bcsr_interface.c 16 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471
  1. /* StarPU --- Runtime system for heterogeneous multicore architectures.
  2. *
  3. * Copyright (C) 2011-2012,2017 Inria
  4. * Copyright (C) 2008-2018 Université de Bordeaux
  5. * Copyright (C) 2010-2015,2017 CNRS
  6. *
  7. * StarPU is free software; you can redistribute it and/or modify
  8. * it under the terms of the GNU Lesser General Public License as published by
  9. * the Free Software Foundation; either version 2.1 of the License, or (at
  10. * your option) any later version.
  11. *
  12. * StarPU is distributed in the hope that it will be useful, but
  13. * WITHOUT ANY WARRANTY; without even the implied warranty of
  14. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  15. *
  16. * See the GNU Lesser General Public License in COPYING.LGPL for more details.
  17. */
  18. #include <starpu.h>
  19. #include <common/config.h>
  20. #include <datawizard/coherency.h>
  21. #include <datawizard/copy_driver.h>
  22. #include <datawizard/filters.h>
  23. #include <datawizard/memory_nodes.h>
  24. #include <datawizard/malloc.h>
  25. #include <starpu_hash.h>
  26. #include <starpu_cuda.h>
  27. #include <starpu_opencl.h>
  28. #include <drivers/opencl/driver_opencl.h>
  29. /*
  30. * BCSR : blocked CSR, we use blocks of size (r x c)
  31. */
  32. static int copy_any_to_any(void *src_interface, unsigned src_node, void *dst_interface, unsigned dst_node, void *async_data);
  33. static const struct starpu_data_copy_methods bcsr_copy_data_methods_s =
  34. {
  35. .any_to_any = copy_any_to_any,
  36. };
  37. static void register_bcsr_handle(starpu_data_handle_t handle, unsigned home_node, void *data_interface);
  38. static void *bcsr_to_pointer(void *data_interface, unsigned node);
  39. static int bcsr_pointer_is_inside(void *data_interface, unsigned node, void *ptr);
  40. static starpu_ssize_t allocate_bcsr_buffer_on_node(void *data_interface, unsigned dst_node);
  41. static void free_bcsr_buffer_on_node(void *data_interface, unsigned node);
  42. static size_t bcsr_interface_get_size(starpu_data_handle_t handle);
  43. static int bcsr_compare(void *data_interface_a, void *data_interface_b);
  44. static uint32_t footprint_bcsr_interface_crc32(starpu_data_handle_t handle);
  45. static starpu_ssize_t describe(void *data_interface, char *buf, size_t size);
  46. static int pack_data(starpu_data_handle_t handle, unsigned node, void **ptr, starpu_ssize_t *count);
  47. static int unpack_data(starpu_data_handle_t handle, unsigned node, void *ptr, size_t count);
  48. struct starpu_data_interface_ops starpu_interface_bcsr_ops =
  49. {
  50. .register_data_handle = register_bcsr_handle,
  51. .allocate_data_on_node = allocate_bcsr_buffer_on_node,
  52. .free_data_on_node = free_bcsr_buffer_on_node,
  53. .copy_methods = &bcsr_copy_data_methods_s,
  54. .get_size = bcsr_interface_get_size,
  55. .interfaceid = STARPU_BCSR_INTERFACE_ID,
  56. .interface_size = sizeof(struct starpu_bcsr_interface),
  57. .footprint = footprint_bcsr_interface_crc32,
  58. .compare = bcsr_compare,
  59. .describe = describe,
  60. .to_pointer = bcsr_to_pointer,
  61. .pointer_is_inside = bcsr_pointer_is_inside,
  62. .name = "STARPU_BCSR_INTERFACE",
  63. .pack_data = pack_data,
  64. .unpack_data = unpack_data
  65. };
  66. static void *bcsr_to_pointer(void *data_interface, unsigned node)
  67. {
  68. (void) node;
  69. struct starpu_bcsr_interface *bcsr_interface = data_interface;
  70. return (void*) bcsr_interface->nzval;
  71. }
  72. static int bcsr_pointer_is_inside(void *data_interface, unsigned node, void *ptr)
  73. {
  74. (void) node;
  75. struct starpu_bcsr_interface *bcsr_interface = data_interface;
  76. return ((char*) ptr >= (char*) bcsr_interface->nzval &&
  77. (char*) ptr < (char*) bcsr_interface->nzval + bcsr_interface->nnz*bcsr_interface->r*bcsr_interface->c*bcsr_interface->elemsize)
  78. || ((char*) ptr >= (char*) bcsr_interface->colind &&
  79. (char*) ptr < (char*) bcsr_interface->colind + bcsr_interface->nnz*sizeof(uint32_t))
  80. || ((char*) ptr >= (char*) bcsr_interface->rowptr &&
  81. (char*) ptr < (char*) bcsr_interface->rowptr + (bcsr_interface->nrow+1)*sizeof(uint32_t));
  82. }
  83. static void register_bcsr_handle(starpu_data_handle_t handle, unsigned home_node, void *data_interface)
  84. {
  85. struct starpu_bcsr_interface *bcsr_interface = (struct starpu_bcsr_interface *) data_interface;
  86. unsigned node;
  87. for (node = 0; node < STARPU_MAXNODES; node++)
  88. {
  89. struct starpu_bcsr_interface *local_interface = (struct starpu_bcsr_interface *)
  90. starpu_data_get_interface_on_node(handle, node);
  91. if (node == home_node)
  92. {
  93. local_interface->nzval = bcsr_interface->nzval;
  94. local_interface->colind = bcsr_interface->colind;
  95. local_interface->rowptr = bcsr_interface->rowptr;
  96. }
  97. else
  98. {
  99. local_interface->nzval = 0;
  100. local_interface->colind = NULL;
  101. local_interface->rowptr = NULL;
  102. }
  103. local_interface->id = bcsr_interface->id;
  104. local_interface->nnz = bcsr_interface->nnz;
  105. local_interface->nrow = bcsr_interface->nrow;
  106. local_interface->firstentry = bcsr_interface->firstentry;
  107. local_interface->r = bcsr_interface->r;
  108. local_interface->c = bcsr_interface->c;
  109. local_interface->elemsize = bcsr_interface->elemsize;
  110. }
  111. }
  112. void starpu_bcsr_data_register(starpu_data_handle_t *handleptr, int home_node,
  113. uint32_t nnz, uint32_t nrow, uintptr_t nzval, uint32_t *colind,
  114. uint32_t *rowptr, uint32_t firstentry,
  115. uint32_t r, uint32_t c, size_t elemsize)
  116. {
  117. struct starpu_bcsr_interface bcsr_interface =
  118. {
  119. .id = STARPU_BCSR_INTERFACE_ID,
  120. .nzval = nzval,
  121. .colind = colind,
  122. .rowptr = rowptr,
  123. .nnz = nnz,
  124. .nrow = nrow,
  125. .firstentry = firstentry,
  126. .r = r,
  127. .c = c,
  128. .elemsize = elemsize
  129. };
  130. #ifndef STARPU_SIMGRID
  131. if (home_node >= 0 && starpu_node_get_kind(home_node) == STARPU_CPU_RAM)
  132. {
  133. STARPU_ASSERT_ACCESSIBLE(nzval);
  134. STARPU_ASSERT_ACCESSIBLE(nzval + nnz*elemsize*r*c - 1);
  135. STARPU_ASSERT_ACCESSIBLE(colind);
  136. STARPU_ASSERT_ACCESSIBLE((uintptr_t) colind + nnz*sizeof(uint32_t) - 1);
  137. STARPU_ASSERT_ACCESSIBLE(rowptr);
  138. STARPU_ASSERT_ACCESSIBLE((uintptr_t) rowptr + (nrow+1)*sizeof(uint32_t) - 1);
  139. }
  140. #endif
  141. starpu_data_register(handleptr, home_node, &bcsr_interface, &starpu_interface_bcsr_ops);
  142. }
  143. static uint32_t footprint_bcsr_interface_crc32(starpu_data_handle_t handle)
  144. {
  145. uint32_t hash;
  146. hash = starpu_hash_crc32c_be(starpu_bcsr_get_nnz(handle), 0);
  147. hash = starpu_hash_crc32c_be(starpu_bcsr_get_c(handle), hash);
  148. hash = starpu_hash_crc32c_be(starpu_bcsr_get_r(handle), hash);
  149. return hash;
  150. }
  151. static int bcsr_compare(void *data_interface_a, void *data_interface_b)
  152. {
  153. struct starpu_bcsr_interface *bcsr_a = (struct starpu_bcsr_interface *) data_interface_a;
  154. struct starpu_bcsr_interface *bcsr_b = (struct starpu_bcsr_interface *) data_interface_b;
  155. /* Two matricess are considered compatible if they have the same size */
  156. return (bcsr_a->nnz == bcsr_b->nnz)
  157. && (bcsr_a->nrow == bcsr_b->nrow)
  158. && (bcsr_a->r == bcsr_b->r)
  159. && (bcsr_a->c == bcsr_b->c)
  160. && (bcsr_a->elemsize == bcsr_b->elemsize);
  161. }
  162. /* offer an access to the data parameters */
  163. uint32_t starpu_bcsr_get_nnz(starpu_data_handle_t handle)
  164. {
  165. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  166. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  167. #ifdef STARPU_DEBUG
  168. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  169. #endif
  170. return data_interface->nnz;
  171. }
  172. uint32_t starpu_bcsr_get_nrow(starpu_data_handle_t handle)
  173. {
  174. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  175. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  176. #ifdef STARPU_DEBUG
  177. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  178. #endif
  179. return data_interface->nrow;
  180. }
  181. uint32_t starpu_bcsr_get_firstentry(starpu_data_handle_t handle)
  182. {
  183. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  184. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  185. #ifdef STARPU_DEBUG
  186. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  187. #endif
  188. return data_interface->firstentry;
  189. }
  190. uint32_t starpu_bcsr_get_r(starpu_data_handle_t handle)
  191. {
  192. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  193. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  194. #ifdef STARPU_DEBUG
  195. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  196. #endif
  197. return data_interface->r;
  198. }
  199. uint32_t starpu_bcsr_get_c(starpu_data_handle_t handle)
  200. {
  201. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  202. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  203. #ifdef STARPU_DEBUG
  204. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  205. #endif
  206. return data_interface->c;
  207. }
  208. size_t starpu_bcsr_get_elemsize(starpu_data_handle_t handle)
  209. {
  210. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  211. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  212. #ifdef STARPU_DEBUG
  213. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  214. #endif
  215. return data_interface->elemsize;
  216. }
  217. uintptr_t starpu_bcsr_get_local_nzval(starpu_data_handle_t handle)
  218. {
  219. unsigned node;
  220. node = _starpu_memory_node_get_local_key();
  221. STARPU_ASSERT(starpu_data_test_if_allocated_on_node(handle, node));
  222. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  223. starpu_data_get_interface_on_node(handle, node);
  224. #ifdef STARPU_DEBUG
  225. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  226. #endif
  227. return data_interface->nzval;
  228. }
  229. uint32_t *starpu_bcsr_get_local_colind(starpu_data_handle_t handle)
  230. {
  231. int node = handle->home_node;
  232. if (node < 0 || (starpu_node_get_kind(node) != STARPU_CPU_RAM))
  233. node = STARPU_MAIN_RAM;
  234. /* XXX 0 */
  235. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  236. starpu_data_get_interface_on_node(handle, node);
  237. #ifdef STARPU_DEBUG
  238. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  239. #endif
  240. return data_interface->colind;
  241. }
  242. uint32_t *starpu_bcsr_get_local_rowptr(starpu_data_handle_t handle)
  243. {
  244. int node = handle->home_node;
  245. if (node < 0 || (starpu_node_get_kind(node) != STARPU_CPU_RAM))
  246. node = STARPU_MAIN_RAM;
  247. /* XXX 0 */
  248. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  249. starpu_data_get_interface_on_node(handle, node);
  250. #ifdef STARPU_DEBUG
  251. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  252. #endif
  253. return data_interface->rowptr;
  254. }
  255. static size_t bcsr_interface_get_size(starpu_data_handle_t handle)
  256. {
  257. size_t size;
  258. uint32_t nnz = starpu_bcsr_get_nnz(handle);
  259. uint32_t nrow = starpu_bcsr_get_nrow(handle);
  260. uint32_t r = starpu_bcsr_get_r(handle);
  261. uint32_t c = starpu_bcsr_get_c(handle);
  262. size_t elemsize = starpu_bcsr_get_elemsize(handle);
  263. size = nnz*r*c*elemsize + nnz*sizeof(uint32_t) + (nrow+1)*sizeof(uint32_t);
  264. return size;
  265. }
  266. /* memory allocation/deallocation primitives for the BLAS interface */
  267. /* returns the size of the allocated area */
  268. static starpu_ssize_t allocate_bcsr_buffer_on_node(void *data_interface_, unsigned dst_node)
  269. {
  270. uintptr_t addr_nzval, addr_colind, addr_rowptr;
  271. starpu_ssize_t allocated_memory;
  272. /* we need the 3 arrays to be allocated */
  273. struct starpu_bcsr_interface *bcsr_interface = (struct starpu_bcsr_interface *) data_interface_;
  274. uint32_t nnz = bcsr_interface->nnz;
  275. uint32_t nrow = bcsr_interface->nrow;
  276. size_t elemsize = bcsr_interface->elemsize;
  277. uint32_t r = bcsr_interface->r;
  278. uint32_t c = bcsr_interface->c;
  279. addr_nzval = starpu_malloc_on_node(dst_node, nnz*r*c*elemsize);
  280. if (!addr_nzval)
  281. goto fail_nzval;
  282. addr_colind = starpu_malloc_on_node(dst_node, nnz*sizeof(uint32_t));
  283. if (!addr_colind)
  284. goto fail_colind;
  285. addr_rowptr = starpu_malloc_on_node(dst_node, (nrow+1)*sizeof(uint32_t));
  286. if (!addr_rowptr)
  287. goto fail_rowptr;
  288. /* allocation succeeded */
  289. allocated_memory =
  290. nnz*r*c*elemsize + nnz*sizeof(uint32_t) + (nrow+1)*sizeof(uint32_t);
  291. /* update the data properly in consequence */
  292. bcsr_interface->nzval = addr_nzval;
  293. bcsr_interface->colind = (uint32_t*) addr_colind;
  294. bcsr_interface->rowptr = (uint32_t*) addr_rowptr;
  295. return allocated_memory;
  296. fail_rowptr:
  297. starpu_free_on_node(dst_node, addr_colind, nnz*sizeof(uint32_t));
  298. fail_colind:
  299. starpu_free_on_node(dst_node, addr_nzval, nnz*r*c*elemsize);
  300. fail_nzval:
  301. /* allocation failed */
  302. return -ENOMEM;
  303. }
  304. static void free_bcsr_buffer_on_node(void *data_interface, unsigned node)
  305. {
  306. struct starpu_bcsr_interface *bcsr_interface = (struct starpu_bcsr_interface *) data_interface;
  307. uint32_t nnz = bcsr_interface->nnz;
  308. uint32_t nrow = bcsr_interface->nrow;
  309. size_t elemsize = bcsr_interface->elemsize;
  310. uint32_t r = bcsr_interface->r;
  311. uint32_t c = bcsr_interface->c;
  312. starpu_free_on_node(node, bcsr_interface->nzval, nnz*r*c*elemsize);
  313. starpu_free_on_node(node, (uintptr_t) bcsr_interface->colind, nnz*sizeof(uint32_t));
  314. starpu_free_on_node(node, (uintptr_t) bcsr_interface->rowptr, (nrow+1)*sizeof(uint32_t));
  315. }
  316. static int copy_any_to_any(void *src_interface, unsigned src_node, void *dst_interface, unsigned dst_node, void *async_data)
  317. {
  318. struct starpu_bcsr_interface *src_bcsr = (struct starpu_bcsr_interface *) src_interface;
  319. struct starpu_bcsr_interface *dst_bcsr = (struct starpu_bcsr_interface *) dst_interface;
  320. uint32_t nnz = src_bcsr->nnz;
  321. uint32_t nrow = src_bcsr->nrow;
  322. size_t elemsize = src_bcsr->elemsize;
  323. uint32_t r = src_bcsr->r;
  324. uint32_t c = src_bcsr->c;
  325. int ret = 0;
  326. if (starpu_interface_copy(src_bcsr->nzval, 0, src_node, dst_bcsr->nzval, 0, dst_node, nnz*elemsize*r*c, async_data))
  327. ret = -EAGAIN;
  328. 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))
  329. ret = -EAGAIN;
  330. 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))
  331. ret = -EAGAIN;
  332. _STARPU_TRACE_DATA_COPY(src_node, dst_node, nnz*elemsize*r*c + (nnz+nrow+1)*sizeof(uint32_t));
  333. return ret;
  334. }
  335. static starpu_ssize_t describe(void *data_interface, char *buf, size_t size)
  336. {
  337. struct starpu_bcsr_interface *bcsr = (struct starpu_bcsr_interface *) data_interface;
  338. return snprintf(buf, size, "b%ux%ux%ux%ux%u",
  339. (unsigned) bcsr->nnz,
  340. (unsigned) bcsr->nrow,
  341. (unsigned) bcsr->r,
  342. (unsigned) bcsr->c,
  343. (unsigned) bcsr->elemsize);
  344. }
  345. static int pack_data(starpu_data_handle_t handle, unsigned node, void **ptr, starpu_ssize_t *count)
  346. {
  347. STARPU_ASSERT(starpu_data_test_if_allocated_on_node(handle, node));
  348. struct starpu_bcsr_interface *bcsr = (struct starpu_bcsr_interface *) starpu_data_get_interface_on_node(handle, node);
  349. // We first pack colind
  350. *count = bcsr->nnz * sizeof(bcsr->colind[0]);
  351. // Then rowptr
  352. *count += (bcsr->nrow + 1) * sizeof(bcsr->rowptr[0]);
  353. // Then nnzval
  354. *count += bcsr->r * bcsr->c * bcsr->nnz * bcsr->elemsize;
  355. if (ptr != NULL)
  356. {
  357. _starpu_malloc_flags_on_node(node, ptr, *count, 0);
  358. char *tmp = *ptr;
  359. memcpy(tmp, (void*)bcsr->colind, bcsr->nnz * sizeof(bcsr->colind[0]));
  360. tmp += bcsr->nnz * sizeof(bcsr->colind[0]);
  361. memcpy(tmp, (void*)bcsr->rowptr, (bcsr->nrow + 1) * sizeof(bcsr->rowptr[0]));
  362. tmp += (bcsr->nrow + 1) * sizeof(bcsr->rowptr[0]);
  363. memcpy(tmp, (void*)bcsr->nzval, bcsr->r * bcsr->c * bcsr->nnz * bcsr->elemsize);
  364. }
  365. return 0;
  366. }
  367. static int unpack_data(starpu_data_handle_t handle, unsigned node, void *ptr, size_t count)
  368. {
  369. STARPU_ASSERT(starpu_data_test_if_allocated_on_node(handle, node));
  370. struct starpu_bcsr_interface *bcsr = (struct starpu_bcsr_interface *) starpu_data_get_interface_on_node(handle, node);
  371. STARPU_ASSERT(count == (bcsr->nnz * sizeof(bcsr->colind[0]))+((bcsr->nrow + 1) * sizeof(bcsr->rowptr[0]))+(bcsr->r * bcsr->c * bcsr->nnz * bcsr->elemsize));
  372. char *tmp = ptr;
  373. memcpy((void*)bcsr->colind, tmp, bcsr->nnz * sizeof(bcsr->colind[0]));
  374. tmp += bcsr->nnz * sizeof(bcsr->colind[0]);
  375. memcpy((void*)bcsr->rowptr, tmp, (bcsr->nrow + 1) * sizeof(bcsr->rowptr[0]));
  376. tmp += (bcsr->nrow + 1) * sizeof(bcsr->rowptr[0]);
  377. memcpy((void*)bcsr->nzval, tmp, bcsr->r * bcsr->c * bcsr->nnz * bcsr->elemsize);
  378. return 0;
  379. }