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 38 kB

6 years ago
6 years ago
6 years ago
6 years ago
6 years ago
6 years ago
11 years ago
11 years ago
1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306
  1. /*********************************************************************/
  2. /* Copyright 2009, 2010 The University of Texas at Austin. */
  3. /* All rights reserved. */
  4. /* */
  5. /* Redistribution and use in source and binary forms, with or */
  6. /* without modification, are permitted provided that the following */
  7. /* conditions are met: */
  8. /* */
  9. /* 1. Redistributions of source code must retain the above */
  10. /* copyright notice, this list of conditions and the following */
  11. /* disclaimer. */
  12. /* */
  13. /* 2. Redistributions in binary form must reproduce the above */
  14. /* copyright notice, this list of conditions and the following */
  15. /* disclaimer in the documentation and/or other materials */
  16. /* provided with the distribution. */
  17. /* */
  18. /* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
  19. /* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
  20. /* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
  21. /* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
  22. /* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
  23. /* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
  24. /* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
  25. /* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
  26. /* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
  27. /* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
  28. /* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
  29. /* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
  30. /* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
  31. /* POSSIBILITY OF SUCH DAMAGE. */
  32. /* */
  33. /* The views and conclusions contained in the software and */
  34. /* documentation are those of the authors and should not be */
  35. /* interpreted as representing official policies, either expressed */
  36. /* or implied, of The University of Texas at Austin. */
  37. /*********************************************************************/
  38. #include <stdio.h>
  39. #include <string.h>
  40. #include "common.h"
  41. #ifdef BUILD_KERNEL
  42. #include "kernelTS.h"
  43. #endif
  44. #undef DEBUG
  45. static void init_parameter(void);
  46. gotoblas_t TABLE_NAME = {
  47. DTB_DEFAULT_ENTRIES ,
  48. GEMM_DEFAULT_OFFSET_A, GEMM_DEFAULT_OFFSET_B, GEMM_DEFAULT_ALIGN,
  49. 0, 0, 0,
  50. SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N,
  51. #ifdef SGEMM_DEFAULT_UNROLL_MN
  52. SGEMM_DEFAULT_UNROLL_MN,
  53. #else
  54. MAX(SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N),
  55. #endif
  56. #ifdef HAVE_EXCLUSIVE_CACHE
  57. 1,
  58. #else
  59. 0,
  60. #endif
  61. samax_kTS, samin_kTS, smax_kTS, smin_kTS,
  62. isamax_kTS, isamin_kTS, ismax_kTS, ismin_kTS,
  63. snrm2_kTS, sasum_kTS, ssum_kTS, scopy_kTS, sdot_kTS,
  64. dsdot_kTS,
  65. srot_kTS, saxpy_kTS, sscal_kTS, sswap_kTS,
  66. sgemv_nTS, sgemv_tTS, sger_kTS,
  67. ssymv_LTS, ssymv_UTS,
  68. sgemm_kernelTS, sgemm_betaTS,
  69. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  70. sgemm_incopyTS, sgemm_itcopyTS,
  71. #else
  72. sgemm_oncopyTS, sgemm_otcopyTS,
  73. #endif
  74. sgemm_oncopyTS, sgemm_otcopyTS,
  75. strsm_kernel_LNTS, strsm_kernel_LTTS, strsm_kernel_RNTS, strsm_kernel_RTTS,
  76. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  77. strsm_iunucopyTS, strsm_iunncopyTS, strsm_iutucopyTS, strsm_iutncopyTS,
  78. strsm_ilnucopyTS, strsm_ilnncopyTS, strsm_iltucopyTS, strsm_iltncopyTS,
  79. #else
  80. strsm_ounucopyTS, strsm_ounncopyTS, strsm_outucopyTS, strsm_outncopyTS,
  81. strsm_olnucopyTS, strsm_olnncopyTS, strsm_oltucopyTS, strsm_oltncopyTS,
  82. #endif
  83. strsm_ounucopyTS, strsm_ounncopyTS, strsm_outucopyTS, strsm_outncopyTS,
  84. strsm_olnucopyTS, strsm_olnncopyTS, strsm_oltucopyTS, strsm_oltncopyTS,
  85. strmm_kernel_RNTS, strmm_kernel_RTTS, strmm_kernel_LNTS, strmm_kernel_LTTS,
  86. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  87. strmm_iunucopyTS, strmm_iunncopyTS, strmm_iutucopyTS, strmm_iutncopyTS,
  88. strmm_ilnucopyTS, strmm_ilnncopyTS, strmm_iltucopyTS, strmm_iltncopyTS,
  89. #else
  90. strmm_ounucopyTS, strmm_ounncopyTS, strmm_outucopyTS, strmm_outncopyTS,
  91. strmm_olnucopyTS, strmm_olnncopyTS, strmm_oltucopyTS, strmm_oltncopyTS,
  92. #endif
  93. strmm_ounucopyTS, strmm_ounncopyTS, strmm_outucopyTS, strmm_outncopyTS,
  94. strmm_olnucopyTS, strmm_olnncopyTS, strmm_oltucopyTS, strmm_oltncopyTS,
  95. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  96. ssymm_iutcopyTS, ssymm_iltcopyTS,
  97. #else
  98. ssymm_outcopyTS, ssymm_oltcopyTS,
  99. #endif
  100. ssymm_outcopyTS, ssymm_oltcopyTS,
  101. #ifndef NO_LAPACK
  102. sneg_tcopyTS, slaswp_ncopyTS,
  103. #else
  104. NULL,NULL,
  105. #endif
  106. 0, 0, 0,
  107. DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N,
  108. #ifdef DGEMM_DEFAULT_UNROLL_MN
  109. DGEMM_DEFAULT_UNROLL_MN,
  110. #else
  111. MAX(DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N),
  112. #endif
  113. damax_kTS, damin_kTS, dmax_kTS, dmin_kTS,
  114. idamax_kTS, idamin_kTS, idmax_kTS, idmin_kTS,
  115. dnrm2_kTS, dasum_kTS, dsum_kTS, dcopy_kTS, ddot_kTS,
  116. drot_kTS, daxpy_kTS, dscal_kTS, dswap_kTS,
  117. dgemv_nTS, dgemv_tTS, dger_kTS,
  118. dsymv_LTS, dsymv_UTS,
  119. dgemm_kernelTS, dgemm_betaTS,
  120. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  121. dgemm_incopyTS, dgemm_itcopyTS,
  122. #else
  123. dgemm_oncopyTS, dgemm_otcopyTS,
  124. #endif
  125. dgemm_oncopyTS, dgemm_otcopyTS,
  126. dtrsm_kernel_LNTS, dtrsm_kernel_LTTS, dtrsm_kernel_RNTS, dtrsm_kernel_RTTS,
  127. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  128. dtrsm_iunucopyTS, dtrsm_iunncopyTS, dtrsm_iutucopyTS, dtrsm_iutncopyTS,
  129. dtrsm_ilnucopyTS, dtrsm_ilnncopyTS, dtrsm_iltucopyTS, dtrsm_iltncopyTS,
  130. #else
  131. dtrsm_ounucopyTS, dtrsm_ounncopyTS, dtrsm_outucopyTS, dtrsm_outncopyTS,
  132. dtrsm_olnucopyTS, dtrsm_olnncopyTS, dtrsm_oltucopyTS, dtrsm_oltncopyTS,
  133. #endif
  134. dtrsm_ounucopyTS, dtrsm_ounncopyTS, dtrsm_outucopyTS, dtrsm_outncopyTS,
  135. dtrsm_olnucopyTS, dtrsm_olnncopyTS, dtrsm_oltucopyTS, dtrsm_oltncopyTS,
  136. dtrmm_kernel_RNTS, dtrmm_kernel_RTTS, dtrmm_kernel_LNTS, dtrmm_kernel_LTTS,
  137. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  138. dtrmm_iunucopyTS, dtrmm_iunncopyTS, dtrmm_iutucopyTS, dtrmm_iutncopyTS,
  139. dtrmm_ilnucopyTS, dtrmm_ilnncopyTS, dtrmm_iltucopyTS, dtrmm_iltncopyTS,
  140. #else
  141. dtrmm_ounucopyTS, dtrmm_ounncopyTS, dtrmm_outucopyTS, dtrmm_outncopyTS,
  142. dtrmm_olnucopyTS, dtrmm_olnncopyTS, dtrmm_oltucopyTS, dtrmm_oltncopyTS,
  143. #endif
  144. dtrmm_ounucopyTS, dtrmm_ounncopyTS, dtrmm_outucopyTS, dtrmm_outncopyTS,
  145. dtrmm_olnucopyTS, dtrmm_olnncopyTS, dtrmm_oltucopyTS, dtrmm_oltncopyTS,
  146. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  147. dsymm_iutcopyTS, dsymm_iltcopyTS,
  148. #else
  149. dsymm_outcopyTS, dsymm_oltcopyTS,
  150. #endif
  151. dsymm_outcopyTS, dsymm_oltcopyTS,
  152. #ifndef NO_LAPACK
  153. dneg_tcopyTS, dlaswp_ncopyTS,
  154. #else
  155. NULL, NULL,
  156. #endif
  157. #ifdef EXPRECISION
  158. 0, 0, 0,
  159. QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N, MAX(QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N),
  160. qamax_kTS, qamin_kTS, qmax_kTS, qmin_kTS,
  161. iqamax_kTS, iqamin_kTS, iqmax_kTS, iqmin_kTS,
  162. qnrm2_kTS, qasum_kTS, qsum_kTS, qcopy_kTS, qdot_kTS,
  163. qrot_kTS, qaxpy_kTS, qscal_kTS, qswap_kTS,
  164. qgemv_nTS, qgemv_tTS, qger_kTS,
  165. qsymv_LTS, qsymv_UTS,
  166. qgemm_kernelTS, qgemm_betaTS,
  167. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  168. qgemm_incopyTS, qgemm_itcopyTS,
  169. #else
  170. qgemm_oncopyTS, qgemm_otcopyTS,
  171. #endif
  172. qgemm_oncopyTS, qgemm_otcopyTS,
  173. qtrsm_kernel_LNTS, qtrsm_kernel_LTTS, qtrsm_kernel_RNTS, qtrsm_kernel_RTTS,
  174. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  175. qtrsm_iunucopyTS, qtrsm_iunncopyTS, qtrsm_iutucopyTS, qtrsm_iutncopyTS,
  176. qtrsm_ilnucopyTS, qtrsm_ilnncopyTS, qtrsm_iltucopyTS, qtrsm_iltncopyTS,
  177. #else
  178. qtrsm_ounucopyTS, qtrsm_ounncopyTS, qtrsm_outucopyTS, qtrsm_outncopyTS,
  179. qtrsm_olnucopyTS, qtrsm_olnncopyTS, qtrsm_oltucopyTS, qtrsm_oltncopyTS,
  180. #endif
  181. qtrsm_ounucopyTS, qtrsm_ounncopyTS, qtrsm_outucopyTS, qtrsm_outncopyTS,
  182. qtrsm_olnucopyTS, qtrsm_olnncopyTS, qtrsm_oltucopyTS, qtrsm_oltncopyTS,
  183. qtrmm_kernel_RNTS, qtrmm_kernel_RTTS, qtrmm_kernel_LNTS, qtrmm_kernel_LTTS,
  184. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  185. qtrmm_iunucopyTS, qtrmm_iunncopyTS, qtrmm_iutucopyTS, qtrmm_iutncopyTS,
  186. qtrmm_ilnucopyTS, qtrmm_ilnncopyTS, qtrmm_iltucopyTS, qtrmm_iltncopyTS,
  187. #else
  188. qtrmm_ounucopyTS, qtrmm_ounncopyTS, qtrmm_outucopyTS, qtrmm_outncopyTS,
  189. qtrmm_olnucopyTS, qtrmm_olnncopyTS, qtrmm_oltucopyTS, qtrmm_oltncopyTS,
  190. #endif
  191. qtrmm_ounucopyTS, qtrmm_ounncopyTS, qtrmm_outucopyTS, qtrmm_outncopyTS,
  192. qtrmm_olnucopyTS, qtrmm_olnncopyTS, qtrmm_oltucopyTS, qtrmm_oltncopyTS,
  193. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  194. qsymm_iutcopyTS, qsymm_iltcopyTS,
  195. #else
  196. qsymm_outcopyTS, qsymm_oltcopyTS,
  197. #endif
  198. qsymm_outcopyTS, qsymm_oltcopyTS,
  199. #ifndef NO_LAPACK
  200. qneg_tcopyTS, qlaswp_ncopyTS,
  201. #else
  202. NULL, NULL,
  203. #endif
  204. #endif
  205. 0, 0, 0,
  206. CGEMM_DEFAULT_UNROLL_M, CGEMM_DEFAULT_UNROLL_N,
  207. #ifdef CGEMM_DEFAULT_UNROLL_MN
  208. CGEMM_DEFAULT_UNROLL_MN,
  209. #else
  210. MAX(CGEMM_DEFAULT_UNROLL_M, CGEMM_DEFAULT_UNROLL_N),
  211. #endif
  212. camax_kTS, camin_kTS, icamax_kTS, icamin_kTS,
  213. cnrm2_kTS, casum_kTS, csum_kTS, ccopy_kTS,
  214. cdotu_kTS, cdotc_kTS, csrot_kTS,
  215. caxpy_kTS, caxpyc_kTS, cscal_kTS, cswap_kTS,
  216. cgemv_nTS, cgemv_tTS, cgemv_rTS, cgemv_cTS,
  217. cgemv_oTS, cgemv_uTS, cgemv_sTS, cgemv_dTS,
  218. cgeru_kTS, cgerc_kTS, cgerv_kTS, cgerd_kTS,
  219. csymv_LTS, csymv_UTS,
  220. chemv_LTS, chemv_UTS, chemv_MTS, chemv_VTS,
  221. cgemm_kernel_nTS, cgemm_kernel_lTS, cgemm_kernel_rTS, cgemm_kernel_bTS,
  222. cgemm_betaTS,
  223. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  224. cgemm_incopyTS, cgemm_itcopyTS,
  225. #else
  226. cgemm_oncopyTS, cgemm_otcopyTS,
  227. #endif
  228. cgemm_oncopyTS, cgemm_otcopyTS,
  229. ctrsm_kernel_LNTS, ctrsm_kernel_LTTS, ctrsm_kernel_LRTS, ctrsm_kernel_LCTS,
  230. ctrsm_kernel_RNTS, ctrsm_kernel_RTTS, ctrsm_kernel_RRTS, ctrsm_kernel_RCTS,
  231. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  232. ctrsm_iunucopyTS, ctrsm_iunncopyTS, ctrsm_iutucopyTS, ctrsm_iutncopyTS,
  233. ctrsm_ilnucopyTS, ctrsm_ilnncopyTS, ctrsm_iltucopyTS, ctrsm_iltncopyTS,
  234. #else
  235. ctrsm_ounucopyTS, ctrsm_ounncopyTS, ctrsm_outucopyTS, ctrsm_outncopyTS,
  236. ctrsm_olnucopyTS, ctrsm_olnncopyTS, ctrsm_oltucopyTS, ctrsm_oltncopyTS,
  237. #endif
  238. ctrsm_ounucopyTS, ctrsm_ounncopyTS, ctrsm_outucopyTS, ctrsm_outncopyTS,
  239. ctrsm_olnucopyTS, ctrsm_olnncopyTS, ctrsm_oltucopyTS, ctrsm_oltncopyTS,
  240. ctrmm_kernel_RNTS, ctrmm_kernel_RTTS, ctrmm_kernel_RRTS, ctrmm_kernel_RCTS,
  241. ctrmm_kernel_LNTS, ctrmm_kernel_LTTS, ctrmm_kernel_LRTS, ctrmm_kernel_LCTS,
  242. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  243. ctrmm_iunucopyTS, ctrmm_iunncopyTS, ctrmm_iutucopyTS, ctrmm_iutncopyTS,
  244. ctrmm_ilnucopyTS, ctrmm_ilnncopyTS, ctrmm_iltucopyTS, ctrmm_iltncopyTS,
  245. #else
  246. ctrmm_ounucopyTS, ctrmm_ounncopyTS, ctrmm_outucopyTS, ctrmm_outncopyTS,
  247. ctrmm_olnucopyTS, ctrmm_olnncopyTS, ctrmm_oltucopyTS, ctrmm_oltncopyTS,
  248. #endif
  249. ctrmm_ounucopyTS, ctrmm_ounncopyTS, ctrmm_outucopyTS, ctrmm_outncopyTS,
  250. ctrmm_olnucopyTS, ctrmm_olnncopyTS, ctrmm_oltucopyTS, ctrmm_oltncopyTS,
  251. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  252. csymm_iutcopyTS, csymm_iltcopyTS,
  253. #else
  254. csymm_outcopyTS, csymm_oltcopyTS,
  255. #endif
  256. csymm_outcopyTS, csymm_oltcopyTS,
  257. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  258. chemm_iutcopyTS, chemm_iltcopyTS,
  259. #else
  260. chemm_outcopyTS, chemm_oltcopyTS,
  261. #endif
  262. chemm_outcopyTS, chemm_oltcopyTS,
  263. 0, 0, 0,
  264. #if defined(USE_GEMM3M)
  265. #ifdef CGEMM3M_DEFAULT_UNROLL_M
  266. CGEMM3M_DEFAULT_UNROLL_M, CGEMM3M_DEFAULT_UNROLL_N, MAX(CGEMM3M_DEFAULT_UNROLL_M, CGEMM3M_DEFAULT_UNROLL_N),
  267. #else
  268. SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N, MAX(SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N),
  269. #endif
  270. cgemm3m_kernelTS,
  271. cgemm3m_incopybTS, cgemm3m_incopyrTS,
  272. cgemm3m_incopyiTS, cgemm3m_itcopybTS,
  273. cgemm3m_itcopyrTS, cgemm3m_itcopyiTS,
  274. cgemm3m_oncopybTS, cgemm3m_oncopyrTS,
  275. cgemm3m_oncopyiTS, cgemm3m_otcopybTS,
  276. cgemm3m_otcopyrTS, cgemm3m_otcopyiTS,
  277. csymm3m_iucopybTS, csymm3m_ilcopybTS,
  278. csymm3m_iucopyrTS, csymm3m_ilcopyrTS,
  279. csymm3m_iucopyiTS, csymm3m_ilcopyiTS,
  280. csymm3m_oucopybTS, csymm3m_olcopybTS,
  281. csymm3m_oucopyrTS, csymm3m_olcopyrTS,
  282. csymm3m_oucopyiTS, csymm3m_olcopyiTS,
  283. chemm3m_iucopybTS, chemm3m_ilcopybTS,
  284. chemm3m_iucopyrTS, chemm3m_ilcopyrTS,
  285. chemm3m_iucopyiTS, chemm3m_ilcopyiTS,
  286. chemm3m_oucopybTS, chemm3m_olcopybTS,
  287. chemm3m_oucopyrTS, chemm3m_olcopyrTS,
  288. chemm3m_oucopyiTS, chemm3m_olcopyiTS,
  289. #else
  290. 0, 0, 0,
  291. NULL,
  292. NULL, NULL,
  293. NULL, NULL,
  294. NULL, NULL,
  295. NULL, NULL,
  296. NULL, NULL,
  297. NULL, NULL,
  298. NULL, NULL,
  299. NULL, NULL,
  300. NULL, NULL,
  301. NULL, NULL,
  302. NULL, NULL,
  303. NULL, NULL,
  304. NULL, NULL,
  305. NULL, NULL,
  306. NULL, NULL,
  307. NULL, NULL,
  308. NULL, NULL,
  309. NULL, NULL,
  310. #endif
  311. #ifndef NO_LAPACK
  312. cneg_tcopyTS, claswp_ncopyTS,
  313. #else
  314. NULL, NULL,
  315. #endif
  316. 0, 0, 0,
  317. ZGEMM_DEFAULT_UNROLL_M, ZGEMM_DEFAULT_UNROLL_N,
  318. #ifdef ZGEMM_DEFAULT_UNROLL_MN
  319. ZGEMM_DEFAULT_UNROLL_MN,
  320. #else
  321. MAX(ZGEMM_DEFAULT_UNROLL_M, ZGEMM_DEFAULT_UNROLL_N),
  322. #endif
  323. zamax_kTS, zamin_kTS, izamax_kTS, izamin_kTS,
  324. znrm2_kTS, zasum_kTS, zsum_kTS, zcopy_kTS,
  325. zdotu_kTS, zdotc_kTS, zdrot_kTS,
  326. zaxpy_kTS, zaxpyc_kTS, zscal_kTS, zswap_kTS,
  327. zgemv_nTS, zgemv_tTS, zgemv_rTS, zgemv_cTS,
  328. zgemv_oTS, zgemv_uTS, zgemv_sTS, zgemv_dTS,
  329. zgeru_kTS, zgerc_kTS, zgerv_kTS, zgerd_kTS,
  330. zsymv_LTS, zsymv_UTS,
  331. zhemv_LTS, zhemv_UTS, zhemv_MTS, zhemv_VTS,
  332. zgemm_kernel_nTS, zgemm_kernel_lTS, zgemm_kernel_rTS, zgemm_kernel_bTS,
  333. zgemm_betaTS,
  334. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  335. zgemm_incopyTS, zgemm_itcopyTS,
  336. #else
  337. zgemm_oncopyTS, zgemm_otcopyTS,
  338. #endif
  339. zgemm_oncopyTS, zgemm_otcopyTS,
  340. ztrsm_kernel_LNTS, ztrsm_kernel_LTTS, ztrsm_kernel_LRTS, ztrsm_kernel_LCTS,
  341. ztrsm_kernel_RNTS, ztrsm_kernel_RTTS, ztrsm_kernel_RRTS, ztrsm_kernel_RCTS,
  342. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  343. ztrsm_iunucopyTS, ztrsm_iunncopyTS, ztrsm_iutucopyTS, ztrsm_iutncopyTS,
  344. ztrsm_ilnucopyTS, ztrsm_ilnncopyTS, ztrsm_iltucopyTS, ztrsm_iltncopyTS,
  345. #else
  346. ztrsm_ounucopyTS, ztrsm_ounncopyTS, ztrsm_outucopyTS, ztrsm_outncopyTS,
  347. ztrsm_olnucopyTS, ztrsm_olnncopyTS, ztrsm_oltucopyTS, ztrsm_oltncopyTS,
  348. #endif
  349. ztrsm_ounucopyTS, ztrsm_ounncopyTS, ztrsm_outucopyTS, ztrsm_outncopyTS,
  350. ztrsm_olnucopyTS, ztrsm_olnncopyTS, ztrsm_oltucopyTS, ztrsm_oltncopyTS,
  351. ztrmm_kernel_RNTS, ztrmm_kernel_RTTS, ztrmm_kernel_RRTS, ztrmm_kernel_RCTS,
  352. ztrmm_kernel_LNTS, ztrmm_kernel_LTTS, ztrmm_kernel_LRTS, ztrmm_kernel_LCTS,
  353. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  354. ztrmm_iunucopyTS, ztrmm_iunncopyTS, ztrmm_iutucopyTS, ztrmm_iutncopyTS,
  355. ztrmm_ilnucopyTS, ztrmm_ilnncopyTS, ztrmm_iltucopyTS, ztrmm_iltncopyTS,
  356. #else
  357. ztrmm_ounucopyTS, ztrmm_ounncopyTS, ztrmm_outucopyTS, ztrmm_outncopyTS,
  358. ztrmm_olnucopyTS, ztrmm_olnncopyTS, ztrmm_oltucopyTS, ztrmm_oltncopyTS,
  359. #endif
  360. ztrmm_ounucopyTS, ztrmm_ounncopyTS, ztrmm_outucopyTS, ztrmm_outncopyTS,
  361. ztrmm_olnucopyTS, ztrmm_olnncopyTS, ztrmm_oltucopyTS, ztrmm_oltncopyTS,
  362. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  363. zsymm_iutcopyTS, zsymm_iltcopyTS,
  364. #else
  365. zsymm_outcopyTS, zsymm_oltcopyTS,
  366. #endif
  367. zsymm_outcopyTS, zsymm_oltcopyTS,
  368. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  369. zhemm_iutcopyTS, zhemm_iltcopyTS,
  370. #else
  371. zhemm_outcopyTS, zhemm_oltcopyTS,
  372. #endif
  373. zhemm_outcopyTS, zhemm_oltcopyTS,
  374. 0, 0, 0,
  375. #if defined(USE_GEMM3M)
  376. #ifdef ZGEMM3M_DEFAULT_UNROLL_M
  377. ZGEMM3M_DEFAULT_UNROLL_M, ZGEMM3M_DEFAULT_UNROLL_N, MAX(ZGEMM3M_DEFAULT_UNROLL_M, ZGEMM3M_DEFAULT_UNROLL_N),
  378. #else
  379. DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N, MAX(DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N),
  380. #endif
  381. zgemm3m_kernelTS,
  382. zgemm3m_incopybTS, zgemm3m_incopyrTS,
  383. zgemm3m_incopyiTS, zgemm3m_itcopybTS,
  384. zgemm3m_itcopyrTS, zgemm3m_itcopyiTS,
  385. zgemm3m_oncopybTS, zgemm3m_oncopyrTS,
  386. zgemm3m_oncopyiTS, zgemm3m_otcopybTS,
  387. zgemm3m_otcopyrTS, zgemm3m_otcopyiTS,
  388. zsymm3m_iucopybTS, zsymm3m_ilcopybTS,
  389. zsymm3m_iucopyrTS, zsymm3m_ilcopyrTS,
  390. zsymm3m_iucopyiTS, zsymm3m_ilcopyiTS,
  391. zsymm3m_oucopybTS, zsymm3m_olcopybTS,
  392. zsymm3m_oucopyrTS, zsymm3m_olcopyrTS,
  393. zsymm3m_oucopyiTS, zsymm3m_olcopyiTS,
  394. zhemm3m_iucopybTS, zhemm3m_ilcopybTS,
  395. zhemm3m_iucopyrTS, zhemm3m_ilcopyrTS,
  396. zhemm3m_iucopyiTS, zhemm3m_ilcopyiTS,
  397. zhemm3m_oucopybTS, zhemm3m_olcopybTS,
  398. zhemm3m_oucopyrTS, zhemm3m_olcopyrTS,
  399. zhemm3m_oucopyiTS, zhemm3m_olcopyiTS,
  400. #else
  401. 0, 0, 0,
  402. NULL,
  403. NULL, NULL,
  404. NULL, NULL,
  405. NULL, NULL,
  406. NULL, NULL,
  407. NULL, NULL,
  408. NULL, NULL,
  409. NULL, NULL,
  410. NULL, NULL,
  411. NULL, NULL,
  412. NULL, NULL,
  413. NULL, NULL,
  414. NULL, NULL,
  415. NULL, NULL,
  416. NULL, NULL,
  417. NULL, NULL,
  418. NULL, NULL,
  419. NULL, NULL,
  420. NULL, NULL,
  421. #endif
  422. #ifndef NO_LAPACK
  423. zneg_tcopyTS, zlaswp_ncopyTS,
  424. #else
  425. NULL, NULL,
  426. #endif
  427. #ifdef EXPRECISION
  428. 0, 0, 0,
  429. XGEMM_DEFAULT_UNROLL_M, XGEMM_DEFAULT_UNROLL_N, MAX(XGEMM_DEFAULT_UNROLL_M, XGEMM_DEFAULT_UNROLL_N),
  430. xamax_kTS, xamin_kTS, ixamax_kTS, ixamin_kTS,
  431. xnrm2_kTS, xasum_kTS, xsum_kTS, xcopy_kTS,
  432. xdotu_kTS, xdotc_kTS, xqrot_kTS,
  433. xaxpy_kTS, xaxpyc_kTS, xscal_kTS, xswap_kTS,
  434. xgemv_nTS, xgemv_tTS, xgemv_rTS, xgemv_cTS,
  435. xgemv_oTS, xgemv_uTS, xgemv_sTS, xgemv_dTS,
  436. xgeru_kTS, xgerc_kTS, xgerv_kTS, xgerd_kTS,
  437. xsymv_LTS, xsymv_UTS,
  438. xhemv_LTS, xhemv_UTS, xhemv_MTS, xhemv_VTS,
  439. xgemm_kernel_nTS, xgemm_kernel_lTS, xgemm_kernel_rTS, xgemm_kernel_bTS,
  440. xgemm_betaTS,
  441. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  442. xgemm_incopyTS, xgemm_itcopyTS,
  443. #else
  444. xgemm_oncopyTS, xgemm_otcopyTS,
  445. #endif
  446. xgemm_oncopyTS, xgemm_otcopyTS,
  447. xtrsm_kernel_LNTS, xtrsm_kernel_LTTS, xtrsm_kernel_LRTS, xtrsm_kernel_LCTS,
  448. xtrsm_kernel_RNTS, xtrsm_kernel_RTTS, xtrsm_kernel_RRTS, xtrsm_kernel_RCTS,
  449. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  450. xtrsm_iunucopyTS, xtrsm_iunncopyTS, xtrsm_iutucopyTS, xtrsm_iutncopyTS,
  451. xtrsm_ilnucopyTS, xtrsm_ilnncopyTS, xtrsm_iltucopyTS, xtrsm_iltncopyTS,
  452. #else
  453. xtrsm_ounucopyTS, xtrsm_ounncopyTS, xtrsm_outucopyTS, xtrsm_outncopyTS,
  454. xtrsm_olnucopyTS, xtrsm_olnncopyTS, xtrsm_oltucopyTS, xtrsm_oltncopyTS,
  455. #endif
  456. xtrsm_ounucopyTS, xtrsm_ounncopyTS, xtrsm_outucopyTS, xtrsm_outncopyTS,
  457. xtrsm_olnucopyTS, xtrsm_olnncopyTS, xtrsm_oltucopyTS, xtrsm_oltncopyTS,
  458. xtrmm_kernel_RNTS, xtrmm_kernel_RTTS, xtrmm_kernel_RRTS, xtrmm_kernel_RCTS,
  459. xtrmm_kernel_LNTS, xtrmm_kernel_LTTS, xtrmm_kernel_LRTS, xtrmm_kernel_LCTS,
  460. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  461. xtrmm_iunucopyTS, xtrmm_iunncopyTS, xtrmm_iutucopyTS, xtrmm_iutncopyTS,
  462. xtrmm_ilnucopyTS, xtrmm_ilnncopyTS, xtrmm_iltucopyTS, xtrmm_iltncopyTS,
  463. #else
  464. xtrmm_ounucopyTS, xtrmm_ounncopyTS, xtrmm_outucopyTS, xtrmm_outncopyTS,
  465. xtrmm_olnucopyTS, xtrmm_olnncopyTS, xtrmm_oltucopyTS, xtrmm_oltncopyTS,
  466. #endif
  467. xtrmm_ounucopyTS, xtrmm_ounncopyTS, xtrmm_outucopyTS, xtrmm_outncopyTS,
  468. xtrmm_olnucopyTS, xtrmm_olnncopyTS, xtrmm_oltucopyTS, xtrmm_oltncopyTS,
  469. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  470. xsymm_iutcopyTS, xsymm_iltcopyTS,
  471. #else
  472. xsymm_outcopyTS, xsymm_oltcopyTS,
  473. #endif
  474. xsymm_outcopyTS, xsymm_oltcopyTS,
  475. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  476. xhemm_iutcopyTS, xhemm_iltcopyTS,
  477. #else
  478. xhemm_outcopyTS, xhemm_oltcopyTS,
  479. #endif
  480. xhemm_outcopyTS, xhemm_oltcopyTS,
  481. 0, 0, 0,
  482. #if defined(USE_GEMM3M)
  483. QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N, MAX(QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N),
  484. xgemm3m_kernelTS,
  485. xgemm3m_incopybTS, xgemm3m_incopyrTS,
  486. xgemm3m_incopyiTS, xgemm3m_itcopybTS,
  487. xgemm3m_itcopyrTS, xgemm3m_itcopyiTS,
  488. xgemm3m_oncopybTS, xgemm3m_oncopyrTS,
  489. xgemm3m_oncopyiTS, xgemm3m_otcopybTS,
  490. xgemm3m_otcopyrTS, xgemm3m_otcopyiTS,
  491. xsymm3m_iucopybTS, xsymm3m_ilcopybTS,
  492. xsymm3m_iucopyrTS, xsymm3m_ilcopyrTS,
  493. xsymm3m_iucopyiTS, xsymm3m_ilcopyiTS,
  494. xsymm3m_oucopybTS, xsymm3m_olcopybTS,
  495. xsymm3m_oucopyrTS, xsymm3m_olcopyrTS,
  496. xsymm3m_oucopyiTS, xsymm3m_olcopyiTS,
  497. xhemm3m_iucopybTS, xhemm3m_ilcopybTS,
  498. xhemm3m_iucopyrTS, xhemm3m_ilcopyrTS,
  499. xhemm3m_iucopyiTS, xhemm3m_ilcopyiTS,
  500. xhemm3m_oucopybTS, xhemm3m_olcopybTS,
  501. xhemm3m_oucopyrTS, xhemm3m_olcopyrTS,
  502. xhemm3m_oucopyiTS, xhemm3m_olcopyiTS,
  503. #else
  504. 0, 0, 0,
  505. NULL,
  506. NULL, NULL,
  507. NULL, NULL,
  508. NULL, NULL,
  509. NULL, NULL,
  510. NULL, NULL,
  511. NULL, NULL,
  512. NULL, NULL,
  513. NULL, NULL,
  514. NULL, NULL,
  515. NULL, NULL,
  516. NULL, NULL,
  517. NULL, NULL,
  518. NULL, NULL,
  519. NULL, NULL,
  520. NULL, NULL,
  521. NULL, NULL,
  522. NULL, NULL,
  523. NULL, NULL,
  524. #endif
  525. #ifndef NO_LAPACK
  526. xneg_tcopyTS, xlaswp_ncopyTS,
  527. #else
  528. NULL, NULL,
  529. #endif
  530. #endif
  531. init_parameter,
  532. SNUMOPT, DNUMOPT, QNUMOPT,
  533. saxpby_kTS, daxpby_kTS, caxpby_kTS, zaxpby_kTS,
  534. somatcopy_k_cnTS, somatcopy_k_ctTS, somatcopy_k_rnTS, somatcopy_k_rtTS,
  535. domatcopy_k_cnTS, domatcopy_k_ctTS, domatcopy_k_rnTS, domatcopy_k_rtTS,
  536. comatcopy_k_cnTS, comatcopy_k_ctTS, comatcopy_k_rnTS, comatcopy_k_rtTS,
  537. comatcopy_k_cncTS, comatcopy_k_ctcTS, comatcopy_k_rncTS, comatcopy_k_rtcTS,
  538. zomatcopy_k_cnTS, zomatcopy_k_ctTS, zomatcopy_k_rnTS, zomatcopy_k_rtTS,
  539. zomatcopy_k_cncTS, zomatcopy_k_ctcTS, zomatcopy_k_rncTS, zomatcopy_k_rtcTS,
  540. simatcopy_k_cnTS, simatcopy_k_ctTS, simatcopy_k_rnTS, simatcopy_k_rtTS,
  541. dimatcopy_k_cnTS, dimatcopy_k_ctTS, dimatcopy_k_rnTS, dimatcopy_k_rtTS,
  542. cimatcopy_k_cnTS, cimatcopy_k_ctTS, cimatcopy_k_rnTS, cimatcopy_k_rtTS,
  543. cimatcopy_k_cncTS, cimatcopy_k_ctcTS, cimatcopy_k_rncTS, cimatcopy_k_rtcTS,
  544. zimatcopy_k_cnTS, zimatcopy_k_ctTS, zimatcopy_k_rnTS, zimatcopy_k_rtTS,
  545. zimatcopy_k_cncTS, zimatcopy_k_ctcTS, zimatcopy_k_rncTS, zimatcopy_k_rtcTS,
  546. sgeadd_kTS, dgeadd_kTS, cgeadd_kTS, zgeadd_kTS
  547. };
  548. #if defined(ARCH_ARM64)
  549. static void init_parameter(void) {
  550. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  551. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  552. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  553. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  554. TABLE_NAME.sgemm_q = SGEMM_DEFAULT_Q;
  555. TABLE_NAME.dgemm_q = DGEMM_DEFAULT_Q;
  556. TABLE_NAME.cgemm_q = CGEMM_DEFAULT_Q;
  557. TABLE_NAME.zgemm_q = ZGEMM_DEFAULT_Q;
  558. TABLE_NAME.sgemm_r = SGEMM_DEFAULT_R;
  559. TABLE_NAME.dgemm_r = DGEMM_DEFAULT_R;
  560. TABLE_NAME.cgemm_r = CGEMM_DEFAULT_R;
  561. TABLE_NAME.zgemm_r = ZGEMM_DEFAULT_R;
  562. #ifdef EXPRECISION
  563. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  564. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  565. TABLE_NAME.qgemm_q = QGEMM_DEFAULT_Q;
  566. TABLE_NAME.xgemm_q = XGEMM_DEFAULT_Q;
  567. TABLE_NAME.qgemm_r = QGEMM_DEFAULT_R;
  568. TABLE_NAME.xgemm_r = XGEMM_DEFAULT_R;
  569. #endif
  570. #if defined(USE_GEMM3M)
  571. #ifdef CGEMM3M_DEFAULT_P
  572. TABLE_NAME.cgemm3m_p = CGEMM3M_DEFAULT_P;
  573. #else
  574. TABLE_NAME.cgemm3m_p = TABLE_NAME.sgemm_p;
  575. #endif
  576. #ifdef ZGEMM3M_DEFAULT_P
  577. TABLE_NAME.zgemm3m_p = ZGEMM3M_DEFAULT_P;
  578. #else
  579. TABLE_NAME.zgemm3m_p = TABLE_NAME.dgemm_p;
  580. #endif
  581. #ifdef CGEMM3M_DEFAULT_Q
  582. TABLE_NAME.cgemm3m_q = CGEMM3M_DEFAULT_Q;
  583. #else
  584. TABLE_NAME.cgemm3m_q = TABLE_NAME.sgemm_q;
  585. #endif
  586. #ifdef ZGEMM3M_DEFAULT_Q
  587. TABLE_NAME.zgemm3m_q = ZGEMM3M_DEFAULT_Q;
  588. #else
  589. TABLE_NAME.zgemm3m_q = TABLE_NAME.dgemm_q;
  590. #endif
  591. #ifdef CGEMM3M_DEFAULT_R
  592. TABLE_NAME.cgemm3m_r = CGEMM3M_DEFAULT_R;
  593. #else
  594. TABLE_NAME.cgemm3m_r = TABLE_NAME.sgemm_r;
  595. #endif
  596. #ifdef ZGEMM3M_DEFAULT_R
  597. TABLE_NAME.zgemm3m_r = ZGEMM3M_DEFAULT_R;
  598. #else
  599. TABLE_NAME.zgemm3m_r = TABLE_NAME.dgemm_r;
  600. #endif
  601. #ifdef EXPRECISION
  602. TABLE_NAME.xgemm3m_p = TABLE_NAME.qgemm_p;
  603. TABLE_NAME.xgemm3m_q = TABLE_NAME.qgemm_q;
  604. TABLE_NAME.xgemm3m_r = TABLE_NAME.qgemm_r;
  605. #endif
  606. #endif
  607. }
  608. #else // defined(ARCH_ARM64)
  609. #ifdef ARCH_X86
  610. static int get_l2_size_old(void){
  611. int i, eax, ebx, ecx, edx, cpuid_level;
  612. int info[15];
  613. cpuid(2, &eax, &ebx, &ecx, &edx);
  614. info[ 0] = BITMASK(eax, 8, 0xff);
  615. info[ 1] = BITMASK(eax, 16, 0xff);
  616. info[ 2] = BITMASK(eax, 24, 0xff);
  617. info[ 3] = BITMASK(ebx, 0, 0xff);
  618. info[ 4] = BITMASK(ebx, 8, 0xff);
  619. info[ 5] = BITMASK(ebx, 16, 0xff);
  620. info[ 6] = BITMASK(ebx, 24, 0xff);
  621. info[ 7] = BITMASK(ecx, 0, 0xff);
  622. info[ 8] = BITMASK(ecx, 8, 0xff);
  623. info[ 9] = BITMASK(ecx, 16, 0xff);
  624. info[10] = BITMASK(ecx, 24, 0xff);
  625. info[11] = BITMASK(edx, 0, 0xff);
  626. info[12] = BITMASK(edx, 8, 0xff);
  627. info[13] = BITMASK(edx, 16, 0xff);
  628. info[14] = BITMASK(edx, 24, 0xff);
  629. for (i = 0; i < 15; i++){
  630. switch (info[i]){
  631. /* This table is from http://www.sandpile.org/ia32/cpuid.htm */
  632. case 0x1a :
  633. return 96;
  634. case 0x39 :
  635. case 0x3b :
  636. case 0x41 :
  637. case 0x79 :
  638. case 0x81 :
  639. return 128;
  640. case 0x3a :
  641. return 192;
  642. case 0x21 :
  643. case 0x3c :
  644. case 0x42 :
  645. case 0x7a :
  646. case 0x7e :
  647. case 0x82 :
  648. return 256;
  649. case 0x3d :
  650. return 384;
  651. case 0x3e :
  652. case 0x43 :
  653. case 0x7b :
  654. case 0x7f :
  655. case 0x83 :
  656. case 0x86 :
  657. return 512;
  658. case 0x44 :
  659. case 0x78 :
  660. case 0x7c :
  661. case 0x84 :
  662. case 0x87 :
  663. return 1024;
  664. case 0x45 :
  665. case 0x7d :
  666. case 0x85 :
  667. return 2048;
  668. case 0x48 :
  669. return 3184;
  670. case 0x49 :
  671. return 4096;
  672. case 0x4e :
  673. return 6144;
  674. }
  675. }
  676. // return 0;
  677. fprintf (stderr,"OpenBLAS WARNING - could not determine the L2 cache size on this system, assuming 256k\n");
  678. return 256;
  679. }
  680. #endif
  681. static __inline__ int get_l2_size(void){
  682. int eax, ebx, ecx, edx, l2;
  683. cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
  684. l2 = BITMASK(ecx, 16, 0xffff);
  685. #ifndef ARCH_X86
  686. if (l2 <= 0) {
  687. fprintf (stderr,"OpenBLAS WARNING - could not determine the L2 cache size on this system, assuming 256k\n");
  688. return 256;
  689. }
  690. return l2;
  691. #else
  692. if (l2 > 0) return l2;
  693. return get_l2_size_old();
  694. #endif
  695. }
  696. static __inline__ int get_l3_size(void){
  697. int eax, ebx, ecx, edx;
  698. cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
  699. return BITMASK(edx, 18, 0x3fff) * 512;
  700. }
  701. static void init_parameter(void) {
  702. int l2 = get_l2_size();
  703. (void) l2; /* dirty trick to suppress unused variable warning for targets */
  704. /* where the GEMM unrolling parameters do not depend on l2 */
  705. TABLE_NAME.sgemm_q = SGEMM_DEFAULT_Q;
  706. TABLE_NAME.dgemm_q = DGEMM_DEFAULT_Q;
  707. TABLE_NAME.cgemm_q = CGEMM_DEFAULT_Q;
  708. TABLE_NAME.zgemm_q = ZGEMM_DEFAULT_Q;
  709. #ifdef CGEMM3M_DEFAULT_Q
  710. TABLE_NAME.cgemm3m_q = CGEMM3M_DEFAULT_Q;
  711. #else
  712. TABLE_NAME.cgemm3m_q = SGEMM_DEFAULT_Q;
  713. #endif
  714. #ifdef ZGEMM3M_DEFAULT_Q
  715. TABLE_NAME.zgemm3m_q = ZGEMM3M_DEFAULT_Q;
  716. #else
  717. TABLE_NAME.zgemm3m_q = DGEMM_DEFAULT_Q;
  718. #endif
  719. #ifdef EXPRECISION
  720. TABLE_NAME.qgemm_q = QGEMM_DEFAULT_Q;
  721. TABLE_NAME.xgemm_q = XGEMM_DEFAULT_Q;
  722. TABLE_NAME.xgemm3m_q = QGEMM_DEFAULT_Q;
  723. #endif
  724. #if defined(CORE_KATMAI) || defined(CORE_COPPERMINE) || defined(CORE_BANIAS) || defined(CORE_YONAH) || defined(CORE_ATHLON)
  725. #ifdef DEBUG
  726. fprintf(stderr, "Katmai, Coppermine, Banias, Athlon\n");
  727. #endif
  728. TABLE_NAME.sgemm_p = 64 * (l2 >> 7);
  729. TABLE_NAME.dgemm_p = 32 * (l2 >> 7);
  730. TABLE_NAME.cgemm_p = 32 * (l2 >> 7);
  731. TABLE_NAME.zgemm_p = 16 * (l2 >> 7);
  732. #ifdef EXPRECISION
  733. TABLE_NAME.qgemm_p = 16 * (l2 >> 7);
  734. TABLE_NAME.xgemm_p = 8 * (l2 >> 7);
  735. #endif
  736. #endif
  737. #ifdef CORE_NORTHWOOD
  738. #ifdef DEBUG
  739. fprintf(stderr, "Northwood\n");
  740. #endif
  741. TABLE_NAME.sgemm_p = 96 * (l2 >> 7);
  742. TABLE_NAME.dgemm_p = 48 * (l2 >> 7);
  743. TABLE_NAME.cgemm_p = 48 * (l2 >> 7);
  744. TABLE_NAME.zgemm_p = 24 * (l2 >> 7);
  745. #ifdef EXPRECISION
  746. TABLE_NAME.qgemm_p = 24 * (l2 >> 7);
  747. TABLE_NAME.xgemm_p = 12 * (l2 >> 7);
  748. #endif
  749. #endif
  750. #ifdef ATOM
  751. #ifdef DEBUG
  752. fprintf(stderr, "Atom\n");
  753. #endif
  754. TABLE_NAME.sgemm_p = 256;
  755. TABLE_NAME.dgemm_p = 128;
  756. TABLE_NAME.cgemm_p = 128;
  757. TABLE_NAME.zgemm_p = 64;
  758. #ifdef EXPRECISION
  759. TABLE_NAME.qgemm_p = 64;
  760. TABLE_NAME.xgemm_p = 32;
  761. #endif
  762. #endif
  763. #ifdef CORE_PRESCOTT
  764. #ifdef DEBUG
  765. fprintf(stderr, "Prescott\n");
  766. #endif
  767. TABLE_NAME.sgemm_p = 56 * (l2 >> 7);
  768. TABLE_NAME.dgemm_p = 28 * (l2 >> 7);
  769. TABLE_NAME.cgemm_p = 28 * (l2 >> 7);
  770. TABLE_NAME.zgemm_p = 14 * (l2 >> 7);
  771. #ifdef EXPRECISION
  772. TABLE_NAME.qgemm_p = 14 * (l2 >> 7);
  773. TABLE_NAME.xgemm_p = 7 * (l2 >> 7);
  774. #endif
  775. #endif
  776. #ifdef CORE2
  777. #ifdef DEBUG
  778. fprintf(stderr, "Core2\n");
  779. #endif
  780. TABLE_NAME.sgemm_p = 92 * (l2 >> 9) + 8;
  781. TABLE_NAME.dgemm_p = 46 * (l2 >> 9) + 8;
  782. TABLE_NAME.cgemm_p = 46 * (l2 >> 9) + 4;
  783. TABLE_NAME.zgemm_p = 23 * (l2 >> 9) + 4;
  784. #ifdef EXPRECISION
  785. TABLE_NAME.qgemm_p = 92 * (l2 >> 9) + 8;
  786. TABLE_NAME.xgemm_p = 46 * (l2 >> 9) + 4;
  787. #endif
  788. #endif
  789. #ifdef PENRYN
  790. #ifdef DEBUG
  791. fprintf(stderr, "Penryn\n");
  792. #endif
  793. TABLE_NAME.sgemm_p = 42 * (l2 >> 9) + 8;
  794. TABLE_NAME.dgemm_p = 42 * (l2 >> 9) + 8;
  795. TABLE_NAME.cgemm_p = 21 * (l2 >> 9) + 4;
  796. TABLE_NAME.zgemm_p = 21 * (l2 >> 9) + 4;
  797. #ifdef EXPRECISION
  798. TABLE_NAME.qgemm_p = 42 * (l2 >> 9) + 8;
  799. TABLE_NAME.xgemm_p = 21 * (l2 >> 9) + 4;
  800. #endif
  801. #endif
  802. #ifdef DUNNINGTON
  803. #ifdef DEBUG
  804. fprintf(stderr, "Dunnington\n");
  805. #endif
  806. TABLE_NAME.sgemm_p = 42 * (l2 >> 9) + 8;
  807. TABLE_NAME.dgemm_p = 42 * (l2 >> 9) + 8;
  808. TABLE_NAME.cgemm_p = 21 * (l2 >> 9) + 4;
  809. TABLE_NAME.zgemm_p = 21 * (l2 >> 9) + 4;
  810. #ifdef EXPRECISION
  811. TABLE_NAME.qgemm_p = 42 * (l2 >> 9) + 8;
  812. TABLE_NAME.xgemm_p = 21 * (l2 >> 9) + 4;
  813. #endif
  814. #endif
  815. #ifdef NEHALEM
  816. #ifdef DEBUG
  817. fprintf(stderr, "Nehalem\n");
  818. #endif
  819. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  820. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  821. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  822. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  823. #ifdef EXPRECISION
  824. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  825. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  826. #endif
  827. #endif
  828. #ifdef SANDYBRIDGE
  829. #ifdef DEBUG
  830. fprintf(stderr, "Sandybridge\n");
  831. #endif
  832. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  833. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  834. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  835. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  836. #ifdef EXPRECISION
  837. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  838. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  839. #endif
  840. #endif
  841. #ifdef HASWELL
  842. #ifdef DEBUG
  843. fprintf(stderr, "Haswell\n");
  844. #endif
  845. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  846. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  847. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  848. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  849. #ifdef EXPRECISION
  850. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  851. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  852. #endif
  853. #endif
  854. #ifdef SKYLAKEX
  855. #ifdef DEBUG
  856. fprintf(stderr, "SkylakeX\n");
  857. #endif
  858. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  859. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  860. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  861. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  862. #ifdef EXPRECISION
  863. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  864. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  865. #endif
  866. #endif
  867. #ifdef OPTERON
  868. #ifdef DEBUG
  869. fprintf(stderr, "Opteron\n");
  870. #endif
  871. TABLE_NAME.sgemm_p = 224 + 56 * (l2 >> 7);
  872. TABLE_NAME.dgemm_p = 112 + 28 * (l2 >> 7);
  873. TABLE_NAME.cgemm_p = 112 + 28 * (l2 >> 7);
  874. TABLE_NAME.zgemm_p = 56 + 14 * (l2 >> 7);
  875. #ifdef EXPRECISION
  876. TABLE_NAME.qgemm_p = 56 + 14 * (l2 >> 7);
  877. TABLE_NAME.xgemm_p = 28 + 7 * (l2 >> 7);
  878. #endif
  879. #endif
  880. #ifdef BARCELONA
  881. #ifdef DEBUG
  882. fprintf(stderr, "Barcelona\n");
  883. #endif
  884. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  885. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  886. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  887. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  888. #ifdef EXPRECISION
  889. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  890. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  891. #endif
  892. #endif
  893. #ifdef BOBCAT
  894. #ifdef DEBUG
  895. fprintf(stderr, "Bobcate\n");
  896. #endif
  897. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  898. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  899. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  900. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  901. #ifdef EXPRECISION
  902. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  903. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  904. #endif
  905. #endif
  906. #ifdef BULLDOZER
  907. #ifdef DEBUG
  908. fprintf(stderr, "Bulldozer\n");
  909. #endif
  910. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  911. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  912. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  913. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  914. #ifdef EXPRECISION
  915. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  916. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  917. #endif
  918. #endif
  919. #ifdef EXCAVATOR
  920. #ifdef DEBUG
  921. fprintf(stderr, "Excavator\n");
  922. #endif
  923. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  924. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  925. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  926. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  927. #ifdef EXPRECISION
  928. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  929. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  930. #endif
  931. #endif
  932. #ifdef PILEDRIVER
  933. #ifdef DEBUG
  934. fprintf(stderr, "Piledriver\n");
  935. #endif
  936. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  937. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  938. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  939. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  940. #ifdef EXPRECISION
  941. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  942. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  943. #endif
  944. #endif
  945. #ifdef STEAMROLLER
  946. #ifdef DEBUG
  947. fprintf(stderr, "Steamroller\n");
  948. #endif
  949. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  950. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  951. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  952. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  953. #ifdef EXPRECISION
  954. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  955. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  956. #endif
  957. #endif
  958. #ifdef ZEN
  959. #ifdef DEBUG
  960. fprintf(stderr, "Zen\n");
  961. #endif
  962. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  963. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  964. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  965. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  966. #ifdef EXPRECISION
  967. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  968. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  969. #endif
  970. #endif
  971. #ifdef NANO
  972. #ifdef DEBUG
  973. fprintf(stderr, "NANO\n");
  974. #endif
  975. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  976. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  977. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  978. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  979. #ifdef EXPRECISION
  980. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  981. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  982. #endif
  983. #endif
  984. #ifdef CGEMM3M_DEFAULT_P
  985. TABLE_NAME.cgemm3m_p = CGEMM3M_DEFAULT_P;
  986. #else
  987. TABLE_NAME.cgemm3m_p = TABLE_NAME.sgemm_p;
  988. #endif
  989. #ifdef ZGEMM3M_DEFAULT_P
  990. TABLE_NAME.zgemm3m_p = ZGEMM3M_DEFAULT_P;
  991. #else
  992. TABLE_NAME.zgemm3m_p = TABLE_NAME.dgemm_p;
  993. #endif
  994. #ifdef EXPRECISION
  995. TABLE_NAME.xgemm3m_p = TABLE_NAME.qgemm_p;
  996. #endif
  997. TABLE_NAME.sgemm_p = ((TABLE_NAME.sgemm_p + SGEMM_DEFAULT_UNROLL_M - 1)/SGEMM_DEFAULT_UNROLL_M) * SGEMM_DEFAULT_UNROLL_M;
  998. TABLE_NAME.dgemm_p = ((TABLE_NAME.dgemm_p + DGEMM_DEFAULT_UNROLL_M - 1)/DGEMM_DEFAULT_UNROLL_M) * DGEMM_DEFAULT_UNROLL_M;
  999. TABLE_NAME.cgemm_p = ((TABLE_NAME.cgemm_p + CGEMM_DEFAULT_UNROLL_M - 1)/CGEMM_DEFAULT_UNROLL_M) * CGEMM_DEFAULT_UNROLL_M;
  1000. TABLE_NAME.zgemm_p = ((TABLE_NAME.zgemm_p + ZGEMM_DEFAULT_UNROLL_M - 1)/ZGEMM_DEFAULT_UNROLL_M) * ZGEMM_DEFAULT_UNROLL_M;
  1001. #ifdef CGEMM3M_DEFAULT_UNROLL_M
  1002. TABLE_NAME.cgemm3m_p = ((TABLE_NAME.cgemm3m_p + CGEMM3M_DEFAULT_UNROLL_M - 1)/CGEMM3M_DEFAULT_UNROLL_M) * CGEMM3M_DEFAULT_UNROLL_M;
  1003. #else
  1004. TABLE_NAME.cgemm3m_p = ((TABLE_NAME.cgemm3m_p + SGEMM_DEFAULT_UNROLL_M - 1)/SGEMM_DEFAULT_UNROLL_M) * SGEMM_DEFAULT_UNROLL_M;
  1005. #endif
  1006. #ifdef ZGEMM3M_DEFAULT_UNROLL_M
  1007. TABLE_NAME.zgemm3m_p = ((TABLE_NAME.zgemm3m_p + ZGEMM3M_DEFAULT_UNROLL_M - 1)/ZGEMM3M_DEFAULT_UNROLL_M) * ZGEMM3M_DEFAULT_UNROLL_M;
  1008. #else
  1009. TABLE_NAME.zgemm3m_p = ((TABLE_NAME.zgemm3m_p + DGEMM_DEFAULT_UNROLL_M - 1)/DGEMM_DEFAULT_UNROLL_M) * DGEMM_DEFAULT_UNROLL_M;
  1010. #endif
  1011. #ifdef QUAD_PRECISION
  1012. TABLE_NAME.qgemm_p = ((TABLE_NAME.qgemm_p + QGEMM_DEFAULT_UNROLL_M - 1)/QGEMM_DEFAULT_UNROLL_M) * QGEMM_DEFAULT_UNROLL_M;
  1013. TABLE_NAME.xgemm_p = ((TABLE_NAME.xgemm_p + XGEMM_DEFAULT_UNROLL_M - 1)/XGEMM_DEFAULT_UNROLL_M) * XGEMM_DEFAULT_UNROLL_M;
  1014. TABLE_NAME.xgemm3m_p = ((TABLE_NAME.xgemm3m_p + QGEMM_DEFAULT_UNROLL_M - 1)/QGEMM_DEFAULT_UNROLL_M) * QGEMM_DEFAULT_UNROLL_M;
  1015. #endif
  1016. #ifdef DEBUG
  1017. fprintf(stderr, "L2 = %8d DGEMM_P .. %d\n", l2, TABLE_NAME.dgemm_p);
  1018. #endif
  1019. TABLE_NAME.sgemm_r = (((BUFFER_SIZE -
  1020. ((TABLE_NAME.sgemm_p * TABLE_NAME.sgemm_q * 4 + TABLE_NAME.offsetA
  1021. + TABLE_NAME.align) & ~TABLE_NAME.align)
  1022. ) / (TABLE_NAME.sgemm_q * 4) - 15) & ~15);
  1023. TABLE_NAME.dgemm_r = (((BUFFER_SIZE -
  1024. ((TABLE_NAME.dgemm_p * TABLE_NAME.dgemm_q * 8 + TABLE_NAME.offsetA
  1025. + TABLE_NAME.align) & ~TABLE_NAME.align)
  1026. ) / (TABLE_NAME.dgemm_q * 8) - 15) & ~15);
  1027. #ifdef EXPRECISION
  1028. TABLE_NAME.qgemm_r = (((BUFFER_SIZE -
  1029. ((TABLE_NAME.qgemm_p * TABLE_NAME.qgemm_q * 16 + TABLE_NAME.offsetA
  1030. + TABLE_NAME.align) & ~TABLE_NAME.align)
  1031. ) / (TABLE_NAME.qgemm_q * 16) - 15) & ~15);
  1032. #endif
  1033. TABLE_NAME.cgemm_r = (((BUFFER_SIZE -
  1034. ((TABLE_NAME.cgemm_p * TABLE_NAME.cgemm_q * 8 + TABLE_NAME.offsetA
  1035. + TABLE_NAME.align) & ~TABLE_NAME.align)
  1036. ) / (TABLE_NAME.cgemm_q * 8) - 15) & ~15);
  1037. TABLE_NAME.zgemm_r = (((BUFFER_SIZE -
  1038. ((TABLE_NAME.zgemm_p * TABLE_NAME.zgemm_q * 16 + TABLE_NAME.offsetA
  1039. + TABLE_NAME.align) & ~TABLE_NAME.align)
  1040. ) / (TABLE_NAME.zgemm_q * 16) - 15) & ~15);
  1041. TABLE_NAME.cgemm3m_r = (((BUFFER_SIZE -
  1042. ((TABLE_NAME.cgemm3m_p * TABLE_NAME.cgemm3m_q * 8 + TABLE_NAME.offsetA
  1043. + TABLE_NAME.align) & ~TABLE_NAME.align)
  1044. ) / (TABLE_NAME.cgemm3m_q * 8) - 15) & ~15);
  1045. TABLE_NAME.zgemm3m_r = (((BUFFER_SIZE -
  1046. ((TABLE_NAME.zgemm3m_p * TABLE_NAME.zgemm3m_q * 16 + TABLE_NAME.offsetA
  1047. + TABLE_NAME.align) & ~TABLE_NAME.align)
  1048. ) / (TABLE_NAME.zgemm3m_q * 16) - 15) & ~15);
  1049. #ifdef EXPRECISION
  1050. TABLE_NAME.xgemm_r = (((BUFFER_SIZE -
  1051. ((TABLE_NAME.xgemm_p * TABLE_NAME.xgemm_q * 32 + TABLE_NAME.offsetA
  1052. + TABLE_NAME.align) & ~TABLE_NAME.align)
  1053. ) / (TABLE_NAME.xgemm_q * 32) - 15) & ~15);
  1054. TABLE_NAME.xgemm3m_r = (((BUFFER_SIZE -
  1055. ((TABLE_NAME.xgemm3m_p * TABLE_NAME.xgemm3m_q * 32 + TABLE_NAME.offsetA
  1056. + TABLE_NAME.align) & ~TABLE_NAME.align)
  1057. ) / (TABLE_NAME.xgemm3m_q * 32) - 15) & ~15);
  1058. #endif
  1059. }
  1060. #endif //defined(ARCH_ARM64)