@fugood/llama.node 0.3.13 → 0.3.15

package/src/llama.cpp/ggml/src/ggml-cann/ggml-cann.cpp

@@ -796,11 +796,11 @@ static bool need_transform(ggml_type type) {
  * @param buffer The CANN buffer from which to initialize the tensor.
  * @param tensor Pointer to the tensor to be initialized.
  */
-static void ggml_backend_cann_buffer_init_tensor(
+static enum ggml_status ggml_backend_cann_buffer_init_tensor(
     ggml_backend_buffer_t buffer, ggml_tensor* tensor) {
     if (tensor->view_src != NULL && tensor->view_offs == 0) {
         GGML_ASSERT(tensor->view_src->buffer->buft == buffer->buft);
-        return;
+        return GGML_STATUS_SUCCESS;
     }
 
     // TODO: can backend doesn't support quantized yet. Just leave the code
@@ -817,6 +817,7 @@ static void ggml_backend_cann_buffer_init_tensor(
                                   memset_size, 0, memset_size));
         }
     }
+    return GGML_STATUS_SUCCESS;
 }
 
 // TODO: need handle tensor which has paddings.
@@ -1688,11 +1689,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_MUL_MAT: {
             switch (op->src[0]->type) {
                 case GGML_TYPE_Q8_0:
-                    // Current groupsize should not be greater than k-1 in
-                    // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize
-                    if (op->src[0]->ne[0] <= QK8_0) {
-                        return false;
-                    }
                 case GGML_TYPE_F16:
                 case GGML_TYPE_F32:
                 case GGML_TYPE_Q4_0:

package/src/llama.cpp/ggml/src/ggml-cann/kernels/dup.cpp

@@ -1,7 +1,5 @@
 #include "kernel_operator.h"
 
-#include <cmath>
-
 using namespace AscendC;
 
 #define BUFFER_NUM 2
@@ -183,7 +181,7 @@ extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp32(
     copy_to_ub(output_ne_gm, output_ne_ub, 32);
     copy_to_ub(output_nb_gm, output_nb_ub, 32);
 
-    DupByRows<float_t, float_t> op;
+    DupByRows<float, float> op;
     op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub);
     op.dup();
 }
@@ -206,7 +204,7 @@ extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp32_to_fp16(
     copy_to_ub(output_ne_gm, output_ne_ub, 32);
     copy_to_ub(output_nb_gm, output_nb_ub, 32);
 
-    DupByRows<float_t, half> op;
+    DupByRows<float, half> op;
     op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub);
     op.dup_with_cast();
 }
@@ -230,7 +228,7 @@ extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp16_to_fp32(
     copy_to_ub(output_ne_gm, output_ne_ub, 32);
     copy_to_ub(output_nb_gm, output_nb_ub, 32);
 
-    DupByRows<half, float_t> op;
+    DupByRows<half, float> op;
     op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub);
     op.dup_with_cast();
 }

package/src/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt

@@ -111,14 +111,15 @@ 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("int main() { return 0; }" GGML_MACHINE_SUPPORTS_no${tag})
+                        if (GGML_MACHINE_SUPPORTS_no${tag})
+                            set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+no${tag}" PARENT_SCOPE)
+                        endif()
                     endif()
                     set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
                 endfunction()
@@ -126,6 +127,7 @@ 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 +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 ${feature} enabled")
@@ -217,6 +219,10 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             if (GGML_AVX_VNNI)
                 list(APPEND ARCH_DEFINITIONS __AVXVNNI__ GGML_AVX_VNNI)
             endif()
+            if (GGML_BMI2)
+                # MSVC does not define macro __BMI2__
+                list(APPEND ARCH_DEFINITIONS __BMI2__ GGML_BMI2)
+            endif()
         else ()
             if (GGML_NATIVE)
                 list(APPEND ARCH_FLAGS -march=native)
@@ -231,6 +237,10 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                     list(APPEND ARCH_FLAGS -mfma)
                     list(APPEND ARCH_DEFINITIONS GGML_FMA)
                 endif()
+                if (GGML_BMI2)
+                    list(APPEND ARCH_FLAGS -mbmi2)
+                    list(APPEND ARCH_DEFINITIONS GGML_BMI2)
+                endif()
                 if (GGML_AVX)
                     list(APPEND ARCH_FLAGS -mavx)
                     list(APPEND ARCH_DEFINITIONS GGML_AVX)
@@ -277,21 +287,25 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 endif()
             endif()
         endif()
-    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
+    elseif ("${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "ppc64le " OR "${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "powerpc ")
         message(STATUS "PowerPC detected")
-        execute_process(COMMAND bash -c "grep POWER10 /proc/cpuinfo | head -n 1" OUTPUT_VARIABLE POWER10_M)
-        string(FIND "${POWER10_M}" "POWER10" substring_index)
-        if (NOT DEFINED substring_index OR "${substring_index}" STREQUAL "")
-            set(substring_index -1)
+        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()
 
-        if (${substring_index} GREATER_EQUAL 0)
-        list(APPEND ARCH_FLAGS -mcpu=power10)
+        string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M}")
+        string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}")
+
+        if (EXTRACTED_NUMBER GREATER_EQUAL 10)
+            list(APPEND ARCH_FLAGS -mcpu=power10 -mpowerpc64)
+        elseif (EXTRACTED_NUMBER EQUAL 9)
+            list(APPEND ARCH_FLAGS -mcpu=power9 -mpowerpc64)
         elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
-        list(APPEND ARCH_FLAGS -mcpu=powerpc64le)
+            list(APPEND ARCH_FLAGS -mcpu=powerpc64le -mtune=native)
         else()
-            list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
-            # TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
+            list(APPEND ARCH_FLAGS -mcpu=native -mtune=native -mpowerpc64)
         endif()
     elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
         message(STATUS "loongarch64 detected")
@@ -308,6 +322,27 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         if (GGML_RVV)
             list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d)
         endif()
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "s390x")
+        message(STATUS "s390x detected")
+        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 -mtune=z15)
+        elseif (${S390X_M} MATCHES "3931")
+            message(STATUS "z16 target")
+            list(APPEND ARCH_FLAGS -march=z16 -mtune=z16)
+        else()
+            message(STATUS "Unknown target")
+            message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")
+            list(APPEND ARCH_FLAGS -march=native -mtune=native)
+        endif()
+
+        if (GGML_VXE)
+            list(APPEND ARCH_FLAGS -mvx -mzvector)
+        endif()
     else()
         message(STATUS "Unknown architecture")
     endif()
@@ -316,6 +351,94 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         target_compile_definitions(${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/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/
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/)
+
+        set(ARCH_FLAGS_TEMP "${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()
+
+        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})

package/src/llama.cpp/ggml/src/ggml-cpu/amx/amx.cpp

@@ -50,10 +50,11 @@ static void * ggml_backend_amx_buffer_get_base(ggml_backend_buffer_t buffer) {
     return (void *) (buffer->context);
 }
 
-static void ggml_backend_amx_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+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,

package/src/llama.cpp/ggml/src/ggml-cpu/cpu-feats-x86.cpp

@@ -278,6 +278,10 @@ static int ggml_backend_cpu_x86_score() {
     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;

package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp

@@ -4135,10 +4135,11 @@ static const ggml::cpu::tensor_traits * ggml_aarch64_get_optimal_repack_type(con
     return nullptr;
 }
 
-static void ggml_backend_cpu_aarch64_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+static enum ggml_status ggml_backend_cpu_aarch64_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
     tensor->extra = (void *) const_cast<ggml::cpu::tensor_traits *>(ggml_aarch64_get_optimal_repack_type(tensor));
 
     GGML_UNUSED(buffer);
+    return GGML_STATUS_SUCCESS;
 }
 
 static void ggml_backend_cpu_aarch64_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor,

package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -59,6 +59,15 @@ struct ggml_compute_params {
 #endif
 #endif
 
+#if defined(__s390x__) && defined(__VEC__)
+#ifndef __VXE__
+#define __VXE__
+#endif
+#ifndef __VXE2__
+#define __VXE2__
+#endif
+#endif
+
 #if defined(__ARM_FEATURE_SVE)
 #include <arm_sve.h>
 #include <sys/prctl.h>
@@ -359,6 +368,148 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
 #endif
 #endif
 
+#if defined(__VXE__) || defined(__VXE2__)
+#include <vecintrin.h>
+
+#define vec_neg(a)    (-(a))                // Vector Negate
+#define vec_add(a, b) ((a) + (b))           // Vector Add
+#define vec_sub(a, b) ((a) - (b))           // Vector Subtract
+#define vec_mul(a, b) ((a) * (b))           // Vector Multiply
+#define vec_div(a, b) ((a) / (b))           // Vector Divide
+#define vec_sl(a, b)  ((a) << (b))          // Vector Shift Left
+#define vec_sra(a, b) ((a) >> (b))          // Vector Shift Right
+#define vec_sr(a, b)  ((a) >> (b))          // Vector Shift Right Algebraic
+#define vec_slo(a, b) vec_slb(a, (b) << 64) // Vector Shift Left by Octet
+#define vec_sro(a, b) vec_srb(a, (b) << 64) // Vector Shift Right by Octet
+
+#ifndef vec_and
+#define vec_and(a, b) ((a) & (b)) // Vector AND
+#endif
+
+#ifndef vec_or
+#define vec_or(a, b)  ((a) | (b)) // Vector OR
+#endif
+
+#ifndef vec_xor
+#define vec_xor(a, b) ((a) ^ (b)) // Vector XOR
+#endif
+
+typedef signed char char8x16_t __attribute__((vector_size(16)));
+typedef unsigned char uchar8x16_t __attribute__((vector_size(16)));
+
+typedef int8_t  int8x16_t __attribute__((vector_size(16)));
+typedef int16_t int16x8_t __attribute__((vector_size(16)));
+typedef int32_t int32x4_t __attribute__((vector_size(16)));
+
+typedef uint8_t  uint8x16_t __attribute__((vector_size(16)));
+typedef uint16_t uint16x8_t __attribute__((vector_size(16)));
+typedef uint32_t uint32x4_t __attribute__((vector_size(16)));
+
+typedef float float32x4_t __attribute__((vector_size(16)));
+typedef double double64x2_t __attribute((vector_size(16)));
+
+typedef signed long long long64x2_t __attribute((vector_size(16)));
+typedef unsigned long long ulong64x2_t __attribute__((vector_size(16)));
+
+typedef struct ggml_uint8x16x2_t {
+    uint8x16_t val[2];
+} ggml_uint8x16x2_t;
+
+inline static ggml_uint8x16x2_t ggml_vec_xl_u8x2(const uint8_t * ptr) {
+    ggml_uint8x16x2_t res;
+
+    res.val[0] = vec_xl( 0, ptr);
+    res.val[1] = vec_xl(16, ptr);
+
+    return res;
+}
+
+typedef struct ggml_uint8x16x4_t {
+    uint8x16_t val[4];
+} ggml_uint8x16x4_t;
+
+inline static ggml_uint8x16x4_t ggml_vec_xl_u8x4(const uint8_t * ptr) {
+    ggml_uint8x16x4_t res;
+
+    res.val[0] = vec_xl( 0, ptr);
+    res.val[1] = vec_xl(16, ptr);
+    res.val[2] = vec_xl(32, ptr);
+    res.val[3] = vec_xl(48, ptr);
+
+    return res;
+}
+
+typedef struct ggml_int8x16x4_t {
+    int8x16_t val[4];
+} ggml_int8x16x4_t;
+
+inline static ggml_int8x16x4_t ggml_vec_xl_s8x4(const int8_t * ptr) {
+    ggml_int8x16x4_t res;
+
+    res.val[0] = vec_xl( 0, ptr);
+    res.val[1] = vec_xl(16, ptr);
+    res.val[2] = vec_xl(32, ptr);
+    res.val[3] = vec_xl(48, ptr);
+
+    return res;
+}
+
+typedef struct ggml_int16x8x2_t {
+    int16x8_t val[2];
+} ggml_int16x8x2_t;
+
+inline static ggml_int16x8x2_t ggml_vec_xl_s16x2(const int16_t * ptr) {
+    ggml_int16x8x2_t res;
+
+    res.val[0] = vec_xl( 0, ptr);
+    res.val[1] = vec_xl(16, ptr);
+
+    return res;
+}
+
+/*
+    ! WARNING: Very slow. Use vec_perm if possible. Refer to iq4_xs
+    !          or iq4_nl for example implementation.
+*/
+inline static int8x16_t ggml_vec_tbl(int8x16_t a, uint8x16_t b) {
+    int8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
+inline static int16x8_t vec_padd_s16(int16x8_t a, int16x8_t b) {
+    const uchar8x16_t v_maske = {  0,  1,  4,  5,  8,  9, 12, 13,
+                                  16, 17, 20, 21, 24, 25, 28, 29 };
+
+    const int16x8_t v_abo = vec_pack((int32x4_t)a, (int32x4_t)b);
+    const int16x8_t v_abe = vec_perm(a, b, v_maske);
+    return v_abo + v_abe;
+}
+
+inline static int32x4_t ggml_vec_dot(int32x4_t acc, int8x16_t a, int8x16_t b) {
+    const int16x8_t p = vec_mule(a, b) + vec_mulo(a, b);
+    return acc + (vec_unpackh(p) + vec_unpackl(p));
+}
+
+#endif
+
 #if defined(__loongarch_asx)
 /* float type data load instructions */
 static __m128 __lsx_vreplfr2vr_s(const float val) {