bcsr_interface.c 15 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457
  1. /* StarPU --- Runtime system for heterogeneous multicore architectures.
  2. *
  3. * Copyright (C) 2009-2017 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 <datawizard/malloc.h>
  24. #include <starpu_hash.h>
  25. #include <starpu_cuda.h>
  26. #include <starpu_opencl.h>
  27. #include <drivers/opencl/driver_opencl.h>
  28. /*
  29. * BCSR : blocked CSR, we use blocks of size (r x c)
  30. */
  31. static int copy_any_to_any(void *src_interface, unsigned src_node, void *dst_interface, unsigned dst_node, void *async_data);
  32. static const struct starpu_data_copy_methods bcsr_copy_data_methods_s =
  33. {
  34. .any_to_any = copy_any_to_any,
  35. };
  36. static void register_bcsr_handle(starpu_data_handle_t handle, unsigned home_node, void *data_interface);
  37. static void *bcsr_handle_to_pointer(starpu_data_handle_t data_handle, unsigned node);
  38. static starpu_ssize_t allocate_bcsr_buffer_on_node(void *data_interface, unsigned dst_node);
  39. static void free_bcsr_buffer_on_node(void *data_interface, unsigned node);
  40. static size_t bcsr_interface_get_size(starpu_data_handle_t handle);
  41. static int bcsr_compare(void *data_interface_a, void *data_interface_b);
  42. static uint32_t footprint_bcsr_interface_crc32(starpu_data_handle_t handle);
  43. static starpu_ssize_t describe(void *data_interface, char *buf, size_t size);
  44. static int pack_data(starpu_data_handle_t handle, unsigned node, void **ptr, starpu_ssize_t *count);
  45. static int unpack_data(starpu_data_handle_t handle, unsigned node, void *ptr, size_t count);
  46. struct starpu_data_interface_ops starpu_interface_bcsr_ops =
  47. {
  48. .register_data_handle = register_bcsr_handle,
  49. .allocate_data_on_node = allocate_bcsr_buffer_on_node,
  50. .free_data_on_node = free_bcsr_buffer_on_node,
  51. .copy_methods = &bcsr_copy_data_methods_s,
  52. .get_size = bcsr_interface_get_size,
  53. .interfaceid = STARPU_BCSR_INTERFACE_ID,
  54. .interface_size = sizeof(struct starpu_bcsr_interface),
  55. .footprint = footprint_bcsr_interface_crc32,
  56. .compare = bcsr_compare,
  57. .describe = describe,
  58. .handle_to_pointer = bcsr_handle_to_pointer,
  59. .name = "STARPU_BCSR_INTERFACE",
  60. .pack_data = pack_data,
  61. .unpack_data = unpack_data
  62. };
  63. static void *bcsr_handle_to_pointer(starpu_data_handle_t data_handle, unsigned node)
  64. {
  65. STARPU_ASSERT(starpu_data_test_if_allocated_on_node(data_handle, node));
  66. struct starpu_bcsr_interface *bcsr_interface = (struct starpu_bcsr_interface *)
  67. starpu_data_get_interface_on_node(data_handle, node);
  68. return (void*) bcsr_interface->nzval;
  69. }
  70. static void register_bcsr_handle(starpu_data_handle_t handle, unsigned home_node, void *data_interface)
  71. {
  72. struct starpu_bcsr_interface *bcsr_interface = (struct starpu_bcsr_interface *) data_interface;
  73. unsigned node;
  74. for (node = 0; node < STARPU_MAXNODES; node++)
  75. {
  76. struct starpu_bcsr_interface *local_interface = (struct starpu_bcsr_interface *)
  77. starpu_data_get_interface_on_node(handle, node);
  78. if (node == home_node)
  79. {
  80. local_interface->nzval = bcsr_interface->nzval;
  81. local_interface->colind = bcsr_interface->colind;
  82. local_interface->rowptr = bcsr_interface->rowptr;
  83. }
  84. else
  85. {
  86. local_interface->nzval = 0;
  87. local_interface->colind = NULL;
  88. local_interface->rowptr = NULL;
  89. }
  90. local_interface->id = bcsr_interface->id;
  91. local_interface->nnz = bcsr_interface->nnz;
  92. local_interface->nrow = bcsr_interface->nrow;
  93. local_interface->firstentry = bcsr_interface->firstentry;
  94. local_interface->r = bcsr_interface->r;
  95. local_interface->c = bcsr_interface->c;
  96. local_interface->elemsize = bcsr_interface->elemsize;
  97. }
  98. }
  99. void starpu_bcsr_data_register(starpu_data_handle_t *handleptr, int home_node,
  100. uint32_t nnz, uint32_t nrow, uintptr_t nzval, uint32_t *colind,
  101. uint32_t *rowptr, uint32_t firstentry,
  102. uint32_t r, uint32_t c, size_t elemsize)
  103. {
  104. struct starpu_bcsr_interface bcsr_interface =
  105. {
  106. .id = STARPU_BCSR_INTERFACE_ID,
  107. .nzval = nzval,
  108. .colind = colind,
  109. .rowptr = rowptr,
  110. .nnz = nnz,
  111. .nrow = nrow,
  112. .firstentry = firstentry,
  113. .r = r,
  114. .c = c,
  115. .elemsize = elemsize
  116. };
  117. #ifndef STARPU_SIMGRID
  118. if (home_node >= 0 && starpu_node_get_kind(home_node) == STARPU_CPU_RAM)
  119. {
  120. STARPU_ASSERT_ACCESSIBLE(nzval);
  121. STARPU_ASSERT_ACCESSIBLE(nzval + nnz*elemsize*r*c - 1);
  122. STARPU_ASSERT_ACCESSIBLE(colind);
  123. STARPU_ASSERT_ACCESSIBLE((uintptr_t) colind + nnz*sizeof(uint32_t) - 1);
  124. STARPU_ASSERT_ACCESSIBLE(rowptr);
  125. STARPU_ASSERT_ACCESSIBLE((uintptr_t) rowptr + (nrow+1)*sizeof(uint32_t) - 1);
  126. }
  127. #endif
  128. starpu_data_register(handleptr, home_node, &bcsr_interface, &starpu_interface_bcsr_ops);
  129. }
  130. static uint32_t footprint_bcsr_interface_crc32(starpu_data_handle_t handle)
  131. {
  132. uint32_t hash;
  133. hash = starpu_hash_crc32c_be(starpu_bcsr_get_nnz(handle), 0);
  134. hash = starpu_hash_crc32c_be(starpu_bcsr_get_c(handle), hash);
  135. hash = starpu_hash_crc32c_be(starpu_bcsr_get_r(handle), hash);
  136. return hash;
  137. }
  138. static int bcsr_compare(void *data_interface_a, void *data_interface_b)
  139. {
  140. struct starpu_bcsr_interface *bcsr_a = (struct starpu_bcsr_interface *) data_interface_a;
  141. struct starpu_bcsr_interface *bcsr_b = (struct starpu_bcsr_interface *) data_interface_b;
  142. /* Two matricess are considered compatible if they have the same size */
  143. return (bcsr_a->nnz == bcsr_b->nnz)
  144. && (bcsr_a->nrow == bcsr_b->nrow)
  145. && (bcsr_a->r == bcsr_b->r)
  146. && (bcsr_a->c == bcsr_b->c)
  147. && (bcsr_a->elemsize == bcsr_b->elemsize);
  148. }
  149. /* offer an access to the data parameters */
  150. uint32_t starpu_bcsr_get_nnz(starpu_data_handle_t handle)
  151. {
  152. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  153. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  154. #ifdef STARPU_DEBUG
  155. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  156. #endif
  157. return data_interface->nnz;
  158. }
  159. uint32_t starpu_bcsr_get_nrow(starpu_data_handle_t handle)
  160. {
  161. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  162. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  163. #ifdef STARPU_DEBUG
  164. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  165. #endif
  166. return data_interface->nrow;
  167. }
  168. uint32_t starpu_bcsr_get_firstentry(starpu_data_handle_t handle)
  169. {
  170. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  171. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  172. #ifdef STARPU_DEBUG
  173. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  174. #endif
  175. return data_interface->firstentry;
  176. }
  177. uint32_t starpu_bcsr_get_r(starpu_data_handle_t handle)
  178. {
  179. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  180. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  181. #ifdef STARPU_DEBUG
  182. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  183. #endif
  184. return data_interface->r;
  185. }
  186. uint32_t starpu_bcsr_get_c(starpu_data_handle_t handle)
  187. {
  188. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  189. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  190. #ifdef STARPU_DEBUG
  191. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  192. #endif
  193. return data_interface->c;
  194. }
  195. size_t starpu_bcsr_get_elemsize(starpu_data_handle_t handle)
  196. {
  197. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  198. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  199. #ifdef STARPU_DEBUG
  200. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  201. #endif
  202. return data_interface->elemsize;
  203. }
  204. uintptr_t starpu_bcsr_get_local_nzval(starpu_data_handle_t handle)
  205. {
  206. unsigned node;
  207. node = _starpu_memory_node_get_local_key();
  208. STARPU_ASSERT(starpu_data_test_if_allocated_on_node(handle, node));
  209. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  210. starpu_data_get_interface_on_node(handle, node);
  211. #ifdef STARPU_DEBUG
  212. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  213. #endif
  214. return data_interface->nzval;
  215. }
  216. uint32_t *starpu_bcsr_get_local_colind(starpu_data_handle_t handle)
  217. {
  218. int node = handle->home_node;
  219. if (node < 0 || (starpu_node_get_kind(node) != STARPU_CPU_RAM))
  220. node = STARPU_MAIN_RAM;
  221. /* XXX 0 */
  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->colind;
  228. }
  229. uint32_t *starpu_bcsr_get_local_rowptr(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->rowptr;
  241. }
  242. static size_t bcsr_interface_get_size(starpu_data_handle_t handle)
  243. {
  244. size_t size;
  245. uint32_t nnz = starpu_bcsr_get_nnz(handle);
  246. uint32_t nrow = starpu_bcsr_get_nrow(handle);
  247. uint32_t r = starpu_bcsr_get_r(handle);
  248. uint32_t c = starpu_bcsr_get_c(handle);
  249. size_t elemsize = starpu_bcsr_get_elemsize(handle);
  250. size = nnz*r*c*elemsize + nnz*sizeof(uint32_t) + (nrow+1)*sizeof(uint32_t);
  251. return size;
  252. }
  253. /* memory allocation/deallocation primitives for the BLAS interface */
  254. /* returns the size of the allocated area */
  255. static starpu_ssize_t allocate_bcsr_buffer_on_node(void *data_interface_, unsigned dst_node)
  256. {
  257. uintptr_t addr_nzval, addr_colind, addr_rowptr;
  258. starpu_ssize_t allocated_memory;
  259. /* we need the 3 arrays to be allocated */
  260. struct starpu_bcsr_interface *bcsr_interface = (struct starpu_bcsr_interface *) data_interface_;
  261. uint32_t nnz = bcsr_interface->nnz;
  262. uint32_t nrow = bcsr_interface->nrow;
  263. size_t elemsize = bcsr_interface->elemsize;
  264. uint32_t r = bcsr_interface->r;
  265. uint32_t c = bcsr_interface->c;
  266. addr_nzval = starpu_malloc_on_node(dst_node, nnz*r*c*elemsize);
  267. if (!addr_nzval)
  268. goto fail_nzval;
  269. addr_colind = starpu_malloc_on_node(dst_node, nnz*sizeof(uint32_t));
  270. if (!addr_colind)
  271. goto fail_colind;
  272. addr_rowptr = starpu_malloc_on_node(dst_node, (nrow+1)*sizeof(uint32_t));
  273. if (!addr_rowptr)
  274. goto fail_rowptr;
  275. /* allocation succeeded */
  276. allocated_memory =
  277. nnz*r*c*elemsize + nnz*sizeof(uint32_t) + (nrow+1)*sizeof(uint32_t);
  278. /* update the data properly in consequence */
  279. bcsr_interface->nzval = addr_nzval;
  280. bcsr_interface->colind = (uint32_t*) addr_colind;
  281. bcsr_interface->rowptr = (uint32_t*) addr_rowptr;
  282. return allocated_memory;
  283. fail_rowptr:
  284. starpu_free_on_node(dst_node, addr_colind, nnz*sizeof(uint32_t));
  285. fail_colind:
  286. starpu_free_on_node(dst_node, addr_nzval, nnz*r*c*elemsize);
  287. fail_nzval:
  288. /* allocation failed */
  289. return -ENOMEM;
  290. }
  291. static void free_bcsr_buffer_on_node(void *data_interface, unsigned node)
  292. {
  293. struct starpu_bcsr_interface *bcsr_interface = (struct starpu_bcsr_interface *) data_interface;
  294. uint32_t nnz = bcsr_interface->nnz;
  295. uint32_t nrow = bcsr_interface->nrow;
  296. size_t elemsize = bcsr_interface->elemsize;
  297. uint32_t r = bcsr_interface->r;
  298. uint32_t c = bcsr_interface->c;
  299. starpu_free_on_node(node, bcsr_interface->nzval, nnz*r*c*elemsize);
  300. starpu_free_on_node(node, (uintptr_t) bcsr_interface->colind, nnz*sizeof(uint32_t));
  301. starpu_free_on_node(node, (uintptr_t) bcsr_interface->rowptr, (nrow+1)*sizeof(uint32_t));
  302. }
  303. static int copy_any_to_any(void *src_interface, unsigned src_node, void *dst_interface, unsigned dst_node, void *async_data)
  304. {
  305. struct starpu_bcsr_interface *src_bcsr = (struct starpu_bcsr_interface *) src_interface;
  306. struct starpu_bcsr_interface *dst_bcsr = (struct starpu_bcsr_interface *) dst_interface;
  307. uint32_t nnz = src_bcsr->nnz;
  308. uint32_t nrow = src_bcsr->nrow;
  309. size_t elemsize = src_bcsr->elemsize;
  310. uint32_t r = src_bcsr->r;
  311. uint32_t c = src_bcsr->c;
  312. int ret = 0;
  313. if (starpu_interface_copy(src_bcsr->nzval, 0, src_node, dst_bcsr->nzval, 0, dst_node, nnz*elemsize*r*c, async_data))
  314. ret = -EAGAIN;
  315. 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))
  316. ret = -EAGAIN;
  317. 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))
  318. ret = -EAGAIN;
  319. _STARPU_TRACE_DATA_COPY(src_node, dst_node, nnz*elemsize*r*c + (nnz+nrow+1)*sizeof(uint32_t));
  320. return ret;
  321. }
  322. static starpu_ssize_t describe(void *data_interface, char *buf, size_t size)
  323. {
  324. struct starpu_bcsr_interface *bcsr = (struct starpu_bcsr_interface *) data_interface;
  325. return snprintf(buf, size, "b%ux%ux%ux%ux%u",
  326. (unsigned) bcsr->nnz,
  327. (unsigned) bcsr->nrow,
  328. (unsigned) bcsr->r,
  329. (unsigned) bcsr->c,
  330. (unsigned) bcsr->elemsize);
  331. }
  332. static int pack_data(starpu_data_handle_t handle, unsigned node, void **ptr, starpu_ssize_t *count)
  333. {
  334. STARPU_ASSERT(starpu_data_test_if_allocated_on_node(handle, node));
  335. struct starpu_bcsr_interface *bcsr = (struct starpu_bcsr_interface *) starpu_data_get_interface_on_node(handle, node);
  336. // We first pack colind
  337. *count = bcsr->nnz * sizeof(bcsr->colind[0]);
  338. // Then rowptr
  339. *count += (bcsr->nrow + 1) * sizeof(bcsr->rowptr[0]);
  340. // Then nnzval
  341. *count += bcsr->r * bcsr->c * bcsr->nnz * bcsr->elemsize;
  342. if (ptr != NULL)
  343. {
  344. _starpu_malloc_flags_on_node(node, ptr, *count, 0);
  345. char *tmp = *ptr;
  346. memcpy(tmp, (void*)bcsr->colind, bcsr->nnz * sizeof(bcsr->colind[0]));
  347. tmp += bcsr->nnz * sizeof(bcsr->colind[0]);
  348. memcpy(tmp, (void*)bcsr->rowptr, (bcsr->nrow + 1) * sizeof(bcsr->rowptr[0]));
  349. tmp += (bcsr->nrow + 1) * sizeof(bcsr->rowptr[0]);
  350. memcpy(tmp, (void*)bcsr->nzval, bcsr->r * bcsr->c * bcsr->nnz * bcsr->elemsize);
  351. }
  352. return 0;
  353. }
  354. static int unpack_data(starpu_data_handle_t handle, unsigned node, void *ptr, size_t count)
  355. {
  356. STARPU_ASSERT(starpu_data_test_if_allocated_on_node(handle, node));
  357. struct starpu_bcsr_interface *bcsr = (struct starpu_bcsr_interface *) starpu_data_get_interface_on_node(handle, node);
  358. STARPU_ASSERT(count == (bcsr->nnz * sizeof(bcsr->colind[0]))+((bcsr->nrow + 1) * sizeof(bcsr->rowptr[0]))+(bcsr->r * bcsr->c * bcsr->nnz * bcsr->elemsize));
  359. char *tmp = ptr;
  360. memcpy((void*)bcsr->colind, tmp, bcsr->nnz * sizeof(bcsr->colind[0]));
  361. tmp += bcsr->nnz * sizeof(bcsr->colind[0]);
  362. memcpy((void*)bcsr->rowptr, tmp, (bcsr->nrow + 1) * sizeof(bcsr->rowptr[0]));
  363. tmp += (bcsr->nrow + 1) * sizeof(bcsr->rowptr[0]);
  364. memcpy((void*)bcsr->nzval, tmp, bcsr->r * bcsr->c * bcsr->nnz * bcsr->elemsize);
  365. return 0;
  366. }