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.

KERNEL.POWER10 7.2 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241
  1. ifeq ($(HAVE_GAS), 1)
  2. include $(KERNELDIR)/KERNEL.POWER8
  3. else
  4. #SGEMM_BETA = ../generic/gemm_beta.c
  5. #DGEMM_BETA = ../generic/gemm_beta.c
  6. #CGEMM_BETA = ../generic/zgemm_beta.c
  7. #ZGEMM_BETA = ../generic/zgemm_beta.c
  8. SBGEMM_BETA = ../generic/gemm_beta.c
  9. SBGEMMKERNEL = sbgemm_kernel_power10.c
  10. SBGEMMINCOPY = sbgemm_ncopy_16_power10.c
  11. SBGEMMITCOPY = sbgemm_tcopy_16_power10.c
  12. SBGEMMONCOPY = sbgemm_ncopy_8_power10.c
  13. SBGEMMOTCOPY = sbgemm_tcopy_8_power10.c
  14. SBGEMMINCOPYOBJ = sbgemm_incopy$(TSUFFIX).$(SUFFIX)
  15. SBGEMMITCOPYOBJ = sbgemm_itcopy$(TSUFFIX).$(SUFFIX)
  16. SBGEMMONCOPYOBJ = sbgemm_oncopy$(TSUFFIX).$(SUFFIX)
  17. SBGEMMOTCOPYOBJ = sbgemm_otcopy$(TSUFFIX).$(SUFFIX)
  18. STRMMKERNEL = sgemm_kernel_power10.c
  19. DTRMMKERNEL = dgemm_kernel_power10.c
  20. CTRMMKERNEL = cgemm_kernel_power10.S
  21. ZTRMMKERNEL = zgemm_kernel_power10.S
  22. SGEMMKERNEL = sgemm_kernel_power10.c
  23. SGEMMINCOPY = ../generic/gemm_ncopy_16.c
  24. SGEMMITCOPY = sgemm_tcopy_16_power8.S
  25. SGEMMONCOPY = ../generic/gemm_ncopy_8.c
  26. SGEMMOTCOPY = sgemm_tcopy_8_power8.S
  27. SGEMMINCOPYOBJ = sgemm_incopy$(TSUFFIX).$(SUFFIX)
  28. SGEMMITCOPYOBJ = sgemm_itcopy$(TSUFFIX).$(SUFFIX)
  29. SGEMMONCOPYOBJ = sgemm_oncopy$(TSUFFIX).$(SUFFIX)
  30. SGEMMOTCOPYOBJ = sgemm_otcopy$(TSUFFIX).$(SUFFIX)
  31. SGEMM_SMALL_M_PERMIT = gemm_small_kernel_permit_power10.c
  32. SGEMM_SMALL_K_NN = sgemm_small_kernel_nn_power10.c
  33. SGEMM_SMALL_K_B0_NN = sgemm_small_kernel_nn_power10.c
  34. SGEMM_SMALL_K_NT = sgemm_small_kernel_nt_power10.c
  35. SGEMM_SMALL_K_B0_NT = sgemm_small_kernel_nt_power10.c
  36. SGEMM_SMALL_K_TN = sgemm_small_kernel_tn_power10.c
  37. SGEMM_SMALL_K_B0_TN = sgemm_small_kernel_tn_power10.c
  38. SGEMM_SMALL_K_TT = sgemm_small_kernel_tt_power10.c
  39. SGEMM_SMALL_K_B0_TT = sgemm_small_kernel_tt_power10.c
  40. DGEMMKERNEL = dgemm_kernel_power10.c
  41. DGEMMINCOPY =
  42. DGEMMITCOPY =
  43. DGEMMONCOPY = dgemm_ncopy_8_power10.c
  44. DGEMMOTCOPY = ../generic/gemm_tcopy_8.c
  45. DGEMMINCOPYOBJ =
  46. DGEMMITCOPYOBJ =
  47. DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX)
  48. DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX)
  49. DGEMM_SMALL_M_PERMIT = gemm_small_kernel_permit_power10.c
  50. DGEMM_SMALL_K_NT = dgemm_small_kernel_nt_power10.c
  51. DGEMM_SMALL_K_B0_NT = dgemm_small_kernel_nt_power10.c
  52. DGEMM_SMALL_K_NN = dgemm_small_kernel_nn_power10.c
  53. DGEMM_SMALL_K_B0_NN = dgemm_small_kernel_nn_power10.c
  54. DGEMM_SMALL_K_TT = dgemm_small_kernel_tt_power10.c
  55. DGEMM_SMALL_K_B0_TT = dgemm_small_kernel_tt_power10.c
  56. DGEMM_SMALL_K_TN = dgemm_small_kernel_tn_power10.c
  57. DGEMM_SMALL_K_B0_TN = dgemm_small_kernel_tn_power10.c
  58. CGEMMKERNEL = cgemm_kernel_power10.S
  59. #CGEMMKERNEL = cgemm_kernel_8x4_power8.S
  60. CGEMMINCOPY = ../generic/zgemm_ncopy_8.c
  61. CGEMMITCOPY = ../generic/zgemm_tcopy_8.c
  62. CGEMMONCOPY = ../generic/zgemm_ncopy_4.c
  63. CGEMMOTCOPY = ../generic/zgemm_tcopy_4.c
  64. CGEMMONCOPYOBJ = cgemm_oncopy$(TSUFFIX).$(SUFFIX)
  65. CGEMMOTCOPYOBJ = cgemm_otcopy$(TSUFFIX).$(SUFFIX)
  66. CGEMMINCOPYOBJ = cgemm_incopy$(TSUFFIX).$(SUFFIX)
  67. CGEMMITCOPYOBJ = cgemm_itcopy$(TSUFFIX).$(SUFFIX)
  68. ZGEMMKERNEL = zgemm_kernel_power10.S
  69. ZGEMMONCOPY = ../generic/zgemm_ncopy_2.c
  70. ZGEMMOTCOPY = ../generic/zgemm_tcopy_2.c
  71. ZGEMMINCOPY = ../generic/zgemm_ncopy_8.c
  72. ZGEMMITCOPY = zgemm_tcopy_8_power8.S
  73. ZGEMMONCOPYOBJ = zgemm_oncopy$(TSUFFIX).$(SUFFIX)
  74. ZGEMMOTCOPYOBJ = zgemm_otcopy$(TSUFFIX).$(SUFFIX)
  75. ZGEMMINCOPYOBJ = zgemm_incopy$(TSUFFIX).$(SUFFIX)
  76. ZGEMMITCOPYOBJ = zgemm_itcopy$(TSUFFIX).$(SUFFIX)
  77. STRSMKERNEL_LN = trsm_kernel_LN_power10.c
  78. STRSMKERNEL_LT = trsm_kernel_LT_power10.c
  79. STRSMKERNEL_RN = trsm_kernel_RN_power10.c
  80. STRSMKERNEL_RT = trsm_kernel_RT_power10.c
  81. DTRSMKERNEL_LN = trsm_kernel_LN_power10.c
  82. DTRSMKERNEL_LT = trsm_kernel_LT_power10.c
  83. DTRSMKERNEL_RN = trsm_kernel_RN_power10.c
  84. DTRSMKERNEL_RT = trsm_kernel_RT_power10.c
  85. CTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
  86. CTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
  87. CTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
  88. CTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
  89. ZTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
  90. ZTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
  91. ZTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
  92. ZTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
  93. #Todo: CGEMM3MKERNEL should be 4x4 blocksizes.
  94. #CGEMM3MKERNEL = zgemm3m_kernel_8x4_sse3.S
  95. #ZGEMM3MKERNEL = zgemm3m_kernel_4x4_sse3.S
  96. #Pure C for other kernels
  97. #SAMAXKERNEL = ../arm/amax.c
  98. #DAMAXKERNEL = ../arm/amax.c
  99. #CAMAXKERNEL = ../arm/zamax.c
  100. #ZAMAXKERNEL = ../arm/zamax.c
  101. #
  102. #SAMINKERNEL = ../arm/amin.c
  103. #DAMINKERNEL = ../arm/amin.c
  104. #CAMINKERNEL = ../arm/zamin.c
  105. #ZAMINKERNEL = ../arm/zamin.c
  106. #
  107. #SMAXKERNEL = ../arm/max.c
  108. #DMAXKERNEL = ../arm/max.c
  109. #
  110. #SMINKERNEL = ../arm/min.c
  111. #DMINKERNEL = ../arm/min.c
  112. #
  113. ifneq ($(GCCVERSIONGTEQ9),1)
  114. ISAMAXKERNEL = isamax_power9.S
  115. else
  116. ISAMAXKERNEL = isamax.c
  117. endif
  118. IDAMAXKERNEL = idamax.c
  119. ifneq ($(GCCVERSIONGTEQ9),1)
  120. ICAMAXKERNEL = icamax_power9.S
  121. else
  122. ICAMAXKERNEL = icamax.c
  123. endif
  124. IZAMAXKERNEL = izamax.c
  125. #
  126. ifneq ($(GCCVERSIONGTEQ9),1)
  127. ISAMINKERNEL = isamin_power9.S
  128. else
  129. ISAMINKERNEL = isamin.c
  130. endif
  131. IDAMINKERNEL = idamin.c
  132. ifneq ($(GCCVERSIONGTEQ9),1)
  133. ICAMINKERNEL = icamin_power9.S
  134. else
  135. ICAMINKERNEL = icamin.c
  136. endif
  137. IZAMINKERNEL = izamin.c
  138. #
  139. #ISMAXKERNEL = ../arm/imax.c
  140. #IDMAXKERNEL = ../arm/imax.c
  141. #
  142. #ISMINKERNEL = ../arm/imin.c
  143. #IDMINKERNEL = ../arm/imin.c
  144. #
  145. SASUMKERNEL = sasum.c
  146. DASUMKERNEL = dasum.c
  147. CASUMKERNEL = casum.c
  148. ZASUMKERNEL = zasum.c
  149. #
  150. SAXPYKERNEL = saxpy_power10.c
  151. DAXPYKERNEL = daxpy_power10.c
  152. CAXPYKERNEL = caxpy_power10.c
  153. ZAXPYKERNEL = zaxpy_power10.c
  154. #
  155. SCOPYKERNEL = scopy_power10.c
  156. DCOPYKERNEL = dcopy_power10.c
  157. CCOPYKERNEL = ccopy_power10.c
  158. ZCOPYKERNEL = zcopy_power10.c
  159. #
  160. SDOTKERNEL = sdot_power10.c
  161. DDOTKERNEL = ddot_power10.c
  162. DSDOTKERNEL = sdot_power10.c
  163. CDOTKERNEL = cdot.c
  164. ZDOTKERNEL = zdot.c
  165. #
  166. SNRM2KERNEL = ../arm/nrm2.c
  167. DNRM2KERNEL = ../arm/nrm2.c
  168. CNRM2KERNEL = ../arm/znrm2.c
  169. ZNRM2KERNEL = ../arm/znrm2.c
  170. #
  171. SROTKERNEL = srot.c
  172. DROTKERNEL = drot.c
  173. CROTKERNEL = crot.c
  174. ZROTKERNEL = zrot.c
  175. #
  176. SSCALKERNEL = sscal.c
  177. DSCALKERNEL = dscal.c
  178. ifeq ($(C_COMPILER), PGI)
  179. CSCALKERNEL = ../arm/zscal.c
  180. ZSCALKERNEL = ../arm/zscal.c
  181. else
  182. CSCALKERNEL = zscal.c
  183. ZSCALKERNEL = zscal.c
  184. endif
  185. #
  186. SSWAPKERNEL = sswap.c
  187. DSWAPKERNEL = dswap.c
  188. CSWAPKERNEL = cswap.c
  189. ZSWAPKERNEL = zswap.c
  190. #
  191. SGEMVNKERNEL = sgemv_n.c
  192. DGEMVNKERNEL = dgemv_n_power10.c
  193. CGEMVNKERNEL = cgemv_n.c
  194. ZGEMVNKERNEL = zgemv_n_power10.c
  195. #
  196. SGEMVTKERNEL = sgemv_t.c
  197. DGEMVTKERNEL = dgemv_t_power10.c
  198. CGEMVTKERNEL = cgemv_t.c
  199. ZGEMVTKERNEL = zgemv_t_4.c
  200. #SSYMV_U_KERNEL = ../generic/symv_k.c
  201. #SSYMV_L_KERNEL = ../generic/symv_k.c
  202. #DSYMV_U_KERNEL = ../generic/symv_k.c
  203. #DSYMV_L_KERNEL = ../generic/symv_k.c
  204. #QSYMV_U_KERNEL = ../generic/symv_k.c
  205. #QSYMV_L_KERNEL = ../generic/symv_k.c
  206. #CSYMV_U_KERNEL = ../generic/zsymv_k.c
  207. #CSYMV_L_KERNEL = ../generic/zsymv_k.c
  208. #ZSYMV_U_KERNEL = ../generic/zsymv_k.c
  209. #ZSYMV_L_KERNEL = ../generic/zsymv_k.c
  210. #XSYMV_U_KERNEL = ../generic/zsymv_k.c
  211. #XSYMV_L_KERNEL = ../generic/zsymv_k.c
  212. #ZHEMV_U_KERNEL = ../generic/zhemv_k.c
  213. #ZHEMV_L_KERNEL = ../generic/zhemv_k.c
  214. LSAME_KERNEL = ../generic/lsame.c
  215. SCABS_KERNEL = ../generic/cabs.c
  216. DCABS_KERNEL = ../generic/cabs.c
  217. QCABS_KERNEL = ../generic/cabs.c
  218. #Dump kernel
  219. CGEMM3MKERNEL = ../generic/zgemm3mkernel_dump.c
  220. ZGEMM3MKERNEL = ../generic/zgemm3mkernel_dump.c
  221. endif