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_uncopy_8.c 24 kB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226
  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. data01 = *(ao1 + 0);
  80. data02 = *(ao1 + 1);
  81. data03 = *(ao1 + 2);
  82. data04 = *(ao1 + 3);
  83. data05 = *(ao1 + 4);
  84. data06 = *(ao1 + 5);
  85. data07 = *(ao1 + 6);
  86. data08 = *(ao1 + 7);
  87. data09 = *(ao2 + 0);
  88. data10 = *(ao2 + 1);
  89. data11 = *(ao2 + 2);
  90. data12 = *(ao2 + 3);
  91. data13 = *(ao2 + 4);
  92. data14 = *(ao2 + 5);
  93. data15 = *(ao2 + 6);
  94. data16 = *(ao2 + 7);
  95. data17 = *(ao3 + 0);
  96. data18 = *(ao3 + 1);
  97. data19 = *(ao3 + 2);
  98. data20 = *(ao3 + 3);
  99. data21 = *(ao3 + 4);
  100. data22 = *(ao3 + 5);
  101. data23 = *(ao3 + 6);
  102. data24 = *(ao3 + 7);
  103. data25 = *(ao4 + 0);
  104. data26 = *(ao4 + 1);
  105. data27 = *(ao4 + 2);
  106. data28 = *(ao4 + 3);
  107. data29 = *(ao4 + 4);
  108. data30 = *(ao4 + 5);
  109. data31 = *(ao4 + 6);
  110. data32 = *(ao4 + 7);
  111. data33 = *(ao5 + 0);
  112. data34 = *(ao5 + 1);
  113. data35 = *(ao5 + 2);
  114. data36 = *(ao5 + 3);
  115. data37 = *(ao5 + 4);
  116. data38 = *(ao5 + 5);
  117. data39 = *(ao5 + 6);
  118. data40 = *(ao5 + 7);
  119. data41 = *(ao6 + 0);
  120. data42 = *(ao6 + 1);
  121. data43 = *(ao6 + 2);
  122. data44 = *(ao6 + 3);
  123. data45 = *(ao6 + 4);
  124. data46 = *(ao6 + 5);
  125. data47 = *(ao6 + 6);
  126. data48 = *(ao6 + 7);
  127. data49 = *(ao7 + 0);
  128. data50 = *(ao7 + 1);
  129. data51 = *(ao7 + 2);
  130. data52 = *(ao7 + 3);
  131. data53 = *(ao7 + 4);
  132. data54 = *(ao7 + 5);
  133. data55 = *(ao7 + 6);
  134. data56 = *(ao7 + 7);
  135. data57 = *(ao8 + 0);
  136. data58 = *(ao8 + 1);
  137. data59 = *(ao8 + 2);
  138. data60 = *(ao8 + 3);
  139. data61 = *(ao8 + 4);
  140. data62 = *(ao8 + 5);
  141. data63 = *(ao8 + 6);
  142. data64 = *(ao8 + 7);
  143. b[ 0] = data01;
  144. b[ 1] = data09;
  145. b[ 2] = data17;
  146. b[ 3] = data25;
  147. b[ 4] = data33;
  148. b[ 5] = data41;
  149. b[ 6] = data49;
  150. b[ 7] = data57;
  151. b[ 8] = data02;
  152. b[ 9] = data10;
  153. b[10] = data18;
  154. b[11] = data26;
  155. b[12] = data34;
  156. b[13] = data42;
  157. b[14] = data50;
  158. b[15] = data58;
  159. b[16] = data03;
  160. b[17] = data11;
  161. b[18] = data19;
  162. b[19] = data27;
  163. b[20] = data35;
  164. b[21] = data43;
  165. b[22] = data51;
  166. b[23] = data59;
  167. b[24] = data04;
  168. b[25] = data12;
  169. b[26] = data20;
  170. b[27] = data28;
  171. b[28] = data36;
  172. b[29] = data44;
  173. b[30] = data52;
  174. b[31] = data60;
  175. b[32] = data05;
  176. b[33] = data13;
  177. b[34] = data21;
  178. b[35] = data29;
  179. b[36] = data37;
  180. b[37] = data45;
  181. b[38] = data53;
  182. b[39] = data61;
  183. b[40] = data06;
  184. b[41] = data14;
  185. b[42] = data22;
  186. b[43] = data30;
  187. b[44] = data38;
  188. b[45] = data46;
  189. b[46] = data54;
  190. b[47] = data62;
  191. b[48] = data07;
  192. b[49] = data15;
  193. b[50] = data23;
  194. b[51] = data31;
  195. b[52] = data39;
  196. b[53] = data47;
  197. b[54] = data55;
  198. b[55] = data63;
  199. b[56] = data08;
  200. b[57] = data16;
  201. b[58] = data24;
  202. b[59] = data32;
  203. b[60] = data40;
  204. b[61] = data48;
  205. b[62] = data56;
  206. b[63] = data64;
  207. ao1 += 8;
  208. ao2 += 8;
  209. ao3 += 8;
  210. ao4 += 8;
  211. ao5 += 8;
  212. ao6 += 8;
  213. ao7 += 8;
  214. ao8 += 8;
  215. b += 64;
  216. } else
  217. if (X > posY) {
  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. data09 = *(ao2 + 0);
  232. #ifndef UNIT
  233. data10 = *(ao2 + 1);
  234. #endif
  235. data17 = *(ao3 + 0);
  236. data18 = *(ao3 + 1);
  237. #ifndef UNIT
  238. data19 = *(ao3 + 2);
  239. #endif
  240. data25 = *(ao4 + 0);
  241. data26 = *(ao4 + 1);
  242. data27 = *(ao4 + 2);
  243. #ifndef UNIT
  244. data28 = *(ao4 + 3);
  245. #endif
  246. data33 = *(ao5 + 0);
  247. data34 = *(ao5 + 1);
  248. data35 = *(ao5 + 2);
  249. data36 = *(ao5 + 3);
  250. #ifndef UNIT
  251. data37 = *(ao5 + 4);
  252. #endif
  253. data41 = *(ao6 + 0);
  254. data42 = *(ao6 + 1);
  255. data43 = *(ao6 + 2);
  256. data44 = *(ao6 + 3);
  257. data45 = *(ao6 + 4);
  258. #ifndef UNIT
  259. data46 = *(ao6 + 5);
  260. #endif
  261. data49 = *(ao7 + 0);
  262. data50 = *(ao7 + 1);
  263. data51 = *(ao7 + 2);
  264. data52 = *(ao7 + 3);
  265. data53 = *(ao7 + 4);
  266. data54 = *(ao7 + 5);
  267. #ifndef UNIT
  268. data55 = *(ao7 + 6);
  269. #endif
  270. data57 = *(ao8 + 0);
  271. data58 = *(ao8 + 1);
  272. data59 = *(ao8 + 2);
  273. data60 = *(ao8 + 3);
  274. data61 = *(ao8 + 4);
  275. data62 = *(ao8 + 5);
  276. data63 = *(ao8 + 6);
  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] = data09;
  286. b[ 2] = data17;
  287. b[ 3] = data25;
  288. b[ 4] = data33;
  289. b[ 5] = data41;
  290. b[ 6] = data49;
  291. b[ 7] = data57;
  292. b[ 8] = ZERO;
  293. #ifdef UNIT
  294. b[ 9] = ONE;
  295. #else
  296. b[ 9] = data10;
  297. #endif
  298. b[10] = data18;
  299. b[11] = data26;
  300. b[12] = data34;
  301. b[13] = data42;
  302. b[14] = data50;
  303. b[15] = data58;
  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] = data27;
  312. b[20] = data35;
  313. b[21] = data43;
  314. b[22] = data51;
  315. b[23] = data59;
  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] = data36;
  325. b[29] = data44;
  326. b[30] = data52;
  327. b[31] = data60;
  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] = data45;
  338. b[38] = data53;
  339. b[39] = data61;
  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] = data54;
  351. b[47] = data62;
  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] = data63;
  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 * lda;
  377. ao2 += 8 * lda;
  378. ao3 += 8 * lda;
  379. ao4 += 8 * lda;
  380. ao5 += 8 * lda;
  381. ao6 += 8 * lda;
  382. ao7 += 8 * lda;
  383. ao8 += 8 * lda;
  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. data01 = *(ao1 + 0);
  395. data02 = *(ao1 + 1);
  396. data03 = *(ao1 + 2);
  397. data04 = *(ao1 + 3);
  398. data09 = *(ao2 + 0);
  399. data10 = *(ao2 + 1);
  400. data11 = *(ao2 + 2);
  401. data12 = *(ao2 + 3);
  402. data17 = *(ao3 + 0);
  403. data18 = *(ao3 + 1);
  404. data19 = *(ao3 + 2);
  405. data20 = *(ao3 + 3);
  406. data25 = *(ao4 + 0);
  407. data26 = *(ao4 + 1);
  408. data27 = *(ao4 + 2);
  409. data28 = *(ao4 + 3);
  410. data33 = *(ao5 + 0);
  411. data34 = *(ao5 + 1);
  412. data35 = *(ao5 + 2);
  413. data36 = *(ao5 + 3);
  414. data41 = *(ao6 + 0);
  415. data42 = *(ao6 + 1);
  416. data43 = *(ao6 + 2);
  417. data44 = *(ao6 + 3);
  418. data49 = *(ao7 + 0);
  419. data50 = *(ao7 + 1);
  420. data51 = *(ao7 + 2);
  421. data52 = *(ao7 + 3);
  422. data57 = *(ao8 + 0);
  423. data58 = *(ao8 + 1);
  424. data59 = *(ao8 + 2);
  425. data60 = *(ao8 + 3);
  426. b[ 0] = data01;
  427. b[ 1] = data09;
  428. b[ 2] = data17;
  429. b[ 3] = data25;
  430. b[ 4] = data33;
  431. b[ 5] = data41;
  432. b[ 6] = data49;
  433. b[ 7] = data57;
  434. b[ 8] = data02;
  435. b[ 9] = data10;
  436. b[10] = data18;
  437. b[11] = data26;
  438. b[12] = data34;
  439. b[13] = data42;
  440. b[14] = data50;
  441. b[15] = data58;
  442. b[16] = data03;
  443. b[17] = data11;
  444. b[18] = data19;
  445. b[19] = data27;
  446. b[20] = data35;
  447. b[21] = data43;
  448. b[22] = data51;
  449. b[23] = data59;
  450. b[24] = data04;
  451. b[25] = data12;
  452. b[26] = data20;
  453. b[27] = data28;
  454. b[28] = data36;
  455. b[29] = data44;
  456. b[30] = data52;
  457. b[31] = data60;
  458. ao1 += 4;
  459. ao2 += 4;
  460. ao3 += 4;
  461. ao4 += 4;
  462. ao5 += 4;
  463. ao6 += 4;
  464. ao7 += 4;
  465. ao8 += 4;
  466. b += 32;
  467. }
  468. if (m & 2) {
  469. data01 = *(ao1 + 0);
  470. data02 = *(ao1 + 1);
  471. data09 = *(ao2 + 0);
  472. data10 = *(ao2 + 1);
  473. data17 = *(ao3 + 0);
  474. data18 = *(ao3 + 1);
  475. data25 = *(ao4 + 0);
  476. data26 = *(ao4 + 1);
  477. data33 = *(ao5 + 0);
  478. data34 = *(ao5 + 1);
  479. data41 = *(ao6 + 0);
  480. data42 = *(ao6 + 1);
  481. data49 = *(ao7 + 0);
  482. data50 = *(ao7 + 1);
  483. data57 = *(ao8 + 0);
  484. data58 = *(ao8 + 1);
  485. b[ 0] = data01;
  486. b[ 1] = data09;
  487. b[ 2] = data17;
  488. b[ 3] = data25;
  489. b[ 4] = data33;
  490. b[ 5] = data41;
  491. b[ 6] = data49;
  492. b[ 7] = data57;
  493. b[ 8] = data02;
  494. b[ 9] = data10;
  495. b[10] = data18;
  496. b[11] = data26;
  497. b[12] = data34;
  498. b[13] = data42;
  499. b[14] = data50;
  500. b[15] = data58;
  501. ao1 += 2;
  502. ao2 += 2;
  503. ao3 += 2;
  504. ao4 += 2;
  505. ao5 += 2;
  506. ao6 += 2;
  507. ao7 += 2;
  508. ao8 += 2;
  509. b += 16;
  510. }
  511. if (m & 1) {
  512. data01 = *(ao1 + 0);
  513. data09 = *(ao2 + 0);
  514. data17 = *(ao3 + 0);
  515. data25 = *(ao4 + 0);
  516. data33 = *(ao5 + 0);
  517. data41 = *(ao6 + 0);
  518. data49 = *(ao7 + 0);
  519. data57 = *(ao8 + 0);
  520. b[ 0] = data01;
  521. b[ 1] = data09;
  522. b[ 2] = data17;
  523. b[ 3] = data25;
  524. b[ 4] = data33;
  525. b[ 5] = data41;
  526. b[ 6] = data49;
  527. b[ 7] = data57;
  528. b += 8;
  529. }
  530. } else
  531. if (X > posY) {
  532. if (m & 4) {
  533. ao1 += 4 * lda;
  534. ao2 += 4 * lda;
  535. ao3 += 4 * lda;
  536. ao4 += 4 * lda;
  537. b += 32;
  538. }
  539. if (m & 2) {
  540. ao1 += 2 * lda;
  541. b += 16;
  542. }
  543. if (m & 1) {
  544. b += 8;
  545. }
  546. } else {
  547. #ifndef UNIT
  548. data01 = *(ao1 + 0);
  549. #endif
  550. data09 = *(ao2 + 0);
  551. data17 = *(ao3 + 0);
  552. data25 = *(ao4 + 0);
  553. data33 = *(ao5 + 0);
  554. data41 = *(ao6 + 0);
  555. data49 = *(ao7 + 0);
  556. data57 = *(ao8 + 0);
  557. if (i >= 2) {
  558. #ifndef UNIT
  559. data10 = *(ao2 + 1);
  560. #endif
  561. data18 = *(ao3 + 1);
  562. data26 = *(ao4 + 1);
  563. data34 = *(ao5 + 1);
  564. data42 = *(ao6 + 1);
  565. data50 = *(ao7 + 1);
  566. data58 = *(ao8 + 1);
  567. }
  568. if (i >= 3) {
  569. #ifndef UNIT
  570. data19 = *(ao3 + 2);
  571. #endif
  572. data27 = *(ao4 + 2);
  573. data35 = *(ao5 + 2);
  574. data43 = *(ao6 + 2);
  575. data51 = *(ao7 + 2);
  576. data59 = *(ao8 + 2);
  577. }
  578. if (i >= 4) {
  579. #ifndef UNIT
  580. data28 = *(ao4 + 3);
  581. #endif
  582. data36 = *(ao5 + 3);
  583. data44 = *(ao6 + 3);
  584. data52 = *(ao7 + 3);
  585. data60 = *(ao8 + 3);
  586. }
  587. if (i >= 5) {
  588. #ifndef UNIT
  589. data37 = *(ao5 + 4);
  590. #endif
  591. data45 = *(ao6 + 4);
  592. data53 = *(ao7 + 4);
  593. data61 = *(ao8 + 4);
  594. }
  595. if (i >= 6) {
  596. #ifndef UNIT
  597. data46 = *(ao6 + 5);
  598. #endif
  599. data54 = *(ao7 + 5);
  600. data62 = *(ao8 + 5);
  601. }
  602. if (i >= 7) {
  603. #ifndef UNIT
  604. data55 = *(ao7 + 6);
  605. #endif
  606. data63 = *(ao8 + 6);
  607. }
  608. #ifdef UNIT
  609. b[ 0] = ONE;
  610. #else
  611. b[ 0] = data01;
  612. #endif
  613. b[ 1] = data09;
  614. b[ 2] = data17;
  615. b[ 3] = data25;
  616. b[ 4] = data33;
  617. b[ 5] = data41;
  618. b[ 6] = data49;
  619. b[ 7] = data57;
  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] = data18;
  629. b[ 3] = data26;
  630. b[ 4] = data34;
  631. b[ 5] = data42;
  632. b[ 6] = data50;
  633. b[ 7] = data58;
  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] = data27;
  645. b[ 4] = data35;
  646. b[ 5] = data43;
  647. b[ 6] = data51;
  648. b[ 7] = data59;
  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] = data36;
  661. b[ 5] = data44;
  662. b[ 6] = data52;
  663. b[ 7] = data60;
  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] = data45;
  677. b[ 6] = data53;
  678. b[ 7] = data61;
  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] = data54;
  693. b[ 7] = data62;
  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] = data63;
  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 + posX + (posY + 0) * lda;
  721. ao2 = a + posX + (posY + 1) * lda;
  722. ao3 = a + posX + (posY + 2) * lda;
  723. ao4 = a + posX + (posY + 3) * lda;
  724. } else {
  725. ao1 = a + posY + (posX + 0) * lda;
  726. ao2 = a + posY + (posX + 1) * lda;
  727. ao3 = a + posY + (posX + 2) * lda;
  728. ao4 = a + posY + (posX + 3) * lda;
  729. }
  730. i = (m >> 2);
  731. if (i > 0) {
  732. do {
  733. if (X < posY) {
  734. data01 = *(ao1 + 0);
  735. data02 = *(ao1 + 1);
  736. data03 = *(ao1 + 2);
  737. data04 = *(ao1 + 3);
  738. data09 = *(ao2 + 0);
  739. data10 = *(ao2 + 1);
  740. data11 = *(ao2 + 2);
  741. data12 = *(ao2 + 3);
  742. data17 = *(ao3 + 0);
  743. data18 = *(ao3 + 1);
  744. data19 = *(ao3 + 2);
  745. data20 = *(ao3 + 3);
  746. data25 = *(ao4 + 0);
  747. data26 = *(ao4 + 1);
  748. data27 = *(ao4 + 2);
  749. data28 = *(ao4 + 3);
  750. b[ 0] = data01;
  751. b[ 1] = data09;
  752. b[ 2] = data17;
  753. b[ 3] = data25;
  754. b[ 4] = data02;
  755. b[ 5] = data10;
  756. b[ 6] = data18;
  757. b[ 7] = data26;
  758. b[ 8] = data03;
  759. b[ 9] = data11;
  760. b[10] = data19;
  761. b[11] = data27;
  762. b[12] = data04;
  763. b[13] = data12;
  764. b[14] = data20;
  765. b[15] = data28;
  766. ao1 += 4;
  767. ao2 += 4;
  768. ao3 += 4;
  769. ao4 += 4;
  770. b += 16;
  771. } else
  772. if (X > posY) {
  773. ao1 += 4 * lda;
  774. ao2 += 4 * lda;
  775. ao3 += 4 * lda;
  776. ao4 += 4 * lda;
  777. b += 16;
  778. } else {
  779. #ifdef UNIT
  780. data09 = *(ao2 + 0);
  781. data17 = *(ao3 + 0);
  782. data18 = *(ao3 + 1);
  783. data25 = *(ao4 + 0);
  784. data26 = *(ao4 + 1);
  785. data27 = *(ao4 + 2);
  786. b[ 0] = ONE;
  787. b[ 1] = data09;
  788. b[ 2] = data17;
  789. b[ 3] = data25;
  790. b[ 4] = ZERO;
  791. b[ 5] = ONE;
  792. b[ 6] = data18;
  793. b[ 7] = data26;
  794. b[ 8] = ZERO;
  795. b[ 9] = ZERO;
  796. b[10] = ONE;
  797. b[11] = data27;
  798. b[12] = ZERO;
  799. b[13] = ZERO;
  800. b[14] = ZERO;
  801. b[15] = ONE;
  802. #else
  803. data01 = *(ao1 + 0);
  804. data09 = *(ao2 + 0);
  805. data10 = *(ao2 + 1);
  806. data17 = *(ao3 + 0);
  807. data18 = *(ao3 + 1);
  808. data19 = *(ao3 + 2);
  809. data25 = *(ao4 + 0);
  810. data26 = *(ao4 + 1);
  811. data27 = *(ao4 + 2);
  812. data28 = *(ao4 + 3);
  813. b[ 0] = data01;
  814. b[ 1] = data09;
  815. b[ 2] = data17;
  816. b[ 3] = data25;
  817. b[ 4] = ZERO;
  818. b[ 5] = data10;
  819. b[ 6] = data18;
  820. b[ 7] = data26;
  821. b[ 8] = ZERO;
  822. b[ 9] = ZERO;
  823. b[10] = data19;
  824. b[11] = data27;
  825. b[12] = ZERO;
  826. b[13] = ZERO;
  827. b[14] = ZERO;
  828. b[15] = data28;
  829. #endif
  830. ao1 += 4 * lda;
  831. ao2 += 4 * lda;
  832. ao3 += 4 * lda;
  833. ao4 += 4 * lda;
  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. data01 = *(ao1 + 0);
  845. data02 = *(ao1 + 1);
  846. data09 = *(ao2 + 0);
  847. data10 = *(ao2 + 1);
  848. data17 = *(ao3 + 0);
  849. data18 = *(ao3 + 1);
  850. data25 = *(ao4 + 0);
  851. data26 = *(ao4 + 1);
  852. b[ 0] = data01;
  853. b[ 1] = data09;
  854. b[ 2] = data17;
  855. b[ 3] = data25;
  856. b[ 4] = data02;
  857. b[ 5] = data10;
  858. b[ 6] = data18;
  859. b[ 7] = data26;
  860. ao1 += 2;
  861. ao2 += 2;
  862. ao3 += 2;
  863. ao4 += 2;
  864. b += 8;
  865. }
  866. if (m & 1) {
  867. data01 = *(ao1 + 0);
  868. data09 = *(ao2 + 0);
  869. data17 = *(ao3 + 0);
  870. data25 = *(ao4 + 0);
  871. b[ 0] = data01;
  872. b[ 1] = data09;
  873. b[ 2] = data17;
  874. b[ 3] = data25;
  875. b += 4;
  876. }
  877. } else
  878. if (X > posY) {
  879. if (m & 2) {
  880. ao1 += 2 * lda;
  881. b += 8;
  882. }
  883. if (m & 1) {
  884. b += 4;
  885. }
  886. } else {
  887. #ifndef UNIT
  888. data01 = *(ao1 + 0);
  889. #endif
  890. data09 = *(ao2 + 0);
  891. data17 = *(ao3 + 0);
  892. data25 = *(ao4 + 0);
  893. if (i >= 2) {
  894. #ifndef UNIT
  895. data10 = *(ao2 + 1);
  896. #endif
  897. data18 = *(ao3 + 1);
  898. data26 = *(ao4 + 1);
  899. }
  900. if (i >= 3) {
  901. #ifndef UNIT
  902. data19 = *(ao3 + 2);
  903. #endif
  904. data27 = *(ao4 + 2);
  905. }
  906. #ifdef UNIT
  907. b[ 0] = ONE;
  908. #else
  909. b[ 0] = data01;
  910. #endif
  911. b[ 1] = data09;
  912. b[ 2] = data17;
  913. b[ 3] = data25;
  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] = data18;
  923. b[ 3] = data26;
  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] = data27;
  935. b += 4;
  936. }
  937. }
  938. }
  939. posY += 4;
  940. }
  941. if (n & 2){
  942. X = posX;
  943. if (posX <= posY) {
  944. ao1 = a + posX + (posY + 0) * lda;
  945. ao2 = a + posX + (posY + 1) * lda;
  946. } else {
  947. ao1 = a + posY + (posX + 0) * lda;
  948. ao2 = a + posY + (posX + 1) * lda;
  949. }
  950. i = (m >> 1);
  951. if (i > 0) {
  952. do {
  953. if (X < posY) {
  954. data01 = *(ao1 + 0);
  955. data02 = *(ao1 + 1);
  956. data09 = *(ao2 + 0);
  957. data10 = *(ao2 + 1);
  958. b[ 0] = data01;
  959. b[ 1] = data09;
  960. b[ 2] = data02;
  961. b[ 3] = data10;
  962. ao1 += 2;
  963. ao2 += 2;
  964. b += 4;
  965. } else
  966. if (X > posY) {
  967. ao1 += 2 * lda;
  968. ao2 += 2 * lda;
  969. b += 4;
  970. } else {
  971. #ifdef UNIT
  972. data09 = *(ao2 + 0);
  973. b[ 0] = ONE;
  974. b[ 1] = data09;
  975. b[ 2] = ZERO;
  976. b[ 3] = ONE;
  977. #else
  978. data01 = *(ao1 + 0);
  979. data09 = *(ao2 + 0);
  980. data10 = *(ao2 + 1);
  981. b[ 0] = data01;
  982. b[ 1] = data09;
  983. b[ 2] = ZERO;
  984. b[ 3] = data10;
  985. #endif
  986. ao1 += 2 * lda;
  987. ao2 += 2 * lda;
  988. b += 4;
  989. }
  990. X += 2;
  991. i --;
  992. } while (i > 0);
  993. }
  994. if (m & 1) {
  995. if (X < posY) {
  996. data01 = *(ao1 + 0);
  997. data09 = *(ao2 + 0);
  998. b[ 0] = data01;
  999. b[ 1] = data09;
  1000. b += 2;
  1001. } else
  1002. if (X > posY) {
  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 + posX + (posY + 0) * lda;
  1024. } else {
  1025. ao1 = a + posY + (posX + 0) * lda;
  1026. }
  1027. i = m;
  1028. if (m > 0) {
  1029. do {
  1030. if (X < posY) {
  1031. data01 = *(ao1 + 0);
  1032. b[ 0] = data01;
  1033. ao1 += 1;
  1034. b += 1;
  1035. } else
  1036. if (X > posY) {
  1037. ao1 += lda;
  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 += lda;
  1047. b += 1;
  1048. }
  1049. X += 1;
  1050. i --;
  1051. } while (i > 0);
  1052. }
  1053. }
  1054. return 0;
  1055. }