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 112 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638
  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_BFLOAT16
  46. int sbgemm_p, sbgemm_q, sbgemm_r;
  47. int sbgemm_unroll_m, sbgemm_unroll_n, sbgemm_unroll_mn;
  48. int sbgemm_align_k;
  49. void (*sbstobf16_k) (BLASLONG, float *, BLASLONG, bfloat16 *, BLASLONG);
  50. void (*sbdtobf16_k) (BLASLONG, double *, BLASLONG, bfloat16 *, BLASLONG);
  51. void (*sbf16tos_k) (BLASLONG, bfloat16 *, BLASLONG, float *, BLASLONG);
  52. void (*dbf16tod_k) (BLASLONG, bfloat16 *, BLASLONG, double *, BLASLONG);
  53. float (*sbamax_k) (BLASLONG, float *, BLASLONG);
  54. float (*sbamin_k) (BLASLONG, float *, BLASLONG);
  55. float (*sbmax_k) (BLASLONG, float *, BLASLONG);
  56. float (*sbmin_k) (BLASLONG, float *, BLASLONG);
  57. BLASLONG (*isbamax_k)(BLASLONG, float *, BLASLONG);
  58. BLASLONG (*isbamin_k)(BLASLONG, float *, BLASLONG);
  59. BLASLONG (*isbmax_k) (BLASLONG, float *, BLASLONG);
  60. BLASLONG (*isbmin_k) (BLASLONG, float *, BLASLONG);
  61. float (*sbnrm2_k) (BLASLONG, float *, BLASLONG);
  62. float (*sbasum_k) (BLASLONG, float *, BLASLONG);
  63. float (*sbsum_k) (BLASLONG, float *, BLASLONG);
  64. int (*sbcopy_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  65. float (*sbdot_k) (BLASLONG, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG);
  66. double (*dsbdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  67. int (*sbrot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG, float, float);
  68. int (*sbaxpy_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  69. int (*sbscal_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  70. int (*sbswap_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  71. int (*sbgemv_n) (BLASLONG, BLASLONG, float, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, float, float *, BLASLONG);
  72. int (*sbgemv_t) (BLASLONG, BLASLONG, float, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, float, float *, BLASLONG);
  73. int (*sbger_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  74. int (*sbsymv_L) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  75. int (*sbsymv_U) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  76. int (*sbgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, float, bfloat16 *, bfloat16 *, float *, BLASLONG);
  77. int (*sbgemm_beta )(BLASLONG, BLASLONG, BLASLONG, float, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, float *, BLASLONG);
  78. int (*sbgemm_incopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
  79. int (*sbgemm_itcopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
  80. int (*sbgemm_oncopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
  81. int (*sbgemm_otcopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
  82. int (*sbtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  83. int (*sbtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  84. int (*sbtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  85. int (*sbtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  86. int (*sbtrsm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  87. int (*sbtrsm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  88. int (*sbtrsm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  89. int (*sbtrsm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  90. int (*sbtrsm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  91. int (*sbtrsm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  92. int (*sbtrsm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  93. int (*sbtrsm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  94. int (*sbtrsm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  95. int (*sbtrsm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  96. int (*sbtrsm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  97. int (*sbtrsm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  98. int (*sbtrsm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  99. int (*sbtrsm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  100. int (*sbtrsm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  101. int (*sbtrsm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  102. int (*sbtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  103. int (*sbtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  104. int (*sbtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  105. int (*sbtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  106. int (*sbtrmm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  107. int (*sbtrmm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  108. int (*sbtrmm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  109. int (*sbtrmm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  110. int (*sbtrmm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  111. int (*sbtrmm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  112. int (*sbtrmm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  113. int (*sbtrmm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  114. int (*sbtrmm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  115. int (*sbtrmm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  116. int (*sbtrmm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  117. int (*sbtrmm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  118. int (*sbtrmm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  119. int (*sbtrmm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  120. int (*sbtrmm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  121. int (*sbtrmm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  122. int (*sbsymm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  123. int (*sbsymm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  124. int (*sbsymm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  125. int (*sbsymm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  126. int (*sbneg_tcopy) (BLASLONG, BLASLONG, float *, BLASLONG, float *);
  127. int (*sblaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG, blasint *, float *);
  128. #ifdef SMALL_MATRIX_OPT
  129. int (*sbgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float beta);
  130. int (*sbgemm_small_kernel_nn )(BLASLONG m, BLASLONG n, BLASLONG k, bfloat16 * A, BLASLONG lda, float alpha, bfloat16 * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
  131. int (*sbgemm_small_kernel_nt )(BLASLONG m, BLASLONG n, BLASLONG k, bfloat16 * A, BLASLONG lda, float alpha, bfloat16 * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
  132. int (*sbgemm_small_kernel_tn )(BLASLONG m, BLASLONG n, BLASLONG k, bfloat16 * A, BLASLONG lda, float alpha, bfloat16 * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
  133. int (*sbgemm_small_kernel_tt )(BLASLONG m, BLASLONG n, BLASLONG k, bfloat16 * A, BLASLONG lda, float alpha, bfloat16 * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
  134. int (*sbgemm_small_kernel_b0_nn )(BLASLONG m, BLASLONG n, BLASLONG k, bfloat16 * A, BLASLONG lda, float alpha, bfloat16 * B, BLASLONG ldb, float * C, BLASLONG ldc);
  135. int (*sbgemm_small_kernel_b0_nt )(BLASLONG m, BLASLONG n, BLASLONG k, bfloat16 * A, BLASLONG lda, float alpha, bfloat16 * B, BLASLONG ldb, float * C, BLASLONG ldc);
  136. int (*sbgemm_small_kernel_b0_tn )(BLASLONG m, BLASLONG n, BLASLONG k, bfloat16 * A, BLASLONG lda, float alpha, bfloat16 * B, BLASLONG ldb, float * C, BLASLONG ldc);
  137. int (*sbgemm_small_kernel_b0_tt )(BLASLONG m, BLASLONG n, BLASLONG k, bfloat16 * A, BLASLONG lda, float alpha, bfloat16 * B, BLASLONG ldb, float * C, BLASLONG ldc);
  138. #endif
  139. #endif
  140. #if defined(BUILD_SINGLE) || defined(BUILD_COMPLEX)
  141. int sgemm_p, sgemm_q, sgemm_r;
  142. int sgemm_unroll_m, sgemm_unroll_n, sgemm_unroll_mn;
  143. #endif
  144. int exclusive_cache;
  145. #if defined(BUILD_SINGLE) || defined(BUILD_COMPLEX)
  146. float (*samax_k) (BLASLONG, float *, BLASLONG);
  147. float (*samin_k) (BLASLONG, float *, BLASLONG);
  148. float (*smax_k) (BLASLONG, float *, BLASLONG);
  149. float (*smin_k) (BLASLONG, float *, BLASLONG);
  150. BLASLONG (*isamax_k)(BLASLONG, float *, BLASLONG);
  151. BLASLONG (*isamin_k)(BLASLONG, float *, BLASLONG);
  152. BLASLONG (*ismax_k) (BLASLONG, float *, BLASLONG);
  153. BLASLONG (*ismin_k) (BLASLONG, float *, BLASLONG);
  154. float (*snrm2_k) (BLASLONG, float *, BLASLONG);
  155. float (*sasum_k) (BLASLONG, float *, BLASLONG);
  156. #endif
  157. #ifdef BUILD_SINGLE
  158. float (*ssum_k) (BLASLONG, float *, BLASLONG);
  159. #endif
  160. #if defined(BUILD_SINGLE) || defined(BUILD_COMPLEX)
  161. int (*scopy_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  162. float (*sdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  163. //double (*dsdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  164. int (*srot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG, float, float);
  165. int (*saxpy_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  166. int (*sscal_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  167. int (*sswap_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  168. int (*sgemv_n) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  169. int (*sgemv_t) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  170. #endif
  171. #ifdef BUILD_SINGLE
  172. int (*sger_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  173. int (*ssymv_L) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  174. int (*ssymv_U) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  175. #endif
  176. #if defined(BUILD_SINGLE) || defined(BUILD_COMPLEX)
  177. #ifdef ARCH_X86_64
  178. void (*sgemm_direct) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG , float *, BLASLONG , float * , BLASLONG);
  179. int (*sgemm_direct_performant) (BLASLONG M, BLASLONG N, BLASLONG K);
  180. #endif
  181. int (*sgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG);
  182. int (*sgemm_beta )(BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  183. int (*sgemm_incopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  184. int (*sgemm_itcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  185. int (*sgemm_oncopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  186. int (*sgemm_otcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  187. #endif
  188. #ifdef BUILD_SINGLE
  189. #ifdef SMALL_MATRIX_OPT
  190. int (*sgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float beta);
  191. int (*sgemm_small_kernel_nn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
  192. int (*sgemm_small_kernel_nt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
  193. int (*sgemm_small_kernel_tn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
  194. int (*sgemm_small_kernel_tt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
  195. int (*sgemm_small_kernel_b0_nn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  196. int (*sgemm_small_kernel_b0_nt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  197. int (*sgemm_small_kernel_b0_tn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  198. int (*sgemm_small_kernel_b0_tt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  199. #endif
  200. int (*strsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  201. int (*strsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  202. int (*strsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  203. int (*strsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  204. int (*strsm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  205. int (*strsm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  206. int (*strsm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  207. int (*strsm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  208. int (*strsm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  209. int (*strsm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  210. int (*strsm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  211. int (*strsm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  212. int (*strsm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  213. int (*strsm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  214. int (*strsm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  215. int (*strsm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  216. int (*strsm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  217. int (*strsm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  218. int (*strsm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  219. int (*strsm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  220. int (*strmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  221. int (*strmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  222. int (*strmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  223. int (*strmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  224. int (*strmm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  225. int (*strmm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  226. int (*strmm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  227. int (*strmm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  228. int (*strmm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  229. int (*strmm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  230. int (*strmm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  231. int (*strmm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  232. int (*strmm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  233. int (*strmm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  234. int (*strmm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  235. int (*strmm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  236. int (*strmm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  237. int (*strmm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  238. int (*strmm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  239. int (*strmm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  240. int (*ssymm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  241. int (*ssymm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  242. int (*ssymm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  243. int (*ssymm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  244. int (*sneg_tcopy) (BLASLONG, BLASLONG, float *, BLASLONG, float *);
  245. int (*slaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG, blasint *, float *);
  246. #endif
  247. #if defined(BUILD_DOUBLE) || defined(BUILD_COMPLEX16)
  248. int dgemm_p, dgemm_q, dgemm_r;
  249. int dgemm_unroll_m, dgemm_unroll_n, dgemm_unroll_mn;
  250. #endif
  251. #if defined(BUILD_DOUBLE) || defined(BUILD_COMPLEX16)
  252. double (*damax_k) (BLASLONG, double *, BLASLONG);
  253. double (*damin_k) (BLASLONG, double *, BLASLONG);
  254. double (*dmax_k) (BLASLONG, double *, BLASLONG);
  255. double (*dmin_k) (BLASLONG, double *, BLASLONG);
  256. BLASLONG (*idamax_k)(BLASLONG, double *, BLASLONG);
  257. BLASLONG (*idamin_k)(BLASLONG, double *, BLASLONG);
  258. BLASLONG (*idmax_k) (BLASLONG, double *, BLASLONG);
  259. BLASLONG (*idmin_k) (BLASLONG, double *, BLASLONG);
  260. #endif
  261. #if defined(BUILD_DOUBLE) || defined(BUILD_COMPLEX16)
  262. double (*dnrm2_k) (BLASLONG, double *, BLASLONG);
  263. double (*dasum_k) (BLASLONG, double *, BLASLONG);
  264. #endif
  265. #ifdef BUILD_DOUBLE
  266. double (*dsum_k) (BLASLONG, double *, BLASLONG);
  267. #endif
  268. #if defined(BUILD_DOUBLE) || defined(BUILD_COMPLEX16)
  269. int (*dcopy_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
  270. double (*ddot_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
  271. #endif
  272. #if defined (BUILD_SINGLE) || defined(BUILD_DOUBLE)
  273. double (*dsdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  274. #endif
  275. #if defined(BUILD_DOUBLE) || defined(BUILD_COMPLEX16)
  276. int (*drot_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG, double, double);
  277. int (*daxpy_k) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  278. int (*dscal_k) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  279. int (*dswap_k) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  280. int (*dgemv_n) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  281. int (*dgemv_t) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  282. #endif
  283. #ifdef BUILD_DOUBLE
  284. int (*dger_k) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  285. int (*dsymv_L) (BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  286. int (*dsymv_U) (BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  287. #endif
  288. #if defined(BUILD_DOUBLE) || defined(BUILD_COMPLEX16)
  289. int (*dgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG);
  290. int (*dgemm_beta )(BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  291. int (*dgemm_incopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  292. int (*dgemm_itcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  293. int (*dgemm_oncopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  294. int (*dgemm_otcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  295. #endif
  296. #ifdef BUILD_DOUBLE
  297. #ifdef SMALL_MATRIX_OPT
  298. int (*dgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, double alpha, double beta);
  299. int (*dgemm_small_kernel_nn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double beta, double * C, BLASLONG ldc);
  300. int (*dgemm_small_kernel_nt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double beta, double * C, BLASLONG ldc);
  301. int (*dgemm_small_kernel_tn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double beta, double * C, BLASLONG ldc);
  302. int (*dgemm_small_kernel_tt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double beta, double * C, BLASLONG ldc);
  303. int (*dgemm_small_kernel_b0_nn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  304. int (*dgemm_small_kernel_b0_nt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  305. int (*dgemm_small_kernel_b0_tn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  306. int (*dgemm_small_kernel_b0_tt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  307. #endif
  308. int (*dtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  309. int (*dtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  310. int (*dtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  311. int (*dtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  312. int (*dtrsm_iunucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  313. int (*dtrsm_iunncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  314. int (*dtrsm_iutucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  315. int (*dtrsm_iutncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  316. int (*dtrsm_ilnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  317. int (*dtrsm_ilnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  318. int (*dtrsm_iltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  319. int (*dtrsm_iltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  320. int (*dtrsm_ounucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  321. int (*dtrsm_ounncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  322. int (*dtrsm_outucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  323. int (*dtrsm_outncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  324. int (*dtrsm_olnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  325. int (*dtrsm_olnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  326. int (*dtrsm_oltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  327. int (*dtrsm_oltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  328. int (*dtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  329. int (*dtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  330. int (*dtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  331. int (*dtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  332. int (*dtrmm_iunucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  333. int (*dtrmm_iunncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  334. int (*dtrmm_iutucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  335. int (*dtrmm_iutncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  336. int (*dtrmm_ilnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  337. int (*dtrmm_ilnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  338. int (*dtrmm_iltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  339. int (*dtrmm_iltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  340. int (*dtrmm_ounucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  341. int (*dtrmm_ounncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  342. int (*dtrmm_outucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  343. int (*dtrmm_outncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  344. int (*dtrmm_olnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  345. int (*dtrmm_olnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  346. int (*dtrmm_oltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  347. int (*dtrmm_oltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  348. int (*dsymm_iutcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  349. int (*dsymm_iltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  350. int (*dsymm_outcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  351. int (*dsymm_oltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  352. int (*dneg_tcopy) (BLASLONG, BLASLONG, double *, BLASLONG, double *);
  353. int (*dlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, double *, BLASLONG, blasint *, double *);
  354. #endif
  355. #ifdef EXPRECISION
  356. int qgemm_p, qgemm_q, qgemm_r;
  357. int qgemm_unroll_m, qgemm_unroll_n, qgemm_unroll_mn;
  358. xdouble (*qamax_k) (BLASLONG, xdouble *, BLASLONG);
  359. xdouble (*qamin_k) (BLASLONG, xdouble *, BLASLONG);
  360. xdouble (*qmax_k) (BLASLONG, xdouble *, BLASLONG);
  361. xdouble (*qmin_k) (BLASLONG, xdouble *, BLASLONG);
  362. BLASLONG (*iqamax_k)(BLASLONG, xdouble *, BLASLONG);
  363. BLASLONG (*iqamin_k)(BLASLONG, xdouble *, BLASLONG);
  364. BLASLONG (*iqmax_k) (BLASLONG, xdouble *, BLASLONG);
  365. BLASLONG (*iqmin_k) (BLASLONG, xdouble *, BLASLONG);
  366. xdouble (*qnrm2_k) (BLASLONG, xdouble *, BLASLONG);
  367. xdouble (*qasum_k) (BLASLONG, xdouble *, BLASLONG);
  368. xdouble (*qsum_k) (BLASLONG, xdouble *, BLASLONG);
  369. int (*qcopy_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  370. xdouble (*qdot_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  371. int (*qrot_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble);
  372. int (*qaxpy_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  373. int (*qscal_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  374. int (*qswap_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  375. int (*qgemv_n) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  376. int (*qgemv_t) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  377. int (*qger_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  378. int (*qsymv_L) (BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  379. int (*qsymv_U) (BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  380. int (*qgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  381. int (*qgemm_beta )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  382. int (*qgemm_incopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  383. int (*qgemm_itcopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  384. int (*qgemm_oncopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  385. int (*qgemm_otcopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  386. int (*qtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  387. int (*qtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  388. int (*qtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  389. int (*qtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  390. int (*qtrsm_iunucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  391. int (*qtrsm_iunncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  392. int (*qtrsm_iutucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  393. int (*qtrsm_iutncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  394. int (*qtrsm_ilnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  395. int (*qtrsm_ilnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  396. int (*qtrsm_iltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  397. int (*qtrsm_iltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  398. int (*qtrsm_ounucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  399. int (*qtrsm_ounncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  400. int (*qtrsm_outucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  401. int (*qtrsm_outncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  402. int (*qtrsm_olnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  403. int (*qtrsm_olnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  404. int (*qtrsm_oltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  405. int (*qtrsm_oltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  406. int (*qtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  407. int (*qtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  408. int (*qtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  409. int (*qtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  410. int (*qtrmm_iunucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  411. int (*qtrmm_iunncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  412. int (*qtrmm_iutucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  413. int (*qtrmm_iutncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  414. int (*qtrmm_ilnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  415. int (*qtrmm_ilnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  416. int (*qtrmm_iltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  417. int (*qtrmm_iltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  418. int (*qtrmm_ounucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  419. int (*qtrmm_ounncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  420. int (*qtrmm_outucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  421. int (*qtrmm_outncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  422. int (*qtrmm_olnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  423. int (*qtrmm_olnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  424. int (*qtrmm_oltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  425. int (*qtrmm_oltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  426. int (*qsymm_iutcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  427. int (*qsymm_iltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  428. int (*qsymm_outcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  429. int (*qsymm_oltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  430. int (*qneg_tcopy) (BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  431. int (*qlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, xdouble *, BLASLONG, blasint *, xdouble *);
  432. #endif
  433. #ifdef BUILD_COMPLEX
  434. int cgemm_p, cgemm_q, cgemm_r;
  435. int cgemm_unroll_m, cgemm_unroll_n, cgemm_unroll_mn;
  436. float (*camax_k) (BLASLONG, float *, BLASLONG);
  437. float (*camin_k) (BLASLONG, float *, BLASLONG);
  438. BLASLONG (*icamax_k)(BLASLONG, float *, BLASLONG);
  439. BLASLONG (*icamin_k)(BLASLONG, float *, BLASLONG);
  440. float (*cnrm2_k) (BLASLONG, float *, BLASLONG);
  441. float (*casum_k) (BLASLONG, float *, BLASLONG);
  442. float (*csum_k) (BLASLONG, float *, BLASLONG);
  443. int (*ccopy_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  444. openblas_complex_float (*cdotu_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  445. openblas_complex_float (*cdotc_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  446. int (*csrot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG, float, float);
  447. int (*caxpy_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  448. int (*caxpyc_k)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  449. int (*cscal_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  450. int (*cswap_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  451. int (*cgemv_n) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  452. int (*cgemv_t) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  453. int (*cgemv_r) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  454. int (*cgemv_c) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  455. int (*cgemv_o) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  456. int (*cgemv_u) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  457. int (*cgemv_s) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  458. int (*cgemv_d) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  459. int (*cgeru_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  460. int (*cgerc_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  461. int (*cgerv_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  462. int (*cgerd_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  463. int (*csymv_L) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  464. int (*csymv_U) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  465. int (*chemv_L) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  466. int (*chemv_U) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  467. int (*chemv_M) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  468. int (*chemv_V) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  469. int (*cgemm_kernel_n )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  470. int (*cgemm_kernel_l )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  471. int (*cgemm_kernel_r )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  472. int (*cgemm_kernel_b )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  473. int (*cgemm_beta )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  474. int (*cgemm_incopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  475. int (*cgemm_itcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  476. int (*cgemm_oncopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  477. int (*cgemm_otcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  478. #ifdef SMALL_MATRIX_OPT
  479. int (*cgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, float alpha0, float alpha1, float beta0, float beta1);
  480. int (*cgemm_small_kernel_nn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  481. int (*cgemm_small_kernel_nt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  482. int (*cgemm_small_kernel_nr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  483. int (*cgemm_small_kernel_nc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  484. int (*cgemm_small_kernel_tn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  485. int (*cgemm_small_kernel_tt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  486. int (*cgemm_small_kernel_tr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  487. int (*cgemm_small_kernel_tc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  488. int (*cgemm_small_kernel_rn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  489. int (*cgemm_small_kernel_rt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  490. int (*cgemm_small_kernel_rr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  491. int (*cgemm_small_kernel_rc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  492. int (*cgemm_small_kernel_cn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  493. int (*cgemm_small_kernel_ct )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  494. int (*cgemm_small_kernel_cr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  495. int (*cgemm_small_kernel_cc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  496. int (*cgemm_small_kernel_b0_nn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  497. int (*cgemm_small_kernel_b0_nt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  498. int (*cgemm_small_kernel_b0_nr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  499. int (*cgemm_small_kernel_b0_nc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  500. int (*cgemm_small_kernel_b0_tn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  501. int (*cgemm_small_kernel_b0_tt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  502. int (*cgemm_small_kernel_b0_tr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  503. int (*cgemm_small_kernel_b0_tc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  504. int (*cgemm_small_kernel_b0_rn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  505. int (*cgemm_small_kernel_b0_rt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  506. int (*cgemm_small_kernel_b0_rr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  507. int (*cgemm_small_kernel_b0_rc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  508. int (*cgemm_small_kernel_b0_cn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  509. int (*cgemm_small_kernel_b0_ct )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  510. int (*cgemm_small_kernel_b0_cr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  511. int (*cgemm_small_kernel_b0_cc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  512. #endif
  513. int (*ctrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  514. int (*ctrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  515. int (*ctrsm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  516. int (*ctrsm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  517. int (*ctrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  518. int (*ctrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  519. int (*ctrsm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  520. int (*ctrsm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  521. int (*ctrsm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  522. int (*ctrsm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  523. int (*ctrsm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  524. int (*ctrsm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  525. int (*ctrsm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  526. int (*ctrsm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  527. int (*ctrsm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  528. int (*ctrsm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  529. int (*ctrsm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  530. int (*ctrsm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  531. int (*ctrsm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  532. int (*ctrsm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  533. int (*ctrsm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  534. int (*ctrsm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  535. int (*ctrsm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  536. int (*ctrsm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  537. int (*ctrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  538. int (*ctrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  539. int (*ctrmm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  540. int (*ctrmm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  541. int (*ctrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  542. int (*ctrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  543. int (*ctrmm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  544. int (*ctrmm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  545. int (*ctrmm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  546. int (*ctrmm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  547. int (*ctrmm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  548. int (*ctrmm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  549. int (*ctrmm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  550. int (*ctrmm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  551. int (*ctrmm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  552. int (*ctrmm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  553. int (*ctrmm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  554. int (*ctrmm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  555. int (*ctrmm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  556. int (*ctrmm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  557. int (*ctrmm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  558. int (*ctrmm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  559. int (*ctrmm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  560. int (*ctrmm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  561. int (*csymm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  562. int (*csymm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  563. int (*csymm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  564. int (*csymm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  565. int (*chemm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  566. int (*chemm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  567. int (*chemm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  568. int (*chemm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  569. int cgemm3m_p, cgemm3m_q, cgemm3m_r;
  570. int cgemm3m_unroll_m, cgemm3m_unroll_n, cgemm3m_unroll_mn;
  571. int (*cgemm3m_kernel)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  572. int (*cgemm3m_incopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  573. int (*cgemm3m_incopyr)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  574. int (*cgemm3m_incopyi)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  575. int (*cgemm3m_itcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  576. int (*cgemm3m_itcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  577. int (*cgemm3m_itcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  578. int (*cgemm3m_oncopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  579. int (*cgemm3m_oncopyr)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  580. int (*cgemm3m_oncopyi)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  581. int (*cgemm3m_otcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  582. int (*cgemm3m_otcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  583. int (*cgemm3m_otcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  584. int (*csymm3m_iucopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  585. int (*csymm3m_ilcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  586. int (*csymm3m_iucopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  587. int (*csymm3m_ilcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  588. int (*csymm3m_iucopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  589. int (*csymm3m_ilcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  590. int (*csymm3m_oucopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  591. int (*csymm3m_olcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  592. int (*csymm3m_oucopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  593. int (*csymm3m_olcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  594. int (*csymm3m_oucopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  595. int (*csymm3m_olcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  596. int (*chemm3m_iucopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  597. int (*chemm3m_ilcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  598. int (*chemm3m_iucopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  599. int (*chemm3m_ilcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  600. int (*chemm3m_iucopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  601. int (*chemm3m_ilcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  602. int (*chemm3m_oucopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  603. int (*chemm3m_olcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  604. int (*chemm3m_oucopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  605. int (*chemm3m_olcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  606. int (*chemm3m_oucopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  607. int (*chemm3m_olcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  608. int (*cneg_tcopy) (BLASLONG, BLASLONG, float *, BLASLONG, float *);
  609. int (*claswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG, blasint *, float *);
  610. #endif
  611. #ifdef BUILD_COMPLEX16
  612. int zgemm_p, zgemm_q, zgemm_r;
  613. int zgemm_unroll_m, zgemm_unroll_n, zgemm_unroll_mn;
  614. double (*zamax_k) (BLASLONG, double *, BLASLONG);
  615. double (*zamin_k) (BLASLONG, double *, BLASLONG);
  616. BLASLONG (*izamax_k)(BLASLONG, double *, BLASLONG);
  617. BLASLONG (*izamin_k)(BLASLONG, double *, BLASLONG);
  618. double (*znrm2_k) (BLASLONG, double *, BLASLONG);
  619. double (*zasum_k) (BLASLONG, double *, BLASLONG);
  620. double (*zsum_k) (BLASLONG, double *, BLASLONG);
  621. int (*zcopy_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
  622. openblas_complex_double (*zdotu_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
  623. openblas_complex_double (*zdotc_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
  624. int (*zdrot_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG, double, double);
  625. int (*zaxpy_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  626. int (*zaxpyc_k)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  627. int (*zscal_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  628. int (*zswap_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  629. int (*zgemv_n) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  630. int (*zgemv_t) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  631. int (*zgemv_r) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  632. int (*zgemv_c) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  633. int (*zgemv_o) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  634. int (*zgemv_u) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  635. int (*zgemv_s) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  636. int (*zgemv_d) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  637. int (*zgeru_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  638. int (*zgerc_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  639. int (*zgerv_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  640. int (*zgerd_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  641. int (*zsymv_L) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  642. int (*zsymv_U) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  643. int (*zhemv_L) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  644. int (*zhemv_U) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  645. int (*zhemv_M) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  646. int (*zhemv_V) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  647. int (*zgemm_kernel_n )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  648. int (*zgemm_kernel_l )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  649. int (*zgemm_kernel_r )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  650. int (*zgemm_kernel_b )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  651. int (*zgemm_beta )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  652. int (*zgemm_incopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  653. int (*zgemm_itcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  654. int (*zgemm_oncopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  655. int (*zgemm_otcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  656. #ifdef SMALL_MATRIX_OPT
  657. int (*zgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, double alpha0, double alpha1, double beta0, double beta1);
  658. int (*zgemm_small_kernel_nn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  659. int (*zgemm_small_kernel_nt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  660. int (*zgemm_small_kernel_nr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  661. int (*zgemm_small_kernel_nc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  662. int (*zgemm_small_kernel_tn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  663. int (*zgemm_small_kernel_tt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  664. int (*zgemm_small_kernel_tr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  665. int (*zgemm_small_kernel_tc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  666. int (*zgemm_small_kernel_rn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  667. int (*zgemm_small_kernel_rt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  668. int (*zgemm_small_kernel_rr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  669. int (*zgemm_small_kernel_rc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  670. int (*zgemm_small_kernel_cn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  671. int (*zgemm_small_kernel_ct )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  672. int (*zgemm_small_kernel_cr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  673. int (*zgemm_small_kernel_cc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  674. int (*zgemm_small_kernel_b0_nn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  675. int (*zgemm_small_kernel_b0_nt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  676. int (*zgemm_small_kernel_b0_nr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  677. int (*zgemm_small_kernel_b0_nc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  678. int (*zgemm_small_kernel_b0_tn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  679. int (*zgemm_small_kernel_b0_tt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  680. int (*zgemm_small_kernel_b0_tr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  681. int (*zgemm_small_kernel_b0_tc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  682. int (*zgemm_small_kernel_b0_rn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  683. int (*zgemm_small_kernel_b0_rt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  684. int (*zgemm_small_kernel_b0_rr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  685. int (*zgemm_small_kernel_b0_rc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  686. int (*zgemm_small_kernel_b0_cn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  687. int (*zgemm_small_kernel_b0_ct )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  688. int (*zgemm_small_kernel_b0_cr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  689. int (*zgemm_small_kernel_b0_cc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  690. #endif
  691. int (*ztrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  692. int (*ztrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  693. int (*ztrsm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  694. int (*ztrsm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  695. int (*ztrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  696. int (*ztrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  697. int (*ztrsm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  698. int (*ztrsm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  699. int (*ztrsm_iunucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  700. int (*ztrsm_iunncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  701. int (*ztrsm_iutucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  702. int (*ztrsm_iutncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  703. int (*ztrsm_ilnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  704. int (*ztrsm_ilnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  705. int (*ztrsm_iltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  706. int (*ztrsm_iltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  707. int (*ztrsm_ounucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  708. int (*ztrsm_ounncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  709. int (*ztrsm_outucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  710. int (*ztrsm_outncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  711. int (*ztrsm_olnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  712. int (*ztrsm_olnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  713. int (*ztrsm_oltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  714. int (*ztrsm_oltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  715. int (*ztrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  716. int (*ztrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  717. int (*ztrmm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  718. int (*ztrmm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  719. int (*ztrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  720. int (*ztrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  721. int (*ztrmm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  722. int (*ztrmm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  723. int (*ztrmm_iunucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  724. int (*ztrmm_iunncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  725. int (*ztrmm_iutucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  726. int (*ztrmm_iutncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  727. int (*ztrmm_ilnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  728. int (*ztrmm_ilnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  729. int (*ztrmm_iltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  730. int (*ztrmm_iltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  731. int (*ztrmm_ounucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  732. int (*ztrmm_ounncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  733. int (*ztrmm_outucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  734. int (*ztrmm_outncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  735. int (*ztrmm_olnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  736. int (*ztrmm_olnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  737. int (*ztrmm_oltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  738. int (*ztrmm_oltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  739. int (*zsymm_iutcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  740. int (*zsymm_iltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  741. int (*zsymm_outcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  742. int (*zsymm_oltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  743. int (*zhemm_iutcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  744. int (*zhemm_iltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  745. int (*zhemm_outcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  746. int (*zhemm_oltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  747. int zgemm3m_p, zgemm3m_q, zgemm3m_r;
  748. int zgemm3m_unroll_m, zgemm3m_unroll_n, zgemm3m_unroll_mn;
  749. int (*zgemm3m_kernel)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  750. int (*zgemm3m_incopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  751. int (*zgemm3m_incopyr)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  752. int (*zgemm3m_incopyi)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  753. int (*zgemm3m_itcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  754. int (*zgemm3m_itcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  755. int (*zgemm3m_itcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  756. int (*zgemm3m_oncopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  757. int (*zgemm3m_oncopyr)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  758. int (*zgemm3m_oncopyi)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  759. int (*zgemm3m_otcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  760. int (*zgemm3m_otcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  761. int (*zgemm3m_otcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  762. int (*zsymm3m_iucopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  763. int (*zsymm3m_ilcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  764. int (*zsymm3m_iucopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  765. int (*zsymm3m_ilcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  766. int (*zsymm3m_iucopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  767. int (*zsymm3m_ilcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  768. int (*zsymm3m_oucopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  769. int (*zsymm3m_olcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  770. int (*zsymm3m_oucopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  771. int (*zsymm3m_olcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  772. int (*zsymm3m_oucopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  773. int (*zsymm3m_olcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  774. int (*zhemm3m_iucopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  775. int (*zhemm3m_ilcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  776. int (*zhemm3m_iucopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  777. int (*zhemm3m_ilcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  778. int (*zhemm3m_iucopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  779. int (*zhemm3m_ilcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  780. int (*zhemm3m_oucopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  781. int (*zhemm3m_olcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  782. int (*zhemm3m_oucopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  783. int (*zhemm3m_olcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  784. int (*zhemm3m_oucopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  785. int (*zhemm3m_olcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  786. int (*zneg_tcopy) (BLASLONG, BLASLONG, double *, BLASLONG, double *);
  787. int (*zlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, double *, BLASLONG, blasint *, double *);
  788. #endif
  789. #ifdef EXPRECISION
  790. int xgemm_p, xgemm_q, xgemm_r;
  791. int xgemm_unroll_m, xgemm_unroll_n, xgemm_unroll_mn;
  792. xdouble (*xamax_k) (BLASLONG, xdouble *, BLASLONG);
  793. xdouble (*xamin_k) (BLASLONG, xdouble *, BLASLONG);
  794. BLASLONG (*ixamax_k)(BLASLONG, xdouble *, BLASLONG);
  795. BLASLONG (*ixamin_k)(BLASLONG, xdouble *, BLASLONG);
  796. xdouble (*xnrm2_k) (BLASLONG, xdouble *, BLASLONG);
  797. xdouble (*xasum_k) (BLASLONG, xdouble *, BLASLONG);
  798. xdouble (*xsum_k) (BLASLONG, xdouble *, BLASLONG);
  799. int (*xcopy_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  800. openblas_complex_xdouble (*xdotu_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  801. openblas_complex_xdouble (*xdotc_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  802. int (*xqrot_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble);
  803. int (*xaxpy_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  804. int (*xaxpyc_k)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  805. int (*xscal_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  806. int (*xswap_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  807. int (*xgemv_n) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  808. int (*xgemv_t) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  809. int (*xgemv_r) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  810. int (*xgemv_c) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  811. int (*xgemv_o) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  812. int (*xgemv_u) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  813. int (*xgemv_s) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  814. int (*xgemv_d) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  815. int (*xgeru_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  816. int (*xgerc_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  817. int (*xgerv_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  818. int (*xgerd_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  819. int (*xsymv_L) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  820. int (*xsymv_U) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  821. int (*xhemv_L) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  822. int (*xhemv_U) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  823. int (*xhemv_M) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  824. int (*xhemv_V) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  825. int (*xgemm_kernel_n )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  826. int (*xgemm_kernel_l )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  827. int (*xgemm_kernel_r )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  828. int (*xgemm_kernel_b )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  829. int (*xgemm_beta )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  830. int (*xgemm_incopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  831. int (*xgemm_itcopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  832. int (*xgemm_oncopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  833. int (*xgemm_otcopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  834. int (*xtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  835. int (*xtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  836. int (*xtrsm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  837. int (*xtrsm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  838. int (*xtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  839. int (*xtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  840. int (*xtrsm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  841. int (*xtrsm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  842. int (*xtrsm_iunucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  843. int (*xtrsm_iunncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  844. int (*xtrsm_iutucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  845. int (*xtrsm_iutncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  846. int (*xtrsm_ilnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  847. int (*xtrsm_ilnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  848. int (*xtrsm_iltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  849. int (*xtrsm_iltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  850. int (*xtrsm_ounucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  851. int (*xtrsm_ounncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  852. int (*xtrsm_outucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  853. int (*xtrsm_outncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  854. int (*xtrsm_olnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  855. int (*xtrsm_olnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  856. int (*xtrsm_oltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  857. int (*xtrsm_oltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  858. int (*xtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  859. int (*xtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  860. int (*xtrmm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  861. int (*xtrmm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  862. int (*xtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  863. int (*xtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  864. int (*xtrmm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  865. int (*xtrmm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  866. int (*xtrmm_iunucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  867. int (*xtrmm_iunncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  868. int (*xtrmm_iutucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  869. int (*xtrmm_iutncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  870. int (*xtrmm_ilnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  871. int (*xtrmm_ilnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  872. int (*xtrmm_iltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  873. int (*xtrmm_iltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  874. int (*xtrmm_ounucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  875. int (*xtrmm_ounncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  876. int (*xtrmm_outucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  877. int (*xtrmm_outncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  878. int (*xtrmm_olnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  879. int (*xtrmm_olnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  880. int (*xtrmm_oltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  881. int (*xtrmm_oltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  882. int (*xsymm_iutcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  883. int (*xsymm_iltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  884. int (*xsymm_outcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  885. int (*xsymm_oltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  886. int (*xhemm_iutcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  887. int (*xhemm_iltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  888. int (*xhemm_outcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  889. int (*xhemm_oltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  890. int xgemm3m_p, xgemm3m_q, xgemm3m_r;
  891. int xgemm3m_unroll_m, xgemm3m_unroll_n, xgemm3m_unroll_mn;
  892. int (*xgemm3m_kernel)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  893. int (*xgemm3m_incopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  894. int (*xgemm3m_incopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  895. int (*xgemm3m_incopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  896. int (*xgemm3m_itcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  897. int (*xgemm3m_itcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  898. int (*xgemm3m_itcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  899. int (*xgemm3m_oncopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  900. int (*xgemm3m_oncopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  901. int (*xgemm3m_oncopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  902. int (*xgemm3m_otcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  903. int (*xgemm3m_otcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  904. int (*xgemm3m_otcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  905. int (*xsymm3m_iucopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  906. int (*xsymm3m_ilcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  907. int (*xsymm3m_iucopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  908. int (*xsymm3m_ilcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  909. int (*xsymm3m_iucopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  910. int (*xsymm3m_ilcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  911. int (*xsymm3m_oucopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  912. int (*xsymm3m_olcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  913. int (*xsymm3m_oucopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  914. int (*xsymm3m_olcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  915. int (*xsymm3m_oucopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  916. int (*xsymm3m_olcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  917. int (*xhemm3m_iucopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  918. int (*xhemm3m_ilcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  919. int (*xhemm3m_iucopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  920. int (*xhemm3m_ilcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  921. int (*xhemm3m_iucopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  922. int (*xhemm3m_ilcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  923. int (*xhemm3m_oucopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  924. int (*xhemm3m_olcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  925. int (*xhemm3m_oucopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  926. int (*xhemm3m_olcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  927. int (*xhemm3m_oucopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  928. int (*xhemm3m_olcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  929. int (*xneg_tcopy) (BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  930. int (*xlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, xdouble *, BLASLONG, blasint *, xdouble *);
  931. #endif
  932. void (*init)(void);
  933. int snum_opt, dnum_opt, qnum_opt;
  934. #ifdef BUILD_SINGLE
  935. int (*saxpby_k) (BLASLONG, float, float*, BLASLONG,float, float*, BLASLONG);
  936. #endif
  937. #ifdef BUILD_DOUBLE
  938. int (*daxpby_k) (BLASLONG, double, double*, BLASLONG,double, double*, BLASLONG);
  939. #endif
  940. #ifdef BUILD_COMPLEX
  941. int (*caxpby_k) (BLASLONG, float, float, float*, BLASLONG,float,float, float*, BLASLONG);
  942. #endif
  943. #ifdef BUILD_COMPLEX16
  944. int (*zaxpby_k) (BLASLONG, double, double, double*, BLASLONG,double,double, double*, BLASLONG);
  945. #endif
  946. #ifdef BUILD_SINGLE
  947. int (*somatcopy_k_cn) (BLASLONG, BLASLONG, float, float*, BLASLONG, float*, BLASLONG);
  948. int (*somatcopy_k_ct) (BLASLONG, BLASLONG, float, float*, BLASLONG, float*, BLASLONG);
  949. int (*somatcopy_k_rn) (BLASLONG, BLASLONG, float, float*, BLASLONG, float*, BLASLONG);
  950. int (*somatcopy_k_rt) (BLASLONG, BLASLONG, float, float*, BLASLONG, float*, BLASLONG);
  951. #endif
  952. #ifdef BUILD_DOUBLE
  953. int (*domatcopy_k_cn) (BLASLONG, BLASLONG, double, double*, BLASLONG, double*, BLASLONG);
  954. int (*domatcopy_k_ct) (BLASLONG, BLASLONG, double, double*, BLASLONG, double*, BLASLONG);
  955. int (*domatcopy_k_rn) (BLASLONG, BLASLONG, double, double*, BLASLONG, double*, BLASLONG);
  956. int (*domatcopy_k_rt) (BLASLONG, BLASLONG, double, double*, BLASLONG, double*, BLASLONG);
  957. #endif
  958. #ifdef BUILD_COMPLEX
  959. int (*comatcopy_k_cn) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  960. int (*comatcopy_k_ct) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  961. int (*comatcopy_k_rn) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  962. int (*comatcopy_k_rt) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  963. int (*comatcopy_k_cnc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  964. int (*comatcopy_k_ctc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  965. int (*comatcopy_k_rnc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  966. int (*comatcopy_k_rtc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  967. #endif
  968. #ifdef BUILD_COMPLEX16
  969. int (*zomatcopy_k_cn) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  970. int (*zomatcopy_k_ct) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  971. int (*zomatcopy_k_rn) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  972. int (*zomatcopy_k_rt) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  973. int (*zomatcopy_k_cnc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  974. int (*zomatcopy_k_ctc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  975. int (*zomatcopy_k_rnc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  976. int (*zomatcopy_k_rtc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  977. #endif
  978. #ifdef BUILD_SINGLE
  979. int (*simatcopy_k_cn) (BLASLONG, BLASLONG, float, float*, BLASLONG);
  980. int (*simatcopy_k_ct) (BLASLONG, BLASLONG, float, float*, BLASLONG);
  981. int (*simatcopy_k_rn) (BLASLONG, BLASLONG, float, float*, BLASLONG);
  982. int (*simatcopy_k_rt) (BLASLONG, BLASLONG, float, float*, BLASLONG);
  983. #endif
  984. #ifdef BUILD_DOUBLE
  985. int (*dimatcopy_k_cn) (BLASLONG, BLASLONG, double, double*, BLASLONG);
  986. int (*dimatcopy_k_ct) (BLASLONG, BLASLONG, double, double*, BLASLONG);
  987. int (*dimatcopy_k_rn) (BLASLONG, BLASLONG, double, double*, BLASLONG);
  988. int (*dimatcopy_k_rt) (BLASLONG, BLASLONG, double, double*, BLASLONG);
  989. #endif
  990. #ifdef BUILD_COMPLEX
  991. int (*cimatcopy_k_cn) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  992. int (*cimatcopy_k_ct) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  993. int (*cimatcopy_k_rn) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  994. int (*cimatcopy_k_rt) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  995. int (*cimatcopy_k_cnc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  996. int (*cimatcopy_k_ctc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  997. int (*cimatcopy_k_rnc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  998. int (*cimatcopy_k_rtc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  999. #endif
  1000. #ifdef BUILD_COMPLEX16
  1001. int (*zimatcopy_k_cn) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1002. int (*zimatcopy_k_ct) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1003. int (*zimatcopy_k_rn) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1004. int (*zimatcopy_k_rt) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1005. int (*zimatcopy_k_cnc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1006. int (*zimatcopy_k_ctc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1007. int (*zimatcopy_k_rnc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1008. int (*zimatcopy_k_rtc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1009. #endif
  1010. #ifdef BUILD_SINGLE
  1011. int (*sgeadd_k) (BLASLONG, BLASLONG, float, float *, BLASLONG, float, float *, BLASLONG);
  1012. #endif
  1013. #ifdef BUILD_DOUBLE
  1014. int (*dgeadd_k) (BLASLONG, BLASLONG, double, double *, BLASLONG, double, double *, BLASLONG);
  1015. #endif
  1016. #ifdef BUILD_COMPLEX
  1017. int (*cgeadd_k) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float, float, float *, BLASLONG);
  1018. #endif
  1019. #ifdef BUILD_COMPLEX16
  1020. int (*zgeadd_k) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double, double, double *, BLASLONG);
  1021. #endif
  1022. } gotoblas_t;
  1023. extern gotoblas_t *gotoblas;
  1024. #define FUNC_OFFSET(func) (size_t)(&((gotoblas_t *)NULL)->func)
  1025. #define DTB_ENTRIES gotoblas -> dtb_entries
  1026. #define GEMM_OFFSET_A gotoblas -> offsetA
  1027. #define GEMM_OFFSET_B gotoblas -> offsetB
  1028. #define GEMM_ALIGN gotoblas -> align
  1029. #define HAVE_EX_L2 gotoblas -> exclusive_cache
  1030. #ifdef BUILD_BFLOAT16
  1031. #define SBGEMM_P gotoblas -> sbgemm_p
  1032. #define SBGEMM_Q gotoblas -> sbgemm_q
  1033. #define SBGEMM_R gotoblas -> sbgemm_r
  1034. #define SBGEMM_UNROLL_M gotoblas -> sbgemm_unroll_m
  1035. #define SBGEMM_UNROLL_N gotoblas -> sbgemm_unroll_n
  1036. #define SBGEMM_UNROLL_MN gotoblas -> sbgemm_unroll_mn
  1037. #endif
  1038. #if defined (BUILD_SINGLE)
  1039. #define SGEMM_P gotoblas -> sgemm_p
  1040. #define SGEMM_Q gotoblas -> sgemm_q
  1041. #define SGEMM_R gotoblas -> sgemm_r
  1042. #define SGEMM_UNROLL_M gotoblas -> sgemm_unroll_m
  1043. #define SGEMM_UNROLL_N gotoblas -> sgemm_unroll_n
  1044. #define SGEMM_UNROLL_MN gotoblas -> sgemm_unroll_mn
  1045. #endif
  1046. #if defined (BUILD_DOUBLE)
  1047. #define DGEMM_P gotoblas -> dgemm_p
  1048. #define DGEMM_Q gotoblas -> dgemm_q
  1049. #define DGEMM_R gotoblas -> dgemm_r
  1050. #define DGEMM_UNROLL_M gotoblas -> dgemm_unroll_m
  1051. #define DGEMM_UNROLL_N gotoblas -> dgemm_unroll_n
  1052. #define DGEMM_UNROLL_MN gotoblas -> dgemm_unroll_mn
  1053. #endif
  1054. #define QGEMM_P gotoblas -> qgemm_p
  1055. #define QGEMM_Q gotoblas -> qgemm_q
  1056. #define QGEMM_R gotoblas -> qgemm_r
  1057. #define QGEMM_UNROLL_M gotoblas -> qgemm_unroll_m
  1058. #define QGEMM_UNROLL_N gotoblas -> qgemm_unroll_n
  1059. #define QGEMM_UNROLL_MN gotoblas -> qgemm_unroll_mn
  1060. #ifdef BUILD_COMPLEX
  1061. #define CGEMM_P gotoblas -> cgemm_p
  1062. #define CGEMM_Q gotoblas -> cgemm_q
  1063. #define CGEMM_R gotoblas -> cgemm_r
  1064. #define CGEMM_UNROLL_M gotoblas -> cgemm_unroll_m
  1065. #define CGEMM_UNROLL_N gotoblas -> cgemm_unroll_n
  1066. #define CGEMM_UNROLL_MN gotoblas -> cgemm_unroll_mn
  1067. #ifndef BUILD_SINGLE
  1068. #define SGEMM_P gotoblas -> sgemm_p
  1069. #define SGEMM_Q gotoblas -> sgemm_q
  1070. #define SGEMM_R 1024
  1071. #define SGEMM_UNROLL_M gotoblas -> sgemm_unroll_m
  1072. #define SGEMM_UNROLL_N gotoblas -> sgemm_unroll_n
  1073. #define SGEMM_UNROLL_MN gotoblas -> sgemm_unroll_mn
  1074. #endif
  1075. #endif
  1076. #ifdef BUILD_COMPLEX16
  1077. #define ZGEMM_P gotoblas -> zgemm_p
  1078. #define ZGEMM_Q gotoblas -> zgemm_q
  1079. #define ZGEMM_R gotoblas -> zgemm_r
  1080. #define ZGEMM_UNROLL_M gotoblas -> zgemm_unroll_m
  1081. #define ZGEMM_UNROLL_N gotoblas -> zgemm_unroll_n
  1082. #define ZGEMM_UNROLL_MN gotoblas -> zgemm_unroll_mn
  1083. #ifndef BUILD_DOUBLE
  1084. #define DGEMM_P gotoblas -> dgemm_p
  1085. #define DGEMM_Q gotoblas -> dgemm_q
  1086. #define DGEMM_R 1024
  1087. #define DGEMM_UNROLL_M gotoblas -> dgemm_unroll_m
  1088. #define DGEMM_UNROLL_N gotoblas -> dgemm_unroll_n
  1089. #define DGEMM_UNROLL_MN gotoblas -> dgemm_unroll_mn
  1090. #endif
  1091. #endif
  1092. #define XGEMM_P gotoblas -> xgemm_p
  1093. #define XGEMM_Q gotoblas -> xgemm_q
  1094. #define XGEMM_R gotoblas -> xgemm_r
  1095. #define XGEMM_UNROLL_M gotoblas -> xgemm_unroll_m
  1096. #define XGEMM_UNROLL_N gotoblas -> xgemm_unroll_n
  1097. #define XGEMM_UNROLL_MN gotoblas -> xgemm_unroll_mn
  1098. #define CGEMM3M_P gotoblas -> cgemm3m_p
  1099. #define CGEMM3M_Q gotoblas -> cgemm3m_q
  1100. #define CGEMM3M_R gotoblas -> cgemm3m_r
  1101. #define CGEMM3M_UNROLL_M gotoblas -> cgemm3m_unroll_m
  1102. #define CGEMM3M_UNROLL_N gotoblas -> cgemm3m_unroll_n
  1103. #define CGEMM3M_UNROLL_MN gotoblas -> cgemm3m_unroll_mn
  1104. #define ZGEMM3M_P gotoblas -> zgemm3m_p
  1105. #define ZGEMM3M_Q gotoblas -> zgemm3m_q
  1106. #define ZGEMM3M_R gotoblas -> zgemm3m_r
  1107. #define ZGEMM3M_UNROLL_M gotoblas -> zgemm3m_unroll_m
  1108. #define ZGEMM3M_UNROLL_N gotoblas -> zgemm3m_unroll_n
  1109. #define ZGEMM3M_UNROLL_MN gotoblas -> zgemm3m_unroll_mn
  1110. #define XGEMM3M_P gotoblas -> xgemm3m_p
  1111. #define XGEMM3M_Q gotoblas -> xgemm3m_q
  1112. #define XGEMM3M_R gotoblas -> xgemm3m_r
  1113. #define XGEMM3M_UNROLL_M gotoblas -> xgemm3m_unroll_m
  1114. #define XGEMM3M_UNROLL_N gotoblas -> xgemm3m_unroll_n
  1115. #define XGEMM3M_UNROLL_MN gotoblas -> xgemm3m_unroll_mn
  1116. #else
  1117. #define FUNC_OFFSET(func) (size_t)(func)
  1118. #define DTB_ENTRIES DTB_DEFAULT_ENTRIES
  1119. #define GEMM_OFFSET_A GEMM_DEFAULT_OFFSET_A
  1120. #define GEMM_OFFSET_B GEMM_DEFAULT_OFFSET_B
  1121. #define GEMM_ALIGN GEMM_DEFAULT_ALIGN
  1122. #ifdef HAVE_EXCLUSIVE_CACHE
  1123. #define HAVE_EX_L2 1
  1124. #else
  1125. #define HAVE_EX_L2 0
  1126. #endif
  1127. #ifdef BUILD_BFLOAT16
  1128. #define SBGEMM_P SBGEMM_DEFAULT_P
  1129. #define SBGEMM_Q SBGEMM_DEFAULT_Q
  1130. #define SBGEMM_R SBGEMM_DEFAULT_R
  1131. #define SBGEMM_UNROLL_M SBGEMM_DEFAULT_UNROLL_M
  1132. #define SBGEMM_UNROLL_N SBGEMM_DEFAULT_UNROLL_N
  1133. #ifdef SBGEMM_DEFAULT_UNROLL_MN
  1134. #define SBGEMM_UNROLL_MN SBGEMM_DEFAULT_UNROLL_MN
  1135. #else
  1136. #define SBGEMM_UNROLL_MN MAX((SBGEMM_UNROLL_M), (SBGEMM_UNROLL_N))
  1137. #endif
  1138. #endif
  1139. #define SGEMM_P SGEMM_DEFAULT_P
  1140. #define SGEMM_Q SGEMM_DEFAULT_Q
  1141. #define SGEMM_R SGEMM_DEFAULT_R
  1142. #define SGEMM_UNROLL_M SGEMM_DEFAULT_UNROLL_M
  1143. #define SGEMM_UNROLL_N SGEMM_DEFAULT_UNROLL_N
  1144. #ifdef SGEMM_DEFAULT_UNROLL_MN
  1145. #define SGEMM_UNROLL_MN SGEMM_DEFAULT_UNROLL_MN
  1146. #else
  1147. #define SGEMM_UNROLL_MN MAX((SGEMM_UNROLL_M), (SGEMM_UNROLL_N))
  1148. #endif
  1149. #define DGEMM_P DGEMM_DEFAULT_P
  1150. #define DGEMM_Q DGEMM_DEFAULT_Q
  1151. #define DGEMM_R DGEMM_DEFAULT_R
  1152. #define DGEMM_UNROLL_M DGEMM_DEFAULT_UNROLL_M
  1153. #define DGEMM_UNROLL_N DGEMM_DEFAULT_UNROLL_N
  1154. #ifdef DGEMM_DEFAULT_UNROLL_MN
  1155. #define DGEMM_UNROLL_MN DGEMM_DEFAULT_UNROLL_MN
  1156. #else
  1157. #define DGEMM_UNROLL_MN MAX((DGEMM_UNROLL_M), (DGEMM_UNROLL_N))
  1158. #endif
  1159. #define QGEMM_P QGEMM_DEFAULT_P
  1160. #define QGEMM_Q QGEMM_DEFAULT_Q
  1161. #define QGEMM_R QGEMM_DEFAULT_R
  1162. #define QGEMM_UNROLL_M QGEMM_DEFAULT_UNROLL_M
  1163. #define QGEMM_UNROLL_N QGEMM_DEFAULT_UNROLL_N
  1164. #define QGEMM_UNROLL_MN MAX((QGEMM_UNROLL_M), (QGEMM_UNROLL_N))
  1165. #define CGEMM_P CGEMM_DEFAULT_P
  1166. #define CGEMM_Q CGEMM_DEFAULT_Q
  1167. #define CGEMM_R CGEMM_DEFAULT_R
  1168. #define CGEMM_UNROLL_M CGEMM_DEFAULT_UNROLL_M
  1169. #define CGEMM_UNROLL_N CGEMM_DEFAULT_UNROLL_N
  1170. #ifdef CGEMM_DEFAULT_UNROLL_MN
  1171. #define CGEMM_UNROLL_MN CGEMM_DEFAULT_UNROLL_MN
  1172. #else
  1173. #define CGEMM_UNROLL_MN MAX((CGEMM_UNROLL_M), (CGEMM_UNROLL_N))
  1174. #endif
  1175. #define ZGEMM_P ZGEMM_DEFAULT_P
  1176. #define ZGEMM_Q ZGEMM_DEFAULT_Q
  1177. #define ZGEMM_R ZGEMM_DEFAULT_R
  1178. #define ZGEMM_UNROLL_M ZGEMM_DEFAULT_UNROLL_M
  1179. #define ZGEMM_UNROLL_N ZGEMM_DEFAULT_UNROLL_N
  1180. #ifdef ZGEMM_DEFAULT_UNROLL_MN
  1181. #define ZGEMM_UNROLL_MN ZGEMM_DEFAULT_UNROLL_MN
  1182. #else
  1183. #define ZGEMM_UNROLL_MN MAX((ZGEMM_UNROLL_M), (ZGEMM_UNROLL_N))
  1184. #endif
  1185. #define XGEMM_P XGEMM_DEFAULT_P
  1186. #define XGEMM_Q XGEMM_DEFAULT_Q
  1187. #define XGEMM_R XGEMM_DEFAULT_R
  1188. #define XGEMM_UNROLL_M XGEMM_DEFAULT_UNROLL_M
  1189. #define XGEMM_UNROLL_N XGEMM_DEFAULT_UNROLL_N
  1190. #define XGEMM_UNROLL_MN MAX((XGEMM_UNROLL_M), (XGEMM_UNROLL_N))
  1191. #ifdef CGEMM3M_DEFAULT_UNROLL_N
  1192. #define CGEMM3M_P CGEMM3M_DEFAULT_P
  1193. #define CGEMM3M_Q CGEMM3M_DEFAULT_Q
  1194. #define CGEMM3M_R CGEMM3M_DEFAULT_R
  1195. #define CGEMM3M_UNROLL_M CGEMM3M_DEFAULT_UNROLL_M
  1196. #define CGEMM3M_UNROLL_N CGEMM3M_DEFAULT_UNROLL_N
  1197. #define CGEMM3M_UNROLL_MN MAX((CGEMM3M_UNROLL_M), (CGEMM3M_UNROLL_N))
  1198. #else
  1199. #define CGEMM3M_P SGEMM_DEFAULT_P
  1200. #define CGEMM3M_Q SGEMM_DEFAULT_Q
  1201. #define CGEMM3M_R SGEMM_DEFAULT_R
  1202. #define CGEMM3M_UNROLL_M SGEMM_DEFAULT_UNROLL_M
  1203. #define CGEMM3M_UNROLL_N SGEMM_DEFAULT_UNROLL_N
  1204. #define CGEMM3M_UNROLL_MN MAX((CGEMM_UNROLL_M), (CGEMM_UNROLL_N))
  1205. #endif
  1206. #ifdef ZGEMM3M_DEFAULT_UNROLL_N
  1207. #define ZGEMM3M_P ZGEMM3M_DEFAULT_P
  1208. #define ZGEMM3M_Q ZGEMM3M_DEFAULT_Q
  1209. #define ZGEMM3M_R ZGEMM3M_DEFAULT_R
  1210. #define ZGEMM3M_UNROLL_M ZGEMM3M_DEFAULT_UNROLL_M
  1211. #define ZGEMM3M_UNROLL_N ZGEMM3M_DEFAULT_UNROLL_N
  1212. #define ZGEMM3M_UNROLL_MN MAX((ZGEMM_UNROLL_M), (ZGEMM_UNROLL_N))
  1213. #else
  1214. #define ZGEMM3M_P DGEMM_DEFAULT_P
  1215. #define ZGEMM3M_Q DGEMM_DEFAULT_Q
  1216. #define ZGEMM3M_R DGEMM_DEFAULT_R
  1217. #define ZGEMM3M_UNROLL_M DGEMM_DEFAULT_UNROLL_M
  1218. #define ZGEMM3M_UNROLL_N DGEMM_DEFAULT_UNROLL_N
  1219. #define ZGEMM3M_UNROLL_MN MAX((ZGEMM_UNROLL_M), (ZGEMM_UNROLL_N))
  1220. #endif
  1221. #define XGEMM3M_P QGEMM_DEFAULT_P
  1222. #define XGEMM3M_Q QGEMM_DEFAULT_Q
  1223. #define XGEMM3M_R QGEMM_DEFAULT_R
  1224. #define XGEMM3M_UNROLL_M QGEMM_DEFAULT_UNROLL_M
  1225. #define XGEMM3M_UNROLL_N QGEMM_DEFAULT_UNROLL_N
  1226. #define XGEMM3M_UNROLL_MN MAX((QGEMM_UNROLL_M), (QGEMM_UNROLL_N))
  1227. #endif
  1228. #endif
  1229. #ifndef COMPLEX
  1230. #if defined(XDOUBLE)
  1231. #define GEMM_P QGEMM_P
  1232. #define GEMM_Q QGEMM_Q
  1233. #define GEMM_R QGEMM_R
  1234. #define GEMM_UNROLL_M QGEMM_UNROLL_M
  1235. #define GEMM_UNROLL_N QGEMM_UNROLL_N
  1236. #define GEMM_UNROLL_MN QGEMM_UNROLL_MN
  1237. #define GEMM_DEFAULT_P QGEMM_DEFAULT_P
  1238. #define GEMM_DEFAULT_Q QGEMM_DEFAULT_Q
  1239. #define GEMM_DEFAULT_R QGEMM_DEFAULT_R
  1240. #define GEMM_DEFAULT_UNROLL_M QGEMM_DEFAULT_UNROLL_M
  1241. #define GEMM_DEFAULT_UNROLL_N QGEMM_DEFAULT_UNROLL_N
  1242. #elif defined(DOUBLE)
  1243. #define GEMM_P DGEMM_P
  1244. #define GEMM_Q DGEMM_Q
  1245. #define GEMM_R DGEMM_R
  1246. #define GEMM_UNROLL_M DGEMM_UNROLL_M
  1247. #define GEMM_UNROLL_N DGEMM_UNROLL_N
  1248. #define GEMM_UNROLL_MN DGEMM_UNROLL_MN
  1249. #define GEMM_DEFAULT_P DGEMM_DEFAULT_P
  1250. #define GEMM_DEFAULT_Q DGEMM_DEFAULT_Q
  1251. #define GEMM_DEFAULT_R DGEMM_DEFAULT_R
  1252. #define GEMM_DEFAULT_UNROLL_M DGEMM_DEFAULT_UNROLL_M
  1253. #define GEMM_DEFAULT_UNROLL_N DGEMM_DEFAULT_UNROLL_N
  1254. #elif defined(BFLOAT16)
  1255. #define GEMM_P SBGEMM_P
  1256. #define GEMM_Q SBGEMM_Q
  1257. #define GEMM_R SBGEMM_R
  1258. #define GEMM_UNROLL_M SBGEMM_UNROLL_M
  1259. #define GEMM_UNROLL_N SBGEMM_UNROLL_N
  1260. #define GEMM_UNROLL_MN SBGEMM_UNROLL_MN
  1261. #define GEMM_DEFAULT_P SBGEMM_DEFAULT_P
  1262. #define GEMM_DEFAULT_Q SBGEMM_DEFAULT_Q
  1263. #define GEMM_DEFAULT_R SBGEMM_DEFAULT_R
  1264. #define GEMM_DEFAULT_UNROLL_M SBGEMM_DEFAULT_UNROLL_M
  1265. #define GEMM_DEFAULT_UNROLL_N SBGEMM_DEFAULT_UNROLL_N
  1266. #else
  1267. #define GEMM_P SGEMM_P
  1268. #define GEMM_Q SGEMM_Q
  1269. #define GEMM_R SGEMM_R
  1270. #define GEMM_UNROLL_M SGEMM_UNROLL_M
  1271. #define GEMM_UNROLL_N SGEMM_UNROLL_N
  1272. #define GEMM_UNROLL_MN SGEMM_UNROLL_MN
  1273. #define GEMM_DEFAULT_P SGEMM_DEFAULT_P
  1274. #define GEMM_DEFAULT_Q SGEMM_DEFAULT_Q
  1275. #define GEMM_DEFAULT_R SGEMM_DEFAULT_R
  1276. #define GEMM_DEFAULT_UNROLL_M SGEMM_DEFAULT_UNROLL_M
  1277. #define GEMM_DEFAULT_UNROLL_N SGEMM_DEFAULT_UNROLL_N
  1278. #endif
  1279. #else
  1280. #if defined(XDOUBLE)
  1281. #define GEMM_P XGEMM_P
  1282. #define GEMM_Q XGEMM_Q
  1283. #define GEMM_R XGEMM_R
  1284. #define GEMM_UNROLL_M XGEMM_UNROLL_M
  1285. #define GEMM_UNROLL_N XGEMM_UNROLL_N
  1286. #define GEMM_UNROLL_MN XGEMM_UNROLL_MN
  1287. #define GEMM_DEFAULT_P XGEMM_DEFAULT_P
  1288. #define GEMM_DEFAULT_Q XGEMM_DEFAULT_Q
  1289. #define GEMM_DEFAULT_R XGEMM_DEFAULT_R
  1290. #define GEMM_DEFAULT_UNROLL_M XGEMM_DEFAULT_UNROLL_M
  1291. #define GEMM_DEFAULT_UNROLL_N XGEMM_DEFAULT_UNROLL_N
  1292. #elif defined(DOUBLE)
  1293. #define GEMM_P ZGEMM_P
  1294. #define GEMM_Q ZGEMM_Q
  1295. #define GEMM_R ZGEMM_R
  1296. #define GEMM_UNROLL_M ZGEMM_UNROLL_M
  1297. #define GEMM_UNROLL_N ZGEMM_UNROLL_N
  1298. #define GEMM_UNROLL_MN ZGEMM_UNROLL_MN
  1299. #define GEMM_DEFAULT_P ZGEMM_DEFAULT_P
  1300. #define GEMM_DEFAULT_Q ZGEMM_DEFAULT_Q
  1301. #define GEMM_DEFAULT_R ZGEMM_DEFAULT_R
  1302. #define GEMM_DEFAULT_UNROLL_M ZGEMM_DEFAULT_UNROLL_M
  1303. #define GEMM_DEFAULT_UNROLL_N ZGEMM_DEFAULT_UNROLL_N
  1304. #else
  1305. #define GEMM_P CGEMM_P
  1306. #define GEMM_Q CGEMM_Q
  1307. #define GEMM_R CGEMM_R
  1308. #define GEMM_UNROLL_M CGEMM_UNROLL_M
  1309. #define GEMM_UNROLL_N CGEMM_UNROLL_N
  1310. #define GEMM_UNROLL_MN CGEMM_UNROLL_MN
  1311. #define GEMM_DEFAULT_P CGEMM_DEFAULT_P
  1312. #define GEMM_DEFAULT_Q CGEMM_DEFAULT_Q
  1313. #define GEMM_DEFAULT_R CGEMM_DEFAULT_R
  1314. #define GEMM_DEFAULT_UNROLL_M CGEMM_DEFAULT_UNROLL_M
  1315. #define GEMM_DEFAULT_UNROLL_N CGEMM_DEFAULT_UNROLL_N
  1316. #endif
  1317. #endif
  1318. #ifdef XDOUBLE
  1319. #define GEMM3M_UNROLL_M XGEMM3M_UNROLL_M
  1320. #define GEMM3M_UNROLL_N XGEMM3M_UNROLL_N
  1321. #elif defined(DOUBLE)
  1322. #define GEMM3M_UNROLL_M ZGEMM3M_UNROLL_M
  1323. #define GEMM3M_UNROLL_N ZGEMM3M_UNROLL_N
  1324. #else
  1325. #define GEMM3M_UNROLL_M CGEMM3M_UNROLL_M
  1326. #define GEMM3M_UNROLL_N CGEMM3M_UNROLL_N
  1327. #endif
  1328. #ifndef QGEMM_DEFAULT_UNROLL_M
  1329. #define QGEMM_DEFAULT_UNROLL_M 2
  1330. #endif
  1331. #ifndef QGEMM_DEFAULT_UNROLL_N
  1332. #define QGEMM_DEFAULT_UNROLL_N 2
  1333. #endif
  1334. #ifndef XGEMM_DEFAULT_UNROLL_M
  1335. #define XGEMM_DEFAULT_UNROLL_M 2
  1336. #endif
  1337. #ifndef XGEMM_DEFAULT_UNROLL_N
  1338. #define XGEMM_DEFAULT_UNROLL_N 2
  1339. #endif
  1340. #ifndef GEMM_THREAD
  1341. #define GEMM_THREAD gemm_thread_n
  1342. #endif
  1343. #ifndef SBGEMM_DEFAULT_R
  1344. #define SBGEMM_DEFAULT_R (((BUFFER_SIZE - ((SBGEMM_DEFAULT_P * SBGEMM_DEFAULT_Q * 4 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (SBGEMM_DEFAULT_Q * 4) - 15) & ~15UL)
  1345. #endif
  1346. #ifndef SGEMM_DEFAULT_R
  1347. #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)
  1348. #endif
  1349. #ifndef DGEMM_DEFAULT_R
  1350. #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)
  1351. #endif
  1352. #ifndef QGEMM_DEFAULT_R
  1353. #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)
  1354. #endif
  1355. #ifndef CGEMM_DEFAULT_R
  1356. #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)
  1357. #endif
  1358. #ifndef ZGEMM_DEFAULT_R
  1359. #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)
  1360. #endif
  1361. #ifndef XGEMM_DEFAULT_R
  1362. #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)
  1363. #endif
  1364. #ifndef SNUMOPT
  1365. #define SNUMOPT 2
  1366. #endif
  1367. #ifndef DNUMOPT
  1368. #define DNUMOPT 2
  1369. #endif
  1370. #ifndef QNUMOPT
  1371. #define QNUMOPT 1
  1372. #endif
  1373. #ifndef GEMM3M_P
  1374. #ifdef XDOUBLE
  1375. #define GEMM3M_P XGEMM3M_P
  1376. #elif defined(DOUBLE)
  1377. #define GEMM3M_P ZGEMM3M_P
  1378. #else
  1379. #define GEMM3M_P CGEMM3M_P
  1380. #endif
  1381. #endif
  1382. #ifndef GEMM3M_Q
  1383. #ifdef XDOUBLE
  1384. #define GEMM3M_Q XGEMM3M_Q
  1385. #elif defined(DOUBLE)
  1386. #define GEMM3M_Q ZGEMM3M_Q
  1387. #else
  1388. #define GEMM3M_Q CGEMM3M_Q
  1389. #endif
  1390. #endif
  1391. #ifndef GEMM3M_R
  1392. #ifdef XDOUBLE
  1393. #define GEMM3M_R XGEMM3M_R
  1394. #elif defined(DOUBLE)
  1395. #define GEMM3M_R ZGEMM3M_R
  1396. #else
  1397. #define GEMM3M_R CGEMM3M_R
  1398. #endif
  1399. #endif
  1400. #endif