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

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923
  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. #endif
  18. for (j=0; j<bn/2; j+=1)
  19. {
  20. #if defined(TRMMKERNEL) && defined(LEFT)
  21. off = offset;
  22. #endif
  23. C0 = C;
  24. C1 = C0+2*ldc;
  25. ptrba = ba;
  26. for (i=0; i<bm/2; i+=1)
  27. {
  28. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  29. ptrbb = bb;
  30. #else
  31. ptrba += off*2*2;
  32. ptrbb = bb+off*2*2;
  33. #endif
  34. res0 = 0;
  35. res1 = 0;
  36. res2 = 0;
  37. res3 = 0;
  38. res4 = 0;
  39. res5 = 0;
  40. res6 = 0;
  41. res7 = 0;
  42. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  43. temp = bk - off;
  44. #elif defined(LEFT)
  45. temp = off + 2;
  46. #else
  47. temp = off + 2;
  48. #endif
  49. for (k=0; k<temp/4; k+=1)
  50. {
  51. #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
  52. load0 = ptrba[4*0+0];
  53. load1 = ptrbb[4*0+0];
  54. res0 = res0+load0*load1;
  55. load2 = ptrba[4*0+1];
  56. res1 = res1+load2*load1;
  57. load3 = ptrbb[4*0+1];
  58. res0 = res0-load2*load3;
  59. res1 = res1+load0*load3;
  60. load4 = ptrba[4*0+2];
  61. res2 = res2+load4*load1;
  62. load5 = ptrba[4*0+3];
  63. res3 = res3+load5*load1;
  64. res2 = res2-load5*load3;
  65. res3 = res3+load4*load3;
  66. load6 = ptrbb[4*0+2];
  67. res4 = res4+load0*load6;
  68. res5 = res5+load2*load6;
  69. load7 = ptrbb[4*0+3];
  70. res4 = res4-load2*load7;
  71. res5 = res5+load0*load7;
  72. res6 = res6+load4*load6;
  73. res7 = res7+load5*load6;
  74. res6 = res6-load5*load7;
  75. res7 = res7+load4*load7;
  76. load8 = ptrba[4*1+0];
  77. load9 = ptrbb[4*1+0];
  78. res0 = res0+load8*load9;
  79. load10 = ptrba[4*1+1];
  80. res1 = res1+load10*load9;
  81. load11 = ptrbb[4*1+1];
  82. res0 = res0-load10*load11;
  83. res1 = res1+load8*load11;
  84. load12 = ptrba[4*1+2];
  85. res2 = res2+load12*load9;
  86. load13 = ptrba[4*1+3];
  87. res3 = res3+load13*load9;
  88. res2 = res2-load13*load11;
  89. res3 = res3+load12*load11;
  90. load14 = ptrbb[4*1+2];
  91. res4 = res4+load8*load14;
  92. res5 = res5+load10*load14;
  93. load15 = ptrbb[4*1+3];
  94. res4 = res4-load10*load15;
  95. res5 = res5+load8*load15;
  96. res6 = res6+load12*load14;
  97. res7 = res7+load13*load14;
  98. res6 = res6-load13*load15;
  99. res7 = res7+load12*load15;
  100. load0 = ptrba[4*2+0];
  101. load1 = ptrbb[4*2+0];
  102. res0 = res0+load0*load1;
  103. load2 = ptrba[4*2+1];
  104. res1 = res1+load2*load1;
  105. load3 = ptrbb[4*2+1];
  106. res0 = res0-load2*load3;
  107. res1 = res1+load0*load3;
  108. load4 = ptrba[4*2+2];
  109. res2 = res2+load4*load1;
  110. load5 = ptrba[4*2+3];
  111. res3 = res3+load5*load1;
  112. res2 = res2-load5*load3;
  113. res3 = res3+load4*load3;
  114. load6 = ptrbb[4*2+2];
  115. res4 = res4+load0*load6;
  116. res5 = res5+load2*load6;
  117. load7 = ptrbb[4*2+3];
  118. res4 = res4-load2*load7;
  119. res5 = res5+load0*load7;
  120. res6 = res6+load4*load6;
  121. res7 = res7+load5*load6;
  122. res6 = res6-load5*load7;
  123. res7 = res7+load4*load7;
  124. load8 = ptrba[4*3+0];
  125. load9 = ptrbb[4*3+0];
  126. res0 = res0+load8*load9;
  127. load10 = ptrba[4*3+1];
  128. res1 = res1+load10*load9;
  129. load11 = ptrbb[4*3+1];
  130. res0 = res0-load10*load11;
  131. res1 = res1+load8*load11;
  132. load12 = ptrba[4*3+2];
  133. res2 = res2+load12*load9;
  134. load13 = ptrba[4*3+3];
  135. res3 = res3+load13*load9;
  136. res2 = res2-load13*load11;
  137. res3 = res3+load12*load11;
  138. load14 = ptrbb[4*3+2];
  139. res4 = res4+load8*load14;
  140. res5 = res5+load10*load14;
  141. load15 = ptrbb[4*3+3];
  142. res4 = res4-load10*load15;
  143. res5 = res5+load8*load15;
  144. res6 = res6+load12*load14;
  145. res7 = res7+load13*load14;
  146. res6 = res6-load13*load15;
  147. res7 = res7+load12*load15;
  148. #endif
  149. #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
  150. load0 = ptrba[4*0+0];
  151. load1 = ptrbb[4*0+0];
  152. res0 = res0+load0*load1;
  153. load2 = ptrba[4*0+1];
  154. res1 = res1+load2*load1;
  155. load3 = ptrbb[4*0+1];
  156. res0 = res0+load2*load3;
  157. res1 = res1-load0*load3;
  158. load4 = ptrba[4*0+2];
  159. res2 = res2+load4*load1;
  160. load5 = ptrba[4*0+3];
  161. res3 = res3+load5*load1;
  162. res2 = res2+load5*load3;
  163. res3 = res3-load4*load3;
  164. load6 = ptrbb[4*0+2];
  165. res4 = res4+load0*load6;
  166. res5 = res5+load2*load6;
  167. load7 = ptrbb[4*0+3];
  168. res4 = res4+load2*load7;
  169. res5 = res5-load0*load7;
  170. res6 = res6+load4*load6;
  171. res7 = res7+load5*load6;
  172. res6 = res6+load5*load7;
  173. res7 = res7-load4*load7;
  174. load8 = ptrba[4*1+0];
  175. load9 = ptrbb[4*1+0];
  176. res0 = res0+load8*load9;
  177. load10 = ptrba[4*1+1];
  178. res1 = res1+load10*load9;
  179. load11 = ptrbb[4*1+1];
  180. res0 = res0+load10*load11;
  181. res1 = res1-load8*load11;
  182. load12 = ptrba[4*1+2];
  183. res2 = res2+load12*load9;
  184. load13 = ptrba[4*1+3];
  185. res3 = res3+load13*load9;
  186. res2 = res2+load13*load11;
  187. res3 = res3-load12*load11;
  188. load14 = ptrbb[4*1+2];
  189. res4 = res4+load8*load14;
  190. res5 = res5+load10*load14;
  191. load15 = ptrbb[4*1+3];
  192. res4 = res4+load10*load15;
  193. res5 = res5-load8*load15;
  194. res6 = res6+load12*load14;
  195. res7 = res7+load13*load14;
  196. res6 = res6+load13*load15;
  197. res7 = res7-load12*load15;
  198. load0 = ptrba[4*2+0];
  199. load1 = ptrbb[4*2+0];
  200. res0 = res0+load0*load1;
  201. load2 = ptrba[4*2+1];
  202. res1 = res1+load2*load1;
  203. load3 = ptrbb[4*2+1];
  204. res0 = res0+load2*load3;
  205. res1 = res1-load0*load3;
  206. load4 = ptrba[4*2+2];
  207. res2 = res2+load4*load1;
  208. load5 = ptrba[4*2+3];
  209. res3 = res3+load5*load1;
  210. res2 = res2+load5*load3;
  211. res3 = res3-load4*load3;
  212. load6 = ptrbb[4*2+2];
  213. res4 = res4+load0*load6;
  214. res5 = res5+load2*load6;
  215. load7 = ptrbb[4*2+3];
  216. res4 = res4+load2*load7;
  217. res5 = res5-load0*load7;
  218. res6 = res6+load4*load6;
  219. res7 = res7+load5*load6;
  220. res6 = res6+load5*load7;
  221. res7 = res7-load4*load7;
  222. load8 = ptrba[4*3+0];
  223. load9 = ptrbb[4*3+0];
  224. res0 = res0+load8*load9;
  225. load10 = ptrba[4*3+1];
  226. res1 = res1+load10*load9;
  227. load11 = ptrbb[4*3+1];
  228. res0 = res0+load10*load11;
  229. res1 = res1-load8*load11;
  230. load12 = ptrba[4*3+2];
  231. res2 = res2+load12*load9;
  232. load13 = ptrba[4*3+3];
  233. res3 = res3+load13*load9;
  234. res2 = res2+load13*load11;
  235. res3 = res3-load12*load11;
  236. load14 = ptrbb[4*3+2];
  237. res4 = res4+load8*load14;
  238. res5 = res5+load10*load14;
  239. load15 = ptrbb[4*3+3];
  240. res4 = res4+load10*load15;
  241. res5 = res5-load8*load15;
  242. res6 = res6+load12*load14;
  243. res7 = res7+load13*load14;
  244. res6 = res6+load13*load15;
  245. res7 = res7-load12*load15;
  246. #endif
  247. #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
  248. load0 = ptrba[4*0+0];
  249. load1 = ptrbb[4*0+0];
  250. res0 = res0+load0*load1;
  251. load2 = ptrba[4*0+1];
  252. res1 = res1-load2*load1;
  253. load3 = ptrbb[4*0+1];
  254. res0 = res0+load2*load3;
  255. res1 = res1+load0*load3;
  256. load4 = ptrba[4*0+2];
  257. res2 = res2+load4*load1;
  258. load5 = ptrba[4*0+3];
  259. res3 = res3-load5*load1;
  260. res2 = res2+load5*load3;
  261. res3 = res3+load4*load3;
  262. load6 = ptrbb[4*0+2];
  263. res4 = res4+load0*load6;
  264. res5 = res5-load2*load6;
  265. load7 = ptrbb[4*0+3];
  266. res4 = res4+load2*load7;
  267. res5 = res5+load0*load7;
  268. res6 = res6+load4*load6;
  269. res7 = res7-load5*load6;
  270. res6 = res6+load5*load7;
  271. res7 = res7+load4*load7;
  272. load8 = ptrba[4*1+0];
  273. load9 = ptrbb[4*1+0];
  274. res0 = res0+load8*load9;
  275. load10 = ptrba[4*1+1];
  276. res1 = res1-load10*load9;
  277. load11 = ptrbb[4*1+1];
  278. res0 = res0+load10*load11;
  279. res1 = res1+load8*load11;
  280. load12 = ptrba[4*1+2];
  281. res2 = res2+load12*load9;
  282. load13 = ptrba[4*1+3];
  283. res3 = res3-load13*load9;
  284. res2 = res2+load13*load11;
  285. res3 = res3+load12*load11;
  286. load14 = ptrbb[4*1+2];
  287. res4 = res4+load8*load14;
  288. res5 = res5-load10*load14;
  289. load15 = ptrbb[4*1+3];
  290. res4 = res4+load10*load15;
  291. res5 = res5+load8*load15;
  292. res6 = res6+load12*load14;
  293. res7 = res7-load13*load14;
  294. res6 = res6+load13*load15;
  295. res7 = res7+load12*load15;
  296. load0 = ptrba[4*2+0];
  297. load1 = ptrbb[4*2+0];
  298. res0 = res0+load0*load1;
  299. load2 = ptrba[4*2+1];
  300. res1 = res1-load2*load1;
  301. load3 = ptrbb[4*2+1];
  302. res0 = res0+load2*load3;
  303. res1 = res1+load0*load3;
  304. load4 = ptrba[4*2+2];
  305. res2 = res2+load4*load1;
  306. load5 = ptrba[4*2+3];
  307. res3 = res3-load5*load1;
  308. res2 = res2+load5*load3;
  309. res3 = res3+load4*load3;
  310. load6 = ptrbb[4*2+2];
  311. res4 = res4+load0*load6;
  312. res5 = res5-load2*load6;
  313. load7 = ptrbb[4*2+3];
  314. res4 = res4+load2*load7;
  315. res5 = res5+load0*load7;
  316. res6 = res6+load4*load6;
  317. res7 = res7-load5*load6;
  318. res6 = res6+load5*load7;
  319. res7 = res7+load4*load7;
  320. load8 = ptrba[4*3+0];
  321. load9 = ptrbb[4*3+0];
  322. res0 = res0+load8*load9;
  323. load10 = ptrba[4*3+1];
  324. res1 = res1-load10*load9;
  325. load11 = ptrbb[4*3+1];
  326. res0 = res0+load10*load11;
  327. res1 = res1+load8*load11;
  328. load12 = ptrba[4*3+2];
  329. res2 = res2+load12*load9;
  330. load13 = ptrba[4*3+3];
  331. res3 = res3-load13*load9;
  332. res2 = res2+load13*load11;
  333. res3 = res3+load12*load11;
  334. load14 = ptrbb[4*3+2];
  335. res4 = res4+load8*load14;
  336. res5 = res5-load10*load14;
  337. load15 = ptrbb[4*3+3];
  338. res4 = res4+load10*load15;
  339. res5 = res5+load8*load15;
  340. res6 = res6+load12*load14;
  341. res7 = res7-load13*load14;
  342. res6 = res6+load13*load15;
  343. res7 = res7+load12*load15;
  344. #endif
  345. #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
  346. load0 = ptrba[4*0+0];
  347. load1 = ptrbb[4*0+0];
  348. res0 = res0+load0*load1;
  349. load2 = ptrba[4*0+1];
  350. res1 = res1-load2*load1;
  351. load3 = ptrbb[4*0+1];
  352. res0 = res0-load2*load3;
  353. res1 = res1-load0*load3;
  354. load4 = ptrba[4*0+2];
  355. res2 = res2+load4*load1;
  356. load5 = ptrba[4*0+3];
  357. res3 = res3-load5*load1;
  358. res2 = res2-load5*load3;
  359. res3 = res3-load4*load3;
  360. load6 = ptrbb[4*0+2];
  361. res4 = res4+load0*load6;
  362. res5 = res5-load2*load6;
  363. load7 = ptrbb[4*0+3];
  364. res4 = res4-load2*load7;
  365. res5 = res5-load0*load7;
  366. res6 = res6+load4*load6;
  367. res7 = res7-load5*load6;
  368. res6 = res6-load5*load7;
  369. res7 = res7-load4*load7;
  370. load8 = ptrba[4*1+0];
  371. load9 = ptrbb[4*1+0];
  372. res0 = res0+load8*load9;
  373. load10 = ptrba[4*1+1];
  374. res1 = res1-load10*load9;
  375. load11 = ptrbb[4*1+1];
  376. res0 = res0-load10*load11;
  377. res1 = res1-load8*load11;
  378. load12 = ptrba[4*1+2];
  379. res2 = res2+load12*load9;
  380. load13 = ptrba[4*1+3];
  381. res3 = res3-load13*load9;
  382. res2 = res2-load13*load11;
  383. res3 = res3-load12*load11;
  384. load14 = ptrbb[4*1+2];
  385. res4 = res4+load8*load14;
  386. res5 = res5-load10*load14;
  387. load15 = ptrbb[4*1+3];
  388. res4 = res4-load10*load15;
  389. res5 = res5-load8*load15;
  390. res6 = res6+load12*load14;
  391. res7 = res7-load13*load14;
  392. res6 = res6-load13*load15;
  393. res7 = res7-load12*load15;
  394. load0 = ptrba[4*2+0];
  395. load1 = ptrbb[4*2+0];
  396. res0 = res0+load0*load1;
  397. load2 = ptrba[4*2+1];
  398. res1 = res1-load2*load1;
  399. load3 = ptrbb[4*2+1];
  400. res0 = res0-load2*load3;
  401. res1 = res1-load0*load3;
  402. load4 = ptrba[4*2+2];
  403. res2 = res2+load4*load1;
  404. load5 = ptrba[4*2+3];
  405. res3 = res3-load5*load1;
  406. res2 = res2-load5*load3;
  407. res3 = res3-load4*load3;
  408. load6 = ptrbb[4*2+2];
  409. res4 = res4+load0*load6;
  410. res5 = res5-load2*load6;
  411. load7 = ptrbb[4*2+3];
  412. res4 = res4-load2*load7;
  413. res5 = res5-load0*load7;
  414. res6 = res6+load4*load6;
  415. res7 = res7-load5*load6;
  416. res6 = res6-load5*load7;
  417. res7 = res7-load4*load7;
  418. load8 = ptrba[4*3+0];
  419. load9 = ptrbb[4*3+0];
  420. res0 = res0+load8*load9;
  421. load10 = ptrba[4*3+1];
  422. res1 = res1-load10*load9;
  423. load11 = ptrbb[4*3+1];
  424. res0 = res0-load10*load11;
  425. res1 = res1-load8*load11;
  426. load12 = ptrba[4*3+2];
  427. res2 = res2+load12*load9;
  428. load13 = ptrba[4*3+3];
  429. res3 = res3-load13*load9;
  430. res2 = res2-load13*load11;
  431. res3 = res3-load12*load11;
  432. load14 = ptrbb[4*3+2];
  433. res4 = res4+load8*load14;
  434. res5 = res5-load10*load14;
  435. load15 = ptrbb[4*3+3];
  436. res4 = res4-load10*load15;
  437. res5 = res5-load8*load15;
  438. res6 = res6+load12*load14;
  439. res7 = res7-load13*load14;
  440. res6 = res6-load13*load15;
  441. res7 = res7-load12*load15;
  442. #endif
  443. ptrba = ptrba+16;
  444. ptrbb = ptrbb+16;
  445. }
  446. for (k=0; k<(temp&3); k+=1)
  447. {
  448. #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
  449. load0 = ptrba[4*0+0];
  450. load1 = ptrbb[4*0+0];
  451. res0 = res0+load0*load1;
  452. load2 = ptrba[4*0+1];
  453. res1 = res1+load2*load1;
  454. load3 = ptrbb[4*0+1];
  455. res0 = res0-load2*load3;
  456. res1 = res1+load0*load3;
  457. load4 = ptrba[4*0+2];
  458. res2 = res2+load4*load1;
  459. load5 = ptrba[4*0+3];
  460. res3 = res3+load5*load1;
  461. res2 = res2-load5*load3;
  462. res3 = res3+load4*load3;
  463. load6 = ptrbb[4*0+2];
  464. res4 = res4+load0*load6;
  465. res5 = res5+load2*load6;
  466. load7 = ptrbb[4*0+3];
  467. res4 = res4-load2*load7;
  468. res5 = res5+load0*load7;
  469. res6 = res6+load4*load6;
  470. res7 = res7+load5*load6;
  471. res6 = res6-load5*load7;
  472. res7 = res7+load4*load7;
  473. #endif
  474. #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
  475. load0 = ptrba[4*0+0];
  476. load1 = ptrbb[4*0+0];
  477. res0 = res0+load0*load1;
  478. load2 = ptrba[4*0+1];
  479. res1 = res1+load2*load1;
  480. load3 = ptrbb[4*0+1];
  481. res0 = res0+load2*load3;
  482. res1 = res1-load0*load3;
  483. load4 = ptrba[4*0+2];
  484. res2 = res2+load4*load1;
  485. load5 = ptrba[4*0+3];
  486. res3 = res3+load5*load1;
  487. res2 = res2+load5*load3;
  488. res3 = res3-load4*load3;
  489. load6 = ptrbb[4*0+2];
  490. res4 = res4+load0*load6;
  491. res5 = res5+load2*load6;
  492. load7 = ptrbb[4*0+3];
  493. res4 = res4+load2*load7;
  494. res5 = res5-load0*load7;
  495. res6 = res6+load4*load6;
  496. res7 = res7+load5*load6;
  497. res6 = res6+load5*load7;
  498. res7 = res7-load4*load7;
  499. #endif
  500. #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
  501. load0 = ptrba[4*0+0];
  502. load1 = ptrbb[4*0+0];
  503. res0 = res0+load0*load1;
  504. load2 = ptrba[4*0+1];
  505. res1 = res1-load2*load1;
  506. load3 = ptrbb[4*0+1];
  507. res0 = res0+load2*load3;
  508. res1 = res1+load0*load3;
  509. load4 = ptrba[4*0+2];
  510. res2 = res2+load4*load1;
  511. load5 = ptrba[4*0+3];
  512. res3 = res3-load5*load1;
  513. res2 = res2+load5*load3;
  514. res3 = res3+load4*load3;
  515. load6 = ptrbb[4*0+2];
  516. res4 = res4+load0*load6;
  517. res5 = res5-load2*load6;
  518. load7 = ptrbb[4*0+3];
  519. res4 = res4+load2*load7;
  520. res5 = res5+load0*load7;
  521. res6 = res6+load4*load6;
  522. res7 = res7-load5*load6;
  523. res6 = res6+load5*load7;
  524. res7 = res7+load4*load7;
  525. #endif
  526. #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
  527. load0 = ptrba[4*0+0];
  528. load1 = ptrbb[4*0+0];
  529. res0 = res0+load0*load1;
  530. load2 = ptrba[4*0+1];
  531. res1 = res1-load2*load1;
  532. load3 = ptrbb[4*0+1];
  533. res0 = res0-load2*load3;
  534. res1 = res1-load0*load3;
  535. load4 = ptrba[4*0+2];
  536. res2 = res2+load4*load1;
  537. load5 = ptrba[4*0+3];
  538. res3 = res3-load5*load1;
  539. res2 = res2-load5*load3;
  540. res3 = res3-load4*load3;
  541. load6 = ptrbb[4*0+2];
  542. res4 = res4+load0*load6;
  543. res5 = res5-load2*load6;
  544. load7 = ptrbb[4*0+3];
  545. res4 = res4-load2*load7;
  546. res5 = res5-load0*load7;
  547. res6 = res6+load4*load6;
  548. res7 = res7-load5*load6;
  549. res6 = res6-load5*load7;
  550. res7 = res7-load4*load7;
  551. #endif
  552. ptrba = ptrba+4;
  553. ptrbb = ptrbb+4;
  554. }
  555. load0 = res0*alphar-res1*alphai;
  556. load1 = res1*alphar+res0*alphai;
  557. C0[0] = load0;
  558. C0[1] = load1;
  559. load2 = res2*alphar-res3*alphai;
  560. load3 = res3*alphar+res2*alphai;
  561. C0[2] = load2;
  562. C0[3] = load3;
  563. load4 = res4*alphar-res5*alphai;
  564. load5 = res5*alphar+res4*alphai;
  565. C1[0] = load4;
  566. C1[1] = load5;
  567. load6 = res6*alphar-res7*alphai;
  568. load7 = res7*alphar+res6*alphai;
  569. C1[2] = load6;
  570. C1[3] = load7;
  571. #if ( defined(LEFT) && defined(TRANSA)) || \
  572. (!defined(LEFT) && !defined(TRANSA))
  573. temp = bk - off;
  574. #ifdef LEFT
  575. temp -= 2;
  576. #else
  577. temp -= 2;
  578. #endif
  579. ptrba += temp*2*2;
  580. ptrbb += temp*2*2;
  581. #endif
  582. #ifdef LEFT
  583. off += 2;
  584. #endif
  585. C0 = C0+4;
  586. C1 = C1+4;
  587. }
  588. for (i=0; i<(bm&1); i+=1)
  589. {
  590. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  591. ptrbb = bb;
  592. #else
  593. ptrba += off*2;
  594. ptrbb = bb + off*2*2;
  595. #endif
  596. res0 = 0;
  597. res1 = 0;
  598. res2 = 0;
  599. res3 = 0;
  600. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  601. temp = bk - off;
  602. #elif defined(LEFT)
  603. temp = off+1;
  604. #else
  605. temp = off+2;
  606. #endif
  607. for (k=0; k<temp; k+=1)
  608. {
  609. #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
  610. load0 = ptrba[2*0+0];
  611. load1 = ptrbb[4*0+0];
  612. res0 = res0+load0*load1;
  613. load2 = ptrba[2*0+1];
  614. res1 = res1+load2*load1;
  615. load3 = ptrbb[4*0+1];
  616. res0 = res0-load2*load3;
  617. res1 = res1+load0*load3;
  618. load4 = ptrbb[4*0+2];
  619. res2 = res2+load0*load4;
  620. res3 = res3+load2*load4;
  621. load5 = ptrbb[4*0+3];
  622. res2 = res2-load2*load5;
  623. res3 = res3+load0*load5;
  624. #endif
  625. #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
  626. load0 = ptrba[2*0+0];
  627. load1 = ptrbb[4*0+0];
  628. res0 = res0+load0*load1;
  629. load2 = ptrba[2*0+1];
  630. res1 = res1+load2*load1;
  631. load3 = ptrbb[4*0+1];
  632. res0 = res0+load2*load3;
  633. res1 = res1-load0*load3;
  634. load4 = ptrbb[4*0+2];
  635. res2 = res2+load0*load4;
  636. res3 = res3+load2*load4;
  637. load5 = ptrbb[4*0+3];
  638. res2 = res2+load2*load5;
  639. res3 = res3-load0*load5;
  640. #endif
  641. #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
  642. load0 = ptrba[2*0+0];
  643. load1 = ptrbb[4*0+0];
  644. res0 = res0+load0*load1;
  645. load2 = ptrba[2*0+1];
  646. res1 = res1-load2*load1;
  647. load3 = ptrbb[4*0+1];
  648. res0 = res0+load2*load3;
  649. res1 = res1+load0*load3;
  650. load4 = ptrbb[4*0+2];
  651. res2 = res2+load0*load4;
  652. res3 = res3-load2*load4;
  653. load5 = ptrbb[4*0+3];
  654. res2 = res2+load2*load5;
  655. res3 = res3+load0*load5;
  656. #endif
  657. #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
  658. load0 = ptrba[2*0+0];
  659. load1 = ptrbb[4*0+0];
  660. res0 = res0+load0*load1;
  661. load2 = ptrba[2*0+1];
  662. res1 = res1-load2*load1;
  663. load3 = ptrbb[4*0+1];
  664. res0 = res0-load2*load3;
  665. res1 = res1-load0*load3;
  666. load4 = ptrbb[4*0+2];
  667. res2 = res2+load0*load4;
  668. res3 = res3-load2*load4;
  669. load5 = ptrbb[4*0+3];
  670. res2 = res2-load2*load5;
  671. res3 = res3-load0*load5;
  672. #endif
  673. ptrba = ptrba+2;
  674. ptrbb = ptrbb+4;
  675. }
  676. load0 = res0*alphar-res1*alphai;
  677. load1 = res1*alphar+res0*alphai;
  678. C0[0] = load0;
  679. C0[1] = load1;
  680. load2 = res2*alphar-res3*alphai;
  681. load3 = res3*alphar+res2*alphai;
  682. C1[0] = load2;
  683. C1[1] = load3;
  684. #if ( defined(LEFT) && defined(TRANSA)) || \
  685. (!defined(LEFT) && !defined(TRANSA))
  686. temp = bk - off;
  687. #ifdef LEFT
  688. temp -= 1;
  689. #else
  690. temp -= 2;
  691. #endif
  692. ptrba += temp*2;
  693. ptrbb += temp*2*2;
  694. #endif
  695. #ifdef LEFT
  696. off += 1;
  697. #endif
  698. C0 = C0+2;
  699. C1 = C1+2;
  700. }
  701. #if defined(TRMMKERNEL) && !defined(LEFT)
  702. off += 2;
  703. #endif
  704. k = (bk<<2);
  705. bb = bb+k;
  706. i = (ldc<<2);
  707. C = C+i;
  708. }
  709. for (j=0; j<(bn&1); j+=1)
  710. {
  711. C0 = C;
  712. #if defined(TRMMKERNEL) && defined(LEFT)
  713. off = offset;
  714. #endif
  715. ptrba = ba;
  716. for (i=0; i<bm/2; i+=1)
  717. {
  718. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  719. ptrbb = bb;
  720. #else
  721. ptrba += off*2*2;
  722. ptrbb = bb+off*2;
  723. #endif
  724. res0 = 0;
  725. res1 = 0;
  726. res2 = 0;
  727. res3 = 0;
  728. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  729. temp = bk - off;
  730. #elif defined(LEFT)
  731. temp = off + 2;
  732. #else
  733. temp = off + 1;
  734. #endif
  735. for (k=0; k<temp; k+=1)
  736. {
  737. #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
  738. load0 = ptrba[4*0+0];
  739. load1 = ptrbb[2*0+0];
  740. res0 = res0+load0*load1;
  741. load2 = ptrba[4*0+1];
  742. res1 = res1+load2*load1;
  743. load3 = ptrbb[2*0+1];
  744. res0 = res0-load2*load3;
  745. res1 = res1+load0*load3;
  746. load4 = ptrba[4*0+2];
  747. res2 = res2+load4*load1;
  748. load5 = ptrba[4*0+3];
  749. res3 = res3+load5*load1;
  750. res2 = res2-load5*load3;
  751. res3 = res3+load4*load3;
  752. #endif
  753. #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
  754. load0 = ptrba[4*0+0];
  755. load1 = ptrbb[2*0+0];
  756. res0 = res0+load0*load1;
  757. load2 = ptrba[4*0+1];
  758. res1 = res1+load2*load1;
  759. load3 = ptrbb[2*0+1];
  760. res0 = res0+load2*load3;
  761. res1 = res1-load0*load3;
  762. load4 = ptrba[4*0+2];
  763. res2 = res2+load4*load1;
  764. load5 = ptrba[4*0+3];
  765. res3 = res3+load5*load1;
  766. res2 = res2+load5*load3;
  767. res3 = res3-load4*load3;
  768. #endif
  769. #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
  770. load0 = ptrba[4*0+0];
  771. load1 = ptrbb[2*0+0];
  772. res0 = res0+load0*load1;
  773. load2 = ptrba[4*0+1];
  774. res1 = res1-load2*load1;
  775. load3 = ptrbb[2*0+1];
  776. res0 = res0+load2*load3;
  777. res1 = res1+load0*load3;
  778. load4 = ptrba[4*0+2];
  779. res2 = res2+load4*load1;
  780. load5 = ptrba[4*0+3];
  781. res3 = res3-load5*load1;
  782. res2 = res2+load5*load3;
  783. res3 = res3+load4*load3;
  784. #endif
  785. #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
  786. load0 = ptrba[4*0+0];
  787. load1 = ptrbb[2*0+0];
  788. res0 = res0+load0*load1;
  789. load2 = ptrba[4*0+1];
  790. res1 = res1-load2*load1;
  791. load3 = ptrbb[2*0+1];
  792. res0 = res0-load2*load3;
  793. res1 = res1-load0*load3;
  794. load4 = ptrba[4*0+2];
  795. res2 = res2+load4*load1;
  796. load5 = ptrba[4*0+3];
  797. res3 = res3-load5*load1;
  798. res2 = res2-load5*load3;
  799. res3 = res3-load4*load3;
  800. #endif
  801. ptrba = ptrba+4;
  802. ptrbb = ptrbb+2;
  803. }
  804. load0 = res0*alphar-res1*alphai;
  805. load1 = res1*alphar+res0*alphai;
  806. C0[0] = load0;
  807. C0[1] = load1;
  808. load2 = res2*alphar-res3*alphai;
  809. load3 = res3*alphar+res2*alphai;
  810. C0[2] = load2;
  811. C0[3] = load3;
  812. #if ( defined(LEFT) && defined(TRANSA)) || \
  813. (!defined(LEFT) && !defined(TRANSA))
  814. temp = bk-off;
  815. #ifdef LEFT
  816. temp -= 2;
  817. #else
  818. temp -= 1;
  819. #endif
  820. ptrba += temp*2*2;
  821. ptrbb += temp*2;
  822. #endif
  823. #ifdef LEFT
  824. off += 2;
  825. #endif
  826. C0 = C0+4;
  827. }
  828. for (i=0; i<(bm&1); i+=1)
  829. {
  830. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  831. ptrbb = bb;
  832. #else
  833. ptrba += off*2;
  834. ptrbb = bb + off*2;
  835. #endif
  836. res0 = 0;
  837. res1 = 0;
  838. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  839. temp = bk-off;
  840. #elif defined(LEFT)
  841. temp = off + 1;
  842. #else
  843. temp = off + 1;
  844. #endif
  845. for (k=0; k<temp; k+=1)
  846. {
  847. #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
  848. load0 = ptrba[2*0+0];
  849. load1 = ptrbb[2*0+0];
  850. res0 = res0+load0*load1;
  851. load2 = ptrba[2*0+1];
  852. res1 = res1+load2*load1;
  853. load3 = ptrbb[2*0+1];
  854. res0 = res0-load2*load3;
  855. res1 = res1+load0*load3;
  856. #endif
  857. #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
  858. load0 = ptrba[2*0+0];
  859. load1 = ptrbb[2*0+0];
  860. res0 = res0+load0*load1;
  861. load2 = ptrba[2*0+1];
  862. res1 = res1+load2*load1;
  863. load3 = ptrbb[2*0+1];
  864. res0 = res0+load2*load3;
  865. res1 = res1-load0*load3;
  866. #endif
  867. #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
  868. load0 = ptrba[2*0+0];
  869. load1 = ptrbb[2*0+0];
  870. res0 = res0+load0*load1;
  871. load2 = ptrba[2*0+1];
  872. res1 = res1-load2*load1;
  873. load3 = ptrbb[2*0+1];
  874. res0 = res0+load2*load3;
  875. res1 = res1+load0*load3;
  876. #endif
  877. #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
  878. load0 = ptrba[2*0+0];
  879. load1 = ptrbb[2*0+0];
  880. res0 = res0+load0*load1;
  881. load2 = ptrba[2*0+1];
  882. res1 = res1-load2*load1;
  883. load3 = ptrbb[2*0+1];
  884. res0 = res0-load2*load3;
  885. res1 = res1-load0*load3;
  886. #endif
  887. ptrba = ptrba+2;
  888. ptrbb = ptrbb+2;
  889. }
  890. load0 = res0*alphar-res1*alphai;
  891. load1 = res1*alphar+res0*alphai;
  892. C0[0] = load0;
  893. C0[1] = load1;
  894. #if ( defined(LEFT) && defined(TRANSA)) || \
  895. (!defined(LEFT) && !defined(TRANSA))
  896. temp = bk - off;
  897. #ifdef LEFT
  898. temp -= 1;
  899. #else
  900. temp -= 1;
  901. #endif
  902. ptrba += temp*2;
  903. ptrbb += temp*2;
  904. #endif
  905. #ifdef LEFT
  906. off += 1;
  907. #endif
  908. C0 = C0+2;
  909. }
  910. k = (bk<<1);
  911. bb = bb+k;
  912. i = (ldc<<1);
  913. C = C+i;
  914. }
  915. return 0;
  916. }