OSchip
/
OpenBLAS
Not watched
1
0
Fork 0
Code
Releases 66 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
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.
843 Commits
51 Branches
78 MB
368 Download
Tree: f331cb1a76
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'f331cb1a76'
${ noResults }
OpenBLAS/kernel/x86/KERNEL

KERNEL 5.9 kB
Raw Normal View History

Import GotoBLAS2 1.13 BSD version codes.
15 years ago
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
replaced files for sdot, sgemv_n and sgemv_t for bug #348
12 years ago
Import GotoBLAS2 1.13 BSD version codes.
15 years ago
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
replaced files for sdot, sgemv_n and sgemv_t for bug #348
12 years ago
Import GotoBLAS2 1.13 BSD version codes.
15 years ago
​
​
replaced files for sdot, sgemv_n and sgemv_t for bug #348
12 years ago
Import GotoBLAS2 1.13 BSD version codes.
15 years ago
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
refs #80. Used GEMV SSE2 kernels on x86.
14 years ago
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
Import GotoBLAS2 1.13 BSD version codes.
15 years ago
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414 
  1. GEMVDEP = ../l2param.h

  2. 

  3. ifdef HAVE_SSE

  4. 

  5. ifndef SAMAXKERNEL

  6. SAMAXKERNEL = amax_sse.S

  7. endif

  8. 

  9. ifndef CAMAXKERNEL

  10. CAMAXKERNEL = zamax_sse.S

  11. endif

  12. 

  13. ifndef SAMINKERNEL

  14. SAMINKERNEL = amax_sse.S

  15. endif

  16. 

  17. ifndef CAMINKERNEL

  18. CAMINKERNEL = zamax_sse.S

  19. endif

  20. 

  21. ifndef ISAMAXKERNEL

  22. ISAMAXKERNEL = iamax_sse.S

  23. endif

  24. 

  25. ifndef ICAMAXKERNEL

  26. ICAMAXKERNEL = izamax_sse.S

  27. endif

  28. 

  29. ifndef ISAMINKERNEL

  30. ISAMINKERNEL = iamax_sse.S

  31. endif

  32. 

  33. ifndef ICAMINKERNEL

  34. ICAMINKERNEL = izamax_sse.S

  35. endif

  36. 

  37. ifndef ISMAXKERNEL

  38. ISMAXKERNEL = iamax_sse.S

  39. endif

  40. 

  41. ifndef ISMINKERNEL

  42. ISMINKERNEL = iamax_sse.S

  43. endif

  44. 

  45. ifndef SMAXKERNEL

  46. SMAXKERNEL = amax_sse.S

  47. endif

  48. 

  49. ifndef SMINKERNEL

  50. SMINKERNEL = amax_sse.S

  51. endif

  52. 

  53. ifndef SASUMKERNEL

  54. SASUMKERNEL = asum_sse.S

  55. endif

  56. 

  57. ifndef CASUMKERNEL

  58. CASUMKERNEL = zasum_sse.S

  59. endif

  60. 

  61. ifndef SDOTKERNEL

  62. SDOTKERNEL = ../arm/dot.c

  63. endif

  64. 

  65. ifndef CDOTKERNEL

  66. CDOTKERNEL = zdot_sse.S

  67. endif

  68. 

  69. ifndef SCOPYKERNEL

  70. SCOPYKERNEL = copy_sse.S

  71. endif

  72. 

  73. ifndef CCOPYKERNEL

  74. CCOPYKERNEL = zcopy_sse.S

  75. endif

  76. 

  77. ifndef SSACALKERNEL

  78. SSCALKERNEL = scal_sse.S

  79. endif

  80. 

  81. ifndef CSACALKERNEL

  82. CSCALKERNEL = zscal_sse.S

  83. endif

  84. 

  85. ifndef SAXPYKERNEL

  86. SAXPYKERNEL = axpy_sse.S

  87. endif

  88. 

  89. ifndef CAXPYKERNEL

  90. CAXPYKERNEL = zaxpy_sse.S

  91. endif

  92. 

  93. ifndef SROTKERNEL

  94. SROTKERNEL = rot_sse.S

  95. endif

  96. 

  97. ifndef CROTKERNEL

  98. CROTKERNEL = zrot_sse.S

  99. endif

  100. 

  101. ifndef SSWAPKERNEL

  102. SSWAPKERNEL =  swap_sse.S

  103. endif

  104. 

  105. ifndef CSWAPKERNEL

  106. CSWAPKERNEL = zswap_sse.S

  107. endif

  108. 

  109. ifndef SGEMVNKERNEL

  110. SGEMVNKERNEL = ../arm/gemv_n.c

  111. endif

  112. 

  113. ifndef SGEMVTKERNEL

  114. SGEMVTKERNEL = ../arm/gemv_t.c

  115. endif

  116. 

  117. ifndef CGEMVNKERNEL

  118. CGEMVNKERNEL = zgemv_n_sse.S

  119. endif

  120. 

  121. ifndef CGEMVTKERNEL

  122. CGEMVTKERNEL = zgemv_t_sse.S

  123. endif

  124. 

  125. endif

  126. 

  127. 

  128. ifdef HAVE_SSE2

  129. 

  130. ifndef DAMAXKERNEL

  131. DAMAXKERNEL = amax_sse2.S

  132. endif

  133. 

  134. ifndef ZAMAXKERNEL

  135. ZAMAXKERNEL = zamax_sse2.S

  136. endif

  137. 

  138. ifndef DAMINKERNEL

  139. DAMINKERNEL = amax_sse2.S

  140. endif

  141. 

  142. ifndef ZAMINKERNEL

  143. ZAMINKERNEL = zamax_sse2.S

  144. endif

  145. 

  146. ifndef IDAMAXKERNEL

  147. IDAMAXKERNEL = iamax_sse2.S

  148. endif

  149. 

  150. ifndef IZAMAXKERNEL

  151. IZAMAXKERNEL = izamax_sse2.S

  152. endif

  153. 

  154. ifndef IDAMINKERNEL

  155. IDAMINKERNEL = iamax_sse2.S

  156. endif

  157. 

  158. ifndef IZAMINKERNEL

  159. IZAMINKERNEL = izamax_sse2.S

  160. endif

  161. 

  162. ifndef IDMAXKERNEL

  163. IDMAXKERNEL = iamax_sse2.S

  164. endif

  165. 

  166. ifndef IDMINKERNEL

  167. IDMINKERNEL = iamax_sse2.S

  168. endif

  169. 

  170. ifndef DMAXKERNEL

  171. DMAXKERNEL = amax_sse2.S

  172. endif

  173. 

  174. ifndef DMINKERNEL

  175. DMINKERNEL = amax_sse2.S

  176. endif

  177. 

  178. ifndef DDOTKERNEL

  179. DDOTKERNEL = dot_sse2.S

  180. endif

  181. 

  182. ifndef ZDOTKERNEL

  183. ZDOTKERNEL = zdot_sse2.S

  184. endif

  185. 

  186. ifndef DCOPYKERNEL

  187. # DCOPYKERNEL = copy_sse2.S

  188. endif

  189. 

  190. ifndef ZCOPYKERNEL

  191. ZCOPYKERNEL = zcopy_sse2.S

  192. endif

  193. 

  194. ifndef DSACALKERNEL

  195. DSCALKERNEL = scal_sse2.S

  196. endif

  197. 

  198. ifndef ZSACALKERNEL

  199. ZSCALKERNEL = zscal_sse2.S

  200. endif

  201. 

  202. ifndef DASUMKERNEL

  203. DASUMKERNEL = asum_sse2.S

  204. endif

  205. 

  206. ifndef ZASUMKERNEL

  207. ZASUMKERNEL = zasum_sse2.S

  208. endif

  209. 

  210. ifndef DAXPYKERNEL

  211. DAXPYKERNEL = axpy_sse2.S

  212. endif

  213. 

  214. ifndef ZAXPYKERNEL

  215. ZAXPYKERNEL = zaxpy_sse2.S

  216. endif

  217. 

  218. ifndef SNRM2KERNEL

  219. SNRM2KERNEL = nrm2_sse.S

  220. endif

  221. 

  222. ifndef CNRM2KERNEL

  223. CNRM2KERNEL = znrm2_sse.S

  224. endif

  225. 

  226. ifndef DROTKERNEL

  227. DROTKERNEL = rot_sse2.S

  228. endif

  229. 

  230. ifndef ZROTKERNEL

  231. ZROTKERNEL = zrot_sse2.S

  232. endif

  233. 

  234. ifndef DSWAPKERNEL

  235. DSWAPKERNEL =  swap_sse2.S

  236. endif

  237. 

  238. ifndef ZSWAPKERNEL

  239. ZSWAPKERNEL = zswap_sse2.S

  240. endif

  241. 

  242. ifndef DGEMVNKERNEL

  243. DGEMVNKERNEL = gemv_n_sse2.S

  244. endif

  245. 

  246. ifndef DGEMVTKERNEL

  247. DGEMVTKERNEL = gemv_t_sse2.S

  248. endif

  249. 

  250. ifndef ZGEMVNKERNEL

  251. ZGEMVNKERNEL = zgemv_n_sse2.S

  252. endif

  253. 

  254. ifndef ZGEMVTKERNEL

  255. ZGEMVTKERNEL = zgemv_t_sse2.S

  256. endif

  257. 

  258. endif

  259. 

  260. 

  261. ifndef SAMINKERNEL

  262. SAMINKERNEL = amax.S

  263. endif

  264. 

  265. ifndef DAMINKERNEL

  266. DAMINKERNEL = amax.S

  267. endif

  268. 

  269. ifndef QAMINKERNEL

  270. QAMINKERNEL = amax.S

  271. endif

  272. 

  273. ifndef CAMINKERNEL

  274. CAMINKERNEL = zamax.S

  275. endif

  276. 

  277. ifndef ZAMINKERNEL

  278. ZAMINKERNEL = zamax.S

  279. endif

  280. 

  281. ifndef XAMINKERNEL

  282. XAMINKERNEL = zamax.S

  283. endif

  284. 

  285. ifndef SMAXKERNEL

  286. SMAXKERNEL = amax.S

  287. endif

  288. 

  289. ifndef DMAXKERNEL

  290. DMAXKERNEL = amax.S

  291. endif

  292. 

  293. ifndef QMAXKERNEL

  294. QMAXKERNEL = amax.S

  295. endif

  296. 

  297. ifndef SMINKERNEL

  298. SMINKERNEL = amax.S

  299. endif

  300. 

  301. ifndef DMINKERNEL

  302. DMINKERNEL = amax.S

  303. endif

  304. 

  305. ifndef QMINKERNEL

  306. QMINKERNEL = amax.S

  307. endif

  308. 

  309. ifndef ISAMINKERNEL

  310. ISAMINKERNEL = iamax.S

  311. endif

  312. 

  313. ifndef IDAMINKERNEL

  314. IDAMINKERNEL = iamax.S

  315. endif

  316. 

  317. ifndef IQAMINKERNEL

  318. IQAMINKERNEL = iamax.S

  319. endif

  320. 

  321. ifndef ICAMINKERNEL

  322. ICAMINKERNEL = izamax.S

  323. endif

  324. 

  325. ifndef IZAMINKERNEL

  326. IZAMINKERNEL = izamax.S

  327. endif

  328. 

  329. ifndef IXAMINKERNEL

  330. IXAMINKERNEL = izamax.S

  331. endif

  332. 

  333. ifndef ISMINKERNEL

  334. ISMINKERNEL = iamax.S

  335. endif

  336. 

  337. ifndef IDMINKERNEL

  338. IDMINKERNEL = iamax.S

  339. endif

  340. 

  341. ifndef IQMINKERNEL

  342. IQMINKERNEL = iamax.S

  343. endif

  344. 

  345. ifndef QDOTKERNEL

  346. QDOTKERNEL = qdot.S

  347. endif

  348. 

  349. ifndef XDOTKERNEL

  350. XDOTKERNEL = xdot.S

  351. endif

  352. 

  353. ifndef QAXPYKERNEL

  354. QAXPYKERNEL = qaxpy.S

  355. endif

  356. 

  357. ifndef XAXPYKERNEL

  358. XAXPYKERNEL = xaxpy.S

  359. endif

  360. 

  361. ifndef QGEMVNKERNEL

  362. QGEMVNKERNEL = qgemv_n.S

  363. endif

  364. 

  365. ifndef QGEMVTKERNEL

  366. QGEMVTKERNEL = qgemv_t.S

  367. endif

  368. 

  369. ifndef XGEMVNKERNEL

  370. XGEMVNKERNEL = xgemv_n.S

  371. endif

  372. 

  373. ifndef XGEMVTKERNEL

  374. XGEMVTKERNEL = xgemv_t.S

  375. endif

  376. 

  377. QGEMMKERNEL    =  qgemm_kernel_2x2.S

  378. QGEMMINCOPY    =

  379. QGEMMITCOPY    =

  380. QGEMMONCOPY    =  ../generic/gemm_ncopy_2.c

  381. QGEMMOTCOPY    =  ../generic/gemm_tcopy_2.c

  382. QGEMMINCOPYOBJ =

  383. QGEMMITCOPYOBJ =

  384. QGEMMONCOPYOBJ =  qgemm_oncopy$(TSUFFIX).$(SUFFIX)

  385. QGEMMOTCOPYOBJ =  qgemm_otcopy$(TSUFFIX).$(SUFFIX)

  386. 

  387. XGEMMKERNEL    =  xgemm_kernel_1x1.S

  388. XGEMMINCOPY    =

  389. XGEMMITCOPY    =

  390. XGEMMONCOPY    =  ../generic/zgemm_ncopy_1.c

  391. XGEMMOTCOPY    =  ../generic/zgemm_tcopy_1.c

  392. XGEMMINCOPYOBJ =

  393. XGEMMITCOPYOBJ =

  394. XGEMMONCOPYOBJ =  xgemm_oncopy$(TSUFFIX).$(SUFFIX)

  395. XGEMMOTCOPYOBJ =  xgemm_otcopy$(TSUFFIX).$(SUFFIX)

  396. 

  397. SGEMM_BETA = gemm_beta.S

  398. DGEMM_BETA = gemm_beta.S

  399. QGEMM_BETA = ../generic/gemm_beta.c

  400. CGEMM_BETA = zgemm_beta.S

  401. ZGEMM_BETA = zgemm_beta.S

  402. XGEMM_BETA = ../generic/zgemm_beta.c

  403. 

  404. QTRSMKERNEL_LN	=  qtrsm_kernel_LN_2x2.S

  405. QTRSMKERNEL_LT	=  qtrsm_kernel_LT_2x2.S

  406. QTRSMKERNEL_RN	=  qtrsm_kernel_LT_2x2.S

  407. QTRSMKERNEL_RT	=  qtrsm_kernel_RT_2x2.S

  408. 

  409. XTRSMKERNEL_LN	=  xtrsm_kernel_LT_1x1.S

  410. XTRSMKERNEL_LT	=  xtrsm_kernel_LT_1x1.S

  411. XTRSMKERNEL_RN	=  xtrsm_kernel_LT_1x1.S

  412. XTRSMKERNEL_RT	=  xtrsm_kernel_LT_1x1.S

  413. 

  414. XGEMM3MKERNEL    =  xgemm3m_kernel_2x2.S

OpenBLAS is an optimized BLAS library based on GotoBLAS2 1.13 BSD version.