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.

amax.S 10 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523
  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 N r3
  41. #define X r4
  42. #define INCX r5
  43. #define PREA r8
  44. #define FZERO f1
  45. #define STACKSIZE 160
  46. PROLOGUE
  47. PROFCODE
  48. addi SP, SP, -STACKSIZE
  49. li r0, 0
  50. stfd f14, 0(SP)
  51. stfd f15, 8(SP)
  52. stfd f16, 16(SP)
  53. stfd f17, 24(SP)
  54. stfd f18, 32(SP)
  55. stfd f19, 40(SP)
  56. stfd f20, 48(SP)
  57. stfd f21, 56(SP)
  58. stfd f22, 64(SP)
  59. stfd f23, 72(SP)
  60. stfd f24, 80(SP)
  61. stfd f25, 88(SP)
  62. stfd f26, 96(SP)
  63. stfd f27, 104(SP)
  64. stfd f28, 112(SP)
  65. stfd f29, 120(SP)
  66. stfd f30, 128(SP)
  67. stfd f31, 136(SP)
  68. stw r0, 144(SP)
  69. lfs FZERO,144(SP)
  70. #ifdef F_INTERFACE
  71. LDINT N, 0(N)
  72. LDINT INCX, 0(INCX)
  73. #endif
  74. slwi INCX, INCX, BASE_SHIFT
  75. li PREA, L1_PREFETCHSIZE
  76. cmpwi cr0, N, 0
  77. ble- LL(9999)
  78. cmpwi cr0, INCX, 0
  79. ble- LL(9999)
  80. LFD f1, 0 * SIZE(X)
  81. add X, X, INCX
  82. fabs f0, f1
  83. fabs f2, f1
  84. fabs f3, f1
  85. fabs f4, f1
  86. fabs f5, f1
  87. fabs f6, f1
  88. fabs f7, f1
  89. fabs f1, f1
  90. subi N, N, 1
  91. cmpwi cr0, INCX, SIZE
  92. bne- cr0, LL(100)
  93. srawi. r0, N, 4
  94. mtspr CTR, r0
  95. beq- cr0, LL(50)
  96. LFD f24, 0 * SIZE(X)
  97. LFD f25, 1 * SIZE(X)
  98. LFD f26, 2 * SIZE(X)
  99. LFD f27, 3 * SIZE(X)
  100. LFD f28, 4 * SIZE(X)
  101. LFD f29, 5 * SIZE(X)
  102. LFD f30, 6 * SIZE(X)
  103. LFD f31, 7 * SIZE(X)
  104. fabs f8, f24
  105. fabs f9, f25
  106. fabs f10, f26
  107. fabs f11, f27
  108. LFD f24, 8 * SIZE(X)
  109. LFD f25, 9 * SIZE(X)
  110. LFD f26, 10 * SIZE(X)
  111. LFD f27, 11 * SIZE(X)
  112. fabs f12, f28
  113. fabs f13, f29
  114. fabs f14, f30
  115. fabs f15, f31
  116. LFD f28, 12 * SIZE(X)
  117. LFD f29, 13 * SIZE(X)
  118. LFD f30, 14 * SIZE(X)
  119. LFD f31, 15 * SIZE(X)
  120. bdz LL(20)
  121. .align 4
  122. LL(10):
  123. fsub f16, f0, f8
  124. fsub f17, f1, f9
  125. fsub f18, f2, f10
  126. fsub f19, f3, f11
  127. fsub f20, f4, f12
  128. fsub f21, f5, f13
  129. fsub f22, f6, f14
  130. fsub f23, f7, f15
  131. fsel f0, f16, f0, f8
  132. fabs f8, f24
  133. fsel f1, f17, f1, f9
  134. fabs f9, f25
  135. fsel f2, f18, f2, f10
  136. fabs f10, f26
  137. fsel f3, f19, f3, f11
  138. fabs f11, f27
  139. LFD f24, 16 * SIZE(X)
  140. LFD f25, 17 * SIZE(X)
  141. LFD f26, 18 * SIZE(X)
  142. LFD f27, 19 * SIZE(X)
  143. fsel f4, f20, f4, f12
  144. fabs f12, f28
  145. fsel f5, f21, f5, f13
  146. fabs f13, f29
  147. fsel f6, f22, f6, f14
  148. fabs f14, f30
  149. fsel f7, f23, f7, f15
  150. fabs f15, f31
  151. LFD f28, 20 * SIZE(X)
  152. LFD f29, 21 * SIZE(X)
  153. LFD f30, 22 * SIZE(X)
  154. LFD f31, 23 * SIZE(X)
  155. fsub f16, f0, f8
  156. fsub f17, f1, f9
  157. fsub f18, f2, f10
  158. fsub f19, f3, f11
  159. fsub f20, f4, f12
  160. fsub f21, f5, f13
  161. fsub f22, f6, f14
  162. fsub f23, f7, f15
  163. fsel f0, f16, f0, f8
  164. fabs f8, f24
  165. fsel f1, f17, f1, f9
  166. fabs f9, f25
  167. fsel f2, f18, f2, f10
  168. fabs f10, f26
  169. fsel f3, f19, f3, f11
  170. fabs f11, f27
  171. LFD f24, 24 * SIZE(X)
  172. LFD f25, 25 * SIZE(X)
  173. LFD f26, 26 * SIZE(X)
  174. LFD f27, 27 * SIZE(X)
  175. fsel f4, f20, f4, f12
  176. fabs f12, f28
  177. fsel f5, f21, f5, f13
  178. fabs f13, f29
  179. fsel f6, f22, f6, f14
  180. fabs f14, f30
  181. fsel f7, f23, f7, f15
  182. fabs f15, f31
  183. LFD f28, 28 * SIZE(X)
  184. LFD f29, 29 * SIZE(X)
  185. LFD f30, 30 * SIZE(X)
  186. LFD f31, 31 * SIZE(X)
  187. #ifndef POWER6
  188. L1_PREFETCH X, PREA
  189. #endif
  190. addi X, X, 16 * SIZE
  191. #ifdef POWER6
  192. L1_PREFETCH X, PREA
  193. #endif
  194. bdnz LL(10)
  195. .align 4
  196. LL(20):
  197. fsub f16, f0, f8
  198. fsub f17, f1, f9
  199. fsub f18, f2, f10
  200. fsub f19, f3, f11
  201. fsub f20, f4, f12
  202. fsub f21, f5, f13
  203. fsub f22, f6, f14
  204. fsub f23, f7, f15
  205. fsel f0, f16, f0, f8
  206. fabs f8, f24
  207. fsel f1, f17, f1, f9
  208. fabs f9, f25
  209. fsel f2, f18, f2, f10
  210. fabs f10, f26
  211. fsel f3, f19, f3, f11
  212. fabs f11, f27
  213. fsel f4, f20, f4, f12
  214. fabs f12, f28
  215. fsel f5, f21, f5, f13
  216. fabs f13, f29
  217. fsel f6, f22, f6, f14
  218. fabs f14, f30
  219. fsel f7, f23, f7, f15
  220. fabs f15, f31
  221. fsub f16, f0, f8
  222. fsub f17, f1, f9
  223. fsub f18, f2, f10
  224. fsub f19, f3, f11
  225. fsub f20, f4, f12
  226. fsub f21, f5, f13
  227. fsub f22, f6, f14
  228. fsub f23, f7, f15
  229. fsel f0, f16, f0, f8
  230. fsel f1, f17, f1, f9
  231. fsel f2, f18, f2, f10
  232. fsel f3, f19, f3, f11
  233. fsel f4, f20, f4, f12
  234. fsel f5, f21, f5, f13
  235. fsel f6, f22, f6, f14
  236. fsel f7, f23, f7, f15
  237. addi X, X, 16 * SIZE
  238. .align 4
  239. LL(50):
  240. andi. r0, N, 15
  241. mtspr CTR, r0
  242. beq LL(999)
  243. .align 4
  244. LL(60):
  245. LFD f8, 0 * SIZE(X)
  246. addi X, X, 1 * SIZE
  247. fabs f8, f8
  248. fsub f16, f1, f8
  249. fsel f1, f16, f1, f8
  250. bdnz LL(60)
  251. b LL(999)
  252. .align 4
  253. LL(100):
  254. sub X, X, INCX
  255. srawi. r0, N, 4
  256. mtspr CTR, r0
  257. beq- LL(150)
  258. LFDUX f24, X, INCX
  259. LFDUX f25, X, INCX
  260. LFDUX f26, X, INCX
  261. LFDUX f27, X, INCX
  262. LFDUX f28, X, INCX
  263. LFDUX f29, X, INCX
  264. LFDUX f30, X, INCX
  265. LFDUX f31, X, INCX
  266. fabs f8, f24
  267. fabs f9, f25
  268. fabs f10, f26
  269. fabs f11, f27
  270. LFDUX f24, X, INCX
  271. LFDUX f25, X, INCX
  272. LFDUX f26, X, INCX
  273. LFDUX f27, X, INCX
  274. fabs f12, f28
  275. fabs f13, f29
  276. fabs f14, f30
  277. fabs f15, f31
  278. LFDUX f28, X, INCX
  279. LFDUX f29, X, INCX
  280. LFDUX f30, X, INCX
  281. LFDUX f31, X, INCX
  282. bdz LL(120)
  283. .align 4
  284. LL(110):
  285. fsub f16, f0, f8
  286. fsub f17, f1, f9
  287. fsub f18, f2, f10
  288. fsub f19, f3, f11
  289. fsub f20, f4, f12
  290. fsub f21, f5, f13
  291. fsub f22, f6, f14
  292. fsub f23, f7, f15
  293. fsel f0, f16, f0, f8
  294. fabs f8, f24
  295. fsel f1, f17, f1, f9
  296. fabs f9, f25
  297. fsel f2, f18, f2, f10
  298. fabs f10, f26
  299. fsel f3, f19, f3, f11
  300. fabs f11, f27
  301. LFDUX f24, X, INCX
  302. LFDUX f25, X, INCX
  303. LFDUX f26, X, INCX
  304. LFDUX f27, X, INCX
  305. fsel f4, f20, f4, f12
  306. fabs f12, f28
  307. fsel f5, f21, f5, f13
  308. fabs f13, f29
  309. fsel f6, f22, f6, f14
  310. fabs f14, f30
  311. fsel f7, f23, f7, f15
  312. fabs f15, f31
  313. LFDUX f28, X, INCX
  314. LFDUX f29, X, INCX
  315. LFDUX f30, X, INCX
  316. LFDUX f31, X, INCX
  317. fsub f16, f0, f8
  318. fsub f17, f1, f9
  319. fsub f18, f2, f10
  320. fsub f19, f3, f11
  321. fsub f20, f4, f12
  322. fsub f21, f5, f13
  323. fsub f22, f6, f14
  324. fsub f23, f7, f15
  325. fsel f0, f16, f0, f8
  326. fabs f8, f24
  327. fsel f1, f17, f1, f9
  328. fabs f9, f25
  329. fsel f2, f18, f2, f10
  330. fabs f10, f26
  331. fsel f3, f19, f3, f11
  332. fabs f11, f27
  333. LFDUX f24, X, INCX
  334. LFDUX f25, X, INCX
  335. LFDUX f26, X, INCX
  336. LFDUX f27, X, INCX
  337. fsel f4, f20, f4, f12
  338. fabs f12, f28
  339. fsel f5, f21, f5, f13
  340. fabs f13, f29
  341. fsel f6, f22, f6, f14
  342. fabs f14, f30
  343. fsel f7, f23, f7, f15
  344. fabs f15, f31
  345. LFDUX f28, X, INCX
  346. LFDUX f29, X, INCX
  347. LFDUX f30, X, INCX
  348. LFDUX f31, X, INCX
  349. bdnz LL(110)
  350. .align 4
  351. LL(120):
  352. fsub f16, f0, f8
  353. fsub f17, f1, f9
  354. fsub f18, f2, f10
  355. fsub f19, f3, f11
  356. fsub f20, f4, f12
  357. fsub f21, f5, f13
  358. fsub f22, f6, f14
  359. fsub f23, f7, f15
  360. fsel f0, f16, f0, f8
  361. fabs f8, f24
  362. fsel f1, f17, f1, f9
  363. fabs f9, f25
  364. fsel f2, f18, f2, f10
  365. fabs f10, f26
  366. fsel f3, f19, f3, f11
  367. fabs f11, f27
  368. fsel f4, f20, f4, f12
  369. fabs f12, f28
  370. fsel f5, f21, f5, f13
  371. fabs f13, f29
  372. fsel f6, f22, f6, f14
  373. fabs f14, f30
  374. fsel f7, f23, f7, f15
  375. fabs f15, f31
  376. fsub f16, f0, f8
  377. fsub f17, f1, f9
  378. fsub f18, f2, f10
  379. fsub f19, f3, f11
  380. fsub f20, f4, f12
  381. fsub f21, f5, f13
  382. fsub f22, f6, f14
  383. fsub f23, f7, f15
  384. fsel f0, f16, f0, f8
  385. fsel f1, f17, f1, f9
  386. fsel f2, f18, f2, f10
  387. fsel f3, f19, f3, f11
  388. fsel f4, f20, f4, f12
  389. fsel f5, f21, f5, f13
  390. fsel f6, f22, f6, f14
  391. fsel f7, f23, f7, f15
  392. .align 4
  393. LL(150):
  394. andi. r0, N, 15
  395. mtspr CTR, r0
  396. beq LL(999)
  397. .align 4
  398. LL(160):
  399. LFDUX f8, X, INCX
  400. fabs f8, f8
  401. fsub f16, f1, f8
  402. fsel f1, f16, f1, f8
  403. bdnz LL(160)
  404. .align 4
  405. LL(999):
  406. fsub f8, f0, f1
  407. fsub f9, f2, f3
  408. fsub f10, f4, f5
  409. fsub f11, f6, f7
  410. fsel f0, f8, f0, f1
  411. fsel f2, f9, f2, f3
  412. fsel f4, f10, f4, f5
  413. fsel f6, f11, f6, f7
  414. fsub f8, f0, f2
  415. fsub f9, f4, f6
  416. fsel f0, f8, f0, f2
  417. fsel f4, f9, f4, f6
  418. fsub f8, f0, f4
  419. fsel f1, f8, f0, f4
  420. .align 4
  421. LL(9999):
  422. lfd f14, 0(SP)
  423. lfd f15, 8(SP)
  424. lfd f16, 16(SP)
  425. lfd f17, 24(SP)
  426. lfd f18, 32(SP)
  427. lfd f19, 40(SP)
  428. lfd f20, 48(SP)
  429. lfd f21, 56(SP)
  430. lfd f22, 64(SP)
  431. lfd f23, 72(SP)
  432. lfd f24, 80(SP)
  433. lfd f25, 88(SP)
  434. lfd f26, 96(SP)
  435. lfd f27, 104(SP)
  436. lfd f28, 112(SP)
  437. lfd f29, 120(SP)
  438. lfd f30, 128(SP)
  439. lfd f31, 136(SP)
  440. addi SP, SP, STACKSIZE
  441. blr
  442. EPILOGUE