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_16.c 31 kB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550
  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, ii;
  42. BLASLONG X;
  43. FLOAT *a01, *a02, *a03 ,*a04, *a05, *a06, *a07, *a08;
  44. FLOAT *a09, *a10, *a11, *a12, *a13, *a14, *a15, *a16;
  45. js = (n >> 4);
  46. if (js > 0){
  47. do {
  48. X = posX;
  49. if (posX <= posY) {
  50. a01 = a + posX + (posY + 0) * lda;
  51. a02 = a + posX + (posY + 1) * lda;
  52. a03 = a + posX + (posY + 2) * lda;
  53. a04 = a + posX + (posY + 3) * lda;
  54. a05 = a + posX + (posY + 4) * lda;
  55. a06 = a + posX + (posY + 5) * lda;
  56. a07 = a + posX + (posY + 6) * lda;
  57. a08 = a + posX + (posY + 7) * lda;
  58. a09 = a + posX + (posY + 8) * lda;
  59. a10 = a + posX + (posY + 9) * lda;
  60. a11 = a + posX + (posY + 10) * lda;
  61. a12 = a + posX + (posY + 11) * lda;
  62. a13 = a + posX + (posY + 12) * lda;
  63. a14 = a + posX + (posY + 13) * lda;
  64. a15 = a + posX + (posY + 14) * lda;
  65. a16 = a + posX + (posY + 15) * lda;
  66. } else {
  67. a01 = a + posY + (posX + 0) * lda;
  68. a02 = a + posY + (posX + 1) * lda;
  69. a03 = a + posY + (posX + 2) * lda;
  70. a04 = a + posY + (posX + 3) * lda;
  71. a05 = a + posY + (posX + 4) * lda;
  72. a06 = a + posY + (posX + 5) * lda;
  73. a07 = a + posY + (posX + 6) * lda;
  74. a08 = a + posY + (posX + 7) * lda;
  75. a09 = a + posY + (posX + 8) * lda;
  76. a10 = a + posY + (posX + 9) * lda;
  77. a11 = a + posY + (posX + 10) * lda;
  78. a12 = a + posY + (posX + 11) * lda;
  79. a13 = a + posY + (posX + 12) * lda;
  80. a14 = a + posY + (posX + 13) * lda;
  81. a15 = a + posY + (posX + 14) * lda;
  82. a16 = a + posY + (posX + 15) * lda;
  83. }
  84. i = (m >> 4);
  85. if (i > 0) {
  86. do {
  87. if (X < posY) {
  88. a01 += 16;
  89. a02 += 16;
  90. a03 += 16;
  91. a04 += 16;
  92. a05 += 16;
  93. a06 += 16;
  94. a07 += 16;
  95. a08 += 16;
  96. a09 += 16;
  97. a10 += 16;
  98. a11 += 16;
  99. a12 += 16;
  100. a13 += 16;
  101. a14 += 16;
  102. a15 += 16;
  103. a16 += 16;
  104. b += 256;
  105. } else
  106. if (X > posY) {
  107. for (ii = 0; ii < 16; ii++){
  108. b[ 0] = *(a01 + 0);
  109. b[ 1] = *(a01 + 1);
  110. b[ 2] = *(a01 + 2);
  111. b[ 3] = *(a01 + 3);
  112. b[ 4] = *(a01 + 4);
  113. b[ 5] = *(a01 + 5);
  114. b[ 6] = *(a01 + 6);
  115. b[ 7] = *(a01 + 7);
  116. b[ 8] = *(a01 + 8);
  117. b[ 9] = *(a01 + 9);
  118. b[ 10] = *(a01 + 10);
  119. b[ 11] = *(a01 + 11);
  120. b[ 12] = *(a01 + 12);
  121. b[ 13] = *(a01 + 13);
  122. b[ 14] = *(a01 + 14);
  123. b[ 15] = *(a01 + 15);
  124. a01 += lda;
  125. b += 16;
  126. }
  127. a02 += 16 * lda;
  128. a03 += 16 * lda;
  129. a04 += 16 * lda;
  130. a05 += 16 * lda;
  131. a06 += 16 * lda;
  132. a07 += 16 * lda;
  133. a08 += 16 * lda;
  134. a09 += 16 * lda;
  135. a10 += 16 * lda;
  136. a11 += 16 * lda;
  137. a12 += 16 * lda;
  138. a13 += 16 * lda;
  139. a14 += 16 * lda;
  140. a15 += 16 * lda;
  141. a16 += 16 * lda;
  142. } else {
  143. #ifdef UNIT
  144. b[ 0] = ONE;
  145. #else
  146. b[ 0] = *(a01 + 0);
  147. #endif
  148. b[ 1] = ZERO;
  149. b[ 2] = ZERO;
  150. b[ 3] = ZERO;
  151. b[ 4] = ZERO;
  152. b[ 5] = ZERO;
  153. b[ 6] = ZERO;
  154. b[ 7] = ZERO;
  155. b[ 8] = ZERO;
  156. b[ 9] = ZERO;
  157. b[ 10] = ZERO;
  158. b[ 11] = ZERO;
  159. b[ 12] = ZERO;
  160. b[ 13] = ZERO;
  161. b[ 14] = ZERO;
  162. b[ 15] = ZERO;
  163. b[ 16] = *(a02 + 0);
  164. #ifdef UNIT
  165. b[ 17] = ONE;
  166. #else
  167. b[ 17] = *(a02 + 1);
  168. #endif
  169. b[ 18] = ZERO;
  170. b[ 19] = ZERO;
  171. b[ 20] = ZERO;
  172. b[ 21] = ZERO;
  173. b[ 22] = ZERO;
  174. b[ 23] = ZERO;
  175. b[ 24] = ZERO;
  176. b[ 25] = ZERO;
  177. b[ 26] = ZERO;
  178. b[ 27] = ZERO;
  179. b[ 28] = ZERO;
  180. b[ 29] = ZERO;
  181. b[ 30] = ZERO;
  182. b[ 31] = ZERO;
  183. b[ 32] = *(a03 + 0);
  184. b[ 33] = *(a03 + 1);
  185. #ifdef UNIT
  186. b[ 34] = ONE;
  187. #else
  188. b[ 34] = *(a03 + 2);
  189. #endif
  190. b[ 35] = ZERO;
  191. b[ 36] = ZERO;
  192. b[ 37] = ZERO;
  193. b[ 38] = ZERO;
  194. b[ 39] = ZERO;
  195. b[ 40] = ZERO;
  196. b[ 41] = ZERO;
  197. b[ 42] = ZERO;
  198. b[ 43] = ZERO;
  199. b[ 44] = ZERO;
  200. b[ 45] = ZERO;
  201. b[ 46] = ZERO;
  202. b[ 47] = ZERO;
  203. b[ 48] = *(a04 + 0);
  204. b[ 49] = *(a04 + 1);
  205. b[ 50] = *(a04 + 2);
  206. #ifdef UNIT
  207. b[ 51] = ONE;
  208. #else
  209. b[ 51] = *(a04 + 3);
  210. #endif
  211. b[ 52] = ZERO;
  212. b[ 53] = ZERO;
  213. b[ 54] = ZERO;
  214. b[ 55] = ZERO;
  215. b[ 56] = ZERO;
  216. b[ 57] = ZERO;
  217. b[ 58] = ZERO;
  218. b[ 59] = ZERO;
  219. b[ 60] = ZERO;
  220. b[ 61] = ZERO;
  221. b[ 62] = ZERO;
  222. b[ 63] = ZERO;
  223. b[ 64] = *(a05 + 0);
  224. b[ 65] = *(a05 + 1);
  225. b[ 66] = *(a05 + 2);
  226. b[ 67] = *(a05 + 3);
  227. #ifdef UNIT
  228. b[ 68] = ONE;
  229. #else
  230. b[ 68] = *(a05 + 4);
  231. #endif
  232. b[ 69] = ZERO;
  233. b[ 70] = ZERO;
  234. b[ 71] = ZERO;
  235. b[ 72] = ZERO;
  236. b[ 73] = ZERO;
  237. b[ 74] = ZERO;
  238. b[ 75] = ZERO;
  239. b[ 76] = ZERO;
  240. b[ 77] = ZERO;
  241. b[ 78] = ZERO;
  242. b[ 79] = ZERO;
  243. b[ 80] = *(a06 + 0);
  244. b[ 81] = *(a06 + 1);
  245. b[ 82] = *(a06 + 2);
  246. b[ 83] = *(a06 + 3);
  247. b[ 84] = *(a06 + 4);
  248. #ifdef UNIT
  249. b[ 85] = ONE;
  250. #else
  251. b[ 85] = *(a06 + 5);
  252. #endif
  253. b[ 86] = ZERO;
  254. b[ 87] = ZERO;
  255. b[ 88] = ZERO;
  256. b[ 89] = ZERO;
  257. b[ 90] = ZERO;
  258. b[ 91] = ZERO;
  259. b[ 92] = ZERO;
  260. b[ 93] = ZERO;
  261. b[ 94] = ZERO;
  262. b[ 95] = ZERO;
  263. b[ 96] = *(a07 + 0);
  264. b[ 97] = *(a07 + 1);
  265. b[ 98] = *(a07 + 2);
  266. b[ 99] = *(a07 + 3);
  267. b[100] = *(a07 + 4);
  268. b[101] = *(a07 + 5);
  269. #ifdef UNIT
  270. b[102] = ONE;
  271. #else
  272. b[102] = *(a07 + 6);
  273. #endif
  274. b[103] = ZERO;
  275. b[104] = ZERO;
  276. b[105] = ZERO;
  277. b[106] = ZERO;
  278. b[107] = ZERO;
  279. b[108] = ZERO;
  280. b[109] = ZERO;
  281. b[110] = ZERO;
  282. b[111] = ZERO;
  283. b[112] = *(a08 + 0);
  284. b[113] = *(a08 + 1);
  285. b[114] = *(a08 + 2);
  286. b[115] = *(a08 + 3);
  287. b[116] = *(a08 + 4);
  288. b[117] = *(a08 + 5);
  289. b[118] = *(a08 + 6);
  290. #ifdef UNIT
  291. b[119] = ONE;
  292. #else
  293. b[119] = *(a08 + 7);
  294. #endif
  295. b[120] = ZERO;
  296. b[121] = ZERO;
  297. b[122] = ZERO;
  298. b[123] = ZERO;
  299. b[124] = ZERO;
  300. b[125] = ZERO;
  301. b[126] = ZERO;
  302. b[127] = ZERO;
  303. b[128] = *(a09 + 0);
  304. b[129] = *(a09 + 1);
  305. b[130] = *(a09 + 2);
  306. b[131] = *(a09 + 3);
  307. b[132] = *(a09 + 4);
  308. b[133] = *(a09 + 5);
  309. b[134] = *(a09 + 6);
  310. b[135] = *(a09 + 7);
  311. #ifdef UNIT
  312. b[136] = ONE;
  313. #else
  314. b[136] = *(a09 + 8);
  315. #endif
  316. b[137] = ZERO;
  317. b[138] = ZERO;
  318. b[139] = ZERO;
  319. b[140] = ZERO;
  320. b[141] = ZERO;
  321. b[142] = ZERO;
  322. b[143] = ZERO;
  323. b[144] = *(a10 + 0);
  324. b[145] = *(a10 + 1);
  325. b[146] = *(a10 + 2);
  326. b[147] = *(a10 + 3);
  327. b[148] = *(a10 + 4);
  328. b[149] = *(a10 + 5);
  329. b[150] = *(a10 + 6);
  330. b[151] = *(a10 + 7);
  331. b[152] = *(a10 + 8);
  332. #ifdef UNIT
  333. b[153] = ONE;
  334. #else
  335. b[153] = *(a10 + 9);
  336. #endif
  337. b[154] = ZERO;
  338. b[155] = ZERO;
  339. b[156] = ZERO;
  340. b[157] = ZERO;
  341. b[158] = ZERO;
  342. b[159] = ZERO;
  343. b[160] = *(a11 + 0);
  344. b[161] = *(a11 + 1);
  345. b[162] = *(a11 + 2);
  346. b[163] = *(a11 + 3);
  347. b[164] = *(a11 + 4);
  348. b[165] = *(a11 + 5);
  349. b[166] = *(a11 + 6);
  350. b[167] = *(a11 + 7);
  351. b[168] = *(a11 + 8);
  352. b[169] = *(a11 + 9);
  353. #ifdef UNIT
  354. b[170] = ONE;
  355. #else
  356. b[170] = *(a11 + 10);
  357. #endif
  358. b[171] = ZERO;
  359. b[172] = ZERO;
  360. b[173] = ZERO;
  361. b[174] = ZERO;
  362. b[175] = ZERO;
  363. b[176] = *(a12 + 0);
  364. b[177] = *(a12 + 1);
  365. b[178] = *(a12 + 2);
  366. b[179] = *(a12 + 3);
  367. b[180] = *(a12 + 4);
  368. b[181] = *(a12 + 5);
  369. b[182] = *(a12 + 6);
  370. b[183] = *(a12 + 7);
  371. b[184] = *(a12 + 8);
  372. b[185] = *(a12 + 9);
  373. b[186] = *(a12 + 10);
  374. #ifdef UNIT
  375. b[187] = ONE;
  376. #else
  377. b[187] = *(a12 + 11);
  378. #endif
  379. b[188] = ZERO;
  380. b[189] = ZERO;
  381. b[190] = ZERO;
  382. b[191] = ZERO;
  383. b[192] = *(a13 + 0);
  384. b[193] = *(a13 + 1);
  385. b[194] = *(a13 + 2);
  386. b[195] = *(a13 + 3);
  387. b[196] = *(a13 + 4);
  388. b[197] = *(a13 + 5);
  389. b[198] = *(a13 + 6);
  390. b[199] = *(a13 + 7);
  391. b[200] = *(a13 + 8);
  392. b[201] = *(a13 + 9);
  393. b[202] = *(a13 + 10);
  394. b[203] = *(a13 + 11);
  395. #ifdef UNIT
  396. b[204] = ONE;
  397. #else
  398. b[204] = *(a13 + 12);
  399. #endif
  400. b[205] = ZERO;
  401. b[206] = ZERO;
  402. b[207] = ZERO;
  403. b[208] = *(a14 + 0);
  404. b[209] = *(a14 + 1);
  405. b[210] = *(a14 + 2);
  406. b[211] = *(a14 + 3);
  407. b[212] = *(a14 + 4);
  408. b[213] = *(a14 + 5);
  409. b[214] = *(a14 + 6);
  410. b[215] = *(a14 + 7);
  411. b[216] = *(a14 + 8);
  412. b[217] = *(a14 + 9);
  413. b[218] = *(a14 + 10);
  414. b[219] = *(a14 + 11);
  415. b[220] = *(a14 + 12);
  416. #ifdef UNIT
  417. b[221] = ONE;
  418. #else
  419. b[221] = *(a14 + 13);
  420. #endif
  421. b[222] = ZERO;
  422. b[223] = ZERO;
  423. b[224] = *(a15 + 0);
  424. b[225] = *(a15 + 1);
  425. b[226] = *(a15 + 2);
  426. b[227] = *(a15 + 3);
  427. b[228] = *(a15 + 4);
  428. b[229] = *(a15 + 5);
  429. b[230] = *(a15 + 6);
  430. b[231] = *(a15 + 7);
  431. b[232] = *(a15 + 8);
  432. b[233] = *(a15 + 9);
  433. b[234] = *(a15 + 10);
  434. b[235] = *(a15 + 11);
  435. b[236] = *(a15 + 12);
  436. b[237] = *(a15 + 13);
  437. #ifdef UNIT
  438. b[238] = ONE;
  439. #else
  440. b[238] = *(a15 + 14);
  441. #endif
  442. b[239] = ZERO;
  443. b[240] = *(a16 + 0);
  444. b[241] = *(a16 + 1);
  445. b[242] = *(a16 + 2);
  446. b[243] = *(a16 + 3);
  447. b[244] = *(a16 + 4);
  448. b[245] = *(a16 + 5);
  449. b[246] = *(a16 + 6);
  450. b[247] = *(a16 + 7);
  451. b[248] = *(a16 + 8);
  452. b[249] = *(a16 + 9);
  453. b[250] = *(a16 + 10);
  454. b[251] = *(a16 + 11);
  455. b[252] = *(a16 + 12);
  456. b[253] = *(a16 + 13);
  457. b[254] = *(a16 + 14);
  458. #ifdef UNIT
  459. b[255] = ONE;
  460. #else
  461. b[255] = *(a16 + 15);
  462. #endif
  463. a01 += 16 * lda;
  464. a02 += 16 * lda;
  465. a03 += 16 * lda;
  466. a04 += 16 * lda;
  467. a05 += 16 * lda;
  468. a06 += 16 * lda;
  469. a07 += 16 * lda;
  470. a08 += 16 * lda;
  471. a09 += 16 * lda;
  472. a10 += 16 * lda;
  473. a11 += 16 * lda;
  474. a12 += 16 * lda;
  475. a13 += 16 * lda;
  476. a14 += 16 * lda;
  477. a15 += 16 * lda;
  478. a16 += 16 * lda;
  479. b += 256;
  480. }
  481. X += 16;
  482. i --;
  483. } while (i > 0);
  484. }
  485. i = (m & 15);
  486. if (i > 0) {
  487. if (X < posY) {
  488. a01 += i;
  489. a02 += i;
  490. a03 += i;
  491. a04 += i;
  492. a05 += i;
  493. a06 += i;
  494. a07 += i;
  495. a08 += i;
  496. a09 += i;
  497. a10 += i;
  498. a11 += i;
  499. a12 += i;
  500. a13 += i;
  501. a14 += i;
  502. a15 += i;
  503. a16 += i;
  504. b += 16 * i;
  505. } else
  506. if (X > posY) {
  507. for (ii = 0; ii < i; ii++){
  508. b[ 0] = *(a01 + 0);
  509. b[ 1] = *(a01 + 1);
  510. b[ 2] = *(a01 + 2);
  511. b[ 3] = *(a01 + 3);
  512. b[ 4] = *(a01 + 4);
  513. b[ 5] = *(a01 + 5);
  514. b[ 6] = *(a01 + 6);
  515. b[ 7] = *(a01 + 7);
  516. b[ 8] = *(a01 + 8);
  517. b[ 9] = *(a01 + 9);
  518. b[ 10] = *(a01 + 10);
  519. b[ 11] = *(a01 + 11);
  520. b[ 12] = *(a01 + 12);
  521. b[ 13] = *(a01 + 13);
  522. b[ 14] = *(a01 + 14);
  523. b[ 15] = *(a01 + 15);
  524. a01 += lda;
  525. a02 += lda;
  526. a03 += lda;
  527. a04 += lda;
  528. a05 += lda;
  529. a06 += lda;
  530. a07 += lda;
  531. a08 += lda;
  532. a09 += lda;
  533. a10 += lda;
  534. a11 += lda;
  535. a12 += lda;
  536. a13 += lda;
  537. a14 += lda;
  538. a15 += lda;
  539. a16 += lda;
  540. b += 16;
  541. }
  542. } else {
  543. #ifdef UNIT
  544. b[ 0] = ONE;
  545. #else
  546. b[ 0] = *(a01 + 0);
  547. #endif
  548. b[ 1] = ZERO;
  549. b[ 2] = ZERO;
  550. b[ 3] = ZERO;
  551. b[ 4] = ZERO;
  552. b[ 5] = ZERO;
  553. b[ 6] = ZERO;
  554. b[ 7] = ZERO;
  555. b[ 8] = ZERO;
  556. b[ 9] = ZERO;
  557. b[10] = ZERO;
  558. b[11] = ZERO;
  559. b[12] = ZERO;
  560. b[13] = ZERO;
  561. b[14] = ZERO;
  562. b[15] = ZERO;
  563. if (i >= 2) {
  564. b[ 0] = *(a02 + 0);
  565. #ifdef UNIT
  566. b[ 1] = ONE;
  567. #else
  568. b[ 1] = *(a02 + 1);
  569. #endif
  570. b[ 2] = ZERO;
  571. b[ 3] = ZERO;
  572. b[ 4] = ZERO;
  573. b[ 5] = ZERO;
  574. b[ 6] = ZERO;
  575. b[ 7] = ZERO;
  576. b[ 8] = ZERO;
  577. b[ 9] = ZERO;
  578. b[10] = ZERO;
  579. b[11] = ZERO;
  580. b[12] = ZERO;
  581. b[13] = ZERO;
  582. b[14] = ZERO;
  583. b[15] = ZERO;
  584. b += 16;
  585. }
  586. if (i >= 3) {
  587. b[ 0] = *(a03 + 0);
  588. b[ 1] = *(a03 + 1);
  589. #ifdef UNIT
  590. b[ 2] = ONE;
  591. #else
  592. b[ 2] = *(a03 + 2);
  593. #endif
  594. b[ 3] = ZERO;
  595. b[ 4] = ZERO;
  596. b[ 5] = ZERO;
  597. b[ 6] = ZERO;
  598. b[ 7] = ZERO;
  599. b[ 8] = ZERO;
  600. b[ 9] = ZERO;
  601. b[10] = ZERO;
  602. b[11] = ZERO;
  603. b[12] = ZERO;
  604. b[13] = ZERO;
  605. b[14] = ZERO;
  606. b[15] = ZERO;
  607. b += 16;
  608. }
  609. if (i >= 4) {
  610. b[ 0] = *(a04 + 0);
  611. b[ 1] = *(a04 + 1);
  612. b[ 2] = *(a04 + 2);
  613. #ifdef UNIT
  614. b[ 3] = ONE;
  615. #else
  616. b[ 3] = *(a04 + 3);
  617. #endif
  618. b[ 4] = ZERO;
  619. b[ 5] = ZERO;
  620. b[ 6] = ZERO;
  621. b[ 7] = ZERO;
  622. b[ 8] = ZERO;
  623. b[ 9] = ZERO;
  624. b[10] = ZERO;
  625. b[11] = ZERO;
  626. b[12] = ZERO;
  627. b[13] = ZERO;
  628. b[14] = ZERO;
  629. b[15] = ZERO;
  630. b[16] = ZERO;
  631. b += 16;
  632. }
  633. if (i >= 5) {
  634. b[ 0] = *(a05 + 0);
  635. b[ 1] = *(a05 + 1);
  636. b[ 2] = *(a05 + 2);
  637. b[ 3] = *(a05 + 3);
  638. #ifdef UNIT
  639. b[ 4] = ONE;
  640. #else
  641. b[ 4] = *(a05 + 4);
  642. #endif
  643. b[ 5] = ZERO;
  644. b[ 6] = ZERO;
  645. b[ 7] = ZERO;
  646. b[ 8] = ZERO;
  647. b[ 9] = ZERO;
  648. b[10] = ZERO;
  649. b[11] = ZERO;
  650. b[12] = ZERO;
  651. b[13] = ZERO;
  652. b[14] = ZERO;
  653. b[15] = ZERO;
  654. b += 16;
  655. }
  656. if (i >= 6) {
  657. b[ 0] = *(a06 + 0);
  658. b[ 1] = *(a06 + 1);
  659. b[ 2] = *(a06 + 2);
  660. b[ 3] = *(a06 + 3);
  661. b[ 4] = *(a06 + 4);
  662. #ifdef UNIT
  663. b[ 5] = ONE;
  664. #else
  665. b[ 5] = *(a06 + 5);
  666. #endif
  667. b[ 6] = ZERO;
  668. b[ 7] = ZERO;
  669. b[ 8] = ZERO;
  670. b[ 9] = ZERO;
  671. b[10] = ZERO;
  672. b[11] = ZERO;
  673. b[12] = ZERO;
  674. b[13] = ZERO;
  675. b[14] = ZERO;
  676. b[15] = ZERO;
  677. b += 16;
  678. }
  679. if (i >= 7) {
  680. b[ 0] = *(a07 + 0);
  681. b[ 1] = *(a07 + 1);
  682. b[ 2] = *(a07 + 2);
  683. b[ 3] = *(a07 + 3);
  684. b[ 4] = *(a07 + 4);
  685. b[ 5] = *(a07 + 5);
  686. #ifdef UNIT
  687. b[ 6] = ONE;
  688. #else
  689. b[ 6] = *(a07 + 6);
  690. #endif
  691. b[ 7] = ZERO;
  692. b[ 8] = ZERO;
  693. b[ 9] = ZERO;
  694. b[10] = ZERO;
  695. b[11] = ZERO;
  696. b[12] = ZERO;
  697. b[13] = ZERO;
  698. b[14] = ZERO;
  699. b[15] = ZERO;
  700. b += 16;
  701. }
  702. if (i >= 8) {
  703. b[ 0] = *(a08 + 0);
  704. b[ 1] = *(a08 + 1);
  705. b[ 2] = *(a08 + 2);
  706. b[ 3] = *(a08 + 3);
  707. b[ 4] = *(a08 + 4);
  708. b[ 5] = *(a08 + 5);
  709. b[ 6] = *(a08 + 6);
  710. #ifdef UNIT
  711. b[ 7] = ONE;
  712. #else
  713. b[ 7] = *(a08 + 7);
  714. #endif
  715. b[ 8] = ZERO;
  716. b[ 9] = ZERO;
  717. b[10] = ZERO;
  718. b[11] = ZERO;
  719. b[12] = ZERO;
  720. b[13] = ZERO;
  721. b[14] = ZERO;
  722. b[15] = ZERO;
  723. b[16] = ZERO;
  724. b += 16;
  725. }
  726. if (i >= 9) {
  727. b[ 0] = *(a09 + 0);
  728. b[ 1] = *(a09 + 1);
  729. b[ 2] = *(a09 + 2);
  730. b[ 3] = *(a09 + 3);
  731. b[ 4] = *(a09 + 4);
  732. b[ 5] = *(a09 + 5);
  733. b[ 6] = *(a09 + 6);
  734. b[ 7] = *(a09 + 7);
  735. #ifdef UNIT
  736. b[ 8] = ONE;
  737. #else
  738. b[ 8] = *(a09 + 8);
  739. #endif
  740. b[ 9] = ZERO;
  741. b[10] = ZERO;
  742. b[11] = ZERO;
  743. b[12] = ZERO;
  744. b[13] = ZERO;
  745. b[14] = ZERO;
  746. b[15] = ZERO;
  747. b += 16;
  748. }
  749. if (i >= 10) {
  750. b[ 0] = *(a10 + 0);
  751. b[ 1] = *(a10 + 1);
  752. b[ 2] = *(a10 + 2);
  753. b[ 3] = *(a10 + 3);
  754. b[ 4] = *(a10 + 4);
  755. b[ 5] = *(a10 + 5);
  756. b[ 6] = *(a10 + 6);
  757. b[ 7] = *(a10 + 7);
  758. b[ 8] = *(a10 + 8);
  759. #ifdef UNIT
  760. b[ 9] = ONE;
  761. #else
  762. b[ 9] = *(a10 + 9);
  763. #endif
  764. b[10] = ZERO;
  765. b[11] = ZERO;
  766. b[12] = ZERO;
  767. b[13] = ZERO;
  768. b[14] = ZERO;
  769. b[15] = ZERO;
  770. b += 16;
  771. }
  772. if (i >= 11) {
  773. b[ 0] = *(a11 + 0);
  774. b[ 1] = *(a11 + 1);
  775. b[ 2] = *(a11 + 2);
  776. b[ 3] = *(a11 + 3);
  777. b[ 4] = *(a11 + 4);
  778. b[ 5] = *(a11 + 5);
  779. b[ 6] = *(a11 + 6);
  780. b[ 7] = *(a11 + 7);
  781. b[ 8] = *(a11 + 8);
  782. b[ 9] = *(a11 + 9);
  783. #ifdef UNIT
  784. b[10] = ONE;
  785. #else
  786. b[10] = *(a11 + 10);
  787. #endif
  788. b[11] = ZERO;
  789. b[12] = ZERO;
  790. b[13] = ZERO;
  791. b[14] = ZERO;
  792. b[15] = ZERO;
  793. b += 16;
  794. }
  795. if (i >= 12) {
  796. b[ 0] = *(a12 + 0);
  797. b[ 1] = *(a12 + 1);
  798. b[ 2] = *(a12 + 2);
  799. b[ 3] = *(a12 + 3);
  800. b[ 4] = *(a12 + 4);
  801. b[ 5] = *(a12 + 5);
  802. b[ 6] = *(a12 + 6);
  803. b[ 7] = *(a12 + 7);
  804. b[ 8] = *(a12 + 8);
  805. b[ 9] = *(a12 + 9);
  806. b[10] = *(a12 + 10);
  807. #ifdef UNIT
  808. b[11] = ONE;
  809. #else
  810. b[11] = *(a12 + 11);
  811. #endif
  812. b[12] = ZERO;
  813. b[13] = ZERO;
  814. b[14] = ZERO;
  815. b[15] = ZERO;
  816. b += 16;
  817. }
  818. if (i >= 13) {
  819. b[ 0] = *(a13 + 0);
  820. b[ 1] = *(a13 + 1);
  821. b[ 2] = *(a13 + 2);
  822. b[ 3] = *(a13 + 3);
  823. b[ 4] = *(a13 + 4);
  824. b[ 5] = *(a13 + 5);
  825. b[ 6] = *(a13 + 6);
  826. b[ 7] = *(a13 + 7);
  827. b[ 8] = *(a13 + 8);
  828. b[ 9] = *(a13 + 9);
  829. b[10] = *(a13 + 10);
  830. b[11] = *(a13 + 11);
  831. #ifdef UNIT
  832. b[12] = ONE;
  833. #else
  834. b[12] = *(a13 + 12);
  835. #endif
  836. b[13] = ZERO;
  837. b[14] = ZERO;
  838. b[15] = ZERO;
  839. b += 16;
  840. }
  841. if (i >= 14) {
  842. b[ 0] = *(a14 + 0);
  843. b[ 1] = *(a14 + 1);
  844. b[ 2] = *(a14 + 2);
  845. b[ 3] = *(a14 + 3);
  846. b[ 4] = *(a14 + 4);
  847. b[ 5] = *(a14 + 5);
  848. b[ 6] = *(a14 + 6);
  849. b[ 7] = *(a14 + 7);
  850. b[ 8] = *(a14 + 8);
  851. b[ 9] = *(a14 + 9);
  852. b[10] = *(a14 + 10);
  853. b[11] = *(a14 + 11);
  854. b[12] = *(a14 + 12);
  855. #ifdef UNIT
  856. b[13] = ONE;
  857. #else
  858. b[13] = *(a14 + 13);
  859. #endif
  860. b[14] = ZERO;
  861. b[15] = ZERO;
  862. b += 16;
  863. }
  864. if (i >= 15) {
  865. b[ 0] = *(a15 + 0);
  866. b[ 1] = *(a15 + 1);
  867. b[ 2] = *(a15 + 2);
  868. b[ 3] = *(a15 + 3);
  869. b[ 4] = *(a15 + 4);
  870. b[ 5] = *(a15 + 5);
  871. b[ 6] = *(a15 + 6);
  872. b[ 7] = *(a15 + 7);
  873. b[ 8] = *(a15 + 8);
  874. b[ 9] = *(a15 + 9);
  875. b[10] = *(a15 + 10);
  876. b[11] = *(a15 + 11);
  877. b[12] = *(a15 + 12);
  878. b[13] = *(a15 + 13);
  879. #ifdef UNIT
  880. b[14] = ONE;
  881. #else
  882. b[14] = *(a15 + 14);
  883. #endif
  884. b[15] = ZERO;
  885. }
  886. }
  887. }
  888. posY += 16;
  889. js --;
  890. } while (js > 0);
  891. } /* End of main loop */
  892. if (n & 8){
  893. X = posX;
  894. if (posX <= posY) {
  895. a01 = a + posX + (posY + 0) * lda;
  896. a02 = a + posX + (posY + 1) * lda;
  897. a03 = a + posX + (posY + 2) * lda;
  898. a04 = a + posX + (posY + 3) * lda;
  899. a05 = a + posX + (posY + 4) * lda;
  900. a06 = a + posX + (posY + 5) * lda;
  901. a07 = a + posX + (posY + 6) * lda;
  902. a08 = a + posX + (posY + 7) * lda;
  903. } else {
  904. a01 = a + posY + (posX + 0) * lda;
  905. a02 = a + posY + (posX + 1) * lda;
  906. a03 = a + posY + (posX + 2) * lda;
  907. a04 = a + posY + (posX + 3) * lda;
  908. a05 = a + posY + (posX + 4) * lda;
  909. a06 = a + posY + (posX + 5) * lda;
  910. a07 = a + posY + (posX + 6) * lda;
  911. a08 = a + posY + (posX + 7) * lda;
  912. }
  913. i = (m >> 3);
  914. if (i > 0) {
  915. do {
  916. if (X < posY) {
  917. a01 += 8;
  918. a02 += 8;
  919. a03 += 8;
  920. a04 += 8;
  921. a05 += 8;
  922. a06 += 8;
  923. a07 += 8;
  924. a08 += 8;
  925. b += 64;
  926. } else
  927. if (X > posY) {
  928. for (ii = 0; ii < 8; ii++){
  929. b[ 0] = *(a01 + 0);
  930. b[ 1] = *(a01 + 1);
  931. b[ 2] = *(a01 + 2);
  932. b[ 3] = *(a01 + 3);
  933. b[ 4] = *(a01 + 4);
  934. b[ 5] = *(a01 + 5);
  935. b[ 6] = *(a01 + 6);
  936. b[ 7] = *(a01 + 7);
  937. a01 += lda;
  938. b += 8;
  939. }
  940. a02 += 8 * lda;
  941. a03 += 8 * lda;
  942. a04 += 8 * lda;
  943. a05 += 8 * lda;
  944. a06 += 8 * lda;
  945. a07 += 8 * lda;
  946. a08 += 8 * lda;
  947. } else {
  948. #ifdef UNIT
  949. b[ 0] = ONE;
  950. #else
  951. b[ 0] = *(a01 + 0);
  952. #endif
  953. b[ 1] = ZERO;
  954. b[ 2] = ZERO;
  955. b[ 3] = ZERO;
  956. b[ 4] = ZERO;
  957. b[ 5] = ZERO;
  958. b[ 6] = ZERO;
  959. b[ 7] = ZERO;
  960. b[ 8] = *(a02 + 0);
  961. #ifdef UNIT
  962. b[ 9] = ONE;
  963. #else
  964. b[ 9] = *(a02 + 1);
  965. #endif
  966. b[ 10] = ZERO;
  967. b[ 11] = ZERO;
  968. b[ 12] = ZERO;
  969. b[ 13] = ZERO;
  970. b[ 14] = ZERO;
  971. b[ 15] = ZERO;
  972. b[ 16] = *(a03 + 0);
  973. b[ 17] = *(a03 + 1);
  974. #ifdef UNIT
  975. b[ 18] = ONE;
  976. #else
  977. b[ 18] = *(a03 + 2);
  978. #endif
  979. b[ 19] = ZERO;
  980. b[ 20] = ZERO;
  981. b[ 21] = ZERO;
  982. b[ 22] = ZERO;
  983. b[ 23] = ZERO;
  984. b[ 24] = *(a04 + 0);
  985. b[ 25] = *(a04 + 1);
  986. b[ 26] = *(a04 + 2);
  987. #ifdef UNIT
  988. b[ 27] = ONE;
  989. #else
  990. b[ 27] = *(a04 + 3);
  991. #endif
  992. b[ 28] = ZERO;
  993. b[ 29] = ZERO;
  994. b[ 30] = ZERO;
  995. b[ 31] = ZERO;
  996. b[ 32] = *(a05 + 0);
  997. b[ 33] = *(a05 + 1);
  998. b[ 34] = *(a05 + 2);
  999. b[ 35] = *(a05 + 3);
  1000. #ifdef UNIT
  1001. b[ 36] = ONE;
  1002. #else
  1003. b[ 36] = *(a05 + 4);
  1004. #endif
  1005. b[ 37] = ZERO;
  1006. b[ 38] = ZERO;
  1007. b[ 39] = ZERO;
  1008. b[ 40] = *(a06 + 0);
  1009. b[ 41] = *(a06 + 1);
  1010. b[ 42] = *(a06 + 2);
  1011. b[ 43] = *(a06 + 3);
  1012. b[ 44] = *(a06 + 4);
  1013. #ifdef UNIT
  1014. b[ 45] = ONE;
  1015. #else
  1016. b[ 45] = *(a06 + 5);
  1017. #endif
  1018. b[ 46] = ZERO;
  1019. b[ 47] = ZERO;
  1020. b[ 48] = *(a07 + 0);
  1021. b[ 49] = *(a07 + 1);
  1022. b[ 50] = *(a07 + 2);
  1023. b[ 51] = *(a07 + 3);
  1024. b[ 52] = *(a07 + 4);
  1025. b[ 53] = *(a07 + 5);
  1026. #ifdef UNIT
  1027. b[ 54] = ONE;
  1028. #else
  1029. b[ 54] = *(a07 + 6);
  1030. #endif
  1031. b[ 55] = ZERO;
  1032. b[ 56] = *(a08 + 0);
  1033. b[ 57] = *(a08 + 1);
  1034. b[ 58] = *(a08 + 2);
  1035. b[ 59] = *(a08 + 3);
  1036. b[ 60] = *(a08 + 4);
  1037. b[ 61] = *(a08 + 5);
  1038. b[ 62] = *(a08 + 6);
  1039. #ifdef UNIT
  1040. b[ 63] = ONE;
  1041. #else
  1042. b[ 63] = *(a08 + 7);
  1043. #endif
  1044. a01 += 8 * lda;
  1045. a02 += 8 * lda;
  1046. a03 += 8 * lda;
  1047. a04 += 8 * lda;
  1048. a05 += 8 * lda;
  1049. a06 += 8 * lda;
  1050. a07 += 8 * lda;
  1051. a08 += 8 * lda;
  1052. b += 64;
  1053. }
  1054. X += 8;
  1055. i --;
  1056. } while (i > 0);
  1057. }
  1058. i = (m & 7);
  1059. if (i > 0) {
  1060. if (X < posY) {
  1061. a01 += i;
  1062. a02 += i;
  1063. a03 += i;
  1064. a04 += i;
  1065. a05 += i;
  1066. a06 += i;
  1067. a07 += i;
  1068. a08 += i;
  1069. b += 8 * i;
  1070. } else
  1071. if (X > posY) {
  1072. for (ii = 0; ii < i; ii++){
  1073. b[ 0] = *(a01 + 0);
  1074. b[ 1] = *(a01 + 1);
  1075. b[ 2] = *(a01 + 2);
  1076. b[ 3] = *(a01 + 3);
  1077. b[ 4] = *(a01 + 4);
  1078. b[ 5] = *(a01 + 5);
  1079. b[ 6] = *(a01 + 6);
  1080. b[ 7] = *(a01 + 7);
  1081. a01 += lda;
  1082. b += 8;
  1083. }
  1084. a02 += i * lda;
  1085. a03 += i * lda;
  1086. a04 += i * lda;
  1087. a05 += i * lda;
  1088. a06 += i * lda;
  1089. a07 += i * lda;
  1090. a08 += i * lda;
  1091. } else {
  1092. #ifdef UNIT
  1093. b[ 0] = ONE;
  1094. #else
  1095. b[ 0] = *(a01 + 0);
  1096. #endif
  1097. b[ 1] = ZERO;
  1098. b[ 2] = ZERO;
  1099. b[ 3] = ZERO;
  1100. b[ 4] = ZERO;
  1101. b[ 5] = ZERO;
  1102. b[ 6] = ZERO;
  1103. b[ 7] = ZERO;
  1104. b += 8;
  1105. if (i >= 2) {
  1106. b[ 0] = *(a02 + 0);
  1107. #ifdef UNIT
  1108. b[ 1] = ONE;
  1109. #else
  1110. b[ 1] = *(a02 + 1);
  1111. #endif
  1112. b[ 2] = ZERO;
  1113. b[ 3] = ZERO;
  1114. b[ 4] = ZERO;
  1115. b[ 5] = ZERO;
  1116. b[ 6] = ZERO;
  1117. b[ 7] = ZERO;
  1118. b += 8;
  1119. }
  1120. if (i >= 3) {
  1121. b[ 0] = *(a03 + 0);
  1122. b[ 1] = *(a03 + 1);
  1123. #ifdef UNIT
  1124. b[ 2] = ONE;
  1125. #else
  1126. b[ 2] = *(a03 + 2);
  1127. #endif
  1128. b[ 3] = ZERO;
  1129. b[ 4] = ZERO;
  1130. b[ 5] = ZERO;
  1131. b[ 6] = ZERO;
  1132. b[ 7] = ZERO;
  1133. b += 8;
  1134. }
  1135. if (i >= 4) {
  1136. b[ 0] = *(a04 + 0);
  1137. b[ 1] = *(a04 + 1);
  1138. b[ 2] = *(a04 + 2);
  1139. #ifdef UNIT
  1140. b[ 3] = ONE;
  1141. #else
  1142. b[ 3] = *(a04 + 3);
  1143. #endif
  1144. b[ 4] = ZERO;
  1145. b[ 5] = ZERO;
  1146. b[ 6] = ZERO;
  1147. b[ 7] = ZERO;
  1148. b += 8;
  1149. }
  1150. if (i >= 5) {
  1151. b[ 0] = *(a05 + 0);
  1152. b[ 1] = *(a05 + 1);
  1153. b[ 2] = *(a05 + 2);
  1154. b[ 3] = *(a05 + 3);
  1155. #ifdef UNIT
  1156. b[ 4] = ONE;
  1157. #else
  1158. b[ 4] = *(a05 + 4);
  1159. #endif
  1160. b[ 5] = ZERO;
  1161. b[ 6] = ZERO;
  1162. b[ 7] = ZERO;
  1163. b += 8;
  1164. }
  1165. if (i >= 6) {
  1166. b[ 0] = *(a06 + 0);
  1167. b[ 1] = *(a06 + 1);
  1168. b[ 2] = *(a06 + 2);
  1169. b[ 3] = *(a06 + 3);
  1170. b[ 4] = *(a06 + 4);
  1171. #ifdef UNIT
  1172. b[ 5] = ONE;
  1173. #else
  1174. b[ 5] = *(a06 + 5);
  1175. #endif
  1176. b[ 6] = ZERO;
  1177. b[ 7] = ZERO;
  1178. b += 8;
  1179. }
  1180. if (i >= 7) {
  1181. b[ 0] = *(a07 + 0);
  1182. b[ 1] = *(a07 + 1);
  1183. b[ 2] = *(a07 + 2);
  1184. b[ 3] = *(a07 + 3);
  1185. b[ 4] = *(a07 + 4);
  1186. b[ 5] = *(a07 + 5);
  1187. #ifdef UNIT
  1188. b[ 6] = ONE;
  1189. #else
  1190. b[ 6] = *(a07 + 6);
  1191. #endif
  1192. b[ 7] = ZERO;
  1193. b += 8;
  1194. }
  1195. }
  1196. }
  1197. posY += 8;
  1198. }
  1199. if (n & 4){
  1200. X = posX;
  1201. if (posX <= posY) {
  1202. a01 = a + posX + (posY + 0) * lda;
  1203. a02 = a + posX + (posY + 1) * lda;
  1204. a03 = a + posX + (posY + 2) * lda;
  1205. a04 = a + posX + (posY + 3) * lda;
  1206. } else {
  1207. a01 = a + posY + (posX + 0) * lda;
  1208. a02 = a + posY + (posX + 1) * lda;
  1209. a03 = a + posY + (posX + 2) * lda;
  1210. a04 = a + posY + (posX + 3) * lda;
  1211. }
  1212. i = (m >> 2);
  1213. if (i > 0) {
  1214. do {
  1215. if (X < posY) {
  1216. a01 += 4;
  1217. a02 += 4;
  1218. a03 += 4;
  1219. a04 += 4;
  1220. b += 16;
  1221. } else
  1222. if (X > posY) {
  1223. for (ii = 0; ii < 4; ii++){
  1224. b[ 0] = *(a01 + 0);
  1225. b[ 1] = *(a01 + 1);
  1226. b[ 2] = *(a01 + 2);
  1227. b[ 3] = *(a01 + 3);
  1228. a01 += lda;
  1229. b += 4;
  1230. }
  1231. a02 += 4 * lda;
  1232. a03 += 4 * lda;
  1233. a04 += 4 * lda;
  1234. } else {
  1235. #ifdef UNIT
  1236. b[ 0] = ONE;
  1237. #else
  1238. b[ 0] = *(a01 + 0);
  1239. #endif
  1240. b[ 1] = ZERO;
  1241. b[ 2] = ZERO;
  1242. b[ 3] = ZERO;
  1243. b[ 4] = *(a02 + 0);
  1244. #ifdef UNIT
  1245. b[ 5] = ONE;
  1246. #else
  1247. b[ 5] = *(a02 + 1);
  1248. #endif
  1249. b[ 6] = ZERO;
  1250. b[ 7] = ZERO;
  1251. b[ 8] = *(a03 + 0);
  1252. b[ 9] = *(a03 + 1);
  1253. #ifdef UNIT
  1254. b[ 10] = ONE;
  1255. #else
  1256. b[ 10] = *(a03 + 2);
  1257. #endif
  1258. b[ 11] = ZERO;
  1259. b[ 12] = *(a04 + 0);
  1260. b[ 13] = *(a04 + 1);
  1261. b[ 14] = *(a04 + 2);
  1262. #ifdef UNIT
  1263. b[ 15] = ONE;
  1264. #else
  1265. b[ 15] = *(a04 + 3);
  1266. #endif
  1267. a01 += 4 * lda;
  1268. a02 += 4 * lda;
  1269. a03 += 4 * lda;
  1270. a04 += 4 * lda;
  1271. b += 16;
  1272. }
  1273. X += 4;
  1274. i --;
  1275. } while (i > 0);
  1276. }
  1277. i = (m & 3);
  1278. if (i > 0) {
  1279. if (X < posY) {
  1280. a01 += i;
  1281. a02 += i;
  1282. a03 += i;
  1283. a04 += i;
  1284. b += 4 * i;
  1285. } else
  1286. if (X > posY) {
  1287. for (ii = 0; ii < i; ii++){
  1288. b[ 0] = *(a01 + 0);
  1289. b[ 1] = *(a01 + 1);
  1290. b[ 2] = *(a01 + 2);
  1291. b[ 3] = *(a01 + 3);
  1292. a01 += lda;
  1293. b += 4;
  1294. }
  1295. a02 += lda;
  1296. a03 += lda;
  1297. a04 += lda;
  1298. } else {
  1299. #ifdef UNIT
  1300. b[ 0] = ONE;
  1301. #else
  1302. b[ 0] = *(a01 + 0);
  1303. #endif
  1304. b[ 1] = ZERO;
  1305. b[ 2] = ZERO;
  1306. b[ 3] = ZERO;
  1307. b += 4;
  1308. if (i >= 2) {
  1309. b[ 0] = *(a02 + 0);
  1310. #ifdef UNIT
  1311. b[ 1] = ONE;
  1312. #else
  1313. b[ 1] = *(a02 + 1);
  1314. #endif
  1315. b[ 2] = ZERO;
  1316. b[ 3] = ZERO;
  1317. b += 4;
  1318. }
  1319. if (i >= 3) {
  1320. b[ 0] = *(a03 + 0);
  1321. b[ 1] = *(a03 + 1);
  1322. #ifdef UNIT
  1323. b[ 2] = ONE;
  1324. #else
  1325. b[ 2] = *(a03 + 2);
  1326. #endif
  1327. b[ 3] = ZERO;
  1328. b += 4;
  1329. }
  1330. }
  1331. }
  1332. posY += 4;
  1333. }
  1334. if (n & 2){
  1335. X = posX;
  1336. if (posX <= posY) {
  1337. a01 = a + posX + (posY + 0) * lda;
  1338. a02 = a + posX + (posY + 1) * lda;
  1339. } else {
  1340. a01 = a + posY + (posX + 0) * lda;
  1341. a02 = a + posY + (posX + 1) * lda;
  1342. }
  1343. i = (m >> 1);
  1344. if (i > 0) {
  1345. do {
  1346. if (X < posY) {
  1347. a01 += 2;
  1348. a02 += 2;
  1349. b += 4;
  1350. } else
  1351. if (X > posY) {
  1352. b[ 0] = *(a01 + 0);
  1353. b[ 1] = *(a01 + 1);
  1354. b[ 2] = *(a02 + 0);
  1355. b[ 3] = *(a02 + 1);
  1356. a01 += 2 * lda;
  1357. a02 += 2 * lda;
  1358. b += 4;
  1359. } else {
  1360. #ifdef UNIT
  1361. b[ 0] = ONE;
  1362. #else
  1363. b[ 0] = *(a01 + 0);
  1364. #endif
  1365. b[ 1] = ZERO;
  1366. b[ 2] = *(a02 + 0);
  1367. #ifdef UNIT
  1368. b[ 3] = ONE;
  1369. #else
  1370. b[ 3] = *(a02 + 1);
  1371. #endif
  1372. a01 += 2 * lda;
  1373. a02 += 2 * lda;
  1374. b += 4;
  1375. }
  1376. X += 2;
  1377. i --;
  1378. } while (i > 0);
  1379. }
  1380. if (m & 1) {
  1381. if (X < posY) {
  1382. a01 ++;
  1383. a02 ++;
  1384. b += 2;
  1385. } else
  1386. if (X > posY) {
  1387. b[ 0] = *(a01 + 0);
  1388. b[ 1] = *(a01 + 1);
  1389. a01 += lda;
  1390. b += 2;
  1391. } else {
  1392. #ifdef UNIT
  1393. b[ 0] = ONE;
  1394. b[ 1] = *(a01 + 1);
  1395. #else
  1396. b[ 0] = *(a01 + 0);
  1397. b[ 1] = *(a01 + 1);
  1398. #endif
  1399. b += 2;
  1400. }
  1401. }
  1402. posY += 2;
  1403. }
  1404. if (n & 1){
  1405. X = posX;
  1406. if (posX <= posY) {
  1407. a01 = a + posX + (posY + 0) * lda;
  1408. } else {
  1409. a01 = a + posY + (posX + 0) * lda;
  1410. }
  1411. i = m;
  1412. if (i > 0) {
  1413. do {
  1414. if (X < posY) {
  1415. a01 += 1;
  1416. b ++;
  1417. } else
  1418. if (X > posY) {
  1419. b[ 0] = *(a01 + 0);
  1420. a01 += lda;
  1421. b ++;
  1422. } else {
  1423. #ifdef UNIT
  1424. b[ 0] = ONE;
  1425. #else
  1426. b[ 0] = *(a01 + 0);
  1427. #endif
  1428. a01 += lda;
  1429. b ++;
  1430. }
  1431. X += 1;
  1432. i --;
  1433. } while (i > 0);
  1434. }
  1435. posY += 1;
  1436. }
  1437. return 0;
  1438. }