xlu_pivot.c 12 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456
  1. /* StarPU --- Runtime system for heterogeneous multicore architectures.
  2. *
  3. * Copyright (C) 2009-2012 Université de Bordeaux 1
  4. * Copyright (C) 2010, 2011, 2012 Centre National de la Recherche Scientifique
  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 "xlu.h"
  18. #include "xlu_kernels.h"
  19. #define TAG11(k) ((starpu_tag_t)( (1ULL<<60) | (unsigned long long)(k)))
  20. #define TAG12(k,i) ((starpu_tag_t)(((2ULL<<60) | (((unsigned long long)(k))<<32) \
  21. | (unsigned long long)(i))))
  22. #define TAG21(k,j) ((starpu_tag_t)(((3ULL<<60) | (((unsigned long long)(k))<<32) \
  23. | (unsigned long long)(j))))
  24. #define TAG22(k,i,j) ((starpu_tag_t)(((4ULL<<60) | ((unsigned long long)(k)<<32) \
  25. | ((unsigned long long)(i)<<16) \
  26. | (unsigned long long)(j))))
  27. #define PIVOT(k,i) ((starpu_tag_t)(((5ULL<<60) | (((unsigned long long)(k))<<32) \
  28. | (unsigned long long)(i))))
  29. static unsigned no_prio = 0;
  30. /*
  31. * Construct the DAG
  32. */
  33. static struct starpu_task *create_task(starpu_tag_t id)
  34. {
  35. struct starpu_task *task = starpu_task_create();
  36. task->cl_arg = NULL;
  37. task->use_tag = 1;
  38. task->tag_id = id;
  39. return task;
  40. }
  41. static int create_task_pivot(starpu_data_handle_t *dataAp, unsigned nblocks,
  42. struct piv_s *piv_description,
  43. unsigned k, unsigned i,
  44. starpu_data_handle_t (* get_block)(starpu_data_handle_t *, unsigned, unsigned, unsigned))
  45. {
  46. int ret;
  47. struct starpu_task *task = create_task(PIVOT(k, i));
  48. task->cl = &cl_pivot;
  49. /* which sub-data is manipulated ? */
  50. task->handles[0] = get_block(dataAp, nblocks, k, i);
  51. task->cl_arg = &piv_description[k];
  52. /* this is an important task */
  53. if (!no_prio && (i == k+1))
  54. task->priority = STARPU_MAX_PRIO;
  55. /* enforce dependencies ... */
  56. if (k == 0)
  57. {
  58. starpu_tag_declare_deps(PIVOT(k, i), 1, TAG11(k));
  59. }
  60. else
  61. {
  62. if (i > k)
  63. {
  64. starpu_tag_declare_deps(PIVOT(k, i), 2, TAG11(k), TAG22(k-1, i, k));
  65. }
  66. else
  67. {
  68. starpu_tag_t *tags = malloc((nblocks - k)*sizeof(starpu_tag_t));
  69. tags[0] = TAG11(k);
  70. unsigned ind, ind2;
  71. for (ind = k + 1, ind2 = 0; ind < nblocks; ind++, ind2++)
  72. {
  73. tags[1 + ind2] = TAG22(k-1, ind, k);
  74. }
  75. /* perhaps we could do better ... :/ */
  76. starpu_tag_declare_deps_array(PIVOT(k, i), (nblocks-k), tags);
  77. free(tags);
  78. }
  79. }
  80. ret = starpu_task_submit(task);
  81. if (ret != -ENODEV) STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
  82. return ret;
  83. }
  84. static struct starpu_task *create_task_11_pivot(starpu_data_handle_t *dataAp, unsigned nblocks,
  85. unsigned k, struct piv_s *piv_description,
  86. starpu_data_handle_t (* get_block)(starpu_data_handle_t *, unsigned, unsigned, unsigned))
  87. {
  88. struct starpu_task *task = create_task(TAG11(k));
  89. task->cl = &cl11_pivot;
  90. task->cl_arg = &piv_description[k];
  91. /* which sub-data is manipulated ? */
  92. task->handles[0] = get_block(dataAp, nblocks, k, k);
  93. /* this is an important task */
  94. if (!no_prio)
  95. task->priority = STARPU_MAX_PRIO;
  96. /* enforce dependencies ... */
  97. if (k > 0)
  98. {
  99. starpu_tag_declare_deps(TAG11(k), 1, TAG22(k-1, k, k));
  100. }
  101. return task;
  102. }
  103. static int create_task_12(starpu_data_handle_t *dataAp, unsigned nblocks, unsigned k, unsigned j,
  104. starpu_data_handle_t (* get_block)(starpu_data_handle_t *, unsigned, unsigned, unsigned))
  105. {
  106. int ret;
  107. /* printf("task 12 k,i = %d,%d TAG = %llx\n", k,i, TAG12(k,i)); */
  108. struct starpu_task *task = create_task(TAG12(k, j));
  109. task->cl = &cl12;
  110. task->cl_arg = (void *)(task->tag_id);
  111. /* which sub-data is manipulated ? */
  112. task->handles[0] = get_block(dataAp, nblocks, k, k);
  113. task->handles[1] = get_block(dataAp, nblocks, j, k);
  114. if (!no_prio && (j == k+1))
  115. {
  116. task->priority = STARPU_MAX_PRIO;
  117. }
  118. /* enforce dependencies ... */
  119. #if 0
  120. starpu_tag_declare_deps(TAG12(k, i), 1, PIVOT(k, i));
  121. #endif
  122. if (k > 0)
  123. {
  124. starpu_tag_declare_deps(TAG12(k, j), 2, TAG11(k), TAG22(k-1, k, j));
  125. }
  126. else
  127. {
  128. starpu_tag_declare_deps(TAG12(k, j), 1, TAG11(k));
  129. }
  130. ret = starpu_task_submit(task);
  131. if (ret != -ENODEV) STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
  132. return ret;
  133. }
  134. static int create_task_21(starpu_data_handle_t *dataAp, unsigned nblocks, unsigned k, unsigned i,
  135. starpu_data_handle_t (* get_block)(starpu_data_handle_t *, unsigned, unsigned, unsigned))
  136. {
  137. int ret;
  138. struct starpu_task *task = create_task(TAG21(k, i));
  139. task->cl = &cl21;
  140. /* which sub-data is manipulated ? */
  141. task->handles[0] = get_block(dataAp, nblocks, k, k);
  142. task->handles[1] = get_block(dataAp, nblocks, k, i);
  143. if (!no_prio && (i == k+1))
  144. {
  145. task->priority = STARPU_MAX_PRIO;
  146. }
  147. task->cl_arg = (void *)(task->tag_id);
  148. /* enforce dependencies ... */
  149. starpu_tag_declare_deps(TAG21(k, i), 1, PIVOT(k, i));
  150. ret = starpu_task_submit(task);
  151. if (ret != -ENODEV) STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
  152. return ret;
  153. }
  154. static int create_task_22(starpu_data_handle_t *dataAp, unsigned nblocks, unsigned k, unsigned i, unsigned j,
  155. starpu_data_handle_t (* get_block)(starpu_data_handle_t *, unsigned, unsigned, unsigned))
  156. {
  157. int ret;
  158. /* printf("task 22 k,i,j = %d,%d,%d TAG = %llx\n", k,i,j, TAG22(k,i,j)); */
  159. struct starpu_task *task = create_task(TAG22(k, i, j));
  160. task->cl = &cl22;
  161. task->cl_arg = (void *)(task->tag_id);
  162. /* which sub-data is manipulated ? */
  163. task->handles[0] = get_block(dataAp, nblocks, k, i); /* produced by TAG21(k, i) */
  164. task->handles[1] = get_block(dataAp, nblocks, j, k); /* produced by TAG12(k, j) */
  165. task->handles[2] = get_block(dataAp, nblocks, j, i); /* produced by TAG22(k-1, i, j) */
  166. if (!no_prio && (i == k + 1) && (j == k +1) )
  167. {
  168. task->priority = STARPU_MAX_PRIO;
  169. }
  170. /* enforce dependencies ... */
  171. if (k > 0)
  172. {
  173. starpu_tag_declare_deps(TAG22(k, i, j), 3, TAG22(k-1, i, j), TAG12(k, j), TAG21(k, i));
  174. }
  175. else
  176. {
  177. starpu_tag_declare_deps(TAG22(k, i, j), 2, TAG12(k, j), TAG21(k, i));
  178. }
  179. ret = starpu_task_submit(task);
  180. if (ret != -ENODEV) STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
  181. return ret;
  182. }
  183. /*
  184. * code to bootstrap the factorization
  185. */
  186. static int dw_codelet_facto_pivot(starpu_data_handle_t *dataAp,
  187. struct piv_s *piv_description,
  188. unsigned nblocks,
  189. starpu_data_handle_t (* get_block)(starpu_data_handle_t *, unsigned, unsigned, unsigned),
  190. double *timing)
  191. {
  192. int ret;
  193. struct timeval start;
  194. struct timeval end;
  195. struct starpu_task *entry_task = NULL;
  196. /* create all the DAG nodes */
  197. unsigned i,j,k;
  198. for (k = 0; k < nblocks; k++)
  199. {
  200. struct starpu_task *task = create_task_11_pivot(dataAp, nblocks, k, piv_description, get_block);
  201. /* we defer the launch of the first task */
  202. if (k == 0)
  203. {
  204. entry_task = task;
  205. }
  206. else
  207. {
  208. ret = starpu_task_submit(task);
  209. if (ret != -ENODEV) STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
  210. return ret;
  211. }
  212. for (i = 0; i < nblocks; i++)
  213. {
  214. if (i != k)
  215. {
  216. ret = create_task_pivot(dataAp, nblocks, piv_description, k, i, get_block);
  217. if (ret == -ENODEV) return ret;
  218. }
  219. }
  220. for (i = k+1; i<nblocks; i++)
  221. {
  222. ret = create_task_12(dataAp, nblocks, k, i, get_block);
  223. if (ret == -ENODEV) return ret;
  224. ret = create_task_21(dataAp, nblocks, k, i, get_block);
  225. if (ret == -ENODEV) return ret;
  226. }
  227. for (i = k+1; i<nblocks; i++)
  228. {
  229. for (j = k+1; j<nblocks; j++)
  230. {
  231. ret = create_task_22(dataAp, nblocks, k, i, j, get_block);
  232. if (ret == -ENODEV) return ret;
  233. }
  234. }
  235. }
  236. /* we wait the last task (TAG11(nblocks - 1)) and all the pivot tasks */
  237. starpu_tag_t *tags = malloc(nblocks*nblocks*sizeof(starpu_tag_t));
  238. unsigned ndeps = 0;
  239. tags[ndeps++] = TAG11(nblocks - 1);
  240. for (j = 0; j < nblocks; j++)
  241. {
  242. for (i = 0; i < j; i++)
  243. {
  244. tags[ndeps++] = PIVOT(j, i);
  245. }
  246. }
  247. /* schedule the codelet */
  248. gettimeofday(&start, NULL);
  249. ret = starpu_task_submit(entry_task);
  250. if (ret != -ENODEV) STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
  251. /* stall the application until the end of computations */
  252. starpu_tag_wait_array(ndeps, tags);
  253. /* starpu_task_wait_for_all(); */
  254. free(tags);
  255. gettimeofday(&end, NULL);
  256. *timing = (double)((end.tv_sec - start.tv_sec)*1000000 + (end.tv_usec - start.tv_usec));
  257. return 0;
  258. }
  259. starpu_data_handle_t get_block_with_striding(starpu_data_handle_t *dataAp,
  260. unsigned nblocks __attribute__((unused)), unsigned j, unsigned i)
  261. {
  262. /* we use filters */
  263. return starpu_data_get_sub_data(*dataAp, 2, j, i);
  264. }
  265. int STARPU_LU(lu_decomposition_pivot)(TYPE *matA, unsigned *ipiv, unsigned size, unsigned ld, unsigned nblocks)
  266. {
  267. starpu_data_handle_t dataA;
  268. /* monitor and partition the A matrix into blocks :
  269. * one block is now determined by 2 unsigned (i,j) */
  270. starpu_matrix_data_register(&dataA, 0, (uintptr_t)matA, ld, size, size, sizeof(TYPE));
  271. /* We already enforce deps by hand */
  272. starpu_data_set_sequential_consistency_flag(dataA, 0);
  273. struct starpu_data_filter f =
  274. {
  275. .filter_func = starpu_matrix_filter_vertical_block,
  276. .nchildren = nblocks
  277. };
  278. struct starpu_data_filter f2 =
  279. {
  280. .filter_func = starpu_matrix_filter_block,
  281. .nchildren = nblocks
  282. };
  283. starpu_data_map_filters(dataA, 2, &f, &f2);
  284. unsigned i;
  285. for (i = 0; i < size; i++)
  286. ipiv[i] = i;
  287. struct piv_s *piv_description = malloc(nblocks*sizeof(struct piv_s));
  288. unsigned block;
  289. for (block = 0; block < nblocks; block++)
  290. {
  291. piv_description[block].piv = ipiv;
  292. piv_description[block].first = block * (size / nblocks);
  293. piv_description[block].last = (block + 1) * (size / nblocks);
  294. }
  295. #if 0
  296. unsigned j;
  297. for (j = 0; j < nblocks; j++)
  298. for (i = 0; i < nblocks; i++)
  299. {
  300. printf("BLOCK %d %d %p\n", i, j, &matA[i*(size/nblocks) + j * (size/nblocks)*ld]);
  301. }
  302. #endif
  303. double timing=0.0;
  304. int ret = dw_codelet_facto_pivot(&dataA, piv_description, nblocks, get_block_with_striding, &timing);
  305. FPRINTF(stderr, "Computation took (in ms)\n");
  306. FPRINTF(stderr, "%2.2f\n", timing/1000);
  307. unsigned n = starpu_matrix_get_nx(dataA);
  308. double flop = (2.0f*n*n*n)/3.0f;
  309. FPRINTF(stderr, "Synthetic GFlops : %2.2f\n", (flop/timing/1000.0f));
  310. /* gather all the data */
  311. starpu_data_unpartition(dataA, 0);
  312. free(piv_description);
  313. return ret;
  314. }
  315. starpu_data_handle_t get_block_with_no_striding(starpu_data_handle_t *dataAp, unsigned nblocks, unsigned j, unsigned i)
  316. {
  317. /* dataAp is an array of data handle */
  318. return dataAp[i+j*nblocks];
  319. }
  320. int STARPU_LU(lu_decomposition_pivot_no_stride)(TYPE **matA, unsigned *ipiv, unsigned size, unsigned ld, unsigned nblocks)
  321. {
  322. starpu_data_handle_t *dataAp = malloc(nblocks*nblocks*sizeof(starpu_data_handle_t));
  323. /* monitor and partition the A matrix into blocks :
  324. * one block is now determined by 2 unsigned (i,j) */
  325. unsigned bi, bj;
  326. for (bj = 0; bj < nblocks; bj++)
  327. for (bi = 0; bi < nblocks; bi++)
  328. {
  329. starpu_matrix_data_register(&dataAp[bi+nblocks*bj], 0,
  330. (uintptr_t)matA[bi+nblocks*bj], size/nblocks,
  331. size/nblocks, size/nblocks, sizeof(TYPE));
  332. /* We already enforce deps by hand */
  333. starpu_data_set_sequential_consistency_flag(dataAp[bi+nblocks*bj], 0);
  334. }
  335. unsigned i;
  336. for (i = 0; i < size; i++)
  337. ipiv[i] = i;
  338. struct piv_s *piv_description = malloc(nblocks*sizeof(struct piv_s));
  339. unsigned block;
  340. for (block = 0; block < nblocks; block++)
  341. {
  342. piv_description[block].piv = ipiv;
  343. piv_description[block].first = block * (size / nblocks);
  344. piv_description[block].last = (block + 1) * (size / nblocks);
  345. }
  346. double timing=0.0;
  347. int ret = dw_codelet_facto_pivot(dataAp, piv_description, nblocks, get_block_with_no_striding, &timing);
  348. FPRINTF(stderr, "Computation took (in ms)\n");
  349. FPRINTF(stderr, "%2.2f\n", timing/1000);
  350. unsigned n = starpu_matrix_get_nx(dataAp[0])*nblocks;
  351. double flop = (2.0f*n*n*n)/3.0f;
  352. FPRINTF(stderr, "Synthetic GFlops : %2.2f\n", (flop/timing/1000.0f));
  353. for (bj = 0; bj < nblocks; bj++)
  354. for (bi = 0; bi < nblocks; bi++)
  355. {
  356. starpu_data_unregister(dataAp[bi+nblocks*bj]);
  357. }
  358. free(dataAp);
  359. return ret;
  360. }