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 28 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893
  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, MAX(SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N),
  51. #ifdef HAVE_EXCLUSIVE_CACHE
  52. 1,
  53. #else
  54. 0,
  55. #endif
  56. samax_kTS, samin_kTS, smax_kTS, smin_kTS,
  57. isamax_kTS, isamin_kTS, ismax_kTS, ismin_kTS,
  58. snrm2_kTS, sasum_kTS, scopy_kTS, sdot_kTS,
  59. dsdot_kTS,
  60. srot_kTS, saxpy_kTS, sscal_kTS, sswap_kTS,
  61. sgemv_nTS, sgemv_tTS, sger_kTS,
  62. ssymv_LTS, ssymv_UTS,
  63. sgemm_kernelTS, sgemm_betaTS,
  64. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  65. sgemm_incopyTS, sgemm_itcopyTS,
  66. #else
  67. sgemm_oncopyTS, sgemm_otcopyTS,
  68. #endif
  69. sgemm_oncopyTS, sgemm_otcopyTS,
  70. strsm_kernel_LNTS, strsm_kernel_LTTS, strsm_kernel_RNTS, strsm_kernel_RTTS,
  71. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  72. strsm_iunucopyTS, strsm_iunncopyTS, strsm_iutucopyTS, strsm_iutncopyTS,
  73. strsm_ilnucopyTS, strsm_ilnncopyTS, strsm_iltucopyTS, strsm_iltncopyTS,
  74. #else
  75. strsm_ounucopyTS, strsm_ounncopyTS, strsm_outucopyTS, strsm_outncopyTS,
  76. strsm_olnucopyTS, strsm_olnncopyTS, strsm_oltucopyTS, strsm_oltncopyTS,
  77. #endif
  78. strsm_ounucopyTS, strsm_ounncopyTS, strsm_outucopyTS, strsm_outncopyTS,
  79. strsm_olnucopyTS, strsm_olnncopyTS, strsm_oltucopyTS, strsm_oltncopyTS,
  80. strmm_kernel_RNTS, strmm_kernel_RTTS, strmm_kernel_LNTS, strmm_kernel_LTTS,
  81. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  82. strmm_iunucopyTS, strmm_iunncopyTS, strmm_iutucopyTS, strmm_iutncopyTS,
  83. strmm_ilnucopyTS, strmm_ilnncopyTS, strmm_iltucopyTS, strmm_iltncopyTS,
  84. #else
  85. strmm_ounucopyTS, strmm_ounncopyTS, strmm_outucopyTS, strmm_outncopyTS,
  86. strmm_olnucopyTS, strmm_olnncopyTS, strmm_oltucopyTS, strmm_oltncopyTS,
  87. #endif
  88. strmm_ounucopyTS, strmm_ounncopyTS, strmm_outucopyTS, strmm_outncopyTS,
  89. strmm_olnucopyTS, strmm_olnncopyTS, strmm_oltucopyTS, strmm_oltncopyTS,
  90. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  91. ssymm_iutcopyTS, ssymm_iltcopyTS,
  92. #else
  93. ssymm_outcopyTS, ssymm_oltcopyTS,
  94. #endif
  95. ssymm_outcopyTS, ssymm_oltcopyTS,
  96. #ifndef NO_LAPACK
  97. sneg_tcopyTS, slaswp_ncopyTS,
  98. #else
  99. NULL,NULL,
  100. #endif
  101. 0, 0, 0,
  102. DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N, MAX(DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N),
  103. damax_kTS, damin_kTS, dmax_kTS, dmin_kTS,
  104. idamax_kTS, idamin_kTS, idmax_kTS, idmin_kTS,
  105. dnrm2_kTS, dasum_kTS, dcopy_kTS, ddot_kTS,
  106. drot_kTS, daxpy_kTS, dscal_kTS, dswap_kTS,
  107. dgemv_nTS, dgemv_tTS, dger_kTS,
  108. dsymv_LTS, dsymv_UTS,
  109. dgemm_kernelTS, dgemm_betaTS,
  110. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  111. dgemm_incopyTS, dgemm_itcopyTS,
  112. #else
  113. dgemm_oncopyTS, dgemm_otcopyTS,
  114. #endif
  115. dgemm_oncopyTS, dgemm_otcopyTS,
  116. dtrsm_kernel_LNTS, dtrsm_kernel_LTTS, dtrsm_kernel_RNTS, dtrsm_kernel_RTTS,
  117. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  118. dtrsm_iunucopyTS, dtrsm_iunncopyTS, dtrsm_iutucopyTS, dtrsm_iutncopyTS,
  119. dtrsm_ilnucopyTS, dtrsm_ilnncopyTS, dtrsm_iltucopyTS, dtrsm_iltncopyTS,
  120. #else
  121. dtrsm_ounucopyTS, dtrsm_ounncopyTS, dtrsm_outucopyTS, dtrsm_outncopyTS,
  122. dtrsm_olnucopyTS, dtrsm_olnncopyTS, dtrsm_oltucopyTS, dtrsm_oltncopyTS,
  123. #endif
  124. dtrsm_ounucopyTS, dtrsm_ounncopyTS, dtrsm_outucopyTS, dtrsm_outncopyTS,
  125. dtrsm_olnucopyTS, dtrsm_olnncopyTS, dtrsm_oltucopyTS, dtrsm_oltncopyTS,
  126. dtrmm_kernel_RNTS, dtrmm_kernel_RTTS, dtrmm_kernel_LNTS, dtrmm_kernel_LTTS,
  127. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  128. dtrmm_iunucopyTS, dtrmm_iunncopyTS, dtrmm_iutucopyTS, dtrmm_iutncopyTS,
  129. dtrmm_ilnucopyTS, dtrmm_ilnncopyTS, dtrmm_iltucopyTS, dtrmm_iltncopyTS,
  130. #else
  131. dtrmm_ounucopyTS, dtrmm_ounncopyTS, dtrmm_outucopyTS, dtrmm_outncopyTS,
  132. dtrmm_olnucopyTS, dtrmm_olnncopyTS, dtrmm_oltucopyTS, dtrmm_oltncopyTS,
  133. #endif
  134. dtrmm_ounucopyTS, dtrmm_ounncopyTS, dtrmm_outucopyTS, dtrmm_outncopyTS,
  135. dtrmm_olnucopyTS, dtrmm_olnncopyTS, dtrmm_oltucopyTS, dtrmm_oltncopyTS,
  136. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  137. dsymm_iutcopyTS, dsymm_iltcopyTS,
  138. #else
  139. dsymm_outcopyTS, dsymm_oltcopyTS,
  140. #endif
  141. dsymm_outcopyTS, dsymm_oltcopyTS,
  142. #ifndef NO_LAPACK
  143. dneg_tcopyTS, dlaswp_ncopyTS,
  144. #else
  145. NULL, NULL,
  146. #endif
  147. #ifdef EXPRECISION
  148. 0, 0, 0,
  149. QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N, MAX(QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N),
  150. qamax_kTS, qamin_kTS, qmax_kTS, qmin_kTS,
  151. iqamax_kTS, iqamin_kTS, iqmax_kTS, iqmin_kTS,
  152. qnrm2_kTS, qasum_kTS, qcopy_kTS, qdot_kTS,
  153. qrot_kTS, qaxpy_kTS, qscal_kTS, qswap_kTS,
  154. qgemv_nTS, qgemv_tTS, qger_kTS,
  155. qsymv_LTS, qsymv_UTS,
  156. qgemm_kernelTS, qgemm_betaTS,
  157. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  158. qgemm_incopyTS, qgemm_itcopyTS,
  159. #else
  160. qgemm_oncopyTS, qgemm_otcopyTS,
  161. #endif
  162. qgemm_oncopyTS, qgemm_otcopyTS,
  163. qtrsm_kernel_LNTS, qtrsm_kernel_LTTS, qtrsm_kernel_RNTS, qtrsm_kernel_RTTS,
  164. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  165. qtrsm_iunucopyTS, qtrsm_iunncopyTS, qtrsm_iutucopyTS, qtrsm_iutncopyTS,
  166. qtrsm_ilnucopyTS, qtrsm_ilnncopyTS, qtrsm_iltucopyTS, qtrsm_iltncopyTS,
  167. #else
  168. qtrsm_ounucopyTS, qtrsm_ounncopyTS, qtrsm_outucopyTS, qtrsm_outncopyTS,
  169. qtrsm_olnucopyTS, qtrsm_olnncopyTS, qtrsm_oltucopyTS, qtrsm_oltncopyTS,
  170. #endif
  171. qtrsm_ounucopyTS, qtrsm_ounncopyTS, qtrsm_outucopyTS, qtrsm_outncopyTS,
  172. qtrsm_olnucopyTS, qtrsm_olnncopyTS, qtrsm_oltucopyTS, qtrsm_oltncopyTS,
  173. qtrmm_kernel_RNTS, qtrmm_kernel_RTTS, qtrmm_kernel_LNTS, qtrmm_kernel_LTTS,
  174. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  175. qtrmm_iunucopyTS, qtrmm_iunncopyTS, qtrmm_iutucopyTS, qtrmm_iutncopyTS,
  176. qtrmm_ilnucopyTS, qtrmm_ilnncopyTS, qtrmm_iltucopyTS, qtrmm_iltncopyTS,
  177. #else
  178. qtrmm_ounucopyTS, qtrmm_ounncopyTS, qtrmm_outucopyTS, qtrmm_outncopyTS,
  179. qtrmm_olnucopyTS, qtrmm_olnncopyTS, qtrmm_oltucopyTS, qtrmm_oltncopyTS,
  180. #endif
  181. qtrmm_ounucopyTS, qtrmm_ounncopyTS, qtrmm_outucopyTS, qtrmm_outncopyTS,
  182. qtrmm_olnucopyTS, qtrmm_olnncopyTS, qtrmm_oltucopyTS, qtrmm_oltncopyTS,
  183. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  184. qsymm_iutcopyTS, qsymm_iltcopyTS,
  185. #else
  186. qsymm_outcopyTS, qsymm_oltcopyTS,
  187. #endif
  188. qsymm_outcopyTS, qsymm_oltcopyTS,
  189. #ifndef NO_LAPACK
  190. qneg_tcopyTS, qlaswp_ncopyTS,
  191. #else
  192. NULL, NULL,
  193. #endif
  194. #endif
  195. 0, 0, 0,
  196. CGEMM_DEFAULT_UNROLL_M, CGEMM_DEFAULT_UNROLL_N, MAX(CGEMM_DEFAULT_UNROLL_M, CGEMM_DEFAULT_UNROLL_N),
  197. camax_kTS, camin_kTS, icamax_kTS, icamin_kTS,
  198. cnrm2_kTS, casum_kTS, ccopy_kTS,
  199. cdotu_kTS, cdotc_kTS, csrot_kTS,
  200. caxpy_kTS, caxpyc_kTS, cscal_kTS, cswap_kTS,
  201. cgemv_nTS, cgemv_tTS, cgemv_rTS, cgemv_cTS,
  202. cgemv_oTS, cgemv_uTS, cgemv_sTS, cgemv_dTS,
  203. cgeru_kTS, cgerc_kTS, cgerv_kTS, cgerd_kTS,
  204. csymv_LTS, csymv_UTS,
  205. chemv_LTS, chemv_UTS, chemv_MTS, chemv_VTS,
  206. cgemm_kernel_nTS, cgemm_kernel_lTS, cgemm_kernel_rTS, cgemm_kernel_bTS,
  207. cgemm_betaTS,
  208. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  209. cgemm_incopyTS, cgemm_itcopyTS,
  210. #else
  211. cgemm_oncopyTS, cgemm_otcopyTS,
  212. #endif
  213. cgemm_oncopyTS, cgemm_otcopyTS,
  214. ctrsm_kernel_LNTS, ctrsm_kernel_LTTS, ctrsm_kernel_LRTS, ctrsm_kernel_LCTS,
  215. ctrsm_kernel_RNTS, ctrsm_kernel_RTTS, ctrsm_kernel_RRTS, ctrsm_kernel_RCTS,
  216. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  217. ctrsm_iunucopyTS, ctrsm_iunncopyTS, ctrsm_iutucopyTS, ctrsm_iutncopyTS,
  218. ctrsm_ilnucopyTS, ctrsm_ilnncopyTS, ctrsm_iltucopyTS, ctrsm_iltncopyTS,
  219. #else
  220. ctrsm_ounucopyTS, ctrsm_ounncopyTS, ctrsm_outucopyTS, ctrsm_outncopyTS,
  221. ctrsm_olnucopyTS, ctrsm_olnncopyTS, ctrsm_oltucopyTS, ctrsm_oltncopyTS,
  222. #endif
  223. ctrsm_ounucopyTS, ctrsm_ounncopyTS, ctrsm_outucopyTS, ctrsm_outncopyTS,
  224. ctrsm_olnucopyTS, ctrsm_olnncopyTS, ctrsm_oltucopyTS, ctrsm_oltncopyTS,
  225. ctrmm_kernel_RNTS, ctrmm_kernel_RTTS, ctrmm_kernel_RRTS, ctrmm_kernel_RCTS,
  226. ctrmm_kernel_LNTS, ctrmm_kernel_LTTS, ctrmm_kernel_LRTS, ctrmm_kernel_LCTS,
  227. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  228. ctrmm_iunucopyTS, ctrmm_iunncopyTS, ctrmm_iutucopyTS, ctrmm_iutncopyTS,
  229. ctrmm_ilnucopyTS, ctrmm_ilnncopyTS, ctrmm_iltucopyTS, ctrmm_iltncopyTS,
  230. #else
  231. ctrmm_ounucopyTS, ctrmm_ounncopyTS, ctrmm_outucopyTS, ctrmm_outncopyTS,
  232. ctrmm_olnucopyTS, ctrmm_olnncopyTS, ctrmm_oltucopyTS, ctrmm_oltncopyTS,
  233. #endif
  234. ctrmm_ounucopyTS, ctrmm_ounncopyTS, ctrmm_outucopyTS, ctrmm_outncopyTS,
  235. ctrmm_olnucopyTS, ctrmm_olnncopyTS, ctrmm_oltucopyTS, ctrmm_oltncopyTS,
  236. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  237. csymm_iutcopyTS, csymm_iltcopyTS,
  238. #else
  239. csymm_outcopyTS, csymm_oltcopyTS,
  240. #endif
  241. csymm_outcopyTS, csymm_oltcopyTS,
  242. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  243. chemm_iutcopyTS, chemm_iltcopyTS,
  244. #else
  245. chemm_outcopyTS, chemm_oltcopyTS,
  246. #endif
  247. chemm_outcopyTS, chemm_oltcopyTS,
  248. cgemm3m_kernelTS,
  249. cgemm3m_incopybTS, cgemm3m_incopyrTS,
  250. cgemm3m_incopyiTS, cgemm3m_itcopybTS,
  251. cgemm3m_itcopyrTS, cgemm3m_itcopyiTS,
  252. cgemm3m_oncopybTS, cgemm3m_oncopyrTS,
  253. cgemm3m_oncopyiTS, cgemm3m_otcopybTS,
  254. cgemm3m_otcopyrTS, cgemm3m_otcopyiTS,
  255. csymm3m_iucopybTS, csymm3m_ilcopybTS,
  256. csymm3m_iucopyrTS, csymm3m_ilcopyrTS,
  257. csymm3m_iucopyiTS, csymm3m_ilcopyiTS,
  258. csymm3m_oucopybTS, csymm3m_olcopybTS,
  259. csymm3m_oucopyrTS, csymm3m_olcopyrTS,
  260. csymm3m_oucopyiTS, csymm3m_olcopyiTS,
  261. chemm3m_iucopybTS, chemm3m_ilcopybTS,
  262. chemm3m_iucopyrTS, chemm3m_ilcopyrTS,
  263. chemm3m_iucopyiTS, chemm3m_ilcopyiTS,
  264. chemm3m_oucopybTS, chemm3m_olcopybTS,
  265. chemm3m_oucopyrTS, chemm3m_olcopyrTS,
  266. chemm3m_oucopyiTS, chemm3m_olcopyiTS,
  267. #ifndef NO_LAPACK
  268. cneg_tcopyTS, claswp_ncopyTS,
  269. #else
  270. NULL, NULL,
  271. #endif
  272. 0, 0, 0,
  273. ZGEMM_DEFAULT_UNROLL_M, ZGEMM_DEFAULT_UNROLL_N, MAX(ZGEMM_DEFAULT_UNROLL_M, ZGEMM_DEFAULT_UNROLL_N),
  274. zamax_kTS, zamin_kTS, izamax_kTS, izamin_kTS,
  275. znrm2_kTS, zasum_kTS, zcopy_kTS,
  276. zdotu_kTS, zdotc_kTS, zdrot_kTS,
  277. zaxpy_kTS, zaxpyc_kTS, zscal_kTS, zswap_kTS,
  278. zgemv_nTS, zgemv_tTS, zgemv_rTS, zgemv_cTS,
  279. zgemv_oTS, zgemv_uTS, zgemv_sTS, zgemv_dTS,
  280. zgeru_kTS, zgerc_kTS, zgerv_kTS, zgerd_kTS,
  281. zsymv_LTS, zsymv_UTS,
  282. zhemv_LTS, zhemv_UTS, zhemv_MTS, zhemv_VTS,
  283. zgemm_kernel_nTS, zgemm_kernel_lTS, zgemm_kernel_rTS, zgemm_kernel_bTS,
  284. zgemm_betaTS,
  285. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  286. zgemm_incopyTS, zgemm_itcopyTS,
  287. #else
  288. zgemm_oncopyTS, zgemm_otcopyTS,
  289. #endif
  290. zgemm_oncopyTS, zgemm_otcopyTS,
  291. ztrsm_kernel_LNTS, ztrsm_kernel_LTTS, ztrsm_kernel_LRTS, ztrsm_kernel_LCTS,
  292. ztrsm_kernel_RNTS, ztrsm_kernel_RTTS, ztrsm_kernel_RRTS, ztrsm_kernel_RCTS,
  293. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  294. ztrsm_iunucopyTS, ztrsm_iunncopyTS, ztrsm_iutucopyTS, ztrsm_iutncopyTS,
  295. ztrsm_ilnucopyTS, ztrsm_ilnncopyTS, ztrsm_iltucopyTS, ztrsm_iltncopyTS,
  296. #else
  297. ztrsm_ounucopyTS, ztrsm_ounncopyTS, ztrsm_outucopyTS, ztrsm_outncopyTS,
  298. ztrsm_olnucopyTS, ztrsm_olnncopyTS, ztrsm_oltucopyTS, ztrsm_oltncopyTS,
  299. #endif
  300. ztrsm_ounucopyTS, ztrsm_ounncopyTS, ztrsm_outucopyTS, ztrsm_outncopyTS,
  301. ztrsm_olnucopyTS, ztrsm_olnncopyTS, ztrsm_oltucopyTS, ztrsm_oltncopyTS,
  302. ztrmm_kernel_RNTS, ztrmm_kernel_RTTS, ztrmm_kernel_RRTS, ztrmm_kernel_RCTS,
  303. ztrmm_kernel_LNTS, ztrmm_kernel_LTTS, ztrmm_kernel_LRTS, ztrmm_kernel_LCTS,
  304. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  305. ztrmm_iunucopyTS, ztrmm_iunncopyTS, ztrmm_iutucopyTS, ztrmm_iutncopyTS,
  306. ztrmm_ilnucopyTS, ztrmm_ilnncopyTS, ztrmm_iltucopyTS, ztrmm_iltncopyTS,
  307. #else
  308. ztrmm_ounucopyTS, ztrmm_ounncopyTS, ztrmm_outucopyTS, ztrmm_outncopyTS,
  309. ztrmm_olnucopyTS, ztrmm_olnncopyTS, ztrmm_oltucopyTS, ztrmm_oltncopyTS,
  310. #endif
  311. ztrmm_ounucopyTS, ztrmm_ounncopyTS, ztrmm_outucopyTS, ztrmm_outncopyTS,
  312. ztrmm_olnucopyTS, ztrmm_olnncopyTS, ztrmm_oltucopyTS, ztrmm_oltncopyTS,
  313. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  314. zsymm_iutcopyTS, zsymm_iltcopyTS,
  315. #else
  316. zsymm_outcopyTS, zsymm_oltcopyTS,
  317. #endif
  318. zsymm_outcopyTS, zsymm_oltcopyTS,
  319. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  320. zhemm_iutcopyTS, zhemm_iltcopyTS,
  321. #else
  322. zhemm_outcopyTS, zhemm_oltcopyTS,
  323. #endif
  324. zhemm_outcopyTS, zhemm_oltcopyTS,
  325. zgemm3m_kernelTS,
  326. zgemm3m_incopybTS, zgemm3m_incopyrTS,
  327. zgemm3m_incopyiTS, zgemm3m_itcopybTS,
  328. zgemm3m_itcopyrTS, zgemm3m_itcopyiTS,
  329. zgemm3m_oncopybTS, zgemm3m_oncopyrTS,
  330. zgemm3m_oncopyiTS, zgemm3m_otcopybTS,
  331. zgemm3m_otcopyrTS, zgemm3m_otcopyiTS,
  332. zsymm3m_iucopybTS, zsymm3m_ilcopybTS,
  333. zsymm3m_iucopyrTS, zsymm3m_ilcopyrTS,
  334. zsymm3m_iucopyiTS, zsymm3m_ilcopyiTS,
  335. zsymm3m_oucopybTS, zsymm3m_olcopybTS,
  336. zsymm3m_oucopyrTS, zsymm3m_olcopyrTS,
  337. zsymm3m_oucopyiTS, zsymm3m_olcopyiTS,
  338. zhemm3m_iucopybTS, zhemm3m_ilcopybTS,
  339. zhemm3m_iucopyrTS, zhemm3m_ilcopyrTS,
  340. zhemm3m_iucopyiTS, zhemm3m_ilcopyiTS,
  341. zhemm3m_oucopybTS, zhemm3m_olcopybTS,
  342. zhemm3m_oucopyrTS, zhemm3m_olcopyrTS,
  343. zhemm3m_oucopyiTS, zhemm3m_olcopyiTS,
  344. #ifndef NO_LAPACK
  345. zneg_tcopyTS, zlaswp_ncopyTS,
  346. #else
  347. NULL, NULL,
  348. #endif
  349. #ifdef EXPRECISION
  350. 0, 0, 0,
  351. XGEMM_DEFAULT_UNROLL_M, XGEMM_DEFAULT_UNROLL_N, MAX(XGEMM_DEFAULT_UNROLL_M, XGEMM_DEFAULT_UNROLL_N),
  352. xamax_kTS, xamin_kTS, ixamax_kTS, ixamin_kTS,
  353. xnrm2_kTS, xasum_kTS, xcopy_kTS,
  354. xdotu_kTS, xdotc_kTS, xqrot_kTS,
  355. xaxpy_kTS, xaxpyc_kTS, xscal_kTS, xswap_kTS,
  356. xgemv_nTS, xgemv_tTS, xgemv_rTS, xgemv_cTS,
  357. xgemv_oTS, xgemv_uTS, xgemv_sTS, xgemv_dTS,
  358. xgeru_kTS, xgerc_kTS, xgerv_kTS, xgerd_kTS,
  359. xsymv_LTS, xsymv_UTS,
  360. xhemv_LTS, xhemv_UTS, xhemv_MTS, xhemv_VTS,
  361. xgemm_kernel_nTS, xgemm_kernel_lTS, xgemm_kernel_rTS, xgemm_kernel_bTS,
  362. xgemm_betaTS,
  363. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  364. xgemm_incopyTS, xgemm_itcopyTS,
  365. #else
  366. xgemm_oncopyTS, xgemm_otcopyTS,
  367. #endif
  368. xgemm_oncopyTS, xgemm_otcopyTS,
  369. xtrsm_kernel_LNTS, xtrsm_kernel_LTTS, xtrsm_kernel_LRTS, xtrsm_kernel_LCTS,
  370. xtrsm_kernel_RNTS, xtrsm_kernel_RTTS, xtrsm_kernel_RRTS, xtrsm_kernel_RCTS,
  371. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  372. xtrsm_iunucopyTS, xtrsm_iunncopyTS, xtrsm_iutucopyTS, xtrsm_iutncopyTS,
  373. xtrsm_ilnucopyTS, xtrsm_ilnncopyTS, xtrsm_iltucopyTS, xtrsm_iltncopyTS,
  374. #else
  375. xtrsm_ounucopyTS, xtrsm_ounncopyTS, xtrsm_outucopyTS, xtrsm_outncopyTS,
  376. xtrsm_olnucopyTS, xtrsm_olnncopyTS, xtrsm_oltucopyTS, xtrsm_oltncopyTS,
  377. #endif
  378. xtrsm_ounucopyTS, xtrsm_ounncopyTS, xtrsm_outucopyTS, xtrsm_outncopyTS,
  379. xtrsm_olnucopyTS, xtrsm_olnncopyTS, xtrsm_oltucopyTS, xtrsm_oltncopyTS,
  380. xtrmm_kernel_RNTS, xtrmm_kernel_RTTS, xtrmm_kernel_RRTS, xtrmm_kernel_RCTS,
  381. xtrmm_kernel_LNTS, xtrmm_kernel_LTTS, xtrmm_kernel_LRTS, xtrmm_kernel_LCTS,
  382. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  383. xtrmm_iunucopyTS, xtrmm_iunncopyTS, xtrmm_iutucopyTS, xtrmm_iutncopyTS,
  384. xtrmm_ilnucopyTS, xtrmm_ilnncopyTS, xtrmm_iltucopyTS, xtrmm_iltncopyTS,
  385. #else
  386. xtrmm_ounucopyTS, xtrmm_ounncopyTS, xtrmm_outucopyTS, xtrmm_outncopyTS,
  387. xtrmm_olnucopyTS, xtrmm_olnncopyTS, xtrmm_oltucopyTS, xtrmm_oltncopyTS,
  388. #endif
  389. xtrmm_ounucopyTS, xtrmm_ounncopyTS, xtrmm_outucopyTS, xtrmm_outncopyTS,
  390. xtrmm_olnucopyTS, xtrmm_olnncopyTS, xtrmm_oltucopyTS, xtrmm_oltncopyTS,
  391. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  392. xsymm_iutcopyTS, xsymm_iltcopyTS,
  393. #else
  394. xsymm_outcopyTS, xsymm_oltcopyTS,
  395. #endif
  396. xsymm_outcopyTS, xsymm_oltcopyTS,
  397. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  398. xhemm_iutcopyTS, xhemm_iltcopyTS,
  399. #else
  400. xhemm_outcopyTS, xhemm_oltcopyTS,
  401. #endif
  402. xhemm_outcopyTS, xhemm_oltcopyTS,
  403. xgemm3m_kernelTS,
  404. xgemm3m_incopybTS, xgemm3m_incopyrTS,
  405. xgemm3m_incopyiTS, xgemm3m_itcopybTS,
  406. xgemm3m_itcopyrTS, xgemm3m_itcopyiTS,
  407. xgemm3m_oncopybTS, xgemm3m_oncopyrTS,
  408. xgemm3m_oncopyiTS, xgemm3m_otcopybTS,
  409. xgemm3m_otcopyrTS, xgemm3m_otcopyiTS,
  410. xsymm3m_iucopybTS, xsymm3m_ilcopybTS,
  411. xsymm3m_iucopyrTS, xsymm3m_ilcopyrTS,
  412. xsymm3m_iucopyiTS, xsymm3m_ilcopyiTS,
  413. xsymm3m_oucopybTS, xsymm3m_olcopybTS,
  414. xsymm3m_oucopyrTS, xsymm3m_olcopyrTS,
  415. xsymm3m_oucopyiTS, xsymm3m_olcopyiTS,
  416. xhemm3m_iucopybTS, xhemm3m_ilcopybTS,
  417. xhemm3m_iucopyrTS, xhemm3m_ilcopyrTS,
  418. xhemm3m_iucopyiTS, xhemm3m_ilcopyiTS,
  419. xhemm3m_oucopybTS, xhemm3m_olcopybTS,
  420. xhemm3m_oucopyrTS, xhemm3m_olcopyrTS,
  421. xhemm3m_oucopyiTS, xhemm3m_olcopyiTS,
  422. #ifndef NO_LAPACK
  423. xneg_tcopyTS, xlaswp_ncopyTS,
  424. #else
  425. NULL, NULL,
  426. #endif
  427. #endif
  428. init_parameter,
  429. SNUMOPT, DNUMOPT, QNUMOPT,
  430. };
  431. #ifdef ARCH_X86
  432. static int get_l2_size_old(void){
  433. int i, eax, ebx, ecx, edx, cpuid_level;
  434. int info[15];
  435. cpuid(2, &eax, &ebx, &ecx, &edx);
  436. info[ 0] = BITMASK(eax, 8, 0xff);
  437. info[ 1] = BITMASK(eax, 16, 0xff);
  438. info[ 2] = BITMASK(eax, 24, 0xff);
  439. info[ 3] = BITMASK(ebx, 0, 0xff);
  440. info[ 4] = BITMASK(ebx, 8, 0xff);
  441. info[ 5] = BITMASK(ebx, 16, 0xff);
  442. info[ 6] = BITMASK(ebx, 24, 0xff);
  443. info[ 7] = BITMASK(ecx, 0, 0xff);
  444. info[ 8] = BITMASK(ecx, 8, 0xff);
  445. info[ 9] = BITMASK(ecx, 16, 0xff);
  446. info[10] = BITMASK(ecx, 24, 0xff);
  447. info[11] = BITMASK(edx, 0, 0xff);
  448. info[12] = BITMASK(edx, 8, 0xff);
  449. info[13] = BITMASK(edx, 16, 0xff);
  450. info[14] = BITMASK(edx, 24, 0xff);
  451. for (i = 0; i < 15; i++){
  452. switch (info[i]){
  453. /* This table is from http://www.sandpile.org/ia32/cpuid.htm */
  454. case 0x1a :
  455. return 96;
  456. case 0x39 :
  457. case 0x3b :
  458. case 0x41 :
  459. case 0x79 :
  460. case 0x81 :
  461. return 128;
  462. case 0x3a :
  463. return 192;
  464. case 0x21 :
  465. case 0x3c :
  466. case 0x42 :
  467. case 0x7a :
  468. case 0x7e :
  469. case 0x82 :
  470. return 256;
  471. case 0x3d :
  472. return 384;
  473. case 0x3e :
  474. case 0x43 :
  475. case 0x7b :
  476. case 0x7f :
  477. case 0x83 :
  478. case 0x86 :
  479. return 512;
  480. case 0x44 :
  481. case 0x78 :
  482. case 0x7c :
  483. case 0x84 :
  484. case 0x87 :
  485. return 1024;
  486. case 0x45 :
  487. case 0x7d :
  488. case 0x85 :
  489. return 2048;
  490. case 0x48 :
  491. return 3184;
  492. case 0x49 :
  493. return 4096;
  494. case 0x4e :
  495. return 6144;
  496. }
  497. }
  498. return 0;
  499. }
  500. #endif
  501. static __inline__ int get_l2_size(void){
  502. int eax, ebx, ecx, edx, l2;
  503. cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
  504. l2 = BITMASK(ecx, 16, 0xffff);
  505. #ifndef ARCH_X86
  506. return l2;
  507. #else
  508. if (l2 > 0) return l2;
  509. return get_l2_size_old();
  510. #endif
  511. }
  512. static __inline__ int get_l3_size(void){
  513. int eax, ebx, ecx, edx;
  514. cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
  515. return BITMASK(edx, 18, 0x3fff) * 512;
  516. }
  517. static void init_parameter(void) {
  518. int l2 = get_l2_size();
  519. TABLE_NAME.sgemm_q = SGEMM_DEFAULT_Q;
  520. TABLE_NAME.dgemm_q = DGEMM_DEFAULT_Q;
  521. TABLE_NAME.cgemm_q = CGEMM_DEFAULT_Q;
  522. TABLE_NAME.zgemm_q = ZGEMM_DEFAULT_Q;
  523. #ifdef EXPRECISION
  524. TABLE_NAME.qgemm_q = QGEMM_DEFAULT_Q;
  525. TABLE_NAME.xgemm_q = XGEMM_DEFAULT_Q;
  526. #endif
  527. #if defined(CORE_KATMAI) || defined(CORE_COPPERMINE) || defined(CORE_BANIAS) || defined(CORE_YONAH) || defined(CORE_ATHLON)
  528. #ifdef DEBUG
  529. fprintf(stderr, "Katmai, Coppermine, Banias, Athlon\n");
  530. #endif
  531. TABLE_NAME.sgemm_p = 64 * (l2 >> 7);
  532. TABLE_NAME.dgemm_p = 32 * (l2 >> 7);
  533. TABLE_NAME.cgemm_p = 32 * (l2 >> 7);
  534. TABLE_NAME.zgemm_p = 16 * (l2 >> 7);
  535. #ifdef EXPRECISION
  536. TABLE_NAME.qgemm_p = 16 * (l2 >> 7);
  537. TABLE_NAME.xgemm_p = 8 * (l2 >> 7);
  538. #endif
  539. #endif
  540. #ifdef CORE_NORTHWOOD
  541. #ifdef DEBUG
  542. fprintf(stderr, "Northwood\n");
  543. #endif
  544. TABLE_NAME.sgemm_p = 96 * (l2 >> 7);
  545. TABLE_NAME.dgemm_p = 48 * (l2 >> 7);
  546. TABLE_NAME.cgemm_p = 48 * (l2 >> 7);
  547. TABLE_NAME.zgemm_p = 24 * (l2 >> 7);
  548. #ifdef EXPRECISION
  549. TABLE_NAME.qgemm_p = 24 * (l2 >> 7);
  550. TABLE_NAME.xgemm_p = 12 * (l2 >> 7);
  551. #endif
  552. #endif
  553. #ifdef ATOM
  554. #ifdef DEBUG
  555. fprintf(stderr, "Atom\n");
  556. #endif
  557. TABLE_NAME.sgemm_p = 256;
  558. TABLE_NAME.dgemm_p = 128;
  559. TABLE_NAME.cgemm_p = 128;
  560. TABLE_NAME.zgemm_p = 64;
  561. #ifdef EXPRECISION
  562. TABLE_NAME.qgemm_p = 64;
  563. TABLE_NAME.xgemm_p = 32;
  564. #endif
  565. #endif
  566. #ifdef CORE_PRESCOTT
  567. #ifdef DEBUG
  568. fprintf(stderr, "Prescott\n");
  569. #endif
  570. TABLE_NAME.sgemm_p = 56 * (l2 >> 7);
  571. TABLE_NAME.dgemm_p = 28 * (l2 >> 7);
  572. TABLE_NAME.cgemm_p = 28 * (l2 >> 7);
  573. TABLE_NAME.zgemm_p = 14 * (l2 >> 7);
  574. #ifdef EXPRECISION
  575. TABLE_NAME.qgemm_p = 14 * (l2 >> 7);
  576. TABLE_NAME.xgemm_p = 7 * (l2 >> 7);
  577. #endif
  578. #endif
  579. #ifdef CORE2
  580. #ifdef DEBUG
  581. fprintf(stderr, "Core2\n");
  582. #endif
  583. TABLE_NAME.sgemm_p = 92 * (l2 >> 9);
  584. TABLE_NAME.dgemm_p = 46 * (l2 >> 9);
  585. TABLE_NAME.cgemm_p = 46 * (l2 >> 9);
  586. TABLE_NAME.zgemm_p = 23 * (l2 >> 9);
  587. #ifdef EXPRECISION
  588. TABLE_NAME.qgemm_p = 92 * (l2 >> 9);
  589. TABLE_NAME.xgemm_p = 46 * (l2 >> 9);
  590. #endif
  591. #endif
  592. #ifdef PENRYN
  593. #ifdef DEBUG
  594. fprintf(stderr, "Penryn\n");
  595. #endif
  596. TABLE_NAME.sgemm_p = 42 * (l2 >> 9) + 8;
  597. TABLE_NAME.dgemm_p = 42 * (l2 >> 9) + 8;
  598. TABLE_NAME.cgemm_p = 21 * (l2 >> 9) + 4;
  599. TABLE_NAME.zgemm_p = 21 * (l2 >> 9) + 4;
  600. #ifdef EXPRECISION
  601. TABLE_NAME.qgemm_p = 42 * (l2 >> 9) + 8;
  602. TABLE_NAME.xgemm_p = 21 * (l2 >> 9) + 4;
  603. #endif
  604. #endif
  605. #ifdef NEHALEM
  606. #ifdef DEBUG
  607. fprintf(stderr, "Nehalem\n");
  608. #endif
  609. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  610. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  611. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  612. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  613. #ifdef EXPRECISION
  614. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  615. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  616. #endif
  617. #endif
  618. #ifdef SANDYBRIDGE
  619. #ifdef DEBUG
  620. fprintf(stderr, "Sandybridge\n");
  621. #endif
  622. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  623. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  624. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  625. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  626. #ifdef EXPRECISION
  627. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  628. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  629. #endif
  630. #endif
  631. #ifdef OPTERON
  632. #ifdef DEBUG
  633. fprintf(stderr, "Opteron\n");
  634. #endif
  635. TABLE_NAME.sgemm_p = 224 + 56 * (l2 >> 7);
  636. TABLE_NAME.dgemm_p = 112 + 28 * (l2 >> 7);
  637. TABLE_NAME.cgemm_p = 112 + 28 * (l2 >> 7);
  638. TABLE_NAME.zgemm_p = 56 + 14 * (l2 >> 7);
  639. #ifdef EXPRECISION
  640. TABLE_NAME.qgemm_p = 56 + 14 * (l2 >> 7);
  641. TABLE_NAME.xgemm_p = 28 + 7 * (l2 >> 7);
  642. #endif
  643. #endif
  644. #ifdef BARCELONA
  645. #ifdef DEBUG
  646. fprintf(stderr, "Barcelona\n");
  647. #endif
  648. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  649. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  650. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  651. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  652. #ifdef EXPRECISION
  653. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  654. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  655. #endif
  656. #endif
  657. #ifdef BOBCAT
  658. #ifdef DEBUG
  659. fprintf(stderr, "Bobcate\n");
  660. #endif
  661. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  662. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  663. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  664. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  665. #ifdef EXPRECISION
  666. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  667. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  668. #endif
  669. #endif
  670. #ifdef BULLDOZER
  671. #ifdef DEBUG
  672. fprintf(stderr, "Bulldozer\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 NANO
  684. #ifdef DEBUG
  685. fprintf(stderr, "NANO\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. TABLE_NAME.sgemm_p = (TABLE_NAME.sgemm_p + SGEMM_DEFAULT_UNROLL_M - 1) & ~(SGEMM_DEFAULT_UNROLL_M - 1);
  697. TABLE_NAME.dgemm_p = (TABLE_NAME.dgemm_p + DGEMM_DEFAULT_UNROLL_M - 1) & ~(DGEMM_DEFAULT_UNROLL_M - 1);
  698. TABLE_NAME.cgemm_p = (TABLE_NAME.cgemm_p + CGEMM_DEFAULT_UNROLL_M - 1) & ~(CGEMM_DEFAULT_UNROLL_M - 1);
  699. TABLE_NAME.zgemm_p = (TABLE_NAME.zgemm_p + ZGEMM_DEFAULT_UNROLL_M - 1) & ~(ZGEMM_DEFAULT_UNROLL_M - 1);
  700. #ifdef QUAD_PRECISION
  701. TABLE_NAME.qgemm_p = (TABLE_NAME.qgemm_p + QGEMM_DEFAULT_UNROLL_M - 1) & ~(QGEMM_DEFAULT_UNROLL_M - 1);
  702. TABLE_NAME.xgemm_p = (TABLE_NAME.xgemm_p + XGEMM_DEFAULT_UNROLL_M - 1) & ~(XGEMM_DEFAULT_UNROLL_M - 1);
  703. #endif
  704. #ifdef DEBUG
  705. fprintf(stderr, "L2 = %8d DGEMM_P .. %d\n", l2, TABLE_NAME.dgemm_p);
  706. #endif
  707. TABLE_NAME.sgemm_r = (((BUFFER_SIZE -
  708. ((TABLE_NAME.sgemm_p * TABLE_NAME.sgemm_q * 4 + TABLE_NAME.offsetA
  709. + TABLE_NAME.align) & ~TABLE_NAME.align)
  710. ) / (TABLE_NAME.sgemm_q * 4) - 15) & ~15);
  711. TABLE_NAME.dgemm_r = (((BUFFER_SIZE -
  712. ((TABLE_NAME.dgemm_p * TABLE_NAME.dgemm_q * 8 + TABLE_NAME.offsetA
  713. + TABLE_NAME.align) & ~TABLE_NAME.align)
  714. ) / (TABLE_NAME.dgemm_q * 8) - 15) & ~15);
  715. #ifdef EXPRECISION
  716. TABLE_NAME.qgemm_r = (((BUFFER_SIZE -
  717. ((TABLE_NAME.qgemm_p * TABLE_NAME.qgemm_q * 16 + TABLE_NAME.offsetA
  718. + TABLE_NAME.align) & ~TABLE_NAME.align)
  719. ) / (TABLE_NAME.qgemm_q * 16) - 15) & ~15);
  720. #endif
  721. TABLE_NAME.cgemm_r = (((BUFFER_SIZE -
  722. ((TABLE_NAME.cgemm_p * TABLE_NAME.cgemm_q * 8 + TABLE_NAME.offsetA
  723. + TABLE_NAME.align) & ~TABLE_NAME.align)
  724. ) / (TABLE_NAME.cgemm_q * 8) - 15) & ~15);
  725. TABLE_NAME.zgemm_r = (((BUFFER_SIZE -
  726. ((TABLE_NAME.zgemm_p * TABLE_NAME.zgemm_q * 16 + TABLE_NAME.offsetA
  727. + TABLE_NAME.align) & ~TABLE_NAME.align)
  728. ) / (TABLE_NAME.zgemm_q * 16) - 15) & ~15);
  729. #ifdef EXPRECISION
  730. TABLE_NAME.xgemm_r = (((BUFFER_SIZE -
  731. ((TABLE_NAME.xgemm_p * TABLE_NAME.xgemm_q * 32 + TABLE_NAME.offsetA
  732. + TABLE_NAME.align) & ~TABLE_NAME.align)
  733. ) / (TABLE_NAME.xgemm_q * 32) - 15) & ~15);
  734. #endif
  735. }