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.

trmm_utcopy_8.c 26 kB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276
  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. int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, FLOAT *b){
  41. BLASLONG i, js;
  42. BLASLONG X;
  43. FLOAT data01, data02, data03, data04, data05, data06, data07, data08;
  44. FLOAT data09, data10, data11, data12, data13, data14, data15, data16;
  45. FLOAT data17, data18, data19, data20, data21, data22, data23, data24;
  46. FLOAT data25, data26, data27, data28, data29, data30, data31, data32;
  47. FLOAT data33, data34, data35, data36, data37, data38, data39, data40;
  48. FLOAT data41, data42, data43, data44, data45, data46, data47, data48;
  49. FLOAT data49, data50, data51, data52, data53, data54, data55, data56;
  50. FLOAT data57, data58, data59, data60, data61, data62, data63, data64;
  51. FLOAT *ao1, *ao2, *ao3, *ao4, *ao5, *ao6, *ao7, *ao8;
  52. js = (n >> 3);
  53. if (js > 0){
  54. do {
  55. X = posX;
  56. if (posX <= posY) {
  57. ao1 = a + posX + (posY + 0) * lda;
  58. ao2 = a + posX + (posY + 1) * lda;
  59. ao3 = a + posX + (posY + 2) * lda;
  60. ao4 = a + posX + (posY + 3) * lda;
  61. ao5 = a + posX + (posY + 4) * lda;
  62. ao6 = a + posX + (posY + 5) * lda;
  63. ao7 = a + posX + (posY + 6) * lda;
  64. ao8 = a + posX + (posY + 7) * lda;
  65. } else {
  66. ao1 = a + posY + (posX + 0) * lda;
  67. ao2 = a + posY + (posX + 1) * lda;
  68. ao3 = a + posY + (posX + 2) * lda;
  69. ao4 = a + posY + (posX + 3) * lda;
  70. ao5 = a + posY + (posX + 4) * lda;
  71. ao6 = a + posY + (posX + 5) * lda;
  72. ao7 = a + posY + (posX + 6) * lda;
  73. ao8 = a + posY + (posX + 7) * lda;
  74. }
  75. i = (m >> 3);
  76. if (i > 0) {
  77. do {
  78. if (X < posY) {
  79. ao1 += 8;
  80. ao2 += 8;
  81. ao3 += 8;
  82. ao4 += 8;
  83. ao5 += 8;
  84. ao6 += 8;
  85. ao7 += 8;
  86. ao8 += 8;
  87. b += 64;
  88. } else
  89. if (X > posY) {
  90. data01 = *(ao1 + 0);
  91. data02 = *(ao1 + 1);
  92. data03 = *(ao1 + 2);
  93. data04 = *(ao1 + 3);
  94. data05 = *(ao1 + 4);
  95. data06 = *(ao1 + 5);
  96. data07 = *(ao1 + 6);
  97. data08 = *(ao1 + 7);
  98. data09 = *(ao2 + 0);
  99. data10 = *(ao2 + 1);
  100. data11 = *(ao2 + 2);
  101. data12 = *(ao2 + 3);
  102. data13 = *(ao2 + 4);
  103. data14 = *(ao2 + 5);
  104. data15 = *(ao2 + 6);
  105. data16 = *(ao2 + 7);
  106. data17 = *(ao3 + 0);
  107. data18 = *(ao3 + 1);
  108. data19 = *(ao3 + 2);
  109. data20 = *(ao3 + 3);
  110. data21 = *(ao3 + 4);
  111. data22 = *(ao3 + 5);
  112. data23 = *(ao3 + 6);
  113. data24 = *(ao3 + 7);
  114. data25 = *(ao4 + 0);
  115. data26 = *(ao4 + 1);
  116. data27 = *(ao4 + 2);
  117. data28 = *(ao4 + 3);
  118. data29 = *(ao4 + 4);
  119. data30 = *(ao4 + 5);
  120. data31 = *(ao4 + 6);
  121. data32 = *(ao4 + 7);
  122. data33 = *(ao5 + 0);
  123. data34 = *(ao5 + 1);
  124. data35 = *(ao5 + 2);
  125. data36 = *(ao5 + 3);
  126. data37 = *(ao5 + 4);
  127. data38 = *(ao5 + 5);
  128. data39 = *(ao5 + 6);
  129. data40 = *(ao5 + 7);
  130. data41 = *(ao6 + 0);
  131. data42 = *(ao6 + 1);
  132. data43 = *(ao6 + 2);
  133. data44 = *(ao6 + 3);
  134. data45 = *(ao6 + 4);
  135. data46 = *(ao6 + 5);
  136. data47 = *(ao6 + 6);
  137. data48 = *(ao6 + 7);
  138. data49 = *(ao7 + 0);
  139. data50 = *(ao7 + 1);
  140. data51 = *(ao7 + 2);
  141. data52 = *(ao7 + 3);
  142. data53 = *(ao7 + 4);
  143. data54 = *(ao7 + 5);
  144. data55 = *(ao7 + 6);
  145. data56 = *(ao7 + 7);
  146. data57 = *(ao8 + 0);
  147. data58 = *(ao8 + 1);
  148. data59 = *(ao8 + 2);
  149. data60 = *(ao8 + 3);
  150. data61 = *(ao8 + 4);
  151. data62 = *(ao8 + 5);
  152. data63 = *(ao8 + 6);
  153. data64 = *(ao8 + 7);
  154. b[ 0] = data01;
  155. b[ 1] = data02;
  156. b[ 2] = data03;
  157. b[ 3] = data04;
  158. b[ 4] = data05;
  159. b[ 5] = data06;
  160. b[ 6] = data07;
  161. b[ 7] = data08;
  162. b[ 8] = data09;
  163. b[ 9] = data10;
  164. b[10] = data11;
  165. b[11] = data12;
  166. b[12] = data13;
  167. b[13] = data14;
  168. b[14] = data15;
  169. b[15] = data16;
  170. b[16] = data17;
  171. b[17] = data18;
  172. b[18] = data19;
  173. b[19] = data20;
  174. b[20] = data21;
  175. b[21] = data22;
  176. b[22] = data23;
  177. b[23] = data24;
  178. b[24] = data25;
  179. b[25] = data26;
  180. b[26] = data27;
  181. b[27] = data28;
  182. b[28] = data29;
  183. b[29] = data30;
  184. b[30] = data31;
  185. b[31] = data32;
  186. b[32] = data33;
  187. b[33] = data34;
  188. b[34] = data35;
  189. b[35] = data36;
  190. b[36] = data37;
  191. b[37] = data38;
  192. b[38] = data39;
  193. b[39] = data40;
  194. b[40] = data41;
  195. b[41] = data42;
  196. b[42] = data43;
  197. b[43] = data44;
  198. b[44] = data45;
  199. b[45] = data46;
  200. b[46] = data47;
  201. b[47] = data48;
  202. b[48] = data49;
  203. b[49] = data50;
  204. b[50] = data51;
  205. b[51] = data52;
  206. b[52] = data53;
  207. b[53] = data54;
  208. b[54] = data55;
  209. b[55] = data56;
  210. b[56] = data57;
  211. b[57] = data58;
  212. b[58] = data59;
  213. b[59] = data60;
  214. b[60] = data61;
  215. b[61] = data62;
  216. b[62] = data63;
  217. b[63] = data64;
  218. ao1 += 8 * lda;
  219. ao2 += 8 * lda;
  220. ao3 += 8 * lda;
  221. ao4 += 8 * lda;
  222. ao5 += 8 * lda;
  223. ao6 += 8 * lda;
  224. ao7 += 8 * lda;
  225. ao8 += 8 * lda;
  226. b += 64;
  227. } else {
  228. #ifdef UNIT
  229. data09 = *(ao2 + 0);
  230. data17 = *(ao3 + 0);
  231. data18 = *(ao3 + 1);
  232. data25 = *(ao4 + 0);
  233. data26 = *(ao4 + 1);
  234. data27 = *(ao4 + 2);
  235. data33 = *(ao5 + 0);
  236. data34 = *(ao5 + 1);
  237. data35 = *(ao5 + 2);
  238. data36 = *(ao5 + 3);
  239. data41 = *(ao6 + 0);
  240. data42 = *(ao6 + 1);
  241. data43 = *(ao6 + 2);
  242. data44 = *(ao6 + 3);
  243. data45 = *(ao6 + 4);
  244. data49 = *(ao7 + 0);
  245. data50 = *(ao7 + 1);
  246. data51 = *(ao7 + 2);
  247. data52 = *(ao7 + 3);
  248. data53 = *(ao7 + 4);
  249. data54 = *(ao7 + 5);
  250. data57 = *(ao8 + 0);
  251. data58 = *(ao8 + 1);
  252. data59 = *(ao8 + 2);
  253. data60 = *(ao8 + 3);
  254. data61 = *(ao8 + 4);
  255. data62 = *(ao8 + 5);
  256. data63 = *(ao8 + 6);
  257. b[ 0] = ONE;
  258. b[ 1] = ZERO;
  259. b[ 2] = ZERO;
  260. b[ 3] = ZERO;
  261. b[ 4] = ZERO;
  262. b[ 5] = ZERO;
  263. b[ 6] = ZERO;
  264. b[ 7] = ZERO;
  265. b[ 8] = data09;
  266. b[ 9] = ONE;
  267. b[10] = ZERO;
  268. b[11] = ZERO;
  269. b[12] = ZERO;
  270. b[13] = ZERO;
  271. b[14] = ZERO;
  272. b[15] = ZERO;
  273. b[16] = data17;
  274. b[17] = data18;
  275. b[18] = ONE;
  276. b[19] = ZERO;
  277. b[20] = ZERO;
  278. b[21] = ZERO;
  279. b[22] = ZERO;
  280. b[23] = ZERO;
  281. b[24] = data25;
  282. b[25] = data26;
  283. b[26] = data27;
  284. b[27] = ONE;
  285. b[28] = ZERO;
  286. b[29] = ZERO;
  287. b[30] = ZERO;
  288. b[31] = ZERO;
  289. b[32] = data33;
  290. b[33] = data34;
  291. b[34] = data35;
  292. b[35] = data36;
  293. b[36] = ONE;
  294. b[37] = ZERO;
  295. b[38] = ZERO;
  296. b[39] = ZERO;
  297. b[40] = data41;
  298. b[41] = data42;
  299. b[42] = data43;
  300. b[43] = data44;
  301. b[44] = data45;
  302. b[45] = ONE;
  303. b[46] = ZERO;
  304. b[47] = ZERO;
  305. b[48] = data49;
  306. b[49] = data50;
  307. b[50] = data51;
  308. b[51] = data52;
  309. b[52] = data53;
  310. b[53] = data54;
  311. b[54] = ONE;
  312. b[55] = ZERO;
  313. b[56] = data57;
  314. b[57] = data58;
  315. b[58] = data59;
  316. b[59] = data60;
  317. b[60] = data61;
  318. b[61] = data62;
  319. b[62] = data63;
  320. b[63] = ONE;
  321. #else
  322. data01 = *(ao1 + 0);
  323. data09 = *(ao2 + 0);
  324. data10 = *(ao2 + 1);
  325. data17 = *(ao3 + 0);
  326. data18 = *(ao3 + 1);
  327. data19 = *(ao3 + 2);
  328. data25 = *(ao4 + 0);
  329. data26 = *(ao4 + 1);
  330. data27 = *(ao4 + 2);
  331. data28 = *(ao4 + 3);
  332. data33 = *(ao5 + 0);
  333. data34 = *(ao5 + 1);
  334. data35 = *(ao5 + 2);
  335. data36 = *(ao5 + 3);
  336. data37 = *(ao5 + 4);
  337. data41 = *(ao6 + 0);
  338. data42 = *(ao6 + 1);
  339. data43 = *(ao6 + 2);
  340. data44 = *(ao6 + 3);
  341. data45 = *(ao6 + 4);
  342. data46 = *(ao6 + 5);
  343. data49 = *(ao7 + 0);
  344. data50 = *(ao7 + 1);
  345. data51 = *(ao7 + 2);
  346. data52 = *(ao7 + 3);
  347. data53 = *(ao7 + 4);
  348. data54 = *(ao7 + 5);
  349. data55 = *(ao7 + 6);
  350. data57 = *(ao8 + 0);
  351. data58 = *(ao8 + 1);
  352. data59 = *(ao8 + 2);
  353. data60 = *(ao8 + 3);
  354. data61 = *(ao8 + 4);
  355. data62 = *(ao8 + 5);
  356. data63 = *(ao8 + 6);
  357. data64 = *(ao8 + 7);
  358. b[ 0] = data01;
  359. b[ 1] = ZERO;
  360. b[ 2] = ZERO;
  361. b[ 3] = ZERO;
  362. b[ 4] = ZERO;
  363. b[ 5] = ZERO;
  364. b[ 6] = ZERO;
  365. b[ 7] = ZERO;
  366. b[ 8] = data09;
  367. b[ 9] = data10;
  368. b[10] = ZERO;
  369. b[11] = ZERO;
  370. b[12] = ZERO;
  371. b[13] = ZERO;
  372. b[14] = ZERO;
  373. b[15] = ZERO;
  374. b[16] = data17;
  375. b[17] = data18;
  376. b[18] = data19;
  377. b[19] = ZERO;
  378. b[20] = ZERO;
  379. b[21] = ZERO;
  380. b[22] = ZERO;
  381. b[23] = ZERO;
  382. b[24] = data25;
  383. b[25] = data26;
  384. b[26] = data27;
  385. b[27] = data28;
  386. b[28] = ZERO;
  387. b[29] = ZERO;
  388. b[30] = ZERO;
  389. b[31] = ZERO;
  390. b[32] = data33;
  391. b[33] = data34;
  392. b[34] = data35;
  393. b[35] = data36;
  394. b[36] = data37;
  395. b[37] = ZERO;
  396. b[38] = ZERO;
  397. b[39] = ZERO;
  398. b[40] = data41;
  399. b[41] = data42;
  400. b[42] = data43;
  401. b[43] = data44;
  402. b[44] = data45;
  403. b[45] = data46;
  404. b[46] = ZERO;
  405. b[47] = ZERO;
  406. b[48] = data49;
  407. b[49] = data50;
  408. b[50] = data51;
  409. b[51] = data52;
  410. b[52] = data53;
  411. b[53] = data54;
  412. b[54] = data55;
  413. b[55] = ZERO;
  414. b[56] = data57;
  415. b[57] = data58;
  416. b[58] = data59;
  417. b[59] = data60;
  418. b[60] = data61;
  419. b[61] = data62;
  420. b[62] = data63;
  421. b[63] = data64;
  422. #endif
  423. ao1 += 8 * lda;
  424. ao2 += 8 * lda;
  425. ao3 += 8 * lda;
  426. ao4 += 8 * lda;
  427. ao5 += 8 * lda;
  428. ao6 += 8 * lda;
  429. ao7 += 8 * lda;
  430. ao8 += 8 * lda;
  431. b += 64;
  432. }
  433. X += 8;
  434. i --;
  435. } while (i > 0);
  436. }
  437. i = (m & 7);
  438. if (i) {
  439. if (X < posY) {
  440. if (m & 4) {
  441. ao1 += 4;
  442. ao2 += 4;
  443. ao3 += 4;
  444. ao4 += 4;
  445. ao5 += 4;
  446. ao6 += 4;
  447. ao7 += 4;
  448. ao8 += 4;
  449. b += 32;
  450. }
  451. if (m & 2) {
  452. ao1 += 2;
  453. ao2 += 2;
  454. ao3 += 2;
  455. ao4 += 2;
  456. ao5 += 2;
  457. ao6 += 2;
  458. ao7 += 2;
  459. ao8 += 2;
  460. b += 16;
  461. }
  462. if (m & 1) {
  463. b += 8;
  464. }
  465. } else
  466. if (X > posY) {
  467. if (m & 4) {
  468. data01 = *(ao1 + 0);
  469. data02 = *(ao1 + 1);
  470. data03 = *(ao1 + 2);
  471. data04 = *(ao1 + 3);
  472. data05 = *(ao1 + 4);
  473. data06 = *(ao1 + 5);
  474. data07 = *(ao1 + 6);
  475. data08 = *(ao1 + 7);
  476. data09 = *(ao2 + 0);
  477. data10 = *(ao2 + 1);
  478. data11 = *(ao2 + 2);
  479. data12 = *(ao2 + 3);
  480. data13 = *(ao2 + 4);
  481. data14 = *(ao2 + 5);
  482. data15 = *(ao2 + 6);
  483. data16 = *(ao2 + 7);
  484. data17 = *(ao3 + 0);
  485. data18 = *(ao3 + 1);
  486. data19 = *(ao3 + 2);
  487. data20 = *(ao3 + 3);
  488. data21 = *(ao3 + 4);
  489. data22 = *(ao3 + 5);
  490. data23 = *(ao3 + 6);
  491. data24 = *(ao3 + 7);
  492. data25 = *(ao4 + 0);
  493. data26 = *(ao4 + 1);
  494. data27 = *(ao4 + 2);
  495. data28 = *(ao4 + 3);
  496. data29 = *(ao4 + 4);
  497. data30 = *(ao4 + 5);
  498. data31 = *(ao4 + 6);
  499. data32 = *(ao4 + 7);
  500. b[ 0] = data01;
  501. b[ 1] = data02;
  502. b[ 2] = data03;
  503. b[ 3] = data04;
  504. b[ 4] = data05;
  505. b[ 5] = data06;
  506. b[ 6] = data07;
  507. b[ 7] = data08;
  508. b[ 8] = data09;
  509. b[ 9] = data10;
  510. b[10] = data11;
  511. b[11] = data12;
  512. b[12] = data13;
  513. b[13] = data14;
  514. b[14] = data15;
  515. b[15] = data16;
  516. b[16] = data17;
  517. b[17] = data18;
  518. b[18] = data19;
  519. b[19] = data20;
  520. b[20] = data21;
  521. b[21] = data22;
  522. b[22] = data23;
  523. b[23] = data24;
  524. b[24] = data25;
  525. b[25] = data26;
  526. b[26] = data27;
  527. b[27] = data28;
  528. b[28] = data29;
  529. b[29] = data30;
  530. b[30] = data31;
  531. b[31] = data32;
  532. ao1 += 4 * lda;
  533. ao2 += 4 * lda;
  534. ao3 += 4 * lda;
  535. ao4 += 4 * lda;
  536. b += 32;
  537. }
  538. if (m & 2) {
  539. data01 = *(ao1 + 0);
  540. data02 = *(ao1 + 1);
  541. data03 = *(ao1 + 2);
  542. data04 = *(ao1 + 3);
  543. data05 = *(ao1 + 4);
  544. data06 = *(ao1 + 5);
  545. data07 = *(ao1 + 6);
  546. data08 = *(ao1 + 7);
  547. data09 = *(ao2 + 0);
  548. data10 = *(ao2 + 1);
  549. data11 = *(ao2 + 2);
  550. data12 = *(ao2 + 3);
  551. data13 = *(ao2 + 4);
  552. data14 = *(ao2 + 5);
  553. data15 = *(ao2 + 6);
  554. data16 = *(ao2 + 7);
  555. b[ 0] = data01;
  556. b[ 1] = data02;
  557. b[ 2] = data03;
  558. b[ 3] = data04;
  559. b[ 4] = data05;
  560. b[ 5] = data06;
  561. b[ 6] = data07;
  562. b[ 7] = data08;
  563. b[ 8] = data09;
  564. b[ 9] = data10;
  565. b[10] = data11;
  566. b[11] = data12;
  567. b[12] = data13;
  568. b[13] = data14;
  569. b[14] = data15;
  570. b[15] = data16;
  571. ao1 += 2 * lda;
  572. b += 16;
  573. }
  574. if (m & 1) {
  575. data01 = *(ao1 + 0);
  576. data02 = *(ao1 + 1);
  577. data03 = *(ao1 + 2);
  578. data04 = *(ao1 + 3);
  579. data05 = *(ao1 + 4);
  580. data06 = *(ao1 + 5);
  581. data07 = *(ao1 + 6);
  582. data08 = *(ao1 + 7);
  583. b[ 0] = data01;
  584. b[ 1] = data02;
  585. b[ 2] = data03;
  586. b[ 3] = data04;
  587. b[ 4] = data05;
  588. b[ 5] = data06;
  589. b[ 6] = data07;
  590. b[ 7] = data08;
  591. b += 8;
  592. }
  593. } else {
  594. #ifndef UNIT
  595. data01 = *(ao1 + 0);
  596. #endif
  597. data09 = *(ao2 + 0);
  598. data17 = *(ao3 + 0);
  599. data25 = *(ao4 + 0);
  600. data33 = *(ao5 + 0);
  601. data41 = *(ao6 + 0);
  602. data49 = *(ao7 + 0);
  603. data57 = *(ao8 + 0);
  604. if (i >= 2) {
  605. #ifndef UNIT
  606. data10 = *(ao2 + 1);
  607. #endif
  608. data18 = *(ao3 + 1);
  609. data26 = *(ao4 + 1);
  610. data34 = *(ao5 + 1);
  611. data42 = *(ao6 + 1);
  612. data50 = *(ao7 + 1);
  613. data58 = *(ao8 + 1);
  614. }
  615. if (i >= 3) {
  616. #ifndef UNIT
  617. data19 = *(ao3 + 2);
  618. #endif
  619. data27 = *(ao4 + 2);
  620. data35 = *(ao5 + 2);
  621. data43 = *(ao6 + 2);
  622. data51 = *(ao7 + 2);
  623. data59 = *(ao8 + 2);
  624. }
  625. if (i >= 4) {
  626. #ifndef UNIT
  627. data28 = *(ao4 + 3);
  628. #endif
  629. data36 = *(ao5 + 3);
  630. data44 = *(ao6 + 3);
  631. data52 = *(ao7 + 3);
  632. data60 = *(ao8 + 3);
  633. }
  634. if (i >= 5) {
  635. #ifndef UNIT
  636. data37 = *(ao5 + 4);
  637. #endif
  638. data45 = *(ao6 + 4);
  639. data53 = *(ao7 + 4);
  640. data61 = *(ao8 + 4);
  641. }
  642. if (i >= 6) {
  643. #ifndef UNIT
  644. data46 = *(ao6 + 5);
  645. #endif
  646. data54 = *(ao7 + 5);
  647. data62 = *(ao8 + 5);
  648. }
  649. if (i >= 7) {
  650. #ifndef UNIT
  651. data55 = *(ao7 + 6);
  652. #endif
  653. data63 = *(ao8 + 6);
  654. }
  655. #ifdef UNIT
  656. b[ 0] = ONE;
  657. #else
  658. b[ 0] = data01;
  659. #endif
  660. b[ 1] = ZERO;
  661. b[ 2] = ZERO;
  662. b[ 3] = ZERO;
  663. b[ 4] = ZERO;
  664. b[ 5] = ZERO;
  665. b[ 6] = ZERO;
  666. b[ 7] = ZERO;
  667. b += 8;
  668. if(i >= 2) {
  669. b[ 0] = data09;
  670. #ifdef UNIT
  671. b[ 1] = ONE;
  672. #else
  673. b[ 1] = data10;
  674. #endif
  675. b[ 2] = ZERO;
  676. b[ 3] = ZERO;
  677. b[ 4] = ZERO;
  678. b[ 5] = ZERO;
  679. b[ 6] = ZERO;
  680. b[ 7] = ZERO;
  681. b += 8;
  682. }
  683. if (i >= 3) {
  684. b[ 0] = data17;
  685. b[ 1] = data18;
  686. #ifdef UNIT
  687. b[ 2] = ONE;
  688. #else
  689. b[ 2] = data19;
  690. #endif
  691. b[ 3] = ZERO;
  692. b[ 4] = ZERO;
  693. b[ 5] = ZERO;
  694. b[ 6] = ZERO;
  695. b[ 7] = ZERO;
  696. b += 8;
  697. }
  698. if (i >= 4) {
  699. b[ 0] = data25;
  700. b[ 1] = data26;
  701. b[ 2] = data27;
  702. #ifdef UNIT
  703. b[ 3] = ONE;
  704. #else
  705. b[ 3] = data28;
  706. #endif
  707. b[ 4] = ZERO;
  708. b[ 5] = ZERO;
  709. b[ 6] = ZERO;
  710. b[ 7] = ZERO;
  711. b += 8;
  712. }
  713. if (i >= 5) {
  714. b[ 0] = data33;
  715. b[ 1] = data34;
  716. b[ 2] = data35;
  717. b[ 3] = data36;
  718. #ifdef UNIT
  719. b[ 4] = ONE;
  720. #else
  721. b[ 4] = data37;
  722. #endif
  723. b[ 5] = ZERO;
  724. b[ 6] = ZERO;
  725. b[ 7] = ZERO;
  726. b += 8;
  727. }
  728. if (i >= 6) {
  729. b[ 0] = data41;
  730. b[ 1] = data42;
  731. b[ 2] = data43;
  732. b[ 3] = data44;
  733. b[ 4] = data45;
  734. #ifdef UNIT
  735. b[ 5] = ONE;
  736. #else
  737. b[ 5] = data46;
  738. #endif
  739. b[ 6] = ZERO;
  740. b[ 7] = ZERO;
  741. b += 8;
  742. }
  743. if (i >= 7) {
  744. b[ 0] = data49;
  745. b[ 1] = data50;
  746. b[ 2] = data51;
  747. b[ 3] = data52;
  748. b[ 4] = data53;
  749. b[ 5] = data54;
  750. #ifdef UNIT
  751. b[ 6] = ONE;
  752. #else
  753. b[ 6] = data55;
  754. #endif
  755. b[ 7] = ZERO;
  756. b += 8;
  757. }
  758. }
  759. }
  760. posY += 8;
  761. js --;
  762. } while (js > 0);
  763. } /* End of main loop */
  764. if (n & 4){
  765. X = posX;
  766. if (posX <= posY) {
  767. ao1 = a + posX + (posY + 0) * lda;
  768. ao2 = a + posX + (posY + 1) * lda;
  769. ao3 = a + posX + (posY + 2) * lda;
  770. ao4 = a + posX + (posY + 3) * lda;
  771. } else {
  772. ao1 = a + posY + (posX + 0) * lda;
  773. ao2 = a + posY + (posX + 1) * lda;
  774. ao3 = a + posY + (posX + 2) * lda;
  775. ao4 = a + posY + (posX + 3) * lda;
  776. }
  777. i = (m >> 2);
  778. if (i > 0) {
  779. do {
  780. if (X < posY) {
  781. ao1 += 4;
  782. ao2 += 4;
  783. ao3 += 4;
  784. ao4 += 4;
  785. b += 16;
  786. } else
  787. if (X > posY) {
  788. data01 = *(ao1 + 0);
  789. data02 = *(ao1 + 1);
  790. data03 = *(ao1 + 2);
  791. data04 = *(ao1 + 3);
  792. data09 = *(ao2 + 0);
  793. data10 = *(ao2 + 1);
  794. data11 = *(ao2 + 2);
  795. data12 = *(ao2 + 3);
  796. data17 = *(ao3 + 0);
  797. data18 = *(ao3 + 1);
  798. data19 = *(ao3 + 2);
  799. data20 = *(ao3 + 3);
  800. data25 = *(ao4 + 0);
  801. data26 = *(ao4 + 1);
  802. data27 = *(ao4 + 2);
  803. data28 = *(ao4 + 3);
  804. ao1 += 4 * lda;
  805. ao2 += 4 * lda;
  806. ao3 += 4 * lda;
  807. ao4 += 4 * lda;
  808. b[ 0] = data01;
  809. b[ 1] = data02;
  810. b[ 2] = data03;
  811. b[ 3] = data04;
  812. b[ 4] = data09;
  813. b[ 5] = data10;
  814. b[ 6] = data11;
  815. b[ 7] = data12;
  816. b[ 8] = data17;
  817. b[ 9] = data18;
  818. b[10] = data19;
  819. b[11] = data20;
  820. b[12] = data25;
  821. b[13] = data26;
  822. b[14] = data27;
  823. b[15] = data28;
  824. b += 16;
  825. } else {
  826. #ifdef UNIT
  827. data09 = *(ao2 + 0);
  828. data17 = *(ao3 + 0);
  829. data18 = *(ao3 + 1);
  830. data25 = *(ao4 + 0);
  831. data26 = *(ao4 + 1);
  832. data27 = *(ao4 + 2);
  833. b[ 0] = ONE;
  834. b[ 1] = ZERO;
  835. b[ 2] = ZERO;
  836. b[ 3] = ZERO;
  837. b[ 4] = data09;
  838. b[ 5] = ONE;
  839. b[ 6] = ZERO;
  840. b[ 7] = ZERO;
  841. b[ 8] = data17;
  842. b[ 9] = data18;
  843. b[10] = ONE;
  844. b[11] = ZERO;
  845. b[12] = data25;
  846. b[13] = data26;
  847. b[14] = data27;
  848. b[15] = ONE;
  849. #else
  850. data01 = *(ao1 + 0);
  851. data09 = *(ao2 + 0);
  852. data10 = *(ao2 + 1);
  853. data17 = *(ao3 + 0);
  854. data18 = *(ao3 + 1);
  855. data19 = *(ao3 + 2);
  856. data25 = *(ao4 + 0);
  857. data26 = *(ao4 + 1);
  858. data27 = *(ao4 + 2);
  859. data28 = *(ao4 + 3);
  860. b[ 0] = data01;
  861. b[ 1] = ZERO;
  862. b[ 2] = ZERO;
  863. b[ 3] = ZERO;
  864. b[ 4] = data09;
  865. b[ 5] = data10;
  866. b[ 6] = ZERO;
  867. b[ 7] = ZERO;
  868. b[ 8] = data17;
  869. b[ 9] = data18;
  870. b[10] = data19;
  871. b[11] = ZERO;
  872. b[12] = data25;
  873. b[13] = data26;
  874. b[14] = data27;
  875. b[15] = data28;
  876. #endif
  877. ao1 += 4 * lda;
  878. ao2 += 4 * lda;
  879. ao3 += 4 * lda;
  880. ao4 += 4 * lda;
  881. b += 16;
  882. }
  883. X += 4;
  884. i --;
  885. } while (i > 0);
  886. }
  887. i = (m & 3);
  888. if (i) {
  889. if (X < posY) {
  890. if (m & 2) {
  891. ao1 += 2;
  892. ao2 += 2;
  893. ao3 += 2;
  894. ao4 += 2;
  895. b += 8;
  896. }
  897. if (m & 1) {
  898. b += 4;
  899. }
  900. } else
  901. if (X > posY) {
  902. if (m & 2) {
  903. data01 = *(ao1 + 0);
  904. data02 = *(ao1 + 1);
  905. data03 = *(ao1 + 2);
  906. data04 = *(ao1 + 3);
  907. data09 = *(ao2 + 0);
  908. data10 = *(ao2 + 1);
  909. data11 = *(ao2 + 2);
  910. data12 = *(ao2 + 3);
  911. b[ 0] = data01;
  912. b[ 1] = data02;
  913. b[ 2] = data03;
  914. b[ 3] = data04;
  915. b[ 4] = data09;
  916. b[ 5] = data10;
  917. b[ 6] = data11;
  918. b[ 7] = data12;
  919. ao1 += 2 * lda;
  920. b += 8;
  921. }
  922. if (m & 1) {
  923. data01 = *(ao1 + 0);
  924. data02 = *(ao1 + 1);
  925. data03 = *(ao1 + 2);
  926. data04 = *(ao1 + 3);
  927. b[ 0] = data01;
  928. b[ 1] = data02;
  929. b[ 2] = data03;
  930. b[ 3] = data04;
  931. b += 4;
  932. }
  933. } else {
  934. #ifndef UNIT
  935. data01 = *(ao1 + 0);
  936. #endif
  937. data09 = *(ao2 + 0);
  938. data17 = *(ao3 + 0);
  939. data25 = *(ao4 + 0);
  940. if (i >= 2) {
  941. #ifndef UNIT
  942. data10 = *(ao2 + 1);
  943. #endif
  944. data18 = *(ao3 + 1);
  945. data26 = *(ao4 + 1);
  946. }
  947. if (i >= 3) {
  948. #ifndef UNIT
  949. data19 = *(ao3 + 2);
  950. #endif
  951. data27 = *(ao4 + 2);
  952. }
  953. #ifndef UNIT
  954. b[ 0] = ONE;
  955. #else
  956. b[ 0] = data01;
  957. #endif
  958. b[ 1] = ZERO;
  959. b[ 2] = ZERO;
  960. b[ 3] = ZERO;
  961. b += 4;
  962. if(i >= 2) {
  963. b[ 0] = data09;
  964. #ifndef UNIT
  965. b[ 1] = ONE;
  966. #else
  967. b[ 1] = data10;
  968. #endif
  969. b[ 2] = ZERO;
  970. b[ 3] = ZERO;
  971. b += 4;
  972. }
  973. if (i >= 3) {
  974. b[ 0] = data17;
  975. b[ 1] = data18;
  976. #ifndef UNIT
  977. b[ 2] = ONE;
  978. #else
  979. b[ 2] = data19;
  980. #endif
  981. b[ 3] = ZERO;
  982. b += 4;
  983. }
  984. }
  985. }
  986. posY += 4;
  987. }
  988. if (n & 2){
  989. X = posX;
  990. if (posX <= posY) {
  991. ao1 = a + posX + (posY + 0) * lda;
  992. ao2 = a + posX + (posY + 1) * lda;
  993. } else {
  994. ao1 = a + posY + (posX + 0) * lda;
  995. ao2 = a + posY + (posX + 1) * lda;
  996. }
  997. i = (m >> 1);
  998. if (i > 0) {
  999. do {
  1000. if (X < posY) {
  1001. ao1 += 2;
  1002. ao2 += 2;
  1003. b += 4;
  1004. } else
  1005. if (X > posY) {
  1006. data01 = *(ao1 + 0);
  1007. data02 = *(ao1 + 1);
  1008. data09 = *(ao2 + 0);
  1009. data10 = *(ao2 + 1);
  1010. ao1 += 2 * lda;
  1011. ao2 += 2 * lda;
  1012. b[ 0] = data01;
  1013. b[ 1] = data02;
  1014. b[ 2] = data09;
  1015. b[ 3] = data10;
  1016. b += 4;
  1017. } else {
  1018. #ifdef UNIT
  1019. data09 = *(ao2 + 0);
  1020. b[ 0] = ONE;
  1021. b[ 1] = ZERO;
  1022. b[ 2] = data09;
  1023. b[ 3] = ONE;
  1024. #else
  1025. data01 = *(ao1 + 0);
  1026. data09 = *(ao2 + 0);
  1027. data10 = *(ao2 + 1);
  1028. b[ 0] = data01;
  1029. b[ 1] = ZERO;
  1030. b[ 2] = data09;
  1031. b[ 3] = data10;
  1032. #endif
  1033. ao1 += 2 * lda;
  1034. ao2 += 2 * lda;
  1035. b += 4;
  1036. }
  1037. X += 2;
  1038. i --;
  1039. } while (i > 0);
  1040. }
  1041. if (m & 1) {
  1042. if (X < posY) {
  1043. b += 2;
  1044. } else
  1045. if (X > posY) {
  1046. data01 = *(ao1 + 0);
  1047. data02 = *(ao1 + 1);
  1048. b[ 0] = data01;
  1049. b[ 1] = data02;
  1050. b += 2;
  1051. } else {
  1052. #ifdef UNIT
  1053. data09 = *(ao2 + 0);
  1054. b[ 0] = ONE;
  1055. b[ 1] = data09;
  1056. #else
  1057. data01 = *(ao1 + 0);
  1058. data09 = *(ao2 + 0);
  1059. b[ 0] = data01;
  1060. b[ 1] = data09;
  1061. #endif
  1062. b += 2;
  1063. }
  1064. }
  1065. posY += 2;
  1066. }
  1067. if (n & 1){
  1068. X = posX;
  1069. if (posX <= posY) {
  1070. ao1 = a + posX + (posY + 0) * lda;
  1071. } else {
  1072. ao1 = a + posY + (posX + 0) * lda;
  1073. }
  1074. i = m;
  1075. if (m > 0) {
  1076. do {
  1077. if (X < posY) {
  1078. ao1 += 1;
  1079. b += 1;
  1080. } else
  1081. if (X > posY) {
  1082. data01 = *(ao1 + 0);
  1083. ao1 += lda;
  1084. b[ 0] = data01;
  1085. b += 1;
  1086. } else {
  1087. #ifdef UNIT
  1088. b[ 0] = ONE;
  1089. #else
  1090. data01 = *(ao1 + 0);
  1091. b[ 0] = data01;
  1092. #endif
  1093. ao1 += lda;
  1094. b += 1;
  1095. }
  1096. X += 1;
  1097. i --;
  1098. } while (i > 0);
  1099. }
  1100. }
  1101. return 0;
  1102. }