lu_example.c 9.4 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455
  1. /* StarPU --- Runtime system for heterogeneous multicore architectures.
  2. *
  3. * Copyright (C) 2009-2020 Université de Bordeaux, CNRS (LaBRI UMR 5800), Inria
  4. *
  5. * StarPU is free software; you can redistribute it and/or modify
  6. * it under the terms of the GNU Lesser General Public License as published by
  7. * the Free Software Foundation; either version 2.1 of the License, or (at
  8. * your option) any later version.
  9. *
  10. * StarPU is distributed in the hope that it will be useful, but
  11. * WITHOUT ANY WARRANTY; without even the implied warranty of
  12. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  13. *
  14. * See the GNU Lesser General Public License in COPYING.LGPL for more details.
  15. */
  16. /* Main body for the LU factorization: matrix initialization and result
  17. * checking */
  18. #include <stdlib.h>
  19. #include <stdio.h>
  20. #include <string.h>
  21. #include <time.h>
  22. #include <math.h>
  23. #include <starpu.h>
  24. #include "xlu.h"
  25. #include "xlu_kernels.h"
  26. static unsigned long size = 0;
  27. static unsigned nblocks = 0;
  28. static unsigned check = 0;
  29. static unsigned pivot = 0;
  30. static unsigned no_stride = 0;
  31. static unsigned profile = 0;
  32. static unsigned no_prio=0;
  33. unsigned bound = 0;
  34. unsigned bounddeps = 0;
  35. unsigned boundprio = 0;
  36. #define FPRINTF(ofile, fmt, ...) do { if (!getenv("STARPU_SSILENT")) {fprintf(ofile, fmt, ## __VA_ARGS__); }} while(0)
  37. TYPE *A, *A_saved;
  38. /* in case we use non-strided blocks */
  39. TYPE **A_blocks;
  40. static void parse_args(int argc, char **argv)
  41. {
  42. int i;
  43. for (i = 1; i < argc; i++)
  44. {
  45. if (strcmp(argv[i], "-size") == 0)
  46. {
  47. char *argptr;
  48. size = strtol(argv[++i], &argptr, 10);
  49. }
  50. else if (strcmp(argv[i], "-nblocks") == 0)
  51. {
  52. char *argptr;
  53. nblocks = strtol(argv[++i], &argptr, 10);
  54. }
  55. #ifndef STARPU_SIMGRID
  56. else if (strcmp(argv[i], "-check") == 0)
  57. {
  58. check = 1;
  59. }
  60. else if (strcmp(argv[i], "-piv") == 0)
  61. {
  62. pivot = 1;
  63. }
  64. else if (strcmp(argv[i], "-no-stride") == 0)
  65. {
  66. no_stride = 1;
  67. }
  68. #endif
  69. else if (strcmp(argv[i], "-profile") == 0)
  70. {
  71. profile = 1;
  72. }
  73. else if (strcmp(argv[i], "-bound") == 0)
  74. {
  75. bound = 1;
  76. }
  77. else if (strcmp(argv[i], "-bounddeps") == 0)
  78. {
  79. bound = 1;
  80. bounddeps = 1;
  81. }
  82. else if (strcmp(argv[i], "-bounddepsprio") == 0)
  83. {
  84. bound = 1;
  85. bounddeps = 1;
  86. boundprio = 1;
  87. }
  88. else if (strcmp(argv[i], "-h") == 0 || strcmp(argv[i], "--help") == 0)
  89. {
  90. fprintf(stderr,"usage: lu [-size n] [-nblocks b] [-piv] [-no-stride] [-profile] [-bound] [-bounddeps] [-bounddepsprio]\n");
  91. fprintf(stderr,"Default is size %lu and nblocks %u\n", size, nblocks);
  92. exit(0);
  93. }
  94. }
  95. }
  96. static void display_matrix(TYPE *m, unsigned n, unsigned ld, char *str)
  97. {
  98. (void)m;
  99. (void)n;
  100. (void)ld;
  101. (void)str;
  102. #if 0
  103. FPRINTF(stderr, "***********\n");
  104. FPRINTF(stderr, "Display matrix %s\n", str);
  105. unsigned i,j;
  106. for (j = 0; j < n; j++)
  107. {
  108. for (i = 0; i < n; i++)
  109. {
  110. FPRINTF(stderr, "%2.2f\t", m[i+j*ld]);
  111. }
  112. FPRINTF(stderr, "\n");
  113. }
  114. FPRINTF(stderr, "***********\n");
  115. #endif
  116. }
  117. void copy_blocks_into_matrix(void)
  118. {
  119. unsigned blocksize = (size/nblocks);
  120. unsigned i, j;
  121. unsigned bi, bj;
  122. for (bj = 0; bj < nblocks; bj++)
  123. for (bi = 0; bi < nblocks; bi++)
  124. {
  125. for (j = 0; j < blocksize; j++)
  126. for (i = 0; i < blocksize; i++)
  127. {
  128. A[(i+bi*blocksize) + (j + bj*blocksize)*size] =
  129. A_blocks[bi+nblocks*bj][i + j * blocksize];
  130. }
  131. starpu_free(A_blocks[bi+nblocks*bj]);
  132. }
  133. }
  134. void copy_matrix_into_blocks(void)
  135. {
  136. unsigned blocksize = (size/nblocks);
  137. unsigned i, j;
  138. unsigned bi, bj;
  139. for (bj = 0; bj < nblocks; bj++)
  140. for (bi = 0; bi < nblocks; bi++)
  141. {
  142. starpu_malloc((void **)&A_blocks[bi+nblocks*bj], (size_t)blocksize*blocksize*sizeof(TYPE));
  143. for (j = 0; j < blocksize; j++)
  144. for (i = 0; i < blocksize; i++)
  145. {
  146. A_blocks[bi+nblocks*bj][i + j * blocksize] =
  147. A[(i+bi*blocksize) + (j + bj*blocksize)*size];
  148. }
  149. }
  150. }
  151. static void init_matrix(void)
  152. {
  153. /* allocate matrix */
  154. #ifdef STARPU_SIMGRID
  155. A = (void*) 1;
  156. #else
  157. starpu_malloc_flags((void **)&A, (size_t)size*size*sizeof(TYPE), STARPU_MALLOC_PINNED|STARPU_MALLOC_SIMULATION_FOLDED);
  158. #endif
  159. STARPU_ASSERT(A);
  160. starpu_srand48((long int)time(NULL));
  161. /* starpu_srand48(0); */
  162. #ifndef STARPU_SIMGRID
  163. /* initialize matrix content */
  164. unsigned long i,j;
  165. for (j = 0; j < size; j++)
  166. {
  167. for (i = 0; i < size; i++)
  168. {
  169. A[i + j*size] = (TYPE)starpu_drand48();
  170. #ifdef COMPLEX_LU
  171. /* also randomize the imaginary component for complex number cases */
  172. A[i + j*size] += (TYPE)(I*starpu_drand48());
  173. #endif
  174. if (i == j)
  175. {
  176. A[i + j*size] += 1;
  177. A[i + j*size] *= 100;
  178. }
  179. }
  180. }
  181. #endif
  182. }
  183. static void save_matrix(void)
  184. {
  185. A_saved = malloc((size_t)size*size*sizeof(TYPE));
  186. STARPU_ASSERT(A_saved);
  187. memcpy(A_saved, A, (size_t)size*size*sizeof(TYPE));
  188. }
  189. static double frobenius_norm(TYPE *v, unsigned n)
  190. {
  191. double sum2 = 0.0;
  192. /* compute sqrt(Sum(|x|^2)) */
  193. unsigned i,j;
  194. for (j = 0; j < n; j++)
  195. for (i = 0; i < n; i++)
  196. {
  197. double a = fabsl((double)v[i+n*j]);
  198. sum2 += a*a;
  199. }
  200. return sqrt(sum2);
  201. }
  202. static void pivot_saved_matrix(unsigned *ipiv)
  203. {
  204. unsigned k;
  205. for (k = 0; k < size; k++)
  206. {
  207. if (k != ipiv[k])
  208. {
  209. /* FPRINTF(stderr, "SWAP %d and %d\n", k, ipiv[k]); */
  210. CPU_SWAP(size, &A_saved[k*size], 1, &A_saved[ipiv[k]*size], 1);
  211. }
  212. }
  213. }
  214. static void check_result(void)
  215. {
  216. unsigned i,j;
  217. TYPE *L, *U;
  218. L = malloc((size_t)size*size*sizeof(TYPE));
  219. U = malloc((size_t)size*size*sizeof(TYPE));
  220. memset(L, 0, size*size*sizeof(TYPE));
  221. memset(U, 0, size*size*sizeof(TYPE));
  222. /* only keep the lower part */
  223. for (j = 0; j < size; j++)
  224. {
  225. for (i = 0; i < j; i++)
  226. {
  227. L[j+i*size] = A[j+i*size];
  228. }
  229. /* diag i = j */
  230. L[j+j*size] = A[j+j*size];
  231. U[j+j*size] = 1.0;
  232. for (i = j+1; i < size; i++)
  233. {
  234. U[j+i*size] = A[j+i*size];
  235. }
  236. }
  237. display_matrix(L, size, size, "L");
  238. display_matrix(U, size, size, "U");
  239. /* now A_err = L, compute L*U */
  240. CPU_TRMM("R", "U", "N", "U", size, size, 1.0f, U, size, L, size);
  241. display_matrix(A_saved, size, size, "P A_saved");
  242. display_matrix(L, size, size, "LU");
  243. /* compute "LU - A" in L*/
  244. CPU_AXPY(size*size, -1.0, A_saved, 1, L, 1);
  245. display_matrix(L, size, size, "Residuals");
  246. #ifdef COMPLEX_LU
  247. double err = CPU_ASUM(size*size, L, 1);
  248. int max = CPU_IAMAX(size*size, L, 1);
  249. TYPE l_max = L[max];
  250. FPRINTF(stderr, "Avg error : %e\n", err/(size*size));
  251. FPRINTF(stderr, "Max error : %e\n", sqrt(creal(l_max)*creal(l_max)+cimag(l_max)*cimag(l_max)));
  252. #else
  253. TYPE err = CPU_ASUM(size*size, L, 1);
  254. int max = CPU_IAMAX(size*size, L, 1);
  255. FPRINTF(stderr, "Avg error : %e\n", err/(size*size));
  256. FPRINTF(stderr, "Max error : %e\n", L[max]);
  257. #endif
  258. double residual = frobenius_norm(L, size);
  259. double matnorm = frobenius_norm(A_saved, size);
  260. FPRINTF(stderr, "||%sA-LU|| / (||A||*N) : %e\n", pivot?"P":"", residual/(matnorm*size));
  261. if (residual/(matnorm*size) > 1e-5)
  262. exit(-1);
  263. free(L);
  264. free(U);
  265. free(A_saved);
  266. }
  267. int main(int argc, char **argv)
  268. {
  269. int ret;
  270. ret = starpu_init(NULL);
  271. if (ret == -ENODEV)
  272. return 77;
  273. STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
  274. int power = starpu_cpu_worker_get_count() + 32 * starpu_cuda_worker_get_count();
  275. int power_cbrt = cbrt(power);
  276. #ifndef STARPU_LONG_CHECK
  277. power_cbrt /= 2;
  278. #endif
  279. if (power_cbrt < 1)
  280. power_cbrt = 1;
  281. #ifdef STARPU_QUICK_CHECK
  282. if (!size)
  283. size = 320*2*power_cbrt;
  284. if (!nblocks)
  285. nblocks = 2*power_cbrt;
  286. #else
  287. if (!size)
  288. size = 960*8*power_cbrt;
  289. if (!nblocks)
  290. nblocks = 8*power_cbrt;
  291. #endif
  292. parse_args(argc, argv);
  293. starpu_cublas_init();
  294. init_matrix();
  295. #ifndef STARPU_SIMGRID
  296. unsigned *ipiv = NULL;
  297. if (check)
  298. save_matrix();
  299. display_matrix(A, size, size, "A");
  300. if (profile)
  301. starpu_profiling_status_set(STARPU_PROFILING_ENABLE);
  302. /* Factorize the matrix (in place) */
  303. if (pivot)
  304. {
  305. ipiv = malloc(size*sizeof(unsigned));
  306. if (no_stride)
  307. {
  308. /* in case the LU decomposition uses non-strided blocks, we _copy_ the matrix into smaller blocks */
  309. A_blocks = malloc(nblocks*nblocks*sizeof(TYPE *));
  310. copy_matrix_into_blocks();
  311. ret = STARPU_LU(lu_decomposition_pivot_no_stride)(A_blocks, ipiv, size, size, nblocks, no_prio);
  312. copy_blocks_into_matrix();
  313. free(A_blocks);
  314. }
  315. else
  316. {
  317. double start;
  318. double end;
  319. start = starpu_timing_now();
  320. ret = STARPU_LU(lu_decomposition_pivot)(A, ipiv, size, size, nblocks, no_prio);
  321. end = starpu_timing_now();
  322. double timing = end - start;
  323. unsigned n = size;
  324. double flop = (2.0f*n*n*n)/3.0f;
  325. FPRINTF(stderr, "Synthetic GFlops (TOTAL) : \n");
  326. FPRINTF(stdout, "%u %6.2f\n", n, (flop/timing/1000.0f));
  327. }
  328. }
  329. else
  330. #endif
  331. {
  332. ret = STARPU_LU(lu_decomposition)(A, size, size, nblocks, no_prio);
  333. }
  334. if (profile)
  335. {
  336. FPRINTF(stderr, "Setting profile\n");
  337. starpu_profiling_status_set(STARPU_PROFILING_DISABLE);
  338. starpu_profiling_bus_helper_display_summary();
  339. }
  340. if (bound)
  341. {
  342. if (bounddeps)
  343. {
  344. FILE *f = fopen("lu.pl", "w");
  345. starpu_bound_print_lp(f);
  346. FPRINTF(stderr,"system printed to lu.pl\n");
  347. fclose(f);
  348. f = fopen("lu.mps", "w");
  349. starpu_bound_print_mps(f);
  350. FPRINTF(stderr,"system printed to lu.mps\n");
  351. fclose(f);
  352. f = fopen("lu.dot", "w");
  353. starpu_bound_print_dot(f);
  354. FPRINTF(stderr,"system printed to lu.mps\n");
  355. fclose(f);
  356. }
  357. }
  358. #ifndef STARPU_SIMGRID
  359. if (check)
  360. {
  361. FPRINTF(stderr, "Checking result\n");
  362. if (pivot)
  363. {
  364. pivot_saved_matrix(ipiv);
  365. }
  366. check_result();
  367. }
  368. if (pivot)
  369. free(ipiv);
  370. #endif
  371. #ifndef STARPU_SIMGRID
  372. starpu_free_flags(A, (size_t)size*size*sizeof(TYPE), STARPU_MALLOC_PINNED|STARPU_MALLOC_SIMULATION_FOLDED);
  373. #endif
  374. starpu_cublas_shutdown();
  375. starpu_shutdown();
  376. if (ret == -ENODEV) return 77; else return 0;
  377. }