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.

common_param.h 93 kB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417
  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. #ifndef COMMON_PARAM_H
  39. #define COMMON_PARAM_H
  40. #ifndef ASSEMBLER
  41. #ifdef DYNAMIC_ARCH
  42. typedef struct {
  43. int dtb_entries;
  44. int offsetA, offsetB, align;
  45. #ifdef BUILD_HALF
  46. int shgemm_p, shgemm_q, shgemm_r;
  47. int shgemm_unroll_m, shgemm_unroll_n, shgemm_unroll_mn;
  48. void (*shstobf16_k) (BLASLONG, float *, BLASLONG, bfloat16 *, BLASLONG);
  49. void (*shdtobf16_k) (BLASLONG, double *, BLASLONG, bfloat16 *, BLASLONG);
  50. void (*sbf16tos_k) (BLASLONG, bfloat16 *, BLASLONG, float *, BLASLONG);
  51. void (*dbf16tod_k) (BLASLONG, bfloat16 *, BLASLONG, double *, BLASLONG);
  52. float (*shamax_k) (BLASLONG, float *, BLASLONG);
  53. float (*shamin_k) (BLASLONG, float *, BLASLONG);
  54. float (*shmax_k) (BLASLONG, float *, BLASLONG);
  55. float (*shmin_k) (BLASLONG, float *, BLASLONG);
  56. BLASLONG (*ishamax_k)(BLASLONG, float *, BLASLONG);
  57. BLASLONG (*ishamin_k)(BLASLONG, float *, BLASLONG);
  58. BLASLONG (*ishmax_k) (BLASLONG, float *, BLASLONG);
  59. BLASLONG (*ishmin_k) (BLASLONG, float *, BLASLONG);
  60. float (*shnrm2_k) (BLASLONG, float *, BLASLONG);
  61. float (*shasum_k) (BLASLONG, float *, BLASLONG);
  62. float (*shsum_k) (BLASLONG, float *, BLASLONG);
  63. int (*shcopy_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  64. float (*shdot_k) (BLASLONG, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG);
  65. double (*dshdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  66. int (*shrot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG, float, float);
  67. int (*shaxpy_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  68. int (*shscal_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  69. int (*shswap_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  70. int (*shgemv_n) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  71. int (*shgemv_t) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  72. int (*shger_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  73. int (*shsymv_L) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  74. int (*shsymv_U) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  75. int (*shgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, float, bfloat16 *, bfloat16 *, float *, BLASLONG);
  76. int (*shgemm_beta )(BLASLONG, BLASLONG, BLASLONG, float, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, float *, BLASLONG);
  77. int (*shgemm_incopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
  78. int (*shgemm_itcopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
  79. int (*shgemm_oncopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
  80. int (*shgemm_otcopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
  81. int (*shtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  82. int (*shtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  83. int (*shtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  84. int (*shtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  85. int (*shtrsm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  86. int (*shtrsm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  87. int (*shtrsm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  88. int (*shtrsm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  89. int (*shtrsm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  90. int (*shtrsm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  91. int (*shtrsm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  92. int (*shtrsm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  93. int (*shtrsm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  94. int (*shtrsm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  95. int (*shtrsm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  96. int (*shtrsm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  97. int (*shtrsm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  98. int (*shtrsm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  99. int (*shtrsm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  100. int (*shtrsm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  101. int (*shtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  102. int (*shtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  103. int (*shtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  104. int (*shtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  105. int (*shtrmm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  106. int (*shtrmm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  107. int (*shtrmm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  108. int (*shtrmm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  109. int (*shtrmm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  110. int (*shtrmm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  111. int (*shtrmm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  112. int (*shtrmm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  113. int (*shtrmm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  114. int (*shtrmm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  115. int (*shtrmm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  116. int (*shtrmm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  117. int (*shtrmm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  118. int (*shtrmm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  119. int (*shtrmm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  120. int (*shtrmm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  121. int (*shsymm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  122. int (*shsymm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  123. int (*shsymm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  124. int (*shsymm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  125. int (*shneg_tcopy) (BLASLONG, BLASLONG, float *, BLASLONG, float *);
  126. int (*shlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG, blasint *, float *);
  127. #endif
  128. int sgemm_p, sgemm_q, sgemm_r;
  129. int sgemm_unroll_m, sgemm_unroll_n, sgemm_unroll_mn;
  130. int exclusive_cache;
  131. float (*samax_k) (BLASLONG, float *, BLASLONG);
  132. float (*samin_k) (BLASLONG, float *, BLASLONG);
  133. float (*smax_k) (BLASLONG, float *, BLASLONG);
  134. float (*smin_k) (BLASLONG, float *, BLASLONG);
  135. BLASLONG (*isamax_k)(BLASLONG, float *, BLASLONG);
  136. BLASLONG (*isamin_k)(BLASLONG, float *, BLASLONG);
  137. BLASLONG (*ismax_k) (BLASLONG, float *, BLASLONG);
  138. BLASLONG (*ismin_k) (BLASLONG, float *, BLASLONG);
  139. float (*snrm2_k) (BLASLONG, float *, BLASLONG);
  140. float (*sasum_k) (BLASLONG, float *, BLASLONG);
  141. float (*ssum_k) (BLASLONG, float *, BLASLONG);
  142. int (*scopy_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  143. float (*sdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  144. double (*dsdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  145. int (*srot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG, float, float);
  146. int (*saxpy_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  147. int (*sscal_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  148. int (*sswap_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  149. int (*sgemv_n) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  150. int (*sgemv_t) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  151. int (*sger_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  152. int (*ssymv_L) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  153. int (*ssymv_U) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  154. #ifdef ARCH_X86_64
  155. void (*sgemm_direct) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG , float *, BLASLONG , float * , BLASLONG);
  156. int (*sgemm_direct_performant) (BLASLONG M, BLASLONG N, BLASLONG K);
  157. #endif
  158. int (*sgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG);
  159. int (*sgemm_beta )(BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  160. int (*sgemm_incopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  161. int (*sgemm_itcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  162. int (*sgemm_oncopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  163. int (*sgemm_otcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  164. int (*strsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  165. int (*strsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  166. int (*strsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  167. int (*strsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  168. int (*strsm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  169. int (*strsm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  170. int (*strsm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  171. int (*strsm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  172. int (*strsm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  173. int (*strsm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  174. int (*strsm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  175. int (*strsm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  176. int (*strsm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  177. int (*strsm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  178. int (*strsm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  179. int (*strsm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  180. int (*strsm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  181. int (*strsm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  182. int (*strsm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  183. int (*strsm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  184. int (*strmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  185. int (*strmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  186. int (*strmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  187. int (*strmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  188. int (*strmm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  189. int (*strmm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  190. int (*strmm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  191. int (*strmm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  192. int (*strmm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  193. int (*strmm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  194. int (*strmm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  195. int (*strmm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  196. int (*strmm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  197. int (*strmm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  198. int (*strmm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  199. int (*strmm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  200. int (*strmm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  201. int (*strmm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  202. int (*strmm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  203. int (*strmm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  204. int (*ssymm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  205. int (*ssymm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  206. int (*ssymm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  207. int (*ssymm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  208. int (*sneg_tcopy) (BLASLONG, BLASLONG, float *, BLASLONG, float *);
  209. int (*slaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG, blasint *, float *);
  210. int dgemm_p, dgemm_q, dgemm_r;
  211. int dgemm_unroll_m, dgemm_unroll_n, dgemm_unroll_mn;
  212. double (*damax_k) (BLASLONG, double *, BLASLONG);
  213. double (*damin_k) (BLASLONG, double *, BLASLONG);
  214. double (*dmax_k) (BLASLONG, double *, BLASLONG);
  215. double (*dmin_k) (BLASLONG, double *, BLASLONG);
  216. BLASLONG (*idamax_k)(BLASLONG, double *, BLASLONG);
  217. BLASLONG (*idamin_k)(BLASLONG, double *, BLASLONG);
  218. BLASLONG (*idmax_k) (BLASLONG, double *, BLASLONG);
  219. BLASLONG (*idmin_k) (BLASLONG, double *, BLASLONG);
  220. double (*dnrm2_k) (BLASLONG, double *, BLASLONG);
  221. double (*dasum_k) (BLASLONG, double *, BLASLONG);
  222. double (*dsum_k) (BLASLONG, double *, BLASLONG);
  223. int (*dcopy_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
  224. double (*ddot_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
  225. int (*drot_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG, double, double);
  226. int (*daxpy_k) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  227. int (*dscal_k) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  228. int (*dswap_k) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  229. int (*dgemv_n) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  230. int (*dgemv_t) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  231. int (*dger_k) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  232. int (*dsymv_L) (BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  233. int (*dsymv_U) (BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  234. int (*dgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG);
  235. int (*dgemm_beta )(BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  236. int (*dgemm_incopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  237. int (*dgemm_itcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  238. int (*dgemm_oncopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  239. int (*dgemm_otcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  240. int (*dtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  241. int (*dtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  242. int (*dtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  243. int (*dtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  244. int (*dtrsm_iunucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  245. int (*dtrsm_iunncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  246. int (*dtrsm_iutucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  247. int (*dtrsm_iutncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  248. int (*dtrsm_ilnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  249. int (*dtrsm_ilnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  250. int (*dtrsm_iltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  251. int (*dtrsm_iltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  252. int (*dtrsm_ounucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  253. int (*dtrsm_ounncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  254. int (*dtrsm_outucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  255. int (*dtrsm_outncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  256. int (*dtrsm_olnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  257. int (*dtrsm_olnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  258. int (*dtrsm_oltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  259. int (*dtrsm_oltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  260. int (*dtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  261. int (*dtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  262. int (*dtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  263. int (*dtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  264. int (*dtrmm_iunucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  265. int (*dtrmm_iunncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  266. int (*dtrmm_iutucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  267. int (*dtrmm_iutncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  268. int (*dtrmm_ilnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  269. int (*dtrmm_ilnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  270. int (*dtrmm_iltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  271. int (*dtrmm_iltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  272. int (*dtrmm_ounucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  273. int (*dtrmm_ounncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  274. int (*dtrmm_outucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  275. int (*dtrmm_outncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  276. int (*dtrmm_olnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  277. int (*dtrmm_olnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  278. int (*dtrmm_oltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  279. int (*dtrmm_oltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  280. int (*dsymm_iutcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  281. int (*dsymm_iltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  282. int (*dsymm_outcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  283. int (*dsymm_oltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  284. int (*dneg_tcopy) (BLASLONG, BLASLONG, double *, BLASLONG, double *);
  285. int (*dlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, double *, BLASLONG, blasint *, double *);
  286. #ifdef EXPRECISION
  287. int qgemm_p, qgemm_q, qgemm_r;
  288. int qgemm_unroll_m, qgemm_unroll_n, qgemm_unroll_mn;
  289. xdouble (*qamax_k) (BLASLONG, xdouble *, BLASLONG);
  290. xdouble (*qamin_k) (BLASLONG, xdouble *, BLASLONG);
  291. xdouble (*qmax_k) (BLASLONG, xdouble *, BLASLONG);
  292. xdouble (*qmin_k) (BLASLONG, xdouble *, BLASLONG);
  293. BLASLONG (*iqamax_k)(BLASLONG, xdouble *, BLASLONG);
  294. BLASLONG (*iqamin_k)(BLASLONG, xdouble *, BLASLONG);
  295. BLASLONG (*iqmax_k) (BLASLONG, xdouble *, BLASLONG);
  296. BLASLONG (*iqmin_k) (BLASLONG, xdouble *, BLASLONG);
  297. xdouble (*qnrm2_k) (BLASLONG, xdouble *, BLASLONG);
  298. xdouble (*qasum_k) (BLASLONG, xdouble *, BLASLONG);
  299. xdouble (*qsum_k) (BLASLONG, xdouble *, BLASLONG);
  300. int (*qcopy_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  301. xdouble (*qdot_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  302. int (*qrot_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble);
  303. int (*qaxpy_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  304. int (*qscal_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  305. int (*qswap_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  306. int (*qgemv_n) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  307. int (*qgemv_t) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  308. int (*qger_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  309. int (*qsymv_L) (BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  310. int (*qsymv_U) (BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  311. int (*qgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  312. int (*qgemm_beta )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  313. int (*qgemm_incopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  314. int (*qgemm_itcopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  315. int (*qgemm_oncopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  316. int (*qgemm_otcopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  317. int (*qtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  318. int (*qtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  319. int (*qtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  320. int (*qtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  321. int (*qtrsm_iunucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  322. int (*qtrsm_iunncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  323. int (*qtrsm_iutucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  324. int (*qtrsm_iutncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  325. int (*qtrsm_ilnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  326. int (*qtrsm_ilnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  327. int (*qtrsm_iltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  328. int (*qtrsm_iltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  329. int (*qtrsm_ounucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  330. int (*qtrsm_ounncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  331. int (*qtrsm_outucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  332. int (*qtrsm_outncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  333. int (*qtrsm_olnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  334. int (*qtrsm_olnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  335. int (*qtrsm_oltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  336. int (*qtrsm_oltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  337. int (*qtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  338. int (*qtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  339. int (*qtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  340. int (*qtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  341. int (*qtrmm_iunucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  342. int (*qtrmm_iunncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  343. int (*qtrmm_iutucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  344. int (*qtrmm_iutncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  345. int (*qtrmm_ilnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  346. int (*qtrmm_ilnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  347. int (*qtrmm_iltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  348. int (*qtrmm_iltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  349. int (*qtrmm_ounucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  350. int (*qtrmm_ounncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  351. int (*qtrmm_outucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  352. int (*qtrmm_outncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  353. int (*qtrmm_olnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  354. int (*qtrmm_olnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  355. int (*qtrmm_oltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  356. int (*qtrmm_oltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  357. int (*qsymm_iutcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  358. int (*qsymm_iltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  359. int (*qsymm_outcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  360. int (*qsymm_oltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  361. int (*qneg_tcopy) (BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  362. int (*qlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, xdouble *, BLASLONG, blasint *, xdouble *);
  363. #endif
  364. int cgemm_p, cgemm_q, cgemm_r;
  365. int cgemm_unroll_m, cgemm_unroll_n, cgemm_unroll_mn;
  366. float (*camax_k) (BLASLONG, float *, BLASLONG);
  367. float (*camin_k) (BLASLONG, float *, BLASLONG);
  368. BLASLONG (*icamax_k)(BLASLONG, float *, BLASLONG);
  369. BLASLONG (*icamin_k)(BLASLONG, float *, BLASLONG);
  370. float (*cnrm2_k) (BLASLONG, float *, BLASLONG);
  371. float (*casum_k) (BLASLONG, float *, BLASLONG);
  372. float (*csum_k) (BLASLONG, float *, BLASLONG);
  373. int (*ccopy_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  374. openblas_complex_float (*cdotu_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  375. openblas_complex_float (*cdotc_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  376. int (*csrot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG, float, float);
  377. int (*caxpy_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  378. int (*caxpyc_k)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  379. int (*cscal_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  380. int (*cswap_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  381. int (*cgemv_n) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  382. int (*cgemv_t) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  383. int (*cgemv_r) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  384. int (*cgemv_c) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  385. int (*cgemv_o) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  386. int (*cgemv_u) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  387. int (*cgemv_s) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  388. int (*cgemv_d) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  389. int (*cgeru_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  390. int (*cgerc_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  391. int (*cgerv_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  392. int (*cgerd_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  393. int (*csymv_L) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  394. int (*csymv_U) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  395. int (*chemv_L) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  396. int (*chemv_U) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  397. int (*chemv_M) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  398. int (*chemv_V) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  399. int (*cgemm_kernel_n )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  400. int (*cgemm_kernel_l )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  401. int (*cgemm_kernel_r )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  402. int (*cgemm_kernel_b )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  403. int (*cgemm_beta )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  404. int (*cgemm_incopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  405. int (*cgemm_itcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  406. int (*cgemm_oncopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  407. int (*cgemm_otcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  408. int (*ctrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  409. int (*ctrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  410. int (*ctrsm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  411. int (*ctrsm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  412. int (*ctrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  413. int (*ctrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  414. int (*ctrsm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  415. int (*ctrsm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  416. int (*ctrsm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  417. int (*ctrsm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  418. int (*ctrsm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  419. int (*ctrsm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  420. int (*ctrsm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  421. int (*ctrsm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  422. int (*ctrsm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  423. int (*ctrsm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  424. int (*ctrsm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  425. int (*ctrsm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  426. int (*ctrsm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  427. int (*ctrsm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  428. int (*ctrsm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  429. int (*ctrsm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  430. int (*ctrsm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  431. int (*ctrsm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  432. int (*ctrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  433. int (*ctrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  434. int (*ctrmm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  435. int (*ctrmm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  436. int (*ctrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  437. int (*ctrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  438. int (*ctrmm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  439. int (*ctrmm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  440. int (*ctrmm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  441. int (*ctrmm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  442. int (*ctrmm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  443. int (*ctrmm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  444. int (*ctrmm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  445. int (*ctrmm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  446. int (*ctrmm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  447. int (*ctrmm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  448. int (*ctrmm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  449. int (*ctrmm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  450. int (*ctrmm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  451. int (*ctrmm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  452. int (*ctrmm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  453. int (*ctrmm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  454. int (*ctrmm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  455. int (*ctrmm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  456. int (*csymm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  457. int (*csymm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  458. int (*csymm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  459. int (*csymm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  460. int (*chemm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  461. int (*chemm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  462. int (*chemm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  463. int (*chemm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  464. int cgemm3m_p, cgemm3m_q, cgemm3m_r;
  465. int cgemm3m_unroll_m, cgemm3m_unroll_n, cgemm3m_unroll_mn;
  466. int (*cgemm3m_kernel)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  467. int (*cgemm3m_incopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  468. int (*cgemm3m_incopyr)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  469. int (*cgemm3m_incopyi)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  470. int (*cgemm3m_itcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  471. int (*cgemm3m_itcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  472. int (*cgemm3m_itcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  473. int (*cgemm3m_oncopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  474. int (*cgemm3m_oncopyr)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  475. int (*cgemm3m_oncopyi)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  476. int (*cgemm3m_otcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  477. int (*cgemm3m_otcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  478. int (*cgemm3m_otcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  479. int (*csymm3m_iucopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  480. int (*csymm3m_ilcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  481. int (*csymm3m_iucopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  482. int (*csymm3m_ilcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  483. int (*csymm3m_iucopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  484. int (*csymm3m_ilcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  485. int (*csymm3m_oucopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  486. int (*csymm3m_olcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  487. int (*csymm3m_oucopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  488. int (*csymm3m_olcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  489. int (*csymm3m_oucopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  490. int (*csymm3m_olcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  491. int (*chemm3m_iucopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  492. int (*chemm3m_ilcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  493. int (*chemm3m_iucopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  494. int (*chemm3m_ilcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  495. int (*chemm3m_iucopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  496. int (*chemm3m_ilcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  497. int (*chemm3m_oucopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  498. int (*chemm3m_olcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  499. int (*chemm3m_oucopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  500. int (*chemm3m_olcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  501. int (*chemm3m_oucopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  502. int (*chemm3m_olcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  503. int (*cneg_tcopy) (BLASLONG, BLASLONG, float *, BLASLONG, float *);
  504. int (*claswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG, blasint *, float *);
  505. int zgemm_p, zgemm_q, zgemm_r;
  506. int zgemm_unroll_m, zgemm_unroll_n, zgemm_unroll_mn;
  507. double (*zamax_k) (BLASLONG, double *, BLASLONG);
  508. double (*zamin_k) (BLASLONG, double *, BLASLONG);
  509. BLASLONG (*izamax_k)(BLASLONG, double *, BLASLONG);
  510. BLASLONG (*izamin_k)(BLASLONG, double *, BLASLONG);
  511. double (*znrm2_k) (BLASLONG, double *, BLASLONG);
  512. double (*zasum_k) (BLASLONG, double *, BLASLONG);
  513. double (*zsum_k) (BLASLONG, double *, BLASLONG);
  514. int (*zcopy_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
  515. openblas_complex_double (*zdotu_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
  516. openblas_complex_double (*zdotc_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
  517. int (*zdrot_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG, double, double);
  518. int (*zaxpy_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  519. int (*zaxpyc_k)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  520. int (*zscal_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  521. int (*zswap_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  522. int (*zgemv_n) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  523. int (*zgemv_t) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  524. int (*zgemv_r) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  525. int (*zgemv_c) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  526. int (*zgemv_o) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  527. int (*zgemv_u) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  528. int (*zgemv_s) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  529. int (*zgemv_d) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  530. int (*zgeru_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  531. int (*zgerc_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  532. int (*zgerv_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  533. int (*zgerd_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  534. int (*zsymv_L) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  535. int (*zsymv_U) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  536. int (*zhemv_L) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  537. int (*zhemv_U) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  538. int (*zhemv_M) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  539. int (*zhemv_V) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  540. int (*zgemm_kernel_n )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  541. int (*zgemm_kernel_l )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  542. int (*zgemm_kernel_r )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  543. int (*zgemm_kernel_b )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  544. int (*zgemm_beta )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  545. int (*zgemm_incopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  546. int (*zgemm_itcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  547. int (*zgemm_oncopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  548. int (*zgemm_otcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  549. int (*ztrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  550. int (*ztrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  551. int (*ztrsm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  552. int (*ztrsm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  553. int (*ztrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  554. int (*ztrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  555. int (*ztrsm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  556. int (*ztrsm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  557. int (*ztrsm_iunucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  558. int (*ztrsm_iunncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  559. int (*ztrsm_iutucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  560. int (*ztrsm_iutncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  561. int (*ztrsm_ilnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  562. int (*ztrsm_ilnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  563. int (*ztrsm_iltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  564. int (*ztrsm_iltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  565. int (*ztrsm_ounucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  566. int (*ztrsm_ounncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  567. int (*ztrsm_outucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  568. int (*ztrsm_outncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  569. int (*ztrsm_olnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  570. int (*ztrsm_olnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  571. int (*ztrsm_oltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  572. int (*ztrsm_oltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  573. int (*ztrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  574. int (*ztrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  575. int (*ztrmm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  576. int (*ztrmm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  577. int (*ztrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  578. int (*ztrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  579. int (*ztrmm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  580. int (*ztrmm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  581. int (*ztrmm_iunucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  582. int (*ztrmm_iunncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  583. int (*ztrmm_iutucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  584. int (*ztrmm_iutncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  585. int (*ztrmm_ilnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  586. int (*ztrmm_ilnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  587. int (*ztrmm_iltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  588. int (*ztrmm_iltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  589. int (*ztrmm_ounucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  590. int (*ztrmm_ounncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  591. int (*ztrmm_outucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  592. int (*ztrmm_outncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  593. int (*ztrmm_olnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  594. int (*ztrmm_olnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  595. int (*ztrmm_oltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  596. int (*ztrmm_oltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  597. int (*zsymm_iutcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  598. int (*zsymm_iltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  599. int (*zsymm_outcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  600. int (*zsymm_oltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  601. int (*zhemm_iutcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  602. int (*zhemm_iltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  603. int (*zhemm_outcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  604. int (*zhemm_oltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  605. int zgemm3m_p, zgemm3m_q, zgemm3m_r;
  606. int zgemm3m_unroll_m, zgemm3m_unroll_n, zgemm3m_unroll_mn;
  607. int (*zgemm3m_kernel)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  608. int (*zgemm3m_incopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  609. int (*zgemm3m_incopyr)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  610. int (*zgemm3m_incopyi)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  611. int (*zgemm3m_itcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  612. int (*zgemm3m_itcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  613. int (*zgemm3m_itcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  614. int (*zgemm3m_oncopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  615. int (*zgemm3m_oncopyr)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  616. int (*zgemm3m_oncopyi)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  617. int (*zgemm3m_otcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  618. int (*zgemm3m_otcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  619. int (*zgemm3m_otcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  620. int (*zsymm3m_iucopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  621. int (*zsymm3m_ilcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  622. int (*zsymm3m_iucopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  623. int (*zsymm3m_ilcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  624. int (*zsymm3m_iucopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  625. int (*zsymm3m_ilcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  626. int (*zsymm3m_oucopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  627. int (*zsymm3m_olcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  628. int (*zsymm3m_oucopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  629. int (*zsymm3m_olcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  630. int (*zsymm3m_oucopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  631. int (*zsymm3m_olcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  632. int (*zhemm3m_iucopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  633. int (*zhemm3m_ilcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  634. int (*zhemm3m_iucopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  635. int (*zhemm3m_ilcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  636. int (*zhemm3m_iucopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  637. int (*zhemm3m_ilcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  638. int (*zhemm3m_oucopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  639. int (*zhemm3m_olcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  640. int (*zhemm3m_oucopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  641. int (*zhemm3m_olcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  642. int (*zhemm3m_oucopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  643. int (*zhemm3m_olcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  644. int (*zneg_tcopy) (BLASLONG, BLASLONG, double *, BLASLONG, double *);
  645. int (*zlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, double *, BLASLONG, blasint *, double *);
  646. #ifdef EXPRECISION
  647. int xgemm_p, xgemm_q, xgemm_r;
  648. int xgemm_unroll_m, xgemm_unroll_n, xgemm_unroll_mn;
  649. xdouble (*xamax_k) (BLASLONG, xdouble *, BLASLONG);
  650. xdouble (*xamin_k) (BLASLONG, xdouble *, BLASLONG);
  651. BLASLONG (*ixamax_k)(BLASLONG, xdouble *, BLASLONG);
  652. BLASLONG (*ixamin_k)(BLASLONG, xdouble *, BLASLONG);
  653. xdouble (*xnrm2_k) (BLASLONG, xdouble *, BLASLONG);
  654. xdouble (*xasum_k) (BLASLONG, xdouble *, BLASLONG);
  655. xdouble (*xsum_k) (BLASLONG, xdouble *, BLASLONG);
  656. int (*xcopy_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  657. openblas_complex_xdouble (*xdotu_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  658. openblas_complex_xdouble (*xdotc_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  659. int (*xqrot_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble);
  660. int (*xaxpy_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  661. int (*xaxpyc_k)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  662. int (*xscal_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  663. int (*xswap_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  664. int (*xgemv_n) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  665. int (*xgemv_t) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  666. int (*xgemv_r) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  667. int (*xgemv_c) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  668. int (*xgemv_o) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  669. int (*xgemv_u) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  670. int (*xgemv_s) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  671. int (*xgemv_d) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  672. int (*xgeru_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  673. int (*xgerc_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  674. int (*xgerv_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  675. int (*xgerd_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  676. int (*xsymv_L) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  677. int (*xsymv_U) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  678. int (*xhemv_L) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  679. int (*xhemv_U) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  680. int (*xhemv_M) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  681. int (*xhemv_V) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  682. int (*xgemm_kernel_n )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  683. int (*xgemm_kernel_l )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  684. int (*xgemm_kernel_r )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  685. int (*xgemm_kernel_b )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  686. int (*xgemm_beta )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  687. int (*xgemm_incopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  688. int (*xgemm_itcopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  689. int (*xgemm_oncopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  690. int (*xgemm_otcopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  691. int (*xtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  692. int (*xtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  693. int (*xtrsm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  694. int (*xtrsm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  695. int (*xtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  696. int (*xtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  697. int (*xtrsm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  698. int (*xtrsm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  699. int (*xtrsm_iunucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  700. int (*xtrsm_iunncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  701. int (*xtrsm_iutucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  702. int (*xtrsm_iutncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  703. int (*xtrsm_ilnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  704. int (*xtrsm_ilnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  705. int (*xtrsm_iltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  706. int (*xtrsm_iltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  707. int (*xtrsm_ounucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  708. int (*xtrsm_ounncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  709. int (*xtrsm_outucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  710. int (*xtrsm_outncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  711. int (*xtrsm_olnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  712. int (*xtrsm_olnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  713. int (*xtrsm_oltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  714. int (*xtrsm_oltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  715. int (*xtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  716. int (*xtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  717. int (*xtrmm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  718. int (*xtrmm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  719. int (*xtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  720. int (*xtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  721. int (*xtrmm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  722. int (*xtrmm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  723. int (*xtrmm_iunucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  724. int (*xtrmm_iunncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  725. int (*xtrmm_iutucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  726. int (*xtrmm_iutncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  727. int (*xtrmm_ilnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  728. int (*xtrmm_ilnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  729. int (*xtrmm_iltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  730. int (*xtrmm_iltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  731. int (*xtrmm_ounucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  732. int (*xtrmm_ounncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  733. int (*xtrmm_outucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  734. int (*xtrmm_outncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  735. int (*xtrmm_olnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  736. int (*xtrmm_olnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  737. int (*xtrmm_oltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  738. int (*xtrmm_oltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  739. int (*xsymm_iutcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  740. int (*xsymm_iltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  741. int (*xsymm_outcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  742. int (*xsymm_oltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  743. int (*xhemm_iutcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  744. int (*xhemm_iltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  745. int (*xhemm_outcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  746. int (*xhemm_oltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  747. int xgemm3m_p, xgemm3m_q, xgemm3m_r;
  748. int xgemm3m_unroll_m, xgemm3m_unroll_n, xgemm3m_unroll_mn;
  749. int (*xgemm3m_kernel)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  750. int (*xgemm3m_incopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  751. int (*xgemm3m_incopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  752. int (*xgemm3m_incopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  753. int (*xgemm3m_itcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  754. int (*xgemm3m_itcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  755. int (*xgemm3m_itcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  756. int (*xgemm3m_oncopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  757. int (*xgemm3m_oncopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  758. int (*xgemm3m_oncopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  759. int (*xgemm3m_otcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  760. int (*xgemm3m_otcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  761. int (*xgemm3m_otcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  762. int (*xsymm3m_iucopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  763. int (*xsymm3m_ilcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  764. int (*xsymm3m_iucopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  765. int (*xsymm3m_ilcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  766. int (*xsymm3m_iucopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  767. int (*xsymm3m_ilcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  768. int (*xsymm3m_oucopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  769. int (*xsymm3m_olcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  770. int (*xsymm3m_oucopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  771. int (*xsymm3m_olcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  772. int (*xsymm3m_oucopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  773. int (*xsymm3m_olcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  774. int (*xhemm3m_iucopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  775. int (*xhemm3m_ilcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  776. int (*xhemm3m_iucopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  777. int (*xhemm3m_ilcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  778. int (*xhemm3m_iucopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  779. int (*xhemm3m_ilcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  780. int (*xhemm3m_oucopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  781. int (*xhemm3m_olcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  782. int (*xhemm3m_oucopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  783. int (*xhemm3m_olcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  784. int (*xhemm3m_oucopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  785. int (*xhemm3m_olcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  786. int (*xneg_tcopy) (BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  787. int (*xlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, xdouble *, BLASLONG, blasint *, xdouble *);
  788. #endif
  789. void (*init)(void);
  790. int snum_opt, dnum_opt, qnum_opt;
  791. int (*saxpby_k) (BLASLONG, float, float*, BLASLONG,float, float*, BLASLONG);
  792. int (*daxpby_k) (BLASLONG, double, double*, BLASLONG,double, double*, BLASLONG);
  793. int (*caxpby_k) (BLASLONG, float, float, float*, BLASLONG,float,float, float*, BLASLONG);
  794. int (*zaxpby_k) (BLASLONG, double, double, double*, BLASLONG,double,double, double*, BLASLONG);
  795. int (*somatcopy_k_cn) (BLASLONG, BLASLONG, float, float*, BLASLONG, float*, BLASLONG);
  796. int (*somatcopy_k_ct) (BLASLONG, BLASLONG, float, float*, BLASLONG, float*, BLASLONG);
  797. int (*somatcopy_k_rn) (BLASLONG, BLASLONG, float, float*, BLASLONG, float*, BLASLONG);
  798. int (*somatcopy_k_rt) (BLASLONG, BLASLONG, float, float*, BLASLONG, float*, BLASLONG);
  799. int (*domatcopy_k_cn) (BLASLONG, BLASLONG, double, double*, BLASLONG, double*, BLASLONG);
  800. int (*domatcopy_k_ct) (BLASLONG, BLASLONG, double, double*, BLASLONG, double*, BLASLONG);
  801. int (*domatcopy_k_rn) (BLASLONG, BLASLONG, double, double*, BLASLONG, double*, BLASLONG);
  802. int (*domatcopy_k_rt) (BLASLONG, BLASLONG, double, double*, BLASLONG, double*, BLASLONG);
  803. int (*comatcopy_k_cn) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  804. int (*comatcopy_k_ct) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  805. int (*comatcopy_k_rn) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  806. int (*comatcopy_k_rt) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  807. int (*comatcopy_k_cnc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  808. int (*comatcopy_k_ctc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  809. int (*comatcopy_k_rnc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  810. int (*comatcopy_k_rtc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  811. int (*zomatcopy_k_cn) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  812. int (*zomatcopy_k_ct) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  813. int (*zomatcopy_k_rn) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  814. int (*zomatcopy_k_rt) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  815. int (*zomatcopy_k_cnc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  816. int (*zomatcopy_k_ctc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  817. int (*zomatcopy_k_rnc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  818. int (*zomatcopy_k_rtc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  819. int (*simatcopy_k_cn) (BLASLONG, BLASLONG, float, float*, BLASLONG);
  820. int (*simatcopy_k_ct) (BLASLONG, BLASLONG, float, float*, BLASLONG);
  821. int (*simatcopy_k_rn) (BLASLONG, BLASLONG, float, float*, BLASLONG);
  822. int (*simatcopy_k_rt) (BLASLONG, BLASLONG, float, float*, BLASLONG);
  823. int (*dimatcopy_k_cn) (BLASLONG, BLASLONG, double, double*, BLASLONG);
  824. int (*dimatcopy_k_ct) (BLASLONG, BLASLONG, double, double*, BLASLONG);
  825. int (*dimatcopy_k_rn) (BLASLONG, BLASLONG, double, double*, BLASLONG);
  826. int (*dimatcopy_k_rt) (BLASLONG, BLASLONG, double, double*, BLASLONG);
  827. int (*cimatcopy_k_cn) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  828. int (*cimatcopy_k_ct) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  829. int (*cimatcopy_k_rn) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  830. int (*cimatcopy_k_rt) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  831. int (*cimatcopy_k_cnc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  832. int (*cimatcopy_k_ctc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  833. int (*cimatcopy_k_rnc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  834. int (*cimatcopy_k_rtc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  835. int (*zimatcopy_k_cn) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  836. int (*zimatcopy_k_ct) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  837. int (*zimatcopy_k_rn) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  838. int (*zimatcopy_k_rt) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  839. int (*zimatcopy_k_cnc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  840. int (*zimatcopy_k_ctc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  841. int (*zimatcopy_k_rnc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  842. int (*zimatcopy_k_rtc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  843. int (*sgeadd_k) (BLASLONG, BLASLONG, float, float *, BLASLONG, float, float *, BLASLONG);
  844. int (*dgeadd_k) (BLASLONG, BLASLONG, double, double *, BLASLONG, double, double *, BLASLONG);
  845. int (*cgeadd_k) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float, float, float *, BLASLONG);
  846. int (*zgeadd_k) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double, double, double *, BLASLONG);
  847. } gotoblas_t;
  848. extern gotoblas_t *gotoblas;
  849. #define DTB_ENTRIES gotoblas -> dtb_entries
  850. #define GEMM_OFFSET_A gotoblas -> offsetA
  851. #define GEMM_OFFSET_B gotoblas -> offsetB
  852. #define GEMM_ALIGN gotoblas -> align
  853. #define HAVE_EX_L2 gotoblas -> exclusive_cache
  854. #ifdef BUILD_HALF
  855. #define SHGEMM_P gotoblas -> shgemm_p
  856. #define SHGEMM_Q gotoblas -> shgemm_q
  857. #define SHGEMM_R gotoblas -> shgemm_r
  858. #define SHGEMM_UNROLL_M gotoblas -> shgemm_unroll_m
  859. #define SHGEMM_UNROLL_N gotoblas -> shgemm_unroll_n
  860. #define SHGEMM_UNROLL_MN gotoblas -> shgemm_unroll_mn
  861. #endif
  862. #define SGEMM_P gotoblas -> sgemm_p
  863. #define SGEMM_Q gotoblas -> sgemm_q
  864. #define SGEMM_R gotoblas -> sgemm_r
  865. #define SGEMM_UNROLL_M gotoblas -> sgemm_unroll_m
  866. #define SGEMM_UNROLL_N gotoblas -> sgemm_unroll_n
  867. #define SGEMM_UNROLL_MN gotoblas -> sgemm_unroll_mn
  868. #define DGEMM_P gotoblas -> dgemm_p
  869. #define DGEMM_Q gotoblas -> dgemm_q
  870. #define DGEMM_R gotoblas -> dgemm_r
  871. #define DGEMM_UNROLL_M gotoblas -> dgemm_unroll_m
  872. #define DGEMM_UNROLL_N gotoblas -> dgemm_unroll_n
  873. #define DGEMM_UNROLL_MN gotoblas -> dgemm_unroll_mn
  874. #define QGEMM_P gotoblas -> qgemm_p
  875. #define QGEMM_Q gotoblas -> qgemm_q
  876. #define QGEMM_R gotoblas -> qgemm_r
  877. #define QGEMM_UNROLL_M gotoblas -> qgemm_unroll_m
  878. #define QGEMM_UNROLL_N gotoblas -> qgemm_unroll_n
  879. #define QGEMM_UNROLL_MN gotoblas -> qgemm_unroll_mn
  880. #define CGEMM_P gotoblas -> cgemm_p
  881. #define CGEMM_Q gotoblas -> cgemm_q
  882. #define CGEMM_R gotoblas -> cgemm_r
  883. #define CGEMM_UNROLL_M gotoblas -> cgemm_unroll_m
  884. #define CGEMM_UNROLL_N gotoblas -> cgemm_unroll_n
  885. #define CGEMM_UNROLL_MN gotoblas -> cgemm_unroll_mn
  886. #define ZGEMM_P gotoblas -> zgemm_p
  887. #define ZGEMM_Q gotoblas -> zgemm_q
  888. #define ZGEMM_R gotoblas -> zgemm_r
  889. #define ZGEMM_UNROLL_M gotoblas -> zgemm_unroll_m
  890. #define ZGEMM_UNROLL_N gotoblas -> zgemm_unroll_n
  891. #define ZGEMM_UNROLL_MN gotoblas -> zgemm_unroll_mn
  892. #define XGEMM_P gotoblas -> xgemm_p
  893. #define XGEMM_Q gotoblas -> xgemm_q
  894. #define XGEMM_R gotoblas -> xgemm_r
  895. #define XGEMM_UNROLL_M gotoblas -> xgemm_unroll_m
  896. #define XGEMM_UNROLL_N gotoblas -> xgemm_unroll_n
  897. #define XGEMM_UNROLL_MN gotoblas -> xgemm_unroll_mn
  898. #define CGEMM3M_P gotoblas -> cgemm3m_p
  899. #define CGEMM3M_Q gotoblas -> cgemm3m_q
  900. #define CGEMM3M_R gotoblas -> cgemm3m_r
  901. #define CGEMM3M_UNROLL_M gotoblas -> cgemm3m_unroll_m
  902. #define CGEMM3M_UNROLL_N gotoblas -> cgemm3m_unroll_n
  903. #define CGEMM3M_UNROLL_MN gotoblas -> cgemm3m_unroll_mn
  904. #define ZGEMM3M_P gotoblas -> zgemm3m_p
  905. #define ZGEMM3M_Q gotoblas -> zgemm3m_q
  906. #define ZGEMM3M_R gotoblas -> zgemm3m_r
  907. #define ZGEMM3M_UNROLL_M gotoblas -> zgemm3m_unroll_m
  908. #define ZGEMM3M_UNROLL_N gotoblas -> zgemm3m_unroll_n
  909. #define ZGEMM3M_UNROLL_MN gotoblas -> zgemm3m_unroll_mn
  910. #define XGEMM3M_P gotoblas -> xgemm3m_p
  911. #define XGEMM3M_Q gotoblas -> xgemm3m_q
  912. #define XGEMM3M_R gotoblas -> xgemm3m_r
  913. #define XGEMM3M_UNROLL_M gotoblas -> xgemm3m_unroll_m
  914. #define XGEMM3M_UNROLL_N gotoblas -> xgemm3m_unroll_n
  915. #define XGEMM3M_UNROLL_MN gotoblas -> xgemm3m_unroll_mn
  916. #else
  917. #define DTB_ENTRIES DTB_DEFAULT_ENTRIES
  918. #define GEMM_OFFSET_A GEMM_DEFAULT_OFFSET_A
  919. #define GEMM_OFFSET_B GEMM_DEFAULT_OFFSET_B
  920. #define GEMM_ALIGN GEMM_DEFAULT_ALIGN
  921. #ifdef HAVE_EXCLUSIVE_CACHE
  922. #define HAVE_EX_L2 1
  923. #else
  924. #define HAVE_EX_L2 0
  925. #endif
  926. #ifdef BUILD_HALF
  927. #define SHGEMM_P SHGEMM_DEFAULT_P
  928. #define SHGEMM_Q SHGEMM_DEFAULT_Q
  929. #define SHGEMM_R SHGEMM_DEFAULT_R
  930. #define SHGEMM_UNROLL_M SHGEMM_DEFAULT_UNROLL_M
  931. #define SHGEMM_UNROLL_N SHGEMM_DEFAULT_UNROLL_N
  932. #ifdef SHGEMM_DEFAULT_UNROLL_MN
  933. #define SHGEMM_UNROLL_MN SHGEMM_DEFAULT_UNROLL_MN
  934. #else
  935. #define SHGEMM_UNROLL_MN MAX((SHGEMM_UNROLL_M), (SHGEMM_UNROLL_N))
  936. #endif
  937. #endif
  938. #define SGEMM_P SGEMM_DEFAULT_P
  939. #define SGEMM_Q SGEMM_DEFAULT_Q
  940. #define SGEMM_R SGEMM_DEFAULT_R
  941. #define SGEMM_UNROLL_M SGEMM_DEFAULT_UNROLL_M
  942. #define SGEMM_UNROLL_N SGEMM_DEFAULT_UNROLL_N
  943. #ifdef SGEMM_DEFAULT_UNROLL_MN
  944. #define SGEMM_UNROLL_MN SGEMM_DEFAULT_UNROLL_MN
  945. #else
  946. #define SGEMM_UNROLL_MN MAX((SGEMM_UNROLL_M), (SGEMM_UNROLL_N))
  947. #endif
  948. #define DGEMM_P DGEMM_DEFAULT_P
  949. #define DGEMM_Q DGEMM_DEFAULT_Q
  950. #define DGEMM_R DGEMM_DEFAULT_R
  951. #define DGEMM_UNROLL_M DGEMM_DEFAULT_UNROLL_M
  952. #define DGEMM_UNROLL_N DGEMM_DEFAULT_UNROLL_N
  953. #ifdef DGEMM_DEFAULT_UNROLL_MN
  954. #define DGEMM_UNROLL_MN DGEMM_DEFAULT_UNROLL_MN
  955. #else
  956. #define DGEMM_UNROLL_MN MAX((DGEMM_UNROLL_M), (DGEMM_UNROLL_N))
  957. #endif
  958. #define QGEMM_P QGEMM_DEFAULT_P
  959. #define QGEMM_Q QGEMM_DEFAULT_Q
  960. #define QGEMM_R QGEMM_DEFAULT_R
  961. #define QGEMM_UNROLL_M QGEMM_DEFAULT_UNROLL_M
  962. #define QGEMM_UNROLL_N QGEMM_DEFAULT_UNROLL_N
  963. #define QGEMM_UNROLL_MN MAX((QGEMM_UNROLL_M), (QGEMM_UNROLL_N))
  964. #define CGEMM_P CGEMM_DEFAULT_P
  965. #define CGEMM_Q CGEMM_DEFAULT_Q
  966. #define CGEMM_R CGEMM_DEFAULT_R
  967. #define CGEMM_UNROLL_M CGEMM_DEFAULT_UNROLL_M
  968. #define CGEMM_UNROLL_N CGEMM_DEFAULT_UNROLL_N
  969. #ifdef CGEMM_DEFAULT_UNROLL_MN
  970. #define CGEMM_UNROLL_MN CGEMM_DEFAULT_UNROLL_MN
  971. #else
  972. #define CGEMM_UNROLL_MN MAX((CGEMM_UNROLL_M), (CGEMM_UNROLL_N))
  973. #endif
  974. #define ZGEMM_P ZGEMM_DEFAULT_P
  975. #define ZGEMM_Q ZGEMM_DEFAULT_Q
  976. #define ZGEMM_R ZGEMM_DEFAULT_R
  977. #define ZGEMM_UNROLL_M ZGEMM_DEFAULT_UNROLL_M
  978. #define ZGEMM_UNROLL_N ZGEMM_DEFAULT_UNROLL_N
  979. #ifdef ZGEMM_DEFAULT_UNROLL_MN
  980. #define ZGEMM_UNROLL_MN ZGEMM_DEFAULT_UNROLL_MN
  981. #else
  982. #define ZGEMM_UNROLL_MN MAX((ZGEMM_UNROLL_M), (ZGEMM_UNROLL_N))
  983. #endif
  984. #define XGEMM_P XGEMM_DEFAULT_P
  985. #define XGEMM_Q XGEMM_DEFAULT_Q
  986. #define XGEMM_R XGEMM_DEFAULT_R
  987. #define XGEMM_UNROLL_M XGEMM_DEFAULT_UNROLL_M
  988. #define XGEMM_UNROLL_N XGEMM_DEFAULT_UNROLL_N
  989. #define XGEMM_UNROLL_MN MAX((XGEMM_UNROLL_M), (XGEMM_UNROLL_N))
  990. #ifdef CGEMM3M_DEFAULT_UNROLL_N
  991. #define CGEMM3M_P CGEMM3M_DEFAULT_P
  992. #define CGEMM3M_Q CGEMM3M_DEFAULT_Q
  993. #define CGEMM3M_R CGEMM3M_DEFAULT_R
  994. #define CGEMM3M_UNROLL_M CGEMM3M_DEFAULT_UNROLL_M
  995. #define CGEMM3M_UNROLL_N CGEMM3M_DEFAULT_UNROLL_N
  996. #define CGEMM3M_UNROLL_MN MAX((CGEMM3M_UNROLL_M), (CGEMM3M_UNROLL_N))
  997. #else
  998. #define CGEMM3M_P SGEMM_DEFAULT_P
  999. #define CGEMM3M_Q SGEMM_DEFAULT_Q
  1000. #define CGEMM3M_R SGEMM_DEFAULT_R
  1001. #define CGEMM3M_UNROLL_M SGEMM_DEFAULT_UNROLL_M
  1002. #define CGEMM3M_UNROLL_N SGEMM_DEFAULT_UNROLL_N
  1003. #define CGEMM3M_UNROLL_MN MAX((CGEMM_UNROLL_M), (CGEMM_UNROLL_N))
  1004. #endif
  1005. #ifdef ZGEMM3M_DEFAULT_UNROLL_N
  1006. #define ZGEMM3M_P ZGEMM3M_DEFAULT_P
  1007. #define ZGEMM3M_Q ZGEMM3M_DEFAULT_Q
  1008. #define ZGEMM3M_R ZGEMM3M_DEFAULT_R
  1009. #define ZGEMM3M_UNROLL_M ZGEMM3M_DEFAULT_UNROLL_M
  1010. #define ZGEMM3M_UNROLL_N ZGEMM3M_DEFAULT_UNROLL_N
  1011. #define ZGEMM3M_UNROLL_MN MAX((ZGEMM_UNROLL_M), (ZGEMM_UNROLL_N))
  1012. #else
  1013. #define ZGEMM3M_P DGEMM_DEFAULT_P
  1014. #define ZGEMM3M_Q DGEMM_DEFAULT_Q
  1015. #define ZGEMM3M_R DGEMM_DEFAULT_R
  1016. #define ZGEMM3M_UNROLL_M DGEMM_DEFAULT_UNROLL_M
  1017. #define ZGEMM3M_UNROLL_N DGEMM_DEFAULT_UNROLL_N
  1018. #define ZGEMM3M_UNROLL_MN MAX((ZGEMM_UNROLL_M), (ZGEMM_UNROLL_N))
  1019. #endif
  1020. #define XGEMM3M_P QGEMM_DEFAULT_P
  1021. #define XGEMM3M_Q QGEMM_DEFAULT_Q
  1022. #define XGEMM3M_R QGEMM_DEFAULT_R
  1023. #define XGEMM3M_UNROLL_M QGEMM_DEFAULT_UNROLL_M
  1024. #define XGEMM3M_UNROLL_N QGEMM_DEFAULT_UNROLL_N
  1025. #define XGEMM3M_UNROLL_MN MAX((QGEMM_UNROLL_M), (QGEMM_UNROLL_N))
  1026. #endif
  1027. #endif
  1028. #ifndef COMPLEX
  1029. #if defined(XDOUBLE)
  1030. #define GEMM_P QGEMM_P
  1031. #define GEMM_Q QGEMM_Q
  1032. #define GEMM_R QGEMM_R
  1033. #define GEMM_UNROLL_M QGEMM_UNROLL_M
  1034. #define GEMM_UNROLL_N QGEMM_UNROLL_N
  1035. #define GEMM_UNROLL_MN QGEMM_UNROLL_MN
  1036. #define GEMM_DEFAULT_P QGEMM_DEFAULT_P
  1037. #define GEMM_DEFAULT_Q QGEMM_DEFAULT_Q
  1038. #define GEMM_DEFAULT_R QGEMM_DEFAULT_R
  1039. #define GEMM_DEFAULT_UNROLL_M QGEMM_DEFAULT_UNROLL_M
  1040. #define GEMM_DEFAULT_UNROLL_N QGEMM_DEFAULT_UNROLL_N
  1041. #elif defined(DOUBLE)
  1042. #define GEMM_P DGEMM_P
  1043. #define GEMM_Q DGEMM_Q
  1044. #define GEMM_R DGEMM_R
  1045. #define GEMM_UNROLL_M DGEMM_UNROLL_M
  1046. #define GEMM_UNROLL_N DGEMM_UNROLL_N
  1047. #define GEMM_UNROLL_MN DGEMM_UNROLL_MN
  1048. #define GEMM_DEFAULT_P DGEMM_DEFAULT_P
  1049. #define GEMM_DEFAULT_Q DGEMM_DEFAULT_Q
  1050. #define GEMM_DEFAULT_R DGEMM_DEFAULT_R
  1051. #define GEMM_DEFAULT_UNROLL_M DGEMM_DEFAULT_UNROLL_M
  1052. #define GEMM_DEFAULT_UNROLL_N DGEMM_DEFAULT_UNROLL_N
  1053. #elif defined(HALF)
  1054. #define GEMM_P SHGEMM_P
  1055. #define GEMM_Q SHGEMM_Q
  1056. #define GEMM_R SHGEMM_R
  1057. #define GEMM_UNROLL_M SHGEMM_UNROLL_M
  1058. #define GEMM_UNROLL_N SHGEMM_UNROLL_N
  1059. #define GEMM_UNROLL_MN SHGEMM_UNROLL_MN
  1060. #define GEMM_DEFAULT_P SHGEMM_DEFAULT_P
  1061. #define GEMM_DEFAULT_Q SHGEMM_DEFAULT_Q
  1062. #define GEMM_DEFAULT_R SHGEMM_DEFAULT_R
  1063. #define GEMM_DEFAULT_UNROLL_M SHGEMM_DEFAULT_UNROLL_M
  1064. #define GEMM_DEFAULT_UNROLL_N SHGEMM_DEFAULT_UNROLL_N
  1065. #else
  1066. #define GEMM_P SGEMM_P
  1067. #define GEMM_Q SGEMM_Q
  1068. #define GEMM_R SGEMM_R
  1069. #define GEMM_UNROLL_M SGEMM_UNROLL_M
  1070. #define GEMM_UNROLL_N SGEMM_UNROLL_N
  1071. #define GEMM_UNROLL_MN SGEMM_UNROLL_MN
  1072. #define GEMM_DEFAULT_P SGEMM_DEFAULT_P
  1073. #define GEMM_DEFAULT_Q SGEMM_DEFAULT_Q
  1074. #define GEMM_DEFAULT_R SGEMM_DEFAULT_R
  1075. #define GEMM_DEFAULT_UNROLL_M SGEMM_DEFAULT_UNROLL_M
  1076. #define GEMM_DEFAULT_UNROLL_N SGEMM_DEFAULT_UNROLL_N
  1077. #endif
  1078. #else
  1079. #if defined(XDOUBLE)
  1080. #define GEMM_P XGEMM_P
  1081. #define GEMM_Q XGEMM_Q
  1082. #define GEMM_R XGEMM_R
  1083. #define GEMM_UNROLL_M XGEMM_UNROLL_M
  1084. #define GEMM_UNROLL_N XGEMM_UNROLL_N
  1085. #define GEMM_UNROLL_MN XGEMM_UNROLL_MN
  1086. #define GEMM_DEFAULT_P XGEMM_DEFAULT_P
  1087. #define GEMM_DEFAULT_Q XGEMM_DEFAULT_Q
  1088. #define GEMM_DEFAULT_R XGEMM_DEFAULT_R
  1089. #define GEMM_DEFAULT_UNROLL_M XGEMM_DEFAULT_UNROLL_M
  1090. #define GEMM_DEFAULT_UNROLL_N XGEMM_DEFAULT_UNROLL_N
  1091. #elif defined(DOUBLE)
  1092. #define GEMM_P ZGEMM_P
  1093. #define GEMM_Q ZGEMM_Q
  1094. #define GEMM_R ZGEMM_R
  1095. #define GEMM_UNROLL_M ZGEMM_UNROLL_M
  1096. #define GEMM_UNROLL_N ZGEMM_UNROLL_N
  1097. #define GEMM_UNROLL_MN ZGEMM_UNROLL_MN
  1098. #define GEMM_DEFAULT_P ZGEMM_DEFAULT_P
  1099. #define GEMM_DEFAULT_Q ZGEMM_DEFAULT_Q
  1100. #define GEMM_DEFAULT_R ZGEMM_DEFAULT_R
  1101. #define GEMM_DEFAULT_UNROLL_M ZGEMM_DEFAULT_UNROLL_M
  1102. #define GEMM_DEFAULT_UNROLL_N ZGEMM_DEFAULT_UNROLL_N
  1103. #else
  1104. #define GEMM_P CGEMM_P
  1105. #define GEMM_Q CGEMM_Q
  1106. #define GEMM_R CGEMM_R
  1107. #define GEMM_UNROLL_M CGEMM_UNROLL_M
  1108. #define GEMM_UNROLL_N CGEMM_UNROLL_N
  1109. #define GEMM_UNROLL_MN CGEMM_UNROLL_MN
  1110. #define GEMM_DEFAULT_P CGEMM_DEFAULT_P
  1111. #define GEMM_DEFAULT_Q CGEMM_DEFAULT_Q
  1112. #define GEMM_DEFAULT_R CGEMM_DEFAULT_R
  1113. #define GEMM_DEFAULT_UNROLL_M CGEMM_DEFAULT_UNROLL_M
  1114. #define GEMM_DEFAULT_UNROLL_N CGEMM_DEFAULT_UNROLL_N
  1115. #endif
  1116. #endif
  1117. #ifdef XDOUBLE
  1118. #define GEMM3M_UNROLL_M XGEMM3M_UNROLL_M
  1119. #define GEMM3M_UNROLL_N XGEMM3M_UNROLL_N
  1120. #elif defined(DOUBLE)
  1121. #define GEMM3M_UNROLL_M ZGEMM3M_UNROLL_M
  1122. #define GEMM3M_UNROLL_N ZGEMM3M_UNROLL_N
  1123. #else
  1124. #define GEMM3M_UNROLL_M CGEMM3M_UNROLL_M
  1125. #define GEMM3M_UNROLL_N CGEMM3M_UNROLL_N
  1126. #endif
  1127. #ifndef QGEMM_DEFAULT_UNROLL_M
  1128. #define QGEMM_DEFAULT_UNROLL_M 2
  1129. #endif
  1130. #ifndef QGEMM_DEFAULT_UNROLL_N
  1131. #define QGEMM_DEFAULT_UNROLL_N 2
  1132. #endif
  1133. #ifndef XGEMM_DEFAULT_UNROLL_M
  1134. #define XGEMM_DEFAULT_UNROLL_M 2
  1135. #endif
  1136. #ifndef XGEMM_DEFAULT_UNROLL_N
  1137. #define XGEMM_DEFAULT_UNROLL_N 2
  1138. #endif
  1139. #ifndef GEMM_THREAD
  1140. #define GEMM_THREAD gemm_thread_n
  1141. #endif
  1142. #ifndef SHGEMM_DEFAULT_R
  1143. #define SHGEMM_DEFAULT_R (((BUFFER_SIZE - ((SHGEMM_DEFAULT_P * SHGEMM_DEFAULT_Q * 4 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (SHGEMM_DEFAULT_Q * 4) - 15) & ~15UL)
  1144. #endif
  1145. #ifndef SGEMM_DEFAULT_R
  1146. #define SGEMM_DEFAULT_R (((BUFFER_SIZE - ((SGEMM_DEFAULT_P * SGEMM_DEFAULT_Q * 4 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (SGEMM_DEFAULT_Q * 4) - 15) & ~15UL)
  1147. #endif
  1148. #ifndef DGEMM_DEFAULT_R
  1149. #define DGEMM_DEFAULT_R (((BUFFER_SIZE - ((DGEMM_DEFAULT_P * DGEMM_DEFAULT_Q * 8 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (DGEMM_DEFAULT_Q * 8) - 15) & ~15UL)
  1150. #endif
  1151. #ifndef QGEMM_DEFAULT_R
  1152. #define QGEMM_DEFAULT_R (((BUFFER_SIZE - ((QGEMM_DEFAULT_P * QGEMM_DEFAULT_Q * 16 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (QGEMM_DEFAULT_Q * 16) - 15) & ~15UL)
  1153. #endif
  1154. #ifndef CGEMM_DEFAULT_R
  1155. #define CGEMM_DEFAULT_R (((BUFFER_SIZE - ((CGEMM_DEFAULT_P * CGEMM_DEFAULT_Q * 8 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (CGEMM_DEFAULT_Q * 8) - 15) & ~15UL)
  1156. #endif
  1157. #ifndef ZGEMM_DEFAULT_R
  1158. #define ZGEMM_DEFAULT_R (((BUFFER_SIZE - ((ZGEMM_DEFAULT_P * ZGEMM_DEFAULT_Q * 16 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (ZGEMM_DEFAULT_Q * 16) - 15) & ~15UL)
  1159. #endif
  1160. #ifndef XGEMM_DEFAULT_R
  1161. #define XGEMM_DEFAULT_R (((BUFFER_SIZE - ((XGEMM_DEFAULT_P * XGEMM_DEFAULT_Q * 32 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (XGEMM_DEFAULT_Q * 32) - 15) & ~15UL)
  1162. #endif
  1163. #ifndef SNUMOPT
  1164. #define SNUMOPT 2
  1165. #endif
  1166. #ifndef DNUMOPT
  1167. #define DNUMOPT 2
  1168. #endif
  1169. #ifndef QNUMOPT
  1170. #define QNUMOPT 1
  1171. #endif
  1172. #ifndef GEMM3M_P
  1173. #ifdef XDOUBLE
  1174. #define GEMM3M_P XGEMM3M_P
  1175. #elif defined(DOUBLE)
  1176. #define GEMM3M_P ZGEMM3M_P
  1177. #else
  1178. #define GEMM3M_P CGEMM3M_P
  1179. #endif
  1180. #endif
  1181. #ifndef GEMM3M_Q
  1182. #ifdef XDOUBLE
  1183. #define GEMM3M_Q XGEMM3M_Q
  1184. #elif defined(DOUBLE)
  1185. #define GEMM3M_Q ZGEMM3M_Q
  1186. #else
  1187. #define GEMM3M_Q CGEMM3M_Q
  1188. #endif
  1189. #endif
  1190. #ifndef GEMM3M_R
  1191. #ifdef XDOUBLE
  1192. #define GEMM3M_R XGEMM3M_R
  1193. #elif defined(DOUBLE)
  1194. #define GEMM3M_R ZGEMM3M_R
  1195. #else
  1196. #define GEMM3M_R CGEMM3M_R
  1197. #endif
  1198. #endif
  1199. #endif