datatypes.c 18 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620
  1. /* StarPU --- Runtime system for heterogeneous multicore architectures.
  2. *
  3. * Copyright (C) 2013-2021 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. #include <starpu_mpi.h>
  17. #include <stdlib.h>
  18. #include "helper.h"
  19. typedef void (*check_func)(starpu_data_handle_t handle_s, starpu_data_handle_t handle_r, int *error);
  20. void send_recv_and_check(int rank, int node, starpu_data_handle_t handle_s, int tag_s, starpu_data_handle_t handle_r, int tag_r, int *error, check_func func)
  21. {
  22. int ret;
  23. MPI_Status status;
  24. if (rank == 0)
  25. {
  26. ret = starpu_mpi_send(handle_s, node, tag_s, MPI_COMM_WORLD);
  27. STARPU_CHECK_RETURN_VALUE(ret, "starpu_mpi_send");
  28. ret = starpu_mpi_recv(handle_r, node, tag_r, MPI_COMM_WORLD, &status);
  29. STARPU_CHECK_RETURN_VALUE(ret, "starpu_mpi_recv");
  30. assert(func);
  31. func(handle_s, handle_r, error);
  32. }
  33. else if (rank == 1)
  34. {
  35. ret = starpu_mpi_recv(handle_s, node, tag_s, MPI_COMM_WORLD, &status);
  36. STARPU_CHECK_RETURN_VALUE(ret, "starpu_mpi_recv");
  37. ret = starpu_mpi_send(handle_s, node, tag_r, MPI_COMM_WORLD);
  38. STARPU_CHECK_RETURN_VALUE(ret, "starpu_mpi_send");
  39. }
  40. }
  41. /*
  42. * Void
  43. */
  44. void check_void(starpu_data_handle_t handle_s, starpu_data_handle_t handle_r, int *error)
  45. {
  46. (void)error;
  47. (void)handle_s;
  48. (void)handle_r;
  49. FPRINTF_MPI(stderr, "Success with void value\n");
  50. }
  51. void exchange_void(int rank, int *error)
  52. {
  53. STARPU_SKIP_IF_VALGRIND;
  54. if (rank == 0)
  55. {
  56. starpu_data_handle_t void_handle[2];
  57. starpu_void_data_register(&void_handle[0]);
  58. starpu_void_data_register(&void_handle[1]);
  59. send_recv_and_check(rank, 1, void_handle[0], 0x42, void_handle[1], 0x1337, error, check_void);
  60. starpu_data_unregister(void_handle[0]);
  61. starpu_data_unregister(void_handle[1]);
  62. }
  63. else if (rank == 1)
  64. {
  65. starpu_data_handle_t void_handle;
  66. starpu_void_data_register(&void_handle);
  67. send_recv_and_check(rank, 0, void_handle, 0x42, NULL, 0x1337, NULL, NULL);
  68. starpu_data_unregister(void_handle);
  69. }
  70. }
  71. /*
  72. * Variable
  73. */
  74. void check_variable(starpu_data_handle_t handle_s, starpu_data_handle_t handle_r, int *error)
  75. {
  76. float *v_s, *v_r;
  77. STARPU_ASSERT(starpu_variable_get_elemsize(handle_s) == starpu_variable_get_elemsize(handle_r));
  78. starpu_data_acquire(handle_s, STARPU_R);
  79. v_s = (float *)starpu_variable_get_local_ptr(handle_s);
  80. starpu_data_release(handle_s);
  81. starpu_data_acquire(handle_r, STARPU_R);
  82. v_r = (float *)starpu_variable_get_local_ptr(handle_r);
  83. starpu_data_release(handle_r);
  84. if (*v_s == *v_r)
  85. {
  86. FPRINTF_MPI(stderr, "Success with variable value: %f == %f\n", *v_s, *v_r);
  87. }
  88. else
  89. {
  90. *error = 1;
  91. FPRINTF_MPI(stderr, "Error with variable value: %f != %f\n", *v_s, *v_r);
  92. }
  93. }
  94. void exchange_variable(int rank, int *error)
  95. {
  96. if (rank == 0)
  97. {
  98. float v = 42.12;
  99. starpu_data_handle_t variable_handle[2];
  100. starpu_variable_data_register(&variable_handle[0], STARPU_MAIN_RAM, (uintptr_t)&v, sizeof(v));
  101. starpu_variable_data_register(&variable_handle[1], -1, (uintptr_t)NULL, sizeof(v));
  102. send_recv_and_check(rank, 1, variable_handle[0], 0x42, variable_handle[1], 0x1337, error, check_variable);
  103. starpu_data_unregister(variable_handle[0]);
  104. starpu_data_unregister(variable_handle[1]);
  105. }
  106. else if (rank == 1)
  107. {
  108. starpu_data_handle_t variable_handle;
  109. starpu_variable_data_register(&variable_handle, -1, (uintptr_t)NULL, sizeof(float));
  110. send_recv_and_check(rank, 0, variable_handle, 0x42, NULL, 0x1337, NULL, NULL);
  111. starpu_data_unregister(variable_handle);
  112. }
  113. }
  114. /*
  115. * Vector
  116. */
  117. void check_vector(starpu_data_handle_t handle_s, starpu_data_handle_t handle_r, int *error)
  118. {
  119. int i;
  120. int nx;
  121. int *v_r, *v_s;
  122. STARPU_ASSERT(starpu_vector_get_elemsize(handle_s) == starpu_vector_get_elemsize(handle_r));
  123. STARPU_ASSERT(starpu_vector_get_nx(handle_s) == starpu_vector_get_nx(handle_r));
  124. nx = starpu_vector_get_nx(handle_r);
  125. v_r = (int *)starpu_vector_get_local_ptr(handle_r);
  126. v_s = (int *)starpu_vector_get_local_ptr(handle_s);
  127. for(i=0 ; i<nx ; i++)
  128. {
  129. if (v_s[i] == v_r[i])
  130. {
  131. FPRINTF_MPI(stderr, "Success with vector[%d] value: %d == %d\n", i, v_s[i], v_r[i]);
  132. }
  133. else
  134. {
  135. *error = 1;
  136. FPRINTF_MPI(stderr, "Error with vector[%d] value: %d != %d\n", i, v_s[i], v_r[i]);
  137. }
  138. }
  139. }
  140. void exchange_vector(int rank, int *error)
  141. {
  142. if (rank == 0)
  143. {
  144. int vector[4] = {1, 2, 3, 4};
  145. starpu_data_handle_t vector_handle[2];
  146. starpu_vector_data_register(&vector_handle[0], STARPU_MAIN_RAM, (uintptr_t)vector, 4, sizeof(vector[0]));
  147. starpu_vector_data_register(&vector_handle[1], -1, (uintptr_t)NULL, 4, sizeof(vector[0]));
  148. send_recv_and_check(rank, 1, vector_handle[0], 0x43, vector_handle[1], 0x2337, error, check_vector);
  149. starpu_data_unregister(vector_handle[0]);
  150. starpu_data_unregister(vector_handle[1]);
  151. }
  152. else if (rank == 1)
  153. {
  154. starpu_data_handle_t vector_handle;
  155. starpu_vector_data_register(&vector_handle, -1, (uintptr_t)NULL, 4, sizeof(int));
  156. send_recv_and_check(rank, 0, vector_handle, 0x43, NULL, 0x2337, NULL, NULL);
  157. starpu_data_unregister(vector_handle);
  158. }
  159. }
  160. /*
  161. * Matrix
  162. */
  163. void check_matrix(starpu_data_handle_t handle_s, starpu_data_handle_t handle_r, int *error)
  164. {
  165. STARPU_ASSERT(starpu_matrix_get_elemsize(handle_s) == starpu_matrix_get_elemsize(handle_r));
  166. STARPU_ASSERT(starpu_matrix_get_nx(handle_s) == starpu_matrix_get_nx(handle_r));
  167. STARPU_ASSERT(starpu_matrix_get_ny(handle_s) == starpu_matrix_get_ny(handle_r));
  168. STARPU_ASSERT(starpu_matrix_get_local_ld(handle_s) == starpu_matrix_get_local_ld(handle_r));
  169. char *matrix_s = (char *)starpu_matrix_get_local_ptr(handle_s);
  170. char *matrix_r = (char *)starpu_matrix_get_local_ptr(handle_r);
  171. int nx = starpu_matrix_get_nx(handle_s);
  172. int ny = starpu_matrix_get_ny(handle_s);
  173. int ldy = starpu_matrix_get_local_ld(handle_s);
  174. int x, y;
  175. for(y=0 ; y<ny ; y++)
  176. {
  177. for(x=0 ; x<nx ; x++)
  178. {
  179. int index=(y*ldy)+x;
  180. if (matrix_s[index] == matrix_r[index])
  181. {
  182. FPRINTF_MPI(stderr, "Success with matrix[%d,%d --> %d] value: %c == %c\n", x, y, index, matrix_s[index], matrix_r[index]);
  183. }
  184. else
  185. {
  186. *error = 1;
  187. FPRINTF_MPI(stderr, "Error with matrix[%d,%d --> %d] value: %c != %c\n", x, y, index, matrix_s[index], matrix_r[index]);
  188. }
  189. }
  190. }
  191. }
  192. void exchange_matrix(int rank, int *error)
  193. {
  194. int nx=3;
  195. int ny=2;
  196. if (rank == 0)
  197. {
  198. char *matrix, n='a';
  199. int x, y;
  200. starpu_data_handle_t matrix_handle[2];
  201. matrix = (char*)malloc(nx*ny*sizeof(char));
  202. assert(matrix);
  203. for(y=0 ; y<ny ; y++)
  204. {
  205. for(x=0 ; x<nx ; x++)
  206. {
  207. matrix[(y*nx)+x] = n++;
  208. }
  209. }
  210. starpu_matrix_data_register(&matrix_handle[0], STARPU_MAIN_RAM, (uintptr_t)matrix, nx, nx, ny, sizeof(char));
  211. starpu_matrix_data_register(&matrix_handle[1], -1, (uintptr_t)NULL, nx, nx, ny, sizeof(char));
  212. send_recv_and_check(rank, 1, matrix_handle[0], 0x75, matrix_handle[1], 0x8555, error, check_matrix);
  213. starpu_data_unregister(matrix_handle[0]);
  214. starpu_data_unregister(matrix_handle[1]);
  215. free(matrix);
  216. }
  217. else if (rank == 1)
  218. {
  219. starpu_data_handle_t matrix_handle;
  220. starpu_matrix_data_register(&matrix_handle, -1, (uintptr_t)NULL, nx, nx, ny, sizeof(char));
  221. send_recv_and_check(rank, 0, matrix_handle, 0x75, NULL, 0x8555, NULL, NULL);
  222. starpu_data_unregister(matrix_handle);
  223. }
  224. }
  225. /*
  226. * Block
  227. */
  228. void check_block(starpu_data_handle_t handle_s, starpu_data_handle_t handle_r, int *error)
  229. {
  230. STARPU_ASSERT(starpu_block_get_elemsize(handle_s) == starpu_block_get_elemsize(handle_r));
  231. STARPU_ASSERT(starpu_block_get_nx(handle_s) == starpu_block_get_nx(handle_r));
  232. STARPU_ASSERT(starpu_block_get_ny(handle_s) == starpu_block_get_ny(handle_r));
  233. STARPU_ASSERT(starpu_block_get_nz(handle_s) == starpu_block_get_nz(handle_r));
  234. STARPU_ASSERT(starpu_block_get_local_ldy(handle_s) == starpu_block_get_local_ldy(handle_r));
  235. STARPU_ASSERT(starpu_block_get_local_ldz(handle_s) == starpu_block_get_local_ldz(handle_r));
  236. starpu_data_acquire(handle_s, STARPU_R);
  237. starpu_data_acquire(handle_r, STARPU_R);
  238. float *block_s = (float *)starpu_block_get_local_ptr(handle_s);
  239. float *block_r = (float *)starpu_block_get_local_ptr(handle_r);
  240. int nx = starpu_block_get_nx(handle_s);
  241. int ny = starpu_block_get_ny(handle_s);
  242. int nz = starpu_block_get_nz(handle_s);
  243. int ldy = starpu_block_get_local_ldy(handle_s);
  244. int ldz = starpu_block_get_local_ldz(handle_s);
  245. int x, y, z;
  246. for(z=0 ; z<nz ; z++)
  247. {
  248. for(y=0 ; y<ny ; y++)
  249. for(x=0 ; x<nx ; x++)
  250. {
  251. int index=(z*ldz)+(y*ldy)+x;
  252. if (block_s[index] == block_r[index])
  253. {
  254. FPRINTF_MPI(stderr, "Success with block[%d,%d,%d --> %d] value: %f == %f\n", x, y, z, index, block_s[index], block_r[index]);
  255. }
  256. else
  257. {
  258. *error = 1;
  259. FPRINTF_MPI(stderr, "Error with block[%d,%d,%d --> %d] value: %f != %f\n", x, y, z, index, block_s[index], block_r[index]);
  260. }
  261. }
  262. }
  263. starpu_data_release(handle_s);
  264. starpu_data_release(handle_r);
  265. }
  266. void exchange_block(int rank, int *error)
  267. {
  268. int nx=3;
  269. int ny=2;
  270. int nz=4;
  271. if (rank == 0)
  272. {
  273. float *block, n=1.0;
  274. int x, y, z;
  275. starpu_data_handle_t block_handle[2];
  276. block = (float*)malloc(nx*ny*nz*sizeof(float));
  277. assert(block);
  278. for(z=0 ; z<nz ; z++)
  279. {
  280. for(y=0 ; y<ny ; y++)
  281. {
  282. for(x=0 ; x<nx ; x++)
  283. {
  284. block[(z*nx*ny)+(y*nx)+x] = n++;
  285. }
  286. }
  287. }
  288. starpu_block_data_register(&block_handle[0], STARPU_MAIN_RAM, (uintptr_t)block, nx, nx*ny, nx, ny, nz, sizeof(float));
  289. starpu_block_data_register(&block_handle[1], -1, (uintptr_t)NULL, nx, nx*ny, nx, ny, nz, sizeof(float));
  290. send_recv_and_check(rank, 1, block_handle[0], 0x73, block_handle[1], 0x8337, error, check_block);
  291. starpu_data_unregister(block_handle[0]);
  292. starpu_data_unregister(block_handle[1]);
  293. free(block);
  294. }
  295. else if (rank == 1)
  296. {
  297. starpu_data_handle_t block_handle;
  298. starpu_block_data_register(&block_handle, -1, (uintptr_t)NULL, nx, nx*ny, nx, ny, nz, sizeof(float));
  299. send_recv_and_check(rank, 0, block_handle, 0x73, NULL, 0x8337, NULL, NULL);
  300. starpu_data_unregister(block_handle);
  301. }
  302. }
  303. /*
  304. * BCSR
  305. */
  306. void check_bcsr(starpu_data_handle_t handle_s, starpu_data_handle_t handle_r, int *error)
  307. {
  308. STARPU_ASSERT(starpu_bcsr_get_elemsize(handle_s) == starpu_bcsr_get_elemsize(handle_r));
  309. STARPU_ASSERT(starpu_bcsr_get_nnz(handle_s) == starpu_bcsr_get_nnz(handle_r));
  310. STARPU_ASSERT(starpu_bcsr_get_nrow(handle_s) == starpu_bcsr_get_nrow(handle_r));
  311. STARPU_ASSERT(starpu_bcsr_get_firstentry(handle_s) == starpu_bcsr_get_firstentry(handle_r));
  312. STARPU_ASSERT(starpu_bcsr_get_r(handle_s) == starpu_bcsr_get_r(handle_r));
  313. STARPU_ASSERT(starpu_bcsr_get_c(handle_s) == starpu_bcsr_get_c(handle_r));
  314. starpu_data_acquire(handle_s, STARPU_R);
  315. starpu_data_acquire(handle_r, STARPU_R);
  316. uint32_t *colind_s = starpu_bcsr_get_local_colind(handle_s);
  317. uint32_t *colind_r = starpu_bcsr_get_local_colind(handle_r);
  318. uint32_t *rowptr_s = starpu_bcsr_get_local_rowptr(handle_s);
  319. uint32_t *rowptr_r = starpu_bcsr_get_local_rowptr(handle_r);
  320. int *bcsr_s = (int *)starpu_bcsr_get_local_nzval(handle_s);
  321. int *bcsr_r = (int *)starpu_bcsr_get_local_nzval(handle_r);
  322. int r = starpu_bcsr_get_r(handle_s);
  323. int c = starpu_bcsr_get_c(handle_s);
  324. int nnz = starpu_bcsr_get_nnz(handle_s);
  325. int nrows = starpu_bcsr_get_nrow(handle_s);
  326. int x;
  327. for(x=0 ; x<nnz ; x++)
  328. {
  329. if (colind_s[x] == colind_r[x])
  330. {
  331. FPRINTF_MPI(stderr, "Success with colind[%d] value: %u == %u\n", x, colind_s[x], colind_r[x]);
  332. }
  333. else
  334. {
  335. *error = 1;
  336. FPRINTF_MPI(stderr, "Error with colind[%d] value: %u != %u\n", x, colind_s[x], colind_r[x]);
  337. }
  338. }
  339. for(x=0 ; x<nrows+1 ; x++)
  340. {
  341. if (rowptr_s[x] == rowptr_r[x])
  342. {
  343. FPRINTF_MPI(stderr, "Success with rowptr[%d] value: %u == %u\n", x, rowptr_s[x], rowptr_r[x]);
  344. }
  345. else
  346. {
  347. *error = 1;
  348. FPRINTF_MPI(stderr, "Error with rowptr[%d] value: %u != %u\n", x, rowptr_s[x], rowptr_r[x]);
  349. }
  350. }
  351. for(x=0 ; x<r*c*nnz ; x++)
  352. {
  353. if (bcsr_s[x] == bcsr_r[x])
  354. {
  355. FPRINTF_MPI(stderr, "Success with bcsr[%d] value: %d == %d\n", x, bcsr_s[x], bcsr_r[x]);
  356. }
  357. else
  358. {
  359. *error = 1;
  360. FPRINTF_MPI(stderr, "Error with bcsr[%d] value: %d != %d\n", x, bcsr_s[x], bcsr_r[x]);
  361. }
  362. }
  363. starpu_data_release(handle_s);
  364. starpu_data_release(handle_r);
  365. }
  366. void exchange_bcsr(int rank, int *error)
  367. {
  368. /*
  369. * We use the following matrix:
  370. *
  371. * +----------------+
  372. * | 0 1 0 0 |
  373. * | 2 3 0 0 |
  374. * | 4 5 8 9 |
  375. * | 6 7 10 11 |
  376. * +----------------+
  377. *
  378. * nzval = [0, 1, 2, 3] ++ [4, 5, 6, 7] ++ [8, 9, 10, 11]
  379. * colind = [0, 0, 1]
  380. * rowptr = [0, 1, 3]
  381. * r = c = 2
  382. */
  383. /* Size of the blocks */
  384. #define BCSR_R 2
  385. #define BCSR_C 2
  386. #define BCSR_NROWS 2
  387. #define BCSR_NNZ_BLOCKS 3 /* out of 4 */
  388. #define BCSR_NZVAL_SIZE (BCSR_R*BCSR_C*BCSR_NNZ_BLOCKS)
  389. if (rank == 0)
  390. {
  391. starpu_data_handle_t bcsr_handle[2];
  392. uint32_t colind[BCSR_NNZ_BLOCKS] = {0, 0, 1};
  393. uint32_t rowptr[BCSR_NROWS+1] = {0, 1, BCSR_NNZ_BLOCKS};
  394. int nzval[BCSR_NZVAL_SIZE] =
  395. {
  396. 0, 1, 2, 3, /* First block */
  397. 4, 5, 6, 7, /* Second block */
  398. 8, 9, 10, 11 /* Third block */
  399. };
  400. starpu_bcsr_data_register(&bcsr_handle[0], STARPU_MAIN_RAM, BCSR_NNZ_BLOCKS, BCSR_NROWS, (uintptr_t) nzval, colind, rowptr, 0, BCSR_R, BCSR_C, sizeof(nzval[0]));
  401. starpu_bcsr_data_register(&bcsr_handle[1], -1, BCSR_NNZ_BLOCKS, BCSR_NROWS, (uintptr_t) NULL, (uint32_t *) NULL, (uint32_t *) NULL, 0, BCSR_R, BCSR_C, sizeof(nzval[0]));
  402. send_recv_and_check(rank, 1, bcsr_handle[0], 0x73, bcsr_handle[1], 0x8337, error, check_bcsr);
  403. starpu_data_unregister(bcsr_handle[0]);
  404. starpu_data_unregister(bcsr_handle[1]);
  405. }
  406. else if (rank == 1)
  407. {
  408. starpu_data_handle_t bcsr_handle;
  409. starpu_bcsr_data_register(&bcsr_handle, -1, BCSR_NNZ_BLOCKS, BCSR_NROWS, (uintptr_t) NULL, (uint32_t *) NULL, (uint32_t *) NULL, 0, BCSR_R, BCSR_C, sizeof(int));
  410. send_recv_and_check(rank, 0, bcsr_handle, 0x73, NULL, 0x8337, NULL, NULL);
  411. starpu_data_unregister(bcsr_handle);
  412. }
  413. }
  414. /*
  415. * CSR
  416. */
  417. void check_csr(starpu_data_handle_t handle_s, starpu_data_handle_t handle_r, int *error)
  418. {
  419. STARPU_ASSERT(starpu_csr_get_elemsize(handle_s) == starpu_csr_get_elemsize(handle_r));
  420. STARPU_ASSERT(starpu_csr_get_nnz(handle_s) == starpu_csr_get_nnz(handle_r));
  421. STARPU_ASSERT(starpu_csr_get_nrow(handle_s) == starpu_csr_get_nrow(handle_r));
  422. STARPU_ASSERT(starpu_csr_get_firstentry(handle_s) == starpu_csr_get_firstentry(handle_r));
  423. starpu_data_acquire(handle_s, STARPU_R);
  424. starpu_data_acquire(handle_r, STARPU_R);
  425. uint32_t *colind_s = starpu_csr_get_local_colind(handle_s);
  426. uint32_t *colind_r = starpu_csr_get_local_colind(handle_r);
  427. uint32_t *rowptr_s = starpu_csr_get_local_rowptr(handle_s);
  428. uint32_t *rowptr_r = starpu_csr_get_local_rowptr(handle_r);
  429. int *csr_s = (int *)starpu_csr_get_local_nzval(handle_s);
  430. int *csr_r = (int *)starpu_csr_get_local_nzval(handle_r);
  431. int nnz = starpu_csr_get_nnz(handle_s);
  432. int nrows = starpu_csr_get_nrow(handle_s);
  433. int x;
  434. for(x=0 ; x<nnz ; x++)
  435. {
  436. if (colind_s[x] == colind_r[x])
  437. {
  438. FPRINTF_MPI(stderr, "Success with colind[%d] value: %u == %u\n", x, colind_s[x], colind_r[x]);
  439. }
  440. else
  441. {
  442. *error = 1;
  443. FPRINTF_MPI(stderr, "Error with colind[%d] value: %u != %u\n", x, colind_s[x], colind_r[x]);
  444. }
  445. }
  446. for(x=0 ; x<nrows+1 ; x++)
  447. {
  448. if (rowptr_s[x] == rowptr_r[x])
  449. {
  450. FPRINTF_MPI(stderr, "Success with rowptr[%d] value: %u == %u\n", x, rowptr_s[x], rowptr_r[x]);
  451. }
  452. else
  453. {
  454. *error = 1;
  455. FPRINTF_MPI(stderr, "Error with rowptr[%d] value: %u != %u\n", x, rowptr_s[x], rowptr_r[x]);
  456. }
  457. }
  458. for(x=0 ; x<nnz ; x++)
  459. {
  460. if (csr_s[x] == csr_r[x])
  461. {
  462. FPRINTF_MPI(stderr, "Success with csr[%d] value: %d == %d\n", x, csr_s[x], csr_r[x]);
  463. }
  464. else
  465. {
  466. *error = 1;
  467. FPRINTF_MPI(stderr, "Error with csr[%d] value: %d != %d\n", x, csr_s[x], csr_r[x]);
  468. }
  469. }
  470. starpu_data_release(handle_s);
  471. starpu_data_release(handle_r);
  472. }
  473. void exchange_csr(int rank, int *error)
  474. {
  475. // the values are completely wrong, we just want to test that the communication is done correctly
  476. #define CSR_NROWS 2
  477. #define CSR_NNZ 5
  478. if (rank == 0)
  479. {
  480. starpu_data_handle_t csr_handle[2];
  481. uint32_t colind[CSR_NNZ] = {0, 1, 2, 3, 4};
  482. uint32_t rowptr[CSR_NROWS+1] = {0, 1, CSR_NNZ};
  483. int nzval[CSR_NNZ] = { 11, 22, 33, 44, 55 };
  484. starpu_csr_data_register(&csr_handle[0], STARPU_MAIN_RAM, CSR_NNZ, CSR_NROWS, (uintptr_t) nzval, colind, rowptr, 0, sizeof(nzval[0]));
  485. starpu_csr_data_register(&csr_handle[1], -1, CSR_NNZ, CSR_NROWS, (uintptr_t) NULL, (uint32_t *) NULL, (uint32_t *) NULL, 0, sizeof(nzval[0]));
  486. send_recv_and_check(rank, 1, csr_handle[0], 0x84, csr_handle[1], 0x8765, error, check_csr);
  487. starpu_data_unregister(csr_handle[0]);
  488. starpu_data_unregister(csr_handle[1]);
  489. }
  490. else if (rank == 1)
  491. {
  492. starpu_data_handle_t csr_handle;
  493. starpu_csr_data_register(&csr_handle, -1, CSR_NNZ, CSR_NROWS, (uintptr_t) NULL, (uint32_t *) NULL, (uint32_t *) NULL, 0, sizeof(int));
  494. send_recv_and_check(rank, 0, csr_handle, 0x84, NULL, 0x8765, NULL, NULL);
  495. starpu_data_unregister(csr_handle);
  496. }
  497. }
  498. int main(int argc, char **argv)
  499. {
  500. int ret, rank, size;
  501. int error=0;
  502. int mpi_init;
  503. MPI_INIT_THREAD(&argc, &argv, MPI_THREAD_SERIALIZED, &mpi_init);
  504. ret = starpu_mpi_init_conf(&argc, &argv, mpi_init, MPI_COMM_WORLD, NULL);
  505. STARPU_CHECK_RETURN_VALUE(ret, "starpu_mpi_init_conf");
  506. starpu_mpi_comm_rank(MPI_COMM_WORLD, &rank);
  507. starpu_mpi_comm_size(MPI_COMM_WORLD, &size);
  508. if (size < 2)
  509. {
  510. if (rank == 0)
  511. FPRINTF(stderr, "We need at least 2 processes.\n");
  512. starpu_mpi_shutdown();
  513. if (!mpi_init)
  514. MPI_Finalize();
  515. return STARPU_TEST_SKIPPED;
  516. }
  517. exchange_void(rank, &error);
  518. exchange_variable(rank, &error);
  519. exchange_vector(rank, &error);
  520. exchange_matrix(rank, &error);
  521. exchange_block(rank, &error);
  522. exchange_bcsr(rank, &error);
  523. exchange_csr(rank, &error);
  524. starpu_mpi_shutdown();
  525. if (!mpi_init)
  526. MPI_Finalize();
  527. return rank == 0 ? error : 0;
  528. }