plu_implicit_example.c 8.1 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370
  1. /* StarPU --- Runtime system for heterogeneous multicore architectures.
  2. *
  3. * Copyright (C) 2010-2011, 2013, 2017 Université de Bordeaux
  4. * Copyright (C) 2010, 2011, 2012, 2013, 2014, 2016, 2017 CNRS
  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 "pxlu.h"
  24. //#include "pxlu_kernels.h"
  25. #ifdef STARPU_HAVE_LIBNUMA
  26. #include <numaif.h>
  27. #endif
  28. static unsigned long size = 4096;
  29. static unsigned nblocks = 16;
  30. static unsigned check = 0;
  31. static int p = 1;
  32. static int q = 1;
  33. static unsigned display = 0;
  34. #ifdef STARPU_HAVE_LIBNUMA
  35. static unsigned numa = 0;
  36. #endif
  37. static size_t allocated_memory = 0;
  38. static size_t allocated_memory_extra = 0;
  39. static starpu_data_handle_t *dataA_handles;
  40. static TYPE **dataA;
  41. int get_block_rank(unsigned i, unsigned j);
  42. static void parse_args(int argc, char **argv)
  43. {
  44. int i;
  45. for (i = 1; i < argc; i++)
  46. {
  47. if (strcmp(argv[i], "-size") == 0)
  48. {
  49. char *argptr;
  50. size = strtol(argv[++i], &argptr, 10);
  51. }
  52. if (strcmp(argv[i], "-nblocks") == 0)
  53. {
  54. char *argptr;
  55. nblocks = strtol(argv[++i], &argptr, 10);
  56. }
  57. if (strcmp(argv[i], "-check") == 0)
  58. {
  59. check = 1;
  60. }
  61. if (strcmp(argv[i], "-display") == 0)
  62. {
  63. display = 1;
  64. }
  65. if (strcmp(argv[i], "-numa") == 0)
  66. {
  67. #ifdef STARPU_HAVE_LIBNUMA
  68. numa = 1;
  69. #else
  70. fprintf(stderr, "Warning: libnuma is not available\n");
  71. #endif
  72. }
  73. if (strcmp(argv[i], "-p") == 0)
  74. {
  75. char *argptr;
  76. p = strtol(argv[++i], &argptr, 10);
  77. }
  78. if (strcmp(argv[i], "-q") == 0)
  79. {
  80. char *argptr;
  81. q = strtol(argv[++i], &argptr, 10);
  82. }
  83. if (strcmp(argv[i], "-h") == 0 || strcmp(argv[i], "-help") == 0 || strcmp(argv[i], "--help") == 0)
  84. {
  85. fprintf(stderr,"usage: %s [-size n] [-nblocks b] [-check] [-display] [-numa] [-p p] [-q q]\n", argv[0]);
  86. fprintf(stderr,"\np * q must be equal to the number of MPI nodes\n");
  87. exit(0);
  88. }
  89. }
  90. }
  91. unsigned STARPU_PLU(display_flag)(void)
  92. {
  93. return display;
  94. }
  95. static void fill_block_with_random(TYPE *blockptr, unsigned psize, unsigned pnblocks)
  96. {
  97. const unsigned block_size = (psize/pnblocks);
  98. unsigned i, j;
  99. for (i = 0; i < block_size; i++)
  100. for (j = 0; j < block_size; j++)
  101. {
  102. blockptr[j+i*block_size] = (TYPE)starpu_drand48();
  103. }
  104. }
  105. static void init_matrix(int rank)
  106. {
  107. #ifdef STARPU_HAVE_LIBNUMA
  108. if (numa)
  109. {
  110. fprintf(stderr, "Using INTERLEAVE policy\n");
  111. unsigned long nodemask = ((1<<0)|(1<<1));
  112. int ret = set_mempolicy(MPOL_INTERLEAVE, &nodemask, 3);
  113. if (ret)
  114. perror("set_mempolicy failed");
  115. }
  116. #endif
  117. /* Allocate a grid of data handles, not all of them have to be allocated later on */
  118. dataA_handles = calloc(nblocks*nblocks, sizeof(starpu_data_handle_t));
  119. dataA = calloc(nblocks*nblocks, sizeof(TYPE *));
  120. allocated_memory_extra += nblocks*nblocks*(sizeof(starpu_data_handle_t) + sizeof(TYPE *));
  121. size_t blocksize = (size_t)(size/nblocks)*(size/nblocks)*sizeof(TYPE);
  122. /* Allocate all the blocks that belong to this mpi node */
  123. unsigned long i,j;
  124. for (j = 0; j < nblocks; j++)
  125. {
  126. for (i = 0; i < nblocks; i++)
  127. {
  128. int block_rank = get_block_rank(i, j);
  129. TYPE **blockptr = &dataA[j+i*nblocks];
  130. // starpu_data_handle_t *handleptr = &dataA_handles[j+nblocks*i];
  131. starpu_data_handle_t *handleptr = &dataA_handles[j+nblocks*i];
  132. if (block_rank == rank)
  133. {
  134. /* This blocks should be treated by the current MPI process */
  135. /* Allocate and fill it */
  136. starpu_malloc((void **)blockptr, blocksize);
  137. allocated_memory += blocksize;
  138. //fprintf(stderr, "Rank %d : fill block (i = %d, j = %d)\n", rank, i, j);
  139. fill_block_with_random(*blockptr, size, nblocks);
  140. //fprintf(stderr, "Rank %d : fill block (i = %d, j = %d)\n", rank, i, j);
  141. if (i == j)
  142. {
  143. unsigned tmp;
  144. for (tmp = 0; tmp < size/nblocks; tmp++)
  145. {
  146. (*blockptr)[tmp*((size/nblocks)+1)] += (TYPE)10*nblocks;
  147. }
  148. }
  149. /* Register it to StarPU */
  150. starpu_matrix_data_register(handleptr, STARPU_MAIN_RAM,
  151. (uintptr_t)*blockptr, size/nblocks,
  152. size/nblocks, size/nblocks, sizeof(TYPE));
  153. }
  154. else
  155. {
  156. starpu_matrix_data_register(handleptr, -1,
  157. 0, size/nblocks,
  158. size/nblocks, size/nblocks, sizeof(TYPE));
  159. *blockptr = STARPU_POISON_PTR;
  160. }
  161. starpu_data_set_coordinates(*handleptr, 2, j, i);
  162. starpu_mpi_data_register(*handleptr, j+i*nblocks, block_rank);
  163. }
  164. }
  165. //display_all_blocks(nblocks, size/nblocks);
  166. }
  167. TYPE *STARPU_PLU(get_block)(unsigned i, unsigned j)
  168. {
  169. return dataA[j+i*nblocks];
  170. }
  171. int get_block_rank(unsigned i, unsigned j)
  172. {
  173. /* Take a 2D block cyclic distribution */
  174. /* NB: p (resp. q) is for "direction" i (resp. j) */
  175. return (j % q) * p + (i % p);
  176. }
  177. starpu_data_handle_t STARPU_PLU(get_block_handle)(unsigned i, unsigned j)
  178. {
  179. return dataA_handles[j+i*nblocks];
  180. }
  181. static void display_grid(int rank, unsigned pnblocks)
  182. {
  183. if (!display)
  184. return;
  185. //if (rank == 0)
  186. {
  187. fprintf(stderr, "2D grid layout (Rank %d): \n", rank);
  188. unsigned i, j;
  189. for (j = 0; j < pnblocks; j++)
  190. {
  191. for (i = 0; i < pnblocks; i++)
  192. {
  193. TYPE *blockptr = STARPU_PLU(get_block)(i, j);
  194. starpu_data_handle_t handle = STARPU_PLU(get_block_handle)(i, j);
  195. fprintf(stderr, "%d (data %p handle %p)", get_block_rank(i, j), blockptr, handle);
  196. }
  197. fprintf(stderr, "\n");
  198. }
  199. }
  200. }
  201. int main(int argc, char **argv)
  202. {
  203. int rank;
  204. int world_size;
  205. starpu_srand48((long int)time(NULL));
  206. parse_args(argc, argv);
  207. int ret = starpu_init(NULL);
  208. STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
  209. ret = starpu_mpi_init(&argc, &argv, 1);
  210. STARPU_CHECK_RETURN_VALUE(ret, "starpu_mpi_init");
  211. starpu_mpi_comm_rank(MPI_COMM_WORLD, &rank);
  212. starpu_mpi_comm_size(MPI_COMM_WORLD, &world_size);
  213. STARPU_ASSERT(p*q == world_size);
  214. starpu_cublas_init();
  215. /*
  216. * Problem Init
  217. */
  218. init_matrix(rank);
  219. fprintf(stderr, "Rank %d: allocated (%d + %d) MB = %d MB\n", rank,
  220. (int)(allocated_memory/(1024*1024)),
  221. (int)(allocated_memory_extra/(1024*1024)),
  222. (int)((allocated_memory+allocated_memory_extra)/(1024*1024)));
  223. display_grid(rank, nblocks);
  224. TYPE *a_r = NULL;
  225. // STARPU_PLU(display_data_content)(a_r, size);
  226. if (check)
  227. {
  228. TYPE *x, *y;
  229. x = calloc(size, sizeof(TYPE));
  230. STARPU_ASSERT(x);
  231. y = calloc(size, sizeof(TYPE));
  232. STARPU_ASSERT(y);
  233. if (rank == 0)
  234. {
  235. unsigned ind;
  236. for (ind = 0; ind < size; ind++)
  237. x[ind] = (TYPE)starpu_drand48();
  238. }
  239. a_r = STARPU_PLU(reconstruct_matrix)(size, nblocks);
  240. if (rank == 0)
  241. STARPU_PLU(display_data_content)(a_r, size);
  242. // STARPU_PLU(compute_ax)(size, x, y, nblocks, rank);
  243. free(x);
  244. free(y);
  245. }
  246. double timing = STARPU_PLU(plu_main)(nblocks, rank, world_size);
  247. /*
  248. * Report performance
  249. */
  250. if (rank == 0)
  251. {
  252. fprintf(stderr, "Computation took: %f ms\n", timing/1000);
  253. unsigned n = size;
  254. double flop = (2.0f*n*n*n)/3.0f;
  255. fprintf(stderr, "Synthetic GFlops : %2.2f\n", (flop/timing/1000.0f));
  256. }
  257. /*
  258. * Test Result Correctness
  259. */
  260. if (check)
  261. {
  262. /*
  263. * Compute || A - LU ||
  264. */
  265. STARPU_PLU(compute_lu_matrix)(size, nblocks, a_r);
  266. #if 0
  267. /*
  268. * Compute || Ax - LUx ||
  269. */
  270. unsigned ind;
  271. y2 = calloc(size, sizeof(TYPE));
  272. STARPU_ASSERT(y);
  273. if (rank == 0)
  274. {
  275. for (ind = 0; ind < size; ind++)
  276. {
  277. y2[ind] = (TYPE)0.0;
  278. }
  279. }
  280. STARPU_PLU(compute_lux)(size, x, y2, nblocks, rank);
  281. /* Compute y2 = y2 - y */
  282. CPU_AXPY(size, -1.0, y, 1, y2, 1);
  283. TYPE err = CPU_ASUM(size, y2, 1);
  284. int max = CPU_IAMAX(size, y2, 1);
  285. fprintf(stderr, "(A - LU)X Avg error : %e\n", err/(size*size));
  286. fprintf(stderr, "(A - LU)X Max error : %e\n", y2[max]);
  287. #endif
  288. }
  289. /*
  290. * Termination
  291. */
  292. starpu_cublas_shutdown();
  293. starpu_mpi_shutdown();
  294. starpu_shutdown();
  295. return 0;
  296. }