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.

param.h 41 kB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732
  1. /*****************************************************************************
  2. Copyright (c) 2011, Lab of Parallel Software and Computational Science,ICSAS
  3. All rights reserved.
  4. Redistribution and use in source and binary forms, with or without
  5. modification, are permitted provided that the following conditions are
  6. met:
  7. 1. Redistributions of source code must retain the above copyright
  8. notice, this list of conditions and the following disclaimer.
  9. 2. Redistributions in binary form must reproduce the above copyright
  10. notice, this list of conditions and the following disclaimer in
  11. the documentation and/or other materials provided with the
  12. distribution.
  13. 3. Neither the name of the ISCAS nor the names of its contributors may
  14. be used to endorse or promote products derived from this software
  15. without specific prior written permission.
  16. THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
  17. AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  18. IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  19. ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
  20. LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  21. DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  22. SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
  23. CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  24. OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
  25. USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  26. **********************************************************************************/
  27. /*********************************************************************/
  28. /* Copyright 2009, 2010 The University of Texas at Austin. */
  29. /* All rights reserved. */
  30. /* */
  31. /* Redistribution and use in source and binary forms, with or */
  32. /* without modification, are permitted provided that the following */
  33. /* conditions are met: */
  34. /* */
  35. /* 1. Redistributions of source code must retain the above */
  36. /* copyright notice, this list of conditions and the following */
  37. /* disclaimer. */
  38. /* */
  39. /* 2. Redistributions in binary form must reproduce the above */
  40. /* copyright notice, this list of conditions and the following */
  41. /* disclaimer in the documentation and/or other materials */
  42. /* provided with the distribution. */
  43. /* */
  44. /* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
  45. /* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
  46. /* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
  47. /* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
  48. /* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
  49. /* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
  50. /* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
  51. /* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
  52. /* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
  53. /* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
  54. /* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
  55. /* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
  56. /* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
  57. /* POSSIBILITY OF SUCH DAMAGE. */
  58. /* */
  59. /* The views and conclusions contained in the software and */
  60. /* documentation are those of the authors and should not be */
  61. /* interpreted as representing official policies, either expressed */
  62. /* or implied, of The University of Texas at Austin. */
  63. /*********************************************************************/
  64. #ifndef PARAM_H
  65. #define PARAM_H
  66. #ifdef OPTERON
  67. #define SNUMOPT 4
  68. #define DNUMOPT 2
  69. #define GEMM_DEFAULT_OFFSET_A 64
  70. #define GEMM_DEFAULT_OFFSET_B 256
  71. #define GEMM_DEFAULT_ALIGN 0x01ffffUL
  72. #define SGEMM_DEFAULT_UNROLL_N 4
  73. #define DGEMM_DEFAULT_UNROLL_N 4
  74. #define QGEMM_DEFAULT_UNROLL_N 2
  75. #define CGEMM_DEFAULT_UNROLL_N 2
  76. #define ZGEMM_DEFAULT_UNROLL_N 2
  77. #define XGEMM_DEFAULT_UNROLL_N 1
  78. #ifdef ARCH_X86
  79. #define SGEMM_DEFAULT_UNROLL_M 4
  80. #define DGEMM_DEFAULT_UNROLL_M 2
  81. #define QGEMM_DEFAULT_UNROLL_M 2
  82. #define CGEMM_DEFAULT_UNROLL_M 2
  83. #define ZGEMM_DEFAULT_UNROLL_M 1
  84. #define XGEMM_DEFAULT_UNROLL_M 1
  85. #else
  86. #define SGEMM_DEFAULT_UNROLL_M 8
  87. #define DGEMM_DEFAULT_UNROLL_M 4
  88. #define QGEMM_DEFAULT_UNROLL_M 2
  89. #define CGEMM_DEFAULT_UNROLL_M 4
  90. #define ZGEMM_DEFAULT_UNROLL_M 2
  91. #define XGEMM_DEFAULT_UNROLL_M 1
  92. #endif
  93. #define SGEMM_DEFAULT_P sgemm_p
  94. #define DGEMM_DEFAULT_P dgemm_p
  95. #define QGEMM_DEFAULT_P qgemm_p
  96. #define CGEMM_DEFAULT_P cgemm_p
  97. #define ZGEMM_DEFAULT_P zgemm_p
  98. #define XGEMM_DEFAULT_P xgemm_p
  99. #define SGEMM_DEFAULT_R sgemm_r
  100. #define DGEMM_DEFAULT_R dgemm_r
  101. #define QGEMM_DEFAULT_R qgemm_r
  102. #define CGEMM_DEFAULT_R cgemm_r
  103. #define ZGEMM_DEFAULT_R zgemm_r
  104. #define XGEMM_DEFAULT_R xgemm_r
  105. #ifdef ALLOC_HUGETLB
  106. #define SGEMM_DEFAULT_Q 248
  107. #define DGEMM_DEFAULT_Q 248
  108. #define QGEMM_DEFAULT_Q 248
  109. #define CGEMM_DEFAULT_Q 248
  110. #define ZGEMM_DEFAULT_Q 248
  111. #define XGEMM_DEFAULT_Q 248
  112. #else
  113. #define SGEMM_DEFAULT_Q 240
  114. #define DGEMM_DEFAULT_Q 240
  115. #define QGEMM_DEFAULT_Q 240
  116. #define CGEMM_DEFAULT_Q 240
  117. #define ZGEMM_DEFAULT_Q 240
  118. #define XGEMM_DEFAULT_Q 240
  119. #endif
  120. #define SYMV_P 16
  121. #define HAVE_EXCLUSIVE_CACHE
  122. #endif
  123. #if defined(BARCELONA) || defined(SHANGHAI)
  124. #define SNUMOPT 8
  125. #define DNUMOPT 4
  126. #define GEMM_DEFAULT_OFFSET_A 64
  127. #define GEMM_DEFAULT_OFFSET_B 832
  128. #define GEMM_DEFAULT_ALIGN 0x0fffUL
  129. #define SGEMM_DEFAULT_UNROLL_N 4
  130. #define DGEMM_DEFAULT_UNROLL_N 4
  131. #define QGEMM_DEFAULT_UNROLL_N 2
  132. #define CGEMM_DEFAULT_UNROLL_N 2
  133. #define ZGEMM_DEFAULT_UNROLL_N 2
  134. #define XGEMM_DEFAULT_UNROLL_N 1
  135. #ifdef ARCH_X86
  136. #define SGEMM_DEFAULT_UNROLL_M 4
  137. #define DGEMM_DEFAULT_UNROLL_M 2
  138. #define QGEMM_DEFAULT_UNROLL_M 2
  139. #define CGEMM_DEFAULT_UNROLL_M 2
  140. #define ZGEMM_DEFAULT_UNROLL_M 1
  141. #define XGEMM_DEFAULT_UNROLL_M 1
  142. #else
  143. #define SGEMM_DEFAULT_UNROLL_M 8
  144. #define DGEMM_DEFAULT_UNROLL_M 4
  145. #define QGEMM_DEFAULT_UNROLL_M 2
  146. #define CGEMM_DEFAULT_UNROLL_M 4
  147. #define ZGEMM_DEFAULT_UNROLL_M 2
  148. #define XGEMM_DEFAULT_UNROLL_M 1
  149. #endif
  150. #if 0
  151. #define SGEMM_DEFAULT_P 496
  152. #define DGEMM_DEFAULT_P 248
  153. #define QGEMM_DEFAULT_P 124
  154. #define CGEMM_DEFAULT_P 248
  155. #define ZGEMM_DEFAULT_P 124
  156. #define XGEMM_DEFAULT_P 62
  157. #define SGEMM_DEFAULT_Q 248
  158. #define DGEMM_DEFAULT_Q 248
  159. #define QGEMM_DEFAULT_Q 248
  160. #define CGEMM_DEFAULT_Q 248
  161. #define ZGEMM_DEFAULT_Q 248
  162. #define XGEMM_DEFAULT_Q 248
  163. #else
  164. #define SGEMM_DEFAULT_P 448
  165. #define DGEMM_DEFAULT_P 224
  166. #define QGEMM_DEFAULT_P 112
  167. #define CGEMM_DEFAULT_P 224
  168. #define ZGEMM_DEFAULT_P 112
  169. #define XGEMM_DEFAULT_P 56
  170. #define SGEMM_DEFAULT_Q 224
  171. #define DGEMM_DEFAULT_Q 224
  172. #define QGEMM_DEFAULT_Q 224
  173. #define CGEMM_DEFAULT_Q 224
  174. #define ZGEMM_DEFAULT_Q 224
  175. #define XGEMM_DEFAULT_Q 224
  176. #endif
  177. #define SGEMM_DEFAULT_R sgemm_r
  178. #define QGEMM_DEFAULT_R qgemm_r
  179. #define DGEMM_DEFAULT_R dgemm_r
  180. #define CGEMM_DEFAULT_R cgemm_r
  181. #define ZGEMM_DEFAULT_R zgemm_r
  182. #define XGEMM_DEFAULT_R xgemm_r
  183. #define SYMV_P 16
  184. #define HAVE_EXCLUSIVE_CACHE
  185. #define GEMM_THREAD gemm_thread_mn
  186. #endif
  187. #ifdef ATHLON
  188. #define SNUMOPT 4
  189. #define DNUMOPT 2
  190. #define GEMM_DEFAULT_OFFSET_A 0
  191. #define GEMM_DEFAULT_OFFSET_B 384
  192. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  193. #define SGEMM_DEFAULT_UNROLL_N 4
  194. #define DGEMM_DEFAULT_UNROLL_N 4
  195. #define QGEMM_DEFAULT_UNROLL_N 2
  196. #define CGEMM_DEFAULT_UNROLL_N 2
  197. #define ZGEMM_DEFAULT_UNROLL_N 2
  198. #define XGEMM_DEFAULT_UNROLL_N 1
  199. #define SGEMM_DEFAULT_UNROLL_M 2
  200. #define DGEMM_DEFAULT_UNROLL_M 1
  201. #define QGEMM_DEFAULT_UNROLL_M 2
  202. #define CGEMM_DEFAULT_UNROLL_M 1
  203. #define ZGEMM_DEFAULT_UNROLL_M 1
  204. #define XGEMM_DEFAULT_UNROLL_M 1
  205. #define SGEMM_DEFAULT_R sgemm_r
  206. #define DGEMM_DEFAULT_R dgemm_r
  207. #define QGEMM_DEFAULT_R qgemm_r
  208. #define CGEMM_DEFAULT_R cgemm_r
  209. #define ZGEMM_DEFAULT_R zgemm_r
  210. #define XGEMM_DEFAULT_R xgemm_r
  211. #define SGEMM_DEFAULT_P 208
  212. #define DGEMM_DEFAULT_P 104
  213. #define QGEMM_DEFAULT_P 56
  214. #define CGEMM_DEFAULT_P 104
  215. #define ZGEMM_DEFAULT_P 56
  216. #define XGEMM_DEFAULT_P 28
  217. #define SGEMM_DEFAULT_Q 208
  218. #define DGEMM_DEFAULT_Q 208
  219. #define QGEMM_DEFAULT_Q 208
  220. #define CGEMM_DEFAULT_Q 208
  221. #define ZGEMM_DEFAULT_Q 208
  222. #define XGEMM_DEFAULT_Q 208
  223. #define SYMV_P 16
  224. #define HAVE_EXCLUSIVE_CACHE
  225. #endif
  226. #ifdef VIAC3
  227. #define SNUMOPT 2
  228. #define DNUMOPT 1
  229. #define GEMM_DEFAULT_OFFSET_A 0
  230. #define GEMM_DEFAULT_OFFSET_B 256
  231. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  232. #define SGEMM_DEFAULT_UNROLL_N 4
  233. #define DGEMM_DEFAULT_UNROLL_N 4
  234. #define QGEMM_DEFAULT_UNROLL_N 2
  235. #define CGEMM_DEFAULT_UNROLL_N 2
  236. #define ZGEMM_DEFAULT_UNROLL_N 2
  237. #define XGEMM_DEFAULT_UNROLL_N 1
  238. #define SGEMM_DEFAULT_UNROLL_M 2
  239. #define DGEMM_DEFAULT_UNROLL_M 1
  240. #define QGEMM_DEFAULT_UNROLL_M 2
  241. #define CGEMM_DEFAULT_UNROLL_M 1
  242. #define ZGEMM_DEFAULT_UNROLL_M 1
  243. #define XGEMM_DEFAULT_UNROLL_M 1
  244. #define SGEMM_DEFAULT_R sgemm_r
  245. #define DGEMM_DEFAULT_R dgemm_r
  246. #define QGEMM_DEFAULT_R qgemm_r
  247. #define CGEMM_DEFAULT_R cgemm_r
  248. #define ZGEMM_DEFAULT_R zgemm_r
  249. #define XGEMM_DEFAULT_R xgemm_r
  250. #define SGEMM_DEFAULT_P 128
  251. #define DGEMM_DEFAULT_P 128
  252. #define QGEMM_DEFAULT_P 128
  253. #define CGEMM_DEFAULT_P 128
  254. #define ZGEMM_DEFAULT_P 128
  255. #define XGEMM_DEFAULT_P 128
  256. #define SGEMM_DEFAULT_Q 512
  257. #define DGEMM_DEFAULT_Q 256
  258. #define QGEMM_DEFAULT_Q 256
  259. #define CGEMM_DEFAULT_Q 256
  260. #define ZGEMM_DEFAULT_Q 128
  261. #define XGEMM_DEFAULT_Q 128
  262. #define SYMV_P 16
  263. #endif
  264. #ifdef NANO
  265. #define SNUMOPT 4
  266. #define DNUMOPT 2
  267. #define GEMM_DEFAULT_OFFSET_A 64
  268. #define GEMM_DEFAULT_OFFSET_B 256
  269. #define GEMM_DEFAULT_ALIGN 0x01ffffUL
  270. #ifdef ARCH_X86
  271. #define SGEMM_DEFAULT_UNROLL_N 4
  272. #define DGEMM_DEFAULT_UNROLL_N 4
  273. #define QGEMM_DEFAULT_UNROLL_N 2
  274. #define CGEMM_DEFAULT_UNROLL_N 2
  275. #define ZGEMM_DEFAULT_UNROLL_N 2
  276. #define XGEMM_DEFAULT_UNROLL_N 1
  277. #define SGEMM_DEFAULT_UNROLL_M 4
  278. #define DGEMM_DEFAULT_UNROLL_M 2
  279. #define QGEMM_DEFAULT_UNROLL_M 2
  280. #define CGEMM_DEFAULT_UNROLL_M 2
  281. #define ZGEMM_DEFAULT_UNROLL_M 1
  282. #define XGEMM_DEFAULT_UNROLL_M 1
  283. #else
  284. #define SGEMM_DEFAULT_UNROLL_N 8
  285. #define DGEMM_DEFAULT_UNROLL_N 4
  286. #define QGEMM_DEFAULT_UNROLL_N 2
  287. #define CGEMM_DEFAULT_UNROLL_N 4
  288. #define ZGEMM_DEFAULT_UNROLL_N 2
  289. #define XGEMM_DEFAULT_UNROLL_N 1
  290. #define SGEMM_DEFAULT_UNROLL_M 4
  291. #define DGEMM_DEFAULT_UNROLL_M 4
  292. #define QGEMM_DEFAULT_UNROLL_M 2
  293. #define CGEMM_DEFAULT_UNROLL_M 2
  294. #define ZGEMM_DEFAULT_UNROLL_M 2
  295. #define XGEMM_DEFAULT_UNROLL_M 1
  296. #endif
  297. #define SGEMM_DEFAULT_P 288
  298. #define DGEMM_DEFAULT_P 288
  299. #define QGEMM_DEFAULT_P 288
  300. #define CGEMM_DEFAULT_P 288
  301. #define ZGEMM_DEFAULT_P 288
  302. #define XGEMM_DEFAULT_P 288
  303. #define SGEMM_DEFAULT_R sgemm_r
  304. #define DGEMM_DEFAULT_R dgemm_r
  305. #define QGEMM_DEFAULT_R qgemm_r
  306. #define CGEMM_DEFAULT_R cgemm_r
  307. #define ZGEMM_DEFAULT_R zgemm_r
  308. #define XGEMM_DEFAULT_R xgemm_r
  309. #define SGEMM_DEFAULT_Q 256
  310. #define DGEMM_DEFAULT_Q 128
  311. #define QGEMM_DEFAULT_Q 64
  312. #define CGEMM_DEFAULT_Q 128
  313. #define ZGEMM_DEFAULT_Q 64
  314. #define XGEMM_DEFAULT_Q 32
  315. #define SYMV_P 16
  316. #define HAVE_EXCLUSIVE_CACHE
  317. #endif
  318. #if defined(PENTIUM) || defined(PENTIUM2) || defined(PENTIUM3)
  319. #ifdef HAVE_SSE
  320. #define SNUMOPT 2
  321. #else
  322. #define SNUMOPT 1
  323. #endif
  324. #define DNUMOPT 1
  325. #define GEMM_DEFAULT_OFFSET_A 0
  326. #define GEMM_DEFAULT_OFFSET_B 0
  327. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  328. #ifdef HAVE_SSE
  329. #define SGEMM_DEFAULT_UNROLL_M 8
  330. #define CGEMM_DEFAULT_UNROLL_M 4
  331. #else
  332. #define SGEMM_DEFAULT_UNROLL_M 4
  333. #define CGEMM_DEFAULT_UNROLL_M 2
  334. #endif
  335. #define DGEMM_DEFAULT_UNROLL_M 2
  336. #define SGEMM_DEFAULT_UNROLL_N 2
  337. #define DGEMM_DEFAULT_UNROLL_N 2
  338. #define QGEMM_DEFAULT_UNROLL_M 2
  339. #define QGEMM_DEFAULT_UNROLL_N 2
  340. #define CGEMM_DEFAULT_UNROLL_N 1
  341. #define ZGEMM_DEFAULT_UNROLL_M 1
  342. #define ZGEMM_DEFAULT_UNROLL_N 1
  343. #define XGEMM_DEFAULT_UNROLL_M 1
  344. #define XGEMM_DEFAULT_UNROLL_N 1
  345. #define SGEMM_DEFAULT_P sgemm_p
  346. #define SGEMM_DEFAULT_Q 256
  347. #define SGEMM_DEFAULT_R sgemm_r
  348. #define DGEMM_DEFAULT_P dgemm_p
  349. #define DGEMM_DEFAULT_Q 256
  350. #define DGEMM_DEFAULT_R dgemm_r
  351. #define QGEMM_DEFAULT_P qgemm_p
  352. #define QGEMM_DEFAULT_Q 256
  353. #define QGEMM_DEFAULT_R qgemm_r
  354. #define CGEMM_DEFAULT_P cgemm_p
  355. #define CGEMM_DEFAULT_Q 256
  356. #define CGEMM_DEFAULT_R cgemm_r
  357. #define ZGEMM_DEFAULT_P zgemm_p
  358. #define ZGEMM_DEFAULT_Q 256
  359. #define ZGEMM_DEFAULT_R zgemm_r
  360. #define XGEMM_DEFAULT_P xgemm_p
  361. #define XGEMM_DEFAULT_Q 256
  362. #define XGEMM_DEFAULT_R xgemm_r
  363. #define SYMV_P 4
  364. #endif
  365. #ifdef PENTIUMM
  366. #define SNUMOPT 2
  367. #define DNUMOPT 1
  368. #define GEMM_DEFAULT_OFFSET_A 0
  369. #define GEMM_DEFAULT_OFFSET_B 0
  370. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  371. #ifdef CORE_YONAH
  372. #define SGEMM_DEFAULT_UNROLL_M 4
  373. #define SGEMM_DEFAULT_UNROLL_N 4
  374. #define DGEMM_DEFAULT_UNROLL_M 2
  375. #define DGEMM_DEFAULT_UNROLL_N 4
  376. #define QGEMM_DEFAULT_UNROLL_M 2
  377. #define QGEMM_DEFAULT_UNROLL_N 2
  378. #define CGEMM_DEFAULT_UNROLL_M 2
  379. #define CGEMM_DEFAULT_UNROLL_N 2
  380. #define ZGEMM_DEFAULT_UNROLL_M 1
  381. #define ZGEMM_DEFAULT_UNROLL_N 2
  382. #define XGEMM_DEFAULT_UNROLL_M 1
  383. #define XGEMM_DEFAULT_UNROLL_N 1
  384. #else
  385. #define SGEMM_DEFAULT_UNROLL_M 8
  386. #define SGEMM_DEFAULT_UNROLL_N 2
  387. #define DGEMM_DEFAULT_UNROLL_M 2
  388. #define DGEMM_DEFAULT_UNROLL_N 2
  389. #define QGEMM_DEFAULT_UNROLL_M 2
  390. #define QGEMM_DEFAULT_UNROLL_N 2
  391. #define CGEMM_DEFAULT_UNROLL_M 4
  392. #define CGEMM_DEFAULT_UNROLL_N 1
  393. #define ZGEMM_DEFAULT_UNROLL_M 1
  394. #define ZGEMM_DEFAULT_UNROLL_N 1
  395. #define XGEMM_DEFAULT_UNROLL_M 1
  396. #define XGEMM_DEFAULT_UNROLL_N 1
  397. #endif
  398. #define SGEMM_DEFAULT_P sgemm_p
  399. #define SGEMM_DEFAULT_Q 256
  400. #define SGEMM_DEFAULT_R sgemm_r
  401. #define DGEMM_DEFAULT_P dgemm_p
  402. #define DGEMM_DEFAULT_Q 256
  403. #define DGEMM_DEFAULT_R dgemm_r
  404. #define QGEMM_DEFAULT_P qgemm_p
  405. #define QGEMM_DEFAULT_Q 256
  406. #define QGEMM_DEFAULT_R qgemm_r
  407. #define CGEMM_DEFAULT_P cgemm_p
  408. #define CGEMM_DEFAULT_Q 256
  409. #define CGEMM_DEFAULT_R cgemm_r
  410. #define ZGEMM_DEFAULT_P zgemm_p
  411. #define ZGEMM_DEFAULT_Q 256
  412. #define ZGEMM_DEFAULT_R zgemm_r
  413. #define XGEMM_DEFAULT_P xgemm_p
  414. #define XGEMM_DEFAULT_Q 256
  415. #define XGEMM_DEFAULT_R xgemm_r
  416. #define SYMV_P 4
  417. #endif
  418. #ifdef CORE_NORTHWOOD
  419. #define SNUMOPT 4
  420. #define DNUMOPT 2
  421. #define GEMM_DEFAULT_OFFSET_A 0
  422. #define GEMM_DEFAULT_OFFSET_B 32
  423. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  424. #define SYMV_P 8
  425. #define SGEMM_DEFAULT_UNROLL_M 8
  426. #define DGEMM_DEFAULT_UNROLL_M 4
  427. #define QGEMM_DEFAULT_UNROLL_M 2
  428. #define CGEMM_DEFAULT_UNROLL_M 4
  429. #define ZGEMM_DEFAULT_UNROLL_M 2
  430. #define XGEMM_DEFAULT_UNROLL_M 1
  431. #define SGEMM_DEFAULT_UNROLL_N 2
  432. #define DGEMM_DEFAULT_UNROLL_N 2
  433. #define QGEMM_DEFAULT_UNROLL_N 2
  434. #define CGEMM_DEFAULT_UNROLL_N 1
  435. #define ZGEMM_DEFAULT_UNROLL_N 1
  436. #define XGEMM_DEFAULT_UNROLL_N 1
  437. #define SGEMM_DEFAULT_P sgemm_p
  438. #define SGEMM_DEFAULT_R sgemm_r
  439. #define DGEMM_DEFAULT_P dgemm_p
  440. #define DGEMM_DEFAULT_R dgemm_r
  441. #define QGEMM_DEFAULT_P qgemm_p
  442. #define QGEMM_DEFAULT_R qgemm_r
  443. #define CGEMM_DEFAULT_P cgemm_p
  444. #define CGEMM_DEFAULT_R cgemm_r
  445. #define ZGEMM_DEFAULT_P zgemm_p
  446. #define ZGEMM_DEFAULT_R zgemm_r
  447. #define XGEMM_DEFAULT_P xgemm_p
  448. #define XGEMM_DEFAULT_R xgemm_r
  449. #define SGEMM_DEFAULT_Q 128
  450. #define DGEMM_DEFAULT_Q 128
  451. #define QGEMM_DEFAULT_Q 128
  452. #define CGEMM_DEFAULT_Q 128
  453. #define ZGEMM_DEFAULT_Q 128
  454. #define XGEMM_DEFAULT_Q 128
  455. #endif
  456. #ifdef CORE_PRESCOTT
  457. #define SNUMOPT 4
  458. #define DNUMOPT 2
  459. #ifndef __64BIT__
  460. #define GEMM_DEFAULT_OFFSET_A 128
  461. #define GEMM_DEFAULT_OFFSET_B 192
  462. #else
  463. #define GEMM_DEFAULT_OFFSET_A 0
  464. #define GEMM_DEFAULT_OFFSET_B 256
  465. #endif
  466. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  467. #define SYMV_P 8
  468. #ifdef ARCH_X86
  469. #define SGEMM_DEFAULT_UNROLL_M 4
  470. #define DGEMM_DEFAULT_UNROLL_M 2
  471. #define QGEMM_DEFAULT_UNROLL_M 2
  472. #define CGEMM_DEFAULT_UNROLL_M 2
  473. #define ZGEMM_DEFAULT_UNROLL_M 1
  474. #define XGEMM_DEFAULT_UNROLL_M 1
  475. #else
  476. #define SGEMM_DEFAULT_UNROLL_M 8
  477. #define DGEMM_DEFAULT_UNROLL_M 4
  478. #define QGEMM_DEFAULT_UNROLL_M 2
  479. #define CGEMM_DEFAULT_UNROLL_M 4
  480. #define ZGEMM_DEFAULT_UNROLL_M 2
  481. #define XGEMM_DEFAULT_UNROLL_M 1
  482. #endif
  483. #define SGEMM_DEFAULT_UNROLL_N 4
  484. #define DGEMM_DEFAULT_UNROLL_N 4
  485. #define QGEMM_DEFAULT_UNROLL_N 2
  486. #define CGEMM_DEFAULT_UNROLL_N 2
  487. #define ZGEMM_DEFAULT_UNROLL_N 2
  488. #define XGEMM_DEFAULT_UNROLL_N 1
  489. #define SGEMM_DEFAULT_P sgemm_p
  490. #define SGEMM_DEFAULT_R sgemm_r
  491. #define DGEMM_DEFAULT_P dgemm_p
  492. #define DGEMM_DEFAULT_R dgemm_r
  493. #define QGEMM_DEFAULT_P qgemm_p
  494. #define QGEMM_DEFAULT_R qgemm_r
  495. #define CGEMM_DEFAULT_P cgemm_p
  496. #define CGEMM_DEFAULT_R cgemm_r
  497. #define ZGEMM_DEFAULT_P zgemm_p
  498. #define ZGEMM_DEFAULT_R zgemm_r
  499. #define XGEMM_DEFAULT_P xgemm_p
  500. #define XGEMM_DEFAULT_R xgemm_r
  501. #define SGEMM_DEFAULT_Q 128
  502. #define DGEMM_DEFAULT_Q 128
  503. #define QGEMM_DEFAULT_Q 128
  504. #define CGEMM_DEFAULT_Q 128
  505. #define ZGEMM_DEFAULT_Q 128
  506. #define XGEMM_DEFAULT_Q 128
  507. #endif
  508. #ifdef CORE2
  509. #define SNUMOPT 8
  510. #define DNUMOPT 4
  511. #define GEMM_DEFAULT_OFFSET_A 448
  512. #define GEMM_DEFAULT_OFFSET_B 128
  513. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  514. #define SYMV_P 8
  515. #define SWITCH_RATIO 4
  516. #ifdef ARCH_X86
  517. #define SGEMM_DEFAULT_UNROLL_M 8
  518. #define DGEMM_DEFAULT_UNROLL_M 4
  519. #define QGEMM_DEFAULT_UNROLL_M 2
  520. #define CGEMM_DEFAULT_UNROLL_M 4
  521. #define ZGEMM_DEFAULT_UNROLL_M 2
  522. #define XGEMM_DEFAULT_UNROLL_M 1
  523. #define SGEMM_DEFAULT_UNROLL_N 2
  524. #define DGEMM_DEFAULT_UNROLL_N 2
  525. #define QGEMM_DEFAULT_UNROLL_N 2
  526. #define CGEMM_DEFAULT_UNROLL_N 1
  527. #define ZGEMM_DEFAULT_UNROLL_N 1
  528. #define XGEMM_DEFAULT_UNROLL_N 1
  529. #define MASK(a, b) ((((a) + (b) - 1) / (b)) * (b))
  530. #else
  531. #define SGEMM_DEFAULT_UNROLL_M 8
  532. #define DGEMM_DEFAULT_UNROLL_M 4
  533. #define QGEMM_DEFAULT_UNROLL_M 2
  534. #define CGEMM_DEFAULT_UNROLL_M 4
  535. #define ZGEMM_DEFAULT_UNROLL_M 2
  536. #define XGEMM_DEFAULT_UNROLL_M 1
  537. #define SGEMM_DEFAULT_UNROLL_N 4
  538. #define DGEMM_DEFAULT_UNROLL_N 4
  539. #define QGEMM_DEFAULT_UNROLL_N 2
  540. #define CGEMM_DEFAULT_UNROLL_N 2
  541. #define ZGEMM_DEFAULT_UNROLL_N 2
  542. #define XGEMM_DEFAULT_UNROLL_N 1
  543. #endif
  544. #define SGEMM_DEFAULT_P sgemm_p
  545. #define SGEMM_DEFAULT_R sgemm_r
  546. #define DGEMM_DEFAULT_P dgemm_p
  547. #define DGEMM_DEFAULT_R dgemm_r
  548. #define QGEMM_DEFAULT_P qgemm_p
  549. #define QGEMM_DEFAULT_R qgemm_r
  550. #define CGEMM_DEFAULT_P cgemm_p
  551. #define CGEMM_DEFAULT_R cgemm_r
  552. #define ZGEMM_DEFAULT_P zgemm_p
  553. #define ZGEMM_DEFAULT_R zgemm_r
  554. #define XGEMM_DEFAULT_P xgemm_p
  555. #define XGEMM_DEFAULT_R xgemm_r
  556. #define SGEMM_DEFAULT_Q 256
  557. #define DGEMM_DEFAULT_Q 256
  558. #define QGEMM_DEFAULT_Q 256
  559. #define CGEMM_DEFAULT_Q 256
  560. #define ZGEMM_DEFAULT_Q 256
  561. #define XGEMM_DEFAULT_Q 256
  562. #endif
  563. #ifdef PENRYN
  564. #define SNUMOPT 8
  565. #define DNUMOPT 4
  566. #define GEMM_DEFAULT_OFFSET_A 128
  567. #define GEMM_DEFAULT_OFFSET_B 0
  568. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  569. #define SYMV_P 8
  570. #define SWITCH_RATIO 4
  571. #ifdef ARCH_X86
  572. #define SGEMM_DEFAULT_UNROLL_M 4
  573. #define DGEMM_DEFAULT_UNROLL_M 2
  574. #define QGEMM_DEFAULT_UNROLL_M 2
  575. #define CGEMM_DEFAULT_UNROLL_M 2
  576. #define ZGEMM_DEFAULT_UNROLL_M 1
  577. #define XGEMM_DEFAULT_UNROLL_M 1
  578. #define SGEMM_DEFAULT_UNROLL_N 4
  579. #define DGEMM_DEFAULT_UNROLL_N 4
  580. #define QGEMM_DEFAULT_UNROLL_N 2
  581. #define CGEMM_DEFAULT_UNROLL_N 2
  582. #define ZGEMM_DEFAULT_UNROLL_N 2
  583. #define XGEMM_DEFAULT_UNROLL_N 1
  584. #else
  585. #define SGEMM_DEFAULT_UNROLL_M 8
  586. #define DGEMM_DEFAULT_UNROLL_M 4
  587. #define QGEMM_DEFAULT_UNROLL_M 2
  588. #define CGEMM_DEFAULT_UNROLL_M 4
  589. #define ZGEMM_DEFAULT_UNROLL_M 2
  590. #define XGEMM_DEFAULT_UNROLL_M 1
  591. #define SGEMM_DEFAULT_UNROLL_N 4
  592. #define DGEMM_DEFAULT_UNROLL_N 4
  593. #define QGEMM_DEFAULT_UNROLL_N 2
  594. #define CGEMM_DEFAULT_UNROLL_N 2
  595. #define ZGEMM_DEFAULT_UNROLL_N 2
  596. #define XGEMM_DEFAULT_UNROLL_N 1
  597. #endif
  598. #define SGEMM_DEFAULT_P sgemm_p
  599. #define SGEMM_DEFAULT_R sgemm_r
  600. #define DGEMM_DEFAULT_P dgemm_p
  601. #define DGEMM_DEFAULT_R dgemm_r
  602. #define QGEMM_DEFAULT_P qgemm_p
  603. #define QGEMM_DEFAULT_R qgemm_r
  604. #define CGEMM_DEFAULT_P cgemm_p
  605. #define CGEMM_DEFAULT_R cgemm_r
  606. #define ZGEMM_DEFAULT_P zgemm_p
  607. #define ZGEMM_DEFAULT_R zgemm_r
  608. #define XGEMM_DEFAULT_P xgemm_p
  609. #define XGEMM_DEFAULT_R xgemm_r
  610. #define SGEMM_DEFAULT_Q 512
  611. #define DGEMM_DEFAULT_Q 256
  612. #define QGEMM_DEFAULT_Q 128
  613. #define CGEMM_DEFAULT_Q 512
  614. #define ZGEMM_DEFAULT_Q 256
  615. #define XGEMM_DEFAULT_Q 128
  616. #define GETRF_FACTOR 0.75
  617. #endif
  618. #ifdef DUNNINGTON
  619. #define SNUMOPT 8
  620. #define DNUMOPT 4
  621. #define GEMM_DEFAULT_OFFSET_A 128
  622. #define GEMM_DEFAULT_OFFSET_B 0
  623. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  624. #define SYMV_P 8
  625. #define SWITCH_RATIO 4
  626. #ifdef ARCH_X86
  627. #define SGEMM_DEFAULT_UNROLL_M 4
  628. #define DGEMM_DEFAULT_UNROLL_M 2
  629. #define QGEMM_DEFAULT_UNROLL_M 2
  630. #define CGEMM_DEFAULT_UNROLL_M 2
  631. #define ZGEMM_DEFAULT_UNROLL_M 1
  632. #define XGEMM_DEFAULT_UNROLL_M 1
  633. #define SGEMM_DEFAULT_UNROLL_N 4
  634. #define DGEMM_DEFAULT_UNROLL_N 4
  635. #define QGEMM_DEFAULT_UNROLL_N 2
  636. #define CGEMM_DEFAULT_UNROLL_N 2
  637. #define ZGEMM_DEFAULT_UNROLL_N 2
  638. #define XGEMM_DEFAULT_UNROLL_N 1
  639. #else
  640. #define SGEMM_DEFAULT_UNROLL_M 8
  641. #define DGEMM_DEFAULT_UNROLL_M 4
  642. #define QGEMM_DEFAULT_UNROLL_M 2
  643. #define CGEMM_DEFAULT_UNROLL_M 4
  644. #define ZGEMM_DEFAULT_UNROLL_M 2
  645. #define XGEMM_DEFAULT_UNROLL_M 1
  646. #define SGEMM_DEFAULT_UNROLL_N 4
  647. #define DGEMM_DEFAULT_UNROLL_N 4
  648. #define QGEMM_DEFAULT_UNROLL_N 2
  649. #define CGEMM_DEFAULT_UNROLL_N 2
  650. #define ZGEMM_DEFAULT_UNROLL_N 2
  651. #define XGEMM_DEFAULT_UNROLL_N 1
  652. #endif
  653. #define SGEMM_DEFAULT_P sgemm_p
  654. #define SGEMM_DEFAULT_R sgemm_r
  655. #define DGEMM_DEFAULT_P dgemm_p
  656. #define DGEMM_DEFAULT_R dgemm_r
  657. #define QGEMM_DEFAULT_P qgemm_p
  658. #define QGEMM_DEFAULT_R qgemm_r
  659. #define CGEMM_DEFAULT_P cgemm_p
  660. #define CGEMM_DEFAULT_R cgemm_r
  661. #define ZGEMM_DEFAULT_P zgemm_p
  662. #define ZGEMM_DEFAULT_R zgemm_r
  663. #define XGEMM_DEFAULT_P xgemm_p
  664. #define XGEMM_DEFAULT_R xgemm_r
  665. #define SGEMM_DEFAULT_Q 768
  666. #define DGEMM_DEFAULT_Q 384
  667. #define QGEMM_DEFAULT_Q 192
  668. #define CGEMM_DEFAULT_Q 768
  669. #define ZGEMM_DEFAULT_Q 384
  670. #define XGEMM_DEFAULT_Q 192
  671. #define GETRF_FACTOR 0.75
  672. #define GEMM_THREAD gemm_thread_mn
  673. #endif
  674. #ifdef NEHALEM
  675. #define SNUMOPT 8
  676. #define DNUMOPT 4
  677. #define GEMM_DEFAULT_OFFSET_A 32
  678. #define GEMM_DEFAULT_OFFSET_B 0
  679. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  680. #define SYMV_P 8
  681. #define SWITCH_RATIO 4
  682. #ifdef ARCH_X86
  683. #define SGEMM_DEFAULT_UNROLL_M 4
  684. #define DGEMM_DEFAULT_UNROLL_M 2
  685. #define QGEMM_DEFAULT_UNROLL_M 2
  686. #define CGEMM_DEFAULT_UNROLL_M 2
  687. #define ZGEMM_DEFAULT_UNROLL_M 1
  688. #define XGEMM_DEFAULT_UNROLL_M 1
  689. #define SGEMM_DEFAULT_UNROLL_N 4
  690. #define DGEMM_DEFAULT_UNROLL_N 4
  691. #define QGEMM_DEFAULT_UNROLL_N 2
  692. #define CGEMM_DEFAULT_UNROLL_N 2
  693. #define ZGEMM_DEFAULT_UNROLL_N 2
  694. #define XGEMM_DEFAULT_UNROLL_N 1
  695. #else
  696. #define SGEMM_DEFAULT_UNROLL_M 4
  697. #define DGEMM_DEFAULT_UNROLL_M 2
  698. #define QGEMM_DEFAULT_UNROLL_M 2
  699. #define CGEMM_DEFAULT_UNROLL_M 2
  700. #define ZGEMM_DEFAULT_UNROLL_M 1
  701. #define XGEMM_DEFAULT_UNROLL_M 1
  702. #define SGEMM_DEFAULT_UNROLL_N 8
  703. #define DGEMM_DEFAULT_UNROLL_N 8
  704. #define QGEMM_DEFAULT_UNROLL_N 2
  705. #define CGEMM_DEFAULT_UNROLL_N 4
  706. #define ZGEMM_DEFAULT_UNROLL_N 4
  707. #define XGEMM_DEFAULT_UNROLL_N 1
  708. #endif
  709. #define SGEMM_DEFAULT_P 504
  710. #define SGEMM_DEFAULT_R sgemm_r
  711. #define DGEMM_DEFAULT_P 504
  712. #define DGEMM_DEFAULT_R dgemm_r
  713. #define QGEMM_DEFAULT_P 504
  714. #define QGEMM_DEFAULT_R qgemm_r
  715. #define CGEMM_DEFAULT_P 252
  716. #define CGEMM_DEFAULT_R cgemm_r
  717. #define ZGEMM_DEFAULT_P 252
  718. #define ZGEMM_DEFAULT_R zgemm_r
  719. #define XGEMM_DEFAULT_P 252
  720. #define XGEMM_DEFAULT_R xgemm_r
  721. #define SGEMM_DEFAULT_Q 512
  722. #define DGEMM_DEFAULT_Q 256
  723. #define QGEMM_DEFAULT_Q 128
  724. #define CGEMM_DEFAULT_Q 512
  725. #define ZGEMM_DEFAULT_Q 256
  726. #define XGEMM_DEFAULT_Q 128
  727. #define GETRF_FACTOR 0.72
  728. #endif
  729. #ifdef SANDYBRIDGE
  730. #define SNUMOPT 8
  731. #define DNUMOPT 4
  732. #define GEMM_DEFAULT_OFFSET_A 32
  733. #define GEMM_DEFAULT_OFFSET_B 0
  734. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  735. #define SYMV_P 8
  736. #define SWITCH_RATIO 4
  737. #ifdef ARCH_X86
  738. #define SGEMM_DEFAULT_UNROLL_M 4
  739. #define DGEMM_DEFAULT_UNROLL_M 2
  740. #define QGEMM_DEFAULT_UNROLL_M 2
  741. #define CGEMM_DEFAULT_UNROLL_M 2
  742. #define ZGEMM_DEFAULT_UNROLL_M 1
  743. #define XGEMM_DEFAULT_UNROLL_M 1
  744. #define SGEMM_DEFAULT_UNROLL_N 4
  745. #define DGEMM_DEFAULT_UNROLL_N 4
  746. #define QGEMM_DEFAULT_UNROLL_N 2
  747. #define CGEMM_DEFAULT_UNROLL_N 2
  748. #define ZGEMM_DEFAULT_UNROLL_N 2
  749. #define XGEMM_DEFAULT_UNROLL_N 1
  750. #else
  751. #define SGEMM_DEFAULT_UNROLL_M 4
  752. #define DGEMM_DEFAULT_UNROLL_M 2
  753. #define QGEMM_DEFAULT_UNROLL_M 2
  754. #define CGEMM_DEFAULT_UNROLL_M 2
  755. #define ZGEMM_DEFAULT_UNROLL_M 1
  756. #define XGEMM_DEFAULT_UNROLL_M 1
  757. #define SGEMM_DEFAULT_UNROLL_N 8
  758. #define DGEMM_DEFAULT_UNROLL_N 8
  759. #define QGEMM_DEFAULT_UNROLL_N 2
  760. #define CGEMM_DEFAULT_UNROLL_N 4
  761. #define ZGEMM_DEFAULT_UNROLL_N 4
  762. #define XGEMM_DEFAULT_UNROLL_N 1
  763. #endif
  764. #define SGEMM_DEFAULT_P 504
  765. #define SGEMM_DEFAULT_R sgemm_r
  766. #define DGEMM_DEFAULT_P 504
  767. #define DGEMM_DEFAULT_R dgemm_r
  768. #define QGEMM_DEFAULT_P 504
  769. #define QGEMM_DEFAULT_R qgemm_r
  770. #define CGEMM_DEFAULT_P 252
  771. #define CGEMM_DEFAULT_R cgemm_r
  772. #define ZGEMM_DEFAULT_P 252
  773. #define ZGEMM_DEFAULT_R zgemm_r
  774. #define XGEMM_DEFAULT_P 252
  775. #define XGEMM_DEFAULT_R xgemm_r
  776. #define SGEMM_DEFAULT_Q 512
  777. #define DGEMM_DEFAULT_Q 256
  778. #define QGEMM_DEFAULT_Q 128
  779. #define CGEMM_DEFAULT_Q 512
  780. #define ZGEMM_DEFAULT_Q 256
  781. #define XGEMM_DEFAULT_Q 128
  782. #define GETRF_FACTOR 0.72
  783. #endif
  784. #ifdef ATOM
  785. #define SNUMOPT 2
  786. #define DNUMOPT 1
  787. #define GEMM_DEFAULT_OFFSET_A 64
  788. #define GEMM_DEFAULT_OFFSET_B 0
  789. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  790. #define SYMV_P 8
  791. #ifdef ARCH_X86
  792. #define SGEMM_DEFAULT_UNROLL_M 4
  793. #define DGEMM_DEFAULT_UNROLL_M 2
  794. #define QGEMM_DEFAULT_UNROLL_M 2
  795. #define CGEMM_DEFAULT_UNROLL_M 2
  796. #define ZGEMM_DEFAULT_UNROLL_M 1
  797. #define XGEMM_DEFAULT_UNROLL_M 1
  798. #else
  799. #define SGEMM_DEFAULT_UNROLL_M 8
  800. #define DGEMM_DEFAULT_UNROLL_M 4
  801. #define QGEMM_DEFAULT_UNROLL_M 2
  802. #define CGEMM_DEFAULT_UNROLL_M 4
  803. #define ZGEMM_DEFAULT_UNROLL_M 2
  804. #define XGEMM_DEFAULT_UNROLL_M 1
  805. #endif
  806. #define SGEMM_DEFAULT_UNROLL_N 4
  807. #define DGEMM_DEFAULT_UNROLL_N 2
  808. #define QGEMM_DEFAULT_UNROLL_N 2
  809. #define CGEMM_DEFAULT_UNROLL_N 2
  810. #define ZGEMM_DEFAULT_UNROLL_N 1
  811. #define XGEMM_DEFAULT_UNROLL_N 1
  812. #define SGEMM_DEFAULT_P sgemm_p
  813. #define SGEMM_DEFAULT_R sgemm_r
  814. #define DGEMM_DEFAULT_P dgemm_p
  815. #define DGEMM_DEFAULT_R dgemm_r
  816. #define QGEMM_DEFAULT_P qgemm_p
  817. #define QGEMM_DEFAULT_R qgemm_r
  818. #define CGEMM_DEFAULT_P cgemm_p
  819. #define CGEMM_DEFAULT_R cgemm_r
  820. #define ZGEMM_DEFAULT_P zgemm_p
  821. #define ZGEMM_DEFAULT_R zgemm_r
  822. #define XGEMM_DEFAULT_P xgemm_p
  823. #define XGEMM_DEFAULT_R xgemm_r
  824. #define SGEMM_DEFAULT_Q 256
  825. #define DGEMM_DEFAULT_Q 256
  826. #define QGEMM_DEFAULT_Q 256
  827. #define CGEMM_DEFAULT_Q 256
  828. #define ZGEMM_DEFAULT_Q 256
  829. #define XGEMM_DEFAULT_Q 256
  830. #endif
  831. #ifdef ITANIUM2
  832. #define SNUMOPT 4
  833. #define DNUMOPT 4
  834. #define GEMM_DEFAULT_OFFSET_A 0
  835. #define GEMM_DEFAULT_OFFSET_B 128
  836. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  837. #define SGEMM_DEFAULT_UNROLL_M 8
  838. #define SGEMM_DEFAULT_UNROLL_N 8
  839. #define DGEMM_DEFAULT_UNROLL_M 8
  840. #define DGEMM_DEFAULT_UNROLL_N 8
  841. #define QGEMM_DEFAULT_UNROLL_M 8
  842. #define QGEMM_DEFAULT_UNROLL_N 8
  843. #define CGEMM_DEFAULT_UNROLL_M 4
  844. #define CGEMM_DEFAULT_UNROLL_N 4
  845. #define ZGEMM_DEFAULT_UNROLL_M 4
  846. #define ZGEMM_DEFAULT_UNROLL_N 4
  847. #define XGEMM_DEFAULT_UNROLL_M 4
  848. #define XGEMM_DEFAULT_UNROLL_N 4
  849. #define SGEMM_DEFAULT_P sgemm_p
  850. #define DGEMM_DEFAULT_P dgemm_p
  851. #define QGEMM_DEFAULT_P qgemm_p
  852. #define CGEMM_DEFAULT_P cgemm_p
  853. #define ZGEMM_DEFAULT_P zgemm_p
  854. #define XGEMM_DEFAULT_P xgemm_p
  855. #define SGEMM_DEFAULT_Q 1024
  856. #define DGEMM_DEFAULT_Q 1024
  857. #define QGEMM_DEFAULT_Q 1024
  858. #define CGEMM_DEFAULT_Q 1024
  859. #define ZGEMM_DEFAULT_Q 1024
  860. #define XGEMM_DEFAULT_Q 1024
  861. #define SGEMM_DEFAULT_R sgemm_r
  862. #define DGEMM_DEFAULT_R dgemm_r
  863. #define QGEMM_DEFAULT_R qgemm_r
  864. #define CGEMM_DEFAULT_R cgemm_r
  865. #define ZGEMM_DEFAULT_R zgemm_r
  866. #define XGEMM_DEFAULT_R xgemm_r
  867. #define SYMV_P 16
  868. #define GETRF_FACTOR 0.65
  869. #endif
  870. #if defined(EV4) || defined(EV5) || defined(EV6)
  871. #ifdef EV4
  872. #define SNUMOPT 1
  873. #define DNUMOPT 1
  874. #else
  875. #define SNUMOPT 2
  876. #define DNUMOPT 2
  877. #endif
  878. #define GEMM_DEFAULT_OFFSET_A 512
  879. #define GEMM_DEFAULT_OFFSET_B 512
  880. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  881. #define SGEMM_DEFAULT_UNROLL_M 4
  882. #define SGEMM_DEFAULT_UNROLL_N 4
  883. #define DGEMM_DEFAULT_UNROLL_M 4
  884. #define DGEMM_DEFAULT_UNROLL_N 4
  885. #define CGEMM_DEFAULT_UNROLL_M 2
  886. #define CGEMM_DEFAULT_UNROLL_N 2
  887. #define ZGEMM_DEFAULT_UNROLL_M 2
  888. #define ZGEMM_DEFAULT_UNROLL_N 2
  889. #define SYMV_P 8
  890. #ifdef EV4
  891. #define SGEMM_DEFAULT_P 32
  892. #define SGEMM_DEFAULT_Q 112
  893. #define SGEMM_DEFAULT_R 256
  894. #define DGEMM_DEFAULT_P 32
  895. #define DGEMM_DEFAULT_Q 56
  896. #define DGEMM_DEFAULT_R 256
  897. #define CGEMM_DEFAULT_P 32
  898. #define CGEMM_DEFAULT_Q 64
  899. #define CGEMM_DEFAULT_R 240
  900. #define ZGEMM_DEFAULT_P 32
  901. #define ZGEMM_DEFAULT_Q 32
  902. #define ZGEMM_DEFAULT_R 240
  903. #endif
  904. #ifdef EV5
  905. #define SGEMM_DEFAULT_P 64
  906. #define SGEMM_DEFAULT_Q 256
  907. #define DGEMM_DEFAULT_P 64
  908. #define DGEMM_DEFAULT_Q 128
  909. #define CGEMM_DEFAULT_P 64
  910. #define CGEMM_DEFAULT_Q 128
  911. #define ZGEMM_DEFAULT_P 64
  912. #define ZGEMM_DEFAULT_Q 64
  913. #endif
  914. #ifdef EV6
  915. #define SGEMM_DEFAULT_P 256
  916. #define SGEMM_DEFAULT_Q 512
  917. #define DGEMM_DEFAULT_P 256
  918. #define DGEMM_DEFAULT_Q 256
  919. #define CGEMM_DEFAULT_P 256
  920. #define CGEMM_DEFAULT_Q 256
  921. #define ZGEMM_DEFAULT_P 128
  922. #define ZGEMM_DEFAULT_Q 256
  923. #endif
  924. #endif
  925. #ifdef CELL
  926. #define SNUMOPT 2
  927. #define DNUMOPT 2
  928. #define GEMM_DEFAULT_OFFSET_A 0
  929. #define GEMM_DEFAULT_OFFSET_B 8192
  930. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  931. #define SGEMM_DEFAULT_UNROLL_M 16
  932. #define SGEMM_DEFAULT_UNROLL_N 4
  933. #define DGEMM_DEFAULT_UNROLL_M 4
  934. #define DGEMM_DEFAULT_UNROLL_N 4
  935. #define CGEMM_DEFAULT_UNROLL_M 8
  936. #define CGEMM_DEFAULT_UNROLL_N 2
  937. #define ZGEMM_DEFAULT_UNROLL_M 2
  938. #define ZGEMM_DEFAULT_UNROLL_N 2
  939. #define SGEMM_DEFAULT_P 128
  940. #define DGEMM_DEFAULT_P 128
  941. #define CGEMM_DEFAULT_P 128
  942. #define ZGEMM_DEFAULT_P 128
  943. #define SGEMM_DEFAULT_Q 512
  944. #define DGEMM_DEFAULT_Q 256
  945. #define CGEMM_DEFAULT_Q 256
  946. #define ZGEMM_DEFAULT_Q 128
  947. #define SYMV_P 4
  948. #endif
  949. #ifdef PPCG4
  950. #define GEMM_DEFAULT_OFFSET_A 0
  951. #define GEMM_DEFAULT_OFFSET_B 1024
  952. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  953. #define SGEMM_DEFAULT_UNROLL_M 16
  954. #define SGEMM_DEFAULT_UNROLL_N 4
  955. #define DGEMM_DEFAULT_UNROLL_M 4
  956. #define DGEMM_DEFAULT_UNROLL_N 4
  957. #define CGEMM_DEFAULT_UNROLL_M 8
  958. #define CGEMM_DEFAULT_UNROLL_N 2
  959. #define ZGEMM_DEFAULT_UNROLL_M 2
  960. #define ZGEMM_DEFAULT_UNROLL_N 2
  961. #define SGEMM_DEFAULT_P 256
  962. #define DGEMM_DEFAULT_P 128
  963. #define CGEMM_DEFAULT_P 128
  964. #define ZGEMM_DEFAULT_P 64
  965. #define SGEMM_DEFAULT_Q 256
  966. #define DGEMM_DEFAULT_Q 256
  967. #define CGEMM_DEFAULT_Q 256
  968. #define ZGEMM_DEFAULT_Q 256
  969. #define SYMV_P 4
  970. #endif
  971. #ifdef PPC970
  972. #define SNUMOPT 4
  973. #define DNUMOPT 4
  974. #define GEMM_DEFAULT_OFFSET_A 2688
  975. #define GEMM_DEFAULT_OFFSET_B 3072
  976. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  977. #define SGEMM_DEFAULT_UNROLL_M 16
  978. #define SGEMM_DEFAULT_UNROLL_N 4
  979. #define DGEMM_DEFAULT_UNROLL_M 4
  980. #define DGEMM_DEFAULT_UNROLL_N 4
  981. #define CGEMM_DEFAULT_UNROLL_M 8
  982. #define CGEMM_DEFAULT_UNROLL_N 2
  983. #define ZGEMM_DEFAULT_UNROLL_M 2
  984. #define ZGEMM_DEFAULT_UNROLL_N 2
  985. #ifdef OS_LINUX
  986. #if L2_SIZE == 1024976
  987. #define SGEMM_DEFAULT_P 320
  988. #define DGEMM_DEFAULT_P 256
  989. #define CGEMM_DEFAULT_P 256
  990. #define ZGEMM_DEFAULT_P 256
  991. #else
  992. #define SGEMM_DEFAULT_P 176
  993. #define DGEMM_DEFAULT_P 176
  994. #define CGEMM_DEFAULT_P 176
  995. #define ZGEMM_DEFAULT_P 176
  996. #endif
  997. #endif
  998. #define SGEMM_DEFAULT_Q 512
  999. #define DGEMM_DEFAULT_Q 256
  1000. #define CGEMM_DEFAULT_Q 256
  1001. #define ZGEMM_DEFAULT_Q 128
  1002. #define SYMV_P 4
  1003. #endif
  1004. #ifdef PPC440
  1005. #define SNUMOPT 2
  1006. #define DNUMOPT 2
  1007. #define GEMM_DEFAULT_OFFSET_A (32 * 0)
  1008. #define GEMM_DEFAULT_OFFSET_B (32 * 0)
  1009. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1010. #define SGEMM_DEFAULT_UNROLL_M 4
  1011. #define SGEMM_DEFAULT_UNROLL_N 4
  1012. #define DGEMM_DEFAULT_UNROLL_M 4
  1013. #define DGEMM_DEFAULT_UNROLL_N 4
  1014. #define CGEMM_DEFAULT_UNROLL_M 2
  1015. #define CGEMM_DEFAULT_UNROLL_N 2
  1016. #define ZGEMM_DEFAULT_UNROLL_M 2
  1017. #define ZGEMM_DEFAULT_UNROLL_N 2
  1018. #define SGEMM_DEFAULT_P 512
  1019. #define DGEMM_DEFAULT_P 512
  1020. #define CGEMM_DEFAULT_P 512
  1021. #define ZGEMM_DEFAULT_P 512
  1022. #define SGEMM_DEFAULT_Q 1024
  1023. #define DGEMM_DEFAULT_Q 512
  1024. #define CGEMM_DEFAULT_Q 512
  1025. #define ZGEMM_DEFAULT_Q 256
  1026. #define SGEMM_DEFAULT_R SGEMM_DEFAULT_P
  1027. #define DGEMM_DEFAULT_R DGEMM_DEFAULT_P
  1028. #define CGEMM_DEFAULT_R CGEMM_DEFAULT_P
  1029. #define ZGEMM_DEFAULT_R ZGEMM_DEFAULT_P
  1030. #define SYMV_P 4
  1031. #endif
  1032. #ifdef PPC440FP2
  1033. #define SNUMOPT 4
  1034. #define DNUMOPT 4
  1035. #define GEMM_DEFAULT_OFFSET_A (32 * 0)
  1036. #define GEMM_DEFAULT_OFFSET_B (32 * 0)
  1037. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1038. #define SGEMM_DEFAULT_UNROLL_M 8
  1039. #define SGEMM_DEFAULT_UNROLL_N 4
  1040. #define DGEMM_DEFAULT_UNROLL_M 8
  1041. #define DGEMM_DEFAULT_UNROLL_N 4
  1042. #define CGEMM_DEFAULT_UNROLL_M 4
  1043. #define CGEMM_DEFAULT_UNROLL_N 2
  1044. #define ZGEMM_DEFAULT_UNROLL_M 4
  1045. #define ZGEMM_DEFAULT_UNROLL_N 2
  1046. #define SGEMM_DEFAULT_P 128
  1047. #define DGEMM_DEFAULT_P 128
  1048. #define CGEMM_DEFAULT_P 128
  1049. #define ZGEMM_DEFAULT_P 128
  1050. #if 1
  1051. #define SGEMM_DEFAULT_Q 4096
  1052. #define DGEMM_DEFAULT_Q 3072
  1053. #define CGEMM_DEFAULT_Q 2048
  1054. #define ZGEMM_DEFAULT_Q 1024
  1055. #else
  1056. #define SGEMM_DEFAULT_Q 512
  1057. #define DGEMM_DEFAULT_Q 256
  1058. #define CGEMM_DEFAULT_Q 256
  1059. #define ZGEMM_DEFAULT_Q 128
  1060. #endif
  1061. #define SYMV_P 4
  1062. #endif
  1063. #if defined(POWER3) || defined(POWER4) || defined(POWER5)
  1064. #define GEMM_DEFAULT_OFFSET_A 0
  1065. #define GEMM_DEFAULT_OFFSET_B 2048
  1066. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1067. #define SGEMM_DEFAULT_UNROLL_M 4
  1068. #define SGEMM_DEFAULT_UNROLL_N 4
  1069. #define DGEMM_DEFAULT_UNROLL_M 4
  1070. #define DGEMM_DEFAULT_UNROLL_N 4
  1071. #define CGEMM_DEFAULT_UNROLL_M 2
  1072. #define CGEMM_DEFAULT_UNROLL_N 2
  1073. #define ZGEMM_DEFAULT_UNROLL_M 2
  1074. #define ZGEMM_DEFAULT_UNROLL_N 2
  1075. #ifdef POWER3
  1076. #define SNUMOPT 4
  1077. #define DNUMOPT 4
  1078. #define SGEMM_DEFAULT_P 256
  1079. #define SGEMM_DEFAULT_Q 432
  1080. #define SGEMM_DEFAULT_R 1012
  1081. #define DGEMM_DEFAULT_P 256
  1082. #define DGEMM_DEFAULT_Q 216
  1083. #define DGEMM_DEFAULT_R 1012
  1084. #define ZGEMM_DEFAULT_P 256
  1085. #define ZGEMM_DEFAULT_Q 104
  1086. #define ZGEMM_DEFAULT_R 1012
  1087. #endif
  1088. #if defined(POWER4)
  1089. #ifdef ALLOC_HUGETLB
  1090. #define SGEMM_DEFAULT_P 184
  1091. #define DGEMM_DEFAULT_P 184
  1092. #define CGEMM_DEFAULT_P 184
  1093. #define ZGEMM_DEFAULT_P 184
  1094. #else
  1095. #define SGEMM_DEFAULT_P 144
  1096. #define DGEMM_DEFAULT_P 144
  1097. #define CGEMM_DEFAULT_P 144
  1098. #define ZGEMM_DEFAULT_P 144
  1099. #endif
  1100. #endif
  1101. #if defined(POWER5)
  1102. #ifdef ALLOC_HUGETLB
  1103. #define SGEMM_DEFAULT_P 512
  1104. #define DGEMM_DEFAULT_P 256
  1105. #define CGEMM_DEFAULT_P 256
  1106. #define ZGEMM_DEFAULT_P 128
  1107. #else
  1108. #define SGEMM_DEFAULT_P 320
  1109. #define DGEMM_DEFAULT_P 160
  1110. #define CGEMM_DEFAULT_P 160
  1111. #define ZGEMM_DEFAULT_P 80
  1112. #endif
  1113. #define SGEMM_DEFAULT_Q 256
  1114. #define CGEMM_DEFAULT_Q 256
  1115. #define DGEMM_DEFAULT_Q 256
  1116. #define ZGEMM_DEFAULT_Q 256
  1117. #endif
  1118. #define SYMV_P 8
  1119. #endif
  1120. #if defined(POWER6)
  1121. #define SNUMOPT 4
  1122. #define DNUMOPT 4
  1123. #define GEMM_DEFAULT_OFFSET_A 384
  1124. #define GEMM_DEFAULT_OFFSET_B 1024
  1125. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1126. #define SGEMM_DEFAULT_UNROLL_M 4
  1127. #define SGEMM_DEFAULT_UNROLL_N 4
  1128. #define DGEMM_DEFAULT_UNROLL_M 4
  1129. #define DGEMM_DEFAULT_UNROLL_N 4
  1130. #define CGEMM_DEFAULT_UNROLL_M 2
  1131. #define CGEMM_DEFAULT_UNROLL_N 4
  1132. #define ZGEMM_DEFAULT_UNROLL_M 2
  1133. #define ZGEMM_DEFAULT_UNROLL_N 4
  1134. #define SGEMM_DEFAULT_P 992
  1135. #define DGEMM_DEFAULT_P 480
  1136. #define CGEMM_DEFAULT_P 488
  1137. #define ZGEMM_DEFAULT_P 248
  1138. #define SGEMM_DEFAULT_Q 504
  1139. #define DGEMM_DEFAULT_Q 504
  1140. #define CGEMM_DEFAULT_Q 400
  1141. #define ZGEMM_DEFAULT_Q 400
  1142. #define SYMV_P 8
  1143. #endif
  1144. #if defined(SPARC) && defined(V7)
  1145. #define SNUMOPT 4
  1146. #define DNUMOPT 4
  1147. #define GEMM_DEFAULT_OFFSET_A 0
  1148. #define GEMM_DEFAULT_OFFSET_B 2048
  1149. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1150. #define SGEMM_DEFAULT_UNROLL_M 2
  1151. #define SGEMM_DEFAULT_UNROLL_N 8
  1152. #define DGEMM_DEFAULT_UNROLL_M 2
  1153. #define DGEMM_DEFAULT_UNROLL_N 8
  1154. #define CGEMM_DEFAULT_UNROLL_M 1
  1155. #define CGEMM_DEFAULT_UNROLL_N 4
  1156. #define ZGEMM_DEFAULT_UNROLL_M 1
  1157. #define ZGEMM_DEFAULT_UNROLL_N 4
  1158. #define SGEMM_DEFAULT_P 256
  1159. #define DGEMM_DEFAULT_P 256
  1160. #define CGEMM_DEFAULT_P 256
  1161. #define ZGEMM_DEFAULT_P 256
  1162. #define SGEMM_DEFAULT_Q 512
  1163. #define DGEMM_DEFAULT_Q 256
  1164. #define CGEMM_DEFAULT_Q 256
  1165. #define ZGEMM_DEFAULT_Q 128
  1166. #define SYMV_P 8
  1167. #define GEMM_THREAD gemm_thread_mn
  1168. #endif
  1169. #if defined(SPARC) && defined(V9)
  1170. #define SNUMOPT 2
  1171. #define DNUMOPT 2
  1172. #define GEMM_DEFAULT_OFFSET_A 0
  1173. #define GEMM_DEFAULT_OFFSET_B 2048
  1174. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1175. #define SGEMM_DEFAULT_UNROLL_M 4
  1176. #define SGEMM_DEFAULT_UNROLL_N 4
  1177. #define DGEMM_DEFAULT_UNROLL_M 4
  1178. #define DGEMM_DEFAULT_UNROLL_N 4
  1179. #define CGEMM_DEFAULT_UNROLL_M 2
  1180. #define CGEMM_DEFAULT_UNROLL_N 2
  1181. #define ZGEMM_DEFAULT_UNROLL_M 2
  1182. #define ZGEMM_DEFAULT_UNROLL_N 2
  1183. #define SGEMM_DEFAULT_P 512
  1184. #define DGEMM_DEFAULT_P 512
  1185. #define CGEMM_DEFAULT_P 512
  1186. #define ZGEMM_DEFAULT_P 512
  1187. #define SGEMM_DEFAULT_Q 1024
  1188. #define DGEMM_DEFAULT_Q 512
  1189. #define CGEMM_DEFAULT_Q 512
  1190. #define ZGEMM_DEFAULT_Q 256
  1191. #define SYMV_P 8
  1192. #endif
  1193. #ifdef SICORTEX
  1194. #define SNUMOPT 2
  1195. #define DNUMOPT 2
  1196. #define GEMM_DEFAULT_OFFSET_A 0
  1197. #define GEMM_DEFAULT_OFFSET_B 0
  1198. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1199. #define SGEMM_DEFAULT_UNROLL_M 2
  1200. #define SGEMM_DEFAULT_UNROLL_N 8
  1201. #define DGEMM_DEFAULT_UNROLL_M 2
  1202. #define DGEMM_DEFAULT_UNROLL_N 8
  1203. #define CGEMM_DEFAULT_UNROLL_M 1
  1204. #define CGEMM_DEFAULT_UNROLL_N 4
  1205. #define ZGEMM_DEFAULT_UNROLL_M 1
  1206. #define ZGEMM_DEFAULT_UNROLL_N 4
  1207. #define SGEMM_DEFAULT_P 108
  1208. #define DGEMM_DEFAULT_P 112
  1209. #define CGEMM_DEFAULT_P 108
  1210. #define ZGEMM_DEFAULT_P 112
  1211. #define SGEMM_DEFAULT_Q 288
  1212. #define DGEMM_DEFAULT_Q 144
  1213. #define CGEMM_DEFAULT_Q 144
  1214. #define ZGEMM_DEFAULT_Q 72
  1215. #define SGEMM_DEFAULT_R 2000
  1216. #define DGEMM_DEFAULT_R 2000
  1217. #define CGEMM_DEFAULT_R 2000
  1218. #define ZGEMM_DEFAULT_R 2000
  1219. #define SYMV_P 16
  1220. #endif
  1221. #ifdef LOONGSON3A
  1222. ////Copy from SICORTEX
  1223. #define SNUMOPT 2
  1224. #define DNUMOPT 2
  1225. #define GEMM_DEFAULT_OFFSET_A 0
  1226. #define GEMM_DEFAULT_OFFSET_B 0
  1227. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1228. #define SGEMM_DEFAULT_UNROLL_M 8
  1229. #define SGEMM_DEFAULT_UNROLL_N 4
  1230. #define DGEMM_DEFAULT_UNROLL_M 4
  1231. #define DGEMM_DEFAULT_UNROLL_N 4
  1232. #define CGEMM_DEFAULT_UNROLL_M 4
  1233. #define CGEMM_DEFAULT_UNROLL_N 2
  1234. #define ZGEMM_DEFAULT_UNROLL_M 2
  1235. #define ZGEMM_DEFAULT_UNROLL_N 2
  1236. #define SGEMM_DEFAULT_P 64
  1237. #define DGEMM_DEFAULT_P 44
  1238. #define CGEMM_DEFAULT_P 64
  1239. #define ZGEMM_DEFAULT_P 32
  1240. #define SGEMM_DEFAULT_Q 192
  1241. #define DGEMM_DEFAULT_Q 92
  1242. #define CGEMM_DEFAULT_Q 128
  1243. #define ZGEMM_DEFAULT_Q 80
  1244. #define SGEMM_DEFAULT_R 640
  1245. #define DGEMM_DEFAULT_R dgemm_r
  1246. #define CGEMM_DEFAULT_R 640
  1247. #define ZGEMM_DEFAULT_R 640
  1248. #define GEMM_OFFSET_A1 0x10000
  1249. #define GEMM_OFFSET_B1 0x100000
  1250. #define SYMV_P 16
  1251. #endif
  1252. #ifdef LOONGSON3B
  1253. #define SNUMOPT 2
  1254. #define DNUMOPT 2
  1255. #define GEMM_DEFAULT_OFFSET_A 0
  1256. #define GEMM_DEFAULT_OFFSET_B 0
  1257. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1258. #define SGEMM_DEFAULT_UNROLL_M 2
  1259. #define SGEMM_DEFAULT_UNROLL_N 2
  1260. #define DGEMM_DEFAULT_UNROLL_M 2
  1261. #define DGEMM_DEFAULT_UNROLL_N 2
  1262. #define CGEMM_DEFAULT_UNROLL_M 2
  1263. #define CGEMM_DEFAULT_UNROLL_N 2
  1264. #define ZGEMM_DEFAULT_UNROLL_M 2
  1265. #define ZGEMM_DEFAULT_UNROLL_N 2
  1266. #define SGEMM_DEFAULT_P 64
  1267. #define DGEMM_DEFAULT_P 24
  1268. #define CGEMM_DEFAULT_P 24
  1269. #define ZGEMM_DEFAULT_P 20
  1270. #define SGEMM_DEFAULT_Q 192
  1271. #define DGEMM_DEFAULT_Q 128
  1272. #define CGEMM_DEFAULT_Q 128
  1273. #define ZGEMM_DEFAULT_Q 64
  1274. #define SGEMM_DEFAULT_R 512
  1275. #define DGEMM_DEFAULT_R 512
  1276. #define CGEMM_DEFAULT_R 512
  1277. #define ZGEMM_DEFAULT_R 512
  1278. #define GEMM_OFFSET_A1 0x10000
  1279. #define GEMM_OFFSET_B1 0x100000
  1280. #define SYMV_P 16
  1281. #endif
  1282. #ifdef GENERIC
  1283. #define SNUMOPT 2
  1284. #define DNUMOPT 2
  1285. #define GEMM_DEFAULT_OFFSET_A 0
  1286. #define GEMM_DEFAULT_OFFSET_B 0
  1287. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1288. #define SGEMM_DEFAULT_UNROLL_N 4
  1289. #define DGEMM_DEFAULT_UNROLL_N 4
  1290. #define QGEMM_DEFAULT_UNROLL_N 2
  1291. #define CGEMM_DEFAULT_UNROLL_N 2
  1292. #define ZGEMM_DEFAULT_UNROLL_N 2
  1293. #define XGEMM_DEFAULT_UNROLL_N 1
  1294. #ifdef ARCH_X86
  1295. #define SGEMM_DEFAULT_UNROLL_M 4
  1296. #define DGEMM_DEFAULT_UNROLL_M 2
  1297. #define QGEMM_DEFAULT_UNROLL_M 2
  1298. #define CGEMM_DEFAULT_UNROLL_M 2
  1299. #define ZGEMM_DEFAULT_UNROLL_M 1
  1300. #define XGEMM_DEFAULT_UNROLL_M 1
  1301. #else
  1302. #define SGEMM_DEFAULT_UNROLL_M 8
  1303. #define DGEMM_DEFAULT_UNROLL_M 4
  1304. #define QGEMM_DEFAULT_UNROLL_M 2
  1305. #define CGEMM_DEFAULT_UNROLL_M 4
  1306. #define ZGEMM_DEFAULT_UNROLL_M 2
  1307. #define XGEMM_DEFAULT_UNROLL_M 1
  1308. #endif
  1309. #define SGEMM_P sgemm_p
  1310. #define DGEMM_P dgemm_p
  1311. #define QGEMM_P qgemm_p
  1312. #define CGEMM_P cgemm_p
  1313. #define ZGEMM_P zgemm_p
  1314. #define XGEMM_P xgemm_p
  1315. #define SGEMM_R sgemm_r
  1316. #define DGEMM_R dgemm_r
  1317. #define QGEMM_R qgemm_r
  1318. #define CGEMM_R cgemm_r
  1319. #define ZGEMM_R zgemm_r
  1320. #define XGEMM_R xgemm_r
  1321. #define SGEMM_Q 128
  1322. #define DGEMM_Q 128
  1323. #define QGEMM_Q 128
  1324. #define CGEMM_Q 128
  1325. #define ZGEMM_Q 128
  1326. #define XGEMM_Q 128
  1327. #define SYMV_P 16
  1328. #endif
  1329. #ifndef QGEMM_DEFAULT_UNROLL_M
  1330. #define QGEMM_DEFAULT_UNROLL_M 2
  1331. #endif
  1332. #ifndef QGEMM_DEFAULT_UNROLL_N
  1333. #define QGEMM_DEFAULT_UNROLL_N 2
  1334. #endif
  1335. #ifndef XGEMM_DEFAULT_UNROLL_M
  1336. #define XGEMM_DEFAULT_UNROLL_M 2
  1337. #endif
  1338. #ifndef XGEMM_DEFAULT_UNROLL_N
  1339. #define XGEMM_DEFAULT_UNROLL_N 2
  1340. #endif
  1341. #ifndef HAVE_SSE2
  1342. #define SHUFPD_0 shufps $0x44,
  1343. #define SHUFPD_1 shufps $0x4e,
  1344. #define SHUFPD_2 shufps $0xe4,
  1345. #define SHUFPD_3 shufps $0xee,
  1346. #endif
  1347. #ifndef SHUFPD_0
  1348. #define SHUFPD_0 shufpd $0,
  1349. #endif
  1350. #ifndef SHUFPD_1
  1351. #define SHUFPD_1 shufpd $1,
  1352. #endif
  1353. #ifndef SHUFPD_2
  1354. #define SHUFPD_2 shufpd $2,
  1355. #endif
  1356. #ifndef SHUFPD_3
  1357. #define SHUFPD_3 shufpd $3,
  1358. #endif
  1359. #ifndef SHUFPS_39
  1360. #define SHUFPS_39 shufps $0x39,
  1361. #endif
  1362. #endif