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.

setparam-ref.c 33 kB

11 years ago
11 years ago
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082
  1. /*********************************************************************/
  2. /* Copyright 2009, 2010 The University of Texas at Austin. */
  3. /* All rights reserved. */
  4. /* */
  5. /* Redistribution and use in source and binary forms, with or */
  6. /* without modification, are permitted provided that the following */
  7. /* conditions are met: */
  8. /* */
  9. /* 1. Redistributions of source code must retain the above */
  10. /* copyright notice, this list of conditions and the following */
  11. /* disclaimer. */
  12. /* */
  13. /* 2. Redistributions in binary form must reproduce the above */
  14. /* copyright notice, this list of conditions and the following */
  15. /* disclaimer in the documentation and/or other materials */
  16. /* provided with the distribution. */
  17. /* */
  18. /* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
  19. /* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
  20. /* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
  21. /* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
  22. /* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
  23. /* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
  24. /* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
  25. /* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
  26. /* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
  27. /* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
  28. /* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
  29. /* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
  30. /* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
  31. /* POSSIBILITY OF SUCH DAMAGE. */
  32. /* */
  33. /* The views and conclusions contained in the software and */
  34. /* documentation are those of the authors and should not be */
  35. /* interpreted as representing official policies, either expressed */
  36. /* or implied, of The University of Texas at Austin. */
  37. /*********************************************************************/
  38. #include <stdio.h>
  39. #include <string.h>
  40. #include "common.h"
  41. #ifdef BUILD_KERNEL
  42. #include "kernelTS.h"
  43. #endif
  44. #undef DEBUG
  45. static void init_parameter(void);
  46. gotoblas_t TABLE_NAME = {
  47. DTB_DEFAULT_ENTRIES ,
  48. GEMM_DEFAULT_OFFSET_A, GEMM_DEFAULT_OFFSET_B, GEMM_DEFAULT_ALIGN,
  49. 0, 0, 0,
  50. SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N,
  51. #ifdef SGEMM_DEFAULT_UNROLL_MN
  52. SGEMM_DEFAULT_UNROLL_MN,
  53. #else
  54. MAX(SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N),
  55. #endif
  56. #ifdef HAVE_EXCLUSIVE_CACHE
  57. 1,
  58. #else
  59. 0,
  60. #endif
  61. samax_kTS, samin_kTS, smax_kTS, smin_kTS,
  62. isamax_kTS, isamin_kTS, ismax_kTS, ismin_kTS,
  63. snrm2_kTS, sasum_kTS, scopy_kTS, sdot_kTS,
  64. dsdot_kTS,
  65. srot_kTS, saxpy_kTS, sscal_kTS, sswap_kTS,
  66. sgemv_nTS, sgemv_tTS, sger_kTS,
  67. ssymv_LTS, ssymv_UTS,
  68. sgemm_kernelTS, sgemm_betaTS,
  69. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  70. sgemm_incopyTS, sgemm_itcopyTS,
  71. #else
  72. sgemm_oncopyTS, sgemm_otcopyTS,
  73. #endif
  74. sgemm_oncopyTS, sgemm_otcopyTS,
  75. strsm_kernel_LNTS, strsm_kernel_LTTS, strsm_kernel_RNTS, strsm_kernel_RTTS,
  76. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  77. strsm_iunucopyTS, strsm_iunncopyTS, strsm_iutucopyTS, strsm_iutncopyTS,
  78. strsm_ilnucopyTS, strsm_ilnncopyTS, strsm_iltucopyTS, strsm_iltncopyTS,
  79. #else
  80. strsm_ounucopyTS, strsm_ounncopyTS, strsm_outucopyTS, strsm_outncopyTS,
  81. strsm_olnucopyTS, strsm_olnncopyTS, strsm_oltucopyTS, strsm_oltncopyTS,
  82. #endif
  83. strsm_ounucopyTS, strsm_ounncopyTS, strsm_outucopyTS, strsm_outncopyTS,
  84. strsm_olnucopyTS, strsm_olnncopyTS, strsm_oltucopyTS, strsm_oltncopyTS,
  85. strmm_kernel_RNTS, strmm_kernel_RTTS, strmm_kernel_LNTS, strmm_kernel_LTTS,
  86. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  87. strmm_iunucopyTS, strmm_iunncopyTS, strmm_iutucopyTS, strmm_iutncopyTS,
  88. strmm_ilnucopyTS, strmm_ilnncopyTS, strmm_iltucopyTS, strmm_iltncopyTS,
  89. #else
  90. strmm_ounucopyTS, strmm_ounncopyTS, strmm_outucopyTS, strmm_outncopyTS,
  91. strmm_olnucopyTS, strmm_olnncopyTS, strmm_oltucopyTS, strmm_oltncopyTS,
  92. #endif
  93. strmm_ounucopyTS, strmm_ounncopyTS, strmm_outucopyTS, strmm_outncopyTS,
  94. strmm_olnucopyTS, strmm_olnncopyTS, strmm_oltucopyTS, strmm_oltncopyTS,
  95. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  96. ssymm_iutcopyTS, ssymm_iltcopyTS,
  97. #else
  98. ssymm_outcopyTS, ssymm_oltcopyTS,
  99. #endif
  100. ssymm_outcopyTS, ssymm_oltcopyTS,
  101. #ifndef NO_LAPACK
  102. sneg_tcopyTS, slaswp_ncopyTS,
  103. #else
  104. NULL,NULL,
  105. #endif
  106. 0, 0, 0,
  107. DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N,
  108. #ifdef DGEMM_DEFAULT_UNROLL_MN
  109. DGEMM_DEFAULT_UNROLL_MN,
  110. #else
  111. MAX(DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N),
  112. #endif
  113. damax_kTS, damin_kTS, dmax_kTS, dmin_kTS,
  114. idamax_kTS, idamin_kTS, idmax_kTS, idmin_kTS,
  115. dnrm2_kTS, dasum_kTS, dcopy_kTS, ddot_kTS,
  116. drot_kTS, daxpy_kTS, dscal_kTS, dswap_kTS,
  117. dgemv_nTS, dgemv_tTS, dger_kTS,
  118. dsymv_LTS, dsymv_UTS,
  119. dgemm_kernelTS, dgemm_betaTS,
  120. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  121. dgemm_incopyTS, dgemm_itcopyTS,
  122. #else
  123. dgemm_oncopyTS, dgemm_otcopyTS,
  124. #endif
  125. dgemm_oncopyTS, dgemm_otcopyTS,
  126. dtrsm_kernel_LNTS, dtrsm_kernel_LTTS, dtrsm_kernel_RNTS, dtrsm_kernel_RTTS,
  127. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  128. dtrsm_iunucopyTS, dtrsm_iunncopyTS, dtrsm_iutucopyTS, dtrsm_iutncopyTS,
  129. dtrsm_ilnucopyTS, dtrsm_ilnncopyTS, dtrsm_iltucopyTS, dtrsm_iltncopyTS,
  130. #else
  131. dtrsm_ounucopyTS, dtrsm_ounncopyTS, dtrsm_outucopyTS, dtrsm_outncopyTS,
  132. dtrsm_olnucopyTS, dtrsm_olnncopyTS, dtrsm_oltucopyTS, dtrsm_oltncopyTS,
  133. #endif
  134. dtrsm_ounucopyTS, dtrsm_ounncopyTS, dtrsm_outucopyTS, dtrsm_outncopyTS,
  135. dtrsm_olnucopyTS, dtrsm_olnncopyTS, dtrsm_oltucopyTS, dtrsm_oltncopyTS,
  136. dtrmm_kernel_RNTS, dtrmm_kernel_RTTS, dtrmm_kernel_LNTS, dtrmm_kernel_LTTS,
  137. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  138. dtrmm_iunucopyTS, dtrmm_iunncopyTS, dtrmm_iutucopyTS, dtrmm_iutncopyTS,
  139. dtrmm_ilnucopyTS, dtrmm_ilnncopyTS, dtrmm_iltucopyTS, dtrmm_iltncopyTS,
  140. #else
  141. dtrmm_ounucopyTS, dtrmm_ounncopyTS, dtrmm_outucopyTS, dtrmm_outncopyTS,
  142. dtrmm_olnucopyTS, dtrmm_olnncopyTS, dtrmm_oltucopyTS, dtrmm_oltncopyTS,
  143. #endif
  144. dtrmm_ounucopyTS, dtrmm_ounncopyTS, dtrmm_outucopyTS, dtrmm_outncopyTS,
  145. dtrmm_olnucopyTS, dtrmm_olnncopyTS, dtrmm_oltucopyTS, dtrmm_oltncopyTS,
  146. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  147. dsymm_iutcopyTS, dsymm_iltcopyTS,
  148. #else
  149. dsymm_outcopyTS, dsymm_oltcopyTS,
  150. #endif
  151. dsymm_outcopyTS, dsymm_oltcopyTS,
  152. #ifndef NO_LAPACK
  153. dneg_tcopyTS, dlaswp_ncopyTS,
  154. #else
  155. NULL, NULL,
  156. #endif
  157. #ifdef EXPRECISION
  158. 0, 0, 0,
  159. QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N, MAX(QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N),
  160. qamax_kTS, qamin_kTS, qmax_kTS, qmin_kTS,
  161. iqamax_kTS, iqamin_kTS, iqmax_kTS, iqmin_kTS,
  162. qnrm2_kTS, qasum_kTS, qcopy_kTS, qdot_kTS,
  163. qrot_kTS, qaxpy_kTS, qscal_kTS, qswap_kTS,
  164. qgemv_nTS, qgemv_tTS, qger_kTS,
  165. qsymv_LTS, qsymv_UTS,
  166. qgemm_kernelTS, qgemm_betaTS,
  167. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  168. qgemm_incopyTS, qgemm_itcopyTS,
  169. #else
  170. qgemm_oncopyTS, qgemm_otcopyTS,
  171. #endif
  172. qgemm_oncopyTS, qgemm_otcopyTS,
  173. qtrsm_kernel_LNTS, qtrsm_kernel_LTTS, qtrsm_kernel_RNTS, qtrsm_kernel_RTTS,
  174. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  175. qtrsm_iunucopyTS, qtrsm_iunncopyTS, qtrsm_iutucopyTS, qtrsm_iutncopyTS,
  176. qtrsm_ilnucopyTS, qtrsm_ilnncopyTS, qtrsm_iltucopyTS, qtrsm_iltncopyTS,
  177. #else
  178. qtrsm_ounucopyTS, qtrsm_ounncopyTS, qtrsm_outucopyTS, qtrsm_outncopyTS,
  179. qtrsm_olnucopyTS, qtrsm_olnncopyTS, qtrsm_oltucopyTS, qtrsm_oltncopyTS,
  180. #endif
  181. qtrsm_ounucopyTS, qtrsm_ounncopyTS, qtrsm_outucopyTS, qtrsm_outncopyTS,
  182. qtrsm_olnucopyTS, qtrsm_olnncopyTS, qtrsm_oltucopyTS, qtrsm_oltncopyTS,
  183. qtrmm_kernel_RNTS, qtrmm_kernel_RTTS, qtrmm_kernel_LNTS, qtrmm_kernel_LTTS,
  184. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  185. qtrmm_iunucopyTS, qtrmm_iunncopyTS, qtrmm_iutucopyTS, qtrmm_iutncopyTS,
  186. qtrmm_ilnucopyTS, qtrmm_ilnncopyTS, qtrmm_iltucopyTS, qtrmm_iltncopyTS,
  187. #else
  188. qtrmm_ounucopyTS, qtrmm_ounncopyTS, qtrmm_outucopyTS, qtrmm_outncopyTS,
  189. qtrmm_olnucopyTS, qtrmm_olnncopyTS, qtrmm_oltucopyTS, qtrmm_oltncopyTS,
  190. #endif
  191. qtrmm_ounucopyTS, qtrmm_ounncopyTS, qtrmm_outucopyTS, qtrmm_outncopyTS,
  192. qtrmm_olnucopyTS, qtrmm_olnncopyTS, qtrmm_oltucopyTS, qtrmm_oltncopyTS,
  193. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  194. qsymm_iutcopyTS, qsymm_iltcopyTS,
  195. #else
  196. qsymm_outcopyTS, qsymm_oltcopyTS,
  197. #endif
  198. qsymm_outcopyTS, qsymm_oltcopyTS,
  199. #ifndef NO_LAPACK
  200. qneg_tcopyTS, qlaswp_ncopyTS,
  201. #else
  202. NULL, NULL,
  203. #endif
  204. #endif
  205. 0, 0, 0,
  206. CGEMM_DEFAULT_UNROLL_M, CGEMM_DEFAULT_UNROLL_N,
  207. #ifdef CGEMM_DEFAULT_UNROLL_MN
  208. CGEMM_DEFAULT_UNROLL_MN,
  209. #else
  210. MAX(CGEMM_DEFAULT_UNROLL_M, CGEMM_DEFAULT_UNROLL_N),
  211. #endif
  212. camax_kTS, camin_kTS, icamax_kTS, icamin_kTS,
  213. cnrm2_kTS, casum_kTS, ccopy_kTS,
  214. cdotu_kTS, cdotc_kTS, csrot_kTS,
  215. caxpy_kTS, caxpyc_kTS, cscal_kTS, cswap_kTS,
  216. cgemv_nTS, cgemv_tTS, cgemv_rTS, cgemv_cTS,
  217. cgemv_oTS, cgemv_uTS, cgemv_sTS, cgemv_dTS,
  218. cgeru_kTS, cgerc_kTS, cgerv_kTS, cgerd_kTS,
  219. csymv_LTS, csymv_UTS,
  220. chemv_LTS, chemv_UTS, chemv_MTS, chemv_VTS,
  221. cgemm_kernel_nTS, cgemm_kernel_lTS, cgemm_kernel_rTS, cgemm_kernel_bTS,
  222. cgemm_betaTS,
  223. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  224. cgemm_incopyTS, cgemm_itcopyTS,
  225. #else
  226. cgemm_oncopyTS, cgemm_otcopyTS,
  227. #endif
  228. cgemm_oncopyTS, cgemm_otcopyTS,
  229. ctrsm_kernel_LNTS, ctrsm_kernel_LTTS, ctrsm_kernel_LRTS, ctrsm_kernel_LCTS,
  230. ctrsm_kernel_RNTS, ctrsm_kernel_RTTS, ctrsm_kernel_RRTS, ctrsm_kernel_RCTS,
  231. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  232. ctrsm_iunucopyTS, ctrsm_iunncopyTS, ctrsm_iutucopyTS, ctrsm_iutncopyTS,
  233. ctrsm_ilnucopyTS, ctrsm_ilnncopyTS, ctrsm_iltucopyTS, ctrsm_iltncopyTS,
  234. #else
  235. ctrsm_ounucopyTS, ctrsm_ounncopyTS, ctrsm_outucopyTS, ctrsm_outncopyTS,
  236. ctrsm_olnucopyTS, ctrsm_olnncopyTS, ctrsm_oltucopyTS, ctrsm_oltncopyTS,
  237. #endif
  238. ctrsm_ounucopyTS, ctrsm_ounncopyTS, ctrsm_outucopyTS, ctrsm_outncopyTS,
  239. ctrsm_olnucopyTS, ctrsm_olnncopyTS, ctrsm_oltucopyTS, ctrsm_oltncopyTS,
  240. ctrmm_kernel_RNTS, ctrmm_kernel_RTTS, ctrmm_kernel_RRTS, ctrmm_kernel_RCTS,
  241. ctrmm_kernel_LNTS, ctrmm_kernel_LTTS, ctrmm_kernel_LRTS, ctrmm_kernel_LCTS,
  242. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  243. ctrmm_iunucopyTS, ctrmm_iunncopyTS, ctrmm_iutucopyTS, ctrmm_iutncopyTS,
  244. ctrmm_ilnucopyTS, ctrmm_ilnncopyTS, ctrmm_iltucopyTS, ctrmm_iltncopyTS,
  245. #else
  246. ctrmm_ounucopyTS, ctrmm_ounncopyTS, ctrmm_outucopyTS, ctrmm_outncopyTS,
  247. ctrmm_olnucopyTS, ctrmm_olnncopyTS, ctrmm_oltucopyTS, ctrmm_oltncopyTS,
  248. #endif
  249. ctrmm_ounucopyTS, ctrmm_ounncopyTS, ctrmm_outucopyTS, ctrmm_outncopyTS,
  250. ctrmm_olnucopyTS, ctrmm_olnncopyTS, ctrmm_oltucopyTS, ctrmm_oltncopyTS,
  251. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  252. csymm_iutcopyTS, csymm_iltcopyTS,
  253. #else
  254. csymm_outcopyTS, csymm_oltcopyTS,
  255. #endif
  256. csymm_outcopyTS, csymm_oltcopyTS,
  257. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  258. chemm_iutcopyTS, chemm_iltcopyTS,
  259. #else
  260. chemm_outcopyTS, chemm_oltcopyTS,
  261. #endif
  262. chemm_outcopyTS, chemm_oltcopyTS,
  263. 0, 0, 0,
  264. #ifdef CGEMM3M_DEFAULT_UNROLL_M
  265. CGEMM3M_DEFAULT_UNROLL_M, CGEMM3M_DEFAULT_UNROLL_N, MAX(CGEMM3M_DEFAULT_UNROLL_M, CGEMM3M_DEFAULT_UNROLL_N),
  266. #else
  267. SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N, MAX(SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N),
  268. #endif
  269. cgemm3m_kernelTS,
  270. cgemm3m_incopybTS, cgemm3m_incopyrTS,
  271. cgemm3m_incopyiTS, cgemm3m_itcopybTS,
  272. cgemm3m_itcopyrTS, cgemm3m_itcopyiTS,
  273. cgemm3m_oncopybTS, cgemm3m_oncopyrTS,
  274. cgemm3m_oncopyiTS, cgemm3m_otcopybTS,
  275. cgemm3m_otcopyrTS, cgemm3m_otcopyiTS,
  276. csymm3m_iucopybTS, csymm3m_ilcopybTS,
  277. csymm3m_iucopyrTS, csymm3m_ilcopyrTS,
  278. csymm3m_iucopyiTS, csymm3m_ilcopyiTS,
  279. csymm3m_oucopybTS, csymm3m_olcopybTS,
  280. csymm3m_oucopyrTS, csymm3m_olcopyrTS,
  281. csymm3m_oucopyiTS, csymm3m_olcopyiTS,
  282. chemm3m_iucopybTS, chemm3m_ilcopybTS,
  283. chemm3m_iucopyrTS, chemm3m_ilcopyrTS,
  284. chemm3m_iucopyiTS, chemm3m_ilcopyiTS,
  285. chemm3m_oucopybTS, chemm3m_olcopybTS,
  286. chemm3m_oucopyrTS, chemm3m_olcopyrTS,
  287. chemm3m_oucopyiTS, chemm3m_olcopyiTS,
  288. #ifndef NO_LAPACK
  289. cneg_tcopyTS, claswp_ncopyTS,
  290. #else
  291. NULL, NULL,
  292. #endif
  293. 0, 0, 0,
  294. ZGEMM_DEFAULT_UNROLL_M, ZGEMM_DEFAULT_UNROLL_N,
  295. #ifdef ZGEMM_DEFAULT_UNROLL_MN
  296. ZGEMM_DEFAULT_UNROLL_MN,
  297. #else
  298. MAX(ZGEMM_DEFAULT_UNROLL_M, ZGEMM_DEFAULT_UNROLL_N),
  299. #endif
  300. zamax_kTS, zamin_kTS, izamax_kTS, izamin_kTS,
  301. znrm2_kTS, zasum_kTS, zcopy_kTS,
  302. zdotu_kTS, zdotc_kTS, zdrot_kTS,
  303. zaxpy_kTS, zaxpyc_kTS, zscal_kTS, zswap_kTS,
  304. zgemv_nTS, zgemv_tTS, zgemv_rTS, zgemv_cTS,
  305. zgemv_oTS, zgemv_uTS, zgemv_sTS, zgemv_dTS,
  306. zgeru_kTS, zgerc_kTS, zgerv_kTS, zgerd_kTS,
  307. zsymv_LTS, zsymv_UTS,
  308. zhemv_LTS, zhemv_UTS, zhemv_MTS, zhemv_VTS,
  309. zgemm_kernel_nTS, zgemm_kernel_lTS, zgemm_kernel_rTS, zgemm_kernel_bTS,
  310. zgemm_betaTS,
  311. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  312. zgemm_incopyTS, zgemm_itcopyTS,
  313. #else
  314. zgemm_oncopyTS, zgemm_otcopyTS,
  315. #endif
  316. zgemm_oncopyTS, zgemm_otcopyTS,
  317. ztrsm_kernel_LNTS, ztrsm_kernel_LTTS, ztrsm_kernel_LRTS, ztrsm_kernel_LCTS,
  318. ztrsm_kernel_RNTS, ztrsm_kernel_RTTS, ztrsm_kernel_RRTS, ztrsm_kernel_RCTS,
  319. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  320. ztrsm_iunucopyTS, ztrsm_iunncopyTS, ztrsm_iutucopyTS, ztrsm_iutncopyTS,
  321. ztrsm_ilnucopyTS, ztrsm_ilnncopyTS, ztrsm_iltucopyTS, ztrsm_iltncopyTS,
  322. #else
  323. ztrsm_ounucopyTS, ztrsm_ounncopyTS, ztrsm_outucopyTS, ztrsm_outncopyTS,
  324. ztrsm_olnucopyTS, ztrsm_olnncopyTS, ztrsm_oltucopyTS, ztrsm_oltncopyTS,
  325. #endif
  326. ztrsm_ounucopyTS, ztrsm_ounncopyTS, ztrsm_outucopyTS, ztrsm_outncopyTS,
  327. ztrsm_olnucopyTS, ztrsm_olnncopyTS, ztrsm_oltucopyTS, ztrsm_oltncopyTS,
  328. ztrmm_kernel_RNTS, ztrmm_kernel_RTTS, ztrmm_kernel_RRTS, ztrmm_kernel_RCTS,
  329. ztrmm_kernel_LNTS, ztrmm_kernel_LTTS, ztrmm_kernel_LRTS, ztrmm_kernel_LCTS,
  330. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  331. ztrmm_iunucopyTS, ztrmm_iunncopyTS, ztrmm_iutucopyTS, ztrmm_iutncopyTS,
  332. ztrmm_ilnucopyTS, ztrmm_ilnncopyTS, ztrmm_iltucopyTS, ztrmm_iltncopyTS,
  333. #else
  334. ztrmm_ounucopyTS, ztrmm_ounncopyTS, ztrmm_outucopyTS, ztrmm_outncopyTS,
  335. ztrmm_olnucopyTS, ztrmm_olnncopyTS, ztrmm_oltucopyTS, ztrmm_oltncopyTS,
  336. #endif
  337. ztrmm_ounucopyTS, ztrmm_ounncopyTS, ztrmm_outucopyTS, ztrmm_outncopyTS,
  338. ztrmm_olnucopyTS, ztrmm_olnncopyTS, ztrmm_oltucopyTS, ztrmm_oltncopyTS,
  339. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  340. zsymm_iutcopyTS, zsymm_iltcopyTS,
  341. #else
  342. zsymm_outcopyTS, zsymm_oltcopyTS,
  343. #endif
  344. zsymm_outcopyTS, zsymm_oltcopyTS,
  345. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  346. zhemm_iutcopyTS, zhemm_iltcopyTS,
  347. #else
  348. zhemm_outcopyTS, zhemm_oltcopyTS,
  349. #endif
  350. zhemm_outcopyTS, zhemm_oltcopyTS,
  351. 0, 0, 0,
  352. #ifdef ZGEMM3M_DEFAULT_UNROLL_M
  353. ZGEMM3M_DEFAULT_UNROLL_M, ZGEMM3M_DEFAULT_UNROLL_N, MAX(ZGEMM3M_DEFAULT_UNROLL_M, ZGEMM3M_DEFAULT_UNROLL_N),
  354. #else
  355. DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N, MAX(DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N),
  356. #endif
  357. zgemm3m_kernelTS,
  358. zgemm3m_incopybTS, zgemm3m_incopyrTS,
  359. zgemm3m_incopyiTS, zgemm3m_itcopybTS,
  360. zgemm3m_itcopyrTS, zgemm3m_itcopyiTS,
  361. zgemm3m_oncopybTS, zgemm3m_oncopyrTS,
  362. zgemm3m_oncopyiTS, zgemm3m_otcopybTS,
  363. zgemm3m_otcopyrTS, zgemm3m_otcopyiTS,
  364. zsymm3m_iucopybTS, zsymm3m_ilcopybTS,
  365. zsymm3m_iucopyrTS, zsymm3m_ilcopyrTS,
  366. zsymm3m_iucopyiTS, zsymm3m_ilcopyiTS,
  367. zsymm3m_oucopybTS, zsymm3m_olcopybTS,
  368. zsymm3m_oucopyrTS, zsymm3m_olcopyrTS,
  369. zsymm3m_oucopyiTS, zsymm3m_olcopyiTS,
  370. zhemm3m_iucopybTS, zhemm3m_ilcopybTS,
  371. zhemm3m_iucopyrTS, zhemm3m_ilcopyrTS,
  372. zhemm3m_iucopyiTS, zhemm3m_ilcopyiTS,
  373. zhemm3m_oucopybTS, zhemm3m_olcopybTS,
  374. zhemm3m_oucopyrTS, zhemm3m_olcopyrTS,
  375. zhemm3m_oucopyiTS, zhemm3m_olcopyiTS,
  376. #ifndef NO_LAPACK
  377. zneg_tcopyTS, zlaswp_ncopyTS,
  378. #else
  379. NULL, NULL,
  380. #endif
  381. #ifdef EXPRECISION
  382. 0, 0, 0,
  383. XGEMM_DEFAULT_UNROLL_M, XGEMM_DEFAULT_UNROLL_N, MAX(XGEMM_DEFAULT_UNROLL_M, XGEMM_DEFAULT_UNROLL_N),
  384. xamax_kTS, xamin_kTS, ixamax_kTS, ixamin_kTS,
  385. xnrm2_kTS, xasum_kTS, xcopy_kTS,
  386. xdotu_kTS, xdotc_kTS, xqrot_kTS,
  387. xaxpy_kTS, xaxpyc_kTS, xscal_kTS, xswap_kTS,
  388. xgemv_nTS, xgemv_tTS, xgemv_rTS, xgemv_cTS,
  389. xgemv_oTS, xgemv_uTS, xgemv_sTS, xgemv_dTS,
  390. xgeru_kTS, xgerc_kTS, xgerv_kTS, xgerd_kTS,
  391. xsymv_LTS, xsymv_UTS,
  392. xhemv_LTS, xhemv_UTS, xhemv_MTS, xhemv_VTS,
  393. xgemm_kernel_nTS, xgemm_kernel_lTS, xgemm_kernel_rTS, xgemm_kernel_bTS,
  394. xgemm_betaTS,
  395. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  396. xgemm_incopyTS, xgemm_itcopyTS,
  397. #else
  398. xgemm_oncopyTS, xgemm_otcopyTS,
  399. #endif
  400. xgemm_oncopyTS, xgemm_otcopyTS,
  401. xtrsm_kernel_LNTS, xtrsm_kernel_LTTS, xtrsm_kernel_LRTS, xtrsm_kernel_LCTS,
  402. xtrsm_kernel_RNTS, xtrsm_kernel_RTTS, xtrsm_kernel_RRTS, xtrsm_kernel_RCTS,
  403. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  404. xtrsm_iunucopyTS, xtrsm_iunncopyTS, xtrsm_iutucopyTS, xtrsm_iutncopyTS,
  405. xtrsm_ilnucopyTS, xtrsm_ilnncopyTS, xtrsm_iltucopyTS, xtrsm_iltncopyTS,
  406. #else
  407. xtrsm_ounucopyTS, xtrsm_ounncopyTS, xtrsm_outucopyTS, xtrsm_outncopyTS,
  408. xtrsm_olnucopyTS, xtrsm_olnncopyTS, xtrsm_oltucopyTS, xtrsm_oltncopyTS,
  409. #endif
  410. xtrsm_ounucopyTS, xtrsm_ounncopyTS, xtrsm_outucopyTS, xtrsm_outncopyTS,
  411. xtrsm_olnucopyTS, xtrsm_olnncopyTS, xtrsm_oltucopyTS, xtrsm_oltncopyTS,
  412. xtrmm_kernel_RNTS, xtrmm_kernel_RTTS, xtrmm_kernel_RRTS, xtrmm_kernel_RCTS,
  413. xtrmm_kernel_LNTS, xtrmm_kernel_LTTS, xtrmm_kernel_LRTS, xtrmm_kernel_LCTS,
  414. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  415. xtrmm_iunucopyTS, xtrmm_iunncopyTS, xtrmm_iutucopyTS, xtrmm_iutncopyTS,
  416. xtrmm_ilnucopyTS, xtrmm_ilnncopyTS, xtrmm_iltucopyTS, xtrmm_iltncopyTS,
  417. #else
  418. xtrmm_ounucopyTS, xtrmm_ounncopyTS, xtrmm_outucopyTS, xtrmm_outncopyTS,
  419. xtrmm_olnucopyTS, xtrmm_olnncopyTS, xtrmm_oltucopyTS, xtrmm_oltncopyTS,
  420. #endif
  421. xtrmm_ounucopyTS, xtrmm_ounncopyTS, xtrmm_outucopyTS, xtrmm_outncopyTS,
  422. xtrmm_olnucopyTS, xtrmm_olnncopyTS, xtrmm_oltucopyTS, xtrmm_oltncopyTS,
  423. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  424. xsymm_iutcopyTS, xsymm_iltcopyTS,
  425. #else
  426. xsymm_outcopyTS, xsymm_oltcopyTS,
  427. #endif
  428. xsymm_outcopyTS, xsymm_oltcopyTS,
  429. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  430. xhemm_iutcopyTS, xhemm_iltcopyTS,
  431. #else
  432. xhemm_outcopyTS, xhemm_oltcopyTS,
  433. #endif
  434. xhemm_outcopyTS, xhemm_oltcopyTS,
  435. 0, 0, 0,
  436. QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N, MAX(QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N),
  437. xgemm3m_kernelTS,
  438. xgemm3m_incopybTS, xgemm3m_incopyrTS,
  439. xgemm3m_incopyiTS, xgemm3m_itcopybTS,
  440. xgemm3m_itcopyrTS, xgemm3m_itcopyiTS,
  441. xgemm3m_oncopybTS, xgemm3m_oncopyrTS,
  442. xgemm3m_oncopyiTS, xgemm3m_otcopybTS,
  443. xgemm3m_otcopyrTS, xgemm3m_otcopyiTS,
  444. xsymm3m_iucopybTS, xsymm3m_ilcopybTS,
  445. xsymm3m_iucopyrTS, xsymm3m_ilcopyrTS,
  446. xsymm3m_iucopyiTS, xsymm3m_ilcopyiTS,
  447. xsymm3m_oucopybTS, xsymm3m_olcopybTS,
  448. xsymm3m_oucopyrTS, xsymm3m_olcopyrTS,
  449. xsymm3m_oucopyiTS, xsymm3m_olcopyiTS,
  450. xhemm3m_iucopybTS, xhemm3m_ilcopybTS,
  451. xhemm3m_iucopyrTS, xhemm3m_ilcopyrTS,
  452. xhemm3m_iucopyiTS, xhemm3m_ilcopyiTS,
  453. xhemm3m_oucopybTS, xhemm3m_olcopybTS,
  454. xhemm3m_oucopyrTS, xhemm3m_olcopyrTS,
  455. xhemm3m_oucopyiTS, xhemm3m_olcopyiTS,
  456. #ifndef NO_LAPACK
  457. xneg_tcopyTS, xlaswp_ncopyTS,
  458. #else
  459. NULL, NULL,
  460. #endif
  461. #endif
  462. init_parameter,
  463. SNUMOPT, DNUMOPT, QNUMOPT,
  464. saxpby_kTS, daxpby_kTS, caxpby_kTS, zaxpby_kTS,
  465. somatcopy_k_cnTS, somatcopy_k_ctTS, somatcopy_k_rnTS, somatcopy_k_rtTS,
  466. domatcopy_k_cnTS, domatcopy_k_ctTS, domatcopy_k_rnTS, domatcopy_k_rtTS,
  467. comatcopy_k_cnTS, comatcopy_k_ctTS, comatcopy_k_rnTS, comatcopy_k_rtTS,
  468. comatcopy_k_cncTS, comatcopy_k_ctcTS, comatcopy_k_rncTS, comatcopy_k_rtcTS,
  469. zomatcopy_k_cnTS, zomatcopy_k_ctTS, zomatcopy_k_rnTS, zomatcopy_k_rtTS,
  470. zomatcopy_k_cncTS, zomatcopy_k_ctcTS, zomatcopy_k_rncTS, zomatcopy_k_rtcTS
  471. };
  472. #ifdef ARCH_X86
  473. static int get_l2_size_old(void){
  474. int i, eax, ebx, ecx, edx, cpuid_level;
  475. int info[15];
  476. cpuid(2, &eax, &ebx, &ecx, &edx);
  477. info[ 0] = BITMASK(eax, 8, 0xff);
  478. info[ 1] = BITMASK(eax, 16, 0xff);
  479. info[ 2] = BITMASK(eax, 24, 0xff);
  480. info[ 3] = BITMASK(ebx, 0, 0xff);
  481. info[ 4] = BITMASK(ebx, 8, 0xff);
  482. info[ 5] = BITMASK(ebx, 16, 0xff);
  483. info[ 6] = BITMASK(ebx, 24, 0xff);
  484. info[ 7] = BITMASK(ecx, 0, 0xff);
  485. info[ 8] = BITMASK(ecx, 8, 0xff);
  486. info[ 9] = BITMASK(ecx, 16, 0xff);
  487. info[10] = BITMASK(ecx, 24, 0xff);
  488. info[11] = BITMASK(edx, 0, 0xff);
  489. info[12] = BITMASK(edx, 8, 0xff);
  490. info[13] = BITMASK(edx, 16, 0xff);
  491. info[14] = BITMASK(edx, 24, 0xff);
  492. for (i = 0; i < 15; i++){
  493. switch (info[i]){
  494. /* This table is from http://www.sandpile.org/ia32/cpuid.htm */
  495. case 0x1a :
  496. return 96;
  497. case 0x39 :
  498. case 0x3b :
  499. case 0x41 :
  500. case 0x79 :
  501. case 0x81 :
  502. return 128;
  503. case 0x3a :
  504. return 192;
  505. case 0x21 :
  506. case 0x3c :
  507. case 0x42 :
  508. case 0x7a :
  509. case 0x7e :
  510. case 0x82 :
  511. return 256;
  512. case 0x3d :
  513. return 384;
  514. case 0x3e :
  515. case 0x43 :
  516. case 0x7b :
  517. case 0x7f :
  518. case 0x83 :
  519. case 0x86 :
  520. return 512;
  521. case 0x44 :
  522. case 0x78 :
  523. case 0x7c :
  524. case 0x84 :
  525. case 0x87 :
  526. return 1024;
  527. case 0x45 :
  528. case 0x7d :
  529. case 0x85 :
  530. return 2048;
  531. case 0x48 :
  532. return 3184;
  533. case 0x49 :
  534. return 4096;
  535. case 0x4e :
  536. return 6144;
  537. }
  538. }
  539. return 0;
  540. }
  541. #endif
  542. static __inline__ int get_l2_size(void){
  543. int eax, ebx, ecx, edx, l2;
  544. cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
  545. l2 = BITMASK(ecx, 16, 0xffff);
  546. #ifndef ARCH_X86
  547. return l2;
  548. #else
  549. if (l2 > 0) return l2;
  550. return get_l2_size_old();
  551. #endif
  552. }
  553. static __inline__ int get_l3_size(void){
  554. int eax, ebx, ecx, edx;
  555. cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
  556. return BITMASK(edx, 18, 0x3fff) * 512;
  557. }
  558. static void init_parameter(void) {
  559. int l2 = get_l2_size();
  560. TABLE_NAME.sgemm_q = SGEMM_DEFAULT_Q;
  561. TABLE_NAME.dgemm_q = DGEMM_DEFAULT_Q;
  562. TABLE_NAME.cgemm_q = CGEMM_DEFAULT_Q;
  563. TABLE_NAME.zgemm_q = ZGEMM_DEFAULT_Q;
  564. #ifdef CGEMM3M_DEFAULT_Q
  565. TABLE_NAME.cgemm3m_q = CGEMM3M_DEFAULT_Q;
  566. #else
  567. TABLE_NAME.cgemm3m_q = SGEMM_DEFAULT_Q;
  568. #endif
  569. #ifdef ZGEMM3M_DEFAULT_Q
  570. TABLE_NAME.zgemm3m_q = ZGEMM3M_DEFAULT_Q;
  571. #else
  572. TABLE_NAME.zgemm3m_q = DGEMM_DEFAULT_Q;
  573. #endif
  574. #ifdef EXPRECISION
  575. TABLE_NAME.qgemm_q = QGEMM_DEFAULT_Q;
  576. TABLE_NAME.xgemm_q = XGEMM_DEFAULT_Q;
  577. TABLE_NAME.xgemm3m_q = QGEMM_DEFAULT_Q;
  578. #endif
  579. #if defined(CORE_KATMAI) || defined(CORE_COPPERMINE) || defined(CORE_BANIAS) || defined(CORE_YONAH) || defined(CORE_ATHLON)
  580. #ifdef DEBUG
  581. fprintf(stderr, "Katmai, Coppermine, Banias, Athlon\n");
  582. #endif
  583. TABLE_NAME.sgemm_p = 64 * (l2 >> 7);
  584. TABLE_NAME.dgemm_p = 32 * (l2 >> 7);
  585. TABLE_NAME.cgemm_p = 32 * (l2 >> 7);
  586. TABLE_NAME.zgemm_p = 16 * (l2 >> 7);
  587. #ifdef EXPRECISION
  588. TABLE_NAME.qgemm_p = 16 * (l2 >> 7);
  589. TABLE_NAME.xgemm_p = 8 * (l2 >> 7);
  590. #endif
  591. #endif
  592. #ifdef CORE_NORTHWOOD
  593. #ifdef DEBUG
  594. fprintf(stderr, "Northwood\n");
  595. #endif
  596. TABLE_NAME.sgemm_p = 96 * (l2 >> 7);
  597. TABLE_NAME.dgemm_p = 48 * (l2 >> 7);
  598. TABLE_NAME.cgemm_p = 48 * (l2 >> 7);
  599. TABLE_NAME.zgemm_p = 24 * (l2 >> 7);
  600. #ifdef EXPRECISION
  601. TABLE_NAME.qgemm_p = 24 * (l2 >> 7);
  602. TABLE_NAME.xgemm_p = 12 * (l2 >> 7);
  603. #endif
  604. #endif
  605. #ifdef ATOM
  606. #ifdef DEBUG
  607. fprintf(stderr, "Atom\n");
  608. #endif
  609. TABLE_NAME.sgemm_p = 256;
  610. TABLE_NAME.dgemm_p = 128;
  611. TABLE_NAME.cgemm_p = 128;
  612. TABLE_NAME.zgemm_p = 64;
  613. #ifdef EXPRECISION
  614. TABLE_NAME.qgemm_p = 64;
  615. TABLE_NAME.xgemm_p = 32;
  616. #endif
  617. #endif
  618. #ifdef CORE_PRESCOTT
  619. #ifdef DEBUG
  620. fprintf(stderr, "Prescott\n");
  621. #endif
  622. TABLE_NAME.sgemm_p = 56 * (l2 >> 7);
  623. TABLE_NAME.dgemm_p = 28 * (l2 >> 7);
  624. TABLE_NAME.cgemm_p = 28 * (l2 >> 7);
  625. TABLE_NAME.zgemm_p = 14 * (l2 >> 7);
  626. #ifdef EXPRECISION
  627. TABLE_NAME.qgemm_p = 14 * (l2 >> 7);
  628. TABLE_NAME.xgemm_p = 7 * (l2 >> 7);
  629. #endif
  630. #endif
  631. #ifdef CORE2
  632. #ifdef DEBUG
  633. fprintf(stderr, "Core2\n");
  634. #endif
  635. TABLE_NAME.sgemm_p = 92 * (l2 >> 9) + 8;
  636. TABLE_NAME.dgemm_p = 46 * (l2 >> 9) + 8;
  637. TABLE_NAME.cgemm_p = 46 * (l2 >> 9) + 4;
  638. TABLE_NAME.zgemm_p = 23 * (l2 >> 9) + 4;
  639. #ifdef EXPRECISION
  640. TABLE_NAME.qgemm_p = 92 * (l2 >> 9) + 8;
  641. TABLE_NAME.xgemm_p = 46 * (l2 >> 9) + 4;
  642. #endif
  643. #endif
  644. #ifdef PENRYN
  645. #ifdef DEBUG
  646. fprintf(stderr, "Penryn\n");
  647. #endif
  648. TABLE_NAME.sgemm_p = 42 * (l2 >> 9) + 8;
  649. TABLE_NAME.dgemm_p = 42 * (l2 >> 9) + 8;
  650. TABLE_NAME.cgemm_p = 21 * (l2 >> 9) + 4;
  651. TABLE_NAME.zgemm_p = 21 * (l2 >> 9) + 4;
  652. #ifdef EXPRECISION
  653. TABLE_NAME.qgemm_p = 42 * (l2 >> 9) + 8;
  654. TABLE_NAME.xgemm_p = 21 * (l2 >> 9) + 4;
  655. #endif
  656. #endif
  657. #ifdef DUNNINGTON
  658. #ifdef DEBUG
  659. fprintf(stderr, "Dunnington\n");
  660. #endif
  661. TABLE_NAME.sgemm_p = 42 * (l2 >> 9) + 8;
  662. TABLE_NAME.dgemm_p = 42 * (l2 >> 9) + 8;
  663. TABLE_NAME.cgemm_p = 21 * (l2 >> 9) + 4;
  664. TABLE_NAME.zgemm_p = 21 * (l2 >> 9) + 4;
  665. #ifdef EXPRECISION
  666. TABLE_NAME.qgemm_p = 42 * (l2 >> 9) + 8;
  667. TABLE_NAME.xgemm_p = 21 * (l2 >> 9) + 4;
  668. #endif
  669. #endif
  670. #ifdef NEHALEM
  671. #ifdef DEBUG
  672. fprintf(stderr, "Nehalem\n");
  673. #endif
  674. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  675. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  676. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  677. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  678. #ifdef EXPRECISION
  679. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  680. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  681. #endif
  682. #endif
  683. #ifdef SANDYBRIDGE
  684. #ifdef DEBUG
  685. fprintf(stderr, "Sandybridge\n");
  686. #endif
  687. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  688. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  689. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  690. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  691. #ifdef EXPRECISION
  692. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  693. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  694. #endif
  695. #endif
  696. #ifdef HASWELL
  697. #ifdef DEBUG
  698. fprintf(stderr, "Haswell\n");
  699. #endif
  700. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  701. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  702. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  703. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  704. #ifdef EXPRECISION
  705. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  706. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  707. #endif
  708. #endif
  709. #ifdef OPTERON
  710. #ifdef DEBUG
  711. fprintf(stderr, "Opteron\n");
  712. #endif
  713. TABLE_NAME.sgemm_p = 224 + 56 * (l2 >> 7);
  714. TABLE_NAME.dgemm_p = 112 + 28 * (l2 >> 7);
  715. TABLE_NAME.cgemm_p = 112 + 28 * (l2 >> 7);
  716. TABLE_NAME.zgemm_p = 56 + 14 * (l2 >> 7);
  717. #ifdef EXPRECISION
  718. TABLE_NAME.qgemm_p = 56 + 14 * (l2 >> 7);
  719. TABLE_NAME.xgemm_p = 28 + 7 * (l2 >> 7);
  720. #endif
  721. #endif
  722. #ifdef BARCELONA
  723. #ifdef DEBUG
  724. fprintf(stderr, "Barcelona\n");
  725. #endif
  726. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  727. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  728. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  729. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  730. #ifdef EXPRECISION
  731. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  732. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  733. #endif
  734. #endif
  735. #ifdef BOBCAT
  736. #ifdef DEBUG
  737. fprintf(stderr, "Bobcate\n");
  738. #endif
  739. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  740. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  741. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  742. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  743. #ifdef EXPRECISION
  744. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  745. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  746. #endif
  747. #endif
  748. #ifdef BULLDOZER
  749. #ifdef DEBUG
  750. fprintf(stderr, "Bulldozer\n");
  751. #endif
  752. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  753. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  754. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  755. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  756. #ifdef EXPRECISION
  757. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  758. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  759. #endif
  760. #endif
  761. #ifdef PILEDRIVER
  762. #ifdef DEBUG
  763. fprintf(stderr, "Piledriver\n");
  764. #endif
  765. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  766. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  767. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  768. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  769. #ifdef EXPRECISION
  770. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  771. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  772. #endif
  773. #endif
  774. #ifdef STEAMROLLER
  775. #ifdef DEBUG
  776. fprintf(stderr, "Steamroller\n");
  777. #endif
  778. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  779. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  780. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  781. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  782. #ifdef EXPRECISION
  783. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  784. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  785. #endif
  786. #endif
  787. #ifdef NANO
  788. #ifdef DEBUG
  789. fprintf(stderr, "NANO\n");
  790. #endif
  791. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  792. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  793. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  794. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  795. #ifdef EXPRECISION
  796. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  797. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  798. #endif
  799. #endif
  800. #ifdef CGEMM3M_DEFAULT_P
  801. TABLE_NAME.cgemm3m_p = CGEMM3M_DEFAULT_P;
  802. #else
  803. TABLE_NAME.cgemm3m_p = TABLE_NAME.sgemm_p;
  804. #endif
  805. #ifdef ZGEMM3M_DEFAULT_P
  806. TABLE_NAME.zgemm3m_p = ZGEMM3M_DEFAULT_P;
  807. #else
  808. TABLE_NAME.zgemm3m_p = TABLE_NAME.dgemm_p;
  809. #endif
  810. #ifdef EXPRECISION
  811. TABLE_NAME.xgemm3m_p = TABLE_NAME.qgemm_p;
  812. #endif
  813. TABLE_NAME.sgemm_p = (TABLE_NAME.sgemm_p + SGEMM_DEFAULT_UNROLL_M - 1) & ~(SGEMM_DEFAULT_UNROLL_M - 1);
  814. TABLE_NAME.dgemm_p = (TABLE_NAME.dgemm_p + DGEMM_DEFAULT_UNROLL_M - 1) & ~(DGEMM_DEFAULT_UNROLL_M - 1);
  815. TABLE_NAME.cgemm_p = (TABLE_NAME.cgemm_p + CGEMM_DEFAULT_UNROLL_M - 1) & ~(CGEMM_DEFAULT_UNROLL_M - 1);
  816. TABLE_NAME.zgemm_p = (TABLE_NAME.zgemm_p + ZGEMM_DEFAULT_UNROLL_M - 1) & ~(ZGEMM_DEFAULT_UNROLL_M - 1);
  817. #ifdef CGEMM3M_DEFAULT_UNROLL_M
  818. TABLE_NAME.cgemm3m_p = (TABLE_NAME.cgemm3m_p + CGEMM3M_DEFAULT_UNROLL_M - 1) & ~(CGEMM3M_DEFAULT_UNROLL_M - 1);
  819. #else
  820. TABLE_NAME.cgemm3m_p = (TABLE_NAME.cgemm3m_p + SGEMM_DEFAULT_UNROLL_M - 1) & ~(SGEMM_DEFAULT_UNROLL_M - 1);
  821. #endif
  822. #ifdef ZGEMM3M_DEFAULT_UNROLL_M
  823. TABLE_NAME.zgemm3m_p = (TABLE_NAME.zgemm3m_p + ZGEMM3M_DEFAULT_UNROLL_M - 1) & ~(ZGEMM3M_DEFAULT_UNROLL_M - 1);
  824. #else
  825. TABLE_NAME.zgemm3m_p = (TABLE_NAME.zgemm3m_p + DGEMM_DEFAULT_UNROLL_M - 1) & ~(DGEMM_DEFAULT_UNROLL_M - 1);
  826. #endif
  827. #ifdef QUAD_PRECISION
  828. TABLE_NAME.qgemm_p = (TABLE_NAME.qgemm_p + QGEMM_DEFAULT_UNROLL_M - 1) & ~(QGEMM_DEFAULT_UNROLL_M - 1);
  829. TABLE_NAME.xgemm_p = (TABLE_NAME.xgemm_p + XGEMM_DEFAULT_UNROLL_M - 1) & ~(XGEMM_DEFAULT_UNROLL_M - 1);
  830. TABLE_NAME.xgemm3m_p = (TABLE_NAME.xgemm3m_p + QGEMM_DEFAULT_UNROLL_M - 1) & ~(QGEMM_DEFAULT_UNROLL_M - 1);
  831. #endif
  832. #ifdef DEBUG
  833. fprintf(stderr, "L2 = %8d DGEMM_P .. %d\n", l2, TABLE_NAME.dgemm_p);
  834. #endif
  835. TABLE_NAME.sgemm_r = (((BUFFER_SIZE -
  836. ((TABLE_NAME.sgemm_p * TABLE_NAME.sgemm_q * 4 + TABLE_NAME.offsetA
  837. + TABLE_NAME.align) & ~TABLE_NAME.align)
  838. ) / (TABLE_NAME.sgemm_q * 4) - 15) & ~15);
  839. TABLE_NAME.dgemm_r = (((BUFFER_SIZE -
  840. ((TABLE_NAME.dgemm_p * TABLE_NAME.dgemm_q * 8 + TABLE_NAME.offsetA
  841. + TABLE_NAME.align) & ~TABLE_NAME.align)
  842. ) / (TABLE_NAME.dgemm_q * 8) - 15) & ~15);
  843. #ifdef EXPRECISION
  844. TABLE_NAME.qgemm_r = (((BUFFER_SIZE -
  845. ((TABLE_NAME.qgemm_p * TABLE_NAME.qgemm_q * 16 + TABLE_NAME.offsetA
  846. + TABLE_NAME.align) & ~TABLE_NAME.align)
  847. ) / (TABLE_NAME.qgemm_q * 16) - 15) & ~15);
  848. #endif
  849. TABLE_NAME.cgemm_r = (((BUFFER_SIZE -
  850. ((TABLE_NAME.cgemm_p * TABLE_NAME.cgemm_q * 8 + TABLE_NAME.offsetA
  851. + TABLE_NAME.align) & ~TABLE_NAME.align)
  852. ) / (TABLE_NAME.cgemm_q * 8) - 15) & ~15);
  853. TABLE_NAME.zgemm_r = (((BUFFER_SIZE -
  854. ((TABLE_NAME.zgemm_p * TABLE_NAME.zgemm_q * 16 + TABLE_NAME.offsetA
  855. + TABLE_NAME.align) & ~TABLE_NAME.align)
  856. ) / (TABLE_NAME.zgemm_q * 16) - 15) & ~15);
  857. TABLE_NAME.cgemm3m_r = (((BUFFER_SIZE -
  858. ((TABLE_NAME.cgemm3m_p * TABLE_NAME.cgemm3m_q * 8 + TABLE_NAME.offsetA
  859. + TABLE_NAME.align) & ~TABLE_NAME.align)
  860. ) / (TABLE_NAME.cgemm3m_q * 8) - 15) & ~15);
  861. TABLE_NAME.zgemm3m_r = (((BUFFER_SIZE -
  862. ((TABLE_NAME.zgemm3m_p * TABLE_NAME.zgemm3m_q * 16 + TABLE_NAME.offsetA
  863. + TABLE_NAME.align) & ~TABLE_NAME.align)
  864. ) / (TABLE_NAME.zgemm3m_q * 16) - 15) & ~15);
  865. #ifdef EXPRECISION
  866. TABLE_NAME.xgemm_r = (((BUFFER_SIZE -
  867. ((TABLE_NAME.xgemm_p * TABLE_NAME.xgemm_q * 32 + TABLE_NAME.offsetA
  868. + TABLE_NAME.align) & ~TABLE_NAME.align)
  869. ) / (TABLE_NAME.xgemm_q * 32) - 15) & ~15);
  870. TABLE_NAME.xgemm3m_r = (((BUFFER_SIZE -
  871. ((TABLE_NAME.xgemm3m_p * TABLE_NAME.xgemm3m_q * 32 + TABLE_NAME.offsetA
  872. + TABLE_NAME.align) & ~TABLE_NAME.align)
  873. ) / (TABLE_NAME.xgemm3m_q * 32) - 15) & ~15);
  874. #endif
  875. }