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.

prebuild.cmake 12 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350
  1. ##
  2. ## Author: Hank Anderson <hank@statease.com>
  3. ## Description: Ported from OpenBLAS/Makefile.prebuild
  4. ## This is triggered by system.cmake and runs before any of the code is built.
  5. ## Creates config.h and Makefile.conf by first running the c_check perl script (which creates those files).
  6. ## Next it runs f_check and appends some fortran information to the files.
  7. ## Then it runs getarch and getarch_2nd for even more environment information.
  8. ## Finally it builds gen_config_h for use at build time to generate config.h.
  9. # CMake vars set by this file:
  10. # CORE
  11. # LIBCORE
  12. # NUM_CORES
  13. # HAVE_MMX
  14. # HAVE_SSE
  15. # HAVE_SSE2
  16. # HAVE_SSE3
  17. # MAKE
  18. # SGEMM_UNROLL_M
  19. # SGEMM_UNROLL_N
  20. # DGEMM_UNROLL_M
  21. # DGEMM_UNROLL_M
  22. # QGEMM_UNROLL_N
  23. # QGEMM_UNROLL_N
  24. # CGEMM_UNROLL_M
  25. # CGEMM_UNROLL_M
  26. # ZGEMM_UNROLL_N
  27. # ZGEMM_UNROLL_N
  28. # XGEMM_UNROLL_M
  29. # XGEMM_UNROLL_N
  30. # CGEMM3M_UNROLL_M
  31. # CGEMM3M_UNROLL_N
  32. # ZGEMM3M_UNROLL_M
  33. # ZGEMM3M_UNROLL_M
  34. # XGEMM3M_UNROLL_N
  35. # XGEMM3M_UNROLL_N
  36. # CPUIDEMU = ../../cpuid/table.o
  37. if (DEFINED CPUIDEMU)
  38. set(EXFLAGS "-DCPUIDEMU -DVENDOR=99")
  39. endif ()
  40. if (BUILD_KERNEL)
  41. # set the C flags for just this file
  42. set(GETARCH2_FLAGS "-DBUILD_KERNEL")
  43. set(TARGET_CONF "config_kernel.h")
  44. set(TARGET_CONF_DIR ${PROJECT_BINARY_DIR}/kernel_config/${TARGET_CORE})
  45. else()
  46. set(TARGET_CONF "config.h")
  47. set(TARGET_CONF_DIR ${PROJECT_BINARY_DIR})
  48. endif ()
  49. set(TARGET_CONF_TEMP "${PROJECT_BINARY_DIR}/${TARGET_CONF}.tmp")
  50. # c_check
  51. set(FU "")
  52. if (APPLE OR (MSVC AND NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang"))
  53. set(FU "_")
  54. endif()
  55. set(COMPILER_ID ${CMAKE_C_COMPILER_ID})
  56. if (${COMPILER_ID} STREQUAL "GNU")
  57. set(COMPILER_ID "GCC")
  58. endif ()
  59. string(TOUPPER ${ARCH} UC_ARCH)
  60. file(WRITE ${TARGET_CONF_TEMP}
  61. "#define OS_${HOST_OS}\t1\n"
  62. "#define ARCH_${UC_ARCH}\t1\n"
  63. "#define C_${COMPILER_ID}\t1\n"
  64. "#define __${BINARY}BIT__\t1\n"
  65. "#define FUNDERSCORE\t${FU}\n")
  66. if (${HOST_OS} STREQUAL "WINDOWSSTORE")
  67. file(APPEND ${TARGET_CONF_TEMP}
  68. "#define OS_WINNT\t1\n")
  69. endif ()
  70. # f_check
  71. if (NOT NOFORTRAN)
  72. include("${PROJECT_SOURCE_DIR}/cmake/f_check.cmake")
  73. endif ()
  74. # Cannot run getarch on target if we are cross-compiling
  75. if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSSTORE"))
  76. # Write to config as getarch would
  77. # TODO: Set up defines that getarch sets up based on every other target
  78. # Perhaps this should be inside a different file as it grows larger
  79. file(APPEND ${TARGET_CONF_TEMP}
  80. "#define ${CORE}\n"
  81. "#define CHAR_CORENAME \"${CORE}\"\n")
  82. if ("${CORE}" STREQUAL "ARMV7")
  83. file(APPEND ${TARGET_CONF_TEMP}
  84. "#define L1_DATA_SIZE\t65536\n"
  85. "#define L1_DATA_LINESIZE\t32\n"
  86. "#define L2_SIZE\t512488\n"
  87. "#define L2_LINESIZE\t32\n"
  88. "#define DTB_DEFAULT_ENTRIES\t64\n"
  89. "#define DTB_SIZE\t4096\n"
  90. "#define L2_ASSOCIATIVE\t4\n"
  91. "#define HAVE_VFPV3\n"
  92. "#define HAVE_VFP\n")
  93. set(SGEMM_UNROLL_M 4)
  94. set(SGEMM_UNROLL_N 4)
  95. set(DGEMM_UNROLL_M 4)
  96. set(DGEMM_UNROLL_N 4)
  97. elseif ("${CORE}" STREQUAL "ARMV8")
  98. file(APPEND ${TARGET_CONF_TEMP}
  99. "#define L1_DATA_SIZE\t32768\n"
  100. "#define L1_DATA_LINESIZE\t64\n"
  101. "#define L2_SIZE\t262144\n"
  102. "#define L2_LINESIZE\t64\n"
  103. "#define DTB_DEFAULT_ENTRIES\t64\n"
  104. "#define DTB_SIZE\t4096\n"
  105. "#define L2_ASSOCIATIVE\t32\n"
  106. "#define ARMV8\n")
  107. set(SGEMM_UNROLL_M 4)
  108. set(SGEMM_UNROLL_N 4)
  109. elseif ("${CORE}" STREQUAL "CORTEXA57" OR "${CORE}" STREQUAL "CORTEXA53")
  110. file(APPEND ${TARGET_CONF_TEMP}
  111. "#define L1_CODE_SIZE\t32768\n"
  112. "#define L1_CODE_LINESIZE\t64\n"
  113. "#define L1_CODE_ASSOCIATIVE\t3\n"
  114. "#define L1_DATA_SIZE\t32768\n"
  115. "#define L1_DATA_LINESIZE\t64\n"
  116. "#define L1_DATA_ASSOCIATIVE\t2\n"
  117. "#define L2_SIZE\t262144\n"
  118. "#define L2_LINESIZE\t64\n"
  119. "#define L2_ASSOCIATIVE\t16\n"
  120. "#define DTB_DEFAULT_ENTRIES\t64\n"
  121. "#define DTB_SIZE\t4096\n"
  122. "#define HAVE_VFPV4\n"
  123. "#define HAVE_VFPV3\n"
  124. "#define HAVE_VFP\n"
  125. "#define HAVE_NEON\n"
  126. "#define ARMV8\n")
  127. set(SGEMM_UNROLL_M 16)
  128. set(SGEMM_UNROLL_N 4)
  129. set(DGEMM_UNROLL_M 8)
  130. set(DGEMM_UNROLL_N 4)
  131. set(CGEMM_UNROLL_M 8)
  132. set(CGEMM_UNROLL_N 4)
  133. set(ZGEMM_UNROLL_M 8)
  134. set(ZGEMM_UNROLL_N 4)
  135. elseif ("${CORE}" STREQUAL "CORTEXA72" OR "${CORE}" STREQUAL "CORTEXA73")
  136. file(APPEND ${TARGET_CONF_TEMP}
  137. "#define L1_CODE_SIZE\t49152\n"
  138. "#define L1_CODE_LINESIZE\t64\n"
  139. "#define L1_CODE_ASSOCIATIVE\t3\n"
  140. "#define L1_DATA_SIZE\t32768\n"
  141. "#define L1_DATA_LINESIZE\t64\n"
  142. "#define L1_DATA_ASSOCIATIVE\t2\n"
  143. "#define L2_SIZE\t524288\n"
  144. "#define L2_LINESIZE\t64\n"
  145. "#define L2_ASSOCIATIVE\t16\n"
  146. "#define DTB_DEFAULT_ENTRIES\t64\n"
  147. "#define DTB_SIZE\t4096\n"
  148. "#define HAVE_VFPV4\n"
  149. "#define HAVE_VFPV3\n"
  150. "#define HAVE_VFP\n"
  151. "#define HAVE_NEON\n"
  152. "#define ARMV8\n")
  153. set(SGEMM_UNROLL_M 16)
  154. set(SGEMM_UNROLL_N 4)
  155. set(DGEMM_UNROLL_M 8)
  156. set(DGEMM_UNROLL_N 4)
  157. set(CGEMM_UNROLL_M 8)
  158. set(CGEMM_UNROLL_N 4)
  159. set(ZGEMM_UNROLL_M 8)
  160. set(ZGEMM_UNROLL_N 4)
  161. elseif ("${CORE}" STREQUAL "FALKOR")
  162. file(APPEND ${TARGET_CONF_TEMP}
  163. "#define L1_CODE_SIZE\t65536\n"
  164. "#define L1_CODE_LINESIZE\t64\n"
  165. "#define L1_CODE_ASSOCIATIVE\t3\n"
  166. "#define L1_DATA_SIZE\t32768\n"
  167. "#define L1_DATA_LINESIZE\t128\n"
  168. "#define L1_DATA_ASSOCIATIVE\t2\n"
  169. "#define L2_SIZE\t524288\n"
  170. "#define L2_LINESIZE\t64\n"
  171. "#define L2_ASSOCIATIVE\t16\n"
  172. "#define DTB_DEFAULT_ENTRIES\t64\n"
  173. "#define DTB_SIZE\t4096\n"
  174. "#define HAVE_VFPV4\n"
  175. "#define HAVE_VFPV3\n"
  176. "#define HAVE_VFP\n"
  177. "#define HAVE_NEON\n"
  178. "#define ARMV8\n")
  179. set(SGEMM_UNROLL_M 16)
  180. set(SGEMM_UNROLL_N 4)
  181. set(DGEMM_UNROLL_M 8)
  182. set(DGEMM_UNROLL_N 4)
  183. set(CGEMM_UNROLL_M 8)
  184. set(CGEMM_UNROLL_N 4)
  185. set(ZGEMM_UNROLL_M 8)
  186. set(ZGEMM_UNROLL_N 4)
  187. elseif ("${CORE}" STREQUAL "THUNDERX)
  188. file(APPEND ${TARGET_CONF_TEMP}
  189. "#define L1_CODE_SIZE\t32768\n"
  190. "#define L1_CODE_LINESIZE\t64\n"
  191. "#define L1_CODE_ASSOCIATIVE\t3\n"
  192. "#define L1_DATA_SIZE\t32768\n"
  193. "#define L1_DATA_LINESIZE\t128\n"
  194. "#define L1_DATA_ASSOCIATIVE\t2\n"
  195. "#define L2_SIZE\t167772164\n"
  196. "#define L2_LINESIZE\t128\n"
  197. "#define L2_ASSOCIATIVE\t16\n"
  198. "#define DTB_DEFAULT_ENTRIES\t64\n"
  199. "#define DTB_SIZE\t4096\n"
  200. "#define HAVE_VFPV4\n"
  201. "#define HAVE_VFPV3\n"
  202. "#define HAVE_VFP\n"
  203. "#define HAVE_NEON\n"
  204. "#define ARMV8\n")
  205. set(SGEMM_UNROLL_M 4)
  206. set(SGEMM_UNROLL_N 4)
  207. set(DGEMM_UNROLL_M 2)
  208. set(DGEMM_UNROLL_N 2)
  209. set(CGEMM_UNROLL_M 2)
  210. set(CGEMM_UNROLL_N 2)
  211. set(ZGEMM_UNROLL_M 2)
  212. set(ZGEMM_UNROLL_N 2)
  213. elseif ("${CORE}" STREQUAL "THUNDERX2T99)
  214. file(APPEND ${TARGET_CONF_TEMP}
  215. "#define L1_CODE_SIZE\t32768\n"
  216. "#define L1_CODE_LINESIZE\t64\n"
  217. "#define L1_CODE_ASSOCIATIVE\t8\n"
  218. "#define L1_DATA_SIZE\t32768\n"
  219. "#define L1_DATA_LINESIZE\t64\n"
  220. "#define L1_DATA_ASSOCIATIVE\t8\n"
  221. "#define L2_SIZE\t262144\n"
  222. "#define L2_LINESIZE\t64\n"
  223. "#define L2_ASSOCIATIVE\t8\n"
  224. "#define L3_SIZE\t33554432\n"
  225. "#define L3_LINESIZE\t64\n"
  226. "#define L3_ASSOCIATIVE\t32\n"
  227. "#define DTB_DEFAULT_ENTRIES\t64\n"
  228. "#define DTB_SIZE\t4096\n"
  229. "#define VULCAN\n")
  230. set(SGEMM_UNROLL_M 16)
  231. set(SGEMM_UNROLL_N 4)
  232. set(DGEMM_UNROLL_M 8)
  233. set(DGEMM_UNROLL_N 4)
  234. set(CGEMM_UNROLL_M 8)
  235. set(CGEMM_UNROLL_N 4)
  236. set(ZGEMM_UNROLL_M 4)
  237. set(ZGEMM_UNROLL_N 4)
  238. endif()
  239. # Or should this actually be NUM_CORES?
  240. if (${NUM_THREADS} GREATER 0)
  241. file(APPEND ${TARGET_CONF_TEMP} "#define NUM_CORES\t${NUM_THREADS}\n")
  242. endif()
  243. # GetArch_2nd
  244. foreach(float_char S;D;Q;C;Z;X)
  245. if (NOT DEFINED ${float_char}GEMM_UNROLL_M)
  246. set(${float_char}GEMM_UNROLL_M 2)
  247. endif()
  248. if (NOT DEFINED ${float_char}GEMM_UNROLL_N)
  249. set(${float_char}GEMM_UNROLL_N 2)
  250. endif()
  251. endforeach()
  252. file(APPEND ${TARGET_CONF_TEMP}
  253. "#define GEMM_MULTITHREAD_THRESHOLD\t${GEMM_MULTITHREAD_THRESHOLD}\n")
  254. # Move to where gen_config_h would place it
  255. file(MAKE_DIRECTORY ${TARGET_CONF_DIR})
  256. file(RENAME ${TARGET_CONF_TEMP} "${TARGET_CONF_DIR}/${TARGET_CONF}")
  257. else(NOT CMAKE_CROSSCOMPILING)
  258. # compile getarch
  259. set(GETARCH_SRC
  260. ${PROJECT_SOURCE_DIR}/getarch.c
  261. ${CPUIDEMU}
  262. )
  263. if ("${CMAKE_C_COMPILER_ID}" STREQUAL "MSVC")
  264. #Use generic for MSVC now
  265. message("MSVC")
  266. set(GETARCH_FLAGS ${GETARCH_FLAGS} -DFORCE_GENERIC)
  267. else()
  268. list(APPEND GETARCH_SRC ${PROJECT_SOURCE_DIR}/cpuid.S)
  269. endif ()
  270. if ("${CMAKE_SYSTEM_NAME}" STREQUAL "WindowsStore")
  271. # disable WindowsStore strict CRT checks
  272. set(GETARCH_FLAGS ${GETARCH_FLAGS} -D_CRT_SECURE_NO_WARNINGS)
  273. endif ()
  274. set(GETARCH_DIR "${PROJECT_BINARY_DIR}/getarch_build")
  275. set(GETARCH_BIN "getarch${CMAKE_EXECUTABLE_SUFFIX}")
  276. file(MAKE_DIRECTORY ${GETARCH_DIR})
  277. configure_file(${TARGET_CONF_TEMP} ${GETARCH_DIR}/${TARGET_CONF} COPYONLY)
  278. if (NOT "${CMAKE_SYSTEM_NAME}" STREQUAL "WindowsStore")
  279. try_compile(GETARCH_RESULT ${GETARCH_DIR}
  280. SOURCES ${GETARCH_SRC}
  281. COMPILE_DEFINITIONS ${EXFLAGS} ${GETARCH_FLAGS} -I${GETARCH_DIR} -I"${PROJECT_SOURCE_DIR}" -I"${PROJECT_BINARY_DIR}"
  282. OUTPUT_VARIABLE GETARCH_LOG
  283. COPY_FILE ${PROJECT_BINARY_DIR}/${GETARCH_BIN}
  284. )
  285. if (NOT ${GETARCH_RESULT})
  286. MESSAGE(FATAL_ERROR "Compiling getarch failed ${GETARCH_LOG}")
  287. endif ()
  288. endif ()
  289. message(STATUS "Running getarch")
  290. # use the cmake binary w/ the -E param to run a shell command in a cross-platform way
  291. execute_process(COMMAND "${PROJECT_BINARY_DIR}/${GETARCH_BIN}" 0 OUTPUT_VARIABLE GETARCH_MAKE_OUT)
  292. execute_process(COMMAND "${PROJECT_BINARY_DIR}/${GETARCH_BIN}" 1 OUTPUT_VARIABLE GETARCH_CONF_OUT)
  293. message(STATUS "GETARCH results:\n${GETARCH_MAKE_OUT}")
  294. # append config data from getarch to the TARGET file and read in CMake vars
  295. file(APPEND ${TARGET_CONF_TEMP} ${GETARCH_CONF_OUT})
  296. ParseGetArchVars(${GETARCH_MAKE_OUT})
  297. set(GETARCH2_DIR "${PROJECT_BINARY_DIR}/getarch2_build")
  298. set(GETARCH2_BIN "getarch_2nd${CMAKE_EXECUTABLE_SUFFIX}")
  299. file(MAKE_DIRECTORY ${GETARCH2_DIR})
  300. configure_file(${TARGET_CONF_TEMP} ${GETARCH2_DIR}/${TARGET_CONF} COPYONLY)
  301. if (NOT "${CMAKE_SYSTEM_NAME}" STREQUAL "WindowsStore")
  302. try_compile(GETARCH2_RESULT ${GETARCH2_DIR}
  303. SOURCES ${PROJECT_SOURCE_DIR}/getarch_2nd.c
  304. COMPILE_DEFINITIONS ${EXFLAGS} ${GETARCH_FLAGS} ${GETARCH2_FLAGS} -I${GETARCH2_DIR} -I"${PROJECT_SOURCE_DIR}" -I"${PROJECT_BINARY_DIR}"
  305. OUTPUT_VARIABLE GETARCH2_LOG
  306. COPY_FILE ${PROJECT_BINARY_DIR}/${GETARCH2_BIN}
  307. )
  308. if (NOT ${GETARCH2_RESULT})
  309. MESSAGE(FATAL_ERROR "Compiling getarch_2nd failed ${GETARCH2_LOG}")
  310. endif ()
  311. endif ()
  312. # use the cmake binary w/ the -E param to run a shell command in a cross-platform way
  313. execute_process(COMMAND "${PROJECT_BINARY_DIR}/${GETARCH2_BIN}" 0 OUTPUT_VARIABLE GETARCH2_MAKE_OUT)
  314. execute_process(COMMAND "${PROJECT_BINARY_DIR}/${GETARCH2_BIN}" 1 OUTPUT_VARIABLE GETARCH2_CONF_OUT)
  315. # append config data from getarch_2nd to the TARGET file and read in CMake vars
  316. file(APPEND ${TARGET_CONF_TEMP} ${GETARCH2_CONF_OUT})
  317. configure_file(${TARGET_CONF_TEMP} ${TARGET_CONF_DIR}/${TARGET_CONF} COPYONLY)
  318. ParseGetArchVars(${GETARCH2_MAKE_OUT})
  319. endif()