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_level3.h 149 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824
  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 ASSEMBLER
  39. #ifdef __CUDACC__
  40. __global__ void cuda_sgemm_kernel(int, int, int, float *, float *, float *);
  41. __global__ void cuda_dgemm_kernel(int, int, int, double *, double *, double *);
  42. #endif
  43. #ifdef __CUDACC__
  44. extern "C" {
  45. #endif
  46. void sgemm_direct(BLASLONG M, BLASLONG N, BLASLONG K,
  47. float * A, BLASLONG strideA,
  48. float * B, BLASLONG strideB,
  49. float * R, BLASLONG strideR);
  50. int sgemm_direct_performant(BLASLONG M, BLASLONG N, BLASLONG K);
  51. int shgemm_beta(BLASLONG, BLASLONG, BLASLONG, float,
  52. bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, float *, BLASLONG);
  53. int sgemm_beta(BLASLONG, BLASLONG, BLASLONG, float,
  54. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  55. int dgemm_beta(BLASLONG, BLASLONG, BLASLONG, double,
  56. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  57. int cgemm_beta(BLASLONG, BLASLONG, BLASLONG, float, float,
  58. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  59. int zgemm_beta(BLASLONG, BLASLONG, BLASLONG, double, double,
  60. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  61. #ifdef EXPRECISION
  62. int qgemm_beta(BLASLONG, BLASLONG, BLASLONG, xdouble,
  63. xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  64. int xgemm_beta(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble,
  65. xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  66. #else
  67. int qgemm_beta(BLASLONG, BLASLONG, BLASLONG, xdouble *,
  68. xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  69. int xgemm_beta(BLASLONG, BLASLONG, BLASLONG, xdouble *,
  70. xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  71. #endif
  72. int shgemm_incopy(BLASLONG m, BLASLONG n, bfloat16 *a, BLASLONG lda, bfloat16 *b);
  73. int shgemm_itcopy(BLASLONG m, BLASLONG n, bfloat16 *a, BLASLONG lda, bfloat16 *b);
  74. int shgemm_oncopy(BLASLONG m, BLASLONG n, bfloat16 *a, BLASLONG lda, bfloat16 *b);
  75. int shgemm_otcopy(BLASLONG m, BLASLONG n, bfloat16 *a, BLASLONG lda, bfloat16 *b);
  76. int sgemm_incopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  77. int sgemm_itcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  78. int sgemm_oncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  79. int sgemm_otcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  80. int dgemm_incopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  81. int dgemm_itcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  82. int dgemm_oncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  83. int dgemm_otcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  84. int cgemm_incopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  85. int cgemm_itcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  86. int cgemm_oncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  87. int cgemm_otcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  88. int zgemm_incopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  89. int zgemm_itcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  90. int zgemm_oncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  91. int zgemm_otcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  92. #ifdef QUAD_PRECISION
  93. int qgemm_incopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xidouble *b);
  94. int qgemm_itcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xidouble *b);
  95. int qgemm_oncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xidouble *b);
  96. int qgemm_otcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xidouble *b);
  97. int xgemm_incopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xidouble *b);
  98. int xgemm_itcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xidouble *b);
  99. int xgemm_oncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xidouble *b);
  100. int xgemm_otcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xidouble *b);
  101. #else
  102. int qgemm_incopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  103. int qgemm_itcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  104. int qgemm_oncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  105. int qgemm_otcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  106. int xgemm_incopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  107. int xgemm_itcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  108. int xgemm_oncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  109. int xgemm_otcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  110. #endif
  111. int strsm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  112. int strsm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  113. int strsm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  114. int strsm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  115. int dtrsm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  116. int dtrsm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  117. int dtrsm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  118. int dtrsm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  119. int qtrsm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  120. int qtrsm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  121. int qtrsm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  122. int qtrsm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  123. int ctrsm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  124. int ctrsm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  125. int ctrsm_kernel_LR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  126. int ctrsm_kernel_LC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  127. int ctrsm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  128. int ctrsm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  129. int ctrsm_kernel_RR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  130. int ctrsm_kernel_RC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  131. int ztrsm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  132. int ztrsm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  133. int ztrsm_kernel_LR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  134. int ztrsm_kernel_LC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  135. int ztrsm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  136. int ztrsm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  137. int ztrsm_kernel_RR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  138. int ztrsm_kernel_RC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  139. int xtrsm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  140. int xtrsm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  141. int xtrsm_kernel_LR(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  142. int xtrsm_kernel_LC(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  143. int xtrsm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  144. int xtrsm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  145. int xtrsm_kernel_RR(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  146. int xtrsm_kernel_RC(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  147. int strmm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  148. int strmm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  149. int strmm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  150. int strmm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  151. int dtrmm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  152. int dtrmm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  153. int dtrmm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  154. int dtrmm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  155. int qtrmm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  156. int qtrmm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  157. int qtrmm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  158. int qtrmm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  159. int ctrmm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  160. int ctrmm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  161. int ctrmm_kernel_RR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  162. int ctrmm_kernel_RC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  163. int ctrmm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  164. int ctrmm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  165. int ctrmm_kernel_LR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  166. int ctrmm_kernel_LC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  167. int ztrmm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  168. int ztrmm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  169. int ztrmm_kernel_RR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  170. int ztrmm_kernel_RC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  171. int ztrmm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  172. int ztrmm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  173. int ztrmm_kernel_LR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  174. int ztrmm_kernel_LC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  175. int xtrmm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  176. int xtrmm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  177. int xtrmm_kernel_RR(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  178. int xtrmm_kernel_RC(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  179. int xtrmm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  180. int xtrmm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  181. int xtrmm_kernel_LR(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  182. int xtrmm_kernel_LC(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  183. int strmm_iunucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  184. int strmm_iunncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  185. int strmm_iutucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  186. int strmm_iutncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  187. int strmm_ounucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  188. int strmm_ounncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  189. int strmm_outucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  190. int strmm_outncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  191. int strmm_ilnucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  192. int strmm_ilnncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  193. int strmm_iltucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  194. int strmm_iltncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  195. int strmm_olnucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  196. int strmm_olnncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  197. int strmm_oltucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  198. int strmm_oltncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  199. int dtrmm_iunucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  200. int dtrmm_iunncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  201. int dtrmm_iutucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  202. int dtrmm_iutncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  203. int dtrmm_ounucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  204. int dtrmm_ounncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  205. int dtrmm_outucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  206. int dtrmm_outncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  207. int dtrmm_ilnucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  208. int dtrmm_ilnncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  209. int dtrmm_iltucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  210. int dtrmm_iltncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  211. int dtrmm_olnucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  212. int dtrmm_olnncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  213. int dtrmm_oltucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  214. int dtrmm_oltncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  215. int qtrmm_iunucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  216. int qtrmm_iunncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  217. int qtrmm_iutucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  218. int qtrmm_iutncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  219. int qtrmm_ounucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  220. int qtrmm_ounncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  221. int qtrmm_outucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  222. int qtrmm_outncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  223. int qtrmm_ilnucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  224. int qtrmm_ilnncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  225. int qtrmm_iltucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  226. int qtrmm_iltncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  227. int qtrmm_olnucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  228. int qtrmm_olnncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  229. int qtrmm_oltucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  230. int qtrmm_oltncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  231. int ctrmm_iunucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  232. int ctrmm_iunncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  233. int ctrmm_iutucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  234. int ctrmm_iutncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  235. int ctrmm_ounucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  236. int ctrmm_ounncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  237. int ctrmm_outucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  238. int ctrmm_outncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  239. int ctrmm_ilnucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  240. int ctrmm_ilnncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  241. int ctrmm_iltucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  242. int ctrmm_iltncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  243. int ctrmm_olnucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  244. int ctrmm_olnncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  245. int ctrmm_oltucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  246. int ctrmm_oltncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  247. int ztrmm_iunucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  248. int ztrmm_iunncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  249. int ztrmm_iutucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  250. int ztrmm_iutncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  251. int ztrmm_ounucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  252. int ztrmm_ounncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  253. int ztrmm_outucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  254. int ztrmm_outncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  255. int ztrmm_ilnucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  256. int ztrmm_ilnncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  257. int ztrmm_iltucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  258. int ztrmm_iltncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  259. int ztrmm_olnucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  260. int ztrmm_olnncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  261. int ztrmm_oltucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  262. int ztrmm_oltncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  263. int xtrmm_iunucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  264. int xtrmm_iunncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  265. int xtrmm_iutucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  266. int xtrmm_iutncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  267. int xtrmm_ounucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  268. int xtrmm_ounncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  269. int xtrmm_outucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  270. int xtrmm_outncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  271. int xtrmm_ilnucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  272. int xtrmm_ilnncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  273. int xtrmm_iltucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  274. int xtrmm_iltncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  275. int xtrmm_olnucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  276. int xtrmm_olnncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  277. int xtrmm_oltucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  278. int xtrmm_oltncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  279. int strsm_iunucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  280. int strsm_iunncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  281. int strsm_iutucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  282. int strsm_iutncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  283. int strsm_ounucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  284. int strsm_ounncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  285. int strsm_outucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  286. int strsm_outncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  287. int strsm_ilnucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  288. int strsm_ilnncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  289. int strsm_iltucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  290. int strsm_iltncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  291. int strsm_olnucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  292. int strsm_olnncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  293. int strsm_oltucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  294. int strsm_oltncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  295. int dtrsm_iunucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  296. int dtrsm_iunncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  297. int dtrsm_iutucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  298. int dtrsm_iutncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  299. int dtrsm_ounucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  300. int dtrsm_ounncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  301. int dtrsm_outucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  302. int dtrsm_outncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  303. int dtrsm_ilnucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  304. int dtrsm_ilnncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  305. int dtrsm_iltucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  306. int dtrsm_iltncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  307. int dtrsm_olnucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  308. int dtrsm_olnncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  309. int dtrsm_oltucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  310. int dtrsm_oltncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  311. int qtrsm_iunucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  312. int qtrsm_iunncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  313. int qtrsm_iutucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  314. int qtrsm_iutncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  315. int qtrsm_ounucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  316. int qtrsm_ounncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  317. int qtrsm_outucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  318. int qtrsm_outncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  319. int qtrsm_ilnucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  320. int qtrsm_ilnncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  321. int qtrsm_iltucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  322. int qtrsm_iltncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  323. int qtrsm_olnucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  324. int qtrsm_olnncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  325. int qtrsm_oltucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  326. int qtrsm_oltncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  327. int ctrsm_iunucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  328. int ctrsm_iunncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  329. int ctrsm_iutucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  330. int ctrsm_iutncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  331. int ctrsm_ounucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  332. int ctrsm_ounncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  333. int ctrsm_outucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  334. int ctrsm_outncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  335. int ctrsm_ilnucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  336. int ctrsm_ilnncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  337. int ctrsm_iltucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  338. int ctrsm_iltncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  339. int ctrsm_olnucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  340. int ctrsm_olnncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  341. int ctrsm_oltucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  342. int ctrsm_oltncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  343. int ztrsm_iunucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  344. int ztrsm_iunncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  345. int ztrsm_iutucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  346. int ztrsm_iutncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  347. int ztrsm_ounucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  348. int ztrsm_ounncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  349. int ztrsm_outucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  350. int ztrsm_outncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  351. int ztrsm_ilnucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  352. int ztrsm_ilnncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  353. int ztrsm_iltucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  354. int ztrsm_iltncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  355. int ztrsm_olnucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  356. int ztrsm_olnncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  357. int ztrsm_oltucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  358. int ztrsm_oltncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  359. int xtrsm_iunucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  360. int xtrsm_iunncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  361. int xtrsm_iutucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  362. int xtrsm_iutncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  363. int xtrsm_ounucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  364. int xtrsm_ounncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  365. int xtrsm_outucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  366. int xtrsm_outncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  367. int xtrsm_ilnucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  368. int xtrsm_ilnncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  369. int xtrsm_iltucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  370. int xtrsm_iltncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  371. int xtrsm_olnucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  372. int xtrsm_olnncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  373. int xtrsm_oltucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  374. int xtrsm_oltncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  375. int ssymm_iutcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  376. int ssymm_outcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  377. int ssymm_iltcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  378. int ssymm_oltcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  379. int dsymm_iutcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  380. int dsymm_outcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  381. int dsymm_iltcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  382. int dsymm_oltcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  383. int qsymm_iutcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  384. int qsymm_outcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  385. int qsymm_iltcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  386. int qsymm_oltcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  387. int csymm_iutcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  388. int csymm_outcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  389. int csymm_iltcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  390. int csymm_oltcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  391. int zsymm_iutcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  392. int zsymm_outcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  393. int zsymm_iltcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  394. int zsymm_oltcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  395. int xsymm_iutcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  396. int xsymm_outcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  397. int xsymm_iltcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  398. int xsymm_oltcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  399. int chemm_iutcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  400. int chemm_outcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  401. int chemm_iltcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  402. int chemm_oltcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  403. int zhemm_iutcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  404. int zhemm_outcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  405. int zhemm_iltcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  406. int zhemm_oltcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  407. int xhemm_iutcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  408. int xhemm_outcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  409. int xhemm_iltcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  410. int xhemm_oltcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  411. int ssyrk_kernel_U(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset);
  412. int ssyrk_kernel_L(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset);
  413. int dsyrk_kernel_U(BLASLONG m, BLASLONG n, BLASLONG k, double alpha, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset);
  414. int dsyrk_kernel_L(BLASLONG m, BLASLONG n, BLASLONG k, double alpha, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset);
  415. int qsyrk_kernel_U(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset);
  416. int qsyrk_kernel_L(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset);
  417. int csyrk_kernel_U(BLASLONG m, BLASLONG n, BLASLONG k, float alpha_r, float alpha_i, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset);
  418. int csyrk_kernel_L(BLASLONG m, BLASLONG n, BLASLONG k, float alpha_r, float alpha_i, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset);
  419. int zsyrk_kernel_U(BLASLONG m, BLASLONG n, BLASLONG k, double alpha_r, double alpha_i, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset);
  420. int zsyrk_kernel_L(BLASLONG m, BLASLONG n, BLASLONG k, double alpha_r, double alpha_i, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset);
  421. int xsyrk_kernel_U(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha_r, xdouble alpha_i, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset);
  422. int xsyrk_kernel_L(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha_r, xdouble alpha_i, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset);
  423. int ssyr2k_kernel_U(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset, int flag);
  424. int ssyr2k_kernel_L(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset, int flag);
  425. int dsyr2k_kernel_U(BLASLONG m, BLASLONG n, BLASLONG k, double alpha, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset, int flag);
  426. int dsyr2k_kernel_L(BLASLONG m, BLASLONG n, BLASLONG k, double alpha, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset, int flag);
  427. int qsyr2k_kernel_U(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset, int flag);
  428. int qsyr2k_kernel_L(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset, int flag);
  429. int csyr2k_kernel_U(BLASLONG m, BLASLONG n, BLASLONG k, float alpha_r, float alpha_i, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset, int flag);
  430. int csyr2k_kernel_L(BLASLONG m, BLASLONG n, BLASLONG k, float alpha_r, float alpha_i, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset, int flag);
  431. int zsyr2k_kernel_U(BLASLONG m, BLASLONG n, BLASLONG k, double alpha_r, double alpha_i, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset, int flag);
  432. int zsyr2k_kernel_L(BLASLONG m, BLASLONG n, BLASLONG k, double alpha_r, double alpha_i, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset, int flag);
  433. int xsyr2k_kernel_U(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha_r, xdouble alpha_i, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset, int flag);
  434. int xsyr2k_kernel_L(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha_r, xdouble alpha_i, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset, int flag);
  435. int cherk_kernel_UN(BLASLONG m, BLASLONG n, BLASLONG k, float alpha_r, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset);
  436. int cherk_kernel_UC(BLASLONG m, BLASLONG n, BLASLONG k, float alpha_r, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset);
  437. int cherk_kernel_LN(BLASLONG m, BLASLONG n, BLASLONG k, float alpha_r, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset);
  438. int cherk_kernel_LC(BLASLONG m, BLASLONG n, BLASLONG k, float alpha_r, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset);
  439. int zherk_kernel_UN(BLASLONG m, BLASLONG n, BLASLONG k, double alpha_r, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset);
  440. int zherk_kernel_UC(BLASLONG m, BLASLONG n, BLASLONG k, double alpha_r, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset);
  441. int zherk_kernel_LN(BLASLONG m, BLASLONG n, BLASLONG k, double alpha_r, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset);
  442. int zherk_kernel_LC(BLASLONG m, BLASLONG n, BLASLONG k, double alpha_r, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset);
  443. int xherk_kernel_UN(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha_r, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset);
  444. int xherk_kernel_UC(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha_r, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset);
  445. int xherk_kernel_LN(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha_r, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset);
  446. int xherk_kernel_LC(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha_r, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset);
  447. int cher2k_kernel_UN(BLASLONG m, BLASLONG n, BLASLONG k, float alpha_r, float alpha_i, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset, int flag);
  448. int cher2k_kernel_UC(BLASLONG m, BLASLONG n, BLASLONG k, float alpha_r, float alpha_i, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset, int flag);
  449. int cher2k_kernel_LN(BLASLONG m, BLASLONG n, BLASLONG k, float alpha_r, float alpha_i, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset, int flag);
  450. int cher2k_kernel_LC(BLASLONG m, BLASLONG n, BLASLONG k, float alpha_r, float alpha_i, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset, int flag);
  451. int zher2k_kernel_UN(BLASLONG m, BLASLONG n, BLASLONG k, double alpha_r, double alpha_i, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset, int flag);
  452. int zher2k_kernel_UC(BLASLONG m, BLASLONG n, BLASLONG k, double alpha_r, double alpha_i, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset, int flag);
  453. int zher2k_kernel_LN(BLASLONG m, BLASLONG n, BLASLONG k, double alpha_r, double alpha_i, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset, int flag);
  454. int zher2k_kernel_LC(BLASLONG m, BLASLONG n, BLASLONG k, double alpha_r, double alpha_i, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset, int flag);
  455. int xher2k_kernel_UN(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha_r, xdouble alpha_i, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset, int flag);
  456. int xher2k_kernel_UC(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha_r, xdouble alpha_i, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset, int flag);
  457. int xher2k_kernel_LN(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha_r, xdouble alpha_i, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset, int flag);
  458. int xher2k_kernel_LC(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha_r, xdouble alpha_i, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset, int flag);
  459. int shgemm_kernel(BLASLONG, BLASLONG, BLASLONG, float, bfloat16 *, bfloat16 *, float *, BLASLONG);
  460. int sgemm_kernel(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG);
  461. int dgemm_kernel(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG);
  462. #ifdef QUAD_PRECISION
  463. int qgemm_kernel(BLASLONG, BLASLONG, BLASLONG, xidouble *, xidouble *, xidouble *, xdouble *, BLASLONG);
  464. #else
  465. int qgemm_kernel(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  466. #endif
  467. int cgemm_kernel_n(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  468. int cgemm_kernel_l(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  469. int cgemm_kernel_r(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  470. int cgemm_kernel_b(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  471. int zgemm_kernel_n(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  472. int zgemm_kernel_l(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  473. int zgemm_kernel_r(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  474. int zgemm_kernel_b(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  475. int xgemm_kernel_n(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  476. int xgemm_kernel_l(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  477. int xgemm_kernel_r(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  478. int xgemm_kernel_b(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  479. int cgemm3m_kernel(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  480. int zgemm3m_kernel(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  481. int xgemm3m_kernel(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  482. int shgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  483. int shgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  484. int shgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  485. int shgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  486. int sgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  487. int sgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  488. int sgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  489. int sgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  490. int dgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  491. int dgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  492. int dgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  493. int dgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  494. #ifdef QUAD_PRECISION
  495. int qgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  496. int qgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  497. int qgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  498. int qgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  499. #else
  500. int qgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  501. int qgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  502. int qgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  503. int qgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  504. #endif
  505. int cgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  506. int cgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  507. int cgemm_nr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  508. int cgemm_nc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  509. int cgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  510. int cgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  511. int cgemm_tr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  512. int cgemm_tc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  513. int cgemm_rn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  514. int cgemm_rt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  515. int cgemm_rr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  516. int cgemm_rc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  517. int cgemm_cn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  518. int cgemm_ct(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  519. int cgemm_cr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  520. int cgemm_cc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  521. int zgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  522. int zgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  523. int zgemm_nr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  524. int zgemm_nc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  525. int zgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  526. int zgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  527. int zgemm_tr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  528. int zgemm_tc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  529. int zgemm_rn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  530. int zgemm_rt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  531. int zgemm_rr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  532. int zgemm_rc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  533. int zgemm_cn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  534. int zgemm_ct(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  535. int zgemm_cr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  536. int zgemm_cc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  537. #ifdef QUAD_PRECISION
  538. int xgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  539. int xgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  540. int xgemm_nr(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  541. int xgemm_nc(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  542. int xgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  543. int xgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  544. int xgemm_tr(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  545. int xgemm_tc(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  546. int xgemm_rn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  547. int xgemm_rt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  548. int xgemm_rr(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  549. int xgemm_rc(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  550. int xgemm_cn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  551. int xgemm_ct(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  552. int xgemm_cr(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  553. int xgemm_cc(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  554. #else
  555. int xgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  556. int xgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  557. int xgemm_nr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  558. int xgemm_nc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  559. int xgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  560. int xgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  561. int xgemm_tr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  562. int xgemm_tc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  563. int xgemm_rn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  564. int xgemm_rt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  565. int xgemm_rr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  566. int xgemm_rc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  567. int xgemm_cn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  568. int xgemm_ct(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  569. int xgemm_cr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  570. int xgemm_cc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  571. #endif
  572. int shgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  573. int shgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  574. int shgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  575. int shgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  576. int sgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  577. int sgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  578. int sgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  579. int sgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  580. int dgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  581. int dgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  582. int dgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  583. int dgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  584. #ifdef QUAD_PRECISION
  585. int qgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  586. int qgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  587. int qgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  588. int qgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  589. #else
  590. int qgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  591. int qgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  592. int qgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  593. int qgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  594. #endif
  595. int cgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  596. int cgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  597. int cgemm_thread_nr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  598. int cgemm_thread_nc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  599. int cgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  600. int cgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  601. int cgemm_thread_tr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  602. int cgemm_thread_tc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  603. int cgemm_thread_rn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  604. int cgemm_thread_rt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  605. int cgemm_thread_rr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  606. int cgemm_thread_rc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  607. int cgemm_thread_cn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  608. int cgemm_thread_ct(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  609. int cgemm_thread_cr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  610. int cgemm_thread_cc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  611. int zgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  612. int zgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  613. int zgemm_thread_nr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  614. int zgemm_thread_nc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  615. int zgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  616. int zgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  617. int zgemm_thread_tr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  618. int zgemm_thread_tc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  619. int zgemm_thread_rn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  620. int zgemm_thread_rt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  621. int zgemm_thread_rr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  622. int zgemm_thread_rc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  623. int zgemm_thread_cn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  624. int zgemm_thread_ct(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  625. int zgemm_thread_cr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  626. int zgemm_thread_cc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  627. int xgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  628. int xgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  629. int xgemm_thread_nr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  630. int xgemm_thread_nc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  631. int xgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  632. int xgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  633. int xgemm_thread_tr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  634. int xgemm_thread_tc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  635. int xgemm_thread_rn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  636. int xgemm_thread_rt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  637. int xgemm_thread_rr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  638. int xgemm_thread_rc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  639. int xgemm_thread_cn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  640. int xgemm_thread_ct(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  641. int xgemm_thread_cr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  642. int xgemm_thread_cc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  643. int cgemm3m_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  644. int cgemm3m_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  645. int cgemm3m_nr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  646. int cgemm3m_nc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  647. int cgemm3m_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  648. int cgemm3m_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  649. int cgemm3m_tr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  650. int cgemm3m_tc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  651. int cgemm3m_rn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  652. int cgemm3m_rt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  653. int cgemm3m_rr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  654. int cgemm3m_rc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  655. int cgemm3m_cn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  656. int cgemm3m_ct(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  657. int cgemm3m_cr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  658. int cgemm3m_cc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  659. int zgemm3m_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  660. int zgemm3m_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  661. int zgemm3m_nr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  662. int zgemm3m_nc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  663. int zgemm3m_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  664. int zgemm3m_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  665. int zgemm3m_tr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  666. int zgemm3m_tc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  667. int zgemm3m_rn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  668. int zgemm3m_rt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  669. int zgemm3m_rr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  670. int zgemm3m_rc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  671. int zgemm3m_cn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  672. int zgemm3m_ct(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  673. int zgemm3m_cr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  674. int zgemm3m_cc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  675. int xgemm3m_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  676. int xgemm3m_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  677. int xgemm3m_nr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  678. int xgemm3m_nc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  679. int xgemm3m_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  680. int xgemm3m_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  681. int xgemm3m_tr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  682. int xgemm3m_tc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  683. int xgemm3m_rn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  684. int xgemm3m_rt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  685. int xgemm3m_rr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  686. int xgemm3m_rc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  687. int xgemm3m_cn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  688. int xgemm3m_ct(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  689. int xgemm3m_cr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  690. int xgemm3m_cc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  691. int cgemm3m_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  692. int cgemm3m_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  693. int cgemm3m_thread_nr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  694. int cgemm3m_thread_nc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  695. int cgemm3m_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  696. int cgemm3m_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  697. int cgemm3m_thread_tr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  698. int cgemm3m_thread_tc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  699. int cgemm3m_thread_rn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  700. int cgemm3m_thread_rt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  701. int cgemm3m_thread_rr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  702. int cgemm3m_thread_rc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  703. int cgemm3m_thread_cn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  704. int cgemm3m_thread_ct(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  705. int cgemm3m_thread_cr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  706. int cgemm3m_thread_cc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  707. int zgemm3m_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  708. int zgemm3m_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  709. int zgemm3m_thread_nr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  710. int zgemm3m_thread_nc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  711. int zgemm3m_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  712. int zgemm3m_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  713. int zgemm3m_thread_tr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  714. int zgemm3m_thread_tc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  715. int zgemm3m_thread_rn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  716. int zgemm3m_thread_rt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  717. int zgemm3m_thread_rr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  718. int zgemm3m_thread_rc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  719. int zgemm3m_thread_cn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  720. int zgemm3m_thread_ct(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  721. int zgemm3m_thread_cr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  722. int zgemm3m_thread_cc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  723. int xgemm3m_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  724. int xgemm3m_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  725. int xgemm3m_thread_nr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  726. int xgemm3m_thread_nc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  727. int xgemm3m_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  728. int xgemm3m_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  729. int xgemm3m_thread_tr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  730. int xgemm3m_thread_tc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  731. int xgemm3m_thread_rn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  732. int xgemm3m_thread_rt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  733. int xgemm3m_thread_rr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  734. int xgemm3m_thread_rc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  735. int xgemm3m_thread_cn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  736. int xgemm3m_thread_ct(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  737. int xgemm3m_thread_cr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  738. int xgemm3m_thread_cc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  739. int cher2m_LNN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  740. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  741. int cher2m_LNT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  742. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  743. int cher2m_LNR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  744. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  745. int cher2m_LNC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  746. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  747. int cher2m_LTN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  748. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  749. int cher2m_LTT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  750. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  751. int cher2m_LTR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  752. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  753. int cher2m_LTC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  754. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  755. int cher2m_LRN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  756. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  757. int cher2m_LRT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  758. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  759. int cher2m_LRR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  760. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  761. int cher2m_LRC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  762. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  763. int cher2m_LCN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  764. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  765. int cher2m_LCT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  766. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  767. int cher2m_LCR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  768. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  769. int cher2m_LCC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  770. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  771. int cher2m_UNN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  772. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  773. int cher2m_UNT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  774. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  775. int cher2m_UNR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  776. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  777. int cher2m_UNC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  778. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  779. int cher2m_UTN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  780. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  781. int cher2m_UTT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  782. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  783. int cher2m_UTR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  784. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  785. int cher2m_UTC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  786. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  787. int cher2m_URN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  788. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  789. int cher2m_URT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  790. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  791. int cher2m_URR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  792. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  793. int cher2m_URC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  794. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  795. int cher2m_UCN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  796. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  797. int cher2m_UCT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  798. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  799. int cher2m_UCR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  800. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  801. int cher2m_UCC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  802. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  803. int zher2m_LNN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  804. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  805. int zher2m_LNT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  806. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  807. int zher2m_LNR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  808. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  809. int zher2m_LNC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  810. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  811. int zher2m_LTN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  812. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  813. int zher2m_LTT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  814. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  815. int zher2m_LTR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  816. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  817. int zher2m_LTC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  818. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  819. int zher2m_LRN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  820. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  821. int zher2m_LRT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  822. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  823. int zher2m_LRR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  824. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  825. int zher2m_LRC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  826. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  827. int zher2m_LCN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  828. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  829. int zher2m_LCT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  830. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  831. int zher2m_LCR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  832. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  833. int zher2m_LCC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  834. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  835. int zher2m_UNN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  836. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  837. int zher2m_UNT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  838. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  839. int zher2m_UNR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  840. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  841. int zher2m_UNC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  842. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  843. int zher2m_UTN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  844. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  845. int zher2m_UTT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  846. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  847. int zher2m_UTR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  848. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  849. int zher2m_UTC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  850. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  851. int zher2m_URN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  852. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  853. int zher2m_URT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  854. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  855. int zher2m_URR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  856. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  857. int zher2m_URC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  858. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  859. int zher2m_UCN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  860. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  861. int zher2m_UCT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  862. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  863. int zher2m_UCR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  864. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  865. int zher2m_UCC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  866. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  867. int strsm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  868. int strsm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  869. int strsm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  870. int strsm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  871. int strsm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  872. int strsm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  873. int strsm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  874. int strsm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  875. int strsm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  876. int strsm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  877. int strsm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  878. int strsm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  879. int strsm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  880. int strsm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  881. int strsm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  882. int strsm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  883. int dtrsm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  884. int dtrsm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  885. int dtrsm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  886. int dtrsm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  887. int dtrsm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  888. int dtrsm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  889. int dtrsm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  890. int dtrsm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  891. int dtrsm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  892. int dtrsm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  893. int dtrsm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  894. int dtrsm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  895. int dtrsm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  896. int dtrsm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  897. int dtrsm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  898. int dtrsm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  899. int qtrsm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  900. int qtrsm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  901. int qtrsm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  902. int qtrsm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  903. int qtrsm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  904. int qtrsm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  905. int qtrsm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  906. int qtrsm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  907. int qtrsm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  908. int qtrsm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  909. int qtrsm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  910. int qtrsm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  911. int qtrsm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  912. int qtrsm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  913. int qtrsm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  914. int qtrsm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  915. int ctrsm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  916. int ctrsm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  917. int ctrsm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  918. int ctrsm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  919. int ctrsm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  920. int ctrsm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  921. int ctrsm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  922. int ctrsm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  923. int ctrsm_LRUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  924. int ctrsm_LRUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  925. int ctrsm_LRLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  926. int ctrsm_LRLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  927. int ctrsm_LCUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  928. int ctrsm_LCUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  929. int ctrsm_LCLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  930. int ctrsm_LCLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  931. int ctrsm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  932. int ctrsm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  933. int ctrsm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  934. int ctrsm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  935. int ctrsm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  936. int ctrsm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  937. int ctrsm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  938. int ctrsm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  939. int ctrsm_RRUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  940. int ctrsm_RRUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  941. int ctrsm_RRLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  942. int ctrsm_RRLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  943. int ctrsm_RCUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  944. int ctrsm_RCUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  945. int ctrsm_RCLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  946. int ctrsm_RCLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  947. int ztrsm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  948. int ztrsm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  949. int ztrsm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  950. int ztrsm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  951. int ztrsm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  952. int ztrsm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  953. int ztrsm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  954. int ztrsm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  955. int ztrsm_LRUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  956. int ztrsm_LRUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  957. int ztrsm_LRLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  958. int ztrsm_LRLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  959. int ztrsm_LCUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  960. int ztrsm_LCUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  961. int ztrsm_LCLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  962. int ztrsm_LCLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  963. int ztrsm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  964. int ztrsm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  965. int ztrsm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  966. int ztrsm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  967. int ztrsm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  968. int ztrsm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  969. int ztrsm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  970. int ztrsm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  971. int ztrsm_RRUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  972. int ztrsm_RRUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  973. int ztrsm_RRLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  974. int ztrsm_RRLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  975. int ztrsm_RCUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  976. int ztrsm_RCUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  977. int ztrsm_RCLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  978. int ztrsm_RCLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  979. int xtrsm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  980. int xtrsm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  981. int xtrsm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  982. int xtrsm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  983. int xtrsm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  984. int xtrsm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  985. int xtrsm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  986. int xtrsm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  987. int xtrsm_LRUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  988. int xtrsm_LRUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  989. int xtrsm_LRLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  990. int xtrsm_LRLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  991. int xtrsm_LCUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  992. int xtrsm_LCUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  993. int xtrsm_LCLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  994. int xtrsm_LCLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  995. int xtrsm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  996. int xtrsm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  997. int xtrsm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  998. int xtrsm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  999. int xtrsm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1000. int xtrsm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1001. int xtrsm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1002. int xtrsm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1003. int xtrsm_RRUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1004. int xtrsm_RRUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1005. int xtrsm_RRLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1006. int xtrsm_RRLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1007. int xtrsm_RCUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1008. int xtrsm_RCUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1009. int xtrsm_RCLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1010. int xtrsm_RCLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1011. int strmm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1012. int strmm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1013. int strmm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1014. int strmm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1015. int strmm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1016. int strmm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1017. int strmm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1018. int strmm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1019. int strmm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1020. int strmm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1021. int strmm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1022. int strmm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1023. int strmm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1024. int strmm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1025. int strmm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1026. int strmm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1027. int dtrmm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1028. int dtrmm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1029. int dtrmm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1030. int dtrmm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1031. int dtrmm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1032. int dtrmm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1033. int dtrmm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1034. int dtrmm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1035. int dtrmm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1036. int dtrmm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1037. int dtrmm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1038. int dtrmm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1039. int dtrmm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1040. int dtrmm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1041. int dtrmm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1042. int dtrmm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1043. int qtrmm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1044. int qtrmm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1045. int qtrmm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1046. int qtrmm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1047. int qtrmm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1048. int qtrmm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1049. int qtrmm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1050. int qtrmm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1051. int qtrmm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1052. int qtrmm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1053. int qtrmm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1054. int qtrmm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1055. int qtrmm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1056. int qtrmm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1057. int qtrmm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1058. int qtrmm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1059. int ctrmm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1060. int ctrmm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1061. int ctrmm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1062. int ctrmm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1063. int ctrmm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1064. int ctrmm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1065. int ctrmm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1066. int ctrmm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1067. int ctrmm_LRUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1068. int ctrmm_LRUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1069. int ctrmm_LRLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1070. int ctrmm_LRLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1071. int ctrmm_LCUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1072. int ctrmm_LCUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1073. int ctrmm_LCLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1074. int ctrmm_LCLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1075. int ctrmm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1076. int ctrmm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1077. int ctrmm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1078. int ctrmm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1079. int ctrmm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1080. int ctrmm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1081. int ctrmm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1082. int ctrmm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1083. int ctrmm_RRUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1084. int ctrmm_RRUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1085. int ctrmm_RRLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1086. int ctrmm_RRLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1087. int ctrmm_RCUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1088. int ctrmm_RCUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1089. int ctrmm_RCLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1090. int ctrmm_RCLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1091. int ztrmm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1092. int ztrmm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1093. int ztrmm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1094. int ztrmm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1095. int ztrmm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1096. int ztrmm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1097. int ztrmm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1098. int ztrmm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1099. int ztrmm_LRUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1100. int ztrmm_LRUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1101. int ztrmm_LRLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1102. int ztrmm_LRLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1103. int ztrmm_LCUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1104. int ztrmm_LCUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1105. int ztrmm_LCLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1106. int ztrmm_LCLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1107. int ztrmm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1108. int ztrmm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1109. int ztrmm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1110. int ztrmm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1111. int ztrmm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1112. int ztrmm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1113. int ztrmm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1114. int ztrmm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1115. int ztrmm_RRUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1116. int ztrmm_RRUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1117. int ztrmm_RRLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1118. int ztrmm_RRLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1119. int ztrmm_RCUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1120. int ztrmm_RCUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1121. int ztrmm_RCLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1122. int ztrmm_RCLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1123. int xtrmm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1124. int xtrmm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1125. int xtrmm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1126. int xtrmm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1127. int xtrmm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1128. int xtrmm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1129. int xtrmm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1130. int xtrmm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1131. int xtrmm_LRUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1132. int xtrmm_LRUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1133. int xtrmm_LRLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1134. int xtrmm_LRLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1135. int xtrmm_LCUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1136. int xtrmm_LCUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1137. int xtrmm_LCLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1138. int xtrmm_LCLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1139. int xtrmm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1140. int xtrmm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1141. int xtrmm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1142. int xtrmm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1143. int xtrmm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1144. int xtrmm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1145. int xtrmm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1146. int xtrmm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1147. int xtrmm_RRUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1148. int xtrmm_RRUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1149. int xtrmm_RRLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1150. int xtrmm_RRLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1151. int xtrmm_RCUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1152. int xtrmm_RCUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1153. int xtrmm_RCLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1154. int xtrmm_RCLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1155. int ssymm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1156. int ssymm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1157. int ssymm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1158. int ssymm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1159. int dsymm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1160. int dsymm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1161. int dsymm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1162. int dsymm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1163. int qsymm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1164. int qsymm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1165. int qsymm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1166. int qsymm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1167. int csymm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1168. int csymm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1169. int csymm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1170. int csymm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1171. int zsymm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1172. int zsymm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1173. int zsymm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1174. int zsymm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1175. int xsymm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1176. int xsymm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1177. int xsymm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1178. int xsymm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1179. int csymm3m_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1180. int csymm3m_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1181. int csymm3m_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1182. int csymm3m_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1183. int zsymm3m_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1184. int zsymm3m_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1185. int zsymm3m_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1186. int zsymm3m_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1187. int xsymm3m_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1188. int xsymm3m_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1189. int xsymm3m_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1190. int xsymm3m_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1191. int csymm3m_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1192. int csymm3m_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1193. int csymm3m_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1194. int csymm3m_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1195. int zsymm3m_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1196. int zsymm3m_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1197. int zsymm3m_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1198. int zsymm3m_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1199. int xsymm3m_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1200. int xsymm3m_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1201. int xsymm3m_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1202. int xsymm3m_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1203. int chemm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1204. int chemm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1205. int chemm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1206. int chemm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1207. int zhemm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1208. int zhemm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1209. int zhemm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1210. int zhemm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1211. int xhemm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1212. int xhemm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1213. int xhemm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1214. int xhemm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1215. int chemm3m_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1216. int chemm3m_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1217. int chemm3m_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1218. int chemm3m_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1219. int zhemm3m_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1220. int zhemm3m_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1221. int zhemm3m_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1222. int zhemm3m_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1223. int xhemm3m_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1224. int xhemm3m_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1225. int xhemm3m_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1226. int xhemm3m_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1227. int chemm3m_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1228. int chemm3m_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1229. int chemm3m_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1230. int chemm3m_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1231. int zhemm3m_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1232. int zhemm3m_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1233. int zhemm3m_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1234. int zhemm3m_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1235. int xhemm3m_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1236. int xhemm3m_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1237. int xhemm3m_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1238. int xhemm3m_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1239. int ssymm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1240. int ssymm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1241. int ssymm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1242. int ssymm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1243. int dsymm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1244. int dsymm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1245. int dsymm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1246. int dsymm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1247. int qsymm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1248. int qsymm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1249. int qsymm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1250. int qsymm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1251. int csymm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1252. int csymm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1253. int csymm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1254. int csymm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1255. int zsymm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1256. int zsymm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1257. int zsymm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1258. int zsymm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1259. int xsymm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1260. int xsymm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1261. int xsymm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1262. int xsymm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1263. int chemm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1264. int chemm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1265. int chemm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1266. int chemm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1267. int zhemm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1268. int zhemm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1269. int zhemm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1270. int zhemm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1271. int xhemm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1272. int xhemm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1273. int xhemm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1274. int xhemm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1275. int ssyrk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1276. int ssyrk_UT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1277. int ssyrk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1278. int ssyrk_LT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1279. int dsyrk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1280. int dsyrk_UT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1281. int dsyrk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1282. int dsyrk_LT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1283. int qsyrk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1284. int qsyrk_UT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1285. int qsyrk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1286. int qsyrk_LT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1287. int csyrk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1288. int csyrk_UT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1289. int csyrk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1290. int csyrk_LT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1291. int zsyrk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1292. int zsyrk_UT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1293. int zsyrk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1294. int zsyrk_LT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1295. int xsyrk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1296. int xsyrk_UT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1297. int xsyrk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1298. int xsyrk_LT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1299. int ssyrk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1300. int ssyrk_thread_UT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1301. int ssyrk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1302. int ssyrk_thread_LT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1303. int dsyrk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1304. int dsyrk_thread_UT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1305. int dsyrk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1306. int dsyrk_thread_LT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1307. int qsyrk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1308. int qsyrk_thread_UT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1309. int qsyrk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1310. int qsyrk_thread_LT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1311. int csyrk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1312. int csyrk_thread_UT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1313. int csyrk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1314. int csyrk_thread_LT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1315. int zsyrk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1316. int zsyrk_thread_UT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1317. int zsyrk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1318. int zsyrk_thread_LT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1319. int xsyrk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1320. int xsyrk_thread_UT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1321. int xsyrk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1322. int xsyrk_thread_LT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1323. int ssyr2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1324. int ssyr2k_UT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1325. int ssyr2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1326. int ssyr2k_LT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1327. int dsyr2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1328. int dsyr2k_UT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1329. int dsyr2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1330. int dsyr2k_LT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1331. int qsyr2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1332. int qsyr2k_UT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1333. int qsyr2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1334. int qsyr2k_LT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1335. int csyr2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1336. int csyr2k_UT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1337. int csyr2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1338. int csyr2k_LT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1339. int zsyr2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1340. int zsyr2k_UT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1341. int zsyr2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1342. int zsyr2k_LT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1343. int xsyr2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1344. int xsyr2k_UT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1345. int xsyr2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1346. int xsyr2k_LT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1347. int cherk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1348. int cherk_UC(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1349. int cherk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1350. int cherk_LC(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1351. int zherk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1352. int zherk_UC(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1353. int zherk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1354. int zherk_LC(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1355. int xherk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1356. int xherk_UC(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1357. int xherk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1358. int xherk_LC(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1359. int cherk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1360. int cherk_thread_UC(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1361. int cherk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1362. int cherk_thread_LC(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1363. int zherk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1364. int zherk_thread_UC(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1365. int zherk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1366. int zherk_thread_LC(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1367. int xherk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1368. int xherk_thread_UC(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1369. int xherk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1370. int xherk_thread_LC(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1371. int cher2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1372. int cher2k_UC(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1373. int cher2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1374. int cher2k_LC(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1375. int zher2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1376. int zher2k_UC(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1377. int zher2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1378. int zher2k_LC(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1379. int xher2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1380. int xher2k_UC(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1381. int xher2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1382. int xher2k_LC(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1383. int sgemt_n(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, int);
  1384. int sgemt_t(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, int);
  1385. int dgemt_n(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, int);
  1386. int dgemt_t(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, int);
  1387. int cgemt_n(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, int);
  1388. int cgemt_t(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, int);
  1389. int cgemt_r(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, int);
  1390. int cgemt_c(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, int);
  1391. int zgemt_n(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, int);
  1392. int zgemt_t(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, int);
  1393. int zgemt_r(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, int);
  1394. int zgemt_c(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, int);
  1395. int sgema_n(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG);
  1396. int sgema_t(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG);
  1397. int dgema_n(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG);
  1398. int dgema_t(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG);
  1399. int cgema_n(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1400. int cgema_t(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1401. int cgema_r(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1402. int cgema_c(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1403. int zgema_n(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1404. int zgema_t(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1405. int zgema_r(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1406. int zgema_c(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1407. int cgemm3m_incopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  1408. int cgemm3m_incopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  1409. int cgemm3m_incopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  1410. int cgemm3m_itcopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  1411. int cgemm3m_itcopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  1412. int cgemm3m_itcopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  1413. int cgemm3m_oncopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float alpha_r, float alpha_i, float *b);
  1414. int cgemm3m_oncopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float alpha_r, float alpha_i, float *b);
  1415. int cgemm3m_oncopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float alpha_r, float alpha_i, float *b);
  1416. int cgemm3m_otcopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float alpha_r, float alpha_i, float *b);
  1417. int cgemm3m_otcopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float alpha_r, float alpha_i, float *b);
  1418. int cgemm3m_otcopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float alpha_r, float alpha_i, float *b);
  1419. int zgemm3m_incopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  1420. int zgemm3m_incopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  1421. int zgemm3m_incopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  1422. int zgemm3m_itcopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  1423. int zgemm3m_itcopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  1424. int zgemm3m_itcopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  1425. int zgemm3m_oncopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double alpha_r, double alpha_i, double *b);
  1426. int zgemm3m_oncopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double alpha_r, double alpha_i, double *b);
  1427. int zgemm3m_oncopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double alpha_r, double alpha_i, double *b);
  1428. int zgemm3m_otcopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double alpha_r, double alpha_i, double *b);
  1429. int zgemm3m_otcopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double alpha_r, double alpha_i, double *b);
  1430. int zgemm3m_otcopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double alpha_r, double alpha_i, double *b);
  1431. int xgemm3m_incopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  1432. int xgemm3m_incopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  1433. int xgemm3m_incopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  1434. int xgemm3m_itcopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  1435. int xgemm3m_itcopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  1436. int xgemm3m_itcopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  1437. int xgemm3m_oncopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1438. int xgemm3m_oncopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1439. int xgemm3m_oncopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1440. int xgemm3m_otcopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1441. int xgemm3m_otcopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1442. int xgemm3m_otcopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1443. int csymm3m_iucopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1444. int csymm3m_ilcopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1445. int csymm3m_iucopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1446. int csymm3m_ilcopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1447. int csymm3m_iucopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1448. int csymm3m_ilcopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1449. int csymm3m_oucopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1450. int csymm3m_olcopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1451. int csymm3m_oucopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1452. int csymm3m_olcopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1453. int csymm3m_oucopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1454. int csymm3m_olcopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1455. int zsymm3m_iucopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1456. int zsymm3m_ilcopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1457. int zsymm3m_iucopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1458. int zsymm3m_ilcopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1459. int zsymm3m_iucopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1460. int zsymm3m_ilcopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1461. int zsymm3m_oucopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1462. int zsymm3m_olcopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1463. int zsymm3m_oucopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1464. int zsymm3m_olcopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1465. int zsymm3m_oucopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1466. int zsymm3m_olcopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1467. int xsymm3m_iucopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1468. int xsymm3m_ilcopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1469. int xsymm3m_iucopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1470. int xsymm3m_ilcopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1471. int xsymm3m_iucopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1472. int xsymm3m_ilcopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1473. int xsymm3m_oucopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1474. int xsymm3m_olcopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1475. int xsymm3m_oucopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1476. int xsymm3m_olcopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1477. int xsymm3m_oucopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1478. int xsymm3m_olcopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1479. int chemm3m_iucopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1480. int chemm3m_ilcopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1481. int chemm3m_iucopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1482. int chemm3m_ilcopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1483. int chemm3m_iucopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1484. int chemm3m_ilcopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1485. int chemm3m_oucopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1486. int chemm3m_olcopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1487. int chemm3m_oucopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1488. int chemm3m_olcopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1489. int chemm3m_oucopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1490. int chemm3m_olcopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1491. int zhemm3m_iucopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1492. int zhemm3m_ilcopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1493. int zhemm3m_iucopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1494. int zhemm3m_ilcopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1495. int zhemm3m_iucopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1496. int zhemm3m_ilcopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1497. int zhemm3m_oucopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1498. int zhemm3m_olcopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1499. int zhemm3m_oucopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1500. int zhemm3m_olcopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1501. int zhemm3m_oucopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1502. int zhemm3m_olcopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1503. int xhemm3m_iucopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1504. int xhemm3m_ilcopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1505. int xhemm3m_iucopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1506. int xhemm3m_ilcopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1507. int xhemm3m_iucopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1508. int xhemm3m_ilcopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1509. int xhemm3m_oucopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1510. int xhemm3m_olcopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1511. int xhemm3m_oucopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1512. int xhemm3m_olcopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1513. int xhemm3m_oucopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1514. int xhemm3m_olcopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1515. int sgemc_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1516. int sgemc_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1517. int sgemc_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1518. int sgemc_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1519. int dgemc_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1520. int dgemc_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1521. int dgemc_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1522. int dgemc_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1523. int qgemc_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1524. int qgemc_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1525. int qgemc_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1526. int qgemc_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1527. int cgemc_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1528. int cgemc_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1529. int cgemc_nr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1530. int cgemc_nc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1531. int cgemc_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1532. int cgemc_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1533. int cgemc_tr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1534. int cgemc_tc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1535. int cgemc_rn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1536. int cgemc_rt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1537. int cgemc_rr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1538. int cgemc_rc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1539. int cgemc_cn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1540. int cgemc_ct(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1541. int cgemc_cr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1542. int cgemc_cc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1543. int zgemc_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1544. int zgemc_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1545. int zgemc_nr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1546. int zgemc_nc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1547. int zgemc_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1548. int zgemc_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1549. int zgemc_tr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1550. int zgemc_tc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1551. int zgemc_rn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1552. int zgemc_rt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1553. int zgemc_rr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1554. int zgemc_rc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1555. int zgemc_cn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1556. int zgemc_ct(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1557. int zgemc_cr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1558. int zgemc_cc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1559. int xgemc_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1560. int xgemc_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1561. int xgemc_nr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1562. int xgemc_nc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1563. int xgemc_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1564. int xgemc_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1565. int xgemc_tr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1566. int xgemc_tc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1567. int xgemc_rn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1568. int xgemc_rt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1569. int xgemc_rr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1570. int xgemc_rc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1571. int xgemc_cn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1572. int xgemc_ct(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1573. int xgemc_cr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1574. int xgemc_cc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1575. int sgemc_oncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b, BLASLONG ldb, float *c);
  1576. int sgemc_otcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b, BLASLONG ldb, float *c);
  1577. int dgemc_oncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b, BLASLONG ldb, double *c);
  1578. int dgemc_otcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b, BLASLONG ldb, double *c);
  1579. int qgemc_oncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b, BLASLONG ldb, xdouble *c);
  1580. int qgemc_otcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b, BLASLONG ldb, xdouble *c);
  1581. int cgemc_oncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b, BLASLONG ldb, float *c);
  1582. int cgemc_otcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b, BLASLONG ldb, float *c);
  1583. int zgemc_oncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b, BLASLONG ldb, double *c);
  1584. int zgemc_otcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b, BLASLONG ldb, double *c);
  1585. int xgemc_oncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b, BLASLONG ldb, xdouble *c);
  1586. int xgemc_otcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b, BLASLONG ldb, xdouble *c);
  1587. int somatcopy_k_cn(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG);
  1588. int somatcopy_k_rn(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG);
  1589. int somatcopy_k_ct(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG);
  1590. int somatcopy_k_rt(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG);
  1591. int simatcopy_k_cn(BLASLONG, BLASLONG, float, float *, BLASLONG);
  1592. int simatcopy_k_rn(BLASLONG, BLASLONG, float, float *, BLASLONG);
  1593. int simatcopy_k_ct(BLASLONG, BLASLONG, float, float *, BLASLONG);
  1594. int simatcopy_k_rt(BLASLONG, BLASLONG, float, float *, BLASLONG);
  1595. int domatcopy_k_cn(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG);
  1596. int domatcopy_k_rn(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG);
  1597. int domatcopy_k_ct(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG);
  1598. int domatcopy_k_rt(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG);
  1599. int dimatcopy_k_cn(BLASLONG, BLASLONG, double, double *, BLASLONG);
  1600. int dimatcopy_k_rn(BLASLONG, BLASLONG, double, double *, BLASLONG);
  1601. int dimatcopy_k_ct(BLASLONG, BLASLONG, double, double *, BLASLONG);
  1602. int dimatcopy_k_rt(BLASLONG, BLASLONG, double, double *, BLASLONG);
  1603. int comatcopy_k_cn(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1604. int comatcopy_k_rn(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1605. int comatcopy_k_ct(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1606. int comatcopy_k_rt(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1607. int cimatcopy_k_cn(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1608. int cimatcopy_k_rn(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1609. int cimatcopy_k_ct(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1610. int cimatcopy_k_rt(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1611. int comatcopy_k_cnc(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1612. int comatcopy_k_rnc(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1613. int comatcopy_k_ctc(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1614. int comatcopy_k_rtc(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1615. int cimatcopy_k_cnc(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1616. int cimatcopy_k_rnc(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1617. int cimatcopy_k_ctc(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1618. int cimatcopy_k_rtc(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1619. int zomatcopy_k_cn(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1620. int zomatcopy_k_rn(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1621. int zomatcopy_k_ct(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1622. int zomatcopy_k_rt(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1623. int zimatcopy_k_cn(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1624. int zimatcopy_k_rn(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1625. int zimatcopy_k_ct(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1626. int zimatcopy_k_rt(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1627. int zomatcopy_k_cnc(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1628. int zomatcopy_k_rnc(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1629. int zomatcopy_k_ctc(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1630. int zomatcopy_k_rtc(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1631. int zimatcopy_k_cnc(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1632. int zimatcopy_k_rnc(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1633. int zimatcopy_k_ctc(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1634. int zimatcopy_k_rtc(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1635. int sgeadd_k(BLASLONG, BLASLONG, float, float*, BLASLONG, float, float *, BLASLONG);
  1636. int dgeadd_k(BLASLONG, BLASLONG, double, double*, BLASLONG, double, double *, BLASLONG);
  1637. int cgeadd_k(BLASLONG, BLASLONG, float, float, float*, BLASLONG, float, float, float *, BLASLONG);
  1638. int zgeadd_k(BLASLONG, BLASLONG, double,double, double*, BLASLONG, double, double, double *, BLASLONG);
  1639. #ifdef __CUDACC__
  1640. }
  1641. #endif
  1642. #endif