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.

ztrmmkernel_2x2.c 24 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925
  1. #include "common.h"
  2. /********************************
  3. ADD1 a*c
  4. ADD2 b*c
  5. ADD3 a*d
  6. ADD4 b*d
  7. *********************************/
  8. int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alphar,FLOAT alphai,FLOAT* ba,FLOAT* bb,
  9. FLOAT* C,BLASLONG ldc, BLASLONG offset)
  10. {
  11. BLASLONG i,j,k;
  12. FLOAT *C0,*C1,*ptrba,*ptrbb;
  13. FLOAT res0,res1,res2,res3,res4,res5,res6,res7,load0,load1,load2,load3,load4,load5,load6,load7,load8,load9,load10,load11,load12,load13,load14,load15;
  14. BLASLONG off, temp;
  15. #if defined(TRMMKERNEL) && !defined(LEFT)
  16. off = -offset;
  17. #else
  18. off = 0;
  19. #endif
  20. for (j=0; j<bn/2; j+=1)
  21. {
  22. #if defined(TRMMKERNEL) && defined(LEFT)
  23. off = offset;
  24. #endif
  25. C0 = C;
  26. C1 = C0+2*ldc;
  27. ptrba = ba;
  28. for (i=0; i<bm/2; i+=1)
  29. {
  30. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  31. ptrbb = bb;
  32. #else
  33. ptrba += off*2*2;
  34. ptrbb = bb+off*2*2;
  35. #endif
  36. res0 = 0;
  37. res1 = 0;
  38. res2 = 0;
  39. res3 = 0;
  40. res4 = 0;
  41. res5 = 0;
  42. res6 = 0;
  43. res7 = 0;
  44. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  45. temp = bk - off;
  46. #elif defined(LEFT)
  47. temp = off + 2;
  48. #else
  49. temp = off + 2;
  50. #endif
  51. for (k=0; k<temp/4; k+=1)
  52. {
  53. #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
  54. load0 = ptrba[4*0+0];
  55. load1 = ptrbb[4*0+0];
  56. res0 = res0+load0*load1;
  57. load2 = ptrba[4*0+1];
  58. res1 = res1+load2*load1;
  59. load3 = ptrbb[4*0+1];
  60. res0 = res0-load2*load3;
  61. res1 = res1+load0*load3;
  62. load4 = ptrba[4*0+2];
  63. res2 = res2+load4*load1;
  64. load5 = ptrba[4*0+3];
  65. res3 = res3+load5*load1;
  66. res2 = res2-load5*load3;
  67. res3 = res3+load4*load3;
  68. load6 = ptrbb[4*0+2];
  69. res4 = res4+load0*load6;
  70. res5 = res5+load2*load6;
  71. load7 = ptrbb[4*0+3];
  72. res4 = res4-load2*load7;
  73. res5 = res5+load0*load7;
  74. res6 = res6+load4*load6;
  75. res7 = res7+load5*load6;
  76. res6 = res6-load5*load7;
  77. res7 = res7+load4*load7;
  78. load8 = ptrba[4*1+0];
  79. load9 = ptrbb[4*1+0];
  80. res0 = res0+load8*load9;
  81. load10 = ptrba[4*1+1];
  82. res1 = res1+load10*load9;
  83. load11 = ptrbb[4*1+1];
  84. res0 = res0-load10*load11;
  85. res1 = res1+load8*load11;
  86. load12 = ptrba[4*1+2];
  87. res2 = res2+load12*load9;
  88. load13 = ptrba[4*1+3];
  89. res3 = res3+load13*load9;
  90. res2 = res2-load13*load11;
  91. res3 = res3+load12*load11;
  92. load14 = ptrbb[4*1+2];
  93. res4 = res4+load8*load14;
  94. res5 = res5+load10*load14;
  95. load15 = ptrbb[4*1+3];
  96. res4 = res4-load10*load15;
  97. res5 = res5+load8*load15;
  98. res6 = res6+load12*load14;
  99. res7 = res7+load13*load14;
  100. res6 = res6-load13*load15;
  101. res7 = res7+load12*load15;
  102. load0 = ptrba[4*2+0];
  103. load1 = ptrbb[4*2+0];
  104. res0 = res0+load0*load1;
  105. load2 = ptrba[4*2+1];
  106. res1 = res1+load2*load1;
  107. load3 = ptrbb[4*2+1];
  108. res0 = res0-load2*load3;
  109. res1 = res1+load0*load3;
  110. load4 = ptrba[4*2+2];
  111. res2 = res2+load4*load1;
  112. load5 = ptrba[4*2+3];
  113. res3 = res3+load5*load1;
  114. res2 = res2-load5*load3;
  115. res3 = res3+load4*load3;
  116. load6 = ptrbb[4*2+2];
  117. res4 = res4+load0*load6;
  118. res5 = res5+load2*load6;
  119. load7 = ptrbb[4*2+3];
  120. res4 = res4-load2*load7;
  121. res5 = res5+load0*load7;
  122. res6 = res6+load4*load6;
  123. res7 = res7+load5*load6;
  124. res6 = res6-load5*load7;
  125. res7 = res7+load4*load7;
  126. load8 = ptrba[4*3+0];
  127. load9 = ptrbb[4*3+0];
  128. res0 = res0+load8*load9;
  129. load10 = ptrba[4*3+1];
  130. res1 = res1+load10*load9;
  131. load11 = ptrbb[4*3+1];
  132. res0 = res0-load10*load11;
  133. res1 = res1+load8*load11;
  134. load12 = ptrba[4*3+2];
  135. res2 = res2+load12*load9;
  136. load13 = ptrba[4*3+3];
  137. res3 = res3+load13*load9;
  138. res2 = res2-load13*load11;
  139. res3 = res3+load12*load11;
  140. load14 = ptrbb[4*3+2];
  141. res4 = res4+load8*load14;
  142. res5 = res5+load10*load14;
  143. load15 = ptrbb[4*3+3];
  144. res4 = res4-load10*load15;
  145. res5 = res5+load8*load15;
  146. res6 = res6+load12*load14;
  147. res7 = res7+load13*load14;
  148. res6 = res6-load13*load15;
  149. res7 = res7+load12*load15;
  150. #endif
  151. #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
  152. load0 = ptrba[4*0+0];
  153. load1 = ptrbb[4*0+0];
  154. res0 = res0+load0*load1;
  155. load2 = ptrba[4*0+1];
  156. res1 = res1+load2*load1;
  157. load3 = ptrbb[4*0+1];
  158. res0 = res0+load2*load3;
  159. res1 = res1-load0*load3;
  160. load4 = ptrba[4*0+2];
  161. res2 = res2+load4*load1;
  162. load5 = ptrba[4*0+3];
  163. res3 = res3+load5*load1;
  164. res2 = res2+load5*load3;
  165. res3 = res3-load4*load3;
  166. load6 = ptrbb[4*0+2];
  167. res4 = res4+load0*load6;
  168. res5 = res5+load2*load6;
  169. load7 = ptrbb[4*0+3];
  170. res4 = res4+load2*load7;
  171. res5 = res5-load0*load7;
  172. res6 = res6+load4*load6;
  173. res7 = res7+load5*load6;
  174. res6 = res6+load5*load7;
  175. res7 = res7-load4*load7;
  176. load8 = ptrba[4*1+0];
  177. load9 = ptrbb[4*1+0];
  178. res0 = res0+load8*load9;
  179. load10 = ptrba[4*1+1];
  180. res1 = res1+load10*load9;
  181. load11 = ptrbb[4*1+1];
  182. res0 = res0+load10*load11;
  183. res1 = res1-load8*load11;
  184. load12 = ptrba[4*1+2];
  185. res2 = res2+load12*load9;
  186. load13 = ptrba[4*1+3];
  187. res3 = res3+load13*load9;
  188. res2 = res2+load13*load11;
  189. res3 = res3-load12*load11;
  190. load14 = ptrbb[4*1+2];
  191. res4 = res4+load8*load14;
  192. res5 = res5+load10*load14;
  193. load15 = ptrbb[4*1+3];
  194. res4 = res4+load10*load15;
  195. res5 = res5-load8*load15;
  196. res6 = res6+load12*load14;
  197. res7 = res7+load13*load14;
  198. res6 = res6+load13*load15;
  199. res7 = res7-load12*load15;
  200. load0 = ptrba[4*2+0];
  201. load1 = ptrbb[4*2+0];
  202. res0 = res0+load0*load1;
  203. load2 = ptrba[4*2+1];
  204. res1 = res1+load2*load1;
  205. load3 = ptrbb[4*2+1];
  206. res0 = res0+load2*load3;
  207. res1 = res1-load0*load3;
  208. load4 = ptrba[4*2+2];
  209. res2 = res2+load4*load1;
  210. load5 = ptrba[4*2+3];
  211. res3 = res3+load5*load1;
  212. res2 = res2+load5*load3;
  213. res3 = res3-load4*load3;
  214. load6 = ptrbb[4*2+2];
  215. res4 = res4+load0*load6;
  216. res5 = res5+load2*load6;
  217. load7 = ptrbb[4*2+3];
  218. res4 = res4+load2*load7;
  219. res5 = res5-load0*load7;
  220. res6 = res6+load4*load6;
  221. res7 = res7+load5*load6;
  222. res6 = res6+load5*load7;
  223. res7 = res7-load4*load7;
  224. load8 = ptrba[4*3+0];
  225. load9 = ptrbb[4*3+0];
  226. res0 = res0+load8*load9;
  227. load10 = ptrba[4*3+1];
  228. res1 = res1+load10*load9;
  229. load11 = ptrbb[4*3+1];
  230. res0 = res0+load10*load11;
  231. res1 = res1-load8*load11;
  232. load12 = ptrba[4*3+2];
  233. res2 = res2+load12*load9;
  234. load13 = ptrba[4*3+3];
  235. res3 = res3+load13*load9;
  236. res2 = res2+load13*load11;
  237. res3 = res3-load12*load11;
  238. load14 = ptrbb[4*3+2];
  239. res4 = res4+load8*load14;
  240. res5 = res5+load10*load14;
  241. load15 = ptrbb[4*3+3];
  242. res4 = res4+load10*load15;
  243. res5 = res5-load8*load15;
  244. res6 = res6+load12*load14;
  245. res7 = res7+load13*load14;
  246. res6 = res6+load13*load15;
  247. res7 = res7-load12*load15;
  248. #endif
  249. #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
  250. load0 = ptrba[4*0+0];
  251. load1 = ptrbb[4*0+0];
  252. res0 = res0+load0*load1;
  253. load2 = ptrba[4*0+1];
  254. res1 = res1-load2*load1;
  255. load3 = ptrbb[4*0+1];
  256. res0 = res0+load2*load3;
  257. res1 = res1+load0*load3;
  258. load4 = ptrba[4*0+2];
  259. res2 = res2+load4*load1;
  260. load5 = ptrba[4*0+3];
  261. res3 = res3-load5*load1;
  262. res2 = res2+load5*load3;
  263. res3 = res3+load4*load3;
  264. load6 = ptrbb[4*0+2];
  265. res4 = res4+load0*load6;
  266. res5 = res5-load2*load6;
  267. load7 = ptrbb[4*0+3];
  268. res4 = res4+load2*load7;
  269. res5 = res5+load0*load7;
  270. res6 = res6+load4*load6;
  271. res7 = res7-load5*load6;
  272. res6 = res6+load5*load7;
  273. res7 = res7+load4*load7;
  274. load8 = ptrba[4*1+0];
  275. load9 = ptrbb[4*1+0];
  276. res0 = res0+load8*load9;
  277. load10 = ptrba[4*1+1];
  278. res1 = res1-load10*load9;
  279. load11 = ptrbb[4*1+1];
  280. res0 = res0+load10*load11;
  281. res1 = res1+load8*load11;
  282. load12 = ptrba[4*1+2];
  283. res2 = res2+load12*load9;
  284. load13 = ptrba[4*1+3];
  285. res3 = res3-load13*load9;
  286. res2 = res2+load13*load11;
  287. res3 = res3+load12*load11;
  288. load14 = ptrbb[4*1+2];
  289. res4 = res4+load8*load14;
  290. res5 = res5-load10*load14;
  291. load15 = ptrbb[4*1+3];
  292. res4 = res4+load10*load15;
  293. res5 = res5+load8*load15;
  294. res6 = res6+load12*load14;
  295. res7 = res7-load13*load14;
  296. res6 = res6+load13*load15;
  297. res7 = res7+load12*load15;
  298. load0 = ptrba[4*2+0];
  299. load1 = ptrbb[4*2+0];
  300. res0 = res0+load0*load1;
  301. load2 = ptrba[4*2+1];
  302. res1 = res1-load2*load1;
  303. load3 = ptrbb[4*2+1];
  304. res0 = res0+load2*load3;
  305. res1 = res1+load0*load3;
  306. load4 = ptrba[4*2+2];
  307. res2 = res2+load4*load1;
  308. load5 = ptrba[4*2+3];
  309. res3 = res3-load5*load1;
  310. res2 = res2+load5*load3;
  311. res3 = res3+load4*load3;
  312. load6 = ptrbb[4*2+2];
  313. res4 = res4+load0*load6;
  314. res5 = res5-load2*load6;
  315. load7 = ptrbb[4*2+3];
  316. res4 = res4+load2*load7;
  317. res5 = res5+load0*load7;
  318. res6 = res6+load4*load6;
  319. res7 = res7-load5*load6;
  320. res6 = res6+load5*load7;
  321. res7 = res7+load4*load7;
  322. load8 = ptrba[4*3+0];
  323. load9 = ptrbb[4*3+0];
  324. res0 = res0+load8*load9;
  325. load10 = ptrba[4*3+1];
  326. res1 = res1-load10*load9;
  327. load11 = ptrbb[4*3+1];
  328. res0 = res0+load10*load11;
  329. res1 = res1+load8*load11;
  330. load12 = ptrba[4*3+2];
  331. res2 = res2+load12*load9;
  332. load13 = ptrba[4*3+3];
  333. res3 = res3-load13*load9;
  334. res2 = res2+load13*load11;
  335. res3 = res3+load12*load11;
  336. load14 = ptrbb[4*3+2];
  337. res4 = res4+load8*load14;
  338. res5 = res5-load10*load14;
  339. load15 = ptrbb[4*3+3];
  340. res4 = res4+load10*load15;
  341. res5 = res5+load8*load15;
  342. res6 = res6+load12*load14;
  343. res7 = res7-load13*load14;
  344. res6 = res6+load13*load15;
  345. res7 = res7+load12*load15;
  346. #endif
  347. #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
  348. load0 = ptrba[4*0+0];
  349. load1 = ptrbb[4*0+0];
  350. res0 = res0+load0*load1;
  351. load2 = ptrba[4*0+1];
  352. res1 = res1-load2*load1;
  353. load3 = ptrbb[4*0+1];
  354. res0 = res0-load2*load3;
  355. res1 = res1-load0*load3;
  356. load4 = ptrba[4*0+2];
  357. res2 = res2+load4*load1;
  358. load5 = ptrba[4*0+3];
  359. res3 = res3-load5*load1;
  360. res2 = res2-load5*load3;
  361. res3 = res3-load4*load3;
  362. load6 = ptrbb[4*0+2];
  363. res4 = res4+load0*load6;
  364. res5 = res5-load2*load6;
  365. load7 = ptrbb[4*0+3];
  366. res4 = res4-load2*load7;
  367. res5 = res5-load0*load7;
  368. res6 = res6+load4*load6;
  369. res7 = res7-load5*load6;
  370. res6 = res6-load5*load7;
  371. res7 = res7-load4*load7;
  372. load8 = ptrba[4*1+0];
  373. load9 = ptrbb[4*1+0];
  374. res0 = res0+load8*load9;
  375. load10 = ptrba[4*1+1];
  376. res1 = res1-load10*load9;
  377. load11 = ptrbb[4*1+1];
  378. res0 = res0-load10*load11;
  379. res1 = res1-load8*load11;
  380. load12 = ptrba[4*1+2];
  381. res2 = res2+load12*load9;
  382. load13 = ptrba[4*1+3];
  383. res3 = res3-load13*load9;
  384. res2 = res2-load13*load11;
  385. res3 = res3-load12*load11;
  386. load14 = ptrbb[4*1+2];
  387. res4 = res4+load8*load14;
  388. res5 = res5-load10*load14;
  389. load15 = ptrbb[4*1+3];
  390. res4 = res4-load10*load15;
  391. res5 = res5-load8*load15;
  392. res6 = res6+load12*load14;
  393. res7 = res7-load13*load14;
  394. res6 = res6-load13*load15;
  395. res7 = res7-load12*load15;
  396. load0 = ptrba[4*2+0];
  397. load1 = ptrbb[4*2+0];
  398. res0 = res0+load0*load1;
  399. load2 = ptrba[4*2+1];
  400. res1 = res1-load2*load1;
  401. load3 = ptrbb[4*2+1];
  402. res0 = res0-load2*load3;
  403. res1 = res1-load0*load3;
  404. load4 = ptrba[4*2+2];
  405. res2 = res2+load4*load1;
  406. load5 = ptrba[4*2+3];
  407. res3 = res3-load5*load1;
  408. res2 = res2-load5*load3;
  409. res3 = res3-load4*load3;
  410. load6 = ptrbb[4*2+2];
  411. res4 = res4+load0*load6;
  412. res5 = res5-load2*load6;
  413. load7 = ptrbb[4*2+3];
  414. res4 = res4-load2*load7;
  415. res5 = res5-load0*load7;
  416. res6 = res6+load4*load6;
  417. res7 = res7-load5*load6;
  418. res6 = res6-load5*load7;
  419. res7 = res7-load4*load7;
  420. load8 = ptrba[4*3+0];
  421. load9 = ptrbb[4*3+0];
  422. res0 = res0+load8*load9;
  423. load10 = ptrba[4*3+1];
  424. res1 = res1-load10*load9;
  425. load11 = ptrbb[4*3+1];
  426. res0 = res0-load10*load11;
  427. res1 = res1-load8*load11;
  428. load12 = ptrba[4*3+2];
  429. res2 = res2+load12*load9;
  430. load13 = ptrba[4*3+3];
  431. res3 = res3-load13*load9;
  432. res2 = res2-load13*load11;
  433. res3 = res3-load12*load11;
  434. load14 = ptrbb[4*3+2];
  435. res4 = res4+load8*load14;
  436. res5 = res5-load10*load14;
  437. load15 = ptrbb[4*3+3];
  438. res4 = res4-load10*load15;
  439. res5 = res5-load8*load15;
  440. res6 = res6+load12*load14;
  441. res7 = res7-load13*load14;
  442. res6 = res6-load13*load15;
  443. res7 = res7-load12*load15;
  444. #endif
  445. ptrba = ptrba+16;
  446. ptrbb = ptrbb+16;
  447. }
  448. for (k=0; k<(temp&3); k+=1)
  449. {
  450. #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
  451. load0 = ptrba[4*0+0];
  452. load1 = ptrbb[4*0+0];
  453. res0 = res0+load0*load1;
  454. load2 = ptrba[4*0+1];
  455. res1 = res1+load2*load1;
  456. load3 = ptrbb[4*0+1];
  457. res0 = res0-load2*load3;
  458. res1 = res1+load0*load3;
  459. load4 = ptrba[4*0+2];
  460. res2 = res2+load4*load1;
  461. load5 = ptrba[4*0+3];
  462. res3 = res3+load5*load1;
  463. res2 = res2-load5*load3;
  464. res3 = res3+load4*load3;
  465. load6 = ptrbb[4*0+2];
  466. res4 = res4+load0*load6;
  467. res5 = res5+load2*load6;
  468. load7 = ptrbb[4*0+3];
  469. res4 = res4-load2*load7;
  470. res5 = res5+load0*load7;
  471. res6 = res6+load4*load6;
  472. res7 = res7+load5*load6;
  473. res6 = res6-load5*load7;
  474. res7 = res7+load4*load7;
  475. #endif
  476. #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
  477. load0 = ptrba[4*0+0];
  478. load1 = ptrbb[4*0+0];
  479. res0 = res0+load0*load1;
  480. load2 = ptrba[4*0+1];
  481. res1 = res1+load2*load1;
  482. load3 = ptrbb[4*0+1];
  483. res0 = res0+load2*load3;
  484. res1 = res1-load0*load3;
  485. load4 = ptrba[4*0+2];
  486. res2 = res2+load4*load1;
  487. load5 = ptrba[4*0+3];
  488. res3 = res3+load5*load1;
  489. res2 = res2+load5*load3;
  490. res3 = res3-load4*load3;
  491. load6 = ptrbb[4*0+2];
  492. res4 = res4+load0*load6;
  493. res5 = res5+load2*load6;
  494. load7 = ptrbb[4*0+3];
  495. res4 = res4+load2*load7;
  496. res5 = res5-load0*load7;
  497. res6 = res6+load4*load6;
  498. res7 = res7+load5*load6;
  499. res6 = res6+load5*load7;
  500. res7 = res7-load4*load7;
  501. #endif
  502. #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
  503. load0 = ptrba[4*0+0];
  504. load1 = ptrbb[4*0+0];
  505. res0 = res0+load0*load1;
  506. load2 = ptrba[4*0+1];
  507. res1 = res1-load2*load1;
  508. load3 = ptrbb[4*0+1];
  509. res0 = res0+load2*load3;
  510. res1 = res1+load0*load3;
  511. load4 = ptrba[4*0+2];
  512. res2 = res2+load4*load1;
  513. load5 = ptrba[4*0+3];
  514. res3 = res3-load5*load1;
  515. res2 = res2+load5*load3;
  516. res3 = res3+load4*load3;
  517. load6 = ptrbb[4*0+2];
  518. res4 = res4+load0*load6;
  519. res5 = res5-load2*load6;
  520. load7 = ptrbb[4*0+3];
  521. res4 = res4+load2*load7;
  522. res5 = res5+load0*load7;
  523. res6 = res6+load4*load6;
  524. res7 = res7-load5*load6;
  525. res6 = res6+load5*load7;
  526. res7 = res7+load4*load7;
  527. #endif
  528. #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
  529. load0 = ptrba[4*0+0];
  530. load1 = ptrbb[4*0+0];
  531. res0 = res0+load0*load1;
  532. load2 = ptrba[4*0+1];
  533. res1 = res1-load2*load1;
  534. load3 = ptrbb[4*0+1];
  535. res0 = res0-load2*load3;
  536. res1 = res1-load0*load3;
  537. load4 = ptrba[4*0+2];
  538. res2 = res2+load4*load1;
  539. load5 = ptrba[4*0+3];
  540. res3 = res3-load5*load1;
  541. res2 = res2-load5*load3;
  542. res3 = res3-load4*load3;
  543. load6 = ptrbb[4*0+2];
  544. res4 = res4+load0*load6;
  545. res5 = res5-load2*load6;
  546. load7 = ptrbb[4*0+3];
  547. res4 = res4-load2*load7;
  548. res5 = res5-load0*load7;
  549. res6 = res6+load4*load6;
  550. res7 = res7-load5*load6;
  551. res6 = res6-load5*load7;
  552. res7 = res7-load4*load7;
  553. #endif
  554. ptrba = ptrba+4;
  555. ptrbb = ptrbb+4;
  556. }
  557. load0 = res0*alphar-res1*alphai;
  558. load1 = res1*alphar+res0*alphai;
  559. C0[0] = load0;
  560. C0[1] = load1;
  561. load2 = res2*alphar-res3*alphai;
  562. load3 = res3*alphar+res2*alphai;
  563. C0[2] = load2;
  564. C0[3] = load3;
  565. load4 = res4*alphar-res5*alphai;
  566. load5 = res5*alphar+res4*alphai;
  567. C1[0] = load4;
  568. C1[1] = load5;
  569. load6 = res6*alphar-res7*alphai;
  570. load7 = res7*alphar+res6*alphai;
  571. C1[2] = load6;
  572. C1[3] = load7;
  573. #if ( defined(LEFT) && defined(TRANSA)) || \
  574. (!defined(LEFT) && !defined(TRANSA))
  575. temp = bk - off;
  576. #ifdef LEFT
  577. temp -= 2;
  578. #else
  579. temp -= 2;
  580. #endif
  581. ptrba += temp*2*2;
  582. ptrbb += temp*2*2;
  583. #endif
  584. #ifdef LEFT
  585. off += 2;
  586. #endif
  587. C0 = C0+4;
  588. C1 = C1+4;
  589. }
  590. for (i=0; i<(bm&1); i+=1)
  591. {
  592. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  593. ptrbb = bb;
  594. #else
  595. ptrba += off*2;
  596. ptrbb = bb + off*2*2;
  597. #endif
  598. res0 = 0;
  599. res1 = 0;
  600. res2 = 0;
  601. res3 = 0;
  602. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  603. temp = bk - off;
  604. #elif defined(LEFT)
  605. temp = off+1;
  606. #else
  607. temp = off+2;
  608. #endif
  609. for (k=0; k<temp; k+=1)
  610. {
  611. #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
  612. load0 = ptrba[2*0+0];
  613. load1 = ptrbb[4*0+0];
  614. res0 = res0+load0*load1;
  615. load2 = ptrba[2*0+1];
  616. res1 = res1+load2*load1;
  617. load3 = ptrbb[4*0+1];
  618. res0 = res0-load2*load3;
  619. res1 = res1+load0*load3;
  620. load4 = ptrbb[4*0+2];
  621. res2 = res2+load0*load4;
  622. res3 = res3+load2*load4;
  623. load5 = ptrbb[4*0+3];
  624. res2 = res2-load2*load5;
  625. res3 = res3+load0*load5;
  626. #endif
  627. #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
  628. load0 = ptrba[2*0+0];
  629. load1 = ptrbb[4*0+0];
  630. res0 = res0+load0*load1;
  631. load2 = ptrba[2*0+1];
  632. res1 = res1+load2*load1;
  633. load3 = ptrbb[4*0+1];
  634. res0 = res0+load2*load3;
  635. res1 = res1-load0*load3;
  636. load4 = ptrbb[4*0+2];
  637. res2 = res2+load0*load4;
  638. res3 = res3+load2*load4;
  639. load5 = ptrbb[4*0+3];
  640. res2 = res2+load2*load5;
  641. res3 = res3-load0*load5;
  642. #endif
  643. #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
  644. load0 = ptrba[2*0+0];
  645. load1 = ptrbb[4*0+0];
  646. res0 = res0+load0*load1;
  647. load2 = ptrba[2*0+1];
  648. res1 = res1-load2*load1;
  649. load3 = ptrbb[4*0+1];
  650. res0 = res0+load2*load3;
  651. res1 = res1+load0*load3;
  652. load4 = ptrbb[4*0+2];
  653. res2 = res2+load0*load4;
  654. res3 = res3-load2*load4;
  655. load5 = ptrbb[4*0+3];
  656. res2 = res2+load2*load5;
  657. res3 = res3+load0*load5;
  658. #endif
  659. #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
  660. load0 = ptrba[2*0+0];
  661. load1 = ptrbb[4*0+0];
  662. res0 = res0+load0*load1;
  663. load2 = ptrba[2*0+1];
  664. res1 = res1-load2*load1;
  665. load3 = ptrbb[4*0+1];
  666. res0 = res0-load2*load3;
  667. res1 = res1-load0*load3;
  668. load4 = ptrbb[4*0+2];
  669. res2 = res2+load0*load4;
  670. res3 = res3-load2*load4;
  671. load5 = ptrbb[4*0+3];
  672. res2 = res2-load2*load5;
  673. res3 = res3-load0*load5;
  674. #endif
  675. ptrba = ptrba+2;
  676. ptrbb = ptrbb+4;
  677. }
  678. load0 = res0*alphar-res1*alphai;
  679. load1 = res1*alphar+res0*alphai;
  680. C0[0] = load0;
  681. C0[1] = load1;
  682. load2 = res2*alphar-res3*alphai;
  683. load3 = res3*alphar+res2*alphai;
  684. C1[0] = load2;
  685. C1[1] = load3;
  686. #if ( defined(LEFT) && defined(TRANSA)) || \
  687. (!defined(LEFT) && !defined(TRANSA))
  688. temp = bk - off;
  689. #ifdef LEFT
  690. temp -= 1;
  691. #else
  692. temp -= 2;
  693. #endif
  694. ptrba += temp*2;
  695. ptrbb += temp*2*2;
  696. #endif
  697. #ifdef LEFT
  698. off += 1;
  699. #endif
  700. C0 = C0+2;
  701. C1 = C1+2;
  702. }
  703. #if defined(TRMMKERNEL) && !defined(LEFT)
  704. off += 2;
  705. #endif
  706. k = (bk<<2);
  707. bb = bb+k;
  708. i = (ldc<<2);
  709. C = C+i;
  710. }
  711. for (j=0; j<(bn&1); j+=1)
  712. {
  713. C0 = C;
  714. #if defined(TRMMKERNEL) && defined(LEFT)
  715. off = offset;
  716. #endif
  717. ptrba = ba;
  718. for (i=0; i<bm/2; i+=1)
  719. {
  720. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  721. ptrbb = bb;
  722. #else
  723. ptrba += off*2*2;
  724. ptrbb = bb+off*2;
  725. #endif
  726. res0 = 0;
  727. res1 = 0;
  728. res2 = 0;
  729. res3 = 0;
  730. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  731. temp = bk - off;
  732. #elif defined(LEFT)
  733. temp = off + 2;
  734. #else
  735. temp = off + 1;
  736. #endif
  737. for (k=0; k<temp; k+=1)
  738. {
  739. #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
  740. load0 = ptrba[4*0+0];
  741. load1 = ptrbb[2*0+0];
  742. res0 = res0+load0*load1;
  743. load2 = ptrba[4*0+1];
  744. res1 = res1+load2*load1;
  745. load3 = ptrbb[2*0+1];
  746. res0 = res0-load2*load3;
  747. res1 = res1+load0*load3;
  748. load4 = ptrba[4*0+2];
  749. res2 = res2+load4*load1;
  750. load5 = ptrba[4*0+3];
  751. res3 = res3+load5*load1;
  752. res2 = res2-load5*load3;
  753. res3 = res3+load4*load3;
  754. #endif
  755. #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
  756. load0 = ptrba[4*0+0];
  757. load1 = ptrbb[2*0+0];
  758. res0 = res0+load0*load1;
  759. load2 = ptrba[4*0+1];
  760. res1 = res1+load2*load1;
  761. load3 = ptrbb[2*0+1];
  762. res0 = res0+load2*load3;
  763. res1 = res1-load0*load3;
  764. load4 = ptrba[4*0+2];
  765. res2 = res2+load4*load1;
  766. load5 = ptrba[4*0+3];
  767. res3 = res3+load5*load1;
  768. res2 = res2+load5*load3;
  769. res3 = res3-load4*load3;
  770. #endif
  771. #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
  772. load0 = ptrba[4*0+0];
  773. load1 = ptrbb[2*0+0];
  774. res0 = res0+load0*load1;
  775. load2 = ptrba[4*0+1];
  776. res1 = res1-load2*load1;
  777. load3 = ptrbb[2*0+1];
  778. res0 = res0+load2*load3;
  779. res1 = res1+load0*load3;
  780. load4 = ptrba[4*0+2];
  781. res2 = res2+load4*load1;
  782. load5 = ptrba[4*0+3];
  783. res3 = res3-load5*load1;
  784. res2 = res2+load5*load3;
  785. res3 = res3+load4*load3;
  786. #endif
  787. #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
  788. load0 = ptrba[4*0+0];
  789. load1 = ptrbb[2*0+0];
  790. res0 = res0+load0*load1;
  791. load2 = ptrba[4*0+1];
  792. res1 = res1-load2*load1;
  793. load3 = ptrbb[2*0+1];
  794. res0 = res0-load2*load3;
  795. res1 = res1-load0*load3;
  796. load4 = ptrba[4*0+2];
  797. res2 = res2+load4*load1;
  798. load5 = ptrba[4*0+3];
  799. res3 = res3-load5*load1;
  800. res2 = res2-load5*load3;
  801. res3 = res3-load4*load3;
  802. #endif
  803. ptrba = ptrba+4;
  804. ptrbb = ptrbb+2;
  805. }
  806. load0 = res0*alphar-res1*alphai;
  807. load1 = res1*alphar+res0*alphai;
  808. C0[0] = load0;
  809. C0[1] = load1;
  810. load2 = res2*alphar-res3*alphai;
  811. load3 = res3*alphar+res2*alphai;
  812. C0[2] = load2;
  813. C0[3] = load3;
  814. #if ( defined(LEFT) && defined(TRANSA)) || \
  815. (!defined(LEFT) && !defined(TRANSA))
  816. temp = bk-off;
  817. #ifdef LEFT
  818. temp -= 2;
  819. #else
  820. temp -= 1;
  821. #endif
  822. ptrba += temp*2*2;
  823. ptrbb += temp*2;
  824. #endif
  825. #ifdef LEFT
  826. off += 2;
  827. #endif
  828. C0 = C0+4;
  829. }
  830. for (i=0; i<(bm&1); i+=1)
  831. {
  832. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  833. ptrbb = bb;
  834. #else
  835. ptrba += off*2;
  836. ptrbb = bb + off*2;
  837. #endif
  838. res0 = 0;
  839. res1 = 0;
  840. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  841. temp = bk-off;
  842. #elif defined(LEFT)
  843. temp = off + 1;
  844. #else
  845. temp = off + 1;
  846. #endif
  847. for (k=0; k<temp; k+=1)
  848. {
  849. #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
  850. load0 = ptrba[2*0+0];
  851. load1 = ptrbb[2*0+0];
  852. res0 = res0+load0*load1;
  853. load2 = ptrba[2*0+1];
  854. res1 = res1+load2*load1;
  855. load3 = ptrbb[2*0+1];
  856. res0 = res0-load2*load3;
  857. res1 = res1+load0*load3;
  858. #endif
  859. #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
  860. load0 = ptrba[2*0+0];
  861. load1 = ptrbb[2*0+0];
  862. res0 = res0+load0*load1;
  863. load2 = ptrba[2*0+1];
  864. res1 = res1+load2*load1;
  865. load3 = ptrbb[2*0+1];
  866. res0 = res0+load2*load3;
  867. res1 = res1-load0*load3;
  868. #endif
  869. #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
  870. load0 = ptrba[2*0+0];
  871. load1 = ptrbb[2*0+0];
  872. res0 = res0+load0*load1;
  873. load2 = ptrba[2*0+1];
  874. res1 = res1-load2*load1;
  875. load3 = ptrbb[2*0+1];
  876. res0 = res0+load2*load3;
  877. res1 = res1+load0*load3;
  878. #endif
  879. #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
  880. load0 = ptrba[2*0+0];
  881. load1 = ptrbb[2*0+0];
  882. res0 = res0+load0*load1;
  883. load2 = ptrba[2*0+1];
  884. res1 = res1-load2*load1;
  885. load3 = ptrbb[2*0+1];
  886. res0 = res0-load2*load3;
  887. res1 = res1-load0*load3;
  888. #endif
  889. ptrba = ptrba+2;
  890. ptrbb = ptrbb+2;
  891. }
  892. load0 = res0*alphar-res1*alphai;
  893. load1 = res1*alphar+res0*alphai;
  894. C0[0] = load0;
  895. C0[1] = load1;
  896. #if ( defined(LEFT) && defined(TRANSA)) || \
  897. (!defined(LEFT) && !defined(TRANSA))
  898. temp = bk - off;
  899. #ifdef LEFT
  900. temp -= 1;
  901. #else
  902. temp -= 1;
  903. #endif
  904. ptrba += temp*2;
  905. ptrbb += temp*2;
  906. #endif
  907. #ifdef LEFT
  908. off += 1;
  909. #endif
  910. C0 = C0+2;
  911. }
  912. k = (bk<<1);
  913. bb = bb+k;
  914. i = (ldc<<1);
  915. C = C+i;
  916. }
  917. return 0;
  918. }