ggml-cpu: Add CPU backend support for KleidiAI library

Author: chaxu01

common/common.cpp

@@ -1099,6 +1099,8 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
         mparams.kv_overrides = params.kv_overrides.data();
     }
 
+    mparams.n_threads = params.cpuparams.n_threads;
+
     return mparams;
 }
 

ggml/CMakeLists.txt

@@ -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})

ggml/include/ggml-backend.h

@@ -189,7 +189,7 @@ extern "C" {
     // Set the number of threads for the backend
     typedef void                         (*ggml_backend_set_n_threads_t)(ggml_backend_t backend, int n_threads);
     // Get additional buffer types provided by the device (returns a NULL-terminated array)
-    typedef ggml_backend_buffer_type_t * (*ggml_backend_dev_get_extra_bufts_t)(ggml_backend_dev_t device);
+    typedef ggml_backend_buffer_type_t * (*ggml_backend_dev_get_extra_bufts_t)(ggml_backend_dev_t device, int n_threads);
     // Set the abort callback for the backend
     typedef void                         (*ggml_backend_set_abort_callback_t)(ggml_backend_t backend, ggml_abort_callback abort_callback, void * abort_callback_data);
     // Get a list of feature flags supported by the backend (returns a NULL-terminated array)

ggml/include/ggml-cpu.h

@@ -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);

ggml/src/ggml-cpu/CMakeLists.txt

@@ -126,6 +126,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 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()
@@ -150,7 +151,7 @@ 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 ${feature} enabled")
@@ -316,6 +317,91 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         target_compile_definitions(${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_SHA "v1.2.0")
+        set(KLEIDIAI_DOWNLOAD_URL "https://gitlab.arm.com/kleidi/kleidiai/-/archive/${KLEIDIAI_COMMIT_SHA}/kleidiai-${KLEIDIAI_COMMIT_SHA}.tar.gz")
+        set(KLEIDIAI_ARCHIVE_MD5  "cebcb660079bf15626e7bdaecd18f49c")
+
+        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/)
+
+        string(FIND ${ARCH_FLAGS} "+dotprod" DOTPROD_ENABLED)
+        string(FIND ${ARCH_FLAGS} "+i8mm" I8MM_ENABLED)
+        string(FIND ${ARCH_FLAGS} "+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)
+        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})

ggml/src/ggml-cpu/ggml-cpu-traits.cpp

@@ -10,7 +10,7 @@ extra_buffer_type::~extra_buffer_type() {}
 }  // namespace ggml::cpu
 
 bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) {
-    for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
+    for (auto extra : ggml_backend_cpu_get_extra_buffers_type(params->nth)) {
         if (extra && extra->context) {
             auto buf_extra     = (ggml::cpu::extra_buffer_type *) extra->context;
             auto tensor_traits = buf_extra->get_tensor_traits(op);
@@ -23,7 +23,7 @@ bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct
 }
 
 bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size) {
-    for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
+    for (auto extra : ggml_backend_cpu_get_extra_buffers_type(n_threads)) {
         if (extra && extra->context) {
             auto buf_extra     = (ggml::cpu::extra_buffer_type *) extra->context;
             auto tensor_traits = buf_extra->get_tensor_traits(op);

ggml/src/ggml-cpu/ggml-cpu-traits.h

@@ -33,6 +33,6 @@ class extra_buffer_type {
 }  // namespace ggml::cpu
 
 // implemented in ggml-cpu.cpp.
-std::vector<ggml_backend_buffer_type_t> & ggml_backend_cpu_get_extra_buffers_type();
+std::vector<ggml_backend_buffer_type_t> & ggml_backend_cpu_get_extra_buffers_type(int n_threads);
 
 #endif

ggml/src/ggml-cpu/ggml-cpu.c

@@ -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
 
 
@@ -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);
@@ -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
@@ -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
@@ -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
     {

ggml/src/ggml-cpu/ggml-cpu.cpp

@@ -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>
@@ -29,8 +33,8 @@
 
 // ggml-backend interface
 
-std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type() {
-    static std::vector<ggml_backend_buffer_type_t> bufts = []() {
+std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type(int n_threads) {
+    static std::vector<ggml_backend_buffer_type_t> bufts = [n_threads]() {
         std::vector<ggml_backend_buffer_type_t> bufts;
 
 #if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
@@ -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(n_threads)) {
+            bufts.push_back(ggml_backend_cpu_kleidiai_buffer_type(n_threads));
+        }
+#endif
+
 #ifdef GGML_USE_CPU_AARCH64
         if (ggml_backend_cpu_aarch64_buffer_type()) {
             bufts.push_back(ggml_backend_cpu_aarch64_buffer_type());
@@ -48,19 +58,21 @@ std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type
         bufts.push_back(NULL);
 
         return bufts;
+
+        GGML_UNUSED(n_threads);
     }();
 
     return bufts;
 }
 
-static ggml_backend_buffer_type_t * ggml_backend_cpu_device_get_extra_buffers_type(ggml_backend_dev_t device) {
-    return ggml_backend_cpu_get_extra_buffers_type().data();
+static ggml_backend_buffer_type_t * ggml_backend_cpu_device_get_extra_buffers_type(ggml_backend_dev_t device, int n_threads) {
+    return ggml_backend_cpu_get_extra_buffers_type(n_threads).data();
 
     GGML_UNUSED(device);
 }
 
 static bool ggml_backend_cpu_is_extra_buffer_type(ggml_backend_buffer_type_t buft) {
-    for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
+    for (auto extra : ggml_backend_cpu_get_extra_buffers_type(-1)) {
         if (extra && extra == buft) return true;
     }
     return false;
@@ -375,7 +387,7 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
     }
 
     // extra_buffer_op?
-    for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
+    for (auto extra : ggml_backend_cpu_get_extra_buffers_type(-1)) {
         if (extra) {
             auto buf_extra = (ggml::cpu::extra_buffer_type*) extra->context;
             if (buf_extra && buf_extra->supports_op(dev, op)) {
@@ -540,6 +552,9 @@ 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" });
         }
@@ -561,6 +576,9 @@ 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_REPACK", "1" });
+    #endif
     #ifdef GGML_USE_CPU_AARCH64
         features.push_back({ "AARCH64_REPACK", "1" });
     #endif