michael-treadgold/win-dictate
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 1 Wiki Activity

Initial commit: Win Dictation - AI Voice to Text for Windows with automatic model selection

Browse Source
This commit is contained in:
Treadgold
2025-12-06 10:28:46 +13:00
commit 81b3d0073e
1666 changed files with 388157 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cpu/CMakeLists.txt
+600
View File
@@ -0,0 +1,600 @@
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}")
endif()
if ("${ARM_MCPU_FLAG}" STREQUAL "")
set(ARM_MCPU_FLAG -mcpu=native)
message(STATUS "ARM -mcpu not found, -mcpu=native will be used")
endif()
include(CheckCXXSourceRuns)
function(check_arm_feature tag code)
set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
set(CMAKE_REQUIRED_FLAGS "${ARM_MCPU_FLAG}+${tag}")
check_cxx_source_runs("${code}" GGML_MACHINE_SUPPORTS_${tag})
if (GGML_MACHINE_SUPPORTS_${tag})
set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+${tag}" PARENT_SCOPE)
else()
set(CMAKE_REQUIRED_FLAGS "${ARM_MCPU_FLAG}+no${tag}")
check_cxx_source_compiles("int main() { return 0; }" GGML_MACHINE_SUPPORTS_no${tag})
if (GGML_MACHINE_SUPPORTS_no${tag})
set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+no${tag}" PARENT_SCOPE)
endif()
endif()
set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
endfunction()
check_arm_feature(dotprod "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }")
check_arm_feature(i8mm "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }")
check_arm_feature(sve "#include <arm_sve.h>\nint main() { svfloat32_t _a, _b; volatile svfloat32_t _c = svadd_f32_z(svptrue_b8(), _a, _b); return 0; }")
check_arm_feature(sme "#include <arm_sme.h>\n__arm_locally_streaming int main() { __asm__ volatile(\"smstart; smstop;\"); return 0; }")
list(APPEND ARCH_FLAGS "${ARM_MCPU_FLAG}${ARM_MCPU_FLAG_FIX}")
else()
if (GGML_CPU_ARM_ARCH)
list(APPEND ARCH_FLAGS -march=${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()
# show enabled features
if (CMAKE_HOST_SYSTEM_NAME STREQUAL "Windows")
set(FEAT_INPUT_FILE "NUL")
else()
set(FEAT_INPUT_FILE "/dev/null")
endif()
execute_process(
COMMAND ${CMAKE_C_COMPILER} ${ARCH_FLAGS} -dM -E -
INPUT_FILE ${FEAT_INPUT_FILE}
OUTPUT_VARIABLE ARM_FEATURE
RESULT_VARIABLE ARM_FEATURE_RESULT
)
if (ARM_FEATURE_RESULT)
message(WARNING "Failed to get ARM features")
else()
foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC SME)
string(FIND "${ARM_FEATURE}" "__ARM_FEATURE_${feature} 1" feature_pos)
if (NOT ${feature_pos} EQUAL -1)
message(STATUS "ARM feature ${feature} enabled")
endif()
endforeach()
endif()
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 -mpowerpc64)
elseif (EXTRACTED_NUMBER EQUAL 9)
list(APPEND ARCH_FLAGS -mcpu=power9 -mpowerpc64)
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_RVV)
if (GGML_XTHEADVECTOR)
list(APPEND ARCH_FLAGS -march=rv64gc_xtheadvector -mabi=lp64d)
elseif (GGML_RV_ZFH)
list(APPEND ARCH_FLAGS -march=rv64gcv_zfhmin -mabi=lp64d)
else()
list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d)
endif()
endif()
elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
message(STATUS "s390x detected")
list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/s390/quants.c)
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")
set(GGML_NNPA OFF)
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=z17)
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()
if (GGML_VXE)
message(STATUS "VX/VXE/VXE2 enabled")
list(APPEND ARCH_FLAGS -mvx -mzvector)
list(APPEND ARCH_DEFINITIONS GGML_VXE)
endif()
if (GGML_NNPA)
message(STATUS "NNPA enabled")
list(APPEND ARCH_DEFINITIONS GGML_NNPA)
endif()
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(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_REPACK)
endif()
if (GGML_CPU_KLEIDIAI)
message(STATUS "Using KleidiAI optimized kernels if applicable")
# Disable the KleidiAI tests
set(KLEIDIAI_BUILD_TESTS OFF)
# Fetch KleidiAI sources:
include(FetchContent)
set(KLEIDIAI_COMMIT_TAG "v1.11.0")
set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
set(KLEIDIAI_ARCHIVE_MD5 "3fe9e5ab964c375c53839296eb71eaa2")
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_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)
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_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)
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)
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)
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_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c)
set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}+sve+sve2")
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()
ggml/src/ggml-cpu/amx/amx.cpp
+221
View File
@@ -0,0 +1,221 @@
#include "amx.h"
#include "common.h"
#include "mmq.h"
#include "ggml-backend-impl.h"
#include "ggml-backend.h"
#include "ggml-impl.h"
#include "ggml-cpu.h"
#include "traits.h"
#if defined(__gnu_linux__)
#include <sys/syscall.h>
#include <unistd.h>
#endif
#include <cstdlib>
#include <cstring>
#include <memory>
#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
// AMX type_trais
namespace ggml::cpu::amx {
class tensor_traits : public ggml::cpu::tensor_traits {
bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
size = ggml_backend_amx_desired_wsize(op);
return true;
}
bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) override {
if (op->op == GGML_OP_MUL_MAT) {
ggml_backend_amx_mul_mat(params, op);
return true;
}
return false;
}
};
static ggml::cpu::tensor_traits * get_tensor_traits(ggml_backend_buffer_t, struct ggml_tensor *) {
static tensor_traits traits;
return &traits;
}
} // namespace ggml::cpu::amx
// AMX buffer interface
static void ggml_backend_amx_buffer_free_buffer(ggml_backend_buffer_t buffer) {
free(buffer->context);
}
static void * ggml_backend_amx_buffer_get_base(ggml_backend_buffer_t buffer) {
return (void *) (buffer->context);
}
static enum ggml_status ggml_backend_amx_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
tensor->extra = (void *) ggml::cpu::amx::get_tensor_traits(buffer, tensor);
GGML_UNUSED(buffer);
return GGML_STATUS_SUCCESS;
}
static void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor,
uint8_t value, size_t offset, size_t size) {
memset((char *) tensor->data + offset, value, size);
GGML_UNUSED(buffer);
}
static void ggml_backend_amx_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor,
const void * data, size_t offset, size_t size) {
if (qtype_has_amx_kernels(tensor->type)) {
GGML_LOG_DEBUG("%s: amx repack tensor %s of type %s\n", __func__, tensor->name, ggml_type_name(tensor->type));
ggml_backend_amx_convert_weight(tensor, data, offset, size);
} else {
memcpy((char *) tensor->data + offset, data, size);
}
GGML_UNUSED(buffer);
}
/*
// need to figure what we need to do with buffer->extra.
static void ggml_backend_amx_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
GGML_ASSERT(!qtype_has_amx_kernels(tensor->type));
memcpy(data, (const char *)tensor->data + offset, size);
GGML_UNUSED(buffer);
}
static bool ggml_backend_amx_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
if (ggml_backend_buffer_is_host(src->buffer)) {
if (qtype_has_amx_kernels(src->type)) {
ggml_backend_amx_convert_weight(dst, src->data, 0, ggml_nbytes(dst));
} else {
memcpy(dst->data, src->data, ggml_nbytes(src));
}
return true;
}
return false;
GGML_UNUSED(buffer);
}
*/
static void ggml_backend_amx_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
memset(buffer->context, value, buffer->size);
}
static ggml_backend_buffer_i ggml_backend_amx_buffer_interface = {
/* .free_buffer = */ ggml_backend_amx_buffer_free_buffer,
/* .get_base = */ ggml_backend_amx_buffer_get_base,
/* .init_tensor = */ ggml_backend_amx_buffer_init_tensor,
/* .memset_tensor = */ ggml_backend_amx_buffer_memset_tensor,
/* .set_tensor = */ ggml_backend_amx_buffer_set_tensor,
/* .get_tensor = */ nullptr,
/* .cpy_tensor = */ nullptr,
/* .clear = */ ggml_backend_amx_buffer_clear,
/* .reset = */ nullptr,
};
static const char * ggml_backend_amx_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "AMX";
GGML_UNUSED(buft);
}
static ggml_backend_buffer_t ggml_backend_amx_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
void * data = ggml_aligned_malloc(size);
if (data == NULL) {
fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size);
return NULL;
}
return ggml_backend_buffer_init(buft, ggml_backend_amx_buffer_interface, data, size);
}
static size_t ggml_backend_amx_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
return TENSOR_ALIGNMENT;
GGML_UNUSED(buft);
}
namespace ggml::cpu::amx {
class extra_buffer_type : ggml::cpu::extra_buffer_type {
bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
// handle only 2d gemm for now
auto is_contiguous_2d = [](const struct ggml_tensor * t) {
return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1;
};
if (op->op == GGML_OP_MUL_MAT && is_contiguous_2d(op->src[0]) && // src0 must be contiguous
is_contiguous_2d(op->src[1]) && // src1 must be contiguous
op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_amx_buffer_type() &&
op->ne[0] % (TILE_N * 2) == 0 && // out_features is 32x
(qtype_has_amx_kernels(op->src[0]->type) || (op->src[0]->type == GGML_TYPE_F16))) {
// src1 must be host buffer
if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
return false;
}
// src1 must be float32
if (op->src[1]->type == GGML_TYPE_F32) {
return true;
}
}
return false;
}
ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {
if (op->op == GGML_OP_MUL_MAT && op->src[0]->buffer &&
op->src[0]->buffer->buft == ggml_backend_amx_buffer_type()) {
return (ggml::cpu::tensor_traits *) op->src[0]->extra;
}
return nullptr;
}
};
} // namespace ggml::cpu::amx
static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
return ggml_backend_amx_get_alloc_size(tensor);
GGML_UNUSED(buft);
}
#define ARCH_GET_XCOMP_PERM 0x1022
#define ARCH_REQ_XCOMP_PERM 0x1023
#define XFEATURE_XTILECFG 17
#define XFEATURE_XTILEDATA 18
static bool ggml_amx_init() {
#if defined(__gnu_linux__)
if (syscall(SYS_arch_prctl, ARCH_REQ_XCOMP_PERM, XFEATURE_XTILEDATA)) {
fprintf(stderr, "AMX is not ready to be used!\n");
return false;
}
return true;
#elif defined(_WIN32)
return true;
#endif
}
ggml_backend_buffer_type_t ggml_backend_amx_buffer_type() {
static struct ggml_backend_buffer_type ggml_backend_buffer_type_amx = {
/* .iface = */ {
/* .get_name = */ ggml_backend_amx_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_amx_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_amx_buffer_type_get_alignment,
/* .get_max_size = */ nullptr, // defaults to SIZE_MAX
/* .get_alloc_size = */ ggml_backend_amx_buffer_type_get_alloc_size,
/* .is_host = */ nullptr,
},
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ new ggml::cpu::amx::extra_buffer_type(),
};
if (!ggml_amx_init()) {
return nullptr;
}
return &ggml_backend_buffer_type_amx;
}
#endif // defined(__AMX_INT8__) && defined(__AVX512VNNI__)
ggml/src/ggml-cpu/amx/amx.h
+8
View File
@@ -0,0 +1,8 @@
#include "ggml-backend.h"
#include "ggml-cpu-impl.h"
// GGML internal header
#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
#endif
ggml/src/ggml-cpu/amx/common.h
+91
View File
@@ -0,0 +1,91 @@
#pragma once
#include "ggml.h"
#include "ggml-cpu-impl.h"
#include <algorithm>
#include <memory>
#include <type_traits>
#if defined(GGML_USE_OPENMP)
#include <omp.h>
#endif
#define TILE_M 16
#define TILE_N 16
#define TILE_K 32
#define VNNI_BLK 4
#define AMX_BLK_SIZE 32
#define TMM0 0
#define TMM1 1
#define TMM2 2
#define TMM3 3
#define TMM4 4
#define TMM5 5
#define TMM6 6
#define TMM7 7
// parallel routines
template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type = 0>
inline T div_up(T x, T y) { return (x + y - 1) / y; }
template <typename T>
inline void balance211(T n, T nth, T ith, T& n_start, T& n_end) {
#if 0
// onednn partition pattern
T& n_my = n_end;
if (nth <= 1 || n == 0) {
n_start = 0;
n_my = n;
} else {
T n1 = div_up(n, nth);
T n2 = n1 - 1;
T T1 = n - n2 * nth;
n_my = ith < T1 ? n1 : n2;
n_start = ith <= T1 ? ith*n1 : T1 * n1 + (ith - T1) * n2;
}
n_end += n_start;
#else
// pytorch aten partition pattern
T n_my = div_up(n, nth);
n_start = ith * n_my;
n_end = std::min(n_start + n_my, n);
#endif
}
template <typename func_t>
inline void parallel_for(int n, const func_t& f) {
#if defined(GGML_USE_OPENMP)
#pragma omp parallel
{
int nth = omp_get_num_threads();
int ith = omp_get_thread_num();
int tbegin, tend;
balance211(n, nth, ith, tbegin, tend);
f(tbegin, tend);
}
#else
f(0, n);
#endif
}
template <typename func_t>
inline void parallel_for_ggml(const ggml_compute_params * params, int n, const func_t & f) {
int tbegin, tend;
balance211(n, params->nth, params->ith, tbegin, tend);
f(tbegin, tend);
}
// quantized types that have AMX support
inline bool qtype_has_amx_kernels(const enum ggml_type type) {
// TODO: fix padding for vnni format
return (type == GGML_TYPE_Q4_0) ||
(type == GGML_TYPE_Q4_1) ||
(type == GGML_TYPE_Q8_0) ||
(type == GGML_TYPE_Q4_K) ||
(type == GGML_TYPE_Q5_K) ||
(type == GGML_TYPE_Q6_K) ||
(type == GGML_TYPE_IQ4_XS);
}
ggml/src/ggml-cpu/amx/mmq.cpp
+2512
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/amx/mmq.h
+10
View File
@@ -0,0 +1,10 @@
#pragma once
#include "common.h"
size_t ggml_backend_amx_desired_wsize(const struct ggml_tensor * dst);
size_t ggml_backend_amx_get_alloc_size(const struct ggml_tensor * tensor);
void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
void ggml_backend_amx_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
ggml/src/ggml-cpu/arch-fallback.h
+184
View File
@@ -0,0 +1,184 @@
#pragma once
// Rename `_generic` functions if no native implementation is available.
// This effectively selects the generic implementation.
#if defined(GGML_CPU_GENERIC)
// quants.c
#define quantize_row_q8_0_generic quantize_row_q8_0
#define quantize_row_q8_1_generic quantize_row_q8_1
#define quantize_row_q8_K_generic quantize_row_q8_K
#define ggml_vec_dot_q4_0_q8_0_generic ggml_vec_dot_q4_0_q8_0
#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
#define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
#define ggml_vec_dot_q3_K_q8_K_generic ggml_vec_dot_q3_K_q8_K
#define ggml_vec_dot_q4_K_q8_K_generic ggml_vec_dot_q4_K_q8_K
#define ggml_vec_dot_q5_K_q8_K_generic ggml_vec_dot_q5_K_q8_K
#define ggml_vec_dot_q6_K_q8_K_generic ggml_vec_dot_q6_K_q8_K
#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64)
// repack.cpp
#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#elif defined(__POWERPC__) || defined(__powerpc__)
// ref: https://github.com/ggml-org/llama.cpp/pull/14146#issuecomment-2972561679
// quants.c
#define quantize_row_q8_K_generic quantize_row_q8_K
#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#elif defined(__loongarch64)
// quants.c
#define quantize_row_q8_K_generic quantize_row_q8_K
#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#elif defined(__riscv)
// quants.c
#define quantize_row_q8_K_generic quantize_row_q8_K
#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#elif defined(__s390x__)
// quants.c
#define quantize_row_q8_K_generic quantize_row_q8_K
#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#elif defined(__wasm__)
// quants.c
#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#endif
ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp
+94
View File
@@ -0,0 +1,94 @@
#include "ggml-backend-impl.h"
#if defined(__aarch64__)
#if defined(__linux__)
#include <sys/auxv.h>
#elif defined(__APPLE__)
#include <sys/sysctl.h>
#endif
#if !defined(HWCAP2_I8MM)
#define HWCAP2_I8MM (1 << 13)
#endif
#if !defined(HWCAP2_SME)
#define HWCAP2_SME (1 << 23)
#endif
struct aarch64_features {
// has_neon not needed, aarch64 has NEON guaranteed
bool has_dotprod = false;
bool has_fp16_va = false;
bool has_sve = false;
bool has_sve2 = false;
bool has_i8mm = false;
bool has_sme = false;
aarch64_features() {
#if defined(__linux__)
uint32_t hwcap = getauxval(AT_HWCAP);
uint32_t hwcap2 = getauxval(AT_HWCAP2);
has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
has_fp16_va = !!(hwcap & HWCAP_FPHP);
has_sve = !!(hwcap & HWCAP_SVE);
has_sve2 = !!(hwcap2 & HWCAP2_SVE2);
has_i8mm = !!(hwcap2 & HWCAP2_I8MM);
has_sme = !!(hwcap2 & HWCAP2_SME);
#elif defined(__APPLE__)
int oldp = 0;
size_t size = sizeof(oldp);
if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, NULL, 0) == 0) {
has_dotprod = static_cast<bool>(oldp);
}
if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) == 0) {
has_i8mm = static_cast<bool>(oldp);
}
if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, NULL, 0) == 0) {
has_sme = static_cast<bool>(oldp);
}
// Apple apparently does not implement SVE yet
#endif
}
};
static int ggml_backend_cpu_aarch64_score() {
int score = 1;
aarch64_features af;
#ifdef GGML_USE_DOTPROD
if (!af.has_dotprod) { return 0; }
score += 1<<1;
#endif
#ifdef GGML_USE_FP16_VECTOR_ARITHMETIC
if (!af.has_fp16_va) { return 0; }
score += 1<<2;
#endif
#ifdef GGML_USE_SVE
if (!af.has_sve) { return 0; }
score += 1<<3;
#endif
#ifdef GGML_USE_MATMUL_INT8
if (!af.has_i8mm) { return 0; }
score += 1<<4;
#endif
#ifdef GGML_USE_SVE2
if (!af.has_sve2) { return 0; }
score += 1<<5;
#endif
#ifdef GGML_USE_SME
if (!af.has_sme) { return 0; }
score += 1<<6;
#endif
return score;
}
GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_aarch64_score)
# endif // defined(__aarch64__)
ggml/src/ggml-cpu/arch/arm/quants.c
+4114
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/arch/arm/repack.cpp
+2163
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/arch/loongarch/quants.c
+2639
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp
+82
View File
@@ -0,0 +1,82 @@
# include "ggml-backend-impl.h"
#if defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__)
#if defined(__linux__)
#include <sys/auxv.h>
#endif
#include <string>
struct powerpc_features {
std::string platform = "";
int power_version = -1;
bool has_vsx = false;
powerpc_features() {
#if defined(__linux__)
unsigned long auxval = getauxval(AT_PLATFORM);
if (auxval) {
platform = std::string(reinterpret_cast<const char*>(auxval));
// TBD: Do systems exist that return this in uppercase?
if (platform.substr(0, 5) == "power") {
// Extractt a numeric suffix, if one exists
int vpos = -1;
for (int i = platform.length() - 1; i >= 0; i--) {
if (std::isdigit(platform[i])) {
vpos = i;
} else {
break;
}
}
if (vpos > -1) {
power_version = std::stoi(platform.substr(vpos));
}
}
}
#endif
if (power_version >= 9) {
has_vsx = true;
}
}
};
static int ggml_backend_cpu_powerpc_score() {
int score = 1;
powerpc_features pf;
// Platform scores
#if defined(GGML_USE_POWER7)
if (pf.power_version < 7) { return 0; }
score += 1<<1;
#endif
#if defined(GGML_USE_POWER8)
if (pf.power_version < 8) { return 0; }
score += 1<<2;
#endif
#if defined(GGML_USE_POWER9)
if (pf.power_version < 9) { return 0; }
score += 1<<3;
#endif
#if defined(GGML_USE_POWER10)
if (pf.power_version < 10) { return 0; }
score += 1<<4;
#endif
#if defined(GGML_USE_POWER11)
if (pf.power_version < 11) { return 0; }
score += 1<<5;
#endif
// Feature scores
#if defined(GGML_USE_VSX)
if (!pf.has_vsx) { return 0; }
score += 1<<6;
#endif
return score;
}
GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_powerpc_score)
#endif // defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__)
ggml/src/ggml-cpu/arch/powerpc/quants.c
+2732
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/arch/riscv/quants.c
+2069
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/arch/riscv/repack.cpp
+397
View File
@@ -0,0 +1,397 @@
#define GGML_COMMON_IMPL_CPP
#define GGML_COMMON_DECL_CPP
#include "ggml-common.h"
#include "ggml-backend-impl.h"
#include "ggml-impl.h"
#include "ggml-cpu.h"
#include "ggml-cpu-impl.h"
#include "simd-mappings.h"
#include "traits.h"
#include <cmath>
#include <cstring>
#include <cassert>
#include <cstdlib> // for qsort
#include <cstdio> // for GGML_ASSERT
#define GGML_CPU_CLANG_WORKAROUND
#include "../../repack.h"
#if defined(__GNUC__)
#pragma GCC diagnostic ignored "-Woverlength-strings"
#endif
#define UNUSED GGML_UNUSED
void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
const int ncols_interleaved = 8;
const int blocklen = 8;
assert (n % qk == 0);
assert (nc % ncols_interleaved == 0);
UNUSED(s);
UNUSED(bs);
UNUSED(vx);
UNUSED(vy);
UNUSED(nr);
UNUSED(nc);
UNUSED(nb);
UNUSED(ncols_interleaved);
UNUSED(blocklen);
#if defined __riscv_v
if (__riscv_vlenb() >= QK4_0) {
const size_t vl = QK4_0;
const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
vfloat32m1_t sumf = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
for (int l = 0; l < nb; l++) {
const int64_t a0 = *(const int64_t *)&a_ptr[l].qs[0];
const int64_t a1 = *(const int64_t *)&a_ptr[l].qs[8];
const int64_t a2 = *(const int64_t *)&a_ptr[l].qs[16];
const int64_t a3 = *(const int64_t *)&a_ptr[l].qs[24];
__asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment constraints
const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a0, vl / 4));
const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a1, vl / 4));
const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a2, vl / 4));
const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a3, vl / 4));
const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_hi_m));
const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
// vector version needs Zvfhmin extension
const float a_scale = GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
const float b_scales[8] = {
GGML_CPU_FP16_TO_FP32(b_ptr[l].d[0]),
GGML_CPU_FP16_TO_FP32(b_ptr[l].d[1]),
GGML_CPU_FP16_TO_FP32(b_ptr[l].d[2]),
GGML_CPU_FP16_TO_FP32(b_ptr[l].d[3]),
GGML_CPU_FP16_TO_FP32(b_ptr[l].d[4]),
GGML_CPU_FP16_TO_FP32(b_ptr[l].d[5]),
GGML_CPU_FP16_TO_FP32(b_ptr[l].d[6]),
GGML_CPU_FP16_TO_FP32(b_ptr[l].d[7])
};
const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scale, vl / 4);
sumf = __riscv_vfmacc_vv_f32m1(sumf, tmp1, b_scales_vec, vl / 4);
}
__riscv_vse32_v_f32m1(s + x * ncols_interleaved, sumf, vl / 4);
}
return;
}
#endif
{
float sumf[8];
int sumi;
const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
}
sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
}
}
}
for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
}
}
}
void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
const int ncols_interleaved = 8;
const int blocklen = 8;
assert (n % qk == 0);
assert (nr % 4 == 0);
assert (nc % ncols_interleaved == 0);
UNUSED(s);
UNUSED(bs);
UNUSED(vx);
UNUSED(vy);
UNUSED(nr);
UNUSED(nc);
UNUSED(nb);
UNUSED(ncols_interleaved);
UNUSED(blocklen);
#if defined __riscv_v
if (__riscv_vlenb() >= QK4_0) {
const size_t vl = QK4_0;
for (int y = 0; y < nr / 4; y++) {
const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
vfloat32m1_t sumf0 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
vfloat32m1_t sumf1 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
vfloat32m1_t sumf2 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
vfloat32m1_t sumf3 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
for (int l = 0; l < nb; l++) {
const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
// vector version needs Zvfhmin extension
const float a_scales[4] = {
GGML_CPU_FP16_TO_FP32(a_ptr[l].d[0]),
GGML_CPU_FP16_TO_FP32(a_ptr[l].d[1]),
GGML_CPU_FP16_TO_FP32(a_ptr[l].d[2]),
GGML_CPU_FP16_TO_FP32(a_ptr[l].d[3])
};
const float b_scales[8] = {
GGML_CPU_FP16_TO_FP32(b_ptr[l].d[0]),
GGML_CPU_FP16_TO_FP32(b_ptr[l].d[1]),
GGML_CPU_FP16_TO_FP32(b_ptr[l].d[2]),
GGML_CPU_FP16_TO_FP32(b_ptr[l].d[3]),
GGML_CPU_FP16_TO_FP32(b_ptr[l].d[4]),
GGML_CPU_FP16_TO_FP32(b_ptr[l].d[5]),
GGML_CPU_FP16_TO_FP32(b_ptr[l].d[6]),
GGML_CPU_FP16_TO_FP32(b_ptr[l].d[7])
};
const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
const int64_t A0 = *(const int64_t *)&a_ptr[l].qs[0];
const int64_t A4 = *(const int64_t *)&a_ptr[l].qs[32];
const int64_t A8 = *(const int64_t *)&a_ptr[l].qs[64];
const int64_t Ac = *(const int64_t *)&a_ptr[l].qs[96];
__asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
vint16m4_t sumi_l0;
{
const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A0, vl / 4));
const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A4, vl / 4));
const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A8, vl / 4));
const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ac, vl / 4));
const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
sumi_l0 = sumi_hi_m;
}
{
const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l0));
const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[0], vl / 4);
sumf0 = __riscv_vfmacc_vv_f32m1(sumf0, tmp1, b_scales_vec, vl / 4);
}
const int64_t A1 = *(const int64_t *)&a_ptr[l].qs[8];
const int64_t A5 = *(const int64_t *)&a_ptr[l].qs[40];
const int64_t A9 = *(const int64_t *)&a_ptr[l].qs[72];
const int64_t Ad = *(const int64_t *)&a_ptr[l].qs[104];
__asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
vint16m4_t sumi_l1;
{
const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A1, vl / 4));
const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A5, vl / 4));
const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A9, vl / 4));
const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ad, vl / 4));
const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
sumi_l1 = sumi_hi_m;
}
{
const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l1));
const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[1], vl / 4);
sumf1 = __riscv_vfmacc_vv_f32m1(sumf1, tmp1, b_scales_vec, vl / 4);
}
const int64_t A2 = *(const int64_t *)&a_ptr[l].qs[16];
const int64_t A6 = *(const int64_t *)&a_ptr[l].qs[48];
const int64_t Aa = *(const int64_t *)&a_ptr[l].qs[80];
const int64_t Ae = *(const int64_t *)&a_ptr[l].qs[112];
__asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
vint16m4_t sumi_l2;
{
const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A2, vl / 4));
const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A6, vl / 4));
const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Aa, vl / 4));
const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ae, vl / 4));
const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
sumi_l2 = sumi_hi_m;
}
{
const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l2));
const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[2], vl / 4);
sumf2 = __riscv_vfmacc_vv_f32m1(sumf2, tmp1, b_scales_vec, vl / 4);
}
const int64_t A3 = *(const int64_t *)&a_ptr[l].qs[24];
const int64_t A7 = *(const int64_t *)&a_ptr[l].qs[56];
const int64_t Ab = *(const int64_t *)&a_ptr[l].qs[88];
const int64_t Af = *(const int64_t *)&a_ptr[l].qs[120];
__asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
vint16m4_t sumi_l3;
{
const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A3, vl / 4));
const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A7, vl / 4));
const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ab, vl / 4));
const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Af, vl / 4));
const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
sumi_l3 = sumi_hi_m;
}
{
const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l3));
const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[3], vl / 4);
sumf3 = __riscv_vfmacc_vv_f32m1(sumf3, tmp1, b_scales_vec, vl / 4);
}
}
__riscv_vse32_v_f32m1(&s[(y * 4 + 0) * bs + x * ncols_interleaved], sumf0, vl / 4);
__riscv_vse32_v_f32m1(&s[(y * 4 + 1) * bs + x * ncols_interleaved], sumf1, vl / 4);
__riscv_vse32_v_f32m1(&s[(y * 4 + 2) * bs + x * ncols_interleaved], sumf2, vl / 4);
__riscv_vse32_v_f32m1(&s[(y * 4 + 3) * bs + x * ncols_interleaved], sumf3, vl / 4);
}
}
return;
}
#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
float sumf[4][8];
int sumi;
for (int y = 0; y < nr / 4; y++) {
const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
}
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
}
sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
}
}
}
}
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++)
s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
}
}
}
}
ggml/src/ggml-cpu/arch/s390/quants.c
+1300
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/arch/wasm/quants.c
+1481
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/arch/x86/cpu-feats.cpp
+327
View File
@@ -0,0 +1,327 @@
#include "ggml-backend-impl.h"
#if defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64))
#ifdef _MSC_VER
#include <intrin.h>
#endif
#include <cstring>
#include <vector>
#include <bitset>
#include <array>
#include <string>
// ref: https://cdrdv2-public.intel.com/782156/325383-sdm-vol-2abcd.pdf
struct cpuid_x86 {
bool SSE3(void) { return f_1_ecx[0]; }
bool PCLMULQDQ(void) { return f_1_ecx[1]; }
bool MONITOR(void) { return f_1_ecx[3]; }
bool SSSE3(void) { return f_1_ecx[9]; }
bool FMA(void) { return f_1_ecx[12]; }
bool CMPXCHG16B(void) { return f_1_ecx[13]; }
bool SSE41(void) { return f_1_ecx[19]; }
bool SSE42(void) { return f_1_ecx[20]; }
bool MOVBE(void) { return f_1_ecx[22]; }
bool POPCNT(void) { return f_1_ecx[23]; }
bool AES(void) { return f_1_ecx[25]; }
bool XSAVE(void) { return f_1_ecx[26]; }
bool OSXSAVE(void) { return f_1_ecx[27]; }
bool AVX(void) { return f_1_ecx[28]; }
bool F16C(void) { return f_1_ecx[29]; }
bool RDRAND(void) { return f_1_ecx[30]; }
bool MSR(void) { return f_1_edx[5]; }
bool CX8(void) { return f_1_edx[8]; }
bool SEP(void) { return f_1_edx[11]; }
bool CMOV(void) { return f_1_edx[15]; }
bool CLFSH(void) { return f_1_edx[19]; }
bool MMX(void) { return f_1_edx[23]; }
bool FXSR(void) { return f_1_edx[24]; }
bool SSE(void) { return f_1_edx[25]; }
bool SSE2(void) { return f_1_edx[26]; }
bool FSGSBASE(void) { return f_7_ebx[0]; }
bool BMI1(void) { return f_7_ebx[3]; }
bool HLE(void) { return is_intel && f_7_ebx[4]; }
bool AVX2(void) { return f_7_ebx[5]; }
bool BMI2(void) { return f_7_ebx[8]; }
bool ERMS(void) { return f_7_ebx[9]; }
bool INVPCID(void) { return f_7_ebx[10]; }
bool RTM(void) { return is_intel && f_7_ebx[11]; }
bool AVX512F(void) { return f_7_ebx[16]; }
bool AVX512DQ(void) { return f_7_ebx[17]; }
bool RDSEED(void) { return f_7_ebx[18]; }
bool ADX(void) { return f_7_ebx[19]; }
bool AVX512PF(void) { return f_7_ebx[26]; }
bool AVX512ER(void) { return f_7_ebx[27]; }
bool AVX512CD(void) { return f_7_ebx[28]; }
bool AVX512BW(void) { return f_7_ebx[30]; }
bool AVX512VL(void) { return f_7_ebx[31]; }
bool SHA(void) { return f_7_ebx[29]; }
bool PREFETCHWT1(void) { return f_7_ecx[0]; }
bool LAHF(void) { return f_81_ecx[0]; }
bool LZCNT(void) { return is_intel && f_81_ecx[5]; }
bool ABM(void) { return is_amd && f_81_ecx[5]; }
bool SSE4a(void) { return is_amd && f_81_ecx[6]; }
bool XOP(void) { return is_amd && f_81_ecx[11]; }
bool TBM(void) { return is_amd && f_81_ecx[21]; }
bool SYSCALL(void) { return is_intel && f_81_edx[11]; }
bool MMXEXT(void) { return is_amd && f_81_edx[22]; }
bool RDTSCP(void) { return is_intel && f_81_edx[27]; }
bool _3DNOWEXT(void) { return is_amd && f_81_edx[30]; }
bool _3DNOW(void) { return is_amd && f_81_edx[31]; }
bool AVX512_VBMI(void) { return f_7_ecx[1]; }
bool AVX512_VNNI(void) { return f_7_ecx[11]; }
bool AVX512_FP16(void) { return f_7_edx[23]; }
bool AVX512_BF16(void) { return f_7_1_eax[5]; }
bool AVX_VNNI(void) { return f_7_1_eax[4]; }
bool AMX_TILE(void) { return f_7_edx[24]; }
bool AMX_INT8(void) { return f_7_edx[25]; }
bool AMX_FP16(void) { return f_7_1_eax[21]; }
bool AMX_BF16(void) { return f_7_edx[22]; }
#ifdef _MSC_VER
static void cpuid(int cpu_info[4], int eax) {
__cpuid(cpu_info, eax);
}
static void cpuidex(int cpu_info[4], int eax, int ecx) {
__cpuidex(cpu_info, eax, ecx);
}
#else
static void cpuid(int cpu_info[4], int eax) {
__asm__ __volatile__(
"cpuid"
: "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
: "a"(eax), "c"(0));
}
static void cpuidex(int cpu_info[4], int eax, int ecx) {
__asm__ __volatile__(
"cpuid"
: "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
: "a"(eax), "c"(ecx));
}
#endif
cpuid_x86() {
std::array<int, 4> cpui;
std::vector<std::array<int, 4>> data;
// calling __cpuid with 0x0 as the function_id argument
// gets the number of the highest valid function ID.
cpuid(cpui.data(), 0);
int n_ids = cpui[0];
for (int i = 0; i <= n_ids; ++i) {
cpuidex(cpui.data(), i, 0);
data.push_back(cpui);
}
// capture vendor string
char vendor[0x20] = {};
*reinterpret_cast<int *>(vendor) = data[0][1];
*reinterpret_cast<int *>(vendor + 4) = data[0][3];
*reinterpret_cast<int *>(vendor + 8) = data[0][2];
this->vendor = vendor;
if (this->vendor == "GenuineIntel") {
is_intel = true;
} else if (this->vendor == "AuthenticAMD") {
is_amd = true;
}
// load bitset with flags for function 0x00000001
if (n_ids >= 1) {
f_1_ecx = data[1][2];
f_1_edx = data[1][3];
}
// load bitset with flags for function 0x00000007
if (n_ids >= 7) {
f_7_ebx = data[7][1];
f_7_ecx = data[7][2];
f_7_edx = data[7][3];
cpuidex(cpui.data(), 7, 1);
f_7_1_eax = cpui[0];
}
// calling __cpuid with 0x80000000 as the function_id argument
// gets the number of the highest valid extended ID.
cpuid(cpui.data(), 0x80000000);
unsigned int n_ex_ids = cpui[0];
std::vector<std::array<int, 4>> ext_data;
for (unsigned int i = 0x80000000; i <= n_ex_ids; ++i) {
cpuidex(cpui.data(), i, 0);
ext_data.push_back(cpui);
}
// load bitset with flags for function 0x80000001
if (n_ex_ids >= 0x80000001) {
f_81_ecx = ext_data[1][2];
f_81_edx = ext_data[1][3];
}
// interpret CPU brand string if reported
char brand[0x40] = {};
if (n_ex_ids >= 0x80000004) {
std::memcpy(brand, ext_data[2].data(), sizeof(cpui));
std::memcpy(brand + 16, ext_data[3].data(), sizeof(cpui));
std::memcpy(brand + 32, ext_data[4].data(), sizeof(cpui));
this->brand = brand;
}
}
bool is_intel = false;
bool is_amd = false;
std::string vendor;
std::string brand;
std::bitset<32> f_1_ecx;
std::bitset<32> f_1_edx;
std::bitset<32> f_7_ebx;
std::bitset<32> f_7_ecx;
std::bitset<32> f_7_edx;
std::bitset<32> f_7_1_eax;
std::bitset<32> f_81_ecx;
std::bitset<32> f_81_edx;
};
#if 0
void test_x86_is() {
cpuid_x86 is;
printf("CPU Vendor: %s\n", is.vendor.c_str());
printf("Brand: %s\n", is.brand.c_str());
printf("is_intel: %d\n", is.is_intel);
printf("is_amd: %d\n", is.is_amd);
printf("sse3: %d\n", is.SSE3());
printf("pclmulqdq: %d\n", is.PCLMULQDQ());
printf("ssse3: %d\n", is.SSSE3());
printf("fma: %d\n", is.FMA());
printf("cmpxchg16b: %d\n", is.CMPXCHG16B());
printf("sse41: %d\n", is.SSE41());
printf("sse42: %d\n", is.SSE42());
printf("movbe: %d\n", is.MOVBE());
printf("popcnt: %d\n", is.POPCNT());
printf("aes: %d\n", is.AES());
printf("xsave: %d\n", is.XSAVE());
printf("osxsave: %d\n", is.OSXSAVE());
printf("avx: %d\n", is.AVX());
printf("f16c: %d\n", is.F16C());
printf("rdrand: %d\n", is.RDRAND());
printf("msr: %d\n", is.MSR());
printf("cx8: %d\n", is.CX8());
printf("sep: %d\n", is.SEP());
printf("cmov: %d\n", is.CMOV());
printf("clflush: %d\n", is.CLFSH());
printf("mmx: %d\n", is.MMX());
printf("fxsr: %d\n", is.FXSR());
printf("sse: %d\n", is.SSE());
printf("sse2: %d\n", is.SSE2());
printf("fsgsbase: %d\n", is.FSGSBASE());
printf("bmi1: %d\n", is.BMI1());
printf("hle: %d\n", is.HLE());
printf("avx2: %d\n", is.AVX2());
printf("bmi2: %d\n", is.BMI2());
printf("erms: %d\n", is.ERMS());
printf("invpcid: %d\n", is.INVPCID());
printf("rtm: %d\n", is.RTM());
printf("avx512f: %d\n", is.AVX512F());
printf("rdseed: %d\n", is.RDSEED());
printf("adx: %d\n", is.ADX());
printf("avx512pf: %d\n", is.AVX512PF());
printf("avx512er: %d\n", is.AVX512ER());
printf("avx512cd: %d\n", is.AVX512CD());
printf("sha: %d\n", is.SHA());
printf("prefetchwt1: %d\n", is.PREFETCHWT1());
printf("lahf: %d\n", is.LAHF());
printf("lzcnt: %d\n", is.LZCNT());
printf("abm: %d\n", is.ABM());
printf("sse4a: %d\n", is.SSE4a());
printf("xop: %d\n", is.XOP());
printf("tbm: %d\n", is.TBM());
printf("syscall: %d\n", is.SYSCALL());
printf("mmxext: %d\n", is.MMXEXT());
printf("rdtscp: %d\n", is.RDTSCP());
printf("3dnowext: %d\n", is._3DNOWEXT());
printf("3dnow: %d\n", is._3DNOW());
printf("avx512_vbmi: %d\n", is.AVX512_VBMI());
printf("avx512_vnni: %d\n", is.AVX512_VNNI());
printf("avx512_fp16: %d\n", is.AVX512_FP16());
printf("avx512_bf16: %d\n", is.AVX512_BF16());
printf("amx_tile: %d\n", is.AMX_TILE());
printf("amx_int8: %d\n", is.AMX_INT8());
printf("amx_fp16: %d\n", is.AMX_FP16());
printf("amx_bf16: %d\n", is.AMX_BF16());
}
#endif
static int ggml_backend_cpu_x86_score() {
// FIXME: this does not check for OS support
int score = 1;
cpuid_x86 is;
#ifdef GGML_FMA
if (!is.FMA()) { return 0; }
score += 1;
#endif
#ifdef GGML_F16C
if (!is.F16C()) { return 0; }
score += 1<<1;
#endif
#ifdef GGML_SSE42
if (!is.SSE42()) { return 0; }
score += 1<<2;
#endif
#ifdef GGML_BMI2
if (!is.BMI2()) { return 0; }
score += 1<<3;
#endif
#ifdef GGML_AVX
if (!is.AVX()) { return 0; }
score += 1<<4;
#endif
#ifdef GGML_AVX2
if (!is.AVX2()) { return 0; }
score += 1<<5;
#endif
#ifdef GGML_AVX_VNNI
if (!is.AVX_VNNI()) { return 0; }
score += 1<<6;
#endif
#ifdef GGML_AVX512
if (!is.AVX512F()) { return 0; }
if (!is.AVX512CD()) { return 0; }
if (!is.AVX512VL()) { return 0; }
if (!is.AVX512DQ()) { return 0; }
if (!is.AVX512BW()) { return 0; }
score += 1<<7;
#endif
#ifdef GGML_AVX512_VBMI
if (!is.AVX512_VBMI()) { return 0; }
score += 1<<8;
#endif
#ifdef GGML_AVX512_BF16
if (!is.AVX512_BF16()) { return 0; }
score += 1<<9;
#endif
#ifdef GGML_AVX512_VNNI
if (!is.AVX512_VNNI()) { return 0; }
score += 1<<10;
#endif
#ifdef GGML_AMX_INT8
if (!is.AMX_INT8()) { return 0; }
score += 1<<11;
#endif
return score;
}
GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_x86_score)
#endif // defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64))
ggml/src/ggml-cpu/arch/x86/quants.c
+4311
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/arch/x86/repack.cpp
+3285
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/binary-ops.cpp
+158
View File
@@ -0,0 +1,158 @@
#include "binary-ops.h"
#if defined(GGML_USE_ACCELERATE)
#include <Accelerate/Accelerate.h>
using vDSP_fn_t = void (*)(const float *, vDSP_Stride, const float *, vDSP_Stride, float *, vDSP_Stride, vDSP_Length);
#endif
static inline float op_add(float a, float b) {
return a + b;
}
static inline float op_sub(float a, float b) {
return a - b;
}
static inline float op_mul(float a, float b) {
return a * b;
}
static inline float op_div(float a, float b) {
return a / b;
}
template <float (*op)(float, float), typename src0_t, typename src1_t, typename dst_t>
static inline void vec_binary_op_contiguous(const int64_t n, dst_t * z, const src0_t * x, const src1_t * y) {
constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32;
constexpr auto src1_to_f32 = type_conversion_table<src1_t>::to_f32;
constexpr auto f32_to_dst = type_conversion_table<dst_t >::from_f32;
for (int i = 0; i < n; i++) {
z[i] = f32_to_dst(op(src0_to_f32(x[i]), src1_to_f32(y[i])));
}
}
template <float (*op)(float, float), typename src0_t, typename src1_t, typename dst_t>
static inline void vec_binary_op_non_contiguous(const int64_t n, const int64_t ne10, const int64_t nb10, dst_t * z, const src0_t * x, const src1_t * y) {
constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32;
constexpr auto src1_to_f32 = type_conversion_table<src1_t>::to_f32;
constexpr auto f32_to_dst = type_conversion_table<dst_t >::from_f32;
for (int i = 0; i < n; i++) {
int i10 = i % ne10;
const src1_t * y_ptr = (const src1_t *)((const char *)y + i10*nb10);
z[i] = f32_to_dst(op(src0_to_f32(x[i]), src1_to_f32(*y_ptr)));
}
}
template <float (*op)(float, float), typename src0_t, typename src1_t, typename dst_t>
static void apply_binary_op(const ggml_compute_params * params, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
GGML_TENSOR_BINARY_OP_LOCALS
GGML_ASSERT( nb0 == sizeof(dst_t));
GGML_ASSERT(nb00 == sizeof(src0_t));
const auto [ir0, ir1] = get_thread_range(params, src0);
const bool is_src1_contiguous = (nb10 == sizeof(src1_t));
if (!is_src1_contiguous) { // broadcast not implemented yet for non-contiguous
GGML_ASSERT(ggml_are_same_shape(src0, src1));
}
#ifdef GGML_USE_ACCELERATE
vDSP_fn_t vDSP_op = nullptr;
// TODO - avoid the f32-only check using type 'trait' lookup tables and row-based src-to-float conversion functions
if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
if (op == op_add) {
vDSP_op = vDSP_vadd;
} else if (op == op_sub) {
vDSP_op = vDSP_vsub;
} else if (op == op_mul) {
vDSP_op = vDSP_vmul;
} else if (op == op_div) {
vDSP_op = vDSP_vdiv;
}
}
#endif
for (int64_t ir = ir0; ir < ir1; ++ir) {
const int64_t i03 = ir/(ne02*ne01);
const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
const int64_t i13 = i03 % ne13;
const int64_t i12 = i02 % ne12;
const int64_t i11 = i01 % ne11;
dst_t * dst_ptr = (dst_t *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 );
const src0_t * src0_ptr = (const src0_t *) ((const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
const src1_t * src1_ptr = (const src1_t *) ((const char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
if (is_src1_contiguous) {
// src1 is broadcastable across src0 and dst in i1, i2, i3
const int64_t nr0 = ne00 / ne10;
for (int64_t r = 0; r < nr0; ++r) {
#ifdef GGML_USE_ACCELERATE
if constexpr (std::is_same_v<src0_t, float> && std::is_same_v<src1_t, float> && std::is_same_v<dst_t, float>) {
if (vDSP_op != nullptr) {
vDSP_op(src1_ptr, 1, src0_ptr + r*ne10, 1, dst_ptr + r*ne10, 1, ne10);
continue;
}
}
#endif
vec_binary_op_contiguous<op>(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr);
}
} else {
vec_binary_op_non_contiguous<op>(ne0, ne10, nb10, dst_ptr, src0_ptr, src1_ptr);
}
}
}
// TODO: Use the 'traits' lookup table (for type conversion fns), instead of a mass of 'if' conditions with long templates
template <float (*op)(float, float)>
static void binary_op(const ggml_compute_params * params, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
/* */ if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { // all f32
apply_binary_op<op, float, float, float>(params, dst);
} else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { // all f16
apply_binary_op<op, ggml_fp16_t, ggml_fp16_t, ggml_fp16_t>(params, dst);
} else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_BF16) { // all bf16
apply_binary_op<op, ggml_bf16_t, ggml_bf16_t, ggml_bf16_t>(params, dst);
} else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_BF16) {
apply_binary_op<op, ggml_bf16_t, float, ggml_bf16_t>(params, dst);
} else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
apply_binary_op<op, ggml_bf16_t, float, float>(params, dst);
} else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
apply_binary_op<op, ggml_fp16_t, float, ggml_fp16_t>(params, dst);
} else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
apply_binary_op<op, ggml_fp16_t, float, float>(params, dst);
} else {
GGML_ABORT("%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__,
ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type));
}
}
void ggml_compute_forward_add_non_quantized(const ggml_compute_params * params, ggml_tensor * dst) {
binary_op<op_add>(params, dst);
}
void ggml_compute_forward_sub(const ggml_compute_params * params, ggml_tensor * dst) {
binary_op<op_sub>(params, dst);
}
void ggml_compute_forward_mul(const ggml_compute_params * params, ggml_tensor * dst) {
binary_op<op_mul>(params, dst);
}
void ggml_compute_forward_div(const ggml_compute_params * params, ggml_tensor * dst) {
binary_op<op_div>(params, dst);
}
ggml/src/ggml-cpu/binary-ops.h
+16
View File
@@ -0,0 +1,16 @@
#pragma once
#include "common.h"
#ifdef __cplusplus
extern "C" {
#endif
void ggml_compute_forward_add_non_quantized(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_sub(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_mul(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_div(const struct ggml_compute_params * params, struct ggml_tensor * dst);
#ifdef __cplusplus
}
#endif
ggml/src/ggml-cpu/common.h
+73
View File
@@ -0,0 +1,73 @@
#pragma once
#include "ggml.h"
#include "traits.h"
#include "ggml-cpu-impl.h"
#include "ggml-impl.h"
#include "simd-mappings.h"
#ifdef __cplusplus
#include <utility>
// convenience functions/macros for use in template calls
// note: these won't be required after the 'traits' lookup table is used.
static inline ggml_fp16_t f32_to_f16(float x) {
return GGML_CPU_FP32_TO_FP16(x);
}
static inline float f16_to_f32(ggml_fp16_t x) {
return GGML_CPU_FP16_TO_FP32(x);
}
static inline ggml_bf16_t f32_to_bf16(float x) {
return GGML_FP32_TO_BF16(x);
}
static inline float bf16_to_f32(ggml_bf16_t x) {
return GGML_BF16_TO_FP32(x);
}
static inline float f32_to_f32(float x) {
return x;
}
// TODO - merge this into the traits table, after using row-based conversions
template <class T>
struct type_conversion_table;
template <>
struct type_conversion_table<ggml_fp16_t> {
static constexpr float (*to_f32)(ggml_fp16_t) = f16_to_f32;
static constexpr ggml_fp16_t (*from_f32)(float) = f32_to_f16;
};
template <>
struct type_conversion_table<float> {
static constexpr float (*to_f32)(float) = f32_to_f32;
static constexpr float (*from_f32)(float) = f32_to_f32;
};
template <>
struct type_conversion_table<ggml_bf16_t> {
static constexpr float (*to_f32)(ggml_bf16_t) = bf16_to_f32;
static constexpr ggml_bf16_t (*from_f32)(float) = f32_to_bf16;
};
static std::pair<int64_t, int64_t> get_thread_range(const struct ggml_compute_params * params, const struct ggml_tensor * src0) {
const int64_t ith = params->ith;
const int64_t nth = params->nth;
const int64_t nr = ggml_nrows(src0);
// rows per thread
const int64_t dr = (nr + nth - 1)/nth;
// row range for this thread
const int64_t ir0 = dr*ith;
const int64_t ir1 = MIN(ir0 + dr, nr);
return {ir0, ir1};
}
#endif
ggml/src/ggml-cpu/ggml-cpu-impl.h
+517
View File
@@ -0,0 +1,517 @@
#pragma once
// GGML CPU internal header
#include "ggml.h"
#include "ggml-impl.h"
#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
//#include <stddef.h>
#include <stdbool.h>
#include <string.h> // memcpy
#include <math.h> // fabsf
#ifdef __cplusplus
extern "C" {
#endif
struct ggml_compute_params {
// ith = thread index, nth = number of threads
int ith, nth;
// work buffer for all threads
size_t wsize;
void * wdata;
struct ggml_threadpool * threadpool;
};
#if defined(_MSC_VER)
#define m512bh(p) p
#define m512i(p) p
#else
#define m512bh(p) (__m512bh)(p)
#define m512i(p) (__m512i)(p)
#endif
// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
#ifndef __FMA__
#define __FMA__
#endif
#ifndef __F16C__
#define __F16C__
#endif
#endif
// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
#ifndef __SSE3__
#define __SSE3__
#endif
#ifndef __SSSE3__
#define __SSSE3__
#endif
#endif
#if defined(__s390x__) && defined(__VEC__)
#ifndef __VXE__
#define __VXE__
#endif // __VXE__
#ifndef __VXE2__
#define __VXE2__
#endif // __VXE2__
#endif // __s390x__ && __VEC__
#if defined(__s390x__) && defined(GGML_NNPA)
#ifndef __NNPA__
#define __NNPA__
#endif // __NNPA__
#endif // __s390x__ && GGML_NNPA
#if defined(__ARM_FEATURE_SVE)
#include <sys/prctl.h>
#endif
#if defined(__ARM_NEON)
// ref: https://github.com/ggml-org/llama.cpp/pull/5404
#ifdef _MSC_VER
#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
#else
#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
#endif // _MSC_VER
#if !defined(__aarch64__)
// 32-bit ARM compatibility
// vaddlvq_s16
// vpaddq_s16
// vpaddq_s32
// vaddvq_s32
// vaddvq_f32
// vmaxvq_f32
// vcvtnq_s32_f32
// vzip1_u8
// vzip2_u8
inline static int32_t vaddlvq_s16(int16x8_t v) {
int32x4_t v0 = vreinterpretq_s32_s64(vpaddlq_s32(vpaddlq_s16(v)));
return vgetq_lane_s32(v0, 0) + vgetq_lane_s32(v0, 2);
}
inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
return vcombine_s16(a0, b0);
}
inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
return vcombine_s32(a0, b0);
}
inline static int32_t vaddvq_s32(int32x4_t v) {
return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
}
inline static float vaddvq_f32(float32x4_t v) {
return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
}
inline static float vmaxvq_f32(float32x4_t v) {
return
MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
}
inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
int32x4_t res;
res[0] = roundf(vgetq_lane_f32(v, 0));
res[1] = roundf(vgetq_lane_f32(v, 1));
res[2] = roundf(vgetq_lane_f32(v, 2));
res[3] = roundf(vgetq_lane_f32(v, 3));
return res;
}
inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
uint8x8_t res;
res[0] = a[0]; res[1] = b[0];
res[2] = a[1]; res[3] = b[1];
res[4] = a[2]; res[5] = b[2];
res[6] = a[3]; res[7] = b[3];
return res;
}
inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
uint8x8_t res;
res[0] = a[4]; res[1] = b[4];
res[2] = a[5]; res[3] = b[5];
res[4] = a[6]; res[5] = b[6];
res[6] = a[7]; res[7] = b[7];
return res;
}
// vld1q_s16_x2
// vld1q_u8_x2
// vld1q_u8_x4
// vld1q_s8_x2
// vld1q_s8_x4
// TODO: double-check these work correctly
typedef struct ggml_int16x8x2_t {
int16x8_t val[2];
} ggml_int16x8x2_t;
inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
ggml_int16x8x2_t res;
res.val[0] = vld1q_s16(ptr + 0);
res.val[1] = vld1q_s16(ptr + 8);
return res;
}
typedef struct ggml_uint8x16x2_t {
uint8x16_t val[2];
} ggml_uint8x16x2_t;
inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
ggml_uint8x16x2_t res;
res.val[0] = vld1q_u8(ptr + 0);
res.val[1] = vld1q_u8(ptr + 16);
return res;
}
typedef struct ggml_uint8x16x4_t {
uint8x16_t val[4];
} ggml_uint8x16x4_t;
inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
ggml_uint8x16x4_t res;
res.val[0] = vld1q_u8(ptr + 0);
res.val[1] = vld1q_u8(ptr + 16);
res.val[2] = vld1q_u8(ptr + 32);
res.val[3] = vld1q_u8(ptr + 48);
return res;
}
typedef struct ggml_int8x16x2_t {
int8x16_t val[2];
} ggml_int8x16x2_t;
inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
ggml_int8x16x2_t res;
res.val[0] = vld1q_s8(ptr + 0);
res.val[1] = vld1q_s8(ptr + 16);
return res;
}
typedef struct ggml_int8x16x4_t {
int8x16_t val[4];
} ggml_int8x16x4_t;
inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
ggml_int8x16x4_t res;
res.val[0] = vld1q_s8(ptr + 0);
res.val[1] = vld1q_s8(ptr + 16);
res.val[2] = vld1q_s8(ptr + 32);
res.val[3] = vld1q_s8(ptr + 48);
return res;
}
// NOTE: not tested
inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
int8x16_t res;
res[ 0] = a[b[ 0]];
res[ 1] = a[b[ 1]];
res[ 2] = a[b[ 2]];
res[ 3] = a[b[ 3]];
res[ 4] = a[b[ 4]];
res[ 5] = a[b[ 5]];
res[ 6] = a[b[ 6]];
res[ 7] = a[b[ 7]];
res[ 8] = a[b[ 8]];
res[ 9] = a[b[ 9]];
res[10] = a[b[10]];
res[11] = a[b[11]];
res[12] = a[b[12]];
res[13] = a[b[13]];
res[14] = a[b[14]];
res[15] = a[b[15]];
return res;
}
// NOTE: not tested
inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
uint8x16_t res;
res[ 0] = a[b[ 0]];
res[ 1] = a[b[ 1]];
res[ 2] = a[b[ 2]];
res[ 3] = a[b[ 3]];
res[ 4] = a[b[ 4]];
res[ 5] = a[b[ 5]];
res[ 6] = a[b[ 6]];
res[ 7] = a[b[ 7]];
res[ 8] = a[b[ 8]];
res[ 9] = a[b[ 9]];
res[10] = a[b[10]];
res[11] = a[b[11]];
res[12] = a[b[12]];
res[13] = a[b[13]];
res[14] = a[b[14]];
res[15] = a[b[15]];
return res;
}
#else
#define ggml_int16x8x2_t int16x8x2_t
#define ggml_uint8x16x2_t uint8x16x2_t
#define ggml_uint8x16x4_t uint8x16x4_t
#define ggml_int8x16x2_t int8x16x2_t
#define ggml_int8x16x4_t int8x16x4_t
#define ggml_vld1q_s16_x2 vld1q_s16_x2
#define ggml_vld1q_u8_x2 vld1q_u8_x2
#define ggml_vld1q_u8_x4 vld1q_u8_x4
#define ggml_vld1q_s8_x2 vld1q_s8_x2
#define ggml_vld1q_s8_x4 vld1q_s8_x4
#define ggml_vqtbl1q_s8 vqtbl1q_s8
#define ggml_vqtbl1q_u8 vqtbl1q_u8
#endif // !defined(__aarch64__)
#if !defined(__ARM_FEATURE_DOTPROD)
inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
}
#else
#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
#endif // !defined(__ARM_FEATURE_DOTPROD)
#endif // defined(__ARM_NEON)
#ifdef __wasm_simd128__
#include <wasm_simd128.h>
#endif
#ifdef __POWER9_VECTOR__
#include <altivec.h>
#endif
#if defined(_MSC_VER) || defined(__MINGW32__)
#include <intrin.h>
#elif defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) || defined(__SSE__)
#include <immintrin.h>
#endif
#ifdef __riscv_v_intrinsic
#include <riscv_vector.h>
#endif
#if defined(__loongarch64)
#if defined(__loongarch_asx)
#include <lasxintrin.h>
#endif
#if defined(__loongarch_sx)
#include <lsxintrin.h>
#endif
#endif
#if defined(__VXE__) || defined(__VXE2__)
#include <vecintrin.h>
#define vec_neg(a) (-(a)) // Vector Negate
#define vec_add(a, b) ((a) + (b)) // Vector Add
#define vec_sub(a, b) ((a) - (b)) // Vector Subtract
#define vec_mul(a, b) ((a) * (b)) // Vector Multiply
#define vec_div(a, b) ((a) / (b)) // Vector Divide
#define vec_sl(a, b) ((a) << (b)) // Vector Shift Left
#define vec_sra(a, b) ((a) >> (b)) // Vector Shift Right
#define vec_sr(a, b) ((a) >> (b)) // Vector Shift Right Algebraic
#define vec_slo(a, b) vec_slb(a, (b) << 64) // Vector Shift Left by Octet
#define vec_sro(a, b) vec_srb(a, (b) << 64) // Vector Shift Right by Octet
#ifndef vec_and
#define vec_and(a, b) ((a) & (b)) // Vector AND
#endif
#ifndef vec_or
#define vec_or(a, b) ((a) | (b)) // Vector OR
#endif
#ifndef vec_xor
#define vec_xor(a, b) ((a) ^ (b)) // Vector XOR
#endif
typedef signed char char8x16_t __attribute__((vector_size(16)));
typedef unsigned char uchar8x16_t __attribute__((vector_size(16)));
typedef int8_t int8x16_t __attribute__((vector_size(16)));
typedef int16_t int16x8_t __attribute__((vector_size(16)));
typedef int32_t int32x4_t __attribute__((vector_size(16)));
typedef uint8_t uint8x16_t __attribute__((vector_size(16)));
typedef uint16_t uint16x8_t __attribute__((vector_size(16)));
typedef uint32_t uint32x4_t __attribute__((vector_size(16)));
typedef float float32x4_t __attribute__((vector_size(16)));
typedef double double64x2_t __attribute__((vector_size(16)));
typedef signed long long long64x2_t __attribute__((vector_size(16)));
typedef unsigned long long ulong64x2_t __attribute__((vector_size(16)));
typedef struct ggml_uint8x16x2_t {
uint8x16_t val[2];
} ggml_uint8x16x2_t;
inline static ggml_uint8x16x2_t ggml_vec_xl_u8x2(const uint8_t * ptr) {
ggml_uint8x16x2_t res;
res.val[0] = vec_xl( 0, ptr);
res.val[1] = vec_xl(16, ptr);
return res;
}
typedef struct ggml_uint8x16x4_t {
uint8x16_t val[4];
} ggml_uint8x16x4_t;
inline static ggml_uint8x16x4_t ggml_vec_xl_u8x4(const uint8_t * ptr) {
ggml_uint8x16x4_t res;
res.val[0] = vec_xl( 0, ptr);
res.val[1] = vec_xl(16, ptr);
res.val[2] = vec_xl(32, ptr);
res.val[3] = vec_xl(48, ptr);
return res;
}
typedef struct ggml_int8x16x4_t {
int8x16_t val[4];
} ggml_int8x16x4_t;
inline static ggml_int8x16x4_t ggml_vec_xl_s8x4(const int8_t * ptr) {
ggml_int8x16x4_t res;
res.val[0] = vec_xl( 0, ptr);
res.val[1] = vec_xl(16, ptr);
res.val[2] = vec_xl(32, ptr);
res.val[3] = vec_xl(48, ptr);
return res;
}
typedef struct ggml_int16x8x2_t {
int16x8_t val[2];
} ggml_int16x8x2_t;
inline static ggml_int16x8x2_t ggml_vec_xl_s16x2(const int16_t * ptr) {
ggml_int16x8x2_t res;
res.val[0] = vec_xl( 0, ptr);
res.val[1] = vec_xl(16, ptr);
return res;
}
/*
! WARNING: Very slow. Use vec_perm if possible. Refer to iq4_xs
! or iq4_nl for example implementation.
*/
inline static int8x16_t ggml_vec_tbl(int8x16_t a, uint8x16_t b) {
int8x16_t res;
res[ 0] = a[b[ 0]];
res[ 1] = a[b[ 1]];
res[ 2] = a[b[ 2]];
res[ 3] = a[b[ 3]];
res[ 4] = a[b[ 4]];
res[ 5] = a[b[ 5]];
res[ 6] = a[b[ 6]];
res[ 7] = a[b[ 7]];
res[ 8] = a[b[ 8]];
res[ 9] = a[b[ 9]];
res[10] = a[b[10]];
res[11] = a[b[11]];
res[12] = a[b[12]];
res[13] = a[b[13]];
res[14] = a[b[14]];
res[15] = a[b[15]];
return res;
}
inline static int16x8_t vec_padd_s16(int16x8_t a, int16x8_t b) {
const uchar8x16_t v_maske = { 0, 1, 4, 5, 8, 9, 12, 13,
16, 17, 20, 21, 24, 25, 28, 29 };
const int16x8_t v_abo = vec_pack((int32x4_t)a, (int32x4_t)b);
const int16x8_t v_abe = vec_perm(a, b, v_maske);
return v_abo + v_abe;
}
inline static int32x4_t ggml_vec_dot(int32x4_t acc, int8x16_t a, int8x16_t b) {
const int16x8_t p = vec_mule(a, b) + vec_mulo(a, b);
return acc + (vec_unpackh(p) + vec_unpackl(p));
}
#endif
#if defined(__loongarch_asx)
/* float type data load instructions */
static __m128 __lsx_vreplfr2vr_s(const float val) {
v4f32 res = {val, val, val, val};
return (__m128)res;
}
static __m256 __lasx_xvreplfr2vr_s(const float val) {
v8f32 res = {val, val, val, val, val, val, val, val};
return (__m256)res;
}
#endif
// TODO: move to ggml-threading
void ggml_barrier(struct ggml_threadpool * tp);
void ggml_threadpool_chunk_set(struct ggml_threadpool * tp, int value);
int ggml_threadpool_chunk_add(struct ggml_threadpool * tp, int value);
#ifdef __cplusplus
}
#endif
ggml/src/ggml-cpu/ggml-cpu.c
+3553
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/ggml-cpu.cpp
+675
View File
@@ -0,0 +1,675 @@
#include "ggml-backend.h"
#include "ggml-backend-impl.h"
#include "ggml-cpu.h"
#include "repack.h"
#include "traits.h"
#include "ggml-impl.h"
#include "amx/amx.h"
#include <cctype>
#include <string>
#include <vector>
#ifdef GGML_USE_CPU_HBM
# include "hbm.h"
#endif
#ifdef GGML_USE_CPU_KLEIDIAI
# include "kleidiai/kleidiai.h"
#endif
#if defined(_WIN32)
# define WIN32_LEAN_AND_MEAN
# ifndef NOMINMAX
# define NOMINMAX
# endif
# include <windows.h>
#else
# include <unistd.h>
#endif
#if defined(__APPLE__)
# include <sys/sysctl.h>
# include <sys/types.h>
#endif
// ggml-backend interface
std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type() {
static std::vector<ggml_backend_buffer_type_t> bufts = []() {
std::vector<ggml_backend_buffer_type_t> bufts;
#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
if (ggml_backend_amx_buffer_type()) {
bufts.push_back(ggml_backend_amx_buffer_type());
}
#endif
#ifdef GGML_USE_CPU_KLEIDIAI
if (ggml_backend_cpu_kleidiai_buffer_type()) {
bufts.push_back(ggml_backend_cpu_kleidiai_buffer_type());
}
#endif
#ifdef GGML_USE_CPU_REPACK
if (ggml_backend_cpu_repack_buffer_type()) {
bufts.push_back(ggml_backend_cpu_repack_buffer_type());
}
#endif
bufts.push_back(NULL);
return bufts;
}();
return bufts;
}
static ggml_backend_buffer_type_t * ggml_backend_cpu_device_get_extra_buffers_type(ggml_backend_dev_t device) {
return ggml_backend_cpu_get_extra_buffers_type().data();
GGML_UNUSED(device);
}
static bool ggml_backend_cpu_is_extra_buffer_type(ggml_backend_buffer_type_t buft) {
for (auto * extra : ggml_backend_cpu_get_extra_buffers_type()) {
if (extra && extra == buft) {
return true;
}
}
return false;
}
// CPU backend - backend (stream)
struct ggml_backend_cpu_context {
int n_threads;
ggml_threadpool_t threadpool;
uint8_t * work_data;
size_t work_size;
ggml_abort_callback abort_callback;
void * abort_callback_data;
};
static const char * ggml_backend_cpu_get_name(ggml_backend_t backend) {
return "CPU";
GGML_UNUSED(backend);
}
static void ggml_backend_cpu_free(ggml_backend_t backend) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
delete[] cpu_ctx->work_data;
delete cpu_ctx;
delete backend;
}
struct ggml_backend_plan_cpu {
struct ggml_cplan cplan;
struct ggml_cgraph cgraph;
};
static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
struct ggml_backend_plan_cpu * cpu_plan = new ggml_backend_plan_cpu;
cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
cpu_plan->cgraph = *cgraph; // FIXME: deep copy
if (cpu_plan->cplan.work_size > 0) {
cpu_plan->cplan.work_data = new uint8_t[cpu_plan->cplan.work_size];
if (cpu_plan->cplan.work_data == NULL) {
delete cpu_plan;
return NULL;
}
}
cpu_plan->cplan.abort_callback = cpu_ctx->abort_callback;
cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
return cpu_plan;
}
static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
delete[] cpu_plan->cplan.work_data;
delete cpu_plan;
GGML_UNUSED(backend);
}
static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
GGML_UNUSED(backend);
}
static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
if (cpu_ctx->work_size < cplan.work_size) {
delete[] cpu_ctx->work_data;
cpu_ctx->work_data = new uint8_t[cplan.work_size];
if (cpu_ctx->work_data == NULL) {
cpu_ctx->work_size = 0;
return GGML_STATUS_ALLOC_FAILED;
}
cpu_ctx->work_size = cplan.work_size;
}
cplan.work_data = (uint8_t *)cpu_ctx->work_data;
cplan.abort_callback = cpu_ctx->abort_callback;
cplan.abort_callback_data = cpu_ctx->abort_callback_data;
return ggml_graph_compute(cgraph, &cplan);
}
static const struct ggml_backend_i ggml_backend_cpu_i = {
/* .get_name = */ ggml_backend_cpu_get_name,
/* .free = */ ggml_backend_cpu_free,
/* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL,
/* .cpy_tensor_async = */ NULL,
/* .synchronize = */ NULL,
/* .graph_plan_create = */ ggml_backend_cpu_graph_plan_create,
/* .graph_plan_free = */ ggml_backend_cpu_graph_plan_free,
/* .graph_plan_update = */ NULL,
/* .graph_plan_compute = */ ggml_backend_cpu_graph_plan_compute,
/* .graph_compute = */ ggml_backend_cpu_graph_compute,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
};
static ggml_guid_t ggml_backend_cpu_guid(void) {
static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 };
return &guid;
}
ggml_backend_t ggml_backend_cpu_init(void) {
// initialize CPU backend now to avoid slowing the first graph computation
ggml_cpu_init();
struct ggml_backend_cpu_context * ctx = new ggml_backend_cpu_context;
if (ctx == NULL) {
return NULL;
}
ctx->n_threads = GGML_DEFAULT_N_THREADS;
ctx->threadpool = NULL;
ctx->work_data = NULL;
ctx->work_size = 0;
ctx->abort_callback = NULL;
ctx->abort_callback_data = NULL;
ggml_backend_t cpu_backend = new ggml_backend {
/* .guid = */ ggml_backend_cpu_guid(),
/* .interface = */ ggml_backend_cpu_i,
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ ctx,
};
if (cpu_backend == NULL) {
delete ctx;
return NULL;
}
return cpu_backend;
}
bool ggml_backend_is_cpu(ggml_backend_t backend) {
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid());
}
void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
ctx->n_threads = n_threads;
}
void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
if (ctx->threadpool && ctx->threadpool != threadpool) {
// already had a different threadpool, pause/suspend it before switching
ggml_threadpool_pause(ctx->threadpool);
}
ctx->threadpool = threadpool;
}
void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
ctx->abort_callback = abort_callback;
ctx->abort_callback_data = abort_callback_data;
}
// CPU backend - device
struct ggml_backend_cpu_device_context {
std::string description = "CPU";
ggml_backend_cpu_device_context() {
#ifdef __APPLE__
size_t len = 0;
if (!sysctlbyname("machdep.cpu.brand_string", NULL, &len, NULL, 0)) {
description.resize(len);
sysctlbyname("machdep.cpu.brand_string", &description[0], &len, NULL, 0); // NOLINT
}
#elif defined(__linux__)
FILE * f = fopen("/proc/cpuinfo", "r");
if (f) {
char buf[1024];
while (fgets(buf, sizeof(buf), f)) {
if (strncmp(buf, "model name", 10) == 0) {
char * p = strchr(buf, ':');
if (p) {
p++;
while (std::isspace(*p)) {
p++;
}
while (std::isspace(p[strlen(p) - 1])) {
p[strlen(p) - 1] = '\0';
}
description = p;
break;
}
}
}
fclose(f);
}
#elif defined(_WIN32)
HKEY hKey;
if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
0,
KEY_READ,
&hKey) == ERROR_SUCCESS) {
DWORD cpu_brand_size = 0;
if (RegQueryValueExA(hKey,
"ProcessorNameString",
NULL,
NULL,
NULL,
&cpu_brand_size) == ERROR_SUCCESS) {
description.resize(cpu_brand_size);
if (RegQueryValueExA(hKey,
"ProcessorNameString",
NULL,
NULL,
(LPBYTE)&description[0], // NOLINT
&cpu_brand_size) == ERROR_SUCCESS) {
if (description.find('\0') != std::string::npos) {
description.resize(description.find('\0'));
}
}
}
RegCloseKey(hKey);
}
#endif
}
};
static const char * ggml_backend_cpu_device_get_name(ggml_backend_dev_t dev) {
return "CPU";
GGML_UNUSED(dev);
}
static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t dev) {
struct ggml_backend_cpu_device_context * ctx = (struct ggml_backend_cpu_device_context *)dev->context;
return ctx->description.c_str();
}
static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
#ifdef _WIN32
MEMORYSTATUSEX status;
status.dwLength = sizeof(status);
GlobalMemoryStatusEx(&status);
*total = status.ullTotalPhys;
*free = status.ullAvailPhys;
#else
long pages = sysconf(_SC_PHYS_PAGES);
long page_size = sysconf(_SC_PAGE_SIZE);
*total = pages * page_size;
*free = *total;
#endif
GGML_UNUSED(dev);
}
static enum ggml_backend_dev_type ggml_backend_cpu_device_get_type(ggml_backend_dev_t dev) {
return GGML_BACKEND_DEVICE_TYPE_CPU;
GGML_UNUSED(dev);
}
static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
props->name = ggml_backend_cpu_device_get_name(dev);
props->description = ggml_backend_cpu_device_get_description(dev);
props->type = ggml_backend_cpu_device_get_type(dev);
ggml_backend_cpu_device_get_memory(dev, &props->memory_free, &props->memory_total);
props->caps = {
/* .async = */ false,
/* .host_buffer = */ false,
/* .buffer_from_host_ptr = */ true,
/* .events = */ false,
};
}
static ggml_backend_t ggml_backend_cpu_device_init_backend(ggml_backend_dev_t dev, const char * params) {
return ggml_backend_cpu_init();
GGML_UNUSED(dev);
GGML_UNUSED(params);
}
static ggml_backend_buffer_type_t ggml_backend_cpu_device_get_buffer_type(ggml_backend_dev_t dev) {
return ggml_backend_cpu_buffer_type();
GGML_UNUSED(dev);
}
static ggml_backend_buffer_t ggml_backend_cpu_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
return ggml_backend_cpu_buffer_from_ptr(ptr, size);
GGML_UNUSED(dev);
GGML_UNUSED(max_tensor_size);
}
static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
const struct ggml_tensor * src0 = op->src[0];
const struct ggml_tensor * src1 = op->src[1];
if (op->op == GGML_OP_NONE || op->op == GGML_OP_RESHAPE || op->op == GGML_OP_VIEW || op->op == GGML_OP_PERMUTE || op->op == GGML_OP_TRANSPOSE) {
return true;
}
// extra_buffer_op?
for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
if (extra) {
auto buf_extra = (ggml::cpu::extra_buffer_type*) extra->context;
if (buf_extra && buf_extra->supports_op(dev, op)) {
return true;
}
}
}
// the other case need host buffer.
for (int i = 0; i < GGML_MAX_SRC; i++) {
if (op->src[i] && op->src[i]->buffer && !ggml_backend_buft_is_host(op->src[i]->buffer->buft)) {
return false;
}
}
switch (op->op) {
case GGML_OP_CPY:
case GGML_OP_SET_ROWS:
return
op->type != GGML_TYPE_IQ3_XXS &&
op->type != GGML_TYPE_IQ3_S &&
op->type != GGML_TYPE_IQ2_XXS &&
op->type != GGML_TYPE_IQ2_XS &&
op->type != GGML_TYPE_IQ2_S &&
op->type != GGML_TYPE_IQ1_S &&
op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
case GGML_OP_MUL_MAT:
return src1->type == GGML_TYPE_F32 || src1->type == ggml_get_type_traits_cpu(src0->type)->vec_dot_type;
case GGML_OP_SOFT_MAX_BACK: {
if (op->src[0]->type != GGML_TYPE_F32 || op->src[1]->type != GGML_TYPE_F32) {
return false;
}
float max_bias = 0.0f;
memcpy(&max_bias, (const float *) op->op_params + 1, sizeof(float));
return max_bias == 0.0f;
}
case GGML_OP_IM2COL_BACK:
return src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32;
case GGML_OP_GET_ROWS_BACK:
return src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16;
case GGML_OP_OUT_PROD:
return (src0->type == GGML_TYPE_F32 || (ggml_is_quantized(src0->type) && src0->ne[2] == src1->ne[2] && src0->ne[3] == src1->ne[3])) &&
src1->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
default:
return true;
}
}
static bool ggml_backend_cpu_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
return ggml_backend_buft_is_host(buft) || ggml_backend_cpu_is_extra_buffer_type(buft);
GGML_UNUSED(dev);
}
static const struct ggml_backend_device_i ggml_backend_cpu_device_i = {
/* .get_name = */ ggml_backend_cpu_device_get_name,
/* .get_description = */ ggml_backend_cpu_device_get_description,
/* .get_memory = */ ggml_backend_cpu_device_get_memory,
/* .get_type = */ ggml_backend_cpu_device_get_type,
/* .get_props = */ ggml_backend_cpu_device_get_props,
/* .init_backend = */ ggml_backend_cpu_device_init_backend,
/* .get_buffer_type = */ ggml_backend_cpu_device_get_buffer_type,
/* .get_host_buffer_type = */ NULL,
/* .buffer_from_host_ptr = */ ggml_backend_cpu_device_buffer_from_host_ptr,
/* .supports_op = */ ggml_backend_cpu_device_supports_op,
/* .supports_buft = */ ggml_backend_cpu_device_supports_buft,
/* .offload_op = */ NULL,
/* .event_new = */ NULL,
/* .event_free = */ NULL,
/* .event_synchronize = */ NULL,
};
// CPU backend - backend (reg)
static const char * ggml_backend_cpu_reg_get_name(ggml_backend_reg_t reg) {
return "CPU";
GGML_UNUSED(reg);
}
static size_t ggml_backend_cpu_reg_get_device_count(ggml_backend_reg_t reg) {
return 1;
GGML_UNUSED(reg);
}
static ggml_backend_dev_t ggml_backend_cpu_reg_get_device(ggml_backend_reg_t reg, size_t index) {
GGML_ASSERT(index == 0);
static ggml_backend_cpu_device_context ctx;
static ggml_backend_device ggml_backend_cpu_device = {
/* .iface = */ ggml_backend_cpu_device_i,
/* .reg = */ reg,
/* .context = */ &ctx,
};
return &ggml_backend_cpu_device;
}
// This is intended to replace the the ggml_cpu_has_* functions when loading the CPU backend dynamically,
// and additionally to allow other backends to expose their own list of features that applications can query using the same API
static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t reg) {
static std::vector<ggml_backend_feature> features = []() {
ggml_cpu_init();
std::vector<ggml_backend_feature> features;
if (ggml_cpu_has_sse3()) {
features.push_back({ "SSE3", "1" });
}
if (ggml_cpu_has_ssse3()) {
features.push_back({ "SSSE3", "1" });
}
if (ggml_cpu_has_avx()) {
features.push_back({ "AVX", "1" });
}
if (ggml_cpu_has_avx_vnni()) {
features.push_back({ "AVX_VNNI", "1" });
}
if (ggml_cpu_has_avx2()) {
features.push_back({ "AVX2", "1" });
}
if (ggml_cpu_has_f16c()) {
features.push_back({ "F16C", "1" });
}
if (ggml_cpu_has_fma()) {
features.push_back({ "FMA", "1" });
}
if (ggml_cpu_has_bmi2()) {
features.push_back({ "BMI2", "1" });
}
if (ggml_cpu_has_avx512()) {
features.push_back({ "AVX512", "1" });
}
if (ggml_cpu_has_avx512_vbmi()) {
features.push_back({ "AVX512_VBMI", "1" });
}
if (ggml_cpu_has_avx512_vnni()) {
features.push_back({ "AVX512_VNNI", "1" });
}
if (ggml_cpu_has_avx512_bf16()) {
features.push_back({ "AVX512_BF16", "1" });
}
if (ggml_cpu_has_amx_int8()) {
features.push_back({ "AMX_INT8", "1" });
}
if (ggml_cpu_has_neon()) {
features.push_back({ "NEON", "1" });
}
if (ggml_cpu_has_arm_fma()) {
features.push_back({ "ARM_FMA", "1" });
}
if (ggml_cpu_has_fp16_va()) {
features.push_back({ "FP16_VA", "1" });
}
if (ggml_cpu_has_matmul_int8()) {
features.push_back({ "MATMUL_INT8", "1" });
}
if (ggml_cpu_has_sve()) {
features.push_back({ "SVE", "1" });
}
if (ggml_cpu_has_dotprod()) {
features.push_back({ "DOTPROD", "1" });
}
if (ggml_cpu_get_sve_cnt() > 0) {
static std::string sve_cnt = std::to_string(ggml_cpu_get_sve_cnt());
features.push_back({ "SVE_CNT", sve_cnt.c_str() });
}
if (ggml_cpu_has_sme()) {
features.push_back({ "SME", "1" });
}
if (ggml_cpu_has_riscv_v()) {
features.push_back({ "RISCV_V", "1" });
}
if (ggml_cpu_has_vsx()) {
features.push_back({ "VSX", "1" });
}
if (ggml_cpu_has_vxe()) {
features.push_back({ "VXE", "1" });
}
if (ggml_cpu_has_nnpa()) {
features.push_back({ "NNPA", "1" });
}
if (ggml_cpu_has_wasm_simd()) {
features.push_back({ "WASM_SIMD", "1" });
}
if (ggml_cpu_has_llamafile()) {
features.push_back({ "LLAMAFILE", "1" });
}
#ifdef GGML_USE_ACCELERATE
features.push_back({ "ACCELERATE", "1" });
#endif
#ifdef GGML_USE_CPU_HBM
features.push_back({ "CPU_HBM", "1" });
#endif
#ifdef GGML_USE_OPENMP
features.push_back({ "OPENMP", "1" });
#endif
#ifdef GGML_USE_CPU_KLEIDIAI
features.push_back({ "KLEIDIAI", "1" });
#endif
#ifdef GGML_USE_CPU_REPACK
features.push_back({ "REPACK", "1" });
#endif
features.push_back({ nullptr, nullptr });
return features;
}();
return features.data();
GGML_UNUSED(reg);
}
static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const char * name) {
if (strcmp(name, "ggml_backend_set_n_threads") == 0) {
ggml_backend_set_n_threads_t fct = ggml_backend_cpu_set_n_threads;
return (void *)fct;
}
if (strcmp(name, "ggml_backend_dev_get_extra_bufts") == 0) {
ggml_backend_dev_get_extra_bufts_t fct = ggml_backend_cpu_device_get_extra_buffers_type;
return (void *)fct;
}
if (strcmp(name, "ggml_backend_get_features") == 0) {
return (void *)ggml_backend_cpu_get_features;
}
if (strcmp(name, "ggml_backend_set_abort_callback") == 0) {
return (void *)ggml_backend_cpu_set_abort_callback;
}
if (strcmp(name, "ggml_backend_cpu_numa_init") == 0) {
return (void *)ggml_numa_init;
}
if (strcmp(name, "ggml_backend_cpu_is_numa") == 0) {
return (void *)ggml_is_numa;
}
// threadpool - TODO: move to ggml-base
if (strcmp(name, "ggml_threadpool_new") == 0) {
return (void *)ggml_threadpool_new;
}
if (strcmp(name, "ggml_threadpool_free") == 0) {
return (void *)ggml_threadpool_free;
}
if (strcmp(name, "ggml_backend_cpu_set_threadpool") == 0) {
return (void *)ggml_backend_cpu_set_threadpool;
}
return NULL;
GGML_UNUSED(reg);
}
static const struct ggml_backend_reg_i ggml_backend_cpu_reg_i = {
/* .get_name = */ ggml_backend_cpu_reg_get_name,
/* .get_device_count = */ ggml_backend_cpu_reg_get_device_count,
/* .get_device = */ ggml_backend_cpu_reg_get_device,
/* .get_proc_address = */ ggml_backend_cpu_get_proc_address,
};
ggml_backend_reg_t ggml_backend_cpu_reg(void) {
// init CPU feature detection
ggml_cpu_init();
static struct ggml_backend_reg ggml_backend_cpu_reg = {
/* .api_version = */ GGML_BACKEND_API_VERSION,
/* .iface = */ ggml_backend_cpu_reg_i,
/* .context = */ NULL,
};
return &ggml_backend_cpu_reg;
}
GGML_BACKEND_DL_IMPL(ggml_backend_cpu_reg)
ggml/src/ggml-cpu/hbm.cpp
+55
View File
@@ -0,0 +1,55 @@
#ifdef GGML_USE_CPU_HBM
#include "ggml-backend.h"
#include "ggml-backend-impl.h"
#include "ggml-cpu.h"
#include "ggml-impl.h"
#include "hbm.h"
// buffer type HBM
#include <hbwmalloc.h>
static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU_HBM";
GGML_UNUSED(buft);
}
static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
hbw_free(buffer->context);
}
static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
size_t size) {
void * ptr;
int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
if (result != 0) {
GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size);
return NULL;
}
ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
buffer->buft = buft;
buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
return buffer;
}
ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cpu_hbm_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_buffer_type_get_alignment,
/* .get_max_size = */ nullptr, // defaults to SIZE_MAX
/* .get_alloc_size = */ nullptr, // defaults to ggml_nbytes
/* .is_host = */ ggml_backend_cpu_buffer_type_is_host,
},
/* .context = */ nullptr,
};
return &ggml_backend_cpu_buffer_type_hbm;
}
#endif
ggml/src/ggml-cpu/hbm.h
+8
View File
@@ -0,0 +1,8 @@
#pragma once
#include "ggml-backend.h"
#include "ggml.h"
// GGML CPU internal header
ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
ggml/src/ggml-cpu/kleidiai/kernels.cpp
+434
View File
@@ -0,0 +1,434 @@
// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
// SPDX-License-Identifier: MIT
//
// KleidiAI micro-kernels
#include "kai_matmul_clamp_f32_qsi8d32p_qsi4c32p_interface.h"
#include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.h"
#include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.h"
#include "kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.h"
#include "kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.h"
#include "kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.h"
#include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.h"
#include "kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.h"
#include "kai_lhs_pack_bf16p2vlx2_f32_sme.h"
#include "kai_lhs_quant_pack_qsi8d32p_f32.h"
#include "kai_lhs_quant_pack_qsi8d32p_f32_neon.h"
#include "kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.h"
#include "kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h"
#include "kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h"
#include "kai_common.h"
#include "simd-mappings.h"
#include "kernels.h"
#define NELEMS(x) sizeof(x) / sizeof(*x)
static const size_t INT4_PER_BYTE = 2;
static const size_t INT4_BITS = 4;
static const int Q4_0_ZERO_POINT = 8;
const size_t INT4_PER_UINT16 = 4;
static void dequantize_row_qsi4c32pscalef16(
const void *packed_data,
int32_t row_idx,
int64_t nc,
float *out,
size_t nr_pack,
size_t packed_row_stride,
size_t kr,
size_t bl,
size_t num_bytes_multiplier
) {
size_t group_idx = row_idx / nr_pack;
size_t row_in_group = row_idx % nr_pack;
const uint8_t *packed_group = (const uint8_t *)packed_data + group_idx * packed_row_stride;
size_t num_blocks = nc / bl;
const uint8_t *block_ptr = packed_group;
for (size_t b = 0; b < num_blocks; ++b) {
uint16_t scale_f16 = *((const uint16_t *)(block_ptr + row_in_group * num_bytes_multiplier));
float scale = GGML_CPU_FP16_TO_FP32(scale_f16);
const uint8_t *segment_ptr = block_ptr + nr_pack * num_bytes_multiplier;
size_t num_segments = bl / kr;
size_t num_bytes_per_segment = kr / INT4_PER_BYTE;
for (size_t s = 0; s < num_segments; ++s) {
const uint8_t *seg_base = segment_ptr + s * nr_pack * num_bytes_per_segment;
const uint8_t *qbytes = seg_base + row_in_group * num_bytes_per_segment;
for (size_t k = 0; k < num_bytes_per_segment; ++k) {
uint8_t byte = qbytes[k] ^ 0x88;
int x0 = (byte & 0x0F) - Q4_0_ZERO_POINT;
int x1 = (byte >> INT4_BITS) - Q4_0_ZERO_POINT;
out[b * bl + s * num_bytes_per_segment + k] = x0 * scale;
out[b * bl + s * num_bytes_per_segment + k + bl/2] = x1 * scale;
}
}
block_ptr += nr_pack * num_bytes_multiplier + num_segments * nr_pack * num_bytes_per_segment;
}
}
static void dequantize_row_qsi4c32ps1s0scalef16(
const void *packed_data,
int32_t row_idx,
int64_t k,
float *out,
size_t nr,
size_t packed_row_stride,
size_t kr,
size_t bl,
size_t num_bytes_multiplier
) {
const size_t num_blocks = k / bl;
const size_t bl4 = bl / INT4_PER_UINT16;
size_t group_idx = row_idx / nr;
size_t row_in_group = row_idx % nr;
const uint8_t *packed_group = (const uint8_t *)packed_data + group_idx * packed_row_stride;
const uint16_t *qdata = (const uint16_t *)packed_group;
const uint16_t *scales = (const uint16_t *)(packed_group + packed_row_stride - (nr * num_blocks * num_bytes_multiplier));
for (size_t block_idx = 0; block_idx < num_blocks; ++block_idx) {
uint16_t scale_f16 = scales[row_in_group + block_idx * nr];
float scale = GGML_CPU_FP16_TO_FP32(scale_f16);
for (size_t bl4_idx = 0; bl4_idx < bl4; ++bl4_idx) {
uint16_t q = qdata[(block_idx * bl4 + bl4_idx) * nr + row_in_group];
for (size_t qidx = 0; qidx < INT4_PER_UINT16; ++qidx) {
int v = ((q >> (qidx * 4)) & 0xF) - Q4_0_ZERO_POINT;
out[block_idx * bl + bl4_idx * INT4_BITS + qidx] = v * scale;
}
}
}
GGML_UNUSED(kr);
}
static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
#if defined(__ARM_FEATURE_SME)
{
/* SME GEMM */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
},
/* SME GEMV */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
},
/* .lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32_neon,
/* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32_neon,
/* .packed_size = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32_neon,
/* .pack_func = */ kai_run_lhs_quant_pack_qsi8d32p_f32_neon,
},
/* .rhs_info = */ {
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
/* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
/* .to_float = */ dequantize_row_qsi4c32ps1s0scalef16,
},
/* .required_cpu = */ CPU_FEATURE_SME,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_Q4_0,
/* .op_type = */ GGML_TYPE_F32,
},
{
/* SME GEMM */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .run_kernel = */ kai_run_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
},
/* SME GEMV */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .run_kernel = */ kai_run_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
},
/* .lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_pack_bf16p2vlx2_f32_sme,
/* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_pack_bf16p2vlx2_f32_sme,
/* .packed_size = */ kai_get_lhs_packed_size_lhs_pack_bf16p2vlx2_f32_sme,
/* .pack_func = */ kai_run_lhs_pack_bf16p2vlx2_f32_sme,
},
/* .rhs_info = */ {
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
/* .packed_stride = */ NULL,
/* .pack_func = */ kai_run_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
/* .to_float = */ NULL,
},
/* .required_cpu = */ CPU_FEATURE_SME,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_F16,
/* .op_type = */ GGML_TYPE_F32,
},
#endif
#if defined(__APPLE__)
#if defined(__ARM_FEATURE_DOTPROD)
{
/* DOTPROD GEMM */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
},
/* DOTPROD GEMV */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
},
/* .lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
/* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
/* .packed_size = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
/* .pack_func = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
},
/* .rhs_info = */ {
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .to_float = */ dequantize_row_qsi4c32pscalef16,
},
/* .required_cpu = */ CPU_FEATURE_DOTPROD,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_Q4_0,
/* .op_type = */ GGML_TYPE_F32,
},
#endif
#if defined(__ARM_FEATURE_MATMUL_INT8)
{
/* i8mm GEMM */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
},
/* i8mm GEMV */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
},
/* .lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
/* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
/* .packed_size = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
/* .pack_func = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
},
/* .rhs_info = */ {
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .to_float = */ dequantize_row_qsi4c32pscalef16,
},
/* .required_cpu = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_Q4_0,
/* .op_type = */ GGML_TYPE_F32,
},
#endif
#else
#if defined(__ARM_FEATURE_MATMUL_INT8)
{
/* i8mm GEMM */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
},
/* i8mm GEMV */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
},
/* .lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
/* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
/* .packed_size = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
/* .pack_func = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
},
/* .rhs_info = */ {
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .to_float = */ dequantize_row_qsi4c32pscalef16,
},
/* .required_cpu = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_Q4_0,
/* .op_type = */ GGML_TYPE_F32,
},
#endif
#if defined(__ARM_FEATURE_DOTPROD)
{
/* DOTPROD GEMM */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
},
/* DOTPROD GEMV */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
},
/* .lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
/* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
/* .packed_size = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
/* .pack_func = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
},
/* .rhs_info = */ {
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .to_float = */ dequantize_row_qsi4c32pscalef16,
},
/* .required_cpu = */ CPU_FEATURE_DOTPROD,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_Q4_0,
/* .op_type = */ GGML_TYPE_F32,
},
#endif
#endif
};
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor) {
ggml_kleidiai_kernels * kernel = nullptr;
if (tensor->op == GGML_OP_MUL_MAT && tensor->src[0] != nullptr && tensor->src[1] != nullptr) {
for (size_t i = 0; i < NELEMS(gemm_gemv_kernels); ++i) {
if ((cpu_features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu &&
gemm_gemv_kernels[i].lhs_type == tensor->src[1]->type &&
gemm_gemv_kernels[i].rhs_type == tensor->src[0]->type &&
gemm_gemv_kernels[i].op_type == tensor->type) {
kernel = &gemm_gemv_kernels[i];
break;
}
}
}
return kernel;
}
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features) {
ggml_kleidiai_kernels * kernels = nullptr;
for (size_t i = 0; i < NELEMS(gemm_gemv_kernels); ++i) {
if ((features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu) {
kernels = &gemm_gemv_kernels[i];
break;
}
}
return kernels;
}
ggml/src/ggml-cpu/kleidiai/kernels.h
+98
View File
@@ -0,0 +1,98 @@
// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
// SPDX-License-Identifier: MIT
//
#pragma once
#include <functional>
#include <variant>
#include "ggml.h"
enum cpu_feature {
CPU_FEATURE_NONE = 0,
CPU_FEATURE_DOTPROD = 1,
CPU_FEATURE_I8MM = 2,
CPU_FEATURE_SVE = 4,
CPU_FEATURE_SME = 8
};
inline cpu_feature& operator|=(cpu_feature& lhs, cpu_feature rhs) {
lhs = static_cast<cpu_feature>(lhs | rhs);
return lhs;
}
inline cpu_feature operator|(cpu_feature lhs, cpu_feature rhs) {
return static_cast<cpu_feature>(static_cast<int>(lhs) | static_cast<int>(rhs));
}
struct kernel_info {
size_t (*get_m_step)(void);
size_t (*get_n_step)(void);
size_t (*get_mr)(void);
size_t (*get_nr)(void);
size_t (*get_kr)(void);
size_t (*get_sr)(void);
std::variant<
std::function<size_t(size_t n_idx, size_t k, size_t bl)>,
std::function<size_t(size_t m_idx, size_t k)>
> get_lhs_offset;
std::variant<
std::function<size_t(size_t n_idx, size_t k, size_t bl)>,
std::function<size_t(size_t n_idx, size_t k)>
> get_rhs_packed_offset;
size_t (*get_dst_offset)(size_t m_idx, size_t n_idx, size_t stride);
size_t (*get_dst_size)(size_t m, size_t n);
std::variant<
std::function<void(size_t m, size_t n, size_t k, size_t bl, const void* lhs_packed, const void* rhs_packed,
float* dst, size_t dst_stride_row, size_t dst_stride_col, float scalar_min, float scalar_max)>,
std::function<void(size_t m, size_t n, size_t k, const void* lhs_packed, const void* rhs_packed, void* dst, size_t dst_stride_row,
size_t dst_stride_col, float clamp_min, float clamp_max)>
> run_kernel;
};
struct lhs_packing_info {
size_t (*get_offset)(size_t m_idx, size_t lhs_stride);
std::variant<
std::function<size_t(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr)>,
std::function<size_t(size_t m_idx, size_t k, size_t mr, size_t kr, size_t sr)>
> get_packed_offset;
std::variant<
std::function<size_t(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr)>,
std::function<size_t(size_t m, size_t k, size_t mr, size_t kr, size_t sr)>
> packed_size;
std::variant<
std::function<void(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const float* lhs,
size_t lhs_stride, void* lhs_packed)>,
std::function<void(size_t m, size_t k, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const void* lhs, size_t lhs_stride,
void* lhs_packed)>
> pack_func;
};
struct rhs_packing_info {
std::variant<
std::function<size_t(size_t n, size_t k, size_t nr, size_t kr, size_t bl)>,
std::function<size_t(size_t n, size_t k)>
> packed_size;
size_t (*packed_stride)(size_t k, size_t nr, size_t kr, size_t bl);
std::variant<
std::function<void(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl, const uint8_t* rhs,
const float* bias, void* rhs_packed, size_t extra_bytes, const struct kai_rhs_pack_qs4cxs1s0_param* params)>,
std::function<void(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t rhs_stride, const void* rhs,
const void* bias, const void* scale, void* rhs_packed, size_t extra_bytes, const void* params)>
> pack_func;
void (*to_float)(const void *packed_data, int32_t row_idx, int64_t nc, float *out, size_t nr_pack, size_t packed_row_stride,
size_t kr, size_t bl, size_t num_bytes_multiplier);
};
struct ggml_kleidiai_kernels {
kernel_info gemm;
kernel_info gemv;
lhs_packing_info lhs_info;
rhs_packing_info rhs_info;
cpu_feature required_cpu;
ggml_type lhs_type;
ggml_type rhs_type;
ggml_type op_type;
};
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor);
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features);
ggml/src/ggml-cpu/kleidiai/kleidiai.cpp
+560
View File
@@ -0,0 +1,560 @@
// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
// SPDX-License-Identifier: MIT
//
#include <arm_neon.h>
#include <assert.h>
#include <atomic>
#include <cfloat>
#include <stdexcept>
#include <stdint.h>
#include <string.h>
#if defined(__linux__)
#include <asm/hwcap.h>
#include <sys/auxv.h>
#elif defined(__APPLE__)
#include <string_view>
#include <sys/sysctl.h>
#include <sys/types.h>
#elif defined(_WIN32)
#include <windows.h>
#include <excpt.h>
#endif
#include "kleidiai.h"
#include "ggml-cpu.h"
#include "ggml-impl.h"
#include "ggml-backend-impl.h"
#include "ggml-threading.h"
#include "traits.h"
#include "kernels.h"
#include "kai_common.h"
#define GGML_COMMON_DECL_CPP
#include "ggml-common.h"
struct ggml_kleidiai_context {
cpu_feature features;
ggml_kleidiai_kernels * kernels;
} static ctx = { CPU_FEATURE_NONE, NULL };
static const char* cpu_feature_to_string(cpu_feature f) {
switch (f) {
case CPU_FEATURE_NONE: return "NONE";
case CPU_FEATURE_DOTPROD: return "DOTPROD";
case CPU_FEATURE_I8MM: return "I8MM";
case CPU_FEATURE_SVE: return "SVE";
case CPU_FEATURE_SME: return "SME";
default: return "UNKNOWN";
}
}
static void init_kleidiai_context(void) {
ggml_critical_section_start();
static bool initialized = false;
if (!initialized) {
initialized = true;
const char *env_var = getenv("GGML_KLEIDIAI_SME");
int sme_enabled = 0;
ctx.features = (ggml_cpu_has_dotprod() ? CPU_FEATURE_DOTPROD : CPU_FEATURE_NONE) |
(ggml_cpu_has_matmul_int8() ? CPU_FEATURE_I8MM : CPU_FEATURE_NONE) |
(ggml_cpu_has_sve() ? CPU_FEATURE_SVE : CPU_FEATURE_NONE);
if (env_var) {
sme_enabled = atoi(env_var);
}
if (sme_enabled != 0) {
ctx.features |= ggml_cpu_has_sme() ? CPU_FEATURE_SME : CPU_FEATURE_NONE;
}
ctx.kernels = ggml_kleidiai_select_kernels_q4_0(ctx.features);
#ifndef NDEBUG
if (ctx.kernels) {
GGML_LOG_DEBUG("kleidiai: using kernel with CPU feature %s\n", cpu_feature_to_string(ctx.kernels->required_cpu));
}
#endif
}
ggml_critical_section_end();
}
static inline int64_t ggml_ne(const ggml_tensor * tensor, int dim) {
GGML_ASSERT(dim >= 0 && dim < GGML_MAX_DIMS);
return tensor->ne[dim];
}
template<typename Ret, typename Variant, typename... Args>
static Ret variant_call(const Variant & var, Args&&... args) {
return std::visit([&](auto&& func) -> Ret {
if constexpr (std::is_invocable_r_v<Ret, decltype(func), Args...>) {
return func(std::forward<Args>(args)...);
} else {
throw std::runtime_error("Invalid function type in variant_call");
}
}, var);
}
namespace ggml::cpu::kleidiai {
static size_t round_down(size_t x, size_t y) {
return y == 0 ? x : x - (x % y);
}
static void transpose_f32kxn_f16nxk(size_t n, size_t k, float * dst, const uint16_t * src, size_t rhs_stride) {
size_t src_stride = rhs_stride / sizeof(uint16_t);
size_t dst_stride = n;
for (size_t k_idx = 0; k_idx < k; ++k_idx) {
for (size_t n_idx = 0; n_idx < n; ++n_idx) {
uint16_t v = *(src + k_idx + n_idx * src_stride);
*(dst + n_idx + k_idx * dst_stride) = kai_cast_f32_f16(v);
}
}
}
class tensor_traits : public ggml::cpu::tensor_traits {
bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
if (op->op != GGML_OP_MUL_MAT) {
return false;
}
ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, op);
GGML_ASSERT(kernels);
kernel_info * kernel = op->src[1]->ne[1] == 1 ? &kernels->gemv : &kernels->gemm;
size_t k = op->src[0]->ne[0];
size_t n = op->src[0]->ne[1];
size_t m = op->src[1]->ne[1];
size_t mr = kernel->get_mr();
size_t kr = kernel->get_kr();
size_t sr = kernel->get_sr();
if (kernels->rhs_type == GGML_TYPE_Q4_0) {
size = variant_call<size_t>(kernels->lhs_info.packed_size, m, k, QK4_0, mr, kr, sr);
} else if (kernels->rhs_type == GGML_TYPE_F16) {
size = variant_call<size_t>(kernels->lhs_info.packed_size, m, k, mr, kr, sr) +
variant_call<size_t>(kernels->rhs_info.packed_size, n, k) +
k * n * sizeof(float) + n * sizeof(float);
} else {
GGML_ASSERT(false);
}
return true;
}
bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * dst) override {
if (dst->op == GGML_OP_MUL_MAT) {
if (dst->src[0]->type == GGML_TYPE_Q4_0) {
return compute_forward_q4_0(params, dst);
} else if (dst->src[0]->type == GGML_TYPE_F16) {
return compute_forward_kv_cache(params, dst);
}
} else if (dst->op == GGML_OP_GET_ROWS) {
if (dst->src[0]->type == GGML_TYPE_Q4_0) {
return compute_forward_get_rows(params, dst);
}
}
return false;
}
bool compute_forward_kv_cache(ggml_compute_params * params, struct ggml_tensor * dst) {
static std::atomic_flag first_to_arrive = ATOMIC_FLAG_INIT;
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
GGML_TENSOR_BINARY_OP_LOCALS
ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst);
GGML_ASSERT(kernels);
kernel_info * kernel = src1->ne[1] == 1 ? &kernels->gemv : &kernels->gemm;
GGML_ASSERT(kernel);
const int nth = params->nth;
const int ith = params->ith;
const int64_t lhs_batch_size0 = ne12;
const int64_t rhs_batch_size0 = ne02;
const int64_t batch_size = rhs_batch_size0;
const int64_t r = lhs_batch_size0 / rhs_batch_size0;
const int64_t m = ne11 * r;
const int64_t n = ne01;
const int64_t k = ne00;
const size_t lhs_stride = src1->nb[1];
const size_t rhs_stride = src0->nb[1];
const size_t dst_stride = dst->nb[1];
const int64_t mr = static_cast<int64_t>(kernel->get_mr());
const int64_t nr = static_cast<int64_t>(kernel->get_nr());
const int64_t kr = static_cast<int64_t>(kernel->get_kr());
const int64_t sr = static_cast<int64_t>(kernel->get_sr());
const size_t lhs_packed_size = variant_call<size_t>(kernels->lhs_info.packed_size, m, k, mr, kr, sr);
const size_t rhs_packed_size = variant_call<size_t>(kernels->rhs_info.packed_size, n, k);
const size_t kxn_size = k * n * sizeof(float);
const size_t bias_size = n * sizeof(float);
const size_t wsize_required = lhs_packed_size + rhs_packed_size + kxn_size + bias_size;
GGML_ASSERT(wsize_required <= params->wsize);
uint8_t * lhs_packed = static_cast<uint8_t *>(params->wdata);
uint8_t * rhs_packed = lhs_packed + lhs_packed_size;
uint8_t * rhs_kxn = rhs_packed + rhs_packed_size;
uint8_t * bias = rhs_kxn + kxn_size;
for (int64_t batch_idx = 0; batch_idx < batch_size; ++batch_idx) {
const uint8_t * lhs_batch = static_cast<const uint8_t *>(src1->data) + batch_idx * m * lhs_stride;
const uint8_t * rhs_batch = static_cast<const uint8_t *>(src0->data) + batch_idx * n * rhs_stride;
uint8_t * dst_batch = static_cast<uint8_t *>(dst->data) + batch_idx * m * dst_stride;
// LHS packing
{
const int64_t m_roundup_mr = kai_roundup(m, mr);
const int64_t num_threads = KAI_MIN(m_roundup_mr / mr, nth);
if (ith < num_threads) {
const int64_t num_m_per_thread0 = round_down(m_roundup_mr / num_threads, mr);
const int64_t num_m_per_threadN_1 = m - (num_threads - 1) * num_m_per_thread0;
const int64_t m_start = ith * num_m_per_thread0;
const int64_t num_m_per_thread = (ith == num_threads - 1) ? num_m_per_threadN_1 : num_m_per_thread0;
const size_t lhs_offset = variant_call<size_t>(kernels->gemm.get_lhs_offset, m_start, lhs_stride);
const size_t lhs_packed_offset = variant_call<size_t>(kernels->lhs_info.get_packed_offset, m_start, k, mr, kr, sr);
const void * src_ptr = static_cast<const uint8_t *>(lhs_batch) + lhs_offset;
void * dst_ptr = static_cast<uint8_t *>(lhs_packed) + lhs_packed_offset;
variant_call<void>(kernels->lhs_info.pack_func, num_m_per_thread, k, mr, kr, sr, 0, src_ptr, lhs_stride, dst_ptr);
}
}
// RHS packing
if (first_to_arrive.test_and_set(std::memory_order_acquire) == false) {
// First thread to reach this point handles RHS packing
memset(bias, 0, n * sizeof(float));
transpose_f32kxn_f16nxk(n, k, reinterpret_cast<float *>(rhs_kxn),
reinterpret_cast<const uint16_t *>(rhs_batch), rhs_stride);
variant_call<void>(kernels->rhs_info.pack_func, 1, n, k, nr, kr, sr, n * sizeof(float),
rhs_kxn, bias, nullptr, rhs_packed, 0, nullptr);
}
ggml_barrier(params->threadpool);
first_to_arrive.clear(std::memory_order_release);
// Perform the matmul
{
const int64_t m_to_process = m;
const int64_t m_start = 0;
const int64_t n_step = static_cast<int64_t>(kernel->get_n_step());
const int64_t num_threads = KAI_MIN(n / n_step, nth);
if (ith < num_threads) {
const int64_t num_n_per_thread0 = round_down(n / num_threads, n_step);
const int64_t num_n_per_threadN_1 = n - (num_threads - 1) * num_n_per_thread0;
const int64_t n_start = ith * num_n_per_thread0;
const int64_t n_to_process = (ith == num_threads - 1) ? num_n_per_threadN_1 : num_n_per_thread0;
const size_t lhs_packed_offset = variant_call<size_t>(kernel->get_lhs_offset, m_start, k);
const size_t rhs_packed_offset = variant_call<size_t>(kernel->get_rhs_packed_offset, n_start, k);
const size_t dst_offset = kernel->get_dst_offset(m_start, n_start, dst_stride);
const void * lhs_ptr = lhs_packed + lhs_packed_offset;
const void * rhs_ptr = rhs_packed + rhs_packed_offset;
float * dst_ptr = reinterpret_cast<float *>(dst_batch + dst_offset);
variant_call<void>(kernel->run_kernel, m_to_process, n_to_process, k, lhs_ptr, rhs_ptr, dst_ptr, dst_stride, sizeof(float), -FLT_MAX, FLT_MAX);
}
}
if (batch_idx != batch_size - 1) {
// This barrier is necessary when the batch size is larger than 1. While processing a batch,
// the work data buffer (params->wdata) is used as temporary storage which means that only
// a single batch can be processed at any given time. No barrier is needed for the last
// batch since GGML inserts a barrier between the execution of every operator.
ggml_barrier(params->threadpool);
}
}
return true;
}
bool compute_forward_q4_0(struct ggml_compute_params * params, struct ggml_tensor * dst) {
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_Q4_0);
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
GGML_TENSOR_BINARY_OP_LOCALS
ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst);
GGML_ASSERT(kernels);
kernel_info * kernel = src1->ne[1] == 1 ? &kernels->gemv : &kernels->gemm;
lhs_packing_info * lhs_info = &kernels->lhs_info;
GGML_ASSERT(kernel);
const int ith = params->ith;
const int nth = params->nth;
const size_t k = ne00;
const size_t m = ne11;
const size_t n = ne01;
size_t mr = kernel->get_mr();
size_t kr = kernel->get_kr();
size_t sr = kernel->get_sr();
const uint8_t * lhs = static_cast<const uint8_t *>(src1->data);
uint8_t * lhs_packed = (uint8_t*)params->wdata;
const uint8_t * rhs_packed = static_cast<const uint8_t *>(src0->data);
const size_t n_step = kernel->get_n_step();
const size_t num_n_per_thread = kai_roundup(kai_roundup(n, nth) / nth, n_step);
const size_t n_start = ith * num_n_per_thread;
size_t n_to_process = num_n_per_thread;
if ((n_start + n_to_process) > n) {
n_to_process = n - n_start;
}
// Calculate number of columns to be processed per thread
const size_t num_m_per_thread = kai_roundup(m, mr * nth) / nth;
const size_t m_start = ith * num_m_per_thread;
size_t m_to_process = num_m_per_thread;
if ((m_start + m_to_process) > m) {
m_to_process = m - m_start;
}
if (m_start < m) {
// Transform LHS
const size_t src_stride = src1->nb[1];
const float * src_ptr = reinterpret_cast<const float *>(lhs + lhs_info->get_offset(m_start, dst->src[1]->nb[1]));
const size_t lhs_packed_offset = variant_call<size_t>(lhs_info->get_packed_offset, m_start, k, QK4_0, mr, kr, sr);
void * lhs_packed_ptr = static_cast<void *>(lhs_packed + lhs_packed_offset);
variant_call<void>(lhs_info->pack_func, m_to_process, k, QK4_0, mr, kr, sr, 0, src_ptr, src_stride, lhs_packed_ptr);
}
ggml_barrier(params->threadpool);
// Perform the operation
const size_t dst_stride = dst->nb[1];
const size_t lhs_packed_offset = variant_call<size_t>(lhs_info->get_packed_offset, 0, k, QK4_0, mr, kr, sr);
const size_t rhs_packed_offset = variant_call<size_t>(kernel->get_rhs_packed_offset, n_start, k, QK4_0);
const size_t dst_offset = kernel->get_dst_offset(0, n_start, dst_stride);
const void * rhs_ptr = static_cast<const void *>(rhs_packed + rhs_packed_offset);
const void* lhs_ptr = (const void*)((const char *)lhs_packed + lhs_packed_offset);
float *dst_ptr = reinterpret_cast<float *>(static_cast<uint8_t *>(dst->data) + dst_offset);
variant_call<void>(kernel->run_kernel, m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride,
sizeof(float), -FLT_MAX, FLT_MAX);
return true;
}
bool compute_forward_get_rows(struct ggml_compute_params * params, struct ggml_tensor * dst) {
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_Q4_0);
GGML_ASSERT(ctx.kernels);
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
GGML_TENSOR_BINARY_OP_LOCALS
rhs_packing_info * rhs_info = &ctx.kernels->rhs_info;
kernel_info * kernel = &ctx.kernels->gemm;
const int64_t nc = ne00;
const int64_t nr = ggml_nelements(src1);
const size_t block_rows = kernel->get_nr();
const size_t kr = kernel->get_kr();
const size_t num_bytes_multiplier = sizeof(uint16_t);
const size_t packed_stride = rhs_info->packed_stride(nc, block_rows, kr, QK4_0);
const int ith = params->ith;
const int nth = params->nth;
const int dr = (nr + nth - 1) / nth;
const int ir0 = dr * ith;
const int ir1 = MIN(ir0 + dr, nr);
for (int64_t i = ir0; i < ir1; ++i) {
GGML_ASSERT(src1->type == GGML_TYPE_I32);
int64_t row_idx = ((const int32_t *)src1->data)[i];
GGML_ASSERT(row_idx >= 0 && row_idx < src0->ne[1]);
float *out = (float *)((char *)dst->data + i * nb1);
rhs_info->to_float(src0->data, row_idx, nc, out, block_rows, packed_stride, kr, QK4_0, num_bytes_multiplier);
}
return true;
}
public:
int repack(struct ggml_tensor * tensor, const void * data, size_t data_size) {
GGML_ASSERT(tensor->type == GGML_TYPE_Q4_0);
GGML_ASSERT(ctx.kernels);
const size_t n = tensor->ne[1];
const size_t k = tensor->ne[0];
size_t nr = ctx.kernels->gemm.get_nr();
size_t kr = ctx.kernels->gemm.get_kr();
size_t sr = ctx.kernels->gemm.get_sr();
struct kai_rhs_pack_qs4cxs1s0_param params;
params.lhs_zero_point = 1;
params.rhs_zero_point = 8;
variant_call<void>(ctx.kernels->rhs_info.pack_func, 1, n, k, nr, kr, sr, QK4_0, (const uint8_t*)data, nullptr, tensor->data, 0, &params);
return 0;
GGML_UNUSED(data_size);
}
};
static ggml::cpu::tensor_traits * get_tensor_traits(ggml_backend_buffer_t, struct ggml_tensor *) {
static tensor_traits traits;
return &traits;
}
} // namespace ggml::cpu::kleidiai
static enum ggml_status ggml_backend_cpu_kleidiai_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
tensor->extra = (void *) ggml::cpu::kleidiai::get_tensor_traits(buffer, tensor);
return GGML_STATUS_SUCCESS;
GGML_UNUSED(buffer);
}
static void ggml_backend_cpu_kleidiai_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor,
const void * data, size_t offset, size_t size) {
GGML_ASSERT(offset == 0);
GGML_ASSERT(size == ggml_nbytes(tensor));
auto tensor_traits = (ggml::cpu::kleidiai::tensor_traits *) tensor->extra;
auto OK = tensor_traits->repack(tensor, data, size);
GGML_ASSERT(OK == 0);
GGML_UNUSED(buffer);
}
static const char * ggml_backend_cpu_kleidiai_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU_KLEIDIAI";
GGML_UNUSED(buft);
}
static ggml_backend_buffer_t ggml_backend_cpu_kleidiai_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
if (buffer == nullptr) {
return nullptr;
}
buffer->buft = buft;
buffer->iface.init_tensor = ggml_backend_cpu_kleidiai_buffer_init_tensor;
buffer->iface.set_tensor = ggml_backend_cpu_kleidiai_buffer_set_tensor;
buffer->iface.get_tensor = nullptr;
buffer->iface.cpy_tensor = nullptr;
return buffer;
}
static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
return TENSOR_ALIGNMENT;
GGML_UNUSED(buft);
}
static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor) {
GGML_ASSERT(tensor->type == GGML_TYPE_Q4_0);
GGML_ASSERT(ctx.kernels);
const size_t n = tensor->ne[1];
const size_t k = tensor->ne[0];
const size_t nr = ctx.kernels->gemm.get_nr();
const size_t kr = ctx.kernels->gemm.get_kr();
return variant_call<size_t>(ctx.kernels->rhs_info.packed_size, n, k, nr, kr, QK4_0);
GGML_UNUSED(buft);
}
namespace ggml::cpu::kleidiai {
class extra_buffer_type : ggml::cpu::extra_buffer_type {
bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
if ((op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_GET_ROWS) &&
op->src[0]->type == GGML_TYPE_Q4_0 &&
op->src[0]->buffer &&
(ggml_n_dims(op->src[0]) == 2) &&
op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type() && ctx.kernels) {
if (op->op == GGML_OP_GET_ROWS && op->src[1]->ne[0] != 8) {
return false;
}
if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
return false;
}
if ((op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_I32) &&
ggml_ne(op->src[1], 2) == 1 && ggml_ne(op->src[1], 3) == 1) {
return true;
}
}
return false;
}
ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {
if (op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_GET_ROWS) {
if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type()) {
return (ggml::cpu::tensor_traits *) op->src[0]->extra;
}
else if (ggml_kleidiai_select_kernels(ctx.features, op) &&
op->src[0]->op == GGML_OP_VIEW &&
(op->src[1]->op == GGML_OP_PERMUTE || op->src[1]->op == GGML_OP_SOFT_MAX) &&
op->src[1]->ne[1] > 1) {
if ((op->src[0]->nb[0] != 2) ||
(op->src[1]->nb[0] != 4) ||
(op->src[0]->nb[1] * op->src[0]->ne[1] != op->src[0]->nb[2]) ||
(op->src[1]->nb[1] * op->src[1]->ne[1] != op->src[1]->nb[2])) {
return nullptr;
}
return ggml::cpu::kleidiai::get_tensor_traits(NULL, NULL);
}
}
return nullptr;
}
};
} // namespace ggml::cpu::kleidiai
ggml_backend_buffer_type_t ggml_backend_cpu_kleidiai_buffer_type(void) {
static ggml::cpu::kleidiai::extra_buffer_type ctx;
static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_kleidiai = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cpu_kleidiai_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_cpu_kleidiai_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_kleidiai_buffer_type_get_alignment,
/* .get_max_size = */ nullptr, // defaults to SIZE_MAX
/* .get_alloc_size = */ ggml_backend_cpu_kleidiai_buffer_type_get_alloc_size,
/* .is_host = */ nullptr,
},
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ &ctx,
};
init_kleidiai_context();
return &ggml_backend_cpu_buffer_type_kleidiai;
}
ggml/src/ggml-cpu/kleidiai/kleidiai.h
+17
View File
@@ -0,0 +1,17 @@
// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
// SPDX-License-Identifier: MIT
//
#pragma once
#include "ggml-alloc.h"
#ifdef __cplusplus
extern "C" {
#endif
ggml_backend_buffer_type_t ggml_backend_cpu_kleidiai_buffer_type(void);
#ifdef __cplusplus
}
#endif
ggml/src/ggml-cpu/llamafile/sgemm.cpp
+2843
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/llamafile/sgemm.h
+19
View File
@@ -0,0 +1,19 @@
#pragma once
#include <stdint.h>
#include <stdbool.h>
#if defined(__VXE__) || defined(__VXE2__)
#include <vecintrin.h>
#endif
#ifdef __cplusplus
extern "C" {
#endif
bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t, int64_t, int64_t,
const void *, int64_t, const void *, int64_t, void *, int64_t,
int, int, int);
#ifdef __cplusplus
}
#endif
ggml/src/ggml-cpu/ops.cpp
+10186
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/ops.h
+118
View File
@@ -0,0 +1,118 @@
#pragma once
#include "ggml.h"
//
// cache line
//
#if defined(__cpp_lib_hardware_interference_size)
#define CACHE_LINE_SIZE std::hardware_destructive_interference_size
#else
#if defined(__POWER9_VECTOR__)
#define CACHE_LINE_SIZE 128
#elif defined(__VXE__) || defined(__VXE2__)
#define CACHE_LINE_SIZE 256
#else
#define CACHE_LINE_SIZE 64
#endif
#endif
static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
// Work buffer size for im2col operations in CONV2D
#define GGML_IM2COL_WORK_SIZE (16 * 1024 * 1024)
#ifdef __cplusplus
extern "C" {
#endif
void ggml_compute_forward_dup(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_add(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_add1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_acc(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_sum(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_sum_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_mean(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_argmax(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_count_equal(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_repeat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_repeat_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_concat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_silu_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_norm(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_rms_norm(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_rms_norm_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_group_norm(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_l2_norm(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_out_prod(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_scale(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_set(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cpy(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cont(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_reshape(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_view(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_permute(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_transpose(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_get_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_get_rows_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_set_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_diag(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_diag_mask_inf(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_diag_mask_zero(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_soft_max(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_soft_max_ext_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_rope(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_rope_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_clamp(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_conv_transpose_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_im2col(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_im2col_back_f32(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_conv_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_conv_transpose_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_conv_2d_dw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_pool_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_pool_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_pool_2d_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_upscale(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_pad(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_pad_reflect_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_roll(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_arange(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_timestep_embedding(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_argsort(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_leaky_relu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_flash_attn_ext(
const struct ggml_compute_params * params,
const struct ggml_tensor * q,
const struct ggml_tensor * k,
const struct ggml_tensor * v,
const struct ggml_tensor * mask,
struct ggml_tensor * dst);
void ggml_compute_forward_flash_attn_back(
const struct ggml_compute_params * params,
const bool masked,
struct ggml_tensor * dst);
void ggml_compute_forward_ssm_conv(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_ssm_scan(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_win_part(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_win_unpart(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_unary(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_glu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_get_rel_pos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_add_rel_pos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_rwkv_wkv6(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_rwkv_wkv7(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_gla(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_map_custom1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_map_custom2(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_map_custom3(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_custom(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cross_entropy_loss(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cross_entropy_loss_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_opt_step_adamw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
#ifdef __cplusplus
}
#endif
ggml/src/ggml-cpu/quants.c
+1158
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/quants.h
+89
View File
@@ -0,0 +1,89 @@
#pragma once
#define GGML_COMMON_DECL_C
#include "ggml-common.h"
#include "ggml.h"
// GGML CPU internal header
#ifdef __cplusplus
extern "C" {
#endif
// Quantization
void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
// Dot product
void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq1_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq1_m_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
// Generic implementation
void quantize_row_q8_0_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void quantize_row_q8_1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void quantize_row_q8_K_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void ggml_vec_dot_q4_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q2_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q3_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q4_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q5_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q6_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq3_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq3_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq1_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq1_m_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
#ifdef __cplusplus
}
#endif
ggml/src/ggml-cpu/repack.cpp
+1570
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/repack.h
+98
View File
@@ -0,0 +1,98 @@
#pragma once
#define GGML_COMMON_DECL_CPP
#include "ggml-common.h"
#include "traits.h"
#include "ggml.h"
// GGML internal header
ggml_backend_buffer_type_t ggml_backend_cpu_repack_buffer_type(void);
template <int K> constexpr int QK_0() {
if constexpr (K == 4) {
return QK4_0;
}
if constexpr (K == 8) {
return QK8_0;
}
return -1;
}
template <int K, int N> struct block {
ggml_half d[N]; // deltas for N qK_0 blocks
int8_t qs[(QK_0<K>() * N * K) / 8]; // quants for N qK_0 blocks
};
// control size
static_assert(sizeof(block<4, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 2, "wrong block<4,4> size/padding");
static_assert(sizeof(block<4, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<4,8> size/padding");
static_assert(sizeof(block<8, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<8,4> size/padding");
static_assert(sizeof(block<8, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<8,8> size/padding");
using block_q4_0x4 = block<4, 4>;
using block_q4_0x8 = block<4, 8>;
using block_q8_0x4 = block<8, 4>;
using block_q8_0x8 = block<8, 8>;
struct block_q4_Kx8 {
ggml_half d[8]; // super-block scale for quantized scales
ggml_half dmin[8]; // super-block scale for quantized mins
uint8_t scales[96]; // scales and mins, quantized with 6 bits
uint8_t qs[1024]; // 4--bit quants
};
static_assert(sizeof(block_q4_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 4, "wrong q4_K block size/padding");
struct block_q8_Kx4 {
float d[4]; // delta
int8_t qs[QK_K * 4]; // quants
int16_t bsums[QK_K / 4]; // sum of quants in groups of 16
};
static_assert(sizeof(block_q8_Kx4) == sizeof(float) * 4 + QK_K * 4 + (QK_K / 4) * sizeof(int16_t), "wrong q8_K block size/padding");
struct block_iq4_nlx4 {
ggml_half d[4]; // deltas for 4 iq4_nl blocks
uint8_t qs[QK4_NL * 2]; // nibbles / quants for 4 iq4_nl blocks
};
static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding");
#if defined(__cplusplus)
extern "C" {
#endif
void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
// Native implementations
void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
#if defined(__cplusplus)
} // extern "C"
#endif
ggml/src/ggml-cpu/simd-mappings.h
+1184
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/traits.cpp
+36
View File
@@ -0,0 +1,36 @@
#include "traits.h"
#include "ggml-backend-impl.h"
#include "ggml-backend.h"
namespace ggml::cpu {
tensor_traits::~tensor_traits() {}
extra_buffer_type::~extra_buffer_type() {}
} // namespace ggml::cpu
bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) {
for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
if (extra && extra->context) {
auto buf_extra = (ggml::cpu::extra_buffer_type *) extra->context;
auto tensor_traits = buf_extra->get_tensor_traits(op);
if (tensor_traits && tensor_traits->compute_forward(params, op)) {
return true;
}
}
}
return false;
}
bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size) {
for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
if (extra && extra->context) {
auto buf_extra = (ggml::cpu::extra_buffer_type *) extra->context;
auto tensor_traits = buf_extra->get_tensor_traits(op);
if (tensor_traits && tensor_traits->work_size(n_threads, op, *size)) {
return true;
}
}
}
return false;
}
ggml/src/ggml-cpu/traits.h
+38
View File
@@ -0,0 +1,38 @@
#pragma once
#include "ggml-backend-impl.h"
#include "ggml-cpu-impl.h"
#include "ggml.h"
#ifdef __cplusplus
# include <vector>
extern "C" {
#endif
// return true if op part of extra "accelerator"
bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op);
bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size);
#ifdef __cplusplus
}
namespace ggml::cpu {
// register in tensor->extra
class tensor_traits {
public:
virtual ~tensor_traits();
virtual bool work_size(int n_threads, const struct ggml_tensor * op, size_t & size) = 0;
virtual bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) = 0;
};
class extra_buffer_type {
public:
virtual ~extra_buffer_type();
virtual bool supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) = 0;
virtual tensor_traits * get_tensor_traits(const struct ggml_tensor * op) = 0;
};
} // namespace ggml::cpu
// implemented in ggml-cpu.cpp.
std::vector<ggml_backend_buffer_type_t> & ggml_backend_cpu_get_extra_buffers_type();
#endif
ggml/src/ggml-cpu/unary-ops.cpp
+186
View File
@@ -0,0 +1,186 @@
#include "unary-ops.h"
static inline float op_abs(float x) {
return fabsf(x);
}
static inline float op_sgn(float x) {
return (x > 0.f) ? 1.f : ((x < 0.f) ? -1.f : 0.f);
}
static inline float op_neg(float x) {
return -x;
}
static inline float op_step(float x) {
return (x > 0.f) ? 1.f : 0.f;
}
static inline float op_tanh(float x) {
return tanhf(x);
}
static inline float op_elu(float x) {
return (x > 0.f) ? x : expm1f(x);
}
static inline float op_relu(float x) {
return (x > 0.f) ? x : 0.f;
}
static inline float op_sigmoid(float x) {
return 1.f / (1.f + expf(-x));
}
static inline float op_hardsigmoid(float x) {
return fminf(1.0f, fmaxf(0.0f, (x + 3.0f) / 6.0f));
}
static inline float op_exp(float x) {
return expf(x);
}
static inline float op_hardswish(float x) {
return x * fminf(1.0f, fmaxf(0.0f, (x + 3.0f) / 6.0f));
}
static inline float op_sqr(float x) {
return x * x;
}
static inline float op_sqrt(float x) {
return sqrtf(x);
}
static inline float op_sin(float x) {
return sinf(x);
}
static inline float op_cos(float x) {
return cosf(x);
}
static inline float op_log(float x) {
return logf(x);
}
template <float (*op)(float), typename src0_t, typename dst_t>
static inline void vec_unary_op(int64_t n, dst_t * y, const src0_t * x) {
constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32;
constexpr auto f32_to_dst = type_conversion_table<dst_t >::from_f32;
for (int i = 0; i < n; i++) {
y[i] = f32_to_dst(op(src0_to_f32(x[i])));
}
}
template <float (*op)(float), typename src0_t, typename dst_t>
static void apply_unary_op(const ggml_compute_params * params, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
GGML_ASSERT(ggml_is_contiguous_1(src0) && ggml_is_contiguous_1(dst) && ggml_are_same_shape(src0, dst));
GGML_TENSOR_UNARY_OP_LOCALS
GGML_ASSERT( nb0 == sizeof(dst_t));
GGML_ASSERT(nb00 == sizeof(src0_t));
const auto [ir0, ir1] = get_thread_range(params, src0);
for (int64_t ir = ir0; ir < ir1; ++ir) {
const int64_t i03 = ir/(ne02*ne01);
const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
dst_t * dst_ptr = (dst_t *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 );
const src0_t * src0_ptr = (const src0_t *) ((const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
vec_unary_op<op>(ne0, dst_ptr, src0_ptr);
}
}
// TODO: Use the 'traits' lookup table (for type conversion fns), instead of a mass of 'if' conditions with long templates
template <float (*op)(float)>
static void unary_op(const ggml_compute_params * params, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
/* */ if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { // all f32
apply_unary_op<op, float, float>(params, dst);
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { // all f16
apply_unary_op<op, ggml_fp16_t, ggml_fp16_t>(params, dst);
} else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_BF16) { // all bf16
apply_unary_op<op, ggml_bf16_t, ggml_bf16_t>(params, dst);
} else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_F32) {
apply_unary_op<op, ggml_bf16_t, float>(params, dst);
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
apply_unary_op<op, ggml_fp16_t, float>(params, dst);
} else {
fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s\n", __func__,
ggml_type_name(dst->type), ggml_type_name(src0->type));
GGML_ABORT("fatal error");
}
}
void ggml_compute_forward_abs(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_abs>(params, dst);
}
void ggml_compute_forward_sgn(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_sgn>(params, dst);
}
void ggml_compute_forward_neg(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_neg>(params, dst);
}
void ggml_compute_forward_step(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_step>(params, dst);
}
void ggml_compute_forward_tanh(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_tanh>(params, dst);
}
void ggml_compute_forward_elu(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_elu>(params, dst);
}
void ggml_compute_forward_relu(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_relu>(params, dst);
}
void ggml_compute_forward_sigmoid(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_sigmoid>(params, dst);
}
void ggml_compute_forward_hardsigmoid(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_hardsigmoid>(params, dst);
}
void ggml_compute_forward_exp(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_exp>(params, dst);
}
void ggml_compute_forward_hardswish(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_hardswish>(params, dst);
}
void ggml_compute_forward_sqr(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_sqr>(params, dst);
}
void ggml_compute_forward_sqrt(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_sqrt>(params, dst);
}
void ggml_compute_forward_sin(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_sin>(params, dst);
}
void ggml_compute_forward_cos(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_cos>(params, dst);
}
void ggml_compute_forward_log(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_log>(params, dst);
}
ggml/src/ggml-cpu/unary-ops.h
+28
View File
@@ -0,0 +1,28 @@
#pragma once
#include "common.h"
#ifdef __cplusplus
extern "C" {
#endif
void ggml_compute_forward_abs(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_sgn(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_neg(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_step(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_tanh(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_elu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_relu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_sigmoid(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_hardsigmoid(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_exp(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_hardswish(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_sqr(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_sqrt(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_sin(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_log(const struct ggml_compute_params * params, struct ggml_tensor * dst);
#ifdef __cplusplus
}
#endif
ggml/src/ggml-cpu/vec.cpp
+348
View File
@@ -0,0 +1,348 @@
#include "vec.h"
#include <cassert>
// precomputed gelu table for f16 (128 KB)
ggml_fp16_t ggml_table_gelu_f16[1 << 16];
// precomputed quick gelu table for f16 (128 KB)
ggml_fp16_t ggml_table_gelu_quick_f16[1 << 16];
void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * GGML_RESTRICT x, size_t bx, const float * GGML_RESTRICT y, size_t by, int nrc) {
assert(nrc == 1);
GGML_UNUSED(nrc);
GGML_UNUSED(bx);
GGML_UNUSED(by);
GGML_UNUSED(bs);
#if defined(GGML_SIMD)
float sumf = 0.0f;
#if defined(__ARM_FEATURE_SVE)
const int sve_register_length = ggml_cpu_get_sve_cnt() * 8;
const int ggml_f32_epr = sve_register_length / 32;//8;//svcntw(); // SVE128:4, SVE256:8, SVE512:16
const int ggml_f32_step = 8 * ggml_f32_epr; // choose 8 SVE registers
const int np = (n & ~(ggml_f32_step - 1));
svfloat32_t sum1 = svdup_n_f32(0.0f);
svfloat32_t sum2 = svdup_n_f32(0.0f);
svfloat32_t sum3 = svdup_n_f32(0.0f);
svfloat32_t sum4 = svdup_n_f32(0.0f);
svfloat32_t sum5 = svdup_n_f32(0.0f);
svfloat32_t sum6 = svdup_n_f32(0.0f);
svfloat32_t sum7 = svdup_n_f32(0.0f);
svfloat32_t sum8 = svdup_n_f32(0.0f);
svfloat32_t ax1,ax2,ax3,ax4,ax5,ax6,ax7,ax8;
svfloat32_t ay1,ay2,ay3,ay4,ay5,ay6,ay7,ay8;
for (int i = 0; i < np; i += ggml_f32_step) {
ax1 = GGML_F32_VEC_LOAD(x + i);
ay1 = GGML_F32_VEC_LOAD(y + i);
sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1);
ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr);
ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
sum2 = GGML_F32_VEC_FMA(sum2, ax2, ay2);
ax3 = GGML_F32_VEC_LOAD(x + i + 2*ggml_f32_epr);
ay3 = GGML_F32_VEC_LOAD(y + i + 2*ggml_f32_epr);
sum3 = GGML_F32_VEC_FMA(sum3, ax3, ay3);
ax4 = GGML_F32_VEC_LOAD(x + i + 3*ggml_f32_epr);
ay4 = GGML_F32_VEC_LOAD(y + i + 3*ggml_f32_epr);
sum4 = GGML_F32_VEC_FMA(sum4, ax4, ay4);
ax5 = GGML_F32_VEC_LOAD(x + i + 4*ggml_f32_epr);
ay5 = GGML_F32_VEC_LOAD(y + i + 4*ggml_f32_epr);
sum5 = GGML_F32_VEC_FMA(sum5, ax5, ay5);
ax6 = GGML_F32_VEC_LOAD(x + i + 5*ggml_f32_epr);
ay6 = GGML_F32_VEC_LOAD(y + i + 5*ggml_f32_epr);
sum6 = GGML_F32_VEC_FMA(sum6, ax6, ay6);
ax7 = GGML_F32_VEC_LOAD(x + i + 6*ggml_f32_epr);
ay7 = GGML_F32_VEC_LOAD(y + i + 6*ggml_f32_epr);
sum7 = GGML_F32_VEC_FMA(sum7, ax7, ay7);
ax8 = GGML_F32_VEC_LOAD(x + i + 7*ggml_f32_epr);
ay8 = GGML_F32_VEC_LOAD(y + i + 7*ggml_f32_epr);
sum8 = GGML_F32_VEC_FMA(sum8, ax8, ay8);
}
// leftovers
// Since 8 unrolls are done in above loop, leftovers lie in range [0, ggml_f32_step] which is handled in below loop
const int np2 = (n & ~(ggml_f32_epr - 1));
for (int i = np; i < np2; i += ggml_f32_epr) {
ax1 = GGML_F32_VEC_LOAD(x + i);
ay1 = GGML_F32_VEC_LOAD(y + i);
sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1);
}
// maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
if (np2 < n) {
svbool_t pg = svwhilelt_b32(np2, n);
ax1 = svld1_f32(pg, x + np2);
ay1 = svld1_f32(pg, y + np2);
sum1 = svmad_f32_m(pg, ax1, ay1, sum1);
}
// reduce sum1,sum2 to sum1
GGML_F32_VEC_REDUCE(sumf, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8);
#else
const int np = (n & ~(GGML_F32_STEP - 1));
GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
GGML_F32_VEC ax[GGML_F32_ARR];
GGML_F32_VEC ay[GGML_F32_ARR];
for (int i = 0; i < np; i += GGML_F32_STEP) {
for (int j = 0; j < GGML_F32_ARR; j++) {
ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
sum[j] = GGML_F32_VEC_FMA(sum[j], ax[j], ay[j]);
}
}
// reduce sum0..sum3 to sum0
GGML_F32_VEC_REDUCE(sumf, sum);
// leftovers
for (int i = np; i < n; ++i) {
sumf += x[i]*y[i];
}
#endif
#else
// scalar
ggml_float sumf = 0.0;
for (int i = 0; i < n; ++i) {
sumf += (ggml_float)(x[i]*y[i]);
}
#endif
*s = sumf;
}
void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t * GGML_RESTRICT x, size_t bx, ggml_bf16_t * GGML_RESTRICT y, size_t by, int nrc) {
assert(nrc == 1);
GGML_UNUSED(nrc);
GGML_UNUSED(bx);
GGML_UNUSED(by);
GGML_UNUSED(bs);
int i = 0;
ggml_float sumf = 0;
#if defined(__AVX512BF16__)
__m512 c1 = _mm512_setzero_ps();
__m512 c2 = _mm512_setzero_ps();
for (; i + 64 <= n; i += 64) {
c1 = _mm512_dpbf16_ps(c1, m512bh(_mm512_loadu_si512((x + i))),
m512bh(_mm512_loadu_si512((y + i))));
c2 = _mm512_dpbf16_ps(c2, m512bh(_mm512_loadu_si512((x + i + 32))),
m512bh(_mm512_loadu_si512((y + i + 32))));
}
sumf += (ggml_float)_mm512_reduce_add_ps(c1);
sumf += (ggml_float)_mm512_reduce_add_ps(c2);
#elif defined(__AVX512F__)
#define LOAD(p) _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_cvtepu16_epi32(_mm256_loadu_si256((const __m256i *)(p))), 16))
__m512 c1 = _mm512_setzero_ps();
__m512 c2 = _mm512_setzero_ps();
for (; i + 32 <= n; i += 32) {
c1 = _mm512_add_ps(_mm512_mul_ps(LOAD(x + i), LOAD(y + i)), c1);
c2 = _mm512_add_ps(_mm512_mul_ps(LOAD(x + i + 16), LOAD(y + i + 16)), c2);
}
sumf += (ggml_float)_mm512_reduce_add_ps(c1);
sumf += (ggml_float)_mm512_reduce_add_ps(c2);
#undef LOAD
#elif defined(__AVX2__) || defined(__AVX__)
#if defined(__AVX2__)
#define LOAD(p) _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16))
#else
#define LOAD(p) _mm256_castsi256_ps(_mm256_insertf128_si256(_mm256_castsi128_si256(_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16)), (_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_bsrli_si128(_mm_loadu_si128((const __m128i *)(p)), 8)), 16)), 1))
#endif
__m256 c1 = _mm256_setzero_ps();
__m256 c2 = _mm256_setzero_ps();
__m256 c3 = _mm256_setzero_ps();
__m256 c4 = _mm256_setzero_ps();
for (; i + 32 <= n; i += 32) {
c1 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i), LOAD(y + i)), c1);
c2 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 8), LOAD(y + i + 8)), c2);
c3 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 16), LOAD(y + i + 16)), c3);
c4 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 24), LOAD(y + i + 24)), c4);
}
__m128 g;
c1 = _mm256_add_ps(_mm256_add_ps(c1, c3),
_mm256_add_ps(c2, c4));
g = _mm_add_ps(_mm256_extractf128_ps(c1, 1),
_mm256_castps256_ps128(c1));
g = _mm_add_ps(g, _mm_movehl_ps(g, g));
g = _mm_add_ss(g, _mm_movehdup_ps(g));
sumf += (ggml_float)_mm_cvtss_f32(g);
#undef LOAD
#endif
for (; i < n; ++i) {
sumf += (ggml_float)(GGML_BF16_TO_FP32(x[i]) *
GGML_BF16_TO_FP32(y[i]));
}
*s = sumf;
}
void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * GGML_RESTRICT x, size_t bx, ggml_fp16_t * GGML_RESTRICT y, size_t by, int nrc) {
assert(nrc == 1);
GGML_UNUSED(nrc);
GGML_UNUSED(bx);
GGML_UNUSED(by);
GGML_UNUSED(bs);
ggml_float sumf = 0.0;
#if defined(GGML_SIMD)
const int np = (n & ~(GGML_F16_STEP - 1));
GGML_F16_VEC sum[GGML_F16_ARR] = { GGML_F16_VEC_ZERO };
GGML_F16_VEC ax[GGML_F16_ARR];
GGML_F16_VEC ay[GGML_F16_ARR];
for (int i = 0; i < np; i += GGML_F16_STEP) {
for (int j = 0; j < GGML_F16_ARR; j++) {
ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j);
ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
sum[j] = GGML_F16_VEC_FMA(sum[j], ax[j], ay[j]);
}
}
// reduce sum0..sum3 to sum0
GGML_F16_VEC_REDUCE(sumf, sum);
// leftovers
for (int i = np; i < n; ++i) {
sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
}
// if you hit this, you are likely running outside the FP range
assert(!isnan(sumf) && !isinf(sumf));
#else
for (int i = 0; i < n; ++i) {
sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
}
#endif
*s = sumf;
}
void ggml_vec_silu_f32(const int n, float * y, const float * x) {
int i = 0;
#if defined(__AVX512F__) && defined(__AVX512DQ__)
for (; i + 15 < n; i += 16) {
_mm512_storeu_ps(y + i, ggml_v_silu(_mm512_loadu_ps(x + i)));
}
#elif defined(__AVX2__) && defined(__FMA__)
for (; i + 7 < n; i += 8) {
_mm256_storeu_ps(y + i, ggml_v_silu(_mm256_loadu_ps(x + i)));
}
#elif defined(__SSE2__)
for (; i + 3 < n; i += 4) {
_mm_storeu_ps(y + i, ggml_v_silu(_mm_loadu_ps(x + i)));
}
#elif defined(__ARM_NEON) && defined(__aarch64__)
for (; i + 3 < n; i += 4) {
vst1q_f32(y + i, ggml_v_silu(vld1q_f32(x + i)));
}
#endif
for (; i < n; ++i) {
y[i] = ggml_silu_f32(x[i]);
}
}
void ggml_vec_swiglu_f32(const int n, float * y, const float * x, const float * g) {
int i = 0;
#if defined(__AVX512F__) && defined(__AVX512DQ__)
for (; i + 15 < n; i += 16) {
_mm512_storeu_ps(y + i, _mm512_mul_ps(ggml_v_silu(_mm512_loadu_ps(x + i)), _mm512_loadu_ps(g + i)));
}
#elif defined(__AVX2__) && defined(__FMA__)
for (; i + 7 < n; i += 8) {
_mm256_storeu_ps(y + i, _mm256_mul_ps(ggml_v_silu(_mm256_loadu_ps(x + i)), _mm256_loadu_ps(g + i)));
}
#elif defined(__SSE2__)
for (; i + 3 < n; i += 4) {
_mm_storeu_ps(y + i, _mm_mul_ps(ggml_v_silu(_mm_loadu_ps(x + i)), _mm_loadu_ps(g + i)));
}
#elif defined(__ARM_NEON) && defined(__aarch64__)
for (; i + 3 < n; i += 4) {
vst1q_f32(y + i, vmulq_f32(ggml_v_silu(vld1q_f32(x + i)), vld1q_f32(g + i)));
}
#endif
for (; i < n; ++i) {
y[i] = ggml_silu_f32(x[i]) * g[i];
}
}
ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float max) {
int i = 0;
ggml_float sum = 0;
#if defined(__AVX512F__) && defined(__AVX512DQ__)
for (; i + 15 < n; i += 16) {
__m512 val = ggml_v_expf(_mm512_sub_ps(_mm512_loadu_ps(x + i),
_mm512_set1_ps(max)));
_mm512_storeu_ps(y + i, val);
sum += (ggml_float)_mm512_reduce_add_ps(val);
}
#elif defined(__AVX2__) && defined(__FMA__)
for (; i + 7 < n; i += 8) {
__m256 val = ggml_v_expf(_mm256_sub_ps(_mm256_loadu_ps(x + i),
_mm256_set1_ps(max)));
_mm256_storeu_ps(y + i, val);
__m128 val2 = _mm_add_ps(_mm256_extractf128_ps(val, 1),
_mm256_castps256_ps128(val));
val2 = _mm_add_ps(val2, _mm_movehl_ps(val2, val2));
val2 = _mm_add_ss(val2, _mm_movehdup_ps(val2));
sum += (ggml_float)_mm_cvtss_f32(val2);
}
#elif defined(__SSE2__)
for (; i + 3 < n; i += 4) {
__m128 val = ggml_v_expf(_mm_sub_ps(_mm_loadu_ps(x + i),
_mm_set1_ps(max)));
_mm_storeu_ps(y + i, val);
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
val = _mm_add_ps(val, _mm_movehl_ps(val, val));
val = _mm_add_ss(val, _mm_movehdup_ps(val));
#else
__m128 tmp = _mm_shuffle_ps(val, val, _MM_SHUFFLE(2, 3, 0, 1));
val = _mm_add_ps(val, tmp);
tmp = _mm_movehl_ps(tmp, val);
val = _mm_add_ss(val, tmp);
#endif
sum += (ggml_float)_mm_cvtss_f32(val);
}
#elif defined(__ARM_NEON) && defined(__aarch64__)
for (; i + 3 < n; i += 4) {
float32x4_t val = ggml_v_expf(vsubq_f32(vld1q_f32(x + i),
vdupq_n_f32(max)));
vst1q_f32(y + i, val);
sum += (ggml_float)vaddvq_f32(val);
}
#endif
for (; i < n; ++i) {
float val = expf(x[i] - max);
sum += (ggml_float)val;
y[i] = val;
}
return sum;
}
ggml_float ggml_vec_log_soft_max_f32(const int n, float * y, const float * x, float max) {
// log(soft_max) = log(soft_max_i / soft_max_sum) = log(soft_max_i) - log(soft_max_sum) = (logit_i - max) - log(soft_max_i)
int i = 0;
ggml_float sum = 0;
for (; i < n; ++i) {
float val = x[i] - max;
y[i] = val;
sum += (ggml_float)expf(val);
}
return sum = (ggml_float)logf(sum);
}
ggml/src/ggml-cpu/vec.h
+1106
View File
File diff suppressed because it is too large Load Diff