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.

system.cmake 21 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746
  1. ##
  2. ## Author: Hank Anderson <hank@statease.com>
  3. ## Description: Ported from OpenBLAS/Makefile.system
  4. ##
  5. set(NETLIB_LAPACK_DIR "${PROJECT_SOURCE_DIR}/lapack-netlib")
  6. # System detection, via CMake.
  7. include("${PROJECT_SOURCE_DIR}/cmake/system_check.cmake")
  8. if(CMAKE_CROSSCOMPILING AND NOT DEFINED TARGET)
  9. # Detect target without running getarch
  10. if (ARM64)
  11. set(TARGET "ARMV8")
  12. elseif(ARM)
  13. set(TARGET "ARMV7") # TODO: Ask compiler which arch this is
  14. else()
  15. message(FATAL_ERROR "When cross compiling, a TARGET is required.")
  16. endif()
  17. endif()
  18. # Other files expect CORE, which is actually TARGET and will become TARGET_CORE for kernel build. Confused yet?
  19. # It seems we are meant to use TARGET as input and CORE internally as kernel.
  20. if(NOT DEFINED CORE AND DEFINED TARGET)
  21. set(CORE ${TARGET})
  22. endif()
  23. # TARGET_CORE will override TARGET which is used in DYNAMIC_ARCH=1.
  24. if (DEFINED TARGET_CORE)
  25. set(TARGET ${TARGET_CORE})
  26. endif ()
  27. # Force fallbacks for 32bit
  28. if (DEFINED BINARY AND DEFINED TARGET AND BINARY EQUAL 32)
  29. message(STATUS "Compiling a ${BINARY}-bit binary.")
  30. set(NO_AVX 1)
  31. if (${TARGET} STREQUAL "HASWELL" OR ${TARGET} STREQUAL "SANDYBRIDGE" OR ${TARGET} STREQUAL "SKYLAKEX" OR ${TARGET} STREQUAL "COOPERLAKE" OR ${TARGET} STREQUAL "SAPPHIRERAPIDS")
  32. set(TARGET "NEHALEM")
  33. endif ()
  34. if (${TARGET} STREQUAL "BULLDOZER" OR ${TARGET} STREQUAL "PILEDRIVER" OR ${TARGET} STREQUAL "ZEN")
  35. set(TARGET "BARCELONA")
  36. endif ()
  37. if (${TARGET} STREQUAL "ARMV8" OR ${TARGET} STREQUAL "CORTEXA57" OR ${TARGET} STREQUAL "CORTEXA53" OR ${TARGET} STREQUAL "CORTEXA55")
  38. set(TARGET "ARMV7")
  39. endif ()
  40. if (${TARGET} STREQUAL "POWER8" OR ${TARGET} STREQUAL "POWER9" OR ${TARGET} STREQUAL "POWER10")
  41. set(TARGET "POWER6")
  42. endif ()
  43. endif ()
  44. if (DEFINED TARGET)
  45. message(STATUS "-- -- -- -- -- -- -- -- -- -- -- -- --")
  46. message(STATUS "Targeting the ${TARGET} architecture.")
  47. set(GETARCH_FLAGS "-DFORCE_${TARGET}")
  48. endif ()
  49. # On x86_64 build getarch with march=native. This is required to detect AVX512 support in getarch.
  50. if (X86_64 AND NOT ${CMAKE_C_COMPILER_ID} STREQUAL "PGI")
  51. set(GETARCH_FLAGS "${GETARCH_FLAGS} -march=native")
  52. endif ()
  53. # On x86 no AVX support is available
  54. if (X86 OR X86_64)
  55. if ((DEFINED BINARY AND BINARY EQUAL 32) OR ("$CMAKE_SIZEOF_VOID_P}" EQUAL "4"))
  56. set(GETARCH_FLAGS "${GETARCH_FLAGS} -DNO_AVX -DNO_AVX2 -DNO_AVX512")
  57. endif ()
  58. endif ()
  59. if (INTERFACE64)
  60. message(STATUS "Using 64-bit integers.")
  61. set(GETARCH_FLAGS "${GETARCH_FLAGS} -DUSE64BITINT")
  62. endif ()
  63. if (NOT DEFINED GEMM_MULTITHREAD_THRESHOLD)
  64. set(GEMM_MULTITHREAD_THRESHOLD 4)
  65. endif ()
  66. message(STATUS "GEMM multithread threshold set to ${GEMM_MULTITHREAD_THRESHOLD}.")
  67. set(GETARCH_FLAGS "${GETARCH_FLAGS} -DGEMM_MULTITHREAD_THRESHOLD=${GEMM_MULTITHREAD_THRESHOLD}")
  68. if (NO_AVX)
  69. message(STATUS "Disabling Advanced Vector Extensions (AVX).")
  70. set(GETARCH_FLAGS "${GETARCH_FLAGS} -DNO_AVX")
  71. endif ()
  72. if (NO_AVX2)
  73. message(STATUS "Disabling Advanced Vector Extensions 2 (AVX2).")
  74. set(GETARCH_FLAGS "${GETARCH_FLAGS} -DNO_AVX2")
  75. endif ()
  76. if (NO_AVX512)
  77. message(STATUS "Disabling Advanced Vector Extensions 512 (AVX512).")
  78. set(GETARCH_FLAGS "${GETARCH_FLAGS} -DNO_AVX512")
  79. endif ()
  80. if (CMAKE_BUILD_TYPE STREQUAL "Debug")
  81. set(GETARCH_FLAGS "${GETARCH_FLAGS} ${CMAKE_C_FLAGS_DEBUG}")
  82. endif ()
  83. if (NOT DEFINED NO_PARALLEL_MAKE)
  84. set(NO_PARALLEL_MAKE 0)
  85. endif ()
  86. set(GETARCH_FLAGS "${GETARCH_FLAGS} -DNO_PARALLEL_MAKE=${NO_PARALLEL_MAKE}")
  87. if (CMAKE_C_COMPILER STREQUAL loongcc)
  88. set(GETARCH_FLAGS "${GETARCH_FLAGS} -static")
  89. endif ()
  90. if (POWER)
  91. set(NO_WARMUP 1)
  92. set(HAVE_GAS 1)
  93. if (CMAKE_ASM_COMPILER_ID STREQUAL "GNU")
  94. set(HAVE_GAS 0)
  95. elseif (CMAKE_ASM_COMPILER_ID STREQUAL "Clang")
  96. set(CCOMMON_OPT "${CCOMMON_OPT} -fno-integrated-as")
  97. set(HAVE_GAS 0)
  98. endif ()
  99. set(GETARCH_FLAGS "${GETARCH_FLAGS} -DHAVE_GAS=${HAVE_GAS}")
  100. endif ()
  101. #if don't use Fortran, it will only compile CBLAS.
  102. if (ONLY_CBLAS)
  103. set(NO_LAPACK 1)
  104. else ()
  105. set(ONLY_CBLAS 0)
  106. endif ()
  107. # N.B. this is NUM_THREAD in Makefile.system which is probably a bug -hpa
  108. if (NOT CMAKE_CROSSCOMPILING)
  109. if (NOT DEFINED NUM_CORES)
  110. include(ProcessorCount)
  111. ProcessorCount(NUM_CORES)
  112. endif()
  113. endif()
  114. if (NOT DEFINED NUM_PARALLEL)
  115. set(NUM_PARALLEL 1)
  116. endif()
  117. if (NOT DEFINED NUM_THREADS)
  118. if (DEFINED NUM_CORES AND NOT NUM_CORES EQUAL 0)
  119. # HT?
  120. set(NUM_THREADS ${NUM_CORES})
  121. else ()
  122. set(NUM_THREADS 0)
  123. endif ()
  124. endif()
  125. if (${NUM_THREADS} LESS 2)
  126. set(USE_THREAD 0)
  127. elseif(NOT DEFINED USE_THREAD)
  128. set(USE_THREAD 1)
  129. endif ()
  130. if (USE_THREAD)
  131. message(STATUS "Multi-threading enabled with ${NUM_THREADS} threads.")
  132. else()
  133. if (${USE_LOCKING})
  134. set(CCOMMON_OPT "${CCOMMON_OPT} -DUSE_LOCKING")
  135. endif ()
  136. endif ()
  137. include("${PROJECT_SOURCE_DIR}/cmake/prebuild.cmake")
  138. if (DEFINED TARGET)
  139. if (${TARGET} STREQUAL COOPERLAKE AND NOT NO_AVX512)
  140. if (${CMAKE_C_COMPILER_ID} STREQUAL "GNU")
  141. execute_process(COMMAND ${CMAKE_C_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
  142. if (${CMAKE_C_COMPILER_VERSION} VERSION_GREATER 10.09)
  143. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=cooperlake")
  144. else()
  145. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=skylake-avx512")
  146. endif()
  147. elseif (${CMAKE_C_COMPILER_ID} STREQUAL "Clang" OR ${CMAKE_C_COMPILER_ID} STREQUAL "AppleClang")
  148. if (${CMAKE_C_COMPILER_VERSION} VERSION_GREATER 8.99)
  149. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=cooperlake")
  150. else()
  151. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=skylake-avx512")
  152. endif()
  153. endif()
  154. endif()
  155. if (${TARGET} STREQUAL SAPPHIRERAPIDS AND NOT NO_AVX512)
  156. if (${CMAKE_C_COMPILER_ID} STREQUAL "GNU")
  157. execute_process(COMMAND ${CMAKE_C_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
  158. if (${CMAKE_C_COMPILER_VERSION} VERSION_GREATER 11.0)
  159. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=sapphirerapids")
  160. else()
  161. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=skylake-avx512")
  162. endif()
  163. elseif (${CMAKE_C_COMPILER_ID} STREQUAL "Clang" OR ${CMAKE_C_COMPILER_ID} STREQUAL "AppleClang")
  164. if (${CMAKE_C_COMPILER_VERSION} VERSION_GREATER 12.0)
  165. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=sapphirerapids")
  166. else()
  167. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=skylake-avx512")
  168. endif()
  169. endif()
  170. endif()
  171. if (${TARGET} STREQUAL SKYLAKEX AND NOT NO_AVX512)
  172. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=skylake-avx512")
  173. endif()
  174. if (${TARGET} STREQUAL HASWELL AND NOT NO_AVX2)
  175. if (${CMAKE_C_COMPILER_ID} STREQUAL "GNU")
  176. execute_process(COMMAND ${CMAKE_C_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
  177. if (${GCC_VERSION} VERSION_GREATER 4.7 OR ${GCC_VERSION} VERSION_EQUAL 4.7)
  178. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mavx2")
  179. endif()
  180. elseif (${CMAKE_C_COMPILER_ID} STREQUAL "CLANG")
  181. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mavx2")
  182. endif()
  183. endif()
  184. if (DEFINED HAVE_AVX)
  185. if (NOT NO_AVX)
  186. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mavx")
  187. endif()
  188. endif()
  189. if (DEFINED HAVE_AVX2)
  190. if (NOT NO_AVX2)
  191. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mavx2")
  192. endif()
  193. endif()
  194. # if (DEFINED HAVE_FMA3)
  195. # if (NOT NO_AVX2)
  196. # set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mfma")
  197. # endif()
  198. # endif()
  199. if (DEFINED HAVE_SSE)
  200. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -msse")
  201. endif()
  202. if (DEFINED HAVE_SSE2)
  203. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -msse2")
  204. endif()
  205. if (DEFINED HAVE_SSE3)
  206. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -msse3")
  207. endif()
  208. if (DEFINED HAVE_SSSE3)
  209. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mssse3")
  210. endif()
  211. if (DEFINED HAVE_SSE4_1)
  212. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -msse4.1")
  213. endif()
  214. if (${TARGET} STREQUAL POWER10)
  215. execute_process(COMMAND ${CMAKE_C_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
  216. if (${GCC_VERSION} VERSION_GREATER 10.2 OR ${GCC_VERSION} VERSION_EQUAL 10.2)
  217. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mcpu=power10 -mtune=power10 -mvsx -fno-fast-math")
  218. else ()
  219. message(FATAL_ERROR "Compiler GCC.${GCC_VERSION} does not support Power10.")
  220. endif()
  221. endif()
  222. if (${TARGET} STREQUAL POWER9)
  223. execute_process(COMMAND ${CMAKE_C_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
  224. if (${GCC_VERSION} VERSION_GREATER 5.0 OR ${GCC_VERSION} VERSION_EQUAL 5.0)
  225. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mcpu=power9 -mtune=power9 -mvsx -fno-fast-math")
  226. else ()
  227. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mcpu=power8 -mtune=power8 -mvsx -fno-fast-math")
  228. message(WARNING "Compiler GCC.${GCC_VERSION} does not support fully Power9.")
  229. endif()
  230. endif()
  231. if (${TARGET} STREQUAL POWER8)
  232. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mcpu=power8 -mtune=power8 -mvsx -fno-fast-math")
  233. endif()
  234. endif()
  235. if (DEFINED BINARY)
  236. message(STATUS "Compiling a ${BINARY}-bit binary.")
  237. endif ()
  238. if (NOT DEFINED NEED_PIC)
  239. set(NEED_PIC 1)
  240. endif ()
  241. # OS dependent settings
  242. include("${PROJECT_SOURCE_DIR}/cmake/os.cmake")
  243. # Architecture dependent settings
  244. include("${PROJECT_SOURCE_DIR}/cmake/arch.cmake")
  245. # C Compiler dependent settings
  246. include("${PROJECT_SOURCE_DIR}/cmake/cc.cmake")
  247. if (INTERFACE64)
  248. set(SUFFIX64 64)
  249. set(SUFFIX64_UNDERSCORE _64)
  250. endif()
  251. if (NOT NOFORTRAN)
  252. # Fortran Compiler dependent settings
  253. include("${PROJECT_SOURCE_DIR}/cmake/fc.cmake")
  254. else ()
  255. if (NOT XXXX)
  256. set(C_LAPACK 1)
  257. if (INTERFACE64)
  258. set (CCOMMON_OPT "${CCOMMON_OPT} -DLAPACK_ILP64")
  259. endif ()
  260. set(TIMER "NONE")
  261. else ()
  262. set (NO_LAPACK 1)
  263. endif ()
  264. endif ()
  265. if (BINARY64)
  266. if (INTERFACE64)
  267. # CCOMMON_OPT += -DUSE64BITINT
  268. endif ()
  269. endif ()
  270. if(EMBEDDED)
  271. set(CCOMMON_OPT "${CCOMMON_OPT} -DOS_EMBEDDED")
  272. set(CCOMMON_OPT "${CCOMMON_OPT} -mthumb -mcpu=cortex-m4 -mfloat-abi=hard -mfpu=fpv4-sp-d16")
  273. endif()
  274. if (NEED_PIC)
  275. if (${CMAKE_C_COMPILER} STREQUAL "IBM")
  276. set(CCOMMON_OPT "${CCOMMON_OPT} -qpic=large")
  277. else ()
  278. set(CCOMMON_OPT "${CCOMMON_OPT} -fPIC")
  279. endif ()
  280. if (NOT NOFORTRAN)
  281. if (${F_COMPILER} STREQUAL "SUN")
  282. set(FCOMMON_OPT "${FCOMMON_OPT} -pic")
  283. else ()
  284. set(FCOMMON_OPT "${FCOMMON_OPT} -fPIC")
  285. endif ()
  286. endif()
  287. endif ()
  288. if (X86_64 OR ${CORE} STREQUAL POWER10)
  289. set(SMALL_MATRIX_OPT TRUE)
  290. endif ()
  291. if (SMALL_MATRIX_OPT)
  292. set(CCOMMON_OPT "${CCOMMON_OPT} -DSMALL_MATRIX_OPT")
  293. endif ()
  294. if (DYNAMIC_ARCH)
  295. if (X86 OR X86_64 OR ARM64 OR POWER)
  296. set(CCOMMON_OPT "${CCOMMON_OPT} -DDYNAMIC_ARCH")
  297. if (DYNAMIC_OLDER)
  298. set(CCOMMON_OPT "${CCOMMON_OPT} -DDYNAMIC_OLDER")
  299. endif ()
  300. else ()
  301. unset (DYNAMIC_ARCH)
  302. message (STATUS "DYNAMIC_ARCH is not supported on the target architecture, removing")
  303. endif ()
  304. endif ()
  305. if (DYNAMIC_LIST)
  306. set(CCOMMON_OPT "${CCOMMON_OPT} -DDYNAMIC_LIST")
  307. foreach(DCORE ${DYNAMIC_LIST})
  308. set(CCOMMON_OPT "${CCOMMON_OPT} -DDYN_${DCORE}")
  309. endforeach ()
  310. endif ()
  311. if (NO_LAPACK)
  312. set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_LAPACK")
  313. #Disable LAPACK C interface
  314. set(NO_LAPACKE 1)
  315. endif ()
  316. if (NO_LAPACKE)
  317. set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_LAPACKE")
  318. endif ()
  319. if (NO_AVX)
  320. set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX")
  321. endif ()
  322. if (X86)
  323. set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX")
  324. endif ()
  325. if (NO_AVX2)
  326. set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX2")
  327. endif ()
  328. if (NO_AVX512)
  329. set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX512")
  330. endif ()
  331. if (USE_THREAD)
  332. # USE_SIMPLE_THREADED_LEVEL3 = 1
  333. # NO_AFFINITY = 1
  334. set(CCOMMON_OPT "${CCOMMON_OPT} -DSMP_SERVER")
  335. if (MIPS64)
  336. if (NOT ${CORE} STREQUAL "LOONGSON3B")
  337. set(USE_SIMPLE_THREADED_LEVEL3 1)
  338. endif ()
  339. endif ()
  340. if (BIGNUMA)
  341. set(CCOMMON_OPT "${CCOMMON_OPT} -DBIGNUMA")
  342. endif ()
  343. endif ()
  344. if (NO_WARMUP)
  345. set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_WARMUP")
  346. endif ()
  347. if (CONSISTENT_FPCSR)
  348. set(CCOMMON_OPT "${CCOMMON_OPT} -DCONSISTENT_FPCSR")
  349. endif ()
  350. if (USE_TLS)
  351. set(CCOMMON_OPT "${CCOMMON_OPT} -DUSE_TLS")
  352. endif ()
  353. # Only for development
  354. # set(CCOMMON_OPT "${CCOMMON_OPT} -DPARAMTEST")
  355. # set(CCOMMON_OPT "${CCOMMON_OPT} -DPREFETCHTEST")
  356. # set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_SWITCHING")
  357. # set(USE_PAPI 1)
  358. if (USE_PAPI)
  359. set(CCOMMON_OPT "${CCOMMON_OPT} -DUSE_PAPI")
  360. set(EXTRALIB "${EXTRALIB} -lpapi -lperfctr")
  361. endif ()
  362. if (DYNAMIC_THREADS)
  363. set(CCOMMON_OPT "${CCOMMON_OPT} -DDYNAMIC_THREADS")
  364. endif ()
  365. set(CCOMMON_OPT "${CCOMMON_OPT} -DMAX_CPU_NUMBER=${NUM_THREADS}")
  366. set(CCOMMON_OPT "${CCOMMON_OPT} -DMAX_PARALLEL_NUMBER=${NUM_PARALLEL}")
  367. if (BUFFERSIZE)
  368. set(CCOMMON_OPT "${CCOMMON_OPT} -DBUFFERSIZE=${BUFFERSIZE}")
  369. endif ()
  370. if (USE_SIMPLE_THREADED_LEVEL3)
  371. set(CCOMMON_OPT "${CCOMMON_OPT} -DUSE_SIMPLE_THREADED_LEVEL3")
  372. endif ()
  373. if (NOT ${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
  374. if (DEFINED MAX_STACK_ALLOC)
  375. if (NOT ${MAX_STACK_ALLOC} EQUAL 0)
  376. set(CCOMMON_OPT "${CCOMMON_OPT} -DMAX_STACK_ALLOC=${MAX_STACK_ALLOC}")
  377. endif ()
  378. else ()
  379. set(CCOMMON_OPT "${CCOMMON_OPT} -DMAX_STACK_ALLOC=2048")
  380. endif ()
  381. endif ()
  382. if (NOT ${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
  383. if (DEFINED BLAS3_MEM_ALLOC_THRESHOLD)
  384. if (NOT ${BLAS3_MEM_ALLOC_THRESHOLD} EQUAL 32)
  385. set(CCOMMON_OPT "${CCOMMON_OPT} -DBLAS3_MEM_ALLOC_THRESHOLD=${BLAS3_MEM_ALLOC_THRESHOLD}")
  386. endif()
  387. endif()
  388. endif()
  389. if (DEFINED LIBNAMESUFFIX)
  390. set(LIBPREFIX "libopenblas_${LIBNAMESUFFIX}")
  391. else ()
  392. set(LIBPREFIX "libopenblas")
  393. endif ()
  394. if (NOT DEFINED SYMBOLPREFIX)
  395. set(SYMBOLPREFIX "")
  396. endif ()
  397. if (NOT DEFINED SYMBOLSUFFIX)
  398. set(SYMBOLSUFFIX "")
  399. endif ()
  400. set(KERNELDIR "${PROJECT_SOURCE_DIR}/kernel/${ARCH}")
  401. # TODO: need to convert these Makefiles
  402. # include ${PROJECT_SOURCE_DIR}/cmake/${ARCH}.cmake
  403. if (${CORE} STREQUAL "PPC440")
  404. set(CCOMMON_OPT "${CCOMMON_OPT} -DALLOC_QALLOC")
  405. endif ()
  406. if (${CORE} STREQUAL "PPC440FP2")
  407. set(STATIC_ALLOCATION 1)
  408. endif ()
  409. if (NOT ${CMAKE_SYSTEM_NAME} STREQUAL "Linux")
  410. set(NO_AFFINITY 1)
  411. endif ()
  412. if (NOT X86_64 AND NOT X86 AND NOT ${CORE} STREQUAL "LOONGSON3B")
  413. set(NO_AFFINITY 1)
  414. endif ()
  415. if (NO_AFFINITY)
  416. set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AFFINITY")
  417. endif ()
  418. if (FUNCTION_PROFILE)
  419. set(CCOMMON_OPT "${CCOMMON_OPT} -DFUNCTION_PROFILE")
  420. endif ()
  421. if (HUGETLB_ALLOCATION)
  422. set(CCOMMON_OPT "${CCOMMON_OPT} -DALLOC_HUGETLB")
  423. endif ()
  424. if (DEFINED HUGETLBFILE_ALLOCATION)
  425. set(CCOMMON_OPT "${CCOMMON_OPT} -DALLOC_HUGETLBFILE -DHUGETLB_FILE_NAME=${HUGETLBFILE_ALLOCATION})")
  426. endif ()
  427. if (STATIC_ALLOCATION)
  428. set(CCOMMON_OPT "${CCOMMON_OPT} -DALLOC_STATIC")
  429. endif ()
  430. if (DEVICEDRIVER_ALLOCATION)
  431. set(CCOMMON_OPT "${CCOMMON_OPT} -DALLOC_DEVICEDRIVER -DDEVICEDRIVER_NAME=\"/dev/mapper\"")
  432. endif ()
  433. if (MIXED_MEMORY_ALLOCATION)
  434. set(CCOMMON_OPT "${CCOMMON_OPT} -DMIXED_MEMORY_ALLOCATION")
  435. endif ()
  436. set(CCOMMON_OPT "${CCOMMON_OPT} -DVERSION=\"\\\"${OpenBLAS_VERSION}\\\"\"")
  437. set(REVISION "-r${OpenBLAS_VERSION}")
  438. set(MAJOR_VERSION ${OpenBLAS_MAJOR_VERSION})
  439. set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${CCOMMON_OPT}")
  440. if (NOT BUILD_SINGLE AND NOT BUILD_DOUBLE AND NOT BUILD_COMPLEX AND NOT BUILD_COMPLEX16)
  441. set (BUILD_SINGLE ON)
  442. set (BUILD_DOUBLE ON)
  443. set (BUILD_COMPLEX ON)
  444. set (BUILD_COMPLEX16 ON)
  445. endif()
  446. if (BUILD_SINGLE)
  447. set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_SINGLE")
  448. endif()
  449. if (BUILD_DOUBLE)
  450. set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_DOUBLE")
  451. endif()
  452. if (BUILD_COMPLEX)
  453. set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_COMPLEX")
  454. endif()
  455. if (BUILD_COMPLEX16)
  456. set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_COMPLEX16")
  457. endif()
  458. if (BUILD_BFLOAT16)
  459. set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_BFLOAT16")
  460. endif()
  461. if(NOT MSVC)
  462. set(CMAKE_ASM_FLAGS "${CMAKE_ASM_FLAGS} ${CCOMMON_OPT}")
  463. endif()
  464. # TODO: not sure what PFLAGS is -hpa
  465. set(PFLAGS "${PFLAGS} ${CCOMMON_OPT} -I${TOPDIR} -DPROFILE ${COMMON_PROF}")
  466. set(CMAKE_Fortran_FLAGS "${CMAKE_Fortran_FLAGS} ${FCOMMON_OPT}")
  467. # TODO: not sure what FPFLAGS is -hpa
  468. set(FPFLAGS "${FPFLAGS} ${FCOMMON_OPT} ${COMMON_PROF}")
  469. #For LAPACK Fortran codes.
  470. set(LAPACK_FFLAGS "${LAPACK_FFLAGS} ${CMAKE_Fortran_FLAGS}")
  471. set(LAPACK_FPFLAGS "${LAPACK_FPFLAGS} ${FPFLAGS}")
  472. #Disable -fopenmp for LAPACK Fortran codes on Windows.
  473. if (${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
  474. set(FILTER_FLAGS "-fopenmp;-mp;-openmp;-xopenmp=parallel")
  475. foreach (FILTER_FLAG ${FILTER_FLAGS})
  476. string(REPLACE ${FILTER_FLAG} "" LAPACK_FFLAGS ${LAPACK_FFLAGS})
  477. string(REPLACE ${FILTER_FLAG} "" LAPACK_FPFLAGS ${LAPACK_FPFLAGS})
  478. endforeach ()
  479. endif ()
  480. if ("${F_COMPILER}" STREQUAL "NAG" OR "${F_COMPILER}" STREQUAL "CRAY")
  481. set(FILTER_FLAGS "-msse3;-mssse3;-msse4.1;-mavx;-mavx2,-mskylake-avx512")
  482. foreach (FILTER_FLAG ${FILTER_FLAGS})
  483. string(REPLACE ${FILTER_FLAG} "" LAPACK_FFLAGS ${LAPACK_FFLAGS})
  484. string(REPLACE ${FILTER_FLAG} "" LAPACK_FPFLAGS ${LAPACK_FPFLAGS})
  485. endforeach ()
  486. endif ()
  487. if ("${F_COMPILER}" STREQUAL "GFORTRAN")
  488. # lapack-netlib is rife with uninitialized warnings -hpa
  489. set(LAPACK_FFLAGS "${LAPACK_FFLAGS} -Wno-maybe-uninitialized")
  490. endif ()
  491. set(LAPACK_CFLAGS "${CMAKE_C_CFLAGS} -DHAVE_LAPACK_CONFIG_H")
  492. if (INTERFACE64)
  493. set(LAPACK_CFLAGS "${LAPACK_CFLAGS} -DLAPACK_ILP64")
  494. endif ()
  495. if (${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
  496. set(LAPACK_CFLAGS "${LAPACK_CFLAGS} -DOPENBLAS_OS_WINDOWS")
  497. endif ()
  498. if (${CMAKE_C_COMPILER} STREQUAL "LSB" OR ${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
  499. set(LAPACK_CFLAGS "${LAPACK_CFLAGS} -DLAPACK_COMPLEX_STRUCTURE")
  500. endif ()
  501. if ("${CMAKE_BUILD_TYPE}" STREQUAL "Release")
  502. if ("${F_COMPILER}" STREQUAL "FLANG")
  503. if (${CMAKE_Fortran_COMPILER_VERSION} VERSION_LESS_EQUAL 3)
  504. set(CMAKE_Fortran_FLAGS_RELEASE "${CMAKE_Fortran_FLAGS_RELEASE} -fno-unroll-loops")
  505. endif ()
  506. endif ()
  507. endif ()
  508. if (NOT DEFINED SUFFIX)
  509. set(SUFFIX o)
  510. endif ()
  511. if (NOT DEFINED PSUFFIX)
  512. set(PSUFFIX po)
  513. endif ()
  514. if (NOT DEFINED LIBSUFFIX)
  515. set(LIBSUFFIX a)
  516. endif ()
  517. if (DYNAMIC_ARCH)
  518. if (USE_THREAD)
  519. set(LIBNAME "${LIBPREFIX}p${REVISION}.${LIBSUFFIX}")
  520. set(LIBNAME_P "${LIBPREFIX}p${REVISION}_p.${LIBSUFFIX}")
  521. else ()
  522. set(LIBNAME "${LIBPREFIX}${REVISION}.${LIBSUFFIX}")
  523. set(LIBNAME_P "${LIBPREFIX}${REVISION}_p.${LIBSUFFIX}")
  524. endif ()
  525. else ()
  526. if (USE_THREAD)
  527. set(LIBNAME "${LIBPREFIX}_${LIBCORE}p${REVISION}.${LIBSUFFIX}")
  528. set(LIBNAME_P "${LIBPREFIX}_${LIBCORE}p${REVISION}_p.${LIBSUFFIX}")
  529. else ()
  530. set(LIBNAME "${LIBPREFIX}_${LIBCORE}${REVISION}.${LIBSUFFIX}")
  531. set(LIBNAME_P "${LIBPREFIX}_${LIBCORE}${REVISION}_p.${LIBSUFFIX}")
  532. endif ()
  533. endif ()
  534. set(LIBDLLNAME "${LIBPREFIX}.dll")
  535. set(LIBSONAME "${LIBNAME}.${LIBSUFFIX}.so")
  536. set(LIBDYNNAME "${LIBNAME}.${LIBSUFFIX}.dylib")
  537. set(LIBDEFNAME "${LIBNAME}.${LIBSUFFIX}.def")
  538. set(LIBEXPNAME "${LIBNAME}.${LIBSUFFIX}.exp")
  539. set(LIBZIPNAME "${LIBNAME}.${LIBSUFFIX}.zip")
  540. set(LIBS "${PROJECT_SOURCE_DIR}/${LIBNAME}")
  541. set(LIBS_P "${PROJECT_SOURCE_DIR}/${LIBNAME_P}")
  542. set(LIB_COMPONENTS BLAS)
  543. if (NOT NO_CBLAS)
  544. set(LIB_COMPONENTS "${LIB_COMPONENTS} CBLAS")
  545. endif ()
  546. if (NOT NO_LAPACK)
  547. set(LIB_COMPONENTS "${LIB_COMPONENTS} LAPACK")
  548. if (NOT NO_LAPACKE)
  549. set(LIB_COMPONENTS "${LIB_COMPONENTS} LAPACKE")
  550. endif ()
  551. if (BUILD_RELAPACK)
  552. set(LIB_COMPONENTS "${LIB_COMPONENTS} ReLAPACK")
  553. endif ()
  554. endif ()
  555. if (ONLY_CBLAS)
  556. set(LIB_COMPONENTS CBLAS)
  557. endif ()
  558. # For GEMM3M
  559. set(USE_GEMM3M 0)
  560. if (DEFINED ARCH)
  561. if (X86 OR X86_64 OR ${ARCH} STREQUAL "ia64" OR MIPS64)
  562. set(USE_GEMM3M 1)
  563. endif ()
  564. if (${CORE} STREQUAL "generic")
  565. set(USE_GEMM3M 0)
  566. endif ()
  567. endif ()
  568. #export OSNAME
  569. #export ARCH
  570. #export CORE
  571. #export LIBCORE
  572. #export PGCPATH
  573. #export CONFIG
  574. #export CC
  575. #export FC
  576. #export BU
  577. #export FU
  578. #export NEED2UNDERSCORES
  579. #export USE_THREAD
  580. #export NUM_THREADS
  581. #export NUM_CORES
  582. #export SMP
  583. #export MAKEFILE_RULE
  584. #export NEED_PIC
  585. #export BINARY
  586. #export BINARY32
  587. #export BINARY64
  588. #export F_COMPILER
  589. #export C_COMPILER
  590. #export USE_OPENMP
  591. #export CROSS
  592. #export CROSS_SUFFIX
  593. #export NOFORTRAN
  594. #export NO_FBLAS
  595. #export EXTRALIB
  596. #export CEXTRALIB
  597. #export FEXTRALIB
  598. #export HAVE_SSE
  599. #export HAVE_SSE2
  600. #export HAVE_SSE3
  601. #export HAVE_SSSE3
  602. #export HAVE_SSE4_1
  603. #export HAVE_SSE4_2
  604. #export HAVE_SSE4A
  605. #export HAVE_SSE5
  606. #export HAVE_AVX
  607. #export HAVE_VFP
  608. #export HAVE_VFPV3
  609. #export HAVE_VFPV4
  610. #export HAVE_NEON
  611. #export KERNELDIR
  612. #export FUNCTION_PROFILE
  613. #export TARGET_CORE
  614. #
  615. #export SBGEMM_UNROLL_M
  616. #export SBGEMM_UNROLL_N
  617. #export SGEMM_UNROLL_M
  618. #export SGEMM_UNROLL_N
  619. #export DGEMM_UNROLL_M
  620. #export DGEMM_UNROLL_N
  621. #export QGEMM_UNROLL_M
  622. #export QGEMM_UNROLL_N
  623. #export CGEMM_UNROLL_M
  624. #export CGEMM_UNROLL_N
  625. #export ZGEMM_UNROLL_M
  626. #export ZGEMM_UNROLL_N
  627. #export XGEMM_UNROLL_M
  628. #export XGEMM_UNROLL_N
  629. #export CGEMM3M_UNROLL_M
  630. #export CGEMM3M_UNROLL_N
  631. #export ZGEMM3M_UNROLL_M
  632. #export ZGEMM3M_UNROLL_N
  633. #export XGEMM3M_UNROLL_M
  634. #export XGEMM3M_UNROLL_N
  635. #if (USE_CUDA)
  636. # export CUDADIR
  637. # export CUCC
  638. # export CUFLAGS
  639. # export CULIB
  640. #endif