diff --git a/kernel/CMakeLists.txt b/kernel/CMakeLists.txt index c0a7543ca..a720f6249 100644 --- a/kernel/CMakeLists.txt +++ b/kernel/CMakeLists.txt @@ -1,106 +1,545 @@ include_directories(${PROJECT_SOURCE_DIR}) -include_directories(${PROJECT_BINARY_DIR}) + +# Makefile +function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS) + set (OPENBLAS_SRC "") + set (ARCH_SUFFIX "") + include("${PROJECT_SOURCE_DIR}/cmake/kernel.cmake") + if (${DYNAMIC_ARCH}) + include("${PROJECT_SOURCE_DIR}/cmake/system.cmake") + endif () + SetDefaultL1() + SetDefaultL2() + SetDefaultL3() + ParseMakefileVars("${KERNELDIR}/KERNEL") + ParseMakefileVars("${KERNELDIR}/KERNEL.${TARGET_CORE}") + + set(KERNEL_INTERFACE common_level1.h common_level2.h common_level3.h) + if(NOT NO_LAPACK) + set(KERNEL_INTERFACE ${KERNEL_INTERFACE} common_lapack.h) + endif () + + if (${ADD_COMMONOBJS}) + if (X86) + if (NOT "${CMAKE_C_COMPILER_ID}" STREQUAL "MSVC") + GenerateNamedObjects("${KERNELDIR}/cpuid.S" "" "" false "" "" true) + else() + GenerateNamedObjects("${KERNELDIR}/cpuid_win.c" "" "" false "" "" true) + endif() + endif () + + # don't use float type name mangling here + GenerateNamedObjects("${KERNELDIR}/${LSAME_KERNEL}" "F_INTERFACE" "lsame" false "" "" true "") + GenerateNamedObjects("${KERNELDIR}/${SCABS_KERNEL}" "COMPLEX;F_INTERFACE" "scabs1" false "" "" true "") + GenerateNamedObjects("${KERNELDIR}/${DCABS_KERNEL}" "DOUBLE;COMPLEX;F_INTERFACE" "dcabs1" false "" "" true "") + endif () + + # Run with no ARCH_SUFFIX for above + set (ARCH_SUFFIX "${TSUFFIX}") + # Makefile.L1 + foreach (float_type ${FLOAT_TYPES}) + # a bit of metaprogramming here to pull out the appropriate KERNEL var + string(SUBSTRING ${float_type} 0 1 float_char) + GenerateNamedObjects("${KERNELDIR}/${${float_char}AMAXKERNEL}" "USE_ABS" "amax_k" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}AMINKERNEL}" "USE_ABS;USE_MIN" "amin_k" false "" "" false ${float_type}) + if (DEFINED ${float_char}MAXKERNEL) + GenerateNamedObjects("${KERNELDIR}/${${float_char}MAXKERNEL}" "" "max_k" false "" "" false ${float_type}) + endif () + if (DEFINED ${float_char}MINKERNEL) + GenerateNamedObjects("${KERNELDIR}/${${float_char}MINKERNEL}" "" "min_k" false "" "" false ${float_type}) + endif () + GenerateNamedObjects("${KERNELDIR}/${I${float_char}AMAXKERNEL}" "USE_ABS" "i*amax_k" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${I${float_char}AMINKERNEL}" "USE_ABS;USE_MIN" "i*amin_k" false "" "" false ${float_type}) + if (DEFINED I${float_char}MAXKERNEL) + GenerateNamedObjects("${KERNELDIR}/${I${float_char}MAXKERNEL}" "" "i*max_k" false "" "" false ${float_type}) + endif () + if (DEFINED I${float_char}MINKERNEL) + GenerateNamedObjects("${KERNELDIR}/${I${float_char}MINKERNEL}" "" "i*min_k" false "" "" false ${float_type}) + endif () + GenerateNamedObjects("${KERNELDIR}/${${float_char}ASUMKERNEL}" "" "asum_k" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}AXPYKERNEL}" "" "axpy_k" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}COPYKERNEL}" "C_INTERFACE" "copy_k" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}NRM2KERNEL}" "" "nrm2_k" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}ROTKERNEL}" "" "rot_k" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}SCALKERNEL}" "" "scal_k" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}SWAPKERNEL}" "" "swap_k" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}AXPBYKERNEL}" "" "axpby_k" false "" "" false ${float_type}) + + if (${float_type} STREQUAL "COMPLEX" OR ${float_type} STREQUAL "ZCOMPLEX") + GenerateNamedObjects("${KERNELDIR}/${${float_char}AXPYKERNEL}" "CONJ" "axpyc_k" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}DOTKERNEL}" "" "dotu_k" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}DOTKERNEL}" "CONJ" "dotc_k" false "" "" false ${float_type}) + else () + GenerateNamedObjects("${KERNELDIR}/${${float_char}DOTKERNEL}" "" "dot_k" false "" "" false ${float_type}) + endif () + + if (${float_type} STREQUAL "COMPLEX") + GenerateNamedObjects("${KERNELDIR}/${${float_char}ROTKERNEL}" "" "srot_k" false "" "" false ${float_type}) + endif() + if (${float_type} STREQUAL "ZCOMPLEX") + GenerateNamedObjects("${KERNELDIR}/${${float_char}ROTKERNEL}" "" "drot_k" false "" "" false ${float_type}) + endif() + + endforeach () + + #dsdot,sdsdot + GenerateNamedObjects("${KERNELDIR}/${DSDOTKERNEL}" "DSDOT" "d*dot_k" false "" "" false "SINGLE") + GenerateNamedObjects("${KERNELDIR}/${DSDOTKERNEL}" "DSDOT" "dsdot_k" false "" "" false "SINGLE") + + # Makefile.L2 + GenerateCombinationObjects("generic/symv_k.c" "LOWER" "U" "" 1 "" "" 3) + GenerateNamedObjects("generic/ger.c" "" "ger_k" false "" "" "" 3) + foreach (float_type ${FLOAT_TYPES}) + string(SUBSTRING ${float_type} 0 1 float_char) + if (${float_type} STREQUAL "COMPLEX" OR ${float_type} STREQUAL "ZCOMPLEX") + GenerateNamedObjects("${KERNELDIR}/${${float_char}GERUKERNEL}" "" "geru_k" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GERCKERNEL}" "CONJ" "gerc_k" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GERUKERNEL}" "XCONJ" "gerv_k" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GERCKERNEL}" "CONJ;XCONJ" "gerd_k" false "" "" false ${float_type}) + + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMVNKERNEL}" "" "gemv_n" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMVTKERNEL}" "TRANSA" "gemv_t" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMVNKERNEL}" "CONJ" "gemv_r" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMVTKERNEL}" "CONJ;TRANSA" "gemv_c" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMVNKERNEL}" "XCONJ" "gemv_o" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMVTKERNEL}" "XCONJ;TRANSA" "gemv_u" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMVNKERNEL}" "XCONJ;CONJ" "gemv_s" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMVTKERNEL}" "XCONJ;CONJ;TRANSA" "gemv_d" false "" "" false ${float_type}) + + GenerateNamedObjects("${KERNELDIR}/${${float_char}HEMV_U_KERNEL}" "HEMV" "hemv_U" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}HEMV_L_KERNEL}" "HEMV;LOWER" "hemv_L" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}HEMV_V_KERNEL}" "HEMV;HEMVREV" "hemv_V" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}HEMV_M_KERNEL}" "HEMV;HEMVREV;LOWER" "hemv_M" false "" "" false ${float_type}) + + else () + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMVNKERNEL}" "" "gemv_n" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMVTKERNEL}" "TRANS" "gemv_t" false "" "" false ${float_type}) + endif () + endforeach () + + # Makefile.L3 + set(USE_TRMM false) + + if (ARM OR ARM64 OR "${TARGET}" STREQUAL "LONGSOON3B" OR "${TARGET}" STREQUAL "GENERIC" OR "${CORE}" STREQUAL "generic" OR "${TARGET}" STREQUAL "HASWELL" OR "${CORE}" STREQUAL "haswell" OR "{CORE}" STREQUAL "zen") + set(USE_TRMM true) + endif () + + foreach (float_type ${FLOAT_TYPES}) + string(SUBSTRING ${float_type} 0 1 float_char) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMMKERNEL}" "" "gemm_kernel" false "" "" false ${float_type}) + + if (${float_char}GEMMINCOPY) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMMINCOPY}" "${float_type}" "${${float_char}GEMMINCOPYOBJ}" false "" "" true ${float_type}) + endif () + + if (${float_char}GEMMITCOPY) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMMITCOPY}" "${float_type}" "${${float_char}GEMMITCOPYOBJ}" false "" "" true ${float_type}) + endif () + + if (${float_char}GEMMONCOPY) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMMONCOPY}" "${float_type}" "${${float_char}GEMMONCOPYOBJ}" false "" "" true ${float_type}) + endif () + + if (${float_char}GEMMOTCOPY) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMMOTCOPY}" "${float_type}" "${${float_char}GEMMOTCOPYOBJ}" false "" "" true ${float_type}) + endif () + + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMM_BETA}" "" "gemm_beta" false "" "" false ${float_type}) + + if (USE_TRMM) + set(TRMM_KERNEL "${${float_char}TRMMKERNEL}") + else () + set(TRMM_KERNEL "${${float_char}GEMMKERNEL}") + endif () + + if (${float_type} STREQUAL "COMPLEX" OR ${float_type} STREQUAL "ZCOMPLEX") + + # just enumerate all these. there is an extra define for these indicating which side is a conjugate (e.g. CN NC NN) that I don't really want to work into GenerateCombinationObjects + + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMMKERNEL}" "NN" "gemm_kernel_n" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMMKERNEL}" "CN" "gemm_kernel_l" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMMKERNEL}" "NC" "gemm_kernel_r" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMMKERNEL}" "CC" "gemm_kernel_b" false "" "" false ${float_type}) + + GenerateNamedObjects("${KERNELDIR}/${TRMM_KERNEL}" "TRMMKERNEL;LEFT;NN" "trmm_kernel_LN" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${TRMM_KERNEL}" "TRMMKERNEL;LEFT;TRANSA;NN" "trmm_kernel_LT" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${TRMM_KERNEL}" "TRMMKERNEL;LEFT;CONJ;CN" "trmm_kernel_LR" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${TRMM_KERNEL}" "TRMMKERNEL;LEFT;TRANSA;CONJ;CN" "trmm_kernel_LC" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${TRMM_KERNEL}" "TRMMKERNEL;NN" "trmm_kernel_RN" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${TRMM_KERNEL}" "TRMMKERNEL;TRANSA;NN" "trmm_kernel_RT" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${TRMM_KERNEL}" "TRMMKERNEL;CONJ;NC" "trmm_kernel_RR" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${TRMM_KERNEL}" "TRMMKERNEL;TRANSA;CONJ;NC" "trmm_kernel_RC" false "" "" false ${float_type}) + + GenerateNamedObjects("${KERNELDIR}/${${float_char}TRSMKERNEL_LN}" "UPPER;LN;TRSMKERNEL;CONJ" "trsm_kernel_LR" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}TRSMKERNEL_LT}" "LT;TRSMKERNEL;CONJ" "trsm_kernel_LC" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}TRSMKERNEL_RN}" "UPPER;RN;TRSMKERNEL;CONJ" "trsm_kernel_RR" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}TRSMKERNEL_RT}" "RT;TRSMKERNEL;CONJ" "trsm_kernel_RC" false "" "" false ${float_type}) -set(LAPACK_SOURCES - getrf/getrf_single.c - potrf/potrf_U_single.c - potrf/potrf_L_single.c - lauum/lauum_U_single.c - lauum/lauum_L_single.c -) + #hemm + GenerateNamedObjects("generic/zhemm_utcopy_${${float_char}GEMM_UNROLL_M}.c" "" "hemm_iutcopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/zhemm_ltcopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER" "hemm_iltcopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/zhemm_utcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER" "hemm_outcopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/zhemm_ltcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER" "hemm_oltcopy" false "" "" false ${float_type}) -# add a 'z' to filename for complex version -set(LAPACK_MANGLED_SOURCES - getf2/getf2_k.c - lauu2/lauu2_U.c - lauu2/lauu2_L.c - potf2/potf2_U.c - potf2/potf2_L.c -) + # symm for c and z + GenerateNamedObjects("generic/zsymm_ucopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER" "symm_outcopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/zsymm_ucopy_${${float_char}GEMM_UNROLL_M}.c" "" "symm_iutcopy" false "" "" false ${float_type}) -# sources that need TRANS set -# this has a 'z' version -set(TRANS_SOURCES - getrs/getrs_single.c -) + GenerateNamedObjects("generic/zsymm_lcopy_${${float_char}GEMM_UNROLL_N}.c" "LOWER;OUTER" "symm_oltcopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/zsymm_lcopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER" "symm_iltcopy" false "" "" false ${float_type}) -# sources that need UNIT set -# these do NOT have a z version -set(UNIT_SOURCES - trtri/trtri_U_single.c - trtri/trtri_L_single.c -) + GenerateNamedObjects("generic/ztrmm_uncopy_${${float_char}GEMM_UNROLL_M}.c" "UNIT" "trmm_iunucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrmm_uncopy_${${float_char}GEMM_UNROLL_M}.c" "" "trmm_iunncopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrmm_uncopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;UNIT" "trmm_ounucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrmm_uncopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER" "trmm_ounncopy" false "" "" false ${float_type}) -# these have a 'z' version -set(UNIT_SOURCES2 - trti2/trti2_U.c - trti2/trti2_L.c -) + GenerateNamedObjects("generic/ztrmm_lncopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER;UNIT" "trmm_ilnucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrmm_lncopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER" "trmm_ilnncopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrmm_lncopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER;UNIT" "trmm_olnucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrmm_lncopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER" "trmm_olnncopy" false "" "" false ${float_type}) -GenerateNamedObjects("${LAPACK_SOURCES}") -GenerateNamedObjects("${LAPACK_MANGLED_SOURCES}" "" "" false "" "" false 3) + GenerateNamedObjects("generic/ztrmm_utcopy_${${float_char}GEMM_UNROLL_M}.c" "UNIT" "trmm_iutucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrmm_utcopy_${${float_char}GEMM_UNROLL_M}.c" "" "trmm_iutncopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrmm_utcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;UNIT" "trmm_outucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrmm_utcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER" "trmm_outncopy" false "" "" false ${float_type}) -GenerateNamedObjects("laswp/generic/laswp_k_4.c" "" "laswp_plus" false "" "" false 3) -GenerateNamedObjects("laswp/generic/laswp_k_4.c" "MINUS" "laswp_minus" false "" "" false 3) + GenerateNamedObjects("generic/ztrmm_ltcopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER;UNIT" "trmm_iltucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrmm_ltcopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER" "trmm_iltncopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrmm_ltcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER;UNIT" "trmm_oltucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrmm_ltcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER" "trmm_oltncopy" false "" "" false ${float_type}) -# dynamic_arch laswp needs arch specific code ? -#foreach(TARGET_CORE ${DYNAMIC_CORE}) -# set(TSUFFIX "_${TARGET_CORE}") -# -#GenerateNamedObjects("laswp/generic/laswp_k_4.c" "" "laswp_plus" false "" ${TSUFFIX} false 3) -#GenerateNamedObjects("laswp/generic/laswp_k_4.c" "MINUS" "laswp_minus" false "" ${TSUFFIX} false 3) -#endforeach () + GenerateNamedObjects("generic/ztrsm_uncopy_${${float_char}GEMM_UNROLL_M}.c" "UNIT" "trsm_iunucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrsm_uncopy_${${float_char}GEMM_UNROLL_M}.c" "" "trsm_iunncopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrsm_uncopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;UNIT" "trsm_ounucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrsm_uncopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER" "trsm_ounncopy" false "" "" false ${float_type}) -if (USE_THREAD) + GenerateNamedObjects("generic/ztrsm_lncopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER;UNIT" "trsm_ilnucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrsm_lncopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER" "trsm_ilnncopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrsm_lncopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER;UNIT" "trsm_olnucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrsm_lncopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER" "trsm_olnncopy" false "" "" false ${float_type}) - if (USE_OPENMP) - set(GETRF_SRC getrf/getrf_parallel_omp.c) - else () - set(GETRF_SRC getrf/getrf_parallel.c) - endif () + GenerateNamedObjects("generic/ztrsm_utcopy_${${float_char}GEMM_UNROLL_M}.c" "UNIT" "trsm_iutucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrsm_utcopy_${${float_char}GEMM_UNROLL_M}.c" "" "trsm_iutncopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrsm_utcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;UNIT" "trsm_outucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrsm_utcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER" "trsm_outncopy" false "" "" false ${float_type}) - # these do not have 'z' versions - set(PARALLEL_SOURCES - ${GETRF_SRC} - lauum/lauum_U_parallel.c - lauum/lauum_L_parallel.c - potrf/potrf_U_parallel.c - potrf/potrf_L_parallel.c - ) + GenerateNamedObjects("generic/ztrsm_ltcopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER;UNIT" "trsm_iltucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrsm_ltcopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER" "trsm_iltncopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrsm_ltcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER;UNIT" "trsm_oltucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/ztrsm_ltcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER" "trsm_oltncopy" false "" "" false ${float_type}) - # this has a z version - list(APPEND TRANS_SOURCES - getrs/getrs_parallel.c - ) + #gemm3m + if (USE_GEMM3M) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMM3MKERNEL}" "NN" "gemm3m_kernel" false "" "" false ${float_type}) - # these do NOT have a z version - list(APPEND UNIT_SOURCES - trtri/trtri_U_parallel.c - trtri/trtri_L_parallel.c - ) + GenerateNamedObjects("generic/zgemm3m_ncopy_${${float_char}GEMM3M_UNROLL_N}.c" "USE_ALPHA" "gemm3m_oncopyb" false "" "" false ${float_type}) + GenerateNamedObjects("generic/zgemm3m_ncopy_${${float_char}GEMM3M_UNROLL_N}.c" "USE_ALPHA;REAL_ONLY" "gemm3m_oncopyr" false "" "" false ${float_type}) + GenerateNamedObjects("generic/zgemm3m_ncopy_${${float_char}GEMM3M_UNROLL_N}.c" "USE_ALPHA;IMAGE_ONLY" "gemm3m_oncopyi" false "" "" false ${float_type}) - GenerateNamedObjects("${PARALLEL_SOURCES}") + GenerateNamedObjects("generic/zgemm3m_tcopy_${${float_char}GEMM3M_UNROLL_N}.c" "USE_ALPHA" "gemm3m_otcopyb" false "" "" false ${float_type}) + GenerateNamedObjects("generic/zgemm3m_tcopy_${${float_char}GEMM3M_UNROLL_N}.c" "USE_ALPHA;REAL_ONLY" "gemm3m_otcopyr" false "" "" false ${float_type}) + GenerateNamedObjects("generic/zgemm3m_tcopy_${${float_char}GEMM3M_UNROLL_N}.c" "USE_ALPHA;IMAGE_ONLY" "gemm3m_otcopyi" false "" "" false ${float_type}) + + GenerateNamedObjects("generic/zgemm3m_ncopy_${${float_char}GEMM3M_UNROLL_M}.c" "ICOPY" "gemm3m_incopyb" false "" "" false ${float_type}) + GenerateNamedObjects("generic/zgemm3m_ncopy_${${float_char}GEMM3M_UNROLL_M}.c" "ICOPY;REAL_ONLY" "gemm3m_incopyr" false "" "" false ${float_type}) + GenerateNamedObjects("generic/zgemm3m_ncopy_${${float_char}GEMM3M_UNROLL_M}.c" "ICOPY;IMAGE_ONLY" "gemm3m_incopyi" false "" "" false ${float_type}) + + GenerateNamedObjects("generic/zgemm3m_tcopy_${${float_char}GEMM3M_UNROLL_M}.c" "ICOPY" "gemm3m_itcopyb" false "" "" false ${float_type}) + GenerateNamedObjects("generic/zgemm3m_tcopy_${${float_char}GEMM3M_UNROLL_M}.c" "ICOPY;REAL_ONLY" "gemm3m_itcopyr" false "" "" false ${float_type}) + GenerateNamedObjects("generic/zgemm3m_tcopy_${${float_char}GEMM3M_UNROLL_M}.c" "ICOPY;IMAGE_ONLY" "gemm3m_itcopyi" false "" "" false ${float_type}) + +#hemm3m and symm3m + foreach(name symm3m hemm3m) + GenerateNamedObjects("generic/z${name}_ucopy_${${float_char}GEMM3M_UNROLL_M}.c" "USE_ALPHA" "${name}_oucopyb.c" false "" "" false ${float_type}) + GenerateNamedObjects("generic/z${name}_ucopy_${${float_char}GEMM3M_UNROLL_M}.c" "USE_ALPHA;REAL_ONLY" "${name}_oucopyr.c" false "" "" false ${float_type}) + GenerateNamedObjects("generic/z${name}_ucopy_${${float_char}GEMM3M_UNROLL_M}.c" "USE_ALPHA;IMAGE_ONLY" "${name}_oucopyi.c" false "" "" false ${float_type}) + GenerateNamedObjects("generic/z${name}_lcopy_${${float_char}GEMM3M_UNROLL_M}.c" "USE_ALPHA" "${name}_olcopyb.c" false "" "" false ${float_type}) + GenerateNamedObjects("generic/z${name}_lcopy_${${float_char}GEMM3M_UNROLL_M}.c" "USE_ALPHA;REAL_ONLY" "${name}_olcopyr.c" false "" "" false ${float_type}) + GenerateNamedObjects("generic/z${name}_lcopy_${${float_char}GEMM3M_UNROLL_M}.c" "USE_ALPHA;IMAGE_ONLY" "${name}_olcopyi.c" false "" "" false ${float_type}) + GenerateNamedObjects("generic/z${name}_ucopy_${${float_char}GEMM3M_UNROLL_M}.c" "" "${name}_iucopyb.c" false "" "" false ${float_type}) + GenerateNamedObjects("generic/z${name}_ucopy_${${float_char}GEMM3M_UNROLL_M}.c" "REAL_ONLY" "${name}_iucopyr.c" false "" "" false ${float_type}) + GenerateNamedObjects("generic/z${name}_ucopy_${${float_char}GEMM3M_UNROLL_M}.c" "IMAGE_ONLY" "${name}_iucopyi.c" false "" "" false ${float_type}) + GenerateNamedObjects("generic/z${name}_lcopy_${${float_char}GEMM3M_UNROLL_M}.c" "" "${name}_ilcopyb.c" false "" "" false ${float_type}) + GenerateNamedObjects("generic/z${name}_lcopy_${${float_char}GEMM3M_UNROLL_M}.c" "REAL_ONLY" "${name}_ilcopyr.c" false "" "" false ${float_type}) + GenerateNamedObjects("generic/z${name}_lcopy_${${float_char}GEMM3M_UNROLL_M}.c" "IMAGE_ONLY" "${name}_ilcopyi.c" false "" "" false ${float_type}) + + endforeach () + + endif() + + else () #For real + GenerateCombinationObjects("${KERNELDIR}/${TRMM_KERNEL}" "LEFT;TRANSA" "R;N" "TRMMKERNEL" 2 "trmm_kernel" false ${float_type}) + + # symm for s and d + GenerateNamedObjects("generic/symm_ucopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER" "symm_outcopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/symm_ucopy_${${float_char}GEMM_UNROLL_M}.c" "" "symm_iutcopy" false "" "" false ${float_type}) + + GenerateNamedObjects("generic/symm_lcopy_${${float_char}GEMM_UNROLL_N}.c" "LOWER;OUTER" "symm_oltcopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/symm_lcopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER" "symm_iltcopy" false "" "" false ${float_type}) + + # These don't use a scheme that is easy to iterate over - the filenames have part of the DEFINE codes in them, for UPPER/TRANS but not for UNIT/OUTER. Also TRANS is not passed in as a define. + # Could simplify it a bit by pairing up by -UUNIT/-DUNIT. + + GenerateNamedObjects("generic/trmm_uncopy_${${float_char}GEMM_UNROLL_M}.c" "UNIT" "trmm_iunucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trmm_uncopy_${${float_char}GEMM_UNROLL_M}.c" "" "trmm_iunncopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trmm_uncopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;UNIT" "trmm_ounucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trmm_uncopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER" "trmm_ounncopy" false "" "" false ${float_type}) + + GenerateNamedObjects("generic/trmm_lncopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER;UNIT" "trmm_ilnucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trmm_lncopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER" "trmm_ilnncopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trmm_lncopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER;UNIT" "trmm_olnucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trmm_lncopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER" "trmm_olnncopy" false "" "" false ${float_type}) + + GenerateNamedObjects("generic/trmm_utcopy_${${float_char}GEMM_UNROLL_M}.c" "UNIT" "trmm_iutucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trmm_utcopy_${${float_char}GEMM_UNROLL_M}.c" "" "trmm_iutncopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trmm_utcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;UNIT" "trmm_outucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trmm_utcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER" "trmm_outncopy" false "" "" false ${float_type}) + + GenerateNamedObjects("generic/trmm_ltcopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER;UNIT" "trmm_iltucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trmm_ltcopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER" "trmm_iltncopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trmm_ltcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER;UNIT" "trmm_oltucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trmm_ltcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER" "trmm_oltncopy" false "" "" false ${float_type}) + + GenerateNamedObjects("generic/trsm_uncopy_${${float_char}GEMM_UNROLL_M}.c" "UNIT" "trsm_iunucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trsm_uncopy_${${float_char}GEMM_UNROLL_M}.c" "" "trsm_iunncopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trsm_uncopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;UNIT" "trsm_ounucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trsm_uncopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER" "trsm_ounncopy" false "" "" false ${float_type}) + + GenerateNamedObjects("generic/trsm_lncopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER;UNIT" "trsm_ilnucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trsm_lncopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER" "trsm_ilnncopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trsm_lncopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER;UNIT" "trsm_olnucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trsm_lncopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER" "trsm_olnncopy" false "" "" false ${float_type}) + + GenerateNamedObjects("generic/trsm_utcopy_${${float_char}GEMM_UNROLL_M}.c" "UNIT" "trsm_iutucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trsm_utcopy_${${float_char}GEMM_UNROLL_M}.c" "" "trsm_iutncopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trsm_utcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;UNIT" "trsm_outucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trsm_utcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER" "trsm_outncopy" false "" "" false ${float_type}) + + GenerateNamedObjects("generic/trsm_ltcopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER;UNIT" "trsm_iltucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trsm_ltcopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER" "trsm_iltncopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trsm_ltcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER;UNIT" "trsm_oltucopy" false "" "" false ${float_type}) + GenerateNamedObjects("generic/trsm_ltcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER" "trsm_oltncopy" false "" "" false ${float_type}) + + endif () + + GenerateNamedObjects("${KERNELDIR}/${${float_char}TRSMKERNEL_LN}" "UPPER;LN;TRSMKERNEL" "trsm_kernel_LN" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}TRSMKERNEL_LT}" "LT;TRSMKERNEL" "trsm_kernel_LT" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}TRSMKERNEL_RN}" "UPPER;RN;TRSMKERNEL" "trsm_kernel_RN" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}TRSMKERNEL_RT}" "RT;TRSMKERNEL" "trsm_kernel_RT" false "" "" false ${float_type}) + + + + if (NOT DEFINED ${float_char}OMATCOPY_CN) + if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C") + set(${float_char}OMATCOPY_CN ../arm/zomatcopy_cn.c) + else () + set(${float_char}OMATCOPY_CN ../arm/omatcopy_cn.c) + endif () + endif () + if (NOT DEFINED ${float_char}OMATCOPY_RN) + if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C") + set(${float_char}OMATCOPY_RN ../arm/zomatcopy_rn.c) + else () + set(${float_char}OMATCOPY_RN ../arm/omatcopy_rn.c) + endif () + endif () + if (NOT DEFINED ${float_char}OMATCOPY_CT) + if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C") + set(${float_char}OMATCOPY_CT ../arm/zomatcopy_ct.c) + else () + set(${float_char}OMATCOPY_CT ../arm/omatcopy_ct.c) + endif () + endif () + if (NOT DEFINED ${float_char}OMATCOPY_RT) + if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C") + set(${float_char}OMATCOPY_RT ../arm/zomatcopy_rt.c) + else () + set(${float_char}OMATCOPY_RT ../arm/omatcopy_rt.c) + endif () + endif () + + GenerateNamedObjects("${KERNELDIR}/${${float_char}OMATCOPY_CN}" "" "omatcopy_k_cn" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}OMATCOPY_RN}" "ROWM" "omatcopy_k_rn" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}OMATCOPY_CT}" "" "omatcopy_k_ct" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}OMATCOPY_RT}" "ROWM" "omatcopy_k_rt" false "" "" false ${float_type}) + + if (NOT DEFINED ${float_char}OMATCOPY_CNC) + if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C") + set(${float_char}OMATCOPY_CNC ../arm/zomatcopy_cnc.c) + endif () + endif () + if (NOT DEFINED ${float_char}OMATCOPY_RNC) + if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C") + set(${float_char}OMATCOPY_RNC ../arm/zomatcopy_rnc.c) + endif () + endif () + if (NOT DEFINED ${float_char}OMATCOPY_CTC) + if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C") + set(${float_char}OMATCOPY_CTC ../arm/zomatcopy_ctc.c) + endif () + endif () + if (NOT DEFINED ${float_char}OMATCOPY_RTC) + if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C") + set(${float_char}OMATCOPY_RTC ../arm/zomatcopy_rtc.c) + endif () + endif () + + if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C") + GenerateNamedObjects("${KERNELDIR}/${${float_char}OMATCOPY_CNC}" "CONJ" "omatcopy_k_cnc" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}OMATCOPY_RNC}" "CONJ;ROWM" "omatcopy_k_rnc" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}OMATCOPY_CTC}" "CONJ" "omatcopy_k_ctc" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}OMATCOPY_RTC}" "CONJ;ROWM" "omatcopy_k_rtc" false "" "" false ${float_type}) + endif() + + #imatcopy + if (NOT DEFINED ${float_char}IMATCOPY_CN) + if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C") + set(${float_char}IMATCOPY_CN ../generic/zimatcopy_cn.c) + else () + set(${float_char}IMATCOPY_CN ../generic/imatcopy_cn.c) + endif () + endif () + + if (NOT DEFINED ${float_char}IMATCOPY_RN) + if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C") + set(${float_char}IMATCOPY_RN ../generic/zimatcopy_rn.c) + else () + set(${float_char}IMATCOPY_RN ../generic/imatcopy_rn.c) + endif () + endif () + + if (NOT DEFINED ${float_char}IMATCOPY_CT) + if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C") + set(${float_char}IMATCOPY_CT ../generic/zimatcopy_ct.c) + else () + set(${float_char}IMATCOPY_CT ../generic/imatcopy_ct.c) + endif () + endif () + + if (NOT DEFINED ${float_char}IMATCOPY_RT) + if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C") + set(${float_char}IMATCOPY_RT ../generic/zimatcopy_rt.c) + else () + set(${float_char}IMATCOPY_RT ../generic/imatcopy_rt.c) + endif () + endif () + + GenerateNamedObjects("${KERNELDIR}/${${float_char}IMATCOPY_CN}" "" "imatcopy_k_cn" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}IMATCOPY_RN}" "ROWM" "imatcopy_k_rn" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}IMATCOPY_CT}" "" "imatcopy_k_ct" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}IMATCOPY_RT}" "ROWM" "imatcopy_k_rt" false "" "" false ${float_type}) + + + if (NOT DEFINED ${float_char}IMATCOPY_CNC) + if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C") + set(${float_char}IMATCOPY_CNC ../generic/zimatcopy_cnc.c) + endif () + endif () + if (NOT DEFINED ${float_char}IMATCOPY_RNC) + if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C") + set(${float_char}IMATCOPY_RNC ../generic/zimatcopy_rnc.c) + endif () + endif () + if (NOT DEFINED ${float_char}IMATCOPY_CTC) + if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C") + set(${float_char}IMATCOPY_CTC ../generic/zimatcopy_ctc.c) + endif () + endif () + if (NOT DEFINED ${float_char}IMATCOPY_RTC) + if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C") + set(${float_char}IMATCOPY_RTC ../generic/zimatcopy_rtc.c) + endif () + endif () + + if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C") + GenerateNamedObjects("${KERNELDIR}/${${float_char}IMATCOPY_CNC}" "CONJ" "imatcopy_k_cnc" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}IMATCOPY_RNC}" "CONJ;ROWM" "imatcopy_k_rnc" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}IMATCOPY_CTC}" "CONJ" "imatcopy_k_ctc" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}IMATCOPY_RTC}" "CONJ;ROWM" "imatcopy_k_rtc" false "" "" false ${float_type}) + endif() + + #geadd + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEADD_KERNEL}" "" "geadd_k" false "" "" false ${float_type}) + endforeach () + + # Makefile.LA + if(NOT NO_LAPACK) + foreach (float_type ${FLOAT_TYPES}) + if (NOT DEFINED ${float_char}NEG_TCOPY) + if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C" OR ${float_char} STREQUAL "X") + set(${float_char}NEG_TCOPY ../generic/zneg_tcopy.c) + else () + set(${float_char}NEG_TCOPY ../generic/neg_tcopy.c) + endif () + endif () + + if (NOT DEFINED ${float_char}LASWP_NCOPY) + if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C" OR ${float_char} STREQUAL "X") + set(${float_char}LASWP_NCOPY ../generic/zlaswp_ncopy.c) + else () + set(${float_char}LASWP_NCOPY ../generic/laswp_ncopy.c) + endif () + endif () + string(SUBSTRING ${float_type} 0 1 float_char) + GenerateNamedObjects("${KERNELDIR}/${${float_char}NEG_TCOPY}_${${float_char}GEMM_UNROLL_M}" "" "neg_tcopy" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}LASWP_NCOPY}_${${float_char}GEMM_UNROLL_N}" "" "laswp_ncopy" false "" "" false ${float_type}) + endforeach() + endif() + + if (${DYNAMIC_ARCH}) + set(SETPARAM_TARGET_DIR ${CMAKE_CURRENT_BINARY_DIR}${CMAKE_FILES_DIRECTORY}) + file(READ ${CMAKE_CURRENT_SOURCE_DIR}/setparam-ref.c SETPARAM_REF_CONTENTS) + string(REPLACE "TS" "${TSUFFIX}" SETPARAM_REF_CONTENTS_NEW "${SETPARAM_REF_CONTENTS}") + file(WRITE ${SETPARAM_TARGET_DIR}/setparam${TSUFFIX}.tmp "${SETPARAM_REF_CONTENTS_NEW}") + configure_file(${SETPARAM_TARGET_DIR}/setparam${TSUFFIX}.tmp ${SETPARAM_TARGET_DIR}/setparam${TSUFFIX}.c COPYONLY) + set(OPENBLAS_SRC ${OPENBLAS_SRC} ${SETPARAM_TARGET_DIR}/setparam${TSUFFIX}.c) + file(REMOVE ${SETPARAM_TARGET_DIR}/setparam${TSUFFIX}.tmp) + + set(KERNEL_TSUFFIX_CONTENTS "") + foreach (KERNEL_INTERFACE_H ${KERNEL_INTERFACE}) + file(READ ${PROJECT_SOURCE_DIR}/${KERNEL_INTERFACE_H} KERNEL_INTERFACE_H_CONTENTS) + string(REGEX REPLACE " *\\(" "${TSUFFIX}(" KERNEL_INTERFACE_H_CONTENTS_NEW "${KERNEL_INTERFACE_H_CONTENTS}") + set(KERNEL_TSUFFIX_CONTENTS "${KERNEL_TSUFFIX_CONTENTS}\n${KERNEL_INTERFACE_H_CONTENTS_NEW}") + endforeach() + file(WRITE ${SETPARAM_TARGET_DIR}/kernel${TSUFFIX}.tmp "${KERNEL_TSUFFIX_CONTENTS}") + configure_file(${SETPARAM_TARGET_DIR}/kernel${TSUFFIX}.tmp ${SETPARAM_TARGET_DIR}/kernel${TSUFFIX}.h COPYONLY) + file(REMOVE ${SETPARAM_TARGET_DIR}/kernel${TSUFFIX}.tmp) + + foreach (float_type ${FLOAT_TYPES}) + # a bit of metaprogramming here to pull out the appropriate KERNEL var + string(SUBSTRING ${float_type} 0 1 float_char) + GenerateNamedObjects("generic/neg_tcopy_${${float_char}GEMM_UNROLL_M}.c" "" "neg_tcopy" false "" ${TSUFFIX} false ${float_type}) + GenerateNamedObjects("generic/laswp_ncopy_${${float_char}GEMM_UNROLL_N}.c" "" "laswp_ncopy" false "" ${TSUFFIX} false ${float_type}) + endforeach () + + + endif () + + add_library(kernel${TSUFFIX} OBJECT ${OPENBLAS_SRC}) + set_target_properties(kernel${TSUFFIX} PROPERTIES COMPILE_FLAGS "${KERNEL_DEFINITIONS}") + get_target_property(KERNEL_INCLUDE_DIRECTORIES kernel${TSUFFIX} INCLUDE_DIRECTORIES) + set_target_properties(kernel${TSUFFIX} PROPERTIES INCLUDE_DIRECTORIES "${KERNEL_INCLUDE_DIRECTORIES};${TARGET_CONF_DIR}") +endfunction () + + +set(ADD_COMMONOBJS 1) +if (${DYNAMIC_ARCH}) + foreach(TARGET_CORE ${DYNAMIC_CORE}) + set(BUILD_KERNEL 1) + set(KDIR "") + set(TSUFFIX "_${TARGET_CORE}") + set(KERNEL_DEFINITIONS "-DBUILD_KERNEL -DTABLE_NAME=gotoblas_${TARGET_CORE} -DTS=${TSUFFIX}") + build_core("${TARGET_CORE}" "${KDIR}" "${TSUFFIX}" "${KERNEL_DEFINITIONS}") + set(ADD_COMMONOBJS 0) + endforeach() +else () + set(TARGET_CONF_DIR ${PROJECT_BINARY_DIR}) + set(TARGET_CORE ${CORE}) + set(KDIR "") + set(TSUFFIX "") + set(KERNEL_DEFINITIONS "") + build_core("${TARGET_CORE}" "${KDIR}" "${TSUFFIX}" "${KERNEL_DEFINITIONS}") endif () -foreach (float_type ${FLOAT_TYPES}) - if (${float_type} STREQUAL "COMPLEX" OR ${float_type} STREQUAL "ZCOMPLEX") - foreach (trans_src ${TRANS_SOURCES}) - string(REGEX MATCH "[a-z]/([a-z]+_)([a-z]+)" op_name ${trans_src}) - string(REPLACE "/" "/z" ztrans_src ${trans_src}) - GenerateNamedObjects("${ztrans_src}" "TRANS=1" "${CMAKE_MATCH_1}N_${CMAKE_MATCH_2}" false "" "" false ${float_type}) - GenerateNamedObjects("${ztrans_src}" "TRANS=2" "${CMAKE_MATCH_1}T_${CMAKE_MATCH_2}" false "" "" false ${float_type}) - GenerateNamedObjects("${ztrans_src}" "TRANS=3" "${CMAKE_MATCH_1}R_${CMAKE_MATCH_2}" false "" "" false ${float_type}) - GenerateNamedObjects("${ztrans_src}" "TRANS=4" "${CMAKE_MATCH_1}C_${CMAKE_MATCH_2}" false "" "" false ${float_type}) - endforeach () - else () - GenerateCombinationObjects("${TRANS_SOURCES}" "TRANS" "N" "" 4 "" false ${float_type}) - endif () -endforeach () - -GenerateCombinationObjects("${UNIT_SOURCES}" "UNIT" "N" "" 4) -GenerateCombinationObjects("${UNIT_SOURCES2}" "UNIT" "N" "" 0 "" "" 3) - -add_library(lapack OBJECT ${OPENBLAS_SRC}) -