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 45 kB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868186918701871187218731874187518761877187818791880188118821883188418851886188718881889189018911892189318941895189618971898189919001901
  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. #endif
  865. #ifdef SANDYBRIDGE
  866. #define SNUMOPT 8
  867. #define DNUMOPT 4
  868. #define GEMM_DEFAULT_OFFSET_A 0
  869. #define GEMM_DEFAULT_OFFSET_B 0
  870. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  871. #define SYMV_P 8
  872. #define SWITCH_RATIO 4
  873. #ifdef ARCH_X86
  874. #define SGEMM_DEFAULT_UNROLL_M 4
  875. #define DGEMM_DEFAULT_UNROLL_M 2
  876. #define QGEMM_DEFAULT_UNROLL_M 2
  877. #define CGEMM_DEFAULT_UNROLL_M 2
  878. #define ZGEMM_DEFAULT_UNROLL_M 1
  879. #define XGEMM_DEFAULT_UNROLL_M 1
  880. #define SGEMM_DEFAULT_UNROLL_N 4
  881. #define DGEMM_DEFAULT_UNROLL_N 4
  882. #define QGEMM_DEFAULT_UNROLL_N 2
  883. #define CGEMM_DEFAULT_UNROLL_N 2
  884. #define ZGEMM_DEFAULT_UNROLL_N 2
  885. #define XGEMM_DEFAULT_UNROLL_N 1
  886. #else
  887. #define SGEMM_DEFAULT_UNROLL_M 8
  888. #define DGEMM_DEFAULT_UNROLL_M 8
  889. #define QGEMM_DEFAULT_UNROLL_M 2
  890. #define CGEMM_DEFAULT_UNROLL_M 8
  891. #define ZGEMM_DEFAULT_UNROLL_M 4
  892. #define XGEMM_DEFAULT_UNROLL_M 1
  893. #define SGEMM_DEFAULT_UNROLL_N 8
  894. #define DGEMM_DEFAULT_UNROLL_N 4
  895. #define QGEMM_DEFAULT_UNROLL_N 2
  896. #define CGEMM_DEFAULT_UNROLL_N 4
  897. #define ZGEMM_DEFAULT_UNROLL_N 4
  898. #define XGEMM_DEFAULT_UNROLL_N 1
  899. #endif
  900. #define SGEMM_DEFAULT_P 512
  901. #define SGEMM_DEFAULT_R sgemm_r
  902. //#define SGEMM_DEFAULT_R 1024
  903. #define DGEMM_DEFAULT_P 512
  904. #define DGEMM_DEFAULT_R dgemm_r
  905. //#define DGEMM_DEFAULT_R 1024
  906. #define QGEMM_DEFAULT_P 504
  907. #define QGEMM_DEFAULT_R qgemm_r
  908. #define CGEMM_DEFAULT_P 128
  909. //#define CGEMM_DEFAULT_R cgemm_r
  910. #define CGEMM_DEFAULT_R 1024
  911. #define ZGEMM_DEFAULT_P 512
  912. #define ZGEMM_DEFAULT_R zgemm_r
  913. //#define ZGEMM_DEFAULT_R 1024
  914. #define XGEMM_DEFAULT_P 252
  915. #define XGEMM_DEFAULT_R xgemm_r
  916. #define SGEMM_DEFAULT_Q 256
  917. #define DGEMM_DEFAULT_Q 256
  918. #define QGEMM_DEFAULT_Q 128
  919. #define CGEMM_DEFAULT_Q 256
  920. #define ZGEMM_DEFAULT_Q 192
  921. #define XGEMM_DEFAULT_Q 128
  922. #define GETRF_FACTOR 0.72
  923. #endif
  924. #ifdef ATOM
  925. #define SNUMOPT 2
  926. #define DNUMOPT 1
  927. #define GEMM_DEFAULT_OFFSET_A 64
  928. #define GEMM_DEFAULT_OFFSET_B 0
  929. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  930. #define SYMV_P 8
  931. #ifdef ARCH_X86
  932. #define SGEMM_DEFAULT_UNROLL_M 4
  933. #define DGEMM_DEFAULT_UNROLL_M 2
  934. #define QGEMM_DEFAULT_UNROLL_M 2
  935. #define CGEMM_DEFAULT_UNROLL_M 2
  936. #define ZGEMM_DEFAULT_UNROLL_M 1
  937. #define XGEMM_DEFAULT_UNROLL_M 1
  938. #else
  939. #define SGEMM_DEFAULT_UNROLL_M 8
  940. #define DGEMM_DEFAULT_UNROLL_M 4
  941. #define QGEMM_DEFAULT_UNROLL_M 2
  942. #define CGEMM_DEFAULT_UNROLL_M 4
  943. #define ZGEMM_DEFAULT_UNROLL_M 2
  944. #define XGEMM_DEFAULT_UNROLL_M 1
  945. #endif
  946. #define SGEMM_DEFAULT_UNROLL_N 4
  947. #define DGEMM_DEFAULT_UNROLL_N 2
  948. #define QGEMM_DEFAULT_UNROLL_N 2
  949. #define CGEMM_DEFAULT_UNROLL_N 2
  950. #define ZGEMM_DEFAULT_UNROLL_N 1
  951. #define XGEMM_DEFAULT_UNROLL_N 1
  952. #define SGEMM_DEFAULT_P sgemm_p
  953. #define SGEMM_DEFAULT_R sgemm_r
  954. #define DGEMM_DEFAULT_P dgemm_p
  955. #define DGEMM_DEFAULT_R dgemm_r
  956. #define QGEMM_DEFAULT_P qgemm_p
  957. #define QGEMM_DEFAULT_R qgemm_r
  958. #define CGEMM_DEFAULT_P cgemm_p
  959. #define CGEMM_DEFAULT_R cgemm_r
  960. #define ZGEMM_DEFAULT_P zgemm_p
  961. #define ZGEMM_DEFAULT_R zgemm_r
  962. #define XGEMM_DEFAULT_P xgemm_p
  963. #define XGEMM_DEFAULT_R xgemm_r
  964. #define SGEMM_DEFAULT_Q 256
  965. #define DGEMM_DEFAULT_Q 256
  966. #define QGEMM_DEFAULT_Q 256
  967. #define CGEMM_DEFAULT_Q 256
  968. #define ZGEMM_DEFAULT_Q 256
  969. #define XGEMM_DEFAULT_Q 256
  970. #endif
  971. #ifdef ITANIUM2
  972. #define SNUMOPT 4
  973. #define DNUMOPT 4
  974. #define GEMM_DEFAULT_OFFSET_A 0
  975. #define GEMM_DEFAULT_OFFSET_B 128
  976. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  977. #define SGEMM_DEFAULT_UNROLL_M 8
  978. #define SGEMM_DEFAULT_UNROLL_N 8
  979. #define DGEMM_DEFAULT_UNROLL_M 8
  980. #define DGEMM_DEFAULT_UNROLL_N 8
  981. #define QGEMM_DEFAULT_UNROLL_M 8
  982. #define QGEMM_DEFAULT_UNROLL_N 8
  983. #define CGEMM_DEFAULT_UNROLL_M 4
  984. #define CGEMM_DEFAULT_UNROLL_N 4
  985. #define ZGEMM_DEFAULT_UNROLL_M 4
  986. #define ZGEMM_DEFAULT_UNROLL_N 4
  987. #define XGEMM_DEFAULT_UNROLL_M 4
  988. #define XGEMM_DEFAULT_UNROLL_N 4
  989. #define SGEMM_DEFAULT_P sgemm_p
  990. #define DGEMM_DEFAULT_P dgemm_p
  991. #define QGEMM_DEFAULT_P qgemm_p
  992. #define CGEMM_DEFAULT_P cgemm_p
  993. #define ZGEMM_DEFAULT_P zgemm_p
  994. #define XGEMM_DEFAULT_P xgemm_p
  995. #define SGEMM_DEFAULT_Q 1024
  996. #define DGEMM_DEFAULT_Q 1024
  997. #define QGEMM_DEFAULT_Q 1024
  998. #define CGEMM_DEFAULT_Q 1024
  999. #define ZGEMM_DEFAULT_Q 1024
  1000. #define XGEMM_DEFAULT_Q 1024
  1001. #define SGEMM_DEFAULT_R sgemm_r
  1002. #define DGEMM_DEFAULT_R dgemm_r
  1003. #define QGEMM_DEFAULT_R qgemm_r
  1004. #define CGEMM_DEFAULT_R cgemm_r
  1005. #define ZGEMM_DEFAULT_R zgemm_r
  1006. #define XGEMM_DEFAULT_R xgemm_r
  1007. #define SYMV_P 16
  1008. #define GETRF_FACTOR 0.65
  1009. #endif
  1010. #if defined(EV4) || defined(EV5) || defined(EV6)
  1011. #ifdef EV4
  1012. #define SNUMOPT 1
  1013. #define DNUMOPT 1
  1014. #else
  1015. #define SNUMOPT 2
  1016. #define DNUMOPT 2
  1017. #endif
  1018. #define GEMM_DEFAULT_OFFSET_A 512
  1019. #define GEMM_DEFAULT_OFFSET_B 512
  1020. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1021. #define SGEMM_DEFAULT_UNROLL_M 4
  1022. #define SGEMM_DEFAULT_UNROLL_N 4
  1023. #define DGEMM_DEFAULT_UNROLL_M 4
  1024. #define DGEMM_DEFAULT_UNROLL_N 4
  1025. #define CGEMM_DEFAULT_UNROLL_M 2
  1026. #define CGEMM_DEFAULT_UNROLL_N 2
  1027. #define ZGEMM_DEFAULT_UNROLL_M 2
  1028. #define ZGEMM_DEFAULT_UNROLL_N 2
  1029. #define SYMV_P 8
  1030. #ifdef EV4
  1031. #define SGEMM_DEFAULT_P 32
  1032. #define SGEMM_DEFAULT_Q 112
  1033. #define SGEMM_DEFAULT_R 256
  1034. #define DGEMM_DEFAULT_P 32
  1035. #define DGEMM_DEFAULT_Q 56
  1036. #define DGEMM_DEFAULT_R 256
  1037. #define CGEMM_DEFAULT_P 32
  1038. #define CGEMM_DEFAULT_Q 64
  1039. #define CGEMM_DEFAULT_R 240
  1040. #define ZGEMM_DEFAULT_P 32
  1041. #define ZGEMM_DEFAULT_Q 32
  1042. #define ZGEMM_DEFAULT_R 240
  1043. #endif
  1044. #ifdef EV5
  1045. #define SGEMM_DEFAULT_P 64
  1046. #define SGEMM_DEFAULT_Q 256
  1047. #define DGEMM_DEFAULT_P 64
  1048. #define DGEMM_DEFAULT_Q 128
  1049. #define CGEMM_DEFAULT_P 64
  1050. #define CGEMM_DEFAULT_Q 128
  1051. #define ZGEMM_DEFAULT_P 64
  1052. #define ZGEMM_DEFAULT_Q 64
  1053. #endif
  1054. #ifdef EV6
  1055. #define SGEMM_DEFAULT_P 256
  1056. #define SGEMM_DEFAULT_Q 512
  1057. #define DGEMM_DEFAULT_P 256
  1058. #define DGEMM_DEFAULT_Q 256
  1059. #define CGEMM_DEFAULT_P 256
  1060. #define CGEMM_DEFAULT_Q 256
  1061. #define ZGEMM_DEFAULT_P 128
  1062. #define ZGEMM_DEFAULT_Q 256
  1063. #endif
  1064. #endif
  1065. #ifdef CELL
  1066. #define SNUMOPT 2
  1067. #define DNUMOPT 2
  1068. #define GEMM_DEFAULT_OFFSET_A 0
  1069. #define GEMM_DEFAULT_OFFSET_B 8192
  1070. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1071. #define SGEMM_DEFAULT_UNROLL_M 16
  1072. #define SGEMM_DEFAULT_UNROLL_N 4
  1073. #define DGEMM_DEFAULT_UNROLL_M 4
  1074. #define DGEMM_DEFAULT_UNROLL_N 4
  1075. #define CGEMM_DEFAULT_UNROLL_M 8
  1076. #define CGEMM_DEFAULT_UNROLL_N 2
  1077. #define ZGEMM_DEFAULT_UNROLL_M 2
  1078. #define ZGEMM_DEFAULT_UNROLL_N 2
  1079. #define SGEMM_DEFAULT_P 128
  1080. #define DGEMM_DEFAULT_P 128
  1081. #define CGEMM_DEFAULT_P 128
  1082. #define ZGEMM_DEFAULT_P 128
  1083. #define SGEMM_DEFAULT_Q 512
  1084. #define DGEMM_DEFAULT_Q 256
  1085. #define CGEMM_DEFAULT_Q 256
  1086. #define ZGEMM_DEFAULT_Q 128
  1087. #define SYMV_P 4
  1088. #endif
  1089. #ifdef PPCG4
  1090. #define GEMM_DEFAULT_OFFSET_A 0
  1091. #define GEMM_DEFAULT_OFFSET_B 1024
  1092. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1093. #define SGEMM_DEFAULT_UNROLL_M 16
  1094. #define SGEMM_DEFAULT_UNROLL_N 4
  1095. #define DGEMM_DEFAULT_UNROLL_M 4
  1096. #define DGEMM_DEFAULT_UNROLL_N 4
  1097. #define CGEMM_DEFAULT_UNROLL_M 8
  1098. #define CGEMM_DEFAULT_UNROLL_N 2
  1099. #define ZGEMM_DEFAULT_UNROLL_M 2
  1100. #define ZGEMM_DEFAULT_UNROLL_N 2
  1101. #define SGEMM_DEFAULT_P 256
  1102. #define DGEMM_DEFAULT_P 128
  1103. #define CGEMM_DEFAULT_P 128
  1104. #define ZGEMM_DEFAULT_P 64
  1105. #define SGEMM_DEFAULT_Q 256
  1106. #define DGEMM_DEFAULT_Q 256
  1107. #define CGEMM_DEFAULT_Q 256
  1108. #define ZGEMM_DEFAULT_Q 256
  1109. #define SYMV_P 4
  1110. #endif
  1111. #ifdef PPC970
  1112. #define SNUMOPT 4
  1113. #define DNUMOPT 4
  1114. #define GEMM_DEFAULT_OFFSET_A 2688
  1115. #define GEMM_DEFAULT_OFFSET_B 3072
  1116. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1117. #define SGEMM_DEFAULT_UNROLL_M 16
  1118. #define SGEMM_DEFAULT_UNROLL_N 4
  1119. #define DGEMM_DEFAULT_UNROLL_M 4
  1120. #define DGEMM_DEFAULT_UNROLL_N 4
  1121. #define CGEMM_DEFAULT_UNROLL_M 8
  1122. #define CGEMM_DEFAULT_UNROLL_N 2
  1123. #define ZGEMM_DEFAULT_UNROLL_M 2
  1124. #define ZGEMM_DEFAULT_UNROLL_N 2
  1125. #ifdef OS_LINUX
  1126. #if L2_SIZE == 1024976
  1127. #define SGEMM_DEFAULT_P 320
  1128. #define DGEMM_DEFAULT_P 256
  1129. #define CGEMM_DEFAULT_P 256
  1130. #define ZGEMM_DEFAULT_P 256
  1131. #else
  1132. #define SGEMM_DEFAULT_P 176
  1133. #define DGEMM_DEFAULT_P 176
  1134. #define CGEMM_DEFAULT_P 176
  1135. #define ZGEMM_DEFAULT_P 176
  1136. #endif
  1137. #endif
  1138. #define SGEMM_DEFAULT_Q 512
  1139. #define DGEMM_DEFAULT_Q 256
  1140. #define CGEMM_DEFAULT_Q 256
  1141. #define ZGEMM_DEFAULT_Q 128
  1142. #define SYMV_P 4
  1143. #endif
  1144. #ifdef PPC440
  1145. #define SNUMOPT 2
  1146. #define DNUMOPT 2
  1147. #define GEMM_DEFAULT_OFFSET_A (32 * 0)
  1148. #define GEMM_DEFAULT_OFFSET_B (32 * 0)
  1149. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1150. #define SGEMM_DEFAULT_UNROLL_M 4
  1151. #define SGEMM_DEFAULT_UNROLL_N 4
  1152. #define DGEMM_DEFAULT_UNROLL_M 4
  1153. #define DGEMM_DEFAULT_UNROLL_N 4
  1154. #define CGEMM_DEFAULT_UNROLL_M 2
  1155. #define CGEMM_DEFAULT_UNROLL_N 2
  1156. #define ZGEMM_DEFAULT_UNROLL_M 2
  1157. #define ZGEMM_DEFAULT_UNROLL_N 2
  1158. #define SGEMM_DEFAULT_P 512
  1159. #define DGEMM_DEFAULT_P 512
  1160. #define CGEMM_DEFAULT_P 512
  1161. #define ZGEMM_DEFAULT_P 512
  1162. #define SGEMM_DEFAULT_Q 1024
  1163. #define DGEMM_DEFAULT_Q 512
  1164. #define CGEMM_DEFAULT_Q 512
  1165. #define ZGEMM_DEFAULT_Q 256
  1166. #define SGEMM_DEFAULT_R SGEMM_DEFAULT_P
  1167. #define DGEMM_DEFAULT_R DGEMM_DEFAULT_P
  1168. #define CGEMM_DEFAULT_R CGEMM_DEFAULT_P
  1169. #define ZGEMM_DEFAULT_R ZGEMM_DEFAULT_P
  1170. #define SYMV_P 4
  1171. #endif
  1172. #ifdef PPC440FP2
  1173. #define SNUMOPT 4
  1174. #define DNUMOPT 4
  1175. #define GEMM_DEFAULT_OFFSET_A (32 * 0)
  1176. #define GEMM_DEFAULT_OFFSET_B (32 * 0)
  1177. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1178. #define SGEMM_DEFAULT_UNROLL_M 8
  1179. #define SGEMM_DEFAULT_UNROLL_N 4
  1180. #define DGEMM_DEFAULT_UNROLL_M 8
  1181. #define DGEMM_DEFAULT_UNROLL_N 4
  1182. #define CGEMM_DEFAULT_UNROLL_M 4
  1183. #define CGEMM_DEFAULT_UNROLL_N 2
  1184. #define ZGEMM_DEFAULT_UNROLL_M 4
  1185. #define ZGEMM_DEFAULT_UNROLL_N 2
  1186. #define SGEMM_DEFAULT_P 128
  1187. #define DGEMM_DEFAULT_P 128
  1188. #define CGEMM_DEFAULT_P 128
  1189. #define ZGEMM_DEFAULT_P 128
  1190. #if 1
  1191. #define SGEMM_DEFAULT_Q 4096
  1192. #define DGEMM_DEFAULT_Q 3072
  1193. #define CGEMM_DEFAULT_Q 2048
  1194. #define ZGEMM_DEFAULT_Q 1024
  1195. #else
  1196. #define SGEMM_DEFAULT_Q 512
  1197. #define DGEMM_DEFAULT_Q 256
  1198. #define CGEMM_DEFAULT_Q 256
  1199. #define ZGEMM_DEFAULT_Q 128
  1200. #endif
  1201. #define SYMV_P 4
  1202. #endif
  1203. #if defined(POWER3) || defined(POWER4) || defined(POWER5)
  1204. #define GEMM_DEFAULT_OFFSET_A 0
  1205. #define GEMM_DEFAULT_OFFSET_B 2048
  1206. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1207. #define SGEMM_DEFAULT_UNROLL_M 4
  1208. #define SGEMM_DEFAULT_UNROLL_N 4
  1209. #define DGEMM_DEFAULT_UNROLL_M 4
  1210. #define DGEMM_DEFAULT_UNROLL_N 4
  1211. #define CGEMM_DEFAULT_UNROLL_M 2
  1212. #define CGEMM_DEFAULT_UNROLL_N 2
  1213. #define ZGEMM_DEFAULT_UNROLL_M 2
  1214. #define ZGEMM_DEFAULT_UNROLL_N 2
  1215. #ifdef POWER3
  1216. #define SNUMOPT 4
  1217. #define DNUMOPT 4
  1218. #define SGEMM_DEFAULT_P 256
  1219. #define SGEMM_DEFAULT_Q 432
  1220. #define SGEMM_DEFAULT_R 1012
  1221. #define DGEMM_DEFAULT_P 256
  1222. #define DGEMM_DEFAULT_Q 216
  1223. #define DGEMM_DEFAULT_R 1012
  1224. #define ZGEMM_DEFAULT_P 256
  1225. #define ZGEMM_DEFAULT_Q 104
  1226. #define ZGEMM_DEFAULT_R 1012
  1227. #endif
  1228. #if defined(POWER4)
  1229. #ifdef ALLOC_HUGETLB
  1230. #define SGEMM_DEFAULT_P 184
  1231. #define DGEMM_DEFAULT_P 184
  1232. #define CGEMM_DEFAULT_P 184
  1233. #define ZGEMM_DEFAULT_P 184
  1234. #else
  1235. #define SGEMM_DEFAULT_P 144
  1236. #define DGEMM_DEFAULT_P 144
  1237. #define CGEMM_DEFAULT_P 144
  1238. #define ZGEMM_DEFAULT_P 144
  1239. #endif
  1240. #endif
  1241. #if defined(POWER5)
  1242. #ifdef ALLOC_HUGETLB
  1243. #define SGEMM_DEFAULT_P 512
  1244. #define DGEMM_DEFAULT_P 256
  1245. #define CGEMM_DEFAULT_P 256
  1246. #define ZGEMM_DEFAULT_P 128
  1247. #else
  1248. #define SGEMM_DEFAULT_P 320
  1249. #define DGEMM_DEFAULT_P 160
  1250. #define CGEMM_DEFAULT_P 160
  1251. #define ZGEMM_DEFAULT_P 80
  1252. #endif
  1253. #define SGEMM_DEFAULT_Q 256
  1254. #define CGEMM_DEFAULT_Q 256
  1255. #define DGEMM_DEFAULT_Q 256
  1256. #define ZGEMM_DEFAULT_Q 256
  1257. #endif
  1258. #define SYMV_P 8
  1259. #endif
  1260. #if defined(POWER6)
  1261. #define SNUMOPT 4
  1262. #define DNUMOPT 4
  1263. #define GEMM_DEFAULT_OFFSET_A 384
  1264. #define GEMM_DEFAULT_OFFSET_B 1024
  1265. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1266. #define SGEMM_DEFAULT_UNROLL_M 4
  1267. #define SGEMM_DEFAULT_UNROLL_N 4
  1268. #define DGEMM_DEFAULT_UNROLL_M 4
  1269. #define DGEMM_DEFAULT_UNROLL_N 4
  1270. #define CGEMM_DEFAULT_UNROLL_M 2
  1271. #define CGEMM_DEFAULT_UNROLL_N 4
  1272. #define ZGEMM_DEFAULT_UNROLL_M 2
  1273. #define ZGEMM_DEFAULT_UNROLL_N 4
  1274. #define SGEMM_DEFAULT_P 992
  1275. #define DGEMM_DEFAULT_P 480
  1276. #define CGEMM_DEFAULT_P 488
  1277. #define ZGEMM_DEFAULT_P 248
  1278. #define SGEMM_DEFAULT_Q 504
  1279. #define DGEMM_DEFAULT_Q 504
  1280. #define CGEMM_DEFAULT_Q 400
  1281. #define ZGEMM_DEFAULT_Q 400
  1282. #define SYMV_P 8
  1283. #endif
  1284. #if defined(SPARC) && defined(V7)
  1285. #define SNUMOPT 4
  1286. #define DNUMOPT 4
  1287. #define GEMM_DEFAULT_OFFSET_A 0
  1288. #define GEMM_DEFAULT_OFFSET_B 2048
  1289. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1290. #define SGEMM_DEFAULT_UNROLL_M 2
  1291. #define SGEMM_DEFAULT_UNROLL_N 8
  1292. #define DGEMM_DEFAULT_UNROLL_M 2
  1293. #define DGEMM_DEFAULT_UNROLL_N 8
  1294. #define CGEMM_DEFAULT_UNROLL_M 1
  1295. #define CGEMM_DEFAULT_UNROLL_N 4
  1296. #define ZGEMM_DEFAULT_UNROLL_M 1
  1297. #define ZGEMM_DEFAULT_UNROLL_N 4
  1298. #define SGEMM_DEFAULT_P 256
  1299. #define DGEMM_DEFAULT_P 256
  1300. #define CGEMM_DEFAULT_P 256
  1301. #define ZGEMM_DEFAULT_P 256
  1302. #define SGEMM_DEFAULT_Q 512
  1303. #define DGEMM_DEFAULT_Q 256
  1304. #define CGEMM_DEFAULT_Q 256
  1305. #define ZGEMM_DEFAULT_Q 128
  1306. #define SYMV_P 8
  1307. #define GEMM_THREAD gemm_thread_mn
  1308. #endif
  1309. #if (defined(SPARC) && defined(V9)) || defined(__sparc_v9__)
  1310. #define SNUMOPT 2
  1311. #define DNUMOPT 2
  1312. #define GEMM_DEFAULT_OFFSET_A 0
  1313. #define GEMM_DEFAULT_OFFSET_B 2048
  1314. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1315. #define SGEMM_DEFAULT_UNROLL_M 4
  1316. #define SGEMM_DEFAULT_UNROLL_N 4
  1317. #define DGEMM_DEFAULT_UNROLL_M 4
  1318. #define DGEMM_DEFAULT_UNROLL_N 4
  1319. #define CGEMM_DEFAULT_UNROLL_M 2
  1320. #define CGEMM_DEFAULT_UNROLL_N 2
  1321. #define ZGEMM_DEFAULT_UNROLL_M 2
  1322. #define ZGEMM_DEFAULT_UNROLL_N 2
  1323. #define SGEMM_DEFAULT_P 512
  1324. #define DGEMM_DEFAULT_P 512
  1325. #define CGEMM_DEFAULT_P 512
  1326. #define ZGEMM_DEFAULT_P 512
  1327. #define SGEMM_DEFAULT_Q 1024
  1328. #define DGEMM_DEFAULT_Q 512
  1329. #define CGEMM_DEFAULT_Q 512
  1330. #define ZGEMM_DEFAULT_Q 256
  1331. #define SYMV_P 8
  1332. #endif
  1333. #ifdef SICORTEX
  1334. #define SNUMOPT 2
  1335. #define DNUMOPT 2
  1336. #define GEMM_DEFAULT_OFFSET_A 0
  1337. #define GEMM_DEFAULT_OFFSET_B 0
  1338. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1339. #define SGEMM_DEFAULT_UNROLL_M 2
  1340. #define SGEMM_DEFAULT_UNROLL_N 8
  1341. #define DGEMM_DEFAULT_UNROLL_M 2
  1342. #define DGEMM_DEFAULT_UNROLL_N 8
  1343. #define CGEMM_DEFAULT_UNROLL_M 1
  1344. #define CGEMM_DEFAULT_UNROLL_N 4
  1345. #define ZGEMM_DEFAULT_UNROLL_M 1
  1346. #define ZGEMM_DEFAULT_UNROLL_N 4
  1347. #define SGEMM_DEFAULT_P 108
  1348. #define DGEMM_DEFAULT_P 112
  1349. #define CGEMM_DEFAULT_P 108
  1350. #define ZGEMM_DEFAULT_P 112
  1351. #define SGEMM_DEFAULT_Q 288
  1352. #define DGEMM_DEFAULT_Q 144
  1353. #define CGEMM_DEFAULT_Q 144
  1354. #define ZGEMM_DEFAULT_Q 72
  1355. #define SGEMM_DEFAULT_R 2000
  1356. #define DGEMM_DEFAULT_R 2000
  1357. #define CGEMM_DEFAULT_R 2000
  1358. #define ZGEMM_DEFAULT_R 2000
  1359. #define SYMV_P 16
  1360. #endif
  1361. #ifdef LOONGSON3A
  1362. ////Copy from SICORTEX
  1363. #define SNUMOPT 2
  1364. #define DNUMOPT 2
  1365. #define GEMM_DEFAULT_OFFSET_A 0
  1366. #define GEMM_DEFAULT_OFFSET_B 0
  1367. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1368. #define SGEMM_DEFAULT_UNROLL_M 8
  1369. #define SGEMM_DEFAULT_UNROLL_N 4
  1370. #define DGEMM_DEFAULT_UNROLL_M 4
  1371. #define DGEMM_DEFAULT_UNROLL_N 4
  1372. #define CGEMM_DEFAULT_UNROLL_M 4
  1373. #define CGEMM_DEFAULT_UNROLL_N 2
  1374. #define ZGEMM_DEFAULT_UNROLL_M 2
  1375. #define ZGEMM_DEFAULT_UNROLL_N 2
  1376. #define SGEMM_DEFAULT_P 64
  1377. #define DGEMM_DEFAULT_P 44
  1378. #define CGEMM_DEFAULT_P 64
  1379. #define ZGEMM_DEFAULT_P 32
  1380. #define SGEMM_DEFAULT_Q 192
  1381. #define DGEMM_DEFAULT_Q 92
  1382. #define CGEMM_DEFAULT_Q 128
  1383. #define ZGEMM_DEFAULT_Q 80
  1384. #define SGEMM_DEFAULT_R 640
  1385. #define DGEMM_DEFAULT_R dgemm_r
  1386. #define CGEMM_DEFAULT_R 640
  1387. #define ZGEMM_DEFAULT_R 640
  1388. #define GEMM_OFFSET_A1 0x10000
  1389. #define GEMM_OFFSET_B1 0x100000
  1390. #define SYMV_P 16
  1391. #endif
  1392. #ifdef LOONGSON3B
  1393. #define SNUMOPT 2
  1394. #define DNUMOPT 2
  1395. #define GEMM_DEFAULT_OFFSET_A 0
  1396. #define GEMM_DEFAULT_OFFSET_B 0
  1397. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1398. #define SGEMM_DEFAULT_UNROLL_M 2
  1399. #define SGEMM_DEFAULT_UNROLL_N 2
  1400. #define DGEMM_DEFAULT_UNROLL_M 2
  1401. #define DGEMM_DEFAULT_UNROLL_N 2
  1402. #define CGEMM_DEFAULT_UNROLL_M 2
  1403. #define CGEMM_DEFAULT_UNROLL_N 2
  1404. #define ZGEMM_DEFAULT_UNROLL_M 2
  1405. #define ZGEMM_DEFAULT_UNROLL_N 2
  1406. #define SGEMM_DEFAULT_P 64
  1407. #define DGEMM_DEFAULT_P 24
  1408. #define CGEMM_DEFAULT_P 24
  1409. #define ZGEMM_DEFAULT_P 20
  1410. #define SGEMM_DEFAULT_Q 192
  1411. #define DGEMM_DEFAULT_Q 128
  1412. #define CGEMM_DEFAULT_Q 128
  1413. #define ZGEMM_DEFAULT_Q 64
  1414. #define SGEMM_DEFAULT_R 512
  1415. #define DGEMM_DEFAULT_R 512
  1416. #define CGEMM_DEFAULT_R 512
  1417. #define ZGEMM_DEFAULT_R 512
  1418. #define GEMM_OFFSET_A1 0x10000
  1419. #define GEMM_OFFSET_B1 0x100000
  1420. #define SYMV_P 16
  1421. #endif
  1422. #ifdef GENERIC
  1423. #define SNUMOPT 2
  1424. #define DNUMOPT 2
  1425. #define GEMM_DEFAULT_OFFSET_A 0
  1426. #define GEMM_DEFAULT_OFFSET_B 0
  1427. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1428. #define SGEMM_DEFAULT_UNROLL_N 4
  1429. #define DGEMM_DEFAULT_UNROLL_N 4
  1430. #define QGEMM_DEFAULT_UNROLL_N 2
  1431. #define CGEMM_DEFAULT_UNROLL_N 2
  1432. #define ZGEMM_DEFAULT_UNROLL_N 2
  1433. #define XGEMM_DEFAULT_UNROLL_N 1
  1434. #ifdef ARCH_X86
  1435. #define SGEMM_DEFAULT_UNROLL_M 4
  1436. #define DGEMM_DEFAULT_UNROLL_M 2
  1437. #define QGEMM_DEFAULT_UNROLL_M 2
  1438. #define CGEMM_DEFAULT_UNROLL_M 2
  1439. #define ZGEMM_DEFAULT_UNROLL_M 1
  1440. #define XGEMM_DEFAULT_UNROLL_M 1
  1441. #else
  1442. #define SGEMM_DEFAULT_UNROLL_M 8
  1443. #define DGEMM_DEFAULT_UNROLL_M 4
  1444. #define QGEMM_DEFAULT_UNROLL_M 2
  1445. #define CGEMM_DEFAULT_UNROLL_M 4
  1446. #define ZGEMM_DEFAULT_UNROLL_M 2
  1447. #define XGEMM_DEFAULT_UNROLL_M 1
  1448. #endif
  1449. #define SGEMM_DEFAULT_P sgemm_p
  1450. #define DGEMM_DEFAULT_P dgemm_p
  1451. #define QGEMM_DEFAULT_P qgemm_p
  1452. #define CGEMM_DEFAULT_P cgemm_p
  1453. #define ZGEMM_DEFAULT_P zgemm_p
  1454. #define XGEMM_DEFAULT_P xgemm_p
  1455. #define SGEMM_DEFAULT_R sgemm_r
  1456. #define DGEMM_DEFAULT_R dgemm_r
  1457. #define QGEMM_DEFAULT_R qgemm_r
  1458. #define CGEMM_DEFAULT_R cgemm_r
  1459. #define ZGEMM_DEFAULT_R zgemm_r
  1460. #define XGEMM_DEFAULT_R xgemm_r
  1461. #define SGEMM_DEFAULT_Q 128
  1462. #define DGEMM_DEFAULT_Q 128
  1463. #define QGEMM_DEFAULT_Q 128
  1464. #define CGEMM_DEFAULT_Q 128
  1465. #define ZGEMM_DEFAULT_Q 128
  1466. #define XGEMM_DEFAULT_Q 128
  1467. #define SYMV_P 16
  1468. #endif
  1469. #ifndef QGEMM_DEFAULT_UNROLL_M
  1470. #define QGEMM_DEFAULT_UNROLL_M 2
  1471. #endif
  1472. #ifndef QGEMM_DEFAULT_UNROLL_N
  1473. #define QGEMM_DEFAULT_UNROLL_N 2
  1474. #endif
  1475. #ifndef XGEMM_DEFAULT_UNROLL_M
  1476. #define XGEMM_DEFAULT_UNROLL_M 2
  1477. #endif
  1478. #ifndef XGEMM_DEFAULT_UNROLL_N
  1479. #define XGEMM_DEFAULT_UNROLL_N 2
  1480. #endif
  1481. #ifndef HAVE_SSE2
  1482. #define SHUFPD_0 shufps $0x44,
  1483. #define SHUFPD_1 shufps $0x4e,
  1484. #define SHUFPD_2 shufps $0xe4,
  1485. #define SHUFPD_3 shufps $0xee,
  1486. #endif
  1487. #ifndef SHUFPD_0
  1488. #define SHUFPD_0 shufpd $0,
  1489. #endif
  1490. #ifndef SHUFPD_1
  1491. #define SHUFPD_1 shufpd $1,
  1492. #endif
  1493. #ifndef SHUFPD_2
  1494. #define SHUFPD_2 shufpd $2,
  1495. #endif
  1496. #ifndef SHUFPD_3
  1497. #define SHUFPD_3 shufpd $3,
  1498. #endif
  1499. #ifndef SHUFPS_39
  1500. #define SHUFPS_39 shufps $0x39,
  1501. #endif
  1502. #endif