ggml-cpu : split arch-specific implementations (llama/13892)

* move ggml-cpu-aarch64 to repack

* split quantize_row_q8_0/1

* split helper functions

* split ggml_vec_dot_q4_0_q8_0

* split ggml_vec_dot_q4_1_q8_1

* split ggml_vec_dot_q5_0_q8_0

* split ggml_vec_dot_q5_1_q8_1

* split ggml_vec_dot_q8_0_q8_0

* split ggml_vec_dot_tq1_0_q8_K

* split ggml_vec_dot_tq2_0_q8_K

* split ggml_vec_dot_q2_K_q8_K

* split ggml_vec_dot_q3_K_q8_K

* split ggml_vec_dot_q4_K_q8_K

* split ggml_vec_dot_q5_K_q8_K

* split ggml_vec_dot_q6_K_q8_K

* split ggml_vec_dot_iq2_xxs_q8_K

* split ggml_vec_dot_iq2_xs_q8_K

* split ggml_vec_dot_iq2_s_q8_K

* split ggml_vec_dot_iq3_xxs_q8_K

* split ggml_vec_dot_iq3_s_q8_K

* split ggml_vec_dot_iq1_s_q8_K

* split ggml_vec_dot_iq1_m_q8_K

* split ggml_vec_dot_iq4_nl_q8_0

* split ggml_vec_dot_iq4_xs_q8_K

* fix typos

* fix missing prototypes

* rename ggml-cpu-quants.c

* rename ggml-cpu-traits

* rename arm folder

* move cpu-feats-x86.cpp

* rename ggml-cpu-hbm

* update arm detection macro in quants.c

* move iq quant tables

* split ggml_quantize_mat_q8_0/K

* split ggml_gemv_*

* split ggml_gemm_*

* rename namespace aarch64 to repack

* use weak aliases to replace test macros

* rename GGML_CPU_AARCH64 to GGML_CPU_REPACK

* rename more aarch64 to repack

* clean up rebase leftover

* fix compilation errors

* remove trailing spaces

* try to fix clang compilation errors

* try to fix clang compilation errors again

* try to fix clang compilation errors, 3rd attempt

* try to fix clang compilation errors, 4th attempt

* try to fix clang compilation errors, 5th attempt

* try to fix clang compilation errors, 6th attempt

* try to fix clang compilation errors, 7th attempt

* try to fix clang compilation errors, 8th attempt

* try to fix clang compilation errors, 9th attempt

* more cleanup

* fix compilation errors

* fix apple targets

* fix a typo in arm version of ggml_vec_dot_q4_K_q8_K

Co-authored-by: Georgi Gerganov <[email protected]>

---------

Co-authored-by: Georgi Gerganov <[email protected]>

ggml/CMakeLists.txt

@@ -105,7 +105,7 @@ message(DEBUG "GGML_NATIVE_DEFAULT : ${GGML_NATIVE_DEFAULT}")
 message(DEBUG "INS_ENB             : ${INS_ENB}")
 
 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_REPACK       "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_SSE42            "ggml: enable SSE 4.2"          ${INS_ENB})
 option(GGML_AVX              "ggml: enable AVX"              ${INS_ENB})

ggml/src/ggml-common.h

@@ -1074,6 +1074,10 @@ GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512)
     0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101,
 GGML_TABLE_END()
 
+GGML_TABLE_BEGIN(int8_t, kvalues_iq4nl, 16)
+    -127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113,
+GGML_TABLE_END()
+
 #define NGRID_IQ1S 2048
 #define IQ1S_DELTA 0.125f
 #define IQ1M_DELTA 0.125f

ggml/src/ggml-cpu/CMakeLists.txt

@@ -10,14 +10,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
     list (APPEND GGML_CPU_SOURCES
         ggml-cpu/ggml-cpu.c
         ggml-cpu/ggml-cpu.cpp
-        ggml-cpu/ggml-cpu-aarch64.cpp
-        ggml-cpu/ggml-cpu-aarch64.h
-        ggml-cpu/ggml-cpu-hbm.cpp
-        ggml-cpu/ggml-cpu-hbm.h
-        ggml-cpu/ggml-cpu-quants.c
-        ggml-cpu/ggml-cpu-quants.h
-        ggml-cpu/ggml-cpu-traits.cpp
-        ggml-cpu/ggml-cpu-traits.h
+        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
@@ -84,6 +84,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
     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()
@@ -167,6 +172,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         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)
@@ -302,7 +312,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             # 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 ${GGML_CPU_NAME}-feats)
-            add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/cpu-feats-x86.cpp)
+            add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/arch/x86/cpu-feats.cpp)
             target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . .. ../include)
             target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARCH_DEFINITIONS})
             target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
@@ -311,6 +321,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         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)
@@ -338,6 +349,8 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         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)
@@ -347,6 +360,10 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "riscv64")
         message(STATUS "riscv64 detected")
+        list(APPEND GGML_CPU_SOURCES
+            ggml-cpu/arch/riscv/quants.c
+            ggml-cpu/arch/riscv/repack.cpp
+            )
         if (GGML_RVV)
             if (GGML_XTHEADVECTOR)
                 list(APPEND ARCH_FLAGS -march=rv64gc_xtheadvector -mabi=lp64d)
@@ -358,6 +375,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
         message(STATUS "s390x detected")
+        list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/s390/quants.c)
         file(READ "/proc/cpuinfo" CPUINFO_CONTENTS)
         string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS})
 
@@ -381,12 +399,16 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         if (GGML_VXE)
             list(APPEND ARCH_FLAGS -mvx -mzvector)
         endif()
+    elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm")
+        message(STATUS "Wasm detected")
+        list (APPEND GGML_CPU_SOURCES ggml-cpu/arch/wasm/quants.c)
     else()
-        message(STATUS "Unknown architecture")
+        message(WARNING "Unknown CPU architecture. Falling back to generic implementations.")
+        list(APPEND ARCH_FLAGS -DGGML_CPU_GENERIC)
     endif()
 
-    if (GGML_CPU_AARCH64)
-        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_AARCH64)
+    if (GGML_CPU_REPACK)
+        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_REPACK)
     endif()
 
     if (GGML_CPU_KLEIDIAI)

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

@@ -5,7 +5,7 @@
 #include "ggml-backend.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 
 #if defined(__gnu_linux__)
 #include <sys/syscall.h>

ggml/src/ggml-cpu/amx/mmq.cpp

@@ -8,7 +8,7 @@
 #include "mmq.h"
 #include "ggml-impl.h"
 #include "ggml-cpu-impl.h"
-#include "ggml-cpu-quants.h"
+#include "quants.h"
 #include "ggml-quants.h"
 #include <algorithm>
 #include <type_traits>

ggml/src/ggml-cpu/arch/arm/repack.cpp

@@ -0,0 +1,2174 @@
+#define GGML_COMMON_IMPL_CPP
+#define GGML_COMMON_DECL_CPP
+#include "ggml-common.h"
+#include "ggml-backend-impl.h"
+
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+#include "ggml-cpu-impl.h"
+#include "traits.h"
+
+#include <cmath>
+#include <cstring>
+#include <cassert>
+#include <cstdlib> // for qsort
+#include <cstdio>  // for GGML_ASSERT
+
+#define GGML_CPU_CLANG_WORKAROUND
+#include "../../repack.h"
+
+#if defined(__GNUC__)
+#pragma GCC diagnostic ignored "-Woverlength-strings"
+#endif
+
+#define UNUSED GGML_UNUSED
+
+void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(QK8_0 == 32);
+    assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
+
+    block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy;
+
+#if defined(__ARM_NEON)
+    float32x4_t srcv[4][8];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[8];
+
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            for (int j = 0; j < 8; j++) srcv[row_iter][j] = vld1q_f32(x + row_iter * k + i * 32 + 4 * j);
+            for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[row_iter][j]);
+
+            for (int j = 0; j < 4; j++) amaxv[2 * j] = vmaxq_f32(asrcv[2 * j], asrcv[2 * j + 1]);
+            for (int j = 0; j < 2; j++) amaxv[4 * j] = vmaxq_f32(amaxv[4 * j], amaxv[4 * j + 2]);
+            for (int j = 0; j < 1; j++) amaxv[8 * j] = vmaxq_f32(amaxv[8 * j], amaxv[8 * j + 4]);
+
+            const float amax = vmaxvq_f32(amaxv[0]);
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < 8; j++) {
+            float32x4_t v = vmulq_n_f32(srcv[0][j], id[0]);
+            int32x4_t vi = vcvtnq_s32_f32(v);
+            y[i].qs[16 * j + 0] = vgetq_lane_s32(vi, 0);
+            y[i].qs[16 * j + 1] = vgetq_lane_s32(vi, 1);
+            y[i].qs[16 * j + 2] = vgetq_lane_s32(vi, 2);
+            y[i].qs[16 * j + 3] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[1][j], id[1]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[16 * j + 4] = vgetq_lane_s32(vi, 0);
+            y[i].qs[16 * j + 5] = vgetq_lane_s32(vi, 1);
+            y[i].qs[16 * j + 6] = vgetq_lane_s32(vi, 2);
+            y[i].qs[16 * j + 7] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[2][j], id[2]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[16 * j + 8] = vgetq_lane_s32(vi, 0);
+            y[i].qs[16 * j + 9] = vgetq_lane_s32(vi, 1);
+            y[i].qs[16 * j + 10] = vgetq_lane_s32(vi, 2);
+            y[i].qs[16 * j + 11] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[3][j], id[3]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[16 * j + 12] = vgetq_lane_s32(vi, 0);
+            y[i].qs[16 * j + 13] = vgetq_lane_s32(vi, 1);
+            y[i].qs[16 * j + 14] = vgetq_lane_s32(vi, 2);
+            y[i].qs[16 * j + 15] = vgetq_lane_s32(vi, 3);
+        }
+    }
+#else
+    // scalar
+    const int blck_size_interleave = 4;
+    float srcv[4][QK8_0];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            float amax = 0.0f; // absolute max
+
+            for (int j = 0; j < QK8_0; j++) {
+                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
+                amax = MAX(amax, fabsf(srcv[row_iter][j]));
+            }
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < QK8_0 * 4; j++) {
+            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
+            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
+            src_offset += (j % blck_size_interleave);
+
+            float x0 = srcv[src_id][src_offset] * id[src_id];
+            y[i].qs[j] = roundf(x0);
+        }
+    }
+#endif
+}
+
+void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(QK8_0 == 32);
+    assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
+
+    block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy;
+
+#if defined(__ARM_NEON)
+    float32x4_t srcv[4][8];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[8];
+
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            for (int j = 0; j < 8; j++) srcv[row_iter][j] = vld1q_f32(x + row_iter * k + i * 32 + 4 * j);
+            for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[row_iter][j]);
+
+            for (int j = 0; j < 4; j++) amaxv[2 * j] = vmaxq_f32(asrcv[2 * j], asrcv[2 * j + 1]);
+            for (int j = 0; j < 2; j++) amaxv[4 * j] = vmaxq_f32(amaxv[4 * j], amaxv[4 * j + 2]);
+            for (int j = 0; j < 1; j++) amaxv[8 * j] = vmaxq_f32(amaxv[8 * j], amaxv[8 * j + 4]);
+
+            const float amax = vmaxvq_f32(amaxv[0]);
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < 4; j++) {
+            float32x4_t v = vmulq_n_f32(srcv[0][2 * j], id[0]);
+            int32x4_t vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 0] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 1] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 2] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 3] = vgetq_lane_s32(vi, 3);
+            v = vmulq_n_f32(srcv[0][2 * j + 1], id[0]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 4] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 5] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 6] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 7] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[1][2 * j], id[1]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 8] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 9] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 10] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 11] = vgetq_lane_s32(vi, 3);
+            v = vmulq_n_f32(srcv[1][2 * j + 1], id[1]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 12] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 13] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 14] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 15] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[2][2 * j], id[2]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 16] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 17] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 18] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 19] = vgetq_lane_s32(vi, 3);
+            v = vmulq_n_f32(srcv[2][2 * j + 1], id[2]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 20] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 21] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 22] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 23] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[3][2 * j], id[3]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 24] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 25] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 26] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 27] = vgetq_lane_s32(vi, 3);
+            v = vmulq_n_f32(srcv[3][2 * j + 1], id[3]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 28] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 29] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 30] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 31] = vgetq_lane_s32(vi, 3);
+        }
+    }
+
+#else
+    // scalar
+    const int blck_size_interleave = 8;
+    float srcv[4][QK8_0];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            float amax = 0.0f; // absolute max
+
+            for (int j = 0; j < QK8_0; j++) {
+                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
+                amax = MAX(amax, fabsf(srcv[row_iter][j]));
+            }
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < QK8_0 * 4; j++) {
+            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
+            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
+            src_offset += (j % blck_size_interleave);
+
+            float x0 = srcv[src_id][src_offset] * id[src_id];
+            y[i].qs[j] = roundf(x0);
+        }
+    }
+#endif
+}
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx;
+
+        for (int c = 0; c < nc; c += ncols_interleaved) {
+            const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+            float32x4_t acc = vdupq_n_f32(0);
+            for (int b = 0; b < nb; b++) {
+                int8x16_t b0 = vld1q_s8((const int8_t *) b_ptr->qs);
+                int8x16_t b1 = vld1q_s8((const int8_t *) b_ptr->qs + 16);
+                int8x16_t b2 = vld1q_s8((const int8_t *) b_ptr->qs + 32);
+                int8x16_t b3 = vld1q_s8((const int8_t *) b_ptr->qs + 48);
+                float16x4_t bd = vld1_f16((const __fp16 *) b_ptr->d);
+
+                int8x16_t a0 = vld1q_s8(a_ptr->qs);
+                int8x16_t a1 = vld1q_s8(a_ptr->qs + qk/2);
+                float16x4_t ad = vld1_dup_f16((const __fp16 *) &a_ptr->d);
+
+                int32x4_t ret = vdupq_n_s32(0);
+
+                ret = vdotq_laneq_s32(ret, b0 << 4, a0, 0);
+                ret = vdotq_laneq_s32(ret, b1 << 4, a0, 1);
+                ret = vdotq_laneq_s32(ret, b2 << 4, a0, 2);
+                ret = vdotq_laneq_s32(ret, b3 << 4, a0, 3);
+
+                ret = vdotq_laneq_s32(ret, b0 & 0xf0U, a1, 0);
+                ret = vdotq_laneq_s32(ret, b1 & 0xf0U, a1, 1);
+                ret = vdotq_laneq_s32(ret, b2 & 0xf0U, a1, 2);
+                ret = vdotq_laneq_s32(ret, b3 & 0xf0U, a1, 3);
+
+                acc = vfmaq_f32(acc, vcvtq_n_f32_s32(ret, 4),
+                                vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd)));
+                a_ptr++;
+                b_ptr++;
+            }
+            vst1q_f32(s, acc);
+            s += ncols_interleaved;
+        }
+        return;
+    }
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    float sumf[4];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                    }
+                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx;
+
+        for (int c = 0; c < nc; c += ncols_interleaved) {
+            const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+            float32x4_t acc = vdupq_n_f32(0);
+            for (int b = 0; b < nb; b++) {
+                int8x16_t b0 = vld1q_s8((const int8_t *) b_ptr->qs);
+                int8x16_t b1 = vld1q_s8((const int8_t *) b_ptr->qs + 16);
+                int8x16_t b2 = vld1q_s8((const int8_t *) b_ptr->qs + 32);
+                int8x16_t b3 = vld1q_s8((const int8_t *) b_ptr->qs + 48);
+                float16x4_t bd = vld1_f16((const __fp16 *) b_ptr->d);
+
+                int8x16_t a0 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs);
+                int8x16_t a1 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 1);
+                int8x16_t a2 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 2);
+                int8x16_t a3 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 3);
+                float16x4_t ad = vld1_dup_f16((const __fp16 *) &a_ptr->d);
+
+                int32x4_t ret0 = vdupq_n_s32(0);
+                int32x4_t ret1 = vdupq_n_s32(0);
+
+                ret0 = vdotq_s32(ret0, b0 << 4, a0);
+                ret1 = vdotq_s32(ret1, b1 << 4, a0);
+                ret0 = vdotq_s32(ret0, b2 << 4, a1);
+                ret1 = vdotq_s32(ret1, b3 << 4, a1);
+
+                ret0 = vdotq_s32(ret0, b0 & 0xf0U, a2);
+                ret1 = vdotq_s32(ret1, b1 & 0xf0U, a2);
+                ret0 = vdotq_s32(ret0, b2 & 0xf0U, a3);
+                ret1 = vdotq_s32(ret1, b3 & 0xf0U, a3);
+
+                int32x4_t ret = vpaddq_s32(ret0, ret1);
+
+                acc = vfmaq_f32(acc, vcvtq_n_f32_s32(ret, 4),
+                        vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd)));
+                a_ptr++;
+                b_ptr++;
+            }
+            vst1q_f32(s, acc);
+            s += ncols_interleaved;
+        }
+        return;
+    }
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    float sumf[4];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                    }
+                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
+#if defined(__ARM_FEATURE_SVE)
+    if (ggml_cpu_has_sve() && ggml_cpu_get_sve_cnt() == QK8_0) {
+        const void * b_ptr = vx;
+        const void * a_ptr = vy;
+        float * res_ptr = s;
+
+        __asm__ __volatile__(
+            "ptrue p0.b\n"
+            "add %x[b_ptr], %x[b_ptr], #0x10\n"
+            "1:"  // Column loop
+            "add x22, %x[a_ptr], #0x2\n"
+            "mov z31.b, #0x0\n"
+            "mov x21, %x[nb]\n"
+            "2:"  // Block loop
+            "ld1b { z30.b }, p0/Z, [%x[b_ptr]]\n"
+            "ld1b { z29.b }, p0/Z, [%x[b_ptr], #1, MUL VL]\n"
+            "mov z28.s, #0x0\n"
+            "mov z27.s, #0x0\n"
+            "ld1rd { z26.d }, p0/Z, [x22]\n"
+            "ld1b { z25.b }, p0/Z, [%x[b_ptr], #2, MUL VL]\n"
+            "sub x20, x22, #0x2\n"
+            "sub x21, x21, #0x1\n"
+            "ld1b { z24.b }, p0/Z, [%x[b_ptr], #3, MUL VL]\n"
+            "ld1rd { z23.d }, p0/Z, [x22, #8]\n"
+            "lsl z22.b, z30.b, #0x4\n"
+            "lsl z16.b, z29.b, #0x4\n"
+            "and z30.b, z30.b, #0xf0\n"
+            "and z29.b, z29.b, #0xf0\n"
+            "ld1rd { z21.d }, p0/Z, [x22, #16]\n"
+            "ld1rd { z20.d }, p0/Z, [x22, #24]\n"
+            "lsl z19.b, z25.b, #0x4\n"
+            "and z25.b, z25.b, #0xf0\n"
+            "ld1rh { z17.h }, p0/Z, [x20]\n"
+            "ld1h { z18.s }, p0/Z, [%x[b_ptr], #-1, MUL VL]\n"
+            "sdot z28.s, z22.b, z26.b\n"
+            "sdot z27.s, z16.b, z26.b\n"
+            "lsl z16.b, z24.b, #0x4\n"
+            "add x22, x22, #0x22\n"
+            "and z24.b, z24.b, #0xf0\n"
+            "add %x[b_ptr], %x[b_ptr], #0x90\n"
+            "fcvt z17.s, p0/m, z17.h\n"
+            "fcvt z18.s, p0/m, z18.h\n"
+            "sdot z28.s, z19.b, z23.b\n"
+            "sdot z27.s, z16.b, z23.b\n"
+            "fmul z18.s, z18.s, z17.s\n"
+            "sdot z28.s, z30.b, z21.b\n"
+            "sdot z27.s, z29.b, z21.b\n"
+            "sdot z28.s, z25.b, z20.b\n"
+            "sdot z27.s, z24.b, z20.b\n"
+            "uzp1 z17.s, z28.s, z27.s\n"
+            "uzp2 z16.s, z28.s, z27.s\n"
+            "add z17.s, z17.s, z16.s\n"
+            "asr z17.s, z17.s, #0x4\n"
+            "scvtf z17.s, p0/m, z17.s\n"
+            "fmla z31.s, p0/M, z17.s, z18.s\n"
+            "cbnz x21, 2b\n"
+            "sub %x[nc], %x[nc], #0x8\n"
+            "st1w { z31.s }, p0, [%x[res_ptr]]\n"
+            "add %x[res_ptr], %x[res_ptr], #0x20\n"
+            "cbnz %x[nc], 1b\n"
+            : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc)
+            : [a_ptr] "r" (a_ptr), [nb] "r" (nb)
+            : "memory", "p0", "x20", "x21", "x22", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+        );
+        return;
+    }
+#endif // #if defined(__ARM_FEATURE_SVE)
+
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
+    {
+        float sumf[8];
+        int sumi;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+
+            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int j = 0; j < ncols_interleaved; j++) {
+                        sumi = 0;
+                        for (int i = 0; i < blocklen; ++i) {
+                            const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                        }
+                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                    }
+                }
+            }
+            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+        const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl);
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        float * res_ptr = s;
+
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+
+            float32x4_t sumf = vdupq_n_f32(0);
+            for (int l = 0; l < nb; l++) {
+                uint8x16_t b_0 = vld1q_u8(b_ptr[l].qs + 0);
+                uint8x16_t b_1 = vld1q_u8(b_ptr[l].qs + 16);
+                uint8x16_t b_2 = vld1q_u8(b_ptr[l].qs + 32);
+                uint8x16_t b_3 = vld1q_u8(b_ptr[l].qs + 48);
+
+                int8x16_t b_0_hi = vqtbl1q_s8(kvalues, b_0 >> 4);
+                int8x16_t b_0_lo = vqtbl1q_s8(kvalues, b_0 & 0x0F);
+                int8x16_t b_1_hi = vqtbl1q_s8(kvalues, b_1 >> 4);
+                int8x16_t b_1_lo = vqtbl1q_s8(kvalues, b_1 & 0x0F);
+                int8x16_t b_2_hi = vqtbl1q_s8(kvalues, b_2 >> 4);
+                int8x16_t b_2_lo = vqtbl1q_s8(kvalues, b_2 & 0x0F);
+                int8x16_t b_3_hi = vqtbl1q_s8(kvalues, b_3 >> 4);
+                int8x16_t b_3_lo = vqtbl1q_s8(kvalues, b_3 & 0x0F);
+
+                int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 0);
+                int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16);
+
+                int32x4_t sumi = vdupq_n_s32(0);
+                sumi = vdotq_laneq_s32(sumi, b_0_lo, a_0, 0);
+                sumi = vdotq_laneq_s32(sumi, b_0_hi, a_1, 0);
+                sumi = vdotq_laneq_s32(sumi, b_1_lo, a_0, 1);
+                sumi = vdotq_laneq_s32(sumi, b_1_hi, a_1, 1);
+                sumi = vdotq_laneq_s32(sumi, b_2_lo, a_0, 2);
+                sumi = vdotq_laneq_s32(sumi, b_2_hi, a_1, 2);
+                sumi = vdotq_laneq_s32(sumi, b_3_lo, a_0, 3);
+                sumi = vdotq_laneq_s32(sumi, b_3_hi, a_1, 3);
+
+                float32x4_t a_d = vcvt_f32_f16(vld1_dup_f16((const float16_t *)&a_ptr[l].d));
+                float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *)b_ptr[l].d));
+                float32x4_t d = a_d * b_d;
+
+                sumf = vmlaq_f32(sumf, d, vcvtq_f32_s32(sumi));
+            }
+
+            vst1q_f32(res_ptr + x * 4, sumf);
+        }
+        return;
+    }
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
+    {
+        float sumf[4];
+        int sumi;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+
+            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int j = 0; j < ncols_interleaved; j++) {
+                        sumi = 0;
+                        for (int i = 0; i < blocklen; ++i) {
+                            const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                            const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
+                        }
+                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                    }
+                }
+            }
+            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
+void ggml_gemm_q4_0_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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
+    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+        const void * b_ptr = vx;
+        const void * a_ptr = vy;
+        float * res_ptr = s;
+        size_t res_stride = bs * sizeof(float);
+
+        __asm__ __volatile__(
+            "mov x10, %x[nr]\n"
+            "mov x9, #0x88\n"
+            "cmp x10, #0x10\n"
+            "mul x9, %x[nb], x9\n"
+            "blt 4f\n"
+            "1:"  // Row loop
+            "add x28, %x[b_ptr], #0x8\n"
+            "mov x27, %x[nc]\n"
+            "add x26, %x[res_ptr], %x[res_stride], LSL #4\n"
+            "2:"  // Column loop
+            "add x25, %x[a_ptr], #0x8\n"
+            "movi v15.16b, #0x0\n"
+            "movi v19.16b, #0x0\n"
+            "mov x24, %x[nb]\n"
+            "add x23, x25, x9\n"
+            "movi v18.16b, #0x0\n"
+            "movi v14.16b, #0x0\n"
+            "add x22, x23, x9\n"
+            "movi v11.16b, #0x0\n"
+            "movi v13.16b, #0x0\n"
+            "add x21, x22, x9\n"
+            "movi v23.16b, #0x0\n"
+            "movi v16.16b, #0x0\n"
+            "movi v25.16b, #0x0\n"
+            "movi v7.16b, #0x0\n"
+            "movi v0.16b, #0x0\n"
+            "movi v4.16b, #0x0\n"
+            "movi v5.16b, #0x0\n"
+            "movi v21.16b, #0x0\n"
+            "movi v8.16b, #0x0\n"
+            "movi v1.16b, #0x0\n"
+            "3:"  // Block loop
+            "ldr q3, [x28, #0x0]\n"
+            "ldr q31, [x25, #0x0]\n"
+            "movi v28.16b, #0x4\n"
+            "movi v10.4s, #0x0\n"
+            "ldr q22, [x28, #0x10]\n"
+            "ldr q6, [x25, #0x10]\n"
+            "movi v29.4s, #0x0\n"
+            "movi v9.4s, #0x0\n"
+            "ldr q27, [x28, #0x20]\n"
+            "ldr q30, [x28, #0x30]\n"
+            "movi v20.4s, #0x0\n"
+            "movi v24.16b, #0xf0\n"
+            "ldr d2, [x25, #-0x8]\n"
+            "ldr d26, [x23, #-0x8]\n"
+            "sshl v12.16b, v3.16b, v28.16b\n"
+            "sub x20, x28, #0x8\n"
+            "ldr d17, [x20, #0x0]\n"
+            "and v3.16b, v3.16b, v24.16b\n"
+            "subs x24, x24, #0x1\n"
+            "add x28, x28, #0x48\n"
+            ".inst 0x4f9fe18a  // sdot v10.4s, v12.16b, v31.4b[0]\n"
+            ".inst 0x4fbfe19d  // sdot v29.4s, v12.16b, v31.4b[1]\n"
+            ".inst 0x4f9fe989  // sdot v9.4s, v12.16b, v31.4b[2]\n"
+            ".inst 0x4fbfe994  // sdot v20.4s, v12.16b, v31.4b[3]\n"
+            "sshl v31.16b, v22.16b, v28.16b\n"
+            "and v22.16b, v22.16b, v24.16b\n"
+            "fcvtl v17.4s, v17.4h\n"
+            "fcvtl v2.4s, v2.4h\n"
+            "fcvtl v26.4s, v26.4h\n"
+            ".inst 0x4f86e3ea  // sdot v10.4s, v31.16b, v6.4b[0]\n"
+            ".inst 0x4fa6e3fd  // sdot v29.4s, v31.16b, v6.4b[1]\n"
+            ".inst 0x4f86ebe9  // sdot v9.4s, v31.16b, v6.4b[2]\n"
+            ".inst 0x4fa6ebf4  // sdot v20.4s, v31.16b, v6.4b[3]\n"
+            "sshl v6.16b, v27.16b, v28.16b\n"
+            "sshl v28.16b, v30.16b, v28.16b\n"
+            "and v27.16b, v27.16b, v24.16b\n"
+            "and v30.16b, v30.16b, v24.16b\n"
+            "ldr q24, [x25, #0x20]\n"
+            ".inst 0x4f98e0ca  // sdot v10.4s, v6.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e0dd  // sdot v29.4s, v6.16b, v24.4b[1]\n"
+            ".inst 0x4f98e8c9  // sdot v9.4s, v6.16b, v24.4b[2]\n"
+            ".inst 0x4fb8e8d4  // sdot v20.4s, v6.16b, v24.4b[3]\n"
+            "ldr q24, [x25, #0x30]\n"
+            ".inst 0x4f98e38a  // sdot v10.4s, v28.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e39d  // sdot v29.4s, v28.16b, v24.4b[1]\n"
+            ".inst 0x4f98eb89  // sdot v9.4s, v28.16b, v24.4b[2]\n"
+            ".inst 0x4fb8eb94  // sdot v20.4s, v28.16b, v24.4b[3]\n"
+            "ldr q24, [x25, #0x40]\n"
+            ".inst 0x4f98e06a  // sdot v10.4s, v3.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e07d  // sdot v29.4s, v3.16b, v24.4b[1]\n"
+            ".inst 0x4f98e869  // sdot v9.4s, v3.16b, v24.4b[2]\n"
+            ".inst 0x4fb8e874  // sdot v20.4s, v3.16b, v24.4b[3]\n"
+            "ldr q24, [x25, #0x50]\n"
+            ".inst 0x4f98e2ca  // sdot v10.4s, v22.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e2dd  // sdot v29.4s, v22.16b, v24.4b[1]\n"
+            ".inst 0x4f98eac9  // sdot v9.4s, v22.16b, v24.4b[2]\n"
+            ".inst 0x4fb8ead4  // sdot v20.4s, v22.16b, v24.4b[3]\n"
+            "ldr q24, [x25, #0x60]\n"
+            ".inst 0x4f98e36a  // sdot v10.4s, v27.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e37d  // sdot v29.4s, v27.16b, v24.4b[1]\n"
+            ".inst 0x4f98eb69  // sdot v9.4s, v27.16b, v24.4b[2]\n"
+            ".inst 0x4fb8eb74  // sdot v20.4s, v27.16b, v24.4b[3]\n"
+            "ldr q24, [x25, #0x70]\n"
+            "add x25, x25, #0x88\n"
+            ".inst 0x4f98e3ca  // sdot v10.4s, v30.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e3dd  // sdot v29.4s, v30.16b, v24.4b[1]\n"
+            ".inst 0x4f98ebc9  // sdot v9.4s, v30.16b, v24.4b[2]\n"
+            ".inst 0x4fb8ebd4  // sdot v20.4s, v30.16b, v24.4b[3]\n"
+            "fmul v24.4s, v17.4s, v2.s[0]\n"
+            "scvtf v10.4s, v10.4s, #0x4\n"
+            "scvtf v29.4s, v29.4s, #0x4\n"
+            "scvtf v9.4s, v9.4s, #0x4\n"
+            "scvtf v20.4s, v20.4s, #0x4\n"
+            "fmla v15.4s, v10.4s, v24.4s\n"
+            "ldr q24, [x23, #0x0]\n"
+            "fmul v10.4s, v17.4s, v2.s[1]\n"
+            "fmla v19.4s, v29.4s, v10.4s\n"
+            "ldr q10, [x23, #0x10]\n"
+            "fmul v29.4s, v17.4s, v2.s[2]\n"
+            "fmul v2.4s, v17.4s, v2.s[3]\n"
+            "fmla v18.4s, v9.4s, v29.4s\n"
+            "movi v9.4s, #0x0\n"
+            "movi v29.4s, #0x0\n"
+            ".inst 0x4f98e189  // sdot v9.4s, v12.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e19d  // sdot v29.4s, v12.16b, v24.4b[1]\n"
+            "fmla v14.4s, v20.4s, v2.4s\n"
+            "movi v20.4s, #0x0\n"
+            "movi v2.4s, #0x0\n"
+            ".inst 0x4f98e994  // sdot v20.4s, v12.16b, v24.4b[2]\n"
+            ".inst 0x4fb8e982  // sdot v2.4s, v12.16b, v24.4b[3]\n"
+            "ldr q24, [x23, #0x20]\n"
+            ".inst 0x4f8ae3e9  // sdot v9.4s, v31.16b, v10.4b[0]\n"
+            ".inst 0x4faae3fd  // sdot v29.4s, v31.16b, v10.4b[1]\n"
+            ".inst 0x4f8aebf4  // sdot v20.4s, v31.16b, v10.4b[2]\n"
+            ".inst 0x4faaebe2  // sdot v2.4s, v31.16b, v10.4b[3]\n"
+            "ldr q10, [x23, #0x30]\n"
+            ".inst 0x4f98e0c9  // sdot v9.4s, v6.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e0dd  // sdot v29.4s, v6.16b, v24.4b[1]\n"
+            ".inst 0x4f98e8d4  // sdot v20.4s, v6.16b, v24.4b[2]\n"
+            ".inst 0x4fb8e8c2  // sdot v2.4s, v6.16b, v24.4b[3]\n"
+            "ldr q24, [x23, #0x40]\n"
+            ".inst 0x4f8ae389  // sdot v9.4s, v28.16b, v10.4b[0]\n"
+            ".inst 0x4faae39d  // sdot v29.4s, v28.16b, v10.4b[1]\n"
+            ".inst 0x4f8aeb94  // sdot v20.4s, v28.16b, v10.4b[2]\n"
+            ".inst 0x4faaeb82  // sdot v2.4s, v28.16b, v10.4b[3]\n"
+            "ldr q10, [x23, #0x50]\n"
+            ".inst 0x4f98e069  // sdot v9.4s, v3.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e07d  // sdot v29.4s, v3.16b, v24.4b[1]\n"
+            ".inst 0x4f98e874  // sdot v20.4s, v3.16b, v24.4b[2]\n"
+            ".inst 0x4fb8e862  // sdot v2.4s, v3.16b, v24.4b[3]\n"
+            "ldr q24, [x23, #0x60]\n"
+            ".inst 0x4f8ae2c9  // sdot v9.4s, v22.16b, v10.4b[0]\n"
+            ".inst 0x4faae2dd  // sdot v29.4s, v22.16b, v10.4b[1]\n"
+            ".inst 0x4f8aead4  // sdot v20.4s, v22.16b, v10.4b[2]\n"
+            ".inst 0x4faaeac2  // sdot v2.4s, v22.16b, v10.4b[3]\n"
+            "ldr q10, [x23, #0x70]\n"
+            "add x23, x23, #0x88\n"
+            ".inst 0x4f98e369  // sdot v9.4s, v27.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e37d  // sdot v29.4s, v27.16b, v24.4b[1]\n"
+            ".inst 0x4f98eb74  // sdot v20.4s, v27.16b, v24.4b[2]\n"
+            ".inst 0x4fb8eb62  // sdot v2.4s, v27.16b, v24.4b[3]\n"
+            "ldr q24, [x22, #0x0]\n"
+            ".inst 0x4f8ae3c9  // sdot v9.4s, v30.16b, v10.4b[0]\n"
+            ".inst 0x4faae3dd  // sdot v29.4s, v30.16b, v10.4b[1]\n"
+            ".inst 0x4f8aebd4  // sdot v20.4s, v30.16b, v10.4b[2]\n"
+            ".inst 0x4faaebc2  // sdot v2.4s, v30.16b, v10.4b[3]\n"
+            "fmul v10.4s, v17.4s, v26.s[0]\n"
+            "scvtf v9.4s, v9.4s, #0x4\n"
+            "scvtf v29.4s, v29.4s, #0x4\n"
+            "scvtf v20.4s, v20.4s, #0x4\n"
+            "scvtf v2.4s, v2.4s, #0x4\n"
+            "fmla v11.4s, v9.4s, v10.4s\n"
+            "ldr q9, [x22, #0x10]\n"
+            "fmul v10.4s, v17.4s, v26.s[1]\n"
+            "fmla v13.4s, v29.4s, v10.4s\n"
+            "ldr d29, [x22, #-0x8]\n"
+            "fmul v10.4s, v17.4s, v26.s[2]\n"
+            "fmul v26.4s, v17.4s, v26.s[3]\n"
+            "fcvtl v29.4s, v29.4h\n"
+            "fmla v23.4s, v20.4s, v10.4s\n"
+            "movi v20.4s, #0x0\n"
+            "movi v10.4s, #0x0\n"
+            "fmla v16.4s, v2.4s, v26.4s\n"
+            "movi v26.4s, #0x0\n"
+            "movi v2.4s, #0x0\n"
+            ".inst 0x4f98e194  // sdot v20.4s, v12.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e18a  // sdot v10.4s, v12.16b, v24.4b[1]\n"
+            ".inst 0x4f98e99a  // sdot v26.4s, v12.16b, v24.4b[2]\n"
+            ".inst 0x4fb8e982  // sdot v2.4s, v12.16b, v24.4b[3]\n"
+            "ldr q24, [x22, #0x20]\n"
+            ".inst 0x4f89e3f4  // sdot v20.4s, v31.16b, v9.4b[0]\n"
+            ".inst 0x4fa9e3ea  // sdot v10.4s, v31.16b, v9.4b[1]\n"
+            ".inst 0x4f89ebfa  // sdot v26.4s, v31.16b, v9.4b[2]\n"
+            ".inst 0x4fa9ebe2  // sdot v2.4s, v31.16b, v9.4b[3]\n"
+            "ldr q9, [x22, #0x30]\n"
+            ".inst 0x4f98e0d4  // sdot v20.4s, v6.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e0ca  // sdot v10.4s, v6.16b, v24.4b[1]\n"
+            ".inst 0x4f98e8da  // sdot v26.4s, v6.16b, v24.4b[2]\n"
+            ".inst 0x4fb8e8c2  // sdot v2.4s, v6.16b, v24.4b[3]\n"
+            "ldr q24, [x22, #0x40]\n"
+            ".inst 0x4f89e394  // sdot v20.4s, v28.16b, v9.4b[0]\n"
+            ".inst 0x4fa9e38a  // sdot v10.4s, v28.16b, v9.4b[1]\n"
+            ".inst 0x4f89eb9a  // sdot v26.4s, v28.16b, v9.4b[2]\n"
+            ".inst 0x4fa9eb82  // sdot v2.4s, v28.16b, v9.4b[3]\n"
+            "ldr q9, [x22, #0x50]\n"
+            ".inst 0x4f98e074  // sdot v20.4s, v3.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e06a  // sdot v10.4s, v3.16b, v24.4b[1]\n"
+            ".inst 0x4f98e87a  // sdot v26.4s, v3.16b, v24.4b[2]\n"
+            ".inst 0x4fb8e862  // sdot v2.4s, v3.16b, v24.4b[3]\n"
+            "ldr q24, [x22, #0x60]\n"
+            ".inst 0x4f89e2d4  // sdot v20.4s, v22.16b, v9.4b[0]\n"
+            ".inst 0x4fa9e2ca  // sdot v10.4s, v22.16b, v9.4b[1]\n"
+            ".inst 0x4f89eada  // sdot v26.4s, v22.16b, v9.4b[2]\n"
+            ".inst 0x4fa9eac2  // sdot v2.4s, v22.16b, v9.4b[3]\n"
+            "ldr q9, [x22, #0x70]\n"
+            "add x22, x22, #0x88\n"
+            ".inst 0x4f98e374  // sdot v20.4s, v27.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e36a  // sdot v10.4s, v27.16b, v24.4b[1]\n"
+            ".inst 0x4f98eb7a  // sdot v26.4s, v27.16b, v24.4b[2]\n"
+            ".inst 0x4fb8eb62  // sdot v2.4s, v27.16b, v24.4b[3]\n"
+            "ldr q24, [x21, #0x0]\n"
+            ".inst 0x4f89e3d4  // sdot v20.4s, v30.16b, v9.4b[0]\n"
+            ".inst 0x4fa9e3ca  // sdot v10.4s, v30.16b, v9.4b[1]\n"
+            ".inst 0x4f89ebda  // sdot v26.4s, v30.16b, v9.4b[2]\n"
+            ".inst 0x4fa9ebc2  // sdot v2.4s, v30.16b, v9.4b[3]\n"
+            "fmul v9.4s, v17.4s, v29.s[0]\n"
+            "scvtf v20.4s, v20.4s, #0x4\n"
+            "scvtf v10.4s, v10.4s, #0x4\n"
+            "scvtf v26.4s, v26.4s, #0x4\n"
+            "scvtf v2.4s, v2.4s, #0x4\n"
+            "fmla v25.4s, v20.4s, v9.4s\n"
+            "ldr q9, [x21, #0x10]\n"
+            "fmul v20.4s, v17.4s, v29.s[1]\n"
+            "fmla v7.4s, v10.4s, v20.4s\n"
+            "ldr d20, [x21, #-0x8]\n"
+            "fmul v10.4s, v17.4s, v29.s[2]\n"
+            "fmul v29.4s, v17.4s, v29.s[3]\n"
+            "fcvtl v20.4s, v20.4h\n"
+            "fmla v0.4s, v26.4s, v10.4s\n"
+            "movi v26.4s, #0x0\n"
+            "movi v10.4s, #0x0\n"
+            "fmla v4.4s, v2.4s, v29.4s\n"
+            "movi v2.4s, #0x0\n"
+            "movi v29.4s, #0x0\n"
+            ".inst 0x4f98e19a  // sdot v26.4s, v12.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e18a  // sdot v10.4s, v12.16b, v24.4b[1]\n"
+            ".inst 0x4f98e982  // sdot v2.4s, v12.16b, v24.4b[2]\n"
+            ".inst 0x4fb8e99d  // sdot v29.4s, v12.16b, v24.4b[3]\n"
+            "ldr q12, [x21, #0x20]\n"
+            "fmul v24.4s, v17.4s, v20.s[0]\n"
+            ".inst 0x4f89e3fa  // sdot v26.4s, v31.16b, v9.4b[0]\n"
+            ".inst 0x4fa9e3ea  // sdot v10.4s, v31.16b, v9.4b[1]\n"
+            ".inst 0x4f89ebe2  // sdot v2.4s, v31.16b, v9.4b[2]\n"
+            ".inst 0x4fa9ebfd  // sdot v29.4s, v31.16b, v9.4b[3]\n"
+            "ldr q9, [x21, #0x30]\n"
+            "fmul v31.4s, v17.4s, v20.s[1]\n"
+            ".inst 0x4f8ce0da  // sdot v26.4s, v6.16b, v12.4b[0]\n"
+            ".inst 0x4face0ca  // sdot v10.4s, v6.16b, v12.4b[1]\n"
+            ".inst 0x4f8ce8c2  // sdot v2.4s, v6.16b, v12.4b[2]\n"
+            ".inst 0x4face8dd  // sdot v29.4s, v6.16b, v12.4b[3]\n"
+            "ldr q12, [x21, #0x40]\n"
+            "fmul v6.4s, v17.4s, v20.s[2]\n"
+            "fmul v20.4s, v17.4s, v20.s[3]\n"
+            ".inst 0x4f89e39a  // sdot v26.4s, v28.16b, v9.4b[0]\n"
+            ".inst 0x4fa9e38a  // sdot v10.4s, v28.16b, v9.4b[1]\n"
+            ".inst 0x4f89eb82  // sdot v2.4s, v28.16b, v9.4b[2]\n"
+            ".inst 0x4fa9eb9d  // sdot v29.4s, v28.16b, v9.4b[3]\n"
+            "ldr q9, [x21, #0x50]\n"
+            ".inst 0x4f8ce07a  // sdot v26.4s, v3.16b, v12.4b[0]\n"
+            ".inst 0x4face06a  // sdot v10.4s, v3.16b, v12.4b[1]\n"
+            ".inst 0x4f8ce862  // sdot v2.4s, v3.16b, v12.4b[2]\n"
+            ".inst 0x4face87d  // sdot v29.4s, v3.16b, v12.4b[3]\n"
+            "ldr q12, [x21, #0x60]\n"
+            ".inst 0x4f89e2da  // sdot v26.4s, v22.16b, v9.4b[0]\n"
+            ".inst 0x4fa9e2ca  // sdot v10.4s, v22.16b, v9.4b[1]\n"
+            ".inst 0x4f89eac2  // sdot v2.4s, v22.16b, v9.4b[2]\n"
+            ".inst 0x4fa9eadd  // sdot v29.4s, v22.16b, v9.4b[3]\n"
+            "ldr q17, [x21, #0x70]\n"
+            "add x21, x21, #0x88\n"
+            ".inst 0x4f8ce37a  // sdot v26.4s, v27.16b, v12.4b[0]\n"
+            ".inst 0x4face36a  // sdot v10.4s, v27.16b, v12.4b[1]\n"
+            ".inst 0x4f8ceb62  // sdot v2.4s, v27.16b, v12.4b[2]\n"
+            ".inst 0x4faceb7d  // sdot v29.4s, v27.16b, v12.4b[3]\n"
+            ".inst 0x4f91e3da  // sdot v26.4s, v30.16b, v17.4b[0]\n"
+            ".inst 0x4fb1e3ca  // sdot v10.4s, v30.16b, v17.4b[1]\n"
+            ".inst 0x4f91ebc2  // sdot v2.4s, v30.16b, v17.4b[2]\n"
+            ".inst 0x4fb1ebdd  // sdot v29.4s, v30.16b, v17.4b[3]\n"
+            "scvtf v26.4s, v26.4s, #0x4\n"
+            "scvtf v10.4s, v10.4s, #0x4\n"
+            "fmla v5.4s, v26.4s, v24.4s\n"
+            "scvtf v2.4s, v2.4s, #0x4\n"
+            "scvtf v29.4s, v29.4s, #0x4\n"
+            "fmla v21.4s, v10.4s, v31.4s\n"
+            "fmla v8.4s, v2.4s, v6.4s\n"
+            "fmla v1.4s, v29.4s, v20.4s\n"
+            "bgt 3b\n"
+            "mov x20, %x[res_ptr]\n"
+            "subs x27, x27, #0x4\n"
+            "add %x[res_ptr], %x[res_ptr], #0x10\n"
+            "str q15, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q19, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q18, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q14, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q11, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q13, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q23, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q16, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q25, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q7, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q0, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q4, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q5, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q21, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q8, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q1, [x20, #0x0]\n"
+            "bne 2b\n"
+            "mov x20, #0x4\n"
+            "sub x10, x10, #0x10\n"
+            "cmp x10, #0x10\n"
+            "mov %x[res_ptr], x26\n"
+            "madd %x[a_ptr], x20, x9, %x[a_ptr]\n"
+            "bge 1b\n"
+            "4:"  // Row loop skip
+            "cbz x10, 9f\n"
+            "5:"  // Row tail: Row loop
+            "add x24, %x[b_ptr], #0x8\n"
+            "mov x23, %x[nc]\n"
+            "add x22, %x[res_ptr], %x[res_stride], LSL #2\n"
+            "6:"  // Row tail: Column loop
+            "movi v15.16b, #0x0\n"
+            "movi v19.16b, #0x0\n"
+            "add x25, %x[a_ptr], #0x8\n"
+            "mov x21, %x[nb]\n"
+            "movi v18.16b, #0x0\n"
+            "movi v14.16b, #0x0\n"
+            "7:"  // Row tail: Block loop
+            "ldr q7, [x24, #0x0]\n"
+            "ldr q5, [x25, #0x0]\n"
+            "movi v9.16b, #0x4\n"
+            "movi v4.4s, #0x0\n"
+            "ldr q3, [x24, #0x10]\n"
+            "ldr q2, [x25, #0x10]\n"
+            "movi v1.4s, #0x0\n"
+            "movi v0.4s, #0x0\n"
+            "ldr q13, [x24, #0x20]\n"
+            "ldr q31, [x25, #0x20]\n"
+            "movi v30.4s, #0x0\n"
+            "movi v29.16b, #0xf0\n"
+            "ldr q28, [x24, #0x30]\n"
+            "ldr q27, [x25, #0x30]\n"
+            "sshl v20.16b, v7.16b, v9.16b\n"
+            "sub x20, x24, #0x8\n"
+            "ldr q26, [x25, #0x40]\n"
+            "ldr q25, [x25, #0x50]\n"
+            "sshl v17.16b, v3.16b, v9.16b\n"
+            "and v7.16b, v7.16b, v29.16b\n"
+            "ldr q24, [x25, #0x60]\n"
+            "ldr q16, [x25, #0x70]\n"
+            "sshl v22.16b, v13.16b, v9.16b\n"
+            "and v3.16b, v3.16b, v29.16b\n"
+            "ldr d21, [x20, #0x0]\n"
+            "ldr d12, [x25, #-0x8]\n"
+            ".inst 0x4f85e284  // sdot v4.4s, v20.16b, v5.4b[0]\n"
+            ".inst 0x4fa5e281  // sdot v1.4s, v20.16b, v5.4b[1]\n"
+            ".inst 0x4f85ea80  // sdot v0.4s, v20.16b, v5.4b[2]\n"
+            ".inst 0x4fa5ea9e  // sdot v30.4s, v20.16b, v5.4b[3]\n"
+            "sshl v9.16b, v28.16b, v9.16b\n"
+            "subs x21, x21, #0x1\n"
+            "and v13.16b, v13.16b, v29.16b\n"
+            "and v28.16b, v28.16b, v29.16b\n"
+            "add x25, x25, #0x88\n"
+            "add x24, x24, #0x48\n"
+            "fcvtl v21.4s, v21.4h\n"
+            "fcvtl v12.4s, v12.4h\n"
+            ".inst 0x4f82e224  // sdot v4.4s, v17.16b, v2.4b[0]\n"
+            ".inst 0x4fa2e221  // sdot v1.4s, v17.16b, v2.4b[1]\n"
+            ".inst 0x4f82ea20  // sdot v0.4s, v17.16b, v2.4b[2]\n"
+            ".inst 0x4fa2ea3e  // sdot v30.4s, v17.16b, v2.4b[3]\n"
+            "fmul v11.4s, v21.4s, v12.s[0]\n"
+            "fmul v23.4s, v21.4s, v12.s[1]\n"
+            "fmul v17.4s, v21.4s, v12.s[2]\n"
+            ".inst 0x4f9fe2c4  // sdot v4.4s, v22.16b, v31.4b[0]\n"
+            "fmul v6.4s, v21.4s, v12.s[3]\n"
+            ".inst 0x4fbfe2c1  // sdot v1.4s, v22.16b, v31.4b[1]\n"
+            ".inst 0x4f9feac0  // sdot v0.4s, v22.16b, v31.4b[2]\n"
+            ".inst 0x4fbfeade  // sdot v30.4s, v22.16b, v31.4b[3]\n"
+            ".inst 0x4f9be124  // sdot v4.4s, v9.16b, v27.4b[0]\n"
+            ".inst 0x4fbbe121  // sdot v1.4s, v9.16b, v27.4b[1]\n"
+            ".inst 0x4f9be920  // sdot v0.4s, v9.16b, v27.4b[2]\n"
+            ".inst 0x4fbbe93e  // sdot v30.4s, v9.16b, v27.4b[3]\n"
+            ".inst 0x4f9ae0e4  // sdot v4.4s, v7.16b, v26.4b[0]\n"
+            ".inst 0x4fbae0e1  // sdot v1.4s, v7.16b, v26.4b[1]\n"
+            ".inst 0x4f9ae8e0  // sdot v0.4s, v7.16b, v26.4b[2]\n"
+            ".inst 0x4fbae8fe  // sdot v30.4s, v7.16b, v26.4b[3]\n"
+            ".inst 0x4f99e064  // sdot v4.4s, v3.16b, v25.4b[0]\n"
+            ".inst 0x4fb9e061  // sdot v1.4s, v3.16b, v25.4b[1]\n"
+            ".inst 0x4f99e860  // sdot v0.4s, v3.16b, v25.4b[2]\n"
+            ".inst 0x4fb9e87e  // sdot v30.4s, v3.16b, v25.4b[3]\n"
+            ".inst 0x4f98e1a4  // sdot v4.4s, v13.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e1a1  // sdot v1.4s, v13.16b, v24.4b[1]\n"
+            ".inst 0x4f98e9a0  // sdot v0.4s, v13.16b, v24.4b[2]\n"
+            ".inst 0x4fb8e9be  // sdot v30.4s, v13.16b, v24.4b[3]\n"
+            ".inst 0x4f90e384  // sdot v4.4s, v28.16b, v16.4b[0]\n"
+            ".inst 0x4fb0e381  // sdot v1.4s, v28.16b, v16.4b[1]\n"
+            ".inst 0x4f90eb80  // sdot v0.4s, v28.16b, v16.4b[2]\n"
+            ".inst 0x4fb0eb9e  // sdot v30.4s, v28.16b, v16.4b[3]\n"
+            "scvtf v4.4s, v4.4s, #0x4\n"
+            "scvtf v1.4s, v1.4s, #0x4\n"
+            "scvtf v0.4s, v0.4s, #0x4\n"
+            "fmla v15.4s, v4.4s, v11.4s\n"
+            "scvtf v30.4s, v30.4s, #0x4\n"
+            "fmla v19.4s, v1.4s, v23.4s\n"
+            "fmla v18.4s, v0.4s, v17.4s\n"
+            "fmla v14.4s, v30.4s, v6.4s\n"
+            "bgt 7b\n"
+            "mov x20, %x[res_ptr]\n"
+            "cmp x10, #0x1\n"
+            "str q15, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "cmp x10, #0x2\n"
+            "str q19, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "cmp x10, #0x3\n"
+            "str q18, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "str q14, [x20, #0x0]\n"
+            "8:"  // Row tail: Accumulator store skip
+            "subs x23, x23, #0x4\n"
+            "add %x[res_ptr], %x[res_ptr], #0x10\n"
+            "bne 6b\n"
+            "subs x10, x10, #0x4\n"
+            "add %x[a_ptr], %x[a_ptr], x9\n"
+            "mov %x[res_ptr], x22\n"
+            "bgt 5b\n"
+            "9:"  // Row tail: Row loop skip
+            : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
+            : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
+            : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+        );
+        return;
+    }
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
+    {
+        float sumf[4][4];
+        int sumi;
+
+        for (int y = 0; y < nr / 4; y++) {
+            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+                for (int m = 0; m < 4; m++) {
+                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+                }
+                for (int l = 0; l < nb; l++) {
+                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                        for (int m = 0; m < 4; m++) {
+                            for (int j = 0; j < ncols_interleaved; j++) {
+                                sumi = 0;
+                                for (int i = 0; i < blocklen; ++i) {
+                                    const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                    const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                                }
+                                sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            }
+                        }
+                    }
+                }
+                for (int m = 0; m < 4; m++) {
+                    for (int j = 0; j < ncols_interleaved; j++)
+                        s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+                }
+            }
+        }
+    }
+}
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+        const void * b_ptr = vx;
+        const void * a_ptr = vy;
+        float * res_ptr = s;
+        size_t res_stride = bs * sizeof(float);
+
+        __asm__ __volatile__(
+            "mov x10, %x[nr]\n"
+            "mov x9, #0x88\n"
+            "cmp x10, #0x10\n"
+            "mul x9, %x[nb], x9\n"
+            "blt 4f\n"
+            "1:"  // Row loop
+            "add x28, %x[b_ptr], #0x8\n"
+            "mov x27, %x[nc]\n"
+            "add x26, %x[res_ptr], %x[res_stride], LSL #4\n"
+            "2:"  // Column loop
+            "add x25, %x[a_ptr], #0x8\n"
+            "movi v2.16b, #0x0\n"
+            "movi v10.16b, #0x0\n"
+            "mov x24, %x[nb]\n"
+            "add x23, x25, x9\n"
+            "movi v12.16b, #0x0\n"
+            "movi v28.16b, #0x0\n"
+            "add x22, x23, x9\n"
+            "movi v11.16b, #0x0\n"
+            "movi v13.16b, #0x0\n"
+            "add x21, x22, x9\n"
+            "movi v22.16b, #0x0\n"
+            "movi v23.16b, #0x0\n"
+            "movi v25.16b, #0x0\n"
+            "movi v5.16b, #0x0\n"
+            "movi v7.16b, #0x0\n"
+            "movi v4.16b, #0x0\n"
+            "movi v6.16b, #0x0\n"
+            "movi v30.16b, #0x0\n"
+            "movi v24.16b, #0x0\n"
+            "movi v14.16b, #0x0\n"
+            "3:"  // Block loop
+            "ldr q21, [x28, #0x0]\n"
+            "ldr q16, [x28, #0x10]\n"
+            "movi v1.16b, #0x4\n"
+            "movi v19.4s, #0x0\n"
+            "ldr q27, [x25, #0x0]\n"
+            "ldr q15, [x25, #0x10]\n"
+            "movi v26.4s, #0x0\n"
+            "movi v18.4s, #0x0\n"
+            "ldr q29, [x28, #0x20]\n"
+            "ldr q3, [x28, #0x30]\n"
+            "movi v17.4s, #0x0\n"
+            "movi v0.16b, #0xf0\n"
+            "ldr d20, [x25, #-0x8]\n"
+            "ldr d9, [x23, #-0x8]\n"
+            "sshl v8.16b, v21.16b, v1.16b\n"
+            "sshl v31.16b, v16.16b, v1.16b\n"
+            "and v21.16b, v21.16b, v0.16b\n"
+            "and v16.16b, v16.16b, v0.16b\n"
+            "sub x20, x28, #0x8\n"
+            "subs x24, x24, #0x1\n"
+            "add x28, x28, #0x48\n"
+            ".inst 0x4e88a773  // smmla v19.4s, v27.16b, v8.16b\n"
+            ".inst 0x4e9fa77a  // smmla v26.4s, v27.16b, v31.16b\n"
+            "ldr q27, [x25, #0x20]\n"
+            ".inst 0x4e88a5f2  // smmla v18.4s, v15.16b, v8.16b\n"
+            ".inst 0x4e9fa5f1  // smmla v17.4s, v15.16b, v31.16b\n"
+            "sshl v15.16b, v29.16b, v1.16b\n"
+            "sshl v1.16b, v3.16b, v1.16b\n"
+            "and v29.16b, v29.16b, v0.16b\n"
+            "and v3.16b, v3.16b, v0.16b\n"
+            "ldr q0, [x25, #0x30]\n"
+            "fcvtl v20.4s, v20.4h\n"
+            ".inst 0x4e8fa773  // smmla v19.4s, v27.16b, v15.16b\n"
+            "fcvtl v9.4s, v9.4h\n"
+            ".inst 0x4e81a77a  // smmla v26.4s, v27.16b, v1.16b\n"
+            "ldr q27, [x25, #0x40]\n"
+            ".inst 0x4e8fa412  // smmla v18.4s, v0.16b, v15.16b\n"
+            ".inst 0x4e81a411  // smmla v17.4s, v0.16b, v1.16b\n"
+            "ldr q0, [x25, #0x50]\n"
+            ".inst 0x4e95a773  // smmla v19.4s, v27.16b, v21.16b\n"
+            ".inst 0x4e90a77a  // smmla v26.4s, v27.16b, v16.16b\n"
+            "ldr q27, [x25, #0x60]\n"
+            ".inst 0x4e95a412  // smmla v18.4s, v0.16b, v21.16b\n"
+            ".inst 0x4e90a411  // smmla v17.4s, v0.16b, v16.16b\n"
+            "ldr q0, [x25, #0x70]\n"
+            "add x25, x25, #0x88\n"
+            ".inst 0x4e9da773  // smmla v19.4s, v27.16b, v29.16b\n"
+            ".inst 0x4e83a77a  // smmla v26.4s, v27.16b, v3.16b\n"
+            "ldr d27, [x20, #0x0]\n"
+            ".inst 0x4e9da412  // smmla v18.4s, v0.16b, v29.16b\n"
+            ".inst 0x4e83a411  // smmla v17.4s, v0.16b, v3.16b\n"
+            "fcvtl v27.4s, v27.4h\n"
+            "uzp1 v0.2d, v19.2d, v26.2d\n"
+            "uzp2 v26.2d, v19.2d, v26.2d\n"
+            "fmul v19.4s, v27.4s, v20.s[0]\n"
+            "scvtf v0.4s, v0.4s, #0x4\n"
+            "scvtf v26.4s, v26.4s, #0x4\n"
+            "fmla v2.4s, v0.4s, v19.4s\n"
+            "ldr q19, [x23, #0x0]\n"
+            "uzp1 v0.2d, v18.2d, v17.2d\n"
+            "uzp2 v18.2d, v18.2d, v17.2d\n"
+            "fmul v17.4s, v27.4s, v20.s[1]\n"
+            "scvtf v0.4s, v0.4s, #0x4\n"
+            "scvtf v18.4s, v18.4s, #0x4\n"
+            "fmla v10.4s, v26.4s, v17.4s\n"
+            "ldr q17, [x23, #0x10]\n"
+            "fmul v26.4s, v27.4s, v20.s[2]\n"
+            "fmul v20.4s, v27.4s, v20.s[3]\n"
+            "fmla v12.4s, v0.4s, v26.4s\n"
+            "ldr d0, [x22, #-0x8]\n"
+            "ldr d26, [x21, #-0x8]\n"
+            "fcvtl v0.4s, v0.4h\n"
+            "fmla v28.4s, v18.4s, v20.4s\n"
+            "movi v20.4s, #0x0\n"
+            "movi v18.4s, #0x0\n"
+            ".inst 0x4e88a674  // smmla v20.4s, v19.16b, v8.16b\n"
+            ".inst 0x4e9fa672  // smmla v18.4s, v19.16b, v31.16b\n"
+            "ldr q19, [x23, #0x20]\n"
+            "fcvtl v26.4s, v26.4h\n"
+            ".inst 0x4e8fa674  // smmla v20.4s, v19.16b, v15.16b\n"
+            ".inst 0x4e81a672  // smmla v18.4s, v19.16b, v1.16b\n"
+            "ldr q19, [x23, #0x40]\n"
+            ".inst 0x4e95a674  // smmla v20.4s, v19.16b, v21.16b\n"
+            ".inst 0x4e90a672  // smmla v18.4s, v19.16b, v16.16b\n"
+            "ldr q19, [x23, #0x60]\n"
+            ".inst 0x4e9da674  // smmla v20.4s, v19.16b, v29.16b\n"
+            ".inst 0x4e83a672  // smmla v18.4s, v19.16b, v3.16b\n"
+            "uzp1 v19.2d, v20.2d, v18.2d\n"
+            "scvtf v19.4s, v19.4s, #0x4\n"
+            "uzp2 v20.2d, v20.2d, v18.2d\n"
+            "fmul v18.4s, v27.4s, v9.s[0]\n"
+            "scvtf v20.4s, v20.4s, #0x4\n"
+            "fmla v11.4s, v19.4s, v18.4s\n"
+            "ldr q18, [x22, #0x0]\n"
+            "fmul v19.4s, v27.4s, v9.s[1]\n"
+            "fmla v13.4s, v20.4s, v19.4s\n"
+            "movi v19.4s, #0x0\n"
+            "movi v20.4s, #0x0\n"
+            ".inst 0x4e88a633  // smmla v19.4s, v17.16b, v8.16b\n"
+            ".inst 0x4e9fa634  // smmla v20.4s, v17.16b, v31.16b\n"
+            "ldr q17, [x23, #0x30]\n"
+            ".inst 0x4e8fa633  // smmla v19.4s, v17.16b, v15.16b\n"
+            ".inst 0x4e81a634  // smmla v20.4s, v17.16b, v1.16b\n"
+            "ldr q17, [x23, #0x50]\n"
+            ".inst 0x4e95a633  // smmla v19.4s, v17.16b, v21.16b\n"
+            ".inst 0x4e90a634  // smmla v20.4s, v17.16b, v16.16b\n"
+            "ldr q17, [x23, #0x70]\n"
+            "add x23, x23, #0x88\n"
+            ".inst 0x4e9da633  // smmla v19.4s, v17.16b, v29.16b\n"
+            ".inst 0x4e83a634  // smmla v20.4s, v17.16b, v3.16b\n"
+            "uzp1 v17.2d, v19.2d, v20.2d\n"
+            "scvtf v17.4s, v17.4s, #0x4\n"
+            "uzp2 v20.2d, v19.2d, v20.2d\n"
+            "fmul v19.4s, v27.4s, v9.s[2]\n"
+            "fmul v9.4s, v27.4s, v9.s[3]\n"
+            "scvtf v20.4s, v20.4s, #0x4\n"
+            "fmla v22.4s, v17.4s, v19.4s\n"
+            "ldr q17, [x22, #0x10]\n"
+            "movi v19.4s, #0x0\n"
+            ".inst 0x4e88a653  // smmla v19.4s, v18.16b, v8.16b\n"
+            "fmla v23.4s, v20.4s, v9.4s\n"
+            "movi v20.4s, #0x0\n"
+            "movi v9.4s, #0x0\n"
+            ".inst 0x4e9fa654  // smmla v20.4s, v18.16b, v31.16b\n"
+            "ldr q18, [x22, #0x20]\n"
+            ".inst 0x4e88a629  // smmla v9.4s, v17.16b, v8.16b\n"
+            ".inst 0x4e8fa653  // smmla v19.4s, v18.16b, v15.16b\n"
+            ".inst 0x4e81a654  // smmla v20.4s, v18.16b, v1.16b\n"
+            "ldr q18, [x22, #0x40]\n"
+            ".inst 0x4e95a653  // smmla v19.4s, v18.16b, v21.16b\n"
+            ".inst 0x4e90a654  // smmla v20.4s, v18.16b, v16.16b\n"
+            "ldr q18, [x22, #0x60]\n"
+            ".inst 0x4e9da653  // smmla v19.4s, v18.16b, v29.16b\n"
+            ".inst 0x4e83a654  // smmla v20.4s, v18.16b, v3.16b\n"
+            "movi v18.4s, #0x0\n"
+            ".inst 0x4e9fa632  // smmla v18.4s, v17.16b, v31.16b\n"
+            "ldr q17, [x22, #0x30]\n"
+            ".inst 0x4e8fa629  // smmla v9.4s, v17.16b, v15.16b\n"
+            ".inst 0x4e81a632  // smmla v18.4s, v17.16b, v1.16b\n"
+            "ldr q17, [x22, #0x50]\n"
+            ".inst 0x4e95a629  // smmla v9.4s, v17.16b, v21.16b\n"
+            ".inst 0x4e90a632  // smmla v18.4s, v17.16b, v16.16b\n"
+            "ldr q17, [x22, #0x70]\n"
+            "add x22, x22, #0x88\n"
+            ".inst 0x4e9da629  // smmla v9.4s, v17.16b, v29.16b\n"
+            ".inst 0x4e83a632  // smmla v18.4s, v17.16b, v3.16b\n"
+            "uzp1 v17.2d, v19.2d, v20.2d\n"
+            "uzp2 v20.2d, v19.2d, v20.2d\n"
+            "fmul v19.4s, v27.4s, v0.s[0]\n"
+            "scvtf v17.4s, v17.4s, #0x4\n"
+            "scvtf v20.4s, v20.4s, #0x4\n"
+            "fmla v25.4s, v17.4s, v19.4s\n"
+            "ldr q19, [x21, #0x0]\n"
+            "fmul v17.4s, v27.4s, v0.s[1]\n"
+            "fmla v5.4s, v20.4s, v17.4s\n"
+            "ldr q17, [x21, #0x10]\n"
+            "uzp1 v20.2d, v9.2d, v18.2d\n"
+            "uzp2 v9.2d, v9.2d, v18.2d\n"
+            "fmul v18.4s, v27.4s, v0.s[2]\n"
+            "fmul v0.4s, v27.4s, v0.s[3]\n"
+            "scvtf v20.4s, v20.4s, #0x4\n"
+            "scvtf v9.4s, v9.4s, #0x4\n"
+            "fmla v7.4s, v20.4s, v18.4s\n"
+            "movi v20.4s, #0x0\n"
+            "movi v18.4s, #0x0\n"
+            ".inst 0x4e88a674  // smmla v20.4s, v19.16b, v8.16b\n"
+            ".inst 0x4e9fa672  // smmla v18.4s, v19.16b, v31.16b\n"
+            "ldr q19, [x21, #0x20]\n"
+            "fmla v4.4s, v9.4s, v0.4s\n"
+            "movi v9.4s, #0x0\n"
+            "movi v0.4s, #0x0\n"
+            ".inst 0x4e88a629  // smmla v9.4s, v17.16b, v8.16b\n"
+            "fmul v8.4s, v27.4s, v26.s[0]\n"
+            ".inst 0x4e9fa620  // smmla v0.4s, v17.16b, v31.16b\n"
+            "ldr q17, [x21, #0x30]\n"
+            ".inst 0x4e8fa674  // smmla v20.4s, v19.16b, v15.16b\n"
+            "fmul v31.4s, v27.4s, v26.s[1]\n"
+            ".inst 0x4e81a672  // smmla v18.4s, v19.16b, v1.16b\n"
+            "ldr q19, [x21, #0x40]\n"
+            ".inst 0x4e8fa629  // smmla v9.4s, v17.16b, v15.16b\n"
+            "fmul v15.4s, v27.4s, v26.s[2]\n"
+            "fmul v27.4s, v27.4s, v26.s[3]\n"
+            ".inst 0x4e81a620  // smmla v0.4s, v17.16b, v1.16b\n"
+            "ldr q1, [x21, #0x50]\n"
+            ".inst 0x4e95a674  // smmla v20.4s, v19.16b, v21.16b\n"
+            ".inst 0x4e90a672  // smmla v18.4s, v19.16b, v16.16b\n"
+            "ldr q26, [x21, #0x60]\n"
+            ".inst 0x4e95a429  // smmla v9.4s, v1.16b, v21.16b\n"
+            ".inst 0x4e90a420  // smmla v0.4s, v1.16b, v16.16b\n"
+            "ldr q21, [x21, #0x70]\n"
+            "add x21, x21, #0x88\n"
+            ".inst 0x4e9da754  // smmla v20.4s, v26.16b, v29.16b\n"
+            ".inst 0x4e83a752  // smmla v18.4s, v26.16b, v3.16b\n"
+            ".inst 0x4e9da6a9  // smmla v9.4s, v21.16b, v29.16b\n"
+            ".inst 0x4e83a6a0  // smmla v0.4s, v21.16b, v3.16b\n"
+            "uzp1 v29.2d, v20.2d, v18.2d\n"
+            "uzp2 v21.2d, v20.2d, v18.2d\n"
+            "scvtf v29.4s, v29.4s, #0x4\n"
+            "uzp1 v18.2d, v9.2d, v0.2d\n"
+            "uzp2 v16.2d, v9.2d, v0.2d\n"
+            "scvtf v21.4s, v21.4s, #0x4\n"
+            "fmla v6.4s, v29.4s, v8.4s\n"
+            "scvtf v18.4s, v18.4s, #0x4\n"
+            "scvtf v16.4s, v16.4s, #0x4\n"
+            "fmla v30.4s, v21.4s, v31.4s\n"
+            "fmla v24.4s, v18.4s, v15.4s\n"
+            "fmla v14.4s, v16.4s, v27.4s\n"
+            "bgt 3b\n"
+            "mov x20, %x[res_ptr]\n"
+            "subs x27, x27, #0x4\n"
+            "add %x[res_ptr], %x[res_ptr], #0x10\n"
+            "str q2, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q10, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q12, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q28, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q11, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q13, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q22, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q23, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q25, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q5, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q7, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q4, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q6, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q30, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q24, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q14, [x20, #0x0]\n"
+            "bne 2b\n"
+            "mov x20, #0x4\n"
+            "sub x10, x10, #0x10\n"
+            "cmp x10, #0x10\n"
+            "mov %x[res_ptr], x26\n"
+            "madd %x[a_ptr], x20, x9, %x[a_ptr]\n"
+            "bge 1b\n"
+            "4:"  // Row loop skip
+            "cbz x10, 9f\n"
+            "5:"  // Row tail: Row loop
+            "add x24, %x[b_ptr], #0x8\n"
+            "mov x23, %x[nc]\n"
+            "add x22, %x[res_ptr], %x[res_stride], LSL #2\n"
+            "6:"  // Row tail: Column loop
+            "movi v2.16b, #0x0\n"
+            "movi v10.16b, #0x0\n"
+            "add x25, %x[a_ptr], #0x8\n"
+            "mov x21, %x[nb]\n"
+            "movi v12.16b, #0x0\n"
+            "movi v28.16b, #0x0\n"
+            "7:"  // Row tail: Block loop
+            "ldr q6, [x24, #0x0]\n"
+            "ldr q5, [x24, #0x10]\n"
+            "movi v17.16b, #0x4\n"
+            "movi v8.4s, #0x0\n"
+            "ldr q4, [x25, #0x0]\n"
+            "ldr q13, [x25, #0x10]\n"
+            "movi v27.4s, #0x0\n"
+            "movi v0.4s, #0x0\n"
+            "ldr q31, [x24, #0x20]\n"
+            "ldr q14, [x24, #0x30]\n"
+            "movi v29.4s, #0x0\n"
+            "movi v22.16b, #0xf0\n"
+            "ldr q11, [x25, #0x20]\n"
+            "ldr q23, [x25, #0x30]\n"
+            "sshl v21.16b, v6.16b, v17.16b\n"
+            "sshl v16.16b, v5.16b, v17.16b\n"
+            "ldr q20, [x25, #0x40]\n"
+            "ldr q26, [x25, #0x50]\n"
+            "and v6.16b, v6.16b, v22.16b\n"
+            "and v5.16b, v5.16b, v22.16b\n"
+            "ldr q25, [x25, #0x60]\n"
+            "ldr q3, [x25, #0x70]\n"
+            "sshl v19.16b, v31.16b, v17.16b\n"
+            "sshl v18.16b, v14.16b, v17.16b\n"
+            "ldr d17, [x25, #-0x8]\n"
+            ".inst 0x4e95a488  // smmla v8.4s, v4.16b, v21.16b\n"
+            ".inst 0x4e90a49b  // smmla v27.4s, v4.16b, v16.16b\n"
+            "and v31.16b, v31.16b, v22.16b\n"
+            ".inst 0x4e95a5a0  // smmla v0.4s, v13.16b, v21.16b\n"
+            ".inst 0x4e90a5bd  // smmla v29.4s, v13.16b, v16.16b\n"
+            "and v14.16b, v14.16b, v22.16b\n"
+            "sub x20, x24, #0x8\n"
+            "ldr d16, [x20, #0x0]\n"
+            "subs x21, x21, #0x1\n"
+            "add x25, x25, #0x88\n"
+            "fcvtl v17.4s, v17.4h\n"
+            "add x24, x24, #0x48\n"
+            ".inst 0x4e93a568  // smmla v8.4s, v11.16b, v19.16b\n"
+            ".inst 0x4e92a57b  // smmla v27.4s, v11.16b, v18.16b\n"
+            ".inst 0x4e93a6e0  // smmla v0.4s, v23.16b, v19.16b\n"
+            ".inst 0x4e92a6fd  // smmla v29.4s, v23.16b, v18.16b\n"
+            "fcvtl v16.4s, v16.4h\n"
+            ".inst 0x4e86a688  // smmla v8.4s, v20.16b, v6.16b\n"
+            ".inst 0x4e85a69b  // smmla v27.4s, v20.16b, v5.16b\n"
+            "fmul v23.4s, v16.4s, v17.s[0]\n"
+            "fmul v21.4s, v16.4s, v17.s[1]\n"
+            "fmul v1.4s, v16.4s, v17.s[2]\n"
+            "fmul v20.4s, v16.4s, v17.s[3]\n"
+            ".inst 0x4e86a740  // smmla v0.4s, v26.16b, v6.16b\n"
+            ".inst 0x4e85a75d  // smmla v29.4s, v26.16b, v5.16b\n"
+            ".inst 0x4e9fa728  // smmla v8.4s, v25.16b, v31.16b\n"
+            ".inst 0x4e8ea73b  // smmla v27.4s, v25.16b, v14.16b\n"
+            ".inst 0x4e9fa460  // smmla v0.4s, v3.16b, v31.16b\n"
+            ".inst 0x4e8ea47d  // smmla v29.4s, v3.16b, v14.16b\n"
+            "uzp1 v19.2d, v8.2d, v27.2d\n"
+            "uzp2 v18.2d, v8.2d, v27.2d\n"
+            "scvtf v19.4s, v19.4s, #0x4\n"
+            "uzp1 v17.2d, v0.2d, v29.2d\n"
+            "uzp2 v16.2d, v0.2d, v29.2d\n"
+            "scvtf v18.4s, v18.4s, #0x4\n"
+            "fmla v2.4s, v19.4s, v23.4s\n"
+            "scvtf v17.4s, v17.4s, #0x4\n"
+            "scvtf v16.4s, v16.4s, #0x4\n"
+            "fmla v10.4s, v18.4s, v21.4s\n"
+            "fmla v12.4s, v17.4s, v1.4s\n"
+            "fmla v28.4s, v16.4s, v20.4s\n"
+            "bgt 7b\n"
+            "mov x20, %x[res_ptr]\n"
+            "cmp x10, #0x1\n"
+            "str q2, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "cmp x10, #0x2\n"
+            "str q10, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "cmp x10, #0x3\n"
+            "str q12, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "str q28, [x20, #0x0]\n"
+            "8:"  // Row tail: Accumulator store skip
+            "subs x23, x23, #0x4\n"
+            "add %x[res_ptr], %x[res_ptr], #0x10\n"
+            "bne 6b\n"
+            "subs x10, x10, #0x4\n"
+            "add %x[a_ptr], %x[a_ptr], x9\n"
+            "mov %x[res_ptr], x22\n"
+            "bgt 5b\n"
+            "9:"  // Row tail: Row loop skip
+            : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
+            : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
+            : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+        );
+        return;
+    }
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    float sumf[4][4];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                        (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+}
+
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
+#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
+    if (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0) {
+        const void * b_ptr = vx;
+        const void * a_ptr = vy;
+        float * res_ptr = s;
+        size_t res_stride = bs * sizeof(float);
+
+        __asm__ __volatile__(
+            "mov x20, #0x4\n"
+            "mov x13, %x[nr]\n"
+            "mov z28.s, #-0x4\n"
+            "mov x12, #0x88\n"
+            "ptrue p1.b\n"
+            "whilelt p0.s, XZR, x20\n"
+            "cmp x13, #0x10\n"
+            "mul x12, %x[nb], x12\n"
+            "blt 4f\n"
+            "1:"  // Row loop
+            "add x11, %x[b_ptr], #0x10\n"
+            "mov x10, %x[nc]\n"
+            "add x9, %x[res_ptr], %x[res_stride], LSL #4\n"
+            "2:"  // Column loop
+            "add x28, %x[a_ptr], #0x8\n"
+            "mov z24.b, #0x0\n"
+            "mov z15.b, #0x0\n"
+            "mov x27, %x[nb]\n"
+            "add x26, x28, x12\n"
+            "mov z12.b, #0x0\n"
+            "mov z0.b, #0x0\n"
+            "add x25, x26, x12\n"
+            "mov z13.b, #0x0\n"
+            "mov z1.b, #0x0\n"
+            "add x24, x25, x12\n"
+            "mov z20.b, #0x0\n"
+            "mov z25.b, #0x0\n"
+            "mov z11.b, #0x0\n"
+            "mov z16.b, #0x0\n"
+            "mov z19.b, #0x0\n"
+            "mov z26.b, #0x0\n"
+            "mov z8.b, #0x0\n"
+            "mov z29.b, #0x0\n"
+            "mov z27.b, #0x0\n"
+            "mov z10.b, #0x0\n"
+            "3:"  // Block loop
+            "ld1b { z30.b }, p1/Z, [x11]\n"
+            "ld1b { z21.b }, p1/Z, [x11, #1, MUL VL]\n"
+            "mov z18.s, #0x0\n"
+            "mov z7.s, #0x0\n"
+            "ld1rqb { z3.b }, p1/Z, [x28]\n"
+            "ld1rqb { z5.b }, p1/Z, [x28, #16]\n"
+            "mov z9.s, #0x0\n"
+            "mov z22.s, #0x0\n"
+            "ld1b { z4.b }, p1/Z, [x11, #2, MUL VL]\n"
+            "ld1b { z17.b }, p1/Z, [x11, #3, MUL VL]\n"
+            "sub x20, x11, #0x10\n"
+            "sub x23, x28, #0x8\n"
+            "lsl z31.b, z30.b, #0x4\n"
+            "lsl z6.b, z21.b, #0x4\n"
+            "ld1h { z23.s }, p1/Z, [x20]\n"
+            "sub x22, x26, #0x8\n"
+            "and z30.b, z30.b, #0xf0\n"
+            "and z21.b, z21.b, #0xf0\n"
+            "sub x21, x25, #0x8\n"
+            "sub x20, x24, #0x8\n"
+            "lsl z14.b, z4.b, #0x4\n"
+            "lsl z2.b, z17.b, #0x4\n"
+            "subs x27, x27, #0x1\n"
+            "add x11, x11, #0x90\n"
+            ".inst 0x451f9872  // smmla z18.s, z3.b, z31.b\n"
+            ".inst 0x45069867  // smmla z7.s, z3.b, z6.b\n"
+            "ld1rqb { z3.b }, p1/Z, [x28, #32]\n"
+            "and z4.b, z4.b, #0xf0\n"
+            ".inst 0x451f98a9  // smmla z9.s, z5.b, z31.b\n"
+            ".inst 0x450698b6  // smmla z22.s, z5.b, z6.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x28, #48]\n"
+            "and z17.b, z17.b, #0xf0\n"
+            "fcvt z23.s, p1/m, z23.h\n"
+            ".inst 0x450e9872  // smmla z18.s, z3.b, z14.b\n"
+            ".inst 0x45029867  // smmla z7.s, z3.b, z2.b\n"
+            "ld1rqb { z3.b }, p1/Z, [x28, #64]\n"
+            ".inst 0x450e98a9  // smmla z9.s, z5.b, z14.b\n"
+            ".inst 0x450298b6  // smmla z22.s, z5.b, z2.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x28, #80]\n"
+            "fscale z23.s, p1/m, z23.s, z28.s\n"
+            ".inst 0x451e9872  // smmla z18.s, z3.b, z30.b\n"
+            ".inst 0x45159867  // smmla z7.s, z3.b, z21.b\n"
+            "ld1rqb { z3.b }, p1/Z, [x28, #96]\n"
+            ".inst 0x451e98a9  // smmla z9.s, z5.b, z30.b\n"
+            ".inst 0x451598b6  // smmla z22.s, z5.b, z21.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x28, #112]\n"
+            "add x28, x28, #0x88\n"
+            ".inst 0x45049872  // smmla z18.s, z3.b, z4.b\n"
+            ".inst 0x45119867  // smmla z7.s, z3.b, z17.b\n"
+            "ld1h { z3.s }, p0/Z, [x23]\n"
+            ".inst 0x450498a9  // smmla z9.s, z5.b, z4.b\n"
+            ".inst 0x451198b6  // smmla z22.s, z5.b, z17.b\n"
+            "fcvt z3.s, p1/m, z3.h\n"
+            "uzp1 z5.d, z18.d, z7.d\n"
+            "uzp2 z18.d, z18.d, z7.d\n"
+            "mov z3.q, z3.q[0]\n"
+            "uzp1 z7.d, z9.d, z22.d\n"
+            "uzp2 z22.d, z9.d, z22.d\n"
+            "fmul z9.s, z23.s, z3.s[0]\n"
+            "scvtf z5.s, p1/m, z5.s\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "scvtf z7.s, p1/m, z7.s\n"
+            "scvtf z22.s, p1/m, z22.s\n"
+            "fmla z24.s, p1/M, z5.s, z9.s\n"
+            "ld1rqb { z5.b }, p1/Z, [x26]\n"
+            "fmul z9.s, z23.s, z3.s[1]\n"
+            "fmla z15.s, p1/M, z18.s, z9.s\n"
+            "ld1rqb { z18.b }, p1/Z, [x26, #16]\n"
+            "fmul z9.s, z23.s, z3.s[2]\n"
+            "fmul z3.s, z23.s, z3.s[3]\n"
+            "fmla z12.s, p1/M, z7.s, z9.s\n"
+            "mov z9.s, #0x0\n"
+            "ld1h { z7.s }, p0/Z, [x22]\n"
+            ".inst 0x451f98a9  // smmla z9.s, z5.b, z31.b\n"
+            "fmla z0.s, p1/M, z22.s, z3.s\n"
+            "mov z22.s, #0x0\n"
+            "ld1h { z3.s }, p0/Z, [x21]\n"
+            ".inst 0x450698b6  // smmla z22.s, z5.b, z6.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x26, #32]\n"
+            "fcvt z7.s, p1/m, z7.h\n"
+            "fcvt z3.s, p1/m, z3.h\n"
+            ".inst 0x450e98a9  // smmla z9.s, z5.b, z14.b\n"
+            ".inst 0x450298b6  // smmla z22.s, z5.b, z2.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x26, #64]\n"
+            "mov z7.q, z7.q[0]\n"
+            "mov z3.q, z3.q[0]\n"
+            ".inst 0x451e98a9  // smmla z9.s, z5.b, z30.b\n"
+            ".inst 0x451598b6  // smmla z22.s, z5.b, z21.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x26, #96]\n"
+            ".inst 0x450498a9  // smmla z9.s, z5.b, z4.b\n"
+            ".inst 0x451198b6  // smmla z22.s, z5.b, z17.b\n"
+            "uzp1 z5.d, z9.d, z22.d\n"
+            "scvtf z5.s, p1/m, z5.s\n"
+            "uzp2 z22.d, z9.d, z22.d\n"
+            "fmul z9.s, z23.s, z7.s[0]\n"
+            "scvtf z22.s, p1/m, z22.s\n"
+            "fmla z13.s, p1/M, z5.s, z9.s\n"
+            "ld1rqb { z9.b }, p1/Z, [x25]\n"
+            "fmul z5.s, z23.s, z7.s[1]\n"
+            "fmla z1.s, p1/M, z22.s, z5.s\n"
+            "mov z5.s, #0x0\n"
+            "mov z22.s, #0x0\n"
+            ".inst 0x451f9a45  // smmla z5.s, z18.b, z31.b\n"
+            ".inst 0x45069a56  // smmla z22.s, z18.b, z6.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x26, #48]\n"
+            ".inst 0x450e9a45  // smmla z5.s, z18.b, z14.b\n"
+            ".inst 0x45029a56  // smmla z22.s, z18.b, z2.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x26, #80]\n"
+            ".inst 0x451e9a45  // smmla z5.s, z18.b, z30.b\n"
+            ".inst 0x45159a56  // smmla z22.s, z18.b, z21.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x26, #112]\n"
+            "add x26, x26, #0x88\n"
+            ".inst 0x45049a45  // smmla z5.s, z18.b, z4.b\n"
+            ".inst 0x45119a56  // smmla z22.s, z18.b, z17.b\n"
+            "uzp1 z18.d, z5.d, z22.d\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "uzp2 z22.d, z5.d, z22.d\n"
+            "fmul z5.s, z23.s, z7.s[2]\n"
+            "fmul z7.s, z23.s, z7.s[3]\n"
+            "scvtf z22.s, p1/m, z22.s\n"
+            "fmla z20.s, p1/M, z18.s, z5.s\n"
+            "ld1rqb { z18.b }, p1/Z, [x25, #16]\n"
+            "ld1h { z5.s }, p0/Z, [x20]\n"
+            "fcvt z5.s, p1/m, z5.h\n"
+            "fmla z25.s, p1/M, z22.s, z7.s\n"
+            "mov z22.s, #0x0\n"
+            "mov z7.s, #0x0\n"
+            ".inst 0x451f9936  // smmla z22.s, z9.b, z31.b\n"
+            ".inst 0x45069927  // smmla z7.s, z9.b, z6.b\n"
+            "ld1rqb { z9.b }, p1/Z, [x25, #32]\n"
+            "mov z5.q, z5.q[0]\n"
+            ".inst 0x450e9936  // smmla z22.s, z9.b, z14.b\n"
+            ".inst 0x45029927  // smmla z7.s, z9.b, z2.b\n"
+            "ld1rqb { z9.b }, p1/Z, [x25, #64]\n"
+            ".inst 0x451e9936  // smmla z22.s, z9.b, z30.b\n"
+            ".inst 0x45159927  // smmla z7.s, z9.b, z21.b\n"
+            "ld1rqb { z9.b }, p1/Z, [x25, #96]\n"
+            ".inst 0x45049936  // smmla z22.s, z9.b, z4.b\n"
+            ".inst 0x45119927  // smmla z7.s, z9.b, z17.b\n"
+            "uzp1 z9.d, z22.d, z7.d\n"
+            "scvtf z9.s, p1/m, z9.s\n"
+            "uzp2 z22.d, z22.d, z7.d\n"
+            "fmul z7.s, z23.s, z3.s[0]\n"
+            "scvtf z22.s, p1/m, z22.s\n"
+            "fmla z11.s, p1/M, z9.s, z7.s\n"
+            "ld1rqb { z9.b }, p1/Z, [x24]\n"
+            "fmul z7.s, z23.s, z3.s[1]\n"
+            "fmla z16.s, p1/M, z22.s, z7.s\n"
+            "mov z22.s, #0x0\n"
+            "mov z7.s, #0x0\n"
+            ".inst 0x451f9a56  // smmla z22.s, z18.b, z31.b\n"
+            ".inst 0x45069a47  // smmla z7.s, z18.b, z6.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x25, #48]\n"
+            ".inst 0x450e9a56  // smmla z22.s, z18.b, z14.b\n"
+            ".inst 0x45029a47  // smmla z7.s, z18.b, z2.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x25, #80]\n"
+            ".inst 0x451e9a56  // smmla z22.s, z18.b, z30.b\n"
+            ".inst 0x45159a47  // smmla z7.s, z18.b, z21.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x25, #112]\n"
+            "add x25, x25, #0x88\n"
+            ".inst 0x45049a56  // smmla z22.s, z18.b, z4.b\n"
+            ".inst 0x45119a47  // smmla z7.s, z18.b, z17.b\n"
+            "uzp1 z18.d, z22.d, z7.d\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "uzp2 z7.d, z22.d, z7.d\n"
+            "fmul z22.s, z23.s, z3.s[2]\n"
+            "fmul z3.s, z23.s, z3.s[3]\n"
+            "scvtf z7.s, p1/m, z7.s\n"
+            "fmla z19.s, p1/M, z18.s, z22.s\n"
+            "ld1rqb { z18.b }, p1/Z, [x24, #16]\n"
+            "fmul z22.s, z23.s, z5.s[0]\n"
+            "fmla z26.s, p1/M, z7.s, z3.s\n"
+            "mov z3.s, #0x0\n"
+            "mov z7.s, #0x0\n"
+            ".inst 0x451f9923  // smmla z3.s, z9.b, z31.b\n"
+            ".inst 0x45069927  // smmla z7.s, z9.b, z6.b\n"
+            "ld1rqb { z9.b }, p1/Z, [x24, #32]\n"
+            ".inst 0x450e9923  // smmla z3.s, z9.b, z14.b\n"
+            ".inst 0x45029927  // smmla z7.s, z9.b, z2.b\n"
+            "mov z9.s, #0x0\n"
+            ".inst 0x451f9a49  // smmla z9.s, z18.b, z31.b\n"
+            "mov z31.s, #0x0\n"
+            ".inst 0x45069a5f  // smmla z31.s, z18.b, z6.b\n"
+            "ld1rqb { z6.b }, p1/Z, [x24, #48]\n"
+            "ld1rqb { z18.b }, p1/Z, [x24, #64]\n"
+            ".inst 0x450e98c9  // smmla z9.s, z6.b, z14.b\n"
+            "fmul z14.s, z23.s, z5.s[1]\n"
+            ".inst 0x450298df  // smmla z31.s, z6.b, z2.b\n"
+            "ld1rqb { z6.b }, p1/Z, [x24, #80]\n"
+            "fmul z2.s, z23.s, z5.s[2]\n"
+            "fmul z23.s, z23.s, z5.s[3]\n"
+            ".inst 0x451e9a43  // smmla z3.s, z18.b, z30.b\n"
+            ".inst 0x45159a47  // smmla z7.s, z18.b, z21.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x24, #96]\n"
+            ".inst 0x451e98c9  // smmla z9.s, z6.b, z30.b\n"
+            ".inst 0x451598df  // smmla z31.s, z6.b, z21.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x24, #112]\n"
+            "add x24, x24, #0x88\n"
+            ".inst 0x450498a3  // smmla z3.s, z5.b, z4.b\n"
+            ".inst 0x451198a7  // smmla z7.s, z5.b, z17.b\n"
+            ".inst 0x45049a49  // smmla z9.s, z18.b, z4.b\n"
+            ".inst 0x45119a5f  // smmla z31.s, z18.b, z17.b\n"
+            "uzp1 z18.d, z3.d, z7.d\n"
+            "uzp2 z5.d, z3.d, z7.d\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "uzp1 z6.d, z9.d, z31.d\n"
+            "uzp2 z9.d, z9.d, z31.d\n"
+            "scvtf z5.s, p1/m, z5.s\n"
+            "fmla z8.s, p1/M, z18.s, z22.s\n"
+            "scvtf z6.s, p1/m, z6.s\n"
+            "scvtf z9.s, p1/m, z9.s\n"
+            "fmla z29.s, p1/M, z5.s, z14.s\n"
+            "fmla z27.s, p1/M, z6.s, z2.s\n"
+            "fmla z10.s, p1/M, z9.s, z23.s\n"
+            "bgt 3b\n"
+            "mov x20, %x[res_ptr]\n"
+            "subs x10, x10, #0x8\n"
+            "add %x[res_ptr], %x[res_ptr], #0x20\n"
+            "st1w { z24.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z15.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z12.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z0.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z13.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z1.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z20.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z25.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z11.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z16.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z19.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z26.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z8.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z29.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z27.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z10.s }, p1, [x20]\n"
+            "bne 2b\n"
+            "mov x20, #0x4\n"
+            "sub x13, x13, #0x10\n"
+            "cmp x13, #0x10\n"
+            "mov %x[res_ptr], x9\n"
+            "madd %x[a_ptr], x20, x12, %x[a_ptr]\n"
+            "bge 1b\n"
+            "4:"  // Row loop skip
+            "cbz x13, 9f\n"
+            "5:"  // Row tail: Row loop
+            "add x25, %x[b_ptr], #0x10\n"
+            "mov x24, %x[nc]\n"
+            "add x23, %x[res_ptr], %x[res_stride], LSL #2\n"
+            "6:"  // Row tail: Column loop
+            "mov z24.b, #0x0\n"
+            "mov z15.b, #0x0\n"
+            "add x28, %x[a_ptr], #0x8\n"
+            "mov x22, %x[nb]\n"
+            "mov z12.b, #0x0\n"
+            "mov z0.b, #0x0\n"
+            "7:"  // Row tail: Block loop
+            "ld1b { z3.b }, p1/Z, [x25]\n"
+            "ld1b { z6.b }, p1/Z, [x25, #1, MUL VL]\n"
+            "mov z2.s, #0x0\n"
+            "mov z25.s, #0x0\n"
+            "ld1rqb { z26.b }, p1/Z, [x28]\n"
+            "ld1rqb { z21.b }, p1/Z, [x28, #16]\n"
+            "mov z27.s, #0x0\n"
+            "mov z19.s, #0x0\n"
+            "ld1b { z29.b }, p1/Z, [x25, #2, MUL VL]\n"
+            "ld1b { z16.b }, p1/Z, [x25, #3, MUL VL]\n"
+            "sub x21, x25, #0x10\n"
+            "sub x20, x28, #0x8\n"
+            "lsl z20.b, z3.b, #0x4\n"
+            "lsl z4.b, z6.b, #0x4\n"
+            "ld1rqb { z10.b }, p1/Z, [x28, #32]\n"
+            "ld1rqb { z23.b }, p1/Z, [x28, #48]\n"
+            "and z3.b, z3.b, #0xf0\n"
+            "and z6.b, z6.b, #0xf0\n"
+            "ld1rqb { z11.b }, p1/Z, [x28, #64]\n"
+            "ld1rqb { z7.b }, p1/Z, [x28, #80]\n"
+            "lsl z8.b, z29.b, #0x4\n"
+            "lsl z14.b, z16.b, #0x4\n"
+            "ld1rqb { z18.b }, p1/Z, [x28, #96]\n"
+            "ld1rqb { z30.b }, p1/Z, [x28, #112]\n"
+            ".inst 0x45149b42  // smmla z2.s, z26.b, z20.b\n"
+            ".inst 0x45049b59  // smmla z25.s, z26.b, z4.b\n"
+            "and z29.b, z29.b, #0xf0\n"
+            "ld1h { z17.s }, p1/Z, [x21]\n"
+            ".inst 0x45149abb  // smmla z27.s, z21.b, z20.b\n"
+            ".inst 0x45049ab3  // smmla z19.s, z21.b, z4.b\n"
+            "and z16.b, z16.b, #0xf0\n"
+            "ld1h { z4.s }, p0/Z, [x20]\n"
+            "subs x22, x22, #0x1\n"
+            "add x28, x28, #0x88\n"
+            "fcvt z17.s, p1/m, z17.h\n"
+            "add x25, x25, #0x90\n"
+            ".inst 0x45089942  // smmla z2.s, z10.b, z8.b\n"
+            ".inst 0x450e9959  // smmla z25.s, z10.b, z14.b\n"
+            "fcvt z4.s, p1/m, z4.h\n"
+            ".inst 0x45089afb  // smmla z27.s, z23.b, z8.b\n"
+            ".inst 0x450e9af3  // smmla z19.s, z23.b, z14.b\n"
+            "fscale z17.s, p1/m, z17.s, z28.s\n"
+            "mov z4.q, z4.q[0]\n"
+            ".inst 0x45039962  // smmla z2.s, z11.b, z3.b\n"
+            ".inst 0x45069979  // smmla z25.s, z11.b, z6.b\n"
+            "fmul z23.s, z17.s, z4.s[0]\n"
+            "fmul z9.s, z17.s, z4.s[1]\n"
+            "fmul z21.s, z17.s, z4.s[2]\n"
+            "fmul z4.s, z17.s, z4.s[3]\n"
+            ".inst 0x450398fb  // smmla z27.s, z7.b, z3.b\n"
+            ".inst 0x450698f3  // smmla z19.s, z7.b, z6.b\n"
+            ".inst 0x451d9a42  // smmla z2.s, z18.b, z29.b\n"
+            ".inst 0x45109a59  // smmla z25.s, z18.b, z16.b\n"
+            ".inst 0x451d9bdb  // smmla z27.s, z30.b, z29.b\n"
+            ".inst 0x45109bd3  // smmla z19.s, z30.b, z16.b\n"
+            "uzp1 z31.d, z2.d, z25.d\n"
+            "uzp2 z13.d, z2.d, z25.d\n"
+            "scvtf z31.s, p1/m, z31.s\n"
+            "uzp1 z17.d, z27.d, z19.d\n"
+            "uzp2 z18.d, z27.d, z19.d\n"
+            "scvtf z13.s, p1/m, z13.s\n"
+            "fmla z24.s, p1/M, z31.s, z23.s\n"
+            "scvtf z17.s, p1/m, z17.s\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "fmla z15.s, p1/M, z13.s, z9.s\n"
+            "fmla z12.s, p1/M, z17.s, z21.s\n"
+            "fmla z0.s, p1/M, z18.s, z4.s\n"
+            "bgt 7b\n"
+            "mov x20, %x[res_ptr]\n"
+            "cmp x13, #0x1\n"
+            "st1w { z24.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "cmp x13, #0x2\n"
+            "st1w { z15.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "cmp x13, #0x3\n"
+            "st1w { z12.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "st1w { z0.s }, p1, [x20]\n"
+            "8:"  // Row tail: Accumulator store skip
+            "subs x24, x24, #0x8\n"
+            "add %x[res_ptr], %x[res_ptr], #0x20\n"
+            "bne 6b\n"
+            "subs x13, x13, #0x4\n"
+            "add %x[a_ptr], %x[a_ptr], x12\n"
+            "mov %x[res_ptr], x23\n"
+            "bgt 5b\n"
+            "9:"  // Row tail: Row loop skip
+            : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
+            : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
+            : "cc", "memory", "p0", "p1", "x9", "x10", "x11", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+        );
+        return;
+    }
+#endif // #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
+
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
+    float sumf[4][8];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+}
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+        const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl);
+
+        for (int y = 0; y < nr / 4; y++) {
+            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+
+                float32x4_t sumf[4];
+                for (int m = 0; m < 4; m++) {
+                    sumf[m] = vdupq_n_f32(0);
+                }
+
+                for (int l = 0; l < nb; l++) {
+                    float32x4_t a_d = vcvt_f32_f16(vld1_f16((const float16_t *)a_ptr[l].d));
+                    float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *)b_ptr[l].d));
+
+                    int32x4_t sumi_0 = vdupq_n_s32(0);
+                    int32x4_t sumi_1 = vdupq_n_s32(0);
+                    int32x4_t sumi_2 = vdupq_n_s32(0);
+                    int32x4_t sumi_3 = vdupq_n_s32(0);
+
+                    for (int k = 0; k < 4; k++) {
+                        int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 16 * k + 0);
+                        int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16 * k + 64);
+
+                        uint8x16_t b = vld1q_u8(b_ptr[l].qs + 16 * k);
+                        int8x16_t b_hi = vqtbl1q_s8(kvalues, b >> 4);
+                        int8x16_t b_lo = vqtbl1q_s8(kvalues, b & 0xF);
+
+                        sumi_0 = vdotq_laneq_s32(sumi_0, b_lo, a_0, 0);
+                        sumi_1 = vdotq_laneq_s32(sumi_1, b_lo, a_0, 1);
+                        sumi_2 = vdotq_laneq_s32(sumi_2, b_lo, a_0, 2);
+                        sumi_3 = vdotq_laneq_s32(sumi_3, b_lo, a_0, 3);
+                        sumi_0 = vdotq_laneq_s32(sumi_0, b_hi, a_1, 0);
+                        sumi_1 = vdotq_laneq_s32(sumi_1, b_hi, a_1, 1);
+                        sumi_2 = vdotq_laneq_s32(sumi_2, b_hi, a_1, 2);
+                        sumi_3 = vdotq_laneq_s32(sumi_3, b_hi, a_1, 3);
+                    }
+
+                    sumf[0] = vmlaq_f32(sumf[0], vmulq_laneq_f32(b_d, a_d, 0), vcvtq_f32_s32(sumi_0));
+                    sumf[1] = vmlaq_f32(sumf[1], vmulq_laneq_f32(b_d, a_d, 1), vcvtq_f32_s32(sumi_1));
+                    sumf[2] = vmlaq_f32(sumf[2], vmulq_laneq_f32(b_d, a_d, 2), vcvtq_f32_s32(sumi_2));
+                    sumf[3] = vmlaq_f32(sumf[3], vmulq_laneq_f32(b_d, a_d, 3), vcvtq_f32_s32(sumi_3));
+                }
+
+                for (int m = 0; m < 4; m++) {
+                    vst1q_f32(s + (y * 4 + m) * bs + x * 4, sumf[m]);
+                }
+            }
+        }
+        return;
+    }
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
+    {
+        float sumf[4][4];
+        int sumi;
+
+        for (int y = 0; y < nr / 4; y++) {
+            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+                for (int m = 0; m < 4; m++) {
+                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+                }
+                for (int l = 0; l < nb; l++) {
+                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                        for (int m = 0; m < 4; m++) {
+                            for (int j = 0; j < ncols_interleaved; j++) {
+                                sumi = 0;
+                                for (int i = 0; i < blocklen; ++i) {
+                                    const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                                    const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
+                                }
+                                sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            }
+                        }
+                    }
+                }
+                for (int m = 0; m < 4; m++) {
+                    for (int j = 0; j < ncols_interleaved; j++)
+                        s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+                }
+            }
+        }
+    }
+}

ggml/src/ggml-cpu/arch/riscv/quants.c

@@ -0,0 +1,2068 @@
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+#include "ggml-quants.h"
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+
+#include "../../quants.h"
+#include "../../ggml-cpu-impl.h"
+
+#include <math.h>
+#include <string.h>
+#include <assert.h>
+#include <float.h>
+#include <stdlib.h> // for qsort
+#include <stdio.h>  // for GGML_ASSERT
+
+#define GROUP_MAX_EPS 1e-15f
+#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
+#define GROUP_MAX_EPS_IQ2_S 1e-8f
+#define GROUP_MAX_EPS_IQ1_M 1e-7f
+#define GROUP_MAX_EPS_IQ1_S 1e-12f
+
+#define UNUSED GGML_UNUSED
+
+void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(QK8_0 == 32);
+    assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
+
+    block_q8_0 * GGML_RESTRICT y = vy;
+
+#if defined(__riscv_v)
+
+    size_t vl = QK8_0;
+
+    for (int i = 0; i < nb; i++) {
+        // load elements
+        vfloat32m8_t v_x   = __riscv_vle32_v_f32m8(x+i*QK8_0, vl);
+
+        vfloat32m8_t vfabs = __riscv_vfabs_v_f32m8(v_x, vl);
+        vfloat32m1_t tmp   = __riscv_vfmv_v_f_f32m1(0.0f, vl);
+        vfloat32m1_t vmax  = __riscv_vfredmax_vs_f32m8_f32m1(vfabs, tmp, vl);
+        float amax = __riscv_vfmv_f_s_f32m1_f32(vmax);
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        vfloat32m8_t x0 = __riscv_vfmul_vf_f32m8(v_x, id, vl);
+
+        // convert to integer
+        vint16m4_t   vi = __riscv_vfncvt_x_f_w_i16m4(x0, vl);
+        vint8m2_t    vs = __riscv_vncvt_x_x_w_i8m2(vi, vl);
+
+        // store result
+        __riscv_vse8_v_i8m2(y[i].qs , vs, vl);
+    }
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_0_ref(x, y, k);
+#endif
+}
+
+void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(k % QK8_1 == 0);
+    const int nb = k / QK8_1;
+
+    block_q8_1 * GGML_RESTRICT y = vy;
+
+#if defined(__riscv_v)
+
+    size_t vl = QK8_1;
+
+    for (int i = 0; i < nb; i++) {
+        // load elements
+        vfloat32m8_t v_x   = __riscv_vle32_v_f32m8(x+i*QK8_1, vl);
+
+        vfloat32m8_t vfabs = __riscv_vfabs_v_f32m8(v_x, vl);
+        vfloat32m1_t tmp   = __riscv_vfmv_v_f_f32m1(0.0, vl);
+        vfloat32m1_t vmax  = __riscv_vfredmax_vs_f32m8_f32m1(vfabs, tmp, vl);
+        float amax = __riscv_vfmv_f_s_f32m1_f32(vmax);
+
+        const float d  = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        vfloat32m8_t x0 = __riscv_vfmul_vf_f32m8(v_x, id, vl);
+
+        // convert to integer
+        vint16m4_t   vi = __riscv_vfncvt_x_f_w_i16m4(x0, vl);
+        vint8m2_t    vs = __riscv_vncvt_x_x_w_i8m2(vi, vl);
+
+        // store result
+        __riscv_vse8_v_i8m2(y[i].qs , vs, vl);
+
+        // compute sum for y[i].s
+        vint16m1_t tmp2 = __riscv_vmv_v_x_i16m1(0, vl);
+        vint16m1_t vwrs = __riscv_vwredsum_vs_i8m2_i16m1(vs, tmp2, vl);
+
+        // set y[i].s
+        int sum = __riscv_vmv_x_s_i16m1_i16(vwrs);
+        y[i].s = GGML_FP32_TO_FP16(sum*d);
+    }
+
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_1_ref(x, y, k);
+#endif
+}
+
+//===================================== Dot products =================================
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__riscv_v)
+    size_t vl = qk / 2;
+
+    for (; ib < nb; ++ib) {
+        // load elements
+        vuint8m1_t tx = __riscv_vle8_v_u8m1(x[ib].qs, vl);
+
+        vint8m1_t y0 = __riscv_vle8_v_i8m1(y[ib].qs, vl);
+        vint8m1_t y1 = __riscv_vle8_v_i8m1(y[ib].qs+16, vl);
+
+        // mask and store lower part of x, and then upper part
+        vuint8m1_t x_a = __riscv_vand_vx_u8m1(tx, 0x0F, vl);
+        vuint8m1_t x_l = __riscv_vsrl_vx_u8m1(tx, 0x04, vl);
+
+        vint8m1_t x_ai = __riscv_vreinterpret_v_u8m1_i8m1(x_a);
+        vint8m1_t x_li = __riscv_vreinterpret_v_u8m1_i8m1(x_l);
+
+        // subtract offset
+        vint8m1_t v0 = __riscv_vsub_vx_i8m1(x_ai, 8, vl);
+        vint8m1_t v1 = __riscv_vsub_vx_i8m1(x_li, 8, vl);
+
+        vint16m2_t vec_mul1 = __riscv_vwmul_vv_i16m2(v0, y0, vl);
+        vint16m2_t vec_mul2 = __riscv_vwmacc_vv_i16m2(vec_mul1, v1, y1, vl);
+
+        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
+        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m2_i32m1(vec_mul2, vec_zero, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
+
+        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+    }
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
+            const int v1 = (x[ib].qs[j] >>   4) - 8;
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+    }
+
+    *s = sumf;
+}
+
+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) {
+    const int qk = QK8_1;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__riscv_v)
+    size_t vl = qk / 2;
+
+    for (; ib < nb; ++ib) {
+        // load elements
+        vuint8m1_t tx = __riscv_vle8_v_u8m1(x[ib].qs, vl);
+
+        vint8m1_t y0 = __riscv_vle8_v_i8m1(y[ib].qs, vl);
+        vint8m1_t y1 = __riscv_vle8_v_i8m1(y[ib].qs+16, vl);
+
+        // mask and store lower part of x, and then upper part
+        vuint8m1_t x_a = __riscv_vand_vx_u8m1(tx, 0x0F, vl);
+        vuint8m1_t x_l = __riscv_vsrl_vx_u8m1(tx, 0x04, vl);
+
+        vint8m1_t v0 = __riscv_vreinterpret_v_u8m1_i8m1(x_a);
+        vint8m1_t v1 = __riscv_vreinterpret_v_u8m1_i8m1(x_l);
+
+        vint16m2_t vec_mul1 = __riscv_vwmul_vv_i16m2(v0, y0, vl);
+        vint16m2_t vec_mul2 = __riscv_vwmacc_vv_i16m2(vec_mul1, v1, y1, vl);
+
+        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
+        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m2_i32m1(vec_mul2, vec_zero, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
+
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+    }
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F);
+            const int v1 = (x[ib].qs[j] >>   4);
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+#if defined(__riscv_v)
+    size_t vl;
+    size_t vlenb = __riscv_vlenb();
+
+    for (; ib < nb; ++ib) {
+        vl = qk / 2;
+        vuint8m1_t v0 = __riscv_vle8_v_u8m1(x[ib].qs, vl);
+        vint8m1_t v0l = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(v0, 0x0F, vl));
+        vint8m1_t v0h = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(v0, 4, vl));
+        vint8m2_t v0c;
+        if (vlenb == 16) {
+            v0c = __riscv_vcreate_v_i8m1_i8m2(v0l, v0h);
+        } else {
+            v0l = __riscv_vslideup_vx_i8m1(v0l, v0h, 16, 32);
+            v0c = __riscv_vlmul_ext_v_i8m1_i8m2(v0l);
+        }
+
+        vl = qk;
+        vbool4_t qh = __riscv_vlm_v_b4(x[ib].qh, vl);
+        qh = __riscv_vmnand_mm_b4(qh, qh, vl);
+        vint8m2_t v0f = __riscv_vsub_vx_i8m2_mu(qh, v0c, v0c, 0x10, vl);
+        vint8m2_t v1 = __riscv_vle8_v_i8m2(y[ib].qs, vl);
+        vint16m4_t mul = __riscv_vwmul_vv_i16m4(v0f, v1, vl);
+        vint32m1_t zero = __riscv_vmv_v_x_i32m1(0, vl);
+        vint32m1_t sum = __riscv_vwredsum_vs_i16m4_i32m1(mul, zero, vl);
+        int32_t sumi = __riscv_vmv_x_s_i32m1_i32(sum);
+
+        sumf += (GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+    }
+
+#endif
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
+            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
+
+            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
+            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+    }
+
+    *s = sumf;
+}
+
+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) {
+    const int qk = QK8_1;
+    const int nb = n / qk;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_1);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+#if defined(__riscv_v)
+    size_t vl;
+    size_t vlenb = __riscv_vlenb();
+
+    for (; ib < nb; ++ib) {
+        vl = qk / 2;
+        vuint8m1_t v0 = __riscv_vle8_v_u8m1(x[ib].qs, vl);
+        vint8m1_t v0l = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(v0, 0x0F, vl));
+        vint8m1_t v0h = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(v0, 4, vl));
+        vint8m2_t v0c;
+        if (vlenb == 16) {
+            v0c = __riscv_vcreate_v_i8m1_i8m2(v0l, v0h);
+        } else {
+            v0l = __riscv_vslideup_vx_i8m1(v0l, v0h, 16, 32);
+            v0c = __riscv_vlmul_ext_v_i8m1_i8m2(v0l);
+        }
+
+        vl = qk;
+        vbool4_t qh = __riscv_vlm_v_b4(x[ib].qh, vl);
+        vint8m2_t v0f = __riscv_vor_vx_i8m2_mu(qh, v0c, v0c, 0x10, vl);
+        vint8m2_t v1 = __riscv_vle8_v_i8m2(y[ib].qs, vl);
+        vint16m4_t mul = __riscv_vwmul_vv_i16m4(v0f, v1, vl);
+        vint32m1_t zero = __riscv_vmv_v_x_i32m1(0, vl);
+        vint32m1_t sum = __riscv_vwredsum_vs_i16m4_i32m1(mul, zero, vl);
+        int32_t sumi = __riscv_vmv_x_s_i32m1_i32(sum);
+
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+    }
+
+#endif
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
+            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
+
+            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
+            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q8_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__riscv_v)
+    size_t vl = qk;
+
+    for (; ib < nb; ++ib) {
+        // load elements
+        vint8m2_t bx_0 = __riscv_vle8_v_i8m2(x[ib].qs, vl);
+        vint8m2_t by_0 = __riscv_vle8_v_i8m2(y[ib].qs, vl);
+
+        vint16m4_t vw_mul = __riscv_vwmul_vv_i16m4(bx_0, by_0, vl);
+
+        vint32m1_t v_zero = __riscv_vmv_v_x_i32m1(0, vl);
+        vint32m1_t v_sum = __riscv_vwredsum_vs_i16m4_i32m1(vw_mul, v_zero, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(v_sum);
+
+        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+    }
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi = 0;
+
+        for (int j = 0; j < qk; j++) {
+            sumi += x[ib].qs[j]*y[ib].qs[j];
+        }
+
+        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+    }
+
+    *s = sumf;
+}
+
+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) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q2_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __riscv_xtheadvector
+
+    float sumf = 0;
+    uint8_t atmp[16];
+
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * q2 = x[i].qs;
+        const  int8_t * q8 = y[i].qs;
+        const uint8_t * sc = x[i].scales;
+        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        uint8_t *patmp = atmp;
+        int vsums;
+        int tmp;
+        __asm__ __volatile__(
+            "th.vsetvli zero, %[vl16], e8, m1\n\t"
+            "th.vmv.v.x v8, zero\n\t"
+            "th.vlb.v v1, (%[sc])\n\t"
+            "th.vand.vi v0, v1, 0xF\n\t"
+            "th.vsrl.vi v1, v1, 4\n\t"
+            "th.vsb.v v0, (%[scale])\n\t"
+            "th.vwaddu.vx v16, v1, zero\n\t"
+            "th.vsetvli zero, %[vl16], e16, m2\n\t"
+            "th.vlh.v v2, (%[bsums])\n\t"
+            "th.vwmul.vv v4, v16, v2\n\t"
+            "th.vsetvli zero, %[vl16], e32, m4\n\t"
+            "th.vredsum.vs v8, v4, v8\n\t"
+            "th.vmv.x.s %[vsums], v8"
+            : [tmp] "=&r" (tmp), [vsums] "=&r" (vsums)
+            : [sc] "r" (sc), [scale] "r" (atmp), [bsums] "r" (y[i].bsums)
+            , [vl16] "r" (16)
+            : "memory"
+            , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+            , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
+            , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
+            , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
+        );
+        sumf += dmin * vsums;
+        int isum = 0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            __asm__ __volatile__(
+                "th.vsetvli zero, %[vl32], e8, m2\n\t"
+                "th.vlb.v v0, (%[q2])\n\t"
+                "th.vsrl.vi v2, v0, 2\n\t"
+                "th.vsrl.vi v4, v0, 4\n\t"
+                "th.vsrl.vi v6, v0, 6\n\t"
+                "th.vand.vi v0, v0, 0x3\n\t"
+                "th.vand.vi v2, v2, 0x3\n\t"
+                "th.vand.vi v4, v4, 0x3\n\t"
+                "th.vsetvli zero, %[vl128], e8, m8\n\t"
+                "th.vlb.v v8, (%[q8])\n\t"
+                "th.vsetvli zero, %[vl64], e8, m4\n\t"
+                "th.vwmul.vv v16, v0, v8\n\t"
+                "th.vwmul.vv v24, v4, v12\n\t"
+                "th.vsetvli zero, %[vl16], e16, m2\n\t"
+                "th.vmv.v.x v0, zero\n\t"
+                "th.vwredsum.vs v10, v16, v0\n\t"
+                "th.vwredsum.vs v9, v18, v0\n\t"
+                "th.vwredsum.vs v8, v20, v0\n\t"
+                "th.vwredsum.vs v7, v22, v0\n\t"
+                "th.vwredsum.vs v11, v24, v0\n\t"
+                "th.vwredsum.vs v12, v26, v0\n\t"
+                "th.vwredsum.vs v13, v28, v0\n\t"
+                "th.vwredsum.vs v14, v30, v0\n\t"
+                "li %[tmp], 4\n\t"
+                "th.vsetvli zero, %[tmp], e32, m1\n\t"
+                "th.vslideup.vi v10, v9, 1\n\t"
+                "th.vslideup.vi v8, v7, 1\n\t"
+                "th.vslideup.vi v11, v12, 1\n\t"
+                "th.vslideup.vi v13, v14, 1\n\t"
+                "th.vslideup.vi v10, v8, 2\n\t"
+                "th.vslideup.vi v11, v13, 2\n\t"
+                "li %[tmp], 8\n\t"
+                "th.vsetvli zero, %[tmp], e32, m2\n\t"
+                "th.vlbu.v v12, (%[scale])\n\t"
+                "th.vmul.vv v10, v10, v12\n\t"
+                "th.vredsum.vs v0, v10, v0\n\t"
+                "th.vmv.x.s %[tmp], v0\n\t"
+                "add %[isum], %[isum], %[tmp]"
+                : [tmp] "=&r" (tmp), [isum] "+&r" (isum)
+                : [q2] "r" (q2), [scale] "r" (patmp), [q8] "r" (q8)
+                , [vl16] "r" (16), [vl32] "r" (32), [vl64] "r" (64), [vl128] "r" (128)
+                : "memory"
+                , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+                , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
+                , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
+                , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
+            );
+            q2 += 32; q8 += 128; patmp += 8;
+        }
+
+        sumf += dall * isum;
+    }
+
+    *s = sumf;
+
+#elif defined __riscv_v
+
+    float sumf = 0;
+    uint8_t atmp[16];
+
+    const int vector_length = __riscv_vlenb() * 8;
+    uint8_t temp_01[32] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                            1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 };
+
+    switch (vector_length) {
+    case 256:
+        for (int i = 0; i < nb; ++i) {
+            const uint8_t * q2 = x[i].qs;
+            const int8_t *  q8 = y[i].qs;
+            const uint8_t * sc = x[i].scales;
+
+            const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+            const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+            size_t vl = 16;
+
+            vuint8m1_t scales = __riscv_vle8_v_u8m1(sc, vl);
+            vuint8m1_t aux    = __riscv_vand_vx_u8m1(scales, 0x0F, vl);
+
+            vint16m1_t q8sums = __riscv_vle16_v_i16m1(y[i].bsums, vl);
+
+            vuint8mf2_t scales_2 = __riscv_vle8_v_u8mf2(sc, vl);
+            vuint8mf2_t mins8    = __riscv_vsrl_vx_u8mf2(scales_2, 0x4, vl);
+            vint16m1_t  mins     = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(mins8, vl));
+            vint32m2_t  prod     = __riscv_vwmul_vv_i32m2(q8sums, mins, vl);
+            vint32m1_t  vsums    = __riscv_vredsum_vs_i32m2_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
+
+            sumf += dmin * __riscv_vmv_x_s_i32m1_i32(vsums);
+
+            vl = 32;
+
+            vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+            vuint8m1_t v_b   = __riscv_vle8_v_u8m1(temp_01, vl);
+
+            uint8_t is   = 0;
+            int     isum = 0;
+
+            for (int j = 0; j < QK_K / 128; ++j) {
+                // load Q2
+                vuint8m1_t q2_x = __riscv_vle8_v_u8m1(q2, vl);
+
+                vuint8m1_t q2_0 = __riscv_vand_vx_u8m1(q2_x, 0x03, vl);
+                vuint8m1_t q2_1 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x2, vl), 0x03, vl);
+                vuint8m1_t q2_2 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x4, vl), 0x03, vl);
+                vuint8m1_t q2_3 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x6, vl), 0x03, vl);
+
+                // duplicate scale elements for product
+                vuint8m1_t sc0 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 0 + is, vl), vl);
+                vuint8m1_t sc1 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 2 + is, vl), vl);
+                vuint8m1_t sc2 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 4 + is, vl), vl);
+                vuint8m1_t sc3 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 6 + is, vl), vl);
+
+                vint16m2_t p0 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_0, sc0, vl));
+                vint16m2_t p1 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_1, sc1, vl));
+                vint16m2_t p2 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_2, sc2, vl));
+                vint16m2_t p3 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_3, sc3, vl));
+
+                // load Q8
+                vint8m1_t q8_0 = __riscv_vle8_v_i8m1(q8, vl);
+                vint8m1_t q8_1 = __riscv_vle8_v_i8m1(q8 + 32, vl);
+                vint8m1_t q8_2 = __riscv_vle8_v_i8m1(q8 + 64, vl);
+                vint8m1_t q8_3 = __riscv_vle8_v_i8m1(q8 + 96, vl);
+
+                vint32m4_t s0 = __riscv_vwmul_vv_i32m4(p0, __riscv_vwcvt_x_x_v_i16m2(q8_0, vl), vl);
+                vint32m4_t s1 = __riscv_vwmul_vv_i32m4(p1, __riscv_vwcvt_x_x_v_i16m2(q8_1, vl), vl);
+                vint32m4_t s2 = __riscv_vwmul_vv_i32m4(p2, __riscv_vwcvt_x_x_v_i16m2(q8_2, vl), vl);
+                vint32m4_t s3 = __riscv_vwmul_vv_i32m4(p3, __riscv_vwcvt_x_x_v_i16m2(q8_3, vl), vl);
+
+                vint32m1_t isum0 = __riscv_vredsum_vs_i32m4_i32m1(__riscv_vadd_vv_i32m4(s0, s1, vl), vzero, vl);
+                vint32m1_t isum1 = __riscv_vredsum_vs_i32m4_i32m1(__riscv_vadd_vv_i32m4(s2, s3, vl), isum0, vl);
+
+                isum += __riscv_vmv_x_s_i32m1_i32(isum1);
+
+                q2 += 32;
+                q8 += 128;
+                is = 8;
+            }
+
+            sumf += dall * isum;
+        }
+        break;
+    case 128:
+        for (int i = 0; i < nb; ++i) {
+            const uint8_t * q2 = x[i].qs;
+            const  int8_t * q8 = y[i].qs;
+            const uint8_t * sc = x[i].scales;
+            const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+            const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+            uint8_t *patmp = atmp;
+            int vsums;
+            int tmp;
+            __asm__ __volatile__(
+                "vsetivli zero, 16, e8, m1\n\t"
+                "vmv.v.x v8, zero\n\t"
+                "vle8.v v1, (%[sc])\n\t"
+                "vand.vi v0, v1, 0xF\n\t"
+                "vsrl.vi v1, v1, 4\n\t"
+                "vse8.v v0, (%[scale])\n\t"
+                "vsetivli zero, 16, e16, m2\n\t"
+                "vle16.v v2, (%[bsums])\n\t"
+                "vzext.vf2 v0, v1\n\t"
+                "vwmul.vv v4, v0, v2\n\t"
+                "vsetivli zero, 16, e32, m4\n\t"
+                "vredsum.vs v8, v4, v8\n\t"
+                "vmv.x.s %[vsums], v8"
+                : [tmp] "=&r" (tmp), [vsums] "=&r" (vsums)
+                : [sc] "r" (sc), [scale] "r" (atmp), [bsums] "r" (y[i].bsums)
+                : "memory"
+                , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+                , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
+                , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
+                , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
+            );
+            sumf += dmin * vsums;
+            int isum = 0;
+
+            for (int j = 0; j < QK_K/128; ++j) {
+                __asm__ __volatile__(
+                    "vsetvli zero, %[vl32], e8, m2\n\t"
+                    "vle8.v v0, (%[q2])\n\t"
+                    "vsrl.vi v2, v0, 2\n\t"
+                    "vsrl.vi v4, v0, 4\n\t"
+                    "vsrl.vi v6, v0, 6\n\t"
+                    "vand.vi v0, v0, 0x3\n\t"
+                    "vand.vi v2, v2, 0x3\n\t"
+                    "vand.vi v4, v4, 0x3\n\t"
+                    "vsetvli zero, %[vl128], e8, m8\n\t"
+                    "vle8.v v8, (%[q8])\n\t"
+                    "vsetvli zero, %[vl64], e8, m4\n\t"
+                    "vwmul.vv v16, v0, v8\n\t"
+                    "vwmul.vv v24, v4, v12\n\t"
+                    "vsetivli zero, 16, e16, m2\n\t"
+                    "vmv.v.x v0, zero\n\t"
+                    "vwredsum.vs v10, v16, v0\n\t"
+                    "vwredsum.vs v9, v18, v0\n\t"
+                    "vwredsum.vs v8, v20, v0\n\t"
+                    "vwredsum.vs v7, v22, v0\n\t"
+                    "vwredsum.vs v11, v24, v0\n\t"
+                    "vwredsum.vs v12, v26, v0\n\t"
+                    "vwredsum.vs v13, v28, v0\n\t"
+                    "vwredsum.vs v14, v30, v0\n\t"
+                    "vsetivli zero, 4, e32, m1\n\t"
+                    "vslideup.vi v10, v9, 1\n\t"
+                    "vslideup.vi v8, v7, 1\n\t"
+                    "vslideup.vi v11, v12, 1\n\t"
+                    "vslideup.vi v13, v14, 1\n\t"
+                    "vslideup.vi v10, v8, 2\n\t"
+                    "vslideup.vi v11, v13, 2\n\t"
+                    "vsetivli zero, 8, e32, m2\n\t"
+                    "vle8.v v15, (%[scale])\n\t"
+                    "vzext.vf4 v12, v15\n\t"
+                    "vmul.vv v10, v10, v12\n\t"
+                    "vredsum.vs v0, v10, v0\n\t"
+                    "vmv.x.s %[tmp], v0\n\t"
+                    "add %[isum], %[isum], %[tmp]"
+                    : [tmp] "=&r" (tmp), [isum] "+&r" (isum)
+                    : [q2] "r" (q2), [scale] "r" (patmp), [q8] "r" (q8)
+                    , [vl32] "r" (32), [vl64] "r" (64), [vl128] "r" (128)
+                    : "memory"
+                    , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+                    , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
+                    , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
+                    , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
+                );
+                q2 += 32; q8 += 128; patmp += 8;
+            }
+
+            sumf += dall * isum;
+        }
+        break;
+    default:
+        assert(false && "Unsupported vector length");
+        break;
+    }
+
+    *s = sumf;
+
+#else
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * q2 = x[i].qs;
+        const  int8_t * q8 = y[i].qs;
+        const uint8_t * sc = x[i].scales;
+
+        int summs = 0;
+        for (int j = 0; j < 16; ++j) {
+            summs += y[i].bsums[j] * (sc[j] >> 4);
+        }
+
+        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        int isum = 0;
+        int is = 0;
+        int d;
+        for (int k = 0; k < QK_K/128; ++k) {
+            int shift = 0;
+            for (int j = 0; j < 4; ++j) {
+                d = sc[is++] & 0xF;
+                int isuml = 0;
+                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                d = sc[is++] & 0xF;
+                isuml = 0;
+                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                shift += 2;
+                q8 += 32;
+            }
+            q2 += 32;
+        }
+        sumf += dall * isum - dmin * summs;
+    }
+    *s = sumf;
+#endif
+}
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const uint32_t kmask1 = 0x03030303;
+    const uint32_t kmask2 = 0x0f0f0f0f;
+
+    const block_q3_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __riscv_xtheadvector
+
+    uint32_t utmp[4];
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict qh = x[i].hmask;
+        const  int8_t * restrict q8 = y[i].qs;
+
+        int8_t * scale = (int8_t *)utmp;
+        int tmp;
+        __asm__ __volatile__(
+            "li %[tmp], 12\n\t"
+            "th.vsetvli zero, %[tmp], e8, m1\n\t"
+            "th.vlb.v v0, (%[s6b])\n\t"
+            "th.vmv.v.v v2, v0\n\t"
+            "li %[tmp], 2\n\t"
+            "th.vsetvli zero, %[tmp], e64, m1\n\t"
+            "th.vmv.v.x v9, %[sh]\n\t"\
+            "th.vslidedown.vi v1, v0, 1\n\t"
+            "th.vslide1up.vx v8, v9, zero\n\t" // {0, 0, 4, 4}
+            "th.vslideup.vi v0, v2, 1\n\t" // {aux[0], aux[1], aux[0], aux[1]}
+            "li %[tmp], 4\n\t"
+            "th.vsetvli zero, %[tmp], e32, m1\n\t"
+            "th.vid.v v9\n\t"
+            "th.vmv.x.s %[tmp], v1\n\t"
+            "th.vsll.vi v9, v9, 1\n\t" // {0, 2, 4, 6}
+            "th.vmv.v.x v1, %[tmp]\n\t" // {aux[2], aux[2], aux[2], aux[2]}
+            "th.vsrl.vv v4, v1, v9\n\t"
+            "th.vsrl.vv v2, v0, v8\n\t"
+            "th.vand.vx v5, v4, %[kmask1]\n\t"
+            "th.vand.vx v3, v2, %[kmask2]\n\t"
+            "th.vsll.vi v6, v5, 4\n\t"
+            "th.vor.vv v7, v6, v3\n\t"
+            "li %[tmp], 16\n\t"
+            "th.vsetvli zero, %[tmp], e8, m1\n\t"
+            "th.vsub.vx v0, v7, %[c]\n\t"
+            "th.vsb.v v0, (%[scale])"
+            : [tmp] "=&r" (tmp)
+            : [sh] "r" (0x0000000400000004), [s6b] "r" (x[i].scales), [c] "r" (32)
+            , [scale] "r" (scale), [kmask1] "r" (kmask1), [kmask2] "r" (kmask2)
+            : "memory"
+            , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+            , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
+            , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
+            , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
+        );
+
+        uint8_t m = 1;
+        int isum = 0;
+        for (int j = 0; j < QK_K; j += 128) {
+            __asm__ __volatile__(
+                // fixme: use v0p7 mask layout directly
+                "th.vsetvli zero, %[vl32], e8, m2\n\t"
+                "th.vlb.v v8, (%[q3])\n\t"
+                "th.vsrl.vi v10, v8, 2\n\t"
+                "th.vsrl.vi v12, v8, 4\n\t"
+                "th.vsrl.vi v14, v8, 6\n\t"
+                "th.vand.vi v8, v8, 3\n\t"
+                "th.vand.vi v10, v10, 3\n\t"
+                "th.vand.vi v12, v12, 3\n\t"
+                "th.vlb.v v2, (%[qh])\n\t"
+                "th.vand.vx v4, v2, %[m]\n\t"
+                "slli %[m], %[m], 1\n\t"
+                "th.vmseq.vx v0, v4, zero\n\t"
+                "th.vadd.vi v8, v8, -4, v0.t\n\t"
+                "th.vand.vx v4, v2, %[m]\n\t"
+                "slli %[m], %[m], 1\n\t"
+                "th.vmseq.vx v0, v4, zero\n\t"
+                "th.vadd.vi v10, v10, -4, v0.t\n\t"
+                "th.vand.vx v4, v2, %[m]\n\t"
+                "slli %[m], %[m], 1\n\t"
+                "th.vmseq.vx v0, v4, zero\n\t"
+                "th.vadd.vi v12, v12, -4, v0.t\n\t"
+                "th.vand.vx v4, v2, %[m]\n\t"
+                "slli %[m], %[m], 1\n\t"
+                "th.vmseq.vx v0, v4, zero\n\t"
+                "th.vadd.vi v14, v14, -4, v0.t\n\t"
+                "th.vsetvli zero, %[vl128], e8, m8\n\t"
+                "th.vlb.v v0, (%[q8])\n\t"
+                "th.vsetvli zero, %[vl64], e8, m4\n\t"
+                "th.vwmul.vv v16, v0, v8\n\t"
+                "th.vwmul.vv v24, v4, v12\n\t"
+                "li %[tmp], 16\n\t"
+                "th.vsetvli zero, %[tmp], e16, m2\n\t"
+                "th.vmv.v.x v0, zero\n\t"
+                "th.vwredsum.vs v10, v16, v0\n\t"
+                "th.vwredsum.vs v9, v18, v0\n\t"
+                "th.vwredsum.vs v8, v20, v0\n\t"
+                "th.vwredsum.vs v7, v22, v0\n\t"
+                "th.vwredsum.vs v11, v24, v0\n\t"
+                "th.vwredsum.vs v12, v26, v0\n\t"
+                "th.vwredsum.vs v13, v28, v0\n\t"
+                "th.vwredsum.vs v14, v30, v0\n\t"
+                "li %[tmp], 4\n\t"
+                "th.vsetvli zero, %[tmp], e32, m1\n\t"
+                "th.vslideup.vi v10, v9, 1\n\t"
+                "th.vslideup.vi v8, v7, 1\n\t"
+                "th.vslideup.vi v11, v12, 1\n\t"
+                "th.vslideup.vi v13, v14, 1\n\t"
+                "th.vslideup.vi v10, v8, 2\n\t"
+                "th.vslideup.vi v11, v13, 2\n\t"
+                "li %[tmp], 8\n\t"
+                "th.vsetvli zero, %[tmp], e32, m2\n\t"
+                "th.vlb.v v12, (%[scale])\n\t"
+                "th.vmul.vv v10, v10, v12\n\t"
+                "th.vredsum.vs v0, v10, v0\n\t"
+                "th.vmv.x.s %[tmp], v0\n\t"
+                "add %[isum], %[isum], %[tmp]"
+                : [tmp] "=&r" (tmp), [m] "+&r" (m), [isum] "+&r" (isum)
+                : [vl128] "r" (128), [vl64] "r" (64), [vl32] "r" (32)
+                , [q3] "r" (q3), [qh] "r" (qh), [scale] "r" (scale), [q8] "r" (q8)
+                : "memory"
+                , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+                , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
+                , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
+                , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
+            );
+            q3 += 32;    q8 += 128;   scale += 8;
+        }
+
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        sumf += d * isum;
+    }
+
+    *s = sumf;
+
+#elif defined __riscv_v
+
+    uint32_t utmp[4];
+    float sumf = 0;
+    uint32_t aux[3];
+    const int vector_length = __riscv_vlenb() * 8;
+
+    switch (vector_length) {
+    case 256:
+        for (int i = 0; i < nb; ++i) {
+
+            const uint8_t * GGML_RESTRICT q3 = x[i].qs;
+            const uint8_t * GGML_RESTRICT qh = x[i].hmask;
+            const  int8_t * GGML_RESTRICT q8 = y[i].qs;
+
+            memcpy(aux, x[i].scales, 12);
+            utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
+            utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
+            utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
+            utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
+
+            int8_t * scale = (int8_t *)utmp;
+            for (int j = 0; j < 16; ++j) scale[j] -= 32;
+
+
+            size_t vl = 32;
+            uint8_t m =  1;
+
+            vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+            vuint8m1_t vqh = __riscv_vle8_v_u8m1(qh, vl);
+
+            int sum_t = 0;
+
+            for (int j = 0; j < QK_K; j += 128) {
+
+                vl = 32;
+
+                // load Q3
+                vuint8m1_t q3_x = __riscv_vle8_v_u8m1(q3, vl);
+
+                vint8m1_t q3_0 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q3_x, 0x03, vl));
+                vint8m1_t q3_1 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x2, vl), 0x03 , vl));
+                vint8m1_t q3_2 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x4, vl), 0x03 , vl));
+                vint8m1_t q3_3 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x6, vl), 0x03 , vl));
+
+                // compute mask for subtraction
+                vuint8m1_t qh_m0 = __riscv_vand_vx_u8m1(vqh, m, vl);
+                vbool8_t vmask_0 = __riscv_vmseq_vx_u8m1_b8(qh_m0, 0, vl);
+                vint8m1_t q3_m0 = __riscv_vsub_vx_i8m1_mu(vmask_0, q3_0, q3_0, 0x4, vl);
+                m <<= 1;
+
+                vuint8m1_t qh_m1 = __riscv_vand_vx_u8m1(vqh, m, vl);
+                vbool8_t vmask_1 = __riscv_vmseq_vx_u8m1_b8(qh_m1, 0, vl);
+                vint8m1_t q3_m1 = __riscv_vsub_vx_i8m1_mu(vmask_1, q3_1, q3_1, 0x4, vl);
+                m <<= 1;
+
+                vuint8m1_t qh_m2 = __riscv_vand_vx_u8m1(vqh, m, vl);
+                vbool8_t vmask_2 = __riscv_vmseq_vx_u8m1_b8(qh_m2, 0, vl);
+                vint8m1_t q3_m2 = __riscv_vsub_vx_i8m1_mu(vmask_2, q3_2, q3_2, 0x4, vl);
+                m <<= 1;
+
+                vuint8m1_t qh_m3 = __riscv_vand_vx_u8m1(vqh, m, vl);
+                vbool8_t vmask_3 = __riscv_vmseq_vx_u8m1_b8(qh_m3, 0, vl);
+                vint8m1_t q3_m3 = __riscv_vsub_vx_i8m1_mu(vmask_3, q3_3, q3_3, 0x4, vl);
+                m <<= 1;
+
+                // load Q8 and take product with Q3
+                vint16m2_t a0 = __riscv_vwmul_vv_i16m2(q3_m0, __riscv_vle8_v_i8m1(q8, vl), vl);
+                vint16m2_t a1 = __riscv_vwmul_vv_i16m2(q3_m1, __riscv_vle8_v_i8m1(q8+32, vl), vl);
+                vint16m2_t a2 = __riscv_vwmul_vv_i16m2(q3_m2, __riscv_vle8_v_i8m1(q8+64, vl), vl);
+                vint16m2_t a3 = __riscv_vwmul_vv_i16m2(q3_m3, __riscv_vle8_v_i8m1(q8+96, vl), vl);
+
+                vl = 16;
+
+                // retrieve lane to multiply with scale
+                vint32m2_t aux0_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 0), (scale[0]), vl);
+                vint32m2_t aux0_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 1), (scale[1]), vl);
+                vint32m2_t aux1_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 0), (scale[2]), vl);
+                vint32m2_t aux1_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 1), (scale[3]), vl);
+                vint32m2_t aux2_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a2, 0), (scale[4]), vl);
+                vint32m2_t aux2_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a2, 1), (scale[5]), vl);
+                vint32m2_t aux3_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a3, 0), (scale[6]), vl);
+                vint32m2_t aux3_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a3, 1), (scale[7]), vl);
+
+                vint32m1_t isum0 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux0_0, aux0_1, vl), vzero, vl);
+                vint32m1_t isum1 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux1_0, aux1_1, vl), isum0, vl);
+                vint32m1_t isum2 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux2_0, aux2_1, vl), isum1, vl);
+                vint32m1_t isum3 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux3_0, aux3_1, vl), isum2, vl);
+
+                sum_t +=  __riscv_vmv_x_s_i32m1_i32(isum3);
+
+                q3 += 32;    q8 += 128;   scale += 8;
+
+            }
+
+            const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+
+            sumf += d*sum_t;
+
+        }
+        break;
+    case 128:
+        for (int i = 0; i < nb; ++i) {
+            const uint8_t * restrict q3 = x[i].qs;
+            const uint8_t * restrict qh = x[i].hmask;
+            const  int8_t * restrict q8 = y[i].qs;
+
+            int8_t * scale = (int8_t *)utmp;
+            int tmp;
+            __asm__ __volatile__(
+                "vsetivli zero, 12, e8, m1\n\t"
+                "vle8.v v0, (%[s6b])\n\t"
+                "vmv1r.v v2, v0\n\t"
+                "vsetivli zero, 2, e64, m1\n\t"
+                "vmv.v.x v9, %[sh]\n\t"\
+                "vslidedown.vi v1, v0, 1\n\t"
+                "vslide1up.vx v8, v9, zero\n\t" // {0, 0, 4, 4}
+                "vslideup.vi v0, v2, 1\n\t" // {aux[0], aux[1], aux[0], aux[1]}
+                "vsetivli zero, 4, e32, m1\n\t"
+                "vid.v v9\n\t"
+                "vmv.x.s %[tmp], v1\n\t"
+                "vsll.vi v9, v9, 1\n\t" // {0, 2, 4, 6}
+                "vmv.v.x v1, %[tmp]\n\t" // {aux[2], aux[2], aux[2], aux[2]}
+                "vsrl.vv v4, v1, v9\n\t"
+                "vsrl.vv v2, v0, v8\n\t"
+                "vand.vx v5, v4, %[kmask1]\n\t"
+                "vand.vx v3, v2, %[kmask2]\n\t"
+                "vsll.vi v6, v5, 4\n\t"
+                "vor.vv v7, v6, v3\n\t"
+                "vsetivli zero, 16, e8, m1\n\t"
+                "vsub.vx v0, v7, %[c]\n\t"
+                "vse8.v v0, (%[scale])"
+                : [tmp] "=&r" (tmp)
+                : [sh] "r" (0x0000000400000004), [s6b] "r" (x[i].scales), [c] "r" (32)
+                , [scale] "r" (scale), [kmask1] "r" (kmask1), [kmask2] "r" (kmask2)
+                : "memory"
+                , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+                , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
+                , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
+                , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
+            );
+
+            uint8_t m = 1;
+            int isum = 0;
+            for (int j = 0; j < QK_K; j += 128) {
+                __asm__ __volatile__(
+                    "vsetvli zero, %[vl32], e8, m2, ta, mu\n\t"
+                    "vle8.v v8, (%[q3])\n\t"
+                    "vsrl.vi v10, v8, 2\n\t"
+                    "vsrl.vi v12, v8, 4\n\t"
+                    "vsrl.vi v14, v8, 6\n\t"
+                    "vand.vi v8, v8, 3\n\t"
+                    "vand.vi v10, v10, 3\n\t"
+                    "vand.vi v12, v12, 3\n\t"
+                    "vle8.v v2, (%[qh])\n\t"
+                    "vand.vx v4, v2, %[m]\n\t"
+                    "slli %[m], %[m], 1\n\t"
+                    "vmseq.vx v0, v4, zero\n\t"
+                    "vadd.vi v8, v8, -4, v0.t\n\t"
+                    "vand.vx v4, v2, %[m]\n\t"
+                    "slli %[m], %[m], 1\n\t"
+                    "vmseq.vx v0, v4, zero\n\t"
+                    "vadd.vi v10, v10, -4, v0.t\n\t"
+                    "vand.vx v4, v2, %[m]\n\t"
+                    "slli %[m], %[m], 1\n\t"
+                    "vmseq.vx v0, v4, zero\n\t"
+                    "vadd.vi v12, v12, -4, v0.t\n\t"
+                    "vand.vx v4, v2, %[m]\n\t"
+                    "slli %[m], %[m], 1\n\t"
+                    "vmseq.vx v0, v4, zero\n\t"
+                    "vadd.vi v14, v14, -4, v0.t\n\t"
+                    "vsetvli zero, %[vl128], e8, m8\n\t"
+                    "vle8.v v0, (%[q8])\n\t"
+                    "vsetvli zero, %[vl64], e8, m4\n\t"
+                    "vwmul.vv v16, v0, v8\n\t"
+                    "vwmul.vv v24, v4, v12\n\t"
+                    "vsetivli zero, 16, e16, m2\n\t"
+                    "vmv.v.x v0, zero\n\t"
+                    "vwredsum.vs v10, v16, v0\n\t"
+                    "vwredsum.vs v9, v18, v0\n\t"
+                    "vwredsum.vs v8, v20, v0\n\t"
+                    "vwredsum.vs v7, v22, v0\n\t"
+                    "vwredsum.vs v11, v24, v0\n\t"
+                    "vwredsum.vs v12, v26, v0\n\t"
+                    "vwredsum.vs v13, v28, v0\n\t"
+                    "vwredsum.vs v14, v30, v0\n\t"
+                    "vsetivli zero, 4, e32, m1\n\t"
+                    "vslideup.vi v10, v9, 1\n\t"
+                    "vslideup.vi v8, v7, 1\n\t"
+                    "vslideup.vi v11, v12, 1\n\t"
+                    "vslideup.vi v13, v14, 1\n\t"
+                    "vslideup.vi v10, v8, 2\n\t"
+                    "vslideup.vi v11, v13, 2\n\t"
+                    "vsetivli zero, 8, e32, m2\n\t"
+                    "vle8.v v15, (%[scale])\n\t"
+                    "vsext.vf4 v12, v15\n\t"
+                    "vmul.vv v10, v10, v12\n\t"
+                    "vredsum.vs v0, v10, v0\n\t"
+                    "vmv.x.s %[tmp], v0\n\t"
+                    "add %[isum], %[isum], %[tmp]"
+                    : [tmp] "=&r" (tmp), [m] "+&r" (m), [isum] "+&r" (isum)
+                    : [vl128] "r" (128), [vl64] "r" (64), [vl32] "r" (32)
+                    , [q3] "r" (q3), [qh] "r" (qh), [scale] "r" (scale), [q8] "r" (q8)
+                    : "memory"
+                    , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+                    , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
+                    , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
+                    , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
+                );
+                q3 += 32;    q8 += 128;   scale += 8;
+            }
+
+            const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+            sumf += d * isum;
+        }
+        break;
+    default:
+        assert(false && "Unsupported vector length");
+        break;
+    }
+
+    *s = sumf;
+
+#else
+    // scalar version
+    // This function is written like this so the compiler can manage to vectorize most of it
+    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
+    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
+    // The ideal situation would be if we could just write the code once, and the compiler would
+    // automatically produce the best possible set of machine instructions, instead of us having to manually
+    // write vectorized versions for AVX, ARM_NEON, etc.
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    uint32_t auxs[4];
+    const int8_t * scales = (const int8_t*)auxs;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
+        const uint8_t * GGML_RESTRICT hm = x[i].hmask;
+        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * GGML_RESTRICT a = aux8;
+        uint8_t m = 1;
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            q3 += 32;
+        }
+        a = aux8;
+
+        memcpy(auxs, x[i].scales, 12);
+        uint32_t tmp = auxs[2];
+        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
+        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
+        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
+        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
+        for (int j = 0; j < QK_K/16; ++j) {
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+
+#endif
+
+}
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#if defined __riscv_xtheadvector
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        int tmp, tmp2, sumi;
+        __asm__ __volatile__(
+            "li %[t1], 12\n\t"
+            "th.vsetvli zero, %[t1], e8, m1\n\t"
+            "th.vlb.v v1, (%[s6b])\n\t" // {aux[0], aux[1], aux[2]}
+            "li %[t1], 4\n\t"
+            "th.vsetvli zero, %[t1], e32, m1\n\t"
+            "th.vslidedown.vi v2, v1, 2\n\t"
+            "th.vmv.v.v v3, v2\n\t"
+            "th.vslideup.vi v2, v3, 1\n\t" // {aux[2], aux[2]}
+            "li %[t1], 2\n\t"
+            "th.vsetvli zero, %[t1], e32, m1\n\t"
+            "th.vmv.v.i v4, 4\n\t"
+            "th.vand.vx v8, v1, %[kmask1]\n\t"
+            "th.vslide1up.vx v5, v4, zero\n\t" // {0, 4}
+            "th.vsrl.vi v6, v1, 6\n\t"
+            "th.vsrl.vv v7, v2, v5\n\t"
+            "th.vand.vx v0, v6, %[kmask3]\n\t"
+            "th.vand.vx v2, v7, %[kmask2]\n\t"
+            "th.vsll.vi v6, v0, 4\n\t"
+            "li %[t2], 8\n\t"
+            "addi %[t1], %[utmp], 4\n\t"
+            "th.vor.vv v1, v6, v2\n\t"
+            "th.vssw.v v8, (%[utmp]), %[t2]\n\t"
+            "th.vssw.v v1, (%[t1]), %[t2]\n\t"
+            "th.vsetvli zero, zero, e32, m2\n\t" // vl == 8
+            "th.vlw.v v2, (%[bsums])\n\t"
+            "th.vsetvli zero, %[t2], e16, m1\n\t"
+            "th.vnsrl.vi v0, v2, 0\n\t"
+            "th.vnsrl.vi v1, v2, 16\n\t"
+            "th.vadd.vv v2, v0, v1\n\t"
+            "th.vlbu.v v4, (%[mins])\n\t"
+            "th.vwmul.vv v6, v4, v2\n\t"
+            "th.vmv.v.x v0, zero\n\t"
+            "th.vsetvli zero, %[t2], e32, m2\n\t"
+            "th.vredsum.vs v0, v6, v0\n\t"
+            "th.vmv.x.s %[sumi], v0"
+            : [t1] "=&r" (tmp), [t2] "=&r" (tmp2), [sumi] "=&r" (sumi)
+            : [bsums] "r" (y[i].bsums), [mins] "r" (mins), [utmp] "r" (utmp)
+            , [s6b] "r" (x[i].scales), [kmask1] "r" (kmask1)
+            , [kmask2] "r" (kmask2), [kmask3] "r" (kmask3)
+            : "memory"
+            , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+            , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
+            , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
+            , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
+        );
+        sumf -= dmin * sumi;
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        sumi = 0;
+        const uint8_t * scale = scales;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            int vl128 = 128, vl64 = 64, vl32 = 32;
+            __asm__ __volatile__(
+                "th.vsetvli zero, %[vl128], e8, m8\n\t"
+                "th.vlb.v v8, (%[q8])\n\t"
+                "th.vsetvli zero, %[vl64], e8, m4\n\t"
+                "th.vlb.v v0, (%[q4])\n\t"
+                "th.vsrl.vi v4, v0, 4\n\t"
+                "th.vand.vi v0, v0, 0xF\n\t"
+                "th.vsetvli zero, %[vl32], e8, m2\n\t"
+                "th.vwmul.vv v28, v6, v14\n\t"
+                "th.vwmul.vv v20, v4, v10\n\t"
+                "th.vwmul.vv v24, v2, v12\n\t"
+                "th.vwmul.vv v16, v0, v8\n\t"
+                "li %[tmp], 4\n\t"
+                "th.vsetvli zero, %[tmp], e32, m1\n\t"
+                "th.vlbu.v v1, (%[scale])\n\t"
+                "th.vmv.v.x v0, zero\n\t"
+                "th.vsetvli zero, %[vl32], e16, m4\n\t"
+                "th.vwredsum.vs v6, v24, v0\n\t"
+                "th.vwredsum.vs v7, v28, v0\n\t"
+                "th.vwredsum.vs v4, v16, v0\n\t"
+                "th.vwredsum.vs v5, v20, v0\n\t"
+                "th.vsetvli zero, %[tmp], e32, m1\n\t"
+                "th.vslideup.vi v6, v7, 1\n\t"
+                "th.vslideup.vi v4, v5, 1\n\t"
+                "th.vslideup.vi v4, v6, 2\n\t"
+                "th.vmul.vv v8, v4, v1\n\t"
+                "th.vredsum.vs v0, v8, v0\n\t"
+                "th.vmv.x.s %[tmp], v0\n\t"
+                "add %[sumi], %[sumi], %[tmp]"
+                : [tmp] "=&r" (tmp), [sumi] "+&r" (sumi)
+                : [vl128] "r" (vl128), [vl64] "r" (vl64), [vl32] "r" (vl32)
+                , [q4] "r" (q4), [q8] "r" (q8), [scale] "r" (scale)
+                : "memory"
+                , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+                , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
+                , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
+                , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
+            );
+
+            q4 += 64;    q8 += 128;    scale += 4;
+        }
+
+        sumf += d * sumi;
+
+    }
+
+    *s = sumf;
+
+#elif defined __riscv_v
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    float sumf = 0;
+    const int vector_length = __riscv_vlenb() * 8;
+
+    switch (vector_length) {
+    case 256:
+        for (int i = 0; i < nb; ++i) {
+
+            size_t vl = 8;
+
+            const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+            const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+            vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl);
+            vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl);
+            vint16mf2_t q8sums   = __riscv_vadd_vv_i16mf2(q8sums_0, q8sums_1, vl);
+
+            memcpy(utmp, x[i].scales, 12);
+            utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+            const uint32_t uaux = utmp[1] & kmask1;
+            utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+            utmp[2] = uaux;
+            utmp[0] &= kmask1;
+
+            vuint8mf4_t mins8  = __riscv_vle8_v_u8mf4(mins, vl);
+            vint16mf2_t v_mins = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vzext_vf2_u16mf2(mins8, vl));
+            vint32m1_t  prod   = __riscv_vwmul_vv_i32m1(q8sums, v_mins, vl);
+
+            vint32m1_t sumi = __riscv_vredsum_vs_i32m1_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
+            sumf -= dmin * __riscv_vmv_x_s_i32m1_i32(sumi);
+
+            const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+            const int8_t  * GGML_RESTRICT q8 = y[i].qs;
+
+            vl = 32;
+
+            int32_t sum_1 = 0;
+            int32_t sum_2 = 0;
+
+            vint16m1_t vzero = __riscv_vmv_v_x_i16m1(0, 1);
+
+            for (int j = 0; j < QK_K/64; ++j) {
+                // load Q4
+                vuint8m1_t q4_x = __riscv_vle8_v_u8m1(q4, vl);
+
+                // load Q8 and multiply it with lower Q4 nibble
+                vint8m1_t  q8_0 = __riscv_vle8_v_i8m1(q8, vl);
+                vint8m1_t  q4_0 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q4_x, 0x0F, vl));
+                vint16m2_t qv_0 = __riscv_vwmul_vv_i16m2(q4_0, q8_0, vl);
+                vint16m1_t vs_0 = __riscv_vredsum_vs_i16m2_i16m1(qv_0, vzero, vl);
+
+                sum_1 += __riscv_vmv_x_s_i16m1_i16(vs_0) * scales[2*j+0];
+
+                // load Q8 and multiply it with upper Q4 nibble
+                vint8m1_t  q8_1 = __riscv_vle8_v_i8m1(q8+32, vl);
+                vint8m1_t  q4_1 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q4_x, 0x04, vl));
+                vint16m2_t qv_1 = __riscv_vwmul_vv_i16m2(q4_1, q8_1, vl);
+                vint16m1_t vs_1 = __riscv_vredsum_vs_i16m2_i16m1(qv_1, vzero, vl);
+
+                sum_2 += __riscv_vmv_x_s_i16m1_i16(vs_1) * scales[2*j+1];
+
+                q4 += 32;    q8 += 64;
+
+            }
+
+            sumf += d*(sum_1 + sum_2);
+
+        }
+        break;
+    case 128:
+        for (int i = 0; i < nb; ++i) {
+            const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+            const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+            int tmp, tmp2, sumi;
+            __asm__ __volatile__(
+                "vsetivli zero, 12, e8, m1\n\t"
+                "vle8.v v1, (%[s6b])\n\t" // {aux[0], aux[1], aux[2]}
+                "vsetivli zero, 4, e32, m1\n\t"
+                "vslidedown.vi v2, v1, 2\n\t"
+                "vmv1r.v v3, v2\n\t"
+                "vslideup.vi v2, v3, 1\n\t" // {aux[2], aux[2]}
+                "vsetivli zero, 2, e32, m1\n\t"
+                "vmv.v.i v4, 4\n\t"
+                "vand.vx v8, v1, %[kmask1]\n\t"
+                "vslide1up.vx v5, v4, zero\n\t" // {0, 4}
+                "vsrl.vi v6, v1, 6\n\t"
+                "vsrl.vv v7, v2, v5\n\t"
+                "vand.vx v0, v6, %[kmask3]\n\t"
+                "vand.vx v2, v7, %[kmask2]\n\t"
+                "vsll.vi v6, v0, 4\n\t"
+                "li %[t2], 8\n\t"
+                "addi %[t1], %[utmp], 4\n\t"
+                "vor.vv v1, v6, v2\n\t"
+                "vsse32.v v8, (%[utmp]), %[t2]\n\t"
+                "vsse32.v v1, (%[t1]), %[t2]\n\t"
+                "vsetivli zero, 8, e16, m1\n\t"
+                "vle32.v v2, (%[bsums])\n\t"
+                "vnsrl.wi v0, v2, 0\n\t"
+                "vnsrl.wi v1, v2, 16\n\t"
+                "vadd.vv v2, v0, v1\n\t"
+                "vle8.v v3, (%[mins])\n\t"
+                "vzext.vf2 v4, v3\n\t"
+                "vwmul.vv v6, v4, v2\n\t"
+                "vmv.v.x v0, zero\n\t"
+                "vsetivli zero, 8, e32, m2\n\t"
+                "vredsum.vs v0, v6, v0\n\t"
+                "vmv.x.s %[sumi], v0"
+                : [t1] "=&r" (tmp), [t2] "=&r" (tmp2), [sumi] "=&r" (sumi)
+                : [bsums] "r" (y[i].bsums), [mins] "r" (mins), [utmp] "r" (utmp)
+                , [s6b] "r" (x[i].scales), [kmask1] "r" (kmask1)
+                , [kmask2] "r" (kmask2), [kmask3] "r" (kmask3)
+                : "memory"
+                , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+                , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
+                , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
+                , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
+            );
+            sumf -= dmin * sumi;
+
+            const uint8_t * restrict q4 = x[i].qs;
+            const int8_t  * restrict q8 = y[i].qs;
+
+            sumi = 0;
+            const uint8_t * scale = scales;
+
+            for (int j = 0; j < QK_K/128; ++j) {
+                int vl128 = 128, vl64 = 64, vl32 = 32;
+                __asm__ __volatile__(
+                    "vsetvli zero, %[vl128], e8, m8\n\t"
+                    "vle8.v v8, (%[q8])\n\t"
+                    "vsetvli zero, %[vl64], e8, m4\n\t"
+                    "vle8.v v0, (%[q4])\n\t"
+                    "vsrl.vi v4, v0, 4\n\t"
+                    "vand.vi v0, v0, 0xF\n\t"
+                    "vsetvli zero, %[vl32], e8, m2\n\t"
+                    "vwmul.vv v28, v6, v14\n\t"
+                    "vwmul.vv v20, v4, v10\n\t"
+                    "vwmul.vv v24, v2, v12\n\t"
+                    "vwmul.vv v16, v0, v8\n\t"
+                    "vsetivli zero, 4, e32, m1\n\t"
+                    "vle8.v v2, (%[scale])\n\t"
+                    "vmv.v.x v0, zero\n\t"
+                    "vzext.vf4 v1, v2\n\t"
+                    "vsetvli zero, %[vl32], e16, m4\n\t"
+                    "vwredsum.vs v6, v24, v0\n\t"
+                    "vwredsum.vs v7, v28, v0\n\t"
+                    "vwredsum.vs v4, v16, v0\n\t"
+                    "vwredsum.vs v5, v20, v0\n\t"
+                    "vsetivli zero, 4, e32, m1\n\t"
+                    "vslideup.vi v6, v7, 1\n\t"
+                    "vslideup.vi v4, v5, 1\n\t"
+                    "vslideup.vi v4, v6, 2\n\t"
+                    "vmul.vv v8, v4, v1\n\t"
+                    "vredsum.vs v0, v8, v0\n\t"
+                    "vmv.x.s %[tmp], v0\n\t"
+                    "add %[sumi], %[sumi], %[tmp]"
+                    : [tmp] "=&r" (tmp), [sumi] "+&r" (sumi)
+                    : [vl128] "r" (vl128), [vl64] "r" (vl64), [vl32] "r" (vl32)
+                    , [q4] "r" (q4), [q8] "r" (q8), [scale] "r" (scale)
+                    : "memory"
+                    , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+                    , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
+                    , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
+                    , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
+                );
+
+                q4 += 64;    q8 += 128;    scale += 4;
+            }
+
+            sumf += d * sumi;
+        }
+        break;
+    default:
+        assert(false && "Unsupported vector length");
+        break;
+    }
+
+    *s = sumf;
+
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * GGML_RESTRICT a = aux8;
+        for (int j = 0; j < QK_K/64; ++j) {
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
+            a += 32;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            a += 32; q4 += 32;
+        }
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#if defined __riscv_v
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    float sumf = 0;
+    float sums = 0.0;
+
+    size_t vl;
+
+    for (int i = 0; i < nb; ++i) {
+
+        vl = 8;
+
+        const uint8_t * GGML_RESTRICT q5 = x[i].qs;
+        const uint8_t * GGML_RESTRICT hm = x[i].qh;
+        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
+
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+
+        vint16m1_t q8sums_0 = __riscv_vlse16_v_i16m1(y[i].bsums, 4, vl);
+        vint16m1_t q8sums_1 = __riscv_vlse16_v_i16m1(y[i].bsums+1, 4, vl);
+        vint16m1_t q8sums = __riscv_vadd_vv_i16m1(q8sums_0, q8sums_1, vl);
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        vuint8mf2_t mins8 = __riscv_vle8_v_u8mf2(mins, vl);
+        vint16m1_t v_mins = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(mins8, vl));
+        vint32m2_t prod = __riscv_vwmul_vv_i32m2(q8sums, v_mins, vl);
+
+        vint32m1_t sumi = __riscv_vredsum_vs_i32m2_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
+        sumf -= dmin * __riscv_vmv_x_s_i32m1_i32(sumi);
+
+        vl = 32;
+        int32_t aux32 = 0;
+        int is = 0;
+
+        uint8_t m = 1;
+        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+        vuint8m2_t vqh = __riscv_vle8_v_u8m2(hm, vl);
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            // load Q5 and Q8
+            vuint8m2_t q5_x = __riscv_vle8_v_u8m2(q5, vl);
+            vint8m2_t  q8_y1 = __riscv_vle8_v_i8m2(q8, vl);
+            vint8m2_t  q8_y2 = __riscv_vle8_v_i8m2(q8+32, vl);
+
+            // compute mask for addition
+            vint8m2_t q5_a = __riscv_vreinterpret_v_u8m2_i8m2(__riscv_vand_vx_u8m2(q5_x, 0x0F, vl));
+            vuint8m2_t qh_m1 = __riscv_vand_vx_u8m2(vqh, m, vl);
+            vbool4_t vmask_1 = __riscv_vmsne_vx_u8m2_b4(qh_m1, 0, vl);
+            vint8m2_t q5_m1 = __riscv_vadd_vx_i8m2_mu(vmask_1, q5_a, q5_a, 16, vl);
+            m <<= 1;
+
+            vint8m2_t q5_l = __riscv_vreinterpret_v_u8m2_i8m2(__riscv_vsrl_vx_u8m2(q5_x, 0x04, vl));
+            vuint8m2_t qh_m2 = __riscv_vand_vx_u8m2(vqh, m, vl);
+            vbool4_t vmask_2 = __riscv_vmsne_vx_u8m2_b4(qh_m2, 0, vl);
+            vint8m2_t q5_m2 = __riscv_vadd_vx_i8m2_mu(vmask_2, q5_l, q5_l, 16, vl);
+            m <<= 1;
+
+            vint16m4_t v0 = __riscv_vwmul_vv_i16m4(q5_m1, q8_y1, vl);
+            vint16m4_t v1 = __riscv_vwmul_vv_i16m4(q5_m2, q8_y2, vl);
+
+            vint32m8_t vs1 = __riscv_vwmul_vx_i32m8(v0, scales[is++], vl);
+            vint32m8_t vs2 = __riscv_vwmul_vx_i32m8(v1, scales[is++], vl);
+
+            vint32m1_t vacc1 = __riscv_vredsum_vs_i32m8_i32m1(vs1, vzero, vl);
+            vint32m1_t vacc2 = __riscv_vredsum_vs_i32m8_i32m1(vs2, vacc1, vl);
+
+            aux32 += __riscv_vmv_x_s_i32m1_i32(vacc2);
+            q5 += 32;    q8 += 64;
+
+        }
+
+        sums += aux32 * d;
+
+    }
+
+    *s = sumf+sums;
+
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+        const uint8_t * GGML_RESTRICT hm = x[i].qh;
+        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * GGML_RESTRICT a = aux8;
+        uint8_t m = 1;
+        for (int j = 0; j < QK_K/64; ++j) {
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            q4 += 32;
+        }
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q6_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __riscv_xtheadvector
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+
+        const uint8_t * restrict q6 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const  int8_t * restrict q8 = y[i].qs;
+
+        const int8_t * restrict scale = x[i].scales;
+
+        int sum_t = 0;
+        int t0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            __asm__ __volatile__(
+                "th.vsetvli zero, %[vl32], e8, m2\n\t" // vl == 32
+                "th.vlb.v v4, (%[qh])\n\t"
+                "th.vsll.vi v0, v4, 4\n\t"
+                "th.vsll.vi v2, v4, 2\n\t"
+                "th.vsrl.vi v6, v4, 2\n\t"
+                "th.vsetvli zero, %[vl64], e8, m4\n\t" // vl == 64
+                "th.vlb.v v8, (%[q6])\n\t"
+                "th.vsrl.vi v12, v8, 4\n\t"
+                "th.vand.vi v8, v8, 0xF\n\t"
+                "th.vsetvli zero, %[vl128], e8, m8\n\t" // vl == 128
+                "th.vand.vx v0, v0, %[mask]\n\t"
+                "th.vor.vv v8, v8, v0\n\t"
+                "th.vlb.v v0, (%[q8])\n\t"
+                "th.vsub.vx v8, v8, %[vl32]\n\t"
+                "th.vsetvli zero, %[vl64], e8, m4\n\t" // vl == 64
+                "th.vwmul.vv v16, v0, v8\n\t"
+                "th.vwmul.vv v24, v4, v12\n\t"
+                "li %[t0], 16\n\t"
+                "th.vsetvli zero, %[t0], e16, m2\n\t" // vl == 16
+                "th.vmv.v.x v0, zero\n\t"
+                "th.vwredsum.vs v10, v16, v0\n\t"
+                "th.vwredsum.vs v9, v18, v0\n\t"
+                "th.vwredsum.vs v8, v20, v0\n\t"
+                "th.vwredsum.vs v7, v22, v0\n\t"
+                "th.vwredsum.vs v11, v24, v0\n\t"
+                "th.vwredsum.vs v12, v26, v0\n\t"
+                "th.vwredsum.vs v13, v28, v0\n\t"
+                "th.vwredsum.vs v14, v30, v0\n\t"
+                "li %[t0], 4\n\t"
+                "th.vsetvli zero, %[t0], e32, m1\n\t" // vl == 4
+                "th.vslideup.vi v10, v9, 1\n\t"
+                "th.vslideup.vi v8, v7, 1\n\t"
+                "th.vslideup.vi v11, v12, 1\n\t"
+                "th.vslideup.vi v13, v14, 1\n\t"
+                "th.vslideup.vi v10, v8, 2\n\t"
+                "th.vslideup.vi v11, v13, 2\n\t"
+                "li %[t0], 8\n\t"
+                "th.vsetvli zero, %[t0], e32, m2\n\t" // vl == 8
+                "th.vlb.v v4, (%[scale])\n\t"
+                "th.vmul.vv v2, v4, v10\n\t"
+                "th.vredsum.vs v0, v2, v0\n\t"
+                "th.vmv.x.s %[t0], v0\n\t"
+                "add %[sumi], %[sumi], %[t0]"
+                : [sumi] "+&r" (sum_t), [t0] "=&r" (t0)
+                : [qh] "r" (qh), [q6] "r" (q6), [q8] "r" (q8), [scale] "r" (scale)
+                , [vl32] "r" (32), [vl64] "r" (64), [vl128] "r" (128)
+                , [mask] "r" (0x30)
+                : "memory"
+                , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+                , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
+                , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
+                , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
+            );
+            q6 += 64;   qh += 32;   q8 += 128;   scale += 8;
+        }
+
+        sumf += d * sum_t;
+
+    }
+
+    *s = sumf;
+
+#elif defined __riscv_v
+
+    float sumf = 0;
+    const int vector_length = __riscv_vlenb() * 8;
+
+    switch (vector_length) {
+    case 256:
+        for (int i = 0; i < nb; ++i) {
+
+            const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+
+            const uint8_t * GGML_RESTRICT q6 = x[i].ql;
+            const uint8_t * GGML_RESTRICT qh = x[i].qh;
+            const  int8_t * GGML_RESTRICT q8 = y[i].qs;
+
+            const int8_t * GGML_RESTRICT scale = x[i].scales;
+
+            size_t vl;
+
+            vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+
+            int sum_t = 0;
+            int is = 0;
+
+            for (int j = 0; j < QK_K/128; ++j) {
+
+                vl = 32;
+
+                // load qh
+                vuint8m1_t qh_x = __riscv_vle8_v_u8m1(qh, vl);
+
+                // load Q6
+                vuint8m1_t q6_0 = __riscv_vle8_v_u8m1(q6, vl);
+                vuint8m1_t q6_1 = __riscv_vle8_v_u8m1(q6+32, vl);
+
+                vuint8m1_t q6a_0 = __riscv_vand_vx_u8m1(q6_0, 0x0F, vl);
+                vuint8m1_t q6a_1 = __riscv_vand_vx_u8m1(q6_1, 0x0F, vl);
+                vuint8m1_t q6s_0 = __riscv_vsrl_vx_u8m1(q6_0, 0x04, vl);
+                vuint8m1_t q6s_1 = __riscv_vsrl_vx_u8m1(q6_1, 0x04, vl);
+
+                vuint8m1_t qh_0 = __riscv_vand_vx_u8m1(qh_x, 0x03, vl);
+                vuint8m1_t qh_1 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x2, vl), 0x03 , vl);
+                vuint8m1_t qh_2 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x4, vl), 0x03 , vl);
+                vuint8m1_t qh_3 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x6, vl), 0x03 , vl);
+
+                vuint8m1_t qhi_0 = __riscv_vor_vv_u8m1(q6a_0, __riscv_vsll_vx_u8m1(qh_0, 0x04, vl), vl);
+                vuint8m1_t qhi_1 = __riscv_vor_vv_u8m1(q6a_1, __riscv_vsll_vx_u8m1(qh_1, 0x04, vl), vl);
+                vuint8m1_t qhi_2 = __riscv_vor_vv_u8m1(q6s_0, __riscv_vsll_vx_u8m1(qh_2, 0x04, vl), vl);
+                vuint8m1_t qhi_3 = __riscv_vor_vv_u8m1(q6s_1, __riscv_vsll_vx_u8m1(qh_3, 0x04, vl), vl);
+
+                vint8m1_t a_0 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_0), 32, vl);
+                vint8m1_t a_1 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_1), 32, vl);
+                vint8m1_t a_2 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_2), 32, vl);
+                vint8m1_t a_3 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_3), 32, vl);
+
+                // load Q8 and take product
+                vint16m2_t va_q_0 = __riscv_vwmul_vv_i16m2(a_0, __riscv_vle8_v_i8m1(q8, vl), vl);
+                vint16m2_t va_q_1 = __riscv_vwmul_vv_i16m2(a_1, __riscv_vle8_v_i8m1(q8+32, vl), vl);
+                vint16m2_t va_q_2 = __riscv_vwmul_vv_i16m2(a_2, __riscv_vle8_v_i8m1(q8+64, vl), vl);
+                vint16m2_t va_q_3 = __riscv_vwmul_vv_i16m2(a_3, __riscv_vle8_v_i8m1(q8+96, vl), vl);
+
+                vl = 16;
+
+                vint32m2_t vaux_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_0, 0), scale[is+0], vl);
+                vint32m2_t vaux_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_0, 1), scale[is+1], vl);
+                vint32m2_t vaux_2 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_1, 0), scale[is+2], vl);
+                vint32m2_t vaux_3 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_1, 1), scale[is+3], vl);
+                vint32m2_t vaux_4 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_2, 0), scale[is+4], vl);
+                vint32m2_t vaux_5 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_2, 1), scale[is+5], vl);
+                vint32m2_t vaux_6 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_3, 0), scale[is+6], vl);
+                vint32m2_t vaux_7 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_3, 1), scale[is+7], vl);
+
+                vint32m1_t isum0 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_0, vaux_1, vl), vzero, vl);
+                vint32m1_t isum1 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_2, vaux_3, vl), isum0, vl);
+                vint32m1_t isum2 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_4, vaux_5, vl), isum1, vl);
+                vint32m1_t isum3 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_6, vaux_7, vl), isum2, vl);
+
+                sum_t += __riscv_vmv_x_s_i32m1_i32(isum3);
+
+                q6 += 64;   qh += 32;   q8 += 128;   is=8;
+
+            }
+
+            sumf += d * sum_t;
+
+        }
+        break;
+    case 128:
+        for (int i = 0; i < nb; ++i) {
+
+            const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+
+            const uint8_t * restrict q6 = x[i].ql;
+            const uint8_t * restrict qh = x[i].qh;
+            const  int8_t * restrict q8 = y[i].qs;
+
+            const int8_t * restrict scale = x[i].scales;
+
+            int sum_t = 0;
+            int t0;
+
+            for (int j = 0; j < QK_K/128; ++j) {
+                __asm__ __volatile__(
+                    "vsetvli zero, %[vl32], e8, m2\n\t"
+                    "vle8.v v4, (%[qh])\n\t"
+                    "vsll.vi v0, v4, 4\n\t"
+                    "vsll.vi v2, v4, 2\n\t"
+                    "vsrl.vi v6, v4, 2\n\t"
+                    "vsetvli zero, %[vl64], e8, m4\n\t"
+                    "vle8.v v8, (%[q6])\n\t"
+                    "vsrl.vi v12, v8, 4\n\t"
+                    "vand.vi v8, v8, 0xF\n\t"
+                    "vsetvli zero, %[vl128], e8, m8\n\t"
+                    "vand.vx v0, v0, %[mask]\n\t"
+                    "vor.vv v8, v8, v0\n\t"
+                    "vle8.v v0, (%[q8])\n\t"
+                    "vsub.vx v8, v8, %[vl32]\n\t"
+                    "vsetvli zero, %[vl64], e8, m4\n\t"
+                    "vwmul.vv v16, v0, v8\n\t"
+                    "vwmul.vv v24, v4, v12\n\t"
+                    "vsetivli zero, 16, e16, m2\n\t"
+                    "vmv.v.x v0, zero\n\t"
+                    "vwredsum.vs v10, v16, v0\n\t"
+                    "vwredsum.vs v9, v18, v0\n\t"
+                    "vwredsum.vs v8, v20, v0\n\t"
+                    "vwredsum.vs v7, v22, v0\n\t"
+                    "vwredsum.vs v11, v24, v0\n\t"
+                    "vwredsum.vs v12, v26, v0\n\t"
+                    "vwredsum.vs v13, v28, v0\n\t"
+                    "vwredsum.vs v14, v30, v0\n\t"
+                    "vsetivli zero, 4, e32, m1\n\t"
+                    "vslideup.vi v10, v9, 1\n\t"
+                    "vslideup.vi v8, v7, 1\n\t"
+                    "vslideup.vi v11, v12, 1\n\t"
+                    "vslideup.vi v13, v14, 1\n\t"
+                    "vslideup.vi v10, v8, 2\n\t"
+                    "vslideup.vi v11, v13, 2\n\t"
+                    "vsetivli zero, 8, e32, m2\n\t"
+                    "vle8.v v2, (%[scale])\n\t"
+                    "vsext.vf4 v4, v2\n\t"
+                    "vmul.vv v2, v4, v10\n\t"
+                    "vredsum.vs v0, v2, v0\n\t"
+                    "vmv.x.s %[t0], v0\n\t"
+                    "add %[sumi], %[sumi], %[t0]"
+                    : [sumi] "+&r" (sum_t), [t0] "=&r" (t0)
+                    : [qh] "r" (qh), [q6] "r" (q6), [q8] "r" (q8), [scale] "r" (scale)
+                    , [vl32] "r" (32), [vl64] "r" (64), [vl128] "r" (128)
+                    , [mask] "r" (0x30)
+                    : "memory"
+                    , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+                    , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
+                    , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
+                    , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
+                );
+                q6 += 64;   qh += 32;   q8 += 128;   scale += 8;
+            }
+
+            sumf += d * sum_t;
+
+        }
+        break;
+    default:
+        assert(false && "Unsupported vector length");
+        break;
+    }
+
+    *s = sumf;
+
+#else
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * GGML_RESTRICT a = aux8;
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) {
+                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
+                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
+                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
+                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
+            }
+            a  += 128;
+            q4 += 64;
+            qh += 32;
+        }
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/16; ++j) {
+            int scale = x[i].scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+

ggml/src/ggml-cpu/arch/riscv/repack.cpp

@@ -0,0 +1,396 @@
+#define GGML_COMMON_IMPL_CPP
+#define GGML_COMMON_DECL_CPP
+#include "ggml-common.h"
+#include "ggml-backend-impl.h"
+
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+#include "ggml-cpu-impl.h"
+#include "traits.h"
+
+#include <cmath>
+#include <cstring>
+#include <cassert>
+#include <cstdlib> // for qsort
+#include <cstdio>  // for GGML_ASSERT
+
+#define GGML_CPU_CLANG_WORKAROUND
+#include "../../repack.h"
+
+#if defined(__GNUC__)
+#pragma GCC diagnostic ignored "-Woverlength-strings"
+#endif
+
+#define UNUSED GGML_UNUSED
+
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined __riscv_v
+    if (__riscv_vlenb() >= QK4_0) {
+        const size_t vl = QK4_0;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+
+            vfloat32m1_t sumf = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+            for (int l = 0; l < nb; l++) {
+                const int64_t a0 = *(const int64_t *)&a_ptr[l].qs[0];
+                const int64_t a1 = *(const int64_t *)&a_ptr[l].qs[8];
+                const int64_t a2 = *(const int64_t *)&a_ptr[l].qs[16];
+                const int64_t a3 = *(const int64_t *)&a_ptr[l].qs[24];
+                __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment constraints
+                const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a0, vl / 4));
+                const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a1, vl / 4));
+                const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a2, vl / 4));
+                const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a3, vl / 4));
+
+                const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
+                const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
+                const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
+                const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
+                const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
+                const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
+                const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
+
+                const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_hi_m));
+                const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                // vector version needs Zvfhmin extension
+                const float a_scale = GGML_FP16_TO_FP32(a_ptr[l].d);
+                const float b_scales[8] = {
+                    GGML_FP16_TO_FP32(b_ptr[l].d[0]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[1]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[2]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[3]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[4]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[5]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[6]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[7])
+                };
+                const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
+                const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scale, vl / 4);
+                sumf = __riscv_vfmacc_vv_f32m1(sumf, tmp1, b_scales_vec, vl / 4);
+            }
+            __riscv_vse32_v_f32m1(s + x * ncols_interleaved, sumf, vl / 4);
+        }
+        return;
+    }
+
+#endif
+    {
+        float sumf[8];
+        int sumi;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+
+            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int j = 0; j < ncols_interleaved; j++) {
+                        sumi = 0;
+                        for (int i = 0; i < blocklen; ++i) {
+                            const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                        }
+                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                    }
+                }
+            }
+            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined __riscv_v
+    if (__riscv_vlenb() >= QK4_0) {
+        const size_t vl = QK4_0;
+
+        for (int y = 0; y < nr / 4; y++) {
+            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+                vfloat32m1_t sumf0 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+                vfloat32m1_t sumf1 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+                vfloat32m1_t sumf2 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+                vfloat32m1_t sumf3 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+                for (int l = 0; l < nb; l++) {
+                    const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
+                    const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
+                    const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
+                    const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
+                    const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
+                    const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
+                    const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
+
+                    // vector version needs Zvfhmin extension
+                    const float a_scales[4] = {
+                        GGML_FP16_TO_FP32(a_ptr[l].d[0]),
+                        GGML_FP16_TO_FP32(a_ptr[l].d[1]),
+                        GGML_FP16_TO_FP32(a_ptr[l].d[2]),
+                        GGML_FP16_TO_FP32(a_ptr[l].d[3])
+                    };
+                    const float b_scales[8] = {
+                        GGML_FP16_TO_FP32(b_ptr[l].d[0]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[1]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[2]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[3]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[4]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[5]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[6]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[7])
+                    };
+                    const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
+
+                    const int64_t A0 = *(const int64_t *)&a_ptr[l].qs[0];
+                    const int64_t A4 = *(const int64_t *)&a_ptr[l].qs[32];
+                    const int64_t A8 = *(const int64_t *)&a_ptr[l].qs[64];
+                    const int64_t Ac = *(const int64_t *)&a_ptr[l].qs[96];
+                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
+                    vint16m4_t sumi_l0;
+                    {
+                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A0, vl / 4));
+                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A4, vl / 4));
+                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A8, vl / 4));
+                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ac, vl / 4));
+                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                        sumi_l0 = sumi_hi_m;
+                    }
+
+                    {
+                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l0));
+                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[0], vl / 4);
+                        sumf0 = __riscv_vfmacc_vv_f32m1(sumf0, tmp1, b_scales_vec, vl / 4);
+                    }
+
+                    const int64_t A1 = *(const int64_t *)&a_ptr[l].qs[8];
+                    const int64_t A5 = *(const int64_t *)&a_ptr[l].qs[40];
+                    const int64_t A9 = *(const int64_t *)&a_ptr[l].qs[72];
+                    const int64_t Ad = *(const int64_t *)&a_ptr[l].qs[104];
+                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
+                    vint16m4_t sumi_l1;
+                    {
+                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A1, vl / 4));
+                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A5, vl / 4));
+                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A9, vl / 4));
+                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ad, vl / 4));
+                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                        sumi_l1 = sumi_hi_m;
+                    }
+
+                    {
+                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l1));
+                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[1], vl / 4);
+                        sumf1 = __riscv_vfmacc_vv_f32m1(sumf1, tmp1, b_scales_vec, vl / 4);
+                    }
+
+                    const int64_t A2 = *(const int64_t *)&a_ptr[l].qs[16];
+                    const int64_t A6 = *(const int64_t *)&a_ptr[l].qs[48];
+                    const int64_t Aa = *(const int64_t *)&a_ptr[l].qs[80];
+                    const int64_t Ae = *(const int64_t *)&a_ptr[l].qs[112];
+                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
+                    vint16m4_t sumi_l2;
+                    {
+                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A2, vl / 4));
+                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A6, vl / 4));
+                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Aa, vl / 4));
+                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ae, vl / 4));
+                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                        sumi_l2 = sumi_hi_m;
+                    }
+
+                    {
+                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l2));
+                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[2], vl / 4);
+                        sumf2 = __riscv_vfmacc_vv_f32m1(sumf2, tmp1, b_scales_vec, vl / 4);
+                    }
+
+                    const int64_t A3 = *(const int64_t *)&a_ptr[l].qs[24];
+                    const int64_t A7 = *(const int64_t *)&a_ptr[l].qs[56];
+                    const int64_t Ab = *(const int64_t *)&a_ptr[l].qs[88];
+                    const int64_t Af = *(const int64_t *)&a_ptr[l].qs[120];
+                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
+                    vint16m4_t sumi_l3;
+                    {
+                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A3, vl / 4));
+                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A7, vl / 4));
+                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ab, vl / 4));
+                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Af, vl / 4));
+                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                        sumi_l3 = sumi_hi_m;
+                    }
+
+                    {
+                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l3));
+                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[3], vl / 4);
+                        sumf3 = __riscv_vfmacc_vv_f32m1(sumf3, tmp1, b_scales_vec, vl / 4);
+                    }
+                }
+                __riscv_vse32_v_f32m1(&s[(y * 4 + 0) * bs + x * ncols_interleaved], sumf0, vl / 4);
+                __riscv_vse32_v_f32m1(&s[(y * 4 + 1) * bs + x * ncols_interleaved], sumf1, vl / 4);
+                __riscv_vse32_v_f32m1(&s[(y * 4 + 2) * bs + x * ncols_interleaved], sumf2, vl / 4);
+                __riscv_vse32_v_f32m1(&s[(y * 4 + 3) * bs + x * ncols_interleaved], sumf3, vl / 4);
+            }
+        }
+
+        return;
+    }
+
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
+    float sumf[4][8];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+}

ggml/src/ggml-cpu/arch/s390/quants.c

@@ -0,0 +1,1299 @@
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+#include "ggml-quants.h"
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+
+#include "../../quants.h"
+#include "../../ggml-cpu-impl.h"
+
+#include <math.h>
+#include <string.h>
+#include <assert.h>
+#include <float.h>
+#include <stdlib.h> // for qsort
+#include <stdio.h>  // for GGML_ASSERT
+
+#define GROUP_MAX_EPS 1e-15f
+#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
+#define GROUP_MAX_EPS_IQ2_S 1e-8f
+#define GROUP_MAX_EPS_IQ1_M 1e-7f
+#define GROUP_MAX_EPS_IQ1_S 1e-12f
+
+#define UNUSED GGML_UNUSED
+
+void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(QK8_0 == 32);
+    assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
+
+    block_q8_0 * GGML_RESTRICT y = vy;
+
+#if defined(__VXE__) || defined(__VXE2__)
+    for (int i = 0; i < nb; i++) {
+        __vector float srcv [8];
+        __vector float asrcv[8];
+        __vector float amaxv[8];
+
+        for (int j = 0; j < 8; j++) srcv[j] = vec_xl(0, x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = vec_abs(srcv[j]);
+        for (int j = 0; j < 4; j++) amaxv[2*j] = vec_max(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = vec_max(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = vec_max(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = MAX(MAX(vec_extract(amaxv[0], 0),
+                                   vec_extract(amaxv[0], 1)),
+                               MAX(vec_extract(amaxv[0], 2),
+                                   vec_extract(amaxv[0], 3)));
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f / d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        for (int j = 0; j < 8; j++) {
+            const __vector float v = vec_mul(srcv[j], vec_splats(id));
+            const __vector int32_t vi = vec_signed(v);
+
+            y[i].qs[4*j + 0] = vec_extract(vi, 0);
+            y[i].qs[4*j + 1] = vec_extract(vi, 1);
+            y[i].qs[4*j + 2] = vec_extract(vi, 2);
+            y[i].qs[4*j + 3] = vec_extract(vi, 3);
+        }
+    }
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_0_ref(x, y, k);
+#endif
+}
+
+void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(k % QK8_1 == 0);
+    const int nb = k / QK8_1;
+
+    block_q8_1 * GGML_RESTRICT y = vy;
+
+#if defined(__VXE__) || defined(__VXE2__)
+    for (int i = 0; i < nb; i++) {
+        __vector float srcv [8];
+        __vector float asrcv[8];
+        __vector float amaxv[8];
+
+        for (int j = 0; j < 8; j++) srcv[j] = vec_xl(0, x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = vec_abs(srcv[j]);
+        for (int j = 0; j < 4; j++) amaxv[2*j] = vec_max(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = vec_max(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = vec_max(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = MAX(MAX(vec_extract(amaxv[0], 0),
+                                   vec_extract(amaxv[0], 1)),
+                               MAX(vec_extract(amaxv[0], 2),
+                                   vec_extract(amaxv[0], 3)));
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f / d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        __vector int32_t acc = vec_splats(0);
+
+        for (int j = 0; j < 8; j++) {
+            const __vector float v = vec_mul(srcv[j], vec_splats(id));
+            const __vector int32_t vi = vec_signed(v);
+
+            y[i].qs[4*j + 0] = vec_extract(vi, 0);
+            y[i].qs[4*j + 1] = vec_extract(vi, 1);
+            y[i].qs[4*j + 2] = vec_extract(vi, 2);
+            y[i].qs[4*j + 3] = vec_extract(vi, 3);
+
+            acc = vec_add(acc, vi);
+        }
+
+        y[i].s = GGML_FP32_TO_FP16(d * (acc[0] + acc[1] + acc[2] + acc[3]));
+    }
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_1_ref(x, y, k);
+#endif
+}
+
+
+//===================================== Dot products =================================
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__VXE__) || defined(__VXE2__)
+    __vector float acc = vec_splats(0.0f);
+
+    const __vector uint8_t v_m = vec_splats((const uint8_t)0x0F);
+    const __vector int8_t  v_s = vec_splats( (const int8_t)0x08);
+
+    for (; ib < nb; ++ib) {
+        const __vector uint8_t v_x = vec_xl(0, x[ib].qs);
+        const __vector int8_t v_xl = (const __vector int8_t)(v_x & v_m);
+        const __vector int8_t v_xh = (const __vector int8_t)(v_x >> 4);
+
+        const __vector int8_t v_xls = vec_sub(v_xl, v_s);
+        const __vector int8_t v_xhs = vec_sub(v_xh, v_s);
+
+        const __vector int8_t v_yl = vec_xl(0      , y[ib].qs);
+        const __vector int8_t v_yh = vec_xl(QK8_0/2, y[ib].qs);
+
+        const __vector int16_t v_xylso = vec_mulo(v_xls, v_yl);
+        const __vector int16_t v_xylse = vec_mule(v_xls, v_yl);
+        const __vector int16_t v_xyhso = vec_mulo(v_xhs, v_yh);
+        const __vector int16_t v_xyhse = vec_mule(v_xhs, v_yh);
+
+        __vector int16_t v_xy_ = v_xylso + v_xylse + v_xyhso + v_xyhse; v_xy_ += vec_reve(v_xy_);
+
+        const __vector float v_xy = vec_float(vec_unpackh(v_xy_));
+        const __vector float v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+
+        acc = vec_madd(v_xy, v_d, acc);
+    }
+
+    sumf = acc[0] + acc[1] + acc[2] + acc[3];
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
+            const int v1 = (x[ib].qs[j] >>   4) - 8;
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+    }
+
+    *s = sumf;
+}
+
+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) {
+    const int qk = QK8_1;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__VXE__) || defined(__VXE2__)
+    float summs = 0;
+    float32x4_t acc = vec_splats(0.0f);
+
+    const uint8x16_t v_m = vec_splat_u8(0x0F);
+
+#pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+
+        const uint8x16_t v_x = vec_xl(0, x[ib].qs);
+        const int8x16_t v_xl = (const int8x16_t)(v_x & v_m);
+        const int8x16_t v_xh = (const int8x16_t)(v_x >> 4);
+
+        const int8x16_t v_yl = vec_xl(0      , y[ib].qs);
+        const int8x16_t v_yh = vec_xl(QK8_1/2, y[ib].qs);
+
+        const int32x4_t v_xy_ = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
+        const float32x4_t v_xy = vec_float(v_xy_);
+
+        const float32x4_t v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+
+        acc = vec_madd(v_xy, v_d, acc);
+    }
+
+    sumf = acc[0] + acc[1] + acc[2] + acc[3] + summs;
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F);
+            const int v1 = (x[ib].qs[j] >>   4);
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q8_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__VXE__) || defined(__VXE2__)
+    __vector float acc = vec_splats(0.0f);
+
+#pragma GCC unroll 8
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+        const int8x16_t v_xl = vec_xl(0      , x[ib].qs);
+        const int8x16_t v_xh = vec_xl(QK8_0/2, x[ib].qs);
+        const int8x16_t v_yl = vec_xl(0      , y[ib].qs);
+        const int8x16_t v_yh = vec_xl(QK8_0/2, y[ib].qs);
+
+        const int32x4_t v_xy_ = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
+        const float32x4_t v_xy = vec_float(v_xy_);
+        const float32x4_t v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+
+        acc = vec_madd(v_xy, v_d, acc);
+    }
+
+    sumf = acc[0] + acc[1] + acc[2] + acc[3];
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi = 0;
+
+        for (int j = 0; j < qk; j++) {
+            sumi += x[ib].qs[j]*y[ib].qs[j];
+        }
+
+        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+    }
+
+    *s = sumf;
+}
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const uint32_t kmask1 = 0x03030303;
+    const uint32_t kmask2 = 0x0f0f0f0f;
+
+    const block_q3_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__VXE__) || defined(__VXE2__)
+    uint32_t aux[3];
+    uint32_t utmp[4];
+
+    const int32x4_t v_z = vec_splat_s32(0);
+    const uint8x16_t v_3m = vec_splat_u8(0x03);
+
+    const uint8x16_t v_0c = vec_splat_u8(1);
+    const uint8x16_t v_1c = vec_sl(v_0c, 1);
+    const uint8x16_t v_2c = vec_sl(v_0c, 2);
+    const uint8x16_t v_3c = vec_sl(v_0c, 3);
+
+    uint8x16_t q3h[4];
+    uint8x16_t q3b[2];
+    int8x16_t q3bytes[4];
+    int8x16_t q8bytes[4];
+    uint8x16_t qhbits[2];
+
+    float sum = 0;
+
+    for (int i = 0; i < nb; ++i) {
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict x0l = x[i].qs;
+        const uint8_t * restrict x0h = x[i].hmask;
+        const int8_t  * restrict y0  = y[i].qs;
+
+        qhbits[0] = vec_xl(0 , x0h);
+        qhbits[1] = vec_xl(16, x0h);
+
+        int32_t isum = 0;
+
+        memcpy(aux, x[i].scales, 12);
+        utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
+        utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
+        utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
+        utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
+
+        int8_t * scale = (int8_t *)utmp;
+        for (int j = 0; j < 16; ++j) scale[j] -= 32;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            int32x4_t isum0, isum1, isum2, isum3;
+
+            q3b[0] = vec_xl(0 , x0l);
+            q3b[1] = vec_xl(16, x0l);
+            x0l += 32;
+
+            q8bytes[0] = vec_xl(0  , y0);
+            q8bytes[1] = vec_xl(16 , y0);
+            q8bytes[2] = vec_xl(32 , y0);
+            q8bytes[3] = vec_xl(48 , y0);
+            q8bytes[4] = vec_xl(64 , y0);
+            q8bytes[5] = vec_xl(80 , y0);
+            q8bytes[6] = vec_xl(96 , y0);
+            q8bytes[7] = vec_xl(112, y0);
+            y0 += 128;
+
+            q3h[0] = vec_sl(vec_andc(v_0c, qhbits[0]), 2);
+            q3h[1] = vec_sl(vec_andc(v_0c, qhbits[1]), 2);
+            q3h[2] = vec_sl(vec_andc(v_1c, qhbits[0]), 1);
+            q3h[3] = vec_sl(vec_andc(v_1c, qhbits[1]), 1);
+
+            q3bytes[0] = vec_sub((int8x16_t)vec_and(q3b[0], v_3m), (int8x16_t)q3h[0]);
+            q3bytes[1] = vec_sub((int8x16_t)vec_and(q3b[1], v_3m), (int8x16_t)q3h[1]);
+            q3bytes[2] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[0], 2), v_3m), (int8x16_t)q3h[2]);
+            q3bytes[3] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[1], 2), v_3m), (int8x16_t)q3h[3]);
+
+            isum0 = ggml_vec_dot(v_z, q3bytes[0], q8bytes[0]);
+            isum1 = ggml_vec_dot(v_z, q3bytes[1], q8bytes[1]);
+            isum2 = ggml_vec_dot(v_z, q3bytes[2], q8bytes[2]);
+            isum3 = ggml_vec_dot(v_z, q3bytes[3], q8bytes[3]);
+
+            isum += (isum0[0] + isum0[1] + isum0[2] + isum0[3]) * scale[0];
+            isum += (isum1[0] + isum1[1] + isum1[2] + isum1[3]) * scale[1];
+            isum += (isum2[0] + isum2[1] + isum2[2] + isum2[3]) * scale[2];
+            isum += (isum3[0] + isum3[1] + isum3[2] + isum3[3]) * scale[3];
+
+            scale += 4;
+
+            q3h[0] = vec_andc(v_2c, qhbits[0]);
+            q3h[1] = vec_andc(v_2c, qhbits[1]);
+            q3h[2] = vec_sr(vec_andc(v_3c, qhbits[0]), 1);
+            q3h[3] = vec_sr(vec_andc(v_3c, qhbits[1]), 1);
+
+            q3bytes[0] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[0], 4), v_3m), (int8x16_t)q3h[0]);
+            q3bytes[1] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[1], 4), v_3m), (int8x16_t)q3h[1]);
+            q3bytes[2] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[0], 6), v_3m), (int8x16_t)q3h[2]);
+            q3bytes[3] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[1], 6), v_3m), (int8x16_t)q3h[3]);
+
+            isum0 = ggml_vec_dot(v_z, q3bytes[0], q8bytes[4]);
+            isum1 = ggml_vec_dot(v_z, q3bytes[1], q8bytes[5]);
+            isum2 = ggml_vec_dot(v_z, q3bytes[2], q8bytes[6]);
+            isum3 = ggml_vec_dot(v_z, q3bytes[3], q8bytes[7]);
+
+            isum += (isum0[0] + isum0[1] + isum0[2] + isum0[3]) * scale[0];
+            isum += (isum1[0] + isum1[1] + isum1[2] + isum1[3]) * scale[1];
+            isum += (isum2[0] + isum2[1] + isum2[2] + isum2[3]) * scale[2];
+            isum += (isum3[0] + isum3[1] + isum3[2] + isum3[3]) * scale[3];
+
+            scale += 4;
+
+            if (j == 0) {
+                qhbits[0] = vec_sr(qhbits[0], 4);
+                qhbits[1] = vec_sr(qhbits[1], 4);
+            }
+        }
+
+        sum += d * isum;
+    }
+
+    *s = sum;
+
+#else
+    // scalar version
+    // This function is written like this so the compiler can manage to vectorize most of it
+    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
+    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
+    // The ideal situation would be if we could just write the code once, and the compiler would
+    // automatically produce the best possible set of machine instructions, instead of us having to manually
+    // write vectorized versions for AVX, ARM_NEON, etc.
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    uint32_t auxs[4];
+    const int8_t * scales = (const int8_t*)auxs;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
+        const uint8_t * GGML_RESTRICT hm = x[i].hmask;
+        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * GGML_RESTRICT a = aux8;
+        uint8_t m = 1;
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            q3 += 32;
+        }
+        a = aux8;
+
+        memcpy(auxs, x[i].scales, 12);
+        uint32_t tmp = auxs[2];
+        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
+        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
+        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
+        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
+        for (int j = 0; j < QK_K/16; ++j) {
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+
+#endif
+
+}
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#if defined(__VXE__) || defined(__VXE2__)
+    const uint8x16_t v_lm = vec_splat_u8(0x0F);
+    const int32x4_t v_z = vec_splat_s32(0);
+
+    uint8x16_t v_x[2];
+    int8x16_t  v_xl[2];
+    int8x16_t  v_y[2];
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums);
+        const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums);
+        const int16x8_t v_ysums = vec_padd_s16(v_ysumsl, v_ysumsh);
+
+        memcpy(utmp, x[i].scales, 12);
+
+        uint32x4_t v_mins8 = { 0 };
+        v_mins8 = vec_insert(utmp[1] & kmask1, v_mins8, 0);
+        v_mins8 = vec_insert(((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4), v_mins8, 1);
+
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[0] &= kmask1;
+
+        const int16x8_t v_minsh = (int16x8_t)vec_unpackh((uint8x16_t)v_mins8);
+
+        const int32x4_t v_minso = vec_mulo(v_ysums, v_minsh);
+        const int32x4_t v_minse = vec_mule(v_ysums, v_minsh);
+        const int32x4_t v_mins = v_minso + v_minse;
+        sumf -= dmin * (v_mins[0] + v_mins[1] + v_mins[2] + v_mins[3]);
+
+        const uint8_t * scales = (const uint8_t *)utmp;
+        const uint8_t * GGML_RESTRICT x0 = x[i].qs;
+        const int8_t  * GGML_RESTRICT y0 = y[i].qs;
+
+        int32_t sumi1 = 0;
+        int32_t sumi2 = 0;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            v_x[0] = vec_xl(0 , x0);
+            v_x[1] = vec_xl(16, x0);
+            x0 += 32;
+
+            v_y[0] = vec_xl(0 , y0);
+            v_y[1] = vec_xl(16, y0);
+            y0 += 32;
+
+            v_xl[0] = (int8x16_t)vec_and(v_x[0], v_lm);
+            v_xl[1] = (int8x16_t)vec_and(v_x[1], v_lm);
+
+            const int32x4_t p1 = ggml_vec_dot(ggml_vec_dot(v_z, v_xl[0], v_y[0]), v_xl[1], v_y[1]);
+            sumi1 += (p1[0] + p1[1] + p1[2] + p1[3]) * scales[2*j+0];
+
+            v_y[0] = vec_xl(0 , y0);
+            v_y[1] = vec_xl(16, y0);
+            y0 += 32;
+
+            v_xl[0] = (int8x16_t)vec_sr(v_x[0], 4);
+            v_xl[1] = (int8x16_t)vec_sr(v_x[1], 4);
+
+            const int32x4_t p2 = ggml_vec_dot(ggml_vec_dot(v_z, v_xl[0], v_y[0]), v_xl[1], v_y[1]);
+            sumi2 += (p2[0] + p2[1] + p2[2] + p2[3]) * scales[2*j+1];
+        }
+
+        sumf += d * (sumi1 + sumi2);
+    }
+
+    *s = sumf;
+
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * GGML_RESTRICT a = aux8;
+        for (int j = 0; j < QK_K/64; ++j) {
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
+            a += 32;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            a += 32; q4 += 32;
+        }
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#if defined(__VXE__) || defined(__VXE2__)
+    const uint8x16_t v_lm = vec_splat_u8(0x0F);
+    const uint8x16_t v_1m = vec_splat_u8(0x01);
+    const uint8x16_t v_2m = vec_splat_u8(0x02);
+
+    const int32x4_t v_z = vec_splat_s32(0);
+
+    const uchar8x16_t v_minsm = {
+        0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
+        0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF
+    };
+
+    int8x16_t  q5b[4];
+    uint8x16_t q5h[4];
+
+    uint8x16_t v_xl[2];
+    uint8x16_t v_xh[2];
+    int8x16_t  v_y[4];
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums);
+        const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums);
+        const int16x8_t v_ysums = vec_padd_s16(v_ysumsl, v_ysumsh);
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        const uint8x16_t v_mins16 = vec_xl(0, (const uint8_t *)utmp);
+        const uint8x16_t v_mins8 = vec_perm(v_mins16, v_mins16, v_minsm);
+        const int16x8_t v_minsh = (int16x8_t)vec_unpackh(v_mins8);
+
+        const int32x4_t v_minsho = vec_mulo(v_ysums, v_minsh);
+        const int32x4_t v_minshe = vec_mule(v_ysums, v_minsh);
+        const int32x4_t v_mins = vec_add(v_minsho, v_minshe);
+        const int32_t mins = v_mins[0] + v_mins[1] + v_mins[2] + v_mins[3];
+
+        const uint8_t * scales = (const uint8_t *)utmp;
+        const uint8_t * GGML_RESTRICT x0l = x[i].qs;
+        const uint8_t * GGML_RESTRICT x0h = x[i].qh;
+        const int8_t  * GGML_RESTRICT y0 = y[i].qs;
+
+        v_xh[0] = vec_xl(0 , x0h);
+        v_xh[1] = vec_xl(16, x0h);
+
+        int32_t sumi = 0;
+        for (int j = 0; j < QK_K/64; ++j) {
+            v_xl[0] = vec_xl(0 , x0l);
+            v_xl[1] = vec_xl(16, x0l);
+            x0l += 32;
+
+            v_y[0] = vec_xl(0 , y0);
+            v_y[1] = vec_xl(16, y0);
+            v_y[2] = vec_xl(32, y0);
+            v_y[3] = vec_xl(48, y0);
+            y0 += 64;
+
+            q5h[0] = vec_sl(vec_and(v_1m, v_xh[0]), 4);
+            q5h[1] = vec_sl(vec_and(v_1m, v_xh[1]), 4);
+            q5h[2] = vec_sl(vec_and(v_2m, v_xh[0]), 3);
+            q5h[3] = vec_sl(vec_and(v_2m, v_xh[1]), 3);
+            v_xh[0] = vec_sr(v_xh[0], 2);
+            v_xh[1] = vec_sr(v_xh[1], 2);
+
+            q5b[0] = (int8x16_t)vec_or(vec_and(v_xl[0], v_lm), q5h[0]);
+            q5b[1] = (int8x16_t)vec_or(vec_and(v_xl[1], v_lm), q5h[1]);
+            q5b[2] = (int8x16_t)vec_or(vec_sr(v_xl[0], 4), q5h[2]);
+            q5b[3] = (int8x16_t)vec_or(vec_sr(v_xl[1], 4), q5h[3]);
+
+            int32x4_t sumi0 = ggml_vec_dot(ggml_vec_dot(v_z, q5b[0], v_y[0]), q5b[1], v_y[1]);
+            int32x4_t sumi1 = ggml_vec_dot(ggml_vec_dot(v_z, q5b[2], v_y[2]), q5b[3], v_y[3]);
+
+            sumi += (sumi0[0] + sumi0[1] + sumi0[2] + sumi0[3]) * *scales++;
+            sumi += (sumi1[0] + sumi1[1] + sumi1[2] + sumi1[3]) * *scales++;
+        }
+
+        sumf += d * sumi - dmin * mins;
+    }
+
+    *s = sumf;
+
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+        const uint8_t * GGML_RESTRICT hm = x[i].qh;
+        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * GGML_RESTRICT a = aux8;
+        uint8_t m = 1;
+        for (int j = 0; j < QK_K/64; ++j) {
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            q4 += 32;
+        }
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q6_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__VXE__) || defined(__VXE2__)
+    float sum = 0;
+
+    // Lower 4-bit and upper 2-bit masks
+    const uint8x16_t v_lm = vec_splat_u8(0x0F);
+    const uint8x16_t v_um = vec_splat_u8(0x03);
+
+    const int32x4_t v_z = vec_splat_s32(0);
+
+    int8x16_t  q6b[4];
+    uint8x16_t q6h[4];
+
+    uint8x16_t v_xl[4];
+    uint8x16_t v_xh[2];
+    int8x16_t  v_y[4];
+
+    for (int i = 0; i < nb; ++i) {
+        const float d_all = GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * GGML_RESTRICT x0l = x[i].ql;
+        const uint8_t * GGML_RESTRICT x0h = x[i].qh;
+        const int8_t  * GGML_RESTRICT y0 = y[i].qs;
+
+        const int8_t  * GGML_RESTRICT scale = x[i].scales;
+
+        const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums);
+        const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums);
+
+        const int8x16_t v_scale  = vec_xl(0, scale);
+        const int16x8_t v_scalel = vec_unpackh(v_scale);
+        const int16x8_t v_scaleh = vec_unpackl(v_scale);
+
+        const int32x4_t v_minslo = vec_mulo(v_ysumsl, v_scalel);
+        const int32x4_t v_minsle = vec_mule(v_ysumsl, v_scalel);
+        const int32x4_t v_minsho = vec_mulo(v_ysumsh, v_scaleh);
+        const int32x4_t v_minshe = vec_mule(v_ysumsh, v_scaleh);
+        const int32x4_t v_mins = v_minslo + v_minsle + v_minsho + v_minshe;
+
+        const int32_t mins = v_mins[0] + v_mins[1] + v_mins[2] + v_mins[3];
+
+        int32_t isum = 0;
+        for (int j = 0; j < QK_K/128; ++j) {
+            // Load model upper 2 bits
+            v_xh[0] = vec_xl(0 , x0h);
+            v_xh[1] = vec_xl(16, x0h);
+            x0h += 32;
+
+            // Load model lower 4 bits
+            v_xl[0] = vec_xl(0 , x0l);
+            v_xl[1] = vec_xl(16, x0l);
+            v_xl[2] = vec_xl(32, x0l);
+            v_xl[3] = vec_xl(48, x0l);
+            x0l += 64;
+
+            // Load activation quants
+            v_y[0] = vec_xl(0 , y0);
+            v_y[1] = vec_xl(16, y0);
+            v_y[2] = vec_xl(32, y0);
+            v_y[3] = vec_xl(48, y0);
+            y0 += 64;
+
+            q6h[0] = vec_sl(vec_and(v_um, v_xh[0]), 4);
+            q6h[1] = vec_sl(vec_and(v_um, v_xh[1]), 4);
+            uint8x16_t shifted = vec_sr(v_xh[0], 2);
+            q6h[2] = vec_sl(vec_and(v_um, shifted), 4);
+            shifted = vec_sr(v_xh[1], 2);
+            q6h[3] = vec_sl(vec_and(v_um, shifted), 4);
+
+            q6b[0] = (int8x16_t)(vec_or(vec_and(v_xl[0], v_lm), q6h[0]));
+            q6b[1] = (int8x16_t)(vec_or(vec_and(v_xl[1], v_lm), q6h[1]));
+            q6b[2] = (int8x16_t)(vec_or(vec_and(v_xl[2], v_lm), q6h[2]));
+            q6b[3] = (int8x16_t)(vec_or(vec_and(v_xl[3], v_lm), q6h[3]));
+
+            int32x4_t summs0 = ggml_vec_dot(v_z, q6b[0], v_y[0]);
+            int32x4_t summs1 = ggml_vec_dot(v_z, q6b[1], v_y[1]);
+            int32x4_t summs2 = ggml_vec_dot(v_z, q6b[2], v_y[2]);
+            int32x4_t summs3 = ggml_vec_dot(v_z, q6b[3], v_y[3]);
+
+            isum += (summs0[0] + summs0[1] + summs0[2] + summs0[3]) * scale[0] +
+                    (summs1[0] + summs1[1] + summs1[2] + summs1[3]) * scale[1] +
+                    (summs2[0] + summs2[1] + summs2[2] + summs2[3]) * scale[2] +
+                    (summs3[0] + summs3[1] + summs3[2] + summs3[3]) * scale[3];
+
+            scale += 4;
+
+
+            // Load activation quants
+            v_y[0] = vec_xl(0 , y0);
+            v_y[1] = vec_xl(16, y0);
+            v_y[2] = vec_xl(32, y0);
+            v_y[3] = vec_xl(48, y0);
+            y0 += 64;
+
+            shifted = vec_sr(v_xh[0], 4);
+            q6h[0] = vec_sl(vec_and(v_um, shifted), 4);
+            shifted = vec_sr(v_xh[1], 4);
+            q6h[1] = vec_sl(vec_and(v_um, shifted), 4);
+            shifted = vec_sr(v_xh[0], 6);
+            q6h[2] = vec_sl(vec_and(v_um, shifted), 4);
+            shifted = vec_sr(v_xh[1], 6);
+            q6h[3] = vec_sl(vec_and(v_um, shifted), 4);
+
+            q6b[0] = (int8x16_t)(vec_or(vec_sr(v_xl[0], 4), q6h[0]));
+            q6b[1] = (int8x16_t)(vec_or(vec_sr(v_xl[1], 4), q6h[1]));
+            q6b[2] = (int8x16_t)(vec_or(vec_sr(v_xl[2], 4), q6h[2]));
+            q6b[3] = (int8x16_t)(vec_or(vec_sr(v_xl[3], 4), q6h[3]));
+
+            summs0 = ggml_vec_dot(v_z, q6b[0], v_y[0]);
+            summs1 = ggml_vec_dot(v_z, q6b[1], v_y[1]);
+            summs2 = ggml_vec_dot(v_z, q6b[2], v_y[2]);
+            summs3 = ggml_vec_dot(v_z, q6b[3], v_y[3]);
+
+            isum += (summs0[0] + summs0[1] + summs0[2] + summs0[3]) * scale[0] +
+                    (summs1[0] + summs1[1] + summs1[2] + summs1[3]) * scale[1] +
+                    (summs2[0] + summs2[1] + summs2[2] + summs2[3]) * scale[2] +
+                    (summs3[0] + summs3[1] + summs3[2] + summs3[3]) * scale[3];
+
+            scale += 4;
+        }
+
+        sum += d_all * y[i].d * (isum - 32 * mins);
+    }
+
+    *s = sum;
+
+#else
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * GGML_RESTRICT a = aux8;
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) {
+                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
+                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
+                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
+                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
+            }
+            a  += 128;
+            q4 += 64;
+            qh += 32;
+        }
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/16; ++j) {
+            int scale = x[i].scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+// #if defined(__VXE__) || defined(__VXE2__)
+// static const int8_t keven_signs_q2xs[1024] = {
+//      1,  1,  1,  1,  1,  1,  1,  1, -1,  1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1,  1,
+//      1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1,  1,  1, -1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1, -1,
+//      1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1, -1,
+//      1,  1, -1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1,  1,
+//      1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1, -1,
+//      1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1,  1,
+//      1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1,  1,
+//      1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1,  1,  1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1, -1,
+//      1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1, -1,
+//      1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1,  1,
+//      1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1,  1,
+//      1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1, -1,
+//      1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1,  1,
+//      1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1, -1,
+//      1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1, -1,
+//      1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1,  1,
+//      1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1, -1,
+//      1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1,  1,
+//      1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1,  1,
+//      1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1, -1,
+//      1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1,  1,
+//      1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1, -1,
+//      1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1, -1,
+//      1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1,  1,
+//      1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1,  1,
+//      1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1, -1,
+//      1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1, -1,
+//      1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1,  1,
+//      1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1, -1,
+//      1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1,  1,
+//      1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1,  1,
+//      1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1,  1,  1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1, -1,
+// };
+// #endif
+
+// 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) {
+//     assert(n % QK_K == 0);
+//     assert(nrc == 1);
+//     UNUSED(nrc);
+//     UNUSED(bx);
+//     UNUSED(by);
+//     UNUSED(bs);
+
+//     const block_iq2_xxs * GGML_RESTRICT x = vx;
+//     const block_q8_K    * GGML_RESTRICT y = vy;
+
+//     const int nb = n / QK_K;
+
+// #if defined(__VXE__) || defined(__VXE2__)
+//    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+//    uint32_t aux32[4];
+//    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+//    float sumf = 0;
+
+//    for (int i = 0; i < nb; ++i) {
+//        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+//        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
+//        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
+
+//        float sumf1 = 0, sumf2 = 0;
+
+//        for (int ib32 = 0; ib32 < QK_K/32; ib += 2) {
+//            int8x16_t q8b0 = vec_xl( 0, q8);
+//            int8x16_t qb81 = vec_xl(16, q8);
+//            int8x16_t q8b2 = vec_xl(32, q8);
+//            int8x16_t q8b3 = vec_xl(48, q8);
+//            q8 += 64;
+
+//            memcpy(aux32, q2, 4 * sizeof(uint32_t));
+//            q2 += 8;
+
+//            int8x16_t q2u0 = { *(const int64_t *)(iq2xxs_grid + aux8[ 0]), *(const int64_t *)(iq2xxs_grid + aux8[ 1]) };
+//            int8x16_t q2u1 = { *(const int64_t *)(iq2xxs_grid + aux8[ 2]), *(const int64_t *)(iq2xxs_grid + aux8[ 3]) };
+//            int8x16_t q2u2 = { *(const int64_t *)(iq2xxs_grid + aux8[ 8]), *(const int64_t *)(iq2xxs_grid + aux8[ 9]) };
+//            int8x16_t q2u3 = { *(const int64_t *)(iq2xxs_grid + aux8[10]), *(const int64_t *)(iq2xxs_grid + aux8[11]) };
+
+//            int8x16_t q2s0 = { *(const int64_t *)(signs64 + ((aux32[1] >>  0) & 127)), *(const int64_t *)(signs64 + ((aux32[1] >>  7) & 127)) };
+//            int8x16_t q2s1 = { *(const int64_t *)(signs64 + ((aux32[1] >> 14) & 127)), *(const int64_t *)(signs64 + ((aux32[1] >> 21) & 127)) };
+//            int8x16_t q2s2 = { *(const int64_t *)(signs64 + ((aux32[3] >>  0) & 127)), *(const int64_t *)(signs64 + ((aux32[3] >>  7) & 127)) };
+//            int8x16_t q2s3 = { *(const int64_t *)(signs64 + ((aux32[3] >> 14) & 127)), *(const int64_t *)(signs64 + ((aux32[3] >> 21) & 127)) };
+
+//            q2u0 = vec_mul(q2u0, q2s0);
+//            q2u1 = vec_mul(q2u1, q2s1);
+//            q2u2 = vec_mul(q2u2, q2s2);
+//            q2u3 = vec_mul(q2u3, q2s3);
+
+//            const int32x4_t p1 = ggml_vec_dot(ggml_vec_dot(vec_splat_s32(0), q2u0, q8b0), q2u1, q8b1);
+//            const int32x4_t p2 = ggml_vec_dot(ggml_vec_dot(vec_splat_s32(0), q2u2, q8b2), q2u3, q8b3);
+
+//            sumf1 += (p1[0] + p1[1] + p1[2] + p1[3]) * (0.5f + (aux32[1] >> 28));
+//            sumf2 += (p2[0] + p2[1] + p2[2] + p2[3]) * (0.5f + (aux32[3] >> 28));
+//        }
+
+//        sumf += d * (sumf1 + sumf2);
+//    }
+
+//    *s = 0.25f * sumf;
+
+// #else
+
+//     uint32_t aux32[2];
+//     const uint8_t * aux8 = (const uint8_t *)aux32;
+
+//     float sumf = 0.f;
+//     for (int i = 0; i < nb; ++i) {
+//         const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+//         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
+//         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
+//         int32_t bsum = 0;
+//         for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+//             memcpy(aux32, q2, 2*sizeof(uint32_t));
+//             q2 += 4;
+//             const uint32_t ls = 2*(aux32[1] >> 28) + 1;
+//             int32_t sumi = 0;
+//             for (int l = 0; l < 4; ++l) {
+//                 const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
+//                 const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
+//                 for (int j = 0; j < 8; ++j) {
+//                     sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+//                 }
+//                 q8 += 8;
+//             }
+//             bsum += sumi * ls;
+//         }
+//         sumf += d * bsum;
+//     }
+//     *s = 0.125f * sumf;
+// #endif
+// }
+
+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) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK4_NL == 0);
+    static_assert(QK4_NL == QK8_0, "QK4_NL and QK8_0 must be the same");
+
+    const block_iq4_nl * GGML_RESTRICT x = vx;
+    const block_q8_0   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK4_NL;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__VXE__) || defined(__VXE2__)
+    const int8x16_t v_k = vec_xl(0, kvalues_iq4nl);
+    const uint8x16_t v_m = vec_splat_u8(0x0F);
+
+    for (; ib < nb; ++ib) {
+        const block_iq4_nl * GGML_RESTRICT x0 = &x[ib];
+        const block_q8_0   * GGML_RESTRICT y0 = &y[ib];
+
+        const uint8x16_t v_x = vec_xl(0, x0->qs);
+        int8x16_t v_xl = (int8x16_t)vec_and(v_x, v_m);
+        int8x16_t v_xh = (int8x16_t)vec_sr(v_x, 4);
+
+        v_xl = vec_perm(v_k, v_k, (uchar8x16_t)v_xl);
+        v_xh = vec_perm(v_k, v_k, (uchar8x16_t)v_xh);
+
+        const int8x16_t v_yl = vec_xl(0      , y0->qs);
+        const int8x16_t v_yh = vec_xl(QK8_0/2, y0->qs);
+        const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
+
+        sumf += GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d) * (v_xy[0] + v_xy[1] + v_xy[2] + v_xy[3]);
+    }
+
+#endif
+    for (; ib < nb; ++ib) {
+        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        int sumi1 = 0, sumi2 = 0;
+        for (int j = 0; j < QK4_NL/2; ++j) {
+            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
+            sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >>  4];
+        }
+        sumf += d * (sumi1 + sumi2);
+    }
+    *s = sumf;
+}
+
+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) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK_K == 0);
+
+    const block_iq4_xs * GGML_RESTRICT x = vx;
+    const block_q8_K   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__VXE__) || defined(__VXE2__)
+    const int8x16_t v_k = vec_xl(0, kvalues_iq4nl);
+    const uint8x16_t v_m = vec_splat_u8(0x0F);
+
+    float sumf = 0;
+
+    for (int ibl = 0; ibl < nb; ++ibl) {
+        const uint8_t * GGML_RESTRICT q4 = x[ibl].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[ibl].qs;
+
+        uint16_t h = x[ibl].scales_h;
+
+        int sumi1 = 0, sumi2 = 0;
+        for (int ib = 0; ib < QK_K/64; ++ib) {
+            const uint8x16_t v_x0 = vec_xl(0       , q4);
+            const uint8x16_t v_x1 = vec_xl(QK4_NL/2, q4);
+            q4 += 32;
+
+            int8x16_t v_x0l = (int8x16_t)vec_and(v_x0, v_m);
+            int8x16_t v_x0h = (int8x16_t)vec_sr(v_x0, 4);
+            int8x16_t v_x1l = (int8x16_t)vec_and(v_x1, v_m);
+            int8x16_t v_x1h = (int8x16_t)vec_sr(v_x1, 4);
+
+            v_x0l = vec_perm(v_k, v_k, (uchar8x16_t)v_x0l);
+            v_x0h = vec_perm(v_k, v_k, (uchar8x16_t)v_x0h);
+            v_x1l = vec_perm(v_k, v_k, (uchar8x16_t)v_x1l);
+            v_x1h = vec_perm(v_k, v_k, (uchar8x16_t)v_x1h);
+
+            const int8x16_t v_y0 = vec_xl( 0, q8);
+            const int8x16_t v_y1 = vec_xl(16, q8);
+            const int8x16_t v_y2 = vec_xl(32, q8);
+            const int8x16_t v_y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            int32x4_t vsumi0 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x0l, v_y0), v_x0h, v_y1);
+            int32x4_t vsumi1 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x1l, v_y2), v_x1h, v_y3);
+
+            int ls1 = ((x[ibl].scales_l[ib] & 0xF) | ((h << 4) & 0x30)) - 32;
+            int ls2 = ((x[ibl].scales_l[ib] >>  4) | ((h << 2) & 0x30)) - 32;
+
+            h >>= 4;
+
+            sumi1 += (vsumi0[0] + vsumi0[1] + vsumi0[2] + vsumi0[3]) * ls1;
+            sumi2 += (vsumi1[0] + vsumi1[1] + vsumi1[2] + vsumi1[3]) * ls2;
+        }
+
+        sumf += GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
+    }
+
+    *s = sumf;
+
+#else
+    float sumf = 0;
+    for (int ibl = 0; ibl < nb; ++ibl) {
+        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        uint16_t h = x[ibl].scales_h;
+        const uint8_t * qs = x[ibl].qs;
+        const int8_t  * q8 = y[ibl].qs;
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
+            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
+            h >>= 4;
+            const float d1 = d4d8*(ls1 - 32);
+            const float d2 = d4d8*(ls2 - 32);
+            int sumi1 = 0, sumi2 = 0;
+            for (int j = 0; j < 16; ++j) {
+                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
+            }
+            sumf += d1 * (sumi1 + sumi2);
+            qs += 16;
+            q8 += 32;
+            sumi1 = sumi2 = 0;
+            for (int j = 0; j < 16; ++j) {
+                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
+            }
+            sumf += d2 * (sumi1 + sumi2);
+            qs += 16;
+            q8 += 32;
+        }
+    }
+    *s = sumf;
+#endif
+}
+

ggml/src/ggml-cpu/arch/wasm/quants.c

@@ -0,0 +1,1480 @@
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+#include "ggml-quants.h"
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+
+#include "../../quants.h"
+#include "../../ggml-cpu-impl.h"
+
+#include <math.h>
+#include <string.h>
+#include <assert.h>
+#include <float.h>
+#include <stdlib.h> // for qsort
+#include <stdio.h>  // for GGML_ASSERT
+
+#define GROUP_MAX_EPS 1e-15f
+#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
+#define GROUP_MAX_EPS_IQ2_S 1e-8f
+#define GROUP_MAX_EPS_IQ1_M 1e-7f
+#define GROUP_MAX_EPS_IQ1_S 1e-12f
+
+#define UNUSED GGML_UNUSED
+
+#if defined(__wasm_simd128__)
+#define B1(c,s,n)  0x ## n ## c ,  0x ## n ## s
+#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)
+#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s)
+#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s)
+#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s)
+#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s)
+#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s)
+#define B8(c,s  ) B7(c,s,     c), B7(c,s,     s)
+
+// precomputed tables for expanding 8bits to 8 bytes:
+static const uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b) << 4
+static const uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4
+#endif
+
+void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(QK8_0 == 32);
+    assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
+
+    block_q8_0 * GGML_RESTRICT y = vy;
+
+#if defined __wasm_simd128__
+    for (int i = 0; i < nb; i++) {
+        v128_t srcv [8];
+        v128_t asrcv[8];
+        v128_t amaxv[8];
+
+        for (int j = 0; j < 8; j++) srcv[j]  = wasm_v128_load(x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]);
+
+        for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = MAX(MAX(wasm_f32x4_extract_lane(amaxv[0], 0),
+                                   wasm_f32x4_extract_lane(amaxv[0], 1)),
+                               MAX(wasm_f32x4_extract_lane(amaxv[0], 2),
+                                   wasm_f32x4_extract_lane(amaxv[0], 3)));
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        for (int j = 0; j < 8; j++) {
+            const v128_t v  = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
+            const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v);
+
+            y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0);
+            y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1);
+            y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2);
+            y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3);
+        }
+    }
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_0_ref(x, y, k);
+#endif
+}
+
+void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(k % QK8_1 == 0);
+    const int nb = k / QK8_1;
+
+    block_q8_1 * GGML_RESTRICT y = vy;
+#if defined __wasm_simd128__
+    for (int i = 0; i < nb; i++) {
+        v128_t srcv [8];
+        v128_t asrcv[8];
+        v128_t amaxv[8];
+
+        for (int j = 0; j < 8; j++) srcv[j]  = wasm_v128_load(x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]);
+
+        for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = MAX(MAX(wasm_f32x4_extract_lane(amaxv[0], 0),
+                                   wasm_f32x4_extract_lane(amaxv[0], 1)),
+                               MAX(wasm_f32x4_extract_lane(amaxv[0], 2),
+                                   wasm_f32x4_extract_lane(amaxv[0], 3)));
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        v128_t accv = wasm_i32x4_splat(0);
+
+        for (int j = 0; j < 8; j++) {
+            const v128_t v  = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
+            const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v);
+
+            y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0);
+            y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1);
+            y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2);
+            y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3);
+
+            accv = wasm_i32x4_add(accv, vi);
+        }
+
+        y[i].s = GGML_FP32_TO_FP16(
+                d * (wasm_i32x4_extract_lane(accv, 0) +
+                     wasm_i32x4_extract_lane(accv, 1) +
+                     wasm_i32x4_extract_lane(accv, 2) +
+                     wasm_i32x4_extract_lane(accv, 3)));
+    }
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_1_ref(x, y, k);
+#endif
+}
+
+//===================================== Q8_K ==============================================
+
+void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+#ifdef __wasm_simd128__
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
+    block_q8_K * GGML_RESTRICT yc = y; // Cast to proper type
+
+    for (int i = 0; i < nb; i++) {
+        const float * x_block = x + i * QK_K;
+
+        v128_t min_vec = wasm_v128_load(x_block);
+        v128_t max_vec = min_vec;
+
+        for (int j = 4; j < QK_K; j += 4) {
+            v128_t x_vec = wasm_v128_load(x_block + j);
+            max_vec = wasm_f32x4_pmax(max_vec, x_vec);
+            min_vec = wasm_f32x4_pmin(min_vec, x_vec);
+        }
+        max_vec = wasm_f32x4_pmax(max_vec, wasm_i32x4_shuffle(max_vec, max_vec, 2, 3, 0, 1));
+        max_vec = wasm_f32x4_pmax(max_vec, wasm_i32x4_shuffle(max_vec, max_vec, 1, 0, 3, 2));
+        min_vec = wasm_f32x4_pmin(min_vec, wasm_i32x4_shuffle(min_vec, min_vec, 2, 3, 0, 1));
+        min_vec = wasm_f32x4_pmin(min_vec, wasm_i32x4_shuffle(min_vec, min_vec, 1, 0, 3, 2));
+        float max = wasm_f32x4_extract_lane(max_vec, 0);
+        float min = wasm_f32x4_extract_lane(min_vec, 0);
+        float amax = -min > max ? min : max;
+
+        if (amax == 0.0f) {
+            yc[i].d = 0.0f;
+            const v128_t zero = wasm_i8x16_splat(0);
+            for (int j = 0; j < QK_K; j += 16) {
+                wasm_v128_store(yc[i].qs + j, zero);
+            }
+            continue;
+        }
+
+        const float iscale = -127.0f / amax;
+        const v128_t scale_vec = wasm_f32x4_splat(iscale);
+
+        // Process 16 elements per iteration
+        for (int j = 0, jb = 0; j < QK_K; j += 16, jb++) {
+            // Load and quantize 16 floats
+            v128_t x0 = wasm_v128_load(x_block + j);
+            v128_t x1 = wasm_v128_load(x_block + j + 4);
+            v128_t x2 = wasm_v128_load(x_block + j + 8);
+            v128_t x3 = wasm_v128_load(x_block + j + 12);
+
+            v128_t q0 = wasm_f32x4_nearest(wasm_f32x4_mul(x0, scale_vec));
+            v128_t q1 = wasm_f32x4_nearest(wasm_f32x4_mul(x1, scale_vec));
+            v128_t q2 = wasm_f32x4_nearest(wasm_f32x4_mul(x2, scale_vec));
+            v128_t q3 = wasm_f32x4_nearest(wasm_f32x4_mul(x3, scale_vec));
+
+            // Convert to i32 with saturation
+            v128_t i0 = wasm_i32x4_trunc_sat_f32x4(q0);
+            v128_t i1 = wasm_i32x4_trunc_sat_f32x4(q1);
+            v128_t i2 = wasm_i32x4_trunc_sat_f32x4(q2);
+            v128_t i3 = wasm_i32x4_trunc_sat_f32x4(q3);
+
+            // Pack into 16 i8 values
+            v128_t i8 = wasm_i8x16_narrow_i16x8(
+                wasm_i16x8_narrow_i32x4(i0, i1),
+                wasm_i16x8_narrow_i32x4(i2, i3)
+            );
+            wasm_v128_store(yc[i].qs + j, i8);
+
+            // Calculate bsums using SIMD
+            v128_t sum16 = wasm_i16x8_add(
+                wasm_i16x8_extend_low_i8x16(i8),
+                wasm_i16x8_extend_high_i8x16(i8)
+            );
+            v128_t sum32 = wasm_i32x4_add(
+                wasm_i32x4_extend_low_i16x8(sum16),
+                wasm_i32x4_extend_high_i16x8(sum16)
+            );
+            sum32 = wasm_i32x4_add(sum32, wasm_i32x4_shuffle(sum32, sum32, 2, 3, 0, 1));
+            sum32 = wasm_i32x4_add(sum32, wasm_i32x4_shuffle(sum32, sum32, 1, 0, 3, 2));
+            yc[i].bsums[jb] = wasm_i32x4_extract_lane(sum32, 0);
+        }
+
+        yc[i].d = 1.0f / iscale;
+    }
+#else
+    quantize_row_q8_K_ref(x, y, k);
+#endif
+}
+
+
+//===================================== Dot products =================================
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined __wasm_simd128__
+    v128_t sumv = wasm_f32x4_splat(0.0f);
+
+    const v128_t m4b = wasm_i8x16_splat(0x0F);
+    const v128_t s8b = wasm_i8x16_splat(0x8);
+
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q4_0 * GGML_RESTRICT x0 = &x[ib];
+        const block_q4_0 * GGML_RESTRICT x1 = &x[ib + 1];
+        const block_q8_0 * GGML_RESTRICT y0 = &y[ib];
+        const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1];
+
+        // Load and process x0
+        v128_t v0_0 = wasm_v128_load(x0->qs);
+        v128_t v0_0l = wasm_v128_and(v0_0, m4b);
+        v128_t v0_0h = wasm_u8x16_shr(v0_0, 4);
+        v128_t v0_0ls = wasm_i8x16_sub(v0_0l, s8b);
+        v128_t v0_0hs = wasm_i8x16_sub(v0_0h, s8b);
+
+        // Load y0 vectors
+        v128_t y0_l = wasm_v128_load(y0->qs);
+        v128_t y0_h = wasm_v128_load(y0->qs + 16);
+
+        // Extend to i16x8 and compute dot products
+        v128_t dx0l = wasm_i16x8_extend_low_i8x16(v0_0ls);
+        v128_t dx0h = wasm_i16x8_extend_high_i8x16(v0_0ls);
+        v128_t dx0hl = wasm_i16x8_extend_low_i8x16(v0_0hs);
+        v128_t dx0hh = wasm_i16x8_extend_high_i8x16(v0_0hs);
+
+        v128_t dy0ll = wasm_i16x8_extend_low_i8x16(y0_l);
+        v128_t dy0lh = wasm_i16x8_extend_high_i8x16(y0_l);
+        v128_t dy0hl = wasm_i16x8_extend_low_i8x16(y0_h);
+        v128_t dy0hh = wasm_i16x8_extend_high_i8x16(y0_h);
+
+        v128_t dp0 = wasm_i32x4_add(
+            wasm_i32x4_add(
+                wasm_i32x4_dot_i16x8(dx0l, dy0ll),
+                wasm_i32x4_dot_i16x8(dx0h, dy0lh)
+            ),
+            wasm_i32x4_add(
+                wasm_i32x4_dot_i16x8(dx0hl, dy0hl),
+                wasm_i32x4_dot_i16x8(dx0hh, dy0hh)
+            )
+        );
+
+        // Load and process x1
+        v128_t v0_1 = wasm_v128_load(x1->qs);
+        v128_t v0_1l = wasm_v128_and(v0_1, m4b);
+        v128_t v0_1h = wasm_u8x16_shr(v0_1, 4);
+        v128_t v0_1ls = wasm_i8x16_sub(v0_1l, s8b);
+        v128_t v0_1hs = wasm_i8x16_sub(v0_1h, s8b);
+
+        // Load y1 vectors
+        v128_t y1_l = wasm_v128_load(y1->qs);
+        v128_t y1_h = wasm_v128_load(y1->qs + 16);
+
+        // Extend to i16x8 and compute dot products
+        v128_t dx1l = wasm_i16x8_extend_low_i8x16(v0_1ls);
+        v128_t dx1h = wasm_i16x8_extend_high_i8x16(v0_1ls);
+        v128_t dx1hl = wasm_i16x8_extend_low_i8x16(v0_1hs);
+        v128_t dx1hh = wasm_i16x8_extend_high_i8x16(v0_1hs);
+
+        v128_t dy1ll = wasm_i16x8_extend_low_i8x16(y1_l);
+        v128_t dy1lh = wasm_i16x8_extend_high_i8x16(y1_l);
+        v128_t dy1hl = wasm_i16x8_extend_low_i8x16(y1_h);
+        v128_t dy1hh = wasm_i16x8_extend_high_i8x16(y1_h);
+
+        v128_t dp1 = wasm_i32x4_add(
+            wasm_i32x4_add(
+                wasm_i32x4_dot_i16x8(dx1l, dy1ll),
+                wasm_i32x4_dot_i16x8(dx1h, dy1lh)
+            ),
+            wasm_i32x4_add(
+                wasm_i32x4_dot_i16x8(dx1hl, dy1hl),
+                wasm_i32x4_dot_i16x8(dx1hh, dy1hh)
+            )
+        );
+
+        // Accumulate results with scaling
+        float scale0 = GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d);
+        float scale1 = GGML_FP16_TO_FP32(x1->d) * GGML_FP16_TO_FP32(y1->d);
+
+        sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(dp0), wasm_f32x4_splat(scale0)));
+        sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(dp1), wasm_f32x4_splat(scale1)));
+    }
+
+    sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
+           wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
+            const int v1 = (x[ib].qs[j] >>   4) - 8;
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+    }
+
+    *s = sumf;
+}
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+#if defined __wasm_simd128__
+    v128_t sumv = wasm_f32x4_splat(0.0f);
+
+    uint32_t qh_;
+    uint64_t tmp[4];
+
+    // TODO: check if unrolling this is better
+    for (; ib < nb; ++ib) {
+        const block_q5_0 * GGML_RESTRICT x0 = &x[ib];
+        const block_q8_0 * GGML_RESTRICT y0 = &y[ib];
+
+        const v128_t m4b  = wasm_i8x16_splat(0x0F);
+
+        // extract the 5th bit
+        memcpy(&qh_, x0->qh, sizeof(qh_));
+
+        tmp[0] = table_b2b_1[(qh_ >>  0) & 0xFF];
+        tmp[1] = table_b2b_1[(qh_ >>  8) & 0xFF];
+        tmp[2] = table_b2b_1[(qh_ >> 16) & 0xFF];
+        tmp[3] = table_b2b_1[(qh_ >> 24)       ];
+
+        const v128_t qhl = wasm_v128_load(tmp + 0);
+        const v128_t qhh = wasm_v128_load(tmp + 2);
+
+        const v128_t v0 = wasm_v128_load(x0->qs);
+
+        // 4-bit -> 8-bit
+        const v128_t v0l = wasm_v128_and (v0, m4b);
+        const v128_t v0h = wasm_u8x16_shr(v0, 4);
+
+        // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero)
+        const v128_t v0lf = wasm_i8x16_sub(v0l, qhl);
+        const v128_t v0hf = wasm_i8x16_sub(v0h, qhh);
+
+        // load y
+        const v128_t v1l = wasm_v128_load(y0->qs);
+        const v128_t v1h = wasm_v128_load(y0->qs + 16);
+
+        // int8x16 -> int16x8
+        const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
+        const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
+        const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
+        const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
+
+        const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
+        const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
+        const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
+        const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
+
+        // dot product
+        sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(
+                        wasm_i32x4_add(
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
+                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
+                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
+                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
+    }
+
+    sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
+           wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
+
+#endif
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
+            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
+
+            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
+            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+    }
+
+    *s = sumf;
+}
+
+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) {
+    const int qk = QK8_1;
+    const int nb = n / qk;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_1);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+#if defined __wasm_simd128__
+    v128_t sumv = wasm_f32x4_splat(0.0f);
+
+    float summs = 0.0f;
+
+    uint32_t qh_;
+    uint64_t tmp[4];
+
+    // TODO: check if unrolling this is better
+    for (; ib < nb; ++ib) {
+        const block_q5_1 * GGML_RESTRICT x0 = &x[ib];
+        const block_q8_1 * GGML_RESTRICT y0 = &y[ib];
+
+        summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
+
+        const v128_t m4b = wasm_i8x16_splat(0x0F);
+
+        // extract the 5th bit
+        memcpy(&qh_, x0->qh, sizeof(qh_));
+
+        tmp[0] = table_b2b_0[(qh_ >>  0) & 0xFF];
+        tmp[1] = table_b2b_0[(qh_ >>  8) & 0xFF];
+        tmp[2] = table_b2b_0[(qh_ >> 16) & 0xFF];
+        tmp[3] = table_b2b_0[(qh_ >> 24)       ];
+
+        const v128_t qhl = wasm_v128_load(tmp + 0);
+        const v128_t qhh = wasm_v128_load(tmp + 2);
+
+        const v128_t v0 = wasm_v128_load(x0->qs);
+
+        // 4-bit -> 8-bit
+        const v128_t v0l = wasm_v128_and (v0, m4b);
+        const v128_t v0h = wasm_u8x16_shr(v0, 4);
+
+        // add high bit
+        const v128_t v0lf = wasm_v128_or(v0l, qhl);
+        const v128_t v0hf = wasm_v128_or(v0h, qhh);
+
+        // load y
+        const v128_t v1l = wasm_v128_load(y0->qs);
+        const v128_t v1h = wasm_v128_load(y0->qs + 16);
+
+        // int8x16 -> int16x8
+        const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
+        const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
+        const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
+        const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
+
+        const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
+        const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
+        const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
+        const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
+
+        // dot product
+        sumv = wasm_f32x4_add(sumv,
+                wasm_f32x4_mul(wasm_f32x4_convert_i32x4(wasm_i32x4_add(
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
+                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
+                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
+                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
+    }
+
+    sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
+           wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs;
+
+#endif
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
+            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
+
+            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
+            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q8_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined __wasm_simd128__
+    v128_t sumv = wasm_f32x4_splat(0.0f);
+
+    for (; ib < nb; ++ib) {
+        const block_q8_0 * GGML_RESTRICT x0 = &x[ib];
+        const block_q8_0 * GGML_RESTRICT y0 = &y[ib];
+
+        const v128_t x0_0 = wasm_v128_load(x0->qs);
+        const v128_t x0_1 = wasm_v128_load(x0->qs + 16);
+        const v128_t y0_0 = wasm_v128_load(y0->qs);
+        const v128_t y0_1 = wasm_v128_load(y0->qs + 16);
+
+        // Extend 8-bit to 16-bit
+        const v128_t x0_0l = wasm_i16x8_extend_low_i8x16(x0_0);
+        const v128_t x0_0h = wasm_i16x8_extend_high_i8x16(x0_0);
+        const v128_t x0_1l = wasm_i16x8_extend_low_i8x16(x0_1);
+        const v128_t x0_1h = wasm_i16x8_extend_high_i8x16(x0_1);
+
+        const v128_t y0_0l = wasm_i16x8_extend_low_i8x16(y0_0);
+        const v128_t y0_0h = wasm_i16x8_extend_high_i8x16(y0_0);
+        const v128_t y0_1l = wasm_i16x8_extend_low_i8x16(y0_1);
+        const v128_t y0_1h = wasm_i16x8_extend_high_i8x16(y0_1);
+
+        // Compute dot products
+        const v128_t dx0_0 = wasm_i32x4_dot_i16x8(x0_0l, y0_0l);
+        const v128_t dx0_1 = wasm_i32x4_dot_i16x8(x0_0h, y0_0h);
+        const v128_t dx1_0 = wasm_i32x4_dot_i16x8(x0_1l, y0_1l);
+        const v128_t dx1_1 = wasm_i32x4_dot_i16x8(x0_1h, y0_1h);
+
+        // Sum all dot products
+        const v128_t sum_dots = wasm_i32x4_add(wasm_i32x4_add(dx0_0, dx0_1), wasm_i32x4_add(dx1_0, dx1_1));
+
+        // Convert to float and accumulate
+        const float scale = GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d);
+        sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(sum_dots), wasm_f32x4_splat(scale)));
+    }
+
+    sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
+           wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi = 0;
+
+        for (int j = 0; j < qk; j++) {
+            sumi += x[ib].qs[j]*y[ib].qs[j];
+        }
+
+        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+    }
+
+    *s = sumf;
+}
+
+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) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q2_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __wasm_simd128__
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * q2 = x[i].qs;
+        const int8_t * q8 = y[i].qs;
+        const uint8_t * sc = x[i].scales;
+
+        // Vectorized summs calculation
+        v128_t summs_vec = wasm_i32x4_splat(0);
+        {
+            v128_t sc_vec = wasm_v128_load(sc);
+            v128_t sc_upper = wasm_u8x16_shr(sc_vec, 4);
+
+            v128_t sc_low = wasm_u16x8_extend_low_u8x16(sc_upper);
+            v128_t sc_high = wasm_u16x8_extend_high_u8x16(sc_upper);
+
+            v128_t bsums1 = wasm_v128_load(&y[i].bsums[0]);
+            v128_t bsums2 = wasm_v128_load(&y[i].bsums[8]);
+
+            summs_vec = wasm_i32x4_add(
+                wasm_i32x4_add(wasm_i32x4_dot_i16x8(sc_low, bsums1),
+                               wasm_i32x4_dot_i16x8(sc_high, bsums2)),
+                summs_vec
+            );
+
+            summs_vec = wasm_i32x4_add(summs_vec, wasm_i32x4_shuffle(summs_vec, summs_vec, 2, 3, 0, 1));
+            summs_vec = wasm_i32x4_add(summs_vec, wasm_i32x4_shuffle(summs_vec, summs_vec, 1, 0, 3, 2));
+        }
+        int32_t summs = wasm_i32x4_extract_lane(summs_vec, 0);
+
+        // Vectorized isum calculation
+        int32_t isum = 0;
+        const uint8_t * sc_ptr = sc;
+        const int k_iters = QK_K/128;
+
+        for (int k = 0; k < k_iters; ++k) {
+            v128_t isum_vec = wasm_i32x4_splat(0);
+            int shift = 0;
+
+            for (int j = 0; j < 4; ++j) {
+                const int d0 = (sc_ptr[0] & 0xF);
+                const int d1 = (sc_ptr[1] & 0xF);
+                sc_ptr += 2;
+
+                // Process first 16 elements
+                v128_t q2_0 = wasm_v128_load(q2);
+                v128_t q8_0 = wasm_v128_load(q8);
+                v128_t q2_shift_0 = wasm_u8x16_shr(q2_0, shift);
+                v128_t q2_bits_0 = wasm_v128_and(q2_shift_0, wasm_i8x16_splat(0x03));
+
+                // Process next 16 elements
+                v128_t q2_1 = wasm_v128_load(q2 + 16);
+                v128_t q8_1 = wasm_v128_load(q8 + 16);
+                v128_t q2_shift_1 = wasm_u8x16_shr(q2_1, shift);
+                v128_t q2_bits_1 = wasm_v128_and(q2_shift_1, wasm_i8x16_splat(0x03));
+
+                // Calculate dot products
+                v128_t p0 = wasm_i32x4_dot_i16x8(
+                    wasm_i16x8_extend_low_i8x16(q8_0),
+                    wasm_i16x8_extend_low_i8x16(q2_bits_0)
+                );
+                v128_t p1 = wasm_i32x4_dot_i16x8(
+                    wasm_i16x8_extend_high_i8x16(q8_0),
+                    wasm_i16x8_extend_high_i8x16(q2_bits_0)
+                );
+                v128_t p2 = wasm_i32x4_dot_i16x8(
+                    wasm_i16x8_extend_low_i8x16(q8_1),
+                    wasm_i16x8_extend_low_i8x16(q2_bits_1)
+                );
+                v128_t p3 = wasm_i32x4_dot_i16x8(
+                    wasm_i16x8_extend_high_i8x16(q8_1),
+                    wasm_i16x8_extend_high_i8x16(q2_bits_1)
+                );
+
+                // Accumulate scaled results
+                v128_t scaled = wasm_i32x4_add(
+                    wasm_i32x4_mul(wasm_i32x4_add(p0, p1), wasm_i32x4_splat(d0)),
+                    wasm_i32x4_mul(wasm_i32x4_add(p2, p3), wasm_i32x4_splat(d1))
+                );
+
+                isum_vec = wasm_i32x4_add(isum_vec, scaled);
+                q8 += 32;
+                shift += 2;
+            }
+            q2 += 32;
+
+            // Horizontal sum of isum_vec
+            isum_vec = wasm_i32x4_add(isum_vec, wasm_i32x4_shuffle(isum_vec, isum_vec, 2, 3, 0, 1));
+            isum_vec = wasm_i32x4_add(isum_vec, wasm_i32x4_shuffle(isum_vec, isum_vec, 1, 0, 3, 2));
+            isum += wasm_i32x4_extract_lane(isum_vec, 0);
+        }
+
+        const float dall = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf += dall * isum - dmin * summs;
+    }
+
+    *s = sumf;
+
+#else
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * q2 = x[i].qs;
+        const  int8_t * q8 = y[i].qs;
+        const uint8_t * sc = x[i].scales;
+
+        int summs = 0;
+        for (int j = 0; j < 16; ++j) {
+            summs += y[i].bsums[j] * (sc[j] >> 4);
+        }
+
+        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        int isum = 0;
+        int is = 0;
+        int d;
+        for (int k = 0; k < QK_K/128; ++k) {
+            int shift = 0;
+            for (int j = 0; j < 4; ++j) {
+                d = sc[is++] & 0xF;
+                int isuml = 0;
+                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                d = sc[is++] & 0xF;
+                isuml = 0;
+                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                shift += 2;
+                q8 += 32;
+            }
+            q2 += 32;
+        }
+        sumf += dall * isum - dmin * summs;
+    }
+    *s = sumf;
+#endif
+}
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const uint32_t kmask1 = 0x03030303;
+    const uint32_t kmask2 = 0x0f0f0f0f;
+
+    const block_q3_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __wasm_simd128__
+    int8_t  aux8[QK_K];
+    float   sums[8] = {0};
+    uint32_t auxs[4];
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
+        const uint8_t * GGML_RESTRICT hm = x[i].hmask;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
+
+        // Process blocks with SIMD
+        int8_t * a = aux8;
+        uint8_t m = 1;
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int shift = 0; shift <= 6; shift += 2) {
+                v128_t v_m = wasm_i8x16_splat(m);
+                for (int l = 0; l < 32; l += 16) {
+                    v128_t v_q3 = wasm_v128_load(q3 + l);
+                    v128_t v_shift = wasm_i8x16_shr(v_q3, shift);
+                    v128_t v_low2 = wasm_v128_and(v_shift, wasm_i8x16_splat(0x03));
+
+                    v128_t v_hm = wasm_v128_load(hm + l);
+                    v128_t v_mask = wasm_v128_and(v_hm, v_m);
+                    v_mask = wasm_i8x16_ne(v_mask, wasm_i8x16_splat(0));
+
+                    v_low2 = wasm_i8x16_sub(v_low2, wasm_v128_and(wasm_i8x16_splat(4), wasm_v128_not(v_mask)));
+                    wasm_v128_store(a + l, v_low2);
+                }
+                a += 32;
+                m <<= 1;
+            }
+            q3 += 32;
+        }
+
+        // Extract scales
+        memcpy(auxs, x[i].scales, 12);
+        uint32_t tmp = auxs[2];
+        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
+        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
+        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
+        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
+        const int8_t * scales = (const int8_t *)auxs;
+
+        // SIMD dot product with register accumulators
+        v128_t v_acc0 = wasm_i32x4_splat(0);
+        v128_t v_acc1 = wasm_i32x4_splat(0);
+        a = aux8;
+        for (int j = 0; j < QK_K/16; ++j) {
+            const v128_t v_scale = wasm_i16x8_splat(scales[j] - 32);
+
+            // Process 16 elements per iteration
+            for (int k = 0; k < 2; ++k) {
+                const v128_t v_q8 = wasm_i16x8_load8x8(q8);
+                const v128_t v_a = wasm_i16x8_load8x8(a);
+
+                v128_t v_prod = wasm_i16x8_mul(v_q8, v_a);
+                v_prod = wasm_i16x8_mul(v_prod, v_scale);
+
+                v_acc0 = wasm_i32x4_add(v_acc0, wasm_i32x4_extend_low_i16x8(v_prod));
+                v_acc1 = wasm_i32x4_add(v_acc1, wasm_i32x4_extend_high_i16x8(v_prod));
+
+                q8 += 8;
+                a += 8;
+            }
+        }
+
+        // Accumulate results
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const v128_t v_d = wasm_f32x4_splat(d);
+        v128_t v_sum = wasm_f32x4_add(
+            wasm_f32x4_mul(wasm_f32x4_convert_i32x4(v_acc0), v_d),
+            wasm_f32x4_mul(wasm_f32x4_convert_i32x4(v_acc1), v_d)
+        );
+
+        // Accumulate into sums vector
+        wasm_v128_store(sums, wasm_f32x4_add(wasm_v128_load(sums), v_sum));
+    }
+
+    // Horizontal sum
+    v128_t v_sum = wasm_f32x4_add(wasm_v128_load(sums), wasm_v128_load(sums + 4));
+    sumf = wasm_f32x4_extract_lane(v_sum, 0) +
+           wasm_f32x4_extract_lane(v_sum, 1) +
+           wasm_f32x4_extract_lane(v_sum, 2) +
+           wasm_f32x4_extract_lane(v_sum, 3);
+
+    *s = sumf;
+
+#else
+    // scalar version
+    // This function is written like this so the compiler can manage to vectorize most of it
+    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
+    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
+    // The ideal situation would be if we could just write the code once, and the compiler would
+    // automatically produce the best possible set of machine instructions, instead of us having to manually
+    // write vectorized versions for AVX, ARM_NEON, etc.
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    uint32_t auxs[4];
+    const int8_t * scales = (const int8_t*)auxs;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
+        const uint8_t * GGML_RESTRICT hm = x[i].hmask;
+        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * GGML_RESTRICT a = aux8;
+        uint8_t m = 1;
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            q3 += 32;
+        }
+        a = aux8;
+
+        memcpy(auxs, x[i].scales, 12);
+        uint32_t tmp = auxs[2];
+        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
+        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
+        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
+        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
+        for (int j = 0; j < QK_K/16; ++j) {
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+
+#endif
+
+}
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#if defined __wasm_simd128__
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); // Corrected sign
+
+        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
+
+        // Process scales and mins
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        // Sum mins * q8sums
+        int32_t sumi = 0;
+        const int16_t * GGML_RESTRICT q8sums = y[i].bsums;
+        const uint8_t * m = (const uint8_t *)&utmp[2];
+        for (int j = 0; j < 16; j += 2) {
+            sumi += (q8sums[j] + q8sums[j+1]) * m[j/2];
+        }
+        sumf -= dmin * sumi;
+
+        int32_t sumi1 = 0;
+        int32_t sumi2 = 0;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            // Load 64 4-bit weights (32 bytes)
+            const v128_t q4x0 = wasm_v128_load(q4);
+            const v128_t q4x1 = wasm_v128_load(q4 + 16);
+            q4 += 32;
+
+            // Split into low/high nibbles
+            const v128_t q4l0 = wasm_v128_and(q4x0, wasm_i8x16_splat(0x0F));
+            const v128_t q4h0 = wasm_u8x16_shr(q4x0, 4);
+            const v128_t q4l1 = wasm_v128_and(q4x1, wasm_i8x16_splat(0x0F));
+            const v128_t q4h1 = wasm_u8x16_shr(q4x1, 4);
+
+            // Load 64 8-bit values (64 bytes)
+            const v128_t q8x0 = wasm_v128_load(q8);
+            const v128_t q8x1 = wasm_v128_load(q8 + 16);
+            const v128_t q8x2 = wasm_v128_load(q8 + 32);
+            const v128_t q8x3 = wasm_v128_load(q8 + 48);
+            q8 += 64;
+
+            // Low nibble products
+            v128_t vacc1 = wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_low_i8x16(q4l0),
+                wasm_i16x8_extend_low_i8x16(q8x0)
+            );
+            vacc1 = wasm_i32x4_add(vacc1, wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_high_i8x16(q4l0),
+                wasm_i16x8_extend_high_i8x16(q8x0)
+            ));
+            vacc1 = wasm_i32x4_add(vacc1, wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_low_i8x16(q4l1),
+                wasm_i16x8_extend_low_i8x16(q8x1)
+            ));
+            vacc1 = wasm_i32x4_add(vacc1, wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_high_i8x16(q4l1),
+                wasm_i16x8_extend_high_i8x16(q8x1)
+            ));
+
+            // High nibble products
+            v128_t vacc2 = wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_low_i8x16(q4h0),
+                wasm_i16x8_extend_low_i8x16(q8x2)
+            );
+            vacc2 = wasm_i32x4_add(vacc2, wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_high_i8x16(q4h0),
+                wasm_i16x8_extend_high_i8x16(q8x2)
+            ));
+            vacc2 = wasm_i32x4_add(vacc2, wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_low_i8x16(q4h1),
+                wasm_i16x8_extend_low_i8x16(q8x3)
+            ));
+            vacc2 = wasm_i32x4_add(vacc2, wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_high_i8x16(q4h1),
+                wasm_i16x8_extend_high_i8x16(q8x3)
+            ));
+
+            // Accumulate scaled results
+            int32_t vacc1_sum = wasm_i32x4_extract_lane(vacc1, 0) + wasm_i32x4_extract_lane(vacc1, 1) +
+                                wasm_i32x4_extract_lane(vacc1, 2) + wasm_i32x4_extract_lane(vacc1, 3);
+            sumi1 += vacc1_sum * scales[2*j];
+
+            int32_t vacc2_sum = wasm_i32x4_extract_lane(vacc2, 0) + wasm_i32x4_extract_lane(vacc2, 1) +
+                                wasm_i32x4_extract_lane(vacc2, 2) + wasm_i32x4_extract_lane(vacc2, 3);
+            sumi2 += vacc2_sum * scales[2*j+1];
+        }
+
+        sumf += d * (sumi1 + sumi2);
+    }
+
+    *s = sumf;
+
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * GGML_RESTRICT a = aux8;
+        for (int j = 0; j < QK_K/64; ++j) {
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
+            a += 32;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            a += 32; q4 += 32;
+        }
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#if defined __wasm_simd128__
+    //const uint8_t * scales = (const uint8_t*)&utmp[0];
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); // Fixed sign
+
+        const uint8_t * GGML_RESTRICT q5 = x[i].qs;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
+
+        // Process scales and mins
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        // Sum mins * q8sums
+        int32_t sumi_mins = 0;
+        const int16_t * GGML_RESTRICT q8sums = y[i].bsums;
+        const uint8_t * m = (const uint8_t *)&utmp[2];
+        for (int j = 0; j < 16; j += 2) {
+            sumi_mins += (q8sums[j] + q8sums[j+1]) * m[j/2];
+        }
+        sumf -= dmin * sumi_mins; // Correct subtraction
+
+        v128_t qh0 = wasm_v128_load(qh);
+        v128_t qh1 = wasm_v128_load(qh + 16);
+        const uint8_t * sc = (const uint8_t *)utmp;
+
+        int32_t sumi = 0;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            const int shift = j * 2;
+            v128_t qh_shift0 = wasm_u8x16_shr(qh0, shift);
+            v128_t qh_shift1 = wasm_u8x16_shr(qh1, shift);
+
+            v128_t qh_low0 = wasm_i8x16_shl(wasm_v128_and(qh_shift0, wasm_i8x16_splat(0x01)), 4);
+            v128_t qh_high0 = wasm_i8x16_shl(wasm_v128_and(qh_shift0, wasm_i8x16_splat(0x02)), 3);
+            v128_t qh_low1 = wasm_i8x16_shl(wasm_v128_and(qh_shift1, wasm_i8x16_splat(0x01)), 4);
+            v128_t qh_high1 = wasm_i8x16_shl(wasm_v128_and(qh_shift1, wasm_i8x16_splat(0x02)), 3);
+
+            v128_t q5_0 = wasm_v128_load(q5);
+            v128_t q5_1 = wasm_v128_load(q5 + 16);
+            q5 += 32;
+
+            v128_t q5l_0 = wasm_v128_or(wasm_v128_and(q5_0, wasm_i8x16_splat(0x0F)), qh_low0);
+            v128_t q5h_0 = wasm_v128_or(wasm_u8x16_shr(q5_0, 4), qh_high0);
+            v128_t q5l_1 = wasm_v128_or(wasm_v128_and(q5_1, wasm_i8x16_splat(0x0F)), qh_low1);
+            v128_t q5h_1 = wasm_v128_or(wasm_u8x16_shr(q5_1, 4), qh_high1);
+
+            v128_t q8_0 = wasm_v128_load(q8);
+            v128_t q8_1 = wasm_v128_load(q8 + 16);
+            v128_t q8_2 = wasm_v128_load(q8 + 32);
+            v128_t q8_3 = wasm_v128_load(q8 + 48);
+            q8 += 64;
+
+            // Process low quants
+            v128_t pl0 = wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_low_i8x16(q5l_0),
+                wasm_i16x8_extend_low_i8x16(q8_0)
+            );
+            pl0 = wasm_i32x4_add(pl0, wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_high_i8x16(q5l_0),
+                wasm_i16x8_extend_high_i8x16(q8_0)
+            ));
+            v128_t pl1 = wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_low_i8x16(q5l_1),
+                wasm_i16x8_extend_low_i8x16(q8_1)
+            );
+            pl1 = wasm_i32x4_add(pl1, wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_high_i8x16(q5l_1),
+                wasm_i16x8_extend_high_i8x16(q8_1)
+            ));
+            v128_t sum_low = wasm_i32x4_add(pl0, pl1);
+
+            // Process high quants
+            v128_t ph0 = wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_low_i8x16(q5h_0),
+                wasm_i16x8_extend_low_i8x16(q8_2)
+            );
+            ph0 = wasm_i32x4_add(ph0, wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_high_i8x16(q5h_0),
+                wasm_i16x8_extend_high_i8x16(q8_2)
+            ));
+            v128_t ph1 = wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_low_i8x16(q5h_1),
+                wasm_i16x8_extend_low_i8x16(q8_3)
+            );
+            ph1 = wasm_i32x4_add(ph1, wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_high_i8x16(q5h_1),
+                wasm_i16x8_extend_high_i8x16(q8_3)
+            ));
+            v128_t sum_high = wasm_i32x4_add(ph0, ph1);
+
+            // Accumulate with scale factors
+            int32_t sl = wasm_i32x4_extract_lane(sum_low, 0) + wasm_i32x4_extract_lane(sum_low, 1) +
+                        wasm_i32x4_extract_lane(sum_low, 2) + wasm_i32x4_extract_lane(sum_low, 3);
+            int32_t sh = wasm_i32x4_extract_lane(sum_high, 0) + wasm_i32x4_extract_lane(sum_high, 1) +
+                        wasm_i32x4_extract_lane(sum_high, 2) + wasm_i32x4_extract_lane(sum_high, 3);
+
+            sumi += sl * sc[2*j] + sh * sc[2*j+1];
+        }
+
+        sumf += d * sumi;
+    }
+
+    *s = sumf;
+
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+        const uint8_t * GGML_RESTRICT hm = x[i].qh;
+        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * GGML_RESTRICT a = aux8;
+        uint8_t m = 1;
+        for (int j = 0; j < QK_K/64; ++j) {
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            q4 += 32;
+        }
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q6_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __wasm_simd128__
+    int8_t aux8[QK_K] __attribute__((aligned(16)));
+    int32_t aux32[8] __attribute__((aligned(16))) = {0};
+    float sums[8] __attribute__((aligned(16))) = {0};
+
+    for (int i = 0; i < nb; ++i) {
+        // Unpack 6-bit quantized data into aux8 (unchanged)
+        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        int8_t * a = aux8;
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) {
+                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
+                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
+                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
+                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
+            }
+            a += 128;
+            q4 += 64;
+            qh += 32;
+        }
+
+        const int8_t * GGML_RESTRICT a_ptr = aux8;
+        const int8_t * GGML_RESTRICT q8 = y[i].qs;
+        v128_t acc0 = wasm_i32x4_splat(0);
+        v128_t acc1 = wasm_i32x4_splat(0);
+
+        for (int j = 0; j < QK_K/16; ++j) {
+            const int scale = x[i].scales[j];
+            const v128_t vscale = wasm_i32x4_splat(scale);
+
+            // Load 16 elements from a and q8
+            const v128_t a_vec = wasm_v128_load(a_ptr);
+            const v128_t q8_vec = wasm_v128_load(q8);
+
+            // Process low 8 elements
+            v128_t a_low = wasm_i16x8_extend_low_i8x16(a_vec);
+            v128_t q8_low = wasm_i16x8_extend_low_i8x16(q8_vec);
+            v128_t prod_low = wasm_i16x8_mul(a_low, q8_low);
+            v128_t prod_lo_lo = wasm_i32x4_extend_low_i16x8(prod_low);
+            v128_t prod_lo_hi = wasm_i32x4_extend_high_i16x8(prod_low);
+
+            // Process high 8 elements
+            v128_t a_high = wasm_i16x8_extend_high_i8x16(a_vec);
+            v128_t q8_high = wasm_i16x8_extend_high_i8x16(q8_vec);
+            v128_t prod_high = wasm_i16x8_mul(a_high, q8_high);
+            v128_t prod_hi_lo = wasm_i32x4_extend_low_i16x8(prod_high);
+            v128_t prod_hi_hi = wasm_i32x4_extend_high_i16x8(prod_high);
+
+            // Scale and accumulate
+            prod_lo_lo = wasm_i32x4_mul(prod_lo_lo, vscale);
+            prod_lo_hi = wasm_i32x4_mul(prod_lo_hi, vscale);
+            prod_hi_lo = wasm_i32x4_mul(prod_hi_lo, vscale);
+            prod_hi_hi = wasm_i32x4_mul(prod_hi_hi, vscale);
+
+            acc0 = wasm_i32x4_add(acc0, wasm_i32x4_add(prod_lo_lo, prod_hi_lo));
+            acc1 = wasm_i32x4_add(acc1, wasm_i32x4_add(prod_lo_hi, prod_hi_hi));
+
+            a_ptr += 16;
+            q8 += 16;
+        }
+
+        // Store accumulated results
+        wasm_v128_store(&aux32[0], acc0);
+        wasm_v128_store(&aux32[4], acc1);
+
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) {
+            sums[l] += d * aux32[l];
+        }
+    }
+
+    // Sum final results
+    float sumf = 0;
+    for (int l = 0; l < 8; ++l) {
+        sumf += sums[l];
+    }
+    *s = sumf;
+
+#else
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * GGML_RESTRICT a = aux8;
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) {
+                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
+                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
+                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
+                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
+            }
+            a  += 128;
+            q4 += 64;
+            qh += 32;
+        }
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/16; ++j) {
+            int scale = x[i].scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+

ggml/src/ggml-cpu/arch/x86/cpu-feats.cpp

@@ -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/common.h

@@ -1,7 +1,7 @@
 #pragma once
 
 #include "ggml.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-impl.h"
 

ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -506,3 +506,25 @@ void ggml_barrier(struct ggml_threadpool * tp);
 #ifdef __cplusplus
 }
 #endif
+
+#define GGML_DO_PRAGMA_(x) _Pragma (#x)
+#define GGML_DO_PRAGMA(x) GGML_DO_PRAGMA_(x)
+#if defined(GGML_CPU_GENERIC) || defined(__HIPCC__)
+// Note for Apple targets:
+// - clang: aliases are not supported on darwin
+// - all native kernels need to be implemented in both x86 and arm files
+// - on iOS, tvOS, and visionOS, if cmake cannot determine the target architecture, all `_generic` names are replaced by defines
+# define GGML_WEAK_ALIAS(name, alias)
+#elif defined(__GNUC__)
+// GCC/Clang on *nix
+# define GGML_WEAK_ALIAS(name, alias) GGML_DO_PRAGMA(weak name = alias) // NOLINT
+#elif defined(_MSC_VER) && defined (_WIN64)
+// MSVC
+// Note: C name mangling varies across different calling conventions
+// see https://learn.microsoft.com/en-us/cpp/build/reference/decorated-names?view=msvc-170
+# define GGML_WEAK_ALIAS(name, alias) GGML_DO_PRAGMA(comment(linker, "/alternatename:" #name "=" #alias))
+#else
+# error "Unsupported compiler for GGML_WEAK_ALIAS"
+#endif
+
+#define GGML_CPU_NATIVE_IMPL(name) GGML_WEAK_ALIAS(name, name ## _generic)

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

@@ -3,11 +3,11 @@
 
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
-#include "ggml-cpu-quants.h"
+#include "quants.h"
 #include "ggml-threading.h"
 #include "unary-ops.h"
 #include "binary-ops.h"

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

@@ -1,8 +1,8 @@
 #include "ggml-backend.h"
 #include "ggml-backend-impl.h"
 #include "ggml-cpu.h"
-#include "ggml-cpu-aarch64.h"
-#include "ggml-cpu-traits.h"
+#include "repack.h"
+#include "traits.h"
 #include "ggml-impl.h"
 #include "amx/amx.h"
 
@@ -11,7 +11,7 @@
 #include <vector>
 
 #ifdef GGML_USE_CPU_HBM
-#    include "ggml-cpu-hbm.h"
+#    include "hbm.h"
 #endif
 
 #ifdef GGML_USE_CPU_KLEIDIAI
@@ -51,9 +51,9 @@ std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type
         }
 #endif
 
-#ifdef GGML_USE_CPU_AARCH64
-        if (ggml_backend_cpu_aarch64_buffer_type()) {
-            bufts.push_back(ggml_backend_cpu_aarch64_buffer_type());
+#ifdef GGML_USE_CPU_REPACK
+        if (ggml_backend_cpu_repack_buffer_type()) {
+            bufts.push_back(ggml_backend_cpu_repack_buffer_type());
         }
 #endif
 
@@ -596,8 +596,8 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
     #ifdef GGML_USE_CPU_KLEIDIAI
         features.push_back({ "KLEIDIAI", "1" });
     #endif
-    #ifdef GGML_USE_CPU_AARCH64
-        features.push_back({ "AARCH64_REPACK", "1" });
+    #ifdef GGML_USE_CPU_REPACK
+        features.push_back({ "REPACK", "1" });
     #endif
 
         features.push_back({ nullptr, nullptr });

ggml/src/ggml-cpu/hbm.cpp

@@ -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

@@ -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/kleidiai.cpp

@@ -26,7 +26,7 @@
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"
 #include "ggml-threading.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 
 #include "kernels.h"
 

ggml/src/ggml-cpu/quants.c

@@ -0,0 +1,1179 @@
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+
+#include "ggml-cpu-impl.h"
+#include "ggml-quants.h"
+#include "quants.h"
+
+#include <string.h>
+#include <assert.h>
+#include <float.h>
+#include <stdlib.h> // for qsort
+#include <stdio.h>  // for GGML_ASSERT
+
+#define GROUP_MAX_EPS 1e-15f
+#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
+#define GROUP_MAX_EPS_IQ2_S 1e-8f
+#define GROUP_MAX_EPS_IQ1_M 1e-7f
+#define GROUP_MAX_EPS_IQ1_S 1e-12f
+
+#define UNUSED GGML_UNUSED
+
+void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    quantize_row_q4_0_ref(x, y, k);
+}
+
+void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    quantize_row_q4_1_ref(x, y, k);
+}
+
+void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    quantize_row_q5_0_ref(x, y, k);
+}
+
+void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    quantize_row_q5_1_ref(x, y, k);
+}
+
+void quantize_row_q8_0_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    quantize_row_q8_0_ref(x, y, k);
+}
+GGML_CPU_NATIVE_IMPL(quantize_row_q8_0)
+
+void quantize_row_q8_1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    quantize_row_q8_1_ref(x, y, k);
+}
+GGML_CPU_NATIVE_IMPL(quantize_row_q8_1)
+
+//
+// 2-6 bit quantization in super-blocks
+//
+
+//========================- 2-bit (de)-quantization
+
+void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    quantize_row_q2_K_ref(x, vy, k);
+}
+
+//========================= 3-bit (de)-quantization
+
+void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    quantize_row_q3_K_ref(x, vy, k);
+}
+
+// ====================== 4-bit (de)-quantization
+
+void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(k % QK_K == 0);
+    block_q4_K * GGML_RESTRICT y = vy;
+    quantize_row_q4_K_ref(x, y, k);
+}
+
+// ====================== 5-bit (de)-quantization
+
+void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(k % QK_K == 0);
+    block_q5_K * GGML_RESTRICT y = vy;
+    quantize_row_q5_K_ref(x, y, k);
+}
+
+// ====================== 6-bit (de)-quantization
+
+void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(k % QK_K == 0);
+    block_q6_K * GGML_RESTRICT y = vy;
+    quantize_row_q6_K_ref(x, y, k);
+}
+
+// ====================== Ternary (de)-quantization (BitNet b1.58 and TriLMs)
+
+void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(k % QK_K == 0);
+    block_tq1_0 * GGML_RESTRICT y = vy;
+    quantize_row_tq1_0_ref(x, y, k);
+}
+
+void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(k % QK_K == 0);
+    block_tq2_0 * GGML_RESTRICT y = vy;
+    quantize_row_tq2_0_ref(x, y, k);
+}
+
+//===================================== Q8_K ==============================================
+
+void quantize_row_q8_K_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    quantize_row_q8_K_ref(x, y, k);
+}
+GGML_CPU_NATIVE_IMPL(quantize_row_q8_K)
+
+//===================================== Dot products =================================
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
+            const int v1 = (x[ib].qs[j] >>   4) - 8;
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+    }
+
+    *s = sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q4_0_q8_0)
+
+// TODO: add WASM SIMD
+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) {
+    const int qk = QK8_1;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F);
+            const int v1 = (x[ib].qs[j] >>   4);
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q4_1_q8_1)
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
+            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
+
+            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
+            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+    }
+
+    *s = sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q5_0_q8_0)
+
+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) {
+    const int qk = QK8_1;
+    const int nb = n / qk;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_1);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
+            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
+
+            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
+            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q5_1_q8_1)
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q8_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+    for (; ib < nb; ++ib) {
+        int sumi = 0;
+
+        for (int j = 0; j < qk; j++) {
+            sumi += x[ib].qs[j]*y[ib].qs[j];
+        }
+
+        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+    }
+
+    *s = sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q8_0_q8_0)
+
+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) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_tq1_0 * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    const uint8_t pow3[6] = {1, 3, 9, 27, 81, 243};
+
+    float sumf = 0.0f;
+
+    for (int i = 0; i < nb; ++i) {
+        int sum = 0;
+
+        for (size_t j = 0; j < sizeof(x->qs) - sizeof(x->qs) % 32; j += 32) {
+            for (size_t l = 0; l < 5; ++l) {
+                for (size_t m = 0; m < 32; ++m) {
+                    uint8_t q = x[i].qs[j + m] * pow3[l];
+                    uint16_t xi = ((uint16_t) q * 3) >> 8;
+                    sum += (xi - 1) * y[i].qs[j*5 + l*32 + m];
+                }
+            }
+        }
+        for (size_t j = sizeof(x->qs) - sizeof(x->qs) % 32; j < sizeof(x->qs); j += 16) {
+            for (size_t l = 0; l < 5; ++l) {
+                for (size_t m = 0; m < 16; ++m) {
+                    uint8_t q = x[i].qs[j + m] * pow3[l];
+                    uint16_t xi = ((uint16_t) q * 3) >> 8;
+                    sum += (xi - 1) * y[i].qs[j*5 + l*16 + m];
+                }
+            }
+        }
+
+        for (size_t l = 0; l < 4; ++l) {
+            for (size_t j = 0; j < sizeof(x->qh); ++j) {
+                uint8_t q = x[i].qh[j] * pow3[l];
+                uint16_t xi = ((uint16_t) q * 3) >> 8;
+                sum += (xi - 1) * y[i].qs[sizeof(x->qs)*5 + l*sizeof(x->qh) + j];
+            }
+        }
+
+        sumf += (float) sum * (GGML_FP16_TO_FP32(x[i].d) * y[i].d);
+    }
+
+    *s = sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_tq1_0_q8_K)
+
+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) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_tq2_0 * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+    float sumf = 0.0f;
+
+    for (int i = 0; i < nb; ++i) {
+        int32_t sumi = 0;
+
+        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
+            for (size_t l = 0; l < 4; ++l) {
+                for (size_t k = 0; k < 32; ++k) {
+                    sumi += y[i].qs[j*4 + l*32 + k] * (((x[i].qs[j + k] >> (l*2)) & 3) - 1);
+                }
+            }
+        }
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+
+        sumf += (float) sumi * d;
+    }
+
+    *s = sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_tq2_0_q8_K)
+
+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) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q2_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * q2 = x[i].qs;
+        const  int8_t * q8 = y[i].qs;
+        const uint8_t * sc = x[i].scales;
+
+        int summs = 0;
+        for (int j = 0; j < 16; ++j) {
+            summs += y[i].bsums[j] * (sc[j] >> 4);
+        }
+
+        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        int isum = 0;
+        int is = 0;
+        int d;
+        for (int k = 0; k < QK_K/128; ++k) {
+            int shift = 0;
+            for (int j = 0; j < 4; ++j) {
+                d = sc[is++] & 0xF;
+                int isuml = 0;
+                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                d = sc[is++] & 0xF;
+                isuml = 0;
+                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                shift += 2;
+                q8 += 32;
+            }
+            q2 += 32;
+        }
+        sumf += dall * isum - dmin * summs;
+    }
+    *s = sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q2_K_q8_K)
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const uint32_t kmask1 = 0x03030303;
+    const uint32_t kmask2 = 0x0f0f0f0f;
+
+    const block_q3_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    // scalar version
+    // This function is written like this so the compiler can manage to vectorize most of it
+    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
+    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
+    // The ideal situation would be if we could just write the code once, and the compiler would
+    // automatically produce the best possible set of machine instructions, instead of us having to manually
+    // write vectorized versions for AVX, ARM_NEON, etc.
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    uint32_t auxs[4];
+    const int8_t * scales = (const int8_t*)auxs;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
+        const uint8_t * GGML_RESTRICT hm = x[i].hmask;
+        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * GGML_RESTRICT a = aux8;
+        uint8_t m = 1;
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            q3 += 32;
+        }
+        a = aux8;
+
+        memcpy(auxs, x[i].scales, 12);
+        uint32_t tmp = auxs[2];
+        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
+        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
+        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
+        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
+        for (int j = 0; j < QK_K/16; ++j) {
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q3_K_q8_K)
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * GGML_RESTRICT a = aux8;
+        for (int j = 0; j < QK_K/64; ++j) {
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
+            a += 32;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            a += 32; q4 += 32;
+        }
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q4_K_q8_K)
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+        const uint8_t * GGML_RESTRICT hm = x[i].qh;
+        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * GGML_RESTRICT a = aux8;
+        uint8_t m = 1;
+        for (int j = 0; j < QK_K/64; ++j) {
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            q4 += 32;
+        }
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q5_K_q8_K)
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q6_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * GGML_RESTRICT a = aux8;
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) {
+                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
+                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
+                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
+                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
+            }
+            a  += 128;
+            q4 += 64;
+            qh += 32;
+        }
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/16; ++j) {
+            int scale = x[i].scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q6_K_q8_K)
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_xxs * GGML_RESTRICT x = vx;
+    const block_q8_K    * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    uint32_t aux32[2];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
+        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
+        int32_t bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            memcpy(aux32, q2, 2*sizeof(uint32_t));
+            q2 += 4;
+            const uint32_t ls = 2*(aux32[1] >> 28) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
+                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.125f * sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq2_xxs_q8_K)
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_xs * GGML_RESTRICT x = vx;
+    const block_q8_K   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
+        const uint8_t  * GGML_RESTRICT sc = x[i].scales;
+        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
+        int32_t bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1;
+            const uint16_t ls2 = 2*(sc[ib32] >>  4) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 2; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls1;
+            sumi = 0;
+            for (int l = 2; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls2;
+            q2 += 4;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.125f * sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq2_xs_q8_K)
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const int8_t  * q8 = y[i].qs;
+        const uint8_t * qs = x[i].qs;
+        const uint8_t * qh = x[i].qh;
+        const uint8_t * signs = qs + QK_K/8;
+
+        int bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
+            int ls2 = 1 + 2*(x[i].scales[ib32] >>  4);
+            int sumi1 = 0, sumi2 = 0;
+            for (int l = 0; l < 2; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
+                for (int j = 0; j < 8; ++j) {
+                    sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            for (int l = 2; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
+                for (int j = 0; j < 8; ++j) {
+                    sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += ls1 * sumi1 + ls2 * sumi2;
+            qs += 4;
+            signs += 4;
+        }
+
+        sumf += d * bsum;
+    }
+
+    *s = 0.125f * sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq2_s_q8_K)
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq3_xxs * GGML_RESTRICT x = vx;
+    const block_q8_K    * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    uint32_t aux32;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
+        const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
+        int32_t bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t);
+            const uint32_t ls = 2*(aux32 >> 28) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
+                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
+                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            q3 += 8;
+            bsum += sumi * ls;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.25f * sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq3_xxs_q8_K)
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq3_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * GGML_RESTRICT qs = x[i].qs;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const uint8_t * GGML_RESTRICT signs = x[i].signs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
+        int32_t bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1;
+            const uint32_t ls2 = 2*(x[i].scales[ib32/2] >>  4) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
+                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            qs += 8;
+            signs += 4;
+            bsum += sumi * ls1;
+            sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
+                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            qs += 8;
+            signs += 4;
+            bsum += sumi * ls2;
+        }
+        sumf += d * bsum;
+    }
+    *s = sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq3_s_q8_K)
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq1_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        int sumi = 0, sumi1 = 0;
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            const int ls = 2*((qh[ib] >> 12) & 7) + 1;
+            const int delta = qh[ib] & 0x8000 ? -1 : 1;
+            int lsum = 0;
+            for (int l = 0; l < 4; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
+                for (int j = 0; j < 8; ++j) {
+                    lsum += q8[j] * grid[j];
+                }
+                q8 += 8;
+            }
+            sumi  += ls * lsum;
+            sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
+            qs += 4;
+        }
+
+        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+    }
+
+    *s = sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq1_s_q8_K)
+
+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) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq1_m * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    iq1m_scale_t scale;
+
+    int sum1[2], sum2[2], delta[4];
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        const uint16_t * sc = (const uint16_t *)x[i].scales;
+
+        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+
+        int sumi1 = 0, sumi2 = 0;
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            delta[0] = qh[0] & 0x08 ? -1 : 1;
+            delta[1] = qh[0] & 0x80 ? -1 : 1;
+            delta[2] = qh[1] & 0x08 ? -1 : 1;
+            delta[3] = qh[1] & 0x80 ? -1 : 1;
+            sum1[0] = sum1[1] = sum2[0] = sum2[1] = 0;
+            for (int l = 0; l < 4; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((uint16_t)qh[l/2] << (8 - 4*(l%2))) & 0x700)));
+                int lsum1 = 0, lsum2 = 0;
+                for (int j = 0; j < 8; ++j) {
+                    lsum1 += q8[j] * grid[j];
+                    lsum2 += q8[j];
+                }
+                q8 += 8;
+                sum1[l/2] += lsum1;
+                sum2[l/2] += lsum2*delta[l];
+            }
+
+            const int ls1 = 2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1;
+            const int ls2 = 2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1;
+
+            sumi1 += sum1[0] * ls1 + sum1[1] * ls2;
+            sumi2 += sum2[0] * ls1 + sum2[1] * ls2;
+            qs += 4;
+            qh += 2;
+        }
+
+        sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
+    }
+
+    *s = sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq1_m_q8_K)
+
+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) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK4_NL == 0);
+    static_assert(QK4_NL == QK8_0, "QK4_NL and QK8_0 must be the same");
+
+    const block_iq4_nl * GGML_RESTRICT x = vx;
+    const block_q8_0   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK4_NL;
+
+    int ib = 0;
+    float sumf = 0;
+
+    for (; ib < nb; ++ib) {
+        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        int sumi1 = 0, sumi2 = 0;
+        for (int j = 0; j < QK4_NL/2; ++j) {
+            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
+            sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >>  4];
+        }
+        sumf += d * (sumi1 + sumi2);
+    }
+    *s = sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq4_nl_q8_0)
+
+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) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK_K == 0);
+
+    const block_iq4_xs * GGML_RESTRICT x = vx;
+    const block_q8_K   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    float sumf = 0;
+    for (int ibl = 0; ibl < nb; ++ibl) {
+        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        uint16_t h = x[ibl].scales_h;
+        const uint8_t * qs = x[ibl].qs;
+        const int8_t  * q8 = y[ibl].qs;
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
+            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
+            h >>= 4;
+            const float d1 = d4d8*(ls1 - 32);
+            const float d2 = d4d8*(ls2 - 32);
+            int sumi1 = 0, sumi2 = 0;
+            for (int j = 0; j < 16; ++j) {
+                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
+            }
+            sumf += d1 * (sumi1 + sumi2);
+            qs += 16;
+            q8 += 32;
+            sumi1 = sumi2 = 0;
+            for (int j = 0; j < 16; ++j) {
+                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
+            }
+            sumf += d2 * (sumi1 + sumi2);
+            qs += 16;
+            q8 += 32;
+        }
+    }
+    *s = sumf;
+}
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq4_xs_q8_K)
+
+// ============================ 4-bit non-linear quants
+
+void quantize_row_iq4_nl(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    assert(k % QK4_NL == 0);
+    quantize_row_iq4_nl_ref(x, y, k);
+}
+
+void quantize_row_iq4_xs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    assert(k % QK_K == 0);
+    quantize_iq4_xs(x, y, 1, k, NULL);
+}

ggml/src/ggml-cpu/quants.h

@@ -0,0 +1,116 @@
+#pragma once
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
+#include "ggml.h"
+
+// GGML CPU internal header
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Quantization
+void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+// Dot product
+void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_m_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+// Generic implementation
+void quantize_row_q8_0_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void quantize_row_q8_1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void quantize_row_q8_K_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void ggml_vec_dot_q4_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q2_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q3_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy,  size_t by, int nrc);
+void ggml_vec_dot_q6_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_m_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+#if defined(GGML_CPU_GENERIC)
+#define quantize_row_q8_0_generic quantize_row_q8_0
+#define quantize_row_q8_1_generic quantize_row_q8_1
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_q4_0_q8_0_generic ggml_vec_dot_q4_0_q8_0
+#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
+#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
+#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
+#define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
+#define ggml_vec_dot_q3_K_q8_K_generic ggml_vec_dot_q3_K_q8_K
+#define ggml_vec_dot_q4_K_q8_K_generic ggml_vec_dot_q4_K_q8_K
+#define ggml_vec_dot_q5_K_q8_K_generic ggml_vec_dot_q5_K_q8_K
+#define ggml_vec_dot_q6_K_q8_K_generic ggml_vec_dot_q6_K_q8_K
+#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
+#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
+#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
+#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
+#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
+#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
+#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
+#endif
+
+#ifdef __cplusplus
+}
+#endif

ggml/src/ggml-cpu/repack.cpp

@@ -0,0 +1,1566 @@
+#define GGML_COMMON_IMPL_CPP
+#define GGML_COMMON_DECL_CPP
+#include "ggml-common.h"
+#include "ggml-backend-impl.h"
+
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+#include "ggml-cpu-impl.h"
+#include "traits.h"
+
+#include <cmath>
+#include <cstring>
+#include <cassert>
+#include <cstdlib> // for qsort
+#include <cstdio>  // for GGML_ASSERT
+
+#include "repack.h"
+
+#if defined(__GNUC__)
+#pragma GCC diagnostic ignored "-Woverlength-strings"
+#endif
+
+#define UNUSED GGML_UNUSED
+
+static inline int nearest_int(float fval) {
+    assert(fabsf(fval) <= 4194303.f);
+    float val = fval + 12582912.f;
+    int i; memcpy(&i, &val, sizeof(int));
+    return (i & 0x007fffff) - 0x00400000;
+}
+
+// Functions to create the interleaved data layout formats
+
+// interleave 4 block_q4_0s in blocks of blck_size_interleave
+// returns an interleaved block_q4_0x4
+// in the interleaved block_q4_0x4, place deltas for 4 block_q4_0 blocks
+// first, then interleave quants from 4 block_q4_0s in blocks of blck_size_interleave
+//
+// - in                  : an array of block_q4_0 pointers
+// - blck_size_interleave : the block_q4_0 quants bytes are interleaved in blocks of
+//                         blck_size_interleave bytes
+// - xor_mask            : the mask to convert the nibbles in block_q4_0 quants bytes
+//                         from bias offset form to pure sign form (this saves subtract
+//                         operations durin unpacking)
+//
+
+extern "C" {
+
+void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(QK8_0 == 32);
+    assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
+
+    block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy;
+
+    // scalar
+    const int blck_size_interleave = 4;
+    float srcv[4][QK8_0];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            float amax = 0.0f; // absolute max
+
+            for (int j = 0; j < QK8_0; j++) {
+                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
+                amax = MAX(amax, fabsf(srcv[row_iter][j]));
+            }
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < QK8_0 * 4; j++) {
+            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
+            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
+            src_offset += (j % blck_size_interleave);
+
+            float x0 = srcv[src_id][src_offset] * id[src_id];
+            y[i].qs[j] = roundf(x0);
+        }
+    }
+}
+GGML_CPU_NATIVE_IMPL(ggml_quantize_mat_q8_0_4x4)
+
+void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(QK8_0 == 32);
+    assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
+
+    block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy;
+
+    // scalar
+    const int blck_size_interleave = 8;
+    float srcv[4][QK8_0];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            float amax = 0.0f; // absolute max
+
+            for (int j = 0; j < QK8_0; j++) {
+                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
+                amax = MAX(amax, fabsf(srcv[row_iter][j]));
+            }
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < QK8_0 * 4; j++) {
+            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
+            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
+            src_offset += (j % blck_size_interleave);
+
+            float x0 = srcv[src_id][src_offset] * id[src_id];
+            y[i].qs[j] = roundf(x0);
+        }
+    }
+}
+GGML_CPU_NATIVE_IMPL(ggml_quantize_mat_q8_0_4x8)
+
+void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(QK_K == 256);
+    assert(k % QK_K == 0);
+    const int nb = k / QK_K;
+
+    block_q8_Kx4 * GGML_RESTRICT y = (block_q8_Kx4 *) vy;
+
+    // scalar
+    const int blck_size_interleave = 8;
+    float srcv[4][QK_K];
+    float iscale[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            float amax = 0.0f; // absolute max
+            float max = 0;
+
+            for (int j = 0; j < QK_K; j++) {
+                srcv[row_iter][j] = x[row_iter * k + i * QK_K + j];
+                // Update the maximum value of the corresponding super block
+                if(amax < fabsf(srcv[row_iter][j])) {
+                    amax = fabsf(srcv[row_iter][j]);
+                    max = srcv[row_iter][j];
+                }
+            }
+
+            iscale[row_iter] = amax ? -127.f/max : 0;
+
+            y[i].d[row_iter] = amax ? 1/iscale[row_iter] : 0;
+        }
+
+        for (int j = 0; j < QK_K / 4; j++) {
+            y[i].bsums[j] = 0;
+        }
+
+        // Quants values are interleaved in sequence of eight bytes from corresponding super blocks
+        // Bsums values are interleaved in sequence of four bsums from each super block taken for interleaving
+        // i.e first four bsums from the first super block, followed by first four bsums from second super block and so on
+        for (int j = 0; j < QK_K * 4; j++) {
+            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
+            int src_id     = (j % (4 * blck_size_interleave)) / blck_size_interleave;
+            src_offset += (j % blck_size_interleave);
+            int index = (((j & 31) >> 3) << 2) + ((j >> 8) << 4) + ((j >> 6) & 3);
+
+            float x0 = srcv[src_id][src_offset] * iscale[src_id];
+            y[i].qs[j] = nearest_int(x0);
+            y[i].bsums[index] += y[i].qs[j];
+        }
+    }
+}
+GGML_CPU_NATIVE_IMPL(ggml_quantize_mat_q8_K_4x8)
+
+} // extern "C"
+
+template <int64_t INTER_SIZE, ggml_type PARAM_TYPE>
+void ggml_quantize_mat_t(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row);
+
+template <> void ggml_quantize_mat_t<4, GGML_TYPE_Q8_0>(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row) {
+    assert(nrow == 4);
+    UNUSED(nrow);
+    ggml_quantize_mat_q8_0_4x4(x, vy, n_per_row);
+}
+
+template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_0>(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row) {
+    assert(nrow == 4);
+    UNUSED(nrow);
+    ggml_quantize_mat_q8_0_4x8(x, vy, n_per_row);
+}
+
+template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_K>(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row) {
+    assert(nrow == 4);
+    UNUSED(nrow);
+    ggml_quantize_mat_q8_K_4x8(x, vy, n_per_row);
+}
+
+extern "C" {
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    float sumf[4];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                    }
+                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+GGML_CPU_NATIVE_IMPL(ggml_gemv_q4_0_4x4_q8_0)
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    float sumf[4];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                    }
+                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+GGML_CPU_NATIVE_IMPL(ggml_gemv_q4_0_4x8_q8_0)
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    {
+        float sumf[8];
+        int sumi;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+
+            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int j = 0; j < ncols_interleaved; j++) {
+                        sumi = 0;
+                        for (int i = 0; i < blocklen; ++i) {
+                            const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                        }
+                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                    }
+                }
+            }
+            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+GGML_CPU_NATIVE_IMPL(ggml_gemv_q4_0_8x8_q8_0)
+
+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) {
+    const int qk = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    float sumf[8];
+    float sum_minf[8];
+    uint32_t utmp[32];
+    int sumi1;
+    int sumi2;
+    int sumi;
+
+    const block_q8_K * a_ptr = (const block_q8_K *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j] = 0.0;
+            sum_minf[j] = 0.0;
+        }
+        for (int l = 0; l < nb; l++) {
+            for (int sb = 0; sb < 8; sb++) {
+                memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
+                utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
+                const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
+                utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
+                utmp[sb * 4 + 2] = uaux_0;
+                utmp[sb * 4 + 0] &= kmask1;
+            }
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                uint8_t *scales_0 = (uint8_t*) utmp + (k / 4) * 32;
+                uint8_t *scales_1 = (uint8_t*) utmp + (k / 4) * 32 + 16;
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi1 = 0;
+                    sumi2 = 0;
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4);
+                        sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 64 + (k % 4) * blocklen + i]);
+                        sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 64 + (k % 4) * blocklen + i + 32]);
+                        sumi1 = sumi1 * scales_0[j];
+                        sumi2 = sumi2 * scales_1[j];
+                        sumi += sumi1 + sumi2;
+                    }
+                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                }
+            }
+            for (int sb = 0; sb < 8; sb++) {
+                uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16;
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
+        }
+    }
+}
+GGML_CPU_NATIVE_IMPL(ggml_gemv_q4_K_8x8_q8_K)
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    {
+        float sumf[4];
+        int sumi;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+
+            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int j = 0; j < ncols_interleaved; j++) {
+                        sumi = 0;
+                        for (int i = 0; i < blocklen; ++i) {
+                            const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                            const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
+                        }
+                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                    }
+                }
+            }
+            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+GGML_CPU_NATIVE_IMPL(ggml_gemv_iq4_nl_4x4_q8_0)
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    {
+        float sumf[4][4];
+        int sumi;
+
+        for (int y = 0; y < nr / 4; y++) {
+            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+                for (int m = 0; m < 4; m++) {
+                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+                }
+                for (int l = 0; l < nb; l++) {
+                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                        for (int m = 0; m < 4; m++) {
+                            for (int j = 0; j < ncols_interleaved; j++) {
+                                sumi = 0;
+                                for (int i = 0; i < blocklen; ++i) {
+                                    const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                    const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                                }
+                                sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            }
+                        }
+                    }
+                }
+                for (int m = 0; m < 4; m++) {
+                    for (int j = 0; j < ncols_interleaved; j++)
+                        s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+                }
+            }
+        }
+    }
+}
+GGML_CPU_NATIVE_IMPL(ggml_gemm_q4_0_4x4_q8_0)
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    float sumf[4][4];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                        (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+}
+GGML_CPU_NATIVE_IMPL(ggml_gemm_q4_0_4x8_q8_0)
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    float sumf[4][8];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+}
+GGML_CPU_NATIVE_IMPL(ggml_gemm_q4_0_8x8_q8_0)
+
+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) {
+    const int qk = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    float sumf[4][8];
+    float sum_minf[4][8];
+    uint32_t utmp[32];
+    int sumi1;
+    int sumi2;
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumf[m][j] = 0.0;
+                    sum_minf[m][j] = 0.0;
+                }
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int sb = 0; sb < 8; sb++) {
+                    memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
+                    utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
+                    utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
+                    utmp[sb * 4 + 2] = uaux_0;
+                    utmp[sb * 4 + 0] &= kmask1;
+                }
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    uint8_t *scales_0 = (uint8_t*) utmp + (k / 4) * 32;
+                    uint8_t *scales_1 = (uint8_t*) utmp + (k / 4) * 32 + 16;
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi1 = 0;
+                            sumi2 = 0;
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4);
+                                sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i]);
+                                sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i + 128]);
+                                sumi1 = sumi1 * scales_0[j];
+                                sumi2 = sumi2 * scales_1[j];
+                                sumi += sumi1 + sumi2;
+                            }
+                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+                for (int sb = 0; sb < 8; sb++) {
+                    uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16;
+                    for(int m = 0; m < 4; m++) {
+                        const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
+                        for(int j = 0; j < ncols_interleaved; j++) {
+                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
+                }
+            }
+        }
+    }
+}
+GGML_CPU_NATIVE_IMPL(ggml_gemm_q4_K_8x8_q8_K)
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    {
+        float sumf[4][4];
+        int sumi;
+
+        for (int y = 0; y < nr / 4; y++) {
+            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+                for (int m = 0; m < 4; m++) {
+                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+                }
+                for (int l = 0; l < nb; l++) {
+                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                        for (int m = 0; m < 4; m++) {
+                            for (int j = 0; j < ncols_interleaved; j++) {
+                                sumi = 0;
+                                for (int i = 0; i < blocklen; ++i) {
+                                    const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                                    const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
+                                }
+                                sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            }
+                        }
+                    }
+                }
+                for (int m = 0; m < 4; m++) {
+                    for (int j = 0; j < ncols_interleaved; j++)
+                        s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+                }
+            }
+        }
+    }
+}
+GGML_CPU_NATIVE_IMPL(ggml_gemm_iq4_nl_4x4_q8_0)
+
+} // extern "C"
+
+static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
+    block_q4_0x4 out;
+
+    for (int i = 0; i < 4; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    const int end = QK4_0 * 2 / blck_size_interleave;
+
+    if (blck_size_interleave == 8) {
+        const uint64_t xor_mask = 0x8888888888888888ULL;
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            uint64_t elems;
+            // Using memcpy to avoid unaligned memory accesses
+            memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
+            elems ^= xor_mask;
+            memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+        }
+    } else if (blck_size_interleave == 4) {
+        const uint32_t xor_mask = 0x88888888;
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            uint32_t elems;
+            memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint32_t));
+            elems ^= xor_mask;
+            memcpy(&out.qs[dst_offset], &elems, sizeof(uint32_t));
+        }
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    return out;
+}
+
+// interleave 8 block_q4_0s in blocks of blck_size_interleave
+// returns an interleaved block_q4_0x8
+// in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks
+// first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave
+static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave) {
+    block_q4_0x8 out;
+
+    for (int i = 0; i < 8; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    const int end = QK4_0 * 4 / blck_size_interleave;
+    const uint64_t xor_mask = 0x8888888888888888ULL;
+
+    for (int i = 0; i < end; ++i) {
+        int src_id = i % 8;
+        int src_offset = (i / 8) * blck_size_interleave;
+        int dst_offset = i * blck_size_interleave;
+
+        uint64_t elems;
+        memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
+        elems ^= xor_mask;
+        memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+    }
+
+    return out;
+}
+
+static block_q4_Kx8 make_block_q4_Kx8(block_q4_K * in, unsigned int blck_size_interleave) {
+    block_q4_Kx8 out;
+    //Delta(scale) and dmin values of the eight Q4_K structures are copied onto the output interleaved structure
+    for (int i = 0; i < 8; i++) {
+        out.d[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d;
+    }
+
+    for (int i = 0; i < 8; i++) {
+        out.dmin[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.dmin;
+    }
+
+    const int end = QK_K * 4 / blck_size_interleave;
+
+    // Interleave Q4_K quants by taking 8 bytes at a time
+    for (int i = 0; i < end; ++i) {
+        int src_id = i % 8;
+        int src_offset = (i / 8) * blck_size_interleave;
+        int dst_offset = i * blck_size_interleave;
+
+        uint64_t elems;
+        memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
+        memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+    }
+
+    // The below logic is designed so as to unpack and rearrange scales and mins values in Q4_K
+    // Currently the Q4_K structure has 8 scales and 8 mins packed in 12 bytes ( 6 bits for each value)
+    // The output Q4_Kx8 structure has 96 bytes
+    // Every 12 byte is packed such that it contains scales and mins for corresponding sub blocks from Q4_K structure
+    // For eg - First 12 bytes contains 8 scales and 8 mins - each of first sub block from different Q4_K structures
+    uint8_t s[8], m[8];
+
+    for (int i = 0; i < 4; i++) {
+        for (int j = 0; j < 8; j++) {
+            s[j] = in[j].scales[i] & 63;
+            m[j] = in[j].scales[i + 4] & 63;
+        }
+
+        out.scales[i * 12]      = (s[0] & 63) + ((s[4] & 48) << 2);
+        out.scales[i * 12 + 1]  = (s[1] & 63) + ((s[5] & 48) << 2);
+        out.scales[i * 12 + 2]  = (s[2] & 63) + ((s[6] & 48) << 2);
+        out.scales[i * 12 + 3]  = (s[3] & 63) + ((s[7] & 48) << 2);
+        out.scales[i * 12 + 4]  = (m[0] & 63) + ((m[4] & 48) << 2);
+        out.scales[i * 12 + 5]  = (m[1] & 63) + ((m[5] & 48) << 2);
+        out.scales[i * 12 + 6]  = (m[2] & 63) + ((m[6] & 48) << 2);
+        out.scales[i * 12 + 7]  = (m[3] & 63) + ((m[7] & 48) << 2);
+        out.scales[i * 12 + 8]  = (s[4] & 15) + ((m[4] & 15) << 4);
+        out.scales[i * 12 + 9]  = (s[5] & 15) + ((m[5] & 15) << 4);
+        out.scales[i * 12 + 10] = (s[6] & 15) + ((m[6] & 15) << 4);
+        out.scales[i * 12 + 11] = (s[7] & 15) + ((m[7] & 15) << 4);
+
+    }
+
+    for (int i = 0; i < 4; i++) {
+        for (int j = 0; j < 8; j++) {
+            s[j] = ((in[j].scales[i] & 192) >> 2) | (in[j].scales[i+8] & 15);
+            m[j] = ((in[j].scales[i + 4] & 192) >> 2) | ((in[j].scales[i+8] & 240) >> 4);
+        }
+
+        out.scales[i * 12 + 48] = (s[0] & 63) + ((s[4] & 48) << 2);
+        out.scales[i * 12 + 49] = (s[1] & 63) + ((s[5] & 48) << 2);
+        out.scales[i * 12 + 50] = (s[2] & 63) + ((s[6] & 48) << 2);
+        out.scales[i * 12 + 51] = (s[3] & 63) + ((s[7] & 48) << 2);
+        out.scales[i * 12 + 52] = (m[0] & 63) + ((m[4] & 48) << 2);
+        out.scales[i * 12 + 53] = (m[1] & 63) + ((m[5] & 48) << 2);
+        out.scales[i * 12 + 54] = (m[2] & 63) + ((m[6] & 48) << 2);
+        out.scales[i * 12 + 55] = (m[3] & 63) + ((m[7] & 48) << 2);
+        out.scales[i * 12 + 56] = (s[4] & 15) + ((m[4] & 15) << 4);
+        out.scales[i * 12 + 57] = (s[5] & 15) + ((m[5] & 15) << 4);
+        out.scales[i * 12 + 58] = (s[6] & 15) + ((m[6] & 15) << 4);
+        out.scales[i * 12 + 59] = (s[7] & 15) + ((m[7] & 15) << 4);
+
+    }
+
+    return out;
+}
+
+static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
+    GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
+    constexpr int nrows_interleaved = 4;
+
+    block_q4_0x4 * dst = (block_q4_0x4 *)t->data;
+    const block_q4_0 * src = (const block_q4_0 *)data;
+    block_q4_0 dst_tmp[4];
+    int nrow = ggml_nrows(t);
+    int nblocks = t->ne[0] / QK4_0;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0));
+
+    if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_q4_0x4(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+static int repack_q4_K_to_q4_K_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_Q4_K);
+    GGML_ASSERT(interleave_block == 8);
+    constexpr int nrows_interleaved = 8;
+
+    block_q4_Kx8 * dst = (block_q4_Kx8*)t->data;
+    const block_q4_K * src = (const block_q4_K*) data;
+    block_q4_K dst_tmp[8];
+    int nrow = ggml_nrows(t);
+    int nblocks = t->ne[0] / QK_K;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_K));
+
+    if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i  = 0; i < nrows_interleaved; i++ ) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_q4_Kx8(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
+static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
+    GGML_ASSERT(interleave_block == 8);
+    constexpr int nrows_interleaved = 8;
+
+    block_q4_0x8 * dst = (block_q4_0x8*)t->data;
+    const block_q4_0 * src = (const block_q4_0*) data;
+    block_q4_0 dst_tmp[8];
+    int nrow = ggml_nrows(t);
+    int nblocks = t->ne[0] / QK4_0;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0));
+
+    if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i  = 0; i < nrows_interleaved; i++ ) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_q4_0x8(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
+static block_iq4_nlx4 make_block_iq4_nlx4(block_iq4_nl * in, unsigned int blck_size_interleave) {
+    block_iq4_nlx4 out;
+
+    for (int i = 0; i < 4; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    const int end = QK4_NL * 2 / blck_size_interleave;
+
+    // TODO: this branch seems wrong
+    //if (blck_size_interleave == 8) {
+    //    for (int i = 0; i < end; ++i) {
+    //        int src_id = i % 4;
+    //        int src_offset = (i / 4) * blck_size_interleave;
+    //        int dst_offset = i * blck_size_interleave;
+
+    //        // Using memcpy to avoid unaligned memory accesses
+    //        memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint64_t));
+    //    }
+    //} else
+    if (blck_size_interleave == 4) {
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint32_t));
+        }
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    return out;
+}
+
+static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL);
+    //GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
+    GGML_ASSERT(interleave_block == 4);
+
+    block_iq4_nlx4 * dst = (block_iq4_nlx4 *)t->data;
+    const block_iq4_nl * src = (const block_iq4_nl *)data;
+    block_iq4_nl dst_tmp[4];
+    int nrow = ggml_nrows(t);
+    int nrows_interleaved = 4;
+    int nblocks = t->ne[0] / QK4_0;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl));
+
+    if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_iq4_nlx4(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
+namespace ggml::cpu::repack {
+// repack
+template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS>
+int repack(struct ggml_tensor *, const void *, size_t);
+
+// TODO: generalise.
+template <> int repack<block_q4_0, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q4_0_to_q4_0_4_bl(t, 4, data, data_size);
+}
+
+template <> int repack<block_q4_0, 8, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q4_0_to_q4_0_4_bl(t, 8, data, data_size);
+}
+
+template <> int repack<block_q4_0, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q4_0_to_q4_0_8_bl(t, 8, data, data_size);
+}
+
+template <> int repack<block_q4_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q4_K_to_q4_K_8_bl(t, 8, data, data_size);
+}
+
+template <> int repack<block_iq4_nl, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_iq4_nl_to_iq4_nl_4_bl(t, 4, data, data_size);
+}
+
+// TODO: needs to be revisited
+//template <> int repack<block_iq4_nl, 8, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
+//    return repack_iq4_nl_to_iq4_nl_4_bl(t, 8, data, data_size);
+//}
+
+// gemv
+template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE>
+void gemv(int, float *, size_t, const void *, const void *, int, int);
+
+template <> void gemv<block_q4_0, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q4_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemv<block_q4_0, 8, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q4_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemv<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemv<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemv<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+// gemm
+template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE>
+void gemm(int, float *, size_t, const void *, const void *, int, int);
+
+template <> void gemm<block_q4_0, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q4_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemm<block_q4_0, 8, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q4_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemm<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemm<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemm<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+class tensor_traits_base : public ggml::cpu::tensor_traits {
+  public:
+    virtual int repack(struct ggml_tensor * t, const void * data, size_t data_size) = 0;
+};
+
+template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE> class tensor_traits : public tensor_traits_base {
+
+    bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
+        // not realy a GGML_TYPE_Q8_0 but same size.
+        switch (op->op) {
+            case GGML_OP_MUL_MAT:
+                size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1]));
+                return true;
+            case GGML_OP_MUL_MAT_ID:
+                size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1]));
+                size = GGML_PAD(size, sizeof(int64_t));  // + padding for next bloc.
+                size += sizeof(int64_t) * (1+op->src[0]->ne[2]) * op->src[1]->ne[2];
+                return true;
+            default:
+                // GGML_ABORT("fatal error");
+                break;
+        }
+        return false;
+    }
+
+    bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) override {
+        switch (op->op) {
+            case GGML_OP_MUL_MAT:
+                forward_mul_mat(params, op);
+                return true;
+            case GGML_OP_MUL_MAT_ID:
+                forward_mul_mat_id(params, op);
+                return true;
+            default:
+                // GGML_ABORT("fatal error");
+                break;
+        }
+        return false;
+    }
+
+    void forward_mul_mat(ggml_compute_params * params, ggml_tensor * op) {
+        const ggml_tensor * src0 = op->src[0];
+        const ggml_tensor * src1 = op->src[1];
+        ggml_tensor *       dst  = op;
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        const int ith = params->ith;
+        const int nth = params->nth;
+
+        GGML_ASSERT(ne0 == ne01);
+        GGML_ASSERT(ne1 == ne11);
+        GGML_ASSERT(ne2 == ne12);
+        GGML_ASSERT(ne3 == ne13);
+
+        // dst cannot be transposed or permuted
+        GGML_ASSERT(nb0 == sizeof(float));
+        GGML_ASSERT(nb0 <= nb1);
+        GGML_ASSERT(nb1 <= nb2);
+        GGML_ASSERT(nb2 <= nb3);
+
+        GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+        GGML_ASSERT(ggml_n_dims(op->src[0]) == 2);
+        // GGML_ASSERT(ggml_n_dims(op->src[1]) == 2);
+
+        char *       wdata = static_cast<char *>(params->wdata);
+        const size_t nbw1  = ggml_row_size(PARAM_TYPE, ne10);
+
+        assert(params->wsize >= nbw1 * ne11);
+
+        const ggml_from_float_t from_float = ggml_get_type_traits_cpu(PARAM_TYPE)->from_float;
+
+        int64_t i11_processed = 0;
+        for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) {
+            ggml_quantize_mat_t<INTER_SIZE, PARAM_TYPE>((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), 4, ne10);
+        }
+
+        i11_processed = ne11 - ne11 % 4;
+        for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) {
+            from_float((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), ne10);
+        }
+
+        ggml_barrier(params->threadpool);
+
+        const void * src1_wdata      = params->wdata;
+        const size_t src1_col_stride = ggml_row_size(PARAM_TYPE, ne10);
+        int64_t      src0_start      = (ith * ne01) / nth;
+        int64_t      src0_end        = ((ith + 1) * ne01) / nth;
+        src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start;
+        src0_end   = (src0_end   % NB_COLS) ? src0_end   + NB_COLS - (src0_end   % NB_COLS) : src0_end;
+        if (src0_start >= src0_end) {
+            return;
+        }
+
+        // If there are more than three rows in src1, use gemm; otherwise, use gemv.
+        if (ne11 > 3) {
+            gemm<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
+                    (float *) ((char *) dst->data) + src0_start, ne01,
+                    (const char *) src0->data + src0_start * nb01,
+                    (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start);
+        }
+        for (int iter = ne11 - ne11 % 4; iter < ne11; iter++) {
+            gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
+                    (float *) ((char *) dst->data + (iter * nb1)) + src0_start, ne01,
+                    (const char *) src0->data + src0_start * nb01,
+                    (const char *) src1_wdata + (src1_col_stride * iter), 1,
+                    src0_end - src0_start);
+        }
+    }
+
+    void forward_mul_mat_id(ggml_compute_params * params, ggml_tensor * op) {
+        const ggml_tensor * src0 = op->src[0];
+        const ggml_tensor * src1 = op->src[1];
+        const ggml_tensor * ids  = op->src[2];
+        ggml_tensor *       dst  = op;
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        const int ith = params->ith;
+        const int nth = params->nth;
+
+        const ggml_from_float_t from_float = ggml_get_type_traits_cpu(PARAM_TYPE)->from_float;
+
+        // we don't support permuted src0 or src1
+        GGML_ASSERT(nb00 == ggml_type_size(src0->type));
+        GGML_ASSERT(nb10 == ggml_type_size(src1->type));
+
+        // dst cannot be transposed or permuted
+        GGML_ASSERT(nb0 == sizeof(float));
+        GGML_ASSERT(nb0 <= nb1);
+        GGML_ASSERT(nb1 <= nb2);
+        GGML_ASSERT(nb2 <= nb3);
+
+        GGML_ASSERT(ne03 == 1);
+        GGML_ASSERT(ne13 == 1);
+        GGML_ASSERT(ne3  == 1);
+
+        GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+        // row groups
+        const int n_ids = ids->ne[0]; // n_expert_used
+        const int n_as  = ne02;       // n_expert
+
+        const size_t nbw1 = ggml_row_size(PARAM_TYPE, ne10);
+        const size_t nbw2 = nbw1*ne11;
+        const size_t nbw3 = nbw2*ne12;
+
+        struct mmid_row_mapping {
+            int32_t i1;
+            int32_t i2;
+        };
+
+        GGML_ASSERT(params->wsize >= (GGML_PAD(nbw3, sizeof(int64_t)) + n_as * sizeof(int64_t) +
+                                      n_as * ne12 * sizeof(mmid_row_mapping)));
+
+        auto * wdata             = (char *)     params->wdata;
+        auto * wdata_src1_end    = (char *)     wdata + GGML_PAD(nbw3, sizeof(int64_t));
+        auto * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as]
+
+        struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as);  // [n_as][ne12]
+
+        // src1: float32 => param type
+        for (int64_t i12 = 0; i12 < ne12; ++i12) {
+            for (int64_t i11 = ith; i11 < ne11; i11 += nth) {
+                from_float((float *)((char *) src1->data + i12 * nb12 + i11 * nb11),
+                           (void *)               (wdata + i12 * nbw2 + i11 * nbw1),
+                           ne10);
+            }
+        }
+
+#define MMID_MATRIX_ROW(row_id, i1) matrix_rows[(row_id) * ne12 + (i1)]
+
+        if (ith == 0) {
+            // initialize matrix_row_counts
+            memset(matrix_row_counts, 0, n_as * sizeof(int64_t));
+
+            // group rows by src0 matrix
+            for (int32_t iid1 = 0; iid1 < ids->ne[1]; ++iid1) {
+                for (int32_t id = 0; id < n_ids; ++id) {
+                    const int32_t i02 =
+                        *(const int32_t *) ((const char *) ids->data + iid1 * ids->nb[1] + id * ids->nb[0]);
+
+                    GGML_ASSERT(i02 >= 0 && i02 < n_as);
+
+                    MMID_MATRIX_ROW(i02, matrix_row_counts[i02]) = { id, iid1 };
+                    matrix_row_counts[i02] += 1;
+                }
+            }
+        }
+
+        ggml_barrier(params->threadpool);
+
+        // compute each matrix multiplication in sequence
+        for (int cur_a = 0; cur_a < n_as; ++cur_a) {
+            const int64_t cne1 = matrix_row_counts[cur_a];
+
+            if (cne1 == 0) {
+                continue;
+            }
+
+            const auto * src0_cur = (const char *) src0->data + cur_a*nb02;
+
+            //const int64_t nr0 = ne01; // src0 rows
+            const int64_t nr1 = cne1; // src1 rows
+
+            int64_t src0_cur_start = (ith * ne01) / nth;
+            int64_t src0_cur_end   = ((ith + 1) * ne01) / nth;
+
+            src0_cur_start = (src0_cur_start % NB_COLS) ? src0_cur_start + NB_COLS - (src0_cur_start % NB_COLS) : src0_cur_start;
+            src0_cur_end   = (src0_cur_end   % NB_COLS) ? src0_cur_end   + NB_COLS - (src0_cur_end   % NB_COLS) : src0_cur_end;
+
+            if (src0_cur_start >= src0_cur_end) {
+                return;
+            }
+
+            for (int ir1 = 0; ir1 < nr1; ir1++) {
+                struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, ir1);
+
+                const int id = row_mapping.i1; // selected expert index
+
+                const int64_t i11 = id % ne11;
+                const int64_t i12 = row_mapping.i2; // row index in src1
+
+                const int64_t i1 = id;  // selected expert index
+                const int64_t i2 = i12; // row
+
+                const auto * src1_col = (const char *) wdata + (i11 * nbw1 + i12 * nbw2);
+
+                gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
+                        (float *)((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start, ne01,
+                        src0_cur + src0_cur_start * nb01,
+                        src1_col, 1, src0_cur_end - src0_cur_start);
+            }
+        }
+#undef MMID_MATRIX_ROW
+    }
+
+    int repack(struct ggml_tensor * t, const void * data, size_t data_size) override {
+        GGML_LOG_DEBUG("%s: repack tensor %s with %s_%dx%d\n", __func__, t->name, ggml_type_name(t->type),
+                       (int) NB_COLS, (int) INTER_SIZE);
+        return ggml::cpu::repack::repack<BLOC_TYPE, INTER_SIZE, NB_COLS>(t, data, data_size);
+    }
+};
+
+// instance for Q4
+static const tensor_traits<block_q4_0, 4, 4, GGML_TYPE_Q8_0> q4_0_4x4_q8_0;
+static const tensor_traits<block_q4_0, 8, 4, GGML_TYPE_Q8_0> q4_0_4x8_q8_0;
+static const tensor_traits<block_q4_0, 8, 8, GGML_TYPE_Q8_0> q4_0_8x8_q8_0;
+static const tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;
+
+// instance for IQ4
+static const tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0> iq4_nl_4x4_q8_0;
+
+}  // namespace ggml::cpu::repack
+
+static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(const struct ggml_tensor * cur) {
+    if (cur->type == GGML_TYPE_Q4_0) {
+        if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) {
+            if (cur->ne[1] % 8 == 0) {
+                return &ggml::cpu::repack::q4_0_8x8_q8_0;
+            }
+        }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+            if (cur->ne[1] % 4 == 0) {
+                return &ggml::cpu::repack::q4_0_4x8_q8_0;
+            }
+        }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+            if (cur->ne[1] % 4 == 0) {
+                return &ggml::cpu::repack::q4_0_4x4_q8_0;
+            }
+        }
+    } else if (cur->type == GGML_TYPE_Q4_K) {
+        if (ggml_cpu_has_avx2()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &ggml::cpu::repack::q4_K_8x8_q8_K;
+            }
+        }
+    } else if (cur->type == GGML_TYPE_IQ4_NL) {
+        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+            if (cur->ne[1] % 4 == 0) {
+                return &ggml::cpu::repack::iq4_nl_4x4_q8_0;
+            }
+        }
+    }
+
+    return nullptr;
+}
+
+static enum ggml_status ggml_backend_cpu_repack_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    tensor->extra = (void *) const_cast<ggml::cpu::tensor_traits *>(ggml_repack_get_optimal_repack_type(tensor));
+
+    GGML_UNUSED(buffer);
+    return GGML_STATUS_SUCCESS;
+}
+
+static void ggml_backend_cpu_repack_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor,
+                                                       const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(offset == 0);
+    GGML_ASSERT(size == ggml_nbytes(tensor));
+
+    auto tensor_traits = (ggml::cpu::repack::tensor_traits_base *) tensor->extra;
+    auto OK            = tensor_traits->repack(tensor, data, size);
+
+    GGML_ASSERT(OK == 0);
+    GGML_UNUSED(buffer);
+}
+
+static const char * ggml_backend_cpu_repack_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "CPU_REPACK";
+
+    GGML_UNUSED(buft);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_repack_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
+
+    if (buffer == nullptr) {
+        return nullptr;
+    }
+
+    buffer->buft              = buft;
+    buffer->iface.init_tensor = ggml_backend_cpu_repack_buffer_init_tensor;
+    buffer->iface.set_tensor  = ggml_backend_cpu_repack_buffer_set_tensor;
+    buffer->iface.get_tensor  = nullptr;
+    buffer->iface.cpy_tensor  = nullptr;
+    return buffer;
+}
+
+static size_t ggml_backend_cpu_repack_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return TENSOR_ALIGNMENT;
+
+    GGML_UNUSED(buft);
+}
+
+namespace ggml::cpu::repack {
+class extra_buffer_type : ggml::cpu::extra_buffer_type {
+    bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
+        if (    op->op == GGML_OP_MUL_MAT &&
+                op->src[0]->buffer &&
+                (ggml_n_dims(op->src[0]) == 2) &&
+                op->src[0]->buffer->buft == ggml_backend_cpu_repack_buffer_type() &&
+                ggml_repack_get_optimal_repack_type(op->src[0])
+                ) {
+            if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
+                return false;
+            }
+            if (op->src[1]->type == GGML_TYPE_F32) {
+                return true;
+            }
+            //if (op->src[1]->type == GGML_TYPE_Q8_0) {
+            //    return true;
+            //}
+            // may be possible if Q8_0 packed...
+        } else if (op->op == GGML_OP_MUL_MAT_ID
+                && op->src[0]->buffer
+                && (ggml_n_dims(op->src[0]) == 3)
+                && op->src[0]->buffer->buft == ggml_backend_cpu_repack_buffer_type()
+                && ggml_repack_get_optimal_repack_type(op->src[0])
+                ) {
+            if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
+                return false;
+            }
+            if (op->src[1]->type == GGML_TYPE_F32) {
+                return true;
+            }
+            //if (op->src[1]->type == GGML_TYPE_Q8_0) {
+            //    return true;
+            //}
+        }
+        return false;
+    }
+
+    ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {
+        if (op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_MUL_MAT_ID) {
+            if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_repack_buffer_type()) {
+                return (ggml::cpu::tensor_traits *) op->src[0]->extra;
+            }
+        }
+        return nullptr;
+    }
+};
+}  // namespace ggml::cpu::repack
+
+ggml_backend_buffer_type_t ggml_backend_cpu_repack_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_repack = {
+        /* .iface    = */ {
+                           /* .get_name         = */ ggml_backend_cpu_repack_buffer_type_get_name,
+                           /* .alloc_buffer     = */ ggml_backend_cpu_repack_buffer_type_alloc_buffer,
+                           /* .get_alignment    = */ ggml_backend_cpu_repack_buffer_type_get_alignment,
+                           /* .get_max_size     = */ nullptr,  // defaults to SIZE_MAX
+                           /* .get_alloc_size   = */ nullptr,  // defaults to ggml_nbytes
+                           /* .is_host          = */ nullptr,
+                           },
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .context = */ new ggml::cpu::repack::extra_buffer_type(),
+    };
+
+    return &ggml_backend_cpu_buffer_type_repack;
+}

ggml/src/ggml-cpu/repack.h

@@ -0,0 +1,119 @@
+#pragma once
+
+#define GGML_COMMON_DECL_CPP
+#include "ggml-common.h"
+
+#include "traits.h"
+#include "ggml.h"
+
+// GGML internal header
+
+ggml_backend_buffer_type_t ggml_backend_cpu_repack_buffer_type(void);
+
+template <int K> constexpr int QK_0() {
+    if constexpr (K == 4) {
+        return QK4_0;
+    }
+    if constexpr (K == 8) {
+        return QK8_0;
+    }
+    return -1;
+}
+
+template <int K, int N> struct block {
+    ggml_half d[N];                         // deltas for N qK_0 blocks
+    int8_t    qs[(QK_0<K>() * N * K) / 8];  // quants for N qK_0 blocks
+};
+
+// control size
+static_assert(sizeof(block<4, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 2, "wrong block<4,4> size/padding");
+static_assert(sizeof(block<4, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<4,8> size/padding");
+static_assert(sizeof(block<8, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<8,4> size/padding");
+static_assert(sizeof(block<8, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<8,8> size/padding");
+
+using block_q4_0x4 = block<4, 4>;
+using block_q4_0x8 = block<4, 8>;
+using block_q8_0x4 = block<8, 4>;
+using block_q8_0x8 = block<8, 8>;
+
+struct block_q4_Kx8 {
+    ggml_half d[8];      // super-block scale for quantized scales
+    ggml_half dmin[8];   // super-block scale for quantized mins
+    uint8_t scales[96];  // scales and mins, quantized with 6 bits
+    uint8_t qs[1024];    // 4--bit quants
+};
+
+static_assert(sizeof(block_q4_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 4, "wrong q4_K block size/padding");
+
+struct block_q8_Kx4 {
+    float d[4];              // delta
+    int8_t qs[QK_K * 4];     // quants
+    int16_t bsums[QK_K / 4]; // sum of quants in groups of 16
+};
+
+static_assert(sizeof(block_q8_Kx4) == sizeof(float) * 4 + QK_K * 4 + (QK_K / 4) * sizeof(int16_t), "wrong q8_K block size/padding");
+
+struct block_iq4_nlx4 {
+    ggml_half d[4];            // deltas for 4 iq4_nl blocks
+    uint8_t   qs[QK4_NL * 2];  // nibbles / quants for 4 iq4_nl blocks
+};
+
+static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding");
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+// Workaround for clang:
+// clang++ complains: ``error: call to 'ggml_gemm_q4_0_4x4_q8_0' is ambiguous''
+// repro: https://godbolt.org/z/oKdeWKonM (ICE), https://godbolt.org/z/1szq6P36v (ambiguous call)
+#if defined(GGML_CPU_CLANG_WORKAROUND) || !(defined(__GNUC__) && defined(__clang__)) || defined(__HIPCC__)
+void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+#endif // !defined(__clang__)
+
+// Native implementations
+void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+#if defined(GGML_CPU_GENERIC)
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#endif
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif

ggml/src/ggml-cpu/traits.cpp

@@ -0,0 +1,36 @@
+#include "traits.h"
+
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+
+namespace ggml::cpu {
+tensor_traits::~tensor_traits() {}
+
+extra_buffer_type::~extra_buffer_type() {}
+}  // namespace ggml::cpu
+
+bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) {
+    for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
+        if (extra && extra->context) {
+            auto buf_extra     = (ggml::cpu::extra_buffer_type *) extra->context;
+            auto tensor_traits = buf_extra->get_tensor_traits(op);
+            if (tensor_traits && tensor_traits->compute_forward(params, op)) {
+                return true;
+            }
+        }
+    }
+    return false;
+}
+
+bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size) {
+    for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
+        if (extra && extra->context) {
+            auto buf_extra     = (ggml::cpu::extra_buffer_type *) extra->context;
+            auto tensor_traits = buf_extra->get_tensor_traits(op);
+            if (tensor_traits && tensor_traits->work_size(n_threads, op, *size)) {
+                return true;
+            }
+        }
+    }
+    return false;
+}

ggml/src/ggml-cpu/traits.h

@@ -0,0 +1,38 @@
+#pragma once
+#include "ggml-backend-impl.h"
+#include "ggml-cpu-impl.h"
+#include "ggml.h"
+
+#ifdef __cplusplus
+#    include <vector>
+extern "C" {
+#endif
+
+// return true if op part of extra "accelerator"
+bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op);
+bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size);
+
+#ifdef __cplusplus
+}
+
+namespace ggml::cpu {
+// register in tensor->extra
+class tensor_traits {
+  public:
+    virtual ~tensor_traits();
+    virtual bool work_size(int n_threads, const struct ggml_tensor * op, size_t & size)        = 0;
+    virtual bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) = 0;
+};
+
+class extra_buffer_type {
+  public:
+    virtual ~extra_buffer_type();
+    virtual bool            supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) = 0;
+    virtual tensor_traits * get_tensor_traits(const struct ggml_tensor * op)                   = 0;
+};
+}  // namespace ggml::cpu
+
+// implemented in ggml-cpu.cpp.
+std::vector<ggml_backend_buffer_type_t> & ggml_backend_cpu_get_extra_buffers_type();
+
+#endif

ggml/src/ggml-cuda/common.cuh

@@ -466,9 +466,6 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 }
 
-// TODO: move to ggml-common.h
-static constexpr __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
-
 typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
 
 static __device__ __forceinline__ float get_alibi_slope(

ggml/src/ggml-quants.c

@@ -2425,8 +2425,6 @@ void dequantize_row_iq1_m(const block_iq1_m * GGML_RESTRICT x, float * GGML_REST
     }
 }
 
-static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
-
 void dequantize_row_iq4_nl(const block_iq4_nl * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK4_NL == 0);
     const int64_t nb = k / QK4_NL;

ggml/src/ggml-sycl/common.hpp

@@ -149,8 +149,6 @@ typedef sycl::float2 dfloat2;
 
 #define MMVQ_MAX_BATCH_SIZE  8
 
-static const int8_t kvalues_iq4nl[16]={-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
-
 static int g_all_sycl_device_count = -1;
 static bool g_ggml_backend_sycl_buffer_type_initialized = false;