bcsr_interface.c 15 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448
  1. /* StarPU --- Runtime system for heterogeneous multicore architectures.
  2. *
  3. * Copyright (C) 2009-2016 Université de Bordeaux
  4. * Copyright (C) 2010, 2011, 2012, 2013, 2014, 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 <starpu.h>
  18. #include <common/config.h>
  19. #include <datawizard/coherency.h>
  20. #include <datawizard/copy_driver.h>
  21. #include <datawizard/filters.h>
  22. #include <datawizard/memory_nodes.h>
  23. #include <starpu_hash.h>
  24. #include <starpu_cuda.h>
  25. #include <starpu_opencl.h>
  26. #include <drivers/opencl/driver_opencl.h>
  27. /*
  28. * BCSR : blocked CSR, we use blocks of size (r x c)
  29. */
  30. static int copy_any_to_any(void *src_interface, unsigned src_node, void *dst_interface, unsigned dst_node, void *async_data);
  31. static const struct starpu_data_copy_methods bcsr_copy_data_methods_s =
  32. {
  33. .any_to_any = copy_any_to_any,
  34. };
  35. static void register_bcsr_handle(starpu_data_handle_t handle, unsigned home_node, void *data_interface);
  36. static void *bcsr_handle_to_pointer(starpu_data_handle_t data_handle, unsigned node);
  37. static starpu_ssize_t allocate_bcsr_buffer_on_node(void *data_interface, unsigned dst_node);
  38. static void free_bcsr_buffer_on_node(void *data_interface, unsigned node);
  39. static size_t bcsr_interface_get_size(starpu_data_handle_t handle);
  40. static int bcsr_compare(void *data_interface_a, void *data_interface_b);
  41. static uint32_t footprint_bcsr_interface_crc32(starpu_data_handle_t handle);
  42. static starpu_ssize_t describe(void *data_interface, char *buf, size_t size);
  43. static int pack_data(starpu_data_handle_t handle, unsigned node, void **ptr, starpu_ssize_t *count);
  44. static int unpack_data(starpu_data_handle_t handle, unsigned node, void *ptr, size_t count);
  45. struct starpu_data_interface_ops starpu_interface_bcsr_ops =
  46. {
  47. .register_data_handle = register_bcsr_handle,
  48. .allocate_data_on_node = allocate_bcsr_buffer_on_node,
  49. .free_data_on_node = free_bcsr_buffer_on_node,
  50. .copy_methods = &bcsr_copy_data_methods_s,
  51. .get_size = bcsr_interface_get_size,
  52. .interfaceid = STARPU_BCSR_INTERFACE_ID,
  53. .interface_size = sizeof(struct starpu_bcsr_interface),
  54. .footprint = footprint_bcsr_interface_crc32,
  55. .compare = bcsr_compare,
  56. .describe = describe,
  57. .handle_to_pointer = bcsr_handle_to_pointer,
  58. .name = "STARPU_BCSR_INTERFACE",
  59. .pack_data = pack_data,
  60. .unpack_data = unpack_data
  61. };
  62. static void *bcsr_handle_to_pointer(starpu_data_handle_t data_handle, unsigned node)
  63. {
  64. STARPU_ASSERT(starpu_data_test_if_allocated_on_node(data_handle, node));
  65. struct starpu_bcsr_interface *bcsr_interface = (struct starpu_bcsr_interface *)
  66. starpu_data_get_interface_on_node(data_handle, node);
  67. return (void*) bcsr_interface->nzval;
  68. }
  69. static void register_bcsr_handle(starpu_data_handle_t handle, unsigned home_node, void *data_interface)
  70. {
  71. struct starpu_bcsr_interface *bcsr_interface = (struct starpu_bcsr_interface *) data_interface;
  72. unsigned node;
  73. for (node = 0; node < STARPU_MAXNODES; node++)
  74. {
  75. struct starpu_bcsr_interface *local_interface = (struct starpu_bcsr_interface *)
  76. starpu_data_get_interface_on_node(handle, node);
  77. if (node == home_node)
  78. {
  79. local_interface->nzval = bcsr_interface->nzval;
  80. local_interface->colind = bcsr_interface->colind;
  81. local_interface->rowptr = bcsr_interface->rowptr;
  82. }
  83. else
  84. {
  85. local_interface->nzval = 0;
  86. local_interface->colind = NULL;
  87. local_interface->rowptr = NULL;
  88. }
  89. local_interface->id = bcsr_interface->id;
  90. local_interface->nnz = bcsr_interface->nnz;
  91. local_interface->nrow = bcsr_interface->nrow;
  92. local_interface->firstentry = bcsr_interface->firstentry;
  93. local_interface->r = bcsr_interface->r;
  94. local_interface->c = bcsr_interface->c;
  95. local_interface->elemsize = bcsr_interface->elemsize;
  96. }
  97. }
  98. void starpu_bcsr_data_register(starpu_data_handle_t *handleptr, int home_node,
  99. uint32_t nnz, uint32_t nrow, uintptr_t nzval, uint32_t *colind,
  100. uint32_t *rowptr, uint32_t firstentry,
  101. uint32_t r, uint32_t c, size_t elemsize)
  102. {
  103. struct starpu_bcsr_interface bcsr_interface =
  104. {
  105. .id = STARPU_BCSR_INTERFACE_ID,
  106. .nzval = nzval,
  107. .colind = colind,
  108. .rowptr = rowptr,
  109. .nnz = nnz,
  110. .nrow = nrow,
  111. .firstentry = firstentry,
  112. .r = r,
  113. .c = c,
  114. .elemsize = elemsize
  115. };
  116. #ifndef STARPU_SIMGRID
  117. if (home_node == STARPU_MAIN_RAM)
  118. {
  119. STARPU_ASSERT_ACCESSIBLE(nzval);
  120. STARPU_ASSERT_ACCESSIBLE(nzval + nnz*elemsize*r*c - 1);
  121. STARPU_ASSERT_ACCESSIBLE(colind);
  122. STARPU_ASSERT_ACCESSIBLE((uintptr_t) colind + nnz*sizeof(uint32_t) - 1);
  123. STARPU_ASSERT_ACCESSIBLE(rowptr);
  124. STARPU_ASSERT_ACCESSIBLE((uintptr_t) rowptr + (nrow+1)*sizeof(uint32_t) - 1);
  125. }
  126. #endif
  127. starpu_data_register(handleptr, home_node, &bcsr_interface, &starpu_interface_bcsr_ops);
  128. }
  129. static uint32_t footprint_bcsr_interface_crc32(starpu_data_handle_t handle)
  130. {
  131. uint32_t hash;
  132. hash = starpu_hash_crc32c_be(starpu_bcsr_get_nnz(handle), 0);
  133. hash = starpu_hash_crc32c_be(starpu_bcsr_get_c(handle), hash);
  134. hash = starpu_hash_crc32c_be(starpu_bcsr_get_r(handle), hash);
  135. return hash;
  136. }
  137. static int bcsr_compare(void *data_interface_a, void *data_interface_b)
  138. {
  139. struct starpu_bcsr_interface *bcsr_a = (struct starpu_bcsr_interface *) data_interface_a;
  140. struct starpu_bcsr_interface *bcsr_b = (struct starpu_bcsr_interface *) data_interface_b;
  141. /* Two matricess are considered compatible if they have the same size */
  142. return ((bcsr_a->nnz == bcsr_b->nnz)
  143. && (bcsr_a->nrow == bcsr_b->nrow)
  144. && (bcsr_a->r == bcsr_b->r)
  145. && (bcsr_a->c == bcsr_b->c)
  146. && (bcsr_a->elemsize == bcsr_b->elemsize));
  147. }
  148. /* offer an access to the data parameters */
  149. uint32_t starpu_bcsr_get_nnz(starpu_data_handle_t handle)
  150. {
  151. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  152. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  153. #ifdef STARPU_DEBUG
  154. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  155. #endif
  156. return data_interface->nnz;
  157. }
  158. uint32_t starpu_bcsr_get_nrow(starpu_data_handle_t handle)
  159. {
  160. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  161. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  162. #ifdef STARPU_DEBUG
  163. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  164. #endif
  165. return data_interface->nrow;
  166. }
  167. uint32_t starpu_bcsr_get_firstentry(starpu_data_handle_t handle)
  168. {
  169. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  170. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  171. #ifdef STARPU_DEBUG
  172. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  173. #endif
  174. return data_interface->firstentry;
  175. }
  176. uint32_t starpu_bcsr_get_r(starpu_data_handle_t handle)
  177. {
  178. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  179. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  180. #ifdef STARPU_DEBUG
  181. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  182. #endif
  183. return data_interface->r;
  184. }
  185. uint32_t starpu_bcsr_get_c(starpu_data_handle_t handle)
  186. {
  187. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  188. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  189. #ifdef STARPU_DEBUG
  190. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  191. #endif
  192. return data_interface->c;
  193. }
  194. size_t starpu_bcsr_get_elemsize(starpu_data_handle_t handle)
  195. {
  196. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  197. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  198. #ifdef STARPU_DEBUG
  199. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  200. #endif
  201. return data_interface->elemsize;
  202. }
  203. uintptr_t starpu_bcsr_get_local_nzval(starpu_data_handle_t handle)
  204. {
  205. unsigned node;
  206. node = _starpu_memory_node_get_local_key();
  207. STARPU_ASSERT(starpu_data_test_if_allocated_on_node(handle, node));
  208. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  209. starpu_data_get_interface_on_node(handle, node);
  210. #ifdef STARPU_DEBUG
  211. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  212. #endif
  213. return data_interface->nzval;
  214. }
  215. uint32_t *starpu_bcsr_get_local_colind(starpu_data_handle_t handle)
  216. {
  217. /* XXX 0 */
  218. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  219. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  220. #ifdef STARPU_DEBUG
  221. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  222. #endif
  223. return data_interface->colind;
  224. }
  225. uint32_t *starpu_bcsr_get_local_rowptr(starpu_data_handle_t handle)
  226. {
  227. /* XXX 0 */
  228. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  229. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  230. #ifdef STARPU_DEBUG
  231. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  232. #endif
  233. return data_interface->rowptr;
  234. }
  235. static size_t bcsr_interface_get_size(starpu_data_handle_t handle)
  236. {
  237. size_t size;
  238. uint32_t nnz = starpu_bcsr_get_nnz(handle);
  239. uint32_t nrow = starpu_bcsr_get_nrow(handle);
  240. uint32_t r = starpu_bcsr_get_r(handle);
  241. uint32_t c = starpu_bcsr_get_c(handle);
  242. size_t elemsize = starpu_bcsr_get_elemsize(handle);
  243. size = nnz*r*c*elemsize + nnz*sizeof(uint32_t) + (nrow+1)*sizeof(uint32_t);
  244. return size;
  245. }
  246. /* memory allocation/deallocation primitives for the BLAS interface */
  247. /* returns the size of the allocated area */
  248. static starpu_ssize_t allocate_bcsr_buffer_on_node(void *data_interface_, unsigned dst_node)
  249. {
  250. uintptr_t addr_nzval, addr_colind, addr_rowptr;
  251. starpu_ssize_t allocated_memory;
  252. /* we need the 3 arrays to be allocated */
  253. struct starpu_bcsr_interface *bcsr_interface = (struct starpu_bcsr_interface *) data_interface_;
  254. uint32_t nnz = bcsr_interface->nnz;
  255. uint32_t nrow = bcsr_interface->nrow;
  256. size_t elemsize = bcsr_interface->elemsize;
  257. uint32_t r = bcsr_interface->r;
  258. uint32_t c = bcsr_interface->c;
  259. addr_nzval = starpu_malloc_on_node(dst_node, nnz*r*c*elemsize);
  260. if (!addr_nzval)
  261. goto fail_nzval;
  262. addr_colind = starpu_malloc_on_node(dst_node, nnz*sizeof(uint32_t));
  263. if (!addr_colind)
  264. goto fail_colind;
  265. addr_rowptr = starpu_malloc_on_node(dst_node, (nrow+1)*sizeof(uint32_t));
  266. if (!addr_rowptr)
  267. goto fail_rowptr;
  268. /* allocation succeeded */
  269. allocated_memory =
  270. nnz*r*c*elemsize + nnz*sizeof(uint32_t) + (nrow+1)*sizeof(uint32_t);
  271. /* update the data properly in consequence */
  272. bcsr_interface->nzval = addr_nzval;
  273. bcsr_interface->colind = (uint32_t*) addr_colind;
  274. bcsr_interface->rowptr = (uint32_t*) addr_rowptr;
  275. return allocated_memory;
  276. fail_rowptr:
  277. starpu_free_on_node(dst_node, addr_colind, nnz*sizeof(uint32_t));
  278. fail_colind:
  279. starpu_free_on_node(dst_node, addr_nzval, nnz*r*c*elemsize);
  280. fail_nzval:
  281. /* allocation failed */
  282. return -ENOMEM;
  283. }
  284. static void free_bcsr_buffer_on_node(void *data_interface, unsigned node)
  285. {
  286. struct starpu_bcsr_interface *bcsr_interface = (struct starpu_bcsr_interface *) data_interface;
  287. uint32_t nnz = bcsr_interface->nnz;
  288. uint32_t nrow = bcsr_interface->nrow;
  289. size_t elemsize = bcsr_interface->elemsize;
  290. uint32_t r = bcsr_interface->r;
  291. uint32_t c = bcsr_interface->c;
  292. starpu_free_on_node(node, bcsr_interface->nzval, nnz*r*c*elemsize);
  293. starpu_free_on_node(node, (uintptr_t) bcsr_interface->colind, nnz*sizeof(uint32_t));
  294. starpu_free_on_node(node, (uintptr_t) bcsr_interface->rowptr, (nrow+1)*sizeof(uint32_t));
  295. }
  296. static int copy_any_to_any(void *src_interface, unsigned src_node, void *dst_interface, unsigned dst_node, void *async_data)
  297. {
  298. struct starpu_bcsr_interface *src_bcsr = (struct starpu_bcsr_interface *) src_interface;
  299. struct starpu_bcsr_interface *dst_bcsr = (struct starpu_bcsr_interface *) dst_interface;
  300. uint32_t nnz = src_bcsr->nnz;
  301. uint32_t nrow = src_bcsr->nrow;
  302. size_t elemsize = src_bcsr->elemsize;
  303. uint32_t r = src_bcsr->r;
  304. uint32_t c = src_bcsr->c;
  305. int ret = 0;
  306. if (starpu_interface_copy(src_bcsr->nzval, 0, src_node, dst_bcsr->nzval, 0, dst_node, nnz*elemsize*r*c, async_data))
  307. ret = -EAGAIN;
  308. if (starpu_interface_copy((uintptr_t)src_bcsr->colind, 0, src_node, (uintptr_t)dst_bcsr->colind, 0, dst_node, nnz*sizeof(uint32_t), async_data))
  309. ret = -EAGAIN;
  310. if (starpu_interface_copy((uintptr_t)src_bcsr->rowptr, 0, src_node, (uintptr_t)dst_bcsr->rowptr, 0, dst_node, (nrow+1)*sizeof(uint32_t), async_data))
  311. ret = -EAGAIN;
  312. _STARPU_TRACE_DATA_COPY(src_node, dst_node, nnz*elemsize*r*c + (nnz+nrow+1)*sizeof(uint32_t));
  313. return ret;
  314. }
  315. static starpu_ssize_t describe(void *data_interface, char *buf, size_t size)
  316. {
  317. struct starpu_bcsr_interface *bcsr = (struct starpu_bcsr_interface *) data_interface;
  318. return snprintf(buf, size, "b%ux%ux%ux%ux%u",
  319. (unsigned) bcsr->nnz,
  320. (unsigned) bcsr->nrow,
  321. (unsigned) bcsr->r,
  322. (unsigned) bcsr->c,
  323. (unsigned) bcsr->elemsize);
  324. }
  325. static int pack_data(starpu_data_handle_t handle, unsigned node, void **ptr, starpu_ssize_t *count)
  326. {
  327. STARPU_ASSERT(starpu_data_test_if_allocated_on_node(handle, node));
  328. struct starpu_bcsr_interface *bcsr = (struct starpu_bcsr_interface *) starpu_data_get_interface_on_node(handle, node);
  329. // We first pack colind
  330. *count = bcsr->nnz * sizeof(bcsr->colind[0]);
  331. // Then rowptr
  332. *count += (bcsr->nrow + 1) * sizeof(bcsr->rowptr[0]);
  333. // Then nnzval
  334. *count += bcsr->r * bcsr->c * bcsr->nnz * bcsr->elemsize;
  335. if (ptr != NULL)
  336. {
  337. starpu_malloc_flags(ptr, *count, 0);
  338. char *tmp = *ptr;
  339. memcpy(tmp, (void*)bcsr->colind, bcsr->nnz * sizeof(bcsr->colind[0]));
  340. tmp += bcsr->nnz * sizeof(bcsr->colind[0]);
  341. memcpy(tmp, (void*)bcsr->rowptr, (bcsr->nrow + 1) * sizeof(bcsr->rowptr[0]));
  342. tmp += (bcsr->nrow + 1) * sizeof(bcsr->rowptr[0]);
  343. memcpy(tmp, (void*)bcsr->nzval, bcsr->r * bcsr->c * bcsr->nnz * bcsr->elemsize);
  344. }
  345. return 0;
  346. }
  347. static int unpack_data(starpu_data_handle_t handle, unsigned node, void *ptr, size_t count)
  348. {
  349. STARPU_ASSERT(starpu_data_test_if_allocated_on_node(handle, node));
  350. struct starpu_bcsr_interface *bcsr = (struct starpu_bcsr_interface *) starpu_data_get_interface_on_node(handle, node);
  351. STARPU_ASSERT(count == (bcsr->nnz * sizeof(bcsr->colind[0]))+((bcsr->nrow + 1) * sizeof(bcsr->rowptr[0]))+(bcsr->r * bcsr->c * bcsr->nnz * bcsr->elemsize));
  352. char *tmp = ptr;
  353. memcpy((void*)bcsr->colind, tmp, bcsr->nnz * sizeof(bcsr->colind[0]));
  354. tmp += bcsr->nnz * sizeof(bcsr->colind[0]);
  355. memcpy((void*)bcsr->rowptr, tmp, (bcsr->nrow + 1) * sizeof(bcsr->rowptr[0]));
  356. tmp += (bcsr->nrow + 1) * sizeof(bcsr->rowptr[0]);
  357. memcpy((void*)bcsr->nzval, tmp, bcsr->r * bcsr->c * bcsr->nnz * bcsr->elemsize);
  358. return 0;
  359. }