diff options
Diffstat (limited to 'llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt')
| -rw-r--r-- | llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt | 689 |
1 files changed, 689 insertions, 0 deletions
diff --git a/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt b/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt new file mode 100644 index 0000000..7622d0b --- /dev/null +++ b/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt @@ -0,0 +1,689 @@ +function(ggml_add_cpu_backend_features cpu_name arch) + # The feature detection code is compiled as a separate target so that + # it can be built without the architecture flags + # Since multiple variants of the CPU backend may be included in the same + # build, using set_source_files_properties() to set the arch flags is not possible + set(GGML_CPU_FEATS_NAME ${cpu_name}-feats) + add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/arch/${arch}/cpu-feats.cpp) + target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . ../include) + target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARGN}) + target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED) + set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON) + target_link_libraries(${cpu_name} PRIVATE ${GGML_CPU_FEATS_NAME}) +endfunction() + +function(ggml_add_cpu_backend_variant_impl tag_name) + if (tag_name) + set(GGML_CPU_NAME ggml-cpu-${tag_name}) + else() + set(GGML_CPU_NAME ggml-cpu) + endif() + + ggml_add_backend_library(${GGML_CPU_NAME}) + + list (APPEND GGML_CPU_SOURCES + ggml-cpu/ggml-cpu.c + ggml-cpu/ggml-cpu.cpp + ggml-cpu/repack.cpp + ggml-cpu/repack.h + ggml-cpu/hbm.cpp + ggml-cpu/hbm.h + ggml-cpu/quants.c + ggml-cpu/quants.h + ggml-cpu/traits.cpp + ggml-cpu/traits.h + ggml-cpu/amx/amx.cpp + ggml-cpu/amx/amx.h + ggml-cpu/amx/mmq.cpp + ggml-cpu/amx/mmq.h + ggml-cpu/ggml-cpu-impl.h + ggml-cpu/common.h + ggml-cpu/binary-ops.h + ggml-cpu/binary-ops.cpp + ggml-cpu/unary-ops.h + ggml-cpu/unary-ops.cpp + ggml-cpu/simd-mappings.h + ggml-cpu/vec.h + ggml-cpu/vec.cpp + ggml-cpu/ops.h + ggml-cpu/ops.cpp + ) + + target_compile_features(${GGML_CPU_NAME} PRIVATE c_std_11 cxx_std_17) + target_include_directories(${GGML_CPU_NAME} PRIVATE . ggml-cpu) + + if (APPLE AND GGML_ACCELERATE) + find_library(ACCELERATE_FRAMEWORK Accelerate) + if (ACCELERATE_FRAMEWORK) + message(STATUS "Accelerate framework found") + + target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_ACCELERATE) + target_compile_definitions(${GGML_CPU_NAME} PRIVATE ACCELERATE_NEW_LAPACK) + target_compile_definitions(${GGML_CPU_NAME} PRIVATE ACCELERATE_LAPACK_ILP64) + + target_link_libraries(${GGML_CPU_NAME} PRIVATE ${ACCELERATE_FRAMEWORK}) + else() + message(WARNING "Accelerate framework not found") + endif() + endif() + + if (GGML_OPENMP) + find_package(OpenMP) + if (OpenMP_FOUND) + set(GGML_OPENMP_ENABLED "ON" CACHE INTERNAL "") + target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_OPENMP) + + target_link_libraries(${GGML_CPU_NAME} PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX) + else() + set(GGML_OPENMP_ENABLED "OFF" CACHE INTERNAL "") + message(WARNING "OpenMP not found") + endif() + endif() + + if (GGML_LLAMAFILE) + target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_LLAMAFILE) + + list(APPEND GGML_CPU_SOURCES + ggml-cpu/llamafile/sgemm.cpp + ggml-cpu/llamafile/sgemm.h) + endif() + + if (GGML_CPU_HBM) + find_library(memkind memkind REQUIRED) + + message(STATUS "Using memkind for CPU HBM") + + target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_HBM) + + target_link_libraries(${GGML_CPU_NAME} PUBLIC memkind) + endif() + + if (GGML_SYSTEM_ARCH STREQUAL "ARM") + message(STATUS "ARM detected") + list(APPEND GGML_CPU_SOURCES + ggml-cpu/arch/arm/quants.c + ggml-cpu/arch/arm/repack.cpp + ) + + if (MSVC AND NOT CMAKE_C_COMPILER_ID STREQUAL "Clang") + message(FATAL_ERROR "MSVC is not supported for ARM, use clang") + else() + check_cxx_compiler_flag(-mfp16-format=ieee GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E) + if (NOT "${GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "") + list(APPEND ARCH_FLAGS -mfp16-format=ieee) + endif() + + if (GGML_NATIVE) + # -mcpu=native does not always enable all the features in some compilers, + # so we check for them manually and enable them if available + + execute_process( + COMMAND ${CMAKE_C_COMPILER} -mcpu=native -E -v - + INPUT_FILE "/dev/null" + OUTPUT_QUIET + ERROR_VARIABLE ARM_MCPU + RESULT_VARIABLE ARM_MCPU_RESULT + ) + if (NOT ARM_MCPU_RESULT) + string(REGEX MATCH "-mcpu=[^ ']+" ARM_MCPU_FLAG "${ARM_MCPU}") + string(REGEX MATCH "-march=[^ ']+" ARM_MARCH_FLAG "${ARM_MCPU}") + + # on some old GCC we need to read -march= + if (ARM_MARCH_FLAG AND NOT "${ARM_MARCH_FLAG}" STREQUAL "-march=native") + set(ARM_NATIVE_FLAG "${ARM_MARCH_FLAG}") + elseif(ARM_MCPU_FLAG AND NOT "${ARM_MCPU_FLAG}" STREQUAL "-mcpu=native") + set(ARM_NATIVE_FLAG "${ARM_MCPU_FLAG}") + endif() + endif() + + if ("${ARM_NATIVE_FLAG}" STREQUAL "") + set(ARM_NATIVE_FLAG -mcpu=native) + message(WARNING "ARM -march/-mcpu not found, -mcpu=native will be used") + else() + message(STATUS "ARM detected flags: ${ARM_NATIVE_FLAG}") + endif() + + include(CheckCXXSourceRuns) + + macro(check_arm_feature tag feature code) + set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS}) + set(CMAKE_REQUIRED_FLAGS "${ARM_NATIVE_FLAG}+${tag}") + check_cxx_source_runs("${code}" GGML_MACHINE_SUPPORTS_${tag}) + if (GGML_MACHINE_SUPPORTS_${tag}) + set(ARM_NATIVE_FLAG_FIX "${ARM_NATIVE_FLAG_FIX}+${tag}") + else() + set(CMAKE_REQUIRED_FLAGS "${ARM_NATIVE_FLAG}+no${tag}") + check_cxx_source_compiles("int main() { return 0; }" GGML_MACHINE_SUPPORTS_no${tag}) + if (GGML_MACHINE_SUPPORTS_no${tag}) + set(ARM_NATIVE_FLAG_FIX "${ARM_NATIVE_FLAG_FIX}+no${tag}") + list(APPEND ARCH_FLAGS -U__ARM_FEATURE_${feature}) + endif() + endif() + set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE}) + endmacro() + + check_arm_feature(dotprod DOTPROD "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }") + check_arm_feature(i8mm MATMUL_INT8 "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }") + check_arm_feature(sve SVE "#include <arm_sve.h>\nint main() { svfloat32_t _a, _b; volatile svfloat32_t _c = svadd_f32_z(svptrue_b8(), _a, _b); return 0; }") + check_arm_feature(sme SME "#include <arm_sme.h>\n__arm_locally_streaming int main() { __asm__ volatile(\"smstart; smstop;\"); return 0; }") + + list(APPEND ARCH_FLAGS "${ARM_NATIVE_FLAG}${ARM_NATIVE_FLAG_FIX}") + else() + if (GGML_CPU_ARM_ARCH) + list(APPEND ARCH_FLAGS -march=${GGML_CPU_ARM_ARCH}) + elseif(GGML_CPU_ALL_VARIANTS) + # Begin with the lowest baseline + set(ARM_MCPU "armv8-a") + set(ARCH_TAGS "") + set(ARCH_DEFINITIONS "") + + # When a feature is selected, bump the MCPU to the first + # version that supported it + if (GGML_INTERNAL_DOTPROD) + set(ARM_MCPU "armv8.2-a") + set(ARCH_TAGS "${ARCH_TAGS}+dotprod") + list(APPEND ARCH_DEFINITIONS GGML_USE_DOTPROD) + endif() + if (GGML_INTERNAL_FP16_VECTOR_ARITHMETIC) + set(ARM_MCPU "armv8.2-a") + set(ARCH_TAGS "${ARCH_TAGS}+fp16") + list(APPEND ARCH_DEFINITIONS GGML_USE_FP16_VECTOR_ARITHMETIC) + endif() + if (GGML_INTERNAL_SVE) + set(ARM_MCPU "armv8.2-a") + set(ARCH_TAGS "${ARCH_TAGS}+sve") + list(APPEND ARCH_DEFINITIONS GGML_USE_SVE) + endif() + if (GGML_INTERNAL_MATMUL_INT8) + set(ARM_MCPU "armv8.6-a") + set(ARCH_TAGS "${ARCH_TAGS}+i8mm") + list(APPEND ARCH_DEFINITIONS GGML_USE_MATMUL_INT8) + endif() + if (GGML_INTERNAL_SVE2) + set(ARM_MCPU "armv8.6-a") + set(ARCH_TAGS "${ARCH_TAGS}+sve2") + list(APPEND ARCH_DEFINITIONS GGML_USE_SVE2) + endif() + if (GGML_INTERNAL_NOSVE) + set(ARCH_TAGS "${ARCH_TAGS}+nosve") + endif() + if (GGML_INTERNAL_SME) + set(ARM_MCPU "armv9.2-a") + set(ARCH_TAGS "${ARCH_TAGS}+sme") + list(APPEND ARCH_DEFINITIONS GGML_USE_SME) + endif() + list(APPEND ARCH_FLAGS "-march=${ARM_MCPU}${ARCH_TAGS}") + ggml_add_cpu_backend_features(${GGML_CPU_NAME} arm ${ARCH_DEFINITIONS}) + endif() + endif() + + message(STATUS "Checking for ARM features using flags:") + foreach(flag IN LISTS ARCH_FLAGS) + message(STATUS " ${flag}") + endforeach() + + include(CheckCXXSourceCompiles) + set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS}) + string(REPLACE ";" " " ARCH_FLAGS_STR "${ARCH_FLAGS}") + set(CMAKE_REQUIRED_FLAGS "${ARCH_FLAGS_STR}") + foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC SME) + set(ARM_FEATURE "HAVE_${feature}") + check_cxx_source_compiles( + " + #if !defined(__ARM_FEATURE_${feature}) + # error \"Feature ${feature} is not defined\" + #endif + int main() { return 0; } + " + ${ARM_FEATURE} + ) + endforeach() + set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE}) + endif() + elseif (GGML_SYSTEM_ARCH STREQUAL "x86") + message(STATUS "x86 detected") + list(APPEND GGML_CPU_SOURCES + ggml-cpu/arch/x86/quants.c + ggml-cpu/arch/x86/repack.cpp + ) + + if (MSVC) + # instruction set detection for MSVC only + if (GGML_NATIVE) + include(ggml-cpu/cmake/FindSIMD.cmake) + endif () + if (GGML_AVX512) + list(APPEND ARCH_FLAGS /arch:AVX512) + # /arch:AVX512 includes: __AVX512F__, __AVX512CD__, __AVX512BW__, __AVX512DQ__, and __AVX512VL__ + # MSVC has no compile-time flags enabling specific + # AVX512 extensions, neither it defines the + # macros corresponding to the extensions. + # Do it manually. + list(APPEND ARCH_DEFINITIONS GGML_AVX512) + if (GGML_AVX512_VBMI) + list(APPEND ARCH_DEFINITIONS __AVX512VBMI__) + if (CMAKE_C_COMPILER_ID STREQUAL "Clang") + list(APPEND ARCH_FLAGS -mavx512vbmi) + endif() + endif() + if (GGML_AVX512_VNNI) + list(APPEND ARCH_DEFINITIONS __AVX512VNNI__ GGML_AVX512_VNNI) + if (CMAKE_C_COMPILER_ID STREQUAL "Clang") + list(APPEND ARCH_FLAGS -mavx512vnni) + endif() + endif() + if (GGML_AVX512_BF16) + list(APPEND ARCH_DEFINITIONS __AVX512BF16__ GGML_AVX512_BF16) + if (CMAKE_C_COMPILER_ID STREQUAL "Clang") + list(APPEND ARCH_FLAGS -mavx512bf16) + endif() + endif() + if (GGML_AMX_TILE) + list(APPEND ARCH_DEFINITIONS __AMX_TILE__ GGML_AMX_TILE) + endif() + if (GGML_AMX_INT8) + list(APPEND ARCH_DEFINITIONS __AMX_INT8__ GGML_AMX_INT8) + endif() + if (GGML_AMX_BF16) + list(APPEND ARCH_DEFINITIONS __AMX_BF16__ GGML_AMX_BF16) + endif() + elseif (GGML_AVX2) + list(APPEND ARCH_FLAGS /arch:AVX2) + list(APPEND ARCH_DEFINITIONS GGML_AVX2 GGML_FMA GGML_F16C) + elseif (GGML_AVX) + list(APPEND ARCH_FLAGS /arch:AVX) + list(APPEND ARCH_DEFINITIONS GGML_AVX) + elseif (GGML_SSE42) + list(APPEND ARCH_FLAGS /arch:SSE4.2) + list(APPEND ARCH_DEFINITIONS GGML_SSE42) + endif() + if (GGML_AVX_VNNI) + list(APPEND ARCH_DEFINITIONS __AVXVNNI__ GGML_AVX_VNNI) + endif() + if (GGML_BMI2) + # MSVC does not define macro __BMI2__ + list(APPEND ARCH_DEFINITIONS __BMI2__ GGML_BMI2) + endif() + else () + if (GGML_NATIVE) + list(APPEND ARCH_FLAGS -march=native) + else () + if (GGML_SSE42) + list(APPEND ARCH_FLAGS -msse4.2) + list(APPEND ARCH_DEFINITIONS GGML_SSE42) + endif() + if (GGML_F16C) + list(APPEND ARCH_FLAGS -mf16c) + list(APPEND ARCH_DEFINITIONS GGML_F16C) + endif() + if (GGML_FMA) + list(APPEND ARCH_FLAGS -mfma) + list(APPEND ARCH_DEFINITIONS GGML_FMA) + endif() + if (GGML_BMI2) + list(APPEND ARCH_FLAGS -mbmi2) + list(APPEND ARCH_DEFINITIONS GGML_BMI2) + endif() + if (GGML_AVX) + list(APPEND ARCH_FLAGS -mavx) + list(APPEND ARCH_DEFINITIONS GGML_AVX) + endif() + if (GGML_AVX2) + list(APPEND ARCH_FLAGS -mavx2) + list(APPEND ARCH_DEFINITIONS GGML_AVX2) + endif() + if (GGML_AVX_VNNI) + list(APPEND ARCH_FLAGS -mavxvnni) + list(APPEND ARCH_DEFINITIONS GGML_AVX_VNNI) + endif() + if (GGML_AVX512) + list(APPEND ARCH_FLAGS -mavx512f) + list(APPEND ARCH_FLAGS -mavx512cd) + list(APPEND ARCH_FLAGS -mavx512vl) + list(APPEND ARCH_FLAGS -mavx512dq) + list(APPEND ARCH_FLAGS -mavx512bw) + list(APPEND ARCH_DEFINITIONS GGML_AVX512) + endif() + if (GGML_AVX512_VBMI) + list(APPEND ARCH_FLAGS -mavx512vbmi) + list(APPEND ARCH_DEFINITIONS GGML_AVX512_VBMI) + endif() + if (GGML_AVX512_VNNI) + list(APPEND ARCH_FLAGS -mavx512vnni) + list(APPEND ARCH_DEFINITIONS GGML_AVX512_VNNI) + endif() + if (GGML_AVX512_BF16) + list(APPEND ARCH_FLAGS -mavx512bf16) + list(APPEND ARCH_DEFINITIONS GGML_AVX512_BF16) + endif() + if (GGML_AMX_TILE) + list(APPEND ARCH_FLAGS -mamx-tile) + list(APPEND ARCH_DEFINITIONS GGML_AMX_TILE) + endif() + if (GGML_AMX_INT8) + list(APPEND ARCH_FLAGS -mamx-int8) + list(APPEND ARCH_DEFINITIONS GGML_AMX_INT8) + endif() + if (GGML_AMX_BF16) + list(APPEND ARCH_FLAGS -mamx-bf16) + list(APPEND ARCH_DEFINITIONS GGML_AMX_BF16) + endif() + endif() + endif() + + if (GGML_BACKEND_DL) + if (GGML_NATIVE) + # the feature check relies on ARCH_DEFINITIONS, but it is not set with GGML_NATIVE + message(FATAL_ERROR "GGML_NATIVE is not compatible with GGML_BACKEND_DL, consider using GGML_CPU_ALL_VARIANTS") + endif() + ggml_add_cpu_backend_features(${GGML_CPU_NAME} x86 ${ARCH_DEFINITIONS}) + endif() + elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC") + message(STATUS "PowerPC detected") + list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/powerpc/quants.c) + if (GGML_NATIVE) + if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64") + file(READ "/proc/cpuinfo" POWER10_M) + elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "powerpc") + execute_process(COMMAND bash -c "prtconf |grep 'Implementation' | head -n 1" OUTPUT_VARIABLE POWER10_M) + endif() + + string(TOUPPER "${POWER10_M}" POWER10_M_UPPER) + string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M_UPPER}") + string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}") + + if (EXTRACTED_NUMBER GREATER_EQUAL 10) + list(APPEND ARCH_FLAGS -mcpu=power10) + elseif (EXTRACTED_NUMBER EQUAL 9) + list(APPEND ARCH_FLAGS -mcpu=power9) + elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le") + list(APPEND ARCH_FLAGS -mcpu=powerpc64le -mtune=native) + else() + list(APPEND ARCH_FLAGS -mcpu=native -mtune=native -mpowerpc64) + endif() + elseif(GGML_CPU_ALL_VARIANTS) + # Begin with the lowest baseline + set(ARCH_DEFINITIONS "") + + # When a feature is selected, bump the MCPU to the first + # version that supported it + foreach(PVER RANGE 7 11) + if(DEFINED GGML_INTERNAL_POWER${PVER}) + set(POWERPC_MCPU "power${PVER}") + list(APPEND ARCH_DEFINITIONS GGML_USE_POWER${PVER}) + endif() + endforeach() + if (GGML_INTERNAL_VSX) + list(APPEND ARCH_DEFINITIONS GGML_USE_VSX) + list(APPEND ARCH_FLAGS -mvsx) + endif() + + if (DEFINED POWERPC_MCPU) + list(APPEND ARCH_FLAGS -mcpu=${POWERPC_MCPU}) + endif() + ggml_add_cpu_backend_features(${GGML_CPU_NAME} powerpc ${ARCH_DEFINITIONS}) + else() + if (GGML_CPU_POWERPC_CPUTYPE) + list(APPEND ARCH_FLAGS -mcpu=${GGML_CPU_POWERPC_CPUTYPE}) + endif() + endif() + elseif (GGML_SYSTEM_ARCH STREQUAL "loongarch64") + message(STATUS "loongarch64 detected") + list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/loongarch/quants.c) + + list(APPEND ARCH_FLAGS -march=loongarch64) + if (GGML_LASX) + list(APPEND ARCH_FLAGS -mlasx) + endif() + if (GGML_LSX) + list(APPEND ARCH_FLAGS -mlsx) + endif() + elseif (GGML_SYSTEM_ARCH STREQUAL "riscv64") + message(STATUS "riscv64 detected") + list(APPEND GGML_CPU_SOURCES + ggml-cpu/arch/riscv/quants.c + ggml-cpu/arch/riscv/repack.cpp + ) + if (GGML_CPU_RISCV64_SPACEMIT) + target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_RISCV64_SPACEMIT ${RISCV64_SPACEMIT_IME_SPEC}) + list(APPEND GGML_CPU_SOURCES + ggml-cpu/spacemit/ime.cpp + ggml-cpu/spacemit/ime.h + ggml-cpu/spacemit/ime1_kernels.cpp + ggml-cpu/spacemit/ime_kernels.h + ) + endif() + if(NOT GGML_CPU_ALL_VARIANTS) + set(MARCH_STR "rv64gc") + if (GGML_RV_ZFH) + string(APPEND MARCH_STR "_zfh") + endif() + + if (GGML_XTHEADVECTOR) + string(APPEND MARCH_STR "_xtheadvector") + elseif (GGML_RVV) + string(APPEND MARCH_STR "_v") + if (GGML_RV_ZVFH) + string(APPEND MARCH_STR "_zvfh") + endif() + if (GGML_RV_ZVFBFWMA) + string(APPEND MARCH_STR "_zvfbfwma") + endif() + endif() + if (GGML_RV_ZICBOP) + string(APPEND MARCH_STR "_zicbop") + endif() + if (GGML_RV_ZIHINTPAUSE) + string(APPEND MARCH_STR "_zihintpause") + endif() + list(APPEND ARCH_FLAGS "-march=${MARCH_STR}" -mabi=lp64d) + else() + # Begin with the lowest baseline + set(ARCH_DEFINITIONS "") + + if (GGML_INTERNAL_RVV) + message(STATUS "RVV enabled") + list(APPEND ARCH_DEFINITIONS GGML_USE_RVV) + list(APPEND ARCH_FLAGS -march=rv64gc_v -mabi=lp64d) + endif() + + ggml_add_cpu_backend_features(${GGML_CPU_NAME} riscv ${ARCH_DEFINITIONS}) + endif() + elseif (GGML_SYSTEM_ARCH STREQUAL "s390x") + message(STATUS "s390x detected") + list(APPEND GGML_CPU_SOURCES + ggml-cpu/arch/s390/quants.c) + + # for native compilation + if (GGML_NATIVE) + # check machine level to determine target + file(READ "/proc/cpuinfo" CPUINFO_CONTENTS) + string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS}) + + # TODO: Separation to determine activation of VX/VXE/VXE2 + if (${S390X_M} MATCHES "8561|8562") + message(STATUS "z15 target") + list(APPEND ARCH_FLAGS -march=z15) + elseif (${S390X_M} MATCHES "3931") + message(STATUS "z16 target") + list(APPEND ARCH_FLAGS -march=z16) + elseif (${S390X_M} MATCHES "9175|9176") + # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version. + # binutils must also be updated to the latest for the -march=z17 flag to work. Otherwise, use -march=arch15. + message(STATUS "z17 target") + list(APPEND ARCH_FLAGS -march=arch15) + else() + message(STATUS "Unknown target") + message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.") + list(APPEND ARCH_FLAGS -march=native -mtune=native) + endif() + # for cross-compilation + elseif(GGML_CPU_ALL_VARIANTS) + # range through IBM z15 to z17 + # NOTE: update when a new hardware level is released + foreach (ZHW RANGE 15 17) + if(DEFINED GGML_INTERNAL_Z${ZHW}) + message(STATUS "z${ZHW} cross-compile target") + list(APPEND ARCH_FLAGS -march=z${ZHW}) + endif() + endforeach() + endif() + + if (GGML_VXE OR GGML_INTERNAL_VXE2) + message(STATUS "VXE2 enabled") + list(APPEND ARCH_FLAGS -mvx -mzvector) + list(APPEND ARCH_DEFINITIONS GGML_USE_VXE2) + endif() + + if (GGML_INTERNAL_NNPA) + message(STATUS "NNPA enabled") + list(APPEND ARCH_DEFINITIONS GGML_USE_NNPA) + endif() + + ggml_add_cpu_backend_features(${GGML_CPU_NAME} s390 ${ARCH_DEFINITIONS}) + elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm") + message(STATUS "Wasm detected") + list (APPEND GGML_CPU_SOURCES ggml-cpu/arch/wasm/quants.c) + else() + message(WARNING "Unknown CPU architecture. Falling back to generic implementations.") + list(APPEND ARCH_FLAGS -DGGML_CPU_GENERIC) + endif() + + if (GGML_CPU_REPACK) + target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_REPACK) + endif() + + if (GGML_CPU_KLEIDIAI) + message(STATUS "Using KleidiAI optimized kernels if applicable") + + # Disable the KleidiAI tests + set(KLEIDIAI_BUILD_TESTS OFF) + + # Fetch KleidiAI sources: + include(FetchContent) + set(KLEIDIAI_COMMIT_TAG "v1.16.0") + set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz") + set(KLEIDIAI_ARCHIVE_MD5 "0a9e9008adb6031f9e8cf70dff4a3321") + + if (POLICY CMP0135) + cmake_policy(SET CMP0135 NEW) + endif() + + FetchContent_Declare(KleidiAI_Download + URL ${KLEIDIAI_DOWNLOAD_URL} + DOWNLOAD_EXTRACT_TIMESTAMP NEW + URL_HASH MD5=${KLEIDIAI_ARCHIVE_MD5}) + + FetchContent_MakeAvailable(KleidiAI_Download) + FetchContent_GetProperties(KleidiAI_Download + SOURCE_DIR KLEIDIAI_SRC + POPULATED KLEIDIAI_POPULATED) + + if (NOT KLEIDIAI_POPULATED) + message(FATAL_ERROR "KleidiAI source downloaded failed.") + endif() + + add_compile_definitions(GGML_USE_CPU_KLEIDIAI) + + # Remove kleidiai target after fetching it + if (TARGET kleidiai) + set_target_properties(kleidiai PROPERTIES EXCLUDE_FROM_ALL TRUE) + endif() + + list(APPEND GGML_CPU_SOURCES + ggml-cpu/kleidiai/kleidiai.cpp + ggml-cpu/kleidiai/kernels.cpp + ggml-cpu/kleidiai/kleidiai.h + ggml-cpu/kleidiai/kernels.h + ) + + # KleidiAI + include_directories( + ${KLEIDIAI_SRC}/ + ${KLEIDIAI_SRC}/kai/ + ${KLEIDIAI_SRC}/kai/ukernels/ + ${KLEIDIAI_SRC}/kai/ukernels/matmul/ + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/ + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/ + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/ + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/) + + set(ARCH_FLAGS_TEMP "${ARCH_FLAGS}") + if (NOT ARCH_FLAGS_TEMP) + string(REGEX MATCH "-march=[^ ]+" ARCH_FLAGS_TEMP "${CMAKE_C_FLAGS}") + endif() + string(FIND "${ARCH_FLAGS_TEMP}" "+dotprod" DOTPROD_ENABLED) + string(FIND "${ARCH_FLAGS_TEMP}" "+i8mm" I8MM_ENABLED) + string(FIND "${ARCH_FLAGS_TEMP}" "+sme" SME_ENABLED) + string(FIND "${ARCH_FLAGS_TEMP}" "+sve" SVE_ENABLED) + + set(PRIVATE_ARCH_FLAGS ${ARCH_FLAGS_TEMP}) + + list(APPEND GGML_KLEIDIAI_SOURCES + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.c) + + if (NOT DOTPROD_ENABLED MATCHES -1) + list(APPEND GGML_KLEIDIAI_SOURCES + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c) + endif() + + if (NOT I8MM_ENABLED MATCHES -1) + list(APPEND GGML_KLEIDIAI_SOURCES + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c) + endif() + + if (NOT SME_ENABLED MATCHES -1) + list(APPEND GGML_KLEIDIAI_SOURCES + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa_asm.S + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot_asm.S + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa_asm.S + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c + ${KLEIDIAI_SRC}/kai/kai_common_sme_asm.S) + set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}+sve+sve2") + endif() + + if (NOT SVE_ENABLED MATCHES -1) + list(APPEND GGML_KLEIDIAI_SOURCES + ${KLEIDIAI_SRC}/kai/kai_common_sve_asm.S + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod_asm.S + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm_asm.S + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm.c) + endif() + + set_source_files_properties(${GGML_KLEIDIAI_SOURCES} PROPERTIES COMPILE_OPTIONS "${PRIVATE_ARCH_FLAGS}") + list(APPEND GGML_CPU_SOURCES ${GGML_KLEIDIAI_SOURCES}) + endif() + + message(STATUS "Adding CPU backend variant ${GGML_CPU_NAME}: ${ARCH_FLAGS} ${ARCH_DEFINITIONS}") + target_sources(${GGML_CPU_NAME} PRIVATE ${GGML_CPU_SOURCES}) + target_compile_options(${GGML_CPU_NAME} PRIVATE ${ARCH_FLAGS}) + target_compile_definitions(${GGML_CPU_NAME} PRIVATE ${ARCH_DEFINITIONS}) + + if (EMSCRIPTEN) + set_target_properties(${GGML_CPU_NAME} PROPERTIES COMPILE_FLAGS "-msimd128") + endif() + + if (CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM") + # The compiler automatically enables "-ffast-math" which can cause NaNs in tests due to "-fassociative-math" + target_compile_options(${GGML_CPU_NAME} PRIVATE "-fno-associative-math") + endif() +endfunction() |