lu_example.c 8.8 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423
  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 <stdlib.h>
  18. #include <stdio.h>
  19. #include <string.h>
  20. #include <time.h>
  21. #include <math.h>
  22. #include <starpu.h>
  23. #include "xlu.h"
  24. #include "xlu_kernels.h"
  25. static unsigned long size = 4096;
  26. static unsigned nblocks = 16;
  27. static unsigned check = 0;
  28. static unsigned pivot = 0;
  29. static unsigned no_stride = 0;
  30. static unsigned profile = 0;
  31. static unsigned bound = 0;
  32. static unsigned bounddeps = 0;
  33. static unsigned boundprio = 0;
  34. #define FPRINTF(ofile, fmt, args ...) do { if (!getenv("STARPU_SSILENT")) {fprintf(ofile, fmt, ##args); }} while(0)
  35. TYPE *A, *A_saved;
  36. /* in case we use non-strided blocks */
  37. TYPE **A_blocks;
  38. static void parse_args(int argc, char **argv)
  39. {
  40. int i;
  41. for (i = 1; i < argc; i++)
  42. {
  43. if (strcmp(argv[i], "-size") == 0)
  44. {
  45. char *argptr;
  46. size = strtol(argv[++i], &argptr, 10);
  47. }
  48. else if (strcmp(argv[i], "-nblocks") == 0)
  49. {
  50. char *argptr;
  51. nblocks = strtol(argv[++i], &argptr, 10);
  52. }
  53. else if (strcmp(argv[i], "-check") == 0)
  54. {
  55. check = 1;
  56. }
  57. else if (strcmp(argv[i], "-piv") == 0)
  58. {
  59. pivot = 1;
  60. }
  61. else if (strcmp(argv[i], "-no-stride") == 0)
  62. {
  63. no_stride = 1;
  64. }
  65. else if (strcmp(argv[i], "-profile") == 0)
  66. {
  67. profile = 1;
  68. }
  69. else if (strcmp(argv[i], "-bound") == 0)
  70. {
  71. bound = 1;
  72. }
  73. else if (strcmp(argv[i], "-bounddeps") == 0)
  74. {
  75. bound = 1;
  76. bounddeps = 1;
  77. }
  78. else if (strcmp(argv[i], "-bounddepsprio") == 0)
  79. {
  80. bound = 1;
  81. bounddeps = 1;
  82. boundprio = 1;
  83. }
  84. else if (strcmp(argv[i], "-h") == 0 || strcmp(argv[i], "--help") == 0)
  85. {
  86. fprintf(stderr,"usage: lu [-size n] [-nblocks b] [-piv] [-no-stride] [-profile] [-bound] [-bounddeps] [-bounddepsprio]\n");
  87. exit(0);
  88. }
  89. }
  90. }
  91. static void display_matrix(TYPE *m, unsigned n, unsigned ld, char *str)
  92. {
  93. #if 0
  94. FPRINTF(stderr, "***********\n");
  95. FPRINTF(stderr, "Display matrix %s\n", str);
  96. unsigned i,j;
  97. for (j = 0; j < n; j++)
  98. {
  99. for (i = 0; i < n; i++)
  100. {
  101. FPRINTF(stderr, "%2.2f\t", m[i+j*ld]);
  102. }
  103. FPRINTF(stderr, "\n");
  104. }
  105. FPRINTF(stderr, "***********\n");
  106. #endif
  107. }
  108. void copy_blocks_into_matrix(void)
  109. {
  110. unsigned blocksize = (size/nblocks);
  111. unsigned i, j;
  112. unsigned bi, bj;
  113. for (bj = 0; bj < nblocks; bj++)
  114. for (bi = 0; bi < nblocks; bi++)
  115. {
  116. for (j = 0; j < blocksize; j++)
  117. for (i = 0; i < blocksize; i++)
  118. {
  119. A[(i+bi*blocksize) + (j + bj*blocksize)*size] =
  120. A_blocks[bi+nblocks*bj][i + j * blocksize];
  121. }
  122. starpu_free(A_blocks[bi+nblocks*bj]);
  123. }
  124. }
  125. void copy_matrix_into_blocks(void)
  126. {
  127. unsigned blocksize = (size/nblocks);
  128. unsigned i, j;
  129. unsigned bi, bj;
  130. for (bj = 0; bj < nblocks; bj++)
  131. for (bi = 0; bi < nblocks; bi++)
  132. {
  133. starpu_malloc((void **)&A_blocks[bi+nblocks*bj], (size_t)blocksize*blocksize*sizeof(TYPE));
  134. for (j = 0; j < blocksize; j++)
  135. for (i = 0; i < blocksize; i++)
  136. {
  137. A_blocks[bi+nblocks*bj][i + j * blocksize] =
  138. A[(i+bi*blocksize) + (j + bj*blocksize)*size];
  139. }
  140. }
  141. }
  142. static void init_matrix(void)
  143. {
  144. /* allocate matrix */
  145. starpu_malloc((void **)&A, (size_t)size*size*sizeof(TYPE));
  146. STARPU_ASSERT(A);
  147. starpu_srand48((long int)time(NULL));
  148. /* starpu_srand48(0); */
  149. /* initialize matrix content */
  150. unsigned long i,j;
  151. for (j = 0; j < size; j++)
  152. {
  153. for (i = 0; i < size; i++)
  154. {
  155. A[i + j*size] = (TYPE)starpu_drand48();
  156. #ifdef COMPLEX_LU
  157. /* also randomize the imaginary component for complex number cases */
  158. A[i + j*size] += (TYPE)(I*starpu_drand48());
  159. #endif
  160. }
  161. }
  162. }
  163. static void save_matrix(void)
  164. {
  165. A_saved = malloc((size_t)size*size*sizeof(TYPE));
  166. STARPU_ASSERT(A_saved);
  167. memcpy(A_saved, A, (size_t)size*size*sizeof(TYPE));
  168. }
  169. static double frobenius_norm(TYPE *v, unsigned n)
  170. {
  171. double sum2 = 0.0;
  172. /* compute sqrt(Sum(|x|^2)) */
  173. unsigned i,j;
  174. for (j = 0; j < n; j++)
  175. for (i = 0; i < n; i++)
  176. {
  177. double a = fabsl((double)v[i+n*j]);
  178. sum2 += a*a;
  179. }
  180. return sqrt(sum2);
  181. }
  182. static void pivot_saved_matrix(unsigned *ipiv)
  183. {
  184. unsigned k;
  185. for (k = 0; k < size; k++)
  186. {
  187. if (k != ipiv[k])
  188. {
  189. /* FPRINTF(stderr, "SWAP %d and %d\n", k, ipiv[k]); */
  190. CPU_SWAP(size, &A_saved[k*size], 1, &A_saved[ipiv[k]*size], 1);
  191. }
  192. }
  193. }
  194. static void check_result(void)
  195. {
  196. unsigned i,j;
  197. TYPE *L, *U;
  198. L = malloc((size_t)size*size*sizeof(TYPE));
  199. U = malloc((size_t)size*size*sizeof(TYPE));
  200. memset(L, 0, size*size*sizeof(TYPE));
  201. memset(U, 0, size*size*sizeof(TYPE));
  202. /* only keep the lower part */
  203. for (j = 0; j < size; j++)
  204. {
  205. for (i = 0; i < j; i++)
  206. {
  207. L[j+i*size] = A[j+i*size];
  208. }
  209. /* diag i = j */
  210. L[j+j*size] = A[j+j*size];
  211. U[j+j*size] = 1.0;
  212. for (i = j+1; i < size; i++)
  213. {
  214. U[j+i*size] = A[j+i*size];
  215. }
  216. }
  217. display_matrix(L, size, size, "L");
  218. display_matrix(U, size, size, "U");
  219. /* now A_err = L, compute L*U */
  220. CPU_TRMM("R", "U", "N", "U", size, size, 1.0f, U, size, L, size);
  221. display_matrix(A_saved, size, size, "P A_saved");
  222. display_matrix(L, size, size, "LU");
  223. /* compute "LU - A" in L*/
  224. CPU_AXPY(size*size, -1.0, A_saved, 1, L, 1);
  225. display_matrix(L, size, size, "Residuals");
  226. #ifdef COMPLEX_LU
  227. double err = CPU_ASUM(size*size, L, 1);
  228. int max = CPU_IAMAX(size*size, L, 1);
  229. TYPE l_max = L[max];
  230. FPRINTF(stderr, "Avg error : %e\n", err/(size*size));
  231. FPRINTF(stderr, "Max error : %e\n", sqrt(creal(l_max)*creal(l_max)+cimag(l_max)*cimag(l_max)));
  232. #else
  233. TYPE err = CPU_ASUM(size*size, L, 1);
  234. int max = CPU_IAMAX(size*size, L, 1);
  235. FPRINTF(stderr, "Avg error : %e\n", err/(size*size));
  236. FPRINTF(stderr, "Max error : %e\n", L[max]);
  237. #endif
  238. double residual = frobenius_norm(L, size);
  239. double matnorm = frobenius_norm(A_saved, size);
  240. FPRINTF(stderr, "||%sA-LU|| / (||A||*N) : %e\n", pivot?"P":"", residual/(matnorm*size));
  241. if (residual/(matnorm*size) > 1e-5)
  242. exit(-1);
  243. free(L);
  244. free(U);
  245. free(A_saved);
  246. }
  247. int main(int argc, char **argv)
  248. {
  249. int ret;
  250. parse_args(argc, argv);
  251. #ifdef STARPU_QUICK_CHECK
  252. size /= 4;
  253. nblocks /= 4;
  254. #endif
  255. ret = starpu_init(NULL);
  256. if (ret == -ENODEV)
  257. return 77;
  258. STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
  259. starpu_cublas_init();
  260. init_matrix();
  261. unsigned *ipiv = NULL;
  262. if (check)
  263. save_matrix();
  264. display_matrix(A, size, size, "A");
  265. if (bound)
  266. starpu_bound_start(bounddeps, boundprio);
  267. if (profile)
  268. starpu_profiling_status_set(STARPU_PROFILING_ENABLE);
  269. /* Factorize the matrix (in place) */
  270. if (pivot)
  271. {
  272. ipiv = malloc(size*sizeof(unsigned));
  273. if (no_stride)
  274. {
  275. /* in case the LU decomposition uses non-strided blocks, we _copy_ the matrix into smaller blocks */
  276. A_blocks = malloc(nblocks*nblocks*sizeof(TYPE **));
  277. copy_matrix_into_blocks();
  278. ret = STARPU_LU(lu_decomposition_pivot_no_stride)(A_blocks, ipiv, size, size, nblocks);
  279. copy_blocks_into_matrix();
  280. free(A_blocks);
  281. }
  282. else
  283. {
  284. struct timeval start;
  285. struct timeval end;
  286. gettimeofday(&start, NULL);
  287. ret = STARPU_LU(lu_decomposition_pivot)(A, ipiv, size, size, nblocks);
  288. gettimeofday(&end, NULL);
  289. double timing = (double)((end.tv_sec - start.tv_sec)*1000000 + (end.tv_usec - start.tv_usec));
  290. unsigned n = size;
  291. double flop = (2.0f*n*n*n)/3.0f;
  292. FPRINTF(stderr, "Synthetic GFlops (TOTAL) : \n");
  293. FPRINTF(stdout, "%u %6.2f\n", n, (flop/timing/1000.0f));
  294. }
  295. }
  296. else
  297. {
  298. ret = STARPU_LU(lu_decomposition)(A, size, size, nblocks);
  299. }
  300. if (profile)
  301. {
  302. FPRINTF(stderr, "Setting profile\n");
  303. starpu_profiling_status_set(STARPU_PROFILING_DISABLE);
  304. starpu_bus_profiling_helper_display_summary();
  305. }
  306. if (bound)
  307. {
  308. double min;
  309. FPRINTF(stderr, "Setting bound\n");
  310. starpu_bound_stop();
  311. if (bounddeps)
  312. {
  313. FILE *f = fopen("lu.pl", "w");
  314. starpu_bound_print_lp(f);
  315. FPRINTF(stderr,"system printed to lu.pl\n");
  316. fclose(f);
  317. f = fopen("lu.mps", "w");
  318. starpu_bound_print_mps(f);
  319. FPRINTF(stderr,"system printed to lu.mps\n");
  320. fclose(f);
  321. f = fopen("lu.dot", "w");
  322. starpu_bound_print_dot(f);
  323. FPRINTF(stderr,"system printed to lu.mps\n");
  324. fclose(f);
  325. }
  326. else
  327. {
  328. starpu_bound_compute(&min, NULL, 0);
  329. if (min != 0.)
  330. FPRINTF(stderr, "theoretical min: %f ms\n", min);
  331. }
  332. }
  333. if (check)
  334. {
  335. FPRINTF(stderr, "Checking result\n");
  336. if (pivot) {
  337. pivot_saved_matrix(ipiv);
  338. free(ipiv);
  339. }
  340. check_result();
  341. }
  342. starpu_free(A);
  343. FPRINTF(stderr, "Shutting down\n");
  344. starpu_cublas_shutdown();
  345. starpu_shutdown();
  346. if (ret == -ENODEV) return 77; else return 0;
  347. }