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 118 kB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772
  1. /*********************************************************************/
  2. /* Copyright 2009, 2010 The University of Texas at Austin. */
  3. /* Copyright 2023, 2025 The OpenBLAS Project. */
  4. /* All rights reserved. */
  5. /* */
  6. /* Redistribution and use in source and binary forms, with or */
  7. /* without modification, are permitted provided that the following */
  8. /* conditions are met: */
  9. /* */
  10. /* 1. Redistributions of source code must retain the above */
  11. /* copyright notice, this list of conditions and the following */
  12. /* disclaimer. */
  13. /* */
  14. /* 2. Redistributions in binary form must reproduce the above */
  15. /* copyright notice, this list of conditions and the following */
  16. /* disclaimer in the documentation and/or other materials */
  17. /* provided with the distribution. */
  18. /* */
  19. /* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
  20. /* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
  21. /* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
  22. /* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
  23. /* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
  24. /* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
  25. /* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
  26. /* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
  27. /* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
  28. /* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
  29. /* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
  30. /* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
  31. /* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
  32. /* POSSIBILITY OF SUCH DAMAGE. */
  33. /* */
  34. /* The views and conclusions contained in the software and */
  35. /* documentation are those of the authors and should not be */
  36. /* interpreted as representing official policies, either expressed */
  37. /* or implied, of The University of Texas at Austin. */
  38. /*********************************************************************/
  39. #ifndef COMMON_PARAM_H
  40. #define COMMON_PARAM_H
  41. #ifndef ASSEMBLER
  42. #ifdef DYNAMIC_ARCH
  43. typedef struct {
  44. int dtb_entries;
  45. int switch_ratio;
  46. int offsetA, offsetB, align;
  47. #if BUILD_HFLOAT16 == 1
  48. int shgemm_p, shgemm_q, shgemm_r;
  49. int shgemm_unroll_m, shgemm_unroll_n, shgemm_unroll_mn;
  50. int (*shgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, float, hfloat16 *, hfloat16 *, float *, BLASLONG);
  51. int (*shgemm_beta )(BLASLONG, BLASLONG, BLASLONG, float, hfloat16 *, BLASLONG, hfloat16 *, BLASLONG, float *, BLASLONG);
  52. int (*shgemm_incopy )(BLASLONG, BLASLONG, hfloat16 *, BLASLONG, hfloat16 *);
  53. int (*shgemm_itcopy )(BLASLONG, BLASLONG, hfloat16 *, BLASLONG, hfloat16 *);
  54. int (*shgemm_oncopy )(BLASLONG, BLASLONG, hfloat16 *, BLASLONG, hfloat16 *);
  55. int (*shgemm_otcopy )(BLASLONG, BLASLONG, hfloat16 *, BLASLONG, hfloat16 *);
  56. #endif
  57. #if BUILD_BFLOAT16 == 1
  58. int bgemm_p, bgemm_q, bgemm_r;
  59. int bgemm_unroll_m, bgemm_unroll_n, bgemm_unroll_mn;
  60. int bgemm_align_k;
  61. int sbgemm_p, sbgemm_q, sbgemm_r;
  62. int sbgemm_unroll_m, sbgemm_unroll_n, sbgemm_unroll_mn;
  63. int sbgemm_align_k;
  64. int need_amxtile_permission; // 0 default, 1 for device support amx.
  65. void (*sbstobf16_k) (BLASLONG, float *, BLASLONG, bfloat16 *, BLASLONG);
  66. void (*sbdtobf16_k) (BLASLONG, double *, BLASLONG, bfloat16 *, BLASLONG);
  67. void (*sbf16tos_k) (BLASLONG, bfloat16 *, BLASLONG, float *, BLASLONG);
  68. void (*dbf16tod_k) (BLASLONG, bfloat16 *, BLASLONG, double *, BLASLONG);
  69. float (*sbamax_k) (BLASLONG, float *, BLASLONG);
  70. float (*sbamin_k) (BLASLONG, float *, BLASLONG);
  71. float (*sbmax_k) (BLASLONG, float *, BLASLONG);
  72. float (*sbmin_k) (BLASLONG, float *, BLASLONG);
  73. BLASLONG (*isbamax_k)(BLASLONG, float *, BLASLONG);
  74. BLASLONG (*isbamin_k)(BLASLONG, float *, BLASLONG);
  75. BLASLONG (*isbmax_k) (BLASLONG, float *, BLASLONG);
  76. BLASLONG (*isbmin_k) (BLASLONG, float *, BLASLONG);
  77. float (*sbnrm2_k) (BLASLONG, float *, BLASLONG);
  78. float (*sbasum_k) (BLASLONG, float *, BLASLONG);
  79. float (*sbsum_k) (BLASLONG, float *, BLASLONG);
  80. int (*sbcopy_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  81. float (*sbdot_k) (BLASLONG, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG);
  82. double (*dsbdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  83. int (*sbrot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG, float, float);
  84. int (*sbrotm_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  85. int (*bscal_k) (BLASLONG, BLASLONG, BLASLONG, bfloat16, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG);
  86. int (*sbaxpy_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  87. int (*sbscal_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  88. int (*sbswap_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  89. int (*bgemv_n) (BLASLONG, BLASLONG, bfloat16, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, bfloat16, bfloat16 *, BLASLONG);
  90. int (*bgemv_t) (BLASLONG, BLASLONG, bfloat16, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, bfloat16, bfloat16 *, BLASLONG);
  91. int (*sbgemv_n) (BLASLONG, BLASLONG, float, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, float, float *, BLASLONG);
  92. int (*sbgemv_t) (BLASLONG, BLASLONG, float, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, float, float *, BLASLONG);
  93. int (*sbger_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  94. int (*sbsymv_L) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  95. int (*sbsymv_U) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  96. int (*bgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, bfloat16, bfloat16 *, bfloat16 *, bfloat16 *, BLASLONG);
  97. int (*bgemm_beta )(BLASLONG, BLASLONG, BLASLONG, bfloat16, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG);
  98. int (*bgemm_incopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
  99. int (*bgemm_itcopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
  100. int (*bgemm_oncopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
  101. int (*bgemm_otcopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
  102. int (*sbgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, float, bfloat16 *, bfloat16 *, float *, BLASLONG);
  103. int (*sbgemm_beta )(BLASLONG, BLASLONG, BLASLONG, float, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, float *, BLASLONG);
  104. int (*sbgemm_incopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
  105. int (*sbgemm_itcopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
  106. int (*sbgemm_oncopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
  107. int (*sbgemm_otcopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
  108. int (*sbtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  109. int (*sbtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  110. int (*sbtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  111. int (*sbtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  112. int (*sbtrsm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  113. int (*sbtrsm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  114. int (*sbtrsm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  115. int (*sbtrsm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  116. int (*sbtrsm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  117. int (*sbtrsm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  118. int (*sbtrsm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  119. int (*sbtrsm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  120. int (*sbtrsm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  121. int (*sbtrsm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  122. int (*sbtrsm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  123. int (*sbtrsm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  124. int (*sbtrsm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  125. int (*sbtrsm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  126. int (*sbtrsm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  127. int (*sbtrsm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  128. int (*sbtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  129. int (*sbtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  130. int (*sbtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  131. int (*sbtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  132. int (*sbtrmm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  133. int (*sbtrmm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  134. int (*sbtrmm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  135. int (*sbtrmm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  136. int (*sbtrmm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  137. int (*sbtrmm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  138. int (*sbtrmm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  139. int (*sbtrmm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  140. int (*sbtrmm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  141. int (*sbtrmm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  142. int (*sbtrmm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  143. int (*sbtrmm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  144. int (*sbtrmm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  145. int (*sbtrmm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  146. int (*sbtrmm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  147. int (*sbtrmm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  148. int (*sbsymm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  149. int (*sbsymm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  150. int (*sbsymm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  151. int (*sbsymm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  152. int (*sbneg_tcopy) (BLASLONG, BLASLONG, float *, BLASLONG, float *);
  153. int (*sblaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG, blasint *, float *);
  154. #ifdef SMALL_MATRIX_OPT
  155. int (*sbgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float beta);
  156. 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);
  157. 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);
  158. 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);
  159. 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);
  160. 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);
  161. 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);
  162. 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);
  163. 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);
  164. #endif
  165. #endif
  166. #if (BUILD_SINGLE == 1) || (BUILD_DOUBLE == 1) || (BUILD_COMPLEX == 1) || (BUILD_COMPLEX16 == 1)
  167. int sgemm_p, sgemm_q, sgemm_r;
  168. int sgemm_unroll_m, sgemm_unroll_n, sgemm_unroll_mn;
  169. #endif
  170. int exclusive_cache;
  171. #if (BUILD_SINGLE == 1) || (BUILD_COMPLEX == 1)
  172. float (*samax_k) (BLASLONG, float *, BLASLONG);
  173. float (*samin_k) (BLASLONG, float *, BLASLONG);
  174. float (*smax_k) (BLASLONG, float *, BLASLONG);
  175. float (*smin_k) (BLASLONG, float *, BLASLONG);
  176. #endif
  177. #if (BUILD_SINGLE==1) || (BUILD_DOUBLE ==1) || (BUILD_COMPLEX==1)
  178. BLASLONG (*isamax_k)(BLASLONG, float *, BLASLONG);
  179. #endif
  180. #if (BUILD_SINGLE==1) || (BUILD_COMPLEX==1)
  181. BLASLONG (*isamin_k)(BLASLONG, float *, BLASLONG);
  182. BLASLONG (*ismax_k) (BLASLONG, float *, BLASLONG);
  183. BLASLONG (*ismin_k) (BLASLONG, float *, BLASLONG);
  184. float (*snrm2_k) (BLASLONG, float *, BLASLONG);
  185. float (*sasum_k) (BLASLONG, float *, BLASLONG);
  186. #endif
  187. #if (BUILD_SINGLE==1)
  188. float (*ssum_k) (BLASLONG, float *, BLASLONG);
  189. #endif
  190. #if (BUILD_SINGLE==1) || (BUILD_DOUBLE==1) || (BUILD_COMPLEX==1)
  191. int (*scopy_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  192. float (*sdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  193. //double (*dsdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  194. int (*srot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG, float, float);
  195. int (*srotm_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  196. #endif
  197. #if (BUILD_SINGLE==1) || (BUILD_DOUBLE==1) || (BUILD_COMPLEX==1)
  198. int (*saxpy_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  199. #endif
  200. #if (BUILD_SINGLE==1) || (BUILD_DOUBLE==1) || (BUILD_COMPLEX==1) || (BUILD_COMPLEX16==1)
  201. int (*sscal_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  202. #endif
  203. #if (BUILD_SINGLE==1) || (BUILD_DOUBLE==1) || (BUILD_COMPLEX==1)
  204. int (*sswap_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  205. int (*sgemv_n) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  206. int (*sgemv_t) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  207. #endif
  208. #if (BUILD_SINGLE==1)
  209. int (*sger_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  210. int (*ssymv_L) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  211. int (*ssymv_U) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  212. #endif
  213. #if (BUILD_SINGLE==1) || (BUILD_DOUBLE==1) || (BUILD_COMPLEX==1)
  214. #ifdef ARCH_X86_64
  215. void (*sgemm_direct) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG , float *, BLASLONG , float * , BLASLONG);
  216. int (*sgemm_direct_performant) (BLASLONG M, BLASLONG N, BLASLONG K);
  217. #endif
  218. #ifdef ARCH_ARM64
  219. void (*sgemm_direct) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG , float *, BLASLONG , float * , BLASLONG);
  220. #endif
  221. int (*sgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG);
  222. int (*sgemm_beta )(BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  223. int (*sgemm_incopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  224. int (*sgemm_itcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  225. int (*sgemm_oncopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  226. int (*sgemm_otcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  227. #endif
  228. #if (BUILD_SINGLE==1) || (BUILD_DOUBLE==1) || (BUILD_COMPLEX==1)
  229. #ifdef SMALL_MATRIX_OPT
  230. int (*sgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float beta);
  231. 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);
  232. 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);
  233. 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);
  234. 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);
  235. 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);
  236. 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);
  237. 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);
  238. 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);
  239. #endif
  240. int (*strsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  241. int (*strsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  242. int (*strsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  243. int (*strsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  244. int (*strsm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  245. int (*strsm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  246. int (*strsm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  247. int (*strsm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  248. int (*strsm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  249. int (*strsm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  250. int (*strsm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  251. int (*strsm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  252. int (*strsm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  253. int (*strsm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  254. int (*strsm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  255. int (*strsm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  256. int (*strsm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  257. int (*strsm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  258. int (*strsm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  259. int (*strsm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  260. #endif
  261. #if (BUILD_SINGLE==1)
  262. int (*strmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  263. int (*strmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  264. int (*strmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  265. int (*strmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
  266. int (*strmm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  267. int (*strmm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  268. int (*strmm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  269. int (*strmm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  270. int (*strmm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  271. int (*strmm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  272. int (*strmm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  273. int (*strmm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  274. int (*strmm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  275. int (*strmm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  276. int (*strmm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  277. int (*strmm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  278. int (*strmm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  279. int (*strmm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  280. int (*strmm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  281. int (*strmm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  282. int (*ssymm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  283. int (*ssymm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  284. int (*ssymm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  285. int (*ssymm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  286. int (*sneg_tcopy) (BLASLONG, BLASLONG, float *, BLASLONG, float *);
  287. int (*slaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG, blasint *, float *);
  288. #endif
  289. #if (BUILD_DOUBLE==1) || (BUILD_COMPLEX16==1)
  290. int dgemm_p, dgemm_q, dgemm_r;
  291. int dgemm_unroll_m, dgemm_unroll_n, dgemm_unroll_mn;
  292. #endif
  293. #if (BUILD_DOUBLE==1) || (BUILD_COMPLEX16==1)
  294. double (*damax_k) (BLASLONG, double *, BLASLONG);
  295. double (*damin_k) (BLASLONG, double *, BLASLONG);
  296. double (*dmax_k) (BLASLONG, double *, BLASLONG);
  297. double (*dmin_k) (BLASLONG, double *, BLASLONG);
  298. BLASLONG (*idamax_k)(BLASLONG, double *, BLASLONG);
  299. BLASLONG (*idamin_k)(BLASLONG, double *, BLASLONG);
  300. BLASLONG (*idmax_k) (BLASLONG, double *, BLASLONG);
  301. BLASLONG (*idmin_k) (BLASLONG, double *, BLASLONG);
  302. double (*dnrm2_k) (BLASLONG, double *, BLASLONG);
  303. double (*dasum_k) (BLASLONG, double *, BLASLONG);
  304. #endif
  305. #if (BUILD_DOUBLE==1)
  306. double (*dsum_k) (BLASLONG, double *, BLASLONG);
  307. #endif
  308. #if (BUILD_DOUBLE==1) || (BUILD_COMPLEX16==1)
  309. int (*dcopy_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
  310. double (*ddot_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
  311. #endif
  312. #if (BUILD_SINGLE==1) || (BUILD_DOUBLE==1)
  313. double (*dsdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  314. #endif
  315. #if (BUILD_DOUBLE==1) || (BUILD_COMPLEX16==1)
  316. int (*drot_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG, double, double);
  317. int (*drotm_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  318. int (*daxpy_k) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  319. int (*dscal_k) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  320. int (*dswap_k) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  321. int (*dgemv_n) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  322. int (*dgemv_t) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  323. #endif
  324. #if (BUILD_DOUBLE==1)
  325. int (*dger_k) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  326. int (*dsymv_L) (BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  327. int (*dsymv_U) (BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  328. #endif
  329. #if (BUILD_DOUBLE==1) || (BUILD_COMPLEX16==1)
  330. int (*dgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG);
  331. int (*dgemm_beta )(BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  332. int (*dgemm_incopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  333. int (*dgemm_itcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  334. int (*dgemm_oncopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  335. int (*dgemm_otcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  336. #endif
  337. #if (BUILD_DOUBLE==1) || (BUILD_COMPLEX16==1)
  338. #ifdef SMALL_MATRIX_OPT
  339. int (*dgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, double alpha, double beta);
  340. 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);
  341. 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);
  342. 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);
  343. 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);
  344. 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);
  345. 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);
  346. 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);
  347. 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);
  348. #endif
  349. #endif
  350. #if (BUILD_DOUBLE==1)
  351. int (*dtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  352. int (*dtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  353. int (*dtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  354. int (*dtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  355. int (*dtrsm_iunucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  356. int (*dtrsm_iunncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  357. int (*dtrsm_iutucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  358. int (*dtrsm_iutncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  359. int (*dtrsm_ilnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  360. int (*dtrsm_ilnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  361. int (*dtrsm_iltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  362. int (*dtrsm_iltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  363. int (*dtrsm_ounucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  364. int (*dtrsm_ounncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  365. int (*dtrsm_outucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  366. int (*dtrsm_outncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  367. int (*dtrsm_olnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  368. int (*dtrsm_olnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  369. int (*dtrsm_oltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  370. int (*dtrsm_oltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  371. int (*dtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  372. int (*dtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  373. int (*dtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  374. int (*dtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
  375. int (*dtrmm_iunucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  376. int (*dtrmm_iunncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  377. int (*dtrmm_iutucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  378. int (*dtrmm_iutncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  379. int (*dtrmm_ilnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  380. int (*dtrmm_ilnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  381. int (*dtrmm_iltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  382. int (*dtrmm_iltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  383. int (*dtrmm_ounucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  384. int (*dtrmm_ounncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  385. int (*dtrmm_outucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  386. int (*dtrmm_outncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  387. int (*dtrmm_olnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  388. int (*dtrmm_olnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  389. int (*dtrmm_oltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  390. int (*dtrmm_oltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  391. int (*dsymm_iutcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  392. int (*dsymm_iltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  393. int (*dsymm_outcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  394. int (*dsymm_oltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  395. int (*dneg_tcopy) (BLASLONG, BLASLONG, double *, BLASLONG, double *);
  396. int (*dlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, double *, BLASLONG, blasint *, double *);
  397. #endif
  398. #ifdef EXPRECISION
  399. int qgemm_p, qgemm_q, qgemm_r;
  400. int qgemm_unroll_m, qgemm_unroll_n, qgemm_unroll_mn;
  401. xdouble (*qamax_k) (BLASLONG, xdouble *, BLASLONG);
  402. xdouble (*qamin_k) (BLASLONG, xdouble *, BLASLONG);
  403. xdouble (*qmax_k) (BLASLONG, xdouble *, BLASLONG);
  404. xdouble (*qmin_k) (BLASLONG, xdouble *, BLASLONG);
  405. BLASLONG (*iqamax_k)(BLASLONG, xdouble *, BLASLONG);
  406. BLASLONG (*iqamin_k)(BLASLONG, xdouble *, BLASLONG);
  407. BLASLONG (*iqmax_k) (BLASLONG, xdouble *, BLASLONG);
  408. BLASLONG (*iqmin_k) (BLASLONG, xdouble *, BLASLONG);
  409. xdouble (*qnrm2_k) (BLASLONG, xdouble *, BLASLONG);
  410. xdouble (*qasum_k) (BLASLONG, xdouble *, BLASLONG);
  411. xdouble (*qsum_k) (BLASLONG, xdouble *, BLASLONG);
  412. int (*qcopy_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  413. xdouble (*qdot_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  414. int (*qrot_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble);
  415. int (*qrotm_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  416. int (*qaxpy_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  417. int (*qscal_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  418. int (*qswap_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  419. int (*qgemv_n) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  420. int (*qgemv_t) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  421. int (*qger_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  422. int (*qsymv_L) (BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  423. int (*qsymv_U) (BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  424. int (*qgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  425. int (*qgemm_beta )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  426. int (*qgemm_incopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  427. int (*qgemm_itcopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  428. int (*qgemm_oncopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  429. int (*qgemm_otcopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  430. int (*qtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  431. int (*qtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  432. int (*qtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  433. int (*qtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  434. int (*qtrsm_iunucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  435. int (*qtrsm_iunncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  436. int (*qtrsm_iutucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  437. int (*qtrsm_iutncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  438. int (*qtrsm_ilnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  439. int (*qtrsm_ilnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  440. int (*qtrsm_iltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  441. int (*qtrsm_iltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  442. int (*qtrsm_ounucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  443. int (*qtrsm_ounncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  444. int (*qtrsm_outucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  445. int (*qtrsm_outncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  446. int (*qtrsm_olnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  447. int (*qtrsm_olnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  448. int (*qtrsm_oltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  449. int (*qtrsm_oltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  450. int (*qtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  451. int (*qtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  452. int (*qtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  453. int (*qtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  454. int (*qtrmm_iunucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  455. int (*qtrmm_iunncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  456. int (*qtrmm_iutucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  457. int (*qtrmm_iutncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  458. int (*qtrmm_ilnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  459. int (*qtrmm_ilnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  460. int (*qtrmm_iltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  461. int (*qtrmm_iltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  462. int (*qtrmm_ounucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  463. int (*qtrmm_ounncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  464. int (*qtrmm_outucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  465. int (*qtrmm_outncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  466. int (*qtrmm_olnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  467. int (*qtrmm_olnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  468. int (*qtrmm_oltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  469. int (*qtrmm_oltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  470. int (*qsymm_iutcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  471. int (*qsymm_iltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  472. int (*qsymm_outcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  473. int (*qsymm_oltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  474. int (*qneg_tcopy) (BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  475. int (*qlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, xdouble *, BLASLONG, blasint *, xdouble *);
  476. #endif
  477. #if (BUILD_COMPLEX==1)
  478. int cgemm_p, cgemm_q, cgemm_r;
  479. int cgemm_unroll_m, cgemm_unroll_n, cgemm_unroll_mn;
  480. float (*camax_k) (BLASLONG, float *, BLASLONG);
  481. float (*camin_k) (BLASLONG, float *, BLASLONG);
  482. BLASLONG (*icamax_k)(BLASLONG, float *, BLASLONG);
  483. BLASLONG (*icamin_k)(BLASLONG, float *, BLASLONG);
  484. float (*cnrm2_k) (BLASLONG, float *, BLASLONG);
  485. float (*casum_k) (BLASLONG, float *, BLASLONG);
  486. float (*csum_k) (BLASLONG, float *, BLASLONG);
  487. int (*ccopy_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  488. openblas_complex_float (*cdotu_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  489. openblas_complex_float (*cdotc_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
  490. int (*csrot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG, float, float);
  491. int (*caxpy_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  492. int (*caxpyc_k)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  493. int (*cscal_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  494. int (*cswap_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  495. int (*cgemv_n) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  496. int (*cgemv_t) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  497. int (*cgemv_r) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  498. int (*cgemv_c) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  499. int (*cgemv_o) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  500. int (*cgemv_u) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  501. int (*cgemv_s) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  502. int (*cgemv_d) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  503. int (*cgeru_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  504. int (*cgerc_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  505. int (*cgerv_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  506. int (*cgerd_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  507. int (*csymv_L) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  508. int (*csymv_U) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  509. int (*chemv_L) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  510. int (*chemv_U) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  511. int (*chemv_M) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  512. int (*chemv_V) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
  513. int (*cgemm_kernel_n )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  514. int (*cgemm_kernel_l )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  515. int (*cgemm_kernel_r )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  516. int (*cgemm_kernel_b )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  517. int (*cgemm_beta )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
  518. int (*cgemm_incopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  519. int (*cgemm_itcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  520. int (*cgemm_oncopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  521. int (*cgemm_otcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  522. #ifdef SMALL_MATRIX_OPT
  523. int (*cgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, float alpha0, float alpha1, float beta0, float beta1);
  524. 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);
  525. 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);
  526. 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);
  527. 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);
  528. 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);
  529. 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);
  530. 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);
  531. 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);
  532. 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);
  533. 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);
  534. 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);
  535. 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);
  536. 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);
  537. 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);
  538. 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);
  539. 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);
  540. 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);
  541. 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);
  542. 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);
  543. 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);
  544. 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);
  545. 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);
  546. 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);
  547. 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);
  548. 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);
  549. 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);
  550. 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);
  551. 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);
  552. 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);
  553. 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);
  554. 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);
  555. 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);
  556. #endif
  557. int (*ctrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  558. int (*ctrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  559. int (*ctrsm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  560. int (*ctrsm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  561. int (*ctrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  562. int (*ctrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  563. int (*ctrsm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  564. int (*ctrsm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  565. int (*ctrsm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  566. int (*ctrsm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  567. int (*ctrsm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  568. int (*ctrsm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  569. int (*ctrsm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  570. int (*ctrsm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  571. int (*ctrsm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  572. int (*ctrsm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  573. int (*ctrsm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  574. int (*ctrsm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  575. int (*ctrsm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  576. int (*ctrsm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  577. int (*ctrsm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  578. int (*ctrsm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  579. int (*ctrsm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  580. int (*ctrsm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
  581. int (*ctrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  582. int (*ctrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  583. int (*ctrmm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  584. int (*ctrmm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  585. int (*ctrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  586. int (*ctrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  587. int (*ctrmm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  588. int (*ctrmm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
  589. int (*ctrmm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  590. int (*ctrmm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  591. int (*ctrmm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  592. int (*ctrmm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  593. int (*ctrmm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  594. int (*ctrmm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  595. int (*ctrmm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  596. int (*ctrmm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  597. int (*ctrmm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  598. int (*ctrmm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  599. int (*ctrmm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  600. int (*ctrmm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  601. int (*ctrmm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  602. int (*ctrmm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  603. int (*ctrmm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  604. int (*ctrmm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  605. int (*csymm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  606. int (*csymm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  607. int (*csymm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  608. int (*csymm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  609. int (*chemm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  610. int (*chemm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  611. int (*chemm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  612. int (*chemm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  613. int cgemm3m_p, cgemm3m_q, cgemm3m_r;
  614. int cgemm3m_unroll_m, cgemm3m_unroll_n, cgemm3m_unroll_mn;
  615. int (*cgemm3m_kernel)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
  616. int (*cgemm3m_incopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  617. int (*cgemm3m_incopyr)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  618. int (*cgemm3m_incopyi)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  619. int (*cgemm3m_itcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  620. int (*cgemm3m_itcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  621. int (*cgemm3m_itcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
  622. int (*cgemm3m_oncopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  623. int (*cgemm3m_oncopyr)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  624. int (*cgemm3m_oncopyi)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  625. int (*cgemm3m_otcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  626. int (*cgemm3m_otcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  627. int (*cgemm3m_otcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
  628. int (*csymm3m_iucopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  629. int (*csymm3m_ilcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  630. int (*csymm3m_iucopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  631. int (*csymm3m_ilcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  632. int (*csymm3m_iucopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  633. int (*csymm3m_ilcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  634. int (*csymm3m_oucopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  635. int (*csymm3m_olcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  636. int (*csymm3m_oucopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  637. int (*csymm3m_olcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  638. int (*csymm3m_oucopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  639. int (*csymm3m_olcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  640. int (*chemm3m_iucopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  641. int (*chemm3m_ilcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  642. int (*chemm3m_iucopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  643. int (*chemm3m_ilcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  644. int (*chemm3m_iucopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  645. int (*chemm3m_ilcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
  646. int (*chemm3m_oucopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  647. int (*chemm3m_olcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  648. int (*chemm3m_oucopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  649. int (*chemm3m_olcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  650. int (*chemm3m_oucopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  651. int (*chemm3m_olcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
  652. int (*cneg_tcopy) (BLASLONG, BLASLONG, float *, BLASLONG, float *);
  653. int (*claswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG, blasint *, float *);
  654. #endif
  655. #if (BUILD_COMPLEX16 == 1)
  656. int zgemm_p, zgemm_q, zgemm_r;
  657. int zgemm_unroll_m, zgemm_unroll_n, zgemm_unroll_mn;
  658. double (*zamax_k) (BLASLONG, double *, BLASLONG);
  659. double (*zamin_k) (BLASLONG, double *, BLASLONG);
  660. BLASLONG (*izamax_k)(BLASLONG, double *, BLASLONG);
  661. BLASLONG (*izamin_k)(BLASLONG, double *, BLASLONG);
  662. double (*znrm2_k) (BLASLONG, double *, BLASLONG);
  663. double (*zasum_k) (BLASLONG, double *, BLASLONG);
  664. double (*zsum_k) (BLASLONG, double *, BLASLONG);
  665. int (*zcopy_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
  666. openblas_complex_double (*zdotu_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
  667. openblas_complex_double (*zdotc_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
  668. int (*zdrot_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG, double, double);
  669. int (*zaxpy_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  670. int (*zaxpyc_k)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  671. int (*zscal_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  672. int (*zswap_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  673. int (*zgemv_n) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  674. int (*zgemv_t) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  675. int (*zgemv_r) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  676. int (*zgemv_c) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  677. int (*zgemv_o) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  678. int (*zgemv_u) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  679. int (*zgemv_s) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  680. int (*zgemv_d) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  681. int (*zgeru_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  682. int (*zgerc_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  683. int (*zgerv_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  684. int (*zgerd_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  685. int (*zsymv_L) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  686. int (*zsymv_U) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  687. int (*zhemv_L) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  688. int (*zhemv_U) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  689. int (*zhemv_M) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  690. int (*zhemv_V) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
  691. int (*zgemm_kernel_n )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  692. int (*zgemm_kernel_l )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  693. int (*zgemm_kernel_r )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  694. int (*zgemm_kernel_b )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  695. int (*zgemm_beta )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
  696. int (*zgemm_incopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  697. int (*zgemm_itcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  698. int (*zgemm_oncopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  699. int (*zgemm_otcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  700. #ifdef SMALL_MATRIX_OPT
  701. int (*zgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, double alpha0, double alpha1, double beta0, double beta1);
  702. 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);
  703. 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);
  704. 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);
  705. 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);
  706. 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);
  707. 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);
  708. 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);
  709. 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);
  710. 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);
  711. 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);
  712. 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);
  713. 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);
  714. 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);
  715. 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);
  716. 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);
  717. 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);
  718. 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);
  719. 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);
  720. 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);
  721. 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);
  722. 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);
  723. 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);
  724. 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);
  725. 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);
  726. 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);
  727. 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);
  728. 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);
  729. 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);
  730. 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);
  731. 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);
  732. 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);
  733. 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);
  734. #endif
  735. int (*ztrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  736. int (*ztrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  737. int (*ztrsm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  738. int (*ztrsm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  739. int (*ztrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  740. int (*ztrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  741. int (*ztrsm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  742. int (*ztrsm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  743. int (*ztrsm_iunucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  744. int (*ztrsm_iunncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  745. int (*ztrsm_iutucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  746. int (*ztrsm_iutncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  747. int (*ztrsm_ilnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  748. int (*ztrsm_ilnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  749. int (*ztrsm_iltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  750. int (*ztrsm_iltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  751. int (*ztrsm_ounucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  752. int (*ztrsm_ounncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  753. int (*ztrsm_outucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  754. int (*ztrsm_outncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  755. int (*ztrsm_olnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  756. int (*ztrsm_olnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  757. int (*ztrsm_oltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  758. int (*ztrsm_oltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
  759. int (*ztrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  760. int (*ztrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  761. int (*ztrmm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  762. int (*ztrmm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  763. int (*ztrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  764. int (*ztrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  765. int (*ztrmm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  766. int (*ztrmm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
  767. int (*ztrmm_iunucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  768. int (*ztrmm_iunncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  769. int (*ztrmm_iutucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  770. int (*ztrmm_iutncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  771. int (*ztrmm_ilnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  772. int (*ztrmm_ilnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  773. int (*ztrmm_iltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  774. int (*ztrmm_iltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  775. int (*ztrmm_ounucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  776. int (*ztrmm_ounncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  777. int (*ztrmm_outucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  778. int (*ztrmm_outncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  779. int (*ztrmm_olnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  780. int (*ztrmm_olnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  781. int (*ztrmm_oltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  782. int (*ztrmm_oltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  783. int (*zsymm_iutcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  784. int (*zsymm_iltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  785. int (*zsymm_outcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  786. int (*zsymm_oltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  787. int (*zhemm_iutcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  788. int (*zhemm_iltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  789. int (*zhemm_outcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  790. int (*zhemm_oltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  791. int zgemm3m_p, zgemm3m_q, zgemm3m_r;
  792. int zgemm3m_unroll_m, zgemm3m_unroll_n, zgemm3m_unroll_mn;
  793. int (*zgemm3m_kernel)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
  794. int (*zgemm3m_incopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  795. int (*zgemm3m_incopyr)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  796. int (*zgemm3m_incopyi)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  797. int (*zgemm3m_itcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  798. int (*zgemm3m_itcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  799. int (*zgemm3m_itcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
  800. int (*zgemm3m_oncopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  801. int (*zgemm3m_oncopyr)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  802. int (*zgemm3m_oncopyi)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  803. int (*zgemm3m_otcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  804. int (*zgemm3m_otcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  805. int (*zgemm3m_otcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
  806. int (*zsymm3m_iucopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  807. int (*zsymm3m_ilcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  808. int (*zsymm3m_iucopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  809. int (*zsymm3m_ilcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  810. int (*zsymm3m_iucopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  811. int (*zsymm3m_ilcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  812. int (*zsymm3m_oucopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  813. int (*zsymm3m_olcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  814. int (*zsymm3m_oucopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  815. int (*zsymm3m_olcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  816. int (*zsymm3m_oucopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  817. int (*zsymm3m_olcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  818. int (*zhemm3m_iucopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  819. int (*zhemm3m_ilcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  820. int (*zhemm3m_iucopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  821. int (*zhemm3m_ilcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  822. int (*zhemm3m_iucopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  823. int (*zhemm3m_ilcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
  824. int (*zhemm3m_oucopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  825. int (*zhemm3m_olcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  826. int (*zhemm3m_oucopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  827. int (*zhemm3m_olcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  828. int (*zhemm3m_oucopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  829. int (*zhemm3m_olcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
  830. int (*zneg_tcopy) (BLASLONG, BLASLONG, double *, BLASLONG, double *);
  831. int (*zlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, double *, BLASLONG, blasint *, double *);
  832. #endif
  833. #ifdef EXPRECISION
  834. int xgemm_p, xgemm_q, xgemm_r;
  835. int xgemm_unroll_m, xgemm_unroll_n, xgemm_unroll_mn;
  836. xdouble (*xamax_k) (BLASLONG, xdouble *, BLASLONG);
  837. xdouble (*xamin_k) (BLASLONG, xdouble *, BLASLONG);
  838. BLASLONG (*ixamax_k)(BLASLONG, xdouble *, BLASLONG);
  839. BLASLONG (*ixamin_k)(BLASLONG, xdouble *, BLASLONG);
  840. xdouble (*xnrm2_k) (BLASLONG, xdouble *, BLASLONG);
  841. xdouble (*xasum_k) (BLASLONG, xdouble *, BLASLONG);
  842. xdouble (*xsum_k) (BLASLONG, xdouble *, BLASLONG);
  843. int (*xcopy_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  844. openblas_complex_xdouble (*xdotu_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  845. openblas_complex_xdouble (*xdotc_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  846. int (*xqrot_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble);
  847. int (*xaxpy_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  848. int (*xaxpyc_k)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  849. int (*xscal_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  850. int (*xswap_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  851. int (*xgemv_n) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  852. int (*xgemv_t) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  853. int (*xgemv_r) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  854. int (*xgemv_c) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  855. int (*xgemv_o) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  856. int (*xgemv_u) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  857. int (*xgemv_s) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  858. int (*xgemv_d) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  859. int (*xgeru_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  860. int (*xgerc_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  861. int (*xgerv_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  862. int (*xgerd_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  863. int (*xsymv_L) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  864. int (*xsymv_U) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  865. int (*xhemv_L) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  866. int (*xhemv_U) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  867. int (*xhemv_M) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  868. int (*xhemv_V) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
  869. int (*xgemm_kernel_n )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  870. int (*xgemm_kernel_l )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  871. int (*xgemm_kernel_r )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  872. int (*xgemm_kernel_b )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  873. int (*xgemm_beta )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  874. int (*xgemm_incopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  875. int (*xgemm_itcopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  876. int (*xgemm_oncopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  877. int (*xgemm_otcopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  878. int (*xtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  879. int (*xtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  880. int (*xtrsm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  881. int (*xtrsm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  882. int (*xtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  883. int (*xtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  884. int (*xtrsm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  885. int (*xtrsm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  886. int (*xtrsm_iunucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  887. int (*xtrsm_iunncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  888. int (*xtrsm_iutucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  889. int (*xtrsm_iutncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  890. int (*xtrsm_ilnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  891. int (*xtrsm_ilnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  892. int (*xtrsm_iltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  893. int (*xtrsm_iltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  894. int (*xtrsm_ounucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  895. int (*xtrsm_ounncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  896. int (*xtrsm_outucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  897. int (*xtrsm_outncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  898. int (*xtrsm_olnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  899. int (*xtrsm_olnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  900. int (*xtrsm_oltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  901. int (*xtrsm_oltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
  902. int (*xtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  903. int (*xtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  904. int (*xtrmm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  905. int (*xtrmm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  906. int (*xtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  907. int (*xtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  908. int (*xtrmm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  909. int (*xtrmm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
  910. int (*xtrmm_iunucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  911. int (*xtrmm_iunncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  912. int (*xtrmm_iutucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  913. int (*xtrmm_iutncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  914. int (*xtrmm_ilnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  915. int (*xtrmm_ilnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  916. int (*xtrmm_iltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  917. int (*xtrmm_iltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  918. int (*xtrmm_ounucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  919. int (*xtrmm_ounncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  920. int (*xtrmm_outucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  921. int (*xtrmm_outncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  922. int (*xtrmm_olnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  923. int (*xtrmm_olnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  924. int (*xtrmm_oltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  925. int (*xtrmm_oltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  926. int (*xsymm_iutcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  927. int (*xsymm_iltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  928. int (*xsymm_outcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  929. int (*xsymm_oltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  930. int (*xhemm_iutcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  931. int (*xhemm_iltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  932. int (*xhemm_outcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  933. int (*xhemm_oltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  934. int xgemm3m_p, xgemm3m_q, xgemm3m_r;
  935. int xgemm3m_unroll_m, xgemm3m_unroll_n, xgemm3m_unroll_mn;
  936. int (*xgemm3m_kernel)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
  937. int (*xgemm3m_incopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  938. int (*xgemm3m_incopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  939. int (*xgemm3m_incopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  940. int (*xgemm3m_itcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  941. int (*xgemm3m_itcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  942. int (*xgemm3m_itcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  943. int (*xgemm3m_oncopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  944. int (*xgemm3m_oncopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  945. int (*xgemm3m_oncopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  946. int (*xgemm3m_otcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  947. int (*xgemm3m_otcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  948. int (*xgemm3m_otcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
  949. int (*xsymm3m_iucopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  950. int (*xsymm3m_ilcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  951. int (*xsymm3m_iucopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  952. int (*xsymm3m_ilcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  953. int (*xsymm3m_iucopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  954. int (*xsymm3m_ilcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  955. int (*xsymm3m_oucopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  956. int (*xsymm3m_olcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  957. int (*xsymm3m_oucopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  958. int (*xsymm3m_olcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  959. int (*xsymm3m_oucopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  960. int (*xsymm3m_olcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  961. int (*xhemm3m_iucopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  962. int (*xhemm3m_ilcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  963. int (*xhemm3m_iucopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  964. int (*xhemm3m_ilcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  965. int (*xhemm3m_iucopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  966. int (*xhemm3m_ilcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
  967. int (*xhemm3m_oucopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  968. int (*xhemm3m_olcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  969. int (*xhemm3m_oucopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  970. int (*xhemm3m_olcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  971. int (*xhemm3m_oucopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  972. int (*xhemm3m_olcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
  973. int (*xneg_tcopy) (BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
  974. int (*xlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, xdouble *, BLASLONG, blasint *, xdouble *);
  975. #endif
  976. void (*init)(void);
  977. int snum_opt, dnum_opt, qnum_opt;
  978. #if (BUILD_SINGLE==1)
  979. int (*saxpby_k) (BLASLONG, float, float*, BLASLONG,float, float*, BLASLONG);
  980. #endif
  981. #if (BUILD_DOUBLE==1)
  982. int (*daxpby_k) (BLASLONG, double, double*, BLASLONG,double, double*, BLASLONG);
  983. #endif
  984. #if (BUILD_COMPLEX==1)
  985. int (*caxpby_k) (BLASLONG, float, float, float*, BLASLONG,float,float, float*, BLASLONG);
  986. #endif
  987. #if (BUILD_COMPLEX16==1)
  988. int (*zaxpby_k) (BLASLONG, double, double, double*, BLASLONG,double,double, double*, BLASLONG);
  989. #endif
  990. #if (BUILD_SINGLE==1)
  991. int (*somatcopy_k_cn) (BLASLONG, BLASLONG, float, float*, BLASLONG, float*, BLASLONG);
  992. int (*somatcopy_k_ct) (BLASLONG, BLASLONG, float, float*, BLASLONG, float*, BLASLONG);
  993. int (*somatcopy_k_rn) (BLASLONG, BLASLONG, float, float*, BLASLONG, float*, BLASLONG);
  994. int (*somatcopy_k_rt) (BLASLONG, BLASLONG, float, float*, BLASLONG, float*, BLASLONG);
  995. #endif
  996. #if (BUILD_DOUBLE==1)
  997. int (*domatcopy_k_cn) (BLASLONG, BLASLONG, double, double*, BLASLONG, double*, BLASLONG);
  998. int (*domatcopy_k_ct) (BLASLONG, BLASLONG, double, double*, BLASLONG, double*, BLASLONG);
  999. int (*domatcopy_k_rn) (BLASLONG, BLASLONG, double, double*, BLASLONG, double*, BLASLONG);
  1000. int (*domatcopy_k_rt) (BLASLONG, BLASLONG, double, double*, BLASLONG, double*, BLASLONG);
  1001. #endif
  1002. #if (BUILD_COMPLEX==1)
  1003. int (*comatcopy_k_cn) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  1004. int (*comatcopy_k_ct) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  1005. int (*comatcopy_k_rn) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  1006. int (*comatcopy_k_rt) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  1007. int (*comatcopy_k_cnc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  1008. int (*comatcopy_k_ctc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  1009. int (*comatcopy_k_rnc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  1010. int (*comatcopy_k_rtc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
  1011. #endif
  1012. #if (BUILD_COMPLEX16==1)
  1013. int (*zomatcopy_k_cn) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  1014. int (*zomatcopy_k_ct) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  1015. int (*zomatcopy_k_rn) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  1016. int (*zomatcopy_k_rt) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  1017. int (*zomatcopy_k_cnc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  1018. int (*zomatcopy_k_ctc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  1019. int (*zomatcopy_k_rnc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  1020. int (*zomatcopy_k_rtc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
  1021. #endif
  1022. #if (BUILD_SINGLE==1)
  1023. int (*simatcopy_k_cn) (BLASLONG, BLASLONG, float, float*, BLASLONG);
  1024. int (*simatcopy_k_ct) (BLASLONG, BLASLONG, float, float*, BLASLONG);
  1025. int (*simatcopy_k_rn) (BLASLONG, BLASLONG, float, float*, BLASLONG);
  1026. int (*simatcopy_k_rt) (BLASLONG, BLASLONG, float, float*, BLASLONG);
  1027. #endif
  1028. #if (BUILD_DOUBLE==1)
  1029. int (*dimatcopy_k_cn) (BLASLONG, BLASLONG, double, double*, BLASLONG);
  1030. int (*dimatcopy_k_ct) (BLASLONG, BLASLONG, double, double*, BLASLONG);
  1031. int (*dimatcopy_k_rn) (BLASLONG, BLASLONG, double, double*, BLASLONG);
  1032. int (*dimatcopy_k_rt) (BLASLONG, BLASLONG, double, double*, BLASLONG);
  1033. #endif
  1034. #if (BUILD_COMPLEX==1)
  1035. int (*cimatcopy_k_cn) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  1036. int (*cimatcopy_k_ct) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  1037. int (*cimatcopy_k_rn) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  1038. int (*cimatcopy_k_rt) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  1039. int (*cimatcopy_k_cnc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  1040. int (*cimatcopy_k_ctc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  1041. int (*cimatcopy_k_rnc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  1042. int (*cimatcopy_k_rtc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
  1043. #endif
  1044. #if (BUILD_COMPLEX16==1)
  1045. int (*zimatcopy_k_cn) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1046. int (*zimatcopy_k_ct) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1047. int (*zimatcopy_k_rn) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1048. int (*zimatcopy_k_rt) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1049. int (*zimatcopy_k_cnc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1050. int (*zimatcopy_k_ctc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1051. int (*zimatcopy_k_rnc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1052. int (*zimatcopy_k_rtc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
  1053. #endif
  1054. #if (BUILD_SINGLE==1)
  1055. int (*sgeadd_k) (BLASLONG, BLASLONG, float, float *, BLASLONG, float, float *, BLASLONG);
  1056. #endif
  1057. #if (BUILD_DOUBLE==1)
  1058. int (*dgeadd_k) (BLASLONG, BLASLONG, double, double *, BLASLONG, double, double *, BLASLONG);
  1059. #endif
  1060. #if (BUILD_COMPLEX==1)
  1061. int (*cgeadd_k) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float, float, float *, BLASLONG);
  1062. #endif
  1063. #if (BUILD_COMPLEX16==1)
  1064. int (*zgeadd_k) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double, double, double *, BLASLONG);
  1065. #endif
  1066. } gotoblas_t;
  1067. extern gotoblas_t *gotoblas;
  1068. #define FUNC_OFFSET(func) (size_t)(&((gotoblas_t *)NULL)->func)
  1069. #define DTB_ENTRIES gotoblas -> dtb_entries
  1070. #define GEMM_OFFSET_A gotoblas -> offsetA
  1071. #define GEMM_OFFSET_B gotoblas -> offsetB
  1072. #define GEMM_ALIGN gotoblas -> align
  1073. #define HAVE_EX_L2 gotoblas -> exclusive_cache
  1074. #if (BUILD_HFLOAT16==1)
  1075. #define SHGEMM_P gotoblas -> shgemm_p
  1076. #define SHGEMM_Q gotoblas -> shgemm_q
  1077. #define SHGEMM_R gotoblas -> shgemm_r
  1078. #define SHGEMM_UNROLL_M gotoblas -> shgemm_unroll_m
  1079. #define SHGEMM_UNROLL_N gotoblas -> shgemm_unroll_n
  1080. #define SHGEMM_UNROLL_MN gotoblas -> shgemm_unroll_mn
  1081. #endif
  1082. #if (BUILD_BFLOAT16==1)
  1083. #define BGEMM_P gotoblas -> bgemm_p
  1084. #define BGEMM_Q gotoblas -> bgemm_q
  1085. #define BGEMM_R gotoblas -> bgemm_r
  1086. #define BGEMM_UNROLL_M gotoblas -> bgemm_unroll_m
  1087. #define BGEMM_UNROLL_N gotoblas -> bgemm_unroll_n
  1088. #define BGEMM_UNROLL_MN gotoblas -> bgemm_unroll_mn
  1089. #define SBGEMM_P gotoblas -> sbgemm_p
  1090. #define SBGEMM_Q gotoblas -> sbgemm_q
  1091. #define SBGEMM_R gotoblas -> sbgemm_r
  1092. #define SBGEMM_UNROLL_M gotoblas -> sbgemm_unroll_m
  1093. #define SBGEMM_UNROLL_N gotoblas -> sbgemm_unroll_n
  1094. #define SBGEMM_UNROLL_MN gotoblas -> sbgemm_unroll_mn
  1095. #endif
  1096. #if (BUILD_SINGLE==1)
  1097. #define SGEMM_P gotoblas -> sgemm_p
  1098. #define SGEMM_Q gotoblas -> sgemm_q
  1099. #define SGEMM_R gotoblas -> sgemm_r
  1100. #define SGEMM_UNROLL_M gotoblas -> sgemm_unroll_m
  1101. #define SGEMM_UNROLL_N gotoblas -> sgemm_unroll_n
  1102. #define SGEMM_UNROLL_MN gotoblas -> sgemm_unroll_mn
  1103. #endif
  1104. #if (BUILD_DOUBLE==1)
  1105. #define DGEMM_P gotoblas -> dgemm_p
  1106. #define DGEMM_Q gotoblas -> dgemm_q
  1107. #define DGEMM_R gotoblas -> dgemm_r
  1108. #define DGEMM_UNROLL_M gotoblas -> dgemm_unroll_m
  1109. #define DGEMM_UNROLL_N gotoblas -> dgemm_unroll_n
  1110. #define DGEMM_UNROLL_MN gotoblas -> dgemm_unroll_mn
  1111. #if (BUILD_SINGLE != 1)
  1112. #define SGEMM_P gotoblas -> sgemm_p
  1113. #define SGEMM_Q gotoblas -> sgemm_q
  1114. #define SGEMM_R 1024
  1115. #define SGEMM_UNROLL_M gotoblas -> sgemm_unroll_m
  1116. #define SGEMM_UNROLL_N gotoblas -> sgemm_unroll_n
  1117. #define SGEMM_UNROLL_MN gotoblas -> sgemm_unroll_mn
  1118. #endif
  1119. #endif
  1120. #define QGEMM_P gotoblas -> qgemm_p
  1121. #define QGEMM_Q gotoblas -> qgemm_q
  1122. #define QGEMM_R gotoblas -> qgemm_r
  1123. #define QGEMM_UNROLL_M gotoblas -> qgemm_unroll_m
  1124. #define QGEMM_UNROLL_N gotoblas -> qgemm_unroll_n
  1125. #define QGEMM_UNROLL_MN gotoblas -> qgemm_unroll_mn
  1126. #if (BUILD_COMPLEX==1)
  1127. #define CGEMM_P gotoblas -> cgemm_p
  1128. #define CGEMM_Q gotoblas -> cgemm_q
  1129. #define CGEMM_R gotoblas -> cgemm_r
  1130. #define CGEMM_UNROLL_M gotoblas -> cgemm_unroll_m
  1131. #define CGEMM_UNROLL_N gotoblas -> cgemm_unroll_n
  1132. #define CGEMM_UNROLL_MN gotoblas -> cgemm_unroll_mn
  1133. #if (BUILD_SINGLE != 1)
  1134. #define SGEMM_P gotoblas -> sgemm_p
  1135. #define SGEMM_Q gotoblas -> sgemm_q
  1136. #define SGEMM_R 1024
  1137. #define SGEMM_UNROLL_M gotoblas -> sgemm_unroll_m
  1138. #define SGEMM_UNROLL_N gotoblas -> sgemm_unroll_n
  1139. #define SGEMM_UNROLL_MN gotoblas -> sgemm_unroll_mn
  1140. #endif
  1141. #endif
  1142. #if (BUILD_COMPLEX16==1)
  1143. #define ZGEMM_P gotoblas -> zgemm_p
  1144. #define ZGEMM_Q gotoblas -> zgemm_q
  1145. #define ZGEMM_R gotoblas -> zgemm_r
  1146. #define ZGEMM_UNROLL_M gotoblas -> zgemm_unroll_m
  1147. #define ZGEMM_UNROLL_N gotoblas -> zgemm_unroll_n
  1148. #define ZGEMM_UNROLL_MN gotoblas -> zgemm_unroll_mn
  1149. #if (BUILD_DOUBLE != 1)
  1150. #define DGEMM_P gotoblas -> dgemm_p
  1151. #define DGEMM_Q gotoblas -> dgemm_q
  1152. #define DGEMM_R 1024
  1153. #define DGEMM_UNROLL_M gotoblas -> dgemm_unroll_m
  1154. #define DGEMM_UNROLL_N gotoblas -> dgemm_unroll_n
  1155. #define DGEMM_UNROLL_MN gotoblas -> dgemm_unroll_mn
  1156. #endif
  1157. #if (BUILD_COMPLEX != 1)
  1158. #define CGEMM_P gotoblas -> cgemm_p
  1159. #define CGEMM_Q gotoblas -> cgemm_q
  1160. #define CGEMM_R gotoblas -> cgemm_r
  1161. #define CGEMM_UNROLL_M gotoblas -> cgemm_unroll_m
  1162. #define CGEMM_UNROLL_N gotoblas -> cgemm_unroll_n
  1163. #define CGEMM_UNROLL_MN gotoblas -> cgemm_unroll_mn
  1164. #endif
  1165. #endif
  1166. #define XGEMM_P gotoblas -> xgemm_p
  1167. #define XGEMM_Q gotoblas -> xgemm_q
  1168. #define XGEMM_R gotoblas -> xgemm_r
  1169. #define XGEMM_UNROLL_M gotoblas -> xgemm_unroll_m
  1170. #define XGEMM_UNROLL_N gotoblas -> xgemm_unroll_n
  1171. #define XGEMM_UNROLL_MN gotoblas -> xgemm_unroll_mn
  1172. #define CGEMM3M_P gotoblas -> cgemm3m_p
  1173. #define CGEMM3M_Q gotoblas -> cgemm3m_q
  1174. #define CGEMM3M_R gotoblas -> cgemm3m_r
  1175. #define CGEMM3M_UNROLL_M gotoblas -> cgemm3m_unroll_m
  1176. #define CGEMM3M_UNROLL_N gotoblas -> cgemm3m_unroll_n
  1177. #define CGEMM3M_UNROLL_MN gotoblas -> cgemm3m_unroll_mn
  1178. #define ZGEMM3M_P gotoblas -> zgemm3m_p
  1179. #define ZGEMM3M_Q gotoblas -> zgemm3m_q
  1180. #define ZGEMM3M_R gotoblas -> zgemm3m_r
  1181. #define ZGEMM3M_UNROLL_M gotoblas -> zgemm3m_unroll_m
  1182. #define ZGEMM3M_UNROLL_N gotoblas -> zgemm3m_unroll_n
  1183. #define ZGEMM3M_UNROLL_MN gotoblas -> zgemm3m_unroll_mn
  1184. #define XGEMM3M_P gotoblas -> xgemm3m_p
  1185. #define XGEMM3M_Q gotoblas -> xgemm3m_q
  1186. #define XGEMM3M_R gotoblas -> xgemm3m_r
  1187. #define XGEMM3M_UNROLL_M gotoblas -> xgemm3m_unroll_m
  1188. #define XGEMM3M_UNROLL_N gotoblas -> xgemm3m_unroll_n
  1189. #define XGEMM3M_UNROLL_MN gotoblas -> xgemm3m_unroll_mn
  1190. #else
  1191. #define FUNC_OFFSET(func) (size_t)(func)
  1192. #define DTB_ENTRIES DTB_DEFAULT_ENTRIES
  1193. #define GEMM_OFFSET_A GEMM_DEFAULT_OFFSET_A
  1194. #define GEMM_OFFSET_B GEMM_DEFAULT_OFFSET_B
  1195. #define GEMM_ALIGN GEMM_DEFAULT_ALIGN
  1196. #ifdef HAVE_EXCLUSIVE_CACHE
  1197. #define HAVE_EX_L2 1
  1198. #else
  1199. #define HAVE_EX_L2 0
  1200. #endif
  1201. #if (BUILD_HFLOAT16 == 1)
  1202. #define SHGEMM_P SHGEMM_DEFAULT_P
  1203. #define SHGEMM_Q SHGEMM_DEFAULT_Q
  1204. #define SHGEMM_R SHGEMM_DEFAULT_R
  1205. #define SHGEMM_UNROLL_M SHGEMM_DEFAULT_UNROLL_M
  1206. #define SHGEMM_UNROLL_N SHGEMM_DEFAULT_UNROLL_N
  1207. #ifdef SHGEMM_DEFAULT_UNROLL_MN
  1208. #define SHGEMM_UNROLL_MN SHGEMM_DEFAULT_UNROLL_MN
  1209. #else
  1210. #define SHGEMM_UNROLL_MN MAX((SHGEMM_UNROLL_M), (SHGEMM_UNROLL_N))
  1211. #endif
  1212. #endif
  1213. #if (BUILD_BFLOAT16 == 1)
  1214. #define BGEMM_P BGEMM_DEFAULT_P
  1215. #define BGEMM_Q BGEMM_DEFAULT_Q
  1216. #define BGEMM_R BGEMM_DEFAULT_R
  1217. #define BGEMM_UNROLL_M BGEMM_DEFAULT_UNROLL_M
  1218. #define BGEMM_UNROLL_N BGEMM_DEFAULT_UNROLL_N
  1219. #ifdef BGEMM_DEFAULT_UNROLL_MN
  1220. #define BGEMM_UNROLL_MN BGEMM_DEFAULT_UNROLL_MN
  1221. #else
  1222. #define BGEMM_UNROLL_MN MAX((BGEMM_UNROLL_M), (BGEMM_UNROLL_N))
  1223. #endif
  1224. #define SBGEMM_P SBGEMM_DEFAULT_P
  1225. #define SBGEMM_Q SBGEMM_DEFAULT_Q
  1226. #define SBGEMM_R SBGEMM_DEFAULT_R
  1227. #define SBGEMM_UNROLL_M SBGEMM_DEFAULT_UNROLL_M
  1228. #define SBGEMM_UNROLL_N SBGEMM_DEFAULT_UNROLL_N
  1229. #ifdef SBGEMM_DEFAULT_UNROLL_MN
  1230. #define SBGEMM_UNROLL_MN SBGEMM_DEFAULT_UNROLL_MN
  1231. #else
  1232. #define SBGEMM_UNROLL_MN MAX((SBGEMM_UNROLL_M), (SBGEMM_UNROLL_N))
  1233. #endif
  1234. #endif
  1235. #define SGEMM_P SGEMM_DEFAULT_P
  1236. #define SGEMM_Q SGEMM_DEFAULT_Q
  1237. #define SGEMM_R SGEMM_DEFAULT_R
  1238. #define SGEMM_UNROLL_M SGEMM_DEFAULT_UNROLL_M
  1239. #define SGEMM_UNROLL_N SGEMM_DEFAULT_UNROLL_N
  1240. #ifdef SGEMM_DEFAULT_UNROLL_MN
  1241. #define SGEMM_UNROLL_MN SGEMM_DEFAULT_UNROLL_MN
  1242. #else
  1243. #define SGEMM_UNROLL_MN MAX((SGEMM_UNROLL_M), (SGEMM_UNROLL_N))
  1244. #endif
  1245. #define DGEMM_P DGEMM_DEFAULT_P
  1246. #define DGEMM_Q DGEMM_DEFAULT_Q
  1247. #define DGEMM_R DGEMM_DEFAULT_R
  1248. #define DGEMM_UNROLL_M DGEMM_DEFAULT_UNROLL_M
  1249. #define DGEMM_UNROLL_N DGEMM_DEFAULT_UNROLL_N
  1250. #ifdef DGEMM_DEFAULT_UNROLL_MN
  1251. #define DGEMM_UNROLL_MN DGEMM_DEFAULT_UNROLL_MN
  1252. #else
  1253. #define DGEMM_UNROLL_MN MAX((DGEMM_UNROLL_M), (DGEMM_UNROLL_N))
  1254. #endif
  1255. #define QGEMM_P QGEMM_DEFAULT_P
  1256. #define QGEMM_Q QGEMM_DEFAULT_Q
  1257. #define QGEMM_R QGEMM_DEFAULT_R
  1258. #define QGEMM_UNROLL_M QGEMM_DEFAULT_UNROLL_M
  1259. #define QGEMM_UNROLL_N QGEMM_DEFAULT_UNROLL_N
  1260. #define QGEMM_UNROLL_MN MAX((QGEMM_UNROLL_M), (QGEMM_UNROLL_N))
  1261. #define CGEMM_P CGEMM_DEFAULT_P
  1262. #define CGEMM_Q CGEMM_DEFAULT_Q
  1263. #define CGEMM_R CGEMM_DEFAULT_R
  1264. #define CGEMM_UNROLL_M CGEMM_DEFAULT_UNROLL_M
  1265. #define CGEMM_UNROLL_N CGEMM_DEFAULT_UNROLL_N
  1266. #ifdef CGEMM_DEFAULT_UNROLL_MN
  1267. #define CGEMM_UNROLL_MN CGEMM_DEFAULT_UNROLL_MN
  1268. #else
  1269. #define CGEMM_UNROLL_MN MAX((CGEMM_UNROLL_M), (CGEMM_UNROLL_N))
  1270. #endif
  1271. #define ZGEMM_P ZGEMM_DEFAULT_P
  1272. #define ZGEMM_Q ZGEMM_DEFAULT_Q
  1273. #define ZGEMM_R ZGEMM_DEFAULT_R
  1274. #define ZGEMM_UNROLL_M ZGEMM_DEFAULT_UNROLL_M
  1275. #define ZGEMM_UNROLL_N ZGEMM_DEFAULT_UNROLL_N
  1276. #ifdef ZGEMM_DEFAULT_UNROLL_MN
  1277. #define ZGEMM_UNROLL_MN ZGEMM_DEFAULT_UNROLL_MN
  1278. #else
  1279. #define ZGEMM_UNROLL_MN MAX((ZGEMM_UNROLL_M), (ZGEMM_UNROLL_N))
  1280. #endif
  1281. #define XGEMM_P XGEMM_DEFAULT_P
  1282. #define XGEMM_Q XGEMM_DEFAULT_Q
  1283. #define XGEMM_R XGEMM_DEFAULT_R
  1284. #define XGEMM_UNROLL_M XGEMM_DEFAULT_UNROLL_M
  1285. #define XGEMM_UNROLL_N XGEMM_DEFAULT_UNROLL_N
  1286. #define XGEMM_UNROLL_MN MAX((XGEMM_UNROLL_M), (XGEMM_UNROLL_N))
  1287. #ifdef CGEMM3M_DEFAULT_UNROLL_N
  1288. #define CGEMM3M_P CGEMM3M_DEFAULT_P
  1289. #define CGEMM3M_Q CGEMM3M_DEFAULT_Q
  1290. #define CGEMM3M_R CGEMM3M_DEFAULT_R
  1291. #define CGEMM3M_UNROLL_M CGEMM3M_DEFAULT_UNROLL_M
  1292. #define CGEMM3M_UNROLL_N CGEMM3M_DEFAULT_UNROLL_N
  1293. #define CGEMM3M_UNROLL_MN MAX((CGEMM3M_UNROLL_M), (CGEMM3M_UNROLL_N))
  1294. #else
  1295. #define CGEMM3M_P SGEMM_DEFAULT_P
  1296. #define CGEMM3M_Q SGEMM_DEFAULT_Q
  1297. #define CGEMM3M_R SGEMM_DEFAULT_R
  1298. #define CGEMM3M_UNROLL_M SGEMM_DEFAULT_UNROLL_M
  1299. #define CGEMM3M_UNROLL_N SGEMM_DEFAULT_UNROLL_N
  1300. #define CGEMM3M_UNROLL_MN MAX((CGEMM_UNROLL_M), (CGEMM_UNROLL_N))
  1301. #endif
  1302. #ifdef ZGEMM3M_DEFAULT_UNROLL_N
  1303. #define ZGEMM3M_P ZGEMM3M_DEFAULT_P
  1304. #define ZGEMM3M_Q ZGEMM3M_DEFAULT_Q
  1305. #define ZGEMM3M_R ZGEMM3M_DEFAULT_R
  1306. #define ZGEMM3M_UNROLL_M ZGEMM3M_DEFAULT_UNROLL_M
  1307. #define ZGEMM3M_UNROLL_N ZGEMM3M_DEFAULT_UNROLL_N
  1308. #define ZGEMM3M_UNROLL_MN MAX((ZGEMM_UNROLL_M), (ZGEMM_UNROLL_N))
  1309. #else
  1310. #define ZGEMM3M_P DGEMM_DEFAULT_P
  1311. #define ZGEMM3M_Q DGEMM_DEFAULT_Q
  1312. #define ZGEMM3M_R DGEMM_DEFAULT_R
  1313. #define ZGEMM3M_UNROLL_M DGEMM_DEFAULT_UNROLL_M
  1314. #define ZGEMM3M_UNROLL_N DGEMM_DEFAULT_UNROLL_N
  1315. #define ZGEMM3M_UNROLL_MN MAX((ZGEMM_UNROLL_M), (ZGEMM_UNROLL_N))
  1316. #endif
  1317. #define XGEMM3M_P QGEMM_DEFAULT_P
  1318. #define XGEMM3M_Q QGEMM_DEFAULT_Q
  1319. #define XGEMM3M_R QGEMM_DEFAULT_R
  1320. #define XGEMM3M_UNROLL_M QGEMM_DEFAULT_UNROLL_M
  1321. #define XGEMM3M_UNROLL_N QGEMM_DEFAULT_UNROLL_N
  1322. #define XGEMM3M_UNROLL_MN MAX((QGEMM_UNROLL_M), (QGEMM_UNROLL_N))
  1323. #endif
  1324. #endif
  1325. #ifndef COMPLEX
  1326. #if defined(XDOUBLE)
  1327. #define GEMM_P QGEMM_P
  1328. #define GEMM_Q QGEMM_Q
  1329. #define GEMM_R QGEMM_R
  1330. #define GEMM_UNROLL_M QGEMM_UNROLL_M
  1331. #define GEMM_UNROLL_N QGEMM_UNROLL_N
  1332. #define GEMM_UNROLL_MN QGEMM_UNROLL_MN
  1333. #define GEMM_DEFAULT_P QGEMM_DEFAULT_P
  1334. #define GEMM_DEFAULT_Q QGEMM_DEFAULT_Q
  1335. #define GEMM_DEFAULT_R QGEMM_DEFAULT_R
  1336. #define GEMM_DEFAULT_UNROLL_M QGEMM_DEFAULT_UNROLL_M
  1337. #define GEMM_DEFAULT_UNROLL_N QGEMM_DEFAULT_UNROLL_N
  1338. #elif defined(DOUBLE)
  1339. #define GEMM_P DGEMM_P
  1340. #define GEMM_Q DGEMM_Q
  1341. #define GEMM_R DGEMM_R
  1342. #define GEMM_UNROLL_M DGEMM_UNROLL_M
  1343. #define GEMM_UNROLL_N DGEMM_UNROLL_N
  1344. #define GEMM_UNROLL_MN DGEMM_UNROLL_MN
  1345. #define GEMM_DEFAULT_P DGEMM_DEFAULT_P
  1346. #define GEMM_DEFAULT_Q DGEMM_DEFAULT_Q
  1347. #define GEMM_DEFAULT_R DGEMM_DEFAULT_R
  1348. #define GEMM_DEFAULT_UNROLL_M DGEMM_DEFAULT_UNROLL_M
  1349. #define GEMM_DEFAULT_UNROLL_N DGEMM_DEFAULT_UNROLL_N
  1350. #elif defined(HFLOAT16)
  1351. #define GEMM_P SHGEMM_P
  1352. #define GEMM_Q SHGEMM_Q
  1353. #define GEMM_R SHGEMM_R
  1354. #define GEMM_UNROLL_M SHGEMM_UNROLL_M
  1355. #define GEMM_UNROLL_N SHGEMM_UNROLL_N
  1356. #define GEMM_UNROLL_MN SHGEMM_UNROLL_MN
  1357. #define GEMM_DEFAULT_P SHGEMM_DEFAULT_P
  1358. #define GEMM_DEFAULT_Q SHGEMM_DEFAULT_Q
  1359. #define GEMM_DEFAULT_R SHGEMM_DEFAULT_R
  1360. #define GEMM_DEFAULT_UNROLL_M SHGEMM_DEFAULT_UNROLL_M
  1361. #define GEMM_DEFAULT_UNROLL_N SHGEMM_DEFAULT_UNROLL_N
  1362. #elif defined(BFLOAT16) && defined(BGEMM)
  1363. #define GEMM_P BGEMM_P
  1364. #define GEMM_Q BGEMM_Q
  1365. #define GEMM_R BGEMM_R
  1366. #define GEMM_UNROLL_M BGEMM_UNROLL_M
  1367. #define GEMM_UNROLL_N BGEMM_UNROLL_N
  1368. #define GEMM_UNROLL_MN BGEMM_UNROLL_MN
  1369. #define GEMM_DEFAULT_P BGEMM_DEFAULT_P
  1370. #define GEMM_DEFAULT_Q BGEMM_DEFAULT_Q
  1371. #define GEMM_DEFAULT_R BGEMM_DEFAULT_R
  1372. #define GEMM_DEFAULT_UNROLL_M BGEMM_DEFAULT_UNROLL_M
  1373. #define GEMM_DEFAULT_UNROLL_N BGEMM_DEFAULT_UNROLL_N
  1374. #elif defined(BFLOAT16)
  1375. #define GEMM_P SBGEMM_P
  1376. #define GEMM_Q SBGEMM_Q
  1377. #define GEMM_R SBGEMM_R
  1378. #define GEMM_UNROLL_M SBGEMM_UNROLL_M
  1379. #define GEMM_UNROLL_N SBGEMM_UNROLL_N
  1380. #define GEMM_UNROLL_MN SBGEMM_UNROLL_MN
  1381. #define GEMM_DEFAULT_P SBGEMM_DEFAULT_P
  1382. #define GEMM_DEFAULT_Q SBGEMM_DEFAULT_Q
  1383. #define GEMM_DEFAULT_R SBGEMM_DEFAULT_R
  1384. #define GEMM_DEFAULT_UNROLL_M SBGEMM_DEFAULT_UNROLL_M
  1385. #define GEMM_DEFAULT_UNROLL_N SBGEMM_DEFAULT_UNROLL_N
  1386. #else
  1387. #define GEMM_P SGEMM_P
  1388. #define GEMM_Q SGEMM_Q
  1389. #define GEMM_R SGEMM_R
  1390. #define GEMM_UNROLL_M SGEMM_UNROLL_M
  1391. #define GEMM_UNROLL_N SGEMM_UNROLL_N
  1392. #define GEMM_UNROLL_MN SGEMM_UNROLL_MN
  1393. #define GEMM_DEFAULT_P SGEMM_DEFAULT_P
  1394. #define GEMM_DEFAULT_Q SGEMM_DEFAULT_Q
  1395. #define GEMM_DEFAULT_R SGEMM_DEFAULT_R
  1396. #define GEMM_DEFAULT_UNROLL_M SGEMM_DEFAULT_UNROLL_M
  1397. #define GEMM_DEFAULT_UNROLL_N SGEMM_DEFAULT_UNROLL_N
  1398. #endif
  1399. #else
  1400. #if defined(XDOUBLE)
  1401. #define GEMM_P XGEMM_P
  1402. #define GEMM_Q XGEMM_Q
  1403. #define GEMM_R XGEMM_R
  1404. #define GEMM_UNROLL_M XGEMM_UNROLL_M
  1405. #define GEMM_UNROLL_N XGEMM_UNROLL_N
  1406. #define GEMM_UNROLL_MN XGEMM_UNROLL_MN
  1407. #define GEMM_DEFAULT_P XGEMM_DEFAULT_P
  1408. #define GEMM_DEFAULT_Q XGEMM_DEFAULT_Q
  1409. #define GEMM_DEFAULT_R XGEMM_DEFAULT_R
  1410. #define GEMM_DEFAULT_UNROLL_M XGEMM_DEFAULT_UNROLL_M
  1411. #define GEMM_DEFAULT_UNROLL_N XGEMM_DEFAULT_UNROLL_N
  1412. #elif defined(DOUBLE)
  1413. #define GEMM_P ZGEMM_P
  1414. #define GEMM_Q ZGEMM_Q
  1415. #define GEMM_R ZGEMM_R
  1416. #define GEMM_UNROLL_M ZGEMM_UNROLL_M
  1417. #define GEMM_UNROLL_N ZGEMM_UNROLL_N
  1418. #define GEMM_UNROLL_MN ZGEMM_UNROLL_MN
  1419. #define GEMM_DEFAULT_P ZGEMM_DEFAULT_P
  1420. #define GEMM_DEFAULT_Q ZGEMM_DEFAULT_Q
  1421. #define GEMM_DEFAULT_R ZGEMM_DEFAULT_R
  1422. #define GEMM_DEFAULT_UNROLL_M ZGEMM_DEFAULT_UNROLL_M
  1423. #define GEMM_DEFAULT_UNROLL_N ZGEMM_DEFAULT_UNROLL_N
  1424. #else
  1425. #define GEMM_P CGEMM_P
  1426. #define GEMM_Q CGEMM_Q
  1427. #define GEMM_R CGEMM_R
  1428. #define GEMM_UNROLL_M CGEMM_UNROLL_M
  1429. #define GEMM_UNROLL_N CGEMM_UNROLL_N
  1430. #define GEMM_UNROLL_MN CGEMM_UNROLL_MN
  1431. #define GEMM_DEFAULT_P CGEMM_DEFAULT_P
  1432. #define GEMM_DEFAULT_Q CGEMM_DEFAULT_Q
  1433. #define GEMM_DEFAULT_R CGEMM_DEFAULT_R
  1434. #define GEMM_DEFAULT_UNROLL_M CGEMM_DEFAULT_UNROLL_M
  1435. #define GEMM_DEFAULT_UNROLL_N CGEMM_DEFAULT_UNROLL_N
  1436. #endif
  1437. #endif
  1438. #ifdef XDOUBLE
  1439. #define GEMM3M_UNROLL_M XGEMM3M_UNROLL_M
  1440. #define GEMM3M_UNROLL_N XGEMM3M_UNROLL_N
  1441. #elif defined(DOUBLE)
  1442. #define GEMM3M_UNROLL_M ZGEMM3M_UNROLL_M
  1443. #define GEMM3M_UNROLL_N ZGEMM3M_UNROLL_N
  1444. #else
  1445. #define GEMM3M_UNROLL_M CGEMM3M_UNROLL_M
  1446. #define GEMM3M_UNROLL_N CGEMM3M_UNROLL_N
  1447. #endif
  1448. #ifndef QGEMM_DEFAULT_UNROLL_M
  1449. #define QGEMM_DEFAULT_UNROLL_M 2
  1450. #endif
  1451. #ifndef QGEMM_DEFAULT_UNROLL_N
  1452. #define QGEMM_DEFAULT_UNROLL_N 2
  1453. #endif
  1454. #ifndef XGEMM_DEFAULT_UNROLL_M
  1455. #define XGEMM_DEFAULT_UNROLL_M 2
  1456. #endif
  1457. #ifndef XGEMM_DEFAULT_UNROLL_N
  1458. #define XGEMM_DEFAULT_UNROLL_N 2
  1459. #endif
  1460. #ifndef GEMM_THREAD
  1461. #define GEMM_THREAD gemm_thread_n
  1462. #endif
  1463. #ifndef SBGEMM_DEFAULT_R
  1464. #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)
  1465. #endif
  1466. #ifndef SGEMM_DEFAULT_R
  1467. #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)
  1468. #endif
  1469. #ifndef DGEMM_DEFAULT_R
  1470. #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)
  1471. #endif
  1472. #ifndef QGEMM_DEFAULT_R
  1473. #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)
  1474. #endif
  1475. #ifndef CGEMM_DEFAULT_R
  1476. #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)
  1477. #endif
  1478. #ifndef ZGEMM_DEFAULT_R
  1479. #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)
  1480. #endif
  1481. #ifndef XGEMM_DEFAULT_R
  1482. #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)
  1483. #endif
  1484. #ifndef SNUMOPT
  1485. #define SNUMOPT 2
  1486. #endif
  1487. #ifndef DNUMOPT
  1488. #define DNUMOPT 2
  1489. #endif
  1490. #ifndef QNUMOPT
  1491. #define QNUMOPT 1
  1492. #endif
  1493. #ifndef GEMM3M_P
  1494. #ifdef XDOUBLE
  1495. #define GEMM3M_P XGEMM3M_P
  1496. #elif defined(DOUBLE)
  1497. #define GEMM3M_P ZGEMM3M_P
  1498. #else
  1499. #define GEMM3M_P CGEMM3M_P
  1500. #endif
  1501. #endif
  1502. #ifndef GEMM3M_Q
  1503. #ifdef XDOUBLE
  1504. #define GEMM3M_Q XGEMM3M_Q
  1505. #elif defined(DOUBLE)
  1506. #define GEMM3M_Q ZGEMM3M_Q
  1507. #else
  1508. #define GEMM3M_Q CGEMM3M_Q
  1509. #endif
  1510. #endif
  1511. #ifndef GEMM3M_R
  1512. #ifdef XDOUBLE
  1513. #define GEMM3M_R XGEMM3M_R
  1514. #elif defined(DOUBLE)
  1515. #define GEMM3M_R ZGEMM3M_R
  1516. #else
  1517. #define GEMM3M_R CGEMM3M_R
  1518. #endif
  1519. #endif
  1520. #endif