Skip to content

ggml-cpu: Add CPU backend support for KleidiAI library #11390

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 10 commits into from
Feb 20, 2025
Merged

ggml-cpu: Add CPU backend support for KleidiAI library #11390

Show file tree
Hide file tree
Changes from 5 commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
6adca19
ggml-cpu: Add CPU backend support for KleidiAI library
chaxu01 Jan 24, 2025
119d3bf
Add environmental variable GGML_KLEIDIAI_SME
chaxu01 Jan 30, 2025
f4eb1b3
Add support for multithread LHS conversion
chaxu01 Feb 4, 2025
3e08f37
Switch kernel selection order to dotprod and i8mm
chaxu01 Feb 5, 2025
9edd107
updates for review comments
chaxu01 Feb 7, 2025
e04880f
More updates for review comments
chaxu01 Feb 7, 2025
ca0c8b6
Reorganize and rename KleidiAI files
chaxu01 Feb 10, 2025
02315a8
Move ggml-cpu-traits.h to source file
chaxu01 Feb 10, 2025
09436f4
Update cmake for SME build and add alignment for SME
chaxu01 Feb 12, 2025
710552f
Remove append GGML_USE_CPU_KLEIDIAI to the GGML_CDEF_PUBLIC list
chaxu01 Feb 20, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions ggml/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -101,6 +101,7 @@ endif()

option(GGML_CPU_HBM "ggml: use memkind for CPU HBM" OFF)
option(GGML_CPU_AARCH64 "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
option(GGML_CPU_KLEIDIAI "ggml: use KleidiAI optimized kernels if applicable" OFF)
option(GGML_AVX "ggml: enable AVX" ${INS_ENB})
option(GGML_AVX_VNNI "ggml: enable AVX-VNNI" OFF)
option(GGML_AVX2 "ggml: enable AVX2" ${INS_ENB})
Expand Down
1 change: 1 addition & 0 deletions ggml/include/ggml-cpu.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -95,6 +95,7 @@ extern "C" {
GGML_BACKEND_API int ggml_cpu_has_matmul_int8(void);
GGML_BACKEND_API int ggml_cpu_has_sve (void);
GGML_BACKEND_API int ggml_cpu_get_sve_cnt (void); // sve vector length in bytes
GGML_BACKEND_API int ggml_cpu_has_sme (void);
// other
GGML_BACKEND_API int ggml_cpu_has_riscv_v (void);
GGML_BACKEND_API int ggml_cpu_has_vsx (void);
Expand Down
103 changes: 97 additions & 6 deletions ggml/src/ggml-cpu/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -111,21 +111,23 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
function(check_arm_feature tag code)
set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
set(CMAKE_REQUIRED_FLAGS "${ARM_MCPU_FLAG}+${tag}")
check_cxx_source_runs(
"${code}"
GGML_MACHINE_SUPPORTS_${tag}
)
check_cxx_source_runs("${code}" GGML_MACHINE_SUPPORTS_${tag})
if (GGML_MACHINE_SUPPORTS_${tag})
set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+${tag}" PARENT_SCOPE)
else()
set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+no${tag}" PARENT_SCOPE)
set(CMAKE_REQUIRED_FLAGS "${ARM_MCPU_FLAG}+no${tag}")
check_cxx_source_compiles("${code}" GGML_MACHINE_SUPPORTS_no${tag})
if (GGML_MACHINE_SUPPORTS_no${tag})
set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+no${tag}" PARENT_SCOPE)
endif()
endif()
set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
endfunction()

check_arm_feature(dotprod "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }")
check_arm_feature(i8mm "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }")
check_arm_feature(sve "#include <arm_sve.h>\nint main() { svfloat32_t _a, _b; volatile svfloat32_t _c = svadd_f32_z(svptrue_b8(), _a, _b); return 0; }")
check_arm_feature(sme "#include <arm_sme.h>\n__arm_locally_streaming int main() { __asm__ volatile(\"smstart; smstop;\"); return 0; }")

list(APPEND ARCH_FLAGS "${ARM_MCPU_FLAG}${ARM_MCPU_FLAG_FIX}")
else()
Expand All @@ -150,7 +152,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
if (ARM_FEATURE_RESULT)
message(WARNING "Failed to get ARM features")
else()
foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC)
foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC SME)
string(FIND "${ARM_FEATURE}" "__ARM_FEATURE_${feature} 1" feature_pos)
if (NOT ${feature_pos} EQUAL -1)
message(STATUS "ARM feature ${feature} enabled")
Expand Down Expand Up @@ -316,6 +318,95 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_AARCH64)
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 "v1.3.0")
set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
set(KLEIDIAI_ARCHIVE_MD5 "060bd2dc64642b091f461cc8dd7426d9")

if (POLICY CMP0135)
cmake_policy(SET CMP0135 NEW)
endif()

FetchContent_Declare(KleidiAI_Download
URL ${KLEIDIAI_DOWNLOAD_URL}
DOWNLOAD_EXTRACT_TIMESTAMP NEW
URL_HASH MD5=${KLEIDIAI_ARCHIVE_MD5})

FetchContent_MakeAvailable(KleidiAI_Download)
FetchContent_GetProperties(KleidiAI_Download
SOURCE_DIR KLEIDIAI_SRC
POPULATED KLEIDIAI_POPULATED)

if (NOT KLEIDIAI_POPULATED)
message(FATAL_ERROR "KleidiAI source downloaded failed.")
endif()

add_compile_definitions(GGML_USE_CPU_KLEIDIAI)

# Remove kleidiai target after fetching it
if (TARGET kleidiai)
set_target_properties(kleidiai PROPERTIES EXCLUDE_FROM_ALL TRUE)
endif()

list(APPEND GGML_CPU_SOURCES
ggml-cpu/ggml-kleidiai/ggml-kleidiai.cpp
ggml-cpu/ggml-kleidiai/kleidiai_kernels.cpp
ggml-cpu/ggml-kleidiai/ggml-kleidiai.h
ggml-cpu/ggml-kleidiai/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/pack/)

set(ARCH_FLAGS_TEMP "${ARCH_FLAGS}")
if (NOT ARCH_FLAGS_TEMP)
string(REGEX MATCH "-march=[^ ]+" ARCH_FLAGS_TEMP "${CMAKE_C_FLAGS}")
endif()
string(FIND "${ARCH_FLAGS_TEMP}" "+dotprod" DOTPROD_ENABLED)
string(FIND "${ARCH_FLAGS_TEMP}" "+i8mm" I8MM_ENABLED)
string(FIND "${ARCH_FLAGS_TEMP}" "+sme" SME_ENABLED)

set(PRIVATE_ARCH_FLAGS ${ARCH_FLAGS})

list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c)
list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c)
list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c)
list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c)

if (NOT DOTPROD_ENABLED MATCHES -1)
list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c)
list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c)
list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c)
endif()

if (NOT I8MM_ENABLED MATCHES -1)
list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c)
endif()

if (NOT SME_ENABLED MATCHES -1)
list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c)
list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c)
set(PRIVATE_ARCH_FLAGS "${PRIVATE_ARCH_FLAGS}+sve+sve2")
endif()

list(APPEND GGML_CDEF_PUBLIC GGML_USE_CPU_KLEIDIAI)
Copy link
Member

@slaren slaren Feb 19, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

GGML_CDEF_PUBLIC has been removed, and this will have no effect. If it is important to add this definition, it should be done with target_compile_definitions.

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})
Expand Down
33 changes: 29 additions & 4 deletions ggml/src/ggml-cpu/ggml-cpu.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -114,7 +114,8 @@ struct ggml_arm_arch_features_type {
int has_i8mm;
int has_sve;
int sve_cnt;
} ggml_arm_arch_features = {-1, -1, -1, -1, 0};
int has_sme;
} ggml_arm_arch_features = {-1, -1, -1, -1, 0, -1};
#endif


Expand Down Expand Up @@ -2389,15 +2390,20 @@ bool ggml_is_numa(void) {
#define HWCAP2_I8MM (1 << 13)
#endif

#if !defined(HWCAP2_SME)
#define HWCAP2_SME (1 << 23)
#endif

static void ggml_init_arm_arch_features(void) {
#if defined(__linux__) && defined(__aarch64__)
uint32_t hwcap = getauxval(AT_HWCAP);
uint32_t hwcap2 = getauxval(AT_HWCAP2);

ggml_arm_arch_features.has_neon = !!(hwcap & HWCAP_ASIMD);
ggml_arm_arch_features.has_neon = !!(hwcap & HWCAP_ASIMD);
ggml_arm_arch_features.has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
ggml_arm_arch_features.has_i8mm = !!(hwcap2 & HWCAP2_I8MM);
ggml_arm_arch_features.has_sve = !!(hwcap & HWCAP_SVE);
ggml_arm_arch_features.has_i8mm = !!(hwcap2 & HWCAP2_I8MM);
ggml_arm_arch_features.has_sve = !!(hwcap & HWCAP_SVE);
ggml_arm_arch_features.has_sme = !!(hwcap2 & HWCAP2_SME);

#if defined(__ARM_FEATURE_SVE)
ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
Expand All @@ -2420,6 +2426,11 @@ static void ggml_init_arm_arch_features(void) {
}
ggml_arm_arch_features.has_i8mm = oldp;

if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, NULL, 0) != 0) {
oldp = 0;
}
ggml_arm_arch_features.has_sme = oldp;

ggml_arm_arch_features.has_sve = 0;
ggml_arm_arch_features.sve_cnt = 0;
#else
Expand All @@ -2443,6 +2454,12 @@ static void ggml_init_arm_arch_features(void) {
ggml_arm_arch_features.has_sve = 0;
ggml_arm_arch_features.sve_cnt = 0;
#endif

#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_SME2)
ggml_arm_arch_features.has_sme = 1;
#else
ggml_arm_arch_features.has_sme = 0;
#endif
#endif
}
#endif
Expand Down Expand Up @@ -14349,6 +14366,14 @@ int ggml_cpu_get_sve_cnt(void) {
#endif
}

int ggml_cpu_has_sme(void) {
#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SME)
return ggml_arm_arch_features.has_sme;
#else
return 0;
#endif
}

void ggml_cpu_init(void) {
// needed to initialize f16 tables
{
Expand Down
16 changes: 16 additions & 0 deletions ggml/src/ggml-cpu/ggml-cpu.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,6 +14,10 @@
#include "ggml-cpu-hbm.h"
#endif

#ifdef GGML_USE_CPU_KLEIDIAI
#include "ggml-kleidiai/ggml-kleidiai.h"
#endif

#if defined(__APPLE__)
#include <sys/types.h>
#include <sys/sysctl.h>
Expand All @@ -39,6 +43,12 @@ std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_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_AARCH64
if (ggml_backend_cpu_aarch64_buffer_type()) {
bufts.push_back(ggml_backend_cpu_aarch64_buffer_type());
Expand Down Expand Up @@ -540,6 +550,9 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
static std::string sve_cnt = std::to_string(ggml_cpu_get_sve_cnt());
features.push_back({ "SVE_CNT", sve_cnt.c_str() });
}
if (ggml_cpu_has_sme()) {
features.push_back({ "SME", "1" });
}
if (ggml_cpu_has_riscv_v()) {
features.push_back({ "RISCV_V", "1" });
}
Expand All @@ -561,6 +574,9 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
#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_AARCH64
features.push_back({ "AARCH64_REPACK", "1" });
#endif
Expand Down
Loading
Loading