You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

laswp_ncopy_4.c 10 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503
  1. /*********************************************************************/
  2. /* Copyright 2009, 2010 The University of Texas at Austin. */
  3. /* All rights reserved. */
  4. /* */
  5. /* Redistribution and use in source and binary forms, with or */
  6. /* without modification, are permitted provided that the following */
  7. /* conditions are met: */
  8. /* */
  9. /* 1. Redistributions of source code must retain the above */
  10. /* copyright notice, this list of conditions and the following */
  11. /* disclaimer. */
  12. /* */
  13. /* 2. Redistributions in binary form must reproduce the above */
  14. /* copyright notice, this list of conditions and the following */
  15. /* disclaimer in the documentation and/or other materials */
  16. /* provided with the distribution. */
  17. /* */
  18. /* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
  19. /* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
  20. /* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
  21. /* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
  22. /* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
  23. /* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
  24. /* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
  25. /* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
  26. /* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
  27. /* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
  28. /* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
  29. /* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
  30. /* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
  31. /* POSSIBILITY OF SUCH DAMAGE. */
  32. /* */
  33. /* The views and conclusions contained in the software and */
  34. /* documentation are those of the authors and should not be */
  35. /* interpreted as representing official policies, either expressed */
  36. /* or implied, of The University of Texas at Austin. */
  37. /*********************************************************************/
  38. #include <stdio.h>
  39. #include "common.h"
  40. #define PREFETCHSIZE 8
  41. #define a2 (a1 + 1)
  42. #define a4 (a3 + 1)
  43. #define a6 (a5 + 1)
  44. #define a8 (a7 + 1)
  45. int CNAME(BLASLONG n, BLASLONG k1, BLASLONG k2, FLOAT *a, BLASLONG lda, blasint *ipiv, FLOAT *buffer){
  46. BLASLONG i, j, ip1, ip2;
  47. blasint *piv;
  48. FLOAT *a1, *a3, *a5, *a7;
  49. FLOAT *b1, *b2, *b3, *b4;
  50. FLOAT *b5, *b6, *b7, *b8;
  51. FLOAT A1, A2, B1, B2, A3, A4, B3, B4;
  52. FLOAT A5, A6, B5, B6, A7, A8, B7, B8;
  53. a--;
  54. k1 --;
  55. ipiv += k1;
  56. if (n <= 0) return 0;
  57. j = (n >> 2);
  58. if (j > 0) {
  59. do {
  60. piv = ipiv;
  61. a1 = a + k1 + 1;
  62. a3 = a1 + 1 * lda;
  63. a5 = a1 + 2 * lda;
  64. a7 = a1 + 3 * lda;
  65. ip1 = *(piv + 0);
  66. ip2 = *(piv + 1);
  67. piv += 2;
  68. b1 = a + ip1;
  69. b2 = a + ip2;
  70. b3 = b1 + 1 * lda;
  71. b4 = b2 + 1 * lda;
  72. b5 = b1 + 2 * lda;
  73. b6 = b2 + 2 * lda;
  74. b7 = b1 + 3 * lda;
  75. b8 = b2 + 3 * lda;
  76. i = ((k2 - k1) >> 1);
  77. if (i > 0) {
  78. do {
  79. #ifdef __GNUC__
  80. __builtin_prefetch(a1 + PREFETCHSIZE, 0, 0);
  81. __builtin_prefetch(a3 + PREFETCHSIZE, 0, 0);
  82. __builtin_prefetch(a5 + PREFETCHSIZE, 0, 0);
  83. __builtin_prefetch(a7 + PREFETCHSIZE, 0, 0);
  84. #endif
  85. A1 = *a1;
  86. A2 = *a2;
  87. A3 = *a3;
  88. A4 = *a4;
  89. A5 = *a5;
  90. A6 = *a6;
  91. A7 = *a7;
  92. A8 = *a8;
  93. B1 = *b1;
  94. B2 = *b2;
  95. B3 = *b3;
  96. B4 = *b4;
  97. B5 = *b5;
  98. B6 = *b6;
  99. B7 = *b7;
  100. B8 = *b8;
  101. ip1 = *(piv + 0);
  102. ip2 = *(piv + 1);
  103. piv += 2;
  104. if (b1 == a1) {
  105. if (b2 == a2) {
  106. *(buffer + 0) = A1;
  107. *(buffer + 1) = A3;
  108. *(buffer + 2) = A5;
  109. *(buffer + 3) = A7;
  110. *(buffer + 4) = A2;
  111. *(buffer + 5) = A4;
  112. *(buffer + 6) = A6;
  113. *(buffer + 7) = A8;
  114. } else {
  115. *(buffer + 0) = A1;
  116. *(buffer + 1) = A3;
  117. *(buffer + 2) = A5;
  118. *(buffer + 3) = A7;
  119. *(buffer + 4) = B2;
  120. *(buffer + 5) = B4;
  121. *(buffer + 6) = B6;
  122. *(buffer + 7) = B8;
  123. *b2 = A2;
  124. *b4 = A4;
  125. *b6 = A6;
  126. *b8 = A8;
  127. }
  128. } else
  129. if (b1 == a2) {
  130. if (b2 == a2) {
  131. *(buffer + 0) = A2;
  132. *(buffer + 1) = A4;
  133. *(buffer + 2) = A6;
  134. *(buffer + 3) = A8;
  135. *(buffer + 4) = A1;
  136. *(buffer + 5) = A3;
  137. *(buffer + 6) = A5;
  138. *(buffer + 7) = A7;
  139. } else {
  140. *(buffer + 0) = A2;
  141. *(buffer + 1) = A4;
  142. *(buffer + 2) = A6;
  143. *(buffer + 3) = A8;
  144. *(buffer + 4) = B2;
  145. *(buffer + 5) = B4;
  146. *(buffer + 6) = B6;
  147. *(buffer + 7) = B8;
  148. *b2 = A1;
  149. *b4 = A3;
  150. *b6 = A5;
  151. *b8 = A7;
  152. }
  153. } else {
  154. if (b2 == a2) {
  155. *(buffer + 0) = B1;
  156. *(buffer + 1) = B3;
  157. *(buffer + 2) = B5;
  158. *(buffer + 3) = B7;
  159. *(buffer + 4) = A2;
  160. *(buffer + 5) = A4;
  161. *(buffer + 6) = A6;
  162. *(buffer + 7) = A8;
  163. *b1 = A1;
  164. *b3 = A3;
  165. *b5 = A5;
  166. *b7 = A7;
  167. } else
  168. if (b2 == b1) {
  169. *(buffer + 0) = B1;
  170. *(buffer + 1) = B3;
  171. *(buffer + 2) = B5;
  172. *(buffer + 3) = B7;
  173. *(buffer + 4) = A1;
  174. *(buffer + 5) = A3;
  175. *(buffer + 6) = A5;
  176. *(buffer + 7) = A7;
  177. *b1 = A2;
  178. *b3 = A4;
  179. *b5 = A6;
  180. *b7 = A8;
  181. } else {
  182. *(buffer + 0) = B1;
  183. *(buffer + 1) = B3;
  184. *(buffer + 2) = B5;
  185. *(buffer + 3) = B7;
  186. *(buffer + 4) = B2;
  187. *(buffer + 5) = B4;
  188. *(buffer + 6) = B6;
  189. *(buffer + 7) = B8;
  190. *b1 = A1;
  191. *b2 = A2;
  192. *b3 = A3;
  193. *b4 = A4;
  194. *b5 = A5;
  195. *b6 = A6;
  196. *b7 = A7;
  197. *b8 = A8;
  198. }
  199. }
  200. buffer += 8;
  201. b1 = a + ip1;
  202. b2 = a + ip2;
  203. b3 = b1 + 1 * lda;
  204. b4 = b2 + 1 * lda;
  205. b5 = b1 + 2 * lda;
  206. b6 = b2 + 2 * lda;
  207. b7 = b1 + 3 * lda;
  208. b8 = b2 + 3 * lda;
  209. a1 += 2;
  210. a3 += 2;
  211. a5 += 2;
  212. a7 += 2;
  213. i --;
  214. } while (i > 0);
  215. }
  216. i = ((k2 - k1) & 1);
  217. if (i > 0) {
  218. A1 = *a1;
  219. B1 = *b1;
  220. A3 = *a3;
  221. B3 = *b3;
  222. A5 = *a5;
  223. B5 = *b5;
  224. A7 = *a7;
  225. B7 = *b7;
  226. if (a1 == b1) {
  227. *(buffer + 0) = A1;
  228. *(buffer + 1) = A3;
  229. *(buffer + 2) = A5;
  230. *(buffer + 3) = A7;
  231. } else {
  232. *(buffer + 0) = B1;
  233. *(buffer + 1) = B3;
  234. *(buffer + 2) = B5;
  235. *(buffer + 3) = B7;
  236. *b1 = A1;
  237. *b3 = A3;
  238. *b5 = A5;
  239. *b7 = A7;
  240. }
  241. buffer += 4;
  242. }
  243. a += 4 * lda;
  244. j --;
  245. } while (j > 0);
  246. }
  247. if (n & 2) {
  248. piv = ipiv;
  249. a1 = a + k1 + 1;
  250. a3 = a1 + 1 * lda;
  251. ip1 = *(piv + 0);
  252. ip2 = *(piv + 1);
  253. piv += 2;
  254. b1 = a + ip1;
  255. b2 = a + ip2;
  256. b3 = b1 + 1 * lda;
  257. b4 = b2 + 1 * lda;
  258. i = ((k2 - k1) >> 1);
  259. if (i > 0) {
  260. do {
  261. A1 = *a1;
  262. A2 = *a2;
  263. A3 = *a3;
  264. A4 = *a4;
  265. B1 = *b1;
  266. B2 = *b2;
  267. B3 = *b3;
  268. B4 = *b4;
  269. ip1 = *(piv + 0);
  270. ip2 = *(piv + 1);
  271. piv += 2;
  272. if (b1 == a1) {
  273. if (b2 == a2) {
  274. *(buffer + 0) = A1;
  275. *(buffer + 1) = A3;
  276. *(buffer + 2) = A2;
  277. *(buffer + 3) = A4;
  278. } else {
  279. *(buffer + 0) = A1;
  280. *(buffer + 1) = A3;
  281. *(buffer + 2) = B2;
  282. *(buffer + 3) = B4;
  283. *b2 = A2;
  284. *b4 = A4;
  285. }
  286. } else
  287. if (b1 == a2) {
  288. if (b2 == a2) {
  289. *(buffer + 0) = A2;
  290. *(buffer + 1) = A4;
  291. *(buffer + 2) = A1;
  292. *(buffer + 3) = A3;
  293. } else {
  294. *(buffer + 0) = A2;
  295. *(buffer + 1) = A4;
  296. *(buffer + 2) = B2;
  297. *(buffer + 3) = B4;
  298. *b2 = A1;
  299. *b4 = A3;
  300. }
  301. } else {
  302. if (b2 == a2) {
  303. *(buffer + 0) = B1;
  304. *(buffer + 1) = B3;
  305. *(buffer + 2) = A2;
  306. *(buffer + 3) = A4;
  307. *b1 = A1;
  308. *b3 = A3;
  309. } else
  310. if (b2 == b1) {
  311. *(buffer + 0) = B1;
  312. *(buffer + 1) = B3;
  313. *(buffer + 2) = A1;
  314. *(buffer + 3) = A3;
  315. *b1 = A2;
  316. *b3 = A4;
  317. } else {
  318. *(buffer + 0) = B1;
  319. *(buffer + 1) = B3;
  320. *(buffer + 2) = B2;
  321. *(buffer + 3) = B4;
  322. *b1 = A1;
  323. *b2 = A2;
  324. *b3 = A3;
  325. *b4 = A4;
  326. }
  327. }
  328. buffer += 4;
  329. b1 = a + ip1;
  330. b2 = a + ip2;
  331. b3 = b1 + 1 * lda;
  332. b4 = b2 + 1 * lda;
  333. a1 += 2;
  334. a3 += 2;
  335. i --;
  336. } while (i > 0);
  337. }
  338. i = ((k2 - k1) & 1);
  339. if (i > 0) {
  340. A1 = *a1;
  341. B1 = *b1;
  342. A3 = *a3;
  343. B3 = *b3;
  344. if (a1 == b1) {
  345. *(buffer + 0) = A1;
  346. *(buffer + 1) = A3;
  347. } else {
  348. *(buffer + 0) = B1;
  349. *(buffer + 1) = B3;
  350. *b1 = A1;
  351. *b3 = A3;
  352. }
  353. buffer += 2;
  354. }
  355. a += 2 * lda;
  356. }
  357. if (n & 1) {
  358. piv = ipiv;
  359. a1 = a + k1 + 1;
  360. ip1 = *(piv + 0);
  361. ip2 = *(piv + 1);
  362. piv += 2;
  363. b1 = a + ip1;
  364. b2 = a + ip2;
  365. i = ((k2 - k1) >> 1);
  366. if (i > 0) {
  367. do {
  368. A1 = *a1;
  369. A2 = *a2;
  370. B1 = *b1;
  371. B2 = *b2;
  372. ip1 = *(piv + 0);
  373. ip2 = *(piv + 1);
  374. piv += 2;
  375. if (b1 == a1) {
  376. if (b2 == a2) {
  377. *(buffer + 0) = A1;
  378. *(buffer + 1) = A2;
  379. } else {
  380. *(buffer + 0) = A1;
  381. *(buffer + 1) = B2;
  382. *b2 = A2;
  383. }
  384. } else
  385. if (b1 == a2) {
  386. if (b2 == a2) {
  387. *(buffer + 0) = A2;
  388. *(buffer + 1) = A1;
  389. } else {
  390. *(buffer + 0) = A2;
  391. *(buffer + 1) = B2;
  392. *b2 = A1;
  393. }
  394. } else {
  395. if (b2 == a2) {
  396. *(buffer + 0) = B1;
  397. *(buffer + 1) = A2;
  398. *b1 = A1;
  399. } else
  400. if (b2 == b1) {
  401. *(buffer + 0) = B1;
  402. *(buffer + 1) = A1;
  403. *b1 = A2;
  404. } else {
  405. *(buffer + 0) = B1;
  406. *(buffer + 1) = B2;
  407. *b1 = A1;
  408. *b2 = A2;
  409. }
  410. }
  411. buffer += 2;
  412. b1 = a + ip1;
  413. b2 = a + ip2;
  414. a1 += 2;
  415. i --;
  416. } while (i > 0);
  417. }
  418. i = ((k2 - k1) & 1);
  419. if (i > 0) {
  420. A1 = *a1;
  421. B1 = *b1;
  422. if (a1 == b1) {
  423. *(buffer + 0) = A1;
  424. } else {
  425. *(buffer + 0) = B1;
  426. *b1 = A1;
  427. }
  428. }
  429. }
  430. return 0;
  431. }