list, int)) != 0) {
  145. if (arg_type==STARPU_R || arg_type==STARPU_W || arg_type==STARPU_RW || arg_type == STARPU_SCRATCH) {
  146. starpu_data_handle data = va_arg(varg_list, starpu_data_handle);
  147. if (arg_type & STARPU_W) {
  148. if (!data) {
  149. /* We don't have anything allocated for this.
  150. * The application knows we won't do anything
  151. * about this task */
  152. /* Yes, the app could actually not call
  153. * insert_task at all itself, this is just a
  154. * safeguard. */
  155. _STARPU_MPI_DEBUG("oh oh\n");
  156. _STARPU_MPI_LOG_OUT();
  157. return -EINVAL;
  158. }
  159. int mpi_rank = starpu_data_get_rank(data);
  160. if (mpi_rank == me) {
  161. if (do_execute == 0) {
  162. inconsistent_execute = 1;
  163. }
  164. else {
  165. do_execute = 1;
  166. }
  167. }
  168. else if (mpi_rank != -1) {
  169. if (do_execute == 1) {
  170. inconsistent_execute = 1;
  171. }
  172. else {
  173. do_execute = 0;
  174. dest = mpi_rank;
  175. /* That's the rank which needs the data to be sent to */
  176. }
  177. }
  178. else {
  179. _STARPU_ERROR("rank invalid\n");
  180. }
  181. }
  182. }
  183. else if (arg_type==STARPU_VALUE) {
  184. va_arg(varg_list, void *);
  185. va_arg(varg_list, size_t);
  186. }
  187. else if (arg_type==STARPU_CALLBACK) {
  188. va_arg(varg_list, void (*)(void *));
  189. }
  190. else if (arg_type==STARPU_CALLBACK_WITH_ARG) {
  191. va_arg(varg_list, void (*)(void *));
  192. va_arg(varg_list, void *);
  193. }
  194. else if (arg_type==STARPU_CALLBACK_ARG) {
  195. va_arg(varg_list, void *);
  196. }
  197. else if (arg_type==STARPU_PRIORITY) {
  198. va_arg(varg_list, int);
  199. }
  200. else if (arg_type==STARPU_EXECUTE_ON_NODE) {
  201. va_arg(varg_list, int);
  202. }
  203. else if (arg_type==STARPU_EXECUTE_ON_DATA) {
  204. va_arg(varg_list, int);
  205. }
  206. }
  207. va_end(varg_list);
  208. assert(do_execute != -1 && "StarPU needs to see a W data which will tell it where to execute the task");
  209. if (inconsistent_execute == 1) {
  210. if (execute == -1) {
  211. _STARPU_ERROR("Different tasks are owning W data. Needs to specify which one is to execute the codelet, using STARPU_EXECUTE_ON_NODE or STARPU_EXECUTE_ON_DATA\n");
  212. }
  213. else {
  214. do_execute = (me == execute);
  215. dest = execute;
  216. }
  217. }
  218. else if (execute != -1) {
  219. _STARPU_MPI_DEBUG("Property STARPU_EXECUTE_ON_NODE or STARPU_EXECUTE_ON_DATA ignored as W data are all owned by the same task\n");
  220. }
  221. /* Send and receive data as requested */
  222. va_start(varg_list, codelet);
  223. while ((arg_type = va_arg(varg_list, int)) != 0) {
  224. if (arg_type==STARPU_R || arg_type==STARPU_W || arg_type==STARPU_RW || arg_type == STARPU_SCRATCH) {
  225. starpu_data_handle data = va_arg(varg_list, starpu_data_handle);
  226. if (data && arg_type & STARPU_R) {
  227. int mpi_rank = starpu_data_get_rank(data);
  228. int mpi_tag = starpu_data_get_tag(data);
  229. STARPU_ASSERT(mpi_tag >= 0 && "StarPU needs to be told the MPI rank of this data, using starpu_data_set_rank");
  230. /* The task needs to read this data */
  231. if (do_execute && mpi_rank != me && mpi_rank != -1) {
  232. /* I will have to execute but I don't have the data, receive */
  233. #ifdef MPI_CACHE
  234. uint32_t key = _starpu_crc32_be((uintptr_t)data, 0);
  235. void *already_received = _starpu_htbl_search_32(received_data[mpi_rank], key);
  236. if (!already_received) {
  237. _starpu_htbl_insert_32(&received_data[mpi_rank], key, data);
  238. }
  239. else {
  240. _STARPU_MPI_DEBUG("Do not receive data %p from node %d as it is already available\n", data, mpi_rank);
  241. }
  242. if (!already_received)
  243. #endif
  244. {
  245. _STARPU_MPI_DEBUG("Receive data %p from %d\n", data, mpi_rank);
  246. starpu_mpi_irecv_detached(data, mpi_rank, mpi_tag, comm, NULL, NULL);
  247. }
  248. }
  249. if (!do_execute && mpi_rank == me) {
  250. /* Somebody else will execute it, and I have the data, send it. */
  251. #ifdef MPI_CACHE
  252. uint32_t key = _starpu_crc32_be((uintptr_t)data, 0);
  253. void *already_sent = _starpu_htbl_search_32(sent_data[dest], key);
  254. if (!already_sent) {
  255. _starpu_htbl_insert_32(&sent_data[dest], key, data);
  256. }
  257. else {
  258. _STARPU_MPI_DEBUG("Do not sent data %p to node %d as it has already been sent\n", data, dest);
  259. }
  260. if (!already_sent)
  261. #endif
  262. {
  263. _STARPU_MPI_DEBUG("Send data %p to %d\n", data, dest);
  264. starpu_mpi_isend_detached(data, dest, mpi_tag, comm, NULL, NULL);
  265. }
  266. }
  267. }
  268. }
  269. else if (arg_type==STARPU_VALUE) {
  270. va_arg(varg_list, void *);
  271. va_arg(varg_list, size_t);
  272. }
  273. else if (arg_type==STARPU_CALLBACK) {
  274. va_arg(varg_list, void (*)(void *));
  275. }
  276. else if (arg_type==STARPU_CALLBACK_WITH_ARG) {
  277. va_arg(varg_list, void (*)(void *));
  278. va_arg(varg_list, void *);
  279. }
  280. else if (arg_type==STARPU_CALLBACK_ARG) {
  281. va_arg(varg_list, void *);
  282. }
  283. else if (arg_type==STARPU_PRIORITY) {
  284. va_arg(varg_list, int);
  285. }
  286. else if (arg_type==STARPU_EXECUTE_ON_NODE) {
  287. va_arg(varg_list, int);
  288. }
  289. else if (arg_type==STARPU_EXECUTE_ON_DATA) {
  290. va_arg(varg_list, starpu_data_handle);
  291. }
  292. }
  293. va_end(varg_list);
  294. if (do_execute) {
  295. _STARPU_MPI_DEBUG("Execution of the codelet %p\n", codelet);
  296. va_start(varg_list, codelet);
  297. struct starpu_task *task = starpu_task_create();
  298. int ret = _starpu_insert_task_create_and_submit(arg_buffer, codelet, &task, varg_list);
  299. _STARPU_MPI_DEBUG("ret: %d\n", ret);
  300. STARPU_ASSERT(ret==0);
  301. }
  302. if (inconsistent_execute) {
  303. va_start(varg_list, codelet);
  304. while ((arg_type = va_arg(varg_list, int)) != 0) {
  305. if (arg_type==STARPU_R || arg_type==STARPU_W || arg_type==STARPU_RW || arg_type == STARPU_SCRATCH) {
  306. starpu_data_handle data = va_arg(varg_list, starpu_data_handle);
  307. if (arg_type & STARPU_W) {
  308. int mpi_rank = starpu_data_get_rank(data);
  309. int mpi_tag = starpu_data_get_tag(data);
  310. STARPU_ASSERT(mpi_tag >= 0 && "StarPU needs to be told the MPI rank of this data, using starpu_data_set_rank");
  311. if (mpi_rank == me) {
  312. if (execute != -1 && me != execute) {
  313. _STARPU_MPI_DEBUG("Receive data %p back from the task %d which executed the codelet ...\n", data, dest);
  314. starpu_mpi_irecv_detached(data, dest, mpi_tag, comm, NULL, NULL);
  315. }
  316. }
  317. else if (do_execute) {
  318. _STARPU_MPI_DEBUG("Send data %p back to its owner %d...\n", data, mpi_rank);
  319. starpu_mpi_isend_detached(data, mpi_rank, mpi_tag, comm, NULL, NULL);
  320. }
  321. }
  322. }
  323. else if (arg_type==STARPU_VALUE) {
  324. va_arg(varg_list, void *);
  325. va_arg(varg_list, size_t);
  326. }
  327. else if (arg_type==STARPU_CALLBACK) {
  328. va_arg(varg_list, void (*)(void *));
  329. }
  330. else if (arg_type==STARPU_CALLBACK_WITH_ARG) {
  331. va_arg(varg_list, void (*)(void *));
  332. va_arg(varg_list, void *);
  333. }
  334. else if (arg_type==STARPU_CALLBACK_ARG) {
  335. va_arg(varg_list, void *);
  336. }
  337. else if (arg_type==STARPU_PRIORITY) {
  338. va_arg(varg_list, int);
  339. }
  340. else if (arg_type==STARPU_EXECUTE_ON_NODE) {
  341. va_arg(varg_list, int);
  342. }
  343. else if (arg_type==STARPU_EXECUTE_ON_DATA) {
  344. va_arg(varg_list, starpu_data_handle);
  345. }
  346. }
  347. va_end(varg_list);
  348. }
  349. va_start(varg_list, codelet);
  350. while ((arg_type = va_arg(varg_list, int)) != 0) {
  351. if (arg_type==STARPU_R || arg_type==STARPU_W || arg_type==STARPU_RW || arg_type == STARPU_SCRATCH) {
  352. starpu_data_handle data = va_arg(varg_list, starpu_data_handle);
  353. #ifdef MPI_CACHE
  354. if (arg_type & STARPU_W) {
  355. uint32_t key = _starpu_crc32_be((uintptr_t)data, 0);
  356. if (do_execute) {
  357. /* Note that all copies I've sent to neighbours are now invalid */
  358. int n, size;
  359. MPI_Comm_size(comm, &size);
  360. for(n=0 ; n<size ; n++) {
  361. void *already_sent = _starpu_htbl_search_32(sent_data[n], key);
  362. if (already_sent) {
  363. _STARPU_MPI_DEBUG("Posting request to clear send cache for data %p\n", data);
  364. _starpu_mpi_clear_cache_request(data, n, _STARPU_MPI_CLEAR_SENT_DATA);
  365. }
  366. }
  367. }
  368. else {
  369. int mpi_rank = starpu_data_get_rank(data);
  370. void *already_received = _starpu_htbl_search_32(received_data[mpi_rank], key);
  371. if (already_received) {
  372. /* Somebody else will write to the data, so discard our cached copy if any */
  373. /* TODO: starpu_mpi could just remember itself. */
  374. _STARPU_MPI_DEBUG("Posting request to clear receive cache for data %p\n", data);
  375. _starpu_mpi_clear_cache_request(data, mpi_rank, _STARPU_MPI_CLEAR_RECEIVED_DATA);
  376. _starpu_data_deallocate(data);
  377. }
  378. }
  379. }
  380. #else
  381. /* We allocated a temporary buffer for the received data, now drop it */
  382. if ((arg_type & STARPU_R) && do_execute) {
  383. int mpi_rank = starpu_data_get_rank(data);
  384. if (mpi_rank != me && mpi_rank != -1) {
  385. _starpu_data_deallocate(data);
  386. }
  387. }
  388. #endif
  389. }
  390. else if (arg_type==STARPU_VALUE) {
  391. va_arg(varg_list, void *);
  392. va_arg(varg_list, size_t);
  393. }
  394. else if (arg_type==STARPU_CALLBACK) {
  395. va_arg(varg_list, void (*)(void *));
  396. }
  397. else if (arg_type==STARPU_CALLBACK_WITH_ARG) {
  398. va_arg(varg_list, void (*)(void *));
  399. va_arg(varg_list, void *);
  400. }
  401. else if (arg_type==STARPU_CALLBACK_ARG) {
  402. va_arg(varg_list, void *);
  403. }
  404. else if (arg_type==STARPU_PRIORITY) {
  405. va_arg(varg_list, int);
  406. }
  407. else if (arg_type==STARPU_EXECUTE_ON_NODE) {
  408. va_arg(varg_list, int);
  409. }
  410. else if (arg_type==STARPU_EXECUTE_ON_DATA) {
  411. va_arg(varg_list, starpu_data_handle);
  412. }
  413. }
  414. va_end(varg_list);
  415. _STARPU_MPI_LOG_OUT();
  416. return 0;
  417. }
  418. void starpu_mpi_get_data_on_node(MPI_Comm comm, starpu_data_handle data_handle, int node)
  419. {
  420. int me, rank;
  421. rank = starpu_data_get_rank(data_handle);
  422. MPI_Comm_rank(comm, &me);
  423. if (node == rank) return;
  424. if (me == node)
  425. {
  426. starpu_mpi_irecv_detached(data_handle, rank, 42, comm, NULL, NULL);
  427. }
  428. else if (me == rank)
  429. {
  430. starpu_mpi_isend_detached(data_handle, node, 42, comm, NULL, NULL);
  431. }
  432. starpu_task_wait_for_all();
  433. }