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.

zgemv_n.S 15 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777
  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. #define ASSEMBLER
  39. #include "common.h"
  40. #define M $4
  41. #define N $5
  42. #define A $9
  43. #define LDA $10
  44. #define X $11
  45. #define INCX $2
  46. #define Y $6
  47. #define INCY $7
  48. #define BUFFER $8
  49. #define YORIG $3
  50. #define XX $12
  51. #define YY $13
  52. #define I $14
  53. #define J $15
  54. #define AO1 $16
  55. #define AO2 $17
  56. #define ALPHA_R $f15
  57. #define ALPHA_I $f16
  58. #define a1 $f0
  59. #define a2 $f1
  60. #define a3 $f2
  61. #define a4 $f3
  62. #define a5 $f4
  63. #define a6 $f5
  64. #define a7 $f6
  65. #define a8 $f7
  66. #define x1 $f8
  67. #define x2 $f9
  68. #define x3 $f10
  69. #define x4 $f11
  70. #define y1 $f12
  71. #define y2 $f13
  72. #define y3 $f14
  73. #define y4 $f17
  74. #define t1 $f18
  75. #define t2 $f19
  76. #define t3 $f20
  77. #define t4 $f21
  78. #define t5 $f22
  79. #define t6 $f23
  80. #define t7 $f24
  81. #define t8 $f25
  82. #if !defined(CONJ) && !defined(XCONJ)
  83. #define MADD1 MADD
  84. #define MADD2 MADD
  85. #define MADD3 NMSUB
  86. #define MADD4 MADD
  87. #endif
  88. #if defined(CONJ) && !defined(XCONJ)
  89. #define MADD1 MADD
  90. #define MADD2 MADD
  91. #define MADD3 MADD
  92. #define MADD4 NMSUB
  93. #endif
  94. #if !defined(CONJ) && defined(XCONJ)
  95. #define MADD1 MADD
  96. #define MADD2 NMSUB
  97. #define MADD3 MADD
  98. #define MADD4 MADD
  99. #endif
  100. #if defined(CONJ) && defined(XCONJ)
  101. #define MADD1 MADD
  102. #define MADD2 NMSUB
  103. #define MADD3 NMSUB
  104. #define MADD4 NMSUB
  105. #endif
  106. PROLOGUE
  107. LDARG INCX, 0($sp)
  108. LDARG Y, 8($sp)
  109. LDARG INCY, 16($sp)
  110. LDARG BUFFER, 24($sp)
  111. #ifndef __64BIT__
  112. daddiu $sp, $sp, -64
  113. #else
  114. daddiu $sp, $sp, -32
  115. #endif
  116. SDARG $16, 0($sp)
  117. SDARG $17, 8($sp)
  118. sdc1 $f24, 16($sp)
  119. sdc1 $f25, 24($sp)
  120. #ifndef __64BIT__
  121. sdc1 $f20, 32($sp)
  122. sdc1 $f21, 40($sp)
  123. sdc1 $f22, 48($sp)
  124. sdc1 $f23, 56($sp)
  125. #endif
  126. dsll LDA, LDA, ZBASE_SHIFT
  127. blez M, .L999
  128. dsll INCX, INCX, ZBASE_SHIFT
  129. blez N, .L999
  130. dsll INCY, INCY, ZBASE_SHIFT
  131. li YORIG, 2 * SIZE
  132. beq INCY, YORIG, .L10
  133. move YORIG, Y
  134. dsra I, M, 2
  135. move YORIG, BUFFER
  136. move XX, Y
  137. blez I, .L05
  138. move YY, BUFFER
  139. .align 3
  140. .L02:
  141. LD a1, 0 * SIZE(XX)
  142. LD a2, 1 * SIZE(XX)
  143. daddu XX, XX, INCY
  144. LD a3, 0 * SIZE(XX)
  145. LD a4, 1 * SIZE(XX)
  146. daddu XX, XX, INCY
  147. LD a5, 0 * SIZE(XX)
  148. LD a6, 1 * SIZE(XX)
  149. daddu XX, XX, INCY
  150. LD a7, 0 * SIZE(XX)
  151. LD a8, 1 * SIZE(XX)
  152. daddu XX, XX, INCY
  153. daddiu I, I, -1
  154. daddiu YY, YY, 8 * SIZE
  155. ST a1, -8 * SIZE(YY)
  156. ST a2, -7 * SIZE(YY)
  157. ST a3, -6 * SIZE(YY)
  158. ST a4, -5 * SIZE(YY)
  159. ST a5, -4 * SIZE(YY)
  160. ST a6, -3 * SIZE(YY)
  161. ST a7, -2 * SIZE(YY)
  162. bgtz I, .L02
  163. ST a8, -1 * SIZE(YY)
  164. .align 3
  165. .L05:
  166. andi I, M, 3
  167. blez I, .L10
  168. NOP
  169. .align 3
  170. .L06:
  171. LD a1, 0 * SIZE(XX)
  172. LD a2, 1 * SIZE(XX)
  173. daddu XX, XX, INCY
  174. daddiu I, I, -1
  175. ST a1, 0 * SIZE(YY)
  176. ST a2, 1 * SIZE(YY)
  177. bgtz I, .L06
  178. daddiu YY, YY, 2 * SIZE
  179. .align 3
  180. .L10:
  181. dsra J, N, 1
  182. blez J, .L20
  183. NOP
  184. .align 3
  185. .L11:
  186. LD x1, 0 * SIZE(X)
  187. LD x2, 1 * SIZE(X)
  188. daddu X, X, INCX
  189. LD x3, 0 * SIZE(X)
  190. LD x4, 1 * SIZE(X)
  191. daddu X, X, INCX
  192. MUL a1, ALPHA_R, x1
  193. move AO1, A
  194. MUL a2, ALPHA_I, x1
  195. daddu AO2, A, LDA
  196. MUL a3, ALPHA_R, x3
  197. daddu A, AO2, LDA
  198. MUL a4, ALPHA_I, x3
  199. #ifndef XCONJ
  200. NMSUB x1, a1, ALPHA_I, x2
  201. MADD x2, a2, ALPHA_R, x2
  202. NMSUB x3, a3, ALPHA_I, x4
  203. MADD x4, a4, ALPHA_R, x4
  204. #else
  205. MADD x1, a1, ALPHA_I, x2
  206. MSUB x2, a2, ALPHA_R, x2
  207. MADD x3, a3, ALPHA_I, x4
  208. MSUB x4, a4, ALPHA_R, x4
  209. #endif
  210. dsra I, M, 2
  211. blez I, .L15
  212. move YY, YORIG
  213. LD y1, 0 * SIZE(YY)
  214. LD a1, 0 * SIZE(AO1)
  215. LD y2, 1 * SIZE(YY)
  216. LD a3, 2 * SIZE(AO1)
  217. LD y3, 2 * SIZE(YY)
  218. LD a2, 1 * SIZE(AO1)
  219. LD y4, 3 * SIZE(YY)
  220. LD a4, 3 * SIZE(AO1)
  221. LD a5, 0 * SIZE(AO2)
  222. LD a6, 1 * SIZE(AO2)
  223. LD a7, 2 * SIZE(AO2)
  224. LD a8, 3 * SIZE(AO2)
  225. MADD1 t1, y1, x1, a1
  226. LD y1, 4 * SIZE(YY)
  227. MADD2 t2, y2, x2, a1
  228. LD a1, 4 * SIZE(AO1)
  229. MADD1 t3, y3, x1, a3
  230. LD y2, 5 * SIZE(YY)
  231. MADD2 t4, y4, x2, a3
  232. LD a3, 6 * SIZE(AO1)
  233. MADD3 t1, t1, x2, a2
  234. LD y3, 6 * SIZE(YY)
  235. MADD4 t2, t2, x1, a2
  236. LD a2, 5 * SIZE(AO1)
  237. MADD3 t3, t3, x2, a4
  238. LD y4, 7 * SIZE(YY)
  239. MADD4 t4, t4, x1, a4
  240. LD a4, 7 * SIZE(AO1)
  241. MADD1 t1, t1, x3, a5
  242. NOP
  243. MADD2 t2, t2, x4, a5
  244. LD a5, 4 * SIZE(AO2)
  245. MADD1 t3, t3, x3, a7
  246. NOP
  247. MADD2 t4, t4, x4, a7
  248. LD a7, 6 * SIZE(AO2)
  249. MADD3 t1, t1, x4, a6
  250. NOP
  251. MADD4 t2, t2, x3, a6
  252. LD a6, 5 * SIZE(AO2)
  253. MADD3 t3, t3, x4, a8
  254. daddiu I, I, -1
  255. MADD4 t4, t4, x3, a8
  256. blez I, .L13
  257. LD a8, 7 * SIZE(AO2)
  258. .align 3
  259. .L12:
  260. MADD1 t5, y1, x1, a1
  261. LD y1, 8 * SIZE(YY)
  262. MADD2 t6, y2, x2, a1
  263. LD a1, 8 * SIZE(AO1)
  264. MADD1 t7, y3, x1, a3
  265. LD y2, 9 * SIZE(YY)
  266. MADD2 t8, y4, x2, a3
  267. LD a3, 10 * SIZE(AO1)
  268. MADD3 t5, t5, x2, a2
  269. LD y3, 10 * SIZE(YY)
  270. MADD4 t6, t6, x1, a2
  271. LD a2, 9 * SIZE(AO1)
  272. MADD3 t7, t7, x2, a4
  273. LD y4, 11 * SIZE(YY)
  274. MADD4 t8, t8, x1, a4
  275. LD a4, 11 * SIZE(AO1)
  276. MADD1 t5, t5, x3, a5
  277. ST t1, 0 * SIZE(YY)
  278. MADD2 t6, t6, x4, a5
  279. LD a5, 8 * SIZE(AO2)
  280. MADD1 t7, t7, x3, a7
  281. ST t2, 1 * SIZE(YY)
  282. MADD2 t8, t8, x4, a7
  283. LD a7, 10 * SIZE(AO2)
  284. MADD3 t5, t5, x4, a6
  285. ST t3, 2 * SIZE(YY)
  286. MADD4 t6, t6, x3, a6
  287. LD a6, 9 * SIZE(AO2)
  288. MADD3 t7, t7, x4, a8
  289. ST t4, 3 * SIZE(YY)
  290. MADD4 t8, t8, x3, a8
  291. LD a8, 11 * SIZE(AO2)
  292. MADD1 t1, y1, x1, a1
  293. LD y1, 12 * SIZE(YY)
  294. MADD2 t2, y2, x2, a1
  295. LD a1, 12 * SIZE(AO1)
  296. MADD1 t3, y3, x1, a3
  297. LD y2, 13 * SIZE(YY)
  298. MADD2 t4, y4, x2, a3
  299. LD a3, 14 * SIZE(AO1)
  300. MADD3 t1, t1, x2, a2
  301. LD y3, 14 * SIZE(YY)
  302. MADD4 t2, t2, x1, a2
  303. LD a2, 13 * SIZE(AO1)
  304. MADD3 t3, t3, x2, a4
  305. LD y4, 15 * SIZE(YY)
  306. MADD4 t4, t4, x1, a4
  307. LD a4, 15 * SIZE(AO1)
  308. MADD1 t1, t1, x3, a5
  309. ST t5, 4 * SIZE(YY)
  310. MADD2 t2, t2, x4, a5
  311. LD a5, 12 * SIZE(AO2)
  312. MADD1 t3, t3, x3, a7
  313. ST t6, 5 * SIZE(YY)
  314. MADD2 t4, t4, x4, a7
  315. LD a7, 14 * SIZE(AO2)
  316. MADD3 t1, t1, x4, a6
  317. ST t7, 6 * SIZE(YY)
  318. MADD4 t2, t2, x3, a6
  319. LD a6, 13 * SIZE(AO2)
  320. MADD3 t3, t3, x4, a8
  321. ST t8, 7 * SIZE(YY)
  322. MADD4 t4, t4, x3, a8
  323. LD a8, 15 * SIZE(AO2)
  324. daddiu I, I, -1
  325. daddiu YY, YY, 8 * SIZE
  326. daddiu AO1, AO1, 8 * SIZE
  327. bgtz I, .L12
  328. daddiu AO2, AO2, 8 * SIZE
  329. .align 3
  330. .L13:
  331. ST t1, 0 * SIZE(YY)
  332. MADD1 t1, y1, x1, a1
  333. ST t2, 1 * SIZE(YY)
  334. MADD2 t2, y2, x2, a1
  335. ST t3, 2 * SIZE(YY)
  336. MADD1 t3, y3, x1, a3
  337. ST t4, 3 * SIZE(YY)
  338. MADD2 t4, y4, x2, a3
  339. MADD3 t1, t1, x2, a2
  340. MADD4 t2, t2, x1, a2
  341. MADD3 t3, t3, x2, a4
  342. MADD4 t4, t4, x1, a4
  343. MADD1 t1, t1, x3, a5
  344. MADD2 t2, t2, x4, a5
  345. MADD1 t3, t3, x3, a7
  346. MADD2 t4, t4, x4, a7
  347. MADD3 t1, t1, x4, a6
  348. daddiu AO1, AO1, 8 * SIZE
  349. MADD4 t2, t2, x3, a6
  350. daddiu AO2, AO2, 8 * SIZE
  351. MADD3 t3, t3, x4, a8
  352. daddiu YY, YY, 8 * SIZE
  353. MADD4 t4, t4, x3, a8
  354. NOP
  355. ST t1, -4 * SIZE(YY)
  356. ST t2, -3 * SIZE(YY)
  357. ST t3, -2 * SIZE(YY)
  358. ST t4, -1 * SIZE(YY)
  359. .align 3
  360. .L15:
  361. andi I, M, 2
  362. NOP
  363. blez I, .L16
  364. NOP
  365. LD a1, 0 * SIZE(AO1)
  366. LD y1, 0 * SIZE(YY)
  367. LD a2, 1 * SIZE(AO1)
  368. LD y2, 1 * SIZE(YY)
  369. LD a3, 2 * SIZE(AO1)
  370. LD y3, 2 * SIZE(YY)
  371. LD a4, 3 * SIZE(AO1)
  372. LD y4, 3 * SIZE(YY)
  373. MADD1 t1, y1, x1, a1
  374. LD a5, 0 * SIZE(AO2)
  375. MADD2 t2, y2, x2, a1
  376. LD a6, 1 * SIZE(AO2)
  377. MADD1 t3, y3, x1, a3
  378. LD a7, 2 * SIZE(AO2)
  379. MADD2 t4, y4, x2, a3
  380. LD a8, 3 * SIZE(AO2)
  381. MADD3 t1, t1, x2, a2
  382. MADD4 t2, t2, x1, a2
  383. MADD3 t3, t3, x2, a4
  384. MADD4 t4, t4, x1, a4
  385. MADD1 t1, t1, x3, a5
  386. MADD2 t2, t2, x4, a5
  387. MADD1 t3, t3, x3, a7
  388. MADD2 t4, t4, x4, a7
  389. MADD3 t1, t1, x4, a6
  390. daddiu YY, YY, 4 * SIZE
  391. MADD4 t2, t2, x3, a6
  392. daddiu AO1, AO1, 4 * SIZE
  393. MADD3 t3, t3, x4, a8
  394. daddiu AO2, AO2, 4 * SIZE
  395. MADD4 t4, t4, x3, a8
  396. NOP
  397. ST t1, -4 * SIZE(YY)
  398. ST t2, -3 * SIZE(YY)
  399. ST t3, -2 * SIZE(YY)
  400. ST t4, -1 * SIZE(YY)
  401. .align 3
  402. .L16:
  403. andi I, M, 1
  404. NOP
  405. blez I, .L19
  406. NOP
  407. LD y1, 0 * SIZE(YY)
  408. LD y2, 1 * SIZE(YY)
  409. LD a1, 0 * SIZE(AO1)
  410. LD a2, 1 * SIZE(AO1)
  411. MADD1 t1, y1, x1, a1
  412. LD a5, 0 * SIZE(AO2)
  413. MADD2 t2, y2, x2, a1
  414. LD a6, 1 * SIZE(AO2)
  415. MADD3 t1, t1, x2, a2
  416. MADD4 t2, t2, x1, a2
  417. MADD1 t1, t1, x3, a5
  418. MADD2 t2, t2, x4, a5
  419. MADD3 t1, t1, x4, a6
  420. MADD4 t2, t2, x3, a6
  421. ST t1, 0 * SIZE(YY)
  422. ST t2, 1 * SIZE(YY)
  423. .align 3
  424. .L19:
  425. daddiu J, J, -1
  426. bgtz J, .L11
  427. NOP
  428. .align 3
  429. .L20:
  430. andi J, N, 1
  431. blez J, .L900
  432. NOP
  433. LD x1, 0 * SIZE(X)
  434. LD x2, 1 * SIZE(X)
  435. daddu X, X, INCX
  436. MUL a1, ALPHA_R, x1
  437. move AO1, A
  438. MUL a2, ALPHA_I, x1
  439. #ifndef XCONJ
  440. NMSUB x1, a1, ALPHA_I, x2
  441. MADD x2, a2, ALPHA_R, x2
  442. #else
  443. MADD x1, a1, ALPHA_I, x2
  444. MSUB x2, a2, ALPHA_R, x2
  445. #endif
  446. dsra I, M, 2
  447. blez I, .L25
  448. move YY, YORIG
  449. LD y1, 0 * SIZE(YY)
  450. LD a1, 0 * SIZE(AO1)
  451. LD y2, 1 * SIZE(YY)
  452. LD a3, 2 * SIZE(AO1)
  453. LD y3, 2 * SIZE(YY)
  454. LD a2, 1 * SIZE(AO1)
  455. LD y4, 3 * SIZE(YY)
  456. LD a4, 3 * SIZE(AO1)
  457. MADD1 t1, y1, x1, a1
  458. LD y1, 4 * SIZE(YY)
  459. MADD2 t2, y2, x2, a1
  460. LD a1, 4 * SIZE(AO1)
  461. MADD1 t3, y3, x1, a3
  462. LD y2, 5 * SIZE(YY)
  463. MADD2 t4, y4, x2, a3
  464. LD a3, 6 * SIZE(AO1)
  465. MADD3 t1, t1, x2, a2
  466. LD y3, 6 * SIZE(YY)
  467. MADD4 t2, t2, x1, a2
  468. LD a2, 5 * SIZE(AO1)
  469. MADD3 t3, t3, x2, a4
  470. LD y4, 7 * SIZE(YY)
  471. MADD4 t4, t4, x1, a4
  472. daddiu I, I, -1
  473. blez I, .L23
  474. LD a4, 7 * SIZE(AO1)
  475. .align 3
  476. .L22:
  477. MADD1 t5, y1, x1, a1
  478. LD y1, 8 * SIZE(YY)
  479. MADD2 t6, y2, x2, a1
  480. LD a1, 8 * SIZE(AO1)
  481. MADD1 t7, y3, x1, a3
  482. LD y2, 9 * SIZE(YY)
  483. MADD2 t8, y4, x2, a3
  484. LD a3, 10 * SIZE(AO1)
  485. MADD3 t5, t5, x2, a2
  486. LD y3, 10 * SIZE(YY)
  487. MADD4 t6, t6, x1, a2
  488. LD a2, 9 * SIZE(AO1)
  489. MADD3 t7, t7, x2, a4
  490. LD y4, 11 * SIZE(YY)
  491. MADD4 t8, t8, x1, a4
  492. LD a4, 11 * SIZE(AO1)
  493. ST t1, 0 * SIZE(YY)
  494. ST t2, 1 * SIZE(YY)
  495. ST t3, 2 * SIZE(YY)
  496. ST t4, 3 * SIZE(YY)
  497. MADD1 t1, y1, x1, a1
  498. LD y1, 12 * SIZE(YY)
  499. MADD2 t2, y2, x2, a1
  500. LD a1, 12 * SIZE(AO1)
  501. MADD1 t3, y3, x1, a3
  502. LD y2, 13 * SIZE(YY)
  503. MADD2 t4, y4, x2, a3
  504. LD a3, 14 * SIZE(AO1)
  505. MADD3 t1, t1, x2, a2
  506. LD y3, 14 * SIZE(YY)
  507. MADD4 t2, t2, x1, a2
  508. LD a2, 13 * SIZE(AO1)
  509. MADD3 t3, t3, x2, a4
  510. LD y4, 15 * SIZE(YY)
  511. MADD4 t4, t4, x1, a4
  512. LD a4, 15 * SIZE(AO1)
  513. ST t5, 4 * SIZE(YY)
  514. ST t6, 5 * SIZE(YY)
  515. ST t7, 6 * SIZE(YY)
  516. ST t8, 7 * SIZE(YY)
  517. daddiu I, I, -1
  518. daddiu YY, YY, 8 * SIZE
  519. bgtz I, .L22
  520. daddiu AO1, AO1, 8 * SIZE
  521. .align 3
  522. .L23:
  523. ST t1, 0 * SIZE(YY)
  524. MADD1 t1, y1, x1, a1
  525. ST t2, 1 * SIZE(YY)
  526. MADD2 t2, y2, x2, a1
  527. ST t3, 2 * SIZE(YY)
  528. MADD1 t3, y3, x1, a3
  529. ST t4, 3 * SIZE(YY)
  530. MADD2 t4, y4, x2, a3
  531. MADD3 t1, t1, x2, a2
  532. daddiu AO1, AO1, 8 * SIZE
  533. MADD4 t2, t2, x1, a2
  534. daddiu YY, YY, 8 * SIZE
  535. MADD3 t3, t3, x2, a4
  536. MADD4 t4, t4, x1, a4
  537. ST t1, -4 * SIZE(YY)
  538. ST t2, -3 * SIZE(YY)
  539. ST t3, -2 * SIZE(YY)
  540. ST t4, -1 * SIZE(YY)
  541. .align 3
  542. .L25:
  543. andi I, M, 2
  544. NOP
  545. blez I, .L26
  546. NOP
  547. LD a1, 0 * SIZE(AO1)
  548. LD y1, 0 * SIZE(YY)
  549. LD a2, 1 * SIZE(AO1)
  550. LD y2, 1 * SIZE(YY)
  551. LD a3, 2 * SIZE(AO1)
  552. LD y3, 2 * SIZE(YY)
  553. LD a4, 3 * SIZE(AO1)
  554. LD y4, 3 * SIZE(YY)
  555. MADD1 t1, y1, x1, a1
  556. MADD2 t2, y2, x2, a1
  557. MADD1 t3, y3, x1, a3
  558. MADD2 t4, y4, x2, a3
  559. MADD3 t1, t1, x2, a2
  560. daddiu YY, YY, 4 * SIZE
  561. MADD4 t2, t2, x1, a2
  562. daddiu AO1, AO1, 4 * SIZE
  563. MADD3 t3, t3, x2, a4
  564. MADD4 t4, t4, x1, a4
  565. ST t1, -4 * SIZE(YY)
  566. ST t2, -3 * SIZE(YY)
  567. ST t3, -2 * SIZE(YY)
  568. ST t4, -1 * SIZE(YY)
  569. .align 3
  570. .L26:
  571. andi I, M, 1
  572. NOP
  573. blez I, .L900
  574. NOP
  575. LD y1, 0 * SIZE(YY)
  576. LD y2, 1 * SIZE(YY)
  577. LD a1, 0 * SIZE(AO1)
  578. LD a2, 1 * SIZE(AO1)
  579. MADD1 t1, y1, x1, a1
  580. MADD2 t2, y2, x2, a1
  581. MADD3 t1, t1, x2, a2
  582. MADD4 t2, t2, x1, a2
  583. ST t1, 0 * SIZE(YY)
  584. ST t2, 1 * SIZE(YY)
  585. .align 3
  586. .L900:
  587. li YORIG, 2 * SIZE
  588. beq INCY, YORIG, .L999
  589. dsra I, M, 2
  590. blez I, .L905
  591. move XX, BUFFER
  592. .align 3
  593. .L902:
  594. LD a1, 0 * SIZE(XX)
  595. LD a2, 1 * SIZE(XX)
  596. LD a3, 2 * SIZE(XX)
  597. LD a4, 3 * SIZE(XX)
  598. LD a5, 4 * SIZE(XX)
  599. LD a6, 5 * SIZE(XX)
  600. LD a7, 6 * SIZE(XX)
  601. LD a8, 7 * SIZE(XX)
  602. daddiu I, I, -1
  603. ST a1, 0 * SIZE(Y)
  604. ST a2, 1 * SIZE(Y)
  605. daddu Y, Y, INCY
  606. ST a3, 0 * SIZE(Y)
  607. ST a4, 1 * SIZE(Y)
  608. daddu Y, Y, INCY
  609. ST a5, 0 * SIZE(Y)
  610. ST a6, 1 * SIZE(Y)
  611. daddu Y, Y, INCY
  612. ST a7, 0 * SIZE(Y)
  613. ST a8, 1 * SIZE(Y)
  614. daddu Y, Y, INCY
  615. bgtz I, .L902
  616. daddiu XX, XX, 8 * SIZE
  617. .align 3
  618. .L905:
  619. andi I, M, 3
  620. blez I, .L999
  621. NOP
  622. .align 3
  623. .L906:
  624. LD a1, 0 * SIZE(XX)
  625. LD a2, 1 * SIZE(XX)
  626. daddiu XX, XX, 2 * SIZE
  627. daddiu I, I, -1
  628. ST a1, 0 * SIZE(Y)
  629. ST a2, 1 * SIZE(Y)
  630. bgtz I, .L906
  631. daddu Y, Y, INCY
  632. .align 3
  633. .L999:
  634. LDARG $16, 0($sp)
  635. LDARG $17, 8($sp)
  636. ldc1 $f24, 16($sp)
  637. ldc1 $f25, 24($sp)
  638. #ifndef __64BIT__
  639. ldc1 $f20, 32($sp)
  640. ldc1 $f21, 40($sp)
  641. ldc1 $f22, 48($sp)
  642. ldc1 $f23, 56($sp)
  643. #endif
  644. j $31
  645. #ifdef __64BIT__
  646. daddiu $sp, $sp, 32
  647. #else
  648. daddiu $sp, $sp, 64
  649. #endif
  650. EPILOGUE