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 27 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819
  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. GEMM_DEFAULT_OFFSET_A, GEMM_DEFAULT_OFFSET_B, GEMM_DEFAULT_ALIGN,
  48. 0, 0, 0,
  49. SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N, MAX(SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N),
  50. #ifdef HAVE_EXCLUSIVE_CACHE
  51. 1,
  52. #else
  53. 0,
  54. #endif
  55. samax_kTS, samin_kTS, smax_kTS, smin_kTS,
  56. isamax_kTS, isamin_kTS, ismax_kTS, ismin_kTS,
  57. snrm2_kTS, sasum_kTS, scopy_kTS, sdot_kTS,
  58. dsdot_kTS,
  59. srot_kTS, saxpy_kTS, sscal_kTS, sswap_kTS,
  60. sgemv_nTS, sgemv_tTS, sger_kTS,
  61. ssymv_LTS, ssymv_UTS,
  62. sgemm_kernelTS, sgemm_betaTS,
  63. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  64. sgemm_incopyTS, sgemm_itcopyTS,
  65. #else
  66. sgemm_oncopyTS, sgemm_otcopyTS,
  67. #endif
  68. sgemm_oncopyTS, sgemm_otcopyTS,
  69. strsm_kernel_LNTS, strsm_kernel_LTTS, strsm_kernel_RNTS, strsm_kernel_RTTS,
  70. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  71. strsm_iunucopyTS, strsm_iunncopyTS, strsm_iutucopyTS, strsm_iutncopyTS,
  72. strsm_ilnucopyTS, strsm_ilnncopyTS, strsm_iltucopyTS, strsm_iltncopyTS,
  73. #else
  74. strsm_ounucopyTS, strsm_ounncopyTS, strsm_outucopyTS, strsm_outncopyTS,
  75. strsm_olnucopyTS, strsm_olnncopyTS, strsm_oltucopyTS, strsm_oltncopyTS,
  76. #endif
  77. strsm_ounucopyTS, strsm_ounncopyTS, strsm_outucopyTS, strsm_outncopyTS,
  78. strsm_olnucopyTS, strsm_olnncopyTS, strsm_oltucopyTS, strsm_oltncopyTS,
  79. strmm_kernel_RNTS, strmm_kernel_RTTS, strmm_kernel_LNTS, strmm_kernel_LTTS,
  80. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  81. strmm_iunucopyTS, strmm_iunncopyTS, strmm_iutucopyTS, strmm_iutncopyTS,
  82. strmm_ilnucopyTS, strmm_ilnncopyTS, strmm_iltucopyTS, strmm_iltncopyTS,
  83. #else
  84. strmm_ounucopyTS, strmm_ounncopyTS, strmm_outucopyTS, strmm_outncopyTS,
  85. strmm_olnucopyTS, strmm_olnncopyTS, strmm_oltucopyTS, strmm_oltncopyTS,
  86. #endif
  87. strmm_ounucopyTS, strmm_ounncopyTS, strmm_outucopyTS, strmm_outncopyTS,
  88. strmm_olnucopyTS, strmm_olnncopyTS, strmm_oltucopyTS, strmm_oltncopyTS,
  89. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  90. ssymm_iutcopyTS, ssymm_iltcopyTS,
  91. #else
  92. ssymm_outcopyTS, ssymm_oltcopyTS,
  93. #endif
  94. ssymm_outcopyTS, ssymm_oltcopyTS,
  95. sneg_tcopyTS, slaswp_ncopyTS,
  96. 0, 0, 0,
  97. DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N, MAX(DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N),
  98. damax_kTS, damin_kTS, dmax_kTS, dmin_kTS,
  99. idamax_kTS, idamin_kTS, idmax_kTS, idmin_kTS,
  100. dnrm2_kTS, dasum_kTS, dcopy_kTS, ddot_kTS,
  101. drot_kTS, daxpy_kTS, dscal_kTS, dswap_kTS,
  102. dgemv_nTS, dgemv_tTS, dger_kTS,
  103. dsymv_LTS, dsymv_UTS,
  104. dgemm_kernelTS, dgemm_betaTS,
  105. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  106. dgemm_incopyTS, dgemm_itcopyTS,
  107. #else
  108. dgemm_oncopyTS, dgemm_otcopyTS,
  109. #endif
  110. dgemm_oncopyTS, dgemm_otcopyTS,
  111. dtrsm_kernel_LNTS, dtrsm_kernel_LTTS, dtrsm_kernel_RNTS, dtrsm_kernel_RTTS,
  112. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  113. dtrsm_iunucopyTS, dtrsm_iunncopyTS, dtrsm_iutucopyTS, dtrsm_iutncopyTS,
  114. dtrsm_ilnucopyTS, dtrsm_ilnncopyTS, dtrsm_iltucopyTS, dtrsm_iltncopyTS,
  115. #else
  116. dtrsm_ounucopyTS, dtrsm_ounncopyTS, dtrsm_outucopyTS, dtrsm_outncopyTS,
  117. dtrsm_olnucopyTS, dtrsm_olnncopyTS, dtrsm_oltucopyTS, dtrsm_oltncopyTS,
  118. #endif
  119. dtrsm_ounucopyTS, dtrsm_ounncopyTS, dtrsm_outucopyTS, dtrsm_outncopyTS,
  120. dtrsm_olnucopyTS, dtrsm_olnncopyTS, dtrsm_oltucopyTS, dtrsm_oltncopyTS,
  121. dtrmm_kernel_RNTS, dtrmm_kernel_RTTS, dtrmm_kernel_LNTS, dtrmm_kernel_LTTS,
  122. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  123. dtrmm_iunucopyTS, dtrmm_iunncopyTS, dtrmm_iutucopyTS, dtrmm_iutncopyTS,
  124. dtrmm_ilnucopyTS, dtrmm_ilnncopyTS, dtrmm_iltucopyTS, dtrmm_iltncopyTS,
  125. #else
  126. dtrmm_ounucopyTS, dtrmm_ounncopyTS, dtrmm_outucopyTS, dtrmm_outncopyTS,
  127. dtrmm_olnucopyTS, dtrmm_olnncopyTS, dtrmm_oltucopyTS, dtrmm_oltncopyTS,
  128. #endif
  129. dtrmm_ounucopyTS, dtrmm_ounncopyTS, dtrmm_outucopyTS, dtrmm_outncopyTS,
  130. dtrmm_olnucopyTS, dtrmm_olnncopyTS, dtrmm_oltucopyTS, dtrmm_oltncopyTS,
  131. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  132. dsymm_iutcopyTS, dsymm_iltcopyTS,
  133. #else
  134. dsymm_outcopyTS, dsymm_oltcopyTS,
  135. #endif
  136. dsymm_outcopyTS, dsymm_oltcopyTS,
  137. dneg_tcopyTS, dlaswp_ncopyTS,
  138. #ifdef EXPRECISION
  139. 0, 0, 0,
  140. QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N, MAX(QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N),
  141. qamax_kTS, qamin_kTS, qmax_kTS, qmin_kTS,
  142. iqamax_kTS, iqamin_kTS, iqmax_kTS, iqmin_kTS,
  143. qnrm2_kTS, qasum_kTS, qcopy_kTS, qdot_kTS,
  144. qrot_kTS, qaxpy_kTS, qscal_kTS, qswap_kTS,
  145. qgemv_nTS, qgemv_tTS, qger_kTS,
  146. qsymv_LTS, qsymv_UTS,
  147. qgemm_kernelTS, qgemm_betaTS,
  148. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  149. qgemm_incopyTS, qgemm_itcopyTS,
  150. #else
  151. qgemm_oncopyTS, qgemm_otcopyTS,
  152. #endif
  153. qgemm_oncopyTS, qgemm_otcopyTS,
  154. qtrsm_kernel_LNTS, qtrsm_kernel_LTTS, qtrsm_kernel_RNTS, qtrsm_kernel_RTTS,
  155. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  156. qtrsm_iunucopyTS, qtrsm_iunncopyTS, qtrsm_iutucopyTS, qtrsm_iutncopyTS,
  157. qtrsm_ilnucopyTS, qtrsm_ilnncopyTS, qtrsm_iltucopyTS, qtrsm_iltncopyTS,
  158. #else
  159. qtrsm_ounucopyTS, qtrsm_ounncopyTS, qtrsm_outucopyTS, qtrsm_outncopyTS,
  160. qtrsm_olnucopyTS, qtrsm_olnncopyTS, qtrsm_oltucopyTS, qtrsm_oltncopyTS,
  161. #endif
  162. qtrsm_ounucopyTS, qtrsm_ounncopyTS, qtrsm_outucopyTS, qtrsm_outncopyTS,
  163. qtrsm_olnucopyTS, qtrsm_olnncopyTS, qtrsm_oltucopyTS, qtrsm_oltncopyTS,
  164. qtrmm_kernel_RNTS, qtrmm_kernel_RTTS, qtrmm_kernel_LNTS, qtrmm_kernel_LTTS,
  165. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  166. qtrmm_iunucopyTS, qtrmm_iunncopyTS, qtrmm_iutucopyTS, qtrmm_iutncopyTS,
  167. qtrmm_ilnucopyTS, qtrmm_ilnncopyTS, qtrmm_iltucopyTS, qtrmm_iltncopyTS,
  168. #else
  169. qtrmm_ounucopyTS, qtrmm_ounncopyTS, qtrmm_outucopyTS, qtrmm_outncopyTS,
  170. qtrmm_olnucopyTS, qtrmm_olnncopyTS, qtrmm_oltucopyTS, qtrmm_oltncopyTS,
  171. #endif
  172. qtrmm_ounucopyTS, qtrmm_ounncopyTS, qtrmm_outucopyTS, qtrmm_outncopyTS,
  173. qtrmm_olnucopyTS, qtrmm_olnncopyTS, qtrmm_oltucopyTS, qtrmm_oltncopyTS,
  174. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  175. qsymm_iutcopyTS, qsymm_iltcopyTS,
  176. #else
  177. qsymm_outcopyTS, qsymm_oltcopyTS,
  178. #endif
  179. qsymm_outcopyTS, qsymm_oltcopyTS,
  180. qneg_tcopyTS, qlaswp_ncopyTS,
  181. #endif
  182. 0, 0, 0,
  183. CGEMM_DEFAULT_UNROLL_M, CGEMM_DEFAULT_UNROLL_N, MAX(CGEMM_DEFAULT_UNROLL_M, CGEMM_DEFAULT_UNROLL_N),
  184. camax_kTS, camin_kTS, icamax_kTS, icamin_kTS,
  185. cnrm2_kTS, casum_kTS, ccopy_kTS,
  186. cdotu_kTS, cdotc_kTS, csrot_kTS,
  187. caxpy_kTS, caxpyc_kTS, cscal_kTS, cswap_kTS,
  188. cgemv_nTS, cgemv_tTS, cgemv_rTS, cgemv_cTS,
  189. cgemv_oTS, cgemv_uTS, cgemv_sTS, cgemv_dTS,
  190. cgeru_kTS, cgerc_kTS, cgerv_kTS, cgerd_kTS,
  191. csymv_LTS, csymv_UTS,
  192. chemv_LTS, chemv_UTS, chemv_MTS, chemv_VTS,
  193. cgemm_kernel_nTS, cgemm_kernel_lTS, cgemm_kernel_rTS, cgemm_kernel_bTS,
  194. cgemm_betaTS,
  195. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  196. cgemm_incopyTS, cgemm_itcopyTS,
  197. #else
  198. cgemm_oncopyTS, cgemm_otcopyTS,
  199. #endif
  200. cgemm_oncopyTS, cgemm_otcopyTS,
  201. ctrsm_kernel_LNTS, ctrsm_kernel_LTTS, ctrsm_kernel_LRTS, ctrsm_kernel_LCTS,
  202. ctrsm_kernel_RNTS, ctrsm_kernel_RTTS, ctrsm_kernel_RRTS, ctrsm_kernel_RCTS,
  203. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  204. ctrsm_iunucopyTS, ctrsm_iunncopyTS, ctrsm_iutucopyTS, ctrsm_iutncopyTS,
  205. ctrsm_ilnucopyTS, ctrsm_ilnncopyTS, ctrsm_iltucopyTS, ctrsm_iltncopyTS,
  206. #else
  207. ctrsm_ounucopyTS, ctrsm_ounncopyTS, ctrsm_outucopyTS, ctrsm_outncopyTS,
  208. ctrsm_olnucopyTS, ctrsm_olnncopyTS, ctrsm_oltucopyTS, ctrsm_oltncopyTS,
  209. #endif
  210. ctrsm_ounucopyTS, ctrsm_ounncopyTS, ctrsm_outucopyTS, ctrsm_outncopyTS,
  211. ctrsm_olnucopyTS, ctrsm_olnncopyTS, ctrsm_oltucopyTS, ctrsm_oltncopyTS,
  212. ctrmm_kernel_RNTS, ctrmm_kernel_RTTS, ctrmm_kernel_RRTS, ctrmm_kernel_RCTS,
  213. ctrmm_kernel_LNTS, ctrmm_kernel_LTTS, ctrmm_kernel_LRTS, ctrmm_kernel_LCTS,
  214. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  215. ctrmm_iunucopyTS, ctrmm_iunncopyTS, ctrmm_iutucopyTS, ctrmm_iutncopyTS,
  216. ctrmm_ilnucopyTS, ctrmm_ilnncopyTS, ctrmm_iltucopyTS, ctrmm_iltncopyTS,
  217. #else
  218. ctrmm_ounucopyTS, ctrmm_ounncopyTS, ctrmm_outucopyTS, ctrmm_outncopyTS,
  219. ctrmm_olnucopyTS, ctrmm_olnncopyTS, ctrmm_oltucopyTS, ctrmm_oltncopyTS,
  220. #endif
  221. ctrmm_ounucopyTS, ctrmm_ounncopyTS, ctrmm_outucopyTS, ctrmm_outncopyTS,
  222. ctrmm_olnucopyTS, ctrmm_olnncopyTS, ctrmm_oltucopyTS, ctrmm_oltncopyTS,
  223. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  224. csymm_iutcopyTS, csymm_iltcopyTS,
  225. #else
  226. csymm_outcopyTS, csymm_oltcopyTS,
  227. #endif
  228. csymm_outcopyTS, csymm_oltcopyTS,
  229. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  230. chemm_iutcopyTS, chemm_iltcopyTS,
  231. #else
  232. chemm_outcopyTS, chemm_oltcopyTS,
  233. #endif
  234. chemm_outcopyTS, chemm_oltcopyTS,
  235. cgemm3m_kernelTS,
  236. cgemm3m_incopybTS, cgemm3m_incopyrTS,
  237. cgemm3m_incopyiTS, cgemm3m_itcopybTS,
  238. cgemm3m_itcopyrTS, cgemm3m_itcopyiTS,
  239. cgemm3m_oncopybTS, cgemm3m_oncopyrTS,
  240. cgemm3m_oncopyiTS, cgemm3m_otcopybTS,
  241. cgemm3m_otcopyrTS, cgemm3m_otcopyiTS,
  242. csymm3m_iucopybTS, csymm3m_ilcopybTS,
  243. csymm3m_iucopyrTS, csymm3m_ilcopyrTS,
  244. csymm3m_iucopyiTS, csymm3m_ilcopyiTS,
  245. csymm3m_oucopybTS, csymm3m_olcopybTS,
  246. csymm3m_oucopyrTS, csymm3m_olcopyrTS,
  247. csymm3m_oucopyiTS, csymm3m_olcopyiTS,
  248. chemm3m_iucopybTS, chemm3m_ilcopybTS,
  249. chemm3m_iucopyrTS, chemm3m_ilcopyrTS,
  250. chemm3m_iucopyiTS, chemm3m_ilcopyiTS,
  251. chemm3m_oucopybTS, chemm3m_olcopybTS,
  252. chemm3m_oucopyrTS, chemm3m_olcopyrTS,
  253. chemm3m_oucopyiTS, chemm3m_olcopyiTS,
  254. cneg_tcopyTS, claswp_ncopyTS,
  255. 0, 0, 0,
  256. ZGEMM_DEFAULT_UNROLL_M, ZGEMM_DEFAULT_UNROLL_N, MAX(ZGEMM_DEFAULT_UNROLL_M, ZGEMM_DEFAULT_UNROLL_N),
  257. zamax_kTS, zamin_kTS, izamax_kTS, izamin_kTS,
  258. znrm2_kTS, zasum_kTS, zcopy_kTS,
  259. zdotu_kTS, zdotc_kTS, zdrot_kTS,
  260. zaxpy_kTS, zaxpyc_kTS, zscal_kTS, zswap_kTS,
  261. zgemv_nTS, zgemv_tTS, zgemv_rTS, zgemv_cTS,
  262. zgemv_oTS, zgemv_uTS, zgemv_sTS, zgemv_dTS,
  263. zgeru_kTS, zgerc_kTS, zgerv_kTS, zgerd_kTS,
  264. zsymv_LTS, zsymv_UTS,
  265. zhemv_LTS, zhemv_UTS, zhemv_MTS, zhemv_VTS,
  266. zgemm_kernel_nTS, zgemm_kernel_lTS, zgemm_kernel_rTS, zgemm_kernel_bTS,
  267. zgemm_betaTS,
  268. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  269. zgemm_incopyTS, zgemm_itcopyTS,
  270. #else
  271. zgemm_oncopyTS, zgemm_otcopyTS,
  272. #endif
  273. zgemm_oncopyTS, zgemm_otcopyTS,
  274. ztrsm_kernel_LNTS, ztrsm_kernel_LTTS, ztrsm_kernel_LRTS, ztrsm_kernel_LCTS,
  275. ztrsm_kernel_RNTS, ztrsm_kernel_RTTS, ztrsm_kernel_RRTS, ztrsm_kernel_RCTS,
  276. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  277. ztrsm_iunucopyTS, ztrsm_iunncopyTS, ztrsm_iutucopyTS, ztrsm_iutncopyTS,
  278. ztrsm_ilnucopyTS, ztrsm_ilnncopyTS, ztrsm_iltucopyTS, ztrsm_iltncopyTS,
  279. #else
  280. ztrsm_ounucopyTS, ztrsm_ounncopyTS, ztrsm_outucopyTS, ztrsm_outncopyTS,
  281. ztrsm_olnucopyTS, ztrsm_olnncopyTS, ztrsm_oltucopyTS, ztrsm_oltncopyTS,
  282. #endif
  283. ztrsm_ounucopyTS, ztrsm_ounncopyTS, ztrsm_outucopyTS, ztrsm_outncopyTS,
  284. ztrsm_olnucopyTS, ztrsm_olnncopyTS, ztrsm_oltucopyTS, ztrsm_oltncopyTS,
  285. ztrmm_kernel_RNTS, ztrmm_kernel_RTTS, ztrmm_kernel_RRTS, ztrmm_kernel_RCTS,
  286. ztrmm_kernel_LNTS, ztrmm_kernel_LTTS, ztrmm_kernel_LRTS, ztrmm_kernel_LCTS,
  287. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  288. ztrmm_iunucopyTS, ztrmm_iunncopyTS, ztrmm_iutucopyTS, ztrmm_iutncopyTS,
  289. ztrmm_ilnucopyTS, ztrmm_ilnncopyTS, ztrmm_iltucopyTS, ztrmm_iltncopyTS,
  290. #else
  291. ztrmm_ounucopyTS, ztrmm_ounncopyTS, ztrmm_outucopyTS, ztrmm_outncopyTS,
  292. ztrmm_olnucopyTS, ztrmm_olnncopyTS, ztrmm_oltucopyTS, ztrmm_oltncopyTS,
  293. #endif
  294. ztrmm_ounucopyTS, ztrmm_ounncopyTS, ztrmm_outucopyTS, ztrmm_outncopyTS,
  295. ztrmm_olnucopyTS, ztrmm_olnncopyTS, ztrmm_oltucopyTS, ztrmm_oltncopyTS,
  296. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  297. zsymm_iutcopyTS, zsymm_iltcopyTS,
  298. #else
  299. zsymm_outcopyTS, zsymm_oltcopyTS,
  300. #endif
  301. zsymm_outcopyTS, zsymm_oltcopyTS,
  302. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  303. zhemm_iutcopyTS, zhemm_iltcopyTS,
  304. #else
  305. zhemm_outcopyTS, zhemm_oltcopyTS,
  306. #endif
  307. zhemm_outcopyTS, zhemm_oltcopyTS,
  308. zgemm3m_kernelTS,
  309. zgemm3m_incopybTS, zgemm3m_incopyrTS,
  310. zgemm3m_incopyiTS, zgemm3m_itcopybTS,
  311. zgemm3m_itcopyrTS, zgemm3m_itcopyiTS,
  312. zgemm3m_oncopybTS, zgemm3m_oncopyrTS,
  313. zgemm3m_oncopyiTS, zgemm3m_otcopybTS,
  314. zgemm3m_otcopyrTS, zgemm3m_otcopyiTS,
  315. zsymm3m_iucopybTS, zsymm3m_ilcopybTS,
  316. zsymm3m_iucopyrTS, zsymm3m_ilcopyrTS,
  317. zsymm3m_iucopyiTS, zsymm3m_ilcopyiTS,
  318. zsymm3m_oucopybTS, zsymm3m_olcopybTS,
  319. zsymm3m_oucopyrTS, zsymm3m_olcopyrTS,
  320. zsymm3m_oucopyiTS, zsymm3m_olcopyiTS,
  321. zhemm3m_iucopybTS, zhemm3m_ilcopybTS,
  322. zhemm3m_iucopyrTS, zhemm3m_ilcopyrTS,
  323. zhemm3m_iucopyiTS, zhemm3m_ilcopyiTS,
  324. zhemm3m_oucopybTS, zhemm3m_olcopybTS,
  325. zhemm3m_oucopyrTS, zhemm3m_olcopyrTS,
  326. zhemm3m_oucopyiTS, zhemm3m_olcopyiTS,
  327. zneg_tcopyTS, zlaswp_ncopyTS,
  328. #ifdef EXPRECISION
  329. 0, 0, 0,
  330. XGEMM_DEFAULT_UNROLL_M, XGEMM_DEFAULT_UNROLL_N, MAX(XGEMM_DEFAULT_UNROLL_M, XGEMM_DEFAULT_UNROLL_N),
  331. xamax_kTS, xamin_kTS, ixamax_kTS, ixamin_kTS,
  332. xnrm2_kTS, xasum_kTS, xcopy_kTS,
  333. xdotu_kTS, xdotc_kTS, xqrot_kTS,
  334. xaxpy_kTS, xaxpyc_kTS, xscal_kTS, xswap_kTS,
  335. xgemv_nTS, xgemv_tTS, xgemv_rTS, xgemv_cTS,
  336. xgemv_oTS, xgemv_uTS, xgemv_sTS, xgemv_dTS,
  337. xgeru_kTS, xgerc_kTS, xgerv_kTS, xgerd_kTS,
  338. xsymv_LTS, xsymv_UTS,
  339. xhemv_LTS, xhemv_UTS, xhemv_MTS, xhemv_VTS,
  340. xgemm_kernel_nTS, xgemm_kernel_lTS, xgemm_kernel_rTS, xgemm_kernel_bTS,
  341. xgemm_betaTS,
  342. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  343. xgemm_incopyTS, xgemm_itcopyTS,
  344. #else
  345. xgemm_oncopyTS, xgemm_otcopyTS,
  346. #endif
  347. xgemm_oncopyTS, xgemm_otcopyTS,
  348. xtrsm_kernel_LNTS, xtrsm_kernel_LTTS, xtrsm_kernel_LRTS, xtrsm_kernel_LCTS,
  349. xtrsm_kernel_RNTS, xtrsm_kernel_RTTS, xtrsm_kernel_RRTS, xtrsm_kernel_RCTS,
  350. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  351. xtrsm_iunucopyTS, xtrsm_iunncopyTS, xtrsm_iutucopyTS, xtrsm_iutncopyTS,
  352. xtrsm_ilnucopyTS, xtrsm_ilnncopyTS, xtrsm_iltucopyTS, xtrsm_iltncopyTS,
  353. #else
  354. xtrsm_ounucopyTS, xtrsm_ounncopyTS, xtrsm_outucopyTS, xtrsm_outncopyTS,
  355. xtrsm_olnucopyTS, xtrsm_olnncopyTS, xtrsm_oltucopyTS, xtrsm_oltncopyTS,
  356. #endif
  357. xtrsm_ounucopyTS, xtrsm_ounncopyTS, xtrsm_outucopyTS, xtrsm_outncopyTS,
  358. xtrsm_olnucopyTS, xtrsm_olnncopyTS, xtrsm_oltucopyTS, xtrsm_oltncopyTS,
  359. xtrmm_kernel_RNTS, xtrmm_kernel_RTTS, xtrmm_kernel_RRTS, xtrmm_kernel_RCTS,
  360. xtrmm_kernel_LNTS, xtrmm_kernel_LTTS, xtrmm_kernel_LRTS, xtrmm_kernel_LCTS,
  361. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  362. xtrmm_iunucopyTS, xtrmm_iunncopyTS, xtrmm_iutucopyTS, xtrmm_iutncopyTS,
  363. xtrmm_ilnucopyTS, xtrmm_ilnncopyTS, xtrmm_iltucopyTS, xtrmm_iltncopyTS,
  364. #else
  365. xtrmm_ounucopyTS, xtrmm_ounncopyTS, xtrmm_outucopyTS, xtrmm_outncopyTS,
  366. xtrmm_olnucopyTS, xtrmm_olnncopyTS, xtrmm_oltucopyTS, xtrmm_oltncopyTS,
  367. #endif
  368. xtrmm_ounucopyTS, xtrmm_ounncopyTS, xtrmm_outucopyTS, xtrmm_outncopyTS,
  369. xtrmm_olnucopyTS, xtrmm_olnncopyTS, xtrmm_oltucopyTS, xtrmm_oltncopyTS,
  370. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  371. xsymm_iutcopyTS, xsymm_iltcopyTS,
  372. #else
  373. xsymm_outcopyTS, xsymm_oltcopyTS,
  374. #endif
  375. xsymm_outcopyTS, xsymm_oltcopyTS,
  376. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  377. xhemm_iutcopyTS, xhemm_iltcopyTS,
  378. #else
  379. xhemm_outcopyTS, xhemm_oltcopyTS,
  380. #endif
  381. xhemm_outcopyTS, xhemm_oltcopyTS,
  382. xgemm3m_kernelTS,
  383. xgemm3m_incopybTS, xgemm3m_incopyrTS,
  384. xgemm3m_incopyiTS, xgemm3m_itcopybTS,
  385. xgemm3m_itcopyrTS, xgemm3m_itcopyiTS,
  386. xgemm3m_oncopybTS, xgemm3m_oncopyrTS,
  387. xgemm3m_oncopyiTS, xgemm3m_otcopybTS,
  388. xgemm3m_otcopyrTS, xgemm3m_otcopyiTS,
  389. xsymm3m_iucopybTS, xsymm3m_ilcopybTS,
  390. xsymm3m_iucopyrTS, xsymm3m_ilcopyrTS,
  391. xsymm3m_iucopyiTS, xsymm3m_ilcopyiTS,
  392. xsymm3m_oucopybTS, xsymm3m_olcopybTS,
  393. xsymm3m_oucopyrTS, xsymm3m_olcopyrTS,
  394. xsymm3m_oucopyiTS, xsymm3m_olcopyiTS,
  395. xhemm3m_iucopybTS, xhemm3m_ilcopybTS,
  396. xhemm3m_iucopyrTS, xhemm3m_ilcopyrTS,
  397. xhemm3m_iucopyiTS, xhemm3m_ilcopyiTS,
  398. xhemm3m_oucopybTS, xhemm3m_olcopybTS,
  399. xhemm3m_oucopyrTS, xhemm3m_olcopyrTS,
  400. xhemm3m_oucopyiTS, xhemm3m_olcopyiTS,
  401. xneg_tcopyTS, xlaswp_ncopyTS,
  402. #endif
  403. init_parameter,
  404. SNUMOPT, DNUMOPT, QNUMOPT,
  405. };
  406. #ifdef ARCH_X86
  407. static int get_l2_size_old(void){
  408. int i, eax, ebx, ecx, edx, cpuid_level;
  409. int info[15];
  410. cpuid(2, &eax, &ebx, &ecx, &edx);
  411. info[ 0] = BITMASK(eax, 8, 0xff);
  412. info[ 1] = BITMASK(eax, 16, 0xff);
  413. info[ 2] = BITMASK(eax, 24, 0xff);
  414. info[ 3] = BITMASK(ebx, 0, 0xff);
  415. info[ 4] = BITMASK(ebx, 8, 0xff);
  416. info[ 5] = BITMASK(ebx, 16, 0xff);
  417. info[ 6] = BITMASK(ebx, 24, 0xff);
  418. info[ 7] = BITMASK(ecx, 0, 0xff);
  419. info[ 8] = BITMASK(ecx, 8, 0xff);
  420. info[ 9] = BITMASK(ecx, 16, 0xff);
  421. info[10] = BITMASK(ecx, 24, 0xff);
  422. info[11] = BITMASK(edx, 0, 0xff);
  423. info[12] = BITMASK(edx, 8, 0xff);
  424. info[13] = BITMASK(edx, 16, 0xff);
  425. info[14] = BITMASK(edx, 24, 0xff);
  426. for (i = 0; i < 15; i++){
  427. switch (info[i]){
  428. /* This table is from http://www.sandpile.org/ia32/cpuid.htm */
  429. case 0x1a :
  430. return 96;
  431. case 0x39 :
  432. case 0x3b :
  433. case 0x41 :
  434. case 0x79 :
  435. case 0x81 :
  436. return 128;
  437. case 0x3a :
  438. return 192;
  439. case 0x21 :
  440. case 0x3c :
  441. case 0x42 :
  442. case 0x7a :
  443. case 0x7e :
  444. case 0x82 :
  445. return 256;
  446. case 0x3d :
  447. return 384;
  448. case 0x3e :
  449. case 0x43 :
  450. case 0x7b :
  451. case 0x7f :
  452. case 0x83 :
  453. case 0x86 :
  454. return 512;
  455. case 0x44 :
  456. case 0x78 :
  457. case 0x7c :
  458. case 0x84 :
  459. case 0x87 :
  460. return 1024;
  461. case 0x45 :
  462. case 0x7d :
  463. case 0x85 :
  464. return 2048;
  465. case 0x48 :
  466. return 3184;
  467. case 0x49 :
  468. return 4096;
  469. case 0x4e :
  470. return 6144;
  471. }
  472. }
  473. return 0;
  474. }
  475. #endif
  476. static __inline__ int get_l2_size(void){
  477. int eax, ebx, ecx, edx, l2;
  478. cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
  479. l2 = BITMASK(ecx, 16, 0xffff);
  480. #ifndef ARCH_X86
  481. return l2;
  482. #else
  483. if (l2 > 0) return l2;
  484. return get_l2_size_old();
  485. #endif
  486. }
  487. static __inline__ int get_l3_size(void){
  488. int eax, ebx, ecx, edx;
  489. cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
  490. return BITMASK(edx, 18, 0x3fff) * 512;
  491. }
  492. static void init_parameter(void) {
  493. int l2 = get_l2_size();
  494. TABLE_NAME.sgemm_q = SGEMM_DEFAULT_Q;
  495. TABLE_NAME.dgemm_q = DGEMM_DEFAULT_Q;
  496. TABLE_NAME.cgemm_q = CGEMM_DEFAULT_Q;
  497. TABLE_NAME.zgemm_q = ZGEMM_DEFAULT_Q;
  498. #ifdef EXPRECISION
  499. TABLE_NAME.qgemm_q = QGEMM_DEFAULT_Q;
  500. TABLE_NAME.xgemm_q = XGEMM_DEFAULT_Q;
  501. #endif
  502. #if defined(CORE_KATMAI) || defined(CORE_COPPERMINE) || defined(CORE_BANIAS) || defined(CORE_YONAH)
  503. #ifdef DEBUG
  504. fprintf(stderr, "Katmai, Coppermine, Banias\n");
  505. #endif
  506. TABLE_NAME.sgemm_p = 64 * (l2 >> 7);
  507. TABLE_NAME.dgemm_p = 32 * (l2 >> 7);
  508. TABLE_NAME.cgemm_p = 32 * (l2 >> 7);
  509. TABLE_NAME.zgemm_p = 16 * (l2 >> 7);
  510. #ifdef EXPRECISION
  511. TABLE_NAME.qgemm_p = 16 * (l2 >> 7);
  512. TABLE_NAME.xgemm_p = 8 * (l2 >> 7);
  513. #endif
  514. #endif
  515. #ifdef CORE_NORTHWOOD
  516. #ifdef DEBUG
  517. fprintf(stderr, "Northwood\n");
  518. #endif
  519. TABLE_NAME.sgemm_p = 96 * (l2 >> 7);
  520. TABLE_NAME.dgemm_p = 48 * (l2 >> 7);
  521. TABLE_NAME.cgemm_p = 48 * (l2 >> 7);
  522. TABLE_NAME.zgemm_p = 24 * (l2 >> 7);
  523. #ifdef EXPRECISION
  524. TABLE_NAME.qgemm_p = 24 * (l2 >> 7);
  525. TABLE_NAME.xgemm_p = 12 * (l2 >> 7);
  526. #endif
  527. #endif
  528. #ifdef ATOM
  529. #ifdef DEBUG
  530. fprintf(stderr, "Atom\n");
  531. #endif
  532. TABLE_NAME.sgemm_p = 256;
  533. TABLE_NAME.dgemm_p = 128;
  534. TABLE_NAME.cgemm_p = 128;
  535. TABLE_NAME.zgemm_p = 64;
  536. #ifdef EXPRECISION
  537. TABLE_NAME.qgemm_p = 64;
  538. TABLE_NAME.xgemm_p = 32;
  539. #endif
  540. #endif
  541. #ifdef CORE_PRESCOTT
  542. #ifdef DEBUG
  543. fprintf(stderr, "Prescott\n");
  544. #endif
  545. TABLE_NAME.sgemm_p = 56 * (l2 >> 7);
  546. TABLE_NAME.dgemm_p = 28 * (l2 >> 7);
  547. TABLE_NAME.cgemm_p = 28 * (l2 >> 7);
  548. TABLE_NAME.zgemm_p = 14 * (l2 >> 7);
  549. #ifdef EXPRECISION
  550. TABLE_NAME.qgemm_p = 14 * (l2 >> 7);
  551. TABLE_NAME.xgemm_p = 7 * (l2 >> 7);
  552. #endif
  553. #endif
  554. #ifdef CORE2
  555. #ifdef DEBUG
  556. fprintf(stderr, "Core2\n");
  557. #endif
  558. TABLE_NAME.sgemm_p = 92 * (l2 >> 9);
  559. TABLE_NAME.dgemm_p = 46 * (l2 >> 9);
  560. TABLE_NAME.cgemm_p = 46 * (l2 >> 9);
  561. TABLE_NAME.zgemm_p = 23 * (l2 >> 9);
  562. #ifdef EXPRECISION
  563. TABLE_NAME.qgemm_p = 92 * (l2 >> 9);
  564. TABLE_NAME.xgemm_p = 46 * (l2 >> 9);
  565. #endif
  566. #endif
  567. #ifdef PENRYN
  568. #ifdef DEBUG
  569. fprintf(stderr, "Penryn\n");
  570. #endif
  571. TABLE_NAME.sgemm_p = 42 * (l2 >> 9) + 8;
  572. TABLE_NAME.dgemm_p = 42 * (l2 >> 9) + 8;
  573. TABLE_NAME.cgemm_p = 21 * (l2 >> 9) + 4;
  574. TABLE_NAME.zgemm_p = 21 * (l2 >> 9) + 4;
  575. #ifdef EXPRECISION
  576. TABLE_NAME.qgemm_p = 42 * (l2 >> 9) + 8;
  577. TABLE_NAME.xgemm_p = 21 * (l2 >> 9) + 4;
  578. #endif
  579. #endif
  580. #ifdef NEHALEM
  581. #ifdef DEBUG
  582. fprintf(stderr, "Nehalem\n");
  583. #endif
  584. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  585. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  586. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  587. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  588. #ifdef EXPRECISION
  589. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  590. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  591. #endif
  592. #endif
  593. #ifdef OPTERON
  594. #ifdef DEBUG
  595. fprintf(stderr, "Opteron\n");
  596. #endif
  597. TABLE_NAME.sgemm_p = 224 + 56 * (l2 >> 7);
  598. TABLE_NAME.dgemm_p = 112 + 28 * (l2 >> 7);
  599. TABLE_NAME.cgemm_p = 112 + 28 * (l2 >> 7);
  600. TABLE_NAME.zgemm_p = 56 + 14 * (l2 >> 7);
  601. #ifdef EXPRECISION
  602. TABLE_NAME.qgemm_p = 56 + 14 * (l2 >> 7);
  603. TABLE_NAME.xgemm_p = 28 + 7 * (l2 >> 7);
  604. #endif
  605. #endif
  606. #ifdef BARCELONA
  607. #ifdef DEBUG
  608. fprintf(stderr, "Barcelona\n");
  609. #endif
  610. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  611. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  612. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  613. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  614. #ifdef EXPRECISION
  615. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  616. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  617. #endif
  618. #endif
  619. #ifdef NANO
  620. #ifdef DEBUG
  621. fprintf(stderr, "NANO\n");
  622. #endif
  623. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  624. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  625. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  626. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  627. #ifdef EXPRECISION
  628. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  629. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  630. #endif
  631. #endif
  632. TABLE_NAME.sgemm_p = (TABLE_NAME.sgemm_p + SGEMM_DEFAULT_UNROLL_M - 1) & ~(SGEMM_DEFAULT_UNROLL_M - 1);
  633. TABLE_NAME.dgemm_p = (TABLE_NAME.dgemm_p + DGEMM_DEFAULT_UNROLL_M - 1) & ~(DGEMM_DEFAULT_UNROLL_M - 1);
  634. TABLE_NAME.cgemm_p = (TABLE_NAME.cgemm_p + CGEMM_DEFAULT_UNROLL_M - 1) & ~(CGEMM_DEFAULT_UNROLL_M - 1);
  635. TABLE_NAME.zgemm_p = (TABLE_NAME.zgemm_p + ZGEMM_DEFAULT_UNROLL_M - 1) & ~(ZGEMM_DEFAULT_UNROLL_M - 1);
  636. #ifdef QUAD_PRECISION
  637. TABLE_NAME.qgemm_p = (TABLE_NAME.qgemm_p + QGEMM_DEFAULT_UNROLL_M - 1) & ~(QGEMM_DEFAULT_UNROLL_M - 1);
  638. TABLE_NAME.xgemm_p = (TABLE_NAME.xgemm_p + XGEMM_DEFAULT_UNROLL_M - 1) & ~(XGEMM_DEFAULT_UNROLL_M - 1);
  639. #endif
  640. #ifdef DEBUG
  641. fprintf(stderr, "L2 = %8d DGEMM_P .. %d\n", l2, TABLE_NAME.dgemm_p);
  642. #endif
  643. TABLE_NAME.sgemm_r = (((BUFFER_SIZE -
  644. ((TABLE_NAME.sgemm_p * TABLE_NAME.sgemm_q * 4 + TABLE_NAME.offsetA
  645. + TABLE_NAME.align) & ~TABLE_NAME.align)
  646. ) / (TABLE_NAME.sgemm_q * 4) - 15) & ~15);
  647. TABLE_NAME.dgemm_r = (((BUFFER_SIZE -
  648. ((TABLE_NAME.dgemm_p * TABLE_NAME.dgemm_q * 8 + TABLE_NAME.offsetA
  649. + TABLE_NAME.align) & ~TABLE_NAME.align)
  650. ) / (TABLE_NAME.dgemm_q * 8) - 15) & ~15);
  651. #ifdef EXPRECISION
  652. TABLE_NAME.qgemm_r = (((BUFFER_SIZE -
  653. ((TABLE_NAME.qgemm_p * TABLE_NAME.qgemm_q * 16 + TABLE_NAME.offsetA
  654. + TABLE_NAME.align) & ~TABLE_NAME.align)
  655. ) / (TABLE_NAME.qgemm_q * 16) - 15) & ~15);
  656. #endif
  657. TABLE_NAME.cgemm_r = (((BUFFER_SIZE -
  658. ((TABLE_NAME.cgemm_p * TABLE_NAME.cgemm_q * 8 + TABLE_NAME.offsetA
  659. + TABLE_NAME.align) & ~TABLE_NAME.align)
  660. ) / (TABLE_NAME.cgemm_q * 8) - 15) & ~15);
  661. TABLE_NAME.zgemm_r = (((BUFFER_SIZE -
  662. ((TABLE_NAME.zgemm_p * TABLE_NAME.zgemm_q * 16 + TABLE_NAME.offsetA
  663. + TABLE_NAME.align) & ~TABLE_NAME.align)
  664. ) / (TABLE_NAME.zgemm_q * 16) - 15) & ~15);
  665. #ifdef EXPRECISION
  666. TABLE_NAME.xgemm_r = (((BUFFER_SIZE -
  667. ((TABLE_NAME.xgemm_p * TABLE_NAME.xgemm_q * 32 + TABLE_NAME.offsetA
  668. + TABLE_NAME.align) & ~TABLE_NAME.align)
  669. ) / (TABLE_NAME.xgemm_q * 32) - 15) & ~15);
  670. #endif
  671. }

OpenBLAS is an optimized BLAS library based on GotoBLAS2 1.13 BSD version.

Contributors (1)