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_ltcopy_8.c 24 kB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219
  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 + posY + (posX + 0) * lda;
  58. ao2 = a + posY + (posX + 1) * lda;
  59. ao3 = a + posY + (posX + 2) * lda;
  60. ao4 = a + posY + (posX + 3) * lda;
  61. ao5 = a + posY + (posX + 4) * lda;
  62. ao6 = a + posY + (posX + 5) * lda;
  63. ao7 = a + posY + (posX + 6) * lda;
  64. ao8 = a + posY + (posX + 7) * lda;
  65. } else {
  66. ao1 = a + posX + (posY + 0) * lda;
  67. ao2 = a + posX + (posY + 1) * lda;
  68. ao3 = a + posX + (posY + 2) * lda;
  69. ao4 = a + posX + (posY + 3) * lda;
  70. ao5 = a + posX + (posY + 4) * lda;
  71. ao6 = a + posX + (posY + 5) * lda;
  72. ao7 = a + posX + (posY + 6) * lda;
  73. ao8 = a + posX + (posY + 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. #ifndef UNIT
  229. data01 = *(ao1 + 0);
  230. #endif
  231. data02 = *(ao1 + 1);
  232. data03 = *(ao1 + 2);
  233. data04 = *(ao1 + 3);
  234. data05 = *(ao1 + 4);
  235. data06 = *(ao1 + 5);
  236. data07 = *(ao1 + 6);
  237. data08 = *(ao1 + 7);
  238. #ifndef UNIT
  239. data10 = *(ao2 + 1);
  240. #endif
  241. data11 = *(ao2 + 2);
  242. data12 = *(ao2 + 3);
  243. data13 = *(ao2 + 4);
  244. data14 = *(ao2 + 5);
  245. data15 = *(ao2 + 6);
  246. data16 = *(ao2 + 7);
  247. #ifndef UNIT
  248. data19 = *(ao3 + 2);
  249. #endif
  250. data20 = *(ao3 + 3);
  251. data21 = *(ao3 + 4);
  252. data22 = *(ao3 + 5);
  253. data23 = *(ao3 + 6);
  254. data24 = *(ao3 + 7);
  255. #ifndef UNIT
  256. data28 = *(ao4 + 3);
  257. #endif
  258. data29 = *(ao4 + 4);
  259. data30 = *(ao4 + 5);
  260. data31 = *(ao4 + 6);
  261. data32 = *(ao4 + 7);
  262. #ifndef UNIT
  263. data37 = *(ao5 + 4);
  264. #endif
  265. data38 = *(ao5 + 5);
  266. data39 = *(ao5 + 6);
  267. data40 = *(ao5 + 7);
  268. #ifndef UNIT
  269. data46 = *(ao6 + 5);
  270. #endif
  271. data47 = *(ao6 + 6);
  272. data48 = *(ao6 + 7);
  273. #ifndef UNIT
  274. data55 = *(ao7 + 6);
  275. #endif
  276. data56 = *(ao7 + 7);
  277. #ifndef UNIT
  278. data64 = *(ao8 + 7);
  279. #endif
  280. #ifdef UNIT
  281. b[ 0] = ONE;
  282. #else
  283. b[ 0] = data01;
  284. #endif
  285. b[ 1] = data02;
  286. b[ 2] = data03;
  287. b[ 3] = data04;
  288. b[ 4] = data05;
  289. b[ 5] = data06;
  290. b[ 6] = data07;
  291. b[ 7] = data08;
  292. b[ 8] = ZERO;
  293. #ifdef UNIT
  294. b[ 9] = ONE;
  295. #else
  296. b[ 9] = data10;
  297. #endif
  298. b[10] = data11;
  299. b[11] = data12;
  300. b[12] = data13;
  301. b[13] = data14;
  302. b[14] = data15;
  303. b[15] = data16;
  304. b[16] = ZERO;
  305. b[17] = ZERO;
  306. #ifdef UNIT
  307. b[18] = ONE;
  308. #else
  309. b[18] = data19;
  310. #endif
  311. b[19] = data20;
  312. b[20] = data21;
  313. b[21] = data22;
  314. b[22] = data23;
  315. b[23] = data24;
  316. b[24] = ZERO;
  317. b[25] = ZERO;
  318. b[26] = ZERO;
  319. #ifdef UNIT
  320. b[27] = ONE;
  321. #else
  322. b[27] = data28;
  323. #endif
  324. b[28] = data29;
  325. b[29] = data30;
  326. b[30] = data31;
  327. b[31] = data32;
  328. b[32] = ZERO;
  329. b[33] = ZERO;
  330. b[34] = ZERO;
  331. b[35] = ZERO;
  332. #ifdef UNIT
  333. b[36] = ONE;
  334. #else
  335. b[36] = data37;
  336. #endif
  337. b[37] = data38;
  338. b[38] = data39;
  339. b[39] = data40;
  340. b[40] = ZERO;
  341. b[41] = ZERO;
  342. b[42] = ZERO;
  343. b[43] = ZERO;
  344. b[44] = ZERO;
  345. #ifdef UNIT
  346. b[45] = ONE;
  347. #else
  348. b[45] = data46;
  349. #endif
  350. b[46] = data47;
  351. b[47] = data48;
  352. b[48] = ZERO;
  353. b[49] = ZERO;
  354. b[50] = ZERO;
  355. b[51] = ZERO;
  356. b[52] = ZERO;
  357. b[53] = ZERO;
  358. #ifdef UNIT
  359. b[54] = ONE;
  360. #else
  361. b[54] = data55;
  362. #endif
  363. b[55] = data56;
  364. b[56] = ZERO;
  365. b[57] = ZERO;
  366. b[58] = ZERO;
  367. b[59] = ZERO;
  368. b[60] = ZERO;
  369. b[61] = ZERO;
  370. b[62] = ZERO;
  371. #ifdef UNIT
  372. b[63] = ONE;
  373. #else
  374. b[63] = data64;
  375. #endif
  376. ao1 += 8;
  377. ao2 += 8;
  378. ao3 += 8;
  379. ao4 += 8;
  380. ao5 += 8;
  381. ao6 += 8;
  382. ao7 += 8;
  383. ao8 += 8;
  384. b += 64;
  385. }
  386. X += 8;
  387. i --;
  388. } while (i > 0);
  389. }
  390. i = (m & 7);
  391. if (i) {
  392. if (X > posY) {
  393. if (m & 4) {
  394. ao1 += 4;
  395. ao2 += 4;
  396. ao3 += 4;
  397. ao4 += 4;
  398. ao5 += 4;
  399. ao6 += 4;
  400. ao7 += 4;
  401. ao8 += 4;
  402. b += 32;
  403. }
  404. if (m & 2) {
  405. ao1 += 2;
  406. ao2 += 2;
  407. ao3 += 2;
  408. ao4 += 2;
  409. ao5 += 2;
  410. ao6 += 2;
  411. ao7 += 2;
  412. ao8 += 2;
  413. b += 16;
  414. }
  415. if (m & 1) {
  416. b += 8;
  417. }
  418. } else
  419. if (X < posY) {
  420. if (m & 4) {
  421. data01 = *(ao1 + 0);
  422. data02 = *(ao1 + 1);
  423. data03 = *(ao1 + 2);
  424. data04 = *(ao1 + 3);
  425. data05 = *(ao1 + 4);
  426. data06 = *(ao1 + 5);
  427. data07 = *(ao1 + 6);
  428. data08 = *(ao1 + 7);
  429. data09 = *(ao2 + 0);
  430. data10 = *(ao2 + 1);
  431. data11 = *(ao2 + 2);
  432. data12 = *(ao2 + 3);
  433. data13 = *(ao2 + 4);
  434. data14 = *(ao2 + 5);
  435. data15 = *(ao2 + 6);
  436. data16 = *(ao2 + 7);
  437. data17 = *(ao3 + 0);
  438. data18 = *(ao3 + 1);
  439. data19 = *(ao3 + 2);
  440. data20 = *(ao3 + 3);
  441. data21 = *(ao3 + 4);
  442. data22 = *(ao3 + 5);
  443. data23 = *(ao3 + 6);
  444. data24 = *(ao3 + 7);
  445. data25 = *(ao4 + 0);
  446. data26 = *(ao4 + 1);
  447. data27 = *(ao4 + 2);
  448. data28 = *(ao4 + 3);
  449. data29 = *(ao4 + 4);
  450. data30 = *(ao4 + 5);
  451. data31 = *(ao4 + 6);
  452. data32 = *(ao4 + 7);
  453. b[ 0] = data01;
  454. b[ 1] = data02;
  455. b[ 2] = data03;
  456. b[ 3] = data04;
  457. b[ 4] = data05;
  458. b[ 5] = data06;
  459. b[ 6] = data07;
  460. b[ 7] = data08;
  461. b[ 8] = data09;
  462. b[ 9] = data10;
  463. b[10] = data11;
  464. b[11] = data12;
  465. b[12] = data13;
  466. b[13] = data14;
  467. b[14] = data15;
  468. b[15] = data16;
  469. b[16] = data17;
  470. b[17] = data18;
  471. b[18] = data19;
  472. b[19] = data20;
  473. b[20] = data21;
  474. b[21] = data22;
  475. b[22] = data23;
  476. b[23] = data24;
  477. b[24] = data25;
  478. b[25] = data26;
  479. b[26] = data27;
  480. b[27] = data28;
  481. b[28] = data29;
  482. b[29] = data30;
  483. b[30] = data31;
  484. b[31] = data32;
  485. ao1 += 4 * lda;
  486. ao2 += 4 * lda;
  487. ao3 += 4 * lda;
  488. ao4 += 4 * lda;
  489. b += 32;
  490. }
  491. if (m & 2) {
  492. data01 = *(ao1 + 0);
  493. data02 = *(ao1 + 1);
  494. data03 = *(ao1 + 2);
  495. data04 = *(ao1 + 3);
  496. data05 = *(ao1 + 4);
  497. data06 = *(ao1 + 5);
  498. data07 = *(ao1 + 6);
  499. data08 = *(ao1 + 7);
  500. data09 = *(ao2 + 0);
  501. data10 = *(ao2 + 1);
  502. data11 = *(ao2 + 2);
  503. data12 = *(ao2 + 3);
  504. data13 = *(ao2 + 4);
  505. data14 = *(ao2 + 5);
  506. data15 = *(ao2 + 6);
  507. data16 = *(ao2 + 7);
  508. b[ 0] = data01;
  509. b[ 1] = data02;
  510. b[ 2] = data03;
  511. b[ 3] = data04;
  512. b[ 4] = data05;
  513. b[ 5] = data06;
  514. b[ 6] = data07;
  515. b[ 7] = data08;
  516. b[ 8] = data09;
  517. b[ 9] = data10;
  518. b[10] = data11;
  519. b[11] = data12;
  520. b[12] = data13;
  521. b[13] = data14;
  522. b[14] = data15;
  523. b[15] = data16;
  524. ao1 += 2 * lda;
  525. b += 16;
  526. }
  527. if (m & 1) {
  528. data01 = *(ao1 + 0);
  529. data02 = *(ao1 + 1);
  530. data03 = *(ao1 + 2);
  531. data04 = *(ao1 + 3);
  532. data05 = *(ao1 + 4);
  533. data06 = *(ao1 + 5);
  534. data07 = *(ao1 + 6);
  535. data08 = *(ao1 + 7);
  536. b[ 0] = data01;
  537. b[ 1] = data02;
  538. b[ 2] = data03;
  539. b[ 3] = data04;
  540. b[ 4] = data05;
  541. b[ 5] = data06;
  542. b[ 6] = data07;
  543. b[ 7] = data08;
  544. b += 8;
  545. }
  546. } else {
  547. #ifndef UNIT
  548. data01 = *(ao1 + 0);
  549. #endif
  550. data02 = *(ao1 + 1);
  551. data03 = *(ao1 + 2);
  552. data04 = *(ao1 + 3);
  553. data05 = *(ao1 + 4);
  554. data06 = *(ao1 + 5);
  555. data07 = *(ao1 + 6);
  556. data08 = *(ao1 + 7);
  557. if (i >= 2) {
  558. #ifndef UNIT
  559. data10 = *(ao2 + 1);
  560. #endif
  561. data11 = *(ao2 + 2);
  562. data12 = *(ao2 + 3);
  563. data13 = *(ao2 + 4);
  564. data14 = *(ao2 + 5);
  565. data15 = *(ao2 + 6);
  566. data16 = *(ao2 + 7);
  567. }
  568. if (i >= 3) {
  569. #ifndef UNIT
  570. data19 = *(ao3 + 2);
  571. #endif
  572. data20 = *(ao3 + 3);
  573. data21 = *(ao3 + 4);
  574. data22 = *(ao3 + 5);
  575. data23 = *(ao3 + 6);
  576. data24 = *(ao3 + 7);
  577. }
  578. if (i >= 4) {
  579. #ifndef UNIT
  580. data28 = *(ao4 + 3);
  581. #endif
  582. data29 = *(ao4 + 4);
  583. data30 = *(ao4 + 5);
  584. data31 = *(ao4 + 6);
  585. data32 = *(ao4 + 7);
  586. }
  587. if (i >= 5) {
  588. #ifndef UNIT
  589. data37 = *(ao5 + 4);
  590. #endif
  591. data38 = *(ao5 + 5);
  592. data39 = *(ao5 + 6);
  593. data40 = *(ao5 + 7);
  594. }
  595. if (i >= 6) {
  596. #ifndef UNIT
  597. data46 = *(ao6 + 5);
  598. #endif
  599. data47 = *(ao6 + 6);
  600. data48 = *(ao6 + 7);
  601. }
  602. if (i >= 7) {
  603. #ifndef UNIT
  604. data55 = *(ao7 + 6);
  605. #endif
  606. data56 = *(ao7 + 7);
  607. }
  608. #ifdef UNIT
  609. b[ 0] = ONE;
  610. #else
  611. b[ 0] = data01;
  612. #endif
  613. b[ 1] = data02;
  614. b[ 2] = data03;
  615. b[ 3] = data04;
  616. b[ 4] = data05;
  617. b[ 5] = data06;
  618. b[ 6] = data07;
  619. b[ 7] = data08;
  620. b += 8;
  621. if(i >= 2) {
  622. b[ 0] = ZERO;
  623. #ifdef UNIT
  624. b[ 1] = ONE;
  625. #else
  626. b[ 1] = data10;
  627. #endif
  628. b[ 2] = data11;
  629. b[ 3] = data12;
  630. b[ 4] = data13;
  631. b[ 5] = data14;
  632. b[ 6] = data15;
  633. b[ 7] = data16;
  634. b += 8;
  635. }
  636. if (i >= 3) {
  637. b[ 0] = ZERO;
  638. b[ 1] = ZERO;
  639. #ifdef UNIT
  640. b[ 2] = ONE;
  641. #else
  642. b[ 2] = data19;
  643. #endif
  644. b[ 3] = data20;
  645. b[ 4] = data21;
  646. b[ 5] = data22;
  647. b[ 6] = data23;
  648. b[ 7] = data24;
  649. b += 8;
  650. }
  651. if (i >= 4) {
  652. b[ 0] = ZERO;
  653. b[ 1] = ZERO;
  654. b[ 2] = ZERO;
  655. #ifdef UNIT
  656. b[ 3] = ONE;
  657. #else
  658. b[ 3] = data28;
  659. #endif
  660. b[ 4] = data29;
  661. b[ 5] = data30;
  662. b[ 6] = data31;
  663. b[ 7] = data32;
  664. b += 8;
  665. }
  666. if (i >= 5) {
  667. b[ 0] = ZERO;
  668. b[ 1] = ZERO;
  669. b[ 2] = ZERO;
  670. b[ 3] = ZERO;
  671. #ifdef UNIT
  672. b[ 4] = ONE;
  673. #else
  674. b[ 4] = data37;
  675. #endif
  676. b[ 5] = data38;
  677. b[ 6] = data39;
  678. b[ 7] = data40;
  679. b += 8;
  680. }
  681. if (i >= 6) {
  682. b[ 0] = ZERO;
  683. b[ 1] = ZERO;
  684. b[ 2] = ZERO;
  685. b[ 3] = ZERO;
  686. b[ 4] = ZERO;
  687. #ifdef UNIT
  688. b[ 5] = ONE;
  689. #else
  690. b[ 5] = data46;
  691. #endif
  692. b[ 6] = data47;
  693. b[ 7] = data48;
  694. b += 8;
  695. }
  696. if (i >= 7) {
  697. b[ 0] = ZERO;
  698. b[ 1] = ZERO;
  699. b[ 2] = ZERO;
  700. b[ 3] = ZERO;
  701. b[ 4] = ZERO;
  702. b[ 5] = ZERO;
  703. #ifdef UNIT
  704. b[ 6] = ONE;
  705. #else
  706. b[ 6] = data55;
  707. #endif
  708. b[ 7] = data56;
  709. b += 8;
  710. }
  711. }
  712. }
  713. posY += 8;
  714. js --;
  715. } while (js > 0);
  716. } /* End of main loop */
  717. if (n & 4){
  718. X = posX;
  719. if (posX <= posY) {
  720. ao1 = a + posY + (posX + 0) * lda;
  721. ao2 = a + posY + (posX + 1) * lda;
  722. ao3 = a + posY + (posX + 2) * lda;
  723. ao4 = a + posY + (posX + 3) * lda;
  724. } else {
  725. ao1 = a + posX + (posY + 0) * lda;
  726. ao2 = a + posX + (posY + 1) * lda;
  727. ao3 = a + posX + (posY + 2) * lda;
  728. ao4 = a + posX + (posY + 3) * lda;
  729. }
  730. i = (m >> 2);
  731. if (i > 0) {
  732. do {
  733. if (X > posY) {
  734. ao1 += 4;
  735. ao2 += 4;
  736. ao3 += 4;
  737. ao4 += 4;
  738. b += 16;
  739. } else
  740. if (X < posY) {
  741. data01 = *(ao1 + 0);
  742. data02 = *(ao1 + 1);
  743. data03 = *(ao1 + 2);
  744. data04 = *(ao1 + 3);
  745. data09 = *(ao2 + 0);
  746. data10 = *(ao2 + 1);
  747. data11 = *(ao2 + 2);
  748. data12 = *(ao2 + 3);
  749. data17 = *(ao3 + 0);
  750. data18 = *(ao3 + 1);
  751. data19 = *(ao3 + 2);
  752. data20 = *(ao3 + 3);
  753. data25 = *(ao4 + 0);
  754. data26 = *(ao4 + 1);
  755. data27 = *(ao4 + 2);
  756. data28 = *(ao4 + 3);
  757. ao1 += 4 * lda;
  758. ao2 += 4 * lda;
  759. ao3 += 4 * lda;
  760. ao4 += 4 * lda;
  761. b[ 0] = data01;
  762. b[ 1] = data02;
  763. b[ 2] = data03;
  764. b[ 3] = data04;
  765. b[ 4] = data09;
  766. b[ 5] = data10;
  767. b[ 6] = data11;
  768. b[ 7] = data12;
  769. b[ 8] = data17;
  770. b[ 9] = data18;
  771. b[10] = data19;
  772. b[11] = data20;
  773. b[12] = data25;
  774. b[13] = data26;
  775. b[14] = data27;
  776. b[15] = data28;
  777. b += 16;
  778. } else {
  779. #ifdef UNIT
  780. data02 = *(ao1 + 1);
  781. data03 = *(ao1 + 2);
  782. data04 = *(ao1 + 3);
  783. data11 = *(ao2 + 2);
  784. data12 = *(ao2 + 3);
  785. data20 = *(ao3 + 3);
  786. b[ 0] = ONE;
  787. b[ 1] = data02;
  788. b[ 2] = data03;
  789. b[ 3] = data04;
  790. b[ 4] = ZERO;
  791. b[ 5] = ONE;
  792. b[ 6] = data11;
  793. b[ 7] = data12;
  794. b[ 8] = ZERO;
  795. b[ 9] = ZERO;
  796. b[10] = ONE;
  797. b[11] = data20;
  798. b[12] = ZERO;
  799. b[13] = ZERO;
  800. b[14] = ZERO;
  801. b[15] = ONE;
  802. #else
  803. data01 = *(ao1 + 0);
  804. data02 = *(ao1 + 1);
  805. data03 = *(ao1 + 2);
  806. data04 = *(ao1 + 3);
  807. data10 = *(ao2 + 1);
  808. data11 = *(ao2 + 2);
  809. data12 = *(ao2 + 3);
  810. data19 = *(ao3 + 2);
  811. data20 = *(ao3 + 3);
  812. data28 = *(ao4 + 3);
  813. b[ 0] = data01;
  814. b[ 1] = data02;
  815. b[ 2] = data03;
  816. b[ 3] = data04;
  817. b[ 4] = ZERO;
  818. b[ 5] = data10;
  819. b[ 6] = data11;
  820. b[ 7] = data12;
  821. b[ 8] = ZERO;
  822. b[ 9] = ZERO;
  823. b[10] = data19;
  824. b[11] = data20;
  825. b[12] = ZERO;
  826. b[13] = ZERO;
  827. b[14] = ZERO;
  828. b[15] = data28;
  829. #endif
  830. ao1 += 4;
  831. ao2 += 4;
  832. ao3 += 4;
  833. ao4 += 4;
  834. b += 16;
  835. }
  836. X += 4;
  837. i --;
  838. } while (i > 0);
  839. }
  840. i = (m & 3);
  841. if (i) {
  842. if (X > posY) {
  843. if (m & 2) {
  844. ao1 += 2;
  845. ao2 += 2;
  846. ao3 += 2;
  847. ao4 += 2;
  848. b += 8;
  849. }
  850. if (m & 1) {
  851. b += 4;
  852. }
  853. } else
  854. if (X < posY) {
  855. if (m & 2) {
  856. data01 = *(ao1 + 0);
  857. data02 = *(ao1 + 1);
  858. data03 = *(ao1 + 2);
  859. data04 = *(ao1 + 3);
  860. data09 = *(ao2 + 0);
  861. data10 = *(ao2 + 1);
  862. data11 = *(ao2 + 2);
  863. data12 = *(ao2 + 3);
  864. b[ 0] = data01;
  865. b[ 1] = data02;
  866. b[ 2] = data03;
  867. b[ 3] = data04;
  868. b[ 4] = data09;
  869. b[ 5] = data10;
  870. b[ 6] = data11;
  871. b[ 7] = data12;
  872. ao1 += 2 * lda;
  873. b += 8;
  874. }
  875. if (m & 1) {
  876. data01 = *(ao1 + 0);
  877. data02 = *(ao1 + 1);
  878. data03 = *(ao1 + 2);
  879. data04 = *(ao1 + 3);
  880. b[ 0] = data01;
  881. b[ 1] = data02;
  882. b[ 2] = data03;
  883. b[ 3] = data04;
  884. b += 4;
  885. }
  886. } else {
  887. #ifndef UNIT
  888. data01 = *(ao1 + 0);
  889. #endif
  890. data02 = *(ao1 + 1);
  891. data03 = *(ao1 + 2);
  892. data04 = *(ao1 + 3);
  893. if (i >= 2) {
  894. #ifndef UNIT
  895. data10 = *(ao2 + 1);
  896. #endif
  897. data11 = *(ao2 + 2);
  898. data12 = *(ao2 + 3);
  899. }
  900. if (i >= 3) {
  901. #ifndef UNIT
  902. data19 = *(ao3 + 2);
  903. #endif
  904. data20 = *(ao3 + 3);
  905. }
  906. #ifdef UNIT
  907. b[ 0] = ONE;
  908. #else
  909. b[ 0] = data01;
  910. #endif
  911. b[ 1] = data02;
  912. b[ 2] = data03;
  913. b[ 3] = data04;
  914. b += 4;
  915. if(i >= 2) {
  916. b[ 0] = ZERO;
  917. #ifdef UNIT
  918. b[ 1] = ONE;
  919. #else
  920. b[ 1] = data10;
  921. #endif
  922. b[ 2] = data11;
  923. b[ 3] = data12;
  924. b += 4;
  925. }
  926. if (i >= 3) {
  927. b[ 0] = ZERO;
  928. b[ 1] = ZERO;
  929. #ifdef UNIT
  930. b[ 2] = ONE;
  931. #else
  932. b[ 2] = data19;
  933. #endif
  934. b[ 3] = data20;
  935. b += 4;
  936. }
  937. }
  938. }
  939. posY += 4;
  940. }
  941. if (n & 2){
  942. X = posX;
  943. if (posX <= posY) {
  944. ao1 = a + posY + (posX + 0) * lda;
  945. ao2 = a + posY + (posX + 1) * lda;
  946. } else {
  947. ao1 = a + posX + (posY + 0) * lda;
  948. ao2 = a + posX + (posY + 1) * lda;
  949. }
  950. i = (m >> 1);
  951. if (i > 0) {
  952. do {
  953. if (X > posY) {
  954. ao1 += 2;
  955. ao2 += 2;
  956. b += 4;
  957. } else
  958. if (X < posY) {
  959. data01 = *(ao1 + 0);
  960. data02 = *(ao1 + 1);
  961. data09 = *(ao2 + 0);
  962. data10 = *(ao2 + 1);
  963. ao1 += 2 * lda;
  964. ao2 += 2 * lda;
  965. b[ 0] = data01;
  966. b[ 1] = data02;
  967. b[ 2] = data09;
  968. b[ 3] = data10;
  969. b += 4;
  970. } else {
  971. #ifdef UNIT
  972. data02 = *(ao1 + 1);
  973. b[ 0] = ONE;
  974. b[ 1] = data02;
  975. b[ 2] = ZERO;
  976. b[ 3] = ONE;
  977. #else
  978. data01 = *(ao1 + 0);
  979. data02 = *(ao1 + 1);
  980. data10 = *(ao2 + 1);
  981. b[ 0] = data01;
  982. b[ 1] = data02;
  983. b[ 2] = ZERO;
  984. b[ 3] = data10;
  985. #endif
  986. ao1 += 2;
  987. ao2 += 2;
  988. b += 4;
  989. }
  990. X += 2;
  991. i --;
  992. } while (i > 0);
  993. }
  994. if (m & 1) {
  995. if (X > posY) {
  996. b += 2;
  997. } else
  998. if (X < posY) {
  999. data01 = *(ao1 + 0);
  1000. data02 = *(ao1 + 1);
  1001. b[ 0] = data01;
  1002. b[ 1] = data02;
  1003. b += 2;
  1004. } else {
  1005. #ifdef UNIT
  1006. data09 = *(ao2 + 0);
  1007. b[ 0] = ONE;
  1008. b[ 1] = data09;
  1009. #else
  1010. data01 = *(ao1 + 0);
  1011. data09 = *(ao2 + 0);
  1012. b[ 0] = data01;
  1013. b[ 1] = data09;
  1014. #endif
  1015. b += 2;
  1016. }
  1017. }
  1018. posY += 2;
  1019. }
  1020. if (n & 1){
  1021. X = posX;
  1022. if (posX <= posY) {
  1023. ao1 = a + posY + (posX + 0) * lda;
  1024. } else {
  1025. ao1 = a + posX + (posY + 0) * lda;
  1026. }
  1027. i = m;
  1028. if (m > 0) {
  1029. do {
  1030. if (X > posY) {
  1031. ao1 += 1;
  1032. b += 1;
  1033. } else
  1034. if (X < posY) {
  1035. data01 = *(ao1 + 0);
  1036. ao1 += lda;
  1037. b[ 0] = data01;
  1038. b += 1;
  1039. } else {
  1040. #ifdef UNIT
  1041. b[ 0] = ONE;
  1042. #else
  1043. data01 = *(ao1 + 0);
  1044. b[ 0] = data01;
  1045. #endif
  1046. ao1 ++;
  1047. b ++;
  1048. }
  1049. X += 1;
  1050. i --;
  1051. } while (i > 0);
  1052. }
  1053. }
  1054. return 0;
  1055. }