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 163 kB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927
  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 sbgemm_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 sbgemm_incopy(BLASLONG m, BLASLONG n, bfloat16 *a, BLASLONG lda, bfloat16 *b);
  73. int sbgemm_itcopy(BLASLONG m, BLASLONG n, bfloat16 *a, BLASLONG lda, bfloat16 *b);
  74. int sbgemm_oncopy(BLASLONG m, BLASLONG n, bfloat16 *a, BLASLONG lda, bfloat16 *b);
  75. int sbgemm_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 sbgemm_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. #ifdef SMALL_MATRIX_OPT
  468. int sgemm_small_kernel_nn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
  469. int sgemm_small_kernel_nt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
  470. int sgemm_small_kernel_tn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
  471. int sgemm_small_kernel_tt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
  472. int dgemm_small_kernel_nn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double beta, double * C, BLASLONG ldc);
  473. int dgemm_small_kernel_nt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double beta, double * C, BLASLONG ldc);
  474. int dgemm_small_kernel_tn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double beta, double * C, BLASLONG ldc);
  475. int dgemm_small_kernel_tt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double beta, double * C, BLASLONG ldc);
  476. int sgemm_small_kernel_b0_nn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  477. int sgemm_small_kernel_b0_nt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  478. int sgemm_small_kernel_b0_tn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  479. int sgemm_small_kernel_b0_tt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  480. int dgemm_small_kernel_b0_nn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  481. int dgemm_small_kernel_b0_nt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  482. int dgemm_small_kernel_b0_tn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  483. int dgemm_small_kernel_b0_tt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  484. int cgemm_small_kernel_nn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  485. int cgemm_small_kernel_nt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  486. int cgemm_small_kernel_nr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  487. int cgemm_small_kernel_nc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  488. int cgemm_small_kernel_tn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  489. int cgemm_small_kernel_tt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  490. int cgemm_small_kernel_tr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  491. int cgemm_small_kernel_tc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  492. int cgemm_small_kernel_rn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  493. int cgemm_small_kernel_rt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  494. int cgemm_small_kernel_rr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  495. int cgemm_small_kernel_rc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  496. int cgemm_small_kernel_cn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  497. int cgemm_small_kernel_ct(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  498. int cgemm_small_kernel_cr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  499. int cgemm_small_kernel_cc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
  500. int zgemm_small_kernel_nn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  501. int zgemm_small_kernel_nt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  502. int zgemm_small_kernel_nr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  503. int zgemm_small_kernel_nc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  504. int zgemm_small_kernel_tn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  505. int zgemm_small_kernel_tt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  506. int zgemm_small_kernel_tr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  507. int zgemm_small_kernel_tc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  508. int zgemm_small_kernel_rn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  509. int zgemm_small_kernel_rt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  510. int zgemm_small_kernel_rr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  511. int zgemm_small_kernel_rc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  512. int zgemm_small_kernel_cn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  513. int zgemm_small_kernel_ct(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  514. int zgemm_small_kernel_cr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  515. int zgemm_small_kernel_cc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
  516. int cgemm_small_kernel_b0_nn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  517. int cgemm_small_kernel_b0_nt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  518. int cgemm_small_kernel_b0_nr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  519. int cgemm_small_kernel_b0_nc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  520. int cgemm_small_kernel_b0_tn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  521. int cgemm_small_kernel_b0_tt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  522. int cgemm_small_kernel_b0_tr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  523. int cgemm_small_kernel_b0_tc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  524. int cgemm_small_kernel_b0_rn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  525. int cgemm_small_kernel_b0_rt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  526. int cgemm_small_kernel_b0_rr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  527. int cgemm_small_kernel_b0_rc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  528. int cgemm_small_kernel_b0_cn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  529. int cgemm_small_kernel_b0_ct(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  530. int cgemm_small_kernel_b0_cr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  531. int cgemm_small_kernel_b0_cc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
  532. int zgemm_small_kernel_b0_nn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  533. int zgemm_small_kernel_b0_nt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  534. int zgemm_small_kernel_b0_nr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  535. int zgemm_small_kernel_b0_nc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  536. int zgemm_small_kernel_b0_tn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  537. int zgemm_small_kernel_b0_tt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  538. int zgemm_small_kernel_b0_tr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  539. int zgemm_small_kernel_b0_tc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  540. int zgemm_small_kernel_b0_rn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  541. int zgemm_small_kernel_b0_rt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  542. int zgemm_small_kernel_b0_rr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  543. int zgemm_small_kernel_b0_rc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  544. int zgemm_small_kernel_b0_cn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  545. int zgemm_small_kernel_b0_ct(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  546. int zgemm_small_kernel_b0_cr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  547. int zgemm_small_kernel_b0_cc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
  548. #endif
  549. int cgemm_kernel_n(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  550. int cgemm_kernel_l(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  551. int cgemm_kernel_r(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  552. int cgemm_kernel_b(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  553. int zgemm_kernel_n(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  554. int zgemm_kernel_l(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  555. int zgemm_kernel_r(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  556. int zgemm_kernel_b(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  557. int xgemm_kernel_n(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  558. int xgemm_kernel_l(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  559. int xgemm_kernel_r(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  560. int xgemm_kernel_b(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  561. int cgemm3m_kernel(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  562. int zgemm3m_kernel(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  563. int xgemm3m_kernel(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  564. int sbgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  565. int sbgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  566. int sbgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  567. int sbgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  568. int sgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  569. int sgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  570. int sgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  571. int sgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  572. int dgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  573. int dgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  574. int dgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  575. int dgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  576. #ifdef QUAD_PRECISION
  577. int qgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  578. int qgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  579. int qgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  580. int qgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  581. #else
  582. int qgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  583. int qgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  584. int qgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  585. int qgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  586. #endif
  587. int cgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  588. int cgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  589. int cgemm_nr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  590. int cgemm_nc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  591. int cgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  592. int cgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  593. int cgemm_tr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  594. int cgemm_tc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  595. int cgemm_rn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  596. int cgemm_rt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  597. int cgemm_rr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  598. int cgemm_rc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  599. int cgemm_cn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  600. int cgemm_ct(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  601. int cgemm_cr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  602. int cgemm_cc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  603. int zgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  604. int zgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  605. int zgemm_nr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  606. int zgemm_nc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  607. int zgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  608. int zgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  609. int zgemm_tr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  610. int zgemm_tc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  611. int zgemm_rn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  612. int zgemm_rt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  613. int zgemm_rr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  614. int zgemm_rc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  615. int zgemm_cn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  616. int zgemm_ct(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  617. int zgemm_cr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  618. int zgemm_cc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  619. #ifdef QUAD_PRECISION
  620. int xgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  621. int xgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  622. int xgemm_nr(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  623. int xgemm_nc(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  624. int xgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  625. int xgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  626. int xgemm_tr(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  627. int xgemm_tc(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  628. int xgemm_rn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  629. int xgemm_rt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  630. int xgemm_rr(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  631. int xgemm_rc(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  632. int xgemm_cn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  633. int xgemm_ct(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  634. int xgemm_cr(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  635. int xgemm_cc(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  636. #else
  637. int xgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  638. int xgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  639. int xgemm_nr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  640. int xgemm_nc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  641. int xgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  642. int xgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  643. int xgemm_tr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  644. int xgemm_tc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  645. int xgemm_rn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  646. int xgemm_rt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  647. int xgemm_rr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  648. int xgemm_rc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  649. int xgemm_cn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  650. int xgemm_ct(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  651. int xgemm_cr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  652. int xgemm_cc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  653. #endif
  654. int sbgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  655. int sbgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  656. int sbgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  657. int sbgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  658. int sgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  659. int sgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  660. int sgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  661. int sgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  662. int dgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  663. int dgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  664. int dgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  665. int dgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  666. #ifdef QUAD_PRECISION
  667. int qgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  668. int qgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  669. int qgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  670. int qgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  671. #else
  672. int qgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  673. int qgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  674. int qgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  675. int qgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  676. #endif
  677. int cgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  678. int cgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  679. int cgemm_thread_nr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  680. int cgemm_thread_nc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  681. int cgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  682. int cgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  683. int cgemm_thread_tr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  684. int cgemm_thread_tc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  685. int cgemm_thread_rn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  686. int cgemm_thread_rt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  687. int cgemm_thread_rr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  688. int cgemm_thread_rc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  689. int cgemm_thread_cn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  690. int cgemm_thread_ct(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  691. int cgemm_thread_cr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  692. int cgemm_thread_cc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  693. int zgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  694. int zgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  695. int zgemm_thread_nr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  696. int zgemm_thread_nc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  697. int zgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  698. int zgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  699. int zgemm_thread_tr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  700. int zgemm_thread_tc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  701. int zgemm_thread_rn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  702. int zgemm_thread_rt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  703. int zgemm_thread_rr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  704. int zgemm_thread_rc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  705. int zgemm_thread_cn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  706. int zgemm_thread_ct(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  707. int zgemm_thread_cr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  708. int zgemm_thread_cc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  709. int xgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  710. int xgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  711. int xgemm_thread_nr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  712. int xgemm_thread_nc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  713. int xgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  714. int xgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  715. int xgemm_thread_tr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  716. int xgemm_thread_tc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  717. int xgemm_thread_rn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  718. int xgemm_thread_rt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  719. int xgemm_thread_rr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  720. int xgemm_thread_rc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  721. int xgemm_thread_cn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  722. int xgemm_thread_ct(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  723. int xgemm_thread_cr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  724. int xgemm_thread_cc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  725. int cgemm3m_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  726. int cgemm3m_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  727. int cgemm3m_nr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  728. int cgemm3m_nc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  729. int cgemm3m_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  730. int cgemm3m_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  731. int cgemm3m_tr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  732. int cgemm3m_tc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  733. int cgemm3m_rn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  734. int cgemm3m_rt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  735. int cgemm3m_rr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  736. int cgemm3m_rc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  737. int cgemm3m_cn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  738. int cgemm3m_ct(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  739. int cgemm3m_cr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  740. int cgemm3m_cc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  741. int zgemm3m_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  742. int zgemm3m_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  743. int zgemm3m_nr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  744. int zgemm3m_nc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  745. int zgemm3m_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  746. int zgemm3m_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  747. int zgemm3m_tr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  748. int zgemm3m_tc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  749. int zgemm3m_rn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  750. int zgemm3m_rt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  751. int zgemm3m_rr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  752. int zgemm3m_rc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  753. int zgemm3m_cn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  754. int zgemm3m_ct(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  755. int zgemm3m_cr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  756. int zgemm3m_cc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  757. int xgemm3m_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  758. int xgemm3m_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  759. int xgemm3m_nr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  760. int xgemm3m_nc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  761. int xgemm3m_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  762. int xgemm3m_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  763. int xgemm3m_tr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  764. int xgemm3m_tc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  765. int xgemm3m_rn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  766. int xgemm3m_rt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  767. int xgemm3m_rr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  768. int xgemm3m_rc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  769. int xgemm3m_cn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  770. int xgemm3m_ct(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  771. int xgemm3m_cr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  772. int xgemm3m_cc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  773. int cgemm3m_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  774. int cgemm3m_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  775. int cgemm3m_thread_nr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  776. int cgemm3m_thread_nc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  777. int cgemm3m_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  778. int cgemm3m_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  779. int cgemm3m_thread_tr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  780. int cgemm3m_thread_tc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  781. int cgemm3m_thread_rn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  782. int cgemm3m_thread_rt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  783. int cgemm3m_thread_rr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  784. int cgemm3m_thread_rc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  785. int cgemm3m_thread_cn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  786. int cgemm3m_thread_ct(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  787. int cgemm3m_thread_cr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  788. int cgemm3m_thread_cc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  789. int zgemm3m_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  790. int zgemm3m_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  791. int zgemm3m_thread_nr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  792. int zgemm3m_thread_nc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  793. int zgemm3m_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  794. int zgemm3m_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  795. int zgemm3m_thread_tr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  796. int zgemm3m_thread_tc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  797. int zgemm3m_thread_rn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  798. int zgemm3m_thread_rt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  799. int zgemm3m_thread_rr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  800. int zgemm3m_thread_rc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  801. int zgemm3m_thread_cn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  802. int zgemm3m_thread_ct(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  803. int zgemm3m_thread_cr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  804. int zgemm3m_thread_cc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  805. int xgemm3m_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  806. int xgemm3m_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  807. int xgemm3m_thread_nr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  808. int xgemm3m_thread_nc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  809. int xgemm3m_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  810. int xgemm3m_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  811. int xgemm3m_thread_tr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  812. int xgemm3m_thread_tc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  813. int xgemm3m_thread_rn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  814. int xgemm3m_thread_rt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  815. int xgemm3m_thread_rr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  816. int xgemm3m_thread_rc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  817. int xgemm3m_thread_cn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  818. int xgemm3m_thread_ct(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  819. int xgemm3m_thread_cr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  820. int xgemm3m_thread_cc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  821. int cher2m_LNN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  822. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  823. int cher2m_LNT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  824. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  825. int cher2m_LNR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  826. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  827. int cher2m_LNC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  828. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  829. int cher2m_LTN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  830. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  831. int cher2m_LTT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  832. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  833. int cher2m_LTR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  834. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  835. int cher2m_LTC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  836. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  837. int cher2m_LRN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  838. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  839. int cher2m_LRT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  840. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  841. int cher2m_LRR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  842. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  843. int cher2m_LRC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  844. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  845. int cher2m_LCN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  846. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  847. int cher2m_LCT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  848. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  849. int cher2m_LCR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  850. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  851. int cher2m_LCC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  852. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  853. int cher2m_UNN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  854. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  855. int cher2m_UNT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  856. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  857. int cher2m_UNR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  858. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  859. int cher2m_UNC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  860. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  861. int cher2m_UTN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  862. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  863. int cher2m_UTT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  864. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  865. int cher2m_UTR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  866. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  867. int cher2m_UTC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  868. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  869. int cher2m_URN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  870. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  871. int cher2m_URT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  872. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  873. int cher2m_URR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  874. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  875. int cher2m_URC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  876. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  877. int cher2m_UCN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  878. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  879. int cher2m_UCT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  880. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  881. int cher2m_UCR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  882. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  883. int cher2m_UCC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  884. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  885. int zher2m_LNN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  886. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  887. int zher2m_LNT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  888. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  889. int zher2m_LNR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  890. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  891. int zher2m_LNC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  892. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  893. int zher2m_LTN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  894. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  895. int zher2m_LTT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  896. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  897. int zher2m_LTR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  898. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  899. int zher2m_LTC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  900. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  901. int zher2m_LRN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  902. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  903. int zher2m_LRT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  904. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  905. int zher2m_LRR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  906. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  907. int zher2m_LRC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  908. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  909. int zher2m_LCN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  910. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  911. int zher2m_LCT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  912. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  913. int zher2m_LCR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  914. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  915. int zher2m_LCC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  916. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  917. int zher2m_UNN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  918. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  919. int zher2m_UNT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  920. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  921. int zher2m_UNR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  922. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  923. int zher2m_UNC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  924. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  925. int zher2m_UTN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  926. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  927. int zher2m_UTT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  928. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  929. int zher2m_UTR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  930. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  931. int zher2m_UTC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  932. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  933. int zher2m_URN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  934. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  935. int zher2m_URT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  936. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  937. int zher2m_URR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  938. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  939. int zher2m_URC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  940. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  941. int zher2m_UCN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  942. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  943. int zher2m_UCT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  944. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  945. int zher2m_UCR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  946. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  947. int zher2m_UCC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  948. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  949. int strsm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  950. int strsm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  951. int strsm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  952. int strsm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  953. int strsm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  954. int strsm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  955. int strsm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  956. int strsm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  957. int strsm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  958. int strsm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  959. int strsm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  960. int strsm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  961. int strsm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  962. int strsm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  963. int strsm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  964. int strsm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  965. int dtrsm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  966. int dtrsm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  967. int dtrsm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  968. int dtrsm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  969. int dtrsm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  970. int dtrsm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  971. int dtrsm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  972. int dtrsm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  973. int dtrsm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  974. int dtrsm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  975. int dtrsm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  976. int dtrsm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  977. int dtrsm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  978. int dtrsm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  979. int dtrsm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  980. int dtrsm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  981. int qtrsm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  982. int qtrsm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  983. int qtrsm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  984. int qtrsm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  985. int qtrsm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  986. int qtrsm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  987. int qtrsm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  988. int qtrsm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  989. int qtrsm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  990. int qtrsm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  991. int qtrsm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  992. int qtrsm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  993. int qtrsm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  994. int qtrsm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  995. int qtrsm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  996. int qtrsm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  997. int ctrsm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  998. int ctrsm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  999. int ctrsm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1000. int ctrsm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1001. int ctrsm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1002. int ctrsm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1003. int ctrsm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1004. int ctrsm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1005. int ctrsm_LRUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1006. int ctrsm_LRUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1007. int ctrsm_LRLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1008. int ctrsm_LRLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1009. int ctrsm_LCUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1010. int ctrsm_LCUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1011. int ctrsm_LCLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1012. int ctrsm_LCLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1013. int ctrsm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1014. int ctrsm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1015. int ctrsm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1016. int ctrsm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1017. int ctrsm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1018. int ctrsm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1019. int ctrsm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1020. int ctrsm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1021. int ctrsm_RRUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1022. int ctrsm_RRUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1023. int ctrsm_RRLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1024. int ctrsm_RRLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1025. int ctrsm_RCUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1026. int ctrsm_RCUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1027. int ctrsm_RCLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1028. int ctrsm_RCLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1029. int ztrsm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1030. int ztrsm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1031. int ztrsm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1032. int ztrsm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1033. int ztrsm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1034. int ztrsm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1035. int ztrsm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1036. int ztrsm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1037. int ztrsm_LRUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1038. int ztrsm_LRUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1039. int ztrsm_LRLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1040. int ztrsm_LRLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1041. int ztrsm_LCUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1042. int ztrsm_LCUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1043. int ztrsm_LCLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1044. int ztrsm_LCLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1045. int ztrsm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1046. int ztrsm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1047. int ztrsm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1048. int ztrsm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1049. int ztrsm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1050. int ztrsm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1051. int ztrsm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1052. int ztrsm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1053. int ztrsm_RRUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1054. int ztrsm_RRUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1055. int ztrsm_RRLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1056. int ztrsm_RRLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1057. int ztrsm_RCUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1058. int ztrsm_RCUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1059. int ztrsm_RCLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1060. int ztrsm_RCLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1061. int xtrsm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1062. int xtrsm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1063. int xtrsm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1064. int xtrsm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1065. int xtrsm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1066. int xtrsm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1067. int xtrsm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1068. int xtrsm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1069. int xtrsm_LRUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1070. int xtrsm_LRUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1071. int xtrsm_LRLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1072. int xtrsm_LRLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1073. int xtrsm_LCUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1074. int xtrsm_LCUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1075. int xtrsm_LCLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1076. int xtrsm_LCLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1077. int xtrsm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1078. int xtrsm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1079. int xtrsm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1080. int xtrsm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1081. int xtrsm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1082. int xtrsm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1083. int xtrsm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1084. int xtrsm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1085. int xtrsm_RRUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1086. int xtrsm_RRUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1087. int xtrsm_RRLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1088. int xtrsm_RRLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1089. int xtrsm_RCUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1090. int xtrsm_RCUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1091. int xtrsm_RCLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1092. int xtrsm_RCLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1093. int strmm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1094. int strmm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1095. int strmm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1096. int strmm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1097. int strmm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1098. int strmm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1099. int strmm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1100. int strmm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1101. int strmm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1102. int strmm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1103. int strmm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1104. int strmm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1105. int strmm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1106. int strmm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1107. int strmm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1108. int strmm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1109. int dtrmm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1110. int dtrmm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1111. int dtrmm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1112. int dtrmm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1113. int dtrmm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1114. int dtrmm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1115. int dtrmm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1116. int dtrmm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1117. int dtrmm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1118. int dtrmm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1119. int dtrmm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1120. int dtrmm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1121. int dtrmm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1122. int dtrmm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1123. int dtrmm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1124. int dtrmm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1125. int qtrmm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1126. int qtrmm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1127. int qtrmm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1128. int qtrmm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1129. int qtrmm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1130. int qtrmm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1131. int qtrmm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1132. int qtrmm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1133. int qtrmm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1134. int qtrmm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1135. int qtrmm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1136. int qtrmm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1137. int qtrmm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1138. int qtrmm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1139. int qtrmm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1140. int qtrmm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1141. int ctrmm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1142. int ctrmm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1143. int ctrmm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1144. int ctrmm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1145. int ctrmm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1146. int ctrmm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1147. int ctrmm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1148. int ctrmm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1149. int ctrmm_LRUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1150. int ctrmm_LRUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1151. int ctrmm_LRLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1152. int ctrmm_LRLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1153. int ctrmm_LCUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1154. int ctrmm_LCUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1155. int ctrmm_LCLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1156. int ctrmm_LCLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1157. int ctrmm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1158. int ctrmm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1159. int ctrmm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1160. int ctrmm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1161. int ctrmm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1162. int ctrmm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1163. int ctrmm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1164. int ctrmm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1165. int ctrmm_RRUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1166. int ctrmm_RRUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1167. int ctrmm_RRLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1168. int ctrmm_RRLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1169. int ctrmm_RCUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1170. int ctrmm_RCUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1171. int ctrmm_RCLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1172. int ctrmm_RCLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1173. int ztrmm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1174. int ztrmm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1175. int ztrmm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1176. int ztrmm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1177. int ztrmm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1178. int ztrmm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1179. int ztrmm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1180. int ztrmm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1181. int ztrmm_LRUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1182. int ztrmm_LRUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1183. int ztrmm_LRLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1184. int ztrmm_LRLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1185. int ztrmm_LCUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1186. int ztrmm_LCUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1187. int ztrmm_LCLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1188. int ztrmm_LCLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1189. int ztrmm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1190. int ztrmm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1191. int ztrmm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1192. int ztrmm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1193. int ztrmm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1194. int ztrmm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1195. int ztrmm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1196. int ztrmm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1197. int ztrmm_RRUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1198. int ztrmm_RRUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1199. int ztrmm_RRLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1200. int ztrmm_RRLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1201. int ztrmm_RCUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1202. int ztrmm_RCUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1203. int ztrmm_RCLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1204. int ztrmm_RCLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1205. int xtrmm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1206. int xtrmm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1207. int xtrmm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1208. int xtrmm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1209. int xtrmm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1210. int xtrmm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1211. int xtrmm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1212. int xtrmm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1213. int xtrmm_LRUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1214. int xtrmm_LRUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1215. int xtrmm_LRLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1216. int xtrmm_LRLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1217. int xtrmm_LCUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1218. int xtrmm_LCUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1219. int xtrmm_LCLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1220. int xtrmm_LCLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1221. int xtrmm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1222. int xtrmm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1223. int xtrmm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1224. int xtrmm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1225. int xtrmm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1226. int xtrmm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1227. int xtrmm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1228. int xtrmm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1229. int xtrmm_RRUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1230. int xtrmm_RRUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1231. int xtrmm_RRLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1232. int xtrmm_RRLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1233. int xtrmm_RCUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1234. int xtrmm_RCUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1235. int xtrmm_RCLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1236. int xtrmm_RCLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1237. int ssymm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1238. int ssymm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1239. int ssymm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1240. int ssymm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1241. int dsymm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1242. int dsymm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1243. int dsymm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1244. int dsymm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1245. int qsymm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1246. int qsymm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1247. int qsymm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1248. int qsymm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1249. int csymm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1250. int csymm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1251. int csymm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1252. int csymm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1253. int zsymm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1254. int zsymm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1255. int zsymm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1256. int zsymm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1257. int xsymm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1258. int xsymm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1259. int xsymm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1260. int xsymm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1261. int csymm3m_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1262. int csymm3m_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1263. int csymm3m_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1264. int csymm3m_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1265. int zsymm3m_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1266. int zsymm3m_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1267. int zsymm3m_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1268. int zsymm3m_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1269. int xsymm3m_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1270. int xsymm3m_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1271. int xsymm3m_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1272. int xsymm3m_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1273. int csymm3m_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1274. int csymm3m_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1275. int csymm3m_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1276. int csymm3m_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1277. int zsymm3m_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1278. int zsymm3m_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1279. int zsymm3m_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1280. int zsymm3m_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1281. int xsymm3m_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1282. int xsymm3m_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1283. int xsymm3m_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1284. int xsymm3m_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1285. int chemm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1286. int chemm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1287. int chemm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1288. int chemm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1289. int zhemm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1290. int zhemm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1291. int zhemm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1292. int zhemm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1293. int xhemm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1294. int xhemm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1295. int xhemm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1296. int xhemm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1297. int chemm3m_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1298. int chemm3m_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1299. int chemm3m_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1300. int chemm3m_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1301. int zhemm3m_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1302. int zhemm3m_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1303. int zhemm3m_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1304. int zhemm3m_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1305. int xhemm3m_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1306. int xhemm3m_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1307. int xhemm3m_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1308. int xhemm3m_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1309. int chemm3m_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1310. int chemm3m_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1311. int chemm3m_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1312. int chemm3m_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1313. int zhemm3m_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1314. int zhemm3m_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1315. int zhemm3m_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1316. int zhemm3m_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1317. int xhemm3m_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1318. int xhemm3m_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1319. int xhemm3m_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1320. int xhemm3m_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1321. int ssymm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1322. int ssymm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1323. int ssymm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1324. int ssymm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1325. int dsymm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1326. int dsymm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1327. int dsymm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1328. int dsymm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1329. int qsymm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1330. int qsymm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1331. int qsymm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1332. int qsymm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1333. int csymm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1334. int csymm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1335. int csymm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1336. int csymm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1337. int zsymm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1338. int zsymm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1339. int zsymm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1340. int zsymm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1341. int xsymm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1342. int xsymm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1343. int xsymm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1344. int xsymm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1345. int chemm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1346. int chemm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1347. int chemm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1348. int chemm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1349. int zhemm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1350. int zhemm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1351. int zhemm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1352. int zhemm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1353. int xhemm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1354. int xhemm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1355. int xhemm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1356. int xhemm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1357. int ssyrk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1358. int ssyrk_UT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1359. int ssyrk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1360. int ssyrk_LT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1361. int dsyrk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1362. int dsyrk_UT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1363. int dsyrk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1364. int dsyrk_LT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1365. int qsyrk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1366. int qsyrk_UT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1367. int qsyrk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1368. int qsyrk_LT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1369. int csyrk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1370. int csyrk_UT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1371. int csyrk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1372. int csyrk_LT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1373. int zsyrk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1374. int zsyrk_UT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1375. int zsyrk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1376. int zsyrk_LT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1377. int xsyrk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1378. int xsyrk_UT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1379. int xsyrk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1380. int xsyrk_LT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1381. int ssyrk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1382. int ssyrk_thread_UT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1383. int ssyrk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1384. int ssyrk_thread_LT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1385. int dsyrk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1386. int dsyrk_thread_UT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1387. int dsyrk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1388. int dsyrk_thread_LT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1389. int qsyrk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1390. int qsyrk_thread_UT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1391. int qsyrk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1392. int qsyrk_thread_LT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1393. int csyrk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1394. int csyrk_thread_UT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1395. int csyrk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1396. int csyrk_thread_LT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1397. int zsyrk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1398. int zsyrk_thread_UT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1399. int zsyrk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1400. int zsyrk_thread_LT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1401. int xsyrk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1402. int xsyrk_thread_UT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1403. int xsyrk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1404. int xsyrk_thread_LT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1405. int ssyr2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1406. int ssyr2k_UT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1407. int ssyr2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1408. int ssyr2k_LT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1409. int dsyr2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1410. int dsyr2k_UT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1411. int dsyr2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1412. int dsyr2k_LT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1413. int qsyr2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1414. int qsyr2k_UT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1415. int qsyr2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1416. int qsyr2k_LT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1417. int csyr2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1418. int csyr2k_UT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1419. int csyr2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1420. int csyr2k_LT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1421. int zsyr2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1422. int zsyr2k_UT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1423. int zsyr2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1424. int zsyr2k_LT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1425. int xsyr2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1426. int xsyr2k_UT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1427. int xsyr2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1428. int xsyr2k_LT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1429. int cherk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1430. int cherk_UC(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1431. int cherk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1432. int cherk_LC(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1433. int zherk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1434. int zherk_UC(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1435. int zherk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1436. int zherk_LC(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1437. int xherk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1438. int xherk_UC(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1439. int xherk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1440. int xherk_LC(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1441. int cherk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1442. int cherk_thread_UC(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1443. int cherk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1444. int cherk_thread_LC(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1445. int zherk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1446. int zherk_thread_UC(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1447. int zherk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1448. int zherk_thread_LC(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1449. int xherk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1450. int xherk_thread_UC(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1451. int xherk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1452. int xherk_thread_LC(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1453. int cher2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1454. int cher2k_UC(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1455. int cher2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1456. int cher2k_LC(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1457. int zher2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1458. int zher2k_UC(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1459. int zher2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1460. int zher2k_LC(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1461. int xher2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1462. int xher2k_UC(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1463. int xher2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1464. int xher2k_LC(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1465. int sgemt_n(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, int);
  1466. int sgemt_t(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, int);
  1467. int dgemt_n(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, int);
  1468. int dgemt_t(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, int);
  1469. int cgemt_n(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, int);
  1470. int cgemt_t(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, int);
  1471. int cgemt_r(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, int);
  1472. int cgemt_c(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, int);
  1473. int zgemt_n(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, int);
  1474. int zgemt_t(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, int);
  1475. int zgemt_r(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, int);
  1476. int zgemt_c(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, int);
  1477. int sgema_n(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG);
  1478. int sgema_t(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG);
  1479. int dgema_n(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG);
  1480. int dgema_t(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG);
  1481. int cgema_n(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1482. int cgema_t(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1483. int cgema_r(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1484. int cgema_c(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1485. int zgema_n(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1486. int zgema_t(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1487. int zgema_r(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1488. int zgema_c(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1489. int cgemm3m_incopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  1490. int cgemm3m_incopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  1491. int cgemm3m_incopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  1492. int cgemm3m_itcopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  1493. int cgemm3m_itcopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  1494. int cgemm3m_itcopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  1495. int cgemm3m_oncopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float alpha_r, float alpha_i, float *b);
  1496. int cgemm3m_oncopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float alpha_r, float alpha_i, float *b);
  1497. int cgemm3m_oncopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float alpha_r, float alpha_i, float *b);
  1498. int cgemm3m_otcopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float alpha_r, float alpha_i, float *b);
  1499. int cgemm3m_otcopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float alpha_r, float alpha_i, float *b);
  1500. int cgemm3m_otcopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float alpha_r, float alpha_i, float *b);
  1501. int zgemm3m_incopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  1502. int zgemm3m_incopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  1503. int zgemm3m_incopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  1504. int zgemm3m_itcopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  1505. int zgemm3m_itcopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  1506. int zgemm3m_itcopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  1507. int zgemm3m_oncopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double alpha_r, double alpha_i, double *b);
  1508. int zgemm3m_oncopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double alpha_r, double alpha_i, double *b);
  1509. int zgemm3m_oncopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double alpha_r, double alpha_i, double *b);
  1510. int zgemm3m_otcopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double alpha_r, double alpha_i, double *b);
  1511. int zgemm3m_otcopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double alpha_r, double alpha_i, double *b);
  1512. int zgemm3m_otcopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double alpha_r, double alpha_i, double *b);
  1513. int xgemm3m_incopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  1514. int xgemm3m_incopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  1515. int xgemm3m_incopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  1516. int xgemm3m_itcopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  1517. int xgemm3m_itcopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  1518. int xgemm3m_itcopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  1519. int xgemm3m_oncopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1520. int xgemm3m_oncopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1521. int xgemm3m_oncopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1522. int xgemm3m_otcopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1523. int xgemm3m_otcopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1524. int xgemm3m_otcopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1525. int csymm3m_iucopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1526. int csymm3m_ilcopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1527. int csymm3m_iucopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1528. int csymm3m_ilcopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1529. int csymm3m_iucopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1530. int csymm3m_ilcopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1531. int csymm3m_oucopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1532. int csymm3m_olcopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1533. int csymm3m_oucopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1534. int csymm3m_olcopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1535. int csymm3m_oucopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1536. int csymm3m_olcopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1537. int zsymm3m_iucopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1538. int zsymm3m_ilcopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1539. int zsymm3m_iucopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1540. int zsymm3m_ilcopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1541. int zsymm3m_iucopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1542. int zsymm3m_ilcopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1543. int zsymm3m_oucopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1544. int zsymm3m_olcopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1545. int zsymm3m_oucopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1546. int zsymm3m_olcopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1547. int zsymm3m_oucopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1548. int zsymm3m_olcopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1549. int xsymm3m_iucopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1550. int xsymm3m_ilcopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1551. int xsymm3m_iucopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1552. int xsymm3m_ilcopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1553. int xsymm3m_iucopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1554. int xsymm3m_ilcopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1555. int xsymm3m_oucopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1556. int xsymm3m_olcopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1557. int xsymm3m_oucopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1558. int xsymm3m_olcopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1559. int xsymm3m_oucopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1560. int xsymm3m_olcopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1561. int chemm3m_iucopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1562. int chemm3m_ilcopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1563. int chemm3m_iucopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1564. int chemm3m_ilcopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1565. int chemm3m_iucopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1566. int chemm3m_ilcopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1567. int chemm3m_oucopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1568. int chemm3m_olcopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1569. int chemm3m_oucopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1570. int chemm3m_olcopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1571. int chemm3m_oucopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1572. int chemm3m_olcopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1573. int zhemm3m_iucopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1574. int zhemm3m_ilcopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1575. int zhemm3m_iucopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1576. int zhemm3m_ilcopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1577. int zhemm3m_iucopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1578. int zhemm3m_ilcopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1579. int zhemm3m_oucopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1580. int zhemm3m_olcopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1581. int zhemm3m_oucopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1582. int zhemm3m_olcopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1583. int zhemm3m_oucopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1584. int zhemm3m_olcopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1585. int xhemm3m_iucopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1586. int xhemm3m_ilcopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1587. int xhemm3m_iucopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1588. int xhemm3m_ilcopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1589. int xhemm3m_iucopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1590. int xhemm3m_ilcopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1591. int xhemm3m_oucopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1592. int xhemm3m_olcopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1593. int xhemm3m_oucopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1594. int xhemm3m_olcopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1595. int xhemm3m_oucopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1596. int xhemm3m_olcopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1597. int sgemc_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1598. int sgemc_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1599. int sgemc_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1600. int sgemc_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1601. int dgemc_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1602. int dgemc_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1603. int dgemc_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1604. int dgemc_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1605. int qgemc_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1606. int qgemc_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1607. int qgemc_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1608. int qgemc_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1609. int cgemc_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1610. int cgemc_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1611. int cgemc_nr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1612. int cgemc_nc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1613. int cgemc_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1614. int cgemc_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1615. int cgemc_tr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1616. int cgemc_tc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1617. int cgemc_rn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1618. int cgemc_rt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1619. int cgemc_rr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1620. int cgemc_rc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1621. int cgemc_cn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1622. int cgemc_ct(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1623. int cgemc_cr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1624. int cgemc_cc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1625. int zgemc_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1626. int zgemc_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1627. int zgemc_nr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1628. int zgemc_nc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1629. int zgemc_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1630. int zgemc_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1631. int zgemc_tr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1632. int zgemc_tc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1633. int zgemc_rn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1634. int zgemc_rt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1635. int zgemc_rr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1636. int zgemc_rc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1637. int zgemc_cn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1638. int zgemc_ct(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1639. int zgemc_cr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1640. int zgemc_cc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1641. int xgemc_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1642. int xgemc_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1643. int xgemc_nr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1644. int xgemc_nc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1645. int xgemc_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1646. int xgemc_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1647. int xgemc_tr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1648. int xgemc_tc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1649. int xgemc_rn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1650. int xgemc_rt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1651. int xgemc_rr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1652. int xgemc_rc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1653. int xgemc_cn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1654. int xgemc_ct(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1655. int xgemc_cr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1656. int xgemc_cc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1657. int sgemc_oncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b, BLASLONG ldb, float *c);
  1658. int sgemc_otcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b, BLASLONG ldb, float *c);
  1659. int dgemc_oncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b, BLASLONG ldb, double *c);
  1660. int dgemc_otcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b, BLASLONG ldb, double *c);
  1661. int qgemc_oncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b, BLASLONG ldb, xdouble *c);
  1662. int qgemc_otcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b, BLASLONG ldb, xdouble *c);
  1663. int cgemc_oncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b, BLASLONG ldb, float *c);
  1664. int cgemc_otcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b, BLASLONG ldb, float *c);
  1665. int zgemc_oncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b, BLASLONG ldb, double *c);
  1666. int zgemc_otcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b, BLASLONG ldb, double *c);
  1667. int xgemc_oncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b, BLASLONG ldb, xdouble *c);
  1668. int xgemc_otcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b, BLASLONG ldb, xdouble *c);
  1669. int somatcopy_k_cn(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG);
  1670. int somatcopy_k_rn(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG);
  1671. int somatcopy_k_ct(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG);
  1672. int somatcopy_k_rt(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG);
  1673. int simatcopy_k_cn(BLASLONG, BLASLONG, float, float *, BLASLONG);
  1674. int simatcopy_k_rn(BLASLONG, BLASLONG, float, float *, BLASLONG);
  1675. int simatcopy_k_ct(BLASLONG, BLASLONG, float, float *, BLASLONG);
  1676. int simatcopy_k_rt(BLASLONG, BLASLONG, float, float *, BLASLONG);
  1677. int domatcopy_k_cn(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG);
  1678. int domatcopy_k_rn(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG);
  1679. int domatcopy_k_ct(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG);
  1680. int domatcopy_k_rt(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG);
  1681. int dimatcopy_k_cn(BLASLONG, BLASLONG, double, double *, BLASLONG);
  1682. int dimatcopy_k_rn(BLASLONG, BLASLONG, double, double *, BLASLONG);
  1683. int dimatcopy_k_ct(BLASLONG, BLASLONG, double, double *, BLASLONG);
  1684. int dimatcopy_k_rt(BLASLONG, BLASLONG, double, double *, BLASLONG);
  1685. int comatcopy_k_cn(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1686. int comatcopy_k_rn(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1687. int comatcopy_k_ct(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1688. int comatcopy_k_rt(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1689. int cimatcopy_k_cn(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1690. int cimatcopy_k_rn(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1691. int cimatcopy_k_ct(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1692. int cimatcopy_k_rt(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1693. int comatcopy_k_cnc(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1694. int comatcopy_k_rnc(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1695. int comatcopy_k_ctc(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1696. int comatcopy_k_rtc(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1697. int cimatcopy_k_cnc(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1698. int cimatcopy_k_rnc(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1699. int cimatcopy_k_ctc(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1700. int cimatcopy_k_rtc(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1701. int zomatcopy_k_cn(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1702. int zomatcopy_k_rn(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1703. int zomatcopy_k_ct(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1704. int zomatcopy_k_rt(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1705. int zimatcopy_k_cn(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1706. int zimatcopy_k_rn(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1707. int zimatcopy_k_ct(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1708. int zimatcopy_k_rt(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1709. int zomatcopy_k_cnc(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1710. int zomatcopy_k_rnc(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1711. int zomatcopy_k_ctc(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1712. int zomatcopy_k_rtc(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1713. int zimatcopy_k_cnc(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1714. int zimatcopy_k_rnc(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1715. int zimatcopy_k_ctc(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1716. int zimatcopy_k_rtc(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1717. int sgeadd_k(BLASLONG, BLASLONG, float, float*, BLASLONG, float, float *, BLASLONG);
  1718. int dgeadd_k(BLASLONG, BLASLONG, double, double*, BLASLONG, double, double *, BLASLONG);
  1719. int cgeadd_k(BLASLONG, BLASLONG, float, float, float*, BLASLONG, float, float, float *, BLASLONG);
  1720. int zgeadd_k(BLASLONG, BLASLONG, double,double, double*, BLASLONG, double, double, double *, BLASLONG);
  1721. #ifdef __CUDACC__
  1722. }
  1723. #endif
  1724. #endif