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.

12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939194019411942194319441945194619471948194919501951195219531954195519561957195819591960196119621963196419651966196719681969197019711972197319741975197619771978197919801981198219831984198519861987198819891990199119921993199419951996199719981999200020012002200320042005200620072008200920102011201220132014201520162017201820192020202120222023202420252026202720282029203020312032203320342035203620372038203920402041204220432044204520462047204820492050205120522053205420552056205720582059206020612062206320642065206620672068206920702071207220732074207520762077207820792080208120822083208420852086208720882089209020912092209320942095209620972098209921002101210221032104210521062107210821092110211121122113211421152116211721182119212021212122212321242125212621272128212921302131213221332134213521362137213821392140214121422143214421452146214721482149215021512152215321542155215621572158215921602161216221632164216521662167216821692170217121722173217421752176
  1. /*****************************************************************************
  2. Copyright (c) 2011,2012 Lab of Parallel Software and Computational Science,ISCAS
  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) || defined(BOBCAT)
  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 BULLDOZER
  188. #define SNUMOPT 8
  189. #define DNUMOPT 4
  190. #define GEMM_DEFAULT_OFFSET_A 64
  191. #define GEMM_DEFAULT_OFFSET_B 832
  192. #define GEMM_DEFAULT_ALIGN 0x0fffUL
  193. #define QGEMM_DEFAULT_UNROLL_N 2
  194. #define CGEMM_DEFAULT_UNROLL_N 2
  195. #define ZGEMM_DEFAULT_UNROLL_N 2
  196. #define XGEMM_DEFAULT_UNROLL_N 1
  197. #ifdef ARCH_X86
  198. #define SGEMM_DEFAULT_UNROLL_N 4
  199. #define DGEMM_DEFAULT_UNROLL_N 4
  200. #define SGEMM_DEFAULT_UNROLL_M 4
  201. #define DGEMM_DEFAULT_UNROLL_M 2
  202. #define QGEMM_DEFAULT_UNROLL_M 2
  203. #define CGEMM_DEFAULT_UNROLL_M 2
  204. #define ZGEMM_DEFAULT_UNROLL_M 1
  205. #define XGEMM_DEFAULT_UNROLL_M 1
  206. #else
  207. #define SGEMM_DEFAULT_UNROLL_N 2
  208. #define DGEMM_DEFAULT_UNROLL_N 2
  209. #define SGEMM_DEFAULT_UNROLL_M 16
  210. #define DGEMM_DEFAULT_UNROLL_M 8
  211. #define QGEMM_DEFAULT_UNROLL_M 2
  212. #define CGEMM_DEFAULT_UNROLL_M 4
  213. #define ZGEMM_DEFAULT_UNROLL_M 2
  214. #define XGEMM_DEFAULT_UNROLL_M 1
  215. #define CGEMM3M_DEFAULT_UNROLL_N 4
  216. #define CGEMM3M_DEFAULT_UNROLL_M 8
  217. #define ZGEMM3M_DEFAULT_UNROLL_N 4
  218. #define ZGEMM3M_DEFAULT_UNROLL_M 4
  219. #define GEMV_UNROLL 8
  220. #endif
  221. #if defined(ARCH_X86_64)
  222. #define SGEMM_DEFAULT_P 768
  223. #define DGEMM_DEFAULT_P 384
  224. #else
  225. #define SGEMM_DEFAULT_P 448
  226. #define DGEMM_DEFAULT_P 224
  227. #endif
  228. #define QGEMM_DEFAULT_P 112
  229. #define CGEMM_DEFAULT_P 224
  230. #define ZGEMM_DEFAULT_P 112
  231. #define XGEMM_DEFAULT_P 56
  232. #if defined(ARCH_X86_64)
  233. #define SGEMM_DEFAULT_Q 168
  234. #define DGEMM_DEFAULT_Q 168
  235. #else
  236. #define SGEMM_DEFAULT_Q 224
  237. #define DGEMM_DEFAULT_Q 224
  238. #endif
  239. #define QGEMM_DEFAULT_Q 224
  240. #define CGEMM_DEFAULT_Q 224
  241. #define ZGEMM_DEFAULT_Q 224
  242. #define XGEMM_DEFAULT_Q 224
  243. #define SGEMM_DEFAULT_R sgemm_r
  244. #define QGEMM_DEFAULT_R qgemm_r
  245. #define DGEMM_DEFAULT_R dgemm_r
  246. #define CGEMM_DEFAULT_R cgemm_r
  247. #define ZGEMM_DEFAULT_R zgemm_r
  248. #define XGEMM_DEFAULT_R xgemm_r
  249. #define SYMV_P 16
  250. #define HAVE_EXCLUSIVE_CACHE
  251. #define GEMM_THREAD gemm_thread_mn
  252. #endif
  253. #ifdef PILEDRIVER
  254. #define SNUMOPT 8
  255. #define DNUMOPT 4
  256. #define GEMM_DEFAULT_OFFSET_A 64
  257. #define GEMM_DEFAULT_OFFSET_B 832
  258. #define GEMM_DEFAULT_ALIGN 0x0fffUL
  259. #define QGEMM_DEFAULT_UNROLL_N 2
  260. #define CGEMM_DEFAULT_UNROLL_N 2
  261. #define ZGEMM_DEFAULT_UNROLL_N 2
  262. #define XGEMM_DEFAULT_UNROLL_N 1
  263. #ifdef ARCH_X86
  264. #define SGEMM_DEFAULT_UNROLL_N 4
  265. #define DGEMM_DEFAULT_UNROLL_N 4
  266. #define SGEMM_DEFAULT_UNROLL_M 4
  267. #define DGEMM_DEFAULT_UNROLL_M 2
  268. #define QGEMM_DEFAULT_UNROLL_M 2
  269. #define CGEMM_DEFAULT_UNROLL_M 2
  270. #define ZGEMM_DEFAULT_UNROLL_M 1
  271. #define XGEMM_DEFAULT_UNROLL_M 1
  272. #else
  273. #define SGEMM_DEFAULT_UNROLL_N 2
  274. #define DGEMM_DEFAULT_UNROLL_N 2
  275. #define SGEMM_DEFAULT_UNROLL_M 16
  276. #define DGEMM_DEFAULT_UNROLL_M 8
  277. #define QGEMM_DEFAULT_UNROLL_M 2
  278. #define CGEMM_DEFAULT_UNROLL_M 4
  279. #define ZGEMM_DEFAULT_UNROLL_M 2
  280. #define XGEMM_DEFAULT_UNROLL_M 1
  281. #define CGEMM3M_DEFAULT_UNROLL_N 4
  282. #define CGEMM3M_DEFAULT_UNROLL_M 8
  283. #define ZGEMM3M_DEFAULT_UNROLL_N 4
  284. #define ZGEMM3M_DEFAULT_UNROLL_M 4
  285. #define GEMV_UNROLL 8
  286. #endif
  287. #if defined(ARCH_X86_64)
  288. #define SGEMM_DEFAULT_P 768
  289. #define DGEMM_DEFAULT_P 768
  290. #define ZGEMM_DEFAULT_P 384
  291. #define CGEMM_DEFAULT_P 768
  292. #else
  293. #define SGEMM_DEFAULT_P 448
  294. #define DGEMM_DEFAULT_P 480
  295. #define ZGEMM_DEFAULT_P 112
  296. #define CGEMM_DEFAULT_P 224
  297. #endif
  298. #define QGEMM_DEFAULT_P 112
  299. #define XGEMM_DEFAULT_P 56
  300. #if defined(ARCH_X86_64)
  301. #define SGEMM_DEFAULT_Q 192
  302. #define DGEMM_DEFAULT_Q 168
  303. #define ZGEMM_DEFAULT_Q 168
  304. #define CGEMM_DEFAULT_Q 168
  305. #else
  306. #define SGEMM_DEFAULT_Q 224
  307. #define DGEMM_DEFAULT_Q 224
  308. #define ZGEMM_DEFAULT_Q 224
  309. #define CGEMM_DEFAULT_Q 224
  310. #endif
  311. #define QGEMM_DEFAULT_Q 224
  312. #define XGEMM_DEFAULT_Q 224
  313. #define SGEMM_DEFAULT_R 12288
  314. #define QGEMM_DEFAULT_R qgemm_r
  315. #define DGEMM_DEFAULT_R 12288
  316. #define CGEMM_DEFAULT_R cgemm_r
  317. #define ZGEMM_DEFAULT_R zgemm_r
  318. #define XGEMM_DEFAULT_R xgemm_r
  319. #define SYMV_P 16
  320. #define HAVE_EXCLUSIVE_CACHE
  321. #define GEMM_THREAD gemm_thread_mn
  322. #endif
  323. #ifdef ATHLON
  324. #define SNUMOPT 4
  325. #define DNUMOPT 2
  326. #define GEMM_DEFAULT_OFFSET_A 0
  327. #define GEMM_DEFAULT_OFFSET_B 384
  328. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  329. #define SGEMM_DEFAULT_UNROLL_N 4
  330. #define DGEMM_DEFAULT_UNROLL_N 4
  331. #define QGEMM_DEFAULT_UNROLL_N 2
  332. #define CGEMM_DEFAULT_UNROLL_N 2
  333. #define ZGEMM_DEFAULT_UNROLL_N 2
  334. #define XGEMM_DEFAULT_UNROLL_N 1
  335. #define SGEMM_DEFAULT_UNROLL_M 2
  336. #define DGEMM_DEFAULT_UNROLL_M 1
  337. #define QGEMM_DEFAULT_UNROLL_M 2
  338. #define CGEMM_DEFAULT_UNROLL_M 1
  339. #define ZGEMM_DEFAULT_UNROLL_M 1
  340. #define XGEMM_DEFAULT_UNROLL_M 1
  341. #define SGEMM_DEFAULT_R sgemm_r
  342. #define DGEMM_DEFAULT_R dgemm_r
  343. #define QGEMM_DEFAULT_R qgemm_r
  344. #define CGEMM_DEFAULT_R cgemm_r
  345. #define ZGEMM_DEFAULT_R zgemm_r
  346. #define XGEMM_DEFAULT_R xgemm_r
  347. #define SGEMM_DEFAULT_P 208
  348. #define DGEMM_DEFAULT_P 104
  349. #define QGEMM_DEFAULT_P 56
  350. #define CGEMM_DEFAULT_P 104
  351. #define ZGEMM_DEFAULT_P 56
  352. #define XGEMM_DEFAULT_P 28
  353. #define SGEMM_DEFAULT_Q 208
  354. #define DGEMM_DEFAULT_Q 208
  355. #define QGEMM_DEFAULT_Q 208
  356. #define CGEMM_DEFAULT_Q 208
  357. #define ZGEMM_DEFAULT_Q 208
  358. #define XGEMM_DEFAULT_Q 208
  359. #define SYMV_P 16
  360. #define HAVE_EXCLUSIVE_CACHE
  361. #endif
  362. #ifdef VIAC3
  363. #define SNUMOPT 2
  364. #define DNUMOPT 1
  365. #define GEMM_DEFAULT_OFFSET_A 0
  366. #define GEMM_DEFAULT_OFFSET_B 256
  367. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  368. #define SGEMM_DEFAULT_UNROLL_N 4
  369. #define DGEMM_DEFAULT_UNROLL_N 4
  370. #define QGEMM_DEFAULT_UNROLL_N 2
  371. #define CGEMM_DEFAULT_UNROLL_N 2
  372. #define ZGEMM_DEFAULT_UNROLL_N 2
  373. #define XGEMM_DEFAULT_UNROLL_N 1
  374. #define SGEMM_DEFAULT_UNROLL_M 2
  375. #define DGEMM_DEFAULT_UNROLL_M 1
  376. #define QGEMM_DEFAULT_UNROLL_M 2
  377. #define CGEMM_DEFAULT_UNROLL_M 1
  378. #define ZGEMM_DEFAULT_UNROLL_M 1
  379. #define XGEMM_DEFAULT_UNROLL_M 1
  380. #define SGEMM_DEFAULT_R sgemm_r
  381. #define DGEMM_DEFAULT_R dgemm_r
  382. #define QGEMM_DEFAULT_R qgemm_r
  383. #define CGEMM_DEFAULT_R cgemm_r
  384. #define ZGEMM_DEFAULT_R zgemm_r
  385. #define XGEMM_DEFAULT_R xgemm_r
  386. #define SGEMM_DEFAULT_P 128
  387. #define DGEMM_DEFAULT_P 128
  388. #define QGEMM_DEFAULT_P 128
  389. #define CGEMM_DEFAULT_P 128
  390. #define ZGEMM_DEFAULT_P 128
  391. #define XGEMM_DEFAULT_P 128
  392. #define SGEMM_DEFAULT_Q 512
  393. #define DGEMM_DEFAULT_Q 256
  394. #define QGEMM_DEFAULT_Q 256
  395. #define CGEMM_DEFAULT_Q 256
  396. #define ZGEMM_DEFAULT_Q 128
  397. #define XGEMM_DEFAULT_Q 128
  398. #define SYMV_P 16
  399. #endif
  400. #ifdef NANO
  401. #define SNUMOPT 4
  402. #define DNUMOPT 2
  403. #define GEMM_DEFAULT_OFFSET_A 64
  404. #define GEMM_DEFAULT_OFFSET_B 256
  405. #define GEMM_DEFAULT_ALIGN 0x01ffffUL
  406. #ifdef ARCH_X86
  407. #define SGEMM_DEFAULT_UNROLL_N 4
  408. #define DGEMM_DEFAULT_UNROLL_N 4
  409. #define QGEMM_DEFAULT_UNROLL_N 2
  410. #define CGEMM_DEFAULT_UNROLL_N 2
  411. #define ZGEMM_DEFAULT_UNROLL_N 2
  412. #define XGEMM_DEFAULT_UNROLL_N 1
  413. #define SGEMM_DEFAULT_UNROLL_M 4
  414. #define DGEMM_DEFAULT_UNROLL_M 2
  415. #define QGEMM_DEFAULT_UNROLL_M 2
  416. #define CGEMM_DEFAULT_UNROLL_M 2
  417. #define ZGEMM_DEFAULT_UNROLL_M 1
  418. #define XGEMM_DEFAULT_UNROLL_M 1
  419. #else
  420. #define SGEMM_DEFAULT_UNROLL_N 8
  421. #define DGEMM_DEFAULT_UNROLL_N 4
  422. #define QGEMM_DEFAULT_UNROLL_N 2
  423. #define CGEMM_DEFAULT_UNROLL_N 4
  424. #define ZGEMM_DEFAULT_UNROLL_N 2
  425. #define XGEMM_DEFAULT_UNROLL_N 1
  426. #define SGEMM_DEFAULT_UNROLL_M 4
  427. #define DGEMM_DEFAULT_UNROLL_M 4
  428. #define QGEMM_DEFAULT_UNROLL_M 2
  429. #define CGEMM_DEFAULT_UNROLL_M 2
  430. #define ZGEMM_DEFAULT_UNROLL_M 2
  431. #define XGEMM_DEFAULT_UNROLL_M 1
  432. #endif
  433. #define SGEMM_DEFAULT_P 288
  434. #define DGEMM_DEFAULT_P 288
  435. #define QGEMM_DEFAULT_P 288
  436. #define CGEMM_DEFAULT_P 288
  437. #define ZGEMM_DEFAULT_P 288
  438. #define XGEMM_DEFAULT_P 288
  439. #define SGEMM_DEFAULT_R sgemm_r
  440. #define DGEMM_DEFAULT_R dgemm_r
  441. #define QGEMM_DEFAULT_R qgemm_r
  442. #define CGEMM_DEFAULT_R cgemm_r
  443. #define ZGEMM_DEFAULT_R zgemm_r
  444. #define XGEMM_DEFAULT_R xgemm_r
  445. #define SGEMM_DEFAULT_Q 256
  446. #define DGEMM_DEFAULT_Q 128
  447. #define QGEMM_DEFAULT_Q 64
  448. #define CGEMM_DEFAULT_Q 128
  449. #define ZGEMM_DEFAULT_Q 64
  450. #define XGEMM_DEFAULT_Q 32
  451. #define SYMV_P 16
  452. #define HAVE_EXCLUSIVE_CACHE
  453. #endif
  454. #if defined(PENTIUM) || defined(PENTIUM2) || defined(PENTIUM3)
  455. #ifdef HAVE_SSE
  456. #define SNUMOPT 2
  457. #else
  458. #define SNUMOPT 1
  459. #endif
  460. #define DNUMOPT 1
  461. #define GEMM_DEFAULT_OFFSET_A 0
  462. #define GEMM_DEFAULT_OFFSET_B 0
  463. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  464. #ifdef HAVE_SSE
  465. #define SGEMM_DEFAULT_UNROLL_M 8
  466. #define CGEMM_DEFAULT_UNROLL_M 4
  467. #else
  468. #define SGEMM_DEFAULT_UNROLL_M 4
  469. #define CGEMM_DEFAULT_UNROLL_M 2
  470. #endif
  471. #define DGEMM_DEFAULT_UNROLL_M 2
  472. #define SGEMM_DEFAULT_UNROLL_N 2
  473. #define DGEMM_DEFAULT_UNROLL_N 2
  474. #define QGEMM_DEFAULT_UNROLL_M 2
  475. #define QGEMM_DEFAULT_UNROLL_N 2
  476. #define CGEMM_DEFAULT_UNROLL_N 1
  477. #define ZGEMM_DEFAULT_UNROLL_M 1
  478. #define ZGEMM_DEFAULT_UNROLL_N 1
  479. #define XGEMM_DEFAULT_UNROLL_M 1
  480. #define XGEMM_DEFAULT_UNROLL_N 1
  481. #define SGEMM_DEFAULT_P sgemm_p
  482. #define SGEMM_DEFAULT_Q 256
  483. #define SGEMM_DEFAULT_R sgemm_r
  484. #define DGEMM_DEFAULT_P dgemm_p
  485. #define DGEMM_DEFAULT_Q 256
  486. #define DGEMM_DEFAULT_R dgemm_r
  487. #define QGEMM_DEFAULT_P qgemm_p
  488. #define QGEMM_DEFAULT_Q 256
  489. #define QGEMM_DEFAULT_R qgemm_r
  490. #define CGEMM_DEFAULT_P cgemm_p
  491. #define CGEMM_DEFAULT_Q 256
  492. #define CGEMM_DEFAULT_R cgemm_r
  493. #define ZGEMM_DEFAULT_P zgemm_p
  494. #define ZGEMM_DEFAULT_Q 256
  495. #define ZGEMM_DEFAULT_R zgemm_r
  496. #define XGEMM_DEFAULT_P xgemm_p
  497. #define XGEMM_DEFAULT_Q 256
  498. #define XGEMM_DEFAULT_R xgemm_r
  499. #define SYMV_P 4
  500. #endif
  501. #ifdef PENTIUMM
  502. #define SNUMOPT 2
  503. #define DNUMOPT 1
  504. #define GEMM_DEFAULT_OFFSET_A 0
  505. #define GEMM_DEFAULT_OFFSET_B 0
  506. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  507. #ifdef CORE_YONAH
  508. #define SGEMM_DEFAULT_UNROLL_M 4
  509. #define SGEMM_DEFAULT_UNROLL_N 4
  510. #define DGEMM_DEFAULT_UNROLL_M 2
  511. #define DGEMM_DEFAULT_UNROLL_N 4
  512. #define QGEMM_DEFAULT_UNROLL_M 2
  513. #define QGEMM_DEFAULT_UNROLL_N 2
  514. #define CGEMM_DEFAULT_UNROLL_M 2
  515. #define CGEMM_DEFAULT_UNROLL_N 2
  516. #define ZGEMM_DEFAULT_UNROLL_M 1
  517. #define ZGEMM_DEFAULT_UNROLL_N 2
  518. #define XGEMM_DEFAULT_UNROLL_M 1
  519. #define XGEMM_DEFAULT_UNROLL_N 1
  520. #else
  521. #define SGEMM_DEFAULT_UNROLL_M 8
  522. #define SGEMM_DEFAULT_UNROLL_N 2
  523. #define DGEMM_DEFAULT_UNROLL_M 2
  524. #define DGEMM_DEFAULT_UNROLL_N 2
  525. #define QGEMM_DEFAULT_UNROLL_M 2
  526. #define QGEMM_DEFAULT_UNROLL_N 2
  527. #define CGEMM_DEFAULT_UNROLL_M 4
  528. #define CGEMM_DEFAULT_UNROLL_N 1
  529. #define ZGEMM_DEFAULT_UNROLL_M 1
  530. #define ZGEMM_DEFAULT_UNROLL_N 1
  531. #define XGEMM_DEFAULT_UNROLL_M 1
  532. #define XGEMM_DEFAULT_UNROLL_N 1
  533. #endif
  534. #define SGEMM_DEFAULT_P sgemm_p
  535. #define SGEMM_DEFAULT_Q 256
  536. #define SGEMM_DEFAULT_R sgemm_r
  537. #define DGEMM_DEFAULT_P dgemm_p
  538. #define DGEMM_DEFAULT_Q 256
  539. #define DGEMM_DEFAULT_R dgemm_r
  540. #define QGEMM_DEFAULT_P qgemm_p
  541. #define QGEMM_DEFAULT_Q 256
  542. #define QGEMM_DEFAULT_R qgemm_r
  543. #define CGEMM_DEFAULT_P cgemm_p
  544. #define CGEMM_DEFAULT_Q 256
  545. #define CGEMM_DEFAULT_R cgemm_r
  546. #define ZGEMM_DEFAULT_P zgemm_p
  547. #define ZGEMM_DEFAULT_Q 256
  548. #define ZGEMM_DEFAULT_R zgemm_r
  549. #define XGEMM_DEFAULT_P xgemm_p
  550. #define XGEMM_DEFAULT_Q 256
  551. #define XGEMM_DEFAULT_R xgemm_r
  552. #define SYMV_P 4
  553. #endif
  554. #ifdef CORE_NORTHWOOD
  555. #define SNUMOPT 4
  556. #define DNUMOPT 2
  557. #define GEMM_DEFAULT_OFFSET_A 0
  558. #define GEMM_DEFAULT_OFFSET_B 32
  559. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  560. #define SYMV_P 8
  561. #define SGEMM_DEFAULT_UNROLL_M 8
  562. #define DGEMM_DEFAULT_UNROLL_M 4
  563. #define QGEMM_DEFAULT_UNROLL_M 2
  564. #define CGEMM_DEFAULT_UNROLL_M 4
  565. #define ZGEMM_DEFAULT_UNROLL_M 2
  566. #define XGEMM_DEFAULT_UNROLL_M 1
  567. #define SGEMM_DEFAULT_UNROLL_N 2
  568. #define DGEMM_DEFAULT_UNROLL_N 2
  569. #define QGEMM_DEFAULT_UNROLL_N 2
  570. #define CGEMM_DEFAULT_UNROLL_N 1
  571. #define ZGEMM_DEFAULT_UNROLL_N 1
  572. #define XGEMM_DEFAULT_UNROLL_N 1
  573. #define SGEMM_DEFAULT_P sgemm_p
  574. #define SGEMM_DEFAULT_R sgemm_r
  575. #define DGEMM_DEFAULT_P dgemm_p
  576. #define DGEMM_DEFAULT_R dgemm_r
  577. #define QGEMM_DEFAULT_P qgemm_p
  578. #define QGEMM_DEFAULT_R qgemm_r
  579. #define CGEMM_DEFAULT_P cgemm_p
  580. #define CGEMM_DEFAULT_R cgemm_r
  581. #define ZGEMM_DEFAULT_P zgemm_p
  582. #define ZGEMM_DEFAULT_R zgemm_r
  583. #define XGEMM_DEFAULT_P xgemm_p
  584. #define XGEMM_DEFAULT_R xgemm_r
  585. #define SGEMM_DEFAULT_Q 128
  586. #define DGEMM_DEFAULT_Q 128
  587. #define QGEMM_DEFAULT_Q 128
  588. #define CGEMM_DEFAULT_Q 128
  589. #define ZGEMM_DEFAULT_Q 128
  590. #define XGEMM_DEFAULT_Q 128
  591. #endif
  592. #ifdef CORE_PRESCOTT
  593. #define SNUMOPT 4
  594. #define DNUMOPT 2
  595. #ifndef __64BIT__
  596. #define GEMM_DEFAULT_OFFSET_A 128
  597. #define GEMM_DEFAULT_OFFSET_B 192
  598. #else
  599. #define GEMM_DEFAULT_OFFSET_A 0
  600. #define GEMM_DEFAULT_OFFSET_B 256
  601. #endif
  602. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  603. #define SYMV_P 8
  604. #ifdef ARCH_X86
  605. #define SGEMM_DEFAULT_UNROLL_M 4
  606. #define DGEMM_DEFAULT_UNROLL_M 2
  607. #define QGEMM_DEFAULT_UNROLL_M 2
  608. #define CGEMM_DEFAULT_UNROLL_M 2
  609. #define ZGEMM_DEFAULT_UNROLL_M 1
  610. #define XGEMM_DEFAULT_UNROLL_M 1
  611. #else
  612. #define SGEMM_DEFAULT_UNROLL_M 8
  613. #define DGEMM_DEFAULT_UNROLL_M 4
  614. #define QGEMM_DEFAULT_UNROLL_M 2
  615. #define CGEMM_DEFAULT_UNROLL_M 4
  616. #define ZGEMM_DEFAULT_UNROLL_M 2
  617. #define XGEMM_DEFAULT_UNROLL_M 1
  618. #endif
  619. #define SGEMM_DEFAULT_UNROLL_N 4
  620. #define DGEMM_DEFAULT_UNROLL_N 4
  621. #define QGEMM_DEFAULT_UNROLL_N 2
  622. #define CGEMM_DEFAULT_UNROLL_N 2
  623. #define ZGEMM_DEFAULT_UNROLL_N 2
  624. #define XGEMM_DEFAULT_UNROLL_N 1
  625. #define SGEMM_DEFAULT_P sgemm_p
  626. #define SGEMM_DEFAULT_R sgemm_r
  627. #define DGEMM_DEFAULT_P dgemm_p
  628. #define DGEMM_DEFAULT_R dgemm_r
  629. #define QGEMM_DEFAULT_P qgemm_p
  630. #define QGEMM_DEFAULT_R qgemm_r
  631. #define CGEMM_DEFAULT_P cgemm_p
  632. #define CGEMM_DEFAULT_R cgemm_r
  633. #define ZGEMM_DEFAULT_P zgemm_p
  634. #define ZGEMM_DEFAULT_R zgemm_r
  635. #define XGEMM_DEFAULT_P xgemm_p
  636. #define XGEMM_DEFAULT_R xgemm_r
  637. #define SGEMM_DEFAULT_Q 128
  638. #define DGEMM_DEFAULT_Q 128
  639. #define QGEMM_DEFAULT_Q 128
  640. #define CGEMM_DEFAULT_Q 128
  641. #define ZGEMM_DEFAULT_Q 128
  642. #define XGEMM_DEFAULT_Q 128
  643. #endif
  644. #ifdef CORE2
  645. #define SNUMOPT 8
  646. #define DNUMOPT 4
  647. #define GEMM_DEFAULT_OFFSET_A 448
  648. #define GEMM_DEFAULT_OFFSET_B 128
  649. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  650. #define SYMV_P 8
  651. #define SWITCH_RATIO 4
  652. #ifdef ARCH_X86
  653. #define SGEMM_DEFAULT_UNROLL_M 8
  654. #define DGEMM_DEFAULT_UNROLL_M 4
  655. #define QGEMM_DEFAULT_UNROLL_M 2
  656. #define CGEMM_DEFAULT_UNROLL_M 4
  657. #define ZGEMM_DEFAULT_UNROLL_M 2
  658. #define XGEMM_DEFAULT_UNROLL_M 1
  659. #define SGEMM_DEFAULT_UNROLL_N 2
  660. #define DGEMM_DEFAULT_UNROLL_N 2
  661. #define QGEMM_DEFAULT_UNROLL_N 2
  662. #define CGEMM_DEFAULT_UNROLL_N 1
  663. #define ZGEMM_DEFAULT_UNROLL_N 1
  664. #define XGEMM_DEFAULT_UNROLL_N 1
  665. #define MASK(a, b) ((((a) + (b) - 1) / (b)) * (b))
  666. #else
  667. #define SGEMM_DEFAULT_UNROLL_M 8
  668. #define DGEMM_DEFAULT_UNROLL_M 4
  669. #define QGEMM_DEFAULT_UNROLL_M 2
  670. #define CGEMM_DEFAULT_UNROLL_M 4
  671. #define ZGEMM_DEFAULT_UNROLL_M 2
  672. #define XGEMM_DEFAULT_UNROLL_M 1
  673. #define SGEMM_DEFAULT_UNROLL_N 4
  674. #define DGEMM_DEFAULT_UNROLL_N 4
  675. #define QGEMM_DEFAULT_UNROLL_N 2
  676. #define CGEMM_DEFAULT_UNROLL_N 2
  677. #define ZGEMM_DEFAULT_UNROLL_N 2
  678. #define XGEMM_DEFAULT_UNROLL_N 1
  679. #endif
  680. #define SGEMM_DEFAULT_P sgemm_p
  681. #define SGEMM_DEFAULT_R sgemm_r
  682. #define DGEMM_DEFAULT_P dgemm_p
  683. #define DGEMM_DEFAULT_R dgemm_r
  684. #define QGEMM_DEFAULT_P qgemm_p
  685. #define QGEMM_DEFAULT_R qgemm_r
  686. #define CGEMM_DEFAULT_P cgemm_p
  687. #define CGEMM_DEFAULT_R cgemm_r
  688. #define ZGEMM_DEFAULT_P zgemm_p
  689. #define ZGEMM_DEFAULT_R zgemm_r
  690. #define XGEMM_DEFAULT_P xgemm_p
  691. #define XGEMM_DEFAULT_R xgemm_r
  692. #define SGEMM_DEFAULT_Q 256
  693. #define DGEMM_DEFAULT_Q 256
  694. #define QGEMM_DEFAULT_Q 256
  695. #define CGEMM_DEFAULT_Q 256
  696. #define ZGEMM_DEFAULT_Q 256
  697. #define XGEMM_DEFAULT_Q 256
  698. #endif
  699. #ifdef PENRYN
  700. #define SNUMOPT 8
  701. #define DNUMOPT 4
  702. #define GEMM_DEFAULT_OFFSET_A 128
  703. #define GEMM_DEFAULT_OFFSET_B 0
  704. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  705. #define SYMV_P 8
  706. #define SWITCH_RATIO 4
  707. #ifdef ARCH_X86
  708. #define SGEMM_DEFAULT_UNROLL_M 4
  709. #define DGEMM_DEFAULT_UNROLL_M 2
  710. #define QGEMM_DEFAULT_UNROLL_M 2
  711. #define CGEMM_DEFAULT_UNROLL_M 2
  712. #define ZGEMM_DEFAULT_UNROLL_M 1
  713. #define XGEMM_DEFAULT_UNROLL_M 1
  714. #define SGEMM_DEFAULT_UNROLL_N 4
  715. #define DGEMM_DEFAULT_UNROLL_N 4
  716. #define QGEMM_DEFAULT_UNROLL_N 2
  717. #define CGEMM_DEFAULT_UNROLL_N 2
  718. #define ZGEMM_DEFAULT_UNROLL_N 2
  719. #define XGEMM_DEFAULT_UNROLL_N 1
  720. #else
  721. #define SGEMM_DEFAULT_UNROLL_M 8
  722. #define DGEMM_DEFAULT_UNROLL_M 4
  723. #define QGEMM_DEFAULT_UNROLL_M 2
  724. #define CGEMM_DEFAULT_UNROLL_M 4
  725. #define ZGEMM_DEFAULT_UNROLL_M 2
  726. #define XGEMM_DEFAULT_UNROLL_M 1
  727. #define SGEMM_DEFAULT_UNROLL_N 4
  728. #define DGEMM_DEFAULT_UNROLL_N 4
  729. #define QGEMM_DEFAULT_UNROLL_N 2
  730. #define CGEMM_DEFAULT_UNROLL_N 2
  731. #define ZGEMM_DEFAULT_UNROLL_N 2
  732. #define XGEMM_DEFAULT_UNROLL_N 1
  733. #endif
  734. #define SGEMM_DEFAULT_P sgemm_p
  735. #define SGEMM_DEFAULT_R sgemm_r
  736. #define DGEMM_DEFAULT_P dgemm_p
  737. #define DGEMM_DEFAULT_R dgemm_r
  738. #define QGEMM_DEFAULT_P qgemm_p
  739. #define QGEMM_DEFAULT_R qgemm_r
  740. #define CGEMM_DEFAULT_P cgemm_p
  741. #define CGEMM_DEFAULT_R cgemm_r
  742. #define ZGEMM_DEFAULT_P zgemm_p
  743. #define ZGEMM_DEFAULT_R zgemm_r
  744. #define XGEMM_DEFAULT_P xgemm_p
  745. #define XGEMM_DEFAULT_R xgemm_r
  746. #define SGEMM_DEFAULT_Q 512
  747. #define DGEMM_DEFAULT_Q 256
  748. #define QGEMM_DEFAULT_Q 128
  749. #define CGEMM_DEFAULT_Q 512
  750. #define ZGEMM_DEFAULT_Q 256
  751. #define XGEMM_DEFAULT_Q 128
  752. #define GETRF_FACTOR 0.75
  753. #endif
  754. #ifdef DUNNINGTON
  755. #define SNUMOPT 8
  756. #define DNUMOPT 4
  757. #define GEMM_DEFAULT_OFFSET_A 128
  758. #define GEMM_DEFAULT_OFFSET_B 0
  759. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  760. #define SYMV_P 8
  761. #define SWITCH_RATIO 4
  762. #ifdef ARCH_X86
  763. #define SGEMM_DEFAULT_UNROLL_M 4
  764. #define DGEMM_DEFAULT_UNROLL_M 2
  765. #define QGEMM_DEFAULT_UNROLL_M 2
  766. #define CGEMM_DEFAULT_UNROLL_M 2
  767. #define ZGEMM_DEFAULT_UNROLL_M 1
  768. #define XGEMM_DEFAULT_UNROLL_M 1
  769. #define SGEMM_DEFAULT_UNROLL_N 4
  770. #define DGEMM_DEFAULT_UNROLL_N 4
  771. #define QGEMM_DEFAULT_UNROLL_N 2
  772. #define CGEMM_DEFAULT_UNROLL_N 2
  773. #define ZGEMM_DEFAULT_UNROLL_N 2
  774. #define XGEMM_DEFAULT_UNROLL_N 1
  775. #else
  776. #define SGEMM_DEFAULT_UNROLL_M 8
  777. #define DGEMM_DEFAULT_UNROLL_M 4
  778. #define QGEMM_DEFAULT_UNROLL_M 2
  779. #define CGEMM_DEFAULT_UNROLL_M 4
  780. #define ZGEMM_DEFAULT_UNROLL_M 2
  781. #define XGEMM_DEFAULT_UNROLL_M 1
  782. #define SGEMM_DEFAULT_UNROLL_N 4
  783. #define DGEMM_DEFAULT_UNROLL_N 4
  784. #define QGEMM_DEFAULT_UNROLL_N 2
  785. #define CGEMM_DEFAULT_UNROLL_N 2
  786. #define ZGEMM_DEFAULT_UNROLL_N 2
  787. #define XGEMM_DEFAULT_UNROLL_N 1
  788. #endif
  789. #define SGEMM_DEFAULT_P sgemm_p
  790. #define SGEMM_DEFAULT_R sgemm_r
  791. #define DGEMM_DEFAULT_P dgemm_p
  792. #define DGEMM_DEFAULT_R dgemm_r
  793. #define QGEMM_DEFAULT_P qgemm_p
  794. #define QGEMM_DEFAULT_R qgemm_r
  795. #define CGEMM_DEFAULT_P cgemm_p
  796. #define CGEMM_DEFAULT_R cgemm_r
  797. #define ZGEMM_DEFAULT_P zgemm_p
  798. #define ZGEMM_DEFAULT_R zgemm_r
  799. #define XGEMM_DEFAULT_P xgemm_p
  800. #define XGEMM_DEFAULT_R xgemm_r
  801. #define SGEMM_DEFAULT_Q 768
  802. #define DGEMM_DEFAULT_Q 384
  803. #define QGEMM_DEFAULT_Q 192
  804. #define CGEMM_DEFAULT_Q 768
  805. #define ZGEMM_DEFAULT_Q 384
  806. #define XGEMM_DEFAULT_Q 192
  807. #define GETRF_FACTOR 0.75
  808. #define GEMM_THREAD gemm_thread_mn
  809. #endif
  810. #ifdef NEHALEM
  811. #define SNUMOPT 8
  812. #define DNUMOPT 4
  813. #define GEMM_DEFAULT_OFFSET_A 32
  814. #define GEMM_DEFAULT_OFFSET_B 0
  815. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  816. #define SYMV_P 8
  817. #define SWITCH_RATIO 4
  818. #ifdef ARCH_X86
  819. #define SGEMM_DEFAULT_UNROLL_M 4
  820. #define DGEMM_DEFAULT_UNROLL_M 2
  821. #define QGEMM_DEFAULT_UNROLL_M 2
  822. #define CGEMM_DEFAULT_UNROLL_M 2
  823. #define ZGEMM_DEFAULT_UNROLL_M 1
  824. #define XGEMM_DEFAULT_UNROLL_M 1
  825. #define SGEMM_DEFAULT_UNROLL_N 4
  826. #define DGEMM_DEFAULT_UNROLL_N 4
  827. #define QGEMM_DEFAULT_UNROLL_N 2
  828. #define CGEMM_DEFAULT_UNROLL_N 2
  829. #define ZGEMM_DEFAULT_UNROLL_N 2
  830. #define XGEMM_DEFAULT_UNROLL_N 1
  831. #else
  832. #define SGEMM_DEFAULT_UNROLL_M 4
  833. #define DGEMM_DEFAULT_UNROLL_M 2
  834. #define QGEMM_DEFAULT_UNROLL_M 2
  835. #define CGEMM_DEFAULT_UNROLL_M 2
  836. #define ZGEMM_DEFAULT_UNROLL_M 1
  837. #define XGEMM_DEFAULT_UNROLL_M 1
  838. #define SGEMM_DEFAULT_UNROLL_N 8
  839. #define DGEMM_DEFAULT_UNROLL_N 8
  840. #define QGEMM_DEFAULT_UNROLL_N 2
  841. #define CGEMM_DEFAULT_UNROLL_N 4
  842. #define ZGEMM_DEFAULT_UNROLL_N 4
  843. #define XGEMM_DEFAULT_UNROLL_N 1
  844. #endif
  845. #define SGEMM_DEFAULT_P 504
  846. #define SGEMM_DEFAULT_R sgemm_r
  847. #define DGEMM_DEFAULT_P 504
  848. #define DGEMM_DEFAULT_R dgemm_r
  849. #define QGEMM_DEFAULT_P 504
  850. #define QGEMM_DEFAULT_R qgemm_r
  851. #define CGEMM_DEFAULT_P 252
  852. #define CGEMM_DEFAULT_R cgemm_r
  853. #define ZGEMM_DEFAULT_P 252
  854. #define ZGEMM_DEFAULT_R zgemm_r
  855. #define XGEMM_DEFAULT_P 252
  856. #define XGEMM_DEFAULT_R xgemm_r
  857. #define SGEMM_DEFAULT_Q 512
  858. #define DGEMM_DEFAULT_Q 256
  859. #define QGEMM_DEFAULT_Q 128
  860. #define CGEMM_DEFAULT_Q 512
  861. #define ZGEMM_DEFAULT_Q 256
  862. #define XGEMM_DEFAULT_Q 128
  863. #define GETRF_FACTOR 0.72
  864. #define CGEMM3M_DEFAULT_UNROLL_N 4
  865. #define CGEMM3M_DEFAULT_UNROLL_M 8
  866. #define ZGEMM3M_DEFAULT_UNROLL_N 2
  867. #define ZGEMM3M_DEFAULT_UNROLL_M 8
  868. #endif
  869. #ifdef SANDYBRIDGE
  870. #define SNUMOPT 8
  871. #define DNUMOPT 4
  872. #define GEMM_DEFAULT_OFFSET_A 0
  873. #define GEMM_DEFAULT_OFFSET_B 0
  874. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  875. #define SYMV_P 8
  876. #define SWITCH_RATIO 4
  877. #ifdef ARCH_X86
  878. #define SGEMM_DEFAULT_UNROLL_M 4
  879. #define DGEMM_DEFAULT_UNROLL_M 2
  880. #define QGEMM_DEFAULT_UNROLL_M 2
  881. #define CGEMM_DEFAULT_UNROLL_M 2
  882. #define ZGEMM_DEFAULT_UNROLL_M 1
  883. #define XGEMM_DEFAULT_UNROLL_M 1
  884. #define SGEMM_DEFAULT_UNROLL_N 4
  885. #define DGEMM_DEFAULT_UNROLL_N 4
  886. #define QGEMM_DEFAULT_UNROLL_N 2
  887. #define CGEMM_DEFAULT_UNROLL_N 2
  888. #define ZGEMM_DEFAULT_UNROLL_N 2
  889. #define XGEMM_DEFAULT_UNROLL_N 1
  890. #else
  891. #define SGEMM_DEFAULT_UNROLL_M 16
  892. #define DGEMM_DEFAULT_UNROLL_M 8
  893. #define QGEMM_DEFAULT_UNROLL_M 2
  894. #define CGEMM_DEFAULT_UNROLL_M 8
  895. #define ZGEMM_DEFAULT_UNROLL_M 4
  896. #define XGEMM_DEFAULT_UNROLL_M 1
  897. #define SGEMM_DEFAULT_UNROLL_N 4
  898. #define DGEMM_DEFAULT_UNROLL_N 4
  899. #define QGEMM_DEFAULT_UNROLL_N 2
  900. #define CGEMM_DEFAULT_UNROLL_N 2
  901. #define ZGEMM_DEFAULT_UNROLL_N 4
  902. #define XGEMM_DEFAULT_UNROLL_N 1
  903. #endif
  904. #define SGEMM_DEFAULT_P 768
  905. #define SGEMM_DEFAULT_R sgemm_r
  906. //#define SGEMM_DEFAULT_R 1024
  907. #define DGEMM_DEFAULT_P 512
  908. #define DGEMM_DEFAULT_R dgemm_r
  909. //#define DGEMM_DEFAULT_R 1024
  910. #define QGEMM_DEFAULT_P 504
  911. #define QGEMM_DEFAULT_R qgemm_r
  912. #define CGEMM_DEFAULT_P 384
  913. //#define CGEMM_DEFAULT_R cgemm_r
  914. #define CGEMM_DEFAULT_R 1024
  915. #define ZGEMM_DEFAULT_P 512
  916. #define ZGEMM_DEFAULT_R zgemm_r
  917. //#define ZGEMM_DEFAULT_R 1024
  918. #define XGEMM_DEFAULT_P 252
  919. #define XGEMM_DEFAULT_R xgemm_r
  920. #define SGEMM_DEFAULT_Q 384
  921. #define DGEMM_DEFAULT_Q 256
  922. #define QGEMM_DEFAULT_Q 128
  923. #define CGEMM_DEFAULT_Q 192
  924. #define ZGEMM_DEFAULT_Q 192
  925. #define XGEMM_DEFAULT_Q 128
  926. #define CGEMM3M_DEFAULT_UNROLL_N 4
  927. #define CGEMM3M_DEFAULT_UNROLL_M 8
  928. #define ZGEMM3M_DEFAULT_UNROLL_N 2
  929. #define ZGEMM3M_DEFAULT_UNROLL_M 8
  930. #define GETRF_FACTOR 0.72
  931. #endif
  932. #ifdef HASWELL
  933. #define SNUMOPT 16
  934. #define DNUMOPT 8
  935. #define GEMM_DEFAULT_OFFSET_A 0
  936. #define GEMM_DEFAULT_OFFSET_B 0
  937. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  938. #define SYMV_P 8
  939. #define SWITCH_RATIO 4
  940. #ifdef ARCH_X86
  941. #define SGEMM_DEFAULT_UNROLL_M 4
  942. #define DGEMM_DEFAULT_UNROLL_M 2
  943. #define QGEMM_DEFAULT_UNROLL_M 2
  944. #define CGEMM_DEFAULT_UNROLL_M 2
  945. #define ZGEMM_DEFAULT_UNROLL_M 1
  946. #define XGEMM_DEFAULT_UNROLL_M 1
  947. #define SGEMM_DEFAULT_UNROLL_N 4
  948. #define DGEMM_DEFAULT_UNROLL_N 4
  949. #define QGEMM_DEFAULT_UNROLL_N 2
  950. #define CGEMM_DEFAULT_UNROLL_N 2
  951. #define ZGEMM_DEFAULT_UNROLL_N 2
  952. #define XGEMM_DEFAULT_UNROLL_N 1
  953. #else
  954. #define SGEMM_DEFAULT_UNROLL_M 16
  955. #define DGEMM_DEFAULT_UNROLL_M 4
  956. #define QGEMM_DEFAULT_UNROLL_M 2
  957. #define CGEMM_DEFAULT_UNROLL_M 8
  958. #define ZGEMM_DEFAULT_UNROLL_M 4
  959. #define XGEMM_DEFAULT_UNROLL_M 1
  960. #define SGEMM_DEFAULT_UNROLL_N 4
  961. #define DGEMM_DEFAULT_UNROLL_N 4
  962. #define QGEMM_DEFAULT_UNROLL_N 2
  963. #define CGEMM_DEFAULT_UNROLL_N 2
  964. #define ZGEMM_DEFAULT_UNROLL_N 2
  965. #define XGEMM_DEFAULT_UNROLL_N 1
  966. #endif
  967. #ifdef ARCH_X86
  968. #define SGEMM_DEFAULT_P 512
  969. #define SGEMM_DEFAULT_R sgemm_r
  970. #define DGEMM_DEFAULT_P 512
  971. #define DGEMM_DEFAULT_R dgemm_r
  972. #define QGEMM_DEFAULT_P 504
  973. #define QGEMM_DEFAULT_R qgemm_r
  974. #define CGEMM_DEFAULT_P 128
  975. #define CGEMM_DEFAULT_R 1024
  976. #define ZGEMM_DEFAULT_P 512
  977. #define ZGEMM_DEFAULT_R zgemm_r
  978. #define XGEMM_DEFAULT_P 252
  979. #define XGEMM_DEFAULT_R xgemm_r
  980. #define SGEMM_DEFAULT_Q 256
  981. #define DGEMM_DEFAULT_Q 256
  982. #define QGEMM_DEFAULT_Q 128
  983. #define CGEMM_DEFAULT_Q 256
  984. #define ZGEMM_DEFAULT_Q 192
  985. #define XGEMM_DEFAULT_Q 128
  986. #else
  987. #define SGEMM_DEFAULT_P 768
  988. #define DGEMM_DEFAULT_P 512
  989. #define CGEMM_DEFAULT_P 384
  990. #define ZGEMM_DEFAULT_P 256
  991. #define SGEMM_DEFAULT_Q 384
  992. #ifdef WINDOWS_ABI
  993. #define DGEMM_DEFAULT_Q 128
  994. #else
  995. #define DGEMM_DEFAULT_Q 256
  996. #endif
  997. #define CGEMM_DEFAULT_Q 192
  998. #define ZGEMM_DEFAULT_Q 128
  999. #define SGEMM_DEFAULT_R sgemm_r
  1000. #define DGEMM_DEFAULT_R 13824
  1001. #define CGEMM_DEFAULT_R cgemm_r
  1002. #define ZGEMM_DEFAULT_R zgemm_r
  1003. #define QGEMM_DEFAULT_Q 128
  1004. #define QGEMM_DEFAULT_P 504
  1005. #define QGEMM_DEFAULT_R qgemm_r
  1006. #define XGEMM_DEFAULT_P 252
  1007. #define XGEMM_DEFAULT_R xgemm_r
  1008. #define XGEMM_DEFAULT_Q 128
  1009. #define CGEMM3M_DEFAULT_UNROLL_N 4
  1010. #define CGEMM3M_DEFAULT_UNROLL_M 8
  1011. #define ZGEMM3M_DEFAULT_UNROLL_N 2
  1012. #define ZGEMM3M_DEFAULT_UNROLL_M 8
  1013. #endif
  1014. #endif
  1015. #ifdef ATOM
  1016. #define SNUMOPT 2
  1017. #define DNUMOPT 1
  1018. #define GEMM_DEFAULT_OFFSET_A 64
  1019. #define GEMM_DEFAULT_OFFSET_B 0
  1020. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1021. #define SYMV_P 8
  1022. #ifdef ARCH_X86
  1023. #define SGEMM_DEFAULT_UNROLL_M 4
  1024. #define DGEMM_DEFAULT_UNROLL_M 2
  1025. #define QGEMM_DEFAULT_UNROLL_M 2
  1026. #define CGEMM_DEFAULT_UNROLL_M 2
  1027. #define ZGEMM_DEFAULT_UNROLL_M 1
  1028. #define XGEMM_DEFAULT_UNROLL_M 1
  1029. #else
  1030. #define SGEMM_DEFAULT_UNROLL_M 8
  1031. #define DGEMM_DEFAULT_UNROLL_M 4
  1032. #define QGEMM_DEFAULT_UNROLL_M 2
  1033. #define CGEMM_DEFAULT_UNROLL_M 4
  1034. #define ZGEMM_DEFAULT_UNROLL_M 2
  1035. #define XGEMM_DEFAULT_UNROLL_M 1
  1036. #endif
  1037. #define SGEMM_DEFAULT_UNROLL_N 4
  1038. #define DGEMM_DEFAULT_UNROLL_N 2
  1039. #define QGEMM_DEFAULT_UNROLL_N 2
  1040. #define CGEMM_DEFAULT_UNROLL_N 2
  1041. #define ZGEMM_DEFAULT_UNROLL_N 1
  1042. #define XGEMM_DEFAULT_UNROLL_N 1
  1043. #define SGEMM_DEFAULT_P sgemm_p
  1044. #define SGEMM_DEFAULT_R sgemm_r
  1045. #define DGEMM_DEFAULT_P dgemm_p
  1046. #define DGEMM_DEFAULT_R dgemm_r
  1047. #define QGEMM_DEFAULT_P qgemm_p
  1048. #define QGEMM_DEFAULT_R qgemm_r
  1049. #define CGEMM_DEFAULT_P cgemm_p
  1050. #define CGEMM_DEFAULT_R cgemm_r
  1051. #define ZGEMM_DEFAULT_P zgemm_p
  1052. #define ZGEMM_DEFAULT_R zgemm_r
  1053. #define XGEMM_DEFAULT_P xgemm_p
  1054. #define XGEMM_DEFAULT_R xgemm_r
  1055. #define SGEMM_DEFAULT_Q 256
  1056. #define DGEMM_DEFAULT_Q 256
  1057. #define QGEMM_DEFAULT_Q 256
  1058. #define CGEMM_DEFAULT_Q 256
  1059. #define ZGEMM_DEFAULT_Q 256
  1060. #define XGEMM_DEFAULT_Q 256
  1061. #endif
  1062. #ifdef ITANIUM2
  1063. #define SNUMOPT 4
  1064. #define DNUMOPT 4
  1065. #define GEMM_DEFAULT_OFFSET_A 0
  1066. #define GEMM_DEFAULT_OFFSET_B 128
  1067. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1068. #define SGEMM_DEFAULT_UNROLL_M 8
  1069. #define SGEMM_DEFAULT_UNROLL_N 8
  1070. #define DGEMM_DEFAULT_UNROLL_M 8
  1071. #define DGEMM_DEFAULT_UNROLL_N 8
  1072. #define QGEMM_DEFAULT_UNROLL_M 8
  1073. #define QGEMM_DEFAULT_UNROLL_N 8
  1074. #define CGEMM_DEFAULT_UNROLL_M 4
  1075. #define CGEMM_DEFAULT_UNROLL_N 4
  1076. #define ZGEMM_DEFAULT_UNROLL_M 4
  1077. #define ZGEMM_DEFAULT_UNROLL_N 4
  1078. #define XGEMM_DEFAULT_UNROLL_M 4
  1079. #define XGEMM_DEFAULT_UNROLL_N 4
  1080. #define SGEMM_DEFAULT_P sgemm_p
  1081. #define DGEMM_DEFAULT_P dgemm_p
  1082. #define QGEMM_DEFAULT_P qgemm_p
  1083. #define CGEMM_DEFAULT_P cgemm_p
  1084. #define ZGEMM_DEFAULT_P zgemm_p
  1085. #define XGEMM_DEFAULT_P xgemm_p
  1086. #define SGEMM_DEFAULT_Q 1024
  1087. #define DGEMM_DEFAULT_Q 1024
  1088. #define QGEMM_DEFAULT_Q 1024
  1089. #define CGEMM_DEFAULT_Q 1024
  1090. #define ZGEMM_DEFAULT_Q 1024
  1091. #define XGEMM_DEFAULT_Q 1024
  1092. #define SGEMM_DEFAULT_R sgemm_r
  1093. #define DGEMM_DEFAULT_R dgemm_r
  1094. #define QGEMM_DEFAULT_R qgemm_r
  1095. #define CGEMM_DEFAULT_R cgemm_r
  1096. #define ZGEMM_DEFAULT_R zgemm_r
  1097. #define XGEMM_DEFAULT_R xgemm_r
  1098. #define SYMV_P 16
  1099. #define GETRF_FACTOR 0.65
  1100. #endif
  1101. #if defined(EV4) || defined(EV5) || defined(EV6)
  1102. #ifdef EV4
  1103. #define SNUMOPT 1
  1104. #define DNUMOPT 1
  1105. #else
  1106. #define SNUMOPT 2
  1107. #define DNUMOPT 2
  1108. #endif
  1109. #define GEMM_DEFAULT_OFFSET_A 512
  1110. #define GEMM_DEFAULT_OFFSET_B 512
  1111. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1112. #define SGEMM_DEFAULT_UNROLL_M 4
  1113. #define SGEMM_DEFAULT_UNROLL_N 4
  1114. #define DGEMM_DEFAULT_UNROLL_M 4
  1115. #define DGEMM_DEFAULT_UNROLL_N 4
  1116. #define CGEMM_DEFAULT_UNROLL_M 2
  1117. #define CGEMM_DEFAULT_UNROLL_N 2
  1118. #define ZGEMM_DEFAULT_UNROLL_M 2
  1119. #define ZGEMM_DEFAULT_UNROLL_N 2
  1120. #define SYMV_P 8
  1121. #ifdef EV4
  1122. #define SGEMM_DEFAULT_P 32
  1123. #define SGEMM_DEFAULT_Q 112
  1124. #define SGEMM_DEFAULT_R 256
  1125. #define DGEMM_DEFAULT_P 32
  1126. #define DGEMM_DEFAULT_Q 56
  1127. #define DGEMM_DEFAULT_R 256
  1128. #define CGEMM_DEFAULT_P 32
  1129. #define CGEMM_DEFAULT_Q 64
  1130. #define CGEMM_DEFAULT_R 240
  1131. #define ZGEMM_DEFAULT_P 32
  1132. #define ZGEMM_DEFAULT_Q 32
  1133. #define ZGEMM_DEFAULT_R 240
  1134. #endif
  1135. #ifdef EV5
  1136. #define SGEMM_DEFAULT_P 64
  1137. #define SGEMM_DEFAULT_Q 256
  1138. #define DGEMM_DEFAULT_P 64
  1139. #define DGEMM_DEFAULT_Q 128
  1140. #define CGEMM_DEFAULT_P 64
  1141. #define CGEMM_DEFAULT_Q 128
  1142. #define ZGEMM_DEFAULT_P 64
  1143. #define ZGEMM_DEFAULT_Q 64
  1144. #endif
  1145. #ifdef EV6
  1146. #define SGEMM_DEFAULT_P 256
  1147. #define SGEMM_DEFAULT_Q 512
  1148. #define DGEMM_DEFAULT_P 256
  1149. #define DGEMM_DEFAULT_Q 256
  1150. #define CGEMM_DEFAULT_P 256
  1151. #define CGEMM_DEFAULT_Q 256
  1152. #define ZGEMM_DEFAULT_P 128
  1153. #define ZGEMM_DEFAULT_Q 256
  1154. #endif
  1155. #endif
  1156. #ifdef CELL
  1157. #define SNUMOPT 2
  1158. #define DNUMOPT 2
  1159. #define GEMM_DEFAULT_OFFSET_A 0
  1160. #define GEMM_DEFAULT_OFFSET_B 8192
  1161. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1162. #define SGEMM_DEFAULT_UNROLL_M 16
  1163. #define SGEMM_DEFAULT_UNROLL_N 4
  1164. #define DGEMM_DEFAULT_UNROLL_M 4
  1165. #define DGEMM_DEFAULT_UNROLL_N 4
  1166. #define CGEMM_DEFAULT_UNROLL_M 8
  1167. #define CGEMM_DEFAULT_UNROLL_N 2
  1168. #define ZGEMM_DEFAULT_UNROLL_M 2
  1169. #define ZGEMM_DEFAULT_UNROLL_N 2
  1170. #define SGEMM_DEFAULT_P 128
  1171. #define DGEMM_DEFAULT_P 128
  1172. #define CGEMM_DEFAULT_P 128
  1173. #define ZGEMM_DEFAULT_P 128
  1174. #define SGEMM_DEFAULT_Q 512
  1175. #define DGEMM_DEFAULT_Q 256
  1176. #define CGEMM_DEFAULT_Q 256
  1177. #define ZGEMM_DEFAULT_Q 128
  1178. #define SYMV_P 4
  1179. #endif
  1180. #ifdef PPCG4
  1181. #define GEMM_DEFAULT_OFFSET_A 0
  1182. #define GEMM_DEFAULT_OFFSET_B 1024
  1183. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1184. #define SGEMM_DEFAULT_UNROLL_M 16
  1185. #define SGEMM_DEFAULT_UNROLL_N 4
  1186. #define DGEMM_DEFAULT_UNROLL_M 4
  1187. #define DGEMM_DEFAULT_UNROLL_N 4
  1188. #define CGEMM_DEFAULT_UNROLL_M 8
  1189. #define CGEMM_DEFAULT_UNROLL_N 2
  1190. #define ZGEMM_DEFAULT_UNROLL_M 2
  1191. #define ZGEMM_DEFAULT_UNROLL_N 2
  1192. #define SGEMM_DEFAULT_P 256
  1193. #define DGEMM_DEFAULT_P 128
  1194. #define CGEMM_DEFAULT_P 128
  1195. #define ZGEMM_DEFAULT_P 64
  1196. #define SGEMM_DEFAULT_Q 256
  1197. #define DGEMM_DEFAULT_Q 256
  1198. #define CGEMM_DEFAULT_Q 256
  1199. #define ZGEMM_DEFAULT_Q 256
  1200. #define SYMV_P 4
  1201. #endif
  1202. #ifdef PPC970
  1203. #define SNUMOPT 4
  1204. #define DNUMOPT 4
  1205. #define GEMM_DEFAULT_OFFSET_A 2688
  1206. #define GEMM_DEFAULT_OFFSET_B 3072
  1207. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1208. #define SGEMM_DEFAULT_UNROLL_M 16
  1209. #define SGEMM_DEFAULT_UNROLL_N 4
  1210. #define DGEMM_DEFAULT_UNROLL_M 4
  1211. #define DGEMM_DEFAULT_UNROLL_N 4
  1212. #define CGEMM_DEFAULT_UNROLL_M 8
  1213. #define CGEMM_DEFAULT_UNROLL_N 2
  1214. #define ZGEMM_DEFAULT_UNROLL_M 2
  1215. #define ZGEMM_DEFAULT_UNROLL_N 2
  1216. #ifdef OS_LINUX
  1217. #if L2_SIZE == 1024976
  1218. #define SGEMM_DEFAULT_P 320
  1219. #define DGEMM_DEFAULT_P 256
  1220. #define CGEMM_DEFAULT_P 256
  1221. #define ZGEMM_DEFAULT_P 256
  1222. #else
  1223. #define SGEMM_DEFAULT_P 176
  1224. #define DGEMM_DEFAULT_P 176
  1225. #define CGEMM_DEFAULT_P 176
  1226. #define ZGEMM_DEFAULT_P 176
  1227. #endif
  1228. #endif
  1229. #define SGEMM_DEFAULT_Q 512
  1230. #define DGEMM_DEFAULT_Q 256
  1231. #define CGEMM_DEFAULT_Q 256
  1232. #define ZGEMM_DEFAULT_Q 128
  1233. #define SYMV_P 4
  1234. #endif
  1235. #ifdef PPC440
  1236. #define SNUMOPT 2
  1237. #define DNUMOPT 2
  1238. #define GEMM_DEFAULT_OFFSET_A (32 * 0)
  1239. #define GEMM_DEFAULT_OFFSET_B (32 * 0)
  1240. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1241. #define SGEMM_DEFAULT_UNROLL_M 4
  1242. #define SGEMM_DEFAULT_UNROLL_N 4
  1243. #define DGEMM_DEFAULT_UNROLL_M 4
  1244. #define DGEMM_DEFAULT_UNROLL_N 4
  1245. #define CGEMM_DEFAULT_UNROLL_M 2
  1246. #define CGEMM_DEFAULT_UNROLL_N 2
  1247. #define ZGEMM_DEFAULT_UNROLL_M 2
  1248. #define ZGEMM_DEFAULT_UNROLL_N 2
  1249. #define SGEMM_DEFAULT_P 512
  1250. #define DGEMM_DEFAULT_P 512
  1251. #define CGEMM_DEFAULT_P 512
  1252. #define ZGEMM_DEFAULT_P 512
  1253. #define SGEMM_DEFAULT_Q 1024
  1254. #define DGEMM_DEFAULT_Q 512
  1255. #define CGEMM_DEFAULT_Q 512
  1256. #define ZGEMM_DEFAULT_Q 256
  1257. #define SGEMM_DEFAULT_R SGEMM_DEFAULT_P
  1258. #define DGEMM_DEFAULT_R DGEMM_DEFAULT_P
  1259. #define CGEMM_DEFAULT_R CGEMM_DEFAULT_P
  1260. #define ZGEMM_DEFAULT_R ZGEMM_DEFAULT_P
  1261. #define SYMV_P 4
  1262. #endif
  1263. #ifdef PPC440FP2
  1264. #define SNUMOPT 4
  1265. #define DNUMOPT 4
  1266. #define GEMM_DEFAULT_OFFSET_A (32 * 0)
  1267. #define GEMM_DEFAULT_OFFSET_B (32 * 0)
  1268. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1269. #define SGEMM_DEFAULT_UNROLL_M 8
  1270. #define SGEMM_DEFAULT_UNROLL_N 4
  1271. #define DGEMM_DEFAULT_UNROLL_M 8
  1272. #define DGEMM_DEFAULT_UNROLL_N 4
  1273. #define CGEMM_DEFAULT_UNROLL_M 4
  1274. #define CGEMM_DEFAULT_UNROLL_N 2
  1275. #define ZGEMM_DEFAULT_UNROLL_M 4
  1276. #define ZGEMM_DEFAULT_UNROLL_N 2
  1277. #define SGEMM_DEFAULT_P 128
  1278. #define DGEMM_DEFAULT_P 128
  1279. #define CGEMM_DEFAULT_P 128
  1280. #define ZGEMM_DEFAULT_P 128
  1281. #if 1
  1282. #define SGEMM_DEFAULT_Q 4096
  1283. #define DGEMM_DEFAULT_Q 3072
  1284. #define CGEMM_DEFAULT_Q 2048
  1285. #define ZGEMM_DEFAULT_Q 1024
  1286. #else
  1287. #define SGEMM_DEFAULT_Q 512
  1288. #define DGEMM_DEFAULT_Q 256
  1289. #define CGEMM_DEFAULT_Q 256
  1290. #define ZGEMM_DEFAULT_Q 128
  1291. #endif
  1292. #define SYMV_P 4
  1293. #endif
  1294. #if defined(POWER3) || defined(POWER4) || defined(POWER5)
  1295. #define GEMM_DEFAULT_OFFSET_A 0
  1296. #define GEMM_DEFAULT_OFFSET_B 2048
  1297. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1298. #define SGEMM_DEFAULT_UNROLL_M 4
  1299. #define SGEMM_DEFAULT_UNROLL_N 4
  1300. #define DGEMM_DEFAULT_UNROLL_M 4
  1301. #define DGEMM_DEFAULT_UNROLL_N 4
  1302. #define CGEMM_DEFAULT_UNROLL_M 2
  1303. #define CGEMM_DEFAULT_UNROLL_N 2
  1304. #define ZGEMM_DEFAULT_UNROLL_M 2
  1305. #define ZGEMM_DEFAULT_UNROLL_N 2
  1306. #ifdef POWER3
  1307. #define SNUMOPT 4
  1308. #define DNUMOPT 4
  1309. #define SGEMM_DEFAULT_P 256
  1310. #define SGEMM_DEFAULT_Q 432
  1311. #define SGEMM_DEFAULT_R 1012
  1312. #define DGEMM_DEFAULT_P 256
  1313. #define DGEMM_DEFAULT_Q 216
  1314. #define DGEMM_DEFAULT_R 1012
  1315. #define ZGEMM_DEFAULT_P 256
  1316. #define ZGEMM_DEFAULT_Q 104
  1317. #define ZGEMM_DEFAULT_R 1012
  1318. #endif
  1319. #if defined(POWER4)
  1320. #ifdef ALLOC_HUGETLB
  1321. #define SGEMM_DEFAULT_P 184
  1322. #define DGEMM_DEFAULT_P 184
  1323. #define CGEMM_DEFAULT_P 184
  1324. #define ZGEMM_DEFAULT_P 184
  1325. #else
  1326. #define SGEMM_DEFAULT_P 144
  1327. #define DGEMM_DEFAULT_P 144
  1328. #define CGEMM_DEFAULT_P 144
  1329. #define ZGEMM_DEFAULT_P 144
  1330. #endif
  1331. #endif
  1332. #if defined(POWER5)
  1333. #ifdef ALLOC_HUGETLB
  1334. #define SGEMM_DEFAULT_P 512
  1335. #define DGEMM_DEFAULT_P 256
  1336. #define CGEMM_DEFAULT_P 256
  1337. #define ZGEMM_DEFAULT_P 128
  1338. #else
  1339. #define SGEMM_DEFAULT_P 320
  1340. #define DGEMM_DEFAULT_P 160
  1341. #define CGEMM_DEFAULT_P 160
  1342. #define ZGEMM_DEFAULT_P 80
  1343. #endif
  1344. #define SGEMM_DEFAULT_Q 256
  1345. #define CGEMM_DEFAULT_Q 256
  1346. #define DGEMM_DEFAULT_Q 256
  1347. #define ZGEMM_DEFAULT_Q 256
  1348. #endif
  1349. #define SYMV_P 8
  1350. #endif
  1351. #if defined(POWER6)
  1352. #define SNUMOPT 4
  1353. #define DNUMOPT 4
  1354. #define GEMM_DEFAULT_OFFSET_A 384
  1355. #define GEMM_DEFAULT_OFFSET_B 1024
  1356. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1357. #define SGEMM_DEFAULT_UNROLL_M 4
  1358. #define SGEMM_DEFAULT_UNROLL_N 4
  1359. #define DGEMM_DEFAULT_UNROLL_M 4
  1360. #define DGEMM_DEFAULT_UNROLL_N 4
  1361. #define CGEMM_DEFAULT_UNROLL_M 2
  1362. #define CGEMM_DEFAULT_UNROLL_N 4
  1363. #define ZGEMM_DEFAULT_UNROLL_M 2
  1364. #define ZGEMM_DEFAULT_UNROLL_N 4
  1365. #define SGEMM_DEFAULT_P 992
  1366. #define DGEMM_DEFAULT_P 480
  1367. #define CGEMM_DEFAULT_P 488
  1368. #define ZGEMM_DEFAULT_P 248
  1369. #define SGEMM_DEFAULT_Q 504
  1370. #define DGEMM_DEFAULT_Q 504
  1371. #define CGEMM_DEFAULT_Q 400
  1372. #define ZGEMM_DEFAULT_Q 400
  1373. #define SYMV_P 8
  1374. #endif
  1375. #if defined(SPARC) && defined(V7)
  1376. #define SNUMOPT 4
  1377. #define DNUMOPT 4
  1378. #define GEMM_DEFAULT_OFFSET_A 0
  1379. #define GEMM_DEFAULT_OFFSET_B 2048
  1380. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1381. #define SGEMM_DEFAULT_UNROLL_M 2
  1382. #define SGEMM_DEFAULT_UNROLL_N 8
  1383. #define DGEMM_DEFAULT_UNROLL_M 2
  1384. #define DGEMM_DEFAULT_UNROLL_N 8
  1385. #define CGEMM_DEFAULT_UNROLL_M 1
  1386. #define CGEMM_DEFAULT_UNROLL_N 4
  1387. #define ZGEMM_DEFAULT_UNROLL_M 1
  1388. #define ZGEMM_DEFAULT_UNROLL_N 4
  1389. #define SGEMM_DEFAULT_P 256
  1390. #define DGEMM_DEFAULT_P 256
  1391. #define CGEMM_DEFAULT_P 256
  1392. #define ZGEMM_DEFAULT_P 256
  1393. #define SGEMM_DEFAULT_Q 512
  1394. #define DGEMM_DEFAULT_Q 256
  1395. #define CGEMM_DEFAULT_Q 256
  1396. #define ZGEMM_DEFAULT_Q 128
  1397. #define SYMV_P 8
  1398. #define GEMM_THREAD gemm_thread_mn
  1399. #endif
  1400. #if (defined(SPARC) && defined(V9)) || defined(__sparc_v9__)
  1401. #define SNUMOPT 2
  1402. #define DNUMOPT 2
  1403. #define GEMM_DEFAULT_OFFSET_A 0
  1404. #define GEMM_DEFAULT_OFFSET_B 2048
  1405. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1406. #define SGEMM_DEFAULT_UNROLL_M 4
  1407. #define SGEMM_DEFAULT_UNROLL_N 4
  1408. #define DGEMM_DEFAULT_UNROLL_M 4
  1409. #define DGEMM_DEFAULT_UNROLL_N 4
  1410. #define CGEMM_DEFAULT_UNROLL_M 2
  1411. #define CGEMM_DEFAULT_UNROLL_N 2
  1412. #define ZGEMM_DEFAULT_UNROLL_M 2
  1413. #define ZGEMM_DEFAULT_UNROLL_N 2
  1414. #define SGEMM_DEFAULT_P 512
  1415. #define DGEMM_DEFAULT_P 512
  1416. #define CGEMM_DEFAULT_P 512
  1417. #define ZGEMM_DEFAULT_P 512
  1418. #define SGEMM_DEFAULT_Q 1024
  1419. #define DGEMM_DEFAULT_Q 512
  1420. #define CGEMM_DEFAULT_Q 512
  1421. #define ZGEMM_DEFAULT_Q 256
  1422. #define SYMV_P 8
  1423. #endif
  1424. #ifdef SICORTEX
  1425. #define SNUMOPT 2
  1426. #define DNUMOPT 2
  1427. #define GEMM_DEFAULT_OFFSET_A 0
  1428. #define GEMM_DEFAULT_OFFSET_B 0
  1429. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1430. #define SGEMM_DEFAULT_UNROLL_M 2
  1431. #define SGEMM_DEFAULT_UNROLL_N 8
  1432. #define DGEMM_DEFAULT_UNROLL_M 2
  1433. #define DGEMM_DEFAULT_UNROLL_N 8
  1434. #define CGEMM_DEFAULT_UNROLL_M 1
  1435. #define CGEMM_DEFAULT_UNROLL_N 4
  1436. #define ZGEMM_DEFAULT_UNROLL_M 1
  1437. #define ZGEMM_DEFAULT_UNROLL_N 4
  1438. #define SGEMM_DEFAULT_P 108
  1439. #define DGEMM_DEFAULT_P 112
  1440. #define CGEMM_DEFAULT_P 108
  1441. #define ZGEMM_DEFAULT_P 112
  1442. #define SGEMM_DEFAULT_Q 288
  1443. #define DGEMM_DEFAULT_Q 144
  1444. #define CGEMM_DEFAULT_Q 144
  1445. #define ZGEMM_DEFAULT_Q 72
  1446. #define SGEMM_DEFAULT_R 2000
  1447. #define DGEMM_DEFAULT_R 2000
  1448. #define CGEMM_DEFAULT_R 2000
  1449. #define ZGEMM_DEFAULT_R 2000
  1450. #define SYMV_P 16
  1451. #endif
  1452. #ifdef LOONGSON3A
  1453. ////Copy from SICORTEX
  1454. #define SNUMOPT 2
  1455. #define DNUMOPT 2
  1456. #define GEMM_DEFAULT_OFFSET_A 0
  1457. #define GEMM_DEFAULT_OFFSET_B 0
  1458. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1459. #define SGEMM_DEFAULT_UNROLL_M 8
  1460. #define SGEMM_DEFAULT_UNROLL_N 4
  1461. #define DGEMM_DEFAULT_UNROLL_M 4
  1462. #define DGEMM_DEFAULT_UNROLL_N 4
  1463. #define CGEMM_DEFAULT_UNROLL_M 4
  1464. #define CGEMM_DEFAULT_UNROLL_N 2
  1465. #define ZGEMM_DEFAULT_UNROLL_M 2
  1466. #define ZGEMM_DEFAULT_UNROLL_N 2
  1467. #define SGEMM_DEFAULT_P 64
  1468. #define DGEMM_DEFAULT_P 44
  1469. #define CGEMM_DEFAULT_P 64
  1470. #define ZGEMM_DEFAULT_P 32
  1471. #define SGEMM_DEFAULT_Q 192
  1472. #define DGEMM_DEFAULT_Q 92
  1473. #define CGEMM_DEFAULT_Q 128
  1474. #define ZGEMM_DEFAULT_Q 80
  1475. #define SGEMM_DEFAULT_R 640
  1476. #define DGEMM_DEFAULT_R dgemm_r
  1477. #define CGEMM_DEFAULT_R 640
  1478. #define ZGEMM_DEFAULT_R 640
  1479. #define GEMM_OFFSET_A1 0x10000
  1480. #define GEMM_OFFSET_B1 0x100000
  1481. #define SYMV_P 16
  1482. #endif
  1483. #ifdef LOONGSON3B
  1484. #define SNUMOPT 2
  1485. #define DNUMOPT 2
  1486. #define GEMM_DEFAULT_OFFSET_A 0
  1487. #define GEMM_DEFAULT_OFFSET_B 0
  1488. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1489. #define SGEMM_DEFAULT_UNROLL_M 2
  1490. #define SGEMM_DEFAULT_UNROLL_N 2
  1491. #define DGEMM_DEFAULT_UNROLL_M 2
  1492. #define DGEMM_DEFAULT_UNROLL_N 2
  1493. #define CGEMM_DEFAULT_UNROLL_M 2
  1494. #define CGEMM_DEFAULT_UNROLL_N 2
  1495. #define ZGEMM_DEFAULT_UNROLL_M 2
  1496. #define ZGEMM_DEFAULT_UNROLL_N 2
  1497. #define SGEMM_DEFAULT_P 64
  1498. #define DGEMM_DEFAULT_P 24
  1499. #define CGEMM_DEFAULT_P 24
  1500. #define ZGEMM_DEFAULT_P 20
  1501. #define SGEMM_DEFAULT_Q 192
  1502. #define DGEMM_DEFAULT_Q 128
  1503. #define CGEMM_DEFAULT_Q 128
  1504. #define ZGEMM_DEFAULT_Q 64
  1505. #define SGEMM_DEFAULT_R 512
  1506. #define DGEMM_DEFAULT_R 512
  1507. #define CGEMM_DEFAULT_R 512
  1508. #define ZGEMM_DEFAULT_R 512
  1509. #define GEMM_OFFSET_A1 0x10000
  1510. #define GEMM_OFFSET_B1 0x100000
  1511. #define SYMV_P 16
  1512. #endif
  1513. #ifdef ARMV7
  1514. #define SNUMOPT 2
  1515. #define DNUMOPT 2
  1516. #define GEMM_DEFAULT_OFFSET_A 0
  1517. #define GEMM_DEFAULT_OFFSET_B 0
  1518. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1519. #define SGEMM_DEFAULT_UNROLL_M 4
  1520. #define SGEMM_DEFAULT_UNROLL_N 4
  1521. #define DGEMM_DEFAULT_UNROLL_M 4
  1522. #define DGEMM_DEFAULT_UNROLL_N 4
  1523. #define CGEMM_DEFAULT_UNROLL_M 2
  1524. #define CGEMM_DEFAULT_UNROLL_N 2
  1525. #define ZGEMM_DEFAULT_UNROLL_M 2
  1526. #define ZGEMM_DEFAULT_UNROLL_N 2
  1527. #define SGEMM_DEFAULT_P 128
  1528. #define DGEMM_DEFAULT_P 128
  1529. #define CGEMM_DEFAULT_P 96
  1530. #define ZGEMM_DEFAULT_P 64
  1531. #define SGEMM_DEFAULT_Q 240
  1532. #define DGEMM_DEFAULT_Q 120
  1533. #define CGEMM_DEFAULT_Q 120
  1534. #define ZGEMM_DEFAULT_Q 120
  1535. #define SGEMM_DEFAULT_R 12288
  1536. #define DGEMM_DEFAULT_R 8192
  1537. #define CGEMM_DEFAULT_R 4096
  1538. #define ZGEMM_DEFAULT_R 4096
  1539. #define SYMV_P 16
  1540. #endif
  1541. #if defined(ARMV6)
  1542. #define SNUMOPT 2
  1543. #define DNUMOPT 2
  1544. #define GEMM_DEFAULT_OFFSET_A 0
  1545. #define GEMM_DEFAULT_OFFSET_B 0
  1546. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1547. #define SGEMM_DEFAULT_UNROLL_M 4
  1548. #define SGEMM_DEFAULT_UNROLL_N 2
  1549. #define DGEMM_DEFAULT_UNROLL_M 4
  1550. #define DGEMM_DEFAULT_UNROLL_N 2
  1551. #define CGEMM_DEFAULT_UNROLL_M 2
  1552. #define CGEMM_DEFAULT_UNROLL_N 2
  1553. #define ZGEMM_DEFAULT_UNROLL_M 2
  1554. #define ZGEMM_DEFAULT_UNROLL_N 2
  1555. #define SGEMM_DEFAULT_P 128
  1556. #define DGEMM_DEFAULT_P 128
  1557. #define CGEMM_DEFAULT_P 96
  1558. #define ZGEMM_DEFAULT_P 64
  1559. #define SGEMM_DEFAULT_Q 240
  1560. #define DGEMM_DEFAULT_Q 120
  1561. #define CGEMM_DEFAULT_Q 120
  1562. #define ZGEMM_DEFAULT_Q 120
  1563. #define SGEMM_DEFAULT_R 12288
  1564. #define DGEMM_DEFAULT_R 8192
  1565. #define CGEMM_DEFAULT_R 4096
  1566. #define ZGEMM_DEFAULT_R 4096
  1567. #define SYMV_P 16
  1568. #endif
  1569. #if defined(ARMV8)
  1570. #define SNUMOPT 2
  1571. #define DNUMOPT 2
  1572. #define GEMM_DEFAULT_OFFSET_A 0
  1573. #define GEMM_DEFAULT_OFFSET_B 0
  1574. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1575. #define SGEMM_DEFAULT_UNROLL_M 2
  1576. #define SGEMM_DEFAULT_UNROLL_N 2
  1577. #define DGEMM_DEFAULT_UNROLL_M 2
  1578. #define DGEMM_DEFAULT_UNROLL_N 2
  1579. #define CGEMM_DEFAULT_UNROLL_M 2
  1580. #define CGEMM_DEFAULT_UNROLL_N 2
  1581. #define ZGEMM_DEFAULT_UNROLL_M 2
  1582. #define ZGEMM_DEFAULT_UNROLL_N 2
  1583. #define SGEMM_DEFAULT_P 128
  1584. #define DGEMM_DEFAULT_P 128
  1585. #define CGEMM_DEFAULT_P 96
  1586. #define ZGEMM_DEFAULT_P 64
  1587. #define SGEMM_DEFAULT_Q 240
  1588. #define DGEMM_DEFAULT_Q 120
  1589. #define CGEMM_DEFAULT_Q 120
  1590. #define ZGEMM_DEFAULT_Q 120
  1591. #define SGEMM_DEFAULT_R 12288
  1592. #define DGEMM_DEFAULT_R 8192
  1593. #define CGEMM_DEFAULT_R 4096
  1594. #define ZGEMM_DEFAULT_R 4096
  1595. #define SYMV_P 16
  1596. #endif
  1597. #if defined(ARMV5)
  1598. #define SNUMOPT 2
  1599. #define DNUMOPT 2
  1600. #define GEMM_DEFAULT_OFFSET_A 0
  1601. #define GEMM_DEFAULT_OFFSET_B 0
  1602. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1603. #define SGEMM_DEFAULT_UNROLL_M 2
  1604. #define SGEMM_DEFAULT_UNROLL_N 2
  1605. #define DGEMM_DEFAULT_UNROLL_M 2
  1606. #define DGEMM_DEFAULT_UNROLL_N 2
  1607. #define CGEMM_DEFAULT_UNROLL_M 2
  1608. #define CGEMM_DEFAULT_UNROLL_N 2
  1609. #define ZGEMM_DEFAULT_UNROLL_M 2
  1610. #define ZGEMM_DEFAULT_UNROLL_N 2
  1611. #define SGEMM_DEFAULT_P 128
  1612. #define DGEMM_DEFAULT_P 128
  1613. #define CGEMM_DEFAULT_P 96
  1614. #define ZGEMM_DEFAULT_P 64
  1615. #define SGEMM_DEFAULT_Q 240
  1616. #define DGEMM_DEFAULT_Q 120
  1617. #define CGEMM_DEFAULT_Q 120
  1618. #define ZGEMM_DEFAULT_Q 120
  1619. #define SGEMM_DEFAULT_R 12288
  1620. #define DGEMM_DEFAULT_R 8192
  1621. #define CGEMM_DEFAULT_R 4096
  1622. #define ZGEMM_DEFAULT_R 4096
  1623. #define SYMV_P 16
  1624. #endif
  1625. #ifdef GENERIC
  1626. #define SNUMOPT 2
  1627. #define DNUMOPT 2
  1628. #define GEMM_DEFAULT_OFFSET_A 0
  1629. #define GEMM_DEFAULT_OFFSET_B 0
  1630. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1631. #define SGEMM_DEFAULT_UNROLL_N 4
  1632. #define DGEMM_DEFAULT_UNROLL_N 4
  1633. #define QGEMM_DEFAULT_UNROLL_N 2
  1634. #define CGEMM_DEFAULT_UNROLL_N 2
  1635. #define ZGEMM_DEFAULT_UNROLL_N 2
  1636. #define XGEMM_DEFAULT_UNROLL_N 1
  1637. #ifdef ARCH_X86
  1638. #define SGEMM_DEFAULT_UNROLL_M 4
  1639. #define DGEMM_DEFAULT_UNROLL_M 2
  1640. #define QGEMM_DEFAULT_UNROLL_M 2
  1641. #define CGEMM_DEFAULT_UNROLL_M 2
  1642. #define ZGEMM_DEFAULT_UNROLL_M 1
  1643. #define XGEMM_DEFAULT_UNROLL_M 1
  1644. #else
  1645. #define SGEMM_DEFAULT_UNROLL_M 8
  1646. #define DGEMM_DEFAULT_UNROLL_M 4
  1647. #define QGEMM_DEFAULT_UNROLL_M 2
  1648. #define CGEMM_DEFAULT_UNROLL_M 4
  1649. #define ZGEMM_DEFAULT_UNROLL_M 2
  1650. #define XGEMM_DEFAULT_UNROLL_M 1
  1651. #endif
  1652. #define SGEMM_DEFAULT_P sgemm_p
  1653. #define DGEMM_DEFAULT_P dgemm_p
  1654. #define QGEMM_DEFAULT_P qgemm_p
  1655. #define CGEMM_DEFAULT_P cgemm_p
  1656. #define ZGEMM_DEFAULT_P zgemm_p
  1657. #define XGEMM_DEFAULT_P xgemm_p
  1658. #define SGEMM_DEFAULT_R sgemm_r
  1659. #define DGEMM_DEFAULT_R dgemm_r
  1660. #define QGEMM_DEFAULT_R qgemm_r
  1661. #define CGEMM_DEFAULT_R cgemm_r
  1662. #define ZGEMM_DEFAULT_R zgemm_r
  1663. #define XGEMM_DEFAULT_R xgemm_r
  1664. #define SGEMM_DEFAULT_Q 128
  1665. #define DGEMM_DEFAULT_Q 128
  1666. #define QGEMM_DEFAULT_Q 128
  1667. #define CGEMM_DEFAULT_Q 128
  1668. #define ZGEMM_DEFAULT_Q 128
  1669. #define XGEMM_DEFAULT_Q 128
  1670. #define SYMV_P 16
  1671. #endif
  1672. #ifndef QGEMM_DEFAULT_UNROLL_M
  1673. #define QGEMM_DEFAULT_UNROLL_M 2
  1674. #endif
  1675. #ifndef QGEMM_DEFAULT_UNROLL_N
  1676. #define QGEMM_DEFAULT_UNROLL_N 2
  1677. #endif
  1678. #ifndef XGEMM_DEFAULT_UNROLL_M
  1679. #define XGEMM_DEFAULT_UNROLL_M 2
  1680. #endif
  1681. #ifndef XGEMM_DEFAULT_UNROLL_N
  1682. #define XGEMM_DEFAULT_UNROLL_N 2
  1683. #endif
  1684. #ifndef HAVE_SSE2
  1685. #define SHUFPD_0 shufps $0x44,
  1686. #define SHUFPD_1 shufps $0x4e,
  1687. #define SHUFPD_2 shufps $0xe4,
  1688. #define SHUFPD_3 shufps $0xee,
  1689. #endif
  1690. #ifndef SHUFPD_0
  1691. #define SHUFPD_0 shufpd $0,
  1692. #endif
  1693. #ifndef SHUFPD_1
  1694. #define SHUFPD_1 shufpd $1,
  1695. #endif
  1696. #ifndef SHUFPD_2
  1697. #define SHUFPD_2 shufpd $2,
  1698. #endif
  1699. #ifndef SHUFPD_3
  1700. #define SHUFPD_3 shufpd $3,
  1701. #endif
  1702. #ifndef SHUFPS_39
  1703. #define SHUFPS_39 shufps $0x39,
  1704. #endif
  1705. #endif