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.

gemm_ncopy_8.S 20 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921
  1. /*********************************************************************/
  2. /* Copyright 2005-2010 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. #define ASSEMBLER
  39. #include "common.h"
  40. #define PREFETCHSIZE 42
  41. #define WPREFETCHSIZE 20
  42. #define M %i0
  43. #define N %i1
  44. #define A %i2
  45. #define LDA %i3
  46. #define B %i4
  47. #define A1 %l0
  48. #define A2 %l1
  49. #define A3 %l2
  50. #define A4 %l3
  51. #define A5 %o0
  52. #define A6 %o1
  53. #define A7 %o2
  54. #define A8 %o3
  55. #define I %l4
  56. #define J %l5
  57. #ifdef DOUBLE
  58. #define c01 %f0
  59. #define c02 %f2
  60. #define c03 %f4
  61. #define c04 %f6
  62. #define c05 %f8
  63. #define c06 %f10
  64. #define c07 %f12
  65. #define c08 %f14
  66. #define c09 %f16
  67. #define c10 %f18
  68. #define c11 %f20
  69. #define c12 %f22
  70. #define c13 %f24
  71. #define c14 %f26
  72. #define c15 %f28
  73. #define c16 %f30
  74. #else
  75. #define c01 %f0
  76. #define c02 %f1
  77. #define c03 %f2
  78. #define c04 %f3
  79. #define c05 %f4
  80. #define c06 %f5
  81. #define c07 %f6
  82. #define c08 %f7
  83. #define c09 %f8
  84. #define c10 %f9
  85. #define c11 %f10
  86. #define c12 %f11
  87. #define c13 %f12
  88. #define c14 %f13
  89. #define c15 %f14
  90. #define c16 %f15
  91. #endif
  92. PROLOGUE
  93. SAVESP
  94. sra N, 3, J
  95. cmp J, 0
  96. ble,pn %icc, .LL20
  97. sll LDA, BASE_SHIFT, LDA
  98. .LL11:
  99. add A, LDA, A2
  100. mov A, A1
  101. add A2, LDA, A3
  102. sra M, 3, I
  103. add A3, LDA, A4
  104. cmp I, 0
  105. add A4, LDA, A5
  106. add A5, LDA, A6
  107. add A6, LDA, A7
  108. add A7, LDA, A8
  109. ble,pn %icc, .LL13
  110. add A8, LDA, A
  111. .align 4
  112. .LL12:
  113. prefetch [A1 + (PREFETCHSIZE + 0) * SIZE], 0
  114. LDF [A1 + 0 * SIZE], c01
  115. LDF [A2 + 0 * SIZE], c02
  116. LDF [A3 + 0 * SIZE], c03
  117. LDF [A4 + 0 * SIZE], c04
  118. LDF [A5 + 0 * SIZE], c05
  119. LDF [A6 + 0 * SIZE], c06
  120. LDF [A7 + 0 * SIZE], c07
  121. LDF [A8 + 0 * SIZE], c08
  122. prefetch [A2 + (PREFETCHSIZE + 0) * SIZE], 0
  123. LDF [A1 + 1 * SIZE], c09
  124. LDF [A2 + 1 * SIZE], c10
  125. LDF [A3 + 1 * SIZE], c11
  126. LDF [A4 + 1 * SIZE], c12
  127. LDF [A5 + 1 * SIZE], c13
  128. LDF [A6 + 1 * SIZE], c14
  129. LDF [A7 + 1 * SIZE], c15
  130. LDF [A8 + 1 * SIZE], c16
  131. prefetch [B + (WPREFETCHSIZE + 0) * SIZE], 2
  132. STF c01, [B + 0 * SIZE]
  133. STF c02, [B + 1 * SIZE]
  134. STF c03, [B + 2 * SIZE]
  135. STF c04, [B + 3 * SIZE]
  136. STF c05, [B + 4 * SIZE]
  137. STF c06, [B + 5 * SIZE]
  138. STF c07, [B + 6 * SIZE]
  139. STF c08, [B + 7 * SIZE]
  140. prefetch [B + (WPREFETCHSIZE + 8) * SIZE], 2
  141. STF c09, [B + 8 * SIZE]
  142. STF c10, [B + 9 * SIZE]
  143. STF c11, [B + 10 * SIZE]
  144. STF c12, [B + 11 * SIZE]
  145. STF c13, [B + 12 * SIZE]
  146. STF c14, [B + 13 * SIZE]
  147. STF c15, [B + 14 * SIZE]
  148. STF c16, [B + 15 * SIZE]
  149. prefetch [A3 + (PREFETCHSIZE + 0) * SIZE], 0
  150. LDF [A1 + 2 * SIZE], c01
  151. LDF [A2 + 2 * SIZE], c02
  152. LDF [A3 + 2 * SIZE], c03
  153. LDF [A4 + 2 * SIZE], c04
  154. LDF [A5 + 2 * SIZE], c05
  155. LDF [A6 + 2 * SIZE], c06
  156. LDF [A7 + 2 * SIZE], c07
  157. LDF [A8 + 2 * SIZE], c08
  158. prefetch [A4 + (PREFETCHSIZE + 0) * SIZE], 0
  159. LDF [A1 + 3 * SIZE], c09
  160. LDF [A2 + 3 * SIZE], c10
  161. LDF [A3 + 3 * SIZE], c11
  162. LDF [A4 + 3 * SIZE], c12
  163. LDF [A5 + 3 * SIZE], c13
  164. LDF [A6 + 3 * SIZE], c14
  165. LDF [A7 + 3 * SIZE], c15
  166. LDF [A8 + 3 * SIZE], c16
  167. prefetch [B + (WPREFETCHSIZE + 16) * SIZE], 2
  168. STF c01, [B + 16 * SIZE]
  169. STF c02, [B + 17 * SIZE]
  170. STF c03, [B + 18 * SIZE]
  171. STF c04, [B + 19 * SIZE]
  172. STF c05, [B + 20 * SIZE]
  173. STF c06, [B + 21 * SIZE]
  174. STF c07, [B + 22 * SIZE]
  175. STF c08, [B + 23 * SIZE]
  176. prefetch [B + (WPREFETCHSIZE + 24) * SIZE], 2
  177. STF c09, [B + 24 * SIZE]
  178. STF c10, [B + 25 * SIZE]
  179. STF c11, [B + 26 * SIZE]
  180. STF c12, [B + 27 * SIZE]
  181. STF c13, [B + 28 * SIZE]
  182. STF c14, [B + 29 * SIZE]
  183. STF c15, [B + 30 * SIZE]
  184. STF c16, [B + 31 * SIZE]
  185. prefetch [A5 + (PREFETCHSIZE + 0) * SIZE], 0
  186. LDF [A1 + 4 * SIZE], c01
  187. LDF [A2 + 4 * SIZE], c02
  188. LDF [A3 + 4 * SIZE], c03
  189. LDF [A4 + 4 * SIZE], c04
  190. LDF [A5 + 4 * SIZE], c05
  191. LDF [A6 + 4 * SIZE], c06
  192. LDF [A7 + 4 * SIZE], c07
  193. LDF [A8 + 4 * SIZE], c08
  194. prefetch [A6 + (PREFETCHSIZE + 0) * SIZE], 0
  195. LDF [A1 + 5 * SIZE], c09
  196. LDF [A2 + 5 * SIZE], c10
  197. LDF [A3 + 5 * SIZE], c11
  198. LDF [A4 + 5 * SIZE], c12
  199. LDF [A5 + 5 * SIZE], c13
  200. LDF [A6 + 5 * SIZE], c14
  201. LDF [A7 + 5 * SIZE], c15
  202. LDF [A8 + 5 * SIZE], c16
  203. prefetch [B + (WPREFETCHSIZE + 32) * SIZE], 2
  204. STF c01, [B + 32 * SIZE]
  205. STF c02, [B + 33 * SIZE]
  206. STF c03, [B + 34 * SIZE]
  207. STF c04, [B + 35 * SIZE]
  208. STF c05, [B + 36 * SIZE]
  209. STF c06, [B + 37 * SIZE]
  210. STF c07, [B + 38 * SIZE]
  211. STF c08, [B + 39 * SIZE]
  212. prefetch [B + (WPREFETCHSIZE + 40) * SIZE], 2
  213. STF c09, [B + 40 * SIZE]
  214. STF c10, [B + 41 * SIZE]
  215. STF c11, [B + 42 * SIZE]
  216. STF c12, [B + 43 * SIZE]
  217. STF c13, [B + 44 * SIZE]
  218. STF c14, [B + 45 * SIZE]
  219. STF c15, [B + 46 * SIZE]
  220. STF c16, [B + 47 * SIZE]
  221. prefetch [A7 + (PREFETCHSIZE + 0) * SIZE], 0
  222. LDF [A1 + 6 * SIZE], c01
  223. LDF [A2 + 6 * SIZE], c02
  224. LDF [A3 + 6 * SIZE], c03
  225. LDF [A4 + 6 * SIZE], c04
  226. LDF [A5 + 6 * SIZE], c05
  227. LDF [A6 + 6 * SIZE], c06
  228. LDF [A7 + 6 * SIZE], c07
  229. LDF [A8 + 6 * SIZE], c08
  230. prefetch [A8 + (PREFETCHSIZE + 0) * SIZE], 0
  231. LDF [A1 + 7 * SIZE], c09
  232. LDF [A2 + 7 * SIZE], c10
  233. LDF [A3 + 7 * SIZE], c11
  234. LDF [A4 + 7 * SIZE], c12
  235. LDF [A5 + 7 * SIZE], c13
  236. LDF [A6 + 7 * SIZE], c14
  237. LDF [A7 + 7 * SIZE], c15
  238. LDF [A8 + 7 * SIZE], c16
  239. add A1, 8 * SIZE, A1
  240. add A2, 8 * SIZE, A2
  241. add A3, 8 * SIZE, A3
  242. add A4, 8 * SIZE, A4
  243. prefetch [B + (WPREFETCHSIZE + 48) * SIZE], 2
  244. STF c01, [B + 48 * SIZE]
  245. STF c02, [B + 49 * SIZE]
  246. STF c03, [B + 50 * SIZE]
  247. STF c04, [B + 51 * SIZE]
  248. STF c05, [B + 52 * SIZE]
  249. STF c06, [B + 53 * SIZE]
  250. STF c07, [B + 54 * SIZE]
  251. STF c08, [B + 55 * SIZE]
  252. add A5, 8 * SIZE, A5
  253. add A6, 8 * SIZE, A6
  254. add A7, 8 * SIZE, A7
  255. add A8, 8 * SIZE, A8
  256. prefetch [B + (WPREFETCHSIZE + 56) * SIZE], 2
  257. STF c09, [B + 56 * SIZE]
  258. STF c10, [B + 57 * SIZE]
  259. STF c11, [B + 58 * SIZE]
  260. STF c12, [B + 59 * SIZE]
  261. STF c13, [B + 60 * SIZE]
  262. STF c14, [B + 61 * SIZE]
  263. STF c15, [B + 62 * SIZE]
  264. STF c16, [B + 63 * SIZE]
  265. add I, -1, I
  266. cmp I, 0
  267. bg,pt %icc, .LL12
  268. add B, 64 * SIZE, B
  269. .align 4
  270. .LL13:
  271. and M, 4, I
  272. cmp I, 0
  273. ble,pn %icc, .LL14
  274. nop
  275. LDF [A1 + 0 * SIZE], c01
  276. LDF [A2 + 0 * SIZE], c02
  277. LDF [A3 + 0 * SIZE], c03
  278. LDF [A4 + 0 * SIZE], c04
  279. LDF [A5 + 0 * SIZE], c05
  280. LDF [A6 + 0 * SIZE], c06
  281. LDF [A7 + 0 * SIZE], c07
  282. LDF [A8 + 0 * SIZE], c08
  283. LDF [A1 + 1 * SIZE], c09
  284. LDF [A2 + 1 * SIZE], c10
  285. LDF [A3 + 1 * SIZE], c11
  286. LDF [A4 + 1 * SIZE], c12
  287. LDF [A5 + 1 * SIZE], c13
  288. LDF [A6 + 1 * SIZE], c14
  289. LDF [A7 + 1 * SIZE], c15
  290. LDF [A8 + 1 * SIZE], c16
  291. STF c01, [B + 0 * SIZE]
  292. STF c02, [B + 1 * SIZE]
  293. STF c03, [B + 2 * SIZE]
  294. STF c04, [B + 3 * SIZE]
  295. STF c05, [B + 4 * SIZE]
  296. STF c06, [B + 5 * SIZE]
  297. STF c07, [B + 6 * SIZE]
  298. STF c08, [B + 7 * SIZE]
  299. STF c09, [B + 8 * SIZE]
  300. STF c10, [B + 9 * SIZE]
  301. STF c11, [B + 10 * SIZE]
  302. STF c12, [B + 11 * SIZE]
  303. STF c13, [B + 12 * SIZE]
  304. STF c14, [B + 13 * SIZE]
  305. STF c15, [B + 14 * SIZE]
  306. STF c16, [B + 15 * SIZE]
  307. LDF [A1 + 2 * SIZE], c01
  308. LDF [A2 + 2 * SIZE], c02
  309. LDF [A3 + 2 * SIZE], c03
  310. LDF [A4 + 2 * SIZE], c04
  311. LDF [A5 + 2 * SIZE], c05
  312. LDF [A6 + 2 * SIZE], c06
  313. LDF [A7 + 2 * SIZE], c07
  314. LDF [A8 + 2 * SIZE], c08
  315. LDF [A1 + 3 * SIZE], c09
  316. LDF [A2 + 3 * SIZE], c10
  317. LDF [A3 + 3 * SIZE], c11
  318. LDF [A4 + 3 * SIZE], c12
  319. LDF [A5 + 3 * SIZE], c13
  320. LDF [A6 + 3 * SIZE], c14
  321. LDF [A7 + 3 * SIZE], c15
  322. LDF [A8 + 3 * SIZE], c16
  323. STF c01, [B + 16 * SIZE]
  324. STF c02, [B + 17 * SIZE]
  325. STF c03, [B + 18 * SIZE]
  326. STF c04, [B + 19 * SIZE]
  327. STF c05, [B + 20 * SIZE]
  328. STF c06, [B + 21 * SIZE]
  329. STF c07, [B + 22 * SIZE]
  330. STF c08, [B + 23 * SIZE]
  331. STF c09, [B + 24 * SIZE]
  332. STF c10, [B + 25 * SIZE]
  333. STF c11, [B + 26 * SIZE]
  334. STF c12, [B + 27 * SIZE]
  335. STF c13, [B + 28 * SIZE]
  336. STF c14, [B + 29 * SIZE]
  337. STF c15, [B + 30 * SIZE]
  338. STF c16, [B + 31 * SIZE]
  339. add A1, 4 * SIZE, A1
  340. add A2, 4 * SIZE, A2
  341. add A3, 4 * SIZE, A3
  342. add A4, 4 * SIZE, A4
  343. add A5, 4 * SIZE, A5
  344. add A6, 4 * SIZE, A6
  345. add A7, 4 * SIZE, A7
  346. add A8, 4 * SIZE, A8
  347. add B, 32 * SIZE, B
  348. .align 4
  349. .LL14:
  350. and M, 2, I
  351. cmp I, 0
  352. ble,pn %icc, .LL15
  353. nop
  354. LDF [A1 + 0 * SIZE], c01
  355. LDF [A2 + 0 * SIZE], c02
  356. LDF [A3 + 0 * SIZE], c03
  357. LDF [A4 + 0 * SIZE], c04
  358. LDF [A5 + 0 * SIZE], c05
  359. LDF [A6 + 0 * SIZE], c06
  360. LDF [A7 + 0 * SIZE], c07
  361. LDF [A8 + 0 * SIZE], c08
  362. LDF [A1 + 1 * SIZE], c09
  363. LDF [A2 + 1 * SIZE], c10
  364. LDF [A3 + 1 * SIZE], c11
  365. LDF [A4 + 1 * SIZE], c12
  366. LDF [A5 + 1 * SIZE], c13
  367. LDF [A6 + 1 * SIZE], c14
  368. LDF [A7 + 1 * SIZE], c15
  369. LDF [A8 + 1 * SIZE], c16
  370. STF c01, [B + 0 * SIZE]
  371. STF c02, [B + 1 * SIZE]
  372. STF c03, [B + 2 * SIZE]
  373. STF c04, [B + 3 * SIZE]
  374. STF c05, [B + 4 * SIZE]
  375. STF c06, [B + 5 * SIZE]
  376. STF c07, [B + 6 * SIZE]
  377. STF c08, [B + 7 * SIZE]
  378. STF c09, [B + 8 * SIZE]
  379. STF c10, [B + 9 * SIZE]
  380. STF c11, [B + 10 * SIZE]
  381. STF c12, [B + 11 * SIZE]
  382. STF c13, [B + 12 * SIZE]
  383. STF c14, [B + 13 * SIZE]
  384. STF c15, [B + 14 * SIZE]
  385. STF c16, [B + 15 * SIZE]
  386. add A1, 2 * SIZE, A1
  387. add A2, 2 * SIZE, A2
  388. add A3, 2 * SIZE, A3
  389. add A4, 2 * SIZE, A4
  390. add A5, 2 * SIZE, A5
  391. add A6, 2 * SIZE, A6
  392. add A7, 2 * SIZE, A7
  393. add A8, 2 * SIZE, A8
  394. add B, 16 * SIZE, B
  395. .align 4
  396. .LL15:
  397. and M, 1, I
  398. cmp I, 0
  399. ble,pn %icc, .LL19
  400. nop
  401. LDF [A1 + 0 * SIZE], c01
  402. LDF [A2 + 0 * SIZE], c02
  403. LDF [A3 + 0 * SIZE], c03
  404. LDF [A4 + 0 * SIZE], c04
  405. LDF [A5 + 0 * SIZE], c05
  406. LDF [A6 + 0 * SIZE], c06
  407. LDF [A7 + 0 * SIZE], c07
  408. LDF [A8 + 0 * SIZE], c08
  409. STF c01, [B + 0 * SIZE]
  410. STF c02, [B + 1 * SIZE]
  411. STF c03, [B + 2 * SIZE]
  412. STF c04, [B + 3 * SIZE]
  413. STF c05, [B + 4 * SIZE]
  414. STF c06, [B + 5 * SIZE]
  415. STF c07, [B + 6 * SIZE]
  416. STF c08, [B + 7 * SIZE]
  417. add B, 8 * SIZE, B
  418. .align 4
  419. .LL19:
  420. add J, -1, J
  421. cmp J, 0
  422. bg,pt %icc, .LL11
  423. nop
  424. .align 4
  425. .LL20:
  426. and N, 4, J
  427. cmp J, 0
  428. ble,pn %icc, .LL30
  429. nop
  430. add A, LDA, A2
  431. mov A, A1
  432. add A2, LDA, A3
  433. sra M, 3, I
  434. add A3, LDA, A4
  435. cmp I, 0
  436. ble,pn %icc, .LL23
  437. add A4, LDA, A
  438. .align 4
  439. .LL22:
  440. prefetch [A1 + (PREFETCHSIZE + 0) * SIZE], 0
  441. LDF [A1 + 0 * SIZE], c01
  442. LDF [A2 + 0 * SIZE], c02
  443. LDF [A3 + 0 * SIZE], c03
  444. LDF [A4 + 0 * SIZE], c04
  445. LDF [A1 + 1 * SIZE], c05
  446. LDF [A2 + 1 * SIZE], c06
  447. LDF [A3 + 1 * SIZE], c07
  448. LDF [A4 + 1 * SIZE], c08
  449. prefetch [A2 + (PREFETCHSIZE + 0) * SIZE], 0
  450. LDF [A1 + 2 * SIZE], c09
  451. LDF [A2 + 2 * SIZE], c10
  452. LDF [A3 + 2 * SIZE], c11
  453. LDF [A4 + 2 * SIZE], c12
  454. LDF [A1 + 3 * SIZE], c13
  455. LDF [A2 + 3 * SIZE], c14
  456. LDF [A3 + 3 * SIZE], c15
  457. LDF [A4 + 3 * SIZE], c16
  458. prefetch [B + (WPREFETCHSIZE + 0) * SIZE], 2
  459. STF c01, [B + 0 * SIZE]
  460. STF c02, [B + 1 * SIZE]
  461. STF c03, [B + 2 * SIZE]
  462. STF c04, [B + 3 * SIZE]
  463. STF c05, [B + 4 * SIZE]
  464. STF c06, [B + 5 * SIZE]
  465. STF c07, [B + 6 * SIZE]
  466. STF c08, [B + 7 * SIZE]
  467. prefetch [B + (WPREFETCHSIZE + 8) * SIZE], 2
  468. STF c09, [B + 8 * SIZE]
  469. STF c10, [B + 9 * SIZE]
  470. STF c11, [B + 10 * SIZE]
  471. STF c12, [B + 11 * SIZE]
  472. STF c13, [B + 12 * SIZE]
  473. STF c14, [B + 13 * SIZE]
  474. STF c15, [B + 14 * SIZE]
  475. STF c16, [B + 15 * SIZE]
  476. prefetch [A3 + (PREFETCHSIZE + 0) * SIZE], 0
  477. LDF [A1 + 4 * SIZE], c01
  478. LDF [A2 + 4 * SIZE], c02
  479. LDF [A3 + 4 * SIZE], c03
  480. LDF [A4 + 4 * SIZE], c04
  481. LDF [A1 + 5 * SIZE], c05
  482. LDF [A2 + 5 * SIZE], c06
  483. LDF [A3 + 5 * SIZE], c07
  484. LDF [A4 + 5 * SIZE], c08
  485. prefetch [A4 + (PREFETCHSIZE + 0) * SIZE], 0
  486. LDF [A1 + 6 * SIZE], c09
  487. LDF [A2 + 6 * SIZE], c10
  488. LDF [A3 + 6 * SIZE], c11
  489. LDF [A4 + 6 * SIZE], c12
  490. LDF [A1 + 7 * SIZE], c13
  491. LDF [A2 + 7 * SIZE], c14
  492. LDF [A3 + 7 * SIZE], c15
  493. LDF [A4 + 7 * SIZE], c16
  494. prefetch [B + (WPREFETCHSIZE + 16) * SIZE], 2
  495. STF c01, [B + 16 * SIZE]
  496. STF c02, [B + 17 * SIZE]
  497. STF c03, [B + 18 * SIZE]
  498. STF c04, [B + 19 * SIZE]
  499. STF c05, [B + 20 * SIZE]
  500. STF c06, [B + 21 * SIZE]
  501. STF c07, [B + 22 * SIZE]
  502. STF c08, [B + 23 * SIZE]
  503. prefetch [B + (WPREFETCHSIZE + 24) * SIZE], 2
  504. STF c09, [B + 24 * SIZE]
  505. STF c10, [B + 25 * SIZE]
  506. STF c11, [B + 26 * SIZE]
  507. STF c12, [B + 27 * SIZE]
  508. STF c13, [B + 28 * SIZE]
  509. STF c14, [B + 29 * SIZE]
  510. STF c15, [B + 30 * SIZE]
  511. STF c16, [B + 31 * SIZE]
  512. add A1, 8 * SIZE, A1
  513. add A2, 8 * SIZE, A2
  514. add A3, 8 * SIZE, A3
  515. add A4, 8 * SIZE, A4
  516. add I, -1, I
  517. cmp I, 0
  518. bg,pt %icc, .LL22
  519. add B, 32 * SIZE, B
  520. .align 4
  521. .LL23:
  522. and M, 4, I
  523. cmp I, 0
  524. ble,pn %icc, .LL24
  525. nop
  526. LDF [A1 + 0 * SIZE], c01
  527. LDF [A2 + 0 * SIZE], c02
  528. LDF [A3 + 0 * SIZE], c03
  529. LDF [A4 + 0 * SIZE], c04
  530. LDF [A1 + 1 * SIZE], c05
  531. LDF [A2 + 1 * SIZE], c06
  532. LDF [A3 + 1 * SIZE], c07
  533. LDF [A4 + 1 * SIZE], c08
  534. LDF [A1 + 2 * SIZE], c09
  535. LDF [A2 + 2 * SIZE], c10
  536. LDF [A3 + 2 * SIZE], c11
  537. LDF [A4 + 2 * SIZE], c12
  538. LDF [A1 + 3 * SIZE], c13
  539. LDF [A2 + 3 * SIZE], c14
  540. LDF [A3 + 3 * SIZE], c15
  541. LDF [A4 + 3 * SIZE], c16
  542. STF c01, [B + 0 * SIZE]
  543. STF c02, [B + 1 * SIZE]
  544. STF c03, [B + 2 * SIZE]
  545. STF c04, [B + 3 * SIZE]
  546. STF c05, [B + 4 * SIZE]
  547. STF c06, [B + 5 * SIZE]
  548. STF c07, [B + 6 * SIZE]
  549. STF c08, [B + 7 * SIZE]
  550. STF c09, [B + 8 * SIZE]
  551. STF c10, [B + 9 * SIZE]
  552. STF c11, [B + 10 * SIZE]
  553. STF c12, [B + 11 * SIZE]
  554. STF c13, [B + 12 * SIZE]
  555. STF c14, [B + 13 * SIZE]
  556. STF c15, [B + 14 * SIZE]
  557. STF c16, [B + 15 * SIZE]
  558. add A1, 4 * SIZE, A1
  559. add A2, 4 * SIZE, A2
  560. add A3, 4 * SIZE, A3
  561. add A4, 4 * SIZE, A4
  562. add B, 16 * SIZE, B
  563. .align 4
  564. .LL24:
  565. and M, 2, I
  566. cmp I, 0
  567. ble,pn %icc, .LL25
  568. nop
  569. LDF [A1 + 0 * SIZE], c01
  570. LDF [A2 + 0 * SIZE], c02
  571. LDF [A3 + 0 * SIZE], c03
  572. LDF [A4 + 0 * SIZE], c04
  573. LDF [A1 + 1 * SIZE], c05
  574. LDF [A2 + 1 * SIZE], c06
  575. LDF [A3 + 1 * SIZE], c07
  576. LDF [A4 + 1 * SIZE], c08
  577. STF c01, [B + 0 * SIZE]
  578. STF c02, [B + 1 * SIZE]
  579. STF c03, [B + 2 * SIZE]
  580. STF c04, [B + 3 * SIZE]
  581. STF c05, [B + 4 * SIZE]
  582. STF c06, [B + 5 * SIZE]
  583. STF c07, [B + 6 * SIZE]
  584. STF c08, [B + 7 * SIZE]
  585. add A1, 2 * SIZE, A1
  586. add A2, 2 * SIZE, A2
  587. add A3, 2 * SIZE, A3
  588. add A4, 2 * SIZE, A4
  589. add B, 8 * SIZE, B
  590. .align 4
  591. .LL25:
  592. and M, 1, I
  593. cmp I, 0
  594. ble,pn %icc, .LL30
  595. nop
  596. LDF [A1 + 0 * SIZE], c01
  597. LDF [A2 + 0 * SIZE], c02
  598. LDF [A3 + 0 * SIZE], c03
  599. LDF [A4 + 0 * SIZE], c04
  600. STF c01, [B + 0 * SIZE]
  601. STF c02, [B + 1 * SIZE]
  602. STF c03, [B + 2 * SIZE]
  603. STF c04, [B + 3 * SIZE]
  604. add B, 4 * SIZE, B
  605. .align 4
  606. .LL30:
  607. and N, 2, J
  608. cmp J, 0
  609. ble,pn %icc, .LL40
  610. nop
  611. add A, LDA, A2
  612. mov A, A1
  613. sra M, 3, I
  614. cmp I, 0
  615. ble,pn %icc, .LL33
  616. add A2, LDA, A
  617. .align 4
  618. .LL32:
  619. prefetch [A1 + (PREFETCHSIZE + 0) * SIZE], 0
  620. LDF [A1 + 0 * SIZE], c01
  621. LDF [A2 + 0 * SIZE], c02
  622. LDF [A1 + 1 * SIZE], c03
  623. LDF [A2 + 1 * SIZE], c04
  624. LDF [A1 + 2 * SIZE], c05
  625. LDF [A2 + 2 * SIZE], c06
  626. LDF [A1 + 3 * SIZE], c07
  627. LDF [A2 + 3 * SIZE], c08
  628. prefetch [A2 + (PREFETCHSIZE + 0) * SIZE], 0
  629. LDF [A1 + 4 * SIZE], c09
  630. LDF [A2 + 4 * SIZE], c10
  631. LDF [A1 + 5 * SIZE], c11
  632. LDF [A2 + 5 * SIZE], c12
  633. LDF [A1 + 6 * SIZE], c13
  634. LDF [A2 + 6 * SIZE], c14
  635. LDF [A1 + 7 * SIZE], c15
  636. LDF [A2 + 7 * SIZE], c16
  637. prefetch [B + (WPREFETCHSIZE + 0) * SIZE], 2
  638. STF c01, [B + 0 * SIZE]
  639. STF c02, [B + 1 * SIZE]
  640. STF c03, [B + 2 * SIZE]
  641. STF c04, [B + 3 * SIZE]
  642. STF c05, [B + 4 * SIZE]
  643. STF c06, [B + 5 * SIZE]
  644. STF c07, [B + 6 * SIZE]
  645. STF c08, [B + 7 * SIZE]
  646. prefetch [B + (WPREFETCHSIZE + 8) * SIZE], 2
  647. STF c09, [B + 8 * SIZE]
  648. STF c10, [B + 9 * SIZE]
  649. STF c11, [B + 10 * SIZE]
  650. STF c12, [B + 11 * SIZE]
  651. STF c13, [B + 12 * SIZE]
  652. STF c14, [B + 13 * SIZE]
  653. STF c15, [B + 14 * SIZE]
  654. STF c16, [B + 15 * SIZE]
  655. add A1, 8 * SIZE, A1
  656. add A2, 8 * SIZE, A2
  657. add I, -1, I
  658. cmp I, 0
  659. bg,pt %icc, .LL32
  660. add B, 16 * SIZE, B
  661. .align 4
  662. .LL33:
  663. and M, 4, I
  664. cmp I, 0
  665. ble,pn %icc, .LL34
  666. nop
  667. LDF [A1 + 0 * SIZE], c01
  668. LDF [A2 + 0 * SIZE], c02
  669. LDF [A1 + 1 * SIZE], c03
  670. LDF [A2 + 1 * SIZE], c04
  671. LDF [A1 + 2 * SIZE], c05
  672. LDF [A2 + 2 * SIZE], c06
  673. LDF [A1 + 3 * SIZE], c07
  674. LDF [A2 + 3 * SIZE], c08
  675. STF c01, [B + 0 * SIZE]
  676. STF c02, [B + 1 * SIZE]
  677. STF c03, [B + 2 * SIZE]
  678. STF c04, [B + 3 * SIZE]
  679. STF c05, [B + 4 * SIZE]
  680. STF c06, [B + 5 * SIZE]
  681. STF c07, [B + 6 * SIZE]
  682. STF c08, [B + 7 * SIZE]
  683. add A1, 4 * SIZE, A1
  684. add A2, 4 * SIZE, A2
  685. add B, 8 * SIZE, B
  686. .align 4
  687. .LL34:
  688. and M, 2, I
  689. cmp I, 0
  690. ble,pn %icc, .LL35
  691. nop
  692. LDF [A1 + 0 * SIZE], c01
  693. LDF [A2 + 0 * SIZE], c02
  694. LDF [A1 + 1 * SIZE], c03
  695. LDF [A2 + 1 * SIZE], c04
  696. STF c01, [B + 0 * SIZE]
  697. STF c02, [B + 1 * SIZE]
  698. STF c03, [B + 2 * SIZE]
  699. STF c04, [B + 3 * SIZE]
  700. add A1, 2 * SIZE, A1
  701. add A2, 2 * SIZE, A2
  702. add B, 4 * SIZE, B
  703. .align 4
  704. .LL35:
  705. and M, 1, I
  706. cmp I, 0
  707. ble,pn %icc, .LL40
  708. nop
  709. LDF [A1 + 0 * SIZE], c01
  710. LDF [A2 + 0 * SIZE], c02
  711. STF c01, [B + 0 * SIZE]
  712. STF c02, [B + 1 * SIZE]
  713. add B, 2 * SIZE, B
  714. .align 4
  715. .LL40:
  716. and N, 1, J
  717. cmp J, 0
  718. ble,pn %icc, .LL999
  719. nop
  720. sra M, 3, I
  721. cmp I, 0
  722. ble,pn %icc, .LL43
  723. mov A, A1
  724. .align 4
  725. .LL42:
  726. prefetch [A1 + (PREFETCHSIZE + 0) * SIZE], 0
  727. LDF [A1 + 0 * SIZE], c01
  728. LDF [A1 + 1 * SIZE], c02
  729. LDF [A1 + 2 * SIZE], c03
  730. LDF [A1 + 3 * SIZE], c04
  731. LDF [A1 + 4 * SIZE], c05
  732. LDF [A1 + 5 * SIZE], c06
  733. LDF [A1 + 6 * SIZE], c07
  734. LDF [A1 + 7 * SIZE], c08
  735. prefetch [B + (WPREFETCHSIZE + 0) * SIZE], 2
  736. STF c01, [B + 0 * SIZE]
  737. STF c02, [B + 1 * SIZE]
  738. STF c03, [B + 2 * SIZE]
  739. STF c04, [B + 3 * SIZE]
  740. STF c05, [B + 4 * SIZE]
  741. STF c06, [B + 5 * SIZE]
  742. STF c07, [B + 6 * SIZE]
  743. STF c08, [B + 7 * SIZE]
  744. add A1, 8 * SIZE, A1
  745. add I, -1, I
  746. cmp I, 0
  747. bg,pt %icc, .LL42
  748. add B, 8 * SIZE, B
  749. .align 4
  750. .LL43:
  751. and M, 4, I
  752. cmp I, 0
  753. ble,pn %icc, .LL44
  754. nop
  755. LDF [A1 + 0 * SIZE], c01
  756. LDF [A1 + 1 * SIZE], c02
  757. LDF [A1 + 2 * SIZE], c03
  758. LDF [A1 + 3 * SIZE], c04
  759. STF c01, [B + 0 * SIZE]
  760. STF c02, [B + 1 * SIZE]
  761. STF c03, [B + 2 * SIZE]
  762. STF c04, [B + 3 * SIZE]
  763. add A1, 4 * SIZE, A1
  764. add B, 4 * SIZE, B
  765. .align 4
  766. .LL44:
  767. and M, 2, I
  768. cmp I, 0
  769. ble,pn %icc, .LL45
  770. nop
  771. LDF [A1 + 0 * SIZE], c01
  772. LDF [A1 + 1 * SIZE], c02
  773. STF c01, [B + 0 * SIZE]
  774. STF c02, [B + 1 * SIZE]
  775. add A1, 2 * SIZE, A1
  776. add B, 2 * SIZE, B
  777. .align 4
  778. .LL45:
  779. and M, 1, I
  780. cmp I, 0
  781. ble,pn %icc, .LL999
  782. nop
  783. LDF [A1 + 0 * SIZE], c01
  784. STF c01, [B + 0 * SIZE]
  785. .align 4
  786. .LL999:
  787. return %i7 + 8
  788. clr %o0
  789. EPILOGUE