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_lncopy_8.c 16 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841
  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 + 8) = data02;
  125. *(b + 9) = INV(data10);
  126. *(b + 16) = data03;
  127. *(b + 17) = data11;
  128. *(b + 18) = INV(data19);
  129. *(b + 24) = data04;
  130. *(b + 25) = data12;
  131. *(b + 26) = data20;
  132. *(b + 27) = INV(data28);
  133. *(b + 32) = data05;
  134. *(b + 33) = data13;
  135. *(b + 34) = data21;
  136. *(b + 35) = data29;
  137. *(b + 36) = INV(data37);
  138. *(b + 40) = data06;
  139. *(b + 41) = data14;
  140. *(b + 42) = data22;
  141. *(b + 43) = data30;
  142. *(b + 44) = data38;
  143. *(b + 45) = INV(data46);
  144. *(b + 48) = data07;
  145. *(b + 49) = data15;
  146. *(b + 50) = data23;
  147. *(b + 51) = data31;
  148. *(b + 52) = data39;
  149. *(b + 53) = data47;
  150. *(b + 54) = INV(data55);
  151. *(b + 56) = data08;
  152. *(b + 57) = data16;
  153. *(b + 58) = data24;
  154. *(b + 59) = data32;
  155. *(b + 60) = data40;
  156. *(b + 61) = data48;
  157. *(b + 62) = 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) = data09;
  227. *(b + 2) = data17;
  228. *(b + 3) = data25;
  229. *(b + 4) = data33;
  230. *(b + 5) = data41;
  231. *(b + 6) = data49;
  232. *(b + 7) = data57;
  233. *(b + 8) = data02;
  234. *(b + 9) = data10;
  235. *(b + 10) = data18;
  236. *(b + 11) = data26;
  237. *(b + 12) = data34;
  238. *(b + 13) = data42;
  239. *(b + 14) = data50;
  240. *(b + 15) = data58;
  241. *(b + 16) = data03;
  242. *(b + 17) = data11;
  243. *(b + 18) = data19;
  244. *(b + 19) = data27;
  245. *(b + 20) = data35;
  246. *(b + 21) = data43;
  247. *(b + 22) = data51;
  248. *(b + 23) = data59;
  249. *(b + 24) = data04;
  250. *(b + 25) = data12;
  251. *(b + 26) = data20;
  252. *(b + 27) = data28;
  253. *(b + 28) = data36;
  254. *(b + 29) = data44;
  255. *(b + 30) = data52;
  256. *(b + 31) = data60;
  257. *(b + 32) = data05;
  258. *(b + 33) = data13;
  259. *(b + 34) = data21;
  260. *(b + 35) = data29;
  261. *(b + 36) = data37;
  262. *(b + 37) = data45;
  263. *(b + 38) = data53;
  264. *(b + 39) = data61;
  265. *(b + 40) = data06;
  266. *(b + 41) = data14;
  267. *(b + 42) = data22;
  268. *(b + 43) = data30;
  269. *(b + 44) = data38;
  270. *(b + 45) = data46;
  271. *(b + 46) = data54;
  272. *(b + 47) = data62;
  273. *(b + 48) = data07;
  274. *(b + 49) = data15;
  275. *(b + 50) = data23;
  276. *(b + 51) = data31;
  277. *(b + 52) = data39;
  278. *(b + 53) = data47;
  279. *(b + 54) = data55;
  280. *(b + 55) = data63;
  281. *(b + 56) = data08;
  282. *(b + 57) = data16;
  283. *(b + 58) = data24;
  284. *(b + 59) = data32;
  285. *(b + 60) = data40;
  286. *(b + 61) = data48;
  287. *(b + 62) = data56;
  288. *(b + 63) = data64;
  289. }
  290. a1 += 8;
  291. a2 += 8;
  292. a3 += 8;
  293. a4 += 8;
  294. a5 += 8;
  295. a6 += 8;
  296. a7 += 8;
  297. a8 += 8;
  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. #ifndef UNIT
  311. data10 = *(a2 + 1);
  312. #endif
  313. data11 = *(a2 + 2);
  314. data12 = *(a2 + 3);
  315. #ifndef UNIT
  316. data19 = *(a3 + 2);
  317. #endif
  318. data20 = *(a3 + 3);
  319. #ifndef UNIT
  320. data28 = *(a4 + 3);
  321. #endif
  322. *(b + 0) = INV(data01);
  323. *(b + 8) = data02;
  324. *(b + 9) = INV(data10);
  325. *(b + 16) = data03;
  326. *(b + 17) = data11;
  327. *(b + 18) = INV(data19);
  328. *(b + 24) = data04;
  329. *(b + 25) = data12;
  330. *(b + 26) = data20;
  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. data09 = *(a2 + 0);
  339. data10 = *(a2 + 1);
  340. data11 = *(a2 + 2);
  341. data12 = *(a2 + 3);
  342. data17 = *(a3 + 0);
  343. data18 = *(a3 + 1);
  344. data19 = *(a3 + 2);
  345. data20 = *(a3 + 3);
  346. data25 = *(a4 + 0);
  347. data26 = *(a4 + 1);
  348. data27 = *(a4 + 2);
  349. data28 = *(a4 + 3);
  350. data33 = *(a5 + 0);
  351. data34 = *(a5 + 1);
  352. data35 = *(a5 + 2);
  353. data36 = *(a5 + 3);
  354. data41 = *(a6 + 0);
  355. data42 = *(a6 + 1);
  356. data43 = *(a6 + 2);
  357. data44 = *(a6 + 3);
  358. data49 = *(a7 + 0);
  359. data50 = *(a7 + 1);
  360. data51 = *(a7 + 2);
  361. data52 = *(a7 + 3);
  362. data57 = *(a8 + 0);
  363. data58 = *(a8 + 1);
  364. data59 = *(a8 + 2);
  365. data60 = *(a8 + 3);
  366. *(b + 0) = data01;
  367. *(b + 1) = data09;
  368. *(b + 2) = data17;
  369. *(b + 3) = data25;
  370. *(b + 4) = data33;
  371. *(b + 5) = data41;
  372. *(b + 6) = data49;
  373. *(b + 7) = data57;
  374. *(b + 8) = data02;
  375. *(b + 9) = data10;
  376. *(b + 10) = data18;
  377. *(b + 11) = data26;
  378. *(b + 12) = data34;
  379. *(b + 13) = data42;
  380. *(b + 14) = data50;
  381. *(b + 15) = data58;
  382. *(b + 16) = data03;
  383. *(b + 17) = data11;
  384. *(b + 18) = data19;
  385. *(b + 19) = data27;
  386. *(b + 20) = data35;
  387. *(b + 21) = data43;
  388. *(b + 22) = data51;
  389. *(b + 23) = data59;
  390. *(b + 24) = data04;
  391. *(b + 25) = data12;
  392. *(b + 26) = data20;
  393. *(b + 27) = data28;
  394. *(b + 28) = data36;
  395. *(b + 29) = data44;
  396. *(b + 30) = data52;
  397. *(b + 31) = data60;
  398. }
  399. a1 += 4;
  400. a2 += 4;
  401. a3 += 4;
  402. a4 += 4;
  403. a5 += 4;
  404. a6 += 4;
  405. a7 += 4;
  406. a8 += 4;
  407. b += 32;
  408. ii += 4;
  409. }
  410. if (m & 2) {
  411. if (ii == jj) {
  412. #ifndef UNIT
  413. data01 = *(a1 + 0);
  414. #endif
  415. data02 = *(a1 + 1);
  416. #ifndef UNIT
  417. data10 = *(a2 + 1);
  418. #endif
  419. *(b + 0) = INV(data01);
  420. *(b + 8) = data02;
  421. *(b + 9) = INV(data10);
  422. }
  423. if (ii > jj) {
  424. data01 = *(a1 + 0);
  425. data02 = *(a1 + 1);
  426. data09 = *(a2 + 0);
  427. data10 = *(a2 + 1);
  428. data17 = *(a3 + 0);
  429. data18 = *(a3 + 1);
  430. data25 = *(a4 + 0);
  431. data26 = *(a4 + 1);
  432. data33 = *(a5 + 0);
  433. data34 = *(a5 + 1);
  434. data41 = *(a6 + 0);
  435. data42 = *(a6 + 1);
  436. data49 = *(a7 + 0);
  437. data50 = *(a7 + 1);
  438. data57 = *(a8 + 0);
  439. data58 = *(a8 + 1);
  440. *(b + 0) = data01;
  441. *(b + 1) = data09;
  442. *(b + 2) = data17;
  443. *(b + 3) = data25;
  444. *(b + 4) = data33;
  445. *(b + 5) = data41;
  446. *(b + 6) = data49;
  447. *(b + 7) = data57;
  448. *(b + 8) = data02;
  449. *(b + 9) = data10;
  450. *(b + 10) = data18;
  451. *(b + 11) = data26;
  452. *(b + 12) = data34;
  453. *(b + 13) = data42;
  454. *(b + 14) = data50;
  455. *(b + 15) = data58;
  456. }
  457. a1 += 2;
  458. a2 += 2;
  459. a3 += 2;
  460. a4 += 2;
  461. a5 += 2;
  462. a6 += 2;
  463. a7 += 2;
  464. a8 += 2;
  465. b += 16;
  466. ii += 2;
  467. }
  468. if (m & 1) {
  469. if (ii == jj) {
  470. #ifndef UNIT
  471. data01 = *(a1 + 0);
  472. #endif
  473. *(b + 0) = INV(data01);
  474. }
  475. if (ii > jj) {
  476. data01 = *(a1 + 0);
  477. data09 = *(a2 + 0);
  478. data17 = *(a3 + 0);
  479. data25 = *(a4 + 0);
  480. data33 = *(a5 + 0);
  481. data41 = *(a6 + 0);
  482. data49 = *(a7 + 0);
  483. data57 = *(a8 + 0);
  484. *(b + 0) = data01;
  485. *(b + 1) = data09;
  486. *(b + 2) = data17;
  487. *(b + 3) = data25;
  488. *(b + 4) = data33;
  489. *(b + 5) = data41;
  490. *(b + 6) = data49;
  491. *(b + 7) = data57;
  492. }
  493. b += 8;
  494. }
  495. a += 8 * lda;
  496. jj += 8;
  497. j --;
  498. }
  499. if (n & 4) {
  500. a1 = a + 0 * lda;
  501. a2 = a + 1 * lda;
  502. a3 = a + 2 * lda;
  503. a4 = a + 3 * lda;
  504. ii = 0;
  505. i = (m >> 2);
  506. while (i > 0) {
  507. if (ii == jj) {
  508. #ifndef UNIT
  509. data01 = *(a1 + 0);
  510. #endif
  511. data02 = *(a1 + 1);
  512. data03 = *(a1 + 2);
  513. data04 = *(a1 + 3);
  514. #ifndef UNIT
  515. data10 = *(a2 + 1);
  516. #endif
  517. data11 = *(a2 + 2);
  518. data12 = *(a2 + 3);
  519. #ifndef UNIT
  520. data19 = *(a3 + 2);
  521. #endif
  522. data20 = *(a3 + 3);
  523. #ifndef UNIT
  524. data28 = *(a4 + 3);
  525. #endif
  526. *(b + 0) = INV(data01);
  527. *(b + 4) = data02;
  528. *(b + 5) = INV(data10);
  529. *(b + 8) = data03;
  530. *(b + 9) = data11;
  531. *(b + 10) = INV(data19);
  532. *(b + 12) = data04;
  533. *(b + 13) = data12;
  534. *(b + 14) = data20;
  535. *(b + 15) = INV(data28);
  536. }
  537. if (ii > jj) {
  538. data01 = *(a1 + 0);
  539. data02 = *(a1 + 1);
  540. data03 = *(a1 + 2);
  541. data04 = *(a1 + 3);
  542. data09 = *(a2 + 0);
  543. data10 = *(a2 + 1);
  544. data11 = *(a2 + 2);
  545. data12 = *(a2 + 3);
  546. data17 = *(a3 + 0);
  547. data18 = *(a3 + 1);
  548. data19 = *(a3 + 2);
  549. data20 = *(a3 + 3);
  550. data25 = *(a4 + 0);
  551. data26 = *(a4 + 1);
  552. data27 = *(a4 + 2);
  553. data28 = *(a4 + 3);
  554. *(b + 0) = data01;
  555. *(b + 1) = data09;
  556. *(b + 2) = data17;
  557. *(b + 3) = data25;
  558. *(b + 4) = data02;
  559. *(b + 5) = data10;
  560. *(b + 6) = data18;
  561. *(b + 7) = data26;
  562. *(b + 8) = data03;
  563. *(b + 9) = data11;
  564. *(b + 10) = data19;
  565. *(b + 11) = data27;
  566. *(b + 12) = data04;
  567. *(b + 13) = data12;
  568. *(b + 14) = data20;
  569. *(b + 15) = data28;
  570. }
  571. a1 += 4;
  572. a2 += 4;
  573. a3 += 4;
  574. a4 += 4;
  575. b += 16;
  576. i --;
  577. ii += 4;
  578. }
  579. if (m & 2) {
  580. if (ii == jj) {
  581. #ifndef UNIT
  582. data01 = *(a1 + 0);
  583. #endif
  584. data02 = *(a1 + 1);
  585. #ifndef UNIT
  586. data10 = *(a2 + 1);
  587. #endif
  588. *(b + 0) = INV(data01);
  589. *(b + 4) = data02;
  590. *(b + 5) = INV(data10);
  591. }
  592. if (ii > jj) {
  593. data01 = *(a1 + 0);
  594. data02 = *(a1 + 1);
  595. data09 = *(a2 + 0);
  596. data10 = *(a2 + 1);
  597. data17 = *(a3 + 0);
  598. data18 = *(a3 + 1);
  599. data25 = *(a4 + 0);
  600. data26 = *(a4 + 1);
  601. *(b + 0) = data01;
  602. *(b + 1) = data09;
  603. *(b + 2) = data17;
  604. *(b + 3) = data25;
  605. *(b + 4) = data02;
  606. *(b + 5) = data10;
  607. *(b + 6) = data18;
  608. *(b + 7) = data26;
  609. }
  610. a1 += 2;
  611. a2 += 2;
  612. a3 += 2;
  613. a4 += 2;
  614. b += 8;
  615. ii += 2;
  616. }
  617. if (m & 1) {
  618. if (ii == jj) {
  619. #ifndef UNIT
  620. data01 = *(a1 + 0);
  621. #endif
  622. *(b + 0) = INV(data01);
  623. }
  624. if (ii > jj) {
  625. data01 = *(a1 + 0);
  626. data09 = *(a2 + 0);
  627. data17 = *(a3 + 0);
  628. data25 = *(a4 + 0);
  629. *(b + 0) = data01;
  630. *(b + 1) = data09;
  631. *(b + 2) = data17;
  632. *(b + 3) = data25;
  633. }
  634. b += 4;
  635. }
  636. a += 4 * lda;
  637. jj += 4;
  638. }
  639. if (n & 2) {
  640. a1 = a + 0 * lda;
  641. a2 = a + 1 * lda;
  642. ii = 0;
  643. i = (m >> 1);
  644. while (i > 0) {
  645. if (ii == jj) {
  646. #ifndef UNIT
  647. data01 = *(a1 + 0);
  648. #endif
  649. data02 = *(a1 + 1);
  650. #ifndef UNIT
  651. data10 = *(a2 + 1);
  652. #endif
  653. *(b + 0) = INV(data01);
  654. *(b + 2) = data02;
  655. *(b + 3) = INV(data10);
  656. }
  657. if (ii > jj) {
  658. data01 = *(a1 + 0);
  659. data02 = *(a1 + 1);
  660. data09 = *(a2 + 0);
  661. data10 = *(a2 + 1);
  662. *(b + 0) = data01;
  663. *(b + 1) = data09;
  664. *(b + 2) = data02;
  665. *(b + 3) = data10;
  666. }
  667. a1 += 2;
  668. a2 += 2;
  669. b += 4;
  670. i --;
  671. ii += 2;
  672. }
  673. if (m & 1) {
  674. if (ii == jj) {
  675. #ifndef UNIT
  676. data01 = *(a1 + 0);
  677. #endif
  678. *(b + 0) = INV(data01);
  679. }
  680. if (ii > jj) {
  681. data01 = *(a1 + 0);
  682. data09 = *(a2 + 0);
  683. *(b + 0) = data01;
  684. *(b + 1) = data09;
  685. }
  686. b += 2;
  687. }
  688. a += 2 * lda;
  689. jj += 2;
  690. }
  691. if (n & 1) {
  692. a1 = a + 0 * lda;
  693. ii = 0;
  694. i = m;
  695. while (i > 0) {
  696. if (ii == jj) {
  697. #ifndef UNIT
  698. data01 = *(a1 + 0);
  699. #endif
  700. *(b + 0) = INV(data01);
  701. }
  702. if (ii > jj) {
  703. data01 = *(a1 + 0);
  704. *(b + 0) = data01;
  705. }
  706. a1 += 1;
  707. b += 1;
  708. i --;
  709. ii += 1;
  710. }
  711. }
  712. return 0;
  713. }