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_cell.S 9.6 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458
  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. li PREA, 16 * 20 * SIZE
  79. cmpwi cr0, N, 0
  80. ble- cr0, LL(999)
  81. cmpwi cr0, INCX, SIZE
  82. bne cr0, LL(100)
  83. cmpwi cr0, INCY, SIZE
  84. bne cr0, LL(100)
  85. srawi. r0, N, 4
  86. mtspr CTR, r0
  87. beq- cr0, LL(50)
  88. .align 4
  89. LFD f8, 0 * SIZE(X)
  90. LFD f9, 1 * SIZE(X)
  91. LFD f10, 2 * SIZE(X)
  92. LFD f11, 3 * SIZE(X)
  93. LFD f16, 0 * SIZE(Y)
  94. LFD f17, 1 * SIZE(Y)
  95. LFD f18, 2 * SIZE(Y)
  96. LFD f19, 3 * SIZE(Y)
  97. LFD f12, 4 * SIZE(X)
  98. LFD f13, 5 * SIZE(X)
  99. LFD f14, 6 * SIZE(X)
  100. LFD f15, 7 * SIZE(X)
  101. LFD f20, 4 * SIZE(Y)
  102. LFD f21, 5 * SIZE(Y)
  103. LFD f22, 6 * SIZE(Y)
  104. LFD f23, 7 * SIZE(Y)
  105. bdz LL(20)
  106. .align 4
  107. LL(10):
  108. FMADD f0, f8, f16, f0
  109. LFD f8, 8 * SIZE(X)
  110. LFD f16, 8 * SIZE(Y)
  111. FMADD f1, f9, f17, f1
  112. LFD f9, 9 * SIZE(X)
  113. LFD f17, 9 * SIZE(Y)
  114. FMADD f2, f10, f18, f2
  115. LFD f10, 10 * SIZE(X)
  116. LFD f18, 10 * SIZE(Y)
  117. FMADD f3, f11, f19, f3
  118. LFD f11, 11 * SIZE(X)
  119. LFD f19, 11 * SIZE(Y)
  120. FMADD f4, f12, f20, f4
  121. LFD f12, 12 * SIZE(X)
  122. LFD f20, 12 * SIZE(Y)
  123. FMADD f5, f13, f21, f5
  124. LFD f13, 13 * SIZE(X)
  125. LFD f21, 13 * SIZE(Y)
  126. FMADD f6, f14, f22, f6
  127. LFD f14, 14 * SIZE(X)
  128. LFD f22, 14 * SIZE(Y)
  129. FMADD f7, f15, f23, f7
  130. LFD f15, 15 * SIZE(X)
  131. LFD f23, 15 * SIZE(Y)
  132. FMADD f0, f8, f16, f0
  133. LFD f8, 16 * SIZE(X)
  134. LFD f16, 16 * SIZE(Y)
  135. FMADD f1, f9, f17, f1
  136. LFD f9, 17 * SIZE(X)
  137. LFD f17, 17 * SIZE(Y)
  138. FMADD f2, f10, f18, f2
  139. LFD f10, 18 * SIZE(X)
  140. LFD f18, 18 * SIZE(Y)
  141. FMADD f3, f11, f19, f3
  142. LFD f11, 19 * SIZE(X)
  143. LFD f19, 19 * SIZE(Y)
  144. FMADD f4, f12, f20, f4
  145. LFD f12, 20 * SIZE(X)
  146. LFD f20, 20 * SIZE(Y)
  147. FMADD f5, f13, f21, f5
  148. LFD f13, 21 * SIZE(X)
  149. LFD f21, 21 * SIZE(Y)
  150. FMADD f6, f14, f22, f6
  151. LFD f14, 22 * SIZE(X)
  152. LFD f22, 22 * SIZE(Y)
  153. FMADD f7, f15, f23, f7
  154. LFD f15, 23 * SIZE(X)
  155. LFD f23, 23 * SIZE(Y)
  156. dcbt X, PREA
  157. addi X, X, 16 * SIZE
  158. dcbt Y, PREA
  159. addi Y, Y, 16 * SIZE
  160. bdnz LL(10)
  161. .align 4
  162. LL(20):
  163. FMADD f0, f8, f16, f0
  164. LFD f8, 8 * SIZE(X)
  165. LFD f16, 8 * SIZE(Y)
  166. FMADD f1, f9, f17, f1
  167. LFD f9, 9 * SIZE(X)
  168. LFD f17, 9 * SIZE(Y)
  169. FMADD f2, f10, f18, f2
  170. LFD f10, 10 * SIZE(X)
  171. LFD f18, 10 * SIZE(Y)
  172. FMADD f3, f11, f19, f3
  173. LFD f11, 11 * SIZE(X)
  174. LFD f19, 11 * SIZE(Y)
  175. FMADD f4, f12, f20, f4
  176. LFD f12, 12 * SIZE(X)
  177. LFD f20, 12 * SIZE(Y)
  178. FMADD f5, f13, f21, f5
  179. LFD f13, 13 * SIZE(X)
  180. LFD f21, 13 * SIZE(Y)
  181. FMADD f6, f14, f22, f6
  182. LFD f14, 14 * SIZE(X)
  183. LFD f22, 14 * SIZE(Y)
  184. FMADD f7, f15, f23, f7
  185. LFD f15, 15 * SIZE(X)
  186. LFD f23, 15 * SIZE(Y)
  187. FMADD f0, f8, f16, f0
  188. FMADD f1, f9, f17, f1
  189. FMADD f2, f10, f18, f2
  190. FMADD f3, f11, f19, f3
  191. FMADD f4, f12, f20, f4
  192. FMADD f5, f13, f21, f5
  193. FMADD f6, f14, f22, f6
  194. addi X, X, 16 * SIZE
  195. FMADD f7, f15, f23, f7
  196. addi Y, Y, 16 * SIZE
  197. .align 4
  198. LL(50):
  199. andi. r0, N, 15
  200. mtspr CTR, r0
  201. beq LL(999)
  202. .align 4
  203. LL(60):
  204. LFD f8, 0 * SIZE(X)
  205. LFD f16, 0 * SIZE(Y)
  206. addi X, X, 1 * SIZE
  207. addi Y, Y, 1 * SIZE
  208. FMADD f0, f8, f16, f0
  209. bdnz LL(60)
  210. b LL(999)
  211. .align 4
  212. LL(100):
  213. #ifdef F_INTERFACE
  214. cmpwi cr0, INCX, 0
  215. bge+ LL(102)
  216. subi r0, N, 1
  217. mullw r0, r0, INCX
  218. sub X, X, r0
  219. .align 4
  220. LL(102):
  221. cmpwi cr0, INCY, 0
  222. bge+ LL(104)
  223. subi r0, N, 1
  224. mullw r0, r0, INCY
  225. sub Y, Y, r0
  226. .align 4
  227. LL(104):
  228. #endif
  229. sub X, X, INCX
  230. sub Y, Y, INCY
  231. srawi. r0, N, 4
  232. mtspr CTR, r0
  233. beq- LL(150)
  234. LFDUX f8, X, INCX
  235. LFDUX f16, Y, INCY
  236. LFDUX f9, X, INCX
  237. LFDUX f17, Y, INCY
  238. LFDUX f10, X, INCX
  239. LFDUX f18, Y, INCY
  240. LFDUX f11, X, INCX
  241. LFDUX f19, Y, INCY
  242. LFDUX f12, X, INCX
  243. LFDUX f20, Y, INCY
  244. LFDUX f13, X, INCX
  245. LFDUX f21, Y, INCY
  246. LFDUX f14, X, INCX
  247. LFDUX f22, Y, INCY
  248. LFDUX f15, X, INCX
  249. LFDUX f23, Y, INCY
  250. bdz LL(120)
  251. .align 4
  252. LL(110):
  253. FMADD f0, f8, f16, f0
  254. FMADD f1, f9, f17, f1
  255. FMADD f2, f10, f18, f2
  256. FMADD f3, f11, f19, f3
  257. LFDUX f8, X, INCX
  258. LFDUX f16, Y, INCY
  259. LFDUX f9, X, INCX
  260. LFDUX f17, Y, INCY
  261. LFDUX f10, X, INCX
  262. LFDUX f18, Y, INCY
  263. LFDUX f11, X, INCX
  264. LFDUX f19, Y, INCY
  265. FMADD f4, f12, f20, f4
  266. FMADD f5, f13, f21, f5
  267. FMADD f6, f14, f22, f6
  268. FMADD f7, f15, f23, f7
  269. LFDUX f12, X, INCX
  270. LFDUX f20, Y, INCY
  271. LFDUX f13, X, INCX
  272. LFDUX f21, Y, INCY
  273. LFDUX f14, X, INCX
  274. LFDUX f22, Y, INCY
  275. LFDUX f15, X, INCX
  276. LFDUX f23, Y, INCY
  277. FMADD f0, f8, f16, f0
  278. FMADD f1, f9, f17, f1
  279. FMADD f2, f10, f18, f2
  280. FMADD f3, f11, f19, f3
  281. LFDUX f8, X, INCX
  282. LFDUX f16, Y, INCY
  283. LFDUX f9, X, INCX
  284. LFDUX f17, Y, INCY
  285. LFDUX f10, X, INCX
  286. LFDUX f18, Y, INCY
  287. LFDUX f11, X, INCX
  288. LFDUX f19, Y, INCY
  289. FMADD f4, f12, f20, f4
  290. FMADD f5, f13, f21, f5
  291. FMADD f6, f14, f22, f6
  292. FMADD f7, f15, f23, f7
  293. LFDUX f12, X, INCX
  294. LFDUX f20, Y, INCY
  295. LFDUX f13, X, INCX
  296. LFDUX f21, Y, INCY
  297. LFDUX f14, X, INCX
  298. LFDUX f22, Y, INCY
  299. LFDUX f15, X, INCX
  300. LFDUX f23, Y, INCY
  301. bdnz LL(110)
  302. .align 4
  303. LL(120):
  304. FMADD f0, f8, f16, f0
  305. FMADD f1, f9, f17, f1
  306. FMADD f2, f10, f18, f2
  307. FMADD f3, f11, f19, f3
  308. LFDUX f8, X, INCX
  309. LFDUX f16, Y, INCY
  310. LFDUX f9, X, INCX
  311. LFDUX f17, Y, INCY
  312. LFDUX f10, X, INCX
  313. LFDUX f18, Y, INCY
  314. LFDUX f11, X, INCX
  315. LFDUX f19, Y, INCY
  316. FMADD f4, f12, f20, f4
  317. FMADD f5, f13, f21, f5
  318. FMADD f6, f14, f22, f6
  319. FMADD f7, f15, f23, f7
  320. LFDUX f12, X, INCX
  321. LFDUX f20, Y, INCY
  322. LFDUX f13, X, INCX
  323. LFDUX f21, Y, INCY
  324. LFDUX f14, X, INCX
  325. LFDUX f22, Y, INCY
  326. LFDUX f15, X, INCX
  327. LFDUX f23, Y, INCY
  328. FMADD f0, f8, f16, f0
  329. FMADD f1, f9, f17, f1
  330. FMADD f2, f10, f18, f2
  331. FMADD f3, f11, f19, f3
  332. FMADD f4, f12, f20, f4
  333. FMADD f5, f13, f21, f5
  334. FMADD f6, f14, f22, f6
  335. FMADD f7, f15, f23, f7
  336. .align 4
  337. LL(150):
  338. andi. r0, N, 15
  339. mtspr CTR, r0
  340. beq LL(999)
  341. .align 4
  342. LL(160):
  343. LFDUX f8, X, INCX
  344. LFDUX f16, Y, INCY
  345. FMADD f0, f8, f16, f0
  346. bdnz LL(160)
  347. .align 4
  348. LL(999):
  349. FADD f0, f0, f1
  350. FADD f2, f2, f3
  351. FADD f4, f4, f5
  352. FADD f6, f6, f7
  353. FADD f0, f0, f2
  354. FADD f4, f4, f6
  355. FADD f1, f0, f4
  356. lfd f14, 0(SP)
  357. lfd f15, 8(SP)
  358. lfd f16, 16(SP)
  359. lfd f17, 24(SP)
  360. lfd f18, 32(SP)
  361. lfd f19, 40(SP)
  362. lfd f20, 48(SP)
  363. lfd f21, 56(SP)
  364. lfd f22, 64(SP)
  365. lfd f23, 72(SP)
  366. addi SP, SP, STACKSIZE
  367. blr
  368. EPILOGUE