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 25 kB

2 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836
  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" OR ${CMAKE_C_COMPILER_ID} STREQUAL "NVHPC"))
  51. set(GETARCH_FLAGS "${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 "${CCOMMON_OPT} -DUSE_LOCKING")
  135. endif ()
  136. endif ()
  137. if (C_LAPACK)
  138. if (${CMAKE_C_COMPILER_ID} STREQUAL "GNU")
  139. set(CCOMMON_OPT "${CCOMMON_OPT} -Wno-error=incompatible-pointer-types")
  140. endif ()
  141. endif ()
  142. include("${PROJECT_SOURCE_DIR}/cmake/prebuild.cmake")
  143. if (DEFINED TARGET)
  144. if (${TARGET} STREQUAL COOPERLAKE AND NOT NO_AVX512)
  145. if (${CMAKE_C_COMPILER_ID} STREQUAL "GNU")
  146. if (${CMAKE_C_COMPILER_VERSION} VERSION_GREATER 10.09)
  147. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=cooperlake")
  148. else()
  149. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=skylake-avx512")
  150. endif()
  151. elseif (${CMAKE_C_COMPILER_ID} STREQUAL "Clang" OR ${CMAKE_C_COMPILER_ID} STREQUAL "AppleClang")
  152. if (${CMAKE_C_COMPILER_VERSION} VERSION_GREATER 8.99)
  153. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=cooperlake -mllvm -exhaustive-register-search")
  154. else()
  155. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=skylake-avx512 -mllvm -exhaustive-register-search")
  156. endif()
  157. endif()
  158. endif()
  159. if (${TARGET} STREQUAL SAPPHIRERAPIDS AND NOT NO_AVX512)
  160. if (${CMAKE_C_COMPILER_ID} STREQUAL "GNU")
  161. if (${CMAKE_C_COMPILER_VERSION} VERSION_GREATER 11.0)
  162. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=sapphirerapids")
  163. else()
  164. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=skylake-avx512")
  165. endif()
  166. elseif (${CMAKE_C_COMPILER_ID} STREQUAL "Clang" OR ${CMAKE_C_COMPILER_ID} STREQUAL "AppleClang")
  167. if (${CMAKE_C_COMPILER_VERSION} VERSION_GREATER 12.0)
  168. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=sapphirerapids -mllvm -exhaustive-register-search")
  169. else()
  170. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=skylake-avx512 -mllvm -exhaustive-register-search")
  171. endif()
  172. endif()
  173. endif()
  174. if (${TARGET} STREQUAL SKYLAKEX AND NOT NO_AVX512)
  175. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=skylake-avx512")
  176. if (${CMAKE_C_COMPILER_ID} STREQUAL "Clang" OR ${CMAKE_C_COMPILER_ID} STREQUAL "AppleClang")
  177. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mllvm -exhaustive-register-search")
  178. endif()
  179. endif()
  180. if (((${TARGET} STREQUAL ZEN) AND HAVE_AVX512VL) AND NOT NO_AVX512)
  181. if (${CMAKE_C_COMPILER_ID} STREQUAL "GNU")
  182. if (${CMAKE_C_COMPILER_VERSION} VERSION_GREATER 12.99)
  183. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=znver4")
  184. else()
  185. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=skylake-avx512")
  186. endif()
  187. elseif (${CMAKE_C_COMPILER_ID} STREQUAL "Clang" OR ${CMAKE_C_COMPILER_ID} STREQUAL "AppleClang")
  188. if (${CMAKE_C_COMPILER_VERSION} VERSION_GREATER 15.99)
  189. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=znver4")
  190. else()
  191. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=skylake-avx512")
  192. endif()
  193. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mllvm -exhaustive-register-search")
  194. endif()
  195. endif()
  196. if ((${TARGET} STREQUAL HASWELL OR (${TARGET} STREQUAL ZEN AND NOT HAVE_AVX512VL)) AND NOT NO_AVX2)
  197. if (${CMAKE_C_COMPILER_ID} STREQUAL "GNU")
  198. if (CMAKE_C_COMPILER_VERSION VERSION_GREATER 4.7 OR CMAKE_C_COMPILER_VERSION VERSION_EQUAL 4.7)
  199. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mavx2")
  200. endif()
  201. elseif (${CMAKE_C_COMPILER_ID} STREQUAL "CLANG")
  202. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mavx2 -mfma")
  203. endif()
  204. endif()
  205. if (DEFINED HAVE_AVX)
  206. if (NOT NO_AVX)
  207. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mavx")
  208. endif()
  209. endif()
  210. if (DEFINED HAVE_AVX2)
  211. if (NOT NO_AVX2)
  212. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mavx2")
  213. endif()
  214. endif()
  215. # if (DEFINED HAVE_FMA3)
  216. # if (NOT NO_AVX2)
  217. # set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mfma")
  218. # endif()
  219. # endif()
  220. if (DEFINED HAVE_SSE)
  221. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -msse")
  222. endif()
  223. if (DEFINED HAVE_SSE2)
  224. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -msse2")
  225. endif()
  226. if (DEFINED HAVE_SSE3)
  227. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -msse3")
  228. endif()
  229. if (DEFINED HAVE_SSSE3)
  230. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mssse3")
  231. endif()
  232. if (DEFINED HAVE_SSE4_1)
  233. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -msse4.1")
  234. endif()
  235. if (${TARGET} STREQUAL POWER10)
  236. if (CMAKE_C_COMPILER_VERSION VERSION_GREATER 10.2 OR CMAKE_C_COMPILER_VERSION VERSION_EQUAL 10.2)
  237. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mcpu=power10 -mtune=power10 -mvsx -fno-fast-math")
  238. else ()
  239. message(FATAL_ERROR "Compiler GCC ${CMAKE_C_COMPILER_VERSION} does not support Power10.")
  240. endif()
  241. endif()
  242. if (${TARGET} STREQUAL POWER9)
  243. if (CMAKE_C_COMPILER_VERSION VERSION_GREATER 5.0 OR CMAKE_C_COMPILER_VERSION VERSION_EQUAL 5.0)
  244. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mcpu=power9 -mtune=power9 -mvsx -fno-fast-math")
  245. else ()
  246. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mcpu=power8 -mtune=power8 -mvsx -fno-fast-math")
  247. message(WARNING "Compiler GCC ${CMAKE_C_COMPILER_VERSION} does not support fully Power9.")
  248. endif()
  249. endif()
  250. if (${TARGET} STREQUAL POWER8)
  251. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mcpu=power8 -mtune=power8 -mvsx -fno-fast-math")
  252. endif()
  253. if (${TARGET} STREQUAL NEOVERSEV1)
  254. if (${CMAKE_C_COMPILER_ID} STREQUAL "PGI" AND NOT NO_SVE)
  255. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -Msve_intrinsics -march=armv8.4-a+sve -mtune=neoverse-v1")
  256. else ()
  257. if (CMAKE_C_COMPILER_VERSION VERSION_GREATER 10.4 OR CMAKE_C_COMPILER_VERSION VERSION_EQUAL 10.4)
  258. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=armv8.4-a+sve -mtune=neoverse-v1")
  259. else ()
  260. message(FATAL_ERROR "Compiler ${CMAKE_C_COMPILER} ${CMAKE_C_COMPILER_VERSION} does not support Neoverse V1.")
  261. endif()
  262. endif()
  263. endif()
  264. if (${TARGET} STREQUAL NEOVERSEN2)
  265. if (${CMAKE_C_COMPILER_ID} STREQUAL "PGI" AND NOT NO_SVE)
  266. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -Msve-intrinsics -march=armv8.5-a+sve+sve2+bf16 -mtune=neoverse-n2")
  267. else ()
  268. if (CMAKE_C_COMPILER_VERSION VERSION_GREATER 10.4 OR CMAKE_C_COMPILER_VERSION VERSION_EQUAL 10.4)
  269. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=armv8.5-a+sve+sve2+bf16 -mtune=neoverse-n2")
  270. else ()
  271. message(FATAL_ERROR "Compiler $${CMAKE_C_COMPILER} ${CMAKE_C_COMPILER_VERSION} does not support Neoverse N2.")
  272. endif()
  273. endif()
  274. endif()
  275. if (${TARGET} STREQUAL ARMV8SVE)
  276. if (${CMAKE_C_COMPILER_ID} STREQUAL "PGI" AND NOT NO_SVE)
  277. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -Msve-intrinsics -march=armv8.2-a+sve")
  278. else ()
  279. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=armv8.2-a+sve")
  280. endif()
  281. endif()
  282. if (${TARGET} STREQUAL A64FX)
  283. if (${CMAKE_C_COMPILER_ID} STREQUAL "PGI" AND NOT NO_SVE)
  284. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -Msve-intrinsics -march=armv8.2-a+sve -mtune=a64fx")
  285. else ()
  286. execute_process(COMMAND ${CMAKE_C_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
  287. if (${GCC_VERSION} VERSION_GREATER 10.4 OR ${GCC_VERSION} VERSION_EQUAL 10.4)
  288. set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=armv8.2-a+sve -mtune=a64fx")
  289. else ()
  290. message(FATAL_ERROR "Compiler $${CMAKE_C_COMPILER} {GCC_VERSION} does not support A64FX.")
  291. endif()
  292. endif()
  293. endif()
  294. endif()
  295. if (DEFINED BINARY)
  296. message(STATUS "Compiling a ${BINARY}-bit binary.")
  297. endif ()
  298. if (NOT DEFINED NEED_PIC)
  299. set(NEED_PIC 1)
  300. endif ()
  301. # OS dependent settings
  302. include("${PROJECT_SOURCE_DIR}/cmake/os.cmake")
  303. # Architecture dependent settings
  304. include("${PROJECT_SOURCE_DIR}/cmake/arch.cmake")
  305. # C Compiler dependent settings
  306. include("${PROJECT_SOURCE_DIR}/cmake/cc.cmake")
  307. if (INTERFACE64)
  308. set(SUFFIX64 64)
  309. set(SUFFIX64_UNDERSCORE _64)
  310. endif()
  311. if (NOT NOFORTRAN)
  312. # Fortran Compiler dependent settings
  313. include("${PROJECT_SOURCE_DIR}/cmake/fc.cmake")
  314. else ()
  315. if (NOT XXXX)
  316. set(C_LAPACK 1)
  317. if (INTERFACE64)
  318. set (CCOMMON_OPT "${CCOMMON_OPT} -DLAPACK_ILP64")
  319. endif ()
  320. set(TIMER "NONE")
  321. else ()
  322. set (NO_LAPACK 1)
  323. endif ()
  324. endif ()
  325. if (BINARY64)
  326. if (INTERFACE64)
  327. # CCOMMON_OPT += -DUSE64BITINT
  328. endif ()
  329. endif ()
  330. if(EMBEDDED)
  331. set(CCOMMON_OPT "${CCOMMON_OPT} -DOS_EMBEDDED")
  332. set(CCOMMON_OPT "${CCOMMON_OPT} -mthumb -mcpu=cortex-m4 -mfloat-abi=hard -mfpu=fpv4-sp-d16")
  333. endif()
  334. if (NEED_PIC)
  335. if (${CMAKE_C_COMPILER} STREQUAL "IBM")
  336. set(CCOMMON_OPT "${CCOMMON_OPT} -qpic=large")
  337. else ()
  338. set(CCOMMON_OPT "${CCOMMON_OPT} -fPIC")
  339. endif ()
  340. if (NOT NOFORTRAN)
  341. if (${F_COMPILER} STREQUAL "SUN")
  342. set(FCOMMON_OPT "${FCOMMON_OPT} -pic")
  343. else ()
  344. set(FCOMMON_OPT "${FCOMMON_OPT} -fPIC")
  345. endif ()
  346. endif()
  347. endif ()
  348. if (X86_64 OR ${CORE} STREQUAL POWER10 OR ARM64 OR LOONGARCH64)
  349. set(SMALL_MATRIX_OPT TRUE)
  350. endif ()
  351. if (ARM64)
  352. set(GEMM_GEMV_FORWARD TRUE)
  353. endif ()
  354. if (GEMM_GEMV_FORWARD AND NOT ONLY_CBLAS)
  355. set(CCOMMON_OPT "${CCOMMON_OPT} -DGEMM_GEMV_FORWARD")
  356. endif ()
  357. if (GEMM_GEMV_FORWARD_BF16 AND NOT ONLY_CBLAS)
  358. set(CCOMMON_OPT "${CCOMMON_OPT} -DGEMM_GEMV_FORWARD_BF16")
  359. endif ()
  360. if (SMALL_MATRIX_OPT)
  361. set(CCOMMON_OPT "${CCOMMON_OPT} -DSMALL_MATRIX_OPT")
  362. endif ()
  363. if (DYNAMIC_ARCH)
  364. if (X86 OR X86_64 OR ARM64 OR POWER OR RISCV64 OR LOONGARCH64)
  365. set(CCOMMON_OPT "${CCOMMON_OPT} -DDYNAMIC_ARCH")
  366. if (DYNAMIC_OLDER)
  367. set(CCOMMON_OPT "${CCOMMON_OPT} -DDYNAMIC_OLDER")
  368. endif ()
  369. else ()
  370. unset (DYNAMIC_ARCH)
  371. message (STATUS "DYNAMIC_ARCH is not supported on the target architecture, removing")
  372. endif ()
  373. endif ()
  374. if (DYNAMIC_LIST)
  375. set(CCOMMON_OPT "${CCOMMON_OPT} -DDYNAMIC_LIST")
  376. foreach(DCORE ${DYNAMIC_LIST})
  377. set(CCOMMON_OPT "${CCOMMON_OPT} -DDYN_${DCORE}")
  378. endforeach ()
  379. endif ()
  380. if (NO_LAPACK)
  381. set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_LAPACK")
  382. #Disable LAPACK C interface
  383. set(NO_LAPACKE 1)
  384. endif ()
  385. if (NO_LAPACKE)
  386. set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_LAPACKE")
  387. endif ()
  388. if (NO_AVX)
  389. set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX")
  390. endif ()
  391. if (X86)
  392. set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX")
  393. endif ()
  394. if (NO_AVX2)
  395. set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX2")
  396. endif ()
  397. if (NO_AVX512)
  398. set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX512")
  399. endif ()
  400. if (USE_THREAD)
  401. # USE_SIMPLE_THREADED_LEVEL3 = 1
  402. # NO_AFFINITY = 1
  403. set(CCOMMON_OPT "${CCOMMON_OPT} -DSMP_SERVER")
  404. if (MIPS64)
  405. if (NOT ${CORE} STREQUAL "LOONGSON3B")
  406. set(USE_SIMPLE_THREADED_LEVEL3 1)
  407. endif ()
  408. endif ()
  409. if (BIGNUMA)
  410. set(CCOMMON_OPT "${CCOMMON_OPT} -DBIGNUMA")
  411. endif ()
  412. endif ()
  413. if (NO_WARMUP)
  414. set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_WARMUP")
  415. endif ()
  416. if (CONSISTENT_FPCSR)
  417. set(CCOMMON_OPT "${CCOMMON_OPT} -DCONSISTENT_FPCSR")
  418. endif ()
  419. if (USE_TLS)
  420. set(CCOMMON_OPT "${CCOMMON_OPT} -DUSE_TLS")
  421. endif ()
  422. # Only for development
  423. # set(CCOMMON_OPT "${CCOMMON_OPT} -DPARAMTEST")
  424. # set(CCOMMON_OPT "${CCOMMON_OPT} -DPREFETCHTEST")
  425. # set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_SWITCHING")
  426. # set(USE_PAPI 1)
  427. if (USE_PAPI)
  428. set(CCOMMON_OPT "${CCOMMON_OPT} -DUSE_PAPI")
  429. set(EXTRALIB "${EXTRALIB} -lpapi -lperfctr")
  430. endif ()
  431. if (DYNAMIC_THREADS)
  432. set(CCOMMON_OPT "${CCOMMON_OPT} -DDYNAMIC_THREADS")
  433. endif ()
  434. set(CCOMMON_OPT "${CCOMMON_OPT} -DMAX_CPU_NUMBER=${NUM_THREADS}")
  435. set(CCOMMON_OPT "${CCOMMON_OPT} -DMAX_PARALLEL_NUMBER=${NUM_PARALLEL}")
  436. if (BUFFERSIZE)
  437. set(CCOMMON_OPT "${CCOMMON_OPT} -DBUFFERSIZE=${BUFFERSIZE}")
  438. endif ()
  439. if (USE_SIMPLE_THREADED_LEVEL3)
  440. set(CCOMMON_OPT "${CCOMMON_OPT} -DUSE_SIMPLE_THREADED_LEVEL3")
  441. endif ()
  442. if (NOT ${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
  443. if (DEFINED MAX_STACK_ALLOC)
  444. if (NOT ${MAX_STACK_ALLOC} EQUAL 0)
  445. set(CCOMMON_OPT "${CCOMMON_OPT} -DMAX_STACK_ALLOC=${MAX_STACK_ALLOC}")
  446. endif ()
  447. else ()
  448. set(CCOMMON_OPT "${CCOMMON_OPT} -DMAX_STACK_ALLOC=2048")
  449. endif ()
  450. endif ()
  451. if (NOT ${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
  452. if (DEFINED BLAS3_MEM_ALLOC_THRESHOLD)
  453. if (NOT ${BLAS3_MEM_ALLOC_THRESHOLD} EQUAL 32)
  454. set(CCOMMON_OPT "${CCOMMON_OPT} -DBLAS3_MEM_ALLOC_THRESHOLD=${BLAS3_MEM_ALLOC_THRESHOLD}")
  455. endif()
  456. endif()
  457. endif()
  458. set(LIBPREFIX "lib${LIBNAMEPREFIX}openblas")
  459. if (DEFINED LIBNAMESUFFIX)
  460. set(LIBPREFIX "${LIBNAMEPREFIX}_${LIBNAMESUFFIX}")
  461. endif ()
  462. if (NOT DEFINED SYMBOLPREFIX)
  463. set(SYMBOLPREFIX "")
  464. endif ()
  465. if (NOT DEFINED SYMBOLSUFFIX)
  466. set(SYMBOLSUFFIX "")
  467. endif ()
  468. set(KERNELDIR "${PROJECT_SOURCE_DIR}/kernel/${ARCH}")
  469. # TODO: need to convert these Makefiles
  470. # include ${PROJECT_SOURCE_DIR}/cmake/${ARCH}.cmake
  471. if (${CORE} STREQUAL "PPC440")
  472. set(CCOMMON_OPT "${CCOMMON_OPT} -DALLOC_QALLOC")
  473. endif ()
  474. if (${CORE} STREQUAL "PPC440FP2")
  475. set(STATIC_ALLOCATION 1)
  476. endif ()
  477. if (NOT ${CMAKE_SYSTEM_NAME} STREQUAL "Linux")
  478. set(NO_AFFINITY 1)
  479. endif ()
  480. if (NOT X86_64 AND NOT X86 AND NOT ${CORE} STREQUAL "LOONGSON3B")
  481. set(NO_AFFINITY 1)
  482. endif ()
  483. if (NO_AFFINITY)
  484. set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AFFINITY")
  485. endif ()
  486. if (FUNCTION_PROFILE)
  487. set(CCOMMON_OPT "${CCOMMON_OPT} -DFUNCTION_PROFILE")
  488. endif ()
  489. if (HUGETLB_ALLOCATION)
  490. set(CCOMMON_OPT "${CCOMMON_OPT} -DALLOC_HUGETLB")
  491. endif ()
  492. if (DEFINED HUGETLBFILE_ALLOCATION)
  493. set(CCOMMON_OPT "${CCOMMON_OPT} -DALLOC_HUGETLBFILE -DHUGETLB_FILE_NAME=${HUGETLBFILE_ALLOCATION})")
  494. endif ()
  495. if (STATIC_ALLOCATION)
  496. set(CCOMMON_OPT "${CCOMMON_OPT} -DALLOC_STATIC")
  497. endif ()
  498. if (DEVICEDRIVER_ALLOCATION)
  499. set(CCOMMON_OPT "${CCOMMON_OPT} -DALLOC_DEVICEDRIVER -DDEVICEDRIVER_NAME=\"/dev/mapper\"")
  500. endif ()
  501. if (MIXED_MEMORY_ALLOCATION)
  502. set(CCOMMON_OPT "${CCOMMON_OPT} -DMIXED_MEMORY_ALLOCATION")
  503. endif ()
  504. set(CCOMMON_OPT "${CCOMMON_OPT} -DVERSION=\"\\\"${OpenBLAS_VERSION}\\\"\"")
  505. set(REVISION "-r${OpenBLAS_VERSION}")
  506. set(MAJOR_VERSION ${OpenBLAS_MAJOR_VERSION})
  507. set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${CCOMMON_OPT}")
  508. if (NOT BUILD_SINGLE AND NOT BUILD_DOUBLE AND NOT BUILD_COMPLEX AND NOT BUILD_COMPLEX16)
  509. set (BUILD_SINGLE ON)
  510. set (BUILD_DOUBLE ON)
  511. set (BUILD_COMPLEX ON)
  512. set (BUILD_COMPLEX16 ON)
  513. endif()
  514. if (BUILD_SINGLE)
  515. set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_SINGLE")
  516. endif()
  517. if (BUILD_DOUBLE)
  518. set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_DOUBLE")
  519. endif()
  520. if (BUILD_COMPLEX)
  521. set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_COMPLEX")
  522. endif()
  523. if (BUILD_COMPLEX16)
  524. set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_COMPLEX16")
  525. endif()
  526. if (BUILD_BFLOAT16)
  527. set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_BFLOAT16")
  528. endif()
  529. if(NOT MSVC)
  530. set(CMAKE_ASM_FLAGS "${CMAKE_ASM_FLAGS} ${CCOMMON_OPT}")
  531. endif()
  532. # TODO: not sure what PFLAGS is -hpa
  533. set(PFLAGS "${PFLAGS} ${CCOMMON_OPT} -I${TOPDIR} -DPROFILE ${COMMON_PROF}")
  534. set(CMAKE_Fortran_FLAGS "${CMAKE_Fortran_FLAGS} ${FCOMMON_OPT}")
  535. # TODO: not sure what FPFLAGS is -hpa
  536. set(FPFLAGS "${FPFLAGS} ${FCOMMON_OPT} ${COMMON_PROF}")
  537. #For LAPACK Fortran codes.
  538. set(LAPACK_FFLAGS "${LAPACK_FFLAGS} ${CMAKE_Fortran_FLAGS}" )
  539. if (LAPACK_STRLEN)
  540. set (LAPACK_FFLAGS "${LAPACK_FFLAGS} -DLAPACK_STRLEN=${LAPACK_STRLEN}")
  541. endif()
  542. set(LAPACK_FPFLAGS "${LAPACK_FPFLAGS} ${FPFLAGS}")
  543. #Disable -fopenmp for LAPACK Fortran codes on Windows.
  544. if (${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
  545. set(FILTER_FLAGS "-fopenmp;-mp;-openmp;-xopenmp=parallel")
  546. foreach (FILTER_FLAG ${FILTER_FLAGS})
  547. string(REPLACE ${FILTER_FLAG} "" LAPACK_FFLAGS ${LAPACK_FFLAGS})
  548. string(REPLACE ${FILTER_FLAG} "" LAPACK_FPFLAGS ${LAPACK_FPFLAGS})
  549. endforeach ()
  550. endif ()
  551. if (CMAKE_Fortran_COMPILER)
  552. if ("${F_COMPILER}" STREQUAL "NAG" OR "${F_COMPILER}" STREQUAL "CRAY" OR CMAKE_Fortran_COMPILER_ID MATCHES "LLVMFlang.*")
  553. set(FILTER_FLAGS "-msse3;-mssse3;-msse4.1;-mavx;-mavx2,-mskylake-avx512")
  554. if (CMAKE_Fortran_COMPILER_ID MATCHES "LLVMFlang.*")
  555. message(STATUS "removing fortran flags")
  556. set(FILTER_FLAGS "${FILTER_FLAGS};-m32;-m64")
  557. endif ()
  558. foreach (FILTER_FLAG ${FILTER_FLAGS})
  559. string(REPLACE ${FILTER_FLAG} "" LAPACK_FFLAGS ${LAPACK_FFLAGS})
  560. string(REPLACE ${FILTER_FLAG} "" LAPACK_FPFLAGS ${LAPACK_FPFLAGS})
  561. endforeach ()
  562. endif ()
  563. endif ()
  564. if ("${F_COMPILER}" STREQUAL "GFORTRAN")
  565. # lapack-netlib is rife with uninitialized warnings -hpa
  566. set(LAPACK_FFLAGS "${LAPACK_FFLAGS} -Wno-maybe-uninitialized")
  567. endif ()
  568. set(LAPACK_CFLAGS "${CMAKE_C_CFLAGS} -DHAVE_LAPACK_CONFIG_H")
  569. if (INTERFACE64)
  570. set(LAPACK_CFLAGS "${LAPACK_CFLAGS} -DLAPACK_ILP64")
  571. endif ()
  572. if (${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
  573. set(LAPACK_CFLAGS "${LAPACK_CFLAGS} -DOPENBLAS_OS_WINDOWS")
  574. endif ()
  575. if (${CMAKE_C_COMPILER} STREQUAL "LSB" OR ${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
  576. set(LAPACK_CFLAGS "${LAPACK_CFLAGS} -DLAPACK_COMPLEX_STRUCTURE")
  577. endif ()
  578. if ("${CMAKE_BUILD_TYPE}" STREQUAL "Release")
  579. if ("${F_COMPILER}" STREQUAL "FLANG")
  580. if (${CMAKE_Fortran_COMPILER_VERSION} VERSION_LESS_EQUAL 3)
  581. set(CMAKE_Fortran_FLAGS_RELEASE "${CMAKE_Fortran_FLAGS_RELEASE} -fno-unroll-loops")
  582. endif ()
  583. endif ()
  584. endif ()
  585. if (NOT DEFINED SUFFIX)
  586. set(SUFFIX o)
  587. endif ()
  588. if (NOT DEFINED PSUFFIX)
  589. set(PSUFFIX po)
  590. endif ()
  591. if (NOT DEFINED LIBSUFFIX)
  592. set(LIBSUFFIX a)
  593. endif ()
  594. if (DYNAMIC_ARCH)
  595. if (USE_THREAD)
  596. set(LIBNAME "${LIBPREFIX}p${REVISION}.${LIBSUFFIX}")
  597. set(LIBNAME_P "${LIBPREFIX}p${REVISION}_p.${LIBSUFFIX}")
  598. else ()
  599. set(LIBNAME "${LIBPREFIX}${REVISION}.${LIBSUFFIX}")
  600. set(LIBNAME_P "${LIBPREFIX}${REVISION}_p.${LIBSUFFIX}")
  601. endif ()
  602. else ()
  603. if (USE_THREAD)
  604. set(LIBNAME "${LIBPREFIX}_${LIBCORE}p${REVISION}.${LIBSUFFIX}")
  605. set(LIBNAME_P "${LIBPREFIX}_${LIBCORE}p${REVISION}_p.${LIBSUFFIX}")
  606. else ()
  607. set(LIBNAME "${LIBPREFIX}_${LIBCORE}${REVISION}.${LIBSUFFIX}")
  608. set(LIBNAME_P "${LIBPREFIX}_${LIBCORE}${REVISION}_p.${LIBSUFFIX}")
  609. endif ()
  610. endif ()
  611. if (DEFINED FIXED_LIBNAME)
  612. set (LIBNAME "${LIBPREFIX}.${LIBSUFFIX}")
  613. set (LIBNAME "${LIBPREFIX}_p.${LIBSUFFIX}")
  614. endif()
  615. set(LIBDLLNAME "${LIBPREFIX}.dll")
  616. set(LIBSONAME "${LIBNAME}.${LIBSUFFIX}.so")
  617. set(LIBDYNNAME "${LIBNAME}.${LIBSUFFIX}.dylib")
  618. set(LIBDEFNAME "${LIBNAME}.${LIBSUFFIX}.def")
  619. set(LIBEXPNAME "${LIBNAME}.${LIBSUFFIX}.exp")
  620. set(LIBZIPNAME "${LIBNAME}.${LIBSUFFIX}.zip")
  621. set(LIBS "${PROJECT_SOURCE_DIR}/${LIBNAME}")
  622. set(LIBS_P "${PROJECT_SOURCE_DIR}/${LIBNAME_P}")
  623. set(LIB_COMPONENTS BLAS)
  624. if (NOT NO_CBLAS)
  625. set(LIB_COMPONENTS "${LIB_COMPONENTS} CBLAS")
  626. endif ()
  627. if (NOT NO_LAPACK)
  628. set(LIB_COMPONENTS "${LIB_COMPONENTS} LAPACK")
  629. if (NOT NO_LAPACKE)
  630. set(LIB_COMPONENTS "${LIB_COMPONENTS} LAPACKE")
  631. endif ()
  632. if (BUILD_RELAPACK)
  633. set(LIB_COMPONENTS "${LIB_COMPONENTS} ReLAPACK")
  634. endif ()
  635. endif ()
  636. if (ONLY_CBLAS)
  637. set(LIB_COMPONENTS CBLAS)
  638. endif ()
  639. # For GEMM3M
  640. set(USE_GEMM3M 0)
  641. if (DEFINED ARCH)
  642. if (X86 OR X86_64 OR ${ARCH} STREQUAL "ia64" OR MIPS64)
  643. set(USE_GEMM3M 1)
  644. endif ()
  645. if (${CORE} STREQUAL "generic")
  646. set(USE_GEMM3M 0)
  647. endif ()
  648. endif ()
  649. #export OSNAME
  650. #export ARCH
  651. #export CORE
  652. #export LIBCORE
  653. #export PGCPATH
  654. #export CONFIG
  655. #export CC
  656. #export FC
  657. #export BU
  658. #export FU
  659. #export NEED2UNDERSCORES
  660. #export USE_THREAD
  661. #export NUM_THREADS
  662. #export NUM_CORES
  663. #export SMP
  664. #export MAKEFILE_RULE
  665. #export NEED_PIC
  666. #export BINARY
  667. #export BINARY32
  668. #export BINARY64
  669. #export F_COMPILER
  670. #export C_COMPILER
  671. #export USE_OPENMP
  672. #export CROSS
  673. #export CROSS_SUFFIX
  674. #export NOFORTRAN
  675. #export NO_FBLAS
  676. #export EXTRALIB
  677. #export CEXTRALIB
  678. #export FEXTRALIB
  679. #export HAVE_SSE
  680. #export HAVE_SSE2
  681. #export HAVE_SSE3
  682. #export HAVE_SSSE3
  683. #export HAVE_SSE4_1
  684. #export HAVE_SSE4_2
  685. #export HAVE_SSE4A
  686. #export HAVE_SSE5
  687. #export HAVE_AVX
  688. #export HAVE_VFP
  689. #export HAVE_VFPV3
  690. #export HAVE_VFPV4
  691. #export HAVE_NEON
  692. #export KERNELDIR
  693. #export FUNCTION_PROFILE
  694. #export TARGET_CORE
  695. #
  696. #export SBGEMM_UNROLL_M
  697. #export SBGEMM_UNROLL_N
  698. #export SGEMM_UNROLL_M
  699. #export SGEMM_UNROLL_N
  700. #export DGEMM_UNROLL_M
  701. #export DGEMM_UNROLL_N
  702. #export QGEMM_UNROLL_M
  703. #export QGEMM_UNROLL_N
  704. #export CGEMM_UNROLL_M
  705. #export CGEMM_UNROLL_N
  706. #export ZGEMM_UNROLL_M
  707. #export ZGEMM_UNROLL_N
  708. #export XGEMM_UNROLL_M
  709. #export XGEMM_UNROLL_N
  710. #export CGEMM3M_UNROLL_M
  711. #export CGEMM3M_UNROLL_N
  712. #export ZGEMM3M_UNROLL_M
  713. #export ZGEMM3M_UNROLL_N
  714. #export XGEMM3M_UNROLL_M
  715. #export XGEMM3M_UNROLL_N
  716. #if (USE_CUDA)
  717. # export CUDADIR
  718. # export CUCC
  719. # export CUFLAGS
  720. # export CULIB
  721. #endif