You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

setparam-ref.c 33 kB

11 years ago
11 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083
  1. /*********************************************************************/
  2. /* Copyright 2009, 2010 The University of Texas at Austin. */
  3. /* All rights reserved. */
  4. /* */
  5. /* Redistribution and use in source and binary forms, with or */
  6. /* without modification, are permitted provided that the following */
  7. /* conditions are met: */
  8. /* */
  9. /* 1. Redistributions of source code must retain the above */
  10. /* copyright notice, this list of conditions and the following */
  11. /* disclaimer. */
  12. /* */
  13. /* 2. Redistributions in binary form must reproduce the above */
  14. /* copyright notice, this list of conditions and the following */
  15. /* disclaimer in the documentation and/or other materials */
  16. /* provided with the distribution. */
  17. /* */
  18. /* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
  19. /* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
  20. /* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
  21. /* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
  22. /* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
  23. /* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
  24. /* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
  25. /* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
  26. /* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
  27. /* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
  28. /* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
  29. /* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
  30. /* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
  31. /* POSSIBILITY OF SUCH DAMAGE. */
  32. /* */
  33. /* The views and conclusions contained in the software and */
  34. /* documentation are those of the authors and should not be */
  35. /* interpreted as representing official policies, either expressed */
  36. /* or implied, of The University of Texas at Austin. */
  37. /*********************************************************************/
  38. #include <stdio.h>
  39. #include <string.h>
  40. #include "common.h"
  41. #ifdef BUILD_KERNEL
  42. #include "kernelTS.h"
  43. #endif
  44. #undef DEBUG
  45. static void init_parameter(void);
  46. gotoblas_t TABLE_NAME = {
  47. DTB_DEFAULT_ENTRIES ,
  48. GEMM_DEFAULT_OFFSET_A, GEMM_DEFAULT_OFFSET_B, GEMM_DEFAULT_ALIGN,
  49. 0, 0, 0,
  50. SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N,
  51. #ifdef SGEMM_DEFAULT_UNROLL_MN
  52. SGEMM_DEFAULT_UNROLL_MN,
  53. #else
  54. MAX(SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N),
  55. #endif
  56. #ifdef HAVE_EXCLUSIVE_CACHE
  57. 1,
  58. #else
  59. 0,
  60. #endif
  61. samax_kTS, samin_kTS, smax_kTS, smin_kTS,
  62. isamax_kTS, isamin_kTS, ismax_kTS, ismin_kTS,
  63. snrm2_kTS, sasum_kTS, scopy_kTS, sdot_kTS,
  64. dsdot_kTS,
  65. srot_kTS, saxpy_kTS, sscal_kTS, sswap_kTS,
  66. sgemv_nTS, sgemv_tTS, sger_kTS,
  67. ssymv_LTS, ssymv_UTS,
  68. sgemm_kernelTS, sgemm_betaTS,
  69. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  70. sgemm_incopyTS, sgemm_itcopyTS,
  71. #else
  72. sgemm_oncopyTS, sgemm_otcopyTS,
  73. #endif
  74. sgemm_oncopyTS, sgemm_otcopyTS,
  75. strsm_kernel_LNTS, strsm_kernel_LTTS, strsm_kernel_RNTS, strsm_kernel_RTTS,
  76. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  77. strsm_iunucopyTS, strsm_iunncopyTS, strsm_iutucopyTS, strsm_iutncopyTS,
  78. strsm_ilnucopyTS, strsm_ilnncopyTS, strsm_iltucopyTS, strsm_iltncopyTS,
  79. #else
  80. strsm_ounucopyTS, strsm_ounncopyTS, strsm_outucopyTS, strsm_outncopyTS,
  81. strsm_olnucopyTS, strsm_olnncopyTS, strsm_oltucopyTS, strsm_oltncopyTS,
  82. #endif
  83. strsm_ounucopyTS, strsm_ounncopyTS, strsm_outucopyTS, strsm_outncopyTS,
  84. strsm_olnucopyTS, strsm_olnncopyTS, strsm_oltucopyTS, strsm_oltncopyTS,
  85. strmm_kernel_RNTS, strmm_kernel_RTTS, strmm_kernel_LNTS, strmm_kernel_LTTS,
  86. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  87. strmm_iunucopyTS, strmm_iunncopyTS, strmm_iutucopyTS, strmm_iutncopyTS,
  88. strmm_ilnucopyTS, strmm_ilnncopyTS, strmm_iltucopyTS, strmm_iltncopyTS,
  89. #else
  90. strmm_ounucopyTS, strmm_ounncopyTS, strmm_outucopyTS, strmm_outncopyTS,
  91. strmm_olnucopyTS, strmm_olnncopyTS, strmm_oltucopyTS, strmm_oltncopyTS,
  92. #endif
  93. strmm_ounucopyTS, strmm_ounncopyTS, strmm_outucopyTS, strmm_outncopyTS,
  94. strmm_olnucopyTS, strmm_olnncopyTS, strmm_oltucopyTS, strmm_oltncopyTS,
  95. #if SGEMM_DEFAULT_UNROLL_M != SGEMM_DEFAULT_UNROLL_N
  96. ssymm_iutcopyTS, ssymm_iltcopyTS,
  97. #else
  98. ssymm_outcopyTS, ssymm_oltcopyTS,
  99. #endif
  100. ssymm_outcopyTS, ssymm_oltcopyTS,
  101. #ifndef NO_LAPACK
  102. sneg_tcopyTS, slaswp_ncopyTS,
  103. #else
  104. NULL,NULL,
  105. #endif
  106. 0, 0, 0,
  107. DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N,
  108. #ifdef DGEMM_DEFAULT_UNROLL_MN
  109. DGEMM_DEFAULT_UNROLL_MN,
  110. #else
  111. MAX(DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N),
  112. #endif
  113. damax_kTS, damin_kTS, dmax_kTS, dmin_kTS,
  114. idamax_kTS, idamin_kTS, idmax_kTS, idmin_kTS,
  115. dnrm2_kTS, dasum_kTS, dcopy_kTS, ddot_kTS,
  116. drot_kTS, daxpy_kTS, dscal_kTS, dswap_kTS,
  117. dgemv_nTS, dgemv_tTS, dger_kTS,
  118. dsymv_LTS, dsymv_UTS,
  119. dgemm_kernelTS, dgemm_betaTS,
  120. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  121. dgemm_incopyTS, dgemm_itcopyTS,
  122. #else
  123. dgemm_oncopyTS, dgemm_otcopyTS,
  124. #endif
  125. dgemm_oncopyTS, dgemm_otcopyTS,
  126. dtrsm_kernel_LNTS, dtrsm_kernel_LTTS, dtrsm_kernel_RNTS, dtrsm_kernel_RTTS,
  127. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  128. dtrsm_iunucopyTS, dtrsm_iunncopyTS, dtrsm_iutucopyTS, dtrsm_iutncopyTS,
  129. dtrsm_ilnucopyTS, dtrsm_ilnncopyTS, dtrsm_iltucopyTS, dtrsm_iltncopyTS,
  130. #else
  131. dtrsm_ounucopyTS, dtrsm_ounncopyTS, dtrsm_outucopyTS, dtrsm_outncopyTS,
  132. dtrsm_olnucopyTS, dtrsm_olnncopyTS, dtrsm_oltucopyTS, dtrsm_oltncopyTS,
  133. #endif
  134. dtrsm_ounucopyTS, dtrsm_ounncopyTS, dtrsm_outucopyTS, dtrsm_outncopyTS,
  135. dtrsm_olnucopyTS, dtrsm_olnncopyTS, dtrsm_oltucopyTS, dtrsm_oltncopyTS,
  136. dtrmm_kernel_RNTS, dtrmm_kernel_RTTS, dtrmm_kernel_LNTS, dtrmm_kernel_LTTS,
  137. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  138. dtrmm_iunucopyTS, dtrmm_iunncopyTS, dtrmm_iutucopyTS, dtrmm_iutncopyTS,
  139. dtrmm_ilnucopyTS, dtrmm_ilnncopyTS, dtrmm_iltucopyTS, dtrmm_iltncopyTS,
  140. #else
  141. dtrmm_ounucopyTS, dtrmm_ounncopyTS, dtrmm_outucopyTS, dtrmm_outncopyTS,
  142. dtrmm_olnucopyTS, dtrmm_olnncopyTS, dtrmm_oltucopyTS, dtrmm_oltncopyTS,
  143. #endif
  144. dtrmm_ounucopyTS, dtrmm_ounncopyTS, dtrmm_outucopyTS, dtrmm_outncopyTS,
  145. dtrmm_olnucopyTS, dtrmm_olnncopyTS, dtrmm_oltucopyTS, dtrmm_oltncopyTS,
  146. #if DGEMM_DEFAULT_UNROLL_M != DGEMM_DEFAULT_UNROLL_N
  147. dsymm_iutcopyTS, dsymm_iltcopyTS,
  148. #else
  149. dsymm_outcopyTS, dsymm_oltcopyTS,
  150. #endif
  151. dsymm_outcopyTS, dsymm_oltcopyTS,
  152. #ifndef NO_LAPACK
  153. dneg_tcopyTS, dlaswp_ncopyTS,
  154. #else
  155. NULL, NULL,
  156. #endif
  157. #ifdef EXPRECISION
  158. 0, 0, 0,
  159. QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N, MAX(QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N),
  160. qamax_kTS, qamin_kTS, qmax_kTS, qmin_kTS,
  161. iqamax_kTS, iqamin_kTS, iqmax_kTS, iqmin_kTS,
  162. qnrm2_kTS, qasum_kTS, qcopy_kTS, qdot_kTS,
  163. qrot_kTS, qaxpy_kTS, qscal_kTS, qswap_kTS,
  164. qgemv_nTS, qgemv_tTS, qger_kTS,
  165. qsymv_LTS, qsymv_UTS,
  166. qgemm_kernelTS, qgemm_betaTS,
  167. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  168. qgemm_incopyTS, qgemm_itcopyTS,
  169. #else
  170. qgemm_oncopyTS, qgemm_otcopyTS,
  171. #endif
  172. qgemm_oncopyTS, qgemm_otcopyTS,
  173. qtrsm_kernel_LNTS, qtrsm_kernel_LTTS, qtrsm_kernel_RNTS, qtrsm_kernel_RTTS,
  174. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  175. qtrsm_iunucopyTS, qtrsm_iunncopyTS, qtrsm_iutucopyTS, qtrsm_iutncopyTS,
  176. qtrsm_ilnucopyTS, qtrsm_ilnncopyTS, qtrsm_iltucopyTS, qtrsm_iltncopyTS,
  177. #else
  178. qtrsm_ounucopyTS, qtrsm_ounncopyTS, qtrsm_outucopyTS, qtrsm_outncopyTS,
  179. qtrsm_olnucopyTS, qtrsm_olnncopyTS, qtrsm_oltucopyTS, qtrsm_oltncopyTS,
  180. #endif
  181. qtrsm_ounucopyTS, qtrsm_ounncopyTS, qtrsm_outucopyTS, qtrsm_outncopyTS,
  182. qtrsm_olnucopyTS, qtrsm_olnncopyTS, qtrsm_oltucopyTS, qtrsm_oltncopyTS,
  183. qtrmm_kernel_RNTS, qtrmm_kernel_RTTS, qtrmm_kernel_LNTS, qtrmm_kernel_LTTS,
  184. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  185. qtrmm_iunucopyTS, qtrmm_iunncopyTS, qtrmm_iutucopyTS, qtrmm_iutncopyTS,
  186. qtrmm_ilnucopyTS, qtrmm_ilnncopyTS, qtrmm_iltucopyTS, qtrmm_iltncopyTS,
  187. #else
  188. qtrmm_ounucopyTS, qtrmm_ounncopyTS, qtrmm_outucopyTS, qtrmm_outncopyTS,
  189. qtrmm_olnucopyTS, qtrmm_olnncopyTS, qtrmm_oltucopyTS, qtrmm_oltncopyTS,
  190. #endif
  191. qtrmm_ounucopyTS, qtrmm_ounncopyTS, qtrmm_outucopyTS, qtrmm_outncopyTS,
  192. qtrmm_olnucopyTS, qtrmm_olnncopyTS, qtrmm_oltucopyTS, qtrmm_oltncopyTS,
  193. #if QGEMM_DEFAULT_UNROLL_M != QGEMM_DEFAULT_UNROLL_N
  194. qsymm_iutcopyTS, qsymm_iltcopyTS,
  195. #else
  196. qsymm_outcopyTS, qsymm_oltcopyTS,
  197. #endif
  198. qsymm_outcopyTS, qsymm_oltcopyTS,
  199. #ifndef NO_LAPACK
  200. qneg_tcopyTS, qlaswp_ncopyTS,
  201. #else
  202. NULL, NULL,
  203. #endif
  204. #endif
  205. 0, 0, 0,
  206. CGEMM_DEFAULT_UNROLL_M, CGEMM_DEFAULT_UNROLL_N,
  207. #ifdef CGEMM_DEFAULT_UNROLL_MN
  208. CGEMM_DEFAULT_UNROLL_MN,
  209. #else
  210. MAX(CGEMM_DEFAULT_UNROLL_M, CGEMM_DEFAULT_UNROLL_N),
  211. #endif
  212. camax_kTS, camin_kTS, icamax_kTS, icamin_kTS,
  213. cnrm2_kTS, casum_kTS, ccopy_kTS,
  214. cdotu_kTS, cdotc_kTS, csrot_kTS,
  215. caxpy_kTS, caxpyc_kTS, cscal_kTS, cswap_kTS,
  216. cgemv_nTS, cgemv_tTS, cgemv_rTS, cgemv_cTS,
  217. cgemv_oTS, cgemv_uTS, cgemv_sTS, cgemv_dTS,
  218. cgeru_kTS, cgerc_kTS, cgerv_kTS, cgerd_kTS,
  219. csymv_LTS, csymv_UTS,
  220. chemv_LTS, chemv_UTS, chemv_MTS, chemv_VTS,
  221. cgemm_kernel_nTS, cgemm_kernel_lTS, cgemm_kernel_rTS, cgemm_kernel_bTS,
  222. cgemm_betaTS,
  223. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  224. cgemm_incopyTS, cgemm_itcopyTS,
  225. #else
  226. cgemm_oncopyTS, cgemm_otcopyTS,
  227. #endif
  228. cgemm_oncopyTS, cgemm_otcopyTS,
  229. ctrsm_kernel_LNTS, ctrsm_kernel_LTTS, ctrsm_kernel_LRTS, ctrsm_kernel_LCTS,
  230. ctrsm_kernel_RNTS, ctrsm_kernel_RTTS, ctrsm_kernel_RRTS, ctrsm_kernel_RCTS,
  231. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  232. ctrsm_iunucopyTS, ctrsm_iunncopyTS, ctrsm_iutucopyTS, ctrsm_iutncopyTS,
  233. ctrsm_ilnucopyTS, ctrsm_ilnncopyTS, ctrsm_iltucopyTS, ctrsm_iltncopyTS,
  234. #else
  235. ctrsm_ounucopyTS, ctrsm_ounncopyTS, ctrsm_outucopyTS, ctrsm_outncopyTS,
  236. ctrsm_olnucopyTS, ctrsm_olnncopyTS, ctrsm_oltucopyTS, ctrsm_oltncopyTS,
  237. #endif
  238. ctrsm_ounucopyTS, ctrsm_ounncopyTS, ctrsm_outucopyTS, ctrsm_outncopyTS,
  239. ctrsm_olnucopyTS, ctrsm_olnncopyTS, ctrsm_oltucopyTS, ctrsm_oltncopyTS,
  240. ctrmm_kernel_RNTS, ctrmm_kernel_RTTS, ctrmm_kernel_RRTS, ctrmm_kernel_RCTS,
  241. ctrmm_kernel_LNTS, ctrmm_kernel_LTTS, ctrmm_kernel_LRTS, ctrmm_kernel_LCTS,
  242. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  243. ctrmm_iunucopyTS, ctrmm_iunncopyTS, ctrmm_iutucopyTS, ctrmm_iutncopyTS,
  244. ctrmm_ilnucopyTS, ctrmm_ilnncopyTS, ctrmm_iltucopyTS, ctrmm_iltncopyTS,
  245. #else
  246. ctrmm_ounucopyTS, ctrmm_ounncopyTS, ctrmm_outucopyTS, ctrmm_outncopyTS,
  247. ctrmm_olnucopyTS, ctrmm_olnncopyTS, ctrmm_oltucopyTS, ctrmm_oltncopyTS,
  248. #endif
  249. ctrmm_ounucopyTS, ctrmm_ounncopyTS, ctrmm_outucopyTS, ctrmm_outncopyTS,
  250. ctrmm_olnucopyTS, ctrmm_olnncopyTS, ctrmm_oltucopyTS, ctrmm_oltncopyTS,
  251. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  252. csymm_iutcopyTS, csymm_iltcopyTS,
  253. #else
  254. csymm_outcopyTS, csymm_oltcopyTS,
  255. #endif
  256. csymm_outcopyTS, csymm_oltcopyTS,
  257. #if CGEMM_DEFAULT_UNROLL_M != CGEMM_DEFAULT_UNROLL_N
  258. chemm_iutcopyTS, chemm_iltcopyTS,
  259. #else
  260. chemm_outcopyTS, chemm_oltcopyTS,
  261. #endif
  262. chemm_outcopyTS, chemm_oltcopyTS,
  263. 0, 0, 0,
  264. #ifdef CGEMM3M_DEFAULT_UNROLL_M
  265. CGEMM3M_DEFAULT_UNROLL_M, CGEMM3M_DEFAULT_UNROLL_N, MAX(CGEMM3M_DEFAULT_UNROLL_M, CGEMM3M_DEFAULT_UNROLL_N),
  266. #else
  267. SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N, MAX(SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N),
  268. #endif
  269. cgemm3m_kernelTS,
  270. cgemm3m_incopybTS, cgemm3m_incopyrTS,
  271. cgemm3m_incopyiTS, cgemm3m_itcopybTS,
  272. cgemm3m_itcopyrTS, cgemm3m_itcopyiTS,
  273. cgemm3m_oncopybTS, cgemm3m_oncopyrTS,
  274. cgemm3m_oncopyiTS, cgemm3m_otcopybTS,
  275. cgemm3m_otcopyrTS, cgemm3m_otcopyiTS,
  276. csymm3m_iucopybTS, csymm3m_ilcopybTS,
  277. csymm3m_iucopyrTS, csymm3m_ilcopyrTS,
  278. csymm3m_iucopyiTS, csymm3m_ilcopyiTS,
  279. csymm3m_oucopybTS, csymm3m_olcopybTS,
  280. csymm3m_oucopyrTS, csymm3m_olcopyrTS,
  281. csymm3m_oucopyiTS, csymm3m_olcopyiTS,
  282. chemm3m_iucopybTS, chemm3m_ilcopybTS,
  283. chemm3m_iucopyrTS, chemm3m_ilcopyrTS,
  284. chemm3m_iucopyiTS, chemm3m_ilcopyiTS,
  285. chemm3m_oucopybTS, chemm3m_olcopybTS,
  286. chemm3m_oucopyrTS, chemm3m_olcopyrTS,
  287. chemm3m_oucopyiTS, chemm3m_olcopyiTS,
  288. #ifndef NO_LAPACK
  289. cneg_tcopyTS, claswp_ncopyTS,
  290. #else
  291. NULL, NULL,
  292. #endif
  293. 0, 0, 0,
  294. ZGEMM_DEFAULT_UNROLL_M, ZGEMM_DEFAULT_UNROLL_N,
  295. #ifdef ZGEMM_DEFAULT_UNROLL_MN
  296. ZGEMM_DEFAULT_UNROLL_MN,
  297. #else
  298. MAX(ZGEMM_DEFAULT_UNROLL_M, ZGEMM_DEFAULT_UNROLL_N),
  299. #endif
  300. zamax_kTS, zamin_kTS, izamax_kTS, izamin_kTS,
  301. znrm2_kTS, zasum_kTS, zcopy_kTS,
  302. zdotu_kTS, zdotc_kTS, zdrot_kTS,
  303. zaxpy_kTS, zaxpyc_kTS, zscal_kTS, zswap_kTS,
  304. zgemv_nTS, zgemv_tTS, zgemv_rTS, zgemv_cTS,
  305. zgemv_oTS, zgemv_uTS, zgemv_sTS, zgemv_dTS,
  306. zgeru_kTS, zgerc_kTS, zgerv_kTS, zgerd_kTS,
  307. zsymv_LTS, zsymv_UTS,
  308. zhemv_LTS, zhemv_UTS, zhemv_MTS, zhemv_VTS,
  309. zgemm_kernel_nTS, zgemm_kernel_lTS, zgemm_kernel_rTS, zgemm_kernel_bTS,
  310. zgemm_betaTS,
  311. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  312. zgemm_incopyTS, zgemm_itcopyTS,
  313. #else
  314. zgemm_oncopyTS, zgemm_otcopyTS,
  315. #endif
  316. zgemm_oncopyTS, zgemm_otcopyTS,
  317. ztrsm_kernel_LNTS, ztrsm_kernel_LTTS, ztrsm_kernel_LRTS, ztrsm_kernel_LCTS,
  318. ztrsm_kernel_RNTS, ztrsm_kernel_RTTS, ztrsm_kernel_RRTS, ztrsm_kernel_RCTS,
  319. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  320. ztrsm_iunucopyTS, ztrsm_iunncopyTS, ztrsm_iutucopyTS, ztrsm_iutncopyTS,
  321. ztrsm_ilnucopyTS, ztrsm_ilnncopyTS, ztrsm_iltucopyTS, ztrsm_iltncopyTS,
  322. #else
  323. ztrsm_ounucopyTS, ztrsm_ounncopyTS, ztrsm_outucopyTS, ztrsm_outncopyTS,
  324. ztrsm_olnucopyTS, ztrsm_olnncopyTS, ztrsm_oltucopyTS, ztrsm_oltncopyTS,
  325. #endif
  326. ztrsm_ounucopyTS, ztrsm_ounncopyTS, ztrsm_outucopyTS, ztrsm_outncopyTS,
  327. ztrsm_olnucopyTS, ztrsm_olnncopyTS, ztrsm_oltucopyTS, ztrsm_oltncopyTS,
  328. ztrmm_kernel_RNTS, ztrmm_kernel_RTTS, ztrmm_kernel_RRTS, ztrmm_kernel_RCTS,
  329. ztrmm_kernel_LNTS, ztrmm_kernel_LTTS, ztrmm_kernel_LRTS, ztrmm_kernel_LCTS,
  330. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  331. ztrmm_iunucopyTS, ztrmm_iunncopyTS, ztrmm_iutucopyTS, ztrmm_iutncopyTS,
  332. ztrmm_ilnucopyTS, ztrmm_ilnncopyTS, ztrmm_iltucopyTS, ztrmm_iltncopyTS,
  333. #else
  334. ztrmm_ounucopyTS, ztrmm_ounncopyTS, ztrmm_outucopyTS, ztrmm_outncopyTS,
  335. ztrmm_olnucopyTS, ztrmm_olnncopyTS, ztrmm_oltucopyTS, ztrmm_oltncopyTS,
  336. #endif
  337. ztrmm_ounucopyTS, ztrmm_ounncopyTS, ztrmm_outucopyTS, ztrmm_outncopyTS,
  338. ztrmm_olnucopyTS, ztrmm_olnncopyTS, ztrmm_oltucopyTS, ztrmm_oltncopyTS,
  339. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  340. zsymm_iutcopyTS, zsymm_iltcopyTS,
  341. #else
  342. zsymm_outcopyTS, zsymm_oltcopyTS,
  343. #endif
  344. zsymm_outcopyTS, zsymm_oltcopyTS,
  345. #if ZGEMM_DEFAULT_UNROLL_M != ZGEMM_DEFAULT_UNROLL_N
  346. zhemm_iutcopyTS, zhemm_iltcopyTS,
  347. #else
  348. zhemm_outcopyTS, zhemm_oltcopyTS,
  349. #endif
  350. zhemm_outcopyTS, zhemm_oltcopyTS,
  351. 0, 0, 0,
  352. #ifdef ZGEMM3M_DEFAULT_UNROLL_M
  353. ZGEMM3M_DEFAULT_UNROLL_M, ZGEMM3M_DEFAULT_UNROLL_N, MAX(ZGEMM3M_DEFAULT_UNROLL_M, ZGEMM3M_DEFAULT_UNROLL_N),
  354. #else
  355. DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N, MAX(DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N),
  356. #endif
  357. zgemm3m_kernelTS,
  358. zgemm3m_incopybTS, zgemm3m_incopyrTS,
  359. zgemm3m_incopyiTS, zgemm3m_itcopybTS,
  360. zgemm3m_itcopyrTS, zgemm3m_itcopyiTS,
  361. zgemm3m_oncopybTS, zgemm3m_oncopyrTS,
  362. zgemm3m_oncopyiTS, zgemm3m_otcopybTS,
  363. zgemm3m_otcopyrTS, zgemm3m_otcopyiTS,
  364. zsymm3m_iucopybTS, zsymm3m_ilcopybTS,
  365. zsymm3m_iucopyrTS, zsymm3m_ilcopyrTS,
  366. zsymm3m_iucopyiTS, zsymm3m_ilcopyiTS,
  367. zsymm3m_oucopybTS, zsymm3m_olcopybTS,
  368. zsymm3m_oucopyrTS, zsymm3m_olcopyrTS,
  369. zsymm3m_oucopyiTS, zsymm3m_olcopyiTS,
  370. zhemm3m_iucopybTS, zhemm3m_ilcopybTS,
  371. zhemm3m_iucopyrTS, zhemm3m_ilcopyrTS,
  372. zhemm3m_iucopyiTS, zhemm3m_ilcopyiTS,
  373. zhemm3m_oucopybTS, zhemm3m_olcopybTS,
  374. zhemm3m_oucopyrTS, zhemm3m_olcopyrTS,
  375. zhemm3m_oucopyiTS, zhemm3m_olcopyiTS,
  376. #ifndef NO_LAPACK
  377. zneg_tcopyTS, zlaswp_ncopyTS,
  378. #else
  379. NULL, NULL,
  380. #endif
  381. #ifdef EXPRECISION
  382. 0, 0, 0,
  383. XGEMM_DEFAULT_UNROLL_M, XGEMM_DEFAULT_UNROLL_N, MAX(XGEMM_DEFAULT_UNROLL_M, XGEMM_DEFAULT_UNROLL_N),
  384. xamax_kTS, xamin_kTS, ixamax_kTS, ixamin_kTS,
  385. xnrm2_kTS, xasum_kTS, xcopy_kTS,
  386. xdotu_kTS, xdotc_kTS, xqrot_kTS,
  387. xaxpy_kTS, xaxpyc_kTS, xscal_kTS, xswap_kTS,
  388. xgemv_nTS, xgemv_tTS, xgemv_rTS, xgemv_cTS,
  389. xgemv_oTS, xgemv_uTS, xgemv_sTS, xgemv_dTS,
  390. xgeru_kTS, xgerc_kTS, xgerv_kTS, xgerd_kTS,
  391. xsymv_LTS, xsymv_UTS,
  392. xhemv_LTS, xhemv_UTS, xhemv_MTS, xhemv_VTS,
  393. xgemm_kernel_nTS, xgemm_kernel_lTS, xgemm_kernel_rTS, xgemm_kernel_bTS,
  394. xgemm_betaTS,
  395. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  396. xgemm_incopyTS, xgemm_itcopyTS,
  397. #else
  398. xgemm_oncopyTS, xgemm_otcopyTS,
  399. #endif
  400. xgemm_oncopyTS, xgemm_otcopyTS,
  401. xtrsm_kernel_LNTS, xtrsm_kernel_LTTS, xtrsm_kernel_LRTS, xtrsm_kernel_LCTS,
  402. xtrsm_kernel_RNTS, xtrsm_kernel_RTTS, xtrsm_kernel_RRTS, xtrsm_kernel_RCTS,
  403. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  404. xtrsm_iunucopyTS, xtrsm_iunncopyTS, xtrsm_iutucopyTS, xtrsm_iutncopyTS,
  405. xtrsm_ilnucopyTS, xtrsm_ilnncopyTS, xtrsm_iltucopyTS, xtrsm_iltncopyTS,
  406. #else
  407. xtrsm_ounucopyTS, xtrsm_ounncopyTS, xtrsm_outucopyTS, xtrsm_outncopyTS,
  408. xtrsm_olnucopyTS, xtrsm_olnncopyTS, xtrsm_oltucopyTS, xtrsm_oltncopyTS,
  409. #endif
  410. xtrsm_ounucopyTS, xtrsm_ounncopyTS, xtrsm_outucopyTS, xtrsm_outncopyTS,
  411. xtrsm_olnucopyTS, xtrsm_olnncopyTS, xtrsm_oltucopyTS, xtrsm_oltncopyTS,
  412. xtrmm_kernel_RNTS, xtrmm_kernel_RTTS, xtrmm_kernel_RRTS, xtrmm_kernel_RCTS,
  413. xtrmm_kernel_LNTS, xtrmm_kernel_LTTS, xtrmm_kernel_LRTS, xtrmm_kernel_LCTS,
  414. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  415. xtrmm_iunucopyTS, xtrmm_iunncopyTS, xtrmm_iutucopyTS, xtrmm_iutncopyTS,
  416. xtrmm_ilnucopyTS, xtrmm_ilnncopyTS, xtrmm_iltucopyTS, xtrmm_iltncopyTS,
  417. #else
  418. xtrmm_ounucopyTS, xtrmm_ounncopyTS, xtrmm_outucopyTS, xtrmm_outncopyTS,
  419. xtrmm_olnucopyTS, xtrmm_olnncopyTS, xtrmm_oltucopyTS, xtrmm_oltncopyTS,
  420. #endif
  421. xtrmm_ounucopyTS, xtrmm_ounncopyTS, xtrmm_outucopyTS, xtrmm_outncopyTS,
  422. xtrmm_olnucopyTS, xtrmm_olnncopyTS, xtrmm_oltucopyTS, xtrmm_oltncopyTS,
  423. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  424. xsymm_iutcopyTS, xsymm_iltcopyTS,
  425. #else
  426. xsymm_outcopyTS, xsymm_oltcopyTS,
  427. #endif
  428. xsymm_outcopyTS, xsymm_oltcopyTS,
  429. #if XGEMM_DEFAULT_UNROLL_M != XGEMM_DEFAULT_UNROLL_N
  430. xhemm_iutcopyTS, xhemm_iltcopyTS,
  431. #else
  432. xhemm_outcopyTS, xhemm_oltcopyTS,
  433. #endif
  434. xhemm_outcopyTS, xhemm_oltcopyTS,
  435. 0, 0, 0,
  436. QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N, MAX(QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N),
  437. xgemm3m_kernelTS,
  438. xgemm3m_incopybTS, xgemm3m_incopyrTS,
  439. xgemm3m_incopyiTS, xgemm3m_itcopybTS,
  440. xgemm3m_itcopyrTS, xgemm3m_itcopyiTS,
  441. xgemm3m_oncopybTS, xgemm3m_oncopyrTS,
  442. xgemm3m_oncopyiTS, xgemm3m_otcopybTS,
  443. xgemm3m_otcopyrTS, xgemm3m_otcopyiTS,
  444. xsymm3m_iucopybTS, xsymm3m_ilcopybTS,
  445. xsymm3m_iucopyrTS, xsymm3m_ilcopyrTS,
  446. xsymm3m_iucopyiTS, xsymm3m_ilcopyiTS,
  447. xsymm3m_oucopybTS, xsymm3m_olcopybTS,
  448. xsymm3m_oucopyrTS, xsymm3m_olcopyrTS,
  449. xsymm3m_oucopyiTS, xsymm3m_olcopyiTS,
  450. xhemm3m_iucopybTS, xhemm3m_ilcopybTS,
  451. xhemm3m_iucopyrTS, xhemm3m_ilcopyrTS,
  452. xhemm3m_iucopyiTS, xhemm3m_ilcopyiTS,
  453. xhemm3m_oucopybTS, xhemm3m_olcopybTS,
  454. xhemm3m_oucopyrTS, xhemm3m_olcopyrTS,
  455. xhemm3m_oucopyiTS, xhemm3m_olcopyiTS,
  456. #ifndef NO_LAPACK
  457. xneg_tcopyTS, xlaswp_ncopyTS,
  458. #else
  459. NULL, NULL,
  460. #endif
  461. #endif
  462. init_parameter,
  463. SNUMOPT, DNUMOPT, QNUMOPT,
  464. saxpby_kTS, daxpby_kTS, caxpby_kTS, zaxpby_kTS,
  465. somatcopy_k_cnTS, somatcopy_k_ctTS, somatcopy_k_rnTS, somatcopy_k_rtTS,
  466. domatcopy_k_cnTS, domatcopy_k_ctTS, domatcopy_k_rnTS, domatcopy_k_rtTS,
  467. comatcopy_k_cnTS, comatcopy_k_ctTS, comatcopy_k_rnTS, comatcopy_k_rtTS,
  468. comatcopy_k_cncTS, comatcopy_k_ctcTS, comatcopy_k_rncTS, comatcopy_k_rtcTS,
  469. zomatcopy_k_cnTS, zomatcopy_k_ctTS, zomatcopy_k_rnTS, zomatcopy_k_rtTS,
  470. zomatcopy_k_cncTS, zomatcopy_k_ctcTS, zomatcopy_k_rncTS, zomatcopy_k_rtcTS,
  471. sgeadd_kTS, dgeadd_kTS, cgeadd_kTS, zgeadd_kTS
  472. };
  473. #ifdef ARCH_X86
  474. static int get_l2_size_old(void){
  475. int i, eax, ebx, ecx, edx, cpuid_level;
  476. int info[15];
  477. cpuid(2, &eax, &ebx, &ecx, &edx);
  478. info[ 0] = BITMASK(eax, 8, 0xff);
  479. info[ 1] = BITMASK(eax, 16, 0xff);
  480. info[ 2] = BITMASK(eax, 24, 0xff);
  481. info[ 3] = BITMASK(ebx, 0, 0xff);
  482. info[ 4] = BITMASK(ebx, 8, 0xff);
  483. info[ 5] = BITMASK(ebx, 16, 0xff);
  484. info[ 6] = BITMASK(ebx, 24, 0xff);
  485. info[ 7] = BITMASK(ecx, 0, 0xff);
  486. info[ 8] = BITMASK(ecx, 8, 0xff);
  487. info[ 9] = BITMASK(ecx, 16, 0xff);
  488. info[10] = BITMASK(ecx, 24, 0xff);
  489. info[11] = BITMASK(edx, 0, 0xff);
  490. info[12] = BITMASK(edx, 8, 0xff);
  491. info[13] = BITMASK(edx, 16, 0xff);
  492. info[14] = BITMASK(edx, 24, 0xff);
  493. for (i = 0; i < 15; i++){
  494. switch (info[i]){
  495. /* This table is from http://www.sandpile.org/ia32/cpuid.htm */
  496. case 0x1a :
  497. return 96;
  498. case 0x39 :
  499. case 0x3b :
  500. case 0x41 :
  501. case 0x79 :
  502. case 0x81 :
  503. return 128;
  504. case 0x3a :
  505. return 192;
  506. case 0x21 :
  507. case 0x3c :
  508. case 0x42 :
  509. case 0x7a :
  510. case 0x7e :
  511. case 0x82 :
  512. return 256;
  513. case 0x3d :
  514. return 384;
  515. case 0x3e :
  516. case 0x43 :
  517. case 0x7b :
  518. case 0x7f :
  519. case 0x83 :
  520. case 0x86 :
  521. return 512;
  522. case 0x44 :
  523. case 0x78 :
  524. case 0x7c :
  525. case 0x84 :
  526. case 0x87 :
  527. return 1024;
  528. case 0x45 :
  529. case 0x7d :
  530. case 0x85 :
  531. return 2048;
  532. case 0x48 :
  533. return 3184;
  534. case 0x49 :
  535. return 4096;
  536. case 0x4e :
  537. return 6144;
  538. }
  539. }
  540. return 0;
  541. }
  542. #endif
  543. static __inline__ int get_l2_size(void){
  544. int eax, ebx, ecx, edx, l2;
  545. cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
  546. l2 = BITMASK(ecx, 16, 0xffff);
  547. #ifndef ARCH_X86
  548. return l2;
  549. #else
  550. if (l2 > 0) return l2;
  551. return get_l2_size_old();
  552. #endif
  553. }
  554. static __inline__ int get_l3_size(void){
  555. int eax, ebx, ecx, edx;
  556. cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
  557. return BITMASK(edx, 18, 0x3fff) * 512;
  558. }
  559. static void init_parameter(void) {
  560. int l2 = get_l2_size();
  561. TABLE_NAME.sgemm_q = SGEMM_DEFAULT_Q;
  562. TABLE_NAME.dgemm_q = DGEMM_DEFAULT_Q;
  563. TABLE_NAME.cgemm_q = CGEMM_DEFAULT_Q;
  564. TABLE_NAME.zgemm_q = ZGEMM_DEFAULT_Q;
  565. #ifdef CGEMM3M_DEFAULT_Q
  566. TABLE_NAME.cgemm3m_q = CGEMM3M_DEFAULT_Q;
  567. #else
  568. TABLE_NAME.cgemm3m_q = SGEMM_DEFAULT_Q;
  569. #endif
  570. #ifdef ZGEMM3M_DEFAULT_Q
  571. TABLE_NAME.zgemm3m_q = ZGEMM3M_DEFAULT_Q;
  572. #else
  573. TABLE_NAME.zgemm3m_q = DGEMM_DEFAULT_Q;
  574. #endif
  575. #ifdef EXPRECISION
  576. TABLE_NAME.qgemm_q = QGEMM_DEFAULT_Q;
  577. TABLE_NAME.xgemm_q = XGEMM_DEFAULT_Q;
  578. TABLE_NAME.xgemm3m_q = QGEMM_DEFAULT_Q;
  579. #endif
  580. #if defined(CORE_KATMAI) || defined(CORE_COPPERMINE) || defined(CORE_BANIAS) || defined(CORE_YONAH) || defined(CORE_ATHLON)
  581. #ifdef DEBUG
  582. fprintf(stderr, "Katmai, Coppermine, Banias, Athlon\n");
  583. #endif
  584. TABLE_NAME.sgemm_p = 64 * (l2 >> 7);
  585. TABLE_NAME.dgemm_p = 32 * (l2 >> 7);
  586. TABLE_NAME.cgemm_p = 32 * (l2 >> 7);
  587. TABLE_NAME.zgemm_p = 16 * (l2 >> 7);
  588. #ifdef EXPRECISION
  589. TABLE_NAME.qgemm_p = 16 * (l2 >> 7);
  590. TABLE_NAME.xgemm_p = 8 * (l2 >> 7);
  591. #endif
  592. #endif
  593. #ifdef CORE_NORTHWOOD
  594. #ifdef DEBUG
  595. fprintf(stderr, "Northwood\n");
  596. #endif
  597. TABLE_NAME.sgemm_p = 96 * (l2 >> 7);
  598. TABLE_NAME.dgemm_p = 48 * (l2 >> 7);
  599. TABLE_NAME.cgemm_p = 48 * (l2 >> 7);
  600. TABLE_NAME.zgemm_p = 24 * (l2 >> 7);
  601. #ifdef EXPRECISION
  602. TABLE_NAME.qgemm_p = 24 * (l2 >> 7);
  603. TABLE_NAME.xgemm_p = 12 * (l2 >> 7);
  604. #endif
  605. #endif
  606. #ifdef ATOM
  607. #ifdef DEBUG
  608. fprintf(stderr, "Atom\n");
  609. #endif
  610. TABLE_NAME.sgemm_p = 256;
  611. TABLE_NAME.dgemm_p = 128;
  612. TABLE_NAME.cgemm_p = 128;
  613. TABLE_NAME.zgemm_p = 64;
  614. #ifdef EXPRECISION
  615. TABLE_NAME.qgemm_p = 64;
  616. TABLE_NAME.xgemm_p = 32;
  617. #endif
  618. #endif
  619. #ifdef CORE_PRESCOTT
  620. #ifdef DEBUG
  621. fprintf(stderr, "Prescott\n");
  622. #endif
  623. TABLE_NAME.sgemm_p = 56 * (l2 >> 7);
  624. TABLE_NAME.dgemm_p = 28 * (l2 >> 7);
  625. TABLE_NAME.cgemm_p = 28 * (l2 >> 7);
  626. TABLE_NAME.zgemm_p = 14 * (l2 >> 7);
  627. #ifdef EXPRECISION
  628. TABLE_NAME.qgemm_p = 14 * (l2 >> 7);
  629. TABLE_NAME.xgemm_p = 7 * (l2 >> 7);
  630. #endif
  631. #endif
  632. #ifdef CORE2
  633. #ifdef DEBUG
  634. fprintf(stderr, "Core2\n");
  635. #endif
  636. TABLE_NAME.sgemm_p = 92 * (l2 >> 9) + 8;
  637. TABLE_NAME.dgemm_p = 46 * (l2 >> 9) + 8;
  638. TABLE_NAME.cgemm_p = 46 * (l2 >> 9) + 4;
  639. TABLE_NAME.zgemm_p = 23 * (l2 >> 9) + 4;
  640. #ifdef EXPRECISION
  641. TABLE_NAME.qgemm_p = 92 * (l2 >> 9) + 8;
  642. TABLE_NAME.xgemm_p = 46 * (l2 >> 9) + 4;
  643. #endif
  644. #endif
  645. #ifdef PENRYN
  646. #ifdef DEBUG
  647. fprintf(stderr, "Penryn\n");
  648. #endif
  649. TABLE_NAME.sgemm_p = 42 * (l2 >> 9) + 8;
  650. TABLE_NAME.dgemm_p = 42 * (l2 >> 9) + 8;
  651. TABLE_NAME.cgemm_p = 21 * (l2 >> 9) + 4;
  652. TABLE_NAME.zgemm_p = 21 * (l2 >> 9) + 4;
  653. #ifdef EXPRECISION
  654. TABLE_NAME.qgemm_p = 42 * (l2 >> 9) + 8;
  655. TABLE_NAME.xgemm_p = 21 * (l2 >> 9) + 4;
  656. #endif
  657. #endif
  658. #ifdef DUNNINGTON
  659. #ifdef DEBUG
  660. fprintf(stderr, "Dunnington\n");
  661. #endif
  662. TABLE_NAME.sgemm_p = 42 * (l2 >> 9) + 8;
  663. TABLE_NAME.dgemm_p = 42 * (l2 >> 9) + 8;
  664. TABLE_NAME.cgemm_p = 21 * (l2 >> 9) + 4;
  665. TABLE_NAME.zgemm_p = 21 * (l2 >> 9) + 4;
  666. #ifdef EXPRECISION
  667. TABLE_NAME.qgemm_p = 42 * (l2 >> 9) + 8;
  668. TABLE_NAME.xgemm_p = 21 * (l2 >> 9) + 4;
  669. #endif
  670. #endif
  671. #ifdef NEHALEM
  672. #ifdef DEBUG
  673. fprintf(stderr, "Nehalem\n");
  674. #endif
  675. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  676. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  677. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  678. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  679. #ifdef EXPRECISION
  680. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  681. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  682. #endif
  683. #endif
  684. #ifdef SANDYBRIDGE
  685. #ifdef DEBUG
  686. fprintf(stderr, "Sandybridge\n");
  687. #endif
  688. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  689. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  690. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  691. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  692. #ifdef EXPRECISION
  693. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  694. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  695. #endif
  696. #endif
  697. #ifdef HASWELL
  698. #ifdef DEBUG
  699. fprintf(stderr, "Haswell\n");
  700. #endif
  701. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  702. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  703. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  704. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  705. #ifdef EXPRECISION
  706. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  707. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  708. #endif
  709. #endif
  710. #ifdef OPTERON
  711. #ifdef DEBUG
  712. fprintf(stderr, "Opteron\n");
  713. #endif
  714. TABLE_NAME.sgemm_p = 224 + 56 * (l2 >> 7);
  715. TABLE_NAME.dgemm_p = 112 + 28 * (l2 >> 7);
  716. TABLE_NAME.cgemm_p = 112 + 28 * (l2 >> 7);
  717. TABLE_NAME.zgemm_p = 56 + 14 * (l2 >> 7);
  718. #ifdef EXPRECISION
  719. TABLE_NAME.qgemm_p = 56 + 14 * (l2 >> 7);
  720. TABLE_NAME.xgemm_p = 28 + 7 * (l2 >> 7);
  721. #endif
  722. #endif
  723. #ifdef BARCELONA
  724. #ifdef DEBUG
  725. fprintf(stderr, "Barcelona\n");
  726. #endif
  727. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  728. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  729. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  730. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  731. #ifdef EXPRECISION
  732. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  733. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  734. #endif
  735. #endif
  736. #ifdef BOBCAT
  737. #ifdef DEBUG
  738. fprintf(stderr, "Bobcate\n");
  739. #endif
  740. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  741. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  742. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  743. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  744. #ifdef EXPRECISION
  745. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  746. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  747. #endif
  748. #endif
  749. #ifdef BULLDOZER
  750. #ifdef DEBUG
  751. fprintf(stderr, "Bulldozer\n");
  752. #endif
  753. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  754. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  755. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  756. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  757. #ifdef EXPRECISION
  758. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  759. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  760. #endif
  761. #endif
  762. #ifdef PILEDRIVER
  763. #ifdef DEBUG
  764. fprintf(stderr, "Piledriver\n");
  765. #endif
  766. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  767. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  768. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  769. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  770. #ifdef EXPRECISION
  771. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  772. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  773. #endif
  774. #endif
  775. #ifdef STEAMROLLER
  776. #ifdef DEBUG
  777. fprintf(stderr, "Steamroller\n");
  778. #endif
  779. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  780. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  781. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  782. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  783. #ifdef EXPRECISION
  784. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  785. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  786. #endif
  787. #endif
  788. #ifdef NANO
  789. #ifdef DEBUG
  790. fprintf(stderr, "NANO\n");
  791. #endif
  792. TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
  793. TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
  794. TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
  795. TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
  796. #ifdef EXPRECISION
  797. TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
  798. TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
  799. #endif
  800. #endif
  801. #ifdef CGEMM3M_DEFAULT_P
  802. TABLE_NAME.cgemm3m_p = CGEMM3M_DEFAULT_P;
  803. #else
  804. TABLE_NAME.cgemm3m_p = TABLE_NAME.sgemm_p;
  805. #endif
  806. #ifdef ZGEMM3M_DEFAULT_P
  807. TABLE_NAME.zgemm3m_p = ZGEMM3M_DEFAULT_P;
  808. #else
  809. TABLE_NAME.zgemm3m_p = TABLE_NAME.dgemm_p;
  810. #endif
  811. #ifdef EXPRECISION
  812. TABLE_NAME.xgemm3m_p = TABLE_NAME.qgemm_p;
  813. #endif
  814. TABLE_NAME.sgemm_p = (TABLE_NAME.sgemm_p + SGEMM_DEFAULT_UNROLL_M - 1) & ~(SGEMM_DEFAULT_UNROLL_M - 1);
  815. TABLE_NAME.dgemm_p = (TABLE_NAME.dgemm_p + DGEMM_DEFAULT_UNROLL_M - 1) & ~(DGEMM_DEFAULT_UNROLL_M - 1);
  816. TABLE_NAME.cgemm_p = (TABLE_NAME.cgemm_p + CGEMM_DEFAULT_UNROLL_M - 1) & ~(CGEMM_DEFAULT_UNROLL_M - 1);
  817. TABLE_NAME.zgemm_p = (TABLE_NAME.zgemm_p + ZGEMM_DEFAULT_UNROLL_M - 1) & ~(ZGEMM_DEFAULT_UNROLL_M - 1);
  818. #ifdef CGEMM3M_DEFAULT_UNROLL_M
  819. TABLE_NAME.cgemm3m_p = (TABLE_NAME.cgemm3m_p + CGEMM3M_DEFAULT_UNROLL_M - 1) & ~(CGEMM3M_DEFAULT_UNROLL_M - 1);
  820. #else
  821. TABLE_NAME.cgemm3m_p = (TABLE_NAME.cgemm3m_p + SGEMM_DEFAULT_UNROLL_M - 1) & ~(SGEMM_DEFAULT_UNROLL_M - 1);
  822. #endif
  823. #ifdef ZGEMM3M_DEFAULT_UNROLL_M
  824. TABLE_NAME.zgemm3m_p = (TABLE_NAME.zgemm3m_p + ZGEMM3M_DEFAULT_UNROLL_M - 1) & ~(ZGEMM3M_DEFAULT_UNROLL_M - 1);
  825. #else
  826. TABLE_NAME.zgemm3m_p = (TABLE_NAME.zgemm3m_p + DGEMM_DEFAULT_UNROLL_M - 1) & ~(DGEMM_DEFAULT_UNROLL_M - 1);
  827. #endif
  828. #ifdef QUAD_PRECISION
  829. TABLE_NAME.qgemm_p = (TABLE_NAME.qgemm_p + QGEMM_DEFAULT_UNROLL_M - 1) & ~(QGEMM_DEFAULT_UNROLL_M - 1);
  830. TABLE_NAME.xgemm_p = (TABLE_NAME.xgemm_p + XGEMM_DEFAULT_UNROLL_M - 1) & ~(XGEMM_DEFAULT_UNROLL_M - 1);
  831. TABLE_NAME.xgemm3m_p = (TABLE_NAME.xgemm3m_p + QGEMM_DEFAULT_UNROLL_M - 1) & ~(QGEMM_DEFAULT_UNROLL_M - 1);
  832. #endif
  833. #ifdef DEBUG
  834. fprintf(stderr, "L2 = %8d DGEMM_P .. %d\n", l2, TABLE_NAME.dgemm_p);
  835. #endif
  836. TABLE_NAME.sgemm_r = (((BUFFER_SIZE -
  837. ((TABLE_NAME.sgemm_p * TABLE_NAME.sgemm_q * 4 + TABLE_NAME.offsetA
  838. + TABLE_NAME.align) & ~TABLE_NAME.align)
  839. ) / (TABLE_NAME.sgemm_q * 4) - 15) & ~15);
  840. TABLE_NAME.dgemm_r = (((BUFFER_SIZE -
  841. ((TABLE_NAME.dgemm_p * TABLE_NAME.dgemm_q * 8 + TABLE_NAME.offsetA
  842. + TABLE_NAME.align) & ~TABLE_NAME.align)
  843. ) / (TABLE_NAME.dgemm_q * 8) - 15) & ~15);
  844. #ifdef EXPRECISION
  845. TABLE_NAME.qgemm_r = (((BUFFER_SIZE -
  846. ((TABLE_NAME.qgemm_p * TABLE_NAME.qgemm_q * 16 + TABLE_NAME.offsetA
  847. + TABLE_NAME.align) & ~TABLE_NAME.align)
  848. ) / (TABLE_NAME.qgemm_q * 16) - 15) & ~15);
  849. #endif
  850. TABLE_NAME.cgemm_r = (((BUFFER_SIZE -
  851. ((TABLE_NAME.cgemm_p * TABLE_NAME.cgemm_q * 8 + TABLE_NAME.offsetA
  852. + TABLE_NAME.align) & ~TABLE_NAME.align)
  853. ) / (TABLE_NAME.cgemm_q * 8) - 15) & ~15);
  854. TABLE_NAME.zgemm_r = (((BUFFER_SIZE -
  855. ((TABLE_NAME.zgemm_p * TABLE_NAME.zgemm_q * 16 + TABLE_NAME.offsetA
  856. + TABLE_NAME.align) & ~TABLE_NAME.align)
  857. ) / (TABLE_NAME.zgemm_q * 16) - 15) & ~15);
  858. TABLE_NAME.cgemm3m_r = (((BUFFER_SIZE -
  859. ((TABLE_NAME.cgemm3m_p * TABLE_NAME.cgemm3m_q * 8 + TABLE_NAME.offsetA
  860. + TABLE_NAME.align) & ~TABLE_NAME.align)
  861. ) / (TABLE_NAME.cgemm3m_q * 8) - 15) & ~15);
  862. TABLE_NAME.zgemm3m_r = (((BUFFER_SIZE -
  863. ((TABLE_NAME.zgemm3m_p * TABLE_NAME.zgemm3m_q * 16 + TABLE_NAME.offsetA
  864. + TABLE_NAME.align) & ~TABLE_NAME.align)
  865. ) / (TABLE_NAME.zgemm3m_q * 16) - 15) & ~15);
  866. #ifdef EXPRECISION
  867. TABLE_NAME.xgemm_r = (((BUFFER_SIZE -
  868. ((TABLE_NAME.xgemm_p * TABLE_NAME.xgemm_q * 32 + TABLE_NAME.offsetA
  869. + TABLE_NAME.align) & ~TABLE_NAME.align)
  870. ) / (TABLE_NAME.xgemm_q * 32) - 15) & ~15);
  871. TABLE_NAME.xgemm3m_r = (((BUFFER_SIZE -
  872. ((TABLE_NAME.xgemm3m_p * TABLE_NAME.xgemm3m_q * 32 + TABLE_NAME.offsetA
  873. + TABLE_NAME.align) & ~TABLE_NAME.align)
  874. ) / (TABLE_NAME.xgemm3m_q * 32) - 15) & ~15);
  875. #endif
  876. }