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.

trsm_ltcopy_8.c 18 kB

9 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921
  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. #ifndef UNIT
  41. #define INV(a) (ONE / (a))
  42. #else
  43. #define INV(a) (ONE)
  44. #endif
  45. int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT *b){
  46. BLASLONG i, ii, j, jj;
  47. FLOAT data01, data02, data03, data04, data05, data06, data07, data08;
  48. FLOAT data09, data10, data11, data12, data13, data14, data15, data16;
  49. FLOAT data17, data18, data19, data20, data21, data22, data23, data24;
  50. FLOAT data25, data26, data27, data28, data29, data30, data31, data32;
  51. FLOAT data33, data34, data35, data36, data37, data38, data39, data40;
  52. FLOAT data41, data42, data43, data44, data45, data46, data47, data48;
  53. FLOAT data49, data50, data51, data52, data53, data54, data55, data56;
  54. FLOAT data57, data58, data59, data60, data61, data62, data63, data64;
  55. FLOAT *a1, *a2, *a3, *a4, *a5, *a6, *a7, *a8;
  56. jj = offset;
  57. j = (n >> 3);
  58. while (j > 0){
  59. a1 = a + 0 * lda;
  60. a2 = a + 1 * lda;
  61. a3 = a + 2 * lda;
  62. a4 = a + 3 * lda;
  63. a5 = a + 4 * lda;
  64. a6 = a + 5 * lda;
  65. a7 = a + 6 * lda;
  66. a8 = a + 7 * lda;
  67. ii = 0;
  68. i = (m >> 3);
  69. while (i > 0) {
  70. if (ii == jj) {
  71. #ifndef UNIT
  72. data01 = *(a1 + 0);
  73. #endif
  74. data02 = *(a1 + 1);
  75. data03 = *(a1 + 2);
  76. data04 = *(a1 + 3);
  77. data05 = *(a1 + 4);
  78. data06 = *(a1 + 5);
  79. data07 = *(a1 + 6);
  80. data08 = *(a1 + 7);
  81. #ifndef UNIT
  82. data10 = *(a2 + 1);
  83. #endif
  84. data11 = *(a2 + 2);
  85. data12 = *(a2 + 3);
  86. data13 = *(a2 + 4);
  87. data14 = *(a2 + 5);
  88. data15 = *(a2 + 6);
  89. data16 = *(a2 + 7);
  90. #ifndef UNIT
  91. data19 = *(a3 + 2);
  92. #endif
  93. data20 = *(a3 + 3);
  94. data21 = *(a3 + 4);
  95. data22 = *(a3 + 5);
  96. data23 = *(a3 + 6);
  97. data24 = *(a3 + 7);
  98. #ifndef UNIT
  99. data28 = *(a4 + 3);
  100. #endif
  101. data29 = *(a4 + 4);
  102. data30 = *(a4 + 5);
  103. data31 = *(a4 + 6);
  104. data32 = *(a4 + 7);
  105. #ifndef UNIT
  106. data37 = *(a5 + 4);
  107. #endif
  108. data38 = *(a5 + 5);
  109. data39 = *(a5 + 6);
  110. data40 = *(a5 + 7);
  111. #ifndef UNIT
  112. data46 = *(a6 + 5);
  113. #endif
  114. data47 = *(a6 + 6);
  115. data48 = *(a6 + 7);
  116. #ifndef UNIT
  117. data55 = *(a7 + 6);
  118. #endif
  119. data56 = *(a7 + 7);
  120. #ifndef UNIT
  121. data64 = *(a8 + 7);
  122. #endif
  123. *(b + 0) = INV(data01);
  124. *(b + 1) = data02;
  125. *(b + 2) = data03;
  126. *(b + 3) = data04;
  127. *(b + 4) = data05;
  128. *(b + 5) = data06;
  129. *(b + 6) = data07;
  130. *(b + 7) = data08;
  131. *(b + 9) = INV(data10);
  132. *(b + 10) = data11;
  133. *(b + 11) = data12;
  134. *(b + 12) = data13;
  135. *(b + 13) = data14;
  136. *(b + 14) = data15;
  137. *(b + 15) = data16;
  138. *(b + 18) = INV(data19);
  139. *(b + 19) = data20;
  140. *(b + 20) = data21;
  141. *(b + 21) = data22;
  142. *(b + 22) = data23;
  143. *(b + 23) = data24;
  144. *(b + 27) = INV(data28);
  145. *(b + 28) = data29;
  146. *(b + 29) = data30;
  147. *(b + 30) = data31;
  148. *(b + 31) = data32;
  149. *(b + 36) = INV(data37);
  150. *(b + 37) = data38;
  151. *(b + 38) = data39;
  152. *(b + 39) = data40;
  153. *(b + 45) = INV(data46);
  154. *(b + 46) = data47;
  155. *(b + 47) = data48;
  156. *(b + 54) = INV(data55);
  157. *(b + 55) = data56;
  158. *(b + 63) = INV(data64);
  159. }
  160. if (ii < jj) {
  161. data01 = *(a1 + 0);
  162. data02 = *(a1 + 1);
  163. data03 = *(a1 + 2);
  164. data04 = *(a1 + 3);
  165. data05 = *(a1 + 4);
  166. data06 = *(a1 + 5);
  167. data07 = *(a1 + 6);
  168. data08 = *(a1 + 7);
  169. data09 = *(a2 + 0);
  170. data10 = *(a2 + 1);
  171. data11 = *(a2 + 2);
  172. data12 = *(a2 + 3);
  173. data13 = *(a2 + 4);
  174. data14 = *(a2 + 5);
  175. data15 = *(a2 + 6);
  176. data16 = *(a2 + 7);
  177. data17 = *(a3 + 0);
  178. data18 = *(a3 + 1);
  179. data19 = *(a3 + 2);
  180. data20 = *(a3 + 3);
  181. data21 = *(a3 + 4);
  182. data22 = *(a3 + 5);
  183. data23 = *(a3 + 6);
  184. data24 = *(a3 + 7);
  185. data25 = *(a4 + 0);
  186. data26 = *(a4 + 1);
  187. data27 = *(a4 + 2);
  188. data28 = *(a4 + 3);
  189. data29 = *(a4 + 4);
  190. data30 = *(a4 + 5);
  191. data31 = *(a4 + 6);
  192. data32 = *(a4 + 7);
  193. data33 = *(a5 + 0);
  194. data34 = *(a5 + 1);
  195. data35 = *(a5 + 2);
  196. data36 = *(a5 + 3);
  197. data37 = *(a5 + 4);
  198. data38 = *(a5 + 5);
  199. data39 = *(a5 + 6);
  200. data40 = *(a5 + 7);
  201. data41 = *(a6 + 0);
  202. data42 = *(a6 + 1);
  203. data43 = *(a6 + 2);
  204. data44 = *(a6 + 3);
  205. data45 = *(a6 + 4);
  206. data46 = *(a6 + 5);
  207. data47 = *(a6 + 6);
  208. data48 = *(a6 + 7);
  209. data49 = *(a7 + 0);
  210. data50 = *(a7 + 1);
  211. data51 = *(a7 + 2);
  212. data52 = *(a7 + 3);
  213. data53 = *(a7 + 4);
  214. data54 = *(a7 + 5);
  215. data55 = *(a7 + 6);
  216. data56 = *(a7 + 7);
  217. data57 = *(a8 + 0);
  218. data58 = *(a8 + 1);
  219. data59 = *(a8 + 2);
  220. data60 = *(a8 + 3);
  221. data61 = *(a8 + 4);
  222. data62 = *(a8 + 5);
  223. data63 = *(a8 + 6);
  224. data64 = *(a8 + 7);
  225. *(b + 0) = data01;
  226. *(b + 1) = data02;
  227. *(b + 2) = data03;
  228. *(b + 3) = data04;
  229. *(b + 4) = data05;
  230. *(b + 5) = data06;
  231. *(b + 6) = data07;
  232. *(b + 7) = data08;
  233. *(b + 8) = data09;
  234. *(b + 9) = data10;
  235. *(b + 10) = data11;
  236. *(b + 11) = data12;
  237. *(b + 12) = data13;
  238. *(b + 13) = data14;
  239. *(b + 14) = data15;
  240. *(b + 15) = data16;
  241. *(b + 16) = data17;
  242. *(b + 17) = data18;
  243. *(b + 18) = data19;
  244. *(b + 19) = data20;
  245. *(b + 20) = data21;
  246. *(b + 21) = data22;
  247. *(b + 22) = data23;
  248. *(b + 23) = data24;
  249. *(b + 24) = data25;
  250. *(b + 25) = data26;
  251. *(b + 26) = data27;
  252. *(b + 27) = data28;
  253. *(b + 28) = data29;
  254. *(b + 29) = data30;
  255. *(b + 30) = data31;
  256. *(b + 31) = data32;
  257. *(b + 32) = data33;
  258. *(b + 33) = data34;
  259. *(b + 34) = data35;
  260. *(b + 35) = data36;
  261. *(b + 36) = data37;
  262. *(b + 37) = data38;
  263. *(b + 38) = data39;
  264. *(b + 39) = data40;
  265. *(b + 40) = data41;
  266. *(b + 41) = data42;
  267. *(b + 42) = data43;
  268. *(b + 43) = data44;
  269. *(b + 44) = data45;
  270. *(b + 45) = data46;
  271. *(b + 46) = data47;
  272. *(b + 47) = data48;
  273. *(b + 48) = data49;
  274. *(b + 49) = data50;
  275. *(b + 50) = data51;
  276. *(b + 51) = data52;
  277. *(b + 52) = data53;
  278. *(b + 53) = data54;
  279. *(b + 54) = data55;
  280. *(b + 55) = data56;
  281. *(b + 56) = data57;
  282. *(b + 57) = data58;
  283. *(b + 58) = data59;
  284. *(b + 59) = data60;
  285. *(b + 60) = data61;
  286. *(b + 61) = data62;
  287. *(b + 62) = data63;
  288. *(b + 63) = data64;
  289. }
  290. a1 += 8 * lda;
  291. a2 += 8 * lda;
  292. a3 += 8 * lda;
  293. a4 += 8 * lda;
  294. a5 += 8 * lda;
  295. a6 += 8 * lda;
  296. a7 += 8 * lda;
  297. a8 += 8 * lda;
  298. b += 64;
  299. i --;
  300. ii += 8;
  301. }
  302. if (m & 4) {
  303. if (ii == jj) {
  304. #ifndef UNIT
  305. data01 = *(a1 + 0);
  306. #endif
  307. data02 = *(a1 + 1);
  308. data03 = *(a1 + 2);
  309. data04 = *(a1 + 3);
  310. data05 = *(a1 + 4);
  311. data06 = *(a1 + 5);
  312. data07 = *(a1 + 6);
  313. data08 = *(a1 + 7);
  314. #ifndef UNIT
  315. data10 = *(a2 + 1);
  316. #endif
  317. data11 = *(a2 + 2);
  318. data12 = *(a2 + 3);
  319. data13 = *(a2 + 4);
  320. data14 = *(a2 + 5);
  321. data15 = *(a2 + 6);
  322. data16 = *(a2 + 7);
  323. #ifndef UNIT
  324. data19 = *(a3 + 2);
  325. #endif
  326. data20 = *(a3 + 3);
  327. data21 = *(a3 + 4);
  328. data22 = *(a3 + 5);
  329. data23 = *(a3 + 6);
  330. data24 = *(a3 + 7);
  331. #ifndef UNIT
  332. data28 = *(a4 + 3);
  333. #endif
  334. data29 = *(a4 + 4);
  335. data30 = *(a4 + 5);
  336. data31 = *(a4 + 6);
  337. data32 = *(a4 + 7);
  338. *(b + 0) = INV(data01);
  339. *(b + 1) = data02;
  340. *(b + 2) = data03;
  341. *(b + 3) = data04;
  342. *(b + 4) = data05;
  343. *(b + 5) = data06;
  344. *(b + 6) = data07;
  345. *(b + 7) = data08;
  346. *(b + 9) = INV(data10);
  347. *(b + 10) = data11;
  348. *(b + 11) = data12;
  349. *(b + 12) = data13;
  350. *(b + 13) = data14;
  351. *(b + 14) = data15;
  352. *(b + 15) = data16;
  353. *(b + 18) = INV(data19);
  354. *(b + 19) = data20;
  355. *(b + 20) = data21;
  356. *(b + 21) = data22;
  357. *(b + 22) = data23;
  358. *(b + 23) = data24;
  359. *(b + 27) = INV(data28);
  360. *(b + 28) = data29;
  361. *(b + 29) = data30;
  362. *(b + 30) = data31;
  363. *(b + 31) = data32;
  364. }
  365. if (ii < jj) {
  366. data01 = *(a1 + 0);
  367. data02 = *(a1 + 1);
  368. data03 = *(a1 + 2);
  369. data04 = *(a1 + 3);
  370. data05 = *(a1 + 4);
  371. data06 = *(a1 + 5);
  372. data07 = *(a1 + 6);
  373. data08 = *(a1 + 7);
  374. data09 = *(a2 + 0);
  375. data10 = *(a2 + 1);
  376. data11 = *(a2 + 2);
  377. data12 = *(a2 + 3);
  378. data13 = *(a2 + 4);
  379. data14 = *(a2 + 5);
  380. data15 = *(a2 + 6);
  381. data16 = *(a2 + 7);
  382. data17 = *(a3 + 0);
  383. data18 = *(a3 + 1);
  384. data19 = *(a3 + 2);
  385. data20 = *(a3 + 3);
  386. data21 = *(a3 + 4);
  387. data22 = *(a3 + 5);
  388. data23 = *(a3 + 6);
  389. data24 = *(a3 + 7);
  390. data25 = *(a4 + 0);
  391. data26 = *(a4 + 1);
  392. data27 = *(a4 + 2);
  393. data28 = *(a4 + 3);
  394. data29 = *(a4 + 4);
  395. data30 = *(a4 + 5);
  396. data31 = *(a4 + 6);
  397. data32 = *(a4 + 7);
  398. *(b + 0) = data01;
  399. *(b + 1) = data02;
  400. *(b + 2) = data03;
  401. *(b + 3) = data04;
  402. *(b + 4) = data05;
  403. *(b + 5) = data06;
  404. *(b + 6) = data07;
  405. *(b + 7) = data08;
  406. *(b + 8) = data09;
  407. *(b + 9) = data10;
  408. *(b + 10) = data11;
  409. *(b + 11) = data12;
  410. *(b + 12) = data13;
  411. *(b + 13) = data14;
  412. *(b + 14) = data15;
  413. *(b + 15) = data16;
  414. *(b + 16) = data17;
  415. *(b + 17) = data18;
  416. *(b + 18) = data19;
  417. *(b + 19) = data20;
  418. *(b + 20) = data21;
  419. *(b + 21) = data22;
  420. *(b + 22) = data23;
  421. *(b + 23) = data24;
  422. *(b + 24) = data25;
  423. *(b + 25) = data26;
  424. *(b + 26) = data27;
  425. *(b + 27) = data28;
  426. *(b + 28) = data29;
  427. *(b + 29) = data30;
  428. *(b + 30) = data31;
  429. *(b + 31) = data32;
  430. }
  431. a1 += 4 * lda;
  432. a2 += 4 * lda;
  433. a3 += 4 * lda;
  434. a4 += 4 * lda;
  435. b += 32;
  436. ii += 4;
  437. }
  438. if (m & 2) {
  439. if (ii == jj) {
  440. #ifndef UNIT
  441. data01 = *(a1 + 0);
  442. #endif
  443. data02 = *(a1 + 1);
  444. data03 = *(a1 + 2);
  445. data04 = *(a1 + 3);
  446. data05 = *(a1 + 4);
  447. data06 = *(a1 + 5);
  448. data07 = *(a1 + 6);
  449. data08 = *(a1 + 7);
  450. #ifndef UNIT
  451. data10 = *(a2 + 1);
  452. #endif
  453. data11 = *(a2 + 2);
  454. data12 = *(a2 + 3);
  455. data13 = *(a2 + 4);
  456. data14 = *(a2 + 5);
  457. data15 = *(a2 + 6);
  458. data16 = *(a2 + 7);
  459. *(b + 0) = INV(data01);
  460. *(b + 1) = data02;
  461. *(b + 2) = data03;
  462. *(b + 3) = data04;
  463. *(b + 4) = data05;
  464. *(b + 5) = data06;
  465. *(b + 6) = data07;
  466. *(b + 7) = data08;
  467. *(b + 9) = INV(data10);
  468. *(b + 10) = data11;
  469. *(b + 11) = data12;
  470. *(b + 12) = data13;
  471. *(b + 13) = data14;
  472. *(b + 14) = data15;
  473. *(b + 15) = data16;
  474. }
  475. if (ii < jj) {
  476. data01 = *(a1 + 0);
  477. data02 = *(a1 + 1);
  478. data03 = *(a1 + 2);
  479. data04 = *(a1 + 3);
  480. data05 = *(a1 + 4);
  481. data06 = *(a1 + 5);
  482. data07 = *(a1 + 6);
  483. data08 = *(a1 + 7);
  484. data09 = *(a2 + 0);
  485. data10 = *(a2 + 1);
  486. data11 = *(a2 + 2);
  487. data12 = *(a2 + 3);
  488. data13 = *(a2 + 4);
  489. data14 = *(a2 + 5);
  490. data15 = *(a2 + 6);
  491. data16 = *(a2 + 7);
  492. *(b + 0) = data01;
  493. *(b + 1) = data02;
  494. *(b + 2) = data03;
  495. *(b + 3) = data04;
  496. *(b + 4) = data05;
  497. *(b + 5) = data06;
  498. *(b + 6) = data07;
  499. *(b + 7) = data08;
  500. *(b + 8) = data09;
  501. *(b + 9) = data10;
  502. *(b + 10) = data11;
  503. *(b + 11) = data12;
  504. *(b + 12) = data13;
  505. *(b + 13) = data14;
  506. *(b + 14) = data15;
  507. *(b + 15) = data16;
  508. }
  509. a1 += 2 * lda;
  510. a2 += 2 * lda;
  511. b += 16;
  512. ii += 2;
  513. }
  514. if (m & 1) {
  515. if (ii == jj) {
  516. #ifndef UNIT
  517. data01 = *(a1 + 0);
  518. #endif
  519. data02 = *(a1 + 1);
  520. data03 = *(a1 + 2);
  521. data04 = *(a1 + 3);
  522. data05 = *(a1 + 4);
  523. data06 = *(a1 + 5);
  524. data07 = *(a1 + 6);
  525. data08 = *(a1 + 7);
  526. *(b + 0) = INV(data01);
  527. *(b + 1) = data02;
  528. *(b + 2) = data03;
  529. *(b + 3) = data04;
  530. *(b + 4) = data05;
  531. *(b + 5) = data06;
  532. *(b + 6) = data07;
  533. *(b + 7) = data08;
  534. }
  535. if (ii < jj) {
  536. data01 = *(a1 + 0);
  537. data02 = *(a1 + 1);
  538. data03 = *(a1 + 2);
  539. data04 = *(a1 + 3);
  540. data05 = *(a1 + 4);
  541. data06 = *(a1 + 5);
  542. data07 = *(a1 + 6);
  543. data08 = *(a1 + 7);
  544. *(b + 0) = data01;
  545. *(b + 1) = data02;
  546. *(b + 2) = data03;
  547. *(b + 3) = data04;
  548. *(b + 4) = data05;
  549. *(b + 5) = data06;
  550. *(b + 6) = data07;
  551. *(b + 7) = data08;
  552. }
  553. b += 8;
  554. }
  555. a += 8;
  556. jj += 8;
  557. j --;
  558. }
  559. if (n & 4) {
  560. a1 = a + 0 * lda;
  561. a2 = a + 1 * lda;
  562. a3 = a + 2 * lda;
  563. a4 = a + 3 * lda;
  564. ii = 0;
  565. i = (m >> 2);
  566. while (i > 0) {
  567. if (ii == jj) {
  568. #ifndef UNIT
  569. data01 = *(a1 + 0);
  570. #endif
  571. data02 = *(a1 + 1);
  572. data03 = *(a1 + 2);
  573. data04 = *(a1 + 3);
  574. #ifndef UNIT
  575. data10 = *(a2 + 1);
  576. #endif
  577. data11 = *(a2 + 2);
  578. data12 = *(a2 + 3);
  579. #ifndef UNIT
  580. data19 = *(a3 + 2);
  581. #endif
  582. data20 = *(a3 + 3);
  583. #ifndef UNIT
  584. data28 = *(a4 + 3);
  585. #endif
  586. *(b + 0) = INV(data01);
  587. *(b + 1) = data02;
  588. *(b + 2) = data03;
  589. *(b + 3) = data04;
  590. *(b + 5) = INV(data10);
  591. *(b + 6) = data11;
  592. *(b + 7) = data12;
  593. *(b + 10) = INV(data19);
  594. *(b + 11) = data20;
  595. *(b + 15) = INV(data28);
  596. }
  597. if (ii < jj) {
  598. data01 = *(a1 + 0);
  599. data02 = *(a1 + 1);
  600. data03 = *(a1 + 2);
  601. data04 = *(a1 + 3);
  602. data09 = *(a2 + 0);
  603. data10 = *(a2 + 1);
  604. data11 = *(a2 + 2);
  605. data12 = *(a2 + 3);
  606. data17 = *(a3 + 0);
  607. data18 = *(a3 + 1);
  608. data19 = *(a3 + 2);
  609. data20 = *(a3 + 3);
  610. data25 = *(a4 + 0);
  611. data26 = *(a4 + 1);
  612. data27 = *(a4 + 2);
  613. data28 = *(a4 + 3);
  614. *(b + 0) = data01;
  615. *(b + 1) = data02;
  616. *(b + 2) = data03;
  617. *(b + 3) = data04;
  618. *(b + 4) = data09;
  619. *(b + 5) = data10;
  620. *(b + 6) = data11;
  621. *(b + 7) = data12;
  622. *(b + 8) = data17;
  623. *(b + 9) = data18;
  624. *(b + 10) = data19;
  625. *(b + 11) = data20;
  626. *(b + 12) = data25;
  627. *(b + 13) = data26;
  628. *(b + 14) = data27;
  629. *(b + 15) = data28;
  630. }
  631. a1 += 4 * lda;
  632. a2 += 4 * lda;
  633. a3 += 4 * lda;
  634. a4 += 4 * lda;
  635. b += 16;
  636. i --;
  637. ii += 4;
  638. }
  639. if (m & 2) {
  640. if (ii == jj) {
  641. #ifndef UNIT
  642. data01 = *(a1 + 0);
  643. #endif
  644. data02 = *(a1 + 1);
  645. data03 = *(a1 + 2);
  646. data04 = *(a1 + 3);
  647. #ifndef UNIT
  648. data10 = *(a2 + 1);
  649. #endif
  650. data11 = *(a2 + 2);
  651. data12 = *(a2 + 3);
  652. *(b + 0) = INV(data01);
  653. *(b + 1) = data02;
  654. *(b + 2) = data03;
  655. *(b + 3) = data04;
  656. *(b + 6) = INV(data10);
  657. *(b + 7) = data11;
  658. *(b + 8) = data12;
  659. }
  660. if (ii < jj) {
  661. data01 = *(a1 + 0);
  662. data02 = *(a1 + 1);
  663. data03 = *(a1 + 2);
  664. data04 = *(a1 + 3);
  665. data09 = *(a2 + 0);
  666. data10 = *(a2 + 1);
  667. data11 = *(a2 + 2);
  668. data12 = *(a2 + 3);
  669. *(b + 0) = data01;
  670. *(b + 1) = data02;
  671. *(b + 2) = data03;
  672. *(b + 3) = data04;
  673. *(b + 4) = data09;
  674. *(b + 5) = data10;
  675. *(b + 6) = data11;
  676. *(b + 7) = data12;
  677. }
  678. a1 += 2 * lda;
  679. a2 += 2 * lda;
  680. b += 8;
  681. ii += 2;
  682. }
  683. if (m & 1) {
  684. if (ii == jj) {
  685. #ifndef UNIT
  686. data01 = *(a1 + 0);
  687. #endif
  688. data02 = *(a1 + 1);
  689. data03 = *(a1 + 2);
  690. data04 = *(a1 + 3);
  691. *(b + 0) = INV(data01);
  692. *(b + 1) = data02;
  693. *(b + 2) = data03;
  694. *(b + 3) = data04;
  695. // *(b + 4) = data05;
  696. }
  697. if (ii < jj) {
  698. data01 = *(a1 + 0);
  699. data02 = *(a1 + 1);
  700. data03 = *(a1 + 2);
  701. data04 = *(a1 + 3);
  702. *(b + 0) = data01;
  703. *(b + 1) = data02;
  704. *(b + 2) = data03;
  705. *(b + 3) = data04;
  706. }
  707. b += 4;
  708. }
  709. a += 4;
  710. jj += 4;
  711. }
  712. if (n & 2) {
  713. a1 = a + 0 * lda;
  714. a2 = a + 1 * lda;
  715. ii = 0;
  716. i = (m >> 1);
  717. while (i > 0) {
  718. if (ii == jj) {
  719. #ifndef UNIT
  720. data01 = *(a1 + 0);
  721. #endif
  722. data02 = *(a1 + 1);
  723. #ifndef UNIT
  724. data10 = *(a2 + 1);
  725. #endif
  726. *(b + 0) = INV(data01);
  727. *(b + 1) = data02;
  728. *(b + 3) = INV(data10);
  729. }
  730. if (ii < jj) {
  731. data01 = *(a1 + 0);
  732. data02 = *(a1 + 1);
  733. data09 = *(a2 + 0);
  734. data10 = *(a2 + 1);
  735. *(b + 0) = data01;
  736. *(b + 1) = data02;
  737. *(b + 2) = data09;
  738. *(b + 3) = data10;
  739. }
  740. a1 += 2 * lda;
  741. a2 += 2 * lda;
  742. b += 4;
  743. i --;
  744. ii += 2;
  745. }
  746. if (m & 1) {
  747. if (ii == jj) {
  748. #ifndef UNIT
  749. data01 = *(a1 + 0);
  750. #endif
  751. data02 = *(a1 + 1);
  752. *(b + 0) = INV(data01);
  753. *(b + 1) = data02;
  754. }
  755. if (ii < jj) {
  756. data01 = *(a1 + 0);
  757. data02 = *(a1 + 1);
  758. *(b + 0) = data01;
  759. *(b + 1) = data02;
  760. }
  761. b += 2;
  762. }
  763. a += 2;
  764. jj += 2;
  765. }
  766. if (n & 1) {
  767. a1 = a + 0 * lda;
  768. ii = 0;
  769. i = m;
  770. while (i > 0) {
  771. if (ii == jj) {
  772. #ifndef UNIT
  773. data01 = *(a1 + 0);
  774. #endif
  775. *(b + 0) = INV(data01);
  776. }
  777. if (ii < jj) {
  778. data01 = *(a1 + 0);
  779. *(b + 0) = data01;
  780. }
  781. a1 += lda;
  782. b += 1;
  783. i --;
  784. ii += 1;
  785. }
  786. }
  787. return 0;
  788. }