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 167 kB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868186918701871187218731874187518761877187818791880188118821883188418851886188718881889189018911892189318941895189618971898189919001901190219031904190519061907190819091910191119121913191419151916191719181919192019211922192319241925192619271928192919301931193219331934193519361937193819391940194119421943194419451946194719481949195019511952195319541955195619571958195919601961196219631964196519661967196819691970197119721973197419751976
  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. void sgemm_direct_alpha_beta(BLASLONG M, BLASLONG N, BLASLONG K,
  51. float alpha,
  52. float * A, BLASLONG strideA,
  53. float * B, BLASLONG strideB,
  54. float beta,
  55. float * R, BLASLONG strideR);
  56. int sgemm_direct_performant(BLASLONG M, BLASLONG N, BLASLONG K);
  57. int shgemm_beta(BLASLONG, BLASLONG, BLASLONG, float,
  58. hfloat16 *, BLASLONG, hfloat16 *, BLASLONG, float *, BLASLONG);
  59. int sbgemm_beta(BLASLONG, BLASLONG, BLASLONG, float,
  60. bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, float *, BLASLONG);
  61. int sgemm_beta(BLASLONG, BLASLONG, BLASLONG, float,
  62. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  63. int dgemm_beta(BLASLONG, BLASLONG, BLASLONG, double,
  64. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  65. int cgemm_beta(BLASLONG, BLASLONG, BLASLONG, float, float,
  66. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  67. int zgemm_beta(BLASLONG, BLASLONG, BLASLONG, double, double,
  68. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  69. #ifdef EXPRECISION
  70. int qgemm_beta(BLASLONG, BLASLONG, BLASLONG, xdouble,
  71. xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  72. int xgemm_beta(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble,
  73. xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  74. #else
  75. int qgemm_beta(BLASLONG, BLASLONG, BLASLONG, xdouble *,
  76. xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  77. int xgemm_beta(BLASLONG, BLASLONG, BLASLONG, xdouble *,
  78. xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  79. #endif
  80. int shgemm_incopy(BLASLONG m, BLASLONG n, hfloat16 *a, BLASLONG lda, hfloat16 *b);
  81. int shgemm_itcopy(BLASLONG m, BLASLONG n, hfloat16 *a, BLASLONG lda, hfloat16 *b);
  82. int shgemm_oncopy(BLASLONG m, BLASLONG n, hfloat16 *a, BLASLONG lda, hfloat16 *b);
  83. int shgemm_otcopy(BLASLONG m, BLASLONG n, hfloat16 *a, BLASLONG lda, hfloat16 *b);
  84. int sbgemm_incopy(BLASLONG m, BLASLONG n, bfloat16 *a, BLASLONG lda, bfloat16 *b);
  85. int sbgemm_itcopy(BLASLONG m, BLASLONG n, bfloat16 *a, BLASLONG lda, bfloat16 *b);
  86. int sbgemm_oncopy(BLASLONG m, BLASLONG n, bfloat16 *a, BLASLONG lda, bfloat16 *b);
  87. int sbgemm_otcopy(BLASLONG m, BLASLONG n, bfloat16 *a, BLASLONG lda, bfloat16 *b);
  88. int sgemm_incopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  89. int sgemm_itcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  90. int sgemm_oncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  91. int sgemm_otcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  92. int dgemm_incopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  93. int dgemm_itcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  94. int dgemm_oncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  95. int dgemm_otcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  96. int cgemm_incopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  97. int cgemm_itcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  98. int cgemm_oncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  99. int cgemm_otcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  100. int zgemm_incopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  101. int zgemm_itcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  102. int zgemm_oncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  103. int zgemm_otcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  104. #ifdef QUAD_PRECISION
  105. int qgemm_incopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xidouble *b);
  106. int qgemm_itcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xidouble *b);
  107. int qgemm_oncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xidouble *b);
  108. int qgemm_otcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xidouble *b);
  109. int xgemm_incopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xidouble *b);
  110. int xgemm_itcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xidouble *b);
  111. int xgemm_oncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xidouble *b);
  112. int xgemm_otcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xidouble *b);
  113. #else
  114. int qgemm_incopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  115. int qgemm_itcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  116. int qgemm_oncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  117. int qgemm_otcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  118. int xgemm_incopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  119. int xgemm_itcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  120. int xgemm_oncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  121. int xgemm_otcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  122. #endif
  123. int strsm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  124. int strsm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  125. int strsm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  126. int strsm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  127. int dtrsm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  128. int dtrsm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  129. int dtrsm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  130. int dtrsm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  131. int qtrsm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  132. int qtrsm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  133. int qtrsm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  134. int qtrsm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  135. int ctrsm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  136. int ctrsm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  137. int ctrsm_kernel_LR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  138. int ctrsm_kernel_LC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  139. int ctrsm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  140. int ctrsm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  141. int ctrsm_kernel_RR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  142. int ctrsm_kernel_RC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  143. int ztrsm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  144. int ztrsm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  145. int ztrsm_kernel_LR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  146. int ztrsm_kernel_LC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  147. int ztrsm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  148. int ztrsm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  149. int ztrsm_kernel_RR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  150. int ztrsm_kernel_RC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  151. int xtrsm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  152. int xtrsm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  153. int xtrsm_kernel_LR(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  154. int xtrsm_kernel_LC(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  155. int xtrsm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  156. int xtrsm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  157. int xtrsm_kernel_RR(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  158. int xtrsm_kernel_RC(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  159. int strmm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  160. int strmm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  161. int strmm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  162. int strmm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  163. int dtrmm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  164. int dtrmm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  165. int dtrmm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  166. int dtrmm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  167. int qtrmm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  168. int qtrmm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  169. int qtrmm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  170. int qtrmm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  171. int ctrmm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  172. int ctrmm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  173. int ctrmm_kernel_RR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  174. int ctrmm_kernel_RC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  175. int ctrmm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  176. int ctrmm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  177. int ctrmm_kernel_LR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  178. int ctrmm_kernel_LC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  179. int ztrmm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  180. int ztrmm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  181. int ztrmm_kernel_RR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  182. int ztrmm_kernel_RC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  183. int ztrmm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  184. int ztrmm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  185. int ztrmm_kernel_LR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  186. int ztrmm_kernel_LC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  187. int xtrmm_kernel_RN(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  188. int xtrmm_kernel_RT(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  189. int xtrmm_kernel_RR(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  190. int xtrmm_kernel_RC(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  191. int xtrmm_kernel_LN(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  192. int xtrmm_kernel_LT(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  193. int xtrmm_kernel_LR(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  194. int xtrmm_kernel_LC(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  195. int strmm_iunucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  196. int strmm_iunncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  197. int strmm_iutucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  198. int strmm_iutncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  199. int strmm_ounucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  200. int strmm_ounncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  201. int strmm_outucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  202. int strmm_outncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  203. int strmm_ilnucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  204. int strmm_ilnncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  205. int strmm_iltucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  206. int strmm_iltncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  207. int strmm_olnucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  208. int strmm_olnncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  209. int strmm_oltucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  210. int strmm_oltncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  211. int dtrmm_iunucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  212. int dtrmm_iunncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  213. int dtrmm_iutucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  214. int dtrmm_iutncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  215. int dtrmm_ounucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  216. int dtrmm_ounncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  217. int dtrmm_outucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  218. int dtrmm_outncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  219. int dtrmm_ilnucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  220. int dtrmm_ilnncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  221. int dtrmm_iltucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  222. int dtrmm_iltncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  223. int dtrmm_olnucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  224. int dtrmm_olnncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  225. int dtrmm_oltucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  226. int dtrmm_oltncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  227. int qtrmm_iunucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  228. int qtrmm_iunncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  229. int qtrmm_iutucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  230. int qtrmm_iutncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  231. int qtrmm_ounucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  232. int qtrmm_ounncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  233. int qtrmm_outucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  234. int qtrmm_outncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  235. int qtrmm_ilnucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  236. int qtrmm_ilnncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  237. int qtrmm_iltucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  238. int qtrmm_iltncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  239. int qtrmm_olnucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  240. int qtrmm_olnncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  241. int qtrmm_oltucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  242. int qtrmm_oltncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  243. int ctrmm_iunucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  244. int ctrmm_iunncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  245. int ctrmm_iutucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  246. int ctrmm_iutncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  247. int ctrmm_ounucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  248. int ctrmm_ounncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  249. int ctrmm_outucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  250. int ctrmm_outncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  251. int ctrmm_ilnucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  252. int ctrmm_ilnncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  253. int ctrmm_iltucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  254. int ctrmm_iltncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  255. int ctrmm_olnucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  256. int ctrmm_olnncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  257. int ctrmm_oltucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  258. int ctrmm_oltncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  259. int ztrmm_iunucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  260. int ztrmm_iunncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  261. int ztrmm_iutucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  262. int ztrmm_iutncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  263. int ztrmm_ounucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  264. int ztrmm_ounncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  265. int ztrmm_outucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  266. int ztrmm_outncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  267. int ztrmm_ilnucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  268. int ztrmm_ilnncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  269. int ztrmm_iltucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  270. int ztrmm_iltncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  271. int ztrmm_olnucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  272. int ztrmm_olnncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  273. int ztrmm_oltucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  274. int ztrmm_oltncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  275. int xtrmm_iunucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  276. int xtrmm_iunncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  277. int xtrmm_iutucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  278. int xtrmm_iutncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  279. int xtrmm_ounucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  280. int xtrmm_ounncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  281. int xtrmm_outucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  282. int xtrmm_outncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  283. int xtrmm_ilnucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  284. int xtrmm_ilnncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  285. int xtrmm_iltucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  286. int xtrmm_iltncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  287. int xtrmm_olnucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  288. int xtrmm_olnncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  289. int xtrmm_oltucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  290. int xtrmm_oltncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  291. int strsm_iunucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  292. int strsm_iunncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  293. int strsm_iutucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  294. int strsm_iutncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  295. int strsm_ounucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  296. int strsm_ounncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  297. int strsm_outucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  298. int strsm_outncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  299. int strsm_ilnucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  300. int strsm_ilnncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  301. int strsm_iltucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  302. int strsm_iltncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  303. int strsm_olnucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  304. int strsm_olnncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  305. int strsm_oltucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  306. int strsm_oltncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  307. int dtrsm_iunucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  308. int dtrsm_iunncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  309. int dtrsm_iutucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  310. int dtrsm_iutncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  311. int dtrsm_ounucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  312. int dtrsm_ounncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  313. int dtrsm_outucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  314. int dtrsm_outncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  315. int dtrsm_ilnucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  316. int dtrsm_ilnncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  317. int dtrsm_iltucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  318. int dtrsm_iltncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  319. int dtrsm_olnucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  320. int dtrsm_olnncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  321. int dtrsm_oltucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  322. int dtrsm_oltncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  323. int qtrsm_iunucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  324. int qtrsm_iunncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  325. int qtrsm_iutucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  326. int qtrsm_iutncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  327. int qtrsm_ounucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  328. int qtrsm_ounncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  329. int qtrsm_outucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  330. int qtrsm_outncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  331. int qtrsm_ilnucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  332. int qtrsm_ilnncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  333. int qtrsm_iltucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  334. int qtrsm_iltncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  335. int qtrsm_olnucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  336. int qtrsm_olnncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  337. int qtrsm_oltucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  338. int qtrsm_oltncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  339. int ctrsm_iunucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  340. int ctrsm_iunncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  341. int ctrsm_iutucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  342. int ctrsm_iutncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  343. int ctrsm_ounucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  344. int ctrsm_ounncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  345. int ctrsm_outucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  346. int ctrsm_outncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  347. int ctrsm_ilnucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  348. int ctrsm_ilnncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  349. int ctrsm_iltucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  350. int ctrsm_iltncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  351. int ctrsm_olnucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  352. int ctrsm_olnncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  353. int ctrsm_oltucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  354. int ctrsm_oltncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG offset, float *b);
  355. int ztrsm_iunucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  356. int ztrsm_iunncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  357. int ztrsm_iutucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  358. int ztrsm_iutncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  359. int ztrsm_ounucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  360. int ztrsm_ounncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  361. int ztrsm_outucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  362. int ztrsm_outncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  363. int ztrsm_ilnucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  364. int ztrsm_ilnncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  365. int ztrsm_iltucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  366. int ztrsm_iltncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  367. int ztrsm_olnucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  368. int ztrsm_olnncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  369. int ztrsm_oltucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  370. int ztrsm_oltncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG offset, double *b);
  371. int xtrsm_iunucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  372. int xtrsm_iunncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  373. int xtrsm_iutucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  374. int xtrsm_iutncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  375. int xtrsm_ounucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  376. int xtrsm_ounncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  377. int xtrsm_outucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  378. int xtrsm_outncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  379. int xtrsm_ilnucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  380. int xtrsm_ilnncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  381. int xtrsm_iltucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  382. int xtrsm_iltncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  383. int xtrsm_olnucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  384. int xtrsm_olnncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  385. int xtrsm_oltucopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  386. int xtrsm_oltncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG offset, xdouble *b);
  387. int ssymm_iutcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  388. int ssymm_outcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  389. int ssymm_iltcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  390. int ssymm_oltcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  391. int dsymm_iutcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  392. int dsymm_outcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  393. int dsymm_iltcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  394. int dsymm_oltcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  395. int qsymm_iutcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  396. int qsymm_outcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  397. int qsymm_iltcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  398. int qsymm_oltcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  399. int csymm_iutcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  400. int csymm_outcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  401. int csymm_iltcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  402. int csymm_oltcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  403. int zsymm_iutcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  404. int zsymm_outcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  405. int zsymm_iltcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  406. int zsymm_oltcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  407. int xsymm_iutcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  408. int xsymm_outcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  409. int xsymm_iltcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  410. int xsymm_oltcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  411. int chemm_iutcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  412. int chemm_outcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  413. int chemm_iltcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  414. int chemm_oltcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, float *b);
  415. int zhemm_iutcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  416. int zhemm_outcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  417. int zhemm_iltcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  418. int zhemm_oltcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, double *b);
  419. int xhemm_iutcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  420. int xhemm_outcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  421. int xhemm_iltcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  422. int xhemm_oltcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, xdouble *b);
  423. int ssyrk_kernel_U(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset);
  424. int ssyrk_kernel_L(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset);
  425. int dsyrk_kernel_U(BLASLONG m, BLASLONG n, BLASLONG k, double alpha, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset);
  426. int dsyrk_kernel_L(BLASLONG m, BLASLONG n, BLASLONG k, double alpha, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset);
  427. int qsyrk_kernel_U(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset);
  428. int qsyrk_kernel_L(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset);
  429. 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);
  430. 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);
  431. 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);
  432. 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);
  433. 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);
  434. 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);
  435. int ssyr2k_kernel_U(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset, int flag);
  436. int ssyr2k_kernel_L(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset, int flag);
  437. int dsyr2k_kernel_U(BLASLONG m, BLASLONG n, BLASLONG k, double alpha, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset, int flag);
  438. int dsyr2k_kernel_L(BLASLONG m, BLASLONG n, BLASLONG k, double alpha, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset, int flag);
  439. int qsyr2k_kernel_U(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset, int flag);
  440. int qsyr2k_kernel_L(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset, int flag);
  441. 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);
  442. 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);
  443. 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);
  444. 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);
  445. 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);
  446. 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);
  447. int cherk_kernel_UN(BLASLONG m, BLASLONG n, BLASLONG k, float alpha_r, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset);
  448. int cherk_kernel_UC(BLASLONG m, BLASLONG n, BLASLONG k, float alpha_r, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset);
  449. int cherk_kernel_LN(BLASLONG m, BLASLONG n, BLASLONG k, float alpha_r, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset);
  450. int cherk_kernel_LC(BLASLONG m, BLASLONG n, BLASLONG k, float alpha_r, float *a, float *b, float *c, BLASLONG ldc, BLASLONG offset);
  451. int zherk_kernel_UN(BLASLONG m, BLASLONG n, BLASLONG k, double alpha_r, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset);
  452. int zherk_kernel_UC(BLASLONG m, BLASLONG n, BLASLONG k, double alpha_r, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset);
  453. int zherk_kernel_LN(BLASLONG m, BLASLONG n, BLASLONG k, double alpha_r, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset);
  454. int zherk_kernel_LC(BLASLONG m, BLASLONG n, BLASLONG k, double alpha_r, double *a, double *b, double *c, BLASLONG ldc, BLASLONG offset);
  455. int xherk_kernel_UN(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha_r, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset);
  456. int xherk_kernel_UC(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha_r, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset);
  457. int xherk_kernel_LN(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha_r, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset);
  458. int xherk_kernel_LC(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha_r, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset);
  459. 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);
  460. 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);
  461. 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);
  462. 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);
  463. 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);
  464. 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);
  465. 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);
  466. 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);
  467. 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);
  468. 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);
  469. 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);
  470. 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);
  471. int shgemm_kernel(BLASLONG, BLASLONG, BLASLONG, float, hfloat16 *, hfloat16 *, float *, BLASLONG);
  472. int sbgemm_kernel(BLASLONG, BLASLONG, BLASLONG, float, bfloat16 *, bfloat16 *, float *, BLASLONG);
  473. int sgemm_kernel(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG);
  474. int dgemm_kernel(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG);
  475. #ifdef QUAD_PRECISION
  476. int qgemm_kernel(BLASLONG, BLASLONG, BLASLONG, xidouble *, xidouble *, xidouble *, xdouble *, BLASLONG);
  477. #else
  478. int qgemm_kernel(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  479. #endif
  480. #ifdef SMALL_MATRIX_OPT
  481. int sbgemm_small_matrix_permit(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float beta);
  482. int sbgemm_small_kernel_nn(BLASLONG m, BLASLONG n, BLASLONG k, bfloat16 * A, BLASLONG lda, float alpha, bfloat16 * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
  483. int sbgemm_small_kernel_nt(BLASLONG m, BLASLONG n, BLASLONG k, bfloat16 * A, BLASLONG lda, float alpha, bfloat16 * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
  484. int sbgemm_small_kernel_tn(BLASLONG m, BLASLONG n, BLASLONG k, bfloat16 * A, BLASLONG lda, float alpha, bfloat16 * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
  485. int sbgemm_small_kernel_tt(BLASLONG m, BLASLONG n, BLASLONG k, bfloat16 * A, BLASLONG lda, float alpha, bfloat16 * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
  486. int sgemm_small_matrix_permit(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float beta);
  487. 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);
  488. 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);
  489. 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);
  490. 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);
  491. int dgemm_small_matrix_permit(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, double alpha, double beta);
  492. 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);
  493. 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);
  494. 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);
  495. 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);
  496. int sbgemm_small_kernel_b0_nn(BLASLONG m, BLASLONG n, BLASLONG k, bfloat16 * A, BLASLONG lda, float alpha, bfloat16 * B, BLASLONG ldb, float * C, BLASLONG ldc);
  497. int sbgemm_small_kernel_b0_nt(BLASLONG m, BLASLONG n, BLASLONG k, bfloat16 * A, BLASLONG lda, float alpha, bfloat16 * B, BLASLONG ldb, float * C, BLASLONG ldc);
  498. int sbgemm_small_kernel_b0_tn(BLASLONG m, BLASLONG n, BLASLONG k, bfloat16 * A, BLASLONG lda, float alpha, bfloat16 * B, BLASLONG ldb, float * C, BLASLONG ldc);
  499. int sbgemm_small_kernel_b0_tt(BLASLONG m, BLASLONG n, BLASLONG k, bfloat16 * A, BLASLONG lda, float alpha, bfloat16 * B, BLASLONG ldb, float * C, BLASLONG ldc);
  500. 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);
  501. 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);
  502. 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);
  503. 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);
  504. 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);
  505. 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);
  506. 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);
  507. 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);
  508. int cgemm_small_matrix_permit(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, float alpha0, float alpha1, float beta0, float beta1);
  509. 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);
  510. 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);
  511. 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);
  512. 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);
  513. 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);
  514. 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);
  515. 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);
  516. 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);
  517. 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);
  518. 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);
  519. 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);
  520. 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);
  521. 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);
  522. 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);
  523. 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);
  524. 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);
  525. int zgemm_small_matrix_permit(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, double alpha0, double alpha1, double beta0, double beta1);
  526. 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);
  527. 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);
  528. 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);
  529. 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);
  530. 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);
  531. 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);
  532. 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);
  533. 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);
  534. 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);
  535. 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);
  536. 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);
  537. 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);
  538. 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);
  539. 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);
  540. 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);
  541. 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);
  542. 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);
  543. 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);
  544. 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);
  545. 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);
  546. 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);
  547. 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);
  548. 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);
  549. 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);
  550. 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);
  551. 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);
  552. 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);
  553. 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);
  554. 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);
  555. 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);
  556. 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);
  557. 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);
  558. 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);
  559. 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);
  560. 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);
  561. 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);
  562. 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);
  563. 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);
  564. 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);
  565. 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);
  566. 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);
  567. 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);
  568. 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);
  569. 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);
  570. 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);
  571. 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);
  572. 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);
  573. 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);
  574. #endif
  575. int cgemm_kernel_n(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  576. int cgemm_kernel_l(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  577. int cgemm_kernel_r(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  578. int cgemm_kernel_b(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  579. int zgemm_kernel_n(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  580. int zgemm_kernel_l(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  581. int zgemm_kernel_r(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  582. int zgemm_kernel_b(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  583. int xgemm_kernel_n(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  584. int xgemm_kernel_l(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  585. int xgemm_kernel_r(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  586. int xgemm_kernel_b(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  587. int cgemm3m_kernel(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  588. int zgemm3m_kernel(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  589. int xgemm3m_kernel(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  590. int shgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, hfloat16 *, hfloat16 *, BLASLONG);
  591. int shgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, hfloat16 *, hfloat16 *, BLASLONG);
  592. int shgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, hfloat16 *, hfloat16 *, BLASLONG);
  593. int shgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, hfloat16 *, hfloat16 *, BLASLONG);
  594. int sbgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  595. int sbgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  596. int sbgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  597. int sbgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  598. int sgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  599. int sgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  600. int sgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  601. int sgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  602. int dgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  603. int dgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  604. int dgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  605. int dgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  606. #ifdef QUAD_PRECISION
  607. int qgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  608. int qgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  609. int qgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  610. int qgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  611. #else
  612. int qgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  613. int qgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  614. int qgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  615. int qgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  616. #endif
  617. int cgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  618. int cgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  619. int cgemm_nr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  620. int cgemm_nc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  621. int cgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  622. int cgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  623. int cgemm_tr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  624. int cgemm_tc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  625. int cgemm_rn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  626. int cgemm_rt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  627. int cgemm_rr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  628. int cgemm_rc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  629. int cgemm_cn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  630. int cgemm_ct(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  631. int cgemm_cr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  632. int cgemm_cc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  633. int zgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  634. int zgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  635. int zgemm_nr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  636. int zgemm_nc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  637. int zgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  638. int zgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  639. int zgemm_tr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  640. int zgemm_tc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  641. int zgemm_rn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  642. int zgemm_rt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  643. int zgemm_rr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  644. int zgemm_rc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  645. int zgemm_cn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  646. int zgemm_ct(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  647. int zgemm_cr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  648. int zgemm_cc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  649. #ifdef QUAD_PRECISION
  650. int xgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  651. int xgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  652. int xgemm_nr(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  653. int xgemm_nc(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  654. int xgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  655. int xgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  656. int xgemm_tr(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  657. int xgemm_tc(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  658. int xgemm_rn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  659. int xgemm_rt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  660. int xgemm_rr(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  661. int xgemm_rc(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  662. int xgemm_cn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  663. int xgemm_ct(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  664. int xgemm_cr(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  665. int xgemm_cc(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  666. #else
  667. int xgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  668. int xgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  669. int xgemm_nr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  670. int xgemm_nc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  671. int xgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  672. int xgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  673. int xgemm_tr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  674. int xgemm_tc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  675. int xgemm_rn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  676. int xgemm_rt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  677. int xgemm_rr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  678. int xgemm_rc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  679. int xgemm_cn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  680. int xgemm_ct(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  681. int xgemm_cr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  682. int xgemm_cc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  683. #endif
  684. int shgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, hfloat16 *, hfloat16 *, BLASLONG);
  685. int shgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, hfloat16 *, hfloat16 *, BLASLONG);
  686. int shgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, hfloat16 *, hfloat16 *, BLASLONG);
  687. int shgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, hfloat16 *, hfloat16 *, BLASLONG);
  688. int sbgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  689. int sbgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  690. int sbgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  691. int sbgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
  692. int sgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  693. int sgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  694. int sgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  695. int sgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  696. int dgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  697. int dgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  698. int dgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  699. int dgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  700. #ifdef QUAD_PRECISION
  701. int qgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  702. int qgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  703. int qgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  704. int qgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xidouble *, xidouble *, BLASLONG);
  705. #else
  706. int qgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  707. int qgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  708. int qgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  709. int qgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  710. #endif
  711. int cgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  712. int cgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  713. int cgemm_thread_nr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  714. int cgemm_thread_nc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  715. int cgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  716. int cgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  717. int cgemm_thread_tr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  718. int cgemm_thread_tc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  719. int cgemm_thread_rn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  720. int cgemm_thread_rt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  721. int cgemm_thread_rr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  722. int cgemm_thread_rc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  723. int cgemm_thread_cn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  724. int cgemm_thread_ct(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  725. int cgemm_thread_cr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  726. int cgemm_thread_cc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  727. int zgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  728. int zgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  729. int zgemm_thread_nr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  730. int zgemm_thread_nc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  731. int zgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  732. int zgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  733. int zgemm_thread_tr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  734. int zgemm_thread_tc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  735. int zgemm_thread_rn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  736. int zgemm_thread_rt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  737. int zgemm_thread_rr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  738. int zgemm_thread_rc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  739. int zgemm_thread_cn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  740. int zgemm_thread_ct(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  741. int zgemm_thread_cr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  742. int zgemm_thread_cc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  743. int xgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  744. int xgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  745. int xgemm_thread_nr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  746. int xgemm_thread_nc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  747. int xgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  748. int xgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  749. int xgemm_thread_tr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  750. int xgemm_thread_tc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  751. int xgemm_thread_rn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  752. int xgemm_thread_rt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  753. int xgemm_thread_rr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  754. int xgemm_thread_rc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  755. int xgemm_thread_cn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  756. int xgemm_thread_ct(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  757. int xgemm_thread_cr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  758. int xgemm_thread_cc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  759. int cgemm3m_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  760. int cgemm3m_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  761. int cgemm3m_nr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  762. int cgemm3m_nc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  763. int cgemm3m_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  764. int cgemm3m_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  765. int cgemm3m_tr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  766. int cgemm3m_tc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  767. int cgemm3m_rn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  768. int cgemm3m_rt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  769. int cgemm3m_rr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  770. int cgemm3m_rc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  771. int cgemm3m_cn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  772. int cgemm3m_ct(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  773. int cgemm3m_cr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  774. int cgemm3m_cc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  775. int zgemm3m_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  776. int zgemm3m_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  777. int zgemm3m_nr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  778. int zgemm3m_nc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  779. int zgemm3m_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  780. int zgemm3m_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  781. int zgemm3m_tr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  782. int zgemm3m_tc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  783. int zgemm3m_rn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  784. int zgemm3m_rt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  785. int zgemm3m_rr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  786. int zgemm3m_rc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  787. int zgemm3m_cn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  788. int zgemm3m_ct(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  789. int zgemm3m_cr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  790. int zgemm3m_cc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  791. int xgemm3m_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  792. int xgemm3m_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  793. int xgemm3m_nr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  794. int xgemm3m_nc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  795. int xgemm3m_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  796. int xgemm3m_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  797. int xgemm3m_tr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  798. int xgemm3m_tc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  799. int xgemm3m_rn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  800. int xgemm3m_rt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  801. int xgemm3m_rr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  802. int xgemm3m_rc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  803. int xgemm3m_cn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  804. int xgemm3m_ct(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  805. int xgemm3m_cr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  806. int xgemm3m_cc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  807. int cgemm3m_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  808. int cgemm3m_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  809. int cgemm3m_thread_nr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  810. int cgemm3m_thread_nc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  811. int cgemm3m_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  812. int cgemm3m_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  813. int cgemm3m_thread_tr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  814. int cgemm3m_thread_tc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  815. int cgemm3m_thread_rn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  816. int cgemm3m_thread_rt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  817. int cgemm3m_thread_rr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  818. int cgemm3m_thread_rc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  819. int cgemm3m_thread_cn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  820. int cgemm3m_thread_ct(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  821. int cgemm3m_thread_cr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  822. int cgemm3m_thread_cc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  823. int zgemm3m_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  824. int zgemm3m_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  825. int zgemm3m_thread_nr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  826. int zgemm3m_thread_nc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  827. int zgemm3m_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  828. int zgemm3m_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  829. int zgemm3m_thread_tr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  830. int zgemm3m_thread_tc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  831. int zgemm3m_thread_rn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  832. int zgemm3m_thread_rt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  833. int zgemm3m_thread_rr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  834. int zgemm3m_thread_rc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  835. int zgemm3m_thread_cn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  836. int zgemm3m_thread_ct(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  837. int zgemm3m_thread_cr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  838. int zgemm3m_thread_cc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  839. int xgemm3m_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  840. int xgemm3m_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  841. int xgemm3m_thread_nr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  842. int xgemm3m_thread_nc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  843. int xgemm3m_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  844. int xgemm3m_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  845. int xgemm3m_thread_tr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  846. int xgemm3m_thread_tc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  847. int xgemm3m_thread_rn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  848. int xgemm3m_thread_rt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  849. int xgemm3m_thread_rr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  850. int xgemm3m_thread_rc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  851. int xgemm3m_thread_cn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  852. int xgemm3m_thread_ct(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  853. int xgemm3m_thread_cr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  854. int xgemm3m_thread_cc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  855. int cher2m_LNN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  856. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  857. int cher2m_LNT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  858. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  859. int cher2m_LNR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  860. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  861. int cher2m_LNC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  862. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  863. int cher2m_LTN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  864. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  865. int cher2m_LTT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  866. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  867. int cher2m_LTR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  868. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  869. int cher2m_LTC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  870. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  871. int cher2m_LRN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  872. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  873. int cher2m_LRT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  874. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  875. int cher2m_LRR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  876. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  877. int cher2m_LRC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  878. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  879. int cher2m_LCN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  880. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  881. int cher2m_LCT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  882. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  883. int cher2m_LCR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  884. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  885. int cher2m_LCC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  886. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  887. int cher2m_UNN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  888. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  889. int cher2m_UNT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  890. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  891. int cher2m_UNR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  892. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  893. int cher2m_UNC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  894. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  895. int cher2m_UTN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  896. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  897. int cher2m_UTT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  898. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  899. int cher2m_UTR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  900. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  901. int cher2m_UTC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  902. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  903. int cher2m_URN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  904. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  905. int cher2m_URT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  906. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  907. int cher2m_URR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  908. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  909. int cher2m_URC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  910. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  911. int cher2m_UCN(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  912. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  913. int cher2m_UCT(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  914. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  915. int cher2m_UCR(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  916. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  917. int cher2m_UCC(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG,
  918. float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, BLASLONG);
  919. int zher2m_LNN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  920. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  921. int zher2m_LNT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  922. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  923. int zher2m_LNR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  924. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  925. int zher2m_LNC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  926. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  927. int zher2m_LTN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  928. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  929. int zher2m_LTT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  930. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  931. int zher2m_LTR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  932. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  933. int zher2m_LTC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  934. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  935. int zher2m_LRN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  936. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  937. int zher2m_LRT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  938. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  939. int zher2m_LRR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  940. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  941. int zher2m_LRC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  942. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  943. int zher2m_LCN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  944. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  945. int zher2m_LCT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  946. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  947. int zher2m_LCR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  948. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  949. int zher2m_LCC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  950. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  951. int zher2m_UNN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  952. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  953. int zher2m_UNT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  954. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  955. int zher2m_UNR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  956. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  957. int zher2m_UNC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  958. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  959. int zher2m_UTN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  960. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  961. int zher2m_UTT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  962. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  963. int zher2m_UTR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  964. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  965. int zher2m_UTC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  966. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  967. int zher2m_URN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  968. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  969. int zher2m_URT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  970. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  971. int zher2m_URR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  972. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  973. int zher2m_URC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  974. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  975. int zher2m_UCN(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  976. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  977. int zher2m_UCT(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  978. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  979. int zher2m_UCR(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  980. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  981. int zher2m_UCC(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG,
  982. double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, BLASLONG);
  983. int strsm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  984. int strsm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  985. int strsm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  986. int strsm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  987. int strsm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  988. int strsm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  989. int strsm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  990. int strsm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  991. int strsm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  992. int strsm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  993. int strsm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  994. int strsm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  995. int strsm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  996. int strsm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  997. int strsm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  998. int strsm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  999. int dtrsm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1000. int dtrsm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1001. int dtrsm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1002. int dtrsm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1003. int dtrsm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1004. int dtrsm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1005. int dtrsm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1006. int dtrsm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1007. int dtrsm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1008. int dtrsm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1009. int dtrsm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1010. int dtrsm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1011. int dtrsm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1012. int dtrsm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1013. int dtrsm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1014. int dtrsm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1015. int qtrsm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1016. int qtrsm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1017. int qtrsm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1018. int qtrsm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1019. int qtrsm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1020. int qtrsm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1021. int qtrsm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1022. int qtrsm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1023. int qtrsm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1024. int qtrsm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1025. int qtrsm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1026. int qtrsm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1027. int qtrsm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1028. int qtrsm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1029. int qtrsm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1030. int qtrsm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1031. int ctrsm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1032. int ctrsm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1033. int ctrsm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1034. int ctrsm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1035. int ctrsm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1036. int ctrsm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1037. int ctrsm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1038. int ctrsm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1039. int ctrsm_LRUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1040. int ctrsm_LRUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1041. int ctrsm_LRLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1042. int ctrsm_LRLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1043. int ctrsm_LCUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1044. int ctrsm_LCUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1045. int ctrsm_LCLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1046. int ctrsm_LCLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1047. int ctrsm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1048. int ctrsm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1049. int ctrsm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1050. int ctrsm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1051. int ctrsm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1052. int ctrsm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1053. int ctrsm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1054. int ctrsm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1055. int ctrsm_RRUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1056. int ctrsm_RRUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1057. int ctrsm_RRLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1058. int ctrsm_RRLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1059. int ctrsm_RCUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1060. int ctrsm_RCUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1061. int ctrsm_RCLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1062. int ctrsm_RCLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1063. int ztrsm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1064. int ztrsm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1065. int ztrsm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1066. int ztrsm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1067. int ztrsm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1068. int ztrsm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1069. int ztrsm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1070. int ztrsm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1071. int ztrsm_LRUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1072. int ztrsm_LRUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1073. int ztrsm_LRLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1074. int ztrsm_LRLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1075. int ztrsm_LCUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1076. int ztrsm_LCUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1077. int ztrsm_LCLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1078. int ztrsm_LCLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1079. int ztrsm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1080. int ztrsm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1081. int ztrsm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1082. int ztrsm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1083. int ztrsm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1084. int ztrsm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1085. int ztrsm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1086. int ztrsm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1087. int ztrsm_RRUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1088. int ztrsm_RRUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1089. int ztrsm_RRLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1090. int ztrsm_RRLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1091. int ztrsm_RCUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1092. int ztrsm_RCUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1093. int ztrsm_RCLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1094. int ztrsm_RCLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1095. int xtrsm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1096. int xtrsm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1097. int xtrsm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1098. int xtrsm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1099. int xtrsm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1100. int xtrsm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1101. int xtrsm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1102. int xtrsm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1103. int xtrsm_LRUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1104. int xtrsm_LRUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1105. int xtrsm_LRLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1106. int xtrsm_LRLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1107. int xtrsm_LCUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1108. int xtrsm_LCUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1109. int xtrsm_LCLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1110. int xtrsm_LCLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1111. int xtrsm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1112. int xtrsm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1113. int xtrsm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1114. int xtrsm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1115. int xtrsm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1116. int xtrsm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1117. int xtrsm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1118. int xtrsm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1119. int xtrsm_RRUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1120. int xtrsm_RRUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1121. int xtrsm_RRLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1122. int xtrsm_RRLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1123. int xtrsm_RCUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1124. int xtrsm_RCUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1125. int xtrsm_RCLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1126. int xtrsm_RCLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1127. int strmm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1128. int strmm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1129. int strmm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1130. int strmm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1131. int strmm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1132. int strmm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1133. int strmm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1134. int strmm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1135. int strmm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1136. int strmm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1137. int strmm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1138. int strmm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1139. int strmm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1140. int strmm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1141. int strmm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1142. int strmm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1143. int dtrmm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1144. int dtrmm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1145. int dtrmm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1146. int dtrmm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1147. int dtrmm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1148. int dtrmm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1149. int dtrmm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1150. int dtrmm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1151. int dtrmm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1152. int dtrmm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1153. int dtrmm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1154. int dtrmm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1155. int dtrmm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1156. int dtrmm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1157. int dtrmm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1158. int dtrmm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1159. int qtrmm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1160. int qtrmm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1161. int qtrmm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1162. int qtrmm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1163. int qtrmm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1164. int qtrmm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1165. int qtrmm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1166. int qtrmm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1167. int qtrmm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1168. int qtrmm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1169. int qtrmm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1170. int qtrmm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1171. int qtrmm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1172. int qtrmm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1173. int qtrmm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1174. int qtrmm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1175. int ctrmm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1176. int ctrmm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1177. int ctrmm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1178. int ctrmm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1179. int ctrmm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1180. int ctrmm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1181. int ctrmm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1182. int ctrmm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1183. int ctrmm_LRUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1184. int ctrmm_LRUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1185. int ctrmm_LRLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1186. int ctrmm_LRLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1187. int ctrmm_LCUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1188. int ctrmm_LCUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1189. int ctrmm_LCLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1190. int ctrmm_LCLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1191. int ctrmm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1192. int ctrmm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1193. int ctrmm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1194. int ctrmm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1195. int ctrmm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1196. int ctrmm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1197. int ctrmm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1198. int ctrmm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1199. int ctrmm_RRUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1200. int ctrmm_RRUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1201. int ctrmm_RRLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1202. int ctrmm_RRLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1203. int ctrmm_RCUU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1204. int ctrmm_RCUN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1205. int ctrmm_RCLU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1206. int ctrmm_RCLN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1207. int ztrmm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1208. int ztrmm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1209. int ztrmm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1210. int ztrmm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1211. int ztrmm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1212. int ztrmm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1213. int ztrmm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1214. int ztrmm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1215. int ztrmm_LRUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1216. int ztrmm_LRUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1217. int ztrmm_LRLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1218. int ztrmm_LRLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1219. int ztrmm_LCUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1220. int ztrmm_LCUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1221. int ztrmm_LCLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1222. int ztrmm_LCLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1223. int ztrmm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1224. int ztrmm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1225. int ztrmm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1226. int ztrmm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1227. int ztrmm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1228. int ztrmm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1229. int ztrmm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1230. int ztrmm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1231. int ztrmm_RRUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1232. int ztrmm_RRUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1233. int ztrmm_RRLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1234. int ztrmm_RRLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1235. int ztrmm_RCUU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1236. int ztrmm_RCUN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1237. int ztrmm_RCLU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1238. int ztrmm_RCLN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1239. int xtrmm_LNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1240. int xtrmm_LNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1241. int xtrmm_LNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1242. int xtrmm_LNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1243. int xtrmm_LTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1244. int xtrmm_LTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1245. int xtrmm_LTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1246. int xtrmm_LTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1247. int xtrmm_LRUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1248. int xtrmm_LRUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1249. int xtrmm_LRLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1250. int xtrmm_LRLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1251. int xtrmm_LCUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1252. int xtrmm_LCUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1253. int xtrmm_LCLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1254. int xtrmm_LCLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1255. int xtrmm_RNUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1256. int xtrmm_RNUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1257. int xtrmm_RNLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1258. int xtrmm_RNLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1259. int xtrmm_RTUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1260. int xtrmm_RTUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1261. int xtrmm_RTLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1262. int xtrmm_RTLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1263. int xtrmm_RRUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1264. int xtrmm_RRUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1265. int xtrmm_RRLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1266. int xtrmm_RRLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1267. int xtrmm_RCUU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1268. int xtrmm_RCUN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1269. int xtrmm_RCLU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1270. int xtrmm_RCLN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1271. int ssymm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1272. int ssymm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1273. int ssymm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1274. int ssymm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1275. int dsymm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1276. int dsymm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1277. int dsymm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1278. int dsymm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1279. int qsymm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1280. int qsymm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1281. int qsymm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1282. int qsymm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1283. int csymm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1284. int csymm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1285. int csymm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1286. int csymm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1287. int zsymm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1288. int zsymm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1289. int zsymm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1290. int zsymm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1291. int xsymm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1292. int xsymm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1293. int xsymm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1294. int xsymm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1295. int csymm3m_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1296. int csymm3m_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1297. int csymm3m_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1298. int csymm3m_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1299. int zsymm3m_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1300. int zsymm3m_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1301. int zsymm3m_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1302. int zsymm3m_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1303. int xsymm3m_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1304. int xsymm3m_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1305. int xsymm3m_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1306. int xsymm3m_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1307. int csymm3m_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1308. int csymm3m_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1309. int csymm3m_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1310. int csymm3m_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1311. int zsymm3m_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1312. int zsymm3m_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1313. int zsymm3m_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1314. int zsymm3m_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1315. int xsymm3m_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1316. int xsymm3m_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1317. int xsymm3m_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1318. int xsymm3m_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1319. int chemm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1320. int chemm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1321. int chemm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1322. int chemm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1323. int zhemm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1324. int zhemm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1325. int zhemm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1326. int zhemm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1327. int xhemm_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1328. int xhemm_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1329. int xhemm_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1330. int xhemm_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1331. int chemm3m_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1332. int chemm3m_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1333. int chemm3m_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1334. int chemm3m_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1335. int zhemm3m_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1336. int zhemm3m_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1337. int zhemm3m_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1338. int zhemm3m_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1339. int xhemm3m_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1340. int xhemm3m_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1341. int xhemm3m_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1342. int xhemm3m_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1343. int chemm3m_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1344. int chemm3m_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1345. int chemm3m_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1346. int chemm3m_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1347. int zhemm3m_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1348. int zhemm3m_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1349. int zhemm3m_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1350. int zhemm3m_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1351. int xhemm3m_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1352. int xhemm3m_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1353. int xhemm3m_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1354. int xhemm3m_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1355. int ssymm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1356. int ssymm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1357. int ssymm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1358. int ssymm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1359. int dsymm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1360. int dsymm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1361. int dsymm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1362. int dsymm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1363. int qsymm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1364. int qsymm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1365. int qsymm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1366. int qsymm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1367. int csymm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1368. int csymm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1369. int csymm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1370. int csymm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1371. int zsymm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1372. int zsymm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1373. int zsymm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1374. int zsymm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1375. int xsymm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1376. int xsymm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1377. int xsymm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1378. int xsymm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1379. int chemm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1380. int chemm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1381. int chemm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1382. int chemm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1383. int zhemm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1384. int zhemm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1385. int zhemm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1386. int zhemm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1387. int xhemm_thread_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1388. int xhemm_thread_LL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1389. int xhemm_thread_RU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1390. int xhemm_thread_RL(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1391. int ssyrk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1392. int ssyrk_UT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1393. int ssyrk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1394. int ssyrk_LT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1395. int dsyrk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1396. int dsyrk_UT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1397. int dsyrk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1398. int dsyrk_LT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1399. int qsyrk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1400. int qsyrk_UT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1401. int qsyrk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1402. int qsyrk_LT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1403. int csyrk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1404. int csyrk_UT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1405. int csyrk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1406. int csyrk_LT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1407. int zsyrk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1408. int zsyrk_UT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1409. int zsyrk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1410. int zsyrk_LT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1411. int xsyrk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1412. int xsyrk_UT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1413. int xsyrk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1414. int xsyrk_LT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1415. int ssyrk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1416. int ssyrk_thread_UT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1417. int ssyrk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1418. int ssyrk_thread_LT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1419. int dsyrk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1420. int dsyrk_thread_UT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1421. int dsyrk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1422. int dsyrk_thread_LT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1423. int qsyrk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1424. int qsyrk_thread_UT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1425. int qsyrk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1426. int qsyrk_thread_LT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1427. int csyrk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1428. int csyrk_thread_UT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1429. int csyrk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1430. int csyrk_thread_LT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1431. int zsyrk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1432. int zsyrk_thread_UT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1433. int zsyrk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1434. int zsyrk_thread_LT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1435. int xsyrk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1436. int xsyrk_thread_UT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1437. int xsyrk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1438. int xsyrk_thread_LT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1439. int ssyr2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1440. int ssyr2k_UT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1441. int ssyr2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1442. int ssyr2k_LT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1443. int dsyr2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1444. int dsyr2k_UT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1445. int dsyr2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1446. int dsyr2k_LT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1447. int qsyr2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1448. int qsyr2k_UT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1449. int qsyr2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1450. int qsyr2k_LT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1451. int csyr2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1452. int csyr2k_UT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1453. int csyr2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1454. int csyr2k_LT(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1455. int zsyr2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1456. int zsyr2k_UT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1457. int zsyr2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1458. int zsyr2k_LT(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1459. int xsyr2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1460. int xsyr2k_UT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1461. int xsyr2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1462. int xsyr2k_LT(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1463. int cherk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1464. int cherk_UC(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1465. int cherk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1466. int cherk_LC(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1467. int zherk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1468. int zherk_UC(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1469. int zherk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1470. int zherk_LC(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1471. int xherk_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1472. int xherk_UC(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1473. int xherk_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1474. int xherk_LC(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1475. int cherk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1476. int cherk_thread_UC(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1477. int cherk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1478. int cherk_thread_LC(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1479. int zherk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1480. int zherk_thread_UC(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1481. int zherk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1482. int zherk_thread_LC(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1483. int xherk_thread_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1484. int xherk_thread_UC(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1485. int xherk_thread_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1486. int xherk_thread_LC(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1487. int cher2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1488. int cher2k_UC(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1489. int cher2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1490. int cher2k_LC(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1491. int zher2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1492. int zher2k_UC(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1493. int zher2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1494. int zher2k_LC(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1495. int xher2k_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1496. int xher2k_UC(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1497. int xher2k_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1498. int xher2k_LC(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1499. int sgemt_n(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, int);
  1500. int sgemt_t(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, int);
  1501. int dgemt_n(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, int);
  1502. int dgemt_t(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, int);
  1503. int cgemt_n(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, int);
  1504. int cgemt_t(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, int);
  1505. int cgemt_r(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, int);
  1506. int cgemt_c(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, int);
  1507. int zgemt_n(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, int);
  1508. int zgemt_t(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, int);
  1509. int zgemt_r(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, int);
  1510. int zgemt_c(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, int);
  1511. int sgema_n(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG);
  1512. int sgema_t(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG);
  1513. int dgema_n(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG);
  1514. int dgema_t(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG);
  1515. int cgema_n(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1516. int cgema_t(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1517. int cgema_r(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1518. int cgema_c(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1519. int zgema_n(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1520. int zgema_t(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1521. int zgema_r(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1522. int zgema_c(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1523. int cgemm3m_incopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  1524. int cgemm3m_incopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  1525. int cgemm3m_incopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  1526. int cgemm3m_itcopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  1527. int cgemm3m_itcopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  1528. int cgemm3m_itcopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
  1529. int cgemm3m_oncopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float alpha_r, float alpha_i, float *b);
  1530. int cgemm3m_oncopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float alpha_r, float alpha_i, float *b);
  1531. int cgemm3m_oncopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float alpha_r, float alpha_i, float *b);
  1532. int cgemm3m_otcopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float alpha_r, float alpha_i, float *b);
  1533. int cgemm3m_otcopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float alpha_r, float alpha_i, float *b);
  1534. int cgemm3m_otcopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float alpha_r, float alpha_i, float *b);
  1535. int zgemm3m_incopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  1536. int zgemm3m_incopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  1537. int zgemm3m_incopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  1538. int zgemm3m_itcopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  1539. int zgemm3m_itcopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  1540. int zgemm3m_itcopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b);
  1541. int zgemm3m_oncopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double alpha_r, double alpha_i, double *b);
  1542. int zgemm3m_oncopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double alpha_r, double alpha_i, double *b);
  1543. int zgemm3m_oncopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double alpha_r, double alpha_i, double *b);
  1544. int zgemm3m_otcopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double alpha_r, double alpha_i, double *b);
  1545. int zgemm3m_otcopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double alpha_r, double alpha_i, double *b);
  1546. int zgemm3m_otcopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double alpha_r, double alpha_i, double *b);
  1547. int xgemm3m_incopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  1548. int xgemm3m_incopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  1549. int xgemm3m_incopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  1550. int xgemm3m_itcopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  1551. int xgemm3m_itcopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  1552. int xgemm3m_itcopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b);
  1553. int xgemm3m_oncopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1554. int xgemm3m_oncopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1555. int xgemm3m_oncopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1556. int xgemm3m_otcopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1557. int xgemm3m_otcopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1558. int xgemm3m_otcopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1559. int csymm3m_iucopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1560. int csymm3m_ilcopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1561. int csymm3m_iucopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1562. int csymm3m_ilcopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1563. int csymm3m_iucopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1564. int csymm3m_ilcopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1565. int csymm3m_oucopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1566. int csymm3m_olcopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1567. int csymm3m_oucopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1568. int csymm3m_olcopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1569. int csymm3m_oucopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1570. int csymm3m_olcopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1571. int zsymm3m_iucopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1572. int zsymm3m_ilcopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1573. int zsymm3m_iucopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1574. int zsymm3m_ilcopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1575. int zsymm3m_iucopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1576. int zsymm3m_ilcopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1577. int zsymm3m_oucopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1578. int zsymm3m_olcopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1579. int zsymm3m_oucopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1580. int zsymm3m_olcopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1581. int zsymm3m_oucopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1582. int zsymm3m_olcopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1583. int xsymm3m_iucopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1584. int xsymm3m_ilcopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1585. int xsymm3m_iucopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1586. int xsymm3m_ilcopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1587. int xsymm3m_iucopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1588. int xsymm3m_ilcopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1589. int xsymm3m_oucopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1590. int xsymm3m_olcopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1591. int xsymm3m_oucopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1592. int xsymm3m_olcopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1593. int xsymm3m_oucopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1594. int xsymm3m_olcopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1595. int chemm3m_iucopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1596. int chemm3m_ilcopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1597. int chemm3m_iucopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1598. int chemm3m_ilcopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1599. int chemm3m_iucopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1600. int chemm3m_ilcopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float *b);
  1601. int chemm3m_oucopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1602. int chemm3m_olcopyb(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1603. int chemm3m_oucopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1604. int chemm3m_olcopyr(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1605. int chemm3m_oucopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1606. int chemm3m_olcopyi(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG x, BLASLONG y, float alpha_r, float alpha_i, float *b);
  1607. int zhemm3m_iucopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1608. int zhemm3m_ilcopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1609. int zhemm3m_iucopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1610. int zhemm3m_ilcopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1611. int zhemm3m_iucopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1612. int zhemm3m_ilcopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double *b);
  1613. int zhemm3m_oucopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1614. int zhemm3m_olcopyb(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1615. int zhemm3m_oucopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1616. int zhemm3m_olcopyr(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1617. int zhemm3m_oucopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1618. int zhemm3m_olcopyi(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG x, BLASLONG y, double alpha_r, double alpha_i, double *b);
  1619. int xhemm3m_iucopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1620. int xhemm3m_ilcopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1621. int xhemm3m_iucopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1622. int xhemm3m_ilcopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1623. int xhemm3m_iucopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1624. int xhemm3m_ilcopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble *b);
  1625. int xhemm3m_oucopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1626. int xhemm3m_olcopyb(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1627. int xhemm3m_oucopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1628. int xhemm3m_olcopyr(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1629. int xhemm3m_oucopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1630. int xhemm3m_olcopyi(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, BLASLONG x, BLASLONG y, xdouble alpha_r, xdouble alpha_i, xdouble *b);
  1631. int sgemc_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1632. int sgemc_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1633. int sgemc_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1634. int sgemc_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1635. int dgemc_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1636. int dgemc_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1637. int dgemc_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1638. int dgemc_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1639. int qgemc_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1640. int qgemc_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1641. int qgemc_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1642. int qgemc_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1643. int cgemc_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1644. int cgemc_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1645. int cgemc_nr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1646. int cgemc_nc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1647. int cgemc_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1648. int cgemc_tt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1649. int cgemc_tr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1650. int cgemc_tc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1651. int cgemc_rn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1652. int cgemc_rt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1653. int cgemc_rr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1654. int cgemc_rc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1655. int cgemc_cn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1656. int cgemc_ct(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1657. int cgemc_cr(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1658. int cgemc_cc(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
  1659. int zgemc_nn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1660. int zgemc_nt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1661. int zgemc_nr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1662. int zgemc_nc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1663. int zgemc_tn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1664. int zgemc_tt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1665. int zgemc_tr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1666. int zgemc_tc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1667. int zgemc_rn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1668. int zgemc_rt(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1669. int zgemc_rr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1670. int zgemc_rc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1671. int zgemc_cn(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1672. int zgemc_ct(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1673. int zgemc_cr(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1674. int zgemc_cc(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
  1675. int xgemc_nn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1676. int xgemc_nt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1677. int xgemc_nr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1678. int xgemc_nc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1679. int xgemc_tn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1680. int xgemc_tt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1681. int xgemc_tr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1682. int xgemc_tc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1683. int xgemc_rn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1684. int xgemc_rt(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1685. int xgemc_rr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1686. int xgemc_rc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1687. int xgemc_cn(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1688. int xgemc_ct(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1689. int xgemc_cr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1690. int xgemc_cc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
  1691. int sgemc_oncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b, BLASLONG ldb, float *c);
  1692. int sgemc_otcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b, BLASLONG ldb, float *c);
  1693. int dgemc_oncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b, BLASLONG ldb, double *c);
  1694. int dgemc_otcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b, BLASLONG ldb, double *c);
  1695. int qgemc_oncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b, BLASLONG ldb, xdouble *c);
  1696. int qgemc_otcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b, BLASLONG ldb, xdouble *c);
  1697. int cgemc_oncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b, BLASLONG ldb, float *c);
  1698. int cgemc_otcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b, BLASLONG ldb, float *c);
  1699. int zgemc_oncopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b, BLASLONG ldb, double *c);
  1700. int zgemc_otcopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, double *b, BLASLONG ldb, double *c);
  1701. int xgemc_oncopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b, BLASLONG ldb, xdouble *c);
  1702. int xgemc_otcopy(BLASLONG m, BLASLONG n, xdouble *a, BLASLONG lda, xdouble *b, BLASLONG ldb, xdouble *c);
  1703. int somatcopy_k_cn(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG);
  1704. int somatcopy_k_rn(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG);
  1705. int somatcopy_k_ct(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG);
  1706. int somatcopy_k_rt(BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG);
  1707. int simatcopy_k_cn(BLASLONG, BLASLONG, float, float *, BLASLONG);
  1708. int simatcopy_k_rn(BLASLONG, BLASLONG, float, float *, BLASLONG);
  1709. int simatcopy_k_ct(BLASLONG, BLASLONG, float, float *, BLASLONG);
  1710. int simatcopy_k_rt(BLASLONG, BLASLONG, float, float *, BLASLONG);
  1711. int domatcopy_k_cn(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG);
  1712. int domatcopy_k_rn(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG);
  1713. int domatcopy_k_ct(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG);
  1714. int domatcopy_k_rt(BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG);
  1715. int dimatcopy_k_cn(BLASLONG, BLASLONG, double, double *, BLASLONG);
  1716. int dimatcopy_k_rn(BLASLONG, BLASLONG, double, double *, BLASLONG);
  1717. int dimatcopy_k_ct(BLASLONG, BLASLONG, double, double *, BLASLONG);
  1718. int dimatcopy_k_rt(BLASLONG, BLASLONG, double, double *, BLASLONG);
  1719. int comatcopy_k_cn(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1720. int comatcopy_k_rn(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1721. int comatcopy_k_ct(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1722. int comatcopy_k_rt(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1723. int cimatcopy_k_cn(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1724. int cimatcopy_k_rn(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1725. int cimatcopy_k_ct(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1726. int cimatcopy_k_rt(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1727. int comatcopy_k_cnc(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1728. int comatcopy_k_rnc(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1729. int comatcopy_k_ctc(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1730. int comatcopy_k_rtc(BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG);
  1731. int cimatcopy_k_cnc(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1732. int cimatcopy_k_rnc(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1733. int cimatcopy_k_ctc(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1734. int cimatcopy_k_rtc(BLASLONG, BLASLONG, float, float, float *, BLASLONG);
  1735. int zomatcopy_k_cn(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1736. int zomatcopy_k_rn(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1737. int zomatcopy_k_ct(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1738. int zomatcopy_k_rt(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1739. int zimatcopy_k_cn(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1740. int zimatcopy_k_rn(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1741. int zimatcopy_k_ct(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1742. int zimatcopy_k_rt(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1743. int zomatcopy_k_cnc(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1744. int zomatcopy_k_rnc(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1745. int zomatcopy_k_ctc(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1746. int zomatcopy_k_rtc(BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG);
  1747. int zimatcopy_k_cnc(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1748. int zimatcopy_k_rnc(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1749. int zimatcopy_k_ctc(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1750. int zimatcopy_k_rtc(BLASLONG, BLASLONG, double, double, double *, BLASLONG);
  1751. int sgeadd_k(BLASLONG, BLASLONG, float, float*, BLASLONG, float, float *, BLASLONG);
  1752. int dgeadd_k(BLASLONG, BLASLONG, double, double*, BLASLONG, double, double *, BLASLONG);
  1753. int cgeadd_k(BLASLONG, BLASLONG, float, float, float*, BLASLONG, float, float, float *, BLASLONG);
  1754. int zgeadd_k(BLASLONG, BLASLONG, double,double, double*, BLASLONG, double, double, double *, BLASLONG);
  1755. int sgemm_batch_thread(blas_arg_t * queue, BLASLONG nums);
  1756. int dgemm_batch_thread(blas_arg_t * queue, BLASLONG nums);
  1757. int cgemm_batch_thread(blas_arg_t * queue, BLASLONG nums);
  1758. int zgemm_batch_thread(blas_arg_t * queue, BLASLONG nums);
  1759. int sbgemm_batch_thread(blas_arg_t * queue, BLASLONG nums);
  1760. // int shgemm_batch_thread(blas_arg_t * queue, BLASLONG nums);
  1761. #ifdef __CUDACC__
  1762. }
  1763. #endif
  1764. #endif