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.

dot.S 9.8 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468
  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 Y r6
  44. #define INCY r7
  45. #define PREA r8
  46. #define FZERO f0
  47. #define STACKSIZE 96
  48. PROLOGUE
  49. PROFCODE
  50. addi SP, SP, -STACKSIZE
  51. li r0, 0
  52. stfd f14, 0(SP)
  53. stfd f15, 8(SP)
  54. stfd f16, 16(SP)
  55. stfd f17, 24(SP)
  56. stfd f18, 32(SP)
  57. stfd f19, 40(SP)
  58. stfd f20, 48(SP)
  59. stfd f21, 56(SP)
  60. stfd f22, 64(SP)
  61. stfd f23, 72(SP)
  62. stw r0, 80(SP)
  63. lfs FZERO, 80(SP)
  64. #ifdef F_INTERFACE
  65. LDINT N, 0(N)
  66. LDINT INCX, 0(INCX)
  67. LDINT INCY, 0(INCY)
  68. #endif
  69. slwi INCX, INCX, BASE_SHIFT
  70. slwi INCY, INCY, BASE_SHIFT
  71. fmr f1, FZERO
  72. fmr f2, FZERO
  73. fmr f3, FZERO
  74. fmr f4, FZERO
  75. fmr f5, FZERO
  76. fmr f6, FZERO
  77. fmr f7, FZERO
  78. #ifdef L1_DUALFETCH
  79. li PREA, (L1_PREFETCHSIZE) / 2
  80. #else
  81. li PREA, (L1_PREFETCHSIZE)
  82. #endif
  83. cmpwi cr0, N, 0
  84. ble- cr0, LL(999)
  85. cmpwi cr0, INCX, SIZE
  86. bne cr0, LL(100)
  87. cmpwi cr0, INCY, SIZE
  88. bne cr0, LL(100)
  89. srawi. r0, N, 4
  90. mtspr CTR, r0
  91. beq- cr0, LL(50)
  92. .align 4
  93. LFD f8, 0 * SIZE(X)
  94. LFD f9, 1 * SIZE(X)
  95. LFD f10, 2 * SIZE(X)
  96. LFD f11, 3 * SIZE(X)
  97. LFD f16, 0 * SIZE(Y)
  98. LFD f17, 1 * SIZE(Y)
  99. LFD f18, 2 * SIZE(Y)
  100. LFD f19, 3 * SIZE(Y)
  101. LFD f12, 4 * SIZE(X)
  102. LFD f13, 5 * SIZE(X)
  103. LFD f14, 6 * SIZE(X)
  104. LFD f15, 7 * SIZE(X)
  105. LFD f20, 4 * SIZE(Y)
  106. LFD f21, 5 * SIZE(Y)
  107. LFD f22, 6 * SIZE(Y)
  108. LFD f23, 7 * SIZE(Y)
  109. bdz LL(20)
  110. .align 4
  111. LL(10):
  112. FMADD f0, f8, f16, f0
  113. FMADD f1, f9, f17, f1
  114. FMADD f2, f10, f18, f2
  115. FMADD f3, f11, f19, f3
  116. LFD f8, 8 * SIZE(X)
  117. LFD f9, 9 * SIZE(X)
  118. LFD f10, 10 * SIZE(X)
  119. LFD f11, 11 * SIZE(X)
  120. LFD f16, 8 * SIZE(Y)
  121. LFD f17, 9 * SIZE(Y)
  122. LFD f18, 10 * SIZE(Y)
  123. LFD f19, 11 * SIZE(Y)
  124. FMADD f4, f12, f20, f4
  125. FMADD f5, f13, f21, f5
  126. FMADD f6, f14, f22, f6
  127. FMADD f7, f15, f23, f7
  128. LFD f12, 12 * SIZE(X)
  129. LFD f13, 13 * SIZE(X)
  130. LFD f14, 14 * SIZE(X)
  131. LFD f15, 15 * SIZE(X)
  132. LFD f20, 12 * SIZE(Y)
  133. LFD f21, 13 * SIZE(Y)
  134. LFD f22, 14 * SIZE(Y)
  135. LFD f23, 15 * SIZE(Y)
  136. FMADD f0, f8, f16, f0
  137. FMADD f1, f9, f17, f1
  138. FMADD f2, f10, f18, f2
  139. FMADD f3, f11, f19, f3
  140. LFD f8, 16 * SIZE(X)
  141. LFD f9, 17 * SIZE(X)
  142. LFD f10, 18 * SIZE(X)
  143. LFD f11, 19 * SIZE(X)
  144. LFD f16, 16 * SIZE(Y)
  145. LFD f17, 17 * SIZE(Y)
  146. LFD f18, 18 * SIZE(Y)
  147. LFD f19, 19 * SIZE(Y)
  148. FMADD f4, f12, f20, f4
  149. FMADD f5, f13, f21, f5
  150. FMADD f6, f14, f22, f6
  151. FMADD f7, f15, f23, f7
  152. LFD f12, 20 * SIZE(X)
  153. LFD f13, 21 * SIZE(X)
  154. LFD f14, 22 * SIZE(X)
  155. LFD f15, 23 * SIZE(X)
  156. LFD f20, 20 * SIZE(Y)
  157. LFD f21, 21 * SIZE(Y)
  158. LFD f22, 22 * SIZE(Y)
  159. LFD f23, 23 * SIZE(Y)
  160. #ifndef POWER6
  161. L1_PREFETCH X, PREA
  162. #ifdef L1_DUALFETCH
  163. L1_PREFETCH Y, PREA
  164. #endif
  165. #endif
  166. addi X, X, 16 * SIZE
  167. addi Y, Y, 16 * SIZE
  168. #ifdef POWER6
  169. L1_PREFETCH X, PREA
  170. #ifdef L1_DUALFETCH
  171. L1_PREFETCH Y, PREA
  172. #endif
  173. #endif
  174. bdnz LL(10)
  175. .align 4
  176. LL(20):
  177. FMADD f0, f8, f16, f0
  178. FMADD f1, f9, f17, f1
  179. FMADD f2, f10, f18, f2
  180. FMADD f3, f11, f19, f3
  181. LFD f8, 8 * SIZE(X)
  182. LFD f9, 9 * SIZE(X)
  183. LFD f10, 10 * SIZE(X)
  184. LFD f11, 11 * SIZE(X)
  185. LFD f16, 8 * SIZE(Y)
  186. LFD f17, 9 * SIZE(Y)
  187. LFD f18, 10 * SIZE(Y)
  188. LFD f19, 11 * SIZE(Y)
  189. FMADD f4, f12, f20, f4
  190. FMADD f5, f13, f21, f5
  191. FMADD f6, f14, f22, f6
  192. FMADD f7, f15, f23, f7
  193. LFD f12, 12 * SIZE(X)
  194. LFD f13, 13 * SIZE(X)
  195. LFD f14, 14 * SIZE(X)
  196. LFD f15, 15 * SIZE(X)
  197. LFD f20, 12 * SIZE(Y)
  198. LFD f21, 13 * SIZE(Y)
  199. LFD f22, 14 * SIZE(Y)
  200. LFD f23, 15 * SIZE(Y)
  201. FMADD f0, f8, f16, f0
  202. FMADD f1, f9, f17, f1
  203. FMADD f2, f10, f18, f2
  204. FMADD f3, f11, f19, f3
  205. FMADD f4, f12, f20, f4
  206. FMADD f5, f13, f21, f5
  207. FMADD f6, f14, f22, f6
  208. FMADD f7, f15, f23, f7
  209. addi X, X, 16 * SIZE
  210. addi Y, Y, 16 * SIZE
  211. .align 4
  212. LL(50):
  213. andi. r0, N, 15
  214. mtspr CTR, r0
  215. beq LL(999)
  216. .align 4
  217. LL(60):
  218. LFD f8, 0 * SIZE(X)
  219. LFD f16, 0 * SIZE(Y)
  220. addi X, X, 1 * SIZE
  221. addi Y, Y, 1 * SIZE
  222. FMADD f0, f8, f16, f0
  223. bdnz LL(60)
  224. b LL(999)
  225. .align 4
  226. LL(100):
  227. #ifdef F_INTERFACE
  228. cmpwi cr0, INCX, 0
  229. bge+ LL(102)
  230. subi r0, N, 1
  231. mullw r0, r0, INCX
  232. sub X, X, r0
  233. .align 4
  234. LL(102):
  235. cmpwi cr0, INCY, 0
  236. bge+ LL(104)
  237. subi r0, N, 1
  238. mullw r0, r0, INCY
  239. sub Y, Y, r0
  240. .align 4
  241. LL(104):
  242. #endif
  243. sub X, X, INCX
  244. sub Y, Y, INCY
  245. srawi. r0, N, 4
  246. mtspr CTR, r0
  247. beq- LL(150)
  248. LFDUX f8, X, INCX
  249. LFDUX f16, Y, INCY
  250. LFDUX f9, X, INCX
  251. LFDUX f17, Y, INCY
  252. LFDUX f10, X, INCX
  253. LFDUX f18, Y, INCY
  254. LFDUX f11, X, INCX
  255. LFDUX f19, Y, INCY
  256. LFDUX f12, X, INCX
  257. LFDUX f20, Y, INCY
  258. LFDUX f13, X, INCX
  259. LFDUX f21, Y, INCY
  260. LFDUX f14, X, INCX
  261. LFDUX f22, Y, INCY
  262. LFDUX f15, X, INCX
  263. LFDUX f23, Y, INCY
  264. bdz LL(120)
  265. .align 4
  266. LL(110):
  267. FMADD f0, f8, f16, f0
  268. FMADD f1, f9, f17, f1
  269. FMADD f2, f10, f18, f2
  270. FMADD f3, f11, f19, f3
  271. LFDUX f8, X, INCX
  272. LFDUX f16, Y, INCY
  273. LFDUX f9, X, INCX
  274. LFDUX f17, Y, INCY
  275. LFDUX f10, X, INCX
  276. LFDUX f18, Y, INCY
  277. LFDUX f11, X, INCX
  278. LFDUX f19, Y, INCY
  279. FMADD f4, f12, f20, f4
  280. FMADD f5, f13, f21, f5
  281. FMADD f6, f14, f22, f6
  282. FMADD f7, f15, f23, f7
  283. LFDUX f12, X, INCX
  284. LFDUX f20, Y, INCY
  285. LFDUX f13, X, INCX
  286. LFDUX f21, Y, INCY
  287. LFDUX f14, X, INCX
  288. LFDUX f22, Y, INCY
  289. LFDUX f15, X, INCX
  290. LFDUX f23, Y, INCY
  291. FMADD f0, f8, f16, f0
  292. FMADD f1, f9, f17, f1
  293. FMADD f2, f10, f18, f2
  294. FMADD f3, f11, f19, f3
  295. LFDUX f8, X, INCX
  296. LFDUX f16, Y, INCY
  297. LFDUX f9, X, INCX
  298. LFDUX f17, Y, INCY
  299. LFDUX f10, X, INCX
  300. LFDUX f18, Y, INCY
  301. LFDUX f11, X, INCX
  302. LFDUX f19, Y, INCY
  303. FMADD f4, f12, f20, f4
  304. FMADD f5, f13, f21, f5
  305. FMADD f6, f14, f22, f6
  306. FMADD f7, f15, f23, f7
  307. LFDUX f12, X, INCX
  308. LFDUX f20, Y, INCY
  309. LFDUX f13, X, INCX
  310. LFDUX f21, Y, INCY
  311. LFDUX f14, X, INCX
  312. LFDUX f22, Y, INCY
  313. LFDUX f15, X, INCX
  314. LFDUX f23, Y, INCY
  315. bdnz LL(110)
  316. .align 4
  317. LL(120):
  318. FMADD f0, f8, f16, f0
  319. FMADD f1, f9, f17, f1
  320. FMADD f2, f10, f18, f2
  321. FMADD f3, f11, f19, f3
  322. LFDUX f8, X, INCX
  323. LFDUX f16, Y, INCY
  324. LFDUX f9, X, INCX
  325. LFDUX f17, Y, INCY
  326. LFDUX f10, X, INCX
  327. LFDUX f18, Y, INCY
  328. LFDUX f11, X, INCX
  329. LFDUX f19, Y, INCY
  330. FMADD f4, f12, f20, f4
  331. FMADD f5, f13, f21, f5
  332. FMADD f6, f14, f22, f6
  333. FMADD f7, f15, f23, f7
  334. LFDUX f12, X, INCX
  335. LFDUX f20, Y, INCY
  336. LFDUX f13, X, INCX
  337. LFDUX f21, Y, INCY
  338. LFDUX f14, X, INCX
  339. LFDUX f22, Y, INCY
  340. LFDUX f15, X, INCX
  341. LFDUX f23, Y, INCY
  342. FMADD f0, f8, f16, f0
  343. FMADD f1, f9, f17, f1
  344. FMADD f2, f10, f18, f2
  345. FMADD f3, f11, f19, f3
  346. FMADD f4, f12, f20, f4
  347. FMADD f5, f13, f21, f5
  348. FMADD f6, f14, f22, f6
  349. FMADD f7, f15, f23, f7
  350. .align 4
  351. LL(150):
  352. andi. r0, N, 15
  353. mtspr CTR, r0
  354. beq LL(999)
  355. .align 4
  356. LL(160):
  357. LFDUX f8, X, INCX
  358. LFDUX f16, Y, INCY
  359. FMADD f0, f8, f16, f0
  360. bdnz LL(160)
  361. .align 4
  362. LL(999):
  363. FADD f0, f0, f1
  364. FADD f2, f2, f3
  365. FADD f4, f4, f5
  366. FADD f6, f6, f7
  367. FADD f0, f0, f2
  368. FADD f4, f4, f6
  369. FADD f1, f0, f4
  370. lfd f14, 0(SP)
  371. lfd f15, 8(SP)
  372. lfd f16, 16(SP)
  373. lfd f17, 24(SP)
  374. lfd f18, 32(SP)
  375. lfd f19, 40(SP)
  376. lfd f20, 48(SP)
  377. lfd f21, 56(SP)
  378. lfd f22, 64(SP)
  379. lfd f23, 72(SP)
  380. addi SP, SP, STACKSIZE
  381. blr
  382. EPILOGUE