Austin Schuh | 9a24b37 | 2018-01-28 16:12:29 -0800 | [diff] [blame] | 1 | ################################################################################################### |
| 2 | # # |
| 3 | # This file is part of HPIPM. # |
| 4 | # # |
| 5 | # HPIPM -- High Performance Interior Point Method. # |
| 6 | # Copyright (C) 2017 by Gianluca Frison. # |
| 7 | # Developed at IMTEK (University of Freiburg) under the supervision of Moritz Diehl. # |
| 8 | # All rights reserved. # |
| 9 | # # |
| 10 | # HPMPC is free software; you can redistribute it and/or # |
| 11 | # modify it under the terms of the GNU Lesser General Public # |
| 12 | # License as published by the Free Software Foundation; either # |
| 13 | # version 2.1 of the License, or (at your option) any later version. # |
| 14 | # # |
| 15 | # HPMPC is distributed in the hope that it will be useful, # |
| 16 | # but WITHOUT ANY WARRANTY; without even the implied warranty of # |
| 17 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. # |
| 18 | # See the GNU Lesser General Public License for more details. # |
| 19 | # # |
| 20 | # You should have received a copy of the GNU Lesser General Public # |
| 21 | # License along with HPMPC; if not, write to the Free Software # |
| 22 | # Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA # |
| 23 | # # |
| 24 | # Author: Gianluca Frison, gianluca.frison (at) imtek.uni-freiburg.de # |
| 25 | # # |
| 26 | ################################################################################################### |
| 27 | |
| 28 | cmake_minimum_required(VERSION 2.8.11) |
| 29 | |
| 30 | project(blasfeo) |
| 31 | |
| 32 | enable_language(C ASM) |
| 33 | |
| 34 | # Target architecture |
| 35 | #set(TARGET X64_INTEL_HASWELL) |
| 36 | set(TARGET X64_INTEL_SANDY_BRIDGE CACHE STRING "Target architecture") |
| 37 | #set(TARGET X64_INTEL_CORE) |
| 38 | #set(TARGET X64_AMD_BULLDOZER) |
| 39 | #set(TARGET ARMV8A_ARM_CORTEX_A57) |
| 40 | #set(TARGET ARMV7A_ARM_CORTEX_A15) |
| 41 | #set(TARGET GENERIC) |
| 42 | |
| 43 | # Linear Algebra library |
| 44 | set(LA HIGH_PERFORMANCE CACHE STRING "Linear algebra optimization level") |
| 45 | #set(LA REFERENCE) |
| 46 | #set(LA BLAS) |
| 47 | |
| 48 | # BLAS and LAPACK version (for LA=BLAS in BLASFEO) |
| 49 | set(REF_BLAS 0 CACHE STRING "Reference blas to use") |
| 50 | #set(REF_BLAS OPENBLAS) |
| 51 | #set(REF_BLAS NETLIB) |
| 52 | #set(REF_BLAS MKL) |
| 53 | #set(REF_BLAS BLIS) |
| 54 | #set(REF_BLAS ATLAS) |
| 55 | |
| 56 | # Compile auxiliary functions with external dependencies (for memory allocation and printing) |
| 57 | set(EXT_DEP ON CACHE BOOL "Compile external dependencies in BLASFEO") |
| 58 | |
| 59 | configure_file(${PROJECT_SOURCE_DIR}/blasfeo_target.h.in |
| 60 | ${CMAKE_CURRENT_SOURCE_DIR}/include/blasfeo_target.h @ONLY) |
| 61 | |
| 62 | # C Compiler |
| 63 | # set(CC_COMPILER gcc CACHE STRING "compiler") |
| 64 | #set(CC_COMPILER clang) |
| 65 | #set(CC_COMPILER x86_64-w64-mingw32-gcc) |
| 66 | |
| 67 | # build shared library |
| 68 | #set(BUILD_SHARED_LIBS ON CACHE STRING "Build shared libraries") |
| 69 | |
| 70 | # installation directory |
| 71 | if(CMAKE_INSTALL_PREFIX MATCHES "/usr/local") |
| 72 | set(CMAKE_INSTALL_PREFIX "/opt/blasfeo") |
| 73 | endif() |
| 74 | |
| 75 | # headers installation directory |
| 76 | set(BLASFEO_HEADERS_INSTALLATION_DIRECTORY "include" CACHE STRING "Headers local installation directory") |
| 77 | |
| 78 | # Macro level (code size vs performance in assembly kernels): 0 (no macro), 1 (all macro but gemm kernel), 2 (all macro) |
| 79 | set(MACRO_LEVEL 0) |
| 80 | |
| 81 | # enable runtine checks |
| 82 | set(RUNTIME_CHECKS 0) |
| 83 | #set(RUNTIME_CHECKS 0) |
| 84 | |
| 85 | # compiler flags |
| 86 | if(CMAKE_C_COMPILER_ID MATCHES "GNU" OR CMAKE_C_COMPILER_ID MATCHES "Clang") |
| 87 | set(CMAKE_C_FLAGS "") |
| 88 | set(CMAKE_ASM_FLAGS "") |
| 89 | set(CMAKE_C_FLAGS_RELEASE "") |
| 90 | set(CMAKE_ASM_FLAGS_RELEASE "") |
| 91 | # optimization flags |
| 92 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -O2") |
| 93 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fPIC") |
| 94 | # debugging flags |
| 95 | #set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -g") |
| 96 | #set(CMAKE_ASM_FLAGS "${CMAKE_C_FLAGS} -g") |
| 97 | endif() |
| 98 | |
| 99 | # search directories |
| 100 | #set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -I${BLASFEO_PATH}/include") XXX |
| 101 | |
| 102 | # |
| 103 | if(${LA} MATCHES HIGH_PERFORMANCE) |
| 104 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DLA_HIGH_PERFORMANCE") |
| 105 | endif() |
| 106 | if(${LA} MATCHES REFERENCE) |
| 107 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DLA_REFERENCE") |
| 108 | endif() |
| 109 | if(${LA} MATCHES BLAS) |
| 110 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DLA_BLAS") |
| 111 | endif() |
| 112 | |
| 113 | # |
| 114 | if(${RUNTIME_CHECKS} MATCHES 1) |
| 115 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DDIM_CHECK") |
| 116 | endif() |
| 117 | |
| 118 | # |
| 119 | if(${EXT_DEP}) |
| 120 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DEXT_DEP") |
| 121 | endif() |
| 122 | |
| 123 | # |
| 124 | if(${MACRO_LEVEL} MATCHES 1) |
| 125 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DMACRO_LEVEL=1") |
| 126 | endif() |
| 127 | if(${MACRO_LEVEL} MATCHES 2) |
| 128 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DMACRO_LEVEL=2") |
| 129 | endif() |
| 130 | |
| 131 | # |
| 132 | if(${CMAKE_SYSTEM_NAME} MATCHES "Linux") |
| 133 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DOS_LINUX") |
| 134 | set(CMAKE_ASM_FLAGS "${CMAKE_C_FLAGS} -DOS_LINUX") |
| 135 | endif() |
| 136 | if(${CMAKE_SYSTEM_NAME} MATCHES "Darwin") |
| 137 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DOS_MAC") |
| 138 | set(CMAKE_ASM_FLAGS "${CMAKE_C_FLAGS} -DOS_MAC") |
| 139 | endif() |
| 140 | if(${CMAKE_SYSTEM_NAME} MATCHES "Windows") |
| 141 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DOS_WINDOWS") |
| 142 | set(CMAKE_ASM_FLAGS "${CMAKE_C_FLAGS} -DOS_WINDOWS") |
| 143 | endif() |
| 144 | |
| 145 | # |
| 146 | if(${REF_BLAS} MATCHES 0) |
| 147 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ") |
| 148 | endif(${REF_BLAS} MATCHES 0) |
| 149 | if(${REF_BLAS} MATCHES OPENBLAS) |
| 150 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DREF_BLAS_OPENBLAS -I/opt/openblas/include") |
| 151 | endif(${REF_BLAS} MATCHES OPENBLAS) |
| 152 | if(${REF_BLAS} MATCHES BLIS) |
| 153 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DREF_BLAS_BLIS -std=c99") |
| 154 | endif(${REF_BLAS} MATCHES BLIS) |
| 155 | if(${REF_BLAS} MATCHES NETLIB) |
| 156 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DREF_BLAS_NETLIB") |
| 157 | endif(${REF_BLAS} MATCHES NETLIB) |
| 158 | if(${REF_BLAS} MATCHES MKL) |
| 159 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DREF_BLAS_MKL -m64 -I/opt/intel/mkl/include") |
| 160 | endif(${REF_BLAS} MATCHES MKL) |
| 161 | if(${REF_BLAS} MATCHES ATLAS) |
| 162 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DREF_BLAS_ATLAS") |
| 163 | endif(${REF_BLAS} MATCHES ATLAS) |
| 164 | |
| 165 | # architecture-specific flags |
| 166 | if(${TARGET} MATCHES X64_INTEL_HASWELL) |
| 167 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DTARGET_X64_INTEL_HASWELL") |
| 168 | if(CMAKE_C_COMPILER_ID MATCHES "GNU" OR CMAKE_C_COMPILER_ID MATCHES "Clang") |
| 169 | set() |
| 170 | endif() |
| 171 | endif() |
| 172 | |
| 173 | if(${TARGET} MATCHES X64_INTEL_SANDY_BRIDGE) |
| 174 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DTARGET_X64_INTEL_SANDY_BRIDGE") |
| 175 | if(CMAKE_C_COMPILER_ID MATCHES "GNU" OR CMAKE_C_COMPILER_ID MATCHES "Clang") |
| 176 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -m64 -mavx") |
| 177 | endif() |
| 178 | endif() |
| 179 | |
| 180 | if(${TARGET} MATCHES X64_INTEL_CORE) |
| 181 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DTARGET_X64_INTEL_CORE") |
| 182 | if(CMAKE_C_COMPILER_ID MATCHES "GNU" OR CMAKE_C_COMPILER_ID MATCHES "Clang") |
| 183 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -m64 -msse3") |
| 184 | endif() |
| 185 | endif() |
| 186 | |
| 187 | if(${TARGET} MATCHES X64_AMD_BULLDOZER) |
| 188 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DTARGET_X64_AMD_BULLDOZER") |
| 189 | if(CMAKE_C_COMPILER_ID MATCHES "GNU" OR CMAKE_C_COMPILER_ID MATCHES "Clang") |
| 190 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -m64 -mavx -mfma") |
| 191 | endif() |
| 192 | endif() |
| 193 | |
| 194 | if(${TARGET} MATCHES ARMV8A_ARM_CORTEX_A57) |
| 195 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DTARGET_ARMV8A_ARM_CORTEX_A57") |
| 196 | set(CMAKE_ASM_FLAGS "${CMAKE_C_FLAGS} -DTARGET_ARMV8A_ARM_CORTEX_A57") |
| 197 | if(CMAKE_C_COMPILER_ID MATCHES "GNU" OR CMAKE_C_COMPILER_ID MATCHES "Clang") |
| 198 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -march=armv8-a+crc+crypto+fp+simd") |
| 199 | endif() |
| 200 | endif() |
| 201 | |
| 202 | if(${TARGET} MATCHES ARMV7A_ARM_CORTEX_A15) |
| 203 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DTARGET_ARMV7A_ARM_CORTEX_A15") |
| 204 | set(CMAKE_ASM_FLAGS "${CMAKE_C_FLAGS} -DTARGET_ARMV7A_ARM_CORTEX_A15") |
| 205 | if(CMAKE_C_COMPILER_ID MATCHES "GNU" OR CMAKE_C_COMPILER_ID MATCHES "Clang") |
| 206 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -marm -mfloat-abi=hard -mfpu=neon-vfpv4 -mcpu=cortex-a15") |
| 207 | set(CMAKE_ASM_FLAGS "${CMAKE_C_FLAGS} -mfpu=neon-vfpv4") |
| 208 | endif() |
| 209 | endif() |
| 210 | |
| 211 | if(${TARGET} MATCHES GENERIC) |
| 212 | set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DTARGET_GENERIC") |
| 213 | endif() |
| 214 | |
| 215 | |
| 216 | |
| 217 | # source files |
| 218 | |
| 219 | if(${LA} MATCHES HIGH_PERFORMANCE) |
| 220 | |
| 221 | if(${TARGET} MATCHES X64_INTEL_HASWELL) |
| 222 | |
| 223 | file(GLOB AUX_SRC |
| 224 | ${PROJECT_SOURCE_DIR}/auxiliary/d_aux_lib4.c |
| 225 | ${PROJECT_SOURCE_DIR}/auxiliary/avx/kernel_dgecp_lib4.c |
| 226 | ${PROJECT_SOURCE_DIR}/auxiliary/avx2/kernel_dgetr_lib4.c |
| 227 | ${PROJECT_SOURCE_DIR}/auxiliary/s_aux_lib8.c |
| 228 | ${PROJECT_SOURCE_DIR}/auxiliary/m_aux_lib48.c |
| 229 | ) |
| 230 | |
| 231 | file(GLOB KERNEL_SRC |
| 232 | ${PROJECT_SOURCE_DIR}/kernel/avx2/kernel_dgemm_12x4_lib4.S |
| 233 | ${PROJECT_SOURCE_DIR}/kernel/avx2/kernel_dgemm_8x8_lib4.S |
| 234 | ${PROJECT_SOURCE_DIR}/kernel/avx2/kernel_dgemm_8x4_lib4.S |
| 235 | ${PROJECT_SOURCE_DIR}/kernel/avx2/kernel_dgemm_4x4_lib4.S |
| 236 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_dgemm_diag_lib4.c |
| 237 | ${PROJECT_SOURCE_DIR}/kernel/avx2/kernel_dgemv_8_lib4.S |
| 238 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_dgemv_4_lib4.S |
| 239 | ${PROJECT_SOURCE_DIR}/kernel/avx2/kernel_dsymv_6_lib4.S |
| 240 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dsymv_4_lib4.c |
| 241 | ${PROJECT_SOURCE_DIR}/kernel/avx2/kernel_dgetrf_pivot_4_lib4.c |
| 242 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_dgeqrf_4_lib4.c |
| 243 | ${PROJECT_SOURCE_DIR}/kernel/avx2/kernel_dgebp_lib4.S |
| 244 | ${PROJECT_SOURCE_DIR}/kernel/avx2/kernel_dgelqf_4_lib4.S |
| 245 | ${PROJECT_SOURCE_DIR}/kernel/avx2/kernel_sgemm_24x4_lib8.S |
| 246 | ${PROJECT_SOURCE_DIR}/kernel/avx2/kernel_sgemm_16x4_lib8.S |
| 247 | ${PROJECT_SOURCE_DIR}/kernel/avx2/kernel_sgemm_8x8_lib8.S |
| 248 | ${PROJECT_SOURCE_DIR}/kernel/avx2/kernel_sgemm_8x4_lib8.S |
| 249 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_sgemm_diag_lib8.c |
| 250 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_sgead_lib8.S |
| 251 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_sgecp_lib8.S |
| 252 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_sgetr_lib8.S |
| 253 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_sgesc_lib8.S |
| 254 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_sgemv_8_lib8.S |
| 255 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_sgemv_4_lib8.S |
| 256 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_ssymv_4_lib8.c |
| 257 | ) |
| 258 | |
| 259 | file(GLOB BLAS_SRC |
| 260 | ${PROJECT_SOURCE_DIR}/blas/d_blas1_lib4.c |
| 261 | ${PROJECT_SOURCE_DIR}/blas/d_blas2_lib4.c |
| 262 | ${PROJECT_SOURCE_DIR}/blas/d_blas2_diag_lib.c |
| 263 | ${PROJECT_SOURCE_DIR}/blas/d_blas3_lib4.c |
| 264 | ${PROJECT_SOURCE_DIR}/blas/d_blas3_diag_lib4.c |
| 265 | ${PROJECT_SOURCE_DIR}/blas/d_lapack_lib4.c |
| 266 | ${PROJECT_SOURCE_DIR}/blas/s_blas1_lib8.c |
| 267 | ${PROJECT_SOURCE_DIR}/blas/s_blas2_lib8.c |
| 268 | ${PROJECT_SOURCE_DIR}/blas/s_blas2_diag_lib.c |
| 269 | ${PROJECT_SOURCE_DIR}/blas/s_blas3_lib8.c |
| 270 | ${PROJECT_SOURCE_DIR}/blas/s_blas3_diag_lib8.c |
| 271 | ${PROJECT_SOURCE_DIR}/blas/s_lapack_lib8.c |
| 272 | ) |
| 273 | |
| 274 | endif(${TARGET} MATCHES X64_INTEL_HASWELL) |
| 275 | |
| 276 | if(${TARGET} MATCHES X64_INTEL_SANDY_BRIDGE) |
| 277 | |
| 278 | file(GLOB AUX_SRC |
| 279 | ${PROJECT_SOURCE_DIR}/auxiliary/d_aux_lib4.c |
| 280 | ${PROJECT_SOURCE_DIR}/auxiliary/avx/kernel_dgecp_lib4.c |
| 281 | ${PROJECT_SOURCE_DIR}/auxiliary/avx/kernel_dgetr_lib4.c |
| 282 | ${PROJECT_SOURCE_DIR}/auxiliary/s_aux_lib8.c |
| 283 | ${PROJECT_SOURCE_DIR}/auxiliary/m_aux_lib48.c |
| 284 | ) |
| 285 | |
| 286 | file(GLOB KERNEL_SRC |
| 287 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_dgemm_8x4_lib4.S |
| 288 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_dgemm_4x4_lib4.S |
| 289 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_dgemm_diag_lib4.c |
| 290 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_dgemv_12_lib4.S |
| 291 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_dgemv_8_lib4.S |
| 292 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_dgemv_4_lib4.S |
| 293 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_dsymv_6_lib4.S |
| 294 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dsymv_4_lib4.c |
| 295 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_dgetrf_pivot_4_lib4.c |
| 296 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_dgeqrf_4_lib4.c |
| 297 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_dgebp_lib4.S |
| 298 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_dgelqf_4_lib4.S |
| 299 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_sgemm_16x4_lib8.S |
| 300 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_sgemm_8x8_lib8.S |
| 301 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_sgemm_8x4_lib8.S |
| 302 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_sgemm_diag_lib8.c |
| 303 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_sgead_lib8.S |
| 304 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_sgecp_lib8.S |
| 305 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_sgetr_lib8.S |
| 306 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_sgesc_lib8.S |
| 307 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_sgemv_8_lib8.S |
| 308 | ${PROJECT_SOURCE_DIR}/kernel/avx/kernel_sgemv_4_lib8.S |
| 309 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_ssymv_4_lib8.c |
| 310 | ) |
| 311 | |
| 312 | file(GLOB BLAS_SRC |
| 313 | ${PROJECT_SOURCE_DIR}/blas/d_blas1_lib4.c |
| 314 | ${PROJECT_SOURCE_DIR}/blas/d_blas2_lib4.c |
| 315 | ${PROJECT_SOURCE_DIR}/blas/d_blas2_diag_lib.c |
| 316 | ${PROJECT_SOURCE_DIR}/blas/d_blas3_lib4.c |
| 317 | ${PROJECT_SOURCE_DIR}/blas/d_blas3_diag_lib4.c |
| 318 | ${PROJECT_SOURCE_DIR}/blas/d_lapack_lib4.c |
| 319 | ${PROJECT_SOURCE_DIR}/blas/s_blas1_lib8.c |
| 320 | ${PROJECT_SOURCE_DIR}/blas/s_blas2_lib8.c |
| 321 | ${PROJECT_SOURCE_DIR}/blas/s_blas2_diag_lib.c |
| 322 | ${PROJECT_SOURCE_DIR}/blas/s_blas3_lib8.c |
| 323 | ${PROJECT_SOURCE_DIR}/blas/s_blas3_diag_lib8.c |
| 324 | ${PROJECT_SOURCE_DIR}/blas/s_lapack_lib8.c |
| 325 | ) |
| 326 | |
| 327 | endif(${TARGET} MATCHES X64_INTEL_SANDY_BRIDGE) |
| 328 | |
| 329 | if(${TARGET} MATCHES X64_INTEL_CORE) |
| 330 | |
| 331 | file(GLOB AUX_SRC |
| 332 | ${PROJECT_SOURCE_DIR}/auxiliary/d_aux_lib4.c |
| 333 | ${PROJECT_SOURCE_DIR}/auxiliary/c99/kernel_dgecp_lib4.c |
| 334 | ${PROJECT_SOURCE_DIR}/auxiliary/c99/kernel_dgetr_lib4.c |
| 335 | ${PROJECT_SOURCE_DIR}/auxiliary/s_aux_lib4.c |
| 336 | ${PROJECT_SOURCE_DIR}/auxiliary/c99/kernel_sgetr_lib4.c |
| 337 | ${PROJECT_SOURCE_DIR}/auxiliary/m_aux_lib44.c |
| 338 | ) |
| 339 | |
| 340 | file(GLOB KERNEL_SRC |
| 341 | ${PROJECT_SOURCE_DIR}/kernel/sse3/kernel_dgemm_4x4_lib4.S |
| 342 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgemm_4x4_lib4.c |
| 343 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgemm_diag_lib4.c |
| 344 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgemv_4_lib4.c |
| 345 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dsymv_4_lib4.c |
| 346 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgetrf_pivot_4_lib4.c |
| 347 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgeqrf_4_lib4.c |
| 348 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgemm_4x4_lib4.c |
| 349 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgemm_diag_lib4.c |
| 350 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgemv_4_lib4.c |
| 351 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_ssymv_4_lib4.c |
| 352 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgetrf_pivot_4_lib4.c |
| 353 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgecp_lib4.c |
| 354 | ) |
| 355 | |
| 356 | file(GLOB BLAS_SRC |
| 357 | ${PROJECT_SOURCE_DIR}/blas/d_blas1_lib4.c |
| 358 | ${PROJECT_SOURCE_DIR}/blas/d_blas2_lib4.c |
| 359 | ${PROJECT_SOURCE_DIR}/blas/d_blas2_diag_lib.c |
| 360 | ${PROJECT_SOURCE_DIR}/blas/d_blas3_lib4.c |
| 361 | ${PROJECT_SOURCE_DIR}/blas/d_blas3_diag_lib4.c |
| 362 | ${PROJECT_SOURCE_DIR}/blas/d_lapack_lib4.c |
| 363 | ${PROJECT_SOURCE_DIR}/blas/s_blas1_lib4.c |
| 364 | ${PROJECT_SOURCE_DIR}/blas/s_blas2_lib4.c |
| 365 | ${PROJECT_SOURCE_DIR}/blas/s_blas2_diag_lib.c |
| 366 | ${PROJECT_SOURCE_DIR}/blas/s_blas3_lib4.c |
| 367 | ${PROJECT_SOURCE_DIR}/blas/s_blas3_diag_lib4.c |
| 368 | ${PROJECT_SOURCE_DIR}/blas/s_lapack_lib4.c |
| 369 | ) |
| 370 | |
| 371 | endif(${TARGET} MATCHES X64_INTEL_CORE) |
| 372 | |
| 373 | if(${TARGET} MATCHES X64_AMD_BULLDOZER) |
| 374 | |
| 375 | file(GLOB AUX_SRC |
| 376 | ${PROJECT_SOURCE_DIR}/auxiliary/d_aux_lib4.c |
| 377 | ${PROJECT_SOURCE_DIR}/auxiliary/c99/kernel_dgecp_lib4.c |
| 378 | ${PROJECT_SOURCE_DIR}/auxiliary/c99/kernel_dgetr_lib4.c |
| 379 | ${PROJECT_SOURCE_DIR}/auxiliary/s_aux_lib4.c |
| 380 | ${PROJECT_SOURCE_DIR}/auxiliary/c99/kernel_sgetr_lib4.c |
| 381 | ${PROJECT_SOURCE_DIR}/auxiliary/m_aux_lib44.c |
| 382 | ) |
| 383 | |
| 384 | file(GLOB KERNEL_SRC |
| 385 | ${PROJECT_SOURCE_DIR}/kernel/fma/kernel_dgemm_4x4_lib4.S |
| 386 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgemm_4x4_lib4.c |
| 387 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgemm_diag_lib4.c |
| 388 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgemv_4_lib4.c |
| 389 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dsymv_4_lib4.c |
| 390 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgetrf_pivot_4_lib4.c |
| 391 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgeqrf_4_lib4.c |
| 392 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgemm_4x4_lib4.c |
| 393 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgemm_diag_lib4.c |
| 394 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgemv_4_lib4.c |
| 395 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_ssymv_4_lib4.c |
| 396 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgetrf_pivot_4_lib4.c |
| 397 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgecp_lib4.c |
| 398 | ) |
| 399 | |
| 400 | file(GLOB BLAS_SRC |
| 401 | ${PROJECT_SOURCE_DIR}/blas/d_blas1_lib4.c |
| 402 | ${PROJECT_SOURCE_DIR}/blas/d_blas2_lib4.c |
| 403 | ${PROJECT_SOURCE_DIR}/blas/d_blas2_diag_lib.c |
| 404 | ${PROJECT_SOURCE_DIR}/blas/d_blas3_lib4.c |
| 405 | ${PROJECT_SOURCE_DIR}/blas/d_blas3_diag_lib4.c |
| 406 | ${PROJECT_SOURCE_DIR}/blas/d_lapack_lib4.c |
| 407 | ${PROJECT_SOURCE_DIR}/blas/s_blas1_lib4.c |
| 408 | ${PROJECT_SOURCE_DIR}/blas/s_blas2_lib4.c |
| 409 | ${PROJECT_SOURCE_DIR}/blas/s_blas2_diag_lib.c |
| 410 | ${PROJECT_SOURCE_DIR}/blas/s_blas3_lib4.c |
| 411 | ${PROJECT_SOURCE_DIR}/blas/s_blas3_diag_lib4.c |
| 412 | ${PROJECT_SOURCE_DIR}/blas/s_lapack_lib4.c |
| 413 | ) |
| 414 | |
| 415 | endif(${TARGET} MATCHES X64_AMD_BULLDOZER) |
| 416 | |
| 417 | if(${TARGET} MATCHES ARMV8A_ARM_CORTEX_A57) |
| 418 | |
| 419 | file(GLOB AUX_SRC |
| 420 | ${PROJECT_SOURCE_DIR}/auxiliary/d_aux_lib4.c |
| 421 | ${PROJECT_SOURCE_DIR}/auxiliary/c99/kernel_dgecp_lib4.c |
| 422 | ${PROJECT_SOURCE_DIR}/auxiliary/c99/kernel_dgetr_lib4.c |
| 423 | ${PROJECT_SOURCE_DIR}/auxiliary/s_aux_lib4.c |
| 424 | ${PROJECT_SOURCE_DIR}/auxiliary/c99/kernel_sgetr_lib4.c |
| 425 | ${PROJECT_SOURCE_DIR}/auxiliary/m_aux_lib44.c |
| 426 | ) |
| 427 | |
| 428 | file(GLOB KERNEL_SRC |
| 429 | ${PROJECT_SOURCE_DIR}/kernel/armv8a/kernel_dgemm_8x4_lib4.S |
| 430 | ${PROJECT_SOURCE_DIR}/kernel/armv8a/kernel_dgemm_4x4_lib4.S |
| 431 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgemm_4x4_lib4.c |
| 432 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgemm_diag_lib4.c |
| 433 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgemv_4_lib4.c |
| 434 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dsymv_4_lib4.c |
| 435 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgetrf_pivot_4_lib4.c |
| 436 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgeqrf_4_lib4.c |
| 437 | ${PROJECT_SOURCE_DIR}/kernel/armv8a/kernel_sgemm_16x4_lib4.S |
| 438 | ${PROJECT_SOURCE_DIR}/kernel/armv8a/kernel_sgemm_12x4_lib4.S |
| 439 | ${PROJECT_SOURCE_DIR}/kernel/armv8a/kernel_sgemm_8x8_lib4.S |
| 440 | ${PROJECT_SOURCE_DIR}/kernel/armv8a/kernel_sgemm_8x4_lib4.S |
| 441 | ${PROJECT_SOURCE_DIR}/kernel/armv8a/kernel_sgemm_4x4_lib4.S |
| 442 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgemm_4x4_lib4.c |
| 443 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgemm_diag_lib4.c |
| 444 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgemv_4_lib4.c |
| 445 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_ssymv_4_lib4.c |
| 446 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgetrf_pivot_4_lib4.c |
| 447 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgecp_lib4.c |
| 448 | ) |
| 449 | |
| 450 | file(GLOB BLAS_SRC |
| 451 | ${PROJECT_SOURCE_DIR}/blas/d_blas1_lib4.c |
| 452 | ${PROJECT_SOURCE_DIR}/blas/d_blas2_lib4.c |
| 453 | ${PROJECT_SOURCE_DIR}/blas/d_blas2_diag_lib.c |
| 454 | ${PROJECT_SOURCE_DIR}/blas/d_blas3_lib4.c |
| 455 | ${PROJECT_SOURCE_DIR}/blas/d_blas3_diag_lib4.c |
| 456 | ${PROJECT_SOURCE_DIR}/blas/d_lapack_lib4.c |
| 457 | ${PROJECT_SOURCE_DIR}/blas/s_blas1_lib4.c |
| 458 | ${PROJECT_SOURCE_DIR}/blas/s_blas2_lib4.c |
| 459 | ${PROJECT_SOURCE_DIR}/blas/s_blas2_diag_lib.c |
| 460 | ${PROJECT_SOURCE_DIR}/blas/s_blas3_lib4.c |
| 461 | ${PROJECT_SOURCE_DIR}/blas/s_blas3_diag_lib4.c |
| 462 | ${PROJECT_SOURCE_DIR}/blas/s_lapack_lib4.c |
| 463 | ) |
| 464 | |
| 465 | endif(${TARGET} MATCHES ARMV8A_ARM_CORTEX_A57) |
| 466 | |
| 467 | if(${TARGET} MATCHES ARMV7A_ARM_CORTEX_A15) |
| 468 | |
| 469 | file(GLOB AUX_SRC |
| 470 | ${PROJECT_SOURCE_DIR}/auxiliary/d_aux_lib4.c |
| 471 | ${PROJECT_SOURCE_DIR}/auxiliary/c99/kernel_dgecp_lib4.c |
| 472 | ${PROJECT_SOURCE_DIR}/auxiliary/c99/kernel_dgetr_lib4.c |
| 473 | ${PROJECT_SOURCE_DIR}/auxiliary/s_aux_lib4.c |
| 474 | ${PROJECT_SOURCE_DIR}/auxiliary/c99/kernel_sgetr_lib4.c |
| 475 | ${PROJECT_SOURCE_DIR}/auxiliary/m_aux_lib44.c |
| 476 | ) |
| 477 | |
| 478 | file(GLOB KERNEL_SRC |
| 479 | ${PROJECT_SOURCE_DIR}/kernel/armv7a/kernel_dgemm_4x4_lib4.S |
| 480 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgemm_4x4_lib4.c |
| 481 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgemm_diag_lib4.c |
| 482 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgemv_4_lib4.c |
| 483 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dsymv_4_lib4.c |
| 484 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgetrf_pivot_4_lib4.c |
| 485 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgeqrf_4_lib4.c |
| 486 | ${PROJECT_SOURCE_DIR}/kernel/armv7a/kernel_sgemm_12x4_lib4.S |
| 487 | ${PROJECT_SOURCE_DIR}/kernel/armv7a/kernel_sgemm_8x4_lib4.S |
| 488 | ${PROJECT_SOURCE_DIR}/kernel/armv7a/kernel_sgemm_4x4_lib4.S |
| 489 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgemm_4x4_lib4.c |
| 490 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgemm_diag_lib4.c |
| 491 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgemv_4_lib4.c |
| 492 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_ssymv_4_lib4.c |
| 493 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgetrf_pivot_4_lib4.c |
| 494 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgecp_lib4.c |
| 495 | ) |
| 496 | |
| 497 | file(GLOB BLAS_SRC |
| 498 | ${PROJECT_SOURCE_DIR}/blas/d_blas1_lib4.c |
| 499 | ${PROJECT_SOURCE_DIR}/blas/d_blas2_lib4.c |
| 500 | ${PROJECT_SOURCE_DIR}/blas/d_blas2_diag_lib.c |
| 501 | ${PROJECT_SOURCE_DIR}/blas/d_blas3_lib4.c |
| 502 | ${PROJECT_SOURCE_DIR}/blas/d_blas3_diag_lib4.c |
| 503 | ${PROJECT_SOURCE_DIR}/blas/d_lapack_lib4.c |
| 504 | ${PROJECT_SOURCE_DIR}/blas/s_blas1_lib4.c |
| 505 | ${PROJECT_SOURCE_DIR}/blas/s_blas2_lib4.c |
| 506 | ${PROJECT_SOURCE_DIR}/blas/s_blas2_diag_lib.c |
| 507 | ${PROJECT_SOURCE_DIR}/blas/s_blas3_lib4.c |
| 508 | ${PROJECT_SOURCE_DIR}/blas/s_blas3_diag_lib4.c |
| 509 | ${PROJECT_SOURCE_DIR}/blas/s_lapack_lib4.c |
| 510 | ) |
| 511 | |
| 512 | endif(${TARGET} MATCHES ARMV7A_ARM_CORTEX_A15) |
| 513 | |
| 514 | if(${TARGET} MATCHES GENERIC) |
| 515 | |
| 516 | file(GLOB AUX_SRC |
| 517 | ${PROJECT_SOURCE_DIR}/auxiliary/d_aux_lib4.c |
| 518 | ${PROJECT_SOURCE_DIR}/auxiliary/c99/kernel_dgecp_lib4.c |
| 519 | ${PROJECT_SOURCE_DIR}/auxiliary/c99/kernel_dgetr_lib4.c |
| 520 | ${PROJECT_SOURCE_DIR}/auxiliary/s_aux_lib4.c |
| 521 | ${PROJECT_SOURCE_DIR}/auxiliary/c99/kernel_sgetr_lib4.c |
| 522 | ${PROJECT_SOURCE_DIR}/auxiliary/m_aux_lib44.c |
| 523 | ) |
| 524 | |
| 525 | file(GLOB KERNEL_SRC |
| 526 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgemm_4x4_lib4.c |
| 527 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgemm_diag_lib4.c |
| 528 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgemv_4_lib4.c |
| 529 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dsymv_4_lib4.c |
| 530 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgetrf_pivot_4_lib4.c |
| 531 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_dgeqrf_4_lib4.c |
| 532 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgemm_4x4_lib4.c |
| 533 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgemm_diag_lib4.c |
| 534 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgemv_4_lib4.c |
| 535 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_ssymv_4_lib4.c |
| 536 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgetrf_pivot_4_lib4.c |
| 537 | ${PROJECT_SOURCE_DIR}/kernel/c99/kernel_sgecp_lib4.c |
| 538 | ) |
| 539 | |
| 540 | file(GLOB BLAS_SRC |
| 541 | ${PROJECT_SOURCE_DIR}/blas/d_blas1_lib4.c |
| 542 | ${PROJECT_SOURCE_DIR}/blas/d_blas2_lib4.c |
| 543 | ${PROJECT_SOURCE_DIR}/blas/d_blas2_diag_lib.c |
| 544 | ${PROJECT_SOURCE_DIR}/blas/d_blas3_lib4.c |
| 545 | ${PROJECT_SOURCE_DIR}/blas/d_blas3_diag_lib4.c |
| 546 | ${PROJECT_SOURCE_DIR}/blas/d_lapack_lib4.c |
| 547 | ${PROJECT_SOURCE_DIR}/blas/s_blas1_lib4.c |
| 548 | ${PROJECT_SOURCE_DIR}/blas/s_blas2_lib4.c |
| 549 | ${PROJECT_SOURCE_DIR}/blas/s_blas2_diag_lib.c |
| 550 | ${PROJECT_SOURCE_DIR}/blas/s_blas3_lib4.c |
| 551 | ${PROJECT_SOURCE_DIR}/blas/s_blas3_diag_lib4.c |
| 552 | ${PROJECT_SOURCE_DIR}/blas/s_lapack_lib4.c |
| 553 | ) |
| 554 | |
| 555 | endif(${TARGET} MATCHES GENERIC) |
| 556 | |
| 557 | else(${LA} MATCHES HIGH_PERFORMANCE) # REFERENCE BLAS |
| 558 | |
| 559 | file(GLOB AUX_SRC |
| 560 | ${PROJECT_SOURCE_DIR}/auxiliary/d_aux_lib.c |
| 561 | ${PROJECT_SOURCE_DIR}/auxiliary/s_aux_lib.c |
| 562 | ${PROJECT_SOURCE_DIR}/auxiliary/m_aux_lib.c |
| 563 | ) |
| 564 | |
| 565 | file(GLOB BLAS_SRC |
| 566 | ${PROJECT_SOURCE_DIR}/blas/d_blas1_lib.c |
| 567 | ${PROJECT_SOURCE_DIR}/blas/d_blas2_lib.c |
| 568 | ${PROJECT_SOURCE_DIR}/blas/d_blas2_diag_lib.c |
| 569 | ${PROJECT_SOURCE_DIR}/blas/d_blas3_lib.c |
| 570 | ${PROJECT_SOURCE_DIR}/blas/d_blas3_diag_lib.c |
| 571 | ${PROJECT_SOURCE_DIR}/blas/d_lapack_lib.c |
| 572 | ${PROJECT_SOURCE_DIR}/blas/s_blas1_lib.c |
| 573 | ${PROJECT_SOURCE_DIR}/blas/s_blas2_lib.c |
| 574 | ${PROJECT_SOURCE_DIR}/blas/s_blas2_diag_lib.c |
| 575 | ${PROJECT_SOURCE_DIR}/blas/s_blas3_lib.c |
| 576 | ${PROJECT_SOURCE_DIR}/blas/s_blas3_diag_lib.c |
| 577 | ${PROJECT_SOURCE_DIR}/blas/s_lapack_lib.c |
| 578 | ) |
| 579 | |
| 580 | endif(${LA} MATCHES HIGH_PERFORMANCE) |
| 581 | |
| 582 | if(${EXT_DEP}) |
| 583 | |
| 584 | file(GLOB EXT_SRC |
| 585 | ${PROJECT_SOURCE_DIR}/auxiliary/d_aux_ext_dep_lib.c |
| 586 | ${PROJECT_SOURCE_DIR}/auxiliary/s_aux_ext_dep_lib.c |
| 587 | ${PROJECT_SOURCE_DIR}/auxiliary/v_aux_ext_dep_lib.c |
| 588 | ${PROJECT_SOURCE_DIR}/auxiliary/i_aux_ext_dep_lib.c |
| 589 | ) |
| 590 | |
| 591 | endif() |
| 592 | |
| 593 | set(BLASFEO_SRC ${AUX_SRC} ${KERNEL_SRC} ${BLAS_SRC} ${EXT_SRC}) |
| 594 | |
| 595 | # add library |
| 596 | add_library(blasfeo ${BLASFEO_SRC}) |
| 597 | target_include_directories(blasfeo |
| 598 | PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>) |
| 599 | |
| 600 | install(TARGETS blasfeo EXPORT blasfeoConfig |
| 601 | LIBRARY DESTINATION lib |
| 602 | ARCHIVE DESTINATION lib |
| 603 | RUNTIME DESTINATION bin) |
| 604 | |
| 605 | install(EXPORT blasfeoConfig DESTINATION cmake) |
| 606 | |
| 607 | file(GLOB_RECURSE BLASFEO_HEADERS "include/*.h") |
| 608 | install(FILES ${BLASFEO_HEADERS} DESTINATION ${BLASFEO_HEADERS_INSTALLATION_DIRECTORY}) |
| 609 | |
| 610 | # test problems |
| 611 | # add_subdirectory(test_problems) |