csr_interface.c 14 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399
  1. /* StarPU --- Runtime system for heterogeneous multicore architectures.
  2. *
  3. * Copyright (C) 2008-2018 Université de Bordeaux
  4. * Copyright (C) 2011,2012,2017 Inria
  5. * Copyright (C) 2010 Mehdi Juhoor
  6. * Copyright (C) 2010-2015,2017,2019 CNRS
  7. *
  8. * StarPU is free software; you can redistribute it and/or modify
  9. * it under the terms of the GNU Lesser General Public License as published by
  10. * the Free Software Foundation; either version 2.1 of the License, or (at
  11. * your option) any later version.
  12. *
  13. * StarPU is distributed in the hope that it will be useful, but
  14. * WITHOUT ANY WARRANTY; without even the implied warranty of
  15. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  16. *
  17. * See the GNU Lesser General Public License in COPYING.LGPL for more details.
  18. */
  19. #include <starpu.h>
  20. static int copy_any_to_any(void *src_interface, unsigned src_node, void *dst_interface, unsigned dst_node, void *async_data);
  21. static const struct starpu_data_copy_methods csr_copy_data_methods_s =
  22. {
  23. .any_to_any = copy_any_to_any,
  24. };
  25. static void register_csr_handle(starpu_data_handle_t handle, unsigned home_node, void *data_interface);
  26. static int csr_pointer_is_inside(void *data_interface, unsigned node, void *ptr);
  27. static starpu_ssize_t allocate_csr_buffer_on_node(void *data_interface_, unsigned dst_node);
  28. static void free_csr_buffer_on_node(void *data_interface, unsigned node);
  29. static size_t csr_interface_get_size(starpu_data_handle_t handle);
  30. static int csr_compare(void *data_interface_a, void *data_interface_b);
  31. static uint32_t footprint_csr_interface_crc32(starpu_data_handle_t handle);
  32. static starpu_ssize_t describe(void *data_interface, char *buf, size_t size);
  33. static int pack_data(starpu_data_handle_t handle, unsigned node, void **ptr, starpu_ssize_t *count);
  34. static int unpack_data(starpu_data_handle_t handle, unsigned node, void *ptr, size_t count);
  35. struct starpu_data_interface_ops starpu_interface_csr_ops =
  36. {
  37. .register_data_handle = register_csr_handle,
  38. .allocate_data_on_node = allocate_csr_buffer_on_node,
  39. .free_data_on_node = free_csr_buffer_on_node,
  40. .copy_methods = &csr_copy_data_methods_s,
  41. .get_size = csr_interface_get_size,
  42. .interfaceid = STARPU_CSR_INTERFACE_ID,
  43. .interface_size = sizeof(struct starpu_csr_interface),
  44. .footprint = footprint_csr_interface_crc32,
  45. .compare = csr_compare,
  46. .describe = describe,
  47. .pointer_is_inside = csr_pointer_is_inside,
  48. .name = "STARPU_CSR_INTERFACE",
  49. .pack_data = pack_data,
  50. .unpack_data = unpack_data
  51. };
  52. static int csr_pointer_is_inside(void *data_interface, unsigned node, void *ptr)
  53. {
  54. (void) node;
  55. struct starpu_csr_interface *csr_interface = data_interface;
  56. return ((char*) ptr >= (char*) csr_interface->nzval &&
  57. (char*) ptr < (char*) csr_interface->nzval + csr_interface->nnz*csr_interface->elemsize)
  58. || ((char*) ptr >= (char*) csr_interface->colind &&
  59. (char*) ptr < (char*) csr_interface->colind + csr_interface->nnz*sizeof(uint32_t))
  60. || ((char*) ptr >= (char*) csr_interface->rowptr &&
  61. (char*) ptr < (char*) csr_interface->rowptr + (csr_interface->nrow+1)*sizeof(uint32_t));
  62. }
  63. static void register_csr_handle(starpu_data_handle_t handle, unsigned home_node, void *data_interface)
  64. {
  65. struct starpu_csr_interface *csr_interface = (struct starpu_csr_interface *) data_interface;
  66. unsigned node;
  67. for (node = 0; node < STARPU_MAXNODES; node++)
  68. {
  69. struct starpu_csr_interface *local_interface = (struct starpu_csr_interface *)
  70. starpu_data_get_interface_on_node(handle, node);
  71. if (node == home_node)
  72. {
  73. local_interface->nzval = csr_interface->nzval;
  74. local_interface->colind = csr_interface->colind;
  75. }
  76. else
  77. {
  78. local_interface->nzval = 0;
  79. local_interface->colind = NULL;
  80. }
  81. local_interface->id = csr_interface->id;
  82. local_interface->rowptr = csr_interface->rowptr;
  83. local_interface->nnz = csr_interface->nnz;
  84. local_interface->nrow = csr_interface->nrow;
  85. local_interface->firstentry = csr_interface->firstentry;
  86. local_interface->elemsize = csr_interface->elemsize;
  87. }
  88. }
  89. /* declare a new data with the BLAS interface */
  90. void starpu_csr_data_register(starpu_data_handle_t *handleptr, int home_node,
  91. uint32_t nnz, uint32_t nrow, uintptr_t nzval, uint32_t *colind, uint32_t *rowptr, uint32_t firstentry, size_t elemsize)
  92. {
  93. struct starpu_csr_interface csr_interface =
  94. {
  95. .id = STARPU_CSR_INTERFACE_ID,
  96. .nnz = nnz,
  97. .nrow = nrow,
  98. .nzval = nzval,
  99. .colind = colind,
  100. .rowptr = rowptr,
  101. .firstentry = firstentry,
  102. .elemsize = elemsize
  103. };
  104. #ifndef STARPU_SIMGRID
  105. if (home_node >= 0 && starpu_node_get_kind(home_node) == STARPU_CPU_RAM)
  106. {
  107. STARPU_ASSERT_ACCESSIBLE(nzval);
  108. STARPU_ASSERT_ACCESSIBLE(nzval + nnz*elemsize - 1);
  109. STARPU_ASSERT_ACCESSIBLE(colind);
  110. STARPU_ASSERT_ACCESSIBLE((uintptr_t) colind + nnz*sizeof(uint32_t) - 1);
  111. STARPU_ASSERT_ACCESSIBLE(rowptr);
  112. STARPU_ASSERT_ACCESSIBLE((uintptr_t) rowptr + (nrow+1)*sizeof(uint32_t) - 1);
  113. }
  114. #endif
  115. starpu_data_register(handleptr, home_node, &csr_interface, &starpu_interface_csr_ops);
  116. }
  117. static uint32_t footprint_csr_interface_crc32(starpu_data_handle_t handle)
  118. {
  119. return starpu_hash_crc32c_be(starpu_csr_get_nnz(handle), 0);
  120. }
  121. static int csr_compare(void *data_interface_a, void *data_interface_b)
  122. {
  123. struct starpu_csr_interface *csr_a = (struct starpu_csr_interface *) data_interface_a;
  124. struct starpu_csr_interface *csr_b = (struct starpu_csr_interface *) data_interface_b;
  125. /* Two matricess are considered compatible if they have the same size */
  126. return (csr_a->nnz == csr_b->nnz)
  127. && (csr_a->nrow == csr_b->nrow)
  128. && (csr_a->elemsize == csr_b->elemsize);
  129. }
  130. /* offer an access to the data parameters */
  131. uint32_t starpu_csr_get_nnz(starpu_data_handle_t handle)
  132. {
  133. struct starpu_csr_interface *csr_interface = (struct starpu_csr_interface *)
  134. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  135. #ifdef STARPU_DEBUG
  136. STARPU_ASSERT_MSG(csr_interface->id == STARPU_CSR_INTERFACE_ID, "Error. The given data is not a csr.");
  137. #endif
  138. return csr_interface->nnz;
  139. }
  140. uint32_t starpu_csr_get_nrow(starpu_data_handle_t handle)
  141. {
  142. struct starpu_csr_interface *csr_interface = (struct starpu_csr_interface *)
  143. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  144. #ifdef STARPU_DEBUG
  145. STARPU_ASSERT_MSG(csr_interface->id == STARPU_CSR_INTERFACE_ID, "Error. The given data is not a csr.");
  146. #endif
  147. return csr_interface->nrow;
  148. }
  149. uint32_t starpu_csr_get_firstentry(starpu_data_handle_t handle)
  150. {
  151. struct starpu_csr_interface *csr_interface = (struct starpu_csr_interface *)
  152. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  153. #ifdef STARPU_DEBUG
  154. STARPU_ASSERT_MSG(csr_interface->id == STARPU_CSR_INTERFACE_ID, "Error. The given data is not a csr.");
  155. #endif
  156. return csr_interface->firstentry;
  157. }
  158. size_t starpu_csr_get_elemsize(starpu_data_handle_t handle)
  159. {
  160. struct starpu_csr_interface *csr_interface = (struct starpu_csr_interface *)
  161. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  162. #ifdef STARPU_DEBUG
  163. STARPU_ASSERT_MSG(csr_interface->id == STARPU_CSR_INTERFACE_ID, "Error. The given data is not a csr.");
  164. #endif
  165. return csr_interface->elemsize;
  166. }
  167. uintptr_t starpu_csr_get_local_nzval(starpu_data_handle_t handle)
  168. {
  169. unsigned node;
  170. node = starpu_worker_get_local_memory_node();
  171. STARPU_ASSERT(starpu_data_test_if_allocated_on_node(handle, node));
  172. struct starpu_csr_interface *csr_interface = (struct starpu_csr_interface *)
  173. starpu_data_get_interface_on_node(handle, node);
  174. #ifdef STARPU_DEBUG
  175. STARPU_ASSERT_MSG(csr_interface->id == STARPU_CSR_INTERFACE_ID, "Error. The given data is not a csr.");
  176. #endif
  177. return csr_interface->nzval;
  178. }
  179. uint32_t *starpu_csr_get_local_colind(starpu_data_handle_t handle)
  180. {
  181. unsigned node;
  182. node = starpu_worker_get_local_memory_node();
  183. STARPU_ASSERT(starpu_data_test_if_allocated_on_node(handle, node));
  184. struct starpu_csr_interface *csr_interface = (struct starpu_csr_interface *)
  185. starpu_data_get_interface_on_node(handle, node);
  186. #ifdef STARPU_DEBUG
  187. STARPU_ASSERT_MSG(csr_interface->id == STARPU_CSR_INTERFACE_ID, "Error. The given data is not a csr.");
  188. #endif
  189. return csr_interface->colind;
  190. }
  191. uint32_t *starpu_csr_get_local_rowptr(starpu_data_handle_t handle)
  192. {
  193. unsigned node;
  194. node = starpu_worker_get_local_memory_node();
  195. STARPU_ASSERT(starpu_data_test_if_allocated_on_node(handle, node));
  196. struct starpu_csr_interface *csr_interface = (struct starpu_csr_interface *)
  197. starpu_data_get_interface_on_node(handle, node);
  198. #ifdef STARPU_DEBUG
  199. STARPU_ASSERT_MSG(csr_interface->id == STARPU_CSR_INTERFACE_ID, "Error. The given data is not a csr.");
  200. #endif
  201. return csr_interface->rowptr;
  202. }
  203. static size_t csr_interface_get_size(starpu_data_handle_t handle)
  204. {
  205. size_t size;
  206. uint32_t nnz = starpu_csr_get_nnz(handle);
  207. uint32_t nrow = starpu_csr_get_nrow(handle);
  208. size_t elemsize = starpu_csr_get_elemsize(handle);
  209. size = nnz*elemsize + nnz*sizeof(uint32_t) + (nrow+1)*sizeof(uint32_t);
  210. return size;
  211. }
  212. /* memory allocation/deallocation primitives for the BLAS interface */
  213. /* returns the size of the allocated area */
  214. static starpu_ssize_t allocate_csr_buffer_on_node(void *data_interface_, unsigned dst_node)
  215. {
  216. uintptr_t addr_nzval = 0;
  217. uint32_t *addr_colind = NULL, *addr_rowptr = NULL;
  218. starpu_ssize_t allocated_memory;
  219. /* we need the 3 arrays to be allocated */
  220. struct starpu_csr_interface *csr_interface = (struct starpu_csr_interface *) data_interface_;
  221. uint32_t nnz = csr_interface->nnz;
  222. uint32_t nrow = csr_interface->nrow;
  223. size_t elemsize = csr_interface->elemsize;
  224. addr_nzval = starpu_malloc_on_node(dst_node, nnz*elemsize);
  225. if (!addr_nzval)
  226. goto fail_nzval;
  227. addr_colind = (uint32_t*) starpu_malloc_on_node(dst_node, nnz*sizeof(uint32_t));
  228. if (!addr_colind)
  229. goto fail_colind;
  230. addr_rowptr = (uint32_t*) starpu_malloc_on_node(dst_node, (nrow+1)*sizeof(uint32_t));
  231. if (!addr_rowptr)
  232. goto fail_rowptr;
  233. /* allocation succeeded */
  234. allocated_memory =
  235. nnz*elemsize + nnz*sizeof(uint32_t) + (nrow+1)*sizeof(uint32_t);
  236. /* update the data properly in consequence */
  237. csr_interface->nzval = addr_nzval;
  238. csr_interface->colind = addr_colind;
  239. csr_interface->rowptr = addr_rowptr;
  240. return allocated_memory;
  241. fail_rowptr:
  242. starpu_free_on_node(dst_node, (uintptr_t) addr_colind, nnz*sizeof(uint32_t));
  243. fail_colind:
  244. starpu_free_on_node(dst_node, addr_nzval, nnz*elemsize);
  245. fail_nzval:
  246. /* allocation failed */
  247. return -ENOMEM;
  248. }
  249. static void free_csr_buffer_on_node(void *data_interface, unsigned node)
  250. {
  251. struct starpu_csr_interface *csr_interface = (struct starpu_csr_interface *) data_interface;
  252. uint32_t nnz = csr_interface->nnz;
  253. uint32_t nrow = csr_interface->nrow;
  254. size_t elemsize = csr_interface->elemsize;
  255. starpu_free_on_node(node, csr_interface->nzval, nnz*elemsize);
  256. starpu_free_on_node(node, (uintptr_t) csr_interface->colind, nnz*sizeof(uint32_t));
  257. starpu_free_on_node(node, (uintptr_t) csr_interface->rowptr, (nrow+1)*sizeof(uint32_t));
  258. }
  259. /* as not all platform easily have a BLAS lib installed ... */
  260. static int copy_any_to_any(void *src_interface, unsigned src_node, void *dst_interface, unsigned dst_node, void *async_data)
  261. {
  262. struct starpu_csr_interface *src_csr = (struct starpu_csr_interface *) src_interface;
  263. struct starpu_csr_interface *dst_csr = (struct starpu_csr_interface *) dst_interface;
  264. uint32_t nnz = src_csr->nnz;
  265. uint32_t nrow = src_csr->nrow;
  266. size_t elemsize = src_csr->elemsize;
  267. int ret = 0;
  268. if (starpu_interface_copy(src_csr->nzval, 0, src_node, dst_csr->nzval, 0, dst_node, nnz*elemsize, async_data))
  269. ret = -EAGAIN;
  270. if (starpu_interface_copy((uintptr_t)src_csr->colind, 0, src_node, (uintptr_t)dst_csr->colind, 0, dst_node, nnz*sizeof(uint32_t), async_data))
  271. ret = -EAGAIN;
  272. if (starpu_interface_copy((uintptr_t)src_csr->rowptr, 0, src_node, (uintptr_t)dst_csr->rowptr, 0, dst_node, (nrow+1)*sizeof(uint32_t), async_data))
  273. ret = -EAGAIN;
  274. starpu_interface_data_copy(src_node, dst_node, nnz*elemsize + (nnz+nrow+1)*sizeof(uint32_t));
  275. return ret;
  276. }
  277. static starpu_ssize_t describe(void *data_interface, char *buf, size_t size)
  278. {
  279. struct starpu_csr_interface *csr = (struct starpu_csr_interface *) data_interface;
  280. return snprintf(buf, size, "C%ux%ux%u",
  281. (unsigned) csr->nnz,
  282. (unsigned) csr->nrow,
  283. (unsigned) csr->elemsize);
  284. }
  285. static int pack_data(starpu_data_handle_t handle, unsigned node, void **ptr, starpu_ssize_t *count)
  286. {
  287. STARPU_ASSERT(starpu_data_test_if_allocated_on_node(handle, node));
  288. struct starpu_csr_interface *csr = (struct starpu_csr_interface *) starpu_data_get_interface_on_node(handle, node);
  289. // We first pack colind
  290. *count = csr->nnz * sizeof(csr->colind[0]);
  291. // Then rowptr
  292. *count += (csr->nrow + 1) * sizeof(csr->rowptr[0]);
  293. // Then nnzval
  294. *count += csr->nnz * csr->elemsize;
  295. if (ptr != NULL)
  296. {
  297. *ptr = (void *)starpu_malloc_on_node_flags(node, *count, 0);
  298. char *tmp = *ptr;
  299. memcpy(tmp, (void*)csr->colind, csr->nnz * sizeof(csr->colind[0]));
  300. tmp += csr->nnz * sizeof(csr->colind[0]);
  301. memcpy(tmp, (void*)csr->rowptr, (csr->nrow + 1) * sizeof(csr->rowptr[0]));
  302. tmp += (csr->nrow + 1) * sizeof(csr->rowptr[0]);
  303. memcpy(tmp, (void*)csr->nzval, csr->nnz * csr->elemsize);
  304. }
  305. return 0;
  306. }
  307. static int unpack_data(starpu_data_handle_t handle, unsigned node, void *ptr, size_t count)
  308. {
  309. STARPU_ASSERT(starpu_data_test_if_allocated_on_node(handle, node));
  310. struct starpu_csr_interface *csr = (struct starpu_csr_interface *) starpu_data_get_interface_on_node(handle, node);
  311. STARPU_ASSERT(count == (csr->nnz * sizeof(csr->colind[0]))+((csr->nrow + 1) * sizeof(csr->rowptr[0]))+(csr->nnz * csr->elemsize));
  312. char *tmp = ptr;
  313. memcpy((void*)csr->colind, tmp, csr->nnz * sizeof(csr->colind[0]));
  314. tmp += csr->nnz * sizeof(csr->colind[0]);
  315. memcpy((void*)csr->rowptr, tmp, (csr->nrow + 1) * sizeof(csr->rowptr[0]));
  316. tmp += (csr->nrow + 1) * sizeof(csr->rowptr[0]);
  317. memcpy((void*)csr->nzval, tmp, csr->nnz * csr->elemsize);
  318. starpu_free_on_node_flags(node, (uintptr_t)ptr, count, 0);
  319. return 0;
  320. }