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_tcopy_4.S 9.2 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457
  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 r3
  41. #define N r4
  42. #define A r5
  43. #define LDA r6
  44. #define B r7
  45. #define AO1 r8
  46. #define AO2 r9
  47. #define AO3 r10
  48. #define AO4 r11
  49. #define J r12
  50. #define PREA r14
  51. #define PREB1 r15
  52. #define B1 r16
  53. #define B2 r17
  54. #define B3 r18
  55. #define M4 r19
  56. #define c01 f0
  57. #define c02 f1
  58. #define c03 f2
  59. #define c04 f3
  60. #define c05 f4
  61. #define c06 f5
  62. #define c07 f6
  63. #define c08 f7
  64. #define c09 f8
  65. #define c10 f9
  66. #define c11 f10
  67. #define c12 f11
  68. #define c13 f12
  69. #define c14 f13
  70. #define c15 f14
  71. #define c16 f15
  72. #define STACKSIZE 64
  73. #ifdef CELL
  74. #define PREFETCHSIZE 16
  75. #define PREFETCHWSIZE 48
  76. #endif
  77. #ifdef PPC970
  78. #define PREFETCHSIZE 16
  79. #define PREFETCHWSIZE 48
  80. #endif
  81. #ifdef PPC440
  82. #define PREFETCHSIZE 16
  83. #define PREFETCHWSIZE 48
  84. #endif
  85. #ifdef POWER4
  86. #define PREFETCHSIZE 16
  87. #define PREFETCHWSIZE 48
  88. #endif
  89. #ifdef POWER5
  90. #define PREFETCHSIZE 16
  91. #define PREFETCHWSIZE 48
  92. #endif
  93. #ifdef POWER6
  94. #define PREFETCHSIZE 16
  95. #define PREFETCHWSIZE 48
  96. #endif
  97. #ifdef PPCG4
  98. #define PREFETCHSIZE 16
  99. #define PREFETCHWSIZE 48
  100. #endif
  101. #ifdef POWER8
  102. #define PREFETCHSIZE 16
  103. #define PREFETCHWSIZE 48
  104. #endif
  105. PROLOGUE
  106. PROFCODE
  107. addi SP, SP, -STACKSIZE
  108. li r0, 0
  109. stfd f14, 0(SP)
  110. stfd f15, 8(SP)
  111. #ifdef __64BIT__
  112. std r14, 16(SP)
  113. std r15, 24(SP)
  114. std r16, 32(SP)
  115. std r17, 40(SP)
  116. std r18, 48(SP)
  117. std r19, 56(SP)
  118. #else
  119. stw r14, 16(SP)
  120. stw r15, 20(SP)
  121. stw r16, 24(SP)
  122. stw r17, 28(SP)
  123. stw r18, 32(SP)
  124. stw r19, 36(SP)
  125. #endif
  126. slwi LDA, LDA, BASE_SHIFT
  127. slwi M4, M, 2 + BASE_SHIFT
  128. li PREA, -4
  129. li PREB1, -2
  130. and B2, N, PREA
  131. and B3, N, PREB1
  132. mullw B2, B2, M
  133. mullw B3, B3, M
  134. slwi B2, B2, BASE_SHIFT
  135. slwi B3, B3, BASE_SHIFT
  136. add B2, B2, B
  137. add B3, B3, B
  138. li PREA, PREFETCHSIZE * SIZE
  139. li PREB1, (PREFETCHWSIZE + 0) * SIZE
  140. cmpwi cr0, M, 0
  141. ble- LL(999)
  142. cmpwi cr0, N, 0
  143. ble- LL(999)
  144. srawi. J, M, 2
  145. ble LL(20)
  146. .align 4
  147. LL(10):
  148. mr AO1, A
  149. add AO2, A, LDA
  150. add AO3, AO2, LDA
  151. add AO4, AO3, LDA
  152. add A, AO4, LDA
  153. mr B1, B
  154. addi B, B, 16 * SIZE
  155. srawi. r0, N, 2
  156. mtspr CTR, r0
  157. ble LL(13)
  158. .align 4
  159. LL(12):
  160. LFD c01, 0 * SIZE(AO1)
  161. LFD c02, 1 * SIZE(AO1)
  162. LFD c03, 2 * SIZE(AO1)
  163. LFD c04, 3 * SIZE(AO1)
  164. LFD c05, 0 * SIZE(AO2)
  165. LFD c06, 1 * SIZE(AO2)
  166. LFD c07, 2 * SIZE(AO2)
  167. LFD c08, 3 * SIZE(AO2)
  168. LFD c09, 0 * SIZE(AO3)
  169. LFD c10, 1 * SIZE(AO3)
  170. LFD c11, 2 * SIZE(AO3)
  171. LFD c12, 3 * SIZE(AO3)
  172. LFD c13, 0 * SIZE(AO4)
  173. LFD c14, 1 * SIZE(AO4)
  174. LFD c15, 2 * SIZE(AO4)
  175. LFD c16, 3 * SIZE(AO4)
  176. STFD c01, 0 * SIZE(B1)
  177. STFD c02, 1 * SIZE(B1)
  178. STFD c03, 2 * SIZE(B1)
  179. STFD c04, 3 * SIZE(B1)
  180. STFD c05, 4 * SIZE(B1)
  181. STFD c06, 5 * SIZE(B1)
  182. STFD c07, 6 * SIZE(B1)
  183. STFD c08, 7 * SIZE(B1)
  184. STFD c09, 8 * SIZE(B1)
  185. STFD c10, 9 * SIZE(B1)
  186. STFD c11, 10 * SIZE(B1)
  187. STFD c12, 11 * SIZE(B1)
  188. STFD c13, 12 * SIZE(B1)
  189. STFD c14, 13 * SIZE(B1)
  190. STFD c15, 14 * SIZE(B1)
  191. STFD c16, 15 * SIZE(B1)
  192. #if defined(POWER6) || defined(POWER8)
  193. dcbtst PREA, AO1
  194. dcbtst PREA, AO2
  195. dcbtst PREA, AO3
  196. dcbtst PREA, AO4
  197. #else
  198. dcbt PREA, AO1
  199. dcbt PREA, AO2
  200. dcbt PREA, AO3
  201. dcbt PREA, AO4
  202. #endif
  203. dcbtst PREB1, B
  204. addi AO1, AO1, 4 * SIZE
  205. addi AO2, AO2, 4 * SIZE
  206. addi AO3, AO3, 4 * SIZE
  207. addi AO4, AO4, 4 * SIZE
  208. add B1, B1, M4
  209. bdnz LL(12)
  210. .align 4
  211. LL(13):
  212. andi. r0, N, 2
  213. ble LL(14)
  214. LFD c01, 0 * SIZE(AO1)
  215. LFD c02, 1 * SIZE(AO1)
  216. LFD c03, 0 * SIZE(AO2)
  217. LFD c04, 1 * SIZE(AO2)
  218. LFD c05, 0 * SIZE(AO3)
  219. LFD c06, 1 * SIZE(AO3)
  220. LFD c07, 0 * SIZE(AO4)
  221. LFD c08, 1 * SIZE(AO4)
  222. STFD c01, 0 * SIZE(B2)
  223. STFD c02, 1 * SIZE(B2)
  224. STFD c03, 2 * SIZE(B2)
  225. STFD c04, 3 * SIZE(B2)
  226. STFD c05, 4 * SIZE(B2)
  227. STFD c06, 5 * SIZE(B2)
  228. STFD c07, 6 * SIZE(B2)
  229. STFD c08, 7 * SIZE(B2)
  230. addi AO1, AO1, 2 * SIZE
  231. addi AO2, AO2, 2 * SIZE
  232. addi AO3, AO3, 2 * SIZE
  233. addi AO4, AO4, 2 * SIZE
  234. addi B2, B2, 8 * SIZE
  235. .align 4
  236. LL(14):
  237. andi. r0, N, 1
  238. ble LL(17)
  239. LFD c01, 0 * SIZE(AO1)
  240. LFD c02, 0 * SIZE(AO2)
  241. LFD c03, 0 * SIZE(AO3)
  242. LFD c04, 0 * SIZE(AO4)
  243. STFD c01, 0 * SIZE(B3)
  244. STFD c02, 1 * SIZE(B3)
  245. STFD c03, 2 * SIZE(B3)
  246. STFD c04, 3 * SIZE(B3)
  247. addi B3, B3, 4 * SIZE
  248. .align 4
  249. LL(17):
  250. addic. J, J, -1
  251. bgt LL(10)
  252. .align 4
  253. LL(20):
  254. andi. J, M, 2
  255. ble LL(30)
  256. mr AO1, A
  257. add AO2, A, LDA
  258. add A, AO2, LDA
  259. mr B1, B
  260. addi B, B, 8 * SIZE
  261. srawi. r0, N, 2
  262. mtspr CTR, r0
  263. ble LL(23)
  264. .align 4
  265. LL(22):
  266. LFD c01, 0 * SIZE(AO1)
  267. LFD c02, 1 * SIZE(AO1)
  268. LFD c03, 2 * SIZE(AO1)
  269. LFD c04, 3 * SIZE(AO1)
  270. LFD c05, 0 * SIZE(AO2)
  271. LFD c06, 1 * SIZE(AO2)
  272. LFD c07, 2 * SIZE(AO2)
  273. LFD c08, 3 * SIZE(AO2)
  274. STFD c01, 0 * SIZE(B1)
  275. STFD c02, 1 * SIZE(B1)
  276. STFD c03, 2 * SIZE(B1)
  277. STFD c04, 3 * SIZE(B1)
  278. STFD c05, 4 * SIZE(B1)
  279. STFD c06, 5 * SIZE(B1)
  280. STFD c07, 6 * SIZE(B1)
  281. STFD c08, 7 * SIZE(B1)
  282. addi AO1, AO1, 4 * SIZE
  283. addi AO2, AO2, 4 * SIZE
  284. add B1, B1, M4
  285. bdnz LL(22)
  286. .align 4
  287. LL(23):
  288. andi. r0, N, 2
  289. ble LL(24)
  290. LFD c01, 0 * SIZE(AO1)
  291. LFD c02, 1 * SIZE(AO1)
  292. LFD c03, 0 * SIZE(AO2)
  293. LFD c04, 1 * SIZE(AO2)
  294. STFD c01, 0 * SIZE(B2)
  295. STFD c02, 1 * SIZE(B2)
  296. STFD c03, 2 * SIZE(B2)
  297. STFD c04, 3 * SIZE(B2)
  298. addi AO1, AO1, 2 * SIZE
  299. addi AO2, AO2, 2 * SIZE
  300. addi B2, B2, 4 * SIZE
  301. .align 4
  302. LL(24):
  303. andi. r0, N, 1
  304. ble LL(30)
  305. LFD c01, 0 * SIZE(AO1)
  306. LFD c02, 0 * SIZE(AO2)
  307. STFD c01, 0 * SIZE(B3)
  308. STFD c02, 1 * SIZE(B3)
  309. addi B3, B3, 2 * SIZE
  310. .align 4
  311. LL(30):
  312. andi. J, M, 1
  313. ble LL(999)
  314. mr AO1, A
  315. mr B1, B
  316. srawi. r0, N, 2
  317. mtspr CTR, r0
  318. ble LL(33)
  319. .align 4
  320. LL(32):
  321. LFD c01, 0 * SIZE(AO1)
  322. LFD c02, 1 * SIZE(AO1)
  323. LFD c03, 2 * SIZE(AO1)
  324. LFD c04, 3 * SIZE(AO1)
  325. STFD c01, 0 * SIZE(B1)
  326. STFD c02, 1 * SIZE(B1)
  327. STFD c03, 2 * SIZE(B1)
  328. STFD c04, 3 * SIZE(B1)
  329. addi AO1, AO1, 4 * SIZE
  330. add B1, B1, M4
  331. bdnz LL(32)
  332. .align 4
  333. LL(33):
  334. andi. r0, N, 2
  335. ble LL(34)
  336. LFD c01, 0 * SIZE(AO1)
  337. LFD c02, 1 * SIZE(AO1)
  338. STFD c01, 0 * SIZE(B2)
  339. STFD c02, 1 * SIZE(B2)
  340. addi AO1, AO1, 2 * SIZE
  341. addi B2, B2, 2 * SIZE
  342. .align 4
  343. LL(34):
  344. andi. r0, N, 1
  345. ble LL(999)
  346. LFD c01, 0 * SIZE(AO1)
  347. STFD c01, 0 * SIZE(B3)
  348. .align 4
  349. LL(999):
  350. li r3, 0
  351. lfd f14, 0(SP)
  352. lfd f15, 8(SP)
  353. #ifdef __64BIT__
  354. ld r14, 16(SP)
  355. ld r15, 24(SP)
  356. ld r16, 32(SP)
  357. ld r17, 40(SP)
  358. ld r18, 48(SP)
  359. ld r19, 56(SP)
  360. #else
  361. lwz r14, 16(SP)
  362. lwz r15, 20(SP)
  363. lwz r16, 24(SP)
  364. lwz r17, 28(SP)
  365. lwz r18, 32(SP)
  366. lwz r19, 36(SP)
  367. #endif
  368. addi SP, SP, STACKSIZE
  369. blr
  370. EPILOGUE