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_uncopy_8.c 18 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946
  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 "common.h"
  39. #ifndef UNIT
  40. #define INV(a) (ONE / (a))
  41. #else
  42. #define INV(a) (ONE)
  43. #endif
  44. int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT *b){
  45. BLASLONG i, ii, j, jj;
  46. FLOAT data01, data02, data03, data04, data05, data06, data07, data08;
  47. FLOAT data09, data10, data11, data12, data13, data14, data15, data16;
  48. FLOAT data17, data18, data19, data20, data21, data22, data23, data24;
  49. FLOAT data25, data26, data27, data28, data29, data30, data31, data32;
  50. FLOAT data33, data34, data35, data36, data37, data38, data39, data40;
  51. FLOAT data41, data42, data43, data44, data45, data46, data47, data48;
  52. FLOAT data49, data50, data51, data52, data53, data54, data55, data56;
  53. FLOAT data57, data58, data59, data60, data61, data62, data63, data64;
  54. FLOAT *a1, *a2, *a3, *a4, *a5, *a6, *a7, *a8;
  55. jj = offset;
  56. j = (n >> 3);
  57. while (j > 0){
  58. a1 = a + 0 * lda;
  59. a2 = a + 1 * lda;
  60. a3 = a + 2 * lda;
  61. a4 = a + 3 * lda;
  62. a5 = a + 4 * lda;
  63. a6 = a + 5 * lda;
  64. a7 = a + 6 * lda;
  65. a8 = a + 7 * lda;
  66. ii = 0;
  67. i = (m >> 3);
  68. while (i > 0) {
  69. if (ii == jj) {
  70. #ifndef UNIT
  71. data01 = *(a1 + 0);
  72. #endif
  73. data09 = *(a2 + 0);
  74. #ifndef UNIT
  75. data10 = *(a2 + 1);
  76. #endif
  77. data17 = *(a3 + 0);
  78. data18 = *(a3 + 1);
  79. #ifndef UNIT
  80. data19 = *(a3 + 2);
  81. #endif
  82. data25 = *(a4 + 0);
  83. data26 = *(a4 + 1);
  84. data27 = *(a4 + 2);
  85. #ifndef UNIT
  86. data28 = *(a4 + 3);
  87. #endif
  88. data33 = *(a5 + 0);
  89. data34 = *(a5 + 1);
  90. data35 = *(a5 + 2);
  91. data36 = *(a5 + 3);
  92. #ifndef UNIT
  93. data37 = *(a5 + 4);
  94. #endif
  95. data41 = *(a6 + 0);
  96. data42 = *(a6 + 1);
  97. data43 = *(a6 + 2);
  98. data44 = *(a6 + 3);
  99. data45 = *(a6 + 4);
  100. #ifndef UNIT
  101. data46 = *(a6 + 5);
  102. #endif
  103. data49 = *(a7 + 0);
  104. data50 = *(a7 + 1);
  105. data51 = *(a7 + 2);
  106. data52 = *(a7 + 3);
  107. data53 = *(a7 + 4);
  108. data54 = *(a7 + 5);
  109. #ifndef UNIT
  110. data55 = *(a7 + 6);
  111. #endif
  112. data57 = *(a8 + 0);
  113. data58 = *(a8 + 1);
  114. data59 = *(a8 + 2);
  115. data60 = *(a8 + 3);
  116. data61 = *(a8 + 4);
  117. data62 = *(a8 + 5);
  118. data63 = *(a8 + 6);
  119. #ifndef UNIT
  120. data64 = *(a8 + 7);
  121. #endif
  122. *(b + 0) = INV(data01);
  123. *(b + 1) = data09;
  124. *(b + 2) = data17;
  125. *(b + 3) = data25;
  126. *(b + 4) = data33;
  127. *(b + 5) = data41;
  128. *(b + 6) = data49;
  129. *(b + 7) = data57;
  130. *(b + 9) = INV(data10);
  131. *(b + 10) = data18;
  132. *(b + 11) = data26;
  133. *(b + 12) = data34;
  134. *(b + 13) = data42;
  135. *(b + 14) = data50;
  136. *(b + 15) = data58;
  137. *(b + 18) = INV(data19);
  138. *(b + 19) = data27;
  139. *(b + 20) = data35;
  140. *(b + 21) = data43;
  141. *(b + 22) = data51;
  142. *(b + 23) = data59;
  143. *(b + 27) = INV(data28);
  144. *(b + 28) = data36;
  145. *(b + 29) = data44;
  146. *(b + 30) = data52;
  147. *(b + 31) = data60;
  148. *(b + 36) = INV(data37);
  149. *(b + 37) = data45;
  150. *(b + 38) = data53;
  151. *(b + 39) = data61;
  152. *(b + 45) = INV(data46);
  153. *(b + 46) = data54;
  154. *(b + 47) = data62;
  155. *(b + 54) = INV(data55);
  156. *(b + 55) = data63;
  157. *(b + 63) = INV(data64);
  158. }
  159. if (ii < jj) {
  160. data01 = *(a1 + 0);
  161. data02 = *(a1 + 1);
  162. data03 = *(a1 + 2);
  163. data04 = *(a1 + 3);
  164. data05 = *(a1 + 4);
  165. data06 = *(a1 + 5);
  166. data07 = *(a1 + 6);
  167. data08 = *(a1 + 7);
  168. data09 = *(a2 + 0);
  169. data10 = *(a2 + 1);
  170. data11 = *(a2 + 2);
  171. data12 = *(a2 + 3);
  172. data13 = *(a2 + 4);
  173. data14 = *(a2 + 5);
  174. data15 = *(a2 + 6);
  175. data16 = *(a2 + 7);
  176. data17 = *(a3 + 0);
  177. data18 = *(a3 + 1);
  178. data19 = *(a3 + 2);
  179. data20 = *(a3 + 3);
  180. data21 = *(a3 + 4);
  181. data22 = *(a3 + 5);
  182. data23 = *(a3 + 6);
  183. data24 = *(a3 + 7);
  184. data25 = *(a4 + 0);
  185. data26 = *(a4 + 1);
  186. data27 = *(a4 + 2);
  187. data28 = *(a4 + 3);
  188. data29 = *(a4 + 4);
  189. data30 = *(a4 + 5);
  190. data31 = *(a4 + 6);
  191. data32 = *(a4 + 7);
  192. data33 = *(a5 + 0);
  193. data34 = *(a5 + 1);
  194. data35 = *(a5 + 2);
  195. data36 = *(a5 + 3);
  196. data37 = *(a5 + 4);
  197. data38 = *(a5 + 5);
  198. data39 = *(a5 + 6);
  199. data40 = *(a5 + 7);
  200. data41 = *(a6 + 0);
  201. data42 = *(a6 + 1);
  202. data43 = *(a6 + 2);
  203. data44 = *(a6 + 3);
  204. data45 = *(a6 + 4);
  205. data46 = *(a6 + 5);
  206. data47 = *(a6 + 6);
  207. data48 = *(a6 + 7);
  208. data49 = *(a7 + 0);
  209. data50 = *(a7 + 1);
  210. data51 = *(a7 + 2);
  211. data52 = *(a7 + 3);
  212. data53 = *(a7 + 4);
  213. data54 = *(a7 + 5);
  214. data55 = *(a7 + 6);
  215. data56 = *(a7 + 7);
  216. data57 = *(a8 + 0);
  217. data58 = *(a8 + 1);
  218. data59 = *(a8 + 2);
  219. data60 = *(a8 + 3);
  220. data61 = *(a8 + 4);
  221. data62 = *(a8 + 5);
  222. data63 = *(a8 + 6);
  223. data64 = *(a8 + 7);
  224. *(b + 0) = data01;
  225. *(b + 1) = data09;
  226. *(b + 2) = data17;
  227. *(b + 3) = data25;
  228. *(b + 4) = data33;
  229. *(b + 5) = data41;
  230. *(b + 6) = data49;
  231. *(b + 7) = data57;
  232. *(b + 8) = data02;
  233. *(b + 9) = data10;
  234. *(b + 10) = data18;
  235. *(b + 11) = data26;
  236. *(b + 12) = data34;
  237. *(b + 13) = data42;
  238. *(b + 14) = data50;
  239. *(b + 15) = data58;
  240. *(b + 16) = data03;
  241. *(b + 17) = data11;
  242. *(b + 18) = data19;
  243. *(b + 19) = data27;
  244. *(b + 20) = data35;
  245. *(b + 21) = data43;
  246. *(b + 22) = data51;
  247. *(b + 23) = data59;
  248. *(b + 24) = data04;
  249. *(b + 25) = data12;
  250. *(b + 26) = data20;
  251. *(b + 27) = data28;
  252. *(b + 28) = data36;
  253. *(b + 29) = data44;
  254. *(b + 30) = data52;
  255. *(b + 31) = data60;
  256. *(b + 32) = data05;
  257. *(b + 33) = data13;
  258. *(b + 34) = data21;
  259. *(b + 35) = data29;
  260. *(b + 36) = data37;
  261. *(b + 37) = data45;
  262. *(b + 38) = data53;
  263. *(b + 39) = data61;
  264. *(b + 40) = data06;
  265. *(b + 41) = data14;
  266. *(b + 42) = data22;
  267. *(b + 43) = data30;
  268. *(b + 44) = data38;
  269. *(b + 45) = data46;
  270. *(b + 46) = data54;
  271. *(b + 47) = data62;
  272. *(b + 48) = data07;
  273. *(b + 49) = data15;
  274. *(b + 50) = data23;
  275. *(b + 51) = data31;
  276. *(b + 52) = data39;
  277. *(b + 53) = data47;
  278. *(b + 54) = data55;
  279. *(b + 55) = data63;
  280. *(b + 56) = data08;
  281. *(b + 57) = data16;
  282. *(b + 58) = data24;
  283. *(b + 59) = data32;
  284. *(b + 60) = data40;
  285. *(b + 61) = data48;
  286. *(b + 62) = data56;
  287. *(b + 63) = data64;
  288. }
  289. a1 += 8;
  290. a2 += 8;
  291. a3 += 8;
  292. a4 += 8;
  293. a5 += 8;
  294. a6 += 8;
  295. a7 += 8;
  296. a8 += 8;
  297. b += 64;
  298. i --;
  299. ii += 8;
  300. }
  301. if (m & 4) {
  302. if (ii == jj) {
  303. #ifndef UNIT
  304. data01 = *(a1 + 0);
  305. #endif
  306. data09 = *(a2 + 0);
  307. #ifndef UNIT
  308. data10 = *(a2 + 1);
  309. #endif
  310. data17 = *(a3 + 0);
  311. data18 = *(a3 + 1);
  312. #ifndef UNIT
  313. data19 = *(a3 + 2);
  314. #endif
  315. data25 = *(a4 + 0);
  316. data26 = *(a4 + 1);
  317. data27 = *(a4 + 2);
  318. #ifndef UNIT
  319. data28 = *(a4 + 3);
  320. #endif
  321. data33 = *(a5 + 0);
  322. data34 = *(a5 + 1);
  323. data35 = *(a5 + 2);
  324. data36 = *(a5 + 3);
  325. data41 = *(a6 + 0);
  326. data42 = *(a6 + 1);
  327. data43 = *(a6 + 2);
  328. data44 = *(a6 + 3);
  329. data49 = *(a7 + 0);
  330. data50 = *(a7 + 1);
  331. data51 = *(a7 + 2);
  332. data52 = *(a7 + 3);
  333. data57 = *(a8 + 0);
  334. data58 = *(a8 + 1);
  335. data59 = *(a8 + 2);
  336. data60 = *(a8 + 3);
  337. *(b + 0) = INV(data01);
  338. *(b + 1) = data09;
  339. *(b + 2) = data17;
  340. *(b + 3) = data25;
  341. *(b + 4) = data33;
  342. *(b + 5) = data41;
  343. *(b + 6) = data49;
  344. *(b + 7) = data57;
  345. *(b + 9) = INV(data10);
  346. *(b + 10) = data18;
  347. *(b + 11) = data26;
  348. *(b + 12) = data34;
  349. *(b + 13) = data42;
  350. *(b + 14) = data50;
  351. *(b + 15) = data58;
  352. *(b + 18) = INV(data19);
  353. *(b + 19) = data27;
  354. *(b + 20) = data35;
  355. *(b + 21) = data43;
  356. *(b + 22) = data51;
  357. *(b + 23) = data59;
  358. *(b + 27) = INV(data28);
  359. *(b + 28) = data36;
  360. *(b + 29) = data44;
  361. *(b + 30) = data52;
  362. *(b + 31) = data60;
  363. }
  364. if (ii < jj) {
  365. data01 = *(a1 + 0);
  366. data02 = *(a1 + 1);
  367. data03 = *(a1 + 2);
  368. data04 = *(a1 + 3);
  369. data09 = *(a2 + 0);
  370. data10 = *(a2 + 1);
  371. data11 = *(a2 + 2);
  372. data12 = *(a2 + 3);
  373. data17 = *(a3 + 0);
  374. data18 = *(a3 + 1);
  375. data19 = *(a3 + 2);
  376. data20 = *(a3 + 3);
  377. data25 = *(a4 + 0);
  378. data26 = *(a4 + 1);
  379. data27 = *(a4 + 2);
  380. data28 = *(a4 + 3);
  381. data33 = *(a5 + 0);
  382. data34 = *(a5 + 1);
  383. data35 = *(a5 + 2);
  384. data36 = *(a5 + 3);
  385. data41 = *(a6 + 0);
  386. data42 = *(a6 + 1);
  387. data43 = *(a6 + 2);
  388. data44 = *(a6 + 3);
  389. data49 = *(a7 + 0);
  390. data50 = *(a7 + 1);
  391. data51 = *(a7 + 2);
  392. data52 = *(a7 + 3);
  393. data57 = *(a8 + 0);
  394. data58 = *(a8 + 1);
  395. data59 = *(a8 + 2);
  396. data60 = *(a8 + 3);
  397. *(b + 0) = data01;
  398. *(b + 1) = data09;
  399. *(b + 2) = data17;
  400. *(b + 3) = data25;
  401. *(b + 4) = data33;
  402. *(b + 5) = data41;
  403. *(b + 6) = data49;
  404. *(b + 7) = data57;
  405. *(b + 8) = data02;
  406. *(b + 9) = data10;
  407. *(b + 10) = data18;
  408. *(b + 11) = data26;
  409. *(b + 12) = data34;
  410. *(b + 13) = data42;
  411. *(b + 14) = data50;
  412. *(b + 15) = data58;
  413. *(b + 16) = data03;
  414. *(b + 17) = data11;
  415. *(b + 18) = data19;
  416. *(b + 19) = data27;
  417. *(b + 20) = data35;
  418. *(b + 21) = data43;
  419. *(b + 22) = data51;
  420. *(b + 23) = data59;
  421. *(b + 24) = data04;
  422. *(b + 25) = data12;
  423. *(b + 26) = data20;
  424. *(b + 27) = data28;
  425. *(b + 28) = data36;
  426. *(b + 29) = data44;
  427. *(b + 30) = data52;
  428. *(b + 31) = data60;
  429. }
  430. a1 += 4;
  431. a2 += 4;
  432. a3 += 4;
  433. a4 += 4;
  434. a5 += 4;
  435. a6 += 4;
  436. a7 += 4;
  437. a8 += 4;
  438. b += 32;
  439. ii += 4;
  440. }
  441. if (m & 2) {
  442. if (ii == jj) {
  443. #ifndef UNIT
  444. data01 = *(a1 + 0);
  445. #endif
  446. data09 = *(a2 + 0);
  447. #ifndef UNIT
  448. data10 = *(a2 + 1);
  449. #endif
  450. data17 = *(a3 + 0);
  451. data18 = *(a3 + 1);
  452. data25 = *(a4 + 0);
  453. data26 = *(a4 + 1);
  454. data33 = *(a5 + 0);
  455. data34 = *(a5 + 1);
  456. data41 = *(a6 + 0);
  457. data42 = *(a6 + 1);
  458. data49 = *(a7 + 0);
  459. data50 = *(a7 + 1);
  460. data57 = *(a8 + 0);
  461. data58 = *(a8 + 1);
  462. *(b + 0) = INV(data01);
  463. *(b + 1) = data09;
  464. *(b + 2) = data17;
  465. *(b + 3) = data25;
  466. *(b + 4) = data33;
  467. *(b + 5) = data41;
  468. *(b + 6) = data49;
  469. *(b + 7) = data57;
  470. *(b + 9) = INV(data10);
  471. *(b + 10) = data18;
  472. *(b + 11) = data26;
  473. *(b + 12) = data34;
  474. *(b + 13) = data42;
  475. *(b + 14) = data50;
  476. *(b + 15) = data58;
  477. }
  478. if (ii < jj) {
  479. data01 = *(a1 + 0);
  480. data02 = *(a1 + 1);
  481. data09 = *(a2 + 0);
  482. data10 = *(a2 + 1);
  483. data17 = *(a3 + 0);
  484. data18 = *(a3 + 1);
  485. data25 = *(a4 + 0);
  486. data26 = *(a4 + 1);
  487. data33 = *(a5 + 0);
  488. data34 = *(a5 + 1);
  489. data41 = *(a6 + 0);
  490. data42 = *(a6 + 1);
  491. data49 = *(a7 + 0);
  492. data50 = *(a7 + 1);
  493. data57 = *(a8 + 0);
  494. data58 = *(a8 + 1);
  495. *(b + 0) = data01;
  496. *(b + 1) = data09;
  497. *(b + 2) = data17;
  498. *(b + 3) = data25;
  499. *(b + 4) = data33;
  500. *(b + 5) = data41;
  501. *(b + 6) = data49;
  502. *(b + 7) = data57;
  503. *(b + 8) = data02;
  504. *(b + 9) = data10;
  505. *(b + 10) = data18;
  506. *(b + 11) = data26;
  507. *(b + 12) = data34;
  508. *(b + 13) = data42;
  509. *(b + 14) = data50;
  510. *(b + 15) = data58;
  511. }
  512. a1 += 2;
  513. a2 += 2;
  514. a3 += 2;
  515. a4 += 2;
  516. a5 += 2;
  517. a6 += 2;
  518. a7 += 2;
  519. a8 += 2;
  520. b += 16;
  521. ii += 2;
  522. }
  523. if (m & 1) {
  524. if (ii == jj) {
  525. #ifndef UNIT
  526. data01 = *(a1 + 0);
  527. #endif
  528. data09 = *(a2 + 0);
  529. data17 = *(a3 + 0);
  530. data25 = *(a4 + 0);
  531. data33 = *(a5 + 0);
  532. data41 = *(a6 + 0);
  533. data49 = *(a7 + 0);
  534. data57 = *(a8 + 0);
  535. *(b + 0) = INV(data01);
  536. *(b + 1) = data09;
  537. *(b + 2) = data17;
  538. *(b + 3) = data25;
  539. *(b + 4) = data33;
  540. *(b + 5) = data41;
  541. *(b + 6) = data49;
  542. *(b + 7) = data57;
  543. }
  544. if (ii < jj) {
  545. data01 = *(a1 + 0);
  546. // data02 = *(a1 + 1);
  547. data09 = *(a2 + 0);
  548. // data10 = *(a2 + 1);
  549. data17 = *(a3 + 0);
  550. // data18 = *(a3 + 1);
  551. data25 = *(a4 + 0);
  552. // data26 = *(a4 + 1);
  553. *(b + 0) = data01;
  554. *(b + 1) = data09;
  555. *(b + 2) = data17;
  556. *(b + 3) = data25;
  557. *(b + 4) = data33;
  558. *(b + 5) = data41;
  559. *(b + 6) = data49;
  560. *(b + 7) = data57;
  561. }
  562. b += 8;
  563. // ii += 1;
  564. }
  565. a += 8 * lda;
  566. jj += 8;
  567. j --;
  568. }
  569. if (n & 4) {
  570. a1 = a + 0 * lda;
  571. a2 = a + 1 * lda;
  572. a3 = a + 2 * lda;
  573. a4 = a + 3 * lda;
  574. ii = 0;
  575. i = (m >> 2);
  576. while (i > 0) {
  577. if (ii == jj) {
  578. #ifndef UNIT
  579. data01 = *(a1 + 0);
  580. #endif
  581. data09 = *(a2 + 0);
  582. #ifndef UNIT
  583. data10 = *(a2 + 1);
  584. #endif
  585. data17 = *(a3 + 0);
  586. data18 = *(a3 + 1);
  587. #ifndef UNIT
  588. data19 = *(a3 + 2);
  589. #endif
  590. data25 = *(a4 + 0);
  591. data26 = *(a4 + 1);
  592. data27 = *(a4 + 2);
  593. #ifndef UNIT
  594. data28 = *(a4 + 3);
  595. #endif
  596. *(b + 0) = INV(data01);
  597. *(b + 1) = data09;
  598. *(b + 2) = data17;
  599. *(b + 3) = data25;
  600. *(b + 5) = INV(data10);
  601. *(b + 6) = data18;
  602. *(b + 7) = data26;
  603. *(b + 10) = INV(data19);
  604. *(b + 11) = data27;
  605. *(b + 15) = INV(data28);
  606. }
  607. if (ii < jj) {
  608. data01 = *(a1 + 0);
  609. data02 = *(a1 + 1);
  610. data03 = *(a1 + 2);
  611. data04 = *(a1 + 3);
  612. data09 = *(a2 + 0);
  613. data10 = *(a2 + 1);
  614. data11 = *(a2 + 2);
  615. data12 = *(a2 + 3);
  616. data17 = *(a3 + 0);
  617. data18 = *(a3 + 1);
  618. data19 = *(a3 + 2);
  619. data20 = *(a3 + 3);
  620. data25 = *(a4 + 0);
  621. data26 = *(a4 + 1);
  622. data27 = *(a4 + 2);
  623. data28 = *(a4 + 3);
  624. *(b + 0) = data01;
  625. *(b + 1) = data09;
  626. *(b + 2) = data17;
  627. *(b + 3) = data25;
  628. *(b + 4) = data02;
  629. *(b + 5) = data10;
  630. *(b + 6) = data18;
  631. *(b + 7) = data26;
  632. *(b + 8) = data03;
  633. *(b + 9) = data11;
  634. *(b + 10) = data19;
  635. *(b + 11) = data27;
  636. *(b + 12) = data04;
  637. *(b + 13) = data12;
  638. *(b + 14) = data20;
  639. *(b + 15) = data28;
  640. }
  641. a1 += 4;
  642. a2 += 4;
  643. a3 += 4;
  644. a4 += 4;
  645. b += 16;
  646. i --;
  647. ii += 4;
  648. }
  649. if (m & 2) {
  650. if (ii == jj) {
  651. #ifndef UNIT
  652. data01 = *(a1 + 0);
  653. #endif
  654. data09 = *(a2 + 0);
  655. #ifndef UNIT
  656. data10 = *(a2 + 1);
  657. #endif
  658. data17 = *(a3 + 0);
  659. data18 = *(a3 + 1);
  660. data25 = *(a4 + 0);
  661. data26 = *(a4 + 1);
  662. *(b + 0) = INV(data01);
  663. *(b + 1) = data09;
  664. *(b + 2) = data17;
  665. *(b + 3) = data25;
  666. *(b + 5) = INV(data10);
  667. *(b + 6) = data18;
  668. *(b + 7) = data26;
  669. }
  670. if (ii < jj) {
  671. data01 = *(a1 + 0);
  672. data02 = *(a1 + 1);
  673. data09 = *(a2 + 0);
  674. data10 = *(a2 + 1);
  675. data17 = *(a3 + 0);
  676. data18 = *(a3 + 1);
  677. data25 = *(a4 + 0);
  678. data26 = *(a4 + 1);
  679. *(b + 0) = data01;
  680. *(b + 1) = data09;
  681. *(b + 2) = data17;
  682. *(b + 3) = data25;
  683. *(b + 4) = data02;
  684. *(b + 5) = data10;
  685. *(b + 6) = data18;
  686. *(b + 7) = data26;
  687. }
  688. a1 += 2;
  689. a2 += 2;
  690. a3 += 2;
  691. a4 += 2;
  692. b += 8;
  693. ii += 2;
  694. }
  695. if (m & 1) {
  696. if (ii == jj) {
  697. #ifndef UNIT
  698. data01 = *(a1 + 0);
  699. #endif
  700. data09 = *(a2 + 0);
  701. data17 = *(a3 + 0);
  702. data25 = *(a4 + 0);
  703. *(b + 0) = INV(data01);
  704. *(b + 1) = data09;
  705. *(b + 2) = data17;
  706. *(b + 3) = data25;
  707. }
  708. if (ii < jj) {
  709. data01 = *(a1 + 0);
  710. data09 = *(a2 + 0);
  711. data17 = *(a3 + 0);
  712. data25 = *(a4 + 0);
  713. *(b + 0) = data01;
  714. *(b + 1) = data09;
  715. *(b + 2) = data17;
  716. *(b + 3) = data25;
  717. }
  718. b += 4;
  719. // ii += 1;
  720. }
  721. a += 4 * lda;
  722. jj += 4;
  723. }
  724. if (n & 2) {
  725. a1 = a + 0 * lda;
  726. a2 = a + 1 * lda;
  727. ii = 0;
  728. i = (m >> 1);
  729. while (i > 0) {
  730. if (ii == jj) {
  731. #ifndef UNIT
  732. data01 = *(a1 + 0);
  733. #endif
  734. data09 = *(a2 + 0);
  735. #ifndef UNIT
  736. data10 = *(a2 + 1);
  737. #endif
  738. *(b + 0) = INV(data01);
  739. *(b + 1) = data09;
  740. *(b + 3) = INV(data10);
  741. }
  742. if (ii < jj) {
  743. data01 = *(a1 + 0);
  744. data02 = *(a1 + 1);
  745. data09 = *(a2 + 0);
  746. data10 = *(a2 + 1);
  747. *(b + 0) = data01;
  748. *(b + 1) = data09;
  749. *(b + 2) = data02;
  750. *(b + 3) = data10;
  751. }
  752. a1 += 2;
  753. a2 += 2;
  754. b += 4;
  755. i --;
  756. ii += 2;
  757. }
  758. if (m & 1) {
  759. if (ii == jj) {
  760. #ifndef UNIT
  761. data01 = *(a1 + 0);
  762. #endif
  763. data09 = *(a2 + 0);
  764. *(b + 0) = INV(data01);
  765. *(b + 1) = data09;
  766. }
  767. if (ii < jj) {
  768. data01 = *(a1 + 0);
  769. data09 = *(a2 + 0);
  770. *(b + 0) = data01;
  771. *(b + 1) = data09;
  772. }
  773. b += 2;
  774. // ii += 1;
  775. }
  776. a += 2 * lda;
  777. jj += 2;
  778. }
  779. if (n & 1) {
  780. a1 = a + 0 * lda;
  781. ii = 0;
  782. i = m;
  783. while (i > 0) {
  784. if (ii == jj) {
  785. #ifndef UNIT
  786. data01 = *(a1 + 0);
  787. #endif
  788. *(b + 0) = INV(data01);
  789. }
  790. if (ii < jj) {
  791. data01 = *(a1 + 0);
  792. *(b + 0) = data01;
  793. }
  794. a1 += 1;
  795. b += 1;
  796. i --;
  797. ii ++;
  798. }
  799. }
  800. return 0;
  801. }