lu_example.c 8.2 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396
  1. /* StarPU --- Runtime system for heterogeneous multicore architectures.
  2. *
  3. * Copyright (C) 2009, 2010-2011 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 <starpu_profiling.h>
  24. #include <starpu_bound.h>
  25. #include "xlu.h"
  26. #include "xlu_kernels.h"
  27. static unsigned long size = 4096;
  28. static unsigned nblocks = 16;
  29. static unsigned check = 0;
  30. static unsigned pivot = 0;
  31. static unsigned no_stride = 0;
  32. static unsigned profile = 0;
  33. static unsigned bound = 0;
  34. static unsigned bounddeps = 0;
  35. static unsigned boundprio = 0;
  36. #define FPRINTF(ofile, fmt, args ...) do { if (!getenv("STARPU_SSILENT")) {fprintf(ofile, fmt, ##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. if (strcmp(argv[i], "-nblocks") == 0)
  51. {
  52. char *argptr;
  53. nblocks = strtol(argv[++i], &argptr, 10);
  54. }
  55. if (strcmp(argv[i], "-check") == 0)
  56. {
  57. check = 1;
  58. }
  59. if (strcmp(argv[i], "-piv") == 0)
  60. {
  61. pivot = 1;
  62. }
  63. if (strcmp(argv[i], "-no-stride") == 0)
  64. {
  65. no_stride = 1;
  66. }
  67. if (strcmp(argv[i], "-profile") == 0)
  68. {
  69. profile = 1;
  70. }
  71. if (strcmp(argv[i], "-bound") == 0)
  72. {
  73. bound = 1;
  74. }
  75. if (strcmp(argv[i], "-bounddeps") == 0)
  76. {
  77. bound = 1;
  78. bounddeps = 1;
  79. }
  80. if (strcmp(argv[i], "-bounddepsprio") == 0)
  81. {
  82. bound = 1;
  83. bounddeps = 1;
  84. boundprio = 1;
  85. }
  86. }
  87. }
  88. static void display_matrix(TYPE *m, unsigned n, unsigned ld, char *str)
  89. {
  90. #if 0
  91. FPRINTF(stderr, "***********\n");
  92. FPRINTF(stderr, "Display matrix %s\n", str);
  93. unsigned i,j;
  94. for (j = 0; j < n; j++)
  95. {
  96. for (i = 0; i < n; i++)
  97. {
  98. FPRINTF(stderr, "%2.2f\t", m[i+j*ld]);
  99. }
  100. FPRINTF(stderr, "\n");
  101. }
  102. FPRINTF(stderr, "***********\n");
  103. #endif
  104. }
  105. void copy_blocks_into_matrix(void)
  106. {
  107. unsigned blocksize = (size/nblocks);
  108. unsigned i, j;
  109. unsigned bi, bj;
  110. for (bj = 0; bj < nblocks; bj++)
  111. for (bi = 0; bi < nblocks; bi++)
  112. {
  113. for (j = 0; j < blocksize; j++)
  114. for (i = 0; i < blocksize; i++)
  115. {
  116. A[(i+bi*blocksize) + (j + bj*blocksize)*size] =
  117. A_blocks[bi+nblocks*bj][i + j * blocksize];
  118. }
  119. /* free(A_blocks[bi+nblocks*bj]); */
  120. }
  121. }
  122. void copy_matrix_into_blocks(void)
  123. {
  124. unsigned blocksize = (size/nblocks);
  125. unsigned i, j;
  126. unsigned bi, bj;
  127. for (bj = 0; bj < nblocks; bj++)
  128. for (bi = 0; bi < nblocks; bi++)
  129. {
  130. starpu_malloc((void **)&A_blocks[bi+nblocks*bj], (size_t)blocksize*blocksize*sizeof(TYPE));
  131. for (j = 0; j < blocksize; j++)
  132. for (i = 0; i < blocksize; i++)
  133. {
  134. A_blocks[bi+nblocks*bj][i + j * blocksize] =
  135. A[(i+bi*blocksize) + (j + bj*blocksize)*size];
  136. }
  137. }
  138. }
  139. static void init_matrix(void)
  140. {
  141. /* allocate matrix */
  142. starpu_malloc((void **)&A, (size_t)size*size*sizeof(TYPE));
  143. STARPU_ASSERT(A);
  144. starpu_srand48((long int)time(NULL));
  145. /* starpu_srand48(0); */
  146. /* initialize matrix content */
  147. unsigned long i,j;
  148. for (j = 0; j < size; j++)
  149. {
  150. for (i = 0; i < size; i++)
  151. {
  152. A[i + j*size] = (TYPE)starpu_drand48();
  153. #ifdef COMPLEX_LU
  154. /* also randomize the imaginary component for complex number cases */
  155. A[i + j*size] += (TYPE)(I*starpu_drand48());
  156. #endif
  157. }
  158. }
  159. }
  160. static void save_matrix(void)
  161. {
  162. A_saved = malloc((size_t)size*size*sizeof(TYPE));
  163. STARPU_ASSERT(A_saved);
  164. memcpy(A_saved, A, (size_t)size*size*sizeof(TYPE));
  165. }
  166. static double frobenius_norm(TYPE *v, unsigned n)
  167. {
  168. double sum2 = 0.0;
  169. /* compute sqrt(Sum(|x|^2)) */
  170. unsigned i,j;
  171. for (j = 0; j < n; j++)
  172. for (i = 0; i < n; i++)
  173. {
  174. double a = fabsl((double)v[i+n*j]);
  175. sum2 += a*a;
  176. }
  177. return sqrt(sum2);
  178. }
  179. static void pivot_saved_matrix(unsigned *ipiv)
  180. {
  181. unsigned k;
  182. for (k = 0; k < size; k++)
  183. {
  184. if (k != ipiv[k])
  185. {
  186. /* FPRINTF(stderr, "SWAP %d and %d\n", k, ipiv[k]); */
  187. CPU_SWAP(size, &A_saved[k*size], 1, &A_saved[ipiv[k]*size], 1);
  188. }
  189. }
  190. }
  191. static void check_result(void)
  192. {
  193. unsigned i,j;
  194. TYPE *L, *U;
  195. L = malloc((size_t)size*size*sizeof(TYPE));
  196. U = malloc((size_t)size*size*sizeof(TYPE));
  197. memset(L, 0, size*size*sizeof(TYPE));
  198. memset(U, 0, size*size*sizeof(TYPE));
  199. /* only keep the lower part */
  200. for (j = 0; j < size; j++)
  201. {
  202. for (i = 0; i < j; i++)
  203. {
  204. L[j+i*size] = A[j+i*size];
  205. }
  206. /* diag i = j */
  207. L[j+j*size] = A[j+j*size];
  208. U[j+j*size] = 1.0;
  209. for (i = j+1; i < size; i++)
  210. {
  211. U[j+i*size] = A[j+i*size];
  212. }
  213. }
  214. display_matrix(L, size, size, "L");
  215. display_matrix(U, size, size, "U");
  216. /* now A_err = L, compute L*U */
  217. CPU_TRMM("R", "U", "N", "U", size, size, 1.0f, U, size, L, size);
  218. display_matrix(A_saved, size, size, "P A_saved");
  219. display_matrix(L, size, size, "LU");
  220. /* compute "LU - A" in L*/
  221. CPU_AXPY(size*size, -1.0, A_saved, 1, L, 1);
  222. display_matrix(L, size, size, "Residuals");
  223. #ifdef COMPLEX_LU
  224. double err = CPU_ASUM(size*size, L, 1);
  225. int max = CPU_IAMAX(size*size, L, 1);
  226. TYPE l_max = L[max];
  227. FPRINTF(stderr, "Avg error : %e\n", err/(size*size));
  228. FPRINTF(stderr, "Max error : %e\n", sqrt(creal(l_max)*creal(l_max)+cimag(l_max)*cimag(l_max)));
  229. #else
  230. TYPE err = CPU_ASUM(size*size, L, 1);
  231. int max = CPU_IAMAX(size*size, L, 1);
  232. FPRINTF(stderr, "Avg error : %e\n", err/(size*size));
  233. FPRINTF(stderr, "Max error : %e\n", L[max]);
  234. #endif
  235. double residual = frobenius_norm(L, size);
  236. double matnorm = frobenius_norm(A_saved, size);
  237. FPRINTF(stderr, "||%sA-LU|| / (||A||*N) : %e\n", pivot?"P":"", residual/(matnorm*size));
  238. if (residual/(matnorm*size) > 1e-5)
  239. exit(-1);
  240. }
  241. int main(int argc, char **argv)
  242. {
  243. int ret;
  244. parse_args(argc, argv);
  245. ret = starpu_init(NULL);
  246. STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
  247. starpu_helper_cublas_init();
  248. init_matrix();
  249. unsigned *ipiv = NULL;
  250. if (check)
  251. save_matrix();
  252. display_matrix(A, size, size, "A");
  253. if (bound)
  254. starpu_bound_start(bounddeps, boundprio);
  255. if (profile)
  256. starpu_profiling_status_set(STARPU_PROFILING_ENABLE);
  257. /* Factorize the matrix (in place) */
  258. if (pivot)
  259. {
  260. ipiv = malloc(size*sizeof(unsigned));
  261. if (no_stride)
  262. {
  263. /* in case the LU decomposition uses non-strided blocks, we _copy_ the matrix into smaller blocks */
  264. A_blocks = malloc(nblocks*nblocks*sizeof(TYPE **));
  265. copy_matrix_into_blocks();
  266. STARPU_LU(lu_decomposition_pivot_no_stride)(A_blocks, ipiv, size, size, nblocks);
  267. copy_blocks_into_matrix();
  268. free(A_blocks);
  269. }
  270. else
  271. {
  272. struct timeval start;
  273. struct timeval end;
  274. gettimeofday(&start, NULL);
  275. STARPU_LU(lu_decomposition_pivot)(A, ipiv, size, size, nblocks);
  276. gettimeofday(&end, NULL);
  277. double timing = (double)((end.tv_sec - start.tv_sec)*1000000 + (end.tv_usec - start.tv_usec));
  278. unsigned n = size;
  279. double flop = (2.0f*n*n*n)/3.0f;
  280. FPRINTF(stderr, "Synthetic GFlops (TOTAL) : \n");
  281. FPRINTF(stdout, "%u %6.2f\n", n, (flop/timing/1000.0f));
  282. }
  283. }
  284. else
  285. {
  286. STARPU_LU(lu_decomposition)(A, size, size, nblocks);
  287. }
  288. if (profile)
  289. {
  290. FPRINTF(stderr, "Setting profile\n");
  291. starpu_profiling_status_set(STARPU_PROFILING_DISABLE);
  292. starpu_bus_profiling_helper_display_summary();
  293. }
  294. if (bound)
  295. {
  296. double min;
  297. FPRINTF(stderr, "Setting bound\n");
  298. starpu_bound_stop();
  299. if (bounddeps)
  300. {
  301. FILE *f = fopen("lu.pl", "w");
  302. starpu_bound_print_lp(f);
  303. FPRINTF(stderr,"system printed to lu.pl\n");
  304. }
  305. else
  306. {
  307. starpu_bound_compute(&min, NULL, 0);
  308. if (min != 0.)
  309. FPRINTF(stderr, "theoretical min: %f ms\n", min);
  310. }
  311. }
  312. if (check)
  313. {
  314. FPRINTF(stderr, "Checking result\n");
  315. if (pivot)
  316. pivot_saved_matrix(ipiv);
  317. check_result();
  318. }
  319. FPRINTF(stderr, "Shutting down\n");
  320. starpu_helper_cublas_shutdown();
  321. starpu_shutdown();
  322. return 0;
  323. }