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.

CMakeLists.txt 10 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256
  1. include_directories(${PROJECT_SOURCE_DIR})
  2. include_directories(${PROJECT_BINARY_DIR})
  3. set(BLAS1_SOURCES
  4. copy.c
  5. nrm2.c
  6. )
  7. set(BLAS1_REAL_ONLY_SOURCES
  8. rotm.c rotmg.c # N.B. these do not have complex counterparts
  9. rot.c
  10. asum.c
  11. sum.c
  12. )
  13. # these will have 'z' prepended for the complex version
  14. set(BLAS1_MANGLED_SOURCES
  15. axpy.c swap.c
  16. scal.c
  17. dot.c
  18. rotg.c
  19. axpby.c
  20. )
  21. # TODO: USE_NETLIB_GEMV should switch gemv.c to netlib/*gemv.f
  22. # these all have 'z' sources for complex versions
  23. set(BLAS2_SOURCES
  24. gemv.c ger.c
  25. trsv.c trmv.c
  26. syr2.c gbmv.c
  27. sbmv.c spmv.c
  28. spr2.c
  29. tbsv.c tbmv.c
  30. tpsv.c tpmv.c
  31. )
  32. set(BLAS2_REAL_ONLY_SOURCES
  33. symv.c syr.c spr.c
  34. )
  35. set(BLAS2_COMPLEX_LAPACK_SOURCES
  36. symv.c syr.c spr.c
  37. )
  38. set(BLAS2_COMPLEX_ONLY_MANGLED_SOURCES
  39. hemv.c hbmv.c
  40. her.c her2.c
  41. hpmv.c hpr.c
  42. hpr2.c
  43. )
  44. # these do not have separate 'z' sources
  45. set(BLAS3_SOURCES
  46. gemm.c symm.c
  47. trsm.c syrk.c syr2k.c gemmt.c
  48. )
  49. set(BLAS3_MANGLED_SOURCES
  50. omatcopy.c imatcopy.c
  51. geadd.c
  52. )
  53. # generate the BLAS objs once with and once without cblas
  54. set (CBLAS_FLAGS "")
  55. if (NOT DEFINED NO_FBLAS)
  56. list(APPEND CBLAS_FLAGS 0)
  57. endif ()
  58. if (NOT DEFINED NO_CBLAS)
  59. list(APPEND CBLAS_FLAGS 1)
  60. endif ()
  61. foreach (CBLAS_FLAG ${CBLAS_FLAGS})
  62. # TODO: don't compile complex sources with cblas for now, the naming schemes are all different and they will have to be handled separately from SINGLE/DOUBLE
  63. set(DISABLE_COMPLEX 0)
  64. set(MANGLE_COMPLEX 3)
  65. if (CBLAS_FLAG EQUAL 1)
  66. # set(DISABLE_COMPLEX 1)
  67. # set(MANGLE_COMPLEX 1)
  68. endif ()
  69. GenerateNamedObjects("${BLAS1_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false ${DISABLE_COMPLEX})
  70. GenerateNamedObjects("${BLAS1_REAL_ONLY_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false 1)
  71. GenerateNamedObjects("${BLAS1_MANGLED_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false ${MANGLE_COMPLEX})
  72. GenerateNamedObjects("${BLAS2_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false ${MANGLE_COMPLEX})
  73. GenerateNamedObjects("${BLAS2_REAL_ONLY_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false 1)
  74. if (NOT DEFINED NO_LAPACK)
  75. GenerateNamedObjects("${BLAS2_COMPLEX_LAPACK_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false ${MANGLE_COMPLEX})
  76. endif ()
  77. GenerateNamedObjects("${BLAS2_COMPLEX_ONLY_MANGLED_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false 4)
  78. GenerateNamedObjects("${BLAS3_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false ${DISABLE_COMPLEX})
  79. GenerateNamedObjects("${BLAS3_MANGLED_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false ${MANGLE_COMPLEX})
  80. GenerateNamedObjects("xerbla.c" "" "xerbla" ${CBLAS_FLAG} "" "" true)
  81. #sdsdot, dsdot
  82. if (BUILD_SINGLE OR BUILD_DOUBLE)
  83. GenerateNamedObjects("sdsdot.c" "" "sdsdot" ${CBLAS_FLAG} "" "" true "SINGLE")
  84. if(CBLAS_FLAG EQUAL 1)
  85. GenerateNamedObjects("gemm_batch.c" "" "gemm_batch" ${CBLAS_FLAG} "" "" false)
  86. endif ()
  87. endif ()
  88. if (BUILD_DOUBLE)
  89. GenerateNamedObjects("dsdot.c" "" "dsdot" ${CBLAS_FLAG} "" "" true "SINGLE")
  90. endif ()
  91. # trmm is trsm with a compiler flag set
  92. GenerateNamedObjects("trsm.c" "TRMM" "trmm" ${CBLAS_FLAG})
  93. # gemmtr is gemmt under the name adopted by the Reference BLAS
  94. GenerateNamedObjects("gemm.c" "RNAME" "gemmtr" ${CBLAS_FLAG})
  95. # max and imax are compiled 4 times
  96. GenerateNamedObjects("max.c" "" "" ${CBLAS_FLAG})
  97. GenerateNamedObjects("max.c" "USE_ABS" "amax" ${CBLAS_FLAG})
  98. GenerateNamedObjects("max.c" "USE_ABS;USE_MIN" "amin" ${CBLAS_FLAG})
  99. GenerateNamedObjects("max.c" "USE_MIN" "min" ${CBLAS_FLAG})
  100. GenerateNamedObjects("imax.c" "" "i*max" ${CBLAS_FLAG})
  101. GenerateNamedObjects("imax.c" "USE_ABS" "i*amax" ${CBLAS_FLAG})
  102. GenerateNamedObjects("imax.c" "USE_ABS;USE_MIN" "i*amin" ${CBLAS_FLAG})
  103. GenerateNamedObjects("imax.c" "USE_MIN" "i*min" ${CBLAS_FLAG})
  104. if (BUILD_BFLOAT16)
  105. GenerateNamedObjects("bf16dot.c" "" "sbdot" ${CBLAS_FLAG} "" "" true "BFLOAT16")
  106. GenerateNamedObjects("gemm.c" "" "sbgemm" ${CBLAS_FLAG} "" "" true "BFLOAT16")
  107. GenerateNamedObjects("gemmt.c" "" "sbgemmt" ${CBLAS_FLAG} "" "" true "BFLOAT16")
  108. GenerateNamedObjects("gemmt.c" "RNAME" "sbgemmtr" ${CBLAS_FLAG} "" "" true "BFLOAT16")
  109. GenerateNamedObjects("sbgemv.c" "" "sbgemv" ${CBLAS_FLAG} "" "" true "BFLOAT16")
  110. GenerateNamedObjects("tobf16.c" "SINGLE_PREC" "sbstobf16" ${CBLAS_FLAG} "" "" true "BFLOAT16")
  111. GenerateNamedObjects("tobf16.c" "DOUBLE_PREC" "sbdtobf16" ${CBLAS_FLAG} "" "" true "BFLOAT16")
  112. GenerateNamedObjects("bf16to.c" "SINGLE_PREC" "sbf16tos" ${CBLAS_FLAG} "" "" true "BFLOAT16")
  113. GenerateNamedObjects("bf16to.c" "DOUBLE_PREC" "dbf16tod" ${CBLAS_FLAG} "" "" true "BFLOAT16")
  114. if(CBLAS_FLAG EQUAL 1)
  115. GenerateNamedObjects("gemm_batch.c" "" "sbgemm_batch" ${CBLAS_FLAG} "" "" true "BFLOAT16")
  116. endif ()
  117. endif ()
  118. # complex-specific sources
  119. foreach (float_type ${FLOAT_TYPES})
  120. if (${float_type} STREQUAL "COMPLEX" OR ${float_type} STREQUAL "ZCOMPLEX")
  121. GenerateNamedObjects("zaxpy.c" "CONJ" "axpyc" ${CBLAS_FLAG} "" "" false ${float_type})
  122. GenerateNamedObjects("zger.c" "" "geru" ${CBLAS_FLAG} "" "" false ${float_type})
  123. GenerateNamedObjects("zger.c" "CONJ" "gerc" ${CBLAS_FLAG} "" "" false ${float_type})
  124. GenerateNamedObjects("zdot.c" "CONJ" "dotc" ${CBLAS_FLAG} "" "" false ${float_type})
  125. GenerateNamedObjects("zdot.c" "" "dotu" ${CBLAS_FLAG} "" "" false ${float_type})
  126. GenerateNamedObjects("symm.c" "HEMM" "hemm" ${CBLAS_FLAG} "" "" false ${float_type})
  127. GenerateNamedObjects("syrk.c" "HEMM" "herk" ${CBLAS_FLAG} "" "" false ${float_type})
  128. GenerateNamedObjects("syr2k.c" "HEMM" "her2k" ${CBLAS_FLAG} "" "" false ${float_type})
  129. if (USE_GEMM3M)
  130. GenerateNamedObjects("gemm.c" "GEMM3M" "gemm3m" ${CBLAS_FLAG} "" "" false ${float_type})
  131. endif()
  132. endif ()
  133. if (${float_type} STREQUAL "COMPLEX")
  134. GenerateNamedObjects("zscal.c" "SSCAL" "sscal" ${CBLAS_FLAG} "" "" false "COMPLEX")
  135. GenerateNamedObjects("nrm2.c" "" "scnrm2" ${CBLAS_FLAG} "" "" true "COMPLEX")
  136. GenerateNamedObjects("zrot.c" "" "csrot" ${CBLAS_FLAG} "" "" true "COMPLEX")
  137. GenerateNamedObjects("max.c" "USE_ABS;USE_MIN" "scamin" ${CBLAS_FLAG} "" "" true "COMPLEX")
  138. GenerateNamedObjects("max.c" "USE_ABS" "scamax" ${CBLAS_FLAG} "" "" true "COMPLEX")
  139. GenerateNamedObjects("asum.c" "" "scasum" ${CBLAS_FLAG} "" "" true "COMPLEX")
  140. GenerateNamedObjects("sum.c" "" "scsum" ${CBLAS_FLAG} "" "" true "COMPLEX")
  141. if(CBLAS_FLAG EQUAL 1)
  142. GenerateNamedObjects("gemm_batch.c" "" "cgemm_batch" ${CBLAS_FLAG} "" "" true "COMPLEX")
  143. endif ()
  144. endif ()
  145. if (${float_type} STREQUAL "ZCOMPLEX")
  146. GenerateNamedObjects("zscal.c" "SSCAL" "dscal" ${CBLAS_FLAG} "" "" false "ZCOMPLEX")
  147. GenerateNamedObjects("nrm2.c" "" "dznrm2" ${CBLAS_FLAG} "" "" true "ZCOMPLEX")
  148. GenerateNamedObjects("zrot.c" "" "zdrot" ${CBLAS_FLAG} "" "" true "ZCOMPLEX")
  149. GenerateNamedObjects("max.c" "USE_ABS;USE_MIN" "dzamin" ${CBLAS_FLAG} "" "" true "ZCOMPLEX")
  150. GenerateNamedObjects("max.c" "USE_ABS" "dzamax" ${CBLAS_FLAG} "" "" true "ZCOMPLEX")
  151. GenerateNamedObjects("asum.c" "" "dzasum" ${CBLAS_FLAG} "" "" true "ZCOMPLEX")
  152. GenerateNamedObjects("sum.c" "" "dzsum" ${CBLAS_FLAG} "" "" true "ZCOMPLEX")
  153. if(CBLAS_FLAG EQUAL 1)
  154. GenerateNamedObjects("gemm_batch.c" "" "zgemm_batch" ${CBLAS_FLAG} "" "" true "ZCOMPLEX")
  155. endif ()
  156. endif ()
  157. endforeach ()
  158. endforeach ()
  159. #Special functions for CBLAS
  160. if (NOT DEFINED NO_CBLAS)
  161. foreach (float_type ${FLOAT_TYPES})
  162. if (${float_type} STREQUAL "COMPLEX" OR ${float_type} STREQUAL "ZCOMPLEX")
  163. #cblas_dotc_sub cblas_dotu_sub
  164. GenerateNamedObjects("zdot.c" "FORCE_USE_STACK" "dotu_sub" 1 "" "" false ${float_type})
  165. GenerateNamedObjects("zdot.c" "FORCE_USE_STACK;CONJ" "dotc_sub" 1 "" "" false ${float_type})
  166. endif()
  167. endforeach ()
  168. endif()
  169. if (NOT NO_LAPACK)
  170. set(LAPACK_SOURCES
  171. lapack/gesv.c
  172. )
  173. # prepend z for complex versions
  174. set(LAPACK_MANGLED_SOURCES
  175. lapack/getrf.c lapack/getrs.c lapack/potrf.c lapack/getf2.c
  176. lapack/potf2.c lapack/laswp.c lapack/lauu2.c
  177. lapack/lauum.c lapack/trti2.c lapack/trtri.c
  178. )
  179. GenerateNamedObjects("${LAPACK_SOURCES}")
  180. if (NOT RELAPACK_REPLACE)
  181. GenerateNamedObjects("${LAPACK_MANGLED_SOURCES}" "" "" 0 "" "" 0 3)
  182. else ()
  183. GenerateNamedObjects("lapack/getrs.c" "" "" 0 "" "" 0 3)
  184. GenerateNamedObjects("lapack/getf2.c" "" "" 0 "" "" 0 3)
  185. GenerateNamedObjects("lapack/potf2.c" "" "" 0 "" "" 0 3)
  186. GenerateNamedObjects("lapack/laswp.c" "" "" 0 "" "" 0 3)
  187. GenerateNamedObjects("lapack/lauu2.c" "" "" 0 "" "" 0 3)
  188. GenerateNamedObjects("lapack/trti2.c" "" "" 0 "" "" 0 3)
  189. endif()
  190. endif ()
  191. if ( BUILD_COMPLEX AND NOT BUILD_SINGLE)
  192. GenerateNamedObjects("scal.c" "" "scal" 0 "" "" false "SINGLE")
  193. GenerateNamedObjects("copy.c" "" "copy" 0 "" "" false "SINGLE")
  194. GenerateNamedObjects("dot.c" "" "dot" 0 "" "" false "SINGLE")
  195. GenerateNamedObjects("rot.c" "" "rot" 0 "" "" false "SINGLE")
  196. GenerateNamedObjects("nrm2.c" "" "nrm2" 0 "" "" false "SINGLE")
  197. GenerateNamedObjects("gemv.c" "" "gemv" 0 "" "" false "SINGLE")
  198. GenerateNamedObjects("gemm.c" "" "gemm" 0 "" "" false "SINGLE")
  199. GenerateNamedObjects("gemm_batch.c" "" "gemm_batch" 1 "" "" false "SINGLE")
  200. GenerateNamedObjects("asum.c" "" "asum" 0 "" "" false "SINGLE")
  201. GenerateNamedObjects("swap.c" "" "swap" 0 "" "" false "SINGLE")
  202. GenerateNamedObjects("axpy.c" "" "axpy" 0 "" "" false "SINGLE")
  203. GenerateNamedObjects("imax.c" "USE_ABS" "i*amax" 0 "" "" false "SINGLE")
  204. endif ()
  205. if ( BUILD_COMPLEX16 AND NOT BUILD_DOUBLE)
  206. GenerateNamedObjects("scal.c" "" "scal" 0 "" "" false "DOUBLE")
  207. GenerateNamedObjects("copy.c" "" "copy" 0 "" "" false "DOUBLE")
  208. GenerateNamedObjects("dot.c" "" "dot" 0 "" "" false "DOUBLE")
  209. GenerateNamedObjects("rot.c" "" "rot" 0 "" "" false "DOUBLE")
  210. GenerateNamedObjects("nrm2.c" "" "nrm2" 0 "" "" false "DOUBLE")
  211. GenerateNamedObjects("gemv.c" "" "gemv" 0 "" "" false "DOUBLE")
  212. GenerateNamedObjects("gemm.c" "" "gemm" 0 "" "" false "DOUBLE")
  213. GenerateNamedObjects("gemm_batch.c" "" "gemm_batch" 1 "" "" false "DOUBLE")
  214. GenerateNamedObjects("asum.c" "" "asum" 0 "" "" false "DOUBLE")
  215. GenerateNamedObjects("swap.c" "" "swap" 0 "" "" false "DOUBLE")
  216. GenerateNamedObjects("axpy.c" "" "axpy" 0 "" "" false "DOUBLE")
  217. GenerateNamedObjects("imax.c" "USE_ABS" "i*amax" 0 "" "" false "DOUBLE")
  218. endif ()
  219. add_library(interface OBJECT ${OPENBLAS_SRC})
  220. if (USE_OPENMP)
  221. target_link_libraries(interface OpenMP::OpenMP_C)
  222. endif()