bcsr_interface.c 11 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352
  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 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 starpu_ssize_t allocate_bcsr_buffer_on_node(void *data_interface, unsigned dst_node);
  37. static void free_bcsr_buffer_on_node(void *data_interface, unsigned node);
  38. static size_t bcsr_interface_get_size(starpu_data_handle_t handle);
  39. static int bcsr_compare(void *data_interface_a, void *data_interface_b);
  40. static uint32_t footprint_bcsr_interface_crc32(starpu_data_handle_t handle);
  41. static starpu_ssize_t describe(void *data_interface, char *buf, size_t size);
  42. struct starpu_data_interface_ops starpu_interface_bcsr_ops =
  43. {
  44. .register_data_handle = register_bcsr_handle,
  45. .allocate_data_on_node = allocate_bcsr_buffer_on_node,
  46. .free_data_on_node = free_bcsr_buffer_on_node,
  47. .copy_methods = &bcsr_copy_data_methods_s,
  48. .get_size = bcsr_interface_get_size,
  49. .interfaceid = STARPU_BCSR_INTERFACE_ID,
  50. .interface_size = sizeof(struct starpu_bcsr_interface),
  51. .footprint = footprint_bcsr_interface_crc32,
  52. .compare = bcsr_compare,
  53. .describe = describe
  54. };
  55. static void register_bcsr_handle(starpu_data_handle_t handle, unsigned home_node, void *data_interface)
  56. {
  57. struct starpu_bcsr_interface *bcsr_interface = (struct starpu_bcsr_interface *) data_interface;
  58. unsigned node;
  59. for (node = 0; node < STARPU_MAXNODES; node++)
  60. {
  61. struct starpu_bcsr_interface *local_interface = (struct starpu_bcsr_interface *)
  62. starpu_data_get_interface_on_node(handle, node);
  63. if (node == home_node)
  64. {
  65. local_interface->nzval = bcsr_interface->nzval;
  66. local_interface->colind = bcsr_interface->colind;
  67. local_interface->rowptr = bcsr_interface->rowptr;
  68. }
  69. else
  70. {
  71. local_interface->nzval = 0;
  72. local_interface->colind = NULL;
  73. local_interface->rowptr = NULL;
  74. }
  75. local_interface->id = bcsr_interface->id;
  76. local_interface->nnz = bcsr_interface->nnz;
  77. local_interface->nrow = bcsr_interface->nrow;
  78. local_interface->firstentry = bcsr_interface->firstentry;
  79. local_interface->r = bcsr_interface->r;
  80. local_interface->c = bcsr_interface->c;
  81. local_interface->elemsize = bcsr_interface->elemsize;
  82. }
  83. }
  84. void starpu_bcsr_data_register(starpu_data_handle_t *handleptr, int home_node,
  85. uint32_t nnz, uint32_t nrow, uintptr_t nzval, uint32_t *colind,
  86. uint32_t *rowptr, uint32_t firstentry,
  87. uint32_t r, uint32_t c, size_t elemsize)
  88. {
  89. struct starpu_bcsr_interface bcsr_interface =
  90. {
  91. .id = STARPU_BCSR_INTERFACE_ID,
  92. .nzval = nzval,
  93. .colind = colind,
  94. .rowptr = rowptr,
  95. .nnz = nnz,
  96. .nrow = nrow,
  97. .firstentry = firstentry,
  98. .r = r,
  99. .c = c,
  100. .elemsize = elemsize
  101. };
  102. #ifndef STARPU_SIMGRID
  103. if (home_node == STARPU_MAIN_RAM)
  104. {
  105. STARPU_ASSERT_ACCESSIBLE(nzval);
  106. STARPU_ASSERT_ACCESSIBLE(nzval + nnz*elemsize*r*c - 1);
  107. STARPU_ASSERT_ACCESSIBLE(colind);
  108. STARPU_ASSERT_ACCESSIBLE((uintptr_t) colind + nnz*sizeof(uint32_t) - 1);
  109. STARPU_ASSERT_ACCESSIBLE(rowptr);
  110. STARPU_ASSERT_ACCESSIBLE((uintptr_t) rowptr + (nrow+1)*sizeof(uint32_t) - 1);
  111. }
  112. #endif
  113. starpu_data_register(handleptr, home_node, &bcsr_interface, &starpu_interface_bcsr_ops);
  114. }
  115. static uint32_t footprint_bcsr_interface_crc32(starpu_data_handle_t handle)
  116. {
  117. uint32_t hash;
  118. hash = starpu_hash_crc32c_be(starpu_bcsr_get_nnz(handle), 0);
  119. hash = starpu_hash_crc32c_be(starpu_bcsr_get_c(handle), hash);
  120. hash = starpu_hash_crc32c_be(starpu_bcsr_get_r(handle), hash);
  121. return hash;
  122. }
  123. static int bcsr_compare(void *data_interface_a, void *data_interface_b)
  124. {
  125. struct starpu_bcsr_interface *bcsr_a = (struct starpu_bcsr_interface *) data_interface_a;
  126. struct starpu_bcsr_interface *bcsr_b = (struct starpu_bcsr_interface *) data_interface_b;
  127. /* Two matricess are considered compatible if they have the same size */
  128. return ((bcsr_a->nnz == bcsr_b->nnz)
  129. && (bcsr_a->nrow == bcsr_b->nrow)
  130. && (bcsr_a->r == bcsr_b->r)
  131. && (bcsr_a->c == bcsr_b->c)
  132. && (bcsr_a->elemsize == bcsr_b->elemsize));
  133. }
  134. /* offer an access to the data parameters */
  135. uint32_t starpu_bcsr_get_nnz(starpu_data_handle_t handle)
  136. {
  137. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  138. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  139. return data_interface->nnz;
  140. }
  141. uint32_t starpu_bcsr_get_nrow(starpu_data_handle_t handle)
  142. {
  143. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  144. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  145. return data_interface->nrow;
  146. }
  147. uint32_t starpu_bcsr_get_firstentry(starpu_data_handle_t handle)
  148. {
  149. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  150. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  151. return data_interface->firstentry;
  152. }
  153. uint32_t starpu_bcsr_get_r(starpu_data_handle_t handle)
  154. {
  155. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  156. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  157. return data_interface->r;
  158. }
  159. uint32_t starpu_bcsr_get_c(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. return data_interface->c;
  164. }
  165. size_t starpu_bcsr_get_elemsize(starpu_data_handle_t handle)
  166. {
  167. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  168. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  169. return data_interface->elemsize;
  170. }
  171. uintptr_t starpu_bcsr_get_local_nzval(starpu_data_handle_t handle)
  172. {
  173. unsigned node;
  174. node = _starpu_memory_node_get_local_key();
  175. STARPU_ASSERT(starpu_data_test_if_allocated_on_node(handle, node));
  176. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  177. starpu_data_get_interface_on_node(handle, node);
  178. return data_interface->nzval;
  179. }
  180. uint32_t *starpu_bcsr_get_local_colind(starpu_data_handle_t handle)
  181. {
  182. /* XXX 0 */
  183. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  184. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  185. return data_interface->colind;
  186. }
  187. uint32_t *starpu_bcsr_get_local_rowptr(starpu_data_handle_t handle)
  188. {
  189. /* XXX 0 */
  190. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  191. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  192. return data_interface->rowptr;
  193. }
  194. static size_t bcsr_interface_get_size(starpu_data_handle_t handle)
  195. {
  196. size_t size;
  197. uint32_t nnz = starpu_bcsr_get_nnz(handle);
  198. uint32_t nrow = starpu_bcsr_get_nrow(handle);
  199. uint32_t r = starpu_bcsr_get_r(handle);
  200. uint32_t c = starpu_bcsr_get_c(handle);
  201. size_t elemsize = starpu_bcsr_get_elemsize(handle);
  202. size = nnz*r*c*elemsize + nnz*sizeof(uint32_t) + (nrow+1)*sizeof(uint32_t);
  203. return size;
  204. }
  205. /* memory allocation/deallocation primitives for the BLAS interface */
  206. /* returns the size of the allocated area */
  207. static starpu_ssize_t allocate_bcsr_buffer_on_node(void *data_interface_, unsigned dst_node)
  208. {
  209. uintptr_t addr_nzval, addr_colind, addr_rowptr;
  210. starpu_ssize_t allocated_memory;
  211. /* we need the 3 arrays to be allocated */
  212. struct starpu_bcsr_interface *bcsr_interface = (struct starpu_bcsr_interface *) data_interface_;
  213. uint32_t nnz = bcsr_interface->nnz;
  214. uint32_t nrow = bcsr_interface->nrow;
  215. size_t elemsize = bcsr_interface->elemsize;
  216. uint32_t r = bcsr_interface->r;
  217. uint32_t c = bcsr_interface->c;
  218. addr_nzval = starpu_malloc_on_node(dst_node, nnz*r*c*elemsize);
  219. if (!addr_nzval)
  220. goto fail_nzval;
  221. addr_colind = starpu_malloc_on_node(dst_node, nnz*sizeof(uint32_t));
  222. if (!addr_colind)
  223. goto fail_colind;
  224. addr_rowptr = starpu_malloc_on_node(dst_node, (nrow+1)*sizeof(uint32_t));
  225. if (!addr_rowptr)
  226. goto fail_rowptr;
  227. /* allocation succeeded */
  228. allocated_memory =
  229. nnz*r*c*elemsize + nnz*sizeof(uint32_t) + (nrow+1)*sizeof(uint32_t);
  230. /* update the data properly in consequence */
  231. bcsr_interface->nzval = addr_nzval;
  232. bcsr_interface->colind = (uint32_t*) addr_colind;
  233. bcsr_interface->rowptr = (uint32_t*) addr_rowptr;
  234. return allocated_memory;
  235. fail_rowptr:
  236. starpu_free_on_node(dst_node, addr_colind, nnz*sizeof(uint32_t));
  237. fail_colind:
  238. starpu_free_on_node(dst_node, addr_nzval, nnz*r*c*elemsize);
  239. fail_nzval:
  240. /* allocation failed */
  241. return -ENOMEM;
  242. }
  243. static void free_bcsr_buffer_on_node(void *data_interface, unsigned node)
  244. {
  245. struct starpu_bcsr_interface *bcsr_interface = (struct starpu_bcsr_interface *) data_interface;
  246. uint32_t nnz = bcsr_interface->nnz;
  247. uint32_t nrow = bcsr_interface->nrow;
  248. size_t elemsize = bcsr_interface->elemsize;
  249. uint32_t r = bcsr_interface->r;
  250. uint32_t c = bcsr_interface->c;
  251. starpu_free_on_node(node, bcsr_interface->nzval, nnz*r*c*elemsize);
  252. starpu_free_on_node(node, (uintptr_t) bcsr_interface->colind, nnz*sizeof(uint32_t));
  253. starpu_free_on_node(node, (uintptr_t) bcsr_interface->rowptr, (nrow+1)*sizeof(uint32_t));
  254. }
  255. static int copy_any_to_any(void *src_interface, unsigned src_node, void *dst_interface, unsigned dst_node, void *async_data)
  256. {
  257. struct starpu_bcsr_interface *src_bcsr = (struct starpu_bcsr_interface *) src_interface;
  258. struct starpu_bcsr_interface *dst_bcsr = (struct starpu_bcsr_interface *) dst_interface;
  259. uint32_t nnz = src_bcsr->nnz;
  260. uint32_t nrow = src_bcsr->nrow;
  261. size_t elemsize = src_bcsr->elemsize;
  262. uint32_t r = src_bcsr->r;
  263. uint32_t c = src_bcsr->c;
  264. int ret = 0;
  265. if (starpu_interface_copy(src_bcsr->nzval, 0, src_node, dst_bcsr->nzval, 0, dst_node, nnz*elemsize*r*c, async_data))
  266. ret = -EAGAIN;
  267. 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))
  268. ret = -EAGAIN;
  269. 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))
  270. ret = -EAGAIN;
  271. _STARPU_TRACE_DATA_COPY(src_node, dst_node, nnz*elemsize*r*c + (nnz+nrow+1)*sizeof(uint32_t));
  272. return ret;
  273. }
  274. static starpu_ssize_t describe(void *data_interface, char *buf, size_t size)
  275. {
  276. struct starpu_bcsr_interface *bcsr = (struct starpu_bcsr_interface *) data_interface;
  277. return snprintf(buf, size, "b%ux%ux%ux%ux%u",
  278. (unsigned) bcsr->nnz,
  279. (unsigned) bcsr->nrow,
  280. (unsigned) bcsr->r,
  281. (unsigned) bcsr->c,
  282. (unsigned) bcsr->elemsize);
  283. }