owen/longhaul.cpp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Inital commit

Browse Source
This commit is contained in:
owen
2026-07-29 01:00:10 -05:00
commit 23e8ea90e4
3301 changed files with 1376308 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cpu/CMakeLists.txt
+744
View File
@@ -0,0 +1,744 @@
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(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
# Disable LTO for the feature detection code to prevent cross-module optimization
# from inlining architecture-specific instructions into the score function.
# Without this, LTO can cause SIGILL when loading backends on older CPUs
# (e.g., loading power10 backend on power9 crashes before feature check runs).
target_compile_options(${GGML_CPU_FEATS_NAME} PRIVATE -fno-lto)
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 ${ACCELERATE_FRAMEWORK})
else()
message(WARNING "Accelerate framework not found")
endif()
endif()
if (GGML_OPENMP_ENABLED)
target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_OPENMP)
target_link_libraries(${GGML_CPU_NAME} PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
endif()
if (GGML_LLAMAFILE)
target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_LLAMAFILE)
list(APPEND GGML_CPU_SOURCES
ggml-cpu/llamafile/sgemm.cpp
ggml-cpu/llamafile/sgemm.h)
endif()
if (GGML_CPU_HBM)
find_library(memkind memkind REQUIRED)
message(STATUS "Using memkind for CPU HBM")
target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_HBM)
target_link_libraries(${GGML_CPU_NAME} PUBLIC memkind)
endif()
if (GGML_SYSTEM_ARCH STREQUAL "ARM")
message(STATUS "ARM detected")
list(APPEND GGML_CPU_SOURCES
ggml-cpu/arch/arm/quants.c
ggml-cpu/arch/arm/repack.cpp
)
if (MSVC AND NOT CMAKE_C_COMPILER_ID STREQUAL "Clang")
message(FATAL_ERROR "MSVC is not supported for ARM, use clang")
else()
check_cxx_compiler_flag(-mfp16-format=ieee GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E)
if (NOT "${GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
list(APPEND ARCH_FLAGS -mfp16-format=ieee)
endif()
if (GGML_NATIVE)
# -mcpu=native does not always enable all the features in some compilers,
# so we check for them manually and enable them if available
execute_process(
COMMAND ${CMAKE_C_COMPILER} -mcpu=native -E -v -
INPUT_FILE "/dev/null"
OUTPUT_QUIET
ERROR_VARIABLE ARM_MCPU
RESULT_VARIABLE ARM_MCPU_RESULT
)
if (NOT ARM_MCPU_RESULT)
string(REGEX MATCH "-mcpu=[^ ']+" ARM_MCPU_FLAG "${ARM_MCPU}")
string(REGEX MATCH "-march=[^ ']+" ARM_MARCH_FLAG "${ARM_MCPU}")
# on some old GCC we need to read -march=
if (ARM_MARCH_FLAG AND NOT "${ARM_MARCH_FLAG}" STREQUAL "-march=native")
set(ARM_NATIVE_FLAG "${ARM_MARCH_FLAG}")
elseif(ARM_MCPU_FLAG AND NOT "${ARM_MCPU_FLAG}" STREQUAL "-mcpu=native")
set(ARM_NATIVE_FLAG "${ARM_MCPU_FLAG}")
endif()
endif()
if ("${ARM_NATIVE_FLAG}" STREQUAL "")
set(ARM_NATIVE_FLAG -mcpu=native)
message(WARNING "ARM -march/-mcpu not found, -mcpu=native will be used")
else()
message(STATUS "ARM detected flags: ${ARM_NATIVE_FLAG}")
endif()
include(CheckCXXSourceRuns)
macro(check_arm_feature tag feature code)
set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
set(CMAKE_REQUIRED_FLAGS "${ARM_NATIVE_FLAG}+${tag}")
check_cxx_source_runs("${code}" GGML_MACHINE_SUPPORTS_${tag})
if (GGML_MACHINE_SUPPORTS_${tag})
set(ARM_NATIVE_FLAG_FIX "${ARM_NATIVE_FLAG_FIX}+${tag}")
else()
set(CMAKE_REQUIRED_FLAGS "${ARM_NATIVE_FLAG}+no${tag}")
check_cxx_source_compiles("int main() { return 0; }" GGML_MACHINE_SUPPORTS_no${tag})
if (GGML_MACHINE_SUPPORTS_no${tag})
set(ARM_NATIVE_FLAG_FIX "${ARM_NATIVE_FLAG_FIX}+no${tag}")
list(APPEND ARCH_FLAGS -U__ARM_FEATURE_${feature})
endif()
endif()
set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
endmacro()
check_arm_feature(dotprod DOTPROD "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }")
check_arm_feature(i8mm MATMUL_INT8 "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }")
check_arm_feature(sve SVE "#include <arm_sve.h>\nint main() { svfloat32_t _a, _b; volatile svfloat32_t _c = svadd_f32_z(svptrue_b8(), _a, _b); return 0; }")
check_arm_feature(sme SME "#include <arm_sme.h>\n__arm_locally_streaming int main() { __asm__ volatile(\"smstart; smstop;\"); return 0; }")
list(APPEND ARCH_FLAGS "${ARM_NATIVE_FLAG}${ARM_NATIVE_FLAG_FIX}")
else()
if (GGML_CPU_ARM_ARCH)
list(APPEND ARCH_FLAGS -march=${GGML_CPU_ARM_ARCH})
elseif(GGML_CPU_ALL_VARIANTS)
# Begin with the lowest baseline
set(ARM_MCPU "armv8-a")
set(ARCH_TAGS "")
set(ARCH_DEFINITIONS "")
# When a feature is selected, bump the MCPU to the first
# version that supported it
if (GGML_INTERNAL_DOTPROD)
set(ARM_MCPU "armv8.2-a")
set(ARCH_TAGS "${ARCH_TAGS}+dotprod")
list(APPEND ARCH_DEFINITIONS GGML_USE_DOTPROD)
endif()
if (GGML_INTERNAL_FP16_VECTOR_ARITHMETIC)
set(ARM_MCPU "armv8.2-a")
set(ARCH_TAGS "${ARCH_TAGS}+fp16")
list(APPEND ARCH_DEFINITIONS GGML_USE_FP16_VECTOR_ARITHMETIC)
endif()
if (GGML_INTERNAL_SVE)
set(ARM_MCPU "armv8.2-a")
set(ARCH_TAGS "${ARCH_TAGS}+sve")
list(APPEND ARCH_DEFINITIONS GGML_USE_SVE)
endif()
if (GGML_INTERNAL_MATMUL_INT8)
set(ARM_MCPU "armv8.6-a")
set(ARCH_TAGS "${ARCH_TAGS}+i8mm")
list(APPEND ARCH_DEFINITIONS GGML_USE_MATMUL_INT8)
endif()
if (GGML_INTERNAL_SVE2)
set(ARM_MCPU "armv8.6-a")
set(ARCH_TAGS "${ARCH_TAGS}+sve2")
list(APPEND ARCH_DEFINITIONS GGML_USE_SVE2)
endif()
if (GGML_INTERNAL_NOSVE)
set(ARCH_TAGS "${ARCH_TAGS}+nosve")
endif()
if (GGML_INTERNAL_SME)
set(ARM_MCPU "armv9.2-a")
set(ARCH_TAGS "${ARCH_TAGS}+sme")
list(APPEND ARCH_DEFINITIONS GGML_USE_SME)
endif()
list(APPEND ARCH_FLAGS "-march=${ARM_MCPU}${ARCH_TAGS}")
ggml_add_cpu_backend_features(${GGML_CPU_NAME} arm ${ARCH_DEFINITIONS})
endif()
endif()
message(STATUS "Checking for ARM features using flags:")
foreach(flag IN LISTS ARCH_FLAGS)
message(STATUS " ${flag}")
endforeach()
include(CheckCXXSourceCompiles)
set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
string(REPLACE ";" " " ARCH_FLAGS_STR "${ARCH_FLAGS}")
set(CMAKE_REQUIRED_FLAGS "${ARCH_FLAGS_STR}")
foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC SME)
set(ARM_FEATURE "HAVE_${feature}")
check_cxx_source_compiles(
"
#if !defined(__ARM_FEATURE_${feature})
# error \"Feature ${feature} is not defined\"
#endif
int main() { return 0; }
"
${ARM_FEATURE}
)
endforeach()
set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
endif()
elseif (GGML_SYSTEM_ARCH STREQUAL "x86")
message(STATUS "x86 detected")
list(APPEND GGML_CPU_SOURCES
ggml-cpu/arch/x86/quants.c
ggml-cpu/arch/x86/repack.cpp
)
if (MSVC)
# instruction set detection for MSVC only
if (GGML_NATIVE)
include(ggml-cpu/cmake/FindSIMD.cmake)
endif ()
if (GGML_AVX512)
list(APPEND ARCH_FLAGS /arch:AVX512)
# /arch:AVX512 includes: __AVX512F__, __AVX512CD__, __AVX512BW__, __AVX512DQ__, and __AVX512VL__
# MSVC has no compile-time flags enabling specific
# AVX512 extensions, neither it defines the
# macros corresponding to the extensions.
# Do it manually.
list(APPEND ARCH_DEFINITIONS GGML_AVX512)
if (GGML_AVX512_VBMI)
list(APPEND ARCH_DEFINITIONS __AVX512VBMI__)
if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
list(APPEND ARCH_FLAGS -mavx512vbmi)
endif()
endif()
if (GGML_AVX512_VNNI)
list(APPEND ARCH_DEFINITIONS __AVX512VNNI__ GGML_AVX512_VNNI)
if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
list(APPEND ARCH_FLAGS -mavx512vnni)
endif()
endif()
if (GGML_AVX512_BF16)
list(APPEND ARCH_DEFINITIONS __AVX512BF16__ GGML_AVX512_BF16)
if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
list(APPEND ARCH_FLAGS -mavx512bf16)
endif()
endif()
if (GGML_AMX_TILE)
list(APPEND ARCH_DEFINITIONS __AMX_TILE__ GGML_AMX_TILE)
endif()
if (GGML_AMX_INT8)
list(APPEND ARCH_DEFINITIONS __AMX_INT8__ GGML_AMX_INT8)
endif()
if (GGML_AMX_BF16)
list(APPEND ARCH_DEFINITIONS __AMX_BF16__ GGML_AMX_BF16)
endif()
elseif (GGML_AVX2)
list(APPEND ARCH_FLAGS /arch:AVX2)
list(APPEND ARCH_DEFINITIONS GGML_AVX2 GGML_FMA GGML_F16C)
elseif (GGML_AVX)
list(APPEND ARCH_FLAGS /arch:AVX)
list(APPEND ARCH_DEFINITIONS GGML_AVX)
elseif (GGML_SSE42)
list(APPEND ARCH_FLAGS /arch:SSE4.2)
list(APPEND ARCH_DEFINITIONS GGML_SSE42)
endif()
if (GGML_AVX_VNNI)
list(APPEND ARCH_DEFINITIONS __AVXVNNI__ GGML_AVX_VNNI)
endif()
if (GGML_BMI2)
# MSVC does not define macro __BMI2__
list(APPEND ARCH_DEFINITIONS __BMI2__ GGML_BMI2)
endif()
else ()
if (GGML_NATIVE)
list(APPEND ARCH_FLAGS -march=native)
else ()
if (GGML_SSE42)
list(APPEND ARCH_FLAGS -msse4.2)
list(APPEND ARCH_DEFINITIONS GGML_SSE42)
endif()
if (GGML_F16C)
list(APPEND ARCH_FLAGS -mf16c)
list(APPEND ARCH_DEFINITIONS GGML_F16C)
endif()
if (GGML_FMA)
list(APPEND ARCH_FLAGS -mfma)
list(APPEND ARCH_DEFINITIONS GGML_FMA)
endif()
if (GGML_BMI2)
list(APPEND ARCH_FLAGS -mbmi2)
list(APPEND ARCH_DEFINITIONS GGML_BMI2)
endif()
if (GGML_AVX)
list(APPEND ARCH_FLAGS -mavx)
list(APPEND ARCH_DEFINITIONS GGML_AVX)
endif()
if (GGML_AVX2)
list(APPEND ARCH_FLAGS -mavx2)
list(APPEND ARCH_DEFINITIONS GGML_AVX2)
endif()
if (GGML_AVX_VNNI)
list(APPEND ARCH_FLAGS -mavxvnni)
list(APPEND ARCH_DEFINITIONS GGML_AVX_VNNI)
endif()
if (GGML_AVX512)
list(APPEND ARCH_FLAGS -mavx512f)
list(APPEND ARCH_FLAGS -mavx512cd)
list(APPEND ARCH_FLAGS -mavx512vl)
list(APPEND ARCH_FLAGS -mavx512dq)
list(APPEND ARCH_FLAGS -mavx512bw)
list(APPEND ARCH_DEFINITIONS GGML_AVX512)
endif()
if (GGML_AVX512_VBMI)
list(APPEND ARCH_FLAGS -mavx512vbmi)
list(APPEND ARCH_DEFINITIONS GGML_AVX512_VBMI)
endif()
if (GGML_AVX512_VNNI)
list(APPEND ARCH_FLAGS -mavx512vnni)
list(APPEND ARCH_DEFINITIONS GGML_AVX512_VNNI)
endif()
if (GGML_AVX512_BF16)
list(APPEND ARCH_FLAGS -mavx512bf16)
list(APPEND ARCH_DEFINITIONS GGML_AVX512_BF16)
endif()
if (GGML_AMX_TILE)
list(APPEND ARCH_FLAGS -mamx-tile)
list(APPEND ARCH_DEFINITIONS GGML_AMX_TILE)
endif()
if (GGML_AMX_INT8)
list(APPEND ARCH_FLAGS -mamx-int8)
list(APPEND ARCH_DEFINITIONS GGML_AMX_INT8)
endif()
if (GGML_AMX_BF16)
list(APPEND ARCH_FLAGS -mamx-bf16)
list(APPEND ARCH_DEFINITIONS GGML_AMX_BF16)
endif()
endif()
endif()
if (GGML_BACKEND_DL)
if (GGML_NATIVE)
# the feature check relies on ARCH_DEFINITIONS, but it is not set with GGML_NATIVE
message(FATAL_ERROR "GGML_NATIVE is not compatible with GGML_BACKEND_DL, consider using GGML_CPU_ALL_VARIANTS")
endif()
ggml_add_cpu_backend_features(${GGML_CPU_NAME} x86 ${ARCH_DEFINITIONS})
endif()
elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
message(STATUS "PowerPC detected")
list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/powerpc/quants.c)
if (GGML_NATIVE)
if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
file(READ "/proc/cpuinfo" POWER10_M)
elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "powerpc")
execute_process(COMMAND bash -c "prtconf |grep 'Implementation' | head -n 1" OUTPUT_VARIABLE POWER10_M)
endif()
string(TOUPPER "${POWER10_M}" POWER10_M_UPPER)
string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M_UPPER}")
string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}")
if (EXTRACTED_NUMBER EQUAL 10 OR EXTRACTED_NUMBER EQUAL 11)
list(APPEND ARCH_FLAGS -mcpu=power10)
elseif (EXTRACTED_NUMBER EQUAL 9)
list(APPEND ARCH_FLAGS -mcpu=power9)
elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
list(APPEND ARCH_FLAGS -mcpu=powerpc64le -mtune=native)
else()
list(APPEND ARCH_FLAGS -mcpu=native -mtune=native -mpowerpc64)
endif()
elseif(GGML_CPU_ALL_VARIANTS)
# Begin with the lowest baseline
set(ARCH_DEFINITIONS "")
# When a feature is selected, bump the MCPU to the first
# version that supported it
foreach(PVER RANGE 7 11)
if(DEFINED GGML_INTERNAL_POWER${PVER})
set(POWERPC_MCPU "power${PVER}")
list(APPEND ARCH_DEFINITIONS GGML_USE_POWER${PVER})
endif()
endforeach()
if (GGML_INTERNAL_VSX)
list(APPEND ARCH_DEFINITIONS GGML_USE_VSX)
list(APPEND ARCH_FLAGS -mvsx)
endif()
if (DEFINED POWERPC_MCPU)
list(APPEND ARCH_FLAGS -mcpu=${POWERPC_MCPU})
endif()
ggml_add_cpu_backend_features(${GGML_CPU_NAME} powerpc ${ARCH_DEFINITIONS})
else()
if (GGML_CPU_POWERPC_CPUTYPE)
list(APPEND ARCH_FLAGS -mcpu=${GGML_CPU_POWERPC_CPUTYPE})
endif()
endif()
elseif (GGML_SYSTEM_ARCH STREQUAL "loongarch64")
message(STATUS "loongarch64 detected")
list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/loongarch/quants.c)
list(APPEND ARCH_FLAGS -march=loongarch64)
if (GGML_LASX)
list(APPEND ARCH_FLAGS -mlasx)
endif()
if (GGML_LSX)
list(APPEND ARCH_FLAGS -mlsx)
endif()
elseif (GGML_SYSTEM_ARCH STREQUAL "riscv64")
message(STATUS "riscv64 detected")
list(APPEND GGML_CPU_SOURCES
ggml-cpu/arch/riscv/quants.c
ggml-cpu/arch/riscv/repack.cpp
)
if (GGML_CPU_RISCV64_SPACEMIT)
include(ggml-cpu/cmake/FindSMTIME.cmake)
target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_RISCV64_SPACEMIT ${RISCV64_SPACEMIT_IME_SPEC})
list(APPEND GGML_CPU_SOURCES
ggml-cpu/spacemit/ime.cpp
ggml-cpu/spacemit/ime.h
ggml-cpu/spacemit/spine_mem_pool.cpp
ggml-cpu/spacemit/spine_mem_pool.h
ggml-cpu/spacemit/repack.cpp
ggml-cpu/spacemit/repack.h
ggml-cpu/spacemit/ime_env.cpp
ggml-cpu/spacemit/ime_env.h
ggml-cpu/spacemit/ime1_kernels.cpp
ggml-cpu/spacemit/ime2_kernels.cpp
ggml-cpu/spacemit/ime_kernels.h
ggml-cpu/spacemit/rvv_kernels.cpp
ggml-cpu/spacemit/rvv_kernels.h
)
endif()
if(NOT GGML_CPU_ALL_VARIANTS)
set(MARCH_STR "rv64gc")
if (GGML_RVV)
string(APPEND MARCH_STR "v")
endif()
if (GGML_RV_ZFH)
string(APPEND MARCH_STR "_zfh")
endif()
if (GGML_XTHEADVECTOR)
string(APPEND MARCH_STR "_xtheadvector")
elseif (GGML_RVV)
if (GGML_RV_ZVFH)
string(APPEND MARCH_STR "_zvfh")
endif()
if (GGML_RV_ZVFBFWMA)
string(APPEND MARCH_STR "_zvfbfwma")
endif()
endif()
if (GGML_RV_ZICBOP)
string(APPEND MARCH_STR "_zicbop")
endif()
if (GGML_RV_ZIHINTPAUSE)
string(APPEND MARCH_STR "_zihintpause")
endif()
if (GGML_RV_ZBA)
string(APPEND MARCH_STR "_zba")
endif()
if (GGML_CPU_RISCV64_SPACEMIT)
# `xsmtvdotii' is only required for GCC >= 15.
if (CMAKE_C_COMPILER_ID STREQUAL "GNU" AND
CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 15)
string(APPEND MARCH_STR "_xsmtvdotii")
endif()
endif()
list(APPEND ARCH_FLAGS "-march=${MARCH_STR}" -mabi=lp64d)
else()
# Begin with the lowest baseline
set(ARCH_DEFINITIONS "")
if (GGML_INTERNAL_RVV)
message(STATUS "RVV enabled")
list(APPEND ARCH_DEFINITIONS GGML_USE_RVV)
list(APPEND ARCH_FLAGS -march=rv64gc_v -mabi=lp64d)
endif()
ggml_add_cpu_backend_features(${GGML_CPU_NAME} riscv ${ARCH_DEFINITIONS})
endif()
elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
message(STATUS "s390x detected")
list(APPEND GGML_CPU_SOURCES
ggml-cpu/arch/s390/quants.c)
# for native compilation
if (GGML_NATIVE)
# check machine level to determine target
file(READ "/proc/cpuinfo" CPUINFO_CONTENTS)
string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS})
# TODO: Separation to determine activation of VX/VXE/VXE2
if (${S390X_M} MATCHES "8561|8562")
message(STATUS "z15 target")
list(APPEND ARCH_FLAGS -march=z15)
elseif (${S390X_M} MATCHES "3931")
message(STATUS "z16 target")
list(APPEND ARCH_FLAGS -march=z16)
elseif (${S390X_M} MATCHES "9175|9176")
# NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
# binutils must also be updated to the latest for the -march=z17 flag to work. Otherwise, use -march=arch15.
message(STATUS "z17 target")
list(APPEND ARCH_FLAGS -march=arch15)
else()
message(STATUS "Unknown target")
message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")
list(APPEND ARCH_FLAGS -march=native -mtune=native)
endif()
# for cross-compilation
elseif(GGML_CPU_ALL_VARIANTS)
# range through IBM z15 to z17
# NOTE: update when a new hardware level is released
foreach (ZHW RANGE 15 17)
if(DEFINED GGML_INTERNAL_Z${ZHW})
message(STATUS "z${ZHW} cross-compile target")
list(APPEND ARCH_FLAGS -march=z${ZHW})
endif()
endforeach()
endif()
if (GGML_VXE OR GGML_INTERNAL_VXE2)
message(STATUS "VXE2 enabled")
list(APPEND ARCH_FLAGS -mvx -mzvector)
list(APPEND ARCH_DEFINITIONS GGML_USE_VXE2)
endif()
if (GGML_INTERNAL_NNPA)
message(STATUS "NNPA enabled")
list(APPEND ARCH_DEFINITIONS GGML_USE_NNPA)
endif()
ggml_add_cpu_backend_features(${GGML_CPU_NAME} s390 ${ARCH_DEFINITIONS})
elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm")
message(STATUS "Wasm detected")
list (APPEND GGML_CPU_SOURCES ggml-cpu/arch/wasm/quants.c)
else()
message(WARNING "Unknown CPU architecture. Falling back to generic implementations.")
list(APPEND ARCH_FLAGS -DGGML_CPU_GENERIC)
endif()
if (GGML_CPU_REPACK)
target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_REPACK)
endif()
if (GGML_CPU_KLEIDIAI)
message(STATUS "Using KleidiAI optimized kernels if applicable")
# Disable the KleidiAI tests
set(KLEIDIAI_BUILD_TESTS OFF)
# Fetch KleidiAI sources:
include(FetchContent)
set(KLEIDIAI_COMMIT_TAG "v1.24.0")
set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/releases/download/${KLEIDIAI_COMMIT_TAG}/kleidiai-${KLEIDIAI_COMMIT_TAG}-src.tar.gz")
set(KLEIDIAI_RELEASE_ARCHIVE_MD5 "2f02ebe29573d45813e671eb304f2a00")
set(KLEIDIAI_FETCH_ARGS
URL ${KLEIDIAI_DOWNLOAD_URL}
URL_HASH MD5=${KLEIDIAI_RELEASE_ARCHIVE_MD5}
)
if (CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
list(APPEND KLEIDIAI_FETCH_ARGS DOWNLOAD_EXTRACT_TIMESTAMP NEW)
endif()
if (CMAKE_VERSION VERSION_GREATER_EQUAL "3.28")
FetchContent_Declare(KleidiAI_Download
${KLEIDIAI_FETCH_ARGS}
EXCLUDE_FROM_ALL
)
FetchContent_MakeAvailable(KleidiAI_Download)
FetchContent_GetProperties(KleidiAI_Download SOURCE_DIR KLEIDIAI_SRC)
else()
FetchContent_Declare(KleidiAI_Download
${KLEIDIAI_FETCH_ARGS}
)
FetchContent_GetProperties(KleidiAI_Download
SOURCE_DIR KLEIDIAI_SRC
POPULATED KLEIDIAI_POPULATED
)
if (NOT KLEIDIAI_POPULATED)
FetchContent_Populate(KleidiAI_Download)
FetchContent_GetProperties(KleidiAI_Download SOURCE_DIR KLEIDIAI_SRC)
endif()
endif()
add_compile_definitions(GGML_USE_CPU_KLEIDIAI)
list(APPEND GGML_CPU_SOURCES
ggml-cpu/kleidiai/kleidiai.cpp
ggml-cpu/kleidiai/kernels.cpp
ggml-cpu/kleidiai/kleidiai.h
ggml-cpu/kleidiai/kernels.h
)
# KleidiAI
include_directories(
${KLEIDIAI_SRC}/
${KLEIDIAI_SRC}/kai/
${KLEIDIAI_SRC}/kai/ukernels/
${KLEIDIAI_SRC}/kai/ukernels/matmul/
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/)
set(ARCH_FLAGS_TEMP "${ARCH_FLAGS}")
if (NOT ARCH_FLAGS_TEMP)
string(REGEX MATCH "-march=[^ ]+" ARCH_FLAGS_TEMP "${CMAKE_C_FLAGS}")
endif()
string(FIND "${ARCH_FLAGS_TEMP}" "+dotprod" DOTPROD_ENABLED)
string(FIND "${ARCH_FLAGS_TEMP}" "+i8mm" I8MM_ENABLED)
string(FIND "${ARCH_FLAGS_TEMP}" "+sme" SME_ENABLED)
string(FIND "${ARCH_FLAGS_TEMP}" "+sve" SVE_ENABLED)
set(PRIVATE_ARCH_FLAGS ${ARCH_FLAGS_TEMP})
list(APPEND GGML_KLEIDIAI_SOURCES
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.c)
if (NOT DOTPROD_ENABLED MATCHES -1)
list(APPEND GGML_KLEIDIAI_SOURCES
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c)
endif()
if (NOT I8MM_ENABLED MATCHES -1)
list(APPEND GGML_KLEIDIAI_SOURCES
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c)
endif()
if (NOT SME_ENABLED MATCHES -1)
list(APPEND GGML_KLEIDIAI_SME_SOURCES
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa_asm.S
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot_asm.S
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa_asm.S)
set_source_files_properties(${GGML_KLEIDIAI_SME_SOURCES}
PROPERTIES COMPILE_OPTIONS "-fno-tree-vectorize;${ARCH_FLAGS_TEMP}+sve+sve2+sme")
list(APPEND GGML_CPU_SOURCES ${GGML_KLEIDIAI_SME_SOURCES})
list(APPEND GGML_KLEIDIAI_SME2_SOURCES
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa_asm.S
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot_asm.S
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa_asm.S
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa_asm.S
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa_asm.S
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_f16pmrx2_f32_neon.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_f32p2vlx1_f32_sme.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_f32p2vlx1_f32_sme_asm.S
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme_asm.S
${KLEIDIAI_SRC}/kai/kai_common_sme_asm.S)
set_source_files_properties(${GGML_KLEIDIAI_SME2_SOURCES}
PROPERTIES COMPILE_OPTIONS "-fno-tree-vectorize;${ARCH_FLAGS_TEMP}+sve+sve2+sme2+fp16")
list(APPEND GGML_CPU_SOURCES ${GGML_KLEIDIAI_SME2_SOURCES})
set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}")
endif()
if (NOT SVE_ENABLED MATCHES -1)
list(APPEND GGML_KLEIDIAI_SOURCES
${KLEIDIAI_SRC}/kai/kai_common_sve_asm.S
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod_asm.S
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm_asm.S
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm.c)
endif()
set_source_files_properties(${GGML_KLEIDIAI_SOURCES} PROPERTIES COMPILE_OPTIONS "${PRIVATE_ARCH_FLAGS}")
list(APPEND GGML_CPU_SOURCES ${GGML_KLEIDIAI_SOURCES})
endif()
message(STATUS "Adding CPU backend variant ${GGML_CPU_NAME}: ${ARCH_FLAGS} ${ARCH_DEFINITIONS}")
target_sources(${GGML_CPU_NAME} PRIVATE ${GGML_CPU_SOURCES})
target_compile_options(${GGML_CPU_NAME} PRIVATE ${ARCH_FLAGS})
target_compile_definitions(${GGML_CPU_NAME} PRIVATE ${ARCH_DEFINITIONS})
if (EMSCRIPTEN)
set_target_properties(${GGML_CPU_NAME} PROPERTIES COMPILE_FLAGS "-msimd128")
endif()
if (CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")
# The compiler automatically enables "-ffast-math" which can cause NaNs in tests due to "-fassociative-math"
target_compile_options(${GGML_CPU_NAME} PRIVATE "-fno-associative-math")
endif()
endfunction()
ggml/src/ggml-cpu/amx/amx.cpp
+249
View File
@@ -0,0 +1,249 @@
#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(__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,
/* .set_tensor_2d = */ nullptr,
/* .get_tensor_2d = */ 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 {
if (op->op != GGML_OP_MUL_MAT) {
return false;
}
auto * src0 = op->src[0];
auto * src1 = op->src[1];
if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1)) {
return false;
}
if (!src0->buffer || src0->buffer->buft != ggml_backend_amx_buffer_type()) {
return false;
}
if (src1->buffer && !ggml_backend_buft_is_host(src1->buffer->buft)) {
return false;
}
if (op->ne[0] % (TILE_N * 2)) {
return false;
}
int alignment;
switch (src0->type) {
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q8_0:
alignment = TILE_K;
break;
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q5_K:
case GGML_TYPE_Q6_K:
case GGML_TYPE_IQ4_XS:
alignment = 256; // QK_K
break;
case GGML_TYPE_F16:
alignment = 16;
break;
default:
return false;
}
if (src0->ne[0] % alignment) {
return false;
}
if (src1->type != GGML_TYPE_F32) {
return false;
}
return true;
}
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(__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;
#else
return false;
#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
+115
View File
@@ -0,0 +1,115 @@
#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>
#else
#include <thread>
#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 (n <= 0) {
return;
}
#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
int nth = std::thread::hardware_concurrency();
if (nth <= 1) {
f(0, n);
return;
}
if (nth > n) {
nth = n;
}
std::vector<std::thread> threads;
threads.reserve(nth);
for (int ith = 0; ith < nth; ++ith) {
threads.emplace_back([&f, n, ith, nth] {
int tbegin, tend;
balance211(n, nth, ith, tbegin, tend);
f(tbegin, tend);
});
}
for (auto & t : threads) {
t.join();
}
#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
+2511
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
+353
View File
@@ -0,0 +1,353 @@
#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_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
#define ggml_vec_dot_q2_0_q8_0_generic ggml_vec_dot_q2_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_4x4_generic ggml_quantize_mat_q8_K_4x4
#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_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_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_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
#elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64)
// repack.cpp
#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
// quants.c
#define ggml_vec_dot_q2_0_q8_0_generic ggml_vec_dot_q2_0_q8_0
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_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_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_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_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_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_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
#define ggml_vec_dot_q2_0_q8_0_generic ggml_vec_dot_q2_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_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_4x4_generic ggml_quantize_mat_q8_K_4x4
#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_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_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_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_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
#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
#define ggml_vec_dot_q2_0_q8_0_generic ggml_vec_dot_q2_0_q8_0
// 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_4x4_generic ggml_quantize_mat_q8_K_4x4
#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_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_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_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
#elif defined(__riscv)
// quants.c
#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
#define ggml_vec_dot_q2_0_q8_0_generic ggml_vec_dot_q2_0_q8_0
// repack.cpp
#define ggml_quantize_mat_q8_0_4x1_generic ggml_quantize_mat_q8_0_4x1
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_K_4x1_generic ggml_quantize_mat_q8_K_4x1
#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
#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_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_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_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
#elif defined(__s390x__)
// quants.c
#define quantize_row_q8_K_generic quantize_row_q8_K
#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
#define ggml_vec_dot_q2_0_q8_0_generic ggml_vec_dot_q2_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_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_4x4_generic ggml_quantize_mat_q8_K_4x4
#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_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_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_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
#elif defined(__wasm__)
// quants.c
#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
#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
#define ggml_vec_dot_q2_0_q8_0_generic ggml_vec_dot_q2_0_q8_0
// 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_4x4_generic ggml_quantize_mat_q8_K_4x4
#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_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_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_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
#endif
ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp
+103
View File
@@ -0,0 +1,103 @@
#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_SVE2)
#define HWCAP2_SVE2 (1 << 1)
#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;
bool has_sme2 = 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);
}
if (sysctlbyname("hw.optional.arm.FEAT_SME2", &oldp, &size, NULL, 0) == 0) {
has_sme2 = 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
+4319
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/arch/arm/repack.cpp
+5156
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/arch/loongarch/quants.c
+2309
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
+2304
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/arch/riscv/cpu-feats.cpp
+38
View File
@@ -0,0 +1,38 @@
#include "ggml-backend-impl.h"
#if defined(__riscv) && __riscv_xlen == 64
#include <asm/hwprobe.h>
#include <asm/unistd.h>
#include <unistd.h>
struct riscv64_features {
bool has_rvv = false;
riscv64_features() {
struct riscv_hwprobe probe;
probe.key = RISCV_HWPROBE_KEY_IMA_EXT_0;
probe.value = 0;
int ret = syscall(__NR_riscv_hwprobe, &probe, 1, 0, NULL, 0);
if (0 == ret) {
has_rvv = !!(probe.value & RISCV_HWPROBE_IMA_V);
}
}
};
static int ggml_backend_cpu_riscv64_score() {
int score = 1;
riscv64_features rf;
#ifdef GGML_USE_RVV
if (!rf.has_rvv) { return 0; }
score += 1 << 1;
#endif
return score;
}
GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_riscv64_score)
#endif // __riscv && __riscv_xlen == 64
ggml/src/ggml-cpu/arch/riscv/quants.c
+6596
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/arch/riscv/repack.cpp
+1703
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/arch/s390/cpu-feats.cpp
+50
View File
@@ -0,0 +1,50 @@
#include "ggml-backend-impl.h"
#if defined(__s390x__)
#include <sys/auxv.h>
// find hwcap bits in asm/elf.h
#ifndef HWCAP_VXRS_EXT2
#define HWCAP_VXRS_EXT2 (1 << 15)
#endif
#ifndef HWCAP_NNPA
#define HWCAP_NNPA (1 << 20)
#endif
struct s390x_features {
bool has_vxe2 = false;
bool has_nnpa = false;
s390x_features() {
uint32_t hwcap = getauxval(AT_HWCAP);
// NOTE: use hwcap2 with DFLT for z17 and later
// uint32_t hwcap2 = getauxval(AT_HWCAP2);
has_vxe2 = !!(hwcap & HWCAP_VXRS_EXT2);
has_nnpa = !!(hwcap & HWCAP_NNPA);
}
};
static int ggml_backend_cpu_s390x_score() {
int score = 1;
s390x_features sf;
// IBM z15 / LinuxONE 3
#ifdef GGML_USE_VXE2
if (!sf.has_vxe2) { return 0; }
score += 1 << 1;
#endif
// IBM z16 / LinuxONE 4 and z17 / LinuxONE 5
#ifdef GGML_USE_NNPA
if (!sf.has_nnpa) { return 0; }
score += 1 << 2;
#endif
return score;
}
GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_s390x_score)
#endif // __s390x__
ggml/src/ggml-cpu/arch/s390/quants.c
+1465
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/arch/wasm/quants.c
+1292
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
+4108
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/arch/x86/repack.cpp
+6407
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/binary-ops.cpp
+154
View File
@@ -0,0 +1,154 @@
#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_rows = ggml_is_contiguous_rows(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_rows) {
// 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/cmake/FindSIMD.cmake
+100
View File
@@ -0,0 +1,100 @@
include(CheckCSourceRuns)
set(AVX_CODE "
#include <immintrin.h>
int main()
{
__m256 a;
a = _mm256_set1_ps(0);
return 0;
}
")
set(AVX512_CODE "
#include <immintrin.h>
int main()
{
__m512i a = _mm512_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0);
__m512i b = a;
__mmask64 equality_mask = _mm512_cmp_epi8_mask(a, b, _MM_CMPINT_EQ);
return 0;
}
")
set(AVX2_CODE "
#include <immintrin.h>
int main()
{
__m256i a = {0};
a = _mm256_abs_epi16(a);
__m256i x;
_mm256_extract_epi64(x, 0); // we rely on this in our AVX2 code
return 0;
}
")
set(FMA_CODE "
#include <immintrin.h>
int main()
{
__m256 acc = _mm256_setzero_ps();
const __m256 d = _mm256_setzero_ps();
const __m256 p = _mm256_setzero_ps();
acc = _mm256_fmadd_ps( d, p, acc );
return 0;
}
")
macro(check_sse type flags)
set(__FLAG_I 1)
set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
foreach (__FLAG ${flags})
if (NOT ${type}_FOUND)
set(CMAKE_REQUIRED_FLAGS ${__FLAG})
check_c_source_runs("${${type}_CODE}" HAS_${type}_${__FLAG_I})
if (HAS_${type}_${__FLAG_I})
set(${type}_FOUND TRUE CACHE BOOL "${type} support")
set(${type}_FLAGS "${__FLAG}" CACHE STRING "${type} flags")
endif()
math(EXPR __FLAG_I "${__FLAG_I}+1")
endif()
endforeach()
set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
if (NOT ${type}_FOUND)
set(${type}_FOUND FALSE CACHE BOOL "${type} support")
set(${type}_FLAGS "" CACHE STRING "${type} flags")
endif()
mark_as_advanced(${type}_FOUND ${type}_FLAGS)
endmacro()
# flags are for MSVC only!
check_sse("AVX" " ;/arch:AVX")
if (NOT ${AVX_FOUND})
set(GGML_AVX OFF)
else()
set(GGML_AVX ON)
endif()
check_sse("AVX2" " ;/arch:AVX2")
check_sse("FMA" " ;/arch:AVX2")
if ((NOT ${AVX2_FOUND}) OR (NOT ${FMA_FOUND}))
set(GGML_AVX2 OFF)
else()
set(GGML_AVX2 ON)
endif()
check_sse("AVX512" " ;/arch:AVX512")
if (NOT ${AVX512_FOUND})
set(GGML_AVX512 OFF)
else()
set(GGML_AVX512 ON)
endif()
ggml/src/ggml-cpu/cmake/FindSMTIME.cmake
+32
View File
@@ -0,0 +1,32 @@
include(CheckCSourceRuns)
if (CMAKE_SYSTEM_PROCESSOR MATCHES "^(riscv)" AND GGML_CPU_RISCV64_SPACEMIT)
set(SMT_MARCH_STR "-march=rv64gcv_zfh_zvfh_zba_zicbop")
if (CMAKE_C_COMPILER_ID STREQUAL "GNU" AND
CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 15)
string(APPEND SMT_MARCH_STR "_xsmtvdotii")
endif()
set(CMAKE_REQUIRED_FLAGS "${SMT_MARCH_STR}")
check_c_source_compiles("int main() {__asm__ volatile(\"vmadot v2, v0, v1\");}" SPACEMIT_RISCV_COMPILER_SUPPORT_IME1)
check_c_source_compiles("int main() {__asm__ volatile(\"vmadot v2, v0, v1, i4\");}" SPACEMIT_RISCV_COMPILER_SUPPORT_VMADOT_S4)
check_c_source_compiles("int main() {__asm__ volatile(\"vmadot v2, v0, v1, i8\");}" SPACEMIT_RISCV_COMPILER_SUPPORT_VMADOT_S8)
check_c_source_compiles("int main() {__asm__ volatile(\"vfwmadot v2, v0, v1, fp16\");}" SPACEMIT_RISCV_COMPILER_SUPPORT_VFWMADOT_FP16)
check_c_source_compiles("int main() {__asm__ volatile(\"vmadot.hp v2, v0, v1, v0, 0, i4\");}" SPACEMIT_RISCV_COMPILER_SUPPORT_VFMADOT_S4)
check_c_source_compiles("int main() {__asm__ volatile(\"vmadot.hp v2, v0, v1, v0, 0, i8\");}" SPACEMIT_RISCV_COMPILER_SUPPORT_VFMADOT_S8)
check_c_source_compiles("int main() {__asm__ volatile(\"vmadot1 v2, v0, v1\");}" SPACEMIT_RISCV_COMPILER_SUPPORT_VMADOTN)
check_c_source_compiles("int main() {__asm__ volatile(\"vpack.vv v2, v0, v1, 2\");}" SPACEMIT_RISCV_COMPILER_SUPPORT_VPACK)
check_c_source_compiles("int main() {__asm__ volatile(\"vnspack.vv v2, v0, v1, 2\");}" SPACEMIT_RISCV_COMPILER_SUPPORT_VNPACK)
unset(CMAKE_REQUIRED_FLAGS)
list(APPEND RISCV64_SPACEMIT_IME_SPEC "")
if (SPACEMIT_RISCV_COMPILER_SUPPORT_IME1)
set(RISCV64_SPACEMIT_IME_SPEC "RISCV64_SPACEMIT_IME1")
endif()
if (SPACEMIT_RISCV_COMPILER_SUPPORT_VMADOT_S4 AND SPACEMIT_RISCV_COMPILER_SUPPORT_VPACK AND SPACEMIT_RISCV_COMPILER_SUPPORT_VNPACK)
list(APPEND RISCV64_SPACEMIT_IME_SPEC "RISCV64_SPACEMIT_IME2")
endif()
message("RISCV64_SPACEMIT_IME_SPEC: ${RISCV64_SPACEMIT_IME_SPEC}")
endif()
ggml/src/ggml-cpu/common.h
+95
View File
@@ -0,0 +1,95 @@
#pragma once
#include "ggml.h"
#include "traits.h"
#include "ggml-cpu-impl.h"
#include "ggml-impl.h"
#include "simd-mappings.h"
#define GGML_FA_TILE_Q 64
#define GGML_FA_TILE_KV 64
#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 i32_to_f32(int32_t x) {
return x;
}
static inline int32_t f32_to_i32(float x) {
return 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;
};
template <>
struct type_conversion_table<int32_t> {
static constexpr float (*to_f32)(int32_t) = i32_to_f32;
static constexpr int32_t (*from_f32)(float) = f32_to_i32;
};
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};
}
struct ggml_fa_tile_config {
static constexpr size_t Q = GGML_FA_TILE_Q;
static constexpr size_t KV = GGML_FA_TILE_KV;
};
#endif
ggml/src/ggml-cpu/ggml-cpu-impl.h
+539
View File
@@ -0,0 +1,539 @@
#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;
// use reference implementation
bool use_ref;
};
#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(__ARM_FEATURE_SVE) && defined(__linux__)
#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)
// NOTE: this fallback produces the same total sum as native vdotq_s32 but with different per-lane grouping — do not use when individual lane values matter.
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)
static inline int32x4_t ggml_nvfp4_dot8(const int8x8_t q4_lo, const int8x8_t q8_lo,
const int8x8_t q4_hi, const int8x8_t q8_hi) {
const int16x8_t p_lo = vmull_s8(q4_lo, q8_lo);
const int16x8_t p_hi = vmull_s8(q4_hi, q8_hi);
const int32x4_t sum_lo = vpaddlq_s16(p_lo);
const int32x4_t sum_hi = vpaddlq_s16(p_hi);
return vaddq_s32(sum_lo, sum_hi);
}
#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(__SSE__) || defined(__SSE3__) || defined(__SSSE3__) || defined(__AVX__) || defined(__F16C__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX512BF16__)
#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;
}
/**
* @see https://github.com/ggml-org/llama.cpp/pull/14037
*/
inline static float vec_hsum_f32x4(float32x4_t v) {
float32x4_t v_temp = v + vec_reve(v);
return v_temp[0] + v_temp[1];
}
inline static int32_t vec_hsum_i32x4(int32x4_t v) {
int32x4_t v_temp = v + vec_reve(v);
return v_temp[0] + v_temp[1];
}
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_sx)
/* float type data load instructions */
static __m128 __lsx_vreplfr2vr_s(const float val) {
v4f32 res = {val, val, val, val};
return (__m128)res;
}
#endif
#if defined(__loongarch_asx)
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
+3895
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/ggml-cpu.cpp
+707
View File
@@ -0,0 +1,707 @@
#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
#ifdef GGML_USE_CPU_RISCV64_SPACEMIT
# include "spacemit/ime.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_buffer_types() {
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_RISCV64_SPACEMIT
if (ggml_backend_cpu_riscv64_spacemit_buffer_type()) {
bufts.push_back(ggml_backend_cpu_riscv64_spacemit_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
return bufts;
}();
return bufts;
}
static ggml_backend_buffer_type_t * ggml_backend_cpu_device_get_extra_buffers_type(ggml_backend_dev_t device) {
static std::vector<ggml_backend_buffer_type_t> extra_bufts = [] {
std::vector<ggml_backend_buffer_type_t> bufts = ggml_backend_cpu_get_extra_buffer_types();
bufts.push_back(nullptr);
return bufts;
}();
return extra_bufts.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_buffer_types()) {
if (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;
bool use_ref; // use reference implementation
};
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;
cpu_plan->cplan.use_ref = cpu_ctx->use_ref;
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;
cplan.use_ref = cpu_ctx->use_ref;
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,
/* .set_tensor_2d_async = */ NULL,
/* .get_tensor_2d_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,
/* .graph_optimize = */ 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;
ctx->use_ref = false;
ggml_backend_t cpu_backend = new ggml_backend {
/* .guid = */ ggml_backend_cpu_guid(),
/* .iface = */ 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;
}
void ggml_backend_cpu_set_use_ref(ggml_backend_t backend_cpu, bool use_ref) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
ctx->use_ref = use_ref;
}
// 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" system memory is ill-defined, for practical purposes assume that all of it is free:
*free = *total;
#endif // _WIN32
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;
}
// check extra buffer types
// note: only the first sources are checked for extra buffer types to reduce overhead, increase if necessary
for (int i = 0; i < 4; i++) {
if (op->src[i] && op->src[i]->buffer &&
ggml_backend_cpu_is_extra_buffer_type(op->src[i]->buffer->buft)) {
auto * buf_extra = (ggml::cpu::extra_buffer_type *) op->src[i]->buffer->buft->context;
return buf_extra->supports_op(dev, op);
}
}
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 || src1->type == GGML_TYPE_F16);
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 ||
((src0->type == GGML_TYPE_F16 || 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_sme2()) {
features.push_back({ "SME2", "1" });
}
if (ggml_cpu_has_riscv_v()) {
features.push_back({ "RISCV_V", "1" });
}
if (ggml_cpu_get_rvv_vlen() > 0) {
static std::string rvv_vlen = std::to_string(ggml_cpu_get_rvv_vlen());
features.push_back({ "RVV_VLEN", rvv_vlen.c_str() });
}
if (ggml_cpu_has_vsx()) {
features.push_back({ "VSX", "1" });
}
if (ggml_cpu_has_vxe()) {
features.push_back({ "VXE", "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;
}
if (strcmp(name, "ggml_backend_cpu_set_use_ref") == 0) {
return (void *)ggml_backend_cpu_set_use_ref;
}
// 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
+1130
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/kleidiai/kernels.h
+100
View File
@@ -0,0 +1,100 @@
// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
// SPDX-License-Identifier: MIT
//
#pragma once
#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,
CPU_FEATURE_SME2 = 16
};
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);
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);
size_t (*get_lhs_offset_ex)(size_t m_idx, size_t k, size_t bl);
size_t (*get_rhs_packed_offset_ex)(size_t n_idx, size_t k, size_t bl);
void (*run_kernel_ex)(
size_t m, size_t n, size_t k, size_t bl,
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);
};
struct lhs_packing_info {
size_t (*get_offset)(size_t m_idx, size_t lhs_stride);
size_t (*get_packed_offset_ex)(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr);
size_t (*packed_size_ex)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr);
void (*pack_func_ex)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr,
size_t m_idx_start, const void * lhs, size_t lhs_stride, void * lhs_packed);
};
enum rhs_repack_mode {
RHS_REPACK_PER_KERNEL,
RHS_REPACK_SHARED,
RHS_REPACK_SINGLE_ONLY,
};
struct rhs_packing_info {
size_t (*packed_stride)(size_t k, size_t nr, size_t kr, size_t bl);
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);
size_t (*packed_size_ex)(size_t n, size_t k, size_t nr, size_t kr, size_t bl);
size_t (*packed_stride_ex)(size_t k, size_t nr, size_t kr, size_t bl);
void (*pack_func_ex)(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl,
size_t rhs_stride, const void * rhs, const void * bias, const void * scale, void * rhs_packed, size_t extra_bytes, const void * params);
rhs_repack_mode repack_mode = RHS_REPACK_PER_KERNEL;
};
struct ggml_kleidiai_kernels {
kernel_info gemm;
lhs_packing_info gemm_lhs_info;
kernel_info gemv;
lhs_packing_info gemv_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_kleidiai_kernels * ggml_kleidiai_select_kernels_q8_0(cpu_feature features);
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_f32(cpu_feature features);
ggml/src/ggml-cpu/kleidiai/kleidiai.cpp
+1782
View File
File diff suppressed because it is too large Load Diff
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
+4164
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/llamafile/sgemm.h
+25
View File
@@ -0,0 +1,25 @@
#pragma once
#include <stdint.h>
#include <stdbool.h>
#if defined(__VXE__) || defined(__VXE2__)
#include <vecintrin.h>
#endif
#ifdef _MSC_VER
#define NOINLINE __declspec(noinline)
#else
#define NOINLINE __attribute__((__noinline__))
#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
+12004
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/ops.h
+124
View File
@@ -0,0 +1,124 @@
#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_add_id(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_cumsum(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_mul_fused(const struct ggml_compute_params * params, struct ggml_tensor * dst_rms_norm, struct ggml_tensor * dst_mul);
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_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_im2col_3d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_col2im_1d(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_3d(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_top_k(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_tri(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_fill(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_flash_attn_ext(const struct ggml_compute_params * params, 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_solve_tri(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_gated_delta_net(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_lightning_indexer(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_dsv4_hc_comb(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_dsv4_hc_pre(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_dsv4_hc_post(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);
void ggml_compute_forward_fwht(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_opt_step_sgd(const struct ggml_compute_params * params, struct ggml_tensor * dst);
#ifdef __cplusplus
}
#endif
ggml/src/ggml-cpu/quants.c
+1339
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/quants.h
+106
View File
@@ -0,0 +1,106 @@
#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_q1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
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_mxfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_nvfp4(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_q1_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_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_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_mxfp4_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_nvfp4_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_q1_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_q2_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_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_mxfp4_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_nvfp4_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
+4836
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/repack.h
+245
View File
@@ -0,0 +1,245 @@
#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<4, 16>) == 16 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<4,16> 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");
static_assert(sizeof(block<8, 16>) == 16 * sizeof(ggml_half) + QK8_0 * 16, "wrong block<8,16> size/padding");
using block_q4_0x4 = block<4, 4>;
using block_q4_0x8 = block<4, 8>;
using block_q4_0x16 = block<4, 16>;
using block_q8_0x4 = block<8, 4>;
using block_q8_0x8 = block<8, 8>;
using block_q8_0x16 = block<8, 16>;
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_q4_Kx16 {
ggml_half d[16]; // super-block scale for quantized scales
ggml_half dmin[16]; // super-block scale for quantized mins
uint8_t scales[192]; // scales and mins, quantized with 6 bits
uint8_t qs[2048]; // 4--bit quants
};
static_assert(sizeof(block_q4_Kx16) == sizeof(ggml_half) * 32 + K_SCALE_SIZE * 16 + QK_K * 8, "wrong q4_K block size/padding");
struct block_q2_Kx8 {
ggml_half d[8]; // super-block scale for quantized scales
ggml_half dmin[8]; // super-block scale for quantized mins
uint8_t scales[128]; // scales and mins, quantized with 4 bits
uint8_t qs[512]; // 2--bit quants
};
static_assert(sizeof(block_q2_Kx8) == sizeof(ggml_half) * 16 + QK_K/2 + QK_K * 2, "wrong q2_K block size/padding");
struct block_q2_Kx16 {
ggml_half d[16]; // Super-block scale for quantized scales
ggml_half dmin[16]; // Super-block scale for quantized mins
uint8_t scales[256]; // Sub-block scales (16 cols * 16 sub-blocks)
uint8_t qs[1024]; // Data (16 cols * 64 bytes per block)
};
static_assert(sizeof(block_q2_Kx16) == sizeof(ggml_half) * 32 + QK_K + QK_K * 4, "wrong q2_K block size/padding");
struct block_q5_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 qh[QK_K * 8 / 8]; // high bits of 5-bit quants
uint8_t qs[QK_K * 8 / 2]; // low bits of 5-bit quants (in groups of 4)
};
static_assert(sizeof(block_q5_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 5,
"wrong q5_K block size/padding");
struct block_q6_Kx8 {
ggml_half d[8];
int8_t scales[QK_K / 16 * 8];
uint8_t ql[QK_K / 2 * 8]; // low bits of 6-bit quants (groups of 2)
uint8_t qh[QK_K / 4 * 8]; // high bits of 6-bit quants (groups of 4)
};
static_assert(sizeof(block_q6_Kx8) == sizeof(ggml_half) * 8 + QK_K / 16 * 8 + 3 * QK_K / 4 * 8,
"wrong q6_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");
struct block_iq4_nlx8 {
ggml_half d[8]; // deltas for 8 iq4_nl blocks
uint8_t qs[QK4_NL * 4]; // nibbles / quants for 8 iq4_nl blocks
};
static_assert(sizeof(block_iq4_nlx8) == 8 * sizeof(ggml_half) + QK4_NL * 4, "wrong iq4_nlx8 block size/padding");
struct block_iq4_nlx16 {
ggml_half d[16]; // deltas for 16 iq4_nl blocks
uint8_t qs[QK4_NL * 8]; // nibbles / quants for 16 iq4_nl blocks
};
static_assert(sizeof(block_iq4_nlx16) == 16 * sizeof(ggml_half) + QK4_NL * 8, "wrong iq4_nlx16 block size/padding");
struct block_mxfp4x4 {
uint8_t e[4];
uint8_t qs[QK_MXFP4 * 2];
};
static_assert(sizeof(block_mxfp4x4) == 4 + QK_MXFP4 * 2, "wrong mxfp4x4 block size/padding");
struct block_mxfp4x8 {
uint8_t e[8];
uint8_t qs[QK_MXFP4 * 4];
};
static_assert(sizeof(block_mxfp4x8) == 8 + QK_MXFP4 * 4, "wrong mxfp4x8 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_4x4(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_q2_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_q4_K_8x4_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_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_q5_K_8x4_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_q5_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_q6_K_8x4_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_q6_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_gemv_iq4_nl_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_mxfp4_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_mxfp4_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_q8_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_q8_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_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_q2_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_q4_K_8x4_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_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_q5_K_8x4_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_q5_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_q6_K_8x4_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_q6_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);
void ggml_gemm_iq4_nl_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_mxfp4_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_mxfp4_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_q8_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_q8_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);
#if defined __riscv_zvfh
void ggml_quantize_mat_q8_0_4x1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_quantize_mat_q8_K_4x1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_gemv_q4_0_16x1_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_16x1_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_16x1_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_q8_0_16x1_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_q2_K_16x1_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_q4_0_16x1_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_16x1_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_16x1_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_q8_0_16x1_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_q2_K_16x1_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);
#endif
// 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_4x4_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_q2_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_q4_K_8x4_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_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_q5_K_8x4_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_q5_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_q6_K_8x4_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_q6_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_gemv_iq4_nl_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_mxfp4_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_mxfp4_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_q8_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_q8_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_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_q2_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_q4_K_8x4_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_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_q5_K_8x4_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_q5_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_q6_K_8x4_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_q6_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);
void ggml_gemm_iq4_nl_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_mxfp4_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_mxfp4_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_q8_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_q8_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);
#if defined __riscv_zvfh
void ggml_quantize_mat_q8_0_4x1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_quantize_mat_q8_K_4x1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_gemv_q4_0_16x1_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_16x1_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_q8_0_16x1_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_q2_K_16x1_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_16x1_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_16x1_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_16x1_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_q8_0_16x1_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_q2_K_16x1_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_16x1_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);
#endif
#if defined(__cplusplus)
} // extern "C"
#endif
ggml/src/ggml-cpu/simd-gemm.h
+226
View File
@@ -0,0 +1,226 @@
#pragma once
// Computes C[M x N] += A[M x K] * B[K x N]
#include "simd-mappings.h"
// TODO: add support for sizeless vector types
#if defined(GGML_SIMD) && !defined(__ARM_FEATURE_SVE) && !defined(__riscv_v_intrinsic)
// TODO: untested on avx512
// These are in units of GGML_F32_EPR
#if defined(__AVX512F__) || defined (__ARM_NEON__)
static constexpr int GEMM_RM = 4;
static constexpr int GEMM_RN = 4; // 16+4+1 = 25/32
#elif defined(__AVX2__) || defined(__AVX__)
static constexpr int GEMM_RM = 6;
static constexpr int GEMM_RN = 2; // 12+2+1 = 15/16
#else
static constexpr int GEMM_RM = 2;
static constexpr int GEMM_RN = 2;
#endif
template <int RM, int RN>
static inline void simd_gemm_ukernel(
float * GGML_RESTRICT C,
const float * GGML_RESTRICT A,
const float * GGML_RESTRICT B,
int K, int N)
{
static constexpr int KN = GGML_F32_EPR;
GGML_F32_VEC acc[RM][RN];
for (int64_t i = 0; i < RM; i++) {
for (int r = 0; r < RN; r++) {
acc[i][r] = GGML_F32_VEC_LOAD(C + i * N + r * KN);
}
}
for (int64_t kk = 0; kk < K; kk++) {
GGML_F32_VEC Bv[RN];
for (int r = 0; r < RN; r++) {
Bv[r] = GGML_F32_VEC_LOAD(B + kk * N + r * KN);
}
for (int64_t i = 0; i < RM; i++) {
GGML_F32_VEC p = GGML_F32_VEC_SET1(A[i * K + kk]);
for (int r = 0; r < RN; r++) {
acc[i][r] = GGML_F32_VEC_FMA(acc[i][r], Bv[r], p);
}
}
}
for (int64_t i = 0; i < RM; i++) {
for (int r = 0; r < RN; r++) {
GGML_F32_VEC_STORE(C + i * N + r * KN, acc[i][r]);
}
}
}
// C[M x N] += A[M x K] * B[K x N]
static void simd_gemm(
float * GGML_RESTRICT C,
const float * GGML_RESTRICT A,
const float * GGML_RESTRICT B,
int M, int K, int N)
{
static constexpr int KN = GGML_F32_EPR;
int64_t ii = 0;
for (; ii + GEMM_RM <= M; ii += GEMM_RM) {
int64_t jj = 0;
for (; jj + GEMM_RN * KN <= N; jj += GEMM_RN * KN) {
simd_gemm_ukernel<GEMM_RM, GEMM_RN>(C + jj, A, B + jj, K, N);
}
for (; jj + KN <= N; jj += KN) {
simd_gemm_ukernel<GEMM_RM, 1>(C + jj, A, B + jj, K, N);
}
for (; jj < N; jj++) {
for (int64_t i = 0; i < GEMM_RM; i++) {
float a = C[i * N + jj];
for (int64_t kk = 0; kk < K; kk++) {
a += A[i * K + kk] * B[kk * N + jj];
}
C[i * N + jj] = a;
}
}
A += GEMM_RM * K;
C += GEMM_RM * N;
}
// Tail rows: one at a time
for (; ii < M; ii++) {
int64_t jj = 0;
for (; jj + GEMM_RN * KN <= N; jj += GEMM_RN * KN) {
simd_gemm_ukernel<1, GEMM_RN>(C + jj, A, B + jj, K, N);
}
for (; jj + KN <= N; jj += KN) {
simd_gemm_ukernel<1, 1>(C + jj, A, B + jj, K, N);
}
for (; jj < N; jj++) {
float a = C[jj];
for (int64_t kk = 0; kk < K; kk++) {
a += A[kk] * B[kk * N + jj];
}
C[jj] = a;
}
A += K;
C += N;
}
}
#elif defined(GGML_SIMD) && defined(__riscv_v_intrinsic)
// RM accumulators + 1 B vector = RM + 1 <= 8 => RM <= 7
// Microkernel: C[RM x vl] += A[RM x K] * B[K x N]
template <int RM>
static inline void rvv_simd_gemm_ukernel(
float * GGML_RESTRICT C,
const float * GGML_RESTRICT A,
const float * GGML_RESTRICT B,
int K, int N, size_t vl)
{
static_assert(RM >= 1 && RM <= 7, "RM must be 1..7 for LMUL=4");
vfloat32m4_t acc_0 = __riscv_vle32_v_f32m4(C + 0 * N, vl);
vfloat32m4_t acc_1, acc_2, acc_3, acc_4, acc_5, acc_6;
if constexpr (RM > 1) acc_1 = __riscv_vle32_v_f32m4(C + 1 * N, vl);
if constexpr (RM > 2) acc_2 = __riscv_vle32_v_f32m4(C + 2 * N, vl);
if constexpr (RM > 3) acc_3 = __riscv_vle32_v_f32m4(C + 3 * N, vl);
if constexpr (RM > 4) acc_4 = __riscv_vle32_v_f32m4(C + 4 * N, vl);
if constexpr (RM > 5) acc_5 = __riscv_vle32_v_f32m4(C + 5 * N, vl);
if constexpr (RM > 6) acc_6 = __riscv_vle32_v_f32m4(C + 6 * N, vl);
for (int kk = 0; kk < K; kk++) {
vfloat32m4_t b_0 = __riscv_vle32_v_f32m4(B + kk * N, vl);
acc_0 = __riscv_vfmacc_vf_f32m4(acc_0, A[0 * K + kk], b_0, vl);
if constexpr (RM > 1) acc_1 = __riscv_vfmacc_vf_f32m4(acc_1, A[1 * K + kk], b_0, vl);
if constexpr (RM > 2) acc_2 = __riscv_vfmacc_vf_f32m4(acc_2, A[2 * K + kk], b_0, vl);
if constexpr (RM > 3) acc_3 = __riscv_vfmacc_vf_f32m4(acc_3, A[3 * K + kk], b_0, vl);
if constexpr (RM > 4) acc_4 = __riscv_vfmacc_vf_f32m4(acc_4, A[4 * K + kk], b_0, vl);
if constexpr (RM > 5) acc_5 = __riscv_vfmacc_vf_f32m4(acc_5, A[5 * K + kk], b_0, vl);
if constexpr (RM > 6) acc_6 = __riscv_vfmacc_vf_f32m4(acc_6, A[6 * K + kk], b_0, vl);
}
__riscv_vse32_v_f32m4(C + 0 * N, acc_0, vl);
if constexpr (RM > 1) __riscv_vse32_v_f32m4(C + 1 * N, acc_1, vl);
if constexpr (RM > 2) __riscv_vse32_v_f32m4(C + 2 * N, acc_2, vl);
if constexpr (RM > 3) __riscv_vse32_v_f32m4(C + 3 * N, acc_3, vl);
if constexpr (RM > 4) __riscv_vse32_v_f32m4(C + 4 * N, acc_4, vl);
if constexpr (RM > 5) __riscv_vse32_v_f32m4(C + 5 * N, acc_5, vl);
if constexpr (RM > 6) __riscv_vse32_v_f32m4(C + 6 * N, acc_6, vl);
}
template <int RM>
static inline void rvv_simd_gemm_dispatch_tail(
float * GGML_RESTRICT C,
const float * GGML_RESTRICT A,
const float * GGML_RESTRICT B,
int K, int N, int KN, int remaining_rows)
{
if constexpr (RM > 0) {
if (remaining_rows == RM) {
int64_t jj = 0;
for (; jj + KN <= N; jj += KN) {
rvv_simd_gemm_ukernel<RM>(C + jj, A, B + jj, K, N, KN);
}
if (jj < N) {
rvv_simd_gemm_ukernel<RM>(C + jj, A, B + jj, K, N, N - jj);
}
} else {
rvv_simd_gemm_dispatch_tail<RM - 1>(C, A, B, K, N, KN, remaining_rows);
}
}
}
static constexpr int GEMM_RM = 7;
// C[M x N] += A[M x K] * B[K x N]
static void simd_gemm(
float * GGML_RESTRICT C,
const float * GGML_RESTRICT A,
const float * GGML_RESTRICT B,
int M, int K, int N)
{
const int KN = (int)__riscv_vlenb();
int64_t ii = 0;
for (; ii + GEMM_RM <= M; ii += GEMM_RM) {
int64_t jj = 0;
for (; jj + KN <= N; jj += KN) {
rvv_simd_gemm_ukernel<GEMM_RM>(C + jj, A, B + jj, K, N, KN);
}
if (jj < N) {
rvv_simd_gemm_ukernel<GEMM_RM>(C + jj, A, B + jj, K, N, N - jj);
}
A += GEMM_RM * K;
C += GEMM_RM * N;
}
int remaining_rows = M - ii;
rvv_simd_gemm_dispatch_tail<GEMM_RM - 1>(C, A, B, K, N, KN, remaining_rows);
}
#if defined(__GNUC__) && !defined(__clang__)
#pragma GCC diagnostic pop
#endif
#else // scalar path
static void simd_gemm(
float * GGML_RESTRICT C,
const float * GGML_RESTRICT A,
const float * GGML_RESTRICT B,
int M, int K, int N)
{
for (int64_t i = 0; i < M; i++) {
for (int64_t j = 0; j < N; j++) {
float sum = C[i * N + j];
for (int64_t kk = 0; kk < K; kk++) {
sum += A[i * K + kk] * B[kk * N + j];
}
C[i * N + j] = sum;
}
}
}
#endif // GGML_SIMD
ggml/src/ggml-cpu/simd-mappings.h
+1317
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/spacemit/ime.cpp
+1740
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/spacemit/ime.h
+21
View File
@@ -0,0 +1,21 @@
#pragma once
#include "ggml-alloc.h"
#ifdef __cplusplus
extern "C" {
#endif
ggml_backend_buffer_type_t ggml_backend_cpu_riscv64_spacemit_buffer_type(void);
void ggml_backend_cpu_riscv64_spacemit_set_numa_thread_affinity(int thread_n);
void ggml_backend_cpu_riscv64_spacemit_clear_numa_thread_affinity_threaded(int thread_n);
void * ggml_backend_cpu_riscv64_spacemit_alloc_shared(size_t size, size_t alignment);
void ggml_backend_cpu_riscv64_spacemit_free_shared(void * ptr);
#ifdef __cplusplus
}
#endif
ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp
+1027
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/spacemit/ime2_kernels.cpp
+5768
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/spacemit/ime_env.cpp
+320
View File
@@ -0,0 +1,320 @@
#include "ime_env.h"
#include "ggml-impl.h"
#include "spine_mem_pool.h"
#include <fcntl.h>
#include <unistd.h>
#include <algorithm>
#include <array>
#include <cctype>
#include <fstream>
#include <string>
#include <thread>
#include <unordered_map>
namespace ggml::cpu::riscv64_spacemit {
bool spine_core_info::get_spine_core_info(std::vector<spine_core_info> & result) {
static std::unordered_map<uint64_t, spine_core_arch_id> spine_march_mapping_ = {
{0x8000000058000001, spine_core_arch_id::core_arch_x60 },
{ 0x8000000041000001, spine_core_arch_id::core_arch_a60 },
{ 0x8000000058000002, spine_core_arch_id::core_arch_x100},
{ 0x8000000041000002, spine_core_arch_id::core_arch_a100},
};
result.clear();
std::ifstream file("/proc/cpuinfo");
std::string line;
std::vector<std::array<uint64_t, 2>> cpu_info_list;
uint64_t current_processor = spine_invalid_core_id;
uint64_t current_marchid = 0;
bool has_processor = false;
bool has_marchid = false;
if (!file.is_open()) {
return false;
}
while (std::getline(file, line)) {
if (line.substr(0, 9) == "processor") {
if (has_processor && has_marchid) {
cpu_info_list.push_back({ current_processor, current_marchid });
}
size_t colon_pos = line.find(':');
if (colon_pos != std::string::npos) {
current_processor = std::stoi(line.substr(colon_pos + 1));
has_processor = true;
}
has_marchid = false;
} else if (line.substr(0, 7) == "marchid") {
size_t colon_pos = line.find(':');
if (colon_pos != std::string::npos) {
std::string marchid_str = line.substr(colon_pos + 1);
marchid_str.erase(std::remove_if(marchid_str.begin(), marchid_str.end(), isspace), marchid_str.end());
current_marchid = std::stoull(marchid_str, nullptr, 16);
has_marchid = true;
}
}
}
if (has_processor && has_marchid) {
cpu_info_list.push_back({ current_processor, current_marchid });
}
if (has_processor && has_marchid) {
for (auto & cpu_info : cpu_info_list) {
if (cpu_info[0] != spine_invalid_core_id &&
spine_march_mapping_.find(cpu_info[1]) != spine_march_mapping_.end()) {
auto core_info = spine_core_info();
core_info.core_id = cpu_info[0];
core_info.arch_id = spine_core_arch_id(spine_march_mapping_[cpu_info[1]]);
result.push_back(core_info);
}
}
}
return has_processor && has_marchid;
}
namespace {
uint16_t hex_string_to_u16(const std::string & hex_str) {
try {
size_t pos = 0;
if (hex_str.substr(0, 2) == "0x" || hex_str.substr(0, 2) == "0X") {
pos = 2;
}
unsigned long result = std::stoul(hex_str.substr(pos), nullptr, 16);
if (result > std::numeric_limits<uint16_t>::max()) {
throw std::out_of_range("Converted value is out of range for uint16_t");
}
return static_cast<uint16_t>(result);
} catch (const std::invalid_argument & e) {
throw std::invalid_argument("Invalid hexadecimal string");
} catch (const std::out_of_range & e) {
throw;
}
}
const char * spine_mem_pool_backend_to_string(spine_mem_pool_backend backend) {
switch (backend) {
case spine_mem_pool_backend::none:
return "NONE";
case spine_mem_pool_backend::posix_memalign:
return "POSIX";
case spine_mem_pool_backend::transparent_hugepage:
return "HPAGE";
case spine_mem_pool_backend::hugetlb_1g:
return "HPAGE1GB";
}
return "unknown";
}
spine_mem_pool_backend parse_mem_backend(const char * mem_backend_str) {
if (mem_backend_str == nullptr || mem_backend_str[0] == '\0') {
return spine_mem_pool_backend::transparent_hugepage;
}
std::string value(mem_backend_str);
std::transform(value.begin(), value.end(), value.begin(),
[](unsigned char ch) { return static_cast<char>(std::tolower(ch)); });
if (value == "none") {
return spine_mem_pool_backend::none;
}
if (value == "posix") {
return spine_mem_pool_backend::posix_memalign;
}
if (value == "hpage") {
return spine_mem_pool_backend::transparent_hugepage;
}
if (value == "hpage1gb") {
return spine_mem_pool_backend::hugetlb_1g;
}
throw std::runtime_error("invalid SPACEMIT_MEM_BACKEND: " + value + ", expected NONE, POSIX, HPAGE or HPAGE1GB");
}
} // namespace
spine_env_info::spine_env_info() {
num_cores = static_cast<int>(std::thread::hardware_concurrency());
spine_core_info::get_spine_core_info(core_info_list);
// special for x60 K1
if (core_info_list.size() == 8 && core_info_list[0].arch_id == spine_core_arch_id::core_arch_x60) {
for (int i = 0; i < 4; i++) {
core_info_list[i].arch_id = spine_core_arch_id::core_arch_a60;
}
}
// special for qemu
if (core_info_list.size() == 0) {
char * spine_core_arch_str = getenv("SPACEMIT_CORE_ARCH");
if (spine_core_arch_str != nullptr) {
auto arch_id = hex_string_to_u16(spine_core_arch_str);
for (int i = 0; i < num_cores; i++) {
auto core_info = spine_core_info();
core_info.core_id = i;
core_info.arch_id = spine_core_arch_id{ arch_id };
core_info_list.push_back(core_info);
}
}
}
if (core_info_list.size() == 0) {
throw std::runtime_error(
"Failed to get SPACEMIT_CORE_ARCH from environment or failed to parse it from /proc/cpuinfo");
}
char * spine_perfer_core_arch_str = getenv("SPACEMIT_PERFER_CORE_ARCH");
if (spine_perfer_core_arch_str != nullptr && spine_perfer_core_arch_str != "") {
perfer_core_arch_id = spine_core_arch_id{ hex_string_to_u16(spine_perfer_core_arch_str) };
}
char * spine_perfer_core_id_str = getenv("SPACEMIT_PERFER_CORE_ID");
std::vector<int> perfer_core_id_vec;
if (spine_perfer_core_id_str != nullptr && spine_perfer_core_id_str != "") {
std::string perfer_core_id_str(spine_perfer_core_id_str);
size_t start = 0;
size_t end = 0;
while ((end = perfer_core_id_str.find(',', start)) != std::string::npos) {
std::string core_id_substr = perfer_core_id_str.substr(start, end - start);
perfer_core_id_vec.push_back(std::stoi(core_id_substr));
start = end + 1;
}
std::string core_id_substr = perfer_core_id_str.substr(start);
perfer_core_id_vec.push_back(std::stoi(core_id_substr));
}
perfer_core_ids.reserve(num_cores);
if (perfer_core_arch_id == spine_core_arch_id::core_arch_none) {
for (auto & core_info : core_info_list) {
auto core_arch_id = core_info.arch_id;
auto core_arch_head = (uint16_t) (core_arch_id) >> 12;
if (core_arch_head == 0xA) {
num_perfer_cores++;
perfer_core_arch_id = core_arch_id;
cpu_mask |= (1ULL << core_info.core_id);
perfer_core_ids.push_back(core_info.core_id);
}
}
} else {
for (auto & core_info : core_info_list) {
auto core_arch_id = core_info.arch_id;
if (core_arch_id == perfer_core_arch_id) {
num_perfer_cores++;
cpu_mask |= (1ULL << core_info.core_id);
auto core_arch_head = (uint16_t) (core_arch_id) >> 12;
if (core_arch_head == 0xA) {
perfer_core_ids.push_back(core_info.core_id);
}
}
}
if (num_perfer_cores == 0) {
GGML_ABORT("can not find core with arch id %x for SPACEMIT_PERFER_CORE_ARCH in core info list\n",
(uint16_t) perfer_core_arch_id);
}
}
if (perfer_core_id_vec.size() > 0) {
perfer_core_ids.clear();
cpu_mask = 0;
num_perfer_cores = 0;
for (int core_id : perfer_core_id_vec) {
if (core_id < 0 || core_id >= num_cores) {
GGML_ABORT("invalid core id in SPACEMIT_PERFER_CORE_ID: %d, should be between 0 and %d\n", core_id,
num_cores - 1);
}
auto core_info = core_info_list[core_id];
auto core_arch_id = core_info.arch_id;
if (core_arch_id == perfer_core_arch_id) {
cpu_mask |= (1ULL << core_id);
perfer_core_ids.push_back(core_id);
} else {
GGML_ABORT(
"core id %d in SPACEMIT_PERFER_CORE_ID has arch id %x which does not match "
"SPACEMIT_PERFER_CORE_ARCH %x\n",
core_id, (uint16_t) core_arch_id, (uint16_t) perfer_core_arch_id);
}
}
std::string perfer_core_id_vec_str;
for (int core_id : perfer_core_id_vec) {
perfer_core_id_vec_str += std::to_string(core_id) + ",";
}
perfer_core_id_vec_str.pop_back();
GGML_LOG_DEBUG("SPACEMIT_PERFER_CORE_ID is set, perferred core ids: %s\n", perfer_core_id_vec_str.c_str());
num_perfer_cores = static_cast<int>(perfer_core_id_vec.size());
}
use_ime1 = perfer_core_arch_id == spine_core_arch_id::core_arch_a60 ||
perfer_core_arch_id == spine_core_arch_id::core_arch_x100;
use_ime2 = perfer_core_arch_id == spine_core_arch_id::core_arch_a100;
mem_backend = parse_mem_backend(getenv("SPACEMIT_MEM_BACKEND"));
char * spine_disable_tcm_str = getenv("SPACEMIT_DISABLE_TCM");
auto user_disable_tcm = spine_disable_tcm_str != nullptr && strcmp(spine_disable_tcm_str, "0") != 0;
if (!user_disable_tcm) {
spine_mem_pool_tcm_info tcm_info;
if (spine_mem_pool_tcm_init(&tcm_info)) {
use_tcm = tcm_info.available;
tcm_blk_size = tcm_info.blk_size;
GGML_LOG_DEBUG("CPU_RISCV64_SPACEMIT: tcm is available, blk_size: %zu, blk_num: %zu, is_fake_tcm: %d\n",
tcm_info.blk_size, tcm_info.blk_num, tcm_info.is_fake_tcm);
for (auto & core_info : core_info_list) {
auto core_arch_head = (uint16_t) (core_info.arch_id) >> 12;
if (core_arch_head != 0xA) {
aicpu_id_offset++;
} else {
break;
}
}
}
}
GGML_LOG_DEBUG(
"CPU_RISCV64_SPACEMIT: num_cores: %d, num_perfer_cores: %d, perfer_core_arch_id: %x, exclude_main_thread: %d, "
"use_ime1: %d, use_ime2: %d, mem_backend: %s, cpu_mask: %lx, aicpu_id_offset: %d\n",
num_cores, num_perfer_cores, (uint16_t) perfer_core_arch_id, exclude_main_thread, use_ime1, use_ime2,
spine_mem_pool_backend_to_string(mem_backend), cpu_mask, aicpu_id_offset);
const size_t init_barrier_size = sizeof(spine_barrier_t) * spine_init_barrier_count;
init_barrier =
static_cast<spine_barrier_t *>(spine_mem_pool_shared_mem_alloc(init_barrier_size, alignof(spine_barrier_t)));
if (init_barrier != nullptr) {
init_barrier_is_shared_mem = true;
} else {
GGML_LOG_WARN("CPU_RISCV64_SPACEMIT: failed to allocate init_barrier from shared mem, falling back to heap\n",
__func__);
init_barrier = new spine_barrier_t[spine_init_barrier_count];
}
spine_barrier_init(init_barrier, spine_init_barrier_count, 2);
}
spine_env_info::~spine_env_info() {
if (init_barrier_is_shared_mem) {
spine_mem_pool_shared_mem_free(init_barrier);
} else {
delete[] init_barrier;
}
init_barrier = nullptr;
init_barrier_is_shared_mem = false;
}
spine_env_info global_spine_env_info;
} // namespace ggml::cpu::riscv64_spacemit
ggml/src/ggml-cpu/spacemit/ime_env.h
+55
View File
@@ -0,0 +1,55 @@
#pragma once
#include "spine_barrier.h"
#include "spine_mem_pool.h"
#include <cstddef>
#include <cstdint>
#include <vector>
namespace ggml::cpu::riscv64_spacemit {
constexpr uint64_t spine_invalid_core_id = 0xFFFFFFFF;
constexpr size_t spine_init_barrier_count = 16;
enum class spine_core_arch_id : uint16_t {
core_arch_none = 0,
core_arch_x60 = 0x503C,
core_arch_x100 = 0x5064,
core_arch_x200 = 0x50C8,
core_arch_a60 = 0xA03C,
core_arch_a100 = 0xA064,
core_arch_a200 = 0xA0C8,
};
struct spine_core_info {
uint64_t core_id{ spine_invalid_core_id };
spine_core_arch_id arch_id{ spine_core_arch_id::core_arch_none };
static bool get_spine_core_info(std::vector<spine_core_info> & result);
};
struct spine_env_info {
std::vector<spine_core_info> core_info_list;
std::vector<int> perfer_core_ids;
int aicpu_id_offset{ 0 };
int num_cores{ 0 };
int num_perfer_cores{ 0 };
spine_core_arch_id perfer_core_arch_id{ spine_core_arch_id::core_arch_none };
bool exclude_main_thread{ false };
bool use_ime2{ false };
bool use_ime1{ false };
bool use_tcm{ false };
spine_mem_pool_backend mem_backend{ spine_mem_pool_backend::transparent_hugepage };
uint64_t tcm_blk_size{ 0 };
uint64_t cpu_mask{ 0 };
spine_barrier_t * init_barrier{ nullptr };
bool init_barrier_is_shared_mem{ false };
spine_env_info();
~spine_env_info();
};
extern spine_env_info global_spine_env_info;
} // namespace ggml::cpu::riscv64_spacemit
ggml/src/ggml-cpu/spacemit/ime_kernels.h
+189
View File
@@ -0,0 +1,189 @@
#pragma once
#include <cassert>
#include <cstddef>
#include <functional>
namespace spacemit_kernels {
#define BLOCK_QNK_LEN 256
template <int N> struct nrow_block_q2_k {
// [4bit scale + 4bit zp] * N * 16
uint8_t scales[N * BLOCK_QNK_LEN / 16];
// [b0, b16, b32, b48] [b1, b17, b33, b49] ... [b15, b31, b47, b63]
// [b64, b80, b96, b112] ...[b79, b95, b111, b127]
// [b128, b144, b160, b176] ...[b143, b159, b175, b191]
// [b192, b208, b224, b240] ...[b207, b223, b239, b255]
uint8_t qs[N * BLOCK_QNK_LEN / 4];
uint16_t scales16[N];
uint16_t zeros16[N];
};
template <int N> struct nrow_block_q3_k {
// [8bit scale] * N * 16
int8_t scales[N * 16];
// [b0, b1, b2, b3, b4, b5, b6, b7] ... [b248, b249, b250, b251, b252, b253, b254, b255]
uint8_t hmask[N * BLOCK_QNK_LEN / 8];
// [b0, b16, b32, b48] [b1, b17, b33, b49] ... [b15, b31, b47, b63]
// [b64, b80, b96, b112] ...[b79, b95, b111, b127]
// [b128, b144, b160, b176] ...[b143, b159, b175, b191]
// [b192, b208, b224, b240] ...[b207, b223, b239, b255]
uint8_t qs[N * BLOCK_QNK_LEN / 4];
uint16_t scales16[N];
};
template <int N> struct nrow_block_mxfp4 {
uint8_t e[N];
uint8_t qh[4 * N];
uint8_t qs[16 * N];
};
template <int N> struct __attribute__((packed)) nrow_block_q5_1 {
uint16_t scales16[N];
uint8_t zp[N];
// n0 [bh0, bh1, bh2, bh3, bh4, bh5, bh6, bh7] ....
uint8_t qh[4 * N];
// n0 [b0, b1], [b2, b3] .... [b30, b31]
// n1 [b0, b1], [b2, b3] .... [b30, b31]
uint8_t qs[16 * N];
};
static_assert(sizeof(nrow_block_q5_1<1>) == sizeof(uint8_t) + 22, "wrong nrow_block_q5_1 block size/padding");
template <int N> struct __attribute__((packed)) nrow_block_q5_0 {
uint16_t scales16[N];
// n0 [bh0, bh1, bh2, bh3, bh4, bh5, bh6, bh7] ....
uint8_t qh[4 * N];
// n0 [b0, b1], [b2, b3] .... [b30, b31]
// n1 [b0, b1], [b2, b3] .... [b30, b31]
uint8_t qs[16 * N];
};
static_assert(sizeof(nrow_block_q5_0<1>) == 22, "wrong nrow_block_q5_0 block size/padding");
using gemm_kernel_quantize_def = std::function<
size_t(size_t, const uint8_t *, const uint8_t *, const uint8_t *, float *, size_t, size_t, size_t, size_t)>;
using moe_gemm_kernel_quantize_def = std::function<
size_t(size_t, const uint8_t **, const uint8_t *, const uint8_t *, float **, size_t, size_t, size_t, size_t)>;
namespace ime1 {
size_t gemm_kernel_i8i4(size_t blk_len,
const uint8_t * quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float * c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
void quantize_a_row_i8(size_t blk_len, const float * a_ptr, size_t count_k, uint8_t * quant_a_ptr);
void quantize_a_4row_i8(size_t blk_len, const float * a_ptr, size_t count_k, uint8_t * quant_a_ptr);
} // namespace ime1
namespace ime2 {
size_t gemm_kernel_i8i2k(size_t blk_len,
const uint8_t * quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float * c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
size_t gemm_kernel_i8i3k(size_t blk_len,
const uint8_t * quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float * c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
size_t gemm_kernel_i8i4(size_t blk_len,
const uint8_t * quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float * c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
size_t gemm_kernel_i8i4_hp(size_t blk_len,
const uint8_t * quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float * c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
size_t moe_m2_gemm_kernel_i8i4(size_t blk_len,
const uint8_t ** quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float ** c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
size_t gemm_kernel_i8i8(size_t blk_len,
const uint8_t * quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float * c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
size_t gemm_kernel_i8mxfp4(size_t blk_len,
const uint8_t * quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float * c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
size_t moe_m2_gemm_kernel_i8mxfp4(size_t blk_len,
const uint8_t ** quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float ** c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
size_t gemm_kernel_i8i5(size_t blk_len,
const uint8_t * quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float * c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
size_t moe_m2_gemm_kernel_i8i5(size_t blk_len,
const uint8_t ** quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float ** c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
} // namespace ime2
} // namespace spacemit_kernels
ggml/src/ggml-cpu/spacemit/repack.cpp
+1795
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/spacemit/repack.h
+14
View File
@@ -0,0 +1,14 @@
#pragma once
#include "ggml-common.h"
#include "ggml.h"
#include <cstddef>
#include <cstdint>
namespace ggml::cpu::riscv64_spacemit {
template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS>
int repack(ggml_tensor * t, const void * data, size_t data_size);
} // namespace ggml::cpu::riscv64_spacemit
ggml/src/ggml-cpu/spacemit/rvv_kernels.cpp
+3178
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-cpu/spacemit/rvv_kernels.h
+95
View File
@@ -0,0 +1,95 @@
#pragma once
#include "ggml-cpu-impl.h"
#include <cassert>
#include <cstddef>
#include <cstdint>
#include <functional>
namespace spacemit_kernels {
constexpr auto div_round_up(auto up, auto down) {
return (up + down - 1) / down;
}
// Q8 Blk [f32] [s16] [int8 * blk_len]
// Q8 Blk N [f32 * N] [s16 * N] [int8 * blk_len * N]
constexpr size_t q8_blk_size(size_t blk_len, bool with_blk_sum = false) {
const size_t blk_size = sizeof(float) + blk_len * sizeof(int8_t) + (with_blk_sum ? sizeof(int16_t) : 0);
return blk_size;
}
// Q8 HP row block: K is split into K32 subblocks.
// Each subblock stores [f32 scale] [int8 * 32], with an optional fp16 sum trailer per subblock.
constexpr size_t q8_hp_blk_size(size_t blk_len, bool with_blk_sum = false, bool with_blk_scale = false) {
const size_t subblk_count = div_round_up(blk_len, size_t(32));
const size_t blk_size = blk_len * sizeof(int8_t) + subblk_count * sizeof(_Float16) +
(with_blk_sum ? subblk_count * sizeof(_Float16) : 0) +
(with_blk_scale ? sizeof(_Float16) : 0);
return blk_size;
}
// Q8K Blk [f32] [s16 * (blk_len / 16)] [int8 * blk_len]
// Q8K Blk N [f32 * N] [s16 * (blk_len / 16) * N] [int8 * blk_len * N]
constexpr size_t q8k_blk_size(size_t blk_len) {
const size_t blk_size = sizeof(float) + blk_len * sizeof(int8_t) + sizeof(int16_t) * blk_len / 16;
return blk_size;
}
using quantize_a_row_def = std::function<void(size_t, const float *, size_t, uint8_t *)>;
namespace rvv {
void memcpy1d(void * dst, const void * src, int64_t size);
void memcpy2d(void * dst, int64_t dst_stride, const void * src, int64_t src_stride, int64_t tile_rows, int64_t size);
void forward_flash_attn_ext_f16_one_chunk_vlen1024_vf16(const ggml_compute_params * params,
ggml_tensor * dst,
int ir0,
int ir1,
void * tcm_buffer,
size_t tcm_buffer_size);
void forward_flash_attn_ext_f16_tiled_vlen1024_vf16(const ggml_compute_params * params,
ggml_tensor * dst,
int ir0,
int ir1,
void * tcm_buffer,
size_t tcm_buffer_size);
void forward_rms_norm_f32(ggml_compute_params * params, ggml_tensor * op);
void forward_norm_f32(ggml_compute_params * params, ggml_tensor * op);
void forward_cont_with_permute(ggml_compute_params * params, ggml_tensor * op);
void forward_cpy_with_permute(ggml_compute_params * params, ggml_tensor * op);
template <typename T> void forward_get_rows(ggml_compute_params * params, ggml_tensor * op);
template <typename T> void forward_concat(ggml_compute_params * params, ggml_tensor * op);
template <ggml_op op_type, typename T> void forward_binary(ggml_compute_params * params, ggml_tensor * op);
template <typename T> void forward_sum_rows(const ggml_compute_params * params, ggml_tensor * op);
template <typename T> void forward_repeat_nrows(ggml_compute_params * params, ggml_tensor * op);
template <typename T> void forward_repeat_dim1(ggml_compute_params * params, ggml_tensor * op);
void quantize_a_row_i8(size_t blk_len, const float * a_ptr, size_t count_k, uint8_t * quant_a_ptr);
void quantize_a_4row_i8(size_t blk_len, const float * a_ptr, size_t count_k, uint8_t * quant_a_ptr);
void quantize_a_row_i8_hp(size_t blk_len, const float * a_ptr, size_t count_k, uint8_t * quant_a_ptr);
void quantize_a_4row_i8_hp(size_t blk_len, const float * a_ptr, size_t count_k, uint8_t * quant_a_ptr);
void quantize_a_row_i8k(size_t blk_len, const float * a_ptr, size_t count_k, uint8_t * quant_a_ptr);
void quantize_a_4row_i8k(size_t blk_len, const float * a_ptr, size_t count_k, uint8_t * quant_a_ptr);
} // namespace rvv
} // namespace spacemit_kernels
ggml/src/ggml-cpu/spacemit/spine_barrier.h
+34
View File
@@ -0,0 +1,34 @@
#pragma once
#include <atomic>
#include <cstdint>
#define SPINE_CACHE_LINE 64
#define SPINE_CACHE_ALIGN __attribute__((aligned(SPINE_CACHE_LINE)))
struct spine_barrier_t {
SPINE_CACHE_ALIGN std::atomic<int64_t> pending_;
SPINE_CACHE_ALIGN std::atomic<int64_t> rounds_;
SPINE_CACHE_ALIGN int64_t total_;
};
inline void spine_barrier_wait(spine_barrier_t * b) {
auto cur_round = b->rounds_.load(std::memory_order_acquire);
auto cnt = --b->pending_;
if (cnt == 0) {
b->pending_.store(b->total_);
b->rounds_.store(cur_round + 1);
} else {
while (cur_round == b->rounds_.load(std::memory_order_relaxed)) {
__asm__ volatile("pause " ::: "memory");
}
}
}
inline void spine_barrier_init(spine_barrier_t * b, int num_barriers, uint64_t thread_count) {
for (int i = 0; i < num_barriers; i++) {
b[i].total_ = thread_count;
b[i].pending_.store(thread_count);
b[i].rounds_.store(0);
}
}
ggml/src/ggml-cpu/spacemit/spine_mem_pool.cpp
+760
View File
@@ -0,0 +1,760 @@
#include "spine_mem_pool.h"
#include "common.h"
#include "ime_env.h"
#include "spine_tcm.h"
#include <fcntl.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include <unistd.h>
#include <algorithm>
#include <cerrno>
#include <cstdint>
#include <cstdlib>
#include <limits>
#include <memory>
#include <mutex>
#include <unordered_map>
#include <vector>
namespace ggml::cpu::riscv64_spacemit {
namespace {
constexpr size_t SPINE_MEM_POOL_CHUNK_SIZE = 512ull * 1024ull * 1024ull;
constexpr size_t SPINE_SHARE_MEM_POOL_CHUNK_SIZE = 512ull * 1024ull;
constexpr size_t SPINE_MEM_POOL_1G_REGION_SIZE = 1ull << 30;
constexpr uint64_t HUGETLB_1G_FLAG_REQUIRE_PUD = 1ull << 0;
constexpr char SPINE_MEM_POOL_HUGETLB_1G_DEV[] = "/dev/hugetlb_1g";
constexpr char SPINE_MEM_POOL_TCM_SYNC_MEM_DEV[] = "/dev/tcm_sync_mem";
struct hugetlb_1g_region {
uint64_t size{ 0 };
uint64_t dma_addr{ 0 };
uint64_t flags{ 0 };
uint64_t reserved{ 0 };
};
#define HUGETLB_1G_IOC_MAGIC 'M'
#define HUGETLB_1G_IOC_ALLOC _IOWR(HUGETLB_1G_IOC_MAGIC, 0x00, struct hugetlb_1g_region)
#define HUGETLB_1G_IOC_FREE _IO(HUGETLB_1G_IOC_MAGIC, 0x01)
struct free_block {
size_t offset{ 0 };
size_t size{ 0 };
};
struct pool_chunk {
uint8_t * base{ nullptr };
size_t size{ 0 };
int fd{ -1 };
std::vector<free_block> free_blocks;
};
struct pool_allocation {
void * chunk_base{ nullptr };
size_t chunk_size{ 0 };
void * base{ nullptr };
size_t size{ 0 };
};
bool is_power_of_two(size_t value) {
return value != 0 && (value & (value - 1)) == 0;
}
bool align_up(size_t value, size_t alignment, size_t * aligned_value) {
if (aligned_value == nullptr || alignment == 0) {
return false;
}
const size_t remainder = value % alignment;
if (remainder == 0) {
*aligned_value = value;
return true;
}
const size_t padding = alignment - remainder;
if (value > std::numeric_limits<size_t>::max() - padding) {
return false;
}
*aligned_value = value + padding;
return true;
}
bool align_up_uintptr(uintptr_t value, size_t alignment, uintptr_t * aligned_value) {
if (aligned_value == nullptr || alignment == 0) {
return false;
}
const uintptr_t remainder = value % alignment;
if (remainder == 0) {
*aligned_value = value;
return true;
}
const uintptr_t padding = alignment - remainder;
if (value > std::numeric_limits<uintptr_t>::max() - padding) {
return false;
}
*aligned_value = value + padding;
return true;
}
class spine_mem_pool_manager {
public:
explicit spine_mem_pool_manager(size_t default_chunk_size) : default_chunk_size_(default_chunk_size) {}
virtual ~spine_mem_pool_manager() = default;
void * alloc(size_t size, size_t alignment) {
if (size == 0 || !is_power_of_two(alignment)) {
return nullptr;
}
size_t aligned_size = 0;
if (!align_up(size, alignment, &aligned_size)) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: align_up failed for size %zu alignment %zu\n", __func__, size,
alignment);
return nullptr;
}
pool_allocation allocation;
std::lock_guard<std::mutex> lock(mutex_);
if (!try_alloc_locked(aligned_size, alignment, &allocation)) {
if (!add_chunk_locked(aligned_size, alignment)) {
return nullptr;
}
if (!try_alloc_locked(aligned_size, alignment, &allocation)) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: allocation retry failed for size %zu alignment %zu\n",
__func__, aligned_size, alignment);
return nullptr;
}
}
try {
const auto [allocation_it, inserted] = allocations_.emplace(allocation.base, allocation);
if (!inserted) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: duplicate allocation key %p\n", __func__, allocation.base);
rollback_allocation_locked(allocation);
return nullptr;
}
} catch (const std::bad_alloc &) {
rollback_allocation_locked(allocation);
throw;
}
return allocation.base;
}
void free(void * base) {
if (base == nullptr) {
return;
}
std::lock_guard<std::mutex> lock(mutex_);
auto allocation_it = allocations_.find(base);
if (allocation_it == allocations_.end()) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: unknown allocation %p\n", __func__, base);
return;
}
pool_allocation allocation = allocation_it->second;
allocations_.erase(allocation_it);
auto chunk_it = find_chunk_locked(allocation);
if (chunk_it == chunks_.end()) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: unknown chunk for allocation %p size %zu\n", __func__,
allocation.base, allocation.size);
return;
}
auto * chunk_base = chunk_it->base;
auto * alloc_base = static_cast<uint8_t *>(allocation.base);
if (alloc_base < chunk_base || alloc_base >= chunk_base + chunk_it->size) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: allocation %p out of chunk range %p..%p\n", __func__,
allocation.base, chunk_base, chunk_base + chunk_it->size);
return;
}
const size_t offset = static_cast<size_t>(alloc_base - chunk_base);
if (offset > chunk_it->size || allocation.size > chunk_it->size - offset) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: allocation %p size %zu exceeds chunk size %zu\n", __func__,
allocation.base, allocation.size, chunk_it->size);
return;
}
insert_free_block_locked(*chunk_it, { offset, allocation.size });
maybe_release_empty_chunk_locked(chunk_it);
}
protected:
void release_chunks() {
std::lock_guard<std::mutex> lock(mutex_);
allocations_.clear();
for (auto & chunk : chunks_) {
dealloc_chunk(&chunk);
}
chunks_.clear();
}
size_t default_chunk_size() const { return default_chunk_size_; }
static void clear_chunk(pool_chunk * chunk) {
chunk->base = nullptr;
chunk->size = 0;
chunk->fd = -1;
chunk->free_blocks.clear();
}
virtual bool alloc_chunk(size_t min_size, size_t alignment, void * hint_addr, pool_chunk * chunk) = 0;
virtual void dealloc_chunk(pool_chunk * chunk) = 0;
private:
struct alloc_candidate {
size_t chunk_index{ 0 };
size_t block_index{ 0 };
size_t aligned_offset{ 0 };
uintptr_t address{ std::numeric_limits<uintptr_t>::max() };
bool valid{ false };
};
std::vector<pool_chunk>::iterator find_chunk_locked(const pool_allocation & allocation) {
return std::find_if(chunks_.begin(), chunks_.end(), [&](const pool_chunk & chunk) {
return chunk.base == allocation.chunk_base && chunk.size == allocation.chunk_size;
});
}
bool add_chunk_locked(size_t min_size, size_t alignment) {
pool_chunk chunk;
const size_t chunk_request = default_chunk_size_ == 0 ? min_size : std::max(min_size, default_chunk_size_);
void * hint_addr = nullptr;
for (const auto & existing_chunk : chunks_) {
auto * chunk_end = existing_chunk.base + existing_chunk.size;
if (hint_addr == nullptr || chunk_end > hint_addr) {
hint_addr = chunk_end;
}
}
if (!alloc_chunk(chunk_request, alignment, hint_addr, &chunk)) {
return false;
}
if (chunk.base == nullptr || chunk.size < min_size) {
GGML_LOG_ERROR(
"CPU_RISCV64_SPACEMIT: %s: invalid chunk returned for request size %zu, chunk_base=%p chunk_size=%zu\n",
__func__, min_size, chunk.base, chunk.size);
dealloc_chunk(&chunk);
return false;
}
try {
chunk.free_blocks.push_back({ 0, chunk.size });
chunks_.push_back(std::move(chunk));
} catch (const std::bad_alloc &) {
dealloc_chunk(&chunk);
throw;
}
return true;
}
void rollback_allocation_locked(const pool_allocation & allocation) {
auto chunk_it = find_chunk_locked(allocation);
if (chunk_it == chunks_.end()) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: failed to rollback allocation %p, owning chunk not found\n",
__func__, allocation.base);
return;
}
auto * chunk_base = chunk_it->base;
auto * alloc_base = static_cast<uint8_t *>(allocation.base);
if (alloc_base < chunk_base || alloc_base >= chunk_base + chunk_it->size) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: failed to rollback allocation %p, chunk range is invalid\n",
__func__, allocation.base);
return;
}
const size_t offset = static_cast<size_t>(alloc_base - chunk_base);
if (offset > chunk_it->size || allocation.size > chunk_it->size - offset) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: failed to rollback allocation %p size %zu\n", __func__,
allocation.base, allocation.size);
return;
}
insert_free_block_locked(*chunk_it, { offset, allocation.size });
maybe_release_empty_chunk_locked(chunk_it);
}
bool try_alloc_locked(size_t size, size_t alignment, pool_allocation * allocation) {
alloc_candidate best;
for (size_t chunk_index = 0; chunk_index < chunks_.size(); ++chunk_index) {
const auto & chunk = chunks_[chunk_index];
for (size_t block_index = 0; block_index < chunk.free_blocks.size(); ++block_index) {
const auto & block = chunk.free_blocks[block_index];
uintptr_t aligned_addr = 0;
const auto block_addr = reinterpret_cast<uintptr_t>(chunk.base + block.offset);
if (!align_up_uintptr(block_addr, alignment, &aligned_addr)) {
continue;
}
if (aligned_addr < block_addr) {
continue;
}
const size_t aligned_offset = block.offset + static_cast<size_t>(aligned_addr - block_addr);
const size_t padding = aligned_offset - block.offset;
if (padding > block.size || size > block.size - padding) {
continue;
}
if (!best.valid || aligned_addr < best.address) {
best.chunk_index = chunk_index;
best.block_index = block_index;
best.aligned_offset = aligned_offset;
best.address = aligned_addr;
best.valid = true;
}
}
}
if (!best.valid) {
return false;
}
auto & chunk = chunks_[best.chunk_index];
const free_block block = chunk.free_blocks[best.block_index];
const size_t padding = best.aligned_offset - block.offset;
const size_t alloc_end = best.aligned_offset + size;
const size_t block_end = block.offset + block.size;
chunk.free_blocks.erase(chunk.free_blocks.begin() + best.block_index);
auto insert_it = chunk.free_blocks.begin() + best.block_index;
if (padding != 0) {
insert_it = chunk.free_blocks.insert(insert_it, { block.offset, padding });
++insert_it;
}
if (alloc_end < block_end) {
chunk.free_blocks.insert(insert_it, { alloc_end, block_end - alloc_end });
}
allocation->chunk_base = chunk.base;
allocation->chunk_size = chunk.size;
allocation->base = chunk.base + best.aligned_offset;
allocation->size = size;
return true;
}
void maybe_release_empty_chunk_locked(std::vector<pool_chunk>::iterator chunk_it) {
if (chunk_it->free_blocks.size() != 1) {
return;
}
const auto & block = chunk_it->free_blocks.front();
if (block.offset != 0 || block.size != chunk_it->size) {
return;
}
dealloc_chunk(&*chunk_it);
chunks_.erase(chunk_it);
}
void insert_free_block_locked(pool_chunk & chunk, free_block block) {
auto it = chunk.free_blocks.begin();
while (it != chunk.free_blocks.end() && it->offset < block.offset) {
++it;
}
if (it != chunk.free_blocks.begin()) {
const auto & prev = *(it - 1);
if (prev.offset + prev.size > block.offset) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: overlapping free block at offset %zu size %zu\n", __func__,
block.offset, block.size);
return;
}
}
if (it != chunk.free_blocks.end() && block.offset + block.size > it->offset) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: overlapping next free block at offset %zu size %zu\n", __func__,
block.offset, block.size);
return;
}
it = chunk.free_blocks.insert(it, block);
if (it != chunk.free_blocks.begin()) {
auto prev = it - 1;
if (prev->offset + prev->size == it->offset) {
it->offset = prev->offset;
it->size += prev->size;
it = chunk.free_blocks.erase(prev);
}
}
if (it + 1 != chunk.free_blocks.end() && it->offset + it->size == (it + 1)->offset) {
it->size += (it + 1)->size;
chunk.free_blocks.erase(it + 1);
}
}
std::mutex mutex_;
std::vector<pool_chunk> chunks_;
std::unordered_map<void *, pool_allocation> allocations_;
size_t default_chunk_size_{ 0 };
};
class spine_mem_pool_posix final : public spine_mem_pool_manager {
public:
spine_mem_pool_posix() : spine_mem_pool_manager(0) {}
~spine_mem_pool_posix() override { release_chunks(); }
private:
bool alloc_chunk(size_t min_size, size_t alignment, void * hint_addr, pool_chunk * chunk) override {
(void) hint_addr;
const size_t alloc_alignment = std::max(alignment, sizeof(void *));
void * base = nullptr;
const int rc = posix_memalign(&base, alloc_alignment, min_size);
if (rc != 0) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: posix_memalign failed for size %zu alignment %zu, rc=%d\n",
__func__, min_size, alloc_alignment, rc);
return false;
}
chunk->base = static_cast<uint8_t *>(base);
chunk->size = min_size;
chunk->fd = -1;
return true;
}
void dealloc_chunk(pool_chunk * chunk) override {
std::free(chunk->base);
clear_chunk(chunk);
}
};
class spine_mem_pool_transparent_hugepage final : public spine_mem_pool_manager {
public:
spine_mem_pool_transparent_hugepage() : spine_mem_pool_manager(SPINE_MEM_POOL_CHUNK_SIZE) {}
~spine_mem_pool_transparent_hugepage() override { release_chunks(); }
private:
bool alloc_chunk(size_t min_size, size_t alignment, void * hint_addr, pool_chunk * chunk) override {
(void) alignment;
size_t chunk_size = 0;
if (!align_up(min_size, default_chunk_size(), &chunk_size)) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: failed to round chunk size for %zu\n", __func__, min_size);
return false;
}
void * map_addr = mmap(hint_addr, chunk_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
if (map_addr == MAP_FAILED) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: mmap failed for chunk size %zu, errno=%d\n", __func__, chunk_size,
errno);
return false;
}
if (madvise(map_addr, chunk_size, MADV_HUGEPAGE) != 0) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: madvise(MADV_HUGEPAGE) failed for chunk size %zu, errno=%d\n",
__func__, chunk_size, errno);
munmap(map_addr, chunk_size);
return false;
}
chunk->base = static_cast<uint8_t *>(map_addr);
chunk->size = chunk_size;
chunk->fd = -1;
return true;
}
void dealloc_chunk(pool_chunk * chunk) override {
if (chunk->base != nullptr && chunk->size != 0 && munmap(chunk->base, chunk->size) != 0) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: munmap failed for chunk %p size %zu, errno=%d\n", __func__,
chunk->base, chunk->size, errno);
}
clear_chunk(chunk);
}
};
class spine_mem_pool_hugetlb_1g final : public spine_mem_pool_manager {
public:
spine_mem_pool_hugetlb_1g() : spine_mem_pool_manager(SPINE_MEM_POOL_1G_REGION_SIZE) {}
~spine_mem_pool_hugetlb_1g() override { release_chunks(); }
private:
bool alloc_chunk(size_t min_size, size_t alignment, void * hint_addr, pool_chunk * chunk) override {
(void) alignment;
(void) hint_addr;
size_t region_size = 0;
if (!align_up(min_size, SPINE_MEM_POOL_1G_REGION_SIZE, &region_size)) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: failed to round hugetlb_1g size for %zu\n", __func__, min_size);
return false;
}
const int fd = open(SPINE_MEM_POOL_HUGETLB_1G_DEV, O_RDWR);
if (fd < 0) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: open(%s) failed, errno=%d\n", __func__,
SPINE_MEM_POOL_HUGETLB_1G_DEV, errno);
return false;
}
hugetlb_1g_region region;
region.size = region_size;
region.flags = HUGETLB_1G_FLAG_REQUIRE_PUD;
if (ioctl(fd, HUGETLB_1G_IOC_ALLOC, &region) < 0) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: HUGETLB_1G_IOC_ALLOC failed for size %zu, errno=%d\n", __func__,
region_size, errno);
close(fd);
return false;
}
void * map_addr = mmap(nullptr, region.size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
if (map_addr == MAP_FAILED) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: mmap failed for hugetlb_1g size %llu, errno=%d\n", __func__,
static_cast<unsigned long long>(region.size), errno);
ioctl(fd, HUGETLB_1G_IOC_FREE);
close(fd);
return false;
}
chunk->base = static_cast<uint8_t *>(map_addr);
chunk->size = region.size;
chunk->fd = fd;
return true;
}
void dealloc_chunk(pool_chunk * chunk) override {
if (chunk->base != nullptr && chunk->size != 0 && munmap(chunk->base, chunk->size) != 0) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: munmap failed for hugetlb_1g chunk %p size %zu, errno=%d\n",
__func__, chunk->base, chunk->size, errno);
}
if (chunk->fd >= 0) {
if (ioctl(chunk->fd, HUGETLB_1G_IOC_FREE) < 0) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: HUGETLB_1G_IOC_FREE failed for chunk %p, errno=%d\n",
__func__, chunk->base, errno);
}
close(chunk->fd);
}
clear_chunk(chunk);
}
};
class spine_mem_pool_shared_mem final : public spine_mem_pool_manager {
public:
spine_mem_pool_shared_mem() : spine_mem_pool_manager(SPINE_SHARE_MEM_POOL_CHUNK_SIZE) {}
~spine_mem_pool_shared_mem() override { release_chunks(); }
private:
bool alloc_chunk(size_t min_size, size_t alignment, void * hint_addr, pool_chunk * chunk) override {
(void) alignment;
if (hint_addr != nullptr) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: shared_mem does not support multiple active chunks\n", __func__);
return false;
}
if (min_size > default_chunk_size()) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: shared_mem request %zu exceeds chunk size %zu\n", __func__,
min_size, default_chunk_size());
return false;
}
const int fd = open(SPINE_MEM_POOL_TCM_SYNC_MEM_DEV, O_RDWR | O_SYNC);
if (fd < 0) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: open(%s) failed, errno=%d\n", __func__,
SPINE_MEM_POOL_TCM_SYNC_MEM_DEV, errno);
return false;
}
void * map_addr = mmap(nullptr, default_chunk_size(), PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
if (map_addr == MAP_FAILED) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: mmap failed for %s size %zu, errno=%d\n", __func__,
SPINE_MEM_POOL_TCM_SYNC_MEM_DEV, default_chunk_size(), errno);
close(fd);
return false;
}
chunk->base = static_cast<uint8_t *>(map_addr);
chunk->size = default_chunk_size();
chunk->fd = fd;
return true;
}
void dealloc_chunk(pool_chunk * chunk) override {
if (chunk->base != nullptr && chunk->size != 0 && munmap(chunk->base, chunk->size) != 0) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: munmap failed for shared_mem chunk %p size %zu, errno=%d\n",
__func__, chunk->base, chunk->size, errno);
}
if (chunk->fd >= 0) {
close(chunk->fd);
}
clear_chunk(chunk);
}
};
spine_mem_pool_manager & get_spine_mem_pool_manager() {
static std::once_flag pool_once;
static std::unique_ptr<spine_mem_pool_manager> selected_pool;
static spine_mem_pool_backend selected_backend = spine_mem_pool_backend::none;
spine_mem_pool_backend backend = global_spine_env_info.mem_backend;
if (backend == spine_mem_pool_backend::none) {
backend = spine_mem_pool_backend::transparent_hugepage;
}
std::call_once(pool_once, [&]() {
selected_backend = backend;
switch (selected_backend) {
case spine_mem_pool_backend::posix_memalign:
selected_pool = std::make_unique<spine_mem_pool_posix>();
break;
case spine_mem_pool_backend::transparent_hugepage:
selected_pool = std::make_unique<spine_mem_pool_transparent_hugepage>();
break;
case spine_mem_pool_backend::hugetlb_1g:
selected_pool = std::make_unique<spine_mem_pool_hugetlb_1g>();
break;
case spine_mem_pool_backend::none:
selected_backend = spine_mem_pool_backend::transparent_hugepage;
selected_pool = std::make_unique<spine_mem_pool_transparent_hugepage>();
break;
}
});
if (backend != selected_backend) {
GGML_LOG_ERROR(
"CPU_RISCV64_SPACEMIT: %s: mem pool backend is process-global and mutually exclusive, requested=%d but "
"selected=%d\n",
__func__, static_cast<int>(backend), static_cast<int>(selected_backend));
}
if (selected_pool) {
return *selected_pool;
}
throw std::bad_alloc();
}
spine_mem_pool_manager & get_spine_mem_pool_shared_mem_manager() {
static std::once_flag shared_mem_pool_once;
static std::unique_ptr<spine_mem_pool_shared_mem> shared_mem_pool;
std::call_once(shared_mem_pool_once, [&]() { shared_mem_pool = std::make_unique<spine_mem_pool_shared_mem>(); });
if (shared_mem_pool) {
return *shared_mem_pool;
}
throw std::bad_alloc();
}
} // namespace
bool spine_mem_pool_tcm_init(spine_mem_pool_tcm_info * info) noexcept {
if (info == nullptr) {
return false;
}
*info = {};
if (spine_tcm_open_handle(NULL) != 0 || !spine_tcm_is_available()) {
return false;
}
spine_tcm_mem_info_t mem_info;
if (spine_tcm_mem_info(&mem_info) != 0) {
return false;
}
info->available = true;
info->blk_size = mem_info.blk_size;
info->blk_num = mem_info.blk_num;
info->is_fake_tcm = mem_info.is_fake_tcm != 0;
return true;
}
void * spine_mem_pool_tcm_mem_get(int cpu_id) noexcept {
return spine_tcm_mem_get(cpu_id);
}
void * spine_mem_pool_tcm_mem_wait(int cpu_id) noexcept {
return spine_tcm_mem_try_wait(cpu_id, 1000 * 1000);
}
int spine_mem_pool_tcm_mem_release(int cpu_id) noexcept {
return spine_tcm_mem_release(cpu_id);
}
void * spine_mem_pool_alloc(size_t size, size_t alignment) noexcept {
try {
return get_spine_mem_pool_manager().alloc(size, alignment);
} catch (const std::bad_alloc &) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: bad_alloc while allocating size %zu\n", __func__, size);
return nullptr;
}
}
void * spine_mem_pool_shared_mem_alloc(size_t size, size_t alignment) noexcept {
try {
return get_spine_mem_pool_shared_mem_manager().alloc(size, alignment);
} catch (const std::bad_alloc &) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: bad_alloc while allocating shared memory size %zu\n", __func__, size);
return nullptr;
}
}
void spine_mem_pool_free(void * base) noexcept {
try {
get_spine_mem_pool_manager().free(base);
} catch (const std::bad_alloc &) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: bad_alloc while freeing allocation %p\n", __func__, base);
}
}
void spine_mem_pool_shared_mem_free(void * base) noexcept {
try {
get_spine_mem_pool_shared_mem_manager().free(base);
} catch (const std::bad_alloc &) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: bad_alloc while freeing shared allocation %p\n", __func__, base);
}
}
} // namespace ggml::cpu::riscv64_spacemit
extern "C" {
void * ggml_backend_cpu_riscv64_spacemit_alloc_shared(size_t size, size_t alignment) {
void * result = ggml::cpu::riscv64_spacemit::spine_mem_pool_shared_mem_alloc(size, alignment);
if (result == nullptr) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: failed to allocate shared memory size %zu alignment %zu\n", __func__,
size, alignment);
}
return result;
}
void ggml_backend_cpu_riscv64_spacemit_free_shared(void * ptr) {
ggml::cpu::riscv64_spacemit::spine_mem_pool_shared_mem_free(ptr);
}
}
ggml/src/ggml-cpu/spacemit/spine_mem_pool.h
+32
View File
@@ -0,0 +1,32 @@
#pragma once
#include <cstddef>
#include <cstdint>
namespace ggml::cpu::riscv64_spacemit {
enum class spine_mem_pool_backend : uint8_t {
none,
posix_memalign,
transparent_hugepage,
hugetlb_1g,
};
struct spine_mem_pool_tcm_info {
bool available{ false };
size_t blk_size{ 0 };
size_t blk_num{ 0 };
bool is_fake_tcm{ false };
};
bool spine_mem_pool_tcm_init(spine_mem_pool_tcm_info * info) noexcept;
void * spine_mem_pool_tcm_mem_get(int cpu_id) noexcept;
void * spine_mem_pool_tcm_mem_wait(int cpu_id) noexcept;
int spine_mem_pool_tcm_mem_release(int cpu_id) noexcept;
void * spine_mem_pool_alloc(size_t size, size_t alignment) noexcept;
void * spine_mem_pool_shared_mem_alloc(size_t size, size_t alignment) noexcept;
void spine_mem_pool_free(void * base) noexcept;
void spine_mem_pool_shared_mem_free(void * base) noexcept;
} // namespace ggml::cpu::riscv64_spacemit
ggml/src/ggml-cpu/spacemit/spine_tcm.h
+409
View File
@@ -0,0 +1,409 @@
#ifndef SPINE_TCM_PUBLIC_H_
#define SPINE_TCM_PUBLIC_H_
/*
* spine_tcm public API
*
* Usage:
* 1. Direct link mode
* Define SPINE_TCM_DIRECT_LINK and link against libspine_tcm.so.
*
* if (spine_tcm_is_available()) {
* void *buffer = spine_tcm_mem_get(0);
* spine_tcm_mem_free(0);
* }
*
* 2. Header-only loader mode
* Include this header without linking libspine_tcm.so. The loader first
* tries to reuse a process-global spine_tcm instance and falls back to
* dlopen("libspine_tcm.so") when needed.
*
* spine_tcm_open_handle(NULL); // optional pre-bind
* if (spine_tcm_is_available()) {
* void *buffer = spine_tcm_mem_get(0);
* spine_tcm_mem_free(0);
* }
*/
#include <stddef.h>
#include <stdint.h>
#include <string.h>
#if !defined(SPINE_TCM_BUILD_SHARED) && !defined(SPINE_TCM_DIRECT_LINK)
# include <dlfcn.h>
#endif
#ifdef __cplusplus
extern "C" {
#endif
#if defined(_WIN32)
# if defined(SPINE_TCM_BUILD_SHARED)
# define SPINE_TCM_API __declspec(dllexport)
# else
# define SPINE_TCM_API __declspec(dllimport)
# endif
#else
# define SPINE_TCM_API __attribute__((visibility("default")))
#endif
typedef struct spine_tcm_mem_info {
size_t blk_size;
size_t blk_num;
int is_fake_tcm;
} spine_tcm_mem_info_t;
typedef struct spine_tcm_block_info {
int id;
void * va;
size_t size;
uint64_t phys_addr;
uint64_t cpu_affinity_mask;
int owner_tid;
int is_acquired;
} spine_tcm_block_info_t;
/* Shared-library runtime ABI exported by libspine_tcm.so. */
SPINE_TCM_API const char * spine_tcm_runtime_version(void);
SPINE_TCM_API int spine_tcm_runtime_is_available(void);
SPINE_TCM_API int spine_tcm_runtime_layout_info(spine_tcm_mem_info_t * info);
SPINE_TCM_API int spine_tcm_runtime_mem_info(int id, spine_tcm_block_info_t * info);
SPINE_TCM_API void * spine_tcm_runtime_mem_get(int id);
SPINE_TCM_API int spine_tcm_runtime_mem_free(int id);
SPINE_TCM_API void * spine_tcm_runtime_mem_try_wait(int id, size_t timeout_us);
SPINE_TCM_API int spine_tcm_runtime_mem_release(int id);
SPINE_TCM_API int spine_tcm_runtime_mem_force_release(int id);
SPINE_TCM_API int spine_tcm_runtime_mem_query(int id);
#if defined(SPINE_TCM_DIRECT_LINK)
/* Optional no-op in direct-link mode. */
static inline int spine_tcm_open_handle(const char * so_path) {
(void) so_path;
return 0;
}
static inline const char * spine_tcm_version(void) {
return spine_tcm_runtime_version();
}
/* Returns 1 when the runtime driver is available, otherwise 0. */
static inline int spine_tcm_is_available(void) {
return spine_tcm_runtime_is_available();
}
/* Returns runtime memory geometry and whether the current backend is fake TCM. */
static inline int spine_tcm_mem_info(spine_tcm_mem_info_t * info) {
return spine_tcm_runtime_layout_info(info);
}
/* Returns per-block runtime metadata for the given TCM id. */
static inline int spine_tcm_block_info(int id, spine_tcm_block_info_t * info) {
return spine_tcm_runtime_mem_info(id, info);
}
/* Returns a cached buffer for the given TCM id, or NULL on failure. */
static inline void * spine_tcm_mem_get(int id) {
return spine_tcm_runtime_mem_get(id);
}
/* Releases one reference acquired by spine_tcm_mem_get(id). */
static inline int spine_tcm_mem_free(int id) {
return spine_tcm_runtime_mem_free(id);
}
/* Waits for a TCM block handoff and returns the driver-owned buffer when available. */
static inline void * spine_tcm_mem_try_wait(int id, size_t over_time) {
return spine_tcm_runtime_mem_try_wait(id, over_time);
}
/* Releases a buffer acquired by spine_tcm_mem_try_wait(id, over_time). */
static inline int spine_tcm_mem_release(int id) {
return spine_tcm_runtime_mem_release(id);
}
/* Forces a release for the given TCM id when the backend supports it. */
static inline int spine_tcm_mem_force_release(int id) {
return spine_tcm_runtime_mem_force_release(id);
}
/* Returns whether the given TCM id is currently acquired. */
static inline int spine_tcm_mem_query(int id) {
return spine_tcm_runtime_mem_query(id);
}
#elif !defined(SPINE_TCM_BUILD_SHARED)
typedef struct spine_tcm_handle {
void * module_handle;
int use_global_scope;
int owns_module_handle;
const char * (*runtime_version)(void);
int (*runtime_is_available)(void);
int (*runtime_layout_info)(spine_tcm_mem_info_t * info);
int (*runtime_mem_info)(int id, spine_tcm_block_info_t * info);
void * (*runtime_mem_get)(int id);
int (*runtime_mem_free)(int id);
void * (*runtime_mem_try_wait)(int id, size_t over_time);
int (*runtime_mem_release)(int id);
int (*runtime_mem_force_release)(int id);
int (*runtime_mem_query)(int id);
} spine_tcm_handle_t;
static inline spine_tcm_handle_t * spine_tcm_default_handle(void) {
static spine_tcm_handle_t handle = { 0 };
return &handle;
}
static inline void spine_tcm_handle_reset(spine_tcm_handle_t * handle) {
if (handle != NULL) {
memset(handle, 0, sizeof(*handle));
}
}
static inline int spine_tcm_handle_bind(spine_tcm_handle_t * handle) {
void * symbol_scope = handle->use_global_scope ? RTLD_DEFAULT : handle->module_handle;
handle->runtime_version = (const char * (*) (void) ) dlsym(symbol_scope, "spine_tcm_runtime_version");
handle->runtime_is_available = (int (*)(void)) dlsym(symbol_scope, "spine_tcm_runtime_is_available");
handle->runtime_layout_info =
(int (*)(spine_tcm_mem_info_t *)) dlsym(symbol_scope, "spine_tcm_runtime_layout_info");
handle->runtime_mem_info =
(int (*)(int, spine_tcm_block_info_t *)) dlsym(symbol_scope, "spine_tcm_runtime_mem_info");
handle->runtime_mem_get = (void * (*) (int) ) dlsym(symbol_scope, "spine_tcm_runtime_mem_get");
handle->runtime_mem_free = (int (*)(int)) dlsym(symbol_scope, "spine_tcm_runtime_mem_free");
handle->runtime_mem_try_wait = (void * (*) (int, size_t)) dlsym(symbol_scope, "spine_tcm_runtime_mem_try_wait");
handle->runtime_mem_release = (int (*)(int)) dlsym(symbol_scope, "spine_tcm_runtime_mem_release");
handle->runtime_mem_force_release = (int (*)(int)) dlsym(symbol_scope, "spine_tcm_runtime_mem_force_release");
handle->runtime_mem_query = (int (*)(int)) dlsym(symbol_scope, "spine_tcm_runtime_mem_query");
return handle->runtime_version != NULL && handle->runtime_is_available != NULL &&
handle->runtime_layout_info != NULL && handle->runtime_mem_info != NULL &&
handle->runtime_mem_get != NULL && handle->runtime_mem_free != NULL &&
handle->runtime_mem_try_wait != NULL && handle->runtime_mem_release != NULL &&
handle->runtime_mem_force_release != NULL && handle->runtime_mem_query != NULL ?
0 :
-1;
}
/*
* Try to bind against an already-loaded process-global spine_tcm instance.
* The shared library exports spine_tcm_runtime_marker only for this probe.
*/
static inline int spine_tcm_try_bind_global(spine_tcm_handle_t * handle) {
if (dlsym(RTLD_DEFAULT, "spine_tcm_runtime_marker") == NULL) {
return -1;
}
handle->use_global_scope = 1;
return spine_tcm_handle_bind(handle);
}
/*
* Optional pre-bind entry point.
*
* Behavior:
* - Reuses an already-loaded global spine_tcm instance when available.
* - Otherwise loads the shared library from so_path or the default soname.
* - Repeated calls are safe and return 0 after the first successful bind.
*/
static inline int spine_tcm_open_handle(const char * so_path) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
const char * library = (so_path != NULL && so_path[0] != '\0') ? so_path : "libspine_tcm.so";
if (resolved->module_handle != NULL || resolved->use_global_scope) {
return 0;
}
if (spine_tcm_try_bind_global(resolved) == 0) {
return 0;
}
spine_tcm_handle_reset(resolved);
resolved->module_handle = dlopen(library, RTLD_LAZY | RTLD_GLOBAL);
resolved->owns_module_handle = resolved->module_handle != NULL ? 1 : 0;
if (resolved->module_handle == NULL) {
spine_tcm_handle_reset(resolved);
return -1;
}
if (spine_tcm_handle_bind(resolved) != 0) {
if (resolved->owns_module_handle) {
dlclose(resolved->module_handle);
}
spine_tcm_handle_reset(resolved);
return -1;
}
return 0;
}
/* Returns 1 when the runtime driver is available, otherwise 0. */
static inline int spine_tcm_is_available(void) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
(void) spine_tcm_open_handle(NULL);
}
if ((resolved->module_handle == NULL && !resolved->use_global_scope) || resolved->runtime_is_available == NULL) {
return 0;
}
return resolved->runtime_is_available();
}
/* Returns runtime memory geometry and whether the current backend is fake TCM. */
static inline int spine_tcm_mem_info(spine_tcm_mem_info_t * info) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
(void) spine_tcm_open_handle(NULL);
}
if ((resolved->module_handle == NULL && !resolved->use_global_scope) || resolved->runtime_layout_info == NULL) {
return -1;
}
return resolved->runtime_layout_info(info);
}
static inline const char * spine_tcm_version(void) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
(void) spine_tcm_open_handle(NULL);
}
if ((resolved->module_handle == NULL && !resolved->use_global_scope) || resolved->runtime_version == NULL) {
return "unknown";
}
return resolved->runtime_version();
}
/* Returns per-block runtime metadata for the given TCM id. */
static inline int spine_tcm_block_info(int id, spine_tcm_block_info_t * info) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
(void) spine_tcm_open_handle(NULL);
}
if ((resolved->module_handle == NULL && !resolved->use_global_scope) || resolved->runtime_mem_info == NULL) {
return -1;
}
return resolved->runtime_mem_info(id, info);
}
/* Returns a cached buffer for the given TCM id, or NULL on failure. */
static inline void * spine_tcm_mem_get(int id) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
(void) spine_tcm_open_handle(NULL);
}
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
return NULL;
}
if (resolved->runtime_mem_get == NULL) {
return NULL;
}
return resolved->runtime_mem_get(id);
}
/* Releases one reference acquired by spine_tcm_mem_get(id). */
static inline int spine_tcm_mem_free(int id) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
(void) spine_tcm_open_handle(NULL);
}
if ((resolved->module_handle == NULL && !resolved->use_global_scope) || resolved->runtime_mem_free == NULL) {
return -1;
}
return resolved->runtime_mem_free(id);
}
/* Waits for a TCM block handoff and returns the driver-owned buffer when available. */
static inline void * spine_tcm_mem_try_wait(int id, size_t over_time) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
(void) spine_tcm_open_handle(NULL);
}
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
return NULL;
}
if (resolved->runtime_mem_try_wait == NULL) {
return NULL;
}
return resolved->runtime_mem_try_wait(id, over_time);
}
/* Releases a buffer acquired by spine_tcm_mem_try_wait(id, over_time). */
static inline int spine_tcm_mem_release(int id) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
(void) spine_tcm_open_handle(NULL);
}
if ((resolved->module_handle == NULL && !resolved->use_global_scope) || resolved->runtime_mem_release == NULL) {
return -1;
}
return resolved->runtime_mem_release(id);
}
/* Forces a release for the given TCM id when the backend supports it. */
static inline int spine_tcm_mem_force_release(int id) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
(void) spine_tcm_open_handle(NULL);
}
if ((resolved->module_handle == NULL && !resolved->use_global_scope) ||
resolved->runtime_mem_force_release == NULL) {
return -1;
}
return resolved->runtime_mem_force_release(id);
}
/* Returns whether the given TCM id is currently acquired. */
static inline int spine_tcm_mem_query(int id) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
(void) spine_tcm_open_handle(NULL);
}
if ((resolved->module_handle == NULL && !resolved->use_global_scope) || resolved->runtime_mem_query == NULL) {
return -1;
}
return resolved->runtime_mem_query(id);
}
#else
static inline const char * spine_tcm_version(void) {
return spine_tcm_runtime_version();
}
#endif
#define SPINE_TCM_VERSION (spine_tcm_version())
#ifdef __cplusplus
}
#endif
#endif
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_buffer_types()) {
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_buffer_types()) {
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_buffer_types();
#endif
ggml/src/ggml-cpu/unary-ops.cpp
+337
View File
@@ -0,0 +1,337 @@
#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_xielu(float x, float alpha_n, float alpha_p, float beta, float eps) {
if (x > 0.0f) {
return alpha_p * x * x + beta * x;
} else {
const float min_x_eps = fminf(x, eps);
return (expm1f(min_x_eps) - x) * alpha_n + beta * 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);
}
static inline float op_expm1(float x) {
return expf(x) - 1.0f;
}
static inline float op_softplus(float x) {
return (x > 20.0f) ? x : logf(1.0f + expf(x));
}
static inline float op_floor(float x) {
return floorf(x);
}
static inline float op_ceil(float x) {
return ceilf(x);
}
static inline float op_round(float x) {
return roundf(x);
}
static inline float op_trunc(float x) {
return truncf(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_rows(src0) && ggml_is_contiguous_rows(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");
}
}
template <float (*op)(float, ggml_tensor *)>
static void unary_op_params(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");
}
}
// Extend vec_unary_op to support functors
template <typename Op, typename src0_t, typename dst_t>
static inline void vec_unary_op_functor(int64_t n, dst_t * y, const src0_t * x, Op op) {
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])));
}
}
// Extend apply_unary_op to support functors
template <typename Op, typename src0_t, typename dst_t>
static void apply_unary_op_functor(const ggml_compute_params * params, ggml_tensor * dst, Op op) {
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_functor(ne0, dst_ptr, src0_ptr, op);
}
}
// Generic dispatcher for functors
template <typename Op>
static void unary_op_functor(const ggml_compute_params * params, ggml_tensor * dst, Op op) {
const ggml_tensor * src0 = dst->src[0];
/* */ if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { // all f32
apply_unary_op_functor<Op, float, float>(params, dst, op);
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { // all f16
apply_unary_op_functor<Op, ggml_fp16_t, ggml_fp16_t>(params, dst, op);
} else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_BF16) { // all bf16
apply_unary_op_functor<Op, ggml_bf16_t, ggml_bf16_t>(params, dst, op);
} else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_F32) {
apply_unary_op_functor<Op, ggml_bf16_t, float>(params, dst, op);
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
apply_unary_op_functor<Op, ggml_fp16_t, float>(params, dst, op);
} 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);
}
void ggml_compute_forward_expm1(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_expm1>(params, dst);
}
void ggml_compute_forward_softplus(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_softplus>(params, dst);
}
void ggml_compute_forward_floor(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_floor>(params, dst);
}
void ggml_compute_forward_ceil(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_ceil>(params, dst);
}
void ggml_compute_forward_round(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_round>(params, dst);
}
void ggml_compute_forward_trunc(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_trunc>(params, dst);
}
void ggml_compute_forward_xielu(const ggml_compute_params * params, ggml_tensor * dst) {
const float alpha_n = ggml_get_op_params_f32(dst, 1);
const float alpha_p = ggml_get_op_params_f32(dst, 2);
const float beta = ggml_get_op_params_f32(dst, 3);
const float eps = ggml_get_op_params_f32(dst, 4);
const auto xielu_op_params = [alpha_n, alpha_p, beta, eps](float f) {
return op_xielu(f, alpha_n, alpha_p, beta, eps);
};
unary_op_functor(params, dst, xielu_op_params);
}
ggml/src/ggml-cpu/unary-ops.h
+35
View File
@@ -0,0 +1,35 @@
#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);
void ggml_compute_forward_expm1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_softplus(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_floor(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_ceil(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_round(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_trunc(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_xielu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
#ifdef __cplusplus
}
#endif
ggml/src/ggml-cpu/vec.cpp
+613
View File
@@ -0,0 +1,613 @@
#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 svmla 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 = svmla_f32_m(pg, sum1, ax1, ay1);
}
// reduce sum1,sum2 to sum1
GGML_F32_VEC_REDUCE(sumf, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8);
#elif defined(__riscv_v_intrinsic)
int vl = __riscv_vsetvlmax_e32m8();
vfloat32m1_t vs = __riscv_vfmv_v_f_f32m1(0.0f, 1);
vfloat32m8_t vsum;
vfloat32m8_t ax;
vfloat32m8_t ay;
vsum = __riscv_vfmv_v_f_f32m8_tu(vsum, 0.0f, vl);
for (int i = 0; i < n; i += vl) {
vl = __riscv_vsetvl_e32m8(n - i);
ax = __riscv_vle32_v_f32m8_tu(ax, &x[i], vl);
ay = __riscv_vle32_v_f32m8_tu(ay, &y[i], vl);
vsum = __riscv_vfmacc_vv_f32m8_tu(vsum, ax, ay, vl);
}
vl = __riscv_vsetvlmax_e32m8();
vs = __riscv_vfredusum_vs_f32m8_f32m1(vsum, vs, vl);
sumf += __riscv_vfmv_f_s_f32m1_f32(vs);
#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
#elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfbfwma)
size_t vl = __riscv_vsetvlmax_e32m4();
// initialize accumulators to all zeroes
vfloat32m4_t vsum0 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
vfloat32m4_t vsum1 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
// calculate step size
const size_t epr = __riscv_vsetvlmax_e16m2();
const size_t step = epr * 2;
const int np = (n & ~(step - 1));
// unroll by 2
for (; i < np; i += step) {
vbfloat16m2_t ax0 = __riscv_vle16_v_bf16m2((const __bf16 *)&x[i], epr);
vbfloat16m2_t ay0 = __riscv_vle16_v_bf16m2((const __bf16 *)&y[i], epr);
vsum0 = __riscv_vfwmaccbf16_vv_f32m4(vsum0, ax0, ay0, epr);
__asm__ __volatile__ ("" ::: "memory");
vbfloat16m2_t ax1 = __riscv_vle16_v_bf16m2((const __bf16 *)&x[i + epr], epr);
vbfloat16m2_t ay1 = __riscv_vle16_v_bf16m2((const __bf16 *)&y[i + epr], epr);
vsum1 = __riscv_vfwmaccbf16_vv_f32m4(vsum1, ax1, ay1, epr);
__asm__ __volatile__ ("" ::: "memory");
}
// accumulate in 1 register
vsum0 = __riscv_vfadd_vv_f32m4(vsum0, vsum1, vl);
// leftovers
for (i = np; i < n; i += vl) {
vl = __riscv_vsetvl_e16m2(n - i);
vbfloat16m2_t ax0 = __riscv_vle16_v_bf16m2((const __bf16 *)&x[i], vl);
vbfloat16m2_t ay0 = __riscv_vle16_v_bf16m2((const __bf16 *)&y[i], vl);
vsum0 = __riscv_vfwmaccbf16_vv_f32m4(vsum0, ax0, ay0, vl);
}
// reduce
vl = __riscv_vsetvlmax_e32m4();
vfloat32m1_t redsum = __riscv_vfredusum_vs_f32m4_f32m1(vsum0, __riscv_vfmv_v_f_f32m1(0.0f, 1), vl);
sumf += __riscv_vfmv_f_s_f32m1_f32(redsum);
#elif defined(__POWER9_VECTOR__) || defined(__VXE__) || defined(__VXE2__)
const int np = (n & ~(GGML_BF16_STEP - 1));
if (np > 0) {
GGML_F32_VEC sum[4] = {GGML_F32_VEC_ZERO};
for (; i < np; i += GGML_BF16_STEP) {
GGML_BF16_VEC vx0 = GGML_BF16_VEC_LOAD(x + i);
GGML_BF16_VEC vx1 = GGML_BF16_VEC_LOAD(x + i + 8);
GGML_BF16_VEC vy0 = GGML_BF16_VEC_LOAD(y + i);
GGML_BF16_VEC vy1 = GGML_BF16_VEC_LOAD(y + i + 8);
GGML_BF16_FMA_LO(sum[0], vx0, vy0);
GGML_BF16_FMA_HI(sum[1], vx0, vy0);
GGML_BF16_FMA_LO(sum[2], vx1, vy1);
GGML_BF16_FMA_HI(sum[3], vx1, vy1);
}
GGML_F32x4_REDUCE_4(sumf, sum[0], sum[1], sum[2], sum[3]);
}
#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)
#if defined(__ARM_FEATURE_SVE)
const int ggml_f16_epr = svcnth();
const int ggml_f16_step = 8 * ggml_f16_epr;
const int np = n - (n % ggml_f16_step);
const int np2 = n - (n % ggml_f16_epr);
svfloat32_t sum1_lo = svdup_n_f32(0.0f);
svfloat32_t sum1_hi = svdup_n_f32(0.0f);
svfloat32_t sum2_lo = svdup_n_f32(0.0f);
svfloat32_t sum2_hi = svdup_n_f32(0.0f);
svfloat32_t sum3_lo = svdup_n_f32(0.0f);
svfloat32_t sum3_hi = svdup_n_f32(0.0f);
svfloat32_t sum4_lo = svdup_n_f32(0.0f);
svfloat32_t sum4_hi = svdup_n_f32(0.0f);
for (int i = 0; i < np; i += ggml_f16_step) {
ggml_sve_f16_fma_widened(&sum1_lo, &sum1_hi, GGML_F16x_VEC_LOAD(x + i + 0 * ggml_f16_epr, 0), GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0));
ggml_sve_f16_fma_widened(&sum2_lo, &sum2_hi, GGML_F16x_VEC_LOAD(x + i + 1 * ggml_f16_epr, 1), GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1));
ggml_sve_f16_fma_widened(&sum3_lo, &sum3_hi, GGML_F16x_VEC_LOAD(x + i + 2 * ggml_f16_epr, 2), GGML_F16x_VEC_LOAD(y + i + 2 * ggml_f16_epr, 2));
ggml_sve_f16_fma_widened(&sum4_lo, &sum4_hi, GGML_F16x_VEC_LOAD(x + i + 3 * ggml_f16_epr, 3), GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3));
ggml_sve_f16_fma_widened(&sum1_lo, &sum1_hi, GGML_F16x_VEC_LOAD(x + i + 4 * ggml_f16_epr, 4), GGML_F16x_VEC_LOAD(y + i + 4 * ggml_f16_epr, 4));
ggml_sve_f16_fma_widened(&sum2_lo, &sum2_hi, GGML_F16x_VEC_LOAD(x + i + 5 * ggml_f16_epr, 5), GGML_F16x_VEC_LOAD(y + i + 5 * ggml_f16_epr, 5));
ggml_sve_f16_fma_widened(&sum3_lo, &sum3_hi, GGML_F16x_VEC_LOAD(x + i + 6 * ggml_f16_epr, 6), GGML_F16x_VEC_LOAD(y + i + 6 * ggml_f16_epr, 6));
ggml_sve_f16_fma_widened(&sum4_lo, &sum4_hi, GGML_F16x_VEC_LOAD(x + i + 7 * ggml_f16_epr, 7), GGML_F16x_VEC_LOAD(y + i + 7 * ggml_f16_epr, 7));
}
for (int i = np; i < np2; i += ggml_f16_epr) {
ggml_sve_f16_fma_widened(&sum1_lo, &sum1_hi, GGML_F16x_VEC_LOAD(x + i, 0), GGML_F16x_VEC_LOAD(y + i, 0));
}
if (np2 < n) {
const svbool_t pg = svwhilelt_b16(np2, n);
const svfloat16_t rx = svld1_f16(pg, (const __fp16 *)(x + np2));
const svfloat16_t ry = svld1_f16(pg, (const __fp16 *)(y + np2));
ggml_sve_f16_fma_widened(&sum1_lo, &sum1_hi, rx, ry);
}
sum1_lo = svadd_f32_m(DEFAULT_PG32, sum1_lo, sum2_lo);
sum1_hi = svadd_f32_m(DEFAULT_PG32, sum1_hi, sum2_hi);
sum3_lo = svadd_f32_m(DEFAULT_PG32, sum3_lo, sum4_lo);
sum3_hi = svadd_f32_m(DEFAULT_PG32, sum3_hi, sum4_hi);
sum1_lo = svadd_f32_m(DEFAULT_PG32, sum1_lo, sum3_lo);
sum1_hi = svadd_f32_m(DEFAULT_PG32, sum1_hi, sum3_hi);
sumf = ggml_sve_sum_f32x2(sum1_lo, sum1_hi);
#elif defined(__riscv_v_intrinsic)
#if defined(__riscv_zvfh)
int vl = __riscv_vsetvlmax_e32m2();
vfloat32m1_t vs = __riscv_vfmv_v_f_f32m1(0.0f, 1);
vfloat32m2_t vsum;
vfloat16m1_t ax;
vfloat16m1_t ay;
vsum = __riscv_vreinterpret_v_u32m2_f32m2(__riscv_vmv_v_x_u32m2(0, vl));
for (int i = 0; i < n; i += vl) {
vl = __riscv_vsetvl_e16m1(n - i);
ax = __riscv_vle16_v_f16m1_tu(ax, (const _Float16 *)&x[i], vl);
ay = __riscv_vle16_v_f16m1_tu(ay, (const _Float16 *)&y[i], vl);
vsum = __riscv_vfwmacc_vv_f32m2_tu(vsum, ax, ay, vl);
}
vl = __riscv_vsetvlmax_e32m1();
vfloat32m1_t ac0 = __riscv_vfadd_vv_f32m1(__riscv_vget_v_f32m2_f32m1(vsum, 0), __riscv_vget_v_f32m2_f32m1(vsum, 1), vl);
vs = __riscv_vfredusum_vs_f32m1_f32m1(ac0, vs, vl);
sumf += __riscv_vfmv_f_s_f32m1_f32(vs);
#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 // __riscv_zvfh
#else
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));
#endif
#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 // GGML_SIMD
*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_FEATURE_SVE) && defined(__aarch64__)
const int vlen = svcntw();
for (; i < n; i += vlen) {
const svbool_t pg = svwhilelt_b32_s32(i, n);
svst1_f32(pg, y + i, ggml_v_silu(pg, svld1_f32(pg, 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)));
}
#elif defined(__riscv_v_intrinsic)
for (int vl; i < n; i += vl) {
vl = __riscv_vsetvl_e32m2(n - i);
vfloat32m2_t vx = __riscv_vle32_v_f32m2(&x[i], vl);
vfloat32m2_t vy = ggml_v_silu_m2(vx, vl);
__riscv_vse32_v_f32m2(&y[i], vy, vl);
}
#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_FEATURE_SVE) && defined(__aarch64__)
const int vlen = svcntw();
for (; i < n; i += vlen) {
const svbool_t pg = svwhilelt_b32_s32(i, n);
svst1_f32(pg, y + i, svmul_f32_x(pg, ggml_v_silu(pg, svld1_f32(pg, x + i)), svld1_f32(pg, 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)));
}
#elif defined(__riscv_v_intrinsic)
for (int vl; i < n; i += vl) {
vl = __riscv_vsetvl_e32m2(n - i);
vfloat32m2_t vx = __riscv_vle32_v_f32m2(&x[i], vl);
vfloat32m2_t vg = __riscv_vle32_v_f32m2(&g[i], vl);
vfloat32m2_t vy = __riscv_vfmul_vv_f32m2(ggml_v_silu_m2(vx, vl), vg, vl);
__riscv_vse32_v_f32m2(&y[i], vy, vl);
}
#endif
for (; i < n; ++i) {
y[i] = ggml_silu_f32(x[i]) * g[i];
}
}
ggml_float ggml_vec_cvar_f32(const int n, float * y, const float * x, const float mean) {
int i = 0;
ggml_float sum = 0;
// TODO: optimize to process the remaining elements in groups using the smaller vector sizes from AVX2 and SSE
// ref: https://github.com/ggml-org/llama.cpp/pull/15953#pullrequestreview-3310928344
#if defined(__AVX512F__) && defined(__AVX512DQ__)
for (; i + 15 < n; i += 16) {
__m512 val = _mm512_sub_ps(_mm512_loadu_ps(x + i),
_mm512_set1_ps(mean));
_mm512_storeu_ps(y + i, val);
sum += (ggml_float)_mm512_reduce_add_ps(_mm512_mul_ps(val, val));
}
#elif defined(__AVX2__) && defined(__FMA__)
for (; i + 7 < n; i += 8) {
__m256 val = _mm256_sub_ps(_mm256_loadu_ps(x + i),
_mm256_set1_ps(mean));
_mm256_storeu_ps(y + i, val);
val = _mm256_mul_ps(val,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 = _mm_sub_ps(_mm_loadu_ps(x + i),
_mm_set1_ps(mean));
_mm_storeu_ps(y + i, val);
val = _mm_mul_ps(val, 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 // __AVX__ || __AVX2__ || __AVX512F__
sum += (ggml_float)_mm_cvtss_f32(val);
}
#elif defined(__ARM_NEON) && defined(__aarch64__)
for (; i + 3 < n; i += 4) {
float32x4_t val = vsubq_f32(vld1q_f32(x + i),
vdupq_n_f32(mean));
vst1q_f32(y + i, val);
val = vmulq_f32(val, val);
sum += (ggml_float)vaddvq_f32(val);
}
#elif defined(__VXE__) || defined(__VXE2__)
for (; i + 3 < n; i += 4) {
float32x4_t val = vec_sub(vec_xl(0, x + i), vec_splats(mean));
vec_xst(val, 0, y + i);
val = vec_mul(val, val);
sum += (ggml_float)vec_hsum_f32x4(val);
}
#elif defined(__riscv_v_intrinsic)
vfloat64m1_t vsum = __riscv_vfmv_v_f_f64m1(0, 1);
for (int vl; i < n; i += vl) {
vl = __riscv_vsetvl_e32m2(n - i);
vfloat32m2_t val = __riscv_vfsub_vf_f32m2(__riscv_vle32_v_f32m2(&x[i], vl), mean, vl);
__riscv_vse32_v_f32m2(&y[i], val, vl);
val = __riscv_vfmul_vv_f32m2(val, val, vl);
vsum = __riscv_vfwredusum_vs_f32m2_f64m1(val, vsum, vl);
}
sum = (ggml_float)__riscv_vfmv_f_s_f64m1_f64(vsum);
#endif
for (; i < n; ++i) {
float val = x[i] - mean;
y[i] = val;
val *= val;
sum += (ggml_float)val;
}
return sum/n;
}
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_FEATURE_SVE) && defined(__aarch64__)
const int vlen = svcntw();
for (; i < n; i += vlen) {
const svbool_t pg = svwhilelt_b32_s32(i, n);
svfloat32_t val = ggml_v_expf(pg, svsub_f32_x(pg, svld1_f32(pg, x + i),
svdup_n_f32_x(pg, max)));
svst1_f32(pg, y + i, val);
sum += (ggml_float)svaddv_f32(pg, 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);
}
#elif defined(__riscv_v_intrinsic)
vfloat64m1_t vsum = __riscv_vfmv_v_f_f64m1(0, 1);
for (int avl; i < n; i += avl) {
avl = __riscv_vsetvl_e32m2(n - i);
vfloat32m2_t val = ggml_v_expf_m2(__riscv_vfsub_vf_f32m2(__riscv_vle32_v_f32m2(&x[i], avl), max, avl), avl);
__riscv_vse32_v_f32m2(&y[i], val, avl);
vsum = __riscv_vfwredusum_vs_f32m2_f64m1(val, vsum, avl);
}
return (ggml_float)__riscv_vfmv_f_s_f64m1_f64(vsum);
#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
+1570
View File
File diff suppressed because it is too large Load Diff