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_utcopy_8.c 16 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803
  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. i = (m >> 3);
  68. ii = 0;
  69. while (i > 0) {
  70. if (ii == jj) {
  71. #ifndef UNIT
  72. data01 = *(a1 + 0);
  73. #endif
  74. data09 = *(a2 + 0);
  75. #ifndef UNIT
  76. data10 = *(a2 + 1);
  77. #endif
  78. data17 = *(a3 + 0);
  79. data18 = *(a3 + 1);
  80. #ifndef UNIT
  81. data19 = *(a3 + 2);
  82. #endif
  83. data25 = *(a4 + 0);
  84. data26 = *(a4 + 1);
  85. data27 = *(a4 + 2);
  86. #ifndef UNIT
  87. data28 = *(a4 + 3);
  88. #endif
  89. data33 = *(a5 + 0);
  90. data34 = *(a5 + 1);
  91. data35 = *(a5 + 2);
  92. data36 = *(a5 + 3);
  93. #ifndef UNIT
  94. data37 = *(a5 + 4);
  95. #endif
  96. data41 = *(a6 + 0);
  97. data42 = *(a6 + 1);
  98. data43 = *(a6 + 2);
  99. data44 = *(a6 + 3);
  100. data45 = *(a6 + 4);
  101. #ifndef UNIT
  102. data46 = *(a6 + 5);
  103. #endif
  104. data49 = *(a7 + 0);
  105. data50 = *(a7 + 1);
  106. data51 = *(a7 + 2);
  107. data52 = *(a7 + 3);
  108. data53 = *(a7 + 4);
  109. data54 = *(a7 + 5);
  110. #ifndef UNIT
  111. data55 = *(a7 + 6);
  112. #endif
  113. data57 = *(a8 + 0);
  114. data58 = *(a8 + 1);
  115. data59 = *(a8 + 2);
  116. data60 = *(a8 + 3);
  117. data61 = *(a8 + 4);
  118. data62 = *(a8 + 5);
  119. data63 = *(a8 + 6);
  120. #ifndef UNIT
  121. data64 = *(a8 + 7);
  122. #endif
  123. *(b + 0) = INV(data01);
  124. *(b + 8) = data09;
  125. *(b + 9) = INV(data10);
  126. *(b + 16) = data17;
  127. *(b + 17) = data18;
  128. *(b + 18) = INV(data19);
  129. *(b + 24) = data25;
  130. *(b + 25) = data26;
  131. *(b + 26) = data27;
  132. *(b + 27) = INV(data28);
  133. *(b + 32) = data33;
  134. *(b + 33) = data34;
  135. *(b + 34) = data35;
  136. *(b + 35) = data36;
  137. *(b + 36) = INV(data37);
  138. *(b + 40) = data41;
  139. *(b + 41) = data42;
  140. *(b + 42) = data43;
  141. *(b + 43) = data44;
  142. *(b + 44) = data45;
  143. *(b + 45) = INV(data46);
  144. *(b + 48) = data49;
  145. *(b + 49) = data50;
  146. *(b + 50) = data51;
  147. *(b + 51) = data52;
  148. *(b + 52) = data53;
  149. *(b + 53) = data54;
  150. *(b + 54) = INV(data55);
  151. *(b + 56) = data57;
  152. *(b + 57) = data58;
  153. *(b + 58) = data59;
  154. *(b + 59) = data60;
  155. *(b + 60) = data61;
  156. *(b + 61) = data62;
  157. *(b + 62) = data63;
  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. data09 = *(a2 + 0);
  308. #ifndef UNIT
  309. data10 = *(a2 + 1);
  310. #endif
  311. data17 = *(a3 + 0);
  312. data18 = *(a3 + 1);
  313. #ifndef UNIT
  314. data19 = *(a3 + 2);
  315. #endif
  316. data25 = *(a4 + 0);
  317. data26 = *(a4 + 1);
  318. data27 = *(a4 + 2);
  319. #ifndef UNIT
  320. data28 = *(a4 + 3);
  321. #endif
  322. *(b + 0) = INV(data01);
  323. *(b + 8) = data09;
  324. *(b + 9) = INV(data10);
  325. *(b + 16) = data17;
  326. *(b + 17) = data18;
  327. *(b + 18) = INV(data19);
  328. *(b + 24) = data25;
  329. *(b + 25) = data26;
  330. *(b + 26) = data27;
  331. *(b + 27) = INV(data28);
  332. }
  333. if (ii > jj) {
  334. data01 = *(a1 + 0);
  335. data02 = *(a1 + 1);
  336. data03 = *(a1 + 2);
  337. data04 = *(a1 + 3);
  338. data05 = *(a1 + 4);
  339. data06 = *(a1 + 5);
  340. data07 = *(a1 + 6);
  341. data08 = *(a1 + 7);
  342. data09 = *(a2 + 0);
  343. data10 = *(a2 + 1);
  344. data11 = *(a2 + 2);
  345. data12 = *(a2 + 3);
  346. data13 = *(a2 + 4);
  347. data14 = *(a2 + 5);
  348. data15 = *(a2 + 6);
  349. data16 = *(a2 + 7);
  350. data17 = *(a3 + 0);
  351. data18 = *(a3 + 1);
  352. data19 = *(a3 + 2);
  353. data20 = *(a3 + 3);
  354. data21 = *(a3 + 4);
  355. data22 = *(a3 + 5);
  356. data23 = *(a3 + 6);
  357. data24 = *(a3 + 7);
  358. data25 = *(a4 + 0);
  359. data26 = *(a4 + 1);
  360. data27 = *(a4 + 2);
  361. data28 = *(a4 + 3);
  362. data29 = *(a4 + 4);
  363. data30 = *(a4 + 5);
  364. data31 = *(a4 + 6);
  365. data32 = *(a4 + 7);
  366. *(b + 0) = data01;
  367. *(b + 1) = data02;
  368. *(b + 2) = data03;
  369. *(b + 3) = data04;
  370. *(b + 4) = data05;
  371. *(b + 5) = data06;
  372. *(b + 6) = data07;
  373. *(b + 7) = data08;
  374. *(b + 8) = data09;
  375. *(b + 9) = data10;
  376. *(b + 10) = data11;
  377. *(b + 11) = data12;
  378. *(b + 12) = data13;
  379. *(b + 13) = data14;
  380. *(b + 14) = data15;
  381. *(b + 15) = data16;
  382. *(b + 16) = data17;
  383. *(b + 17) = data18;
  384. *(b + 18) = data19;
  385. *(b + 19) = data20;
  386. *(b + 20) = data21;
  387. *(b + 21) = data22;
  388. *(b + 22) = data23;
  389. *(b + 23) = data24;
  390. *(b + 24) = data25;
  391. *(b + 25) = data26;
  392. *(b + 26) = data27;
  393. *(b + 27) = data28;
  394. *(b + 28) = data29;
  395. *(b + 29) = data30;
  396. *(b + 30) = data31;
  397. *(b + 31) = data32;
  398. }
  399. a1 += 4 * lda;
  400. a2 += 4 * lda;
  401. /* a3 += 4 * lda;
  402. a4 += 4 * lda; */
  403. b += 32;
  404. ii += 4;
  405. }
  406. if (m & 2) {
  407. if (ii == jj) {
  408. #ifndef UNIT
  409. data01 = *(a1 + 0);
  410. #endif
  411. data09 = *(a2 + 0);
  412. #ifndef UNIT
  413. data10 = *(a2 + 1);
  414. #endif
  415. *(b + 0) = INV(data01);
  416. *(b + 8) = data09;
  417. *(b + 9) = INV(data10);
  418. }
  419. if (ii > jj) {
  420. data01 = *(a1 + 0);
  421. data02 = *(a1 + 1);
  422. data03 = *(a1 + 2);
  423. data04 = *(a1 + 3);
  424. data05 = *(a1 + 4);
  425. data06 = *(a1 + 5);
  426. data07 = *(a1 + 6);
  427. data08 = *(a1 + 7);
  428. data09 = *(a2 + 0);
  429. data10 = *(a2 + 1);
  430. data11 = *(a2 + 2);
  431. data12 = *(a2 + 3);
  432. data13 = *(a2 + 4);
  433. data14 = *(a2 + 5);
  434. data15 = *(a2 + 6);
  435. data16 = *(a2 + 7);
  436. *(b + 0) = data01;
  437. *(b + 1) = data02;
  438. *(b + 2) = data03;
  439. *(b + 3) = data04;
  440. *(b + 4) = data05;
  441. *(b + 5) = data06;
  442. *(b + 6) = data07;
  443. *(b + 7) = data08;
  444. *(b + 8) = data09;
  445. *(b + 9) = data10;
  446. *(b + 10) = data11;
  447. *(b + 11) = data12;
  448. *(b + 12) = data13;
  449. *(b + 13) = data14;
  450. *(b + 14) = data15;
  451. *(b + 15) = data16;
  452. }
  453. a1 += 2 * lda;
  454. // a2 += 2 * lda;
  455. b += 16;
  456. ii += 2;
  457. }
  458. if (m & 1) {
  459. if (ii == jj) {
  460. #ifndef UNIT
  461. data01 = *(a1 + 0);
  462. #endif
  463. *(b + 0) = INV(data01);
  464. }
  465. if (ii > jj) {
  466. data01 = *(a1 + 0);
  467. data02 = *(a1 + 1);
  468. data03 = *(a1 + 2);
  469. data04 = *(a1 + 3);
  470. data05 = *(a1 + 4);
  471. data06 = *(a1 + 5);
  472. data07 = *(a1 + 6);
  473. data08 = *(a1 + 7);
  474. *(b + 0) = data01;
  475. *(b + 1) = data02;
  476. *(b + 2) = data03;
  477. *(b + 3) = data04;
  478. *(b + 4) = data05;
  479. *(b + 5) = data06;
  480. *(b + 6) = data07;
  481. *(b + 7) = data08;
  482. }
  483. b += 8;
  484. }
  485. a += 8;
  486. jj += 8;
  487. j --;
  488. }
  489. if (n & 4) {
  490. a1 = a + 0 * lda;
  491. a2 = a + 1 * lda;
  492. a3 = a + 2 * lda;
  493. a4 = a + 3 * lda;
  494. i = (m >> 2);
  495. ii = 0;
  496. while (i > 0) {
  497. if (ii == jj) {
  498. #ifndef UNIT
  499. data01 = *(a1 + 0);
  500. #endif
  501. data09 = *(a2 + 0);
  502. #ifndef UNIT
  503. data10 = *(a2 + 1);
  504. #endif
  505. data17 = *(a3 + 0);
  506. data18 = *(a3 + 1);
  507. #ifndef UNIT
  508. data19 = *(a3 + 2);
  509. #endif
  510. data25 = *(a4 + 0);
  511. data26 = *(a4 + 1);
  512. data27 = *(a4 + 2);
  513. #ifndef UNIT
  514. data28 = *(a4 + 3);
  515. #endif
  516. *(b + 0) = INV(data01);
  517. *(b + 4) = data09;
  518. *(b + 5) = INV(data10);
  519. *(b + 8) = data17;
  520. *(b + 9) = data18;
  521. *(b + 10) = INV(data19);
  522. *(b + 12) = data25;
  523. *(b + 13) = data26;
  524. *(b + 14) = data27;
  525. *(b + 15) = INV(data28);
  526. }
  527. if (ii > jj) {
  528. data01 = *(a1 + 0);
  529. data02 = *(a1 + 1);
  530. data03 = *(a1 + 2);
  531. data04 = *(a1 + 3);
  532. data09 = *(a2 + 0);
  533. data10 = *(a2 + 1);
  534. data11 = *(a2 + 2);
  535. data12 = *(a2 + 3);
  536. data17 = *(a3 + 0);
  537. data18 = *(a3 + 1);
  538. data19 = *(a3 + 2);
  539. data20 = *(a3 + 3);
  540. data25 = *(a4 + 0);
  541. data26 = *(a4 + 1);
  542. data27 = *(a4 + 2);
  543. data28 = *(a4 + 3);
  544. *(b + 0) = data01;
  545. *(b + 1) = data02;
  546. *(b + 2) = data03;
  547. *(b + 3) = data04;
  548. *(b + 4) = data09;
  549. *(b + 5) = data10;
  550. *(b + 6) = data11;
  551. *(b + 7) = data12;
  552. *(b + 8) = data17;
  553. *(b + 9) = data18;
  554. *(b + 10) = data19;
  555. *(b + 11) = data20;
  556. *(b + 12) = data25;
  557. *(b + 13) = data26;
  558. *(b + 14) = data27;
  559. *(b + 15) = data28;
  560. }
  561. a1 += 4 * lda;
  562. a2 += 4 * lda;
  563. a3 += 4 * lda;
  564. a4 += 4 * lda;
  565. b += 16;
  566. i --;
  567. ii += 4;
  568. }
  569. if (m & 2) {
  570. if (ii == jj) {
  571. /* #ifndef UNIT
  572. data01 = *(a1 + 0);
  573. #endif
  574. data09 = *(a2 + 0);
  575. #ifndef UNIT
  576. data10 = *(a2 + 1);
  577. #endif */
  578. }
  579. if (ii > jj) {
  580. data01 = *(a1 + 0);
  581. data02 = *(a1 + 1);
  582. data03 = *(a1 + 2);
  583. data04 = *(a1 + 3);
  584. data09 = *(a2 + 0);
  585. data10 = *(a2 + 1);
  586. data11 = *(a2 + 2);
  587. data12 = *(a2 + 3);
  588. *(b + 0) = data01;
  589. *(b + 1) = data02;
  590. *(b + 2) = data03;
  591. *(b + 3) = data04;
  592. *(b + 4) = data09;
  593. *(b + 5) = data10;
  594. *(b + 6) = data11;
  595. *(b + 7) = data12;
  596. }
  597. a1 += 2 * lda;
  598. // a2 += 2 * lda;
  599. b += 8;
  600. ii += 2;
  601. }
  602. if (m & 1) {
  603. if (ii == jj) {
  604. #ifndef UNIT
  605. data01 = *(a1 + 0);
  606. #endif
  607. *(b + 0) = INV(data01);
  608. }
  609. if (ii > jj) {
  610. data01 = *(a1 + 0);
  611. data02 = *(a1 + 1);
  612. data03 = *(a1 + 2);
  613. data04 = *(a1 + 3);
  614. *(b + 0) = data01;
  615. *(b + 1) = data02;
  616. *(b + 2) = data03;
  617. *(b + 3) = data04;
  618. }
  619. b += 4;
  620. }
  621. a += 4;
  622. jj += 4;
  623. }
  624. if (n & 2) {
  625. a1 = a + 0 * lda;
  626. a2 = a + 1 * lda;
  627. i = (m >> 1);
  628. ii = 0;
  629. while (i > 0) {
  630. if (ii == jj) {
  631. #ifndef UNIT
  632. data01 = *(a1 + 0);
  633. #endif
  634. data09 = *(a2 + 0);
  635. #ifndef UNIT
  636. data10 = *(a2 + 1);
  637. #endif
  638. *(b + 0) = INV(data01);
  639. *(b + 2) = data09;
  640. *(b + 3) = INV(data10);
  641. }
  642. if (ii > jj) {
  643. data01 = *(a1 + 0);
  644. data02 = *(a1 + 1);
  645. data09 = *(a2 + 0);
  646. data10 = *(a2 + 1);
  647. *(b + 0) = data01;
  648. *(b + 1) = data02;
  649. *(b + 2) = data09;
  650. *(b + 3) = data10;
  651. }
  652. a1 += 2 * lda;
  653. a2 += 2 * lda;
  654. b += 4;
  655. i --;
  656. ii += 2;
  657. }
  658. if (m & 1) {
  659. if (ii == jj) {
  660. #ifndef UNIT
  661. data01 = *(a1 + 0);
  662. #endif
  663. *(b + 0) = INV(data01);
  664. }
  665. if (ii > jj) {
  666. data01 = *(a1 + 0);
  667. data02 = *(a1 + 1);
  668. *(b + 0) = data01;
  669. *(b + 1) = data02;
  670. }
  671. b += 2;
  672. }
  673. a += 2;
  674. jj += 2;
  675. }
  676. if (n & 1) {
  677. a1 = a + 0 * lda;
  678. i = m;
  679. ii = 0;
  680. while (i > 0) {
  681. if (ii == jj) {
  682. #ifndef UNIT
  683. data01 = *(a1 + 0);
  684. #endif
  685. *(b + 0) = INV(data01);
  686. }
  687. if (ii > jj) {
  688. data01 = *(a1 + 0);
  689. *(b + 0) = data01;
  690. }
  691. a1 += lda;
  692. b += 1;
  693. i --;
  694. ii += 1;
  695. }
  696. }
  697. return 0;
  698. }