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 35 kB

12 years ago
12 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146
  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 "common.h"
  39. #ifdef BUILD_KERNEL
  40. #include "kernelTS.h"
  41. #endif
  42. #undef DEBUG
  43. static void init_parameter(void);
  44. gotoblas_t TABLE_NAME = {
  45. DTB_DEFAULT_ENTRIES ,
  46. GEMM_DEFAULT_OFFSET_A, GEMM_DEFAULT_OFFSET_B, GEMM_DEFAULT_ALIGN,
  47. 0, 0, 0,
  48. SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N,
  49. #ifdef SGEMM_DEFAULT_UNROLL_MN
  50. SGEMM_DEFAULT_UNROLL_MN,
  51. #else
  52. MAX(SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N),
  53. #endif
  54. #ifdef HAVE_EXCLUSIVE_CACHE
  55. 1,
  56. #else
  57. 0,
  58. #endif
  59. samax_kTS, samin_kTS, smax_kTS, smin_kTS,
  60. isamax_kTS, isamin_kTS, ismax_kTS, ismin_kTS,
  61. snrm2_kTS, sasum_kTS, scopy_kTS, sdot_kTS,
  62. dsdot_kTS,
  63. srot_kTS, saxpy_kTS, sscal_kTS, sswap_kTS,
  64. sgemv_nTS, sgemv_tTS, sger_kTS,
  65. ssymv_LTS, ssymv_UTS,
  66. sgemm_kernelTS, sgemm_betaTS,
  67. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  68. sgemm_incopyTS, sgemm_itcopyTS,
  69. #else
  70. sgemm_oncopyTS, sgemm_otcopyTS,
  71. #endif
  72. sgemm_oncopyTS, sgemm_otcopyTS,
  73. strsm_kernel_LNTS, strsm_kernel_LTTS, strsm_kernel_RNTS, strsm_kernel_RTTS,
  74. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  75. strsm_iunucopyTS, strsm_iunncopyTS, strsm_iutucopyTS, strsm_iutncopyTS,
  76. strsm_ilnucopyTS, strsm_ilnncopyTS, strsm_iltucopyTS, strsm_iltncopyTS,
  77. #else
  78. strsm_ounucopyTS, strsm_ounncopyTS, strsm_outucopyTS, strsm_outncopyTS,
  79. strsm_olnucopyTS, strsm_olnncopyTS, strsm_oltucopyTS, strsm_oltncopyTS,
  80. #endif
  81. strsm_ounucopyTS, strsm_ounncopyTS, strsm_outucopyTS, strsm_outncopyTS,
  82. strsm_olnucopyTS, strsm_olnncopyTS, strsm_oltucopyTS, strsm_oltncopyTS,
  83. strmm_kernel_RNTS, strmm_kernel_RTTS, strmm_kernel_LNTS, strmm_kernel_LTTS,
  84. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  85. strmm_iunucopyTS, strmm_iunncopyTS, strmm_iutucopyTS, strmm_iutncopyTS,
  86. strmm_ilnucopyTS, strmm_ilnncopyTS, strmm_iltucopyTS, strmm_iltncopyTS,
  87. #else
  88. strmm_ounucopyTS, strmm_ounncopyTS, strmm_outucopyTS, strmm_outncopyTS,
  89. strmm_olnucopyTS, strmm_olnncopyTS, strmm_oltucopyTS, strmm_oltncopyTS,
  90. #endif
  91. strmm_ounucopyTS, strmm_ounncopyTS, strmm_outucopyTS, strmm_outncopyTS,
  92. strmm_olnucopyTS, strmm_olnncopyTS, strmm_oltucopyTS, strmm_oltncopyTS,
  93. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  94. ssymm_iutcopyTS, ssymm_iltcopyTS,
  95. #else
  96. ssymm_outcopyTS, ssymm_oltcopyTS,
  97. #endif
  98. ssymm_outcopyTS, ssymm_oltcopyTS,
  99. #ifndef NO_LAPACK
  100. sneg_tcopyTS, slaswp_ncopyTS,
  101. #else
  102. NULL,NULL,
  103. #endif
  104. 0, 0, 0,
  105. DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N,
  106. #ifdef DGEMM_DEFAULT_UNROLL_MN
  107. DGEMM_DEFAULT_UNROLL_MN,
  108. #else
  109. MAX(DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N),
  110. #endif
  111. damax_kTS, damin_kTS, dmax_kTS, dmin_kTS,
  112. idamax_kTS, idamin_kTS, idmax_kTS, idmin_kTS,
  113. dnrm2_kTS, dasum_kTS, dcopy_kTS, ddot_kTS,
  114. drot_kTS, daxpy_kTS, dscal_kTS, dswap_kTS,
  115. dgemv_nTS, dgemv_tTS, dger_kTS,
  116. dsymv_LTS, dsymv_UTS,
  117. dgemm_kernelTS, dgemm_betaTS,
  118. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  119. dgemm_incopyTS, dgemm_itcopyTS,
  120. #else
  121. dgemm_oncopyTS, dgemm_otcopyTS,
  122. #endif
  123. dgemm_oncopyTS, dgemm_otcopyTS,
  124. dtrsm_kernel_LNTS, dtrsm_kernel_LTTS, dtrsm_kernel_RNTS, dtrsm_kernel_RTTS,
  125. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  126. dtrsm_iunucopyTS, dtrsm_iunncopyTS, dtrsm_iutucopyTS, dtrsm_iutncopyTS,
  127. dtrsm_ilnucopyTS, dtrsm_ilnncopyTS, dtrsm_iltucopyTS, dtrsm_iltncopyTS,
  128. #else
  129. dtrsm_ounucopyTS, dtrsm_ounncopyTS, dtrsm_outucopyTS, dtrsm_outncopyTS,
  130. dtrsm_olnucopyTS, dtrsm_olnncopyTS, dtrsm_oltucopyTS, dtrsm_oltncopyTS,
  131. #endif
  132. dtrsm_ounucopyTS, dtrsm_ounncopyTS, dtrsm_outucopyTS, dtrsm_outncopyTS,
  133. dtrsm_olnucopyTS, dtrsm_olnncopyTS, dtrsm_oltucopyTS, dtrsm_oltncopyTS,
  134. dtrmm_kernel_RNTS, dtrmm_kernel_RTTS, dtrmm_kernel_LNTS, dtrmm_kernel_LTTS,
  135. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  136. dtrmm_iunucopyTS, dtrmm_iunncopyTS, dtrmm_iutucopyTS, dtrmm_iutncopyTS,
  137. dtrmm_ilnucopyTS, dtrmm_ilnncopyTS, dtrmm_iltucopyTS, dtrmm_iltncopyTS,
  138. #else
  139. dtrmm_ounucopyTS, dtrmm_ounncopyTS, dtrmm_outucopyTS, dtrmm_outncopyTS,
  140. dtrmm_olnucopyTS, dtrmm_olnncopyTS, dtrmm_oltucopyTS, dtrmm_oltncopyTS,
  141. #endif
  142. dtrmm_ounucopyTS, dtrmm_ounncopyTS, dtrmm_outucopyTS, dtrmm_outncopyTS,
  143. dtrmm_olnucopyTS, dtrmm_olnncopyTS, dtrmm_oltucopyTS, dtrmm_oltncopyTS,
  144. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  145. dsymm_iutcopyTS, dsymm_iltcopyTS,
  146. #else
  147. dsymm_outcopyTS, dsymm_oltcopyTS,
  148. #endif
  149. dsymm_outcopyTS, dsymm_oltcopyTS,
  150. #ifndef NO_LAPACK
  151. dneg_tcopyTS, dlaswp_ncopyTS,
  152. #else
  153. NULL, NULL,
  154. #endif
  155. #ifdef EXPRECISION
  156. 0, 0, 0,
  157. QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N, MAX(QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N),
  158. qamax_kTS, qamin_kTS, qmax_kTS, qmin_kTS,
  159. iqamax_kTS, iqamin_kTS, iqmax_kTS, iqmin_kTS,
  160. qnrm2_kTS, qasum_kTS, qcopy_kTS, qdot_kTS,
  161. qrot_kTS, qaxpy_kTS, qscal_kTS, qswap_kTS,
  162. qgemv_nTS, qgemv_tTS, qger_kTS,
  163. qsymv_LTS, qsymv_UTS,
  164. qgemm_kernelTS, qgemm_betaTS,
  165. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  166. qgemm_incopyTS, qgemm_itcopyTS,
  167. #else
  168. qgemm_oncopyTS, qgemm_otcopyTS,
  169. #endif
  170. qgemm_oncopyTS, qgemm_otcopyTS,
  171. qtrsm_kernel_LNTS, qtrsm_kernel_LTTS, qtrsm_kernel_RNTS, qtrsm_kernel_RTTS,
  172. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  173. qtrsm_iunucopyTS, qtrsm_iunncopyTS, qtrsm_iutucopyTS, qtrsm_iutncopyTS,
  174. qtrsm_ilnucopyTS, qtrsm_ilnncopyTS, qtrsm_iltucopyTS, qtrsm_iltncopyTS,
  175. #else
  176. qtrsm_ounucopyTS, qtrsm_ounncopyTS, qtrsm_outucopyTS, qtrsm_outncopyTS,
  177. qtrsm_olnucopyTS, qtrsm_olnncopyTS, qtrsm_oltucopyTS, qtrsm_oltncopyTS,
  178. #endif
  179. qtrsm_ounucopyTS, qtrsm_ounncopyTS, qtrsm_outucopyTS, qtrsm_outncopyTS,
  180. qtrsm_olnucopyTS, qtrsm_olnncopyTS, qtrsm_oltucopyTS, qtrsm_oltncopyTS,
  181. qtrmm_kernel_RNTS, qtrmm_kernel_RTTS, qtrmm_kernel_LNTS, qtrmm_kernel_LTTS,
  182. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  183. qtrmm_iunucopyTS, qtrmm_iunncopyTS, qtrmm_iutucopyTS, qtrmm_iutncopyTS,
  184. qtrmm_ilnucopyTS, qtrmm_ilnncopyTS, qtrmm_iltucopyTS, qtrmm_iltncopyTS,
  185. #else
  186. qtrmm_ounucopyTS, qtrmm_ounncopyTS, qtrmm_outucopyTS, qtrmm_outncopyTS,
  187. qtrmm_olnucopyTS, qtrmm_olnncopyTS, qtrmm_oltucopyTS, qtrmm_oltncopyTS,
  188. #endif
  189. qtrmm_ounucopyTS, qtrmm_ounncopyTS, qtrmm_outucopyTS, qtrmm_outncopyTS,
  190. qtrmm_olnucopyTS, qtrmm_olnncopyTS, qtrmm_oltucopyTS, qtrmm_oltncopyTS,
  191. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  192. qsymm_iutcopyTS, qsymm_iltcopyTS,
  193. #else
  194. qsymm_outcopyTS, qsymm_oltcopyTS,
  195. #endif
  196. qsymm_outcopyTS, qsymm_oltcopyTS,
  197. #ifndef NO_LAPACK
  198. qneg_tcopyTS, qlaswp_ncopyTS,
  199. #else
  200. NULL, NULL,
  201. #endif
  202. #endif
  203. 0, 0, 0,
  204. CGEMM_DEFAULT_UNROLL_M, CGEMM_DEFAULT_UNROLL_N,
  205. #ifdef CGEMM_DEFAULT_UNROLL_MN
  206. CGEMM_DEFAULT_UNROLL_MN,
  207. #else
  208. MAX(CGEMM_DEFAULT_UNROLL_M, CGEMM_DEFAULT_UNROLL_N),
  209. #endif
  210. camax_kTS, camin_kTS, icamax_kTS, icamin_kTS,
  211. cnrm2_kTS, casum_kTS, ccopy_kTS,
  212. cdotu_kTS, cdotc_kTS, csrot_kTS,
  213. caxpy_kTS, caxpyc_kTS, cscal_kTS, cswap_kTS,
  214. cgemv_nTS, cgemv_tTS, cgemv_rTS, cgemv_cTS,
  215. cgemv_oTS, cgemv_uTS, cgemv_sTS, cgemv_dTS,
  216. cgeru_kTS, cgerc_kTS, cgerv_kTS, cgerd_kTS,
  217. csymv_LTS, csymv_UTS,
  218. chemv_LTS, chemv_UTS, chemv_MTS, chemv_VTS,
  219. cgemm_kernel_nTS, cgemm_kernel_lTS, cgemm_kernel_rTS, cgemm_kernel_bTS,
  220. cgemm_betaTS,
  221. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  222. cgemm_incopyTS, cgemm_itcopyTS,
  223. #else
  224. cgemm_oncopyTS, cgemm_otcopyTS,
  225. #endif
  226. cgemm_oncopyTS, cgemm_otcopyTS,
  227. ctrsm_kernel_LNTS, ctrsm_kernel_LTTS, ctrsm_kernel_LRTS, ctrsm_kernel_LCTS,
  228. ctrsm_kernel_RNTS, ctrsm_kernel_RTTS, ctrsm_kernel_RRTS, ctrsm_kernel_RCTS,
  229. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  230. ctrsm_iunucopyTS, ctrsm_iunncopyTS, ctrsm_iutucopyTS, ctrsm_iutncopyTS,
  231. ctrsm_ilnucopyTS, ctrsm_ilnncopyTS, ctrsm_iltucopyTS, ctrsm_iltncopyTS,
  232. #else
  233. ctrsm_ounucopyTS, ctrsm_ounncopyTS, ctrsm_outucopyTS, ctrsm_outncopyTS,
  234. ctrsm_olnucopyTS, ctrsm_olnncopyTS, ctrsm_oltucopyTS, ctrsm_oltncopyTS,
  235. #endif
  236. ctrsm_ounucopyTS, ctrsm_ounncopyTS, ctrsm_outucopyTS, ctrsm_outncopyTS,
  237. ctrsm_olnucopyTS, ctrsm_olnncopyTS, ctrsm_oltucopyTS, ctrsm_oltncopyTS,
  238. ctrmm_kernel_RNTS, ctrmm_kernel_RTTS, ctrmm_kernel_RRTS, ctrmm_kernel_RCTS,
  239. ctrmm_kernel_LNTS, ctrmm_kernel_LTTS, ctrmm_kernel_LRTS, ctrmm_kernel_LCTS,
  240. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  241. ctrmm_iunucopyTS, ctrmm_iunncopyTS, ctrmm_iutucopyTS, ctrmm_iutncopyTS,
  242. ctrmm_ilnucopyTS, ctrmm_ilnncopyTS, ctrmm_iltucopyTS, ctrmm_iltncopyTS,
  243. #else
  244. ctrmm_ounucopyTS, ctrmm_ounncopyTS, ctrmm_outucopyTS, ctrmm_outncopyTS,
  245. ctrmm_olnucopyTS, ctrmm_olnncopyTS, ctrmm_oltucopyTS, ctrmm_oltncopyTS,
  246. #endif
  247. ctrmm_ounucopyTS, ctrmm_ounncopyTS, ctrmm_outucopyTS, ctrmm_outncopyTS,
  248. ctrmm_olnucopyTS, ctrmm_olnncopyTS, ctrmm_oltucopyTS, ctrmm_oltncopyTS,
  249. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  250. csymm_iutcopyTS, csymm_iltcopyTS,
  251. #else
  252. csymm_outcopyTS, csymm_oltcopyTS,
  253. #endif
  254. csymm_outcopyTS, csymm_oltcopyTS,
  255. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  256. chemm_iutcopyTS, chemm_iltcopyTS,
  257. #else
  258. chemm_outcopyTS, chemm_oltcopyTS,
  259. #endif
  260. chemm_outcopyTS, chemm_oltcopyTS,
  261. 0, 0, 0,
  262. #ifdef CGEMM3M_DEFAULT_UNROLL_M
  263. CGEMM3M_DEFAULT_UNROLL_M, CGEMM3M_DEFAULT_UNROLL_N, MAX(CGEMM3M_DEFAULT_UNROLL_M, CGEMM3M_DEFAULT_UNROLL_N),
  264. #else
  265. SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N, MAX(SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N),
  266. #endif
  267. cgemm3m_kernelTS,
  268. cgemm3m_incopybTS, cgemm3m_incopyrTS,
  269. cgemm3m_incopyiTS, cgemm3m_itcopybTS,
  270. cgemm3m_itcopyrTS, cgemm3m_itcopyiTS,
  271. cgemm3m_oncopybTS, cgemm3m_oncopyrTS,
  272. cgemm3m_oncopyiTS, cgemm3m_otcopybTS,
  273. cgemm3m_otcopyrTS, cgemm3m_otcopyiTS,
  274. csymm3m_iucopybTS, csymm3m_ilcopybTS,
  275. csymm3m_iucopyrTS, csymm3m_ilcopyrTS,
  276. csymm3m_iucopyiTS, csymm3m_ilcopyiTS,
  277. csymm3m_oucopybTS, csymm3m_olcopybTS,
  278. csymm3m_oucopyrTS, csymm3m_olcopyrTS,
  279. csymm3m_oucopyiTS, csymm3m_olcopyiTS,
  280. chemm3m_iucopybTS, chemm3m_ilcopybTS,
  281. chemm3m_iucopyrTS, chemm3m_ilcopyrTS,
  282. chemm3m_iucopyiTS, chemm3m_ilcopyiTS,
  283. chemm3m_oucopybTS, chemm3m_olcopybTS,
  284. chemm3m_oucopyrTS, chemm3m_olcopyrTS,
  285. chemm3m_oucopyiTS, chemm3m_olcopyiTS,
  286. #ifndef NO_LAPACK
  287. cneg_tcopyTS, claswp_ncopyTS,
  288. #else
  289. NULL, NULL,
  290. #endif
  291. 0, 0, 0,
  292. ZGEMM_DEFAULT_UNROLL_M, ZGEMM_DEFAULT_UNROLL_N,
  293. #ifdef ZGEMM_DEFAULT_UNROLL_MN
  294. ZGEMM_DEFAULT_UNROLL_MN,
  295. #else
  296. MAX(ZGEMM_DEFAULT_UNROLL_M, ZGEMM_DEFAULT_UNROLL_N),
  297. #endif
  298. zamax_kTS, zamin_kTS, izamax_kTS, izamin_kTS,
  299. znrm2_kTS, zasum_kTS, zcopy_kTS,
  300. zdotu_kTS, zdotc_kTS, zdrot_kTS,
  301. zaxpy_kTS, zaxpyc_kTS, zscal_kTS, zswap_kTS,
  302. zgemv_nTS, zgemv_tTS, zgemv_rTS, zgemv_cTS,
  303. zgemv_oTS, zgemv_uTS, zgemv_sTS, zgemv_dTS,
  304. zgeru_kTS, zgerc_kTS, zgerv_kTS, zgerd_kTS,
  305. zsymv_LTS, zsymv_UTS,
  306. zhemv_LTS, zhemv_UTS, zhemv_MTS, zhemv_VTS,
  307. zgemm_kernel_nTS, zgemm_kernel_lTS, zgemm_kernel_rTS, zgemm_kernel_bTS,
  308. zgemm_betaTS,
  309. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  310. zgemm_incopyTS, zgemm_itcopyTS,
  311. #else
  312. zgemm_oncopyTS, zgemm_otcopyTS,
  313. #endif
  314. zgemm_oncopyTS, zgemm_otcopyTS,
  315. ztrsm_kernel_LNTS, ztrsm_kernel_LTTS, ztrsm_kernel_LRTS, ztrsm_kernel_LCTS,
  316. ztrsm_kernel_RNTS, ztrsm_kernel_RTTS, ztrsm_kernel_RRTS, ztrsm_kernel_RCTS,
  317. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  318. ztrsm_iunucopyTS, ztrsm_iunncopyTS, ztrsm_iutucopyTS, ztrsm_iutncopyTS,
  319. ztrsm_ilnucopyTS, ztrsm_ilnncopyTS, ztrsm_iltucopyTS, ztrsm_iltncopyTS,
  320. #else
  321. ztrsm_ounucopyTS, ztrsm_ounncopyTS, ztrsm_outucopyTS, ztrsm_outncopyTS,
  322. ztrsm_olnucopyTS, ztrsm_olnncopyTS, ztrsm_oltucopyTS, ztrsm_oltncopyTS,
  323. #endif
  324. ztrsm_ounucopyTS, ztrsm_ounncopyTS, ztrsm_outucopyTS, ztrsm_outncopyTS,
  325. ztrsm_olnucopyTS, ztrsm_olnncopyTS, ztrsm_oltucopyTS, ztrsm_oltncopyTS,
  326. ztrmm_kernel_RNTS, ztrmm_kernel_RTTS, ztrmm_kernel_RRTS, ztrmm_kernel_RCTS,
  327. ztrmm_kernel_LNTS, ztrmm_kernel_LTTS, ztrmm_kernel_LRTS, ztrmm_kernel_LCTS,
  328. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  329. ztrmm_iunucopyTS, ztrmm_iunncopyTS, ztrmm_iutucopyTS, ztrmm_iutncopyTS,
  330. ztrmm_ilnucopyTS, ztrmm_ilnncopyTS, ztrmm_iltucopyTS, ztrmm_iltncopyTS,
  331. #else
  332. ztrmm_ounucopyTS, ztrmm_ounncopyTS, ztrmm_outucopyTS, ztrmm_outncopyTS,
  333. ztrmm_olnucopyTS, ztrmm_olnncopyTS, ztrmm_oltucopyTS, ztrmm_oltncopyTS,
  334. #endif
  335. ztrmm_ounucopyTS, ztrmm_ounncopyTS, ztrmm_outucopyTS, ztrmm_outncopyTS,
  336. ztrmm_olnucopyTS, ztrmm_olnncopyTS, ztrmm_oltucopyTS, ztrmm_oltncopyTS,
  337. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  338. zsymm_iutcopyTS, zsymm_iltcopyTS,
  339. #else
  340. zsymm_outcopyTS, zsymm_oltcopyTS,
  341. #endif
  342. zsymm_outcopyTS, zsymm_oltcopyTS,
  343. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  344. zhemm_iutcopyTS, zhemm_iltcopyTS,
  345. #else
  346. zhemm_outcopyTS, zhemm_oltcopyTS,
  347. #endif
  348. zhemm_outcopyTS, zhemm_oltcopyTS,
  349. 0, 0, 0,
  350. #ifdef ZGEMM3M_DEFAULT_UNROLL_M
  351. ZGEMM3M_DEFAULT_UNROLL_M, ZGEMM3M_DEFAULT_UNROLL_N, MAX(ZGEMM3M_DEFAULT_UNROLL_M, ZGEMM3M_DEFAULT_UNROLL_N),
  352. #else
  353. DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N, MAX(DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N),
  354. #endif
  355. zgemm3m_kernelTS,
  356. zgemm3m_incopybTS, zgemm3m_incopyrTS,
  357. zgemm3m_incopyiTS, zgemm3m_itcopybTS,
  358. zgemm3m_itcopyrTS, zgemm3m_itcopyiTS,
  359. zgemm3m_oncopybTS, zgemm3m_oncopyrTS,
  360. zgemm3m_oncopyiTS, zgemm3m_otcopybTS,
  361. zgemm3m_otcopyrTS, zgemm3m_otcopyiTS,
  362. zsymm3m_iucopybTS, zsymm3m_ilcopybTS,
  363. zsymm3m_iucopyrTS, zsymm3m_ilcopyrTS,
  364. zsymm3m_iucopyiTS, zsymm3m_ilcopyiTS,
  365. zsymm3m_oucopybTS, zsymm3m_olcopybTS,
  366. zsymm3m_oucopyrTS, zsymm3m_olcopyrTS,
  367. zsymm3m_oucopyiTS, zsymm3m_olcopyiTS,
  368. zhemm3m_iucopybTS, zhemm3m_ilcopybTS,
  369. zhemm3m_iucopyrTS, zhemm3m_ilcopyrTS,
  370. zhemm3m_iucopyiTS, zhemm3m_ilcopyiTS,
  371. zhemm3m_oucopybTS, zhemm3m_olcopybTS,
  372. zhemm3m_oucopyrTS, zhemm3m_olcopyrTS,
  373. zhemm3m_oucopyiTS, zhemm3m_olcopyiTS,
  374. #ifndef NO_LAPACK
  375. zneg_tcopyTS, zlaswp_ncopyTS,
  376. #else
  377. NULL, NULL,
  378. #endif
  379. #ifdef EXPRECISION
  380. 0, 0, 0,
  381. XGEMM_DEFAULT_UNROLL_M, XGEMM_DEFAULT_UNROLL_N, MAX(XGEMM_DEFAULT_UNROLL_M, XGEMM_DEFAULT_UNROLL_N),
  382. xamax_kTS, xamin_kTS, ixamax_kTS, ixamin_kTS,
  383. xnrm2_kTS, xasum_kTS, xcopy_kTS,
  384. xdotu_kTS, xdotc_kTS, xqrot_kTS,
  385. xaxpy_kTS, xaxpyc_kTS, xscal_kTS, xswap_kTS,
  386. xgemv_nTS, xgemv_tTS, xgemv_rTS, xgemv_cTS,
  387. xgemv_oTS, xgemv_uTS, xgemv_sTS, xgemv_dTS,
  388. xgeru_kTS, xgerc_kTS, xgerv_kTS, xgerd_kTS,
  389. xsymv_LTS, xsymv_UTS,
  390. xhemv_LTS, xhemv_UTS, xhemv_MTS, xhemv_VTS,
  391. xgemm_kernel_nTS, xgemm_kernel_lTS, xgemm_kernel_rTS, xgemm_kernel_bTS,
  392. xgemm_betaTS,
  393. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  394. xgemm_incopyTS, xgemm_itcopyTS,
  395. #else
  396. xgemm_oncopyTS, xgemm_otcopyTS,
  397. #endif
  398. xgemm_oncopyTS, xgemm_otcopyTS,
  399. xtrsm_kernel_LNTS, xtrsm_kernel_LTTS, xtrsm_kernel_LRTS, xtrsm_kernel_LCTS,
  400. xtrsm_kernel_RNTS, xtrsm_kernel_RTTS, xtrsm_kernel_RRTS, xtrsm_kernel_RCTS,
  401. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  402. xtrsm_iunucopyTS, xtrsm_iunncopyTS, xtrsm_iutucopyTS, xtrsm_iutncopyTS,
  403. xtrsm_ilnucopyTS, xtrsm_ilnncopyTS, xtrsm_iltucopyTS, xtrsm_iltncopyTS,
  404. #else
  405. xtrsm_ounucopyTS, xtrsm_ounncopyTS, xtrsm_outucopyTS, xtrsm_outncopyTS,
  406. xtrsm_olnucopyTS, xtrsm_olnncopyTS, xtrsm_oltucopyTS, xtrsm_oltncopyTS,
  407. #endif
  408. xtrsm_ounucopyTS, xtrsm_ounncopyTS, xtrsm_outucopyTS, xtrsm_outncopyTS,
  409. xtrsm_olnucopyTS, xtrsm_olnncopyTS, xtrsm_oltucopyTS, xtrsm_oltncopyTS,
  410. xtrmm_kernel_RNTS, xtrmm_kernel_RTTS, xtrmm_kernel_RRTS, xtrmm_kernel_RCTS,
  411. xtrmm_kernel_LNTS, xtrmm_kernel_LTTS, xtrmm_kernel_LRTS, xtrmm_kernel_LCTS,
  412. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  413. xtrmm_iunucopyTS, xtrmm_iunncopyTS, xtrmm_iutucopyTS, xtrmm_iutncopyTS,
  414. xtrmm_ilnucopyTS, xtrmm_ilnncopyTS, xtrmm_iltucopyTS, xtrmm_iltncopyTS,
  415. #else
  416. xtrmm_ounucopyTS, xtrmm_ounncopyTS, xtrmm_outucopyTS, xtrmm_outncopyTS,
  417. xtrmm_olnucopyTS, xtrmm_olnncopyTS, xtrmm_oltucopyTS, xtrmm_oltncopyTS,
  418. #endif
  419. xtrmm_ounucopyTS, xtrmm_ounncopyTS, xtrmm_outucopyTS, xtrmm_outncopyTS,
  420. xtrmm_olnucopyTS, xtrmm_olnncopyTS, xtrmm_oltucopyTS, xtrmm_oltncopyTS,
  421. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  422. xsymm_iutcopyTS, xsymm_iltcopyTS,
  423. #else
  424. xsymm_outcopyTS, xsymm_oltcopyTS,
  425. #endif
  426. xsymm_outcopyTS, xsymm_oltcopyTS,
  427. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  428. xhemm_iutcopyTS, xhemm_iltcopyTS,
  429. #else
  430. xhemm_outcopyTS, xhemm_oltcopyTS,
  431. #endif
  432. xhemm_outcopyTS, xhemm_oltcopyTS,
  433. 0, 0, 0,
  434. QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N, MAX(QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N),
  435. xgemm3m_kernelTS,
  436. xgemm3m_incopybTS, xgemm3m_incopyrTS,
  437. xgemm3m_incopyiTS, xgemm3m_itcopybTS,
  438. xgemm3m_itcopyrTS, xgemm3m_itcopyiTS,
  439. xgemm3m_oncopybTS, xgemm3m_oncopyrTS,
  440. xgemm3m_oncopyiTS, xgemm3m_otcopybTS,
  441. xgemm3m_otcopyrTS, xgemm3m_otcopyiTS,
  442. xsymm3m_iucopybTS, xsymm3m_ilcopybTS,
  443. xsymm3m_iucopyrTS, xsymm3m_ilcopyrTS,
  444. xsymm3m_iucopyiTS, xsymm3m_ilcopyiTS,
  445. xsymm3m_oucopybTS, xsymm3m_olcopybTS,
  446. xsymm3m_oucopyrTS, xsymm3m_olcopyrTS,
  447. xsymm3m_oucopyiTS, xsymm3m_olcopyiTS,
  448. xhemm3m_iucopybTS, xhemm3m_ilcopybTS,
  449. xhemm3m_iucopyrTS, xhemm3m_ilcopyrTS,
  450. xhemm3m_iucopyiTS, xhemm3m_ilcopyiTS,
  451. xhemm3m_oucopybTS, xhemm3m_olcopybTS,
  452. xhemm3m_oucopyrTS, xhemm3m_olcopyrTS,
  453. xhemm3m_oucopyiTS, xhemm3m_olcopyiTS,
  454. #ifndef NO_LAPACK
  455. xneg_tcopyTS, xlaswp_ncopyTS,
  456. #else
  457. NULL, NULL,
  458. #endif
  459. #endif
  460. init_parameter,
  461. SNUMOPT, DNUMOPT, QNUMOPT,
  462. saxpby_kTS, daxpby_kTS, caxpby_kTS, zaxpby_kTS,
  463. somatcopy_k_cnTS, somatcopy_k_ctTS, somatcopy_k_rnTS, somatcopy_k_rtTS,
  464. domatcopy_k_cnTS, domatcopy_k_ctTS, domatcopy_k_rnTS, domatcopy_k_rtTS,
  465. comatcopy_k_cnTS, comatcopy_k_ctTS, comatcopy_k_rnTS, comatcopy_k_rtTS,
  466. comatcopy_k_cncTS, comatcopy_k_ctcTS, comatcopy_k_rncTS, comatcopy_k_rtcTS,
  467. zomatcopy_k_cnTS, zomatcopy_k_ctTS, zomatcopy_k_rnTS, zomatcopy_k_rtTS,
  468. zomatcopy_k_cncTS, zomatcopy_k_ctcTS, zomatcopy_k_rncTS, zomatcopy_k_rtcTS,
  469. simatcopy_k_cnTS, simatcopy_k_ctTS, simatcopy_k_rnTS, simatcopy_k_rtTS,
  470. dimatcopy_k_cnTS, dimatcopy_k_ctTS, dimatcopy_k_rnTS, dimatcopy_k_rtTS,
  471. cimatcopy_k_cnTS, cimatcopy_k_ctTS, cimatcopy_k_rnTS, cimatcopy_k_rtTS,
  472. cimatcopy_k_cncTS, cimatcopy_k_ctcTS, cimatcopy_k_rncTS, cimatcopy_k_rtcTS,
  473. zimatcopy_k_cnTS, zimatcopy_k_ctTS, zimatcopy_k_rnTS, zimatcopy_k_rtTS,
  474. zimatcopy_k_cncTS, zimatcopy_k_ctcTS, zimatcopy_k_rncTS, zimatcopy_k_rtcTS,
  475. sgeadd_kTS, dgeadd_kTS, cgeadd_kTS, zgeadd_kTS
  476. };
  477. #ifdef ARCH_X86
  478. static int get_l2_size_old(void){
  479. int i, eax, ebx, ecx, edx, cpuid_level;
  480. int info[15];
  481. cpuid(2, &eax, &ebx, &ecx, &edx);
  482. info[ 0] = BITMASK(eax, 8, 0xff);
  483. info[ 1] = BITMASK(eax, 16, 0xff);
  484. info[ 2] = BITMASK(eax, 24, 0xff);
  485. info[ 3] = BITMASK(ebx, 0, 0xff);
  486. info[ 4] = BITMASK(ebx, 8, 0xff);
  487. info[ 5] = BITMASK(ebx, 16, 0xff);
  488. info[ 6] = BITMASK(ebx, 24, 0xff);
  489. info[ 7] = BITMASK(ecx, 0, 0xff);
  490. info[ 8] = BITMASK(ecx, 8, 0xff);
  491. info[ 9] = BITMASK(ecx, 16, 0xff);
  492. info[10] = BITMASK(ecx, 24, 0xff);
  493. info[11] = BITMASK(edx, 0, 0xff);
  494. info[12] = BITMASK(edx, 8, 0xff);
  495. info[13] = BITMASK(edx, 16, 0xff);
  496. info[14] = BITMASK(edx, 24, 0xff);
  497. for (i = 0; i < 15; i++){
  498. switch (info[i]){
  499. /* This table is from http://www.sandpile.org/ia32/cpuid.htm */
  500. case 0x1a :
  501. return 96;
  502. case 0x39 :
  503. case 0x3b :
  504. case 0x41 :
  505. case 0x79 :
  506. case 0x81 :
  507. return 128;
  508. case 0x3a :
  509. return 192;
  510. case 0x21 :
  511. case 0x3c :
  512. case 0x42 :
  513. case 0x7a :
  514. case 0x7e :
  515. case 0x82 :
  516. return 256;
  517. case 0x3d :
  518. return 384;
  519. case 0x3e :
  520. case 0x43 :
  521. case 0x7b :
  522. case 0x7f :
  523. case 0x83 :
  524. case 0x86 :
  525. return 512;
  526. case 0x44 :
  527. case 0x78 :
  528. case 0x7c :
  529. case 0x84 :
  530. case 0x87 :
  531. return 1024;
  532. case 0x45 :
  533. case 0x7d :
  534. case 0x85 :
  535. return 2048;
  536. case 0x48 :
  537. return 3184;
  538. case 0x49 :
  539. return 4096;
  540. case 0x4e :
  541. return 6144;
  542. }
  543. }
  544. // return 0;
  545. fprintf (stderr,"OpenBLAS WARNING - could not determine the L2 cache size on this system, assuming 256k\n");
  546. return 256;
  547. }
  548. #endif
  549. static __inline__ int get_l2_size(void){
  550. int eax, ebx, ecx, edx, l2;
  551. cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
  552. l2 = BITMASK(ecx, 16, 0xffff);
  553. #ifndef ARCH_X86
  554. if (l2 <= 0) {
  555. fprintf (stderr,"OpenBLAS WARNING - could not determine the L2 cache size on this system, assuming 256k\n");
  556. return 256;
  557. }
  558. return l2;
  559. #else
  560. if (l2 > 0) return l2;
  561. return get_l2_size_old();
  562. #endif
  563. }
  564. static __inline__ int get_l3_size(void){
  565. int eax, ebx, ecx, edx;
  566. cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
  567. return BITMASK(edx, 18, 0x3fff) * 512;
  568. }
  569. static void init_parameter(void) {
  570. int l2 = get_l2_size();
  571. (void) l2; /* dirty trick to suppress unused variable warning for targets */
  572. /* where the GEMM unrolling parameters do not depend on l2 */
  573. TABLE_NAME.sgemm_q = SGEMM_DEFAULT_Q;
  574. TABLE_NAME.dgemm_q = DGEMM_DEFAULT_Q;
  575. TABLE_NAME.cgemm_q = CGEMM_DEFAULT_Q;
  576. TABLE_NAME.zgemm_q = ZGEMM_DEFAULT_Q;
  577. #ifdef CGEMM3M_DEFAULT_Q
  578. TABLE_NAME.cgemm3m_q = CGEMM3M_DEFAULT_Q;
  579. #else
  580. TABLE_NAME.cgemm3m_q = SGEMM_DEFAULT_Q;
  581. #endif
  582. #ifdef ZGEMM3M_DEFAULT_Q
  583. TABLE_NAME.zgemm3m_q = ZGEMM3M_DEFAULT_Q;
  584. #else
  585. TABLE_NAME.zgemm3m_q = DGEMM_DEFAULT_Q;
  586. #endif
  587. #ifdef EXPRECISION
  588. TABLE_NAME.qgemm_q = QGEMM_DEFAULT_Q;
  589. TABLE_NAME.xgemm_q = XGEMM_DEFAULT_Q;
  590. TABLE_NAME.xgemm3m_q = QGEMM_DEFAULT_Q;
  591. #endif
  592. #if defined(CORE_KATMAI) || defined(CORE_COPPERMINE) || defined(CORE_BANIAS) || defined(CORE_YONAH) || defined(CORE_ATHLON)
  593. #ifdef DEBUG
  594. fprintf(stderr, "Katmai, Coppermine, Banias, Athlon\n");
  595. #endif
  596. TABLE_NAME.sgemm_p = 64 * (l2 >> 7);
  597. TABLE_NAME.dgemm_p = 32 * (l2 >> 7);
  598. TABLE_NAME.cgemm_p = 32 * (l2 >> 7);
  599. TABLE_NAME.zgemm_p = 16 * (l2 >> 7);
  600. #ifdef EXPRECISION
  601. TABLE_NAME.qgemm_p = 16 * (l2 >> 7);
  602. TABLE_NAME.xgemm_p = 8 * (l2 >> 7);
  603. #endif
  604. #endif
  605. #ifdef CORE_NORTHWOOD
  606. #ifdef DEBUG
  607. fprintf(stderr, "Northwood\n");
  608. #endif
  609. TABLE_NAME.sgemm_p = 96 * (l2 >> 7);
  610. TABLE_NAME.dgemm_p = 48 * (l2 >> 7);
  611. TABLE_NAME.cgemm_p = 48 * (l2 >> 7);
  612. TABLE_NAME.zgemm_p = 24 * (l2 >> 7);
  613. #ifdef EXPRECISION
  614. TABLE_NAME.qgemm_p = 24 * (l2 >> 7);
  615. TABLE_NAME.xgemm_p = 12 * (l2 >> 7);
  616. #endif
  617. #endif
  618. #ifdef ATOM
  619. #ifdef DEBUG
  620. fprintf(stderr, "Atom\n");
  621. #endif
  622. TABLE_NAME.sgemm_p = 256;
  623. TABLE_NAME.dgemm_p = 128;
  624. TABLE_NAME.cgemm_p = 128;
  625. TABLE_NAME.zgemm_p = 64;
  626. #ifdef EXPRECISION
  627. TABLE_NAME.qgemm_p = 64;
  628. TABLE_NAME.xgemm_p = 32;
  629. #endif
  630. #endif
  631. #ifdef CORE_PRESCOTT
  632. #ifdef DEBUG
  633. fprintf(stderr, "Prescott\n");
  634. #endif
  635. TABLE_NAME.sgemm_p = 56 * (l2 >> 7);
  636. TABLE_NAME.dgemm_p = 28 * (l2 >> 7);
  637. TABLE_NAME.cgemm_p = 28 * (l2 >> 7);
  638. TABLE_NAME.zgemm_p = 14 * (l2 >> 7);
  639. #ifdef EXPRECISION
  640. TABLE_NAME.qgemm_p = 14 * (l2 >> 7);
  641. TABLE_NAME.xgemm_p = 7 * (l2 >> 7);
  642. #endif
  643. #endif
  644. #ifdef CORE2
  645. #ifdef DEBUG
  646. fprintf(stderr, "Core2\n");
  647. #endif
  648. TABLE_NAME.sgemm_p = 92 * (l2 >> 9) + 8;
  649. TABLE_NAME.dgemm_p = 46 * (l2 >> 9) + 8;
  650. TABLE_NAME.cgemm_p = 46 * (l2 >> 9) + 4;
  651. TABLE_NAME.zgemm_p = 23 * (l2 >> 9) + 4;
  652. #ifdef EXPRECISION
  653. TABLE_NAME.qgemm_p = 92 * (l2 >> 9) + 8;
  654. TABLE_NAME.xgemm_p = 46 * (l2 >> 9) + 4;
  655. #endif
  656. #endif
  657. #ifdef PENRYN
  658. #ifdef DEBUG
  659. fprintf(stderr, "Penryn\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 DUNNINGTON
  671. #ifdef DEBUG
  672. fprintf(stderr, "Dunnington\n");
  673. #endif
  674. TABLE_NAME.sgemm_p = 42 * (l2 >> 9) + 8;
  675. TABLE_NAME.dgemm_p = 42 * (l2 >> 9) + 8;
  676. TABLE_NAME.cgemm_p = 21 * (l2 >> 9) + 4;
  677. TABLE_NAME.zgemm_p = 21 * (l2 >> 9) + 4;
  678. #ifdef EXPRECISION
  679. TABLE_NAME.qgemm_p = 42 * (l2 >> 9) + 8;
  680. TABLE_NAME.xgemm_p = 21 * (l2 >> 9) + 4;
  681. #endif
  682. #endif
  683. #ifdef NEHALEM
  684. #ifdef DEBUG
  685. fprintf(stderr, "Nehalem\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 SANDYBRIDGE
  697. #ifdef DEBUG
  698. fprintf(stderr, "Sandybridge\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 HASWELL
  710. #ifdef DEBUG
  711. fprintf(stderr, "Haswell\n");
  712. #endif
  713. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  714. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  715. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  716. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  717. #ifdef EXPRECISION
  718. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  719. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  720. #endif
  721. #endif
  722. #ifdef SKYLAKEX
  723. #ifdef DEBUG
  724. fprintf(stderr, "SkylakeX\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 OPTERON
  736. #ifdef DEBUG
  737. fprintf(stderr, "Opteron\n");
  738. #endif
  739. TABLE_NAME.sgemm_p = 224 + 56 * (l2 >> 7);
  740. TABLE_NAME.dgemm_p = 112 + 28 * (l2 >> 7);
  741. TABLE_NAME.cgemm_p = 112 + 28 * (l2 >> 7);
  742. TABLE_NAME.zgemm_p = 56 + 14 * (l2 >> 7);
  743. #ifdef EXPRECISION
  744. TABLE_NAME.qgemm_p = 56 + 14 * (l2 >> 7);
  745. TABLE_NAME.xgemm_p = 28 + 7 * (l2 >> 7);
  746. #endif
  747. #endif
  748. #ifdef BARCELONA
  749. #ifdef DEBUG
  750. fprintf(stderr, "Barcelona\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 BOBCAT
  762. #ifdef DEBUG
  763. fprintf(stderr, "Bobcate\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 BULLDOZER
  775. #ifdef DEBUG
  776. fprintf(stderr, "Bulldozer\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 EXCAVATOR
  788. #ifdef DEBUG
  789. fprintf(stderr, "Excavator\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 PILEDRIVER
  801. #ifdef DEBUG
  802. fprintf(stderr, "Piledriver\n");
  803. #endif
  804. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  805. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  806. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  807. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  808. #ifdef EXPRECISION
  809. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  810. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  811. #endif
  812. #endif
  813. #ifdef STEAMROLLER
  814. #ifdef DEBUG
  815. fprintf(stderr, "Steamroller\n");
  816. #endif
  817. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  818. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  819. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  820. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  821. #ifdef EXPRECISION
  822. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  823. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  824. #endif
  825. #endif
  826. #ifdef ZEN
  827. #ifdef DEBUG
  828. fprintf(stderr, "Zen\n");
  829. #endif
  830. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  831. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  832. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  833. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  834. #ifdef EXPRECISION
  835. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  836. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  837. #endif
  838. #endif
  839. #ifdef NANO
  840. #ifdef DEBUG
  841. fprintf(stderr, "NANO\n");
  842. #endif
  843. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  844. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  845. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  846. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  847. #ifdef EXPRECISION
  848. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  849. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  850. #endif
  851. #endif
  852. #ifdef CGEMM3M_DEFAULT_P
  853. TABLE_NAME.cgemm3m_p = CGEMM3M_DEFAULT_P;
  854. #else
  855. TABLE_NAME.cgemm3m_p = TABLE_NAME.sgemm_p;
  856. #endif
  857. #ifdef ZGEMM3M_DEFAULT_P
  858. TABLE_NAME.zgemm3m_p = ZGEMM3M_DEFAULT_P;
  859. #else
  860. TABLE_NAME.zgemm3m_p = TABLE_NAME.dgemm_p;
  861. #endif
  862. #ifdef EXPRECISION
  863. TABLE_NAME.xgemm3m_p = TABLE_NAME.qgemm_p;
  864. #endif
  865. TABLE_NAME.sgemm_p = ((TABLE_NAME.sgemm_p + SGEMM_DEFAULT_UNROLL_M - 1)/SGEMM_DEFAULT_UNROLL_M) * SGEMM_DEFAULT_UNROLL_M;
  866. TABLE_NAME.dgemm_p = ((TABLE_NAME.dgemm_p + DGEMM_DEFAULT_UNROLL_M - 1)/DGEMM_DEFAULT_UNROLL_M) * DGEMM_DEFAULT_UNROLL_M;
  867. TABLE_NAME.cgemm_p = ((TABLE_NAME.cgemm_p + CGEMM_DEFAULT_UNROLL_M - 1)/CGEMM_DEFAULT_UNROLL_M) * CGEMM_DEFAULT_UNROLL_M;
  868. TABLE_NAME.zgemm_p = ((TABLE_NAME.zgemm_p + ZGEMM_DEFAULT_UNROLL_M - 1)/ZGEMM_DEFAULT_UNROLL_M) * ZGEMM_DEFAULT_UNROLL_M;
  869. #ifdef CGEMM3M_DEFAULT_UNROLL_M
  870. TABLE_NAME.cgemm3m_p = ((TABLE_NAME.cgemm3m_p + CGEMM3M_DEFAULT_UNROLL_M - 1)/CGEMM3M_DEFAULT_UNROLL_M) * CGEMM3M_DEFAULT_UNROLL_M;
  871. #else
  872. TABLE_NAME.cgemm3m_p = ((TABLE_NAME.cgemm3m_p + SGEMM_DEFAULT_UNROLL_M - 1)/SGEMM_DEFAULT_UNROLL_M) * SGEMM_DEFAULT_UNROLL_M;
  873. #endif
  874. #ifdef ZGEMM3M_DEFAULT_UNROLL_M
  875. TABLE_NAME.zgemm3m_p = ((TABLE_NAME.zgemm3m_p + ZGEMM3M_DEFAULT_UNROLL_M - 1)/ZGEMM3M_DEFAULT_UNROLL_M) * ZGEMM3M_DEFAULT_UNROLL_M;
  876. #else
  877. TABLE_NAME.zgemm3m_p = ((TABLE_NAME.zgemm3m_p + DGEMM_DEFAULT_UNROLL_M - 1)/DGEMM_DEFAULT_UNROLL_M) * DGEMM_DEFAULT_UNROLL_M;
  878. #endif
  879. #ifdef QUAD_PRECISION
  880. TABLE_NAME.qgemm_p = ((TABLE_NAME.qgemm_p + QGEMM_DEFAULT_UNROLL_M - 1)/QGEMM_DEFAULT_UNROLL_M) * QGEMM_DEFAULT_UNROLL_M;
  881. TABLE_NAME.xgemm_p = ((TABLE_NAME.xgemm_p + XGEMM_DEFAULT_UNROLL_M - 1)/XGEMM_DEFAULT_UNROLL_M) * XGEMM_DEFAULT_UNROLL_M;
  882. TABLE_NAME.xgemm3m_p = ((TABLE_NAME.xgemm3m_p + QGEMM_DEFAULT_UNROLL_M - 1)/QGEMM_DEFAULT_UNROLL_M) * QGEMM_DEFAULT_UNROLL_M;
  883. #endif
  884. #ifdef DEBUG
  885. fprintf(stderr, "L2 = %8d DGEMM_P .. %d\n", l2, TABLE_NAME.dgemm_p);
  886. #endif
  887. TABLE_NAME.sgemm_r = (((BUFFER_SIZE -
  888. ((TABLE_NAME.sgemm_p * TABLE_NAME.sgemm_q * 4 + TABLE_NAME.offsetA
  889. + TABLE_NAME.align) & ~TABLE_NAME.align)
  890. ) / (TABLE_NAME.sgemm_q * 4) - 15) & ~15);
  891. TABLE_NAME.dgemm_r = (((BUFFER_SIZE -
  892. ((TABLE_NAME.dgemm_p * TABLE_NAME.dgemm_q * 8 + TABLE_NAME.offsetA
  893. + TABLE_NAME.align) & ~TABLE_NAME.align)
  894. ) / (TABLE_NAME.dgemm_q * 8) - 15) & ~15);
  895. #ifdef EXPRECISION
  896. TABLE_NAME.qgemm_r = (((BUFFER_SIZE -
  897. ((TABLE_NAME.qgemm_p * TABLE_NAME.qgemm_q * 16 + TABLE_NAME.offsetA
  898. + TABLE_NAME.align) & ~TABLE_NAME.align)
  899. ) / (TABLE_NAME.qgemm_q * 16) - 15) & ~15);
  900. #endif
  901. TABLE_NAME.cgemm_r = (((BUFFER_SIZE -
  902. ((TABLE_NAME.cgemm_p * TABLE_NAME.cgemm_q * 8 + TABLE_NAME.offsetA
  903. + TABLE_NAME.align) & ~TABLE_NAME.align)
  904. ) / (TABLE_NAME.cgemm_q * 8) - 15) & ~15);
  905. TABLE_NAME.zgemm_r = (((BUFFER_SIZE -
  906. ((TABLE_NAME.zgemm_p * TABLE_NAME.zgemm_q * 16 + TABLE_NAME.offsetA
  907. + TABLE_NAME.align) & ~TABLE_NAME.align)
  908. ) / (TABLE_NAME.zgemm_q * 16) - 15) & ~15);
  909. TABLE_NAME.cgemm3m_r = (((BUFFER_SIZE -
  910. ((TABLE_NAME.cgemm3m_p * TABLE_NAME.cgemm3m_q * 8 + TABLE_NAME.offsetA
  911. + TABLE_NAME.align) & ~TABLE_NAME.align)
  912. ) / (TABLE_NAME.cgemm3m_q * 8) - 15) & ~15);
  913. TABLE_NAME.zgemm3m_r = (((BUFFER_SIZE -
  914. ((TABLE_NAME.zgemm3m_p * TABLE_NAME.zgemm3m_q * 16 + TABLE_NAME.offsetA
  915. + TABLE_NAME.align) & ~TABLE_NAME.align)
  916. ) / (TABLE_NAME.zgemm3m_q * 16) - 15) & ~15);
  917. #ifdef EXPRECISION
  918. TABLE_NAME.xgemm_r = (((BUFFER_SIZE -
  919. ((TABLE_NAME.xgemm_p * TABLE_NAME.xgemm_q * 32 + TABLE_NAME.offsetA
  920. + TABLE_NAME.align) & ~TABLE_NAME.align)
  921. ) / (TABLE_NAME.xgemm_q * 32) - 15) & ~15);
  922. TABLE_NAME.xgemm3m_r = (((BUFFER_SIZE -
  923. ((TABLE_NAME.xgemm3m_p * TABLE_NAME.xgemm3m_q * 32 + TABLE_NAME.offsetA
  924. + TABLE_NAME.align) & ~TABLE_NAME.align)
  925. ) / (TABLE_NAME.xgemm3m_q * 32) - 15) & ~15);
  926. #endif
  927. }