You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

common_param.h 112 kB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637
  1. /*********************************************************************/
  2. /* Copyright 2009, 2010 The University of Texas at Austin. */
  3. /* All rights reserved. */
  4. /* */
  5. /* Redistribution and use in source and binary forms, with or */
  6. /* without modification, are permitted provided that the following */
  7. /* conditions are met: */
  8. /* */
  9. /* 1. Redistributions of source code must retain the above */
  10. /* copyright notice, this list of conditions and the following */
  11. /* disclaimer. */
  12. /* */
  13. /* 2. Redistributions in binary form must reproduce the above */
  14. /* copyright notice, this list of conditions and the following */
  15. /* disclaimer in the documentation and/or other materials */
  16. /* provided with the distribution. */
  17. /* */
  18. /* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
  19. /* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
  20. /* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
  21. /* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
  22. /* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
  23. /* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
  24. /* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
  25. /* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
  26. /* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
  27. /* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
  28. /* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
  29. /* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
  30. /* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
  31. /* POSSIBILITY OF SUCH DAMAGE. */
  32. /* */
  33. /* The views and conclusions contained in the software and */
  34. /* documentation are those of the authors and should not be */
  35. /* interpreted as representing official policies, either expressed */
  36. /* or implied, of The University of Texas at Austin. */
  37. /*********************************************************************/
  38. #ifndef COMMON_PARAM_H
  39. #define COMMON_PARAM_H
  40. #ifndef ASSEMBLER
  41. #ifdef DYNAMIC_ARCH
  42. typedef struct {
  43. int dtb_entries;
  44. int offsetA, offsetB, align;
  45. #ifdef BUILD_BFLOAT16
  46. int sbgemm_p, sbgemm_q, sbgemm_r;
  47. int sbgemm_unroll_m, sbgemm_unroll_n, sbgemm_unroll_mn;
  48. void (*sbstobf16_k) (BLASLONG, float *, BLASLONG, bfloat16 *, BLASLONG);
  49. void (*sbdtobf16_k) (BLASLONG, double *, BLASLONG, bfloat16 *, BLASLONG);
  50. void (*sbf16tos_k) (BLASLONG, bfloat16 *, BLASLONG, float *, BLASLONG);
  51. void (*dbf16tod_k) (BLASLONG, bfloat16 *, BLASLONG, double *, BLASLONG);
  52. float (*sbamax_k) (BLASLONG, float *, BLASLONG);
  53. float (*sbamin_k) (BLASLONG, float *, BLASLONG);
  54. float (*sbmax_k) (BLASLONG, float *, BLASLONG);
  55. float (*sbmin_k) (BLASLONG, float *, BLASLONG);
  56. BLASLONG (*isbamax_k)(BLASLONG, float *, BLASLONG);
  57. BLASLONG (*isbamin_k)(BLASLONG, float *, BLASLONG);
  58. BLASLONG (*isbmax_k) (BLASLONG, float *, BLASLONG);
  59. BLASLONG (*isbmin_k) (BLASLONG, float *, BLASLONG);
  60. float (*sbnrm2_k) (BLASLONG, float *, BLASLONG);
  61. float (*sbasum_k) (BLASLONG, float *, BLASLONG);
  62. float (*sbsum_k) (BLASLONG, float *, BLASLONG);
  63. int (*sbcopy_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  64. float (*sbdot_k) (BLASLONG, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG);
  65. double (*dsbdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  66. int (*sbrot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG, float, float);
  67. int (*sbaxpy_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  68. int (*sbscal_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  69. int (*sbswap_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  70. int (*sbgemv_n) (BLASLONG, BLASLONG, float, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, float, float *, BLASLONG);
  71. int (*sbgemv_t) (BLASLONG, BLASLONG, float, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, float, float *, BLASLONG);
  72. int (*sbger_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  73. int (*sbsymv_L) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  74. int (*sbsymv_U) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  75. int (*sbgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, float, bfloat16 *, bfloat16 *, float *, BLASLONG);
  76. int (*sbgemm_beta )(BLASLONG, BLASLONG, BLASLONG, float, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, float *, BLASLONG);
  77. int (*sbgemm_incopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
  78. int (*sbgemm_itcopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
  79. int (*sbgemm_oncopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
  80. int (*sbgemm_otcopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
  81. int (*sbtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  82. int (*sbtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  83. int (*sbtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  84. int (*sbtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  85. int (*sbtrsm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  86. int (*sbtrsm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  87. int (*sbtrsm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  88. int (*sbtrsm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  89. int (*sbtrsm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  90. int (*sbtrsm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  91. int (*sbtrsm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  92. int (*sbtrsm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  93. int (*sbtrsm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  94. int (*sbtrsm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  95. int (*sbtrsm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  96. int (*sbtrsm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  97. int (*sbtrsm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  98. int (*sbtrsm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  99. int (*sbtrsm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  100. int (*sbtrsm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  101. int (*sbtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  102. int (*sbtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  103. int (*sbtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  104. int (*sbtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  105. int (*sbtrmm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  106. int (*sbtrmm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  107. int (*sbtrmm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  108. int (*sbtrmm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  109. int (*sbtrmm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  110. int (*sbtrmm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  111. int (*sbtrmm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  112. int (*sbtrmm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  113. int (*sbtrmm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  114. int (*sbtrmm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  115. int (*sbtrmm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  116. int (*sbtrmm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  117. int (*sbtrmm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  118. int (*sbtrmm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  119. int (*sbtrmm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  120. int (*sbtrmm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  121. int (*sbsymm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  122. int (*sbsymm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  123. int (*sbsymm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  124. int (*sbsymm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  125. int (*sbneg_tcopy) (BLASLONG, BLASLONG, float *, BLASLONG, float *);
  126. int (*sblaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG, blasint *, float *);
  127. #ifdef SMALL_MATRIX_OPT
  128. int (*sbgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float beta);
  129. 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);
  130. 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);
  131. 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);
  132. 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);
  133. 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);
  134. 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);
  135. 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);
  136. 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);
  137. #endif
  138. #endif
  139. #if defined(BUILD_SINGLE) || defined(BUILD_COMPLEX)
  140. int sgemm_p, sgemm_q, sgemm_r;
  141. int sgemm_unroll_m, sgemm_unroll_n, sgemm_unroll_mn;
  142. #endif
  143. int exclusive_cache;
  144. #if defined(BUILD_SINGLE) || defined(BUILD_COMPLEX)
  145. float (*samax_k) (BLASLONG, float *, BLASLONG);
  146. float (*samin_k) (BLASLONG, float *, BLASLONG);
  147. float (*smax_k) (BLASLONG, float *, BLASLONG);
  148. float (*smin_k) (BLASLONG, float *, BLASLONG);
  149. BLASLONG (*isamax_k)(BLASLONG, float *, BLASLONG);
  150. BLASLONG (*isamin_k)(BLASLONG, float *, BLASLONG);
  151. BLASLONG (*ismax_k) (BLASLONG, float *, BLASLONG);
  152. BLASLONG (*ismin_k) (BLASLONG, float *, BLASLONG);
  153. float (*snrm2_k) (BLASLONG, float *, BLASLONG);
  154. float (*sasum_k) (BLASLONG, float *, BLASLONG);
  155. #endif
  156. #ifdef BUILD_SINGLE
  157. float (*ssum_k) (BLASLONG, float *, BLASLONG);
  158. #endif
  159. #if defined(BUILD_SINGLE) || defined(BUILD_COMPLEX)
  160. int (*scopy_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  161. float (*sdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  162. //double (*dsdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  163. int (*srot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG, float, float);
  164. int (*saxpy_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  165. int (*sscal_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  166. int (*sswap_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  167. int (*sgemv_n) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  168. int (*sgemv_t) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  169. #endif
  170. #ifdef BUILD_SINGLE
  171. int (*sger_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  172. int (*ssymv_L) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  173. int (*ssymv_U) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  174. #endif
  175. #if defined(BUILD_SINGLE) || defined(BUILD_COMPLEX)
  176. #ifdef ARCH_X86_64
  177. void (*sgemm_direct) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG , float *, BLASLONG , float * , BLASLONG);
  178. int (*sgemm_direct_performant) (BLASLONG M, BLASLONG N, BLASLONG K);
  179. #endif
  180. int (*sgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG);
  181. int (*sgemm_beta )(BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  182. int (*sgemm_incopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  183. int (*sgemm_itcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  184. int (*sgemm_oncopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  185. int (*sgemm_otcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  186. #endif
  187. #ifdef BUILD_SINGLE
  188. #ifdef SMALL_MATRIX_OPT
  189. int (*sgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float beta);
  190. 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);
  191. 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);
  192. 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);
  193. 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);
  194. 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);
  195. 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);
  196. 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);
  197. 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);
  198. #endif
  199. int (*strsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  200. int (*strsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  201. int (*strsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  202. int (*strsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  203. int (*strsm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  204. int (*strsm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  205. int (*strsm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  206. int (*strsm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  207. int (*strsm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  208. int (*strsm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  209. int (*strsm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  210. int (*strsm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  211. int (*strsm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  212. int (*strsm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  213. int (*strsm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  214. int (*strsm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  215. int (*strsm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  216. int (*strsm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  217. int (*strsm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  218. int (*strsm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  219. int (*strmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  220. int (*strmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  221. int (*strmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  222. int (*strmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  223. int (*strmm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  224. int (*strmm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  225. int (*strmm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  226. int (*strmm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  227. int (*strmm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  228. int (*strmm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  229. int (*strmm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  230. int (*strmm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  231. int (*strmm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  232. int (*strmm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  233. int (*strmm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  234. int (*strmm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  235. int (*strmm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  236. int (*strmm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  237. int (*strmm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  238. int (*strmm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  239. int (*ssymm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  240. int (*ssymm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  241. int (*ssymm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  242. int (*ssymm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  243. int (*sneg_tcopy) (BLASLONG, BLASLONG, float *, BLASLONG, float *);
  244. int (*slaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG, blasint *, float *);
  245. #endif
  246. #if defined(BUILD_DOUBLE) || defined(BUILD_COMPLEX16)
  247. int dgemm_p, dgemm_q, dgemm_r;
  248. int dgemm_unroll_m, dgemm_unroll_n, dgemm_unroll_mn;
  249. #endif
  250. #if defined(BUILD_DOUBLE) || defined(BUILD_COMPLEX16)
  251. double (*damax_k) (BLASLONG, double *, BLASLONG);
  252. double (*damin_k) (BLASLONG, double *, BLASLONG);
  253. double (*dmax_k) (BLASLONG, double *, BLASLONG);
  254. double (*dmin_k) (BLASLONG, double *, BLASLONG);
  255. BLASLONG (*idamax_k)(BLASLONG, double *, BLASLONG);
  256. BLASLONG (*idamin_k)(BLASLONG, double *, BLASLONG);
  257. BLASLONG (*idmax_k) (BLASLONG, double *, BLASLONG);
  258. BLASLONG (*idmin_k) (BLASLONG, double *, BLASLONG);
  259. #endif
  260. #if defined(BUILD_DOUBLE) || defined(BUILD_COMPLEX16)
  261. double (*dnrm2_k) (BLASLONG, double *, BLASLONG);
  262. double (*dasum_k) (BLASLONG, double *, BLASLONG);
  263. #endif
  264. #ifdef BUILD_DOUBLE
  265. double (*dsum_k) (BLASLONG, double *, BLASLONG);
  266. #endif
  267. #if defined(BUILD_DOUBLE) || defined(BUILD_COMPLEX16)
  268. int (*dcopy_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
  269. double (*ddot_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
  270. #endif
  271. #if defined (BUILD_SINGLE) || defined(BUILD_DOUBLE)
  272. double (*dsdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  273. #endif
  274. #if defined(BUILD_DOUBLE) || defined(BUILD_COMPLEX16)
  275. int (*drot_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG, double, double);
  276. int (*daxpy_k) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  277. int (*dscal_k) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  278. int (*dswap_k) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  279. int (*dgemv_n) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  280. int (*dgemv_t) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  281. #endif
  282. #ifdef BUILD_DOUBLE
  283. int (*dger_k) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  284. int (*dsymv_L) (BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  285. int (*dsymv_U) (BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  286. #endif
  287. #if defined(BUILD_DOUBLE) || defined(BUILD_COMPLEX16)
  288. int (*dgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG);
  289. int (*dgemm_beta )(BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  290. int (*dgemm_incopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  291. int (*dgemm_itcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  292. int (*dgemm_oncopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  293. int (*dgemm_otcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  294. #endif
  295. #ifdef BUILD_DOUBLE
  296. #ifdef SMALL_MATRIX_OPT
  297. int (*dgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, double alpha, double beta);
  298. 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);
  299. 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);
  300. 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);
  301. 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);
  302. 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);
  303. 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);
  304. 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);
  305. 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);
  306. #endif
  307. int (*dtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  308. int (*dtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  309. int (*dtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  310. int (*dtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  311. int (*dtrsm_iunucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  312. int (*dtrsm_iunncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  313. int (*dtrsm_iutucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  314. int (*dtrsm_iutncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  315. int (*dtrsm_ilnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  316. int (*dtrsm_ilnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  317. int (*dtrsm_iltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  318. int (*dtrsm_iltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  319. int (*dtrsm_ounucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  320. int (*dtrsm_ounncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  321. int (*dtrsm_outucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  322. int (*dtrsm_outncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  323. int (*dtrsm_olnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  324. int (*dtrsm_olnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  325. int (*dtrsm_oltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  326. int (*dtrsm_oltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  327. int (*dtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  328. int (*dtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  329. int (*dtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  330. int (*dtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  331. int (*dtrmm_iunucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  332. int (*dtrmm_iunncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  333. int (*dtrmm_iutucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  334. int (*dtrmm_iutncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  335. int (*dtrmm_ilnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  336. int (*dtrmm_ilnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  337. int (*dtrmm_iltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  338. int (*dtrmm_iltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  339. int (*dtrmm_ounucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  340. int (*dtrmm_ounncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  341. int (*dtrmm_outucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  342. int (*dtrmm_outncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  343. int (*dtrmm_olnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  344. int (*dtrmm_olnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  345. int (*dtrmm_oltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  346. int (*dtrmm_oltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  347. int (*dsymm_iutcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  348. int (*dsymm_iltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  349. int (*dsymm_outcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  350. int (*dsymm_oltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  351. int (*dneg_tcopy) (BLASLONG, BLASLONG, double *, BLASLONG, double *);
  352. int (*dlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, double *, BLASLONG, blasint *, double *);
  353. #endif
  354. #ifdef EXPRECISION
  355. int qgemm_p, qgemm_q, qgemm_r;
  356. int qgemm_unroll_m, qgemm_unroll_n, qgemm_unroll_mn;
  357. xdouble (*qamax_k) (BLASLONG, xdouble *, BLASLONG);
  358. xdouble (*qamin_k) (BLASLONG, xdouble *, BLASLONG);
  359. xdouble (*qmax_k) (BLASLONG, xdouble *, BLASLONG);
  360. xdouble (*qmin_k) (BLASLONG, xdouble *, BLASLONG);
  361. BLASLONG (*iqamax_k)(BLASLONG, xdouble *, BLASLONG);
  362. BLASLONG (*iqamin_k)(BLASLONG, xdouble *, BLASLONG);
  363. BLASLONG (*iqmax_k) (BLASLONG, xdouble *, BLASLONG);
  364. BLASLONG (*iqmin_k) (BLASLONG, xdouble *, BLASLONG);
  365. xdouble (*qnrm2_k) (BLASLONG, xdouble *, BLASLONG);
  366. xdouble (*qasum_k) (BLASLONG, xdouble *, BLASLONG);
  367. xdouble (*qsum_k) (BLASLONG, xdouble *, BLASLONG);
  368. int (*qcopy_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  369. xdouble (*qdot_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  370. int (*qrot_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble);
  371. int (*qaxpy_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  372. int (*qscal_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  373. int (*qswap_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  374. int (*qgemv_n) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  375. int (*qgemv_t) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  376. int (*qger_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  377. int (*qsymv_L) (BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  378. int (*qsymv_U) (BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  379. int (*qgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  380. int (*qgemm_beta )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  381. int (*qgemm_incopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  382. int (*qgemm_itcopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  383. int (*qgemm_oncopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  384. int (*qgemm_otcopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  385. int (*qtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  386. int (*qtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  387. int (*qtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  388. int (*qtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  389. int (*qtrsm_iunucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  390. int (*qtrsm_iunncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  391. int (*qtrsm_iutucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  392. int (*qtrsm_iutncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  393. int (*qtrsm_ilnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  394. int (*qtrsm_ilnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  395. int (*qtrsm_iltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  396. int (*qtrsm_iltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  397. int (*qtrsm_ounucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  398. int (*qtrsm_ounncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  399. int (*qtrsm_outucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  400. int (*qtrsm_outncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  401. int (*qtrsm_olnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  402. int (*qtrsm_olnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  403. int (*qtrsm_oltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  404. int (*qtrsm_oltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  405. int (*qtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  406. int (*qtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  407. int (*qtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  408. int (*qtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  409. int (*qtrmm_iunucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  410. int (*qtrmm_iunncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  411. int (*qtrmm_iutucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  412. int (*qtrmm_iutncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  413. int (*qtrmm_ilnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  414. int (*qtrmm_ilnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  415. int (*qtrmm_iltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  416. int (*qtrmm_iltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  417. int (*qtrmm_ounucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  418. int (*qtrmm_ounncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  419. int (*qtrmm_outucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  420. int (*qtrmm_outncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  421. int (*qtrmm_olnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  422. int (*qtrmm_olnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  423. int (*qtrmm_oltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  424. int (*qtrmm_oltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  425. int (*qsymm_iutcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  426. int (*qsymm_iltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  427. int (*qsymm_outcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  428. int (*qsymm_oltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  429. int (*qneg_tcopy) (BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  430. int (*qlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, xdouble *, BLASLONG, blasint *, xdouble *);
  431. #endif
  432. #ifdef BUILD_COMPLEX
  433. int cgemm_p, cgemm_q, cgemm_r;
  434. int cgemm_unroll_m, cgemm_unroll_n, cgemm_unroll_mn;
  435. float (*camax_k) (BLASLONG, float *, BLASLONG);
  436. float (*camin_k) (BLASLONG, float *, BLASLONG);
  437. BLASLONG (*icamax_k)(BLASLONG, float *, BLASLONG);
  438. BLASLONG (*icamin_k)(BLASLONG, float *, BLASLONG);
  439. float (*cnrm2_k) (BLASLONG, float *, BLASLONG);
  440. float (*casum_k) (BLASLONG, float *, BLASLONG);
  441. float (*csum_k) (BLASLONG, float *, BLASLONG);
  442. int (*ccopy_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  443. openblas_complex_float (*cdotu_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  444. openblas_complex_float (*cdotc_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  445. int (*csrot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG, float, float);
  446. int (*caxpy_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  447. int (*caxpyc_k)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  448. int (*cscal_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  449. int (*cswap_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  450. int (*cgemv_n) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  451. int (*cgemv_t) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  452. int (*cgemv_r) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  453. int (*cgemv_c) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  454. int (*cgemv_o) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  455. int (*cgemv_u) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  456. int (*cgemv_s) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  457. int (*cgemv_d) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  458. int (*cgeru_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  459. int (*cgerc_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  460. int (*cgerv_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  461. int (*cgerd_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  462. int (*csymv_L) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  463. int (*csymv_U) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  464. int (*chemv_L) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  465. int (*chemv_U) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  466. int (*chemv_M) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  467. int (*chemv_V) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  468. int (*cgemm_kernel_n )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  469. int (*cgemm_kernel_l )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  470. int (*cgemm_kernel_r )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  471. int (*cgemm_kernel_b )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  472. int (*cgemm_beta )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  473. int (*cgemm_incopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  474. int (*cgemm_itcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  475. int (*cgemm_oncopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  476. int (*cgemm_otcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  477. #ifdef SMALL_MATRIX_OPT
  478. int (*cgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, float alpha0, float alpha1, float beta0, float beta1);
  479. 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);
  480. 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);
  481. 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);
  482. 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);
  483. 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);
  484. 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);
  485. 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);
  486. 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);
  487. 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);
  488. 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);
  489. 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);
  490. 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);
  491. 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);
  492. 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);
  493. 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);
  494. 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);
  495. 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);
  496. 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);
  497. 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);
  498. 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);
  499. 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);
  500. 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);
  501. 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);
  502. 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);
  503. 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);
  504. 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);
  505. 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);
  506. 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);
  507. 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);
  508. 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);
  509. 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);
  510. 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);
  511. #endif
  512. int (*ctrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  513. int (*ctrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  514. int (*ctrsm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  515. int (*ctrsm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  516. int (*ctrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  517. int (*ctrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  518. int (*ctrsm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  519. int (*ctrsm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  520. int (*ctrsm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  521. int (*ctrsm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  522. int (*ctrsm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  523. int (*ctrsm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  524. int (*ctrsm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  525. int (*ctrsm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  526. int (*ctrsm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  527. int (*ctrsm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  528. int (*ctrsm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  529. int (*ctrsm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  530. int (*ctrsm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  531. int (*ctrsm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  532. int (*ctrsm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  533. int (*ctrsm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  534. int (*ctrsm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  535. int (*ctrsm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  536. int (*ctrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  537. int (*ctrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  538. int (*ctrmm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  539. int (*ctrmm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  540. int (*ctrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  541. int (*ctrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  542. int (*ctrmm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  543. int (*ctrmm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  544. int (*ctrmm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  545. int (*ctrmm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  546. int (*ctrmm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  547. int (*ctrmm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  548. int (*ctrmm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  549. int (*ctrmm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  550. int (*ctrmm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  551. int (*ctrmm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  552. int (*ctrmm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  553. int (*ctrmm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  554. int (*ctrmm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  555. int (*ctrmm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  556. int (*ctrmm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  557. int (*ctrmm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  558. int (*ctrmm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  559. int (*ctrmm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  560. int (*csymm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  561. int (*csymm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  562. int (*csymm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  563. int (*csymm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  564. int (*chemm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  565. int (*chemm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  566. int (*chemm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  567. int (*chemm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  568. int cgemm3m_p, cgemm3m_q, cgemm3m_r;
  569. int cgemm3m_unroll_m, cgemm3m_unroll_n, cgemm3m_unroll_mn;
  570. int (*cgemm3m_kernel)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  571. int (*cgemm3m_incopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  572. int (*cgemm3m_incopyr)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  573. int (*cgemm3m_incopyi)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  574. int (*cgemm3m_itcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  575. int (*cgemm3m_itcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  576. int (*cgemm3m_itcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  577. int (*cgemm3m_oncopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  578. int (*cgemm3m_oncopyr)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  579. int (*cgemm3m_oncopyi)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  580. int (*cgemm3m_otcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  581. int (*cgemm3m_otcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  582. int (*cgemm3m_otcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  583. int (*csymm3m_iucopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  584. int (*csymm3m_ilcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  585. int (*csymm3m_iucopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  586. int (*csymm3m_ilcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  587. int (*csymm3m_iucopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  588. int (*csymm3m_ilcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  589. int (*csymm3m_oucopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  590. int (*csymm3m_olcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  591. int (*csymm3m_oucopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  592. int (*csymm3m_olcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  593. int (*csymm3m_oucopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  594. int (*csymm3m_olcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  595. int (*chemm3m_iucopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  596. int (*chemm3m_ilcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  597. int (*chemm3m_iucopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  598. int (*chemm3m_ilcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  599. int (*chemm3m_iucopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  600. int (*chemm3m_ilcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  601. int (*chemm3m_oucopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  602. int (*chemm3m_olcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  603. int (*chemm3m_oucopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  604. int (*chemm3m_olcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  605. int (*chemm3m_oucopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  606. int (*chemm3m_olcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  607. int (*cneg_tcopy) (BLASLONG, BLASLONG, float *, BLASLONG, float *);
  608. int (*claswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG, blasint *, float *);
  609. #endif
  610. #ifdef BUILD_COMPLEX16
  611. int zgemm_p, zgemm_q, zgemm_r;
  612. int zgemm_unroll_m, zgemm_unroll_n, zgemm_unroll_mn;
  613. double (*zamax_k) (BLASLONG, double *, BLASLONG);
  614. double (*zamin_k) (BLASLONG, double *, BLASLONG);
  615. BLASLONG (*izamax_k)(BLASLONG, double *, BLASLONG);
  616. BLASLONG (*izamin_k)(BLASLONG, double *, BLASLONG);
  617. double (*znrm2_k) (BLASLONG, double *, BLASLONG);
  618. double (*zasum_k) (BLASLONG, double *, BLASLONG);
  619. double (*zsum_k) (BLASLONG, double *, BLASLONG);
  620. int (*zcopy_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
  621. openblas_complex_double (*zdotu_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
  622. openblas_complex_double (*zdotc_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
  623. int (*zdrot_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG, double, double);
  624. int (*zaxpy_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  625. int (*zaxpyc_k)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  626. int (*zscal_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  627. int (*zswap_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  628. int (*zgemv_n) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  629. int (*zgemv_t) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  630. int (*zgemv_r) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  631. int (*zgemv_c) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  632. int (*zgemv_o) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  633. int (*zgemv_u) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  634. int (*zgemv_s) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  635. int (*zgemv_d) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  636. int (*zgeru_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  637. int (*zgerc_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  638. int (*zgerv_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  639. int (*zgerd_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  640. int (*zsymv_L) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  641. int (*zsymv_U) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  642. int (*zhemv_L) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  643. int (*zhemv_U) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  644. int (*zhemv_M) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  645. int (*zhemv_V) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  646. int (*zgemm_kernel_n )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  647. int (*zgemm_kernel_l )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  648. int (*zgemm_kernel_r )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  649. int (*zgemm_kernel_b )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  650. int (*zgemm_beta )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  651. int (*zgemm_incopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  652. int (*zgemm_itcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  653. int (*zgemm_oncopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  654. int (*zgemm_otcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  655. #ifdef SMALL_MATRIX_OPT
  656. int (*zgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, double alpha0, double alpha1, double beta0, double beta1);
  657. 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);
  658. 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);
  659. 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);
  660. 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);
  661. 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);
  662. 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);
  663. 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);
  664. 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);
  665. 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);
  666. 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);
  667. 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);
  668. 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);
  669. 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);
  670. 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);
  671. 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);
  672. 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);
  673. 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);
  674. 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);
  675. 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);
  676. 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);
  677. 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);
  678. 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);
  679. 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);
  680. 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);
  681. 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);
  682. 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);
  683. 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);
  684. 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);
  685. 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);
  686. 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);
  687. 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);
  688. 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);
  689. #endif
  690. int (*ztrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  691. int (*ztrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  692. int (*ztrsm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  693. int (*ztrsm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  694. int (*ztrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  695. int (*ztrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  696. int (*ztrsm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  697. int (*ztrsm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  698. int (*ztrsm_iunucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  699. int (*ztrsm_iunncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  700. int (*ztrsm_iutucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  701. int (*ztrsm_iutncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  702. int (*ztrsm_ilnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  703. int (*ztrsm_ilnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  704. int (*ztrsm_iltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  705. int (*ztrsm_iltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  706. int (*ztrsm_ounucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  707. int (*ztrsm_ounncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  708. int (*ztrsm_outucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  709. int (*ztrsm_outncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  710. int (*ztrsm_olnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  711. int (*ztrsm_olnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  712. int (*ztrsm_oltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  713. int (*ztrsm_oltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  714. int (*ztrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  715. int (*ztrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  716. int (*ztrmm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  717. int (*ztrmm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  718. int (*ztrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  719. int (*ztrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  720. int (*ztrmm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  721. int (*ztrmm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  722. int (*ztrmm_iunucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  723. int (*ztrmm_iunncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  724. int (*ztrmm_iutucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  725. int (*ztrmm_iutncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  726. int (*ztrmm_ilnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  727. int (*ztrmm_ilnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  728. int (*ztrmm_iltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  729. int (*ztrmm_iltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  730. int (*ztrmm_ounucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  731. int (*ztrmm_ounncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  732. int (*ztrmm_outucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  733. int (*ztrmm_outncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  734. int (*ztrmm_olnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  735. int (*ztrmm_olnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  736. int (*ztrmm_oltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  737. int (*ztrmm_oltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  738. int (*zsymm_iutcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  739. int (*zsymm_iltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  740. int (*zsymm_outcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  741. int (*zsymm_oltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  742. int (*zhemm_iutcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  743. int (*zhemm_iltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  744. int (*zhemm_outcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  745. int (*zhemm_oltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  746. int zgemm3m_p, zgemm3m_q, zgemm3m_r;
  747. int zgemm3m_unroll_m, zgemm3m_unroll_n, zgemm3m_unroll_mn;
  748. int (*zgemm3m_kernel)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  749. int (*zgemm3m_incopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  750. int (*zgemm3m_incopyr)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  751. int (*zgemm3m_incopyi)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  752. int (*zgemm3m_itcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  753. int (*zgemm3m_itcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  754. int (*zgemm3m_itcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  755. int (*zgemm3m_oncopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  756. int (*zgemm3m_oncopyr)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  757. int (*zgemm3m_oncopyi)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  758. int (*zgemm3m_otcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  759. int (*zgemm3m_otcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  760. int (*zgemm3m_otcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  761. int (*zsymm3m_iucopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  762. int (*zsymm3m_ilcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  763. int (*zsymm3m_iucopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  764. int (*zsymm3m_ilcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  765. int (*zsymm3m_iucopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  766. int (*zsymm3m_ilcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  767. int (*zsymm3m_oucopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  768. int (*zsymm3m_olcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  769. int (*zsymm3m_oucopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  770. int (*zsymm3m_olcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  771. int (*zsymm3m_oucopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  772. int (*zsymm3m_olcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  773. int (*zhemm3m_iucopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  774. int (*zhemm3m_ilcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  775. int (*zhemm3m_iucopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  776. int (*zhemm3m_ilcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  777. int (*zhemm3m_iucopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  778. int (*zhemm3m_ilcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  779. int (*zhemm3m_oucopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  780. int (*zhemm3m_olcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  781. int (*zhemm3m_oucopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  782. int (*zhemm3m_olcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  783. int (*zhemm3m_oucopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  784. int (*zhemm3m_olcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  785. int (*zneg_tcopy) (BLASLONG, BLASLONG, double *, BLASLONG, double *);
  786. int (*zlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, double *, BLASLONG, blasint *, double *);
  787. #endif
  788. #ifdef EXPRECISION
  789. int xgemm_p, xgemm_q, xgemm_r;
  790. int xgemm_unroll_m, xgemm_unroll_n, xgemm_unroll_mn;
  791. xdouble (*xamax_k) (BLASLONG, xdouble *, BLASLONG);
  792. xdouble (*xamin_k) (BLASLONG, xdouble *, BLASLONG);
  793. BLASLONG (*ixamax_k)(BLASLONG, xdouble *, BLASLONG);
  794. BLASLONG (*ixamin_k)(BLASLONG, xdouble *, BLASLONG);
  795. xdouble (*xnrm2_k) (BLASLONG, xdouble *, BLASLONG);
  796. xdouble (*xasum_k) (BLASLONG, xdouble *, BLASLONG);
  797. xdouble (*xsum_k) (BLASLONG, xdouble *, BLASLONG);
  798. int (*xcopy_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  799. openblas_complex_xdouble (*xdotu_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  800. openblas_complex_xdouble (*xdotc_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  801. int (*xqrot_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble);
  802. int (*xaxpy_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  803. int (*xaxpyc_k)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  804. int (*xscal_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  805. int (*xswap_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  806. int (*xgemv_n) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  807. int (*xgemv_t) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  808. int (*xgemv_r) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  809. int (*xgemv_c) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  810. int (*xgemv_o) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  811. int (*xgemv_u) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  812. int (*xgemv_s) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  813. int (*xgemv_d) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  814. int (*xgeru_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  815. int (*xgerc_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  816. int (*xgerv_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  817. int (*xgerd_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  818. int (*xsymv_L) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  819. int (*xsymv_U) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  820. int (*xhemv_L) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  821. int (*xhemv_U) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  822. int (*xhemv_M) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  823. int (*xhemv_V) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  824. int (*xgemm_kernel_n )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  825. int (*xgemm_kernel_l )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  826. int (*xgemm_kernel_r )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  827. int (*xgemm_kernel_b )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  828. int (*xgemm_beta )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  829. int (*xgemm_incopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  830. int (*xgemm_itcopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  831. int (*xgemm_oncopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  832. int (*xgemm_otcopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  833. int (*xtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  834. int (*xtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  835. int (*xtrsm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  836. int (*xtrsm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  837. int (*xtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  838. int (*xtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  839. int (*xtrsm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  840. int (*xtrsm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  841. int (*xtrsm_iunucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  842. int (*xtrsm_iunncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  843. int (*xtrsm_iutucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  844. int (*xtrsm_iutncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  845. int (*xtrsm_ilnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  846. int (*xtrsm_ilnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  847. int (*xtrsm_iltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  848. int (*xtrsm_iltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  849. int (*xtrsm_ounucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  850. int (*xtrsm_ounncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  851. int (*xtrsm_outucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  852. int (*xtrsm_outncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  853. int (*xtrsm_olnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  854. int (*xtrsm_olnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  855. int (*xtrsm_oltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  856. int (*xtrsm_oltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  857. int (*xtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  858. int (*xtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  859. int (*xtrmm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  860. int (*xtrmm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  861. int (*xtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  862. int (*xtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  863. int (*xtrmm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  864. int (*xtrmm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  865. int (*xtrmm_iunucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  866. int (*xtrmm_iunncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  867. int (*xtrmm_iutucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  868. int (*xtrmm_iutncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  869. int (*xtrmm_ilnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  870. int (*xtrmm_ilnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  871. int (*xtrmm_iltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  872. int (*xtrmm_iltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  873. int (*xtrmm_ounucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  874. int (*xtrmm_ounncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  875. int (*xtrmm_outucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  876. int (*xtrmm_outncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  877. int (*xtrmm_olnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  878. int (*xtrmm_olnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  879. int (*xtrmm_oltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  880. int (*xtrmm_oltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  881. int (*xsymm_iutcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  882. int (*xsymm_iltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  883. int (*xsymm_outcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  884. int (*xsymm_oltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  885. int (*xhemm_iutcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  886. int (*xhemm_iltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  887. int (*xhemm_outcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  888. int (*xhemm_oltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  889. int xgemm3m_p, xgemm3m_q, xgemm3m_r;
  890. int xgemm3m_unroll_m, xgemm3m_unroll_n, xgemm3m_unroll_mn;
  891. int (*xgemm3m_kernel)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  892. int (*xgemm3m_incopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  893. int (*xgemm3m_incopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  894. int (*xgemm3m_incopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  895. int (*xgemm3m_itcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  896. int (*xgemm3m_itcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  897. int (*xgemm3m_itcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  898. int (*xgemm3m_oncopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  899. int (*xgemm3m_oncopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  900. int (*xgemm3m_oncopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  901. int (*xgemm3m_otcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  902. int (*xgemm3m_otcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  903. int (*xgemm3m_otcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  904. int (*xsymm3m_iucopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  905. int (*xsymm3m_ilcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  906. int (*xsymm3m_iucopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  907. int (*xsymm3m_ilcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  908. int (*xsymm3m_iucopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  909. int (*xsymm3m_ilcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  910. int (*xsymm3m_oucopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  911. int (*xsymm3m_olcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  912. int (*xsymm3m_oucopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  913. int (*xsymm3m_olcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  914. int (*xsymm3m_oucopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  915. int (*xsymm3m_olcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  916. int (*xhemm3m_iucopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  917. int (*xhemm3m_ilcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  918. int (*xhemm3m_iucopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  919. int (*xhemm3m_ilcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  920. int (*xhemm3m_iucopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  921. int (*xhemm3m_ilcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  922. int (*xhemm3m_oucopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  923. int (*xhemm3m_olcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  924. int (*xhemm3m_oucopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  925. int (*xhemm3m_olcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  926. int (*xhemm3m_oucopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  927. int (*xhemm3m_olcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  928. int (*xneg_tcopy) (BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  929. int (*xlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, xdouble *, BLASLONG, blasint *, xdouble *);
  930. #endif
  931. void (*init)(void);
  932. int snum_opt, dnum_opt, qnum_opt;
  933. #ifdef BUILD_SINGLE
  934. int (*saxpby_k) (BLASLONG, float, float*, BLASLONG,float, float*, BLASLONG);
  935. #endif
  936. #ifdef BUILD_DOUBLE
  937. int (*daxpby_k) (BLASLONG, double, double*, BLASLONG,double, double*, BLASLONG);
  938. #endif
  939. #ifdef BUILD_COMPLEX
  940. int (*caxpby_k) (BLASLONG, float, float, float*, BLASLONG,float,float, float*, BLASLONG);
  941. #endif
  942. #ifdef BUILD_COMPLEX16
  943. int (*zaxpby_k) (BLASLONG, double, double, double*, BLASLONG,double,double, double*, BLASLONG);
  944. #endif
  945. #ifdef BUILD_SINGLE
  946. int (*somatcopy_k_cn) (BLASLONG, BLASLONG, float, float*, BLASLONG, float*, BLASLONG);
  947. int (*somatcopy_k_ct) (BLASLONG, BLASLONG, float, float*, BLASLONG, float*, BLASLONG);
  948. int (*somatcopy_k_rn) (BLASLONG, BLASLONG, float, float*, BLASLONG, float*, BLASLONG);
  949. int (*somatcopy_k_rt) (BLASLONG, BLASLONG, float, float*, BLASLONG, float*, BLASLONG);
  950. #endif
  951. #ifdef BUILD_DOUBLE
  952. int (*domatcopy_k_cn) (BLASLONG, BLASLONG, double, double*, BLASLONG, double*, BLASLONG);
  953. int (*domatcopy_k_ct) (BLASLONG, BLASLONG, double, double*, BLASLONG, double*, BLASLONG);
  954. int (*domatcopy_k_rn) (BLASLONG, BLASLONG, double, double*, BLASLONG, double*, BLASLONG);
  955. int (*domatcopy_k_rt) (BLASLONG, BLASLONG, double, double*, BLASLONG, double*, BLASLONG);
  956. #endif
  957. #ifdef BUILD_COMPLEX
  958. int (*comatcopy_k_cn) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  959. int (*comatcopy_k_ct) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  960. int (*comatcopy_k_rn) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  961. int (*comatcopy_k_rt) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  962. int (*comatcopy_k_cnc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  963. int (*comatcopy_k_ctc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  964. int (*comatcopy_k_rnc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  965. int (*comatcopy_k_rtc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  966. #endif
  967. #ifdef BUILD_COMPLEX16
  968. int (*zomatcopy_k_cn) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  969. int (*zomatcopy_k_ct) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  970. int (*zomatcopy_k_rn) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  971. int (*zomatcopy_k_rt) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  972. int (*zomatcopy_k_cnc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  973. int (*zomatcopy_k_ctc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  974. int (*zomatcopy_k_rnc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  975. int (*zomatcopy_k_rtc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  976. #endif
  977. #ifdef BUILD_SINGLE
  978. int (*simatcopy_k_cn) (BLASLONG, BLASLONG, float, float*, BLASLONG);
  979. int (*simatcopy_k_ct) (BLASLONG, BLASLONG, float, float*, BLASLONG);
  980. int (*simatcopy_k_rn) (BLASLONG, BLASLONG, float, float*, BLASLONG);
  981. int (*simatcopy_k_rt) (BLASLONG, BLASLONG, float, float*, BLASLONG);
  982. #endif
  983. #ifdef BUILD_DOUBLE
  984. int (*dimatcopy_k_cn) (BLASLONG, BLASLONG, double, double*, BLASLONG);
  985. int (*dimatcopy_k_ct) (BLASLONG, BLASLONG, double, double*, BLASLONG);
  986. int (*dimatcopy_k_rn) (BLASLONG, BLASLONG, double, double*, BLASLONG);
  987. int (*dimatcopy_k_rt) (BLASLONG, BLASLONG, double, double*, BLASLONG);
  988. #endif
  989. #ifdef BUILD_COMPLEX
  990. int (*cimatcopy_k_cn) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  991. int (*cimatcopy_k_ct) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  992. int (*cimatcopy_k_rn) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  993. int (*cimatcopy_k_rt) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  994. int (*cimatcopy_k_cnc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  995. int (*cimatcopy_k_ctc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  996. int (*cimatcopy_k_rnc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  997. int (*cimatcopy_k_rtc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  998. #endif
  999. #ifdef BUILD_COMPLEX16
  1000. int (*zimatcopy_k_cn) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1001. int (*zimatcopy_k_ct) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1002. int (*zimatcopy_k_rn) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1003. int (*zimatcopy_k_rt) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1004. int (*zimatcopy_k_cnc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1005. int (*zimatcopy_k_ctc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1006. int (*zimatcopy_k_rnc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1007. int (*zimatcopy_k_rtc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1008. #endif
  1009. #ifdef BUILD_SINGLE
  1010. int (*sgeadd_k) (BLASLONG, BLASLONG, float, float *, BLASLONG, float, float *, BLASLONG);
  1011. #endif
  1012. #ifdef BUILD_DOUBLE
  1013. int (*dgeadd_k) (BLASLONG, BLASLONG, double, double *, BLASLONG, double, double *, BLASLONG);
  1014. #endif
  1015. #ifdef BUILD_COMPLEX
  1016. int (*cgeadd_k) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float, float, float *, BLASLONG);
  1017. #endif
  1018. #ifdef BUILD_COMPLEX16
  1019. int (*zgeadd_k) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double, double, double *, BLASLONG);
  1020. #endif
  1021. } gotoblas_t;
  1022. extern gotoblas_t *gotoblas;
  1023. #define FUNC_OFFSET(func) (size_t)(&((gotoblas_t *)NULL)->func)
  1024. #define DTB_ENTRIES gotoblas -> dtb_entries
  1025. #define GEMM_OFFSET_A gotoblas -> offsetA
  1026. #define GEMM_OFFSET_B gotoblas -> offsetB
  1027. #define GEMM_ALIGN gotoblas -> align
  1028. #define HAVE_EX_L2 gotoblas -> exclusive_cache
  1029. #ifdef BUILD_BFLOAT16
  1030. #define SBGEMM_P gotoblas -> sbgemm_p
  1031. #define SBGEMM_Q gotoblas -> sbgemm_q
  1032. #define SBGEMM_R gotoblas -> sbgemm_r
  1033. #define SBGEMM_UNROLL_M gotoblas -> sbgemm_unroll_m
  1034. #define SBGEMM_UNROLL_N gotoblas -> sbgemm_unroll_n
  1035. #define SBGEMM_UNROLL_MN gotoblas -> sbgemm_unroll_mn
  1036. #endif
  1037. #if defined (BUILD_SINGLE)
  1038. #define SGEMM_P gotoblas -> sgemm_p
  1039. #define SGEMM_Q gotoblas -> sgemm_q
  1040. #define SGEMM_R gotoblas -> sgemm_r
  1041. #define SGEMM_UNROLL_M gotoblas -> sgemm_unroll_m
  1042. #define SGEMM_UNROLL_N gotoblas -> sgemm_unroll_n
  1043. #define SGEMM_UNROLL_MN gotoblas -> sgemm_unroll_mn
  1044. #endif
  1045. #if defined (BUILD_DOUBLE)
  1046. #define DGEMM_P gotoblas -> dgemm_p
  1047. #define DGEMM_Q gotoblas -> dgemm_q
  1048. #define DGEMM_R gotoblas -> dgemm_r
  1049. #define DGEMM_UNROLL_M gotoblas -> dgemm_unroll_m
  1050. #define DGEMM_UNROLL_N gotoblas -> dgemm_unroll_n
  1051. #define DGEMM_UNROLL_MN gotoblas -> dgemm_unroll_mn
  1052. #endif
  1053. #define QGEMM_P gotoblas -> qgemm_p
  1054. #define QGEMM_Q gotoblas -> qgemm_q
  1055. #define QGEMM_R gotoblas -> qgemm_r
  1056. #define QGEMM_UNROLL_M gotoblas -> qgemm_unroll_m
  1057. #define QGEMM_UNROLL_N gotoblas -> qgemm_unroll_n
  1058. #define QGEMM_UNROLL_MN gotoblas -> qgemm_unroll_mn
  1059. #ifdef BUILD_COMPLEX
  1060. #define CGEMM_P gotoblas -> cgemm_p
  1061. #define CGEMM_Q gotoblas -> cgemm_q
  1062. #define CGEMM_R gotoblas -> cgemm_r
  1063. #define CGEMM_UNROLL_M gotoblas -> cgemm_unroll_m
  1064. #define CGEMM_UNROLL_N gotoblas -> cgemm_unroll_n
  1065. #define CGEMM_UNROLL_MN gotoblas -> cgemm_unroll_mn
  1066. #ifndef BUILD_SINGLE
  1067. #define SGEMM_P gotoblas -> sgemm_p
  1068. #define SGEMM_Q gotoblas -> sgemm_q
  1069. #define SGEMM_R 1024
  1070. #define SGEMM_UNROLL_M gotoblas -> sgemm_unroll_m
  1071. #define SGEMM_UNROLL_N gotoblas -> sgemm_unroll_n
  1072. #define SGEMM_UNROLL_MN gotoblas -> sgemm_unroll_mn
  1073. #endif
  1074. #endif
  1075. #ifdef BUILD_COMPLEX16
  1076. #define ZGEMM_P gotoblas -> zgemm_p
  1077. #define ZGEMM_Q gotoblas -> zgemm_q
  1078. #define ZGEMM_R gotoblas -> zgemm_r
  1079. #define ZGEMM_UNROLL_M gotoblas -> zgemm_unroll_m
  1080. #define ZGEMM_UNROLL_N gotoblas -> zgemm_unroll_n
  1081. #define ZGEMM_UNROLL_MN gotoblas -> zgemm_unroll_mn
  1082. #ifndef BUILD_DOUBLE
  1083. #define DGEMM_P gotoblas -> dgemm_p
  1084. #define DGEMM_Q gotoblas -> dgemm_q
  1085. #define DGEMM_R 1024
  1086. #define DGEMM_UNROLL_M gotoblas -> dgemm_unroll_m
  1087. #define DGEMM_UNROLL_N gotoblas -> dgemm_unroll_n
  1088. #define DGEMM_UNROLL_MN gotoblas -> dgemm_unroll_mn
  1089. #endif
  1090. #endif
  1091. #define XGEMM_P gotoblas -> xgemm_p
  1092. #define XGEMM_Q gotoblas -> xgemm_q
  1093. #define XGEMM_R gotoblas -> xgemm_r
  1094. #define XGEMM_UNROLL_M gotoblas -> xgemm_unroll_m
  1095. #define XGEMM_UNROLL_N gotoblas -> xgemm_unroll_n
  1096. #define XGEMM_UNROLL_MN gotoblas -> xgemm_unroll_mn
  1097. #define CGEMM3M_P gotoblas -> cgemm3m_p
  1098. #define CGEMM3M_Q gotoblas -> cgemm3m_q
  1099. #define CGEMM3M_R gotoblas -> cgemm3m_r
  1100. #define CGEMM3M_UNROLL_M gotoblas -> cgemm3m_unroll_m
  1101. #define CGEMM3M_UNROLL_N gotoblas -> cgemm3m_unroll_n
  1102. #define CGEMM3M_UNROLL_MN gotoblas -> cgemm3m_unroll_mn
  1103. #define ZGEMM3M_P gotoblas -> zgemm3m_p
  1104. #define ZGEMM3M_Q gotoblas -> zgemm3m_q
  1105. #define ZGEMM3M_R gotoblas -> zgemm3m_r
  1106. #define ZGEMM3M_UNROLL_M gotoblas -> zgemm3m_unroll_m
  1107. #define ZGEMM3M_UNROLL_N gotoblas -> zgemm3m_unroll_n
  1108. #define ZGEMM3M_UNROLL_MN gotoblas -> zgemm3m_unroll_mn
  1109. #define XGEMM3M_P gotoblas -> xgemm3m_p
  1110. #define XGEMM3M_Q gotoblas -> xgemm3m_q
  1111. #define XGEMM3M_R gotoblas -> xgemm3m_r
  1112. #define XGEMM3M_UNROLL_M gotoblas -> xgemm3m_unroll_m
  1113. #define XGEMM3M_UNROLL_N gotoblas -> xgemm3m_unroll_n
  1114. #define XGEMM3M_UNROLL_MN gotoblas -> xgemm3m_unroll_mn
  1115. #else
  1116. #define FUNC_OFFSET(func) (size_t)(func)
  1117. #define DTB_ENTRIES DTB_DEFAULT_ENTRIES
  1118. #define GEMM_OFFSET_A GEMM_DEFAULT_OFFSET_A
  1119. #define GEMM_OFFSET_B GEMM_DEFAULT_OFFSET_B
  1120. #define GEMM_ALIGN GEMM_DEFAULT_ALIGN
  1121. #ifdef HAVE_EXCLUSIVE_CACHE
  1122. #define HAVE_EX_L2 1
  1123. #else
  1124. #define HAVE_EX_L2 0
  1125. #endif
  1126. #ifdef BUILD_BFLOAT16
  1127. #define SBGEMM_P SBGEMM_DEFAULT_P
  1128. #define SBGEMM_Q SBGEMM_DEFAULT_Q
  1129. #define SBGEMM_R SBGEMM_DEFAULT_R
  1130. #define SBGEMM_UNROLL_M SBGEMM_DEFAULT_UNROLL_M
  1131. #define SBGEMM_UNROLL_N SBGEMM_DEFAULT_UNROLL_N
  1132. #ifdef SBGEMM_DEFAULT_UNROLL_MN
  1133. #define SBGEMM_UNROLL_MN SBGEMM_DEFAULT_UNROLL_MN
  1134. #else
  1135. #define SBGEMM_UNROLL_MN MAX((SBGEMM_UNROLL_M), (SBGEMM_UNROLL_N))
  1136. #endif
  1137. #endif
  1138. #define SGEMM_P SGEMM_DEFAULT_P
  1139. #define SGEMM_Q SGEMM_DEFAULT_Q
  1140. #define SGEMM_R SGEMM_DEFAULT_R
  1141. #define SGEMM_UNROLL_M SGEMM_DEFAULT_UNROLL_M
  1142. #define SGEMM_UNROLL_N SGEMM_DEFAULT_UNROLL_N
  1143. #ifdef SGEMM_DEFAULT_UNROLL_MN
  1144. #define SGEMM_UNROLL_MN SGEMM_DEFAULT_UNROLL_MN
  1145. #else
  1146. #define SGEMM_UNROLL_MN MAX((SGEMM_UNROLL_M), (SGEMM_UNROLL_N))
  1147. #endif
  1148. #define DGEMM_P DGEMM_DEFAULT_P
  1149. #define DGEMM_Q DGEMM_DEFAULT_Q
  1150. #define DGEMM_R DGEMM_DEFAULT_R
  1151. #define DGEMM_UNROLL_M DGEMM_DEFAULT_UNROLL_M
  1152. #define DGEMM_UNROLL_N DGEMM_DEFAULT_UNROLL_N
  1153. #ifdef DGEMM_DEFAULT_UNROLL_MN
  1154. #define DGEMM_UNROLL_MN DGEMM_DEFAULT_UNROLL_MN
  1155. #else
  1156. #define DGEMM_UNROLL_MN MAX((DGEMM_UNROLL_M), (DGEMM_UNROLL_N))
  1157. #endif
  1158. #define QGEMM_P QGEMM_DEFAULT_P
  1159. #define QGEMM_Q QGEMM_DEFAULT_Q
  1160. #define QGEMM_R QGEMM_DEFAULT_R
  1161. #define QGEMM_UNROLL_M QGEMM_DEFAULT_UNROLL_M
  1162. #define QGEMM_UNROLL_N QGEMM_DEFAULT_UNROLL_N
  1163. #define QGEMM_UNROLL_MN MAX((QGEMM_UNROLL_M), (QGEMM_UNROLL_N))
  1164. #define CGEMM_P CGEMM_DEFAULT_P
  1165. #define CGEMM_Q CGEMM_DEFAULT_Q
  1166. #define CGEMM_R CGEMM_DEFAULT_R
  1167. #define CGEMM_UNROLL_M CGEMM_DEFAULT_UNROLL_M
  1168. #define CGEMM_UNROLL_N CGEMM_DEFAULT_UNROLL_N
  1169. #ifdef CGEMM_DEFAULT_UNROLL_MN
  1170. #define CGEMM_UNROLL_MN CGEMM_DEFAULT_UNROLL_MN
  1171. #else
  1172. #define CGEMM_UNROLL_MN MAX((CGEMM_UNROLL_M), (CGEMM_UNROLL_N))
  1173. #endif
  1174. #define ZGEMM_P ZGEMM_DEFAULT_P
  1175. #define ZGEMM_Q ZGEMM_DEFAULT_Q
  1176. #define ZGEMM_R ZGEMM_DEFAULT_R
  1177. #define ZGEMM_UNROLL_M ZGEMM_DEFAULT_UNROLL_M
  1178. #define ZGEMM_UNROLL_N ZGEMM_DEFAULT_UNROLL_N
  1179. #ifdef ZGEMM_DEFAULT_UNROLL_MN
  1180. #define ZGEMM_UNROLL_MN ZGEMM_DEFAULT_UNROLL_MN
  1181. #else
  1182. #define ZGEMM_UNROLL_MN MAX((ZGEMM_UNROLL_M), (ZGEMM_UNROLL_N))
  1183. #endif
  1184. #define XGEMM_P XGEMM_DEFAULT_P
  1185. #define XGEMM_Q XGEMM_DEFAULT_Q
  1186. #define XGEMM_R XGEMM_DEFAULT_R
  1187. #define XGEMM_UNROLL_M XGEMM_DEFAULT_UNROLL_M
  1188. #define XGEMM_UNROLL_N XGEMM_DEFAULT_UNROLL_N
  1189. #define XGEMM_UNROLL_MN MAX((XGEMM_UNROLL_M), (XGEMM_UNROLL_N))
  1190. #ifdef CGEMM3M_DEFAULT_UNROLL_N
  1191. #define CGEMM3M_P CGEMM3M_DEFAULT_P
  1192. #define CGEMM3M_Q CGEMM3M_DEFAULT_Q
  1193. #define CGEMM3M_R CGEMM3M_DEFAULT_R
  1194. #define CGEMM3M_UNROLL_M CGEMM3M_DEFAULT_UNROLL_M
  1195. #define CGEMM3M_UNROLL_N CGEMM3M_DEFAULT_UNROLL_N
  1196. #define CGEMM3M_UNROLL_MN MAX((CGEMM3M_UNROLL_M), (CGEMM3M_UNROLL_N))
  1197. #else
  1198. #define CGEMM3M_P SGEMM_DEFAULT_P
  1199. #define CGEMM3M_Q SGEMM_DEFAULT_Q
  1200. #define CGEMM3M_R SGEMM_DEFAULT_R
  1201. #define CGEMM3M_UNROLL_M SGEMM_DEFAULT_UNROLL_M
  1202. #define CGEMM3M_UNROLL_N SGEMM_DEFAULT_UNROLL_N
  1203. #define CGEMM3M_UNROLL_MN MAX((CGEMM_UNROLL_M), (CGEMM_UNROLL_N))
  1204. #endif
  1205. #ifdef ZGEMM3M_DEFAULT_UNROLL_N
  1206. #define ZGEMM3M_P ZGEMM3M_DEFAULT_P
  1207. #define ZGEMM3M_Q ZGEMM3M_DEFAULT_Q
  1208. #define ZGEMM3M_R ZGEMM3M_DEFAULT_R
  1209. #define ZGEMM3M_UNROLL_M ZGEMM3M_DEFAULT_UNROLL_M
  1210. #define ZGEMM3M_UNROLL_N ZGEMM3M_DEFAULT_UNROLL_N
  1211. #define ZGEMM3M_UNROLL_MN MAX((ZGEMM_UNROLL_M), (ZGEMM_UNROLL_N))
  1212. #else
  1213. #define ZGEMM3M_P DGEMM_DEFAULT_P
  1214. #define ZGEMM3M_Q DGEMM_DEFAULT_Q
  1215. #define ZGEMM3M_R DGEMM_DEFAULT_R
  1216. #define ZGEMM3M_UNROLL_M DGEMM_DEFAULT_UNROLL_M
  1217. #define ZGEMM3M_UNROLL_N DGEMM_DEFAULT_UNROLL_N
  1218. #define ZGEMM3M_UNROLL_MN MAX((ZGEMM_UNROLL_M), (ZGEMM_UNROLL_N))
  1219. #endif
  1220. #define XGEMM3M_P QGEMM_DEFAULT_P
  1221. #define XGEMM3M_Q QGEMM_DEFAULT_Q
  1222. #define XGEMM3M_R QGEMM_DEFAULT_R
  1223. #define XGEMM3M_UNROLL_M QGEMM_DEFAULT_UNROLL_M
  1224. #define XGEMM3M_UNROLL_N QGEMM_DEFAULT_UNROLL_N
  1225. #define XGEMM3M_UNROLL_MN MAX((QGEMM_UNROLL_M), (QGEMM_UNROLL_N))
  1226. #endif
  1227. #endif
  1228. #ifndef COMPLEX
  1229. #if defined(XDOUBLE)
  1230. #define GEMM_P QGEMM_P
  1231. #define GEMM_Q QGEMM_Q
  1232. #define GEMM_R QGEMM_R
  1233. #define GEMM_UNROLL_M QGEMM_UNROLL_M
  1234. #define GEMM_UNROLL_N QGEMM_UNROLL_N
  1235. #define GEMM_UNROLL_MN QGEMM_UNROLL_MN
  1236. #define GEMM_DEFAULT_P QGEMM_DEFAULT_P
  1237. #define GEMM_DEFAULT_Q QGEMM_DEFAULT_Q
  1238. #define GEMM_DEFAULT_R QGEMM_DEFAULT_R
  1239. #define GEMM_DEFAULT_UNROLL_M QGEMM_DEFAULT_UNROLL_M
  1240. #define GEMM_DEFAULT_UNROLL_N QGEMM_DEFAULT_UNROLL_N
  1241. #elif defined(DOUBLE)
  1242. #define GEMM_P DGEMM_P
  1243. #define GEMM_Q DGEMM_Q
  1244. #define GEMM_R DGEMM_R
  1245. #define GEMM_UNROLL_M DGEMM_UNROLL_M
  1246. #define GEMM_UNROLL_N DGEMM_UNROLL_N
  1247. #define GEMM_UNROLL_MN DGEMM_UNROLL_MN
  1248. #define GEMM_DEFAULT_P DGEMM_DEFAULT_P
  1249. #define GEMM_DEFAULT_Q DGEMM_DEFAULT_Q
  1250. #define GEMM_DEFAULT_R DGEMM_DEFAULT_R
  1251. #define GEMM_DEFAULT_UNROLL_M DGEMM_DEFAULT_UNROLL_M
  1252. #define GEMM_DEFAULT_UNROLL_N DGEMM_DEFAULT_UNROLL_N
  1253. #elif defined(BFLOAT16)
  1254. #define GEMM_P SBGEMM_P
  1255. #define GEMM_Q SBGEMM_Q
  1256. #define GEMM_R SBGEMM_R
  1257. #define GEMM_UNROLL_M SBGEMM_UNROLL_M
  1258. #define GEMM_UNROLL_N SBGEMM_UNROLL_N
  1259. #define GEMM_UNROLL_MN SBGEMM_UNROLL_MN
  1260. #define GEMM_DEFAULT_P SBGEMM_DEFAULT_P
  1261. #define GEMM_DEFAULT_Q SBGEMM_DEFAULT_Q
  1262. #define GEMM_DEFAULT_R SBGEMM_DEFAULT_R
  1263. #define GEMM_DEFAULT_UNROLL_M SBGEMM_DEFAULT_UNROLL_M
  1264. #define GEMM_DEFAULT_UNROLL_N SBGEMM_DEFAULT_UNROLL_N
  1265. #else
  1266. #define GEMM_P SGEMM_P
  1267. #define GEMM_Q SGEMM_Q
  1268. #define GEMM_R SGEMM_R
  1269. #define GEMM_UNROLL_M SGEMM_UNROLL_M
  1270. #define GEMM_UNROLL_N SGEMM_UNROLL_N
  1271. #define GEMM_UNROLL_MN SGEMM_UNROLL_MN
  1272. #define GEMM_DEFAULT_P SGEMM_DEFAULT_P
  1273. #define GEMM_DEFAULT_Q SGEMM_DEFAULT_Q
  1274. #define GEMM_DEFAULT_R SGEMM_DEFAULT_R
  1275. #define GEMM_DEFAULT_UNROLL_M SGEMM_DEFAULT_UNROLL_M
  1276. #define GEMM_DEFAULT_UNROLL_N SGEMM_DEFAULT_UNROLL_N
  1277. #endif
  1278. #else
  1279. #if defined(XDOUBLE)
  1280. #define GEMM_P XGEMM_P
  1281. #define GEMM_Q XGEMM_Q
  1282. #define GEMM_R XGEMM_R
  1283. #define GEMM_UNROLL_M XGEMM_UNROLL_M
  1284. #define GEMM_UNROLL_N XGEMM_UNROLL_N
  1285. #define GEMM_UNROLL_MN XGEMM_UNROLL_MN
  1286. #define GEMM_DEFAULT_P XGEMM_DEFAULT_P
  1287. #define GEMM_DEFAULT_Q XGEMM_DEFAULT_Q
  1288. #define GEMM_DEFAULT_R XGEMM_DEFAULT_R
  1289. #define GEMM_DEFAULT_UNROLL_M XGEMM_DEFAULT_UNROLL_M
  1290. #define GEMM_DEFAULT_UNROLL_N XGEMM_DEFAULT_UNROLL_N
  1291. #elif defined(DOUBLE)
  1292. #define GEMM_P ZGEMM_P
  1293. #define GEMM_Q ZGEMM_Q
  1294. #define GEMM_R ZGEMM_R
  1295. #define GEMM_UNROLL_M ZGEMM_UNROLL_M
  1296. #define GEMM_UNROLL_N ZGEMM_UNROLL_N
  1297. #define GEMM_UNROLL_MN ZGEMM_UNROLL_MN
  1298. #define GEMM_DEFAULT_P ZGEMM_DEFAULT_P
  1299. #define GEMM_DEFAULT_Q ZGEMM_DEFAULT_Q
  1300. #define GEMM_DEFAULT_R ZGEMM_DEFAULT_R
  1301. #define GEMM_DEFAULT_UNROLL_M ZGEMM_DEFAULT_UNROLL_M
  1302. #define GEMM_DEFAULT_UNROLL_N ZGEMM_DEFAULT_UNROLL_N
  1303. #else
  1304. #define GEMM_P CGEMM_P
  1305. #define GEMM_Q CGEMM_Q
  1306. #define GEMM_R CGEMM_R
  1307. #define GEMM_UNROLL_M CGEMM_UNROLL_M
  1308. #define GEMM_UNROLL_N CGEMM_UNROLL_N
  1309. #define GEMM_UNROLL_MN CGEMM_UNROLL_MN
  1310. #define GEMM_DEFAULT_P CGEMM_DEFAULT_P
  1311. #define GEMM_DEFAULT_Q CGEMM_DEFAULT_Q
  1312. #define GEMM_DEFAULT_R CGEMM_DEFAULT_R
  1313. #define GEMM_DEFAULT_UNROLL_M CGEMM_DEFAULT_UNROLL_M
  1314. #define GEMM_DEFAULT_UNROLL_N CGEMM_DEFAULT_UNROLL_N
  1315. #endif
  1316. #endif
  1317. #ifdef XDOUBLE
  1318. #define GEMM3M_UNROLL_M XGEMM3M_UNROLL_M
  1319. #define GEMM3M_UNROLL_N XGEMM3M_UNROLL_N
  1320. #elif defined(DOUBLE)
  1321. #define GEMM3M_UNROLL_M ZGEMM3M_UNROLL_M
  1322. #define GEMM3M_UNROLL_N ZGEMM3M_UNROLL_N
  1323. #else
  1324. #define GEMM3M_UNROLL_M CGEMM3M_UNROLL_M
  1325. #define GEMM3M_UNROLL_N CGEMM3M_UNROLL_N
  1326. #endif
  1327. #ifndef QGEMM_DEFAULT_UNROLL_M
  1328. #define QGEMM_DEFAULT_UNROLL_M 2
  1329. #endif
  1330. #ifndef QGEMM_DEFAULT_UNROLL_N
  1331. #define QGEMM_DEFAULT_UNROLL_N 2
  1332. #endif
  1333. #ifndef XGEMM_DEFAULT_UNROLL_M
  1334. #define XGEMM_DEFAULT_UNROLL_M 2
  1335. #endif
  1336. #ifndef XGEMM_DEFAULT_UNROLL_N
  1337. #define XGEMM_DEFAULT_UNROLL_N 2
  1338. #endif
  1339. #ifndef GEMM_THREAD
  1340. #define GEMM_THREAD gemm_thread_n
  1341. #endif
  1342. #ifndef SBGEMM_DEFAULT_R
  1343. #define SBGEMM_DEFAULT_R (((BUFFER_SIZE - ((SBGEMM_DEFAULT_P * SBGEMM_DEFAULT_Q * 4 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (SBGEMM_DEFAULT_Q * 4) - 15) & ~15UL)
  1344. #endif
  1345. #ifndef SGEMM_DEFAULT_R
  1346. #define SGEMM_DEFAULT_R (((BUFFER_SIZE - ((SGEMM_DEFAULT_P * SGEMM_DEFAULT_Q * 4 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (SGEMM_DEFAULT_Q * 4) - 15) & ~15UL)
  1347. #endif
  1348. #ifndef DGEMM_DEFAULT_R
  1349. #define DGEMM_DEFAULT_R (((BUFFER_SIZE - ((DGEMM_DEFAULT_P * DGEMM_DEFAULT_Q * 8 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (DGEMM_DEFAULT_Q * 8) - 15) & ~15UL)
  1350. #endif
  1351. #ifndef QGEMM_DEFAULT_R
  1352. #define QGEMM_DEFAULT_R (((BUFFER_SIZE - ((QGEMM_DEFAULT_P * QGEMM_DEFAULT_Q * 16 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (QGEMM_DEFAULT_Q * 16) - 15) & ~15UL)
  1353. #endif
  1354. #ifndef CGEMM_DEFAULT_R
  1355. #define CGEMM_DEFAULT_R (((BUFFER_SIZE - ((CGEMM_DEFAULT_P * CGEMM_DEFAULT_Q * 8 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (CGEMM_DEFAULT_Q * 8) - 15) & ~15UL)
  1356. #endif
  1357. #ifndef ZGEMM_DEFAULT_R
  1358. #define ZGEMM_DEFAULT_R (((BUFFER_SIZE - ((ZGEMM_DEFAULT_P * ZGEMM_DEFAULT_Q * 16 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (ZGEMM_DEFAULT_Q * 16) - 15) & ~15UL)
  1359. #endif
  1360. #ifndef XGEMM_DEFAULT_R
  1361. #define XGEMM_DEFAULT_R (((BUFFER_SIZE - ((XGEMM_DEFAULT_P * XGEMM_DEFAULT_Q * 32 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (XGEMM_DEFAULT_Q * 32) - 15) & ~15UL)
  1362. #endif
  1363. #ifndef SNUMOPT
  1364. #define SNUMOPT 2
  1365. #endif
  1366. #ifndef DNUMOPT
  1367. #define DNUMOPT 2
  1368. #endif
  1369. #ifndef QNUMOPT
  1370. #define QNUMOPT 1
  1371. #endif
  1372. #ifndef GEMM3M_P
  1373. #ifdef XDOUBLE
  1374. #define GEMM3M_P XGEMM3M_P
  1375. #elif defined(DOUBLE)
  1376. #define GEMM3M_P ZGEMM3M_P
  1377. #else
  1378. #define GEMM3M_P CGEMM3M_P
  1379. #endif
  1380. #endif
  1381. #ifndef GEMM3M_Q
  1382. #ifdef XDOUBLE
  1383. #define GEMM3M_Q XGEMM3M_Q
  1384. #elif defined(DOUBLE)
  1385. #define GEMM3M_Q ZGEMM3M_Q
  1386. #else
  1387. #define GEMM3M_Q CGEMM3M_Q
  1388. #endif
  1389. #endif
  1390. #ifndef GEMM3M_R
  1391. #ifdef XDOUBLE
  1392. #define GEMM3M_R XGEMM3M_R
  1393. #elif defined(DOUBLE)
  1394. #define GEMM3M_R ZGEMM3M_R
  1395. #else
  1396. #define GEMM3M_R CGEMM3M_R
  1397. #endif
  1398. #endif
  1399. #endif